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 337484 : planner(Query *parse, const char *query_string, int cursorOptions,
316 : ParamListInfo boundParams, ExplainState *es)
317 : {
318 : PlannedStmt *result;
319 :
320 337484 : if (planner_hook)
321 50292 : result = (*planner_hook) (parse, query_string, cursorOptions,
322 : boundParams, es);
323 : else
324 287192 : result = standard_planner(parse, query_string, cursorOptions,
325 : boundParams, es);
326 :
327 333493 : pgstat_report_plan_id(result->planId, false);
328 :
329 333493 : return result;
330 : }
331 :
332 : PlannedStmt *
333 337484 : 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 : *lc;
346 :
347 : /*
348 : * Set up global state for this planner invocation. This data is needed
349 : * across all levels of sub-Query that might exist in the given command,
350 : * so we keep it in a separate struct that's linked to by each per-Query
351 : * PlannerInfo.
352 : */
353 337484 : glob = makeNode(PlannerGlobal);
354 :
355 337484 : glob->boundParams = boundParams;
356 337484 : glob->subplans = NIL;
357 337484 : glob->subpaths = NIL;
358 337484 : glob->subroots = NIL;
359 337484 : glob->rewindPlanIDs = NULL;
360 337484 : glob->finalrtable = NIL;
361 337484 : glob->allRelids = NULL;
362 337484 : glob->prunableRelids = NULL;
363 337484 : glob->finalrteperminfos = NIL;
364 337484 : glob->finalrowmarks = NIL;
365 337484 : glob->resultRelations = NIL;
366 337484 : glob->appendRelations = NIL;
367 337484 : glob->partPruneInfos = NIL;
368 337484 : glob->relationOids = NIL;
369 337484 : glob->invalItems = NIL;
370 337484 : glob->paramExecTypes = NIL;
371 337484 : glob->lastPHId = 0;
372 337484 : glob->lastRowMarkId = 0;
373 337484 : glob->lastPlanNodeId = 0;
374 337484 : glob->transientPlan = false;
375 337484 : glob->dependsOnRole = false;
376 337484 : glob->partition_directory = NULL;
377 337484 : glob->rel_notnullatts_hash = NULL;
378 :
379 : /*
380 : * Assess whether it's feasible to use parallel mode for this query. We
381 : * can't do this in a standalone backend, or if the command will try to
382 : * modify any data, or if this is a cursor operation, or if GUCs are set
383 : * to values that don't permit parallelism, or if parallel-unsafe
384 : * functions are present in the query tree.
385 : *
386 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
387 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
388 : * the command is writing into a completely new table which workers won't
389 : * be able to see. If the workers could see the table, the fact that
390 : * group locking would cause them to ignore the leader's heavyweight GIN
391 : * page locks would make this unsafe. We'll have to fix that somehow if
392 : * we want to allow parallel inserts in general; updates and deletes have
393 : * additional problems especially around combo CIDs.)
394 : *
395 : * For now, we don't try to use parallel mode if we're running inside a
396 : * parallel worker. We might eventually be able to relax this
397 : * restriction, but for now it seems best not to have parallel workers
398 : * trying to create their own parallel workers.
399 : */
400 337484 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
401 319414 : IsUnderPostmaster &&
402 319414 : parse->commandType == CMD_SELECT &&
403 258202 : !parse->hasModifyingCTE &&
404 258079 : max_parallel_workers_per_gather > 0 &&
405 257374 : !IsParallelWorker())
406 : {
407 : /* all the cheap tests pass, so scan the query tree */
408 257334 : glob->maxParallelHazard = max_parallel_hazard(parse);
409 257334 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
410 : }
411 : else
412 : {
413 : /* skip the query tree scan, just assume it's unsafe */
414 80150 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
415 80150 : glob->parallelModeOK = false;
416 : }
417 :
418 : /*
419 : * glob->parallelModeNeeded is normally set to false here and changed to
420 : * true during plan creation if a Gather or Gather Merge plan is actually
421 : * created (cf. create_gather_plan, create_gather_merge_plan).
422 : *
423 : * However, if debug_parallel_query = on or debug_parallel_query =
424 : * regress, then we impose parallel mode whenever it's safe to do so, even
425 : * if the final plan doesn't use parallelism. It's not safe to do so if
426 : * the query contains anything parallel-unsafe; parallelModeOK will be
427 : * false in that case. Note that parallelModeOK can't change after this
428 : * point. Otherwise, everything in the query is either parallel-safe or
429 : * parallel-restricted, and in either case it should be OK to impose
430 : * parallel-mode restrictions. If that ends up breaking something, then
431 : * either some function the user included in the query is incorrectly
432 : * labeled as parallel-safe or parallel-restricted when in reality it's
433 : * parallel-unsafe, or else the query planner itself has a bug.
434 : */
435 559869 : glob->parallelModeNeeded = glob->parallelModeOK &&
436 222385 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
437 :
438 : /* Determine what fraction of the plan is likely to be scanned */
439 337484 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
440 : {
441 : /*
442 : * We have no real idea how many tuples the user will ultimately FETCH
443 : * from a cursor, but it is often the case that he doesn't want 'em
444 : * all, or would prefer a fast-start plan anyway so that he can
445 : * process some of the tuples sooner. Use a GUC parameter to decide
446 : * what fraction to optimize for.
447 : */
448 3222 : tuple_fraction = cursor_tuple_fraction;
449 :
450 : /*
451 : * We document cursor_tuple_fraction as simply being a fraction, which
452 : * means the edge cases 0 and 1 have to be treated specially here. We
453 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
454 : */
455 3222 : if (tuple_fraction >= 1.0)
456 0 : tuple_fraction = 0.0;
457 3222 : else if (tuple_fraction <= 0.0)
458 0 : tuple_fraction = 1e-10;
459 : }
460 : else
461 : {
462 : /* Default assumption is we need all the tuples */
463 334262 : tuple_fraction = 0.0;
464 : }
465 :
466 : /*
467 : * Compute the initial path generation strategy mask.
468 : *
469 : * Some strategies, such as PGS_FOREIGNJOIN, have no corresponding enable_*
470 : * GUC, and so the corresponding bits are always set in the default
471 : * strategy mask.
472 : *
473 : * It may seem surprising that enable_indexscan sets both PGS_INDEXSCAN
474 : * and PGS_INDEXONLYSCAN. However, the historical behavior of this GUC
475 : * corresponds to this exactly: enable_indexscan=off disables both
476 : * index-scan and index-only scan paths, whereas enable_indexonlyscan=off
477 : * converts the index-only scan paths that we would have considered into
478 : * index scan paths.
479 : */
480 337484 : glob->default_pgs_mask = PGS_APPEND | PGS_MERGE_APPEND | PGS_FOREIGNJOIN |
481 : PGS_GATHER | PGS_CONSIDER_NONPARTIAL;
482 337484 : if (enable_tidscan)
483 337484 : glob->default_pgs_mask |= PGS_TIDSCAN;
484 337484 : if (enable_seqscan)
485 320550 : glob->default_pgs_mask |= PGS_SEQSCAN;
486 337484 : if (enable_indexscan)
487 335237 : glob->default_pgs_mask |= PGS_INDEXSCAN | PGS_INDEXONLYSCAN;
488 337484 : if (enable_indexonlyscan)
489 336076 : glob->default_pgs_mask |= PGS_CONSIDER_INDEXONLY;
490 337484 : if (enable_bitmapscan)
491 328343 : glob->default_pgs_mask |= PGS_BITMAPSCAN;
492 337484 : if (enable_mergejoin)
493 : {
494 335577 : glob->default_pgs_mask |= PGS_MERGEJOIN_PLAIN;
495 335577 : if (enable_material)
496 335510 : glob->default_pgs_mask |= PGS_MERGEJOIN_MATERIALIZE;
497 : }
498 337484 : if (enable_nestloop)
499 : {
500 337180 : glob->default_pgs_mask |= PGS_NESTLOOP_PLAIN;
501 337180 : if (enable_material)
502 336898 : glob->default_pgs_mask |= PGS_NESTLOOP_MATERIALIZE;
503 337180 : if (enable_memoize)
504 337061 : glob->default_pgs_mask |= PGS_NESTLOOP_MEMOIZE;
505 : }
506 337484 : if (enable_hashjoin)
507 335242 : glob->default_pgs_mask |= PGS_HASHJOIN;
508 337484 : if (enable_gathermerge)
509 337484 : glob->default_pgs_mask |= PGS_GATHER_MERGE;
510 337484 : if (enable_partitionwise_join)
511 2065 : glob->default_pgs_mask |= PGS_CONSIDER_PARTITIONWISE;
512 :
513 : /* Allow plugins to take control after we've initialized "glob" */
514 337484 : if (planner_setup_hook)
515 88993 : (*planner_setup_hook) (glob, parse, query_string, cursorOptions,
516 : &tuple_fraction, es);
517 :
518 : /* primary planning entry point (may recurse for subqueries) */
519 336716 : root = subquery_planner(glob, parse, NULL, NULL, NULL, false,
520 : tuple_fraction, NULL);
521 :
522 : /* Select best Path and turn it into a Plan */
523 333763 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
524 333763 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
525 :
526 333763 : top_plan = create_plan(root, best_path);
527 :
528 : /*
529 : * If creating a plan for a scrollable cursor, make sure it can run
530 : * backwards on demand. Add a Material node at the top at need.
531 : */
532 333493 : if (cursorOptions & CURSOR_OPT_SCROLL)
533 : {
534 221 : if (!ExecSupportsBackwardScan(top_plan))
535 26 : top_plan = materialize_finished_plan(top_plan);
536 : }
537 :
538 : /*
539 : * Optionally add a Gather node for testing purposes, provided this is
540 : * actually a safe thing to do.
541 : *
542 : * We can add Gather even when top_plan has parallel-safe initPlans, but
543 : * then we have to move the initPlans to the Gather node because of
544 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
545 : * regression tests when debug_parallel_query = regress, because initPlans
546 : * that would normally appear on the top_plan move to the Gather, causing
547 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
548 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
549 : */
550 333493 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
551 157 : top_plan->parallel_safe &&
552 104 : (top_plan->initPlan == NIL ||
553 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
554 : {
555 104 : Gather *gather = makeNode(Gather);
556 : Cost initplan_cost;
557 : bool unsafe_initplans;
558 :
559 104 : gather->plan.targetlist = top_plan->targetlist;
560 104 : gather->plan.qual = NIL;
561 104 : gather->plan.lefttree = top_plan;
562 104 : gather->plan.righttree = NULL;
563 104 : gather->num_workers = 1;
564 104 : gather->single_copy = true;
565 104 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
566 :
567 : /* Transfer any initPlans to the new top node */
568 104 : gather->plan.initPlan = top_plan->initPlan;
569 104 : top_plan->initPlan = NIL;
570 :
571 : /*
572 : * Since this Gather has no parallel-aware descendants to signal to,
573 : * we don't need a rescan Param.
574 : */
575 104 : gather->rescan_param = -1;
576 :
577 : /*
578 : * Ideally we'd use cost_gather here, but setting up dummy path data
579 : * to satisfy it doesn't seem much cleaner than knowing what it does.
580 : */
581 104 : gather->plan.startup_cost = top_plan->startup_cost +
582 : parallel_setup_cost;
583 104 : gather->plan.total_cost = top_plan->total_cost +
584 104 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
585 104 : gather->plan.plan_rows = top_plan->plan_rows;
586 104 : gather->plan.plan_width = top_plan->plan_width;
587 104 : gather->plan.parallel_aware = false;
588 104 : gather->plan.parallel_safe = false;
589 :
590 : /*
591 : * Delete the initplans' cost from top_plan. We needn't add it to the
592 : * Gather node, since the above coding already included it there.
593 : */
594 104 : SS_compute_initplan_cost(gather->plan.initPlan,
595 : &initplan_cost, &unsafe_initplans);
596 104 : top_plan->startup_cost -= initplan_cost;
597 104 : top_plan->total_cost -= initplan_cost;
598 :
599 : /* use parallel mode for parallel plans. */
600 104 : root->glob->parallelModeNeeded = true;
601 :
602 104 : top_plan = &gather->plan;
603 : }
604 :
605 : /*
606 : * If any Params were generated, run through the plan tree and compute
607 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
608 : * set_plan_references' tree traversal, but for now it has to be separate
609 : * because we need to visit subplans before not after main plan.
610 : */
611 333493 : if (glob->paramExecTypes != NIL)
612 : {
613 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
614 148072 : forboth(lp, glob->subplans, lr, glob->subroots)
615 : {
616 33906 : Plan *subplan = (Plan *) lfirst(lp);
617 33906 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
618 :
619 33906 : SS_finalize_plan(subroot, subplan);
620 : }
621 114166 : SS_finalize_plan(root, top_plan);
622 : }
623 :
624 : /* final cleanup of the plan */
625 : Assert(glob->finalrtable == NIL);
626 : Assert(glob->finalrteperminfos == NIL);
627 : Assert(glob->finalrowmarks == NIL);
628 : Assert(glob->resultRelations == NIL);
629 : Assert(glob->appendRelations == NIL);
630 333493 : top_plan = set_plan_references(root, top_plan);
631 : /* ... and the subplans (both regular subplans and initplans) */
632 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
633 367399 : forboth(lp, glob->subplans, lr, glob->subroots)
634 : {
635 33906 : Plan *subplan = (Plan *) lfirst(lp);
636 33906 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
637 :
638 33906 : lfirst(lp) = set_plan_references(subroot, subplan);
639 : }
640 :
641 : /* build the PlannedStmt result */
642 333493 : result = makeNode(PlannedStmt);
643 :
644 333493 : result->commandType = parse->commandType;
645 333493 : result->queryId = parse->queryId;
646 333493 : result->planOrigin = PLAN_STMT_STANDARD;
647 333493 : result->hasReturning = (parse->returningList != NIL);
648 333493 : result->hasModifyingCTE = parse->hasModifyingCTE;
649 333493 : result->canSetTag = parse->canSetTag;
650 333493 : result->transientPlan = glob->transientPlan;
651 333493 : result->dependsOnRole = glob->dependsOnRole;
652 333493 : result->parallelModeNeeded = glob->parallelModeNeeded;
653 333493 : result->planTree = top_plan;
654 333493 : result->partPruneInfos = glob->partPruneInfos;
655 333493 : result->rtable = glob->finalrtable;
656 666986 : result->unprunableRelids = bms_difference(glob->allRelids,
657 333493 : glob->prunableRelids);
658 333493 : result->permInfos = glob->finalrteperminfos;
659 333493 : result->subrtinfos = glob->subrtinfos;
660 333493 : result->appendRelations = glob->appendRelations;
661 333493 : result->subplans = glob->subplans;
662 333493 : result->rewindPlanIDs = glob->rewindPlanIDs;
663 333493 : result->rowMarks = glob->finalrowmarks;
664 :
665 : /*
666 : * Compute resultRelationRelids and rowMarkRelids from resultRelations and
667 : * rowMarks. These can be used for cheap membership checks.
668 : */
669 402420 : foreach(lc, glob->resultRelations)
670 68927 : result->resultRelationRelids = bms_add_member(result->resultRelationRelids,
671 : lfirst_int(lc));
672 344274 : foreach(lc, glob->finalrowmarks)
673 10781 : result->rowMarkRelids = bms_add_member(result->rowMarkRelids,
674 10781 : ((PlanRowMark *) lfirst(lc))->rti);
675 :
676 333493 : result->relationOids = glob->relationOids;
677 333493 : result->invalItems = glob->invalItems;
678 333493 : result->paramExecTypes = glob->paramExecTypes;
679 : /* utilityStmt should be null, but we might as well copy it */
680 333493 : result->utilityStmt = parse->utilityStmt;
681 333493 : result->elidedNodes = glob->elidedNodes;
682 333493 : result->stmt_location = parse->stmt_location;
683 333493 : result->stmt_len = parse->stmt_len;
684 :
685 333493 : result->jitFlags = PGJIT_NONE;
686 333493 : if (jit_enabled && jit_above_cost >= 0 &&
687 0 : top_plan->total_cost > jit_above_cost)
688 : {
689 0 : result->jitFlags |= PGJIT_PERFORM;
690 :
691 : /*
692 : * Decide how much effort should be put into generating better code.
693 : */
694 0 : if (jit_optimize_above_cost >= 0 &&
695 0 : top_plan->total_cost > jit_optimize_above_cost)
696 0 : result->jitFlags |= PGJIT_OPT3;
697 0 : if (jit_inline_above_cost >= 0 &&
698 0 : top_plan->total_cost > jit_inline_above_cost)
699 0 : result->jitFlags |= PGJIT_INLINE;
700 :
701 : /*
702 : * Decide which operations should be JITed.
703 : */
704 0 : if (jit_expressions)
705 0 : result->jitFlags |= PGJIT_EXPR;
706 0 : if (jit_tuple_deforming)
707 0 : result->jitFlags |= PGJIT_DEFORM;
708 : }
709 :
710 : /* Allow plugins to take control before we discard "glob" */
711 333493 : if (planner_shutdown_hook)
712 87454 : (*planner_shutdown_hook) (glob, parse, query_string, result);
713 :
714 333493 : if (glob->partition_directory != NULL)
715 9172 : DestroyPartitionDirectory(glob->partition_directory);
716 :
717 333493 : return result;
718 : }
719 :
720 :
721 : /*--------------------
722 : * subquery_planner
723 : * Invokes the planner on a subquery. We recurse to here for each
724 : * sub-SELECT found in the query tree.
725 : *
726 : * glob is the global state for the current planner run.
727 : * parse is the querytree produced by the parser & rewriter.
728 : * plan_name is the name to assign to this subplan (NULL at the top level).
729 : * parent_root is the immediate parent Query's info (NULL at the top level).
730 : * alternative_root is a previously created PlannerInfo for which this query
731 : * level is an alternative implementation, or else NULL.
732 : * hasRecursion is true if this is a recursive WITH query.
733 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
734 : * tuple_fraction is interpreted as explained for grouping_planner, below.
735 : * setops is used for set operation subqueries to provide the subquery with
736 : * the context in which it's being used so that Paths correctly sorted for the
737 : * set operation can be generated. NULL when not planning a set operation
738 : * child, or when a child of a set op that isn't interested in sorted input.
739 : *
740 : * Basically, this routine does the stuff that should only be done once
741 : * per Query object. It then calls grouping_planner. At one time,
742 : * grouping_planner could be invoked recursively on the same Query object;
743 : * that's not currently true, but we keep the separation between the two
744 : * routines anyway, in case we need it again someday.
745 : *
746 : * subquery_planner will be called recursively to handle sub-Query nodes
747 : * found within the query's expressions and rangetable.
748 : *
749 : * Returns the PlannerInfo struct ("root") that contains all data generated
750 : * while planning the subquery. In particular, the Path(s) attached to
751 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
752 : * cheapest way(s) to implement the query. The top level will select the
753 : * best Path and pass it through createplan.c to produce a finished Plan.
754 : *--------------------
755 : */
756 : PlannerInfo *
757 400998 : subquery_planner(PlannerGlobal *glob, Query *parse, char *plan_name,
758 : PlannerInfo *parent_root, PlannerInfo *alternative_root,
759 : bool hasRecursion, double tuple_fraction,
760 : SetOperationStmt *setops)
761 : {
762 : PlannerInfo *root;
763 : List *newWithCheckOptions;
764 : List *newHaving;
765 : bool hasOuterJoins;
766 : bool hasResultRTEs;
767 : RelOptInfo *final_rel;
768 : ListCell *l;
769 :
770 : /* Create a PlannerInfo data structure for this subquery */
771 400998 : root = makeNode(PlannerInfo);
772 400998 : root->parse = parse;
773 400998 : root->glob = glob;
774 400998 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
775 400998 : root->plan_name = plan_name;
776 400998 : if (alternative_root != NULL)
777 1331 : root->alternative_plan_name = alternative_root->plan_name;
778 : else
779 399667 : root->alternative_plan_name = plan_name;
780 400998 : root->parent_root = parent_root;
781 400998 : root->plan_params = NIL;
782 400998 : root->outer_params = NULL;
783 400998 : root->planner_cxt = CurrentMemoryContext;
784 400998 : root->init_plans = NIL;
785 400998 : root->cte_plan_ids = NIL;
786 400998 : root->multiexpr_params = NIL;
787 400998 : root->join_domains = NIL;
788 400998 : root->eq_classes = NIL;
789 400998 : root->ec_merging_done = false;
790 400998 : root->last_rinfo_serial = 0;
791 400998 : root->all_result_relids =
792 400998 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
793 400998 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
794 400998 : root->append_rel_list = NIL;
795 400998 : root->row_identity_vars = NIL;
796 400998 : root->rowMarks = NIL;
797 400998 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
798 400998 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
799 400998 : root->processed_groupClause = NIL;
800 400998 : root->processed_distinctClause = NIL;
801 400998 : root->processed_tlist = NIL;
802 400998 : root->update_colnos = NIL;
803 400998 : root->grouping_map = NULL;
804 400998 : root->minmax_aggs = NIL;
805 400998 : root->qual_security_level = 0;
806 400998 : root->hasPseudoConstantQuals = false;
807 400998 : root->hasAlternativeSubPlans = false;
808 400998 : root->placeholdersFrozen = false;
809 400998 : root->hasRecursion = hasRecursion;
810 400998 : root->assumeReplanning = false;
811 400998 : if (hasRecursion)
812 692 : root->wt_param_id = assign_special_exec_param(root);
813 : else
814 400306 : root->wt_param_id = -1;
815 400998 : root->non_recursive_path = NULL;
816 :
817 : /*
818 : * Create the top-level join domain. This won't have valid contents until
819 : * deconstruct_jointree fills it in, but the node needs to exist before
820 : * that so we can build EquivalenceClasses referencing it.
821 : */
822 400998 : root->join_domains = list_make1(makeNode(JoinDomain));
823 :
824 : /*
825 : * If there is a WITH list, process each WITH query and either convert it
826 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
827 : */
828 400998 : if (parse->cteList)
829 2211 : SS_process_ctes(root);
830 :
831 : /*
832 : * If it's a MERGE command, transform the joinlist as appropriate.
833 : */
834 400994 : transform_MERGE_to_join(parse);
835 :
836 : /*
837 : * Scan the rangetable for relation RTEs and retrieve the necessary
838 : * catalog information for each relation. Using this information, clear
839 : * the inh flag for any relation that has no children, collect not-null
840 : * attribute numbers for any relation that has column not-null
841 : * constraints, and expand virtual generated columns for any relation that
842 : * contains them. Note that this step does not descend into sublinks and
843 : * subqueries; if we pull up any sublinks or subqueries below, their
844 : * relation RTEs are processed just before pulling them up.
845 : */
846 400994 : parse = root->parse = preprocess_relation_rtes(root);
847 :
848 : /*
849 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
850 : * that we don't need so many special cases to deal with that situation.
851 : */
852 400994 : replace_empty_jointree(parse);
853 :
854 : /*
855 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
856 : * to transform them into joins. Note that this step does not descend
857 : * into subqueries; if we pull up any subqueries below, their SubLinks are
858 : * processed just before pulling them up.
859 : */
860 400994 : if (parse->hasSubLinks)
861 31313 : pull_up_sublinks(root);
862 :
863 : /*
864 : * Scan the rangetable for function RTEs, do const-simplification on them,
865 : * and then inline them if possible (producing subqueries that might get
866 : * pulled up next). Recursion issues here are handled in the same way as
867 : * for SubLinks.
868 : */
869 400994 : preprocess_function_rtes(root);
870 :
871 : /*
872 : * Check to see if any subqueries in the jointree can be merged into this
873 : * query.
874 : */
875 400990 : pull_up_subqueries(root);
876 :
877 : /*
878 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
879 : * do this now because it requires applying pull_up_subqueries to the leaf
880 : * queries of the UNION ALL, which weren't touched above because they
881 : * weren't referenced by the jointree (they will be after we do this).
882 : */
883 400986 : if (parse->setOperations)
884 5490 : flatten_simple_union_all(root);
885 :
886 : /*
887 : * Survey the rangetable to see what kinds of entries are present. We can
888 : * skip some later processing if relevant SQL features are not used; for
889 : * example if there are no JOIN RTEs we can avoid the expense of doing
890 : * flatten_join_alias_vars(). This must be done after we have finished
891 : * adding rangetable entries, of course. (Note: actually, processing of
892 : * inherited or partitioned rels can cause RTEs for their child tables to
893 : * get added later; but those must all be RTE_RELATION entries, so they
894 : * don't invalidate the conclusions drawn here.)
895 : */
896 400986 : root->hasJoinRTEs = false;
897 400986 : root->hasLateralRTEs = false;
898 400986 : root->group_rtindex = 0;
899 400986 : hasOuterJoins = false;
900 400986 : hasResultRTEs = false;
901 1108765 : foreach(l, parse->rtable)
902 : {
903 707779 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
904 :
905 707779 : switch (rte->rtekind)
906 : {
907 74266 : case RTE_JOIN:
908 74266 : root->hasJoinRTEs = true;
909 74266 : if (IS_OUTER_JOIN(rte->jointype))
910 37132 : hasOuterJoins = true;
911 74266 : break;
912 147952 : case RTE_RESULT:
913 147952 : hasResultRTEs = true;
914 147952 : break;
915 4137 : case RTE_GROUP:
916 : Assert(parse->hasGroupRTE);
917 4137 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
918 4137 : break;
919 481424 : default:
920 : /* No work here for other RTE types */
921 481424 : break;
922 : }
923 :
924 707779 : if (rte->lateral)
925 8309 : root->hasLateralRTEs = true;
926 :
927 : /*
928 : * We can also determine the maximum security level required for any
929 : * securityQuals now. Addition of inheritance-child RTEs won't affect
930 : * this, because child tables don't have their own securityQuals; see
931 : * expand_single_inheritance_child().
932 : */
933 707779 : if (rte->securityQuals)
934 2455 : root->qual_security_level = Max(root->qual_security_level,
935 : list_length(rte->securityQuals));
936 : }
937 :
938 : /*
939 : * If we have now verified that the query target relation is
940 : * non-inheriting, mark it as a leaf target.
941 : */
942 400986 : if (parse->resultRelation)
943 : {
944 65256 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
945 :
946 65256 : if (!rte->inh)
947 62940 : root->leaf_result_relids =
948 62940 : bms_make_singleton(parse->resultRelation);
949 : }
950 :
951 : /*
952 : * This would be a convenient time to check access permissions for all
953 : * relations mentioned in the query, since it would be better to fail now,
954 : * before doing any detailed planning. However, for historical reasons,
955 : * we leave this to be done at executor startup.
956 : *
957 : * Note, however, that we do need to check access permissions for any view
958 : * relations mentioned in the query, in order to prevent information being
959 : * leaked by selectivity estimation functions, which only check view owner
960 : * permissions on underlying tables (see all_rows_selectable() and its
961 : * callers). This is a little ugly, because it means that access
962 : * permissions for views will be checked twice, which is another reason
963 : * why it would be better to do all the ACL checks here.
964 : */
965 1107977 : foreach(l, parse->rtable)
966 : {
967 707257 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
968 :
969 707257 : if (rte->perminfoindex != 0 &&
970 382986 : rte->relkind == RELKIND_VIEW)
971 : {
972 : RTEPermissionInfo *perminfo;
973 : bool result;
974 :
975 16499 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
976 16499 : result = ExecCheckOneRelPerms(perminfo);
977 16499 : if (!result)
978 266 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
979 266 : get_rel_name(perminfo->relid));
980 : }
981 : }
982 :
983 : /*
984 : * Preprocess RowMark information. We need to do this after subquery
985 : * pullup, so that all base relations are present.
986 : */
987 400720 : preprocess_rowmarks(root);
988 :
989 : /*
990 : * Set hasHavingQual to remember if HAVING clause is present. Needed
991 : * because preprocess_expression will reduce a constant-true condition to
992 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
993 : */
994 400720 : root->hasHavingQual = (parse->havingQual != NULL);
995 :
996 : /*
997 : * Do expression preprocessing on targetlist and quals, as well as other
998 : * random expressions in the querytree. Note that we do not need to
999 : * handle sort/group expressions explicitly, because they are actually
1000 : * part of the targetlist.
1001 : */
1002 398087 : parse->targetList = (List *)
1003 400720 : preprocess_expression(root, (Node *) parse->targetList,
1004 : EXPRKIND_TARGET);
1005 :
1006 398087 : newWithCheckOptions = NIL;
1007 400655 : foreach(l, parse->withCheckOptions)
1008 : {
1009 2568 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
1010 :
1011 2568 : wco->qual = preprocess_expression(root, wco->qual,
1012 : EXPRKIND_QUAL);
1013 2568 : if (wco->qual != NULL)
1014 2236 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
1015 : }
1016 398087 : parse->withCheckOptions = newWithCheckOptions;
1017 :
1018 398087 : parse->returningList = (List *)
1019 398087 : preprocess_expression(root, (Node *) parse->returningList,
1020 : EXPRKIND_TARGET);
1021 :
1022 398087 : preprocess_qual_conditions(root, (Node *) parse->jointree);
1023 :
1024 398087 : parse->havingQual = preprocess_expression(root, parse->havingQual,
1025 : EXPRKIND_QUAL);
1026 :
1027 400519 : foreach(l, parse->windowClause)
1028 : {
1029 2432 : WindowClause *wc = lfirst_node(WindowClause, l);
1030 :
1031 : /* partitionClause/orderClause are sort/group expressions */
1032 2432 : wc->startOffset = preprocess_expression(root, wc->startOffset,
1033 : EXPRKIND_LIMIT);
1034 2432 : wc->endOffset = preprocess_expression(root, wc->endOffset,
1035 : EXPRKIND_LIMIT);
1036 : }
1037 :
1038 398087 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
1039 : EXPRKIND_LIMIT);
1040 398087 : parse->limitCount = preprocess_expression(root, parse->limitCount,
1041 : EXPRKIND_LIMIT);
1042 :
1043 398087 : if (parse->onConflict)
1044 : {
1045 3528 : parse->onConflict->arbiterElems = (List *)
1046 1764 : preprocess_expression(root,
1047 1764 : (Node *) parse->onConflict->arbiterElems,
1048 : EXPRKIND_ARBITER_ELEM);
1049 3528 : parse->onConflict->arbiterWhere =
1050 1764 : preprocess_expression(root,
1051 1764 : parse->onConflict->arbiterWhere,
1052 : EXPRKIND_QUAL);
1053 3528 : parse->onConflict->onConflictSet = (List *)
1054 1764 : preprocess_expression(root,
1055 1764 : (Node *) parse->onConflict->onConflictSet,
1056 : EXPRKIND_TARGET);
1057 1764 : parse->onConflict->onConflictWhere =
1058 1764 : preprocess_expression(root,
1059 1764 : parse->onConflict->onConflictWhere,
1060 : EXPRKIND_QUAL);
1061 : /* exclRelTlist contains only Vars, so no preprocessing needed */
1062 : }
1063 :
1064 400353 : foreach(l, parse->mergeActionList)
1065 : {
1066 2266 : MergeAction *action = (MergeAction *) lfirst(l);
1067 :
1068 2266 : action->targetList = (List *)
1069 2266 : preprocess_expression(root,
1070 2266 : (Node *) action->targetList,
1071 : EXPRKIND_TARGET);
1072 2266 : action->qual =
1073 2266 : preprocess_expression(root,
1074 : (Node *) action->qual,
1075 : EXPRKIND_QUAL);
1076 : }
1077 :
1078 398087 : parse->mergeJoinCondition =
1079 398087 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1080 :
1081 398087 : root->append_rel_list = (List *)
1082 398087 : preprocess_expression(root, (Node *) root->append_rel_list,
1083 : EXPRKIND_APPINFO);
1084 :
1085 : /* Also need to preprocess expressions within RTEs */
1086 1102044 : foreach(l, parse->rtable)
1087 : {
1088 703957 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1089 : int kind;
1090 : ListCell *lcsq;
1091 :
1092 703957 : if (rte->rtekind == RTE_RELATION)
1093 : {
1094 369591 : if (rte->tablesample)
1095 180 : rte->tablesample = (TableSampleClause *)
1096 180 : preprocess_expression(root,
1097 180 : (Node *) rte->tablesample,
1098 : EXPRKIND_TABLESAMPLE);
1099 : }
1100 334366 : else if (rte->rtekind == RTE_SUBQUERY)
1101 : {
1102 : /*
1103 : * We don't want to do all preprocessing yet on the subquery's
1104 : * expressions, since that will happen when we plan it. But if it
1105 : * contains any join aliases of our level, those have to get
1106 : * expanded now, because planning of the subquery won't do it.
1107 : * That's only possible if the subquery is LATERAL.
1108 : */
1109 63815 : if (rte->lateral && root->hasJoinRTEs)
1110 1595 : rte->subquery = (Query *)
1111 1595 : flatten_join_alias_vars(root, root->parse,
1112 1595 : (Node *) rte->subquery);
1113 : }
1114 270551 : else if (rte->rtekind == RTE_FUNCTION)
1115 : {
1116 : /* Preprocess the function expression(s) fully */
1117 35592 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1118 35592 : rte->functions = (List *)
1119 35592 : preprocess_expression(root, (Node *) rte->functions, kind);
1120 : }
1121 234959 : else if (rte->rtekind == RTE_TABLEFUNC)
1122 : {
1123 : /* Preprocess the function expression(s) fully */
1124 519 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1125 519 : rte->tablefunc = (TableFunc *)
1126 519 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1127 : }
1128 234440 : else if (rte->rtekind == RTE_VALUES)
1129 : {
1130 : /* Preprocess the values lists fully */
1131 6839 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1132 6839 : rte->values_lists = (List *)
1133 6839 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1134 : }
1135 227601 : else if (rte->rtekind == RTE_GROUP)
1136 : {
1137 : /* Preprocess the groupexprs list fully */
1138 4137 : rte->groupexprs = (List *)
1139 4137 : preprocess_expression(root, (Node *) rte->groupexprs,
1140 : EXPRKIND_GROUPEXPR);
1141 : }
1142 :
1143 : /*
1144 : * Process each element of the securityQuals list as if it were a
1145 : * separate qual expression (as indeed it is). We need to do it this
1146 : * way to get proper canonicalization of AND/OR structure. Note that
1147 : * this converts each element into an implicit-AND sublist.
1148 : */
1149 706762 : foreach(lcsq, rte->securityQuals)
1150 : {
1151 2805 : lfirst(lcsq) = preprocess_expression(root,
1152 2805 : (Node *) lfirst(lcsq),
1153 : EXPRKIND_QUAL);
1154 : }
1155 : }
1156 :
1157 : /*
1158 : * Now that we are done preprocessing expressions, and in particular done
1159 : * flattening join alias variables, get rid of the joinaliasvars lists.
1160 : * They no longer match what expressions in the rest of the tree look
1161 : * like, because we have not preprocessed expressions in those lists (and
1162 : * do not want to; for example, expanding a SubLink there would result in
1163 : * a useless unreferenced subplan). Leaving them in place simply creates
1164 : * a hazard for later scans of the tree. We could try to prevent that by
1165 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1166 : * but that doesn't sound very reliable.
1167 : */
1168 398087 : if (root->hasJoinRTEs)
1169 : {
1170 258538 : foreach(l, parse->rtable)
1171 : {
1172 211868 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1173 :
1174 211868 : rte->joinaliasvars = NIL;
1175 : }
1176 : }
1177 :
1178 : /*
1179 : * Replace any Vars in the subquery's targetlist and havingQual that
1180 : * reference GROUP outputs with the underlying grouping expressions.
1181 : *
1182 : * Note that we need to perform this replacement after we've preprocessed
1183 : * the grouping expressions. This is to ensure that there is only one
1184 : * instance of SubPlan for each SubLink contained within the grouping
1185 : * expressions.
1186 : */
1187 398087 : if (parse->hasGroupRTE)
1188 : {
1189 4137 : parse->targetList = (List *)
1190 4137 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1191 4137 : parse->havingQual =
1192 4137 : flatten_group_exprs(root, root->parse, parse->havingQual);
1193 : }
1194 :
1195 : /* Constant-folding might have removed all set-returning functions */
1196 398087 : if (parse->hasTargetSRFs)
1197 10354 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1198 :
1199 : /*
1200 : * If we have grouping sets, expand the groupingSets tree of this query to
1201 : * a flat list of grouping sets. We need to do this before optimizing
1202 : * HAVING, since we can't easily tell if there's an empty grouping set
1203 : * until we have this representation.
1204 : */
1205 398087 : if (parse->groupingSets)
1206 : {
1207 905 : parse->groupingSets =
1208 905 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1209 : }
1210 :
1211 : /*
1212 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1213 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1214 : * volatile functions (since a HAVING clause is supposed to be executed
1215 : * only once per group). We also can't do this if there are any grouping
1216 : * sets and the clause references any columns that are nullable by the
1217 : * grouping sets; the nulled values of those columns are not available
1218 : * before the grouping step. (The test on groupClause might seem wrong,
1219 : * but it's okay: it's just an optimization to avoid running pull_varnos
1220 : * when there cannot be any Vars in the HAVING clause.)
1221 : *
1222 : * Also, it may be that the clause is so expensive to execute that we're
1223 : * better off doing it only once per group, despite the loss of
1224 : * selectivity. This is hard to estimate short of doing the entire
1225 : * planning process twice, so we use a heuristic: clauses containing
1226 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1227 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1228 : * instead of after.
1229 : *
1230 : * If the query has no empty grouping set then we can simply move such a
1231 : * clause into WHERE; any group that fails the clause will not be in the
1232 : * output because none of its tuples will reach the grouping or
1233 : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1234 : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1235 : * clause must be degenerate (variable-free), so we can copy it into WHERE
1236 : * so that query_planner() can use it in a gating Result node. (This could
1237 : * be done better, but it seems not worth optimizing.)
1238 : *
1239 : * Note that a HAVING clause may contain expressions that are not fully
1240 : * preprocessed. This can happen if these expressions are part of
1241 : * grouping items. In such cases, they are replaced with GROUP Vars in
1242 : * the parser and then replaced back after we're done with expression
1243 : * preprocessing on havingQual. This is not an issue if the clause
1244 : * remains in HAVING, because these expressions will be matched to lower
1245 : * target items in setrefs.c. However, if the clause is moved or copied
1246 : * into WHERE, we need to ensure that these expressions are fully
1247 : * preprocessed.
1248 : *
1249 : * Note that both havingQual and parse->jointree->quals are in
1250 : * implicitly-ANDed-list form at this point, even though they are declared
1251 : * as Node *.
1252 : */
1253 398087 : newHaving = NIL;
1254 399187 : foreach(l, (List *) parse->havingQual)
1255 : {
1256 1100 : Node *havingclause = (Node *) lfirst(l);
1257 :
1258 1526 : if (contain_agg_clause(havingclause) ||
1259 852 : contain_volatile_functions(havingclause) ||
1260 426 : contain_subplans(havingclause) ||
1261 526 : (parse->groupClause && parse->groupingSets &&
1262 100 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1263 : {
1264 : /* keep it in HAVING */
1265 734 : newHaving = lappend(newHaving, havingclause);
1266 : }
1267 366 : else if (parse->groupClause &&
1268 336 : (parse->groupingSets == NIL ||
1269 40 : (List *) linitial(parse->groupingSets) != NIL))
1270 326 : {
1271 : /* There is GROUP BY, but no empty grouping set */
1272 : Node *whereclause;
1273 :
1274 : /* Preprocess the HAVING clause fully */
1275 326 : whereclause = preprocess_expression(root, havingclause,
1276 : EXPRKIND_QUAL);
1277 : /* ... and move it to WHERE */
1278 326 : parse->jointree->quals = (Node *)
1279 326 : list_concat((List *) parse->jointree->quals,
1280 : (List *) whereclause);
1281 : }
1282 : else
1283 : {
1284 : /* There is an empty grouping set (perhaps implicitly) */
1285 : Node *whereclause;
1286 :
1287 : /* Preprocess the HAVING clause fully */
1288 40 : whereclause = preprocess_expression(root, copyObject(havingclause),
1289 : EXPRKIND_QUAL);
1290 : /* ... and put a copy in WHERE */
1291 80 : parse->jointree->quals = (Node *)
1292 40 : list_concat((List *) parse->jointree->quals,
1293 : (List *) whereclause);
1294 : /* ... and also keep it in HAVING */
1295 40 : newHaving = lappend(newHaving, havingclause);
1296 : }
1297 : }
1298 398087 : parse->havingQual = (Node *) newHaving;
1299 :
1300 : /*
1301 : * If we have any outer joins, try to reduce them to plain inner joins.
1302 : * This step is most easily done after we've done expression
1303 : * preprocessing.
1304 : */
1305 398087 : if (hasOuterJoins)
1306 25777 : reduce_outer_joins(root);
1307 :
1308 : /*
1309 : * If we have any RTE_RESULT relations, see if they can be deleted from
1310 : * the jointree. We also rely on this processing to flatten single-child
1311 : * FromExprs underneath outer joins. This step is most effectively done
1312 : * after we've done expression preprocessing and outer join reduction.
1313 : */
1314 398087 : if (hasResultRTEs || hasOuterJoins)
1315 169922 : remove_useless_result_rtes(root);
1316 :
1317 : /*
1318 : * Do the main planning.
1319 : */
1320 398087 : grouping_planner(root, tuple_fraction, setops);
1321 :
1322 : /*
1323 : * Capture the set of outer-level param IDs we have access to, for use in
1324 : * extParam/allParam calculations later.
1325 : */
1326 398041 : SS_identify_outer_params(root);
1327 :
1328 : /*
1329 : * If any initPlans were created in this query level, adjust the surviving
1330 : * Paths' costs and parallel-safety flags to account for them. The
1331 : * initPlans won't actually get attached to the plan tree till
1332 : * create_plan() runs, but we must include their effects now.
1333 : */
1334 398041 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1335 398041 : SS_charge_for_initplans(root, final_rel);
1336 :
1337 : /*
1338 : * Make sure we've identified the cheapest Path for the final rel. (By
1339 : * doing this here not in grouping_planner, we include initPlan costs in
1340 : * the decision, though it's unlikely that will change anything.)
1341 : */
1342 398041 : set_cheapest(final_rel);
1343 :
1344 398041 : return root;
1345 : }
1346 :
1347 : /*
1348 : * preprocess_expression
1349 : * Do subquery_planner's preprocessing work for an expression,
1350 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1351 : * conditions), a HAVING clause, or a few other things.
1352 : */
1353 : static Node *
1354 3354872 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1355 : {
1356 : /*
1357 : * Fall out quickly if expression is empty. This occurs often enough to
1358 : * be worth checking. Note that null->null is the correct conversion for
1359 : * implicit-AND result format, too.
1360 : */
1361 3354872 : if (expr == NULL)
1362 2638045 : return NULL;
1363 :
1364 : /*
1365 : * If the query has any join RTEs, replace join alias variables with
1366 : * base-relation variables. We must do this first, since any expressions
1367 : * we may extract from the joinaliasvars lists have not been preprocessed.
1368 : * For example, if we did this after sublink processing, sublinks expanded
1369 : * out from join aliases would not get processed. But we can skip this in
1370 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1371 : * they can't contain any Vars of the current query level.
1372 : */
1373 716827 : if (root->hasJoinRTEs &&
1374 332215 : !(kind == EXPRKIND_RTFUNC ||
1375 165959 : kind == EXPRKIND_VALUES ||
1376 : kind == EXPRKIND_TABLESAMPLE ||
1377 : kind == EXPRKIND_TABLEFUNC))
1378 165944 : expr = flatten_join_alias_vars(root, root->parse, expr);
1379 :
1380 : /*
1381 : * Simplify constant expressions. For function RTEs, this was already
1382 : * done by preprocess_function_rtes. (But note we must do it again for
1383 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1384 : * un-simplified subexpressions inserted by flattening of subqueries or
1385 : * join alias variables.)
1386 : *
1387 : * Note: an essential effect of this is to convert named-argument function
1388 : * calls to positional notation and insert the current actual values of
1389 : * any default arguments for functions. To ensure that happens, we *must*
1390 : * process all expressions here. Previous PG versions sometimes skipped
1391 : * const-simplification if it didn't seem worth the trouble, but we can't
1392 : * do that anymore.
1393 : *
1394 : * Note: this also flattens nested AND and OR expressions into N-argument
1395 : * form. All processing of a qual expression after this point must be
1396 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1397 : * with AND directly under AND, nor OR directly under OR.
1398 : */
1399 716827 : if (kind != EXPRKIND_RTFUNC)
1400 685940 : expr = eval_const_expressions(root, expr);
1401 :
1402 : /*
1403 : * If it's a qual or havingQual, canonicalize it.
1404 : */
1405 714194 : if (kind == EXPRKIND_QUAL)
1406 : {
1407 264558 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1408 :
1409 : #ifdef OPTIMIZER_DEBUG
1410 : printf("After canonicalize_qual()\n");
1411 : pprint(expr);
1412 : #endif
1413 : }
1414 :
1415 : /*
1416 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1417 : * hashfuncid of any that might execute more quickly by using hash lookups
1418 : * instead of a linear search.
1419 : */
1420 714194 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1421 : {
1422 655496 : convert_saop_to_hashed_saop(expr);
1423 : }
1424 :
1425 : /* Expand SubLinks to SubPlans */
1426 714194 : if (root->parse->hasSubLinks)
1427 92004 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1428 :
1429 : /*
1430 : * XXX do not insert anything here unless you have grokked the comments in
1431 : * SS_replace_correlation_vars ...
1432 : */
1433 :
1434 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1435 714194 : if (root->query_level > 1)
1436 149484 : expr = SS_replace_correlation_vars(root, expr);
1437 :
1438 : /*
1439 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1440 : * don't want to do this before eval_const_expressions, since the latter
1441 : * would be unable to simplify a top-level AND correctly. Also,
1442 : * SS_process_sublinks expects explicit-AND format.)
1443 : */
1444 714194 : if (kind == EXPRKIND_QUAL)
1445 264558 : expr = (Node *) make_ands_implicit((Expr *) expr);
1446 :
1447 714194 : return expr;
1448 : }
1449 :
1450 : /*
1451 : * preprocess_qual_conditions
1452 : * Recursively scan the query's jointree and do subquery_planner's
1453 : * preprocessing work on each qual condition found therein.
1454 : */
1455 : static void
1456 1010665 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1457 : {
1458 1010665 : if (jtnode == NULL)
1459 0 : return;
1460 1010665 : if (IsA(jtnode, RangeTblRef))
1461 : {
1462 : /* nothing to do here */
1463 : }
1464 496097 : else if (IsA(jtnode, FromExpr))
1465 : {
1466 413017 : FromExpr *f = (FromExpr *) jtnode;
1467 : ListCell *l;
1468 :
1469 859435 : foreach(l, f->fromlist)
1470 446418 : preprocess_qual_conditions(root, lfirst(l));
1471 :
1472 413017 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1473 : }
1474 83080 : else if (IsA(jtnode, JoinExpr))
1475 : {
1476 83080 : JoinExpr *j = (JoinExpr *) jtnode;
1477 :
1478 83080 : preprocess_qual_conditions(root, j->larg);
1479 83080 : preprocess_qual_conditions(root, j->rarg);
1480 :
1481 83080 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1482 : }
1483 : else
1484 0 : elog(ERROR, "unrecognized node type: %d",
1485 : (int) nodeTag(jtnode));
1486 : }
1487 :
1488 : /*
1489 : * preprocess_phv_expression
1490 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1491 : *
1492 : * If a LATERAL subquery references an output of another subquery, and that
1493 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1494 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1495 : * and later pull it back up during find_lateral_references, which runs after
1496 : * subquery_planner has preprocessed all the expressions that were in the
1497 : * current query level to start with. So we need to preprocess it then.
1498 : */
1499 : Expr *
1500 75 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1501 : {
1502 75 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1503 : }
1504 :
1505 : /*--------------------
1506 : * grouping_planner
1507 : * Perform planning steps related to grouping, aggregation, etc.
1508 : *
1509 : * This function adds all required top-level processing to the scan/join
1510 : * Path(s) produced by query_planner.
1511 : *
1512 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1513 : * tuple_fraction is interpreted as follows:
1514 : * 0: expect all tuples to be retrieved (normal case)
1515 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1516 : * from the plan to be retrieved
1517 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1518 : * expected to be retrieved (ie, a LIMIT specification).
1519 : * setops is used for set operation subqueries to provide the subquery with
1520 : * the context in which it's being used so that Paths correctly sorted for the
1521 : * set operation can be generated. NULL when not planning a set operation
1522 : * child, or when a child of a set op that isn't interested in sorted input.
1523 : *
1524 : * Returns nothing; the useful output is in the Paths we attach to the
1525 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1526 : * root->processed_tlist contains the final processed targetlist.
1527 : *
1528 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1529 : * to leave this to the caller.
1530 : *--------------------
1531 : */
1532 : static void
1533 398087 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1534 : SetOperationStmt *setops)
1535 : {
1536 398087 : Query *parse = root->parse;
1537 398087 : int64 offset_est = 0;
1538 398087 : int64 count_est = 0;
1539 398087 : double limit_tuples = -1.0;
1540 398087 : bool have_postponed_srfs = false;
1541 : PathTarget *final_target;
1542 : List *final_targets;
1543 : List *final_targets_contain_srfs;
1544 : bool final_target_parallel_safe;
1545 : RelOptInfo *current_rel;
1546 : RelOptInfo *final_rel;
1547 : FinalPathExtraData extra;
1548 : ListCell *lc;
1549 :
1550 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1551 398087 : if (parse->limitCount || parse->limitOffset)
1552 : {
1553 3707 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1554 : &offset_est, &count_est);
1555 :
1556 : /*
1557 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1558 : * estimate the effects of using a bounded sort.
1559 : */
1560 3707 : if (count_est > 0 && offset_est >= 0)
1561 3245 : limit_tuples = (double) count_est + (double) offset_est;
1562 : }
1563 :
1564 : /* Make tuple_fraction accessible to lower-level routines */
1565 398087 : root->tuple_fraction = tuple_fraction;
1566 :
1567 398087 : if (parse->setOperations)
1568 : {
1569 : /*
1570 : * Construct Paths for set operations. The results will not need any
1571 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1572 : * special work for recursive unions is the responsibility of
1573 : * plan_set_operations.
1574 : */
1575 4946 : current_rel = plan_set_operations(root);
1576 :
1577 : /*
1578 : * We should not need to call preprocess_targetlist, since we must be
1579 : * in a SELECT query node. Instead, use the processed_tlist returned
1580 : * by plan_set_operations (since this tells whether it returned any
1581 : * resjunk columns!), and transfer any sort key information from the
1582 : * original tlist.
1583 : */
1584 : Assert(parse->commandType == CMD_SELECT);
1585 :
1586 : /* for safety, copy processed_tlist instead of modifying in-place */
1587 4942 : root->processed_tlist =
1588 4942 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1589 : parse->targetList);
1590 :
1591 : /* Also extract the PathTarget form of the setop result tlist */
1592 4942 : final_target = current_rel->cheapest_total_path->pathtarget;
1593 :
1594 : /* And check whether it's parallel safe */
1595 : final_target_parallel_safe =
1596 4942 : is_parallel_safe(root, (Node *) final_target->exprs);
1597 :
1598 : /* The setop result tlist couldn't contain any SRFs */
1599 : Assert(!parse->hasTargetSRFs);
1600 4942 : final_targets = final_targets_contain_srfs = NIL;
1601 :
1602 : /*
1603 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1604 : * checked already, but let's make sure).
1605 : */
1606 4942 : if (parse->rowMarks)
1607 0 : ereport(ERROR,
1608 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1609 : /*------
1610 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1611 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1612 : LCS_asString(linitial_node(RowMarkClause,
1613 : parse->rowMarks)->strength))));
1614 :
1615 : /*
1616 : * Calculate pathkeys that represent result ordering requirements
1617 : */
1618 : Assert(parse->distinctClause == NIL);
1619 4942 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1620 : parse->sortClause,
1621 : root->processed_tlist);
1622 : }
1623 : else
1624 : {
1625 : /* No set operations, do regular planning */
1626 : PathTarget *sort_input_target;
1627 : List *sort_input_targets;
1628 : List *sort_input_targets_contain_srfs;
1629 : bool sort_input_target_parallel_safe;
1630 : PathTarget *grouping_target;
1631 : List *grouping_targets;
1632 : List *grouping_targets_contain_srfs;
1633 : bool grouping_target_parallel_safe;
1634 : PathTarget *scanjoin_target;
1635 : List *scanjoin_targets;
1636 : List *scanjoin_targets_contain_srfs;
1637 : bool scanjoin_target_parallel_safe;
1638 : bool scanjoin_target_same_exprs;
1639 : bool have_grouping;
1640 393141 : WindowFuncLists *wflists = NULL;
1641 393141 : List *activeWindows = NIL;
1642 393141 : grouping_sets_data *gset_data = NULL;
1643 : standard_qp_extra qp_extra;
1644 :
1645 : /* A recursive query should always have setOperations */
1646 : Assert(!root->hasRecursion);
1647 :
1648 : /* Preprocess grouping sets and GROUP BY clause, if any */
1649 393141 : if (parse->groupingSets)
1650 : {
1651 905 : gset_data = preprocess_grouping_sets(root);
1652 : }
1653 392236 : else if (parse->groupClause)
1654 : {
1655 : /* Preprocess regular GROUP BY clause, if any */
1656 3297 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1657 : }
1658 :
1659 : /*
1660 : * Preprocess targetlist. Note that much of the remaining planning
1661 : * work will be done with the PathTarget representation of tlists, but
1662 : * we must also maintain the full representation of the final tlist so
1663 : * that we can transfer its decoration (resnames etc) to the topmost
1664 : * tlist of the finished Plan. This is kept in processed_tlist.
1665 : */
1666 393137 : preprocess_targetlist(root);
1667 :
1668 : /*
1669 : * Mark all the aggregates with resolved aggtranstypes, and detect
1670 : * aggregates that are duplicates or can share transition state. We
1671 : * must do this before slicing and dicing the tlist into various
1672 : * pathtargets, else some copies of the Aggref nodes might escape
1673 : * being marked.
1674 : */
1675 393137 : if (parse->hasAggs)
1676 : {
1677 33136 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1678 33136 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1679 : }
1680 :
1681 : /*
1682 : * Locate any window functions in the tlist. (We don't need to look
1683 : * anywhere else, since expressions used in ORDER BY will be in there
1684 : * too.) Note that they could all have been eliminated by constant
1685 : * folding, in which case we don't need to do any more work.
1686 : */
1687 393137 : if (parse->hasWindowFuncs)
1688 : {
1689 2222 : wflists = find_window_functions((Node *) root->processed_tlist,
1690 2222 : list_length(parse->windowClause));
1691 2222 : if (wflists->numWindowFuncs > 0)
1692 : {
1693 : /*
1694 : * See if any modifications can be made to each WindowClause
1695 : * to allow the executor to execute the WindowFuncs more
1696 : * quickly.
1697 : */
1698 2217 : optimize_window_clauses(root, wflists);
1699 :
1700 : /* Extract the list of windows actually in use. */
1701 2217 : activeWindows = select_active_windows(root, wflists);
1702 :
1703 : /* Make sure they all have names, for EXPLAIN's use. */
1704 2217 : name_active_windows(activeWindows);
1705 : }
1706 : else
1707 5 : parse->hasWindowFuncs = false;
1708 : }
1709 :
1710 : /*
1711 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1712 : * adding logic between here and the query_planner() call. Anything
1713 : * that is needed in MIN/MAX-optimizable cases will have to be
1714 : * duplicated in planagg.c.
1715 : */
1716 393137 : if (parse->hasAggs)
1717 33136 : preprocess_minmax_aggregates(root);
1718 :
1719 : /*
1720 : * Figure out whether there's a hard limit on the number of rows that
1721 : * query_planner's result subplan needs to return. Even if we know a
1722 : * hard limit overall, it doesn't apply if the query has any
1723 : * grouping/aggregation operations, or SRFs in the tlist.
1724 : */
1725 393137 : if (parse->groupClause ||
1726 389004 : parse->groupingSets ||
1727 388939 : parse->distinctClause ||
1728 387025 : parse->hasAggs ||
1729 357434 : parse->hasWindowFuncs ||
1730 355333 : parse->hasTargetSRFs ||
1731 345313 : root->hasHavingQual)
1732 47844 : root->limit_tuples = -1.0;
1733 : else
1734 345293 : root->limit_tuples = limit_tuples;
1735 :
1736 : /* Set up data needed by standard_qp_callback */
1737 393137 : qp_extra.activeWindows = activeWindows;
1738 393137 : qp_extra.gset_data = gset_data;
1739 :
1740 : /*
1741 : * If we're a subquery for a set operation, store the SetOperationStmt
1742 : * in qp_extra.
1743 : */
1744 393137 : qp_extra.setop = setops;
1745 :
1746 : /*
1747 : * Generate the best unsorted and presorted paths for the scan/join
1748 : * portion of this Query, ie the processing represented by the
1749 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1750 : * We also generate (in standard_qp_callback) pathkey representations
1751 : * of the query's sort clause, distinct clause, etc.
1752 : */
1753 393137 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1754 :
1755 : /*
1756 : * Convert the query's result tlist into PathTarget format.
1757 : *
1758 : * Note: this cannot be done before query_planner() has performed
1759 : * appendrel expansion, because that might add resjunk entries to
1760 : * root->processed_tlist. Waiting till afterwards is also helpful
1761 : * because the target width estimates can use per-Var width numbers
1762 : * that were obtained within query_planner().
1763 : */
1764 393103 : final_target = create_pathtarget(root, root->processed_tlist);
1765 : final_target_parallel_safe =
1766 393103 : is_parallel_safe(root, (Node *) final_target->exprs);
1767 :
1768 : /*
1769 : * If ORDER BY was given, consider whether we should use a post-sort
1770 : * projection, and compute the adjusted target for preceding steps if
1771 : * so.
1772 : */
1773 393103 : if (parse->sortClause)
1774 : {
1775 58652 : sort_input_target = make_sort_input_target(root,
1776 : final_target,
1777 : &have_postponed_srfs);
1778 : sort_input_target_parallel_safe =
1779 58652 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1780 : }
1781 : else
1782 : {
1783 334451 : sort_input_target = final_target;
1784 334451 : sort_input_target_parallel_safe = final_target_parallel_safe;
1785 : }
1786 :
1787 : /*
1788 : * If we have window functions to deal with, the output from any
1789 : * grouping step needs to be what the window functions want;
1790 : * otherwise, it should be sort_input_target.
1791 : */
1792 393103 : if (activeWindows)
1793 : {
1794 2217 : grouping_target = make_window_input_target(root,
1795 : final_target,
1796 : activeWindows);
1797 : grouping_target_parallel_safe =
1798 2217 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1799 : }
1800 : else
1801 : {
1802 390886 : grouping_target = sort_input_target;
1803 390886 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1804 : }
1805 :
1806 : /*
1807 : * If we have grouping or aggregation to do, the topmost scan/join
1808 : * plan node must emit what the grouping step wants; otherwise, it
1809 : * should emit grouping_target.
1810 : */
1811 388970 : have_grouping = (parse->groupClause || parse->groupingSets ||
1812 782073 : parse->hasAggs || root->hasHavingQual);
1813 393103 : if (have_grouping)
1814 : {
1815 33840 : scanjoin_target = make_group_input_target(root, final_target);
1816 : scanjoin_target_parallel_safe =
1817 33840 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1818 : }
1819 : else
1820 : {
1821 359263 : scanjoin_target = grouping_target;
1822 359263 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1823 : }
1824 :
1825 : /*
1826 : * If there are any SRFs in the targetlist, we must separate each of
1827 : * these PathTargets into SRF-computing and SRF-free targets. Replace
1828 : * each of the named targets with a SRF-free version, and remember the
1829 : * list of additional projection steps we need to add afterwards.
1830 : */
1831 393103 : if (parse->hasTargetSRFs)
1832 : {
1833 : /* final_target doesn't recompute any SRFs in sort_input_target */
1834 10354 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1835 : &final_targets,
1836 : &final_targets_contain_srfs);
1837 10354 : final_target = linitial_node(PathTarget, final_targets);
1838 : Assert(!linitial_int(final_targets_contain_srfs));
1839 : /* likewise for sort_input_target vs. grouping_target */
1840 10354 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1841 : &sort_input_targets,
1842 : &sort_input_targets_contain_srfs);
1843 10354 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
1844 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1845 : /* likewise for grouping_target vs. scanjoin_target */
1846 10354 : split_pathtarget_at_srfs_grouping(root,
1847 : grouping_target, scanjoin_target,
1848 : &grouping_targets,
1849 : &grouping_targets_contain_srfs);
1850 10354 : grouping_target = linitial_node(PathTarget, grouping_targets);
1851 : Assert(!linitial_int(grouping_targets_contain_srfs));
1852 : /* scanjoin_target will not have any SRFs precomputed for it */
1853 10354 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1854 : &scanjoin_targets,
1855 : &scanjoin_targets_contain_srfs);
1856 10354 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1857 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1858 : }
1859 : else
1860 : {
1861 : /* initialize lists; for most of these, dummy values are OK */
1862 382749 : final_targets = final_targets_contain_srfs = NIL;
1863 382749 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1864 382749 : grouping_targets = grouping_targets_contain_srfs = NIL;
1865 382749 : scanjoin_targets = list_make1(scanjoin_target);
1866 382749 : scanjoin_targets_contain_srfs = NIL;
1867 : }
1868 :
1869 : /* Apply scan/join target. */
1870 393103 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1871 393103 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1872 393103 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1873 : scanjoin_targets_contain_srfs,
1874 : scanjoin_target_parallel_safe,
1875 : scanjoin_target_same_exprs);
1876 :
1877 : /*
1878 : * Save the various upper-rel PathTargets we just computed into
1879 : * root->upper_targets[]. The core code doesn't use this, but it
1880 : * provides a convenient place for extensions to get at the info. For
1881 : * consistency, we save all the intermediate targets, even though some
1882 : * of the corresponding upperrels might not be needed for this query.
1883 : */
1884 393103 : root->upper_targets[UPPERREL_FINAL] = final_target;
1885 393103 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1886 393103 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1887 393103 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1888 393103 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1889 393103 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1890 :
1891 : /*
1892 : * If we have grouping and/or aggregation, consider ways to implement
1893 : * that. We build a new upperrel representing the output of this
1894 : * phase.
1895 : */
1896 393103 : if (have_grouping)
1897 : {
1898 33840 : current_rel = create_grouping_paths(root,
1899 : current_rel,
1900 : grouping_target,
1901 : grouping_target_parallel_safe,
1902 : gset_data);
1903 : /* Fix things up if grouping_target contains SRFs */
1904 33836 : if (parse->hasTargetSRFs)
1905 299 : adjust_paths_for_srfs(root, current_rel,
1906 : grouping_targets,
1907 : grouping_targets_contain_srfs);
1908 : }
1909 :
1910 : /*
1911 : * If we have window functions, consider ways to implement those. We
1912 : * build a new upperrel representing the output of this phase.
1913 : */
1914 393099 : if (activeWindows)
1915 : {
1916 2217 : current_rel = create_window_paths(root,
1917 : current_rel,
1918 : grouping_target,
1919 : sort_input_target,
1920 : sort_input_target_parallel_safe,
1921 : wflists,
1922 : activeWindows);
1923 : /* Fix things up if sort_input_target contains SRFs */
1924 2217 : if (parse->hasTargetSRFs)
1925 10 : adjust_paths_for_srfs(root, current_rel,
1926 : sort_input_targets,
1927 : sort_input_targets_contain_srfs);
1928 : }
1929 :
1930 : /*
1931 : * If there is a DISTINCT clause, consider ways to implement that. We
1932 : * build a new upperrel representing the output of this phase.
1933 : */
1934 393099 : if (parse->distinctClause)
1935 : {
1936 1941 : current_rel = create_distinct_paths(root,
1937 : current_rel,
1938 : sort_input_target);
1939 : }
1940 : } /* end of if (setOperations) */
1941 :
1942 : /*
1943 : * If ORDER BY was given, consider ways to implement that, and generate a
1944 : * new upperrel containing only paths that emit the correct ordering and
1945 : * project the correct final_target. We can apply the original
1946 : * limit_tuples limit in sort costing here, but only if there are no
1947 : * postponed SRFs.
1948 : */
1949 398041 : if (parse->sortClause)
1950 : {
1951 61919 : current_rel = create_ordered_paths(root,
1952 : current_rel,
1953 : final_target,
1954 : final_target_parallel_safe,
1955 : have_postponed_srfs ? -1.0 :
1956 : limit_tuples);
1957 : /* Fix things up if final_target contains SRFs */
1958 61919 : if (parse->hasTargetSRFs)
1959 182 : adjust_paths_for_srfs(root, current_rel,
1960 : final_targets,
1961 : final_targets_contain_srfs);
1962 : }
1963 :
1964 : /*
1965 : * Now we are prepared to build the final-output upperrel.
1966 : */
1967 398041 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1968 :
1969 : /*
1970 : * If the input rel is marked consider_parallel and there's nothing that's
1971 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1972 : * consider_parallel as well. Note that if the query has rowMarks or is
1973 : * not a SELECT, consider_parallel will be false for every relation in the
1974 : * query.
1975 : */
1976 537118 : if (current_rel->consider_parallel &&
1977 278134 : is_parallel_safe(root, parse->limitOffset) &&
1978 139057 : is_parallel_safe(root, parse->limitCount))
1979 139052 : final_rel->consider_parallel = true;
1980 :
1981 : /*
1982 : * If the current_rel belongs to a single FDW, so does the final_rel.
1983 : */
1984 398041 : final_rel->serverid = current_rel->serverid;
1985 398041 : final_rel->userid = current_rel->userid;
1986 398041 : final_rel->useridiscurrent = current_rel->useridiscurrent;
1987 398041 : final_rel->fdwroutine = current_rel->fdwroutine;
1988 :
1989 : /*
1990 : * Generate paths for the final_rel. Insert all surviving paths, with
1991 : * LockRows, Limit, and/or ModifyTable steps added if needed.
1992 : */
1993 812477 : foreach(lc, current_rel->pathlist)
1994 : {
1995 414436 : Path *path = (Path *) lfirst(lc);
1996 :
1997 : /*
1998 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1999 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2000 : * here. If there are only non-locking rowmarks, they should be
2001 : * handled by the ModifyTable node instead. However, root->rowMarks
2002 : * is what goes into the LockRows node.)
2003 : */
2004 414436 : if (parse->rowMarks)
2005 : {
2006 6808 : path = (Path *) create_lockrows_path(root, final_rel, path,
2007 : root->rowMarks,
2008 : assign_special_exec_param(root));
2009 : }
2010 :
2011 : /*
2012 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2013 : */
2014 414436 : if (limit_needed(parse))
2015 : {
2016 4184 : path = (Path *) create_limit_path(root, final_rel, path,
2017 : parse->limitOffset,
2018 : parse->limitCount,
2019 : parse->limitOption,
2020 : offset_est, count_est);
2021 : }
2022 :
2023 : /*
2024 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
2025 : */
2026 414436 : if (parse->commandType != CMD_SELECT)
2027 : {
2028 : Index rootRelation;
2029 64902 : List *resultRelations = NIL;
2030 64902 : List *updateColnosLists = NIL;
2031 64902 : List *withCheckOptionLists = NIL;
2032 64902 : List *returningLists = NIL;
2033 64902 : List *mergeActionLists = NIL;
2034 64902 : List *mergeJoinConditions = NIL;
2035 : List *rowMarks;
2036 :
2037 64902 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
2038 : {
2039 : /* Inherited UPDATE/DELETE/MERGE */
2040 2288 : RelOptInfo *top_result_rel = find_base_rel(root,
2041 : parse->resultRelation);
2042 2288 : int resultRelation = -1;
2043 :
2044 : /* Pass the root result rel forward to the executor. */
2045 2288 : rootRelation = parse->resultRelation;
2046 :
2047 : /* Add only leaf children to ModifyTable. */
2048 6718 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
2049 6718 : resultRelation)) >= 0)
2050 : {
2051 4430 : RelOptInfo *this_result_rel = find_base_rel(root,
2052 : resultRelation);
2053 :
2054 : /*
2055 : * Also exclude any leaf rels that have turned dummy since
2056 : * being added to the list, for example, by being excluded
2057 : * by constraint exclusion.
2058 : */
2059 4430 : if (IS_DUMMY_REL(this_result_rel))
2060 153 : continue;
2061 :
2062 : /* Build per-target-rel lists needed by ModifyTable */
2063 4277 : resultRelations = lappend_int(resultRelations,
2064 : resultRelation);
2065 4277 : if (parse->commandType == CMD_UPDATE)
2066 : {
2067 2933 : List *update_colnos = root->update_colnos;
2068 :
2069 2933 : if (this_result_rel != top_result_rel)
2070 : update_colnos =
2071 2933 : adjust_inherited_attnums_multilevel(root,
2072 : update_colnos,
2073 : this_result_rel->relid,
2074 : top_result_rel->relid);
2075 2933 : updateColnosLists = lappend(updateColnosLists,
2076 : update_colnos);
2077 : }
2078 4277 : if (parse->withCheckOptions)
2079 : {
2080 416 : List *withCheckOptions = parse->withCheckOptions;
2081 :
2082 416 : if (this_result_rel != top_result_rel)
2083 : withCheckOptions = (List *)
2084 416 : adjust_appendrel_attrs_multilevel(root,
2085 : (Node *) withCheckOptions,
2086 : this_result_rel,
2087 : top_result_rel);
2088 416 : withCheckOptionLists = lappend(withCheckOptionLists,
2089 : withCheckOptions);
2090 : }
2091 4277 : if (parse->returningList)
2092 : {
2093 645 : List *returningList = parse->returningList;
2094 :
2095 645 : if (this_result_rel != top_result_rel)
2096 : returningList = (List *)
2097 645 : adjust_appendrel_attrs_multilevel(root,
2098 : (Node *) returningList,
2099 : this_result_rel,
2100 : top_result_rel);
2101 645 : returningLists = lappend(returningLists,
2102 : returningList);
2103 : }
2104 4277 : if (parse->mergeActionList)
2105 : {
2106 : ListCell *l;
2107 425 : List *mergeActionList = NIL;
2108 :
2109 : /*
2110 : * Copy MergeActions and translate stuff that
2111 : * references attribute numbers.
2112 : */
2113 1322 : foreach(l, parse->mergeActionList)
2114 : {
2115 897 : MergeAction *action = lfirst(l),
2116 897 : *leaf_action = copyObject(action);
2117 :
2118 897 : leaf_action->qual =
2119 897 : adjust_appendrel_attrs_multilevel(root,
2120 : (Node *) action->qual,
2121 : this_result_rel,
2122 : top_result_rel);
2123 897 : leaf_action->targetList = (List *)
2124 897 : adjust_appendrel_attrs_multilevel(root,
2125 897 : (Node *) action->targetList,
2126 : this_result_rel,
2127 : top_result_rel);
2128 897 : if (leaf_action->commandType == CMD_UPDATE)
2129 491 : leaf_action->updateColnos =
2130 491 : adjust_inherited_attnums_multilevel(root,
2131 : action->updateColnos,
2132 : this_result_rel->relid,
2133 : top_result_rel->relid);
2134 897 : mergeActionList = lappend(mergeActionList,
2135 : leaf_action);
2136 : }
2137 :
2138 425 : mergeActionLists = lappend(mergeActionLists,
2139 : mergeActionList);
2140 : }
2141 4277 : if (parse->commandType == CMD_MERGE)
2142 : {
2143 425 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2144 :
2145 425 : if (this_result_rel != top_result_rel)
2146 : mergeJoinCondition =
2147 425 : adjust_appendrel_attrs_multilevel(root,
2148 : mergeJoinCondition,
2149 : this_result_rel,
2150 : top_result_rel);
2151 425 : mergeJoinConditions = lappend(mergeJoinConditions,
2152 : mergeJoinCondition);
2153 : }
2154 : }
2155 :
2156 2288 : if (resultRelations == NIL)
2157 : {
2158 : /*
2159 : * We managed to exclude every child rel, so generate a
2160 : * dummy one-relation plan using info for the top target
2161 : * rel (even though that may not be a leaf target).
2162 : * Although it's clear that no data will be updated or
2163 : * deleted, we still need to have a ModifyTable node so
2164 : * that any statement triggers will be executed. (This
2165 : * could be cleaner if we fixed nodeModifyTable.c to allow
2166 : * zero target relations, but that probably wouldn't be a
2167 : * net win.)
2168 : */
2169 28 : resultRelations = list_make1_int(parse->resultRelation);
2170 28 : if (parse->commandType == CMD_UPDATE)
2171 26 : updateColnosLists = list_make1(root->update_colnos);
2172 28 : if (parse->withCheckOptions)
2173 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2174 28 : if (parse->returningList)
2175 15 : returningLists = list_make1(parse->returningList);
2176 28 : if (parse->mergeActionList)
2177 1 : mergeActionLists = list_make1(parse->mergeActionList);
2178 28 : if (parse->commandType == CMD_MERGE)
2179 1 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2180 : }
2181 : }
2182 : else
2183 : {
2184 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2185 62614 : rootRelation = 0; /* there's no separate root rel */
2186 62614 : resultRelations = list_make1_int(parse->resultRelation);
2187 62614 : if (parse->commandType == CMD_UPDATE)
2188 8759 : updateColnosLists = list_make1(root->update_colnos);
2189 62614 : if (parse->withCheckOptions)
2190 889 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2191 62614 : if (parse->returningList)
2192 2169 : returningLists = list_make1(parse->returningList);
2193 62614 : if (parse->mergeActionList)
2194 1286 : mergeActionLists = list_make1(parse->mergeActionList);
2195 62614 : if (parse->commandType == CMD_MERGE)
2196 1286 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2197 : }
2198 :
2199 : /*
2200 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2201 : * will have dealt with fetching non-locked marked rows, else we
2202 : * need to have ModifyTable do that.
2203 : */
2204 64902 : if (parse->rowMarks)
2205 0 : rowMarks = NIL;
2206 : else
2207 64902 : rowMarks = root->rowMarks;
2208 :
2209 : path = (Path *)
2210 64902 : create_modifytable_path(root, final_rel,
2211 : path,
2212 : parse->commandType,
2213 64902 : parse->canSetTag,
2214 64902 : parse->resultRelation,
2215 : rootRelation,
2216 : resultRelations,
2217 : updateColnosLists,
2218 : withCheckOptionLists,
2219 : returningLists,
2220 : rowMarks,
2221 : parse->onConflict,
2222 : mergeActionLists,
2223 : mergeJoinConditions,
2224 : parse->forPortionOf,
2225 : assign_special_exec_param(root));
2226 : }
2227 :
2228 : /* And shove it into final_rel */
2229 414436 : add_path(final_rel, path);
2230 : }
2231 :
2232 : /*
2233 : * Generate partial paths for final_rel, too, if outer query levels might
2234 : * be able to make use of them.
2235 : */
2236 398041 : if (final_rel->consider_parallel && root->query_level > 1 &&
2237 24758 : !limit_needed(parse))
2238 : {
2239 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2240 24697 : foreach(lc, current_rel->partial_pathlist)
2241 : {
2242 112 : Path *partial_path = (Path *) lfirst(lc);
2243 :
2244 112 : add_partial_path(final_rel, partial_path);
2245 : }
2246 : }
2247 :
2248 398041 : extra.limit_needed = limit_needed(parse);
2249 398041 : extra.limit_tuples = limit_tuples;
2250 398041 : extra.count_est = count_est;
2251 398041 : extra.offset_est = offset_est;
2252 :
2253 : /*
2254 : * If there is an FDW that's responsible for all baserels of the query,
2255 : * let it consider adding ForeignPaths.
2256 : */
2257 398041 : if (final_rel->fdwroutine &&
2258 650 : final_rel->fdwroutine->GetForeignUpperPaths)
2259 613 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2260 : current_rel, final_rel,
2261 : &extra);
2262 :
2263 : /* Let extensions possibly add some more paths */
2264 398041 : if (create_upper_paths_hook)
2265 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2266 : current_rel, final_rel, &extra);
2267 :
2268 : /* Note: currently, we leave it to callers to do set_cheapest() */
2269 398041 : }
2270 :
2271 : /*
2272 : * Do preprocessing for groupingSets clause and related data.
2273 : *
2274 : * We expect that parse->groupingSets has already been expanded into a flat
2275 : * list of grouping sets (that is, just integer Lists of ressortgroupref
2276 : * numbers) by expand_grouping_sets(). This function handles the preliminary
2277 : * steps of organizing the grouping sets into lists of rollups, and preparing
2278 : * annotations which will later be filled in with size estimates.
2279 : */
2280 : static grouping_sets_data *
2281 905 : preprocess_grouping_sets(PlannerInfo *root)
2282 : {
2283 905 : Query *parse = root->parse;
2284 : List *sets;
2285 905 : int maxref = 0;
2286 : ListCell *lc_set;
2287 905 : grouping_sets_data *gd = palloc0_object(grouping_sets_data);
2288 :
2289 : /*
2290 : * We don't currently make any attempt to optimize the groupClause when
2291 : * there are grouping sets, so just duplicate it in processed_groupClause.
2292 : */
2293 905 : root->processed_groupClause = parse->groupClause;
2294 :
2295 : /* Detect unhashable and unsortable grouping expressions */
2296 905 : gd->any_hashable = false;
2297 905 : gd->unhashable_refs = NULL;
2298 905 : gd->unsortable_refs = NULL;
2299 905 : gd->unsortable_sets = NIL;
2300 :
2301 905 : if (parse->groupClause)
2302 : {
2303 : ListCell *lc;
2304 :
2305 2651 : foreach(lc, parse->groupClause)
2306 : {
2307 1811 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2308 1811 : Index ref = gc->tleSortGroupRef;
2309 :
2310 1811 : if (ref > maxref)
2311 1771 : maxref = ref;
2312 :
2313 1811 : if (!gc->hashable)
2314 24 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2315 :
2316 1811 : if (!OidIsValid(gc->sortop))
2317 33 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2318 : }
2319 : }
2320 :
2321 : /* Allocate workspace array for remapping */
2322 905 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2323 :
2324 : /*
2325 : * If we have any unsortable sets, we must extract them before trying to
2326 : * prepare rollups. Unsortable sets don't go through
2327 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2328 : * here.
2329 : */
2330 905 : if (!bms_is_empty(gd->unsortable_refs))
2331 : {
2332 33 : List *sortable_sets = NIL;
2333 : ListCell *lc;
2334 :
2335 99 : foreach(lc, parse->groupingSets)
2336 : {
2337 70 : List *gset = (List *) lfirst(lc);
2338 :
2339 70 : if (bms_overlap_list(gd->unsortable_refs, gset))
2340 : {
2341 38 : GroupingSetData *gs = makeNode(GroupingSetData);
2342 :
2343 38 : gs->set = gset;
2344 38 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2345 :
2346 : /*
2347 : * We must enforce here that an unsortable set is hashable;
2348 : * later code assumes this. Parse analysis only checks that
2349 : * every individual column is either hashable or sortable.
2350 : *
2351 : * Note that passing this test doesn't guarantee we can
2352 : * generate a plan; there might be other showstoppers.
2353 : */
2354 38 : if (bms_overlap_list(gd->unhashable_refs, gset))
2355 4 : ereport(ERROR,
2356 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2357 : errmsg("could not implement GROUP BY"),
2358 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2359 : }
2360 : else
2361 32 : sortable_sets = lappend(sortable_sets, gset);
2362 : }
2363 :
2364 29 : if (sortable_sets)
2365 24 : sets = extract_rollup_sets(sortable_sets);
2366 : else
2367 5 : sets = NIL;
2368 : }
2369 : else
2370 872 : sets = extract_rollup_sets(parse->groupingSets);
2371 :
2372 2376 : foreach(lc_set, sets)
2373 : {
2374 1475 : List *current_sets = (List *) lfirst(lc_set);
2375 1475 : RollupData *rollup = makeNode(RollupData);
2376 : GroupingSetData *gs;
2377 :
2378 : /*
2379 : * Reorder the current list of grouping sets into correct prefix
2380 : * order. If only one aggregation pass is needed, try to make the
2381 : * list match the ORDER BY clause; if more than one pass is needed, we
2382 : * don't bother with that.
2383 : *
2384 : * Note that this reorders the sets from smallest-member-first to
2385 : * largest-member-first, and applies the GroupingSetData annotations,
2386 : * though the data will be filled in later.
2387 : */
2388 1475 : current_sets = reorder_grouping_sets(current_sets,
2389 1475 : (list_length(sets) == 1
2390 : ? parse->sortClause
2391 : : NIL));
2392 :
2393 : /*
2394 : * Get the initial (and therefore largest) grouping set.
2395 : */
2396 1475 : gs = linitial_node(GroupingSetData, current_sets);
2397 :
2398 : /*
2399 : * Order the groupClause appropriately. If the first grouping set is
2400 : * empty, then the groupClause must also be empty; otherwise we have
2401 : * to force the groupClause to match that grouping set's order.
2402 : *
2403 : * (The first grouping set can be empty even though parse->groupClause
2404 : * is not empty only if all non-empty grouping sets are unsortable.
2405 : * The groupClauses for hashed grouping sets are built later on.)
2406 : */
2407 1475 : if (gs->set)
2408 1410 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2409 : else
2410 65 : rollup->groupClause = NIL;
2411 :
2412 : /*
2413 : * Is it hashable? We pretend empty sets are hashable even though we
2414 : * actually force them not to be hashed later. But don't bother if
2415 : * there's nothing but empty sets (since in that case we can't hash
2416 : * anything).
2417 : */
2418 1475 : if (gs->set &&
2419 1410 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2420 : {
2421 1390 : rollup->hashable = true;
2422 1390 : gd->any_hashable = true;
2423 : }
2424 :
2425 : /*
2426 : * Now that we've pinned down an order for the groupClause for this
2427 : * list of grouping sets, we need to remap the entries in the grouping
2428 : * sets from sortgrouprefs to plain indices (0-based) into the
2429 : * groupClause for this collection of grouping sets. We keep the
2430 : * original form for later use, though.
2431 : */
2432 1475 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2433 : current_sets,
2434 : gd->tleref_to_colnum_map);
2435 1475 : rollup->gsets_data = current_sets;
2436 :
2437 1475 : gd->rollups = lappend(gd->rollups, rollup);
2438 : }
2439 :
2440 901 : if (gd->unsortable_sets)
2441 : {
2442 : /*
2443 : * We have not yet pinned down a groupclause for this, but we will
2444 : * need index-based lists for estimation purposes. Construct
2445 : * hash_sets_idx based on the entire original groupclause for now.
2446 : */
2447 29 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2448 : gd->unsortable_sets,
2449 : gd->tleref_to_colnum_map);
2450 29 : gd->any_hashable = true;
2451 : }
2452 :
2453 901 : return gd;
2454 : }
2455 :
2456 : /*
2457 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2458 : * (without annotation) mapped to indexes into the given groupclause.
2459 : */
2460 : static List *
2461 4147 : remap_to_groupclause_idx(List *groupClause,
2462 : List *gsets,
2463 : int *tleref_to_colnum_map)
2464 : {
2465 4147 : int ref = 0;
2466 4147 : List *result = NIL;
2467 : ListCell *lc;
2468 :
2469 9861 : foreach(lc, groupClause)
2470 : {
2471 5714 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2472 :
2473 5714 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2474 : }
2475 :
2476 9444 : foreach(lc, gsets)
2477 : {
2478 5297 : List *set = NIL;
2479 : ListCell *lc2;
2480 5297 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2481 :
2482 11720 : foreach(lc2, gs->set)
2483 : {
2484 6423 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2485 : }
2486 :
2487 5297 : result = lappend(result, set);
2488 : }
2489 :
2490 4147 : return result;
2491 : }
2492 :
2493 :
2494 : /*
2495 : * preprocess_rowmarks - set up PlanRowMarks if needed
2496 : */
2497 : static void
2498 400720 : preprocess_rowmarks(PlannerInfo *root)
2499 : {
2500 400720 : Query *parse = root->parse;
2501 : Bitmapset *rels;
2502 : List *prowmarks;
2503 : ListCell *l;
2504 : int i;
2505 :
2506 400720 : if (parse->rowMarks)
2507 : {
2508 : /*
2509 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2510 : * grouping, since grouping renders a reference to individual tuple
2511 : * CTIDs invalid. This is also checked at parse time, but that's
2512 : * insufficient because of rule substitution, query pullup, etc.
2513 : */
2514 6561 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2515 : parse->rowMarks)->strength);
2516 : }
2517 : else
2518 : {
2519 : /*
2520 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2521 : * UPDATE/SHARE.
2522 : */
2523 394159 : if (parse->commandType != CMD_UPDATE &&
2524 383805 : parse->commandType != CMD_DELETE &&
2525 380116 : parse->commandType != CMD_MERGE)
2526 378636 : return;
2527 : }
2528 :
2529 : /*
2530 : * We need to have rowmarks for all base relations except the target. We
2531 : * make a bitmapset of all base rels and then remove the items we don't
2532 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2533 : */
2534 22084 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2535 22084 : if (parse->resultRelation)
2536 15523 : rels = bms_del_member(rels, parse->resultRelation);
2537 :
2538 : /*
2539 : * Convert RowMarkClauses to PlanRowMark representation.
2540 : */
2541 22084 : prowmarks = NIL;
2542 28780 : foreach(l, parse->rowMarks)
2543 : {
2544 6696 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2545 6696 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2546 : PlanRowMark *newrc;
2547 :
2548 : /*
2549 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2550 : * applied to an update/delete target rel. If that ever becomes
2551 : * possible, we should drop the target from the PlanRowMark list.
2552 : */
2553 : Assert(rc->rti != parse->resultRelation);
2554 :
2555 : /*
2556 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2557 : * can't support true locking. Subqueries that got flattened into the
2558 : * main query should be ignored completely. Any that didn't will get
2559 : * ROW_MARK_COPY items in the next loop.
2560 : */
2561 6696 : if (rte->rtekind != RTE_RELATION)
2562 46 : continue;
2563 :
2564 6650 : rels = bms_del_member(rels, rc->rti);
2565 :
2566 6650 : newrc = makeNode(PlanRowMark);
2567 6650 : newrc->rti = newrc->prti = rc->rti;
2568 6650 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2569 6650 : newrc->markType = select_rowmark_type(rte, rc->strength);
2570 6650 : newrc->allMarkTypes = (1 << newrc->markType);
2571 6650 : newrc->strength = rc->strength;
2572 6650 : newrc->waitPolicy = rc->waitPolicy;
2573 6650 : newrc->isParent = false;
2574 :
2575 6650 : prowmarks = lappend(prowmarks, newrc);
2576 : }
2577 :
2578 : /*
2579 : * Now, add rowmarks for any non-target, non-locked base relations.
2580 : */
2581 22084 : i = 0;
2582 52686 : foreach(l, parse->rtable)
2583 : {
2584 30602 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2585 : PlanRowMark *newrc;
2586 :
2587 30602 : i++;
2588 30602 : if (!bms_is_member(i, rels))
2589 27668 : continue;
2590 :
2591 2934 : newrc = makeNode(PlanRowMark);
2592 2934 : newrc->rti = newrc->prti = i;
2593 2934 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2594 2934 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2595 2934 : newrc->allMarkTypes = (1 << newrc->markType);
2596 2934 : newrc->strength = LCS_NONE;
2597 2934 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2598 2934 : newrc->isParent = false;
2599 :
2600 2934 : prowmarks = lappend(prowmarks, newrc);
2601 : }
2602 :
2603 22084 : root->rowMarks = prowmarks;
2604 : }
2605 :
2606 : /*
2607 : * Select RowMarkType to use for a given table
2608 : */
2609 : RowMarkType
2610 11478 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2611 : {
2612 11478 : if (rte->rtekind != RTE_RELATION)
2613 : {
2614 : /* If it's not a table at all, use ROW_MARK_COPY */
2615 1110 : return ROW_MARK_COPY;
2616 : }
2617 10368 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2618 : {
2619 : /* Let the FDW select the rowmark type, if it wants to */
2620 114 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2621 :
2622 114 : if (fdwroutine->GetForeignRowMarkType != NULL)
2623 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2624 : /* Otherwise, use ROW_MARK_COPY by default */
2625 114 : return ROW_MARK_COPY;
2626 : }
2627 : else
2628 : {
2629 : /* Regular table, apply the appropriate lock type */
2630 10254 : switch (strength)
2631 : {
2632 2033 : case LCS_NONE:
2633 :
2634 : /*
2635 : * We don't need a tuple lock, only the ability to re-fetch
2636 : * the row.
2637 : */
2638 2033 : return ROW_MARK_REFERENCE;
2639 : break;
2640 7057 : case LCS_FORKEYSHARE:
2641 7057 : return ROW_MARK_KEYSHARE;
2642 : break;
2643 201 : case LCS_FORSHARE:
2644 201 : return ROW_MARK_SHARE;
2645 : break;
2646 41 : case LCS_FORNOKEYUPDATE:
2647 41 : return ROW_MARK_NOKEYEXCLUSIVE;
2648 : break;
2649 922 : case LCS_FORUPDATE:
2650 922 : return ROW_MARK_EXCLUSIVE;
2651 : break;
2652 : }
2653 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2654 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2655 : }
2656 : }
2657 :
2658 : /*
2659 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2660 : *
2661 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2662 : * results back in *count_est and *offset_est. These variables are set to
2663 : * 0 if the corresponding clause is not present, and -1 if it's present
2664 : * but we couldn't estimate the value for it. (The "0" convention is OK
2665 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2666 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2667 : * usual practice of never estimating less than one row.) These values will
2668 : * be passed to create_limit_path, which see if you change this code.
2669 : *
2670 : * The return value is the suitably adjusted tuple_fraction to use for
2671 : * planning the query. This adjustment is not overridable, since it reflects
2672 : * plan actions that grouping_planner() will certainly take, not assumptions
2673 : * about context.
2674 : */
2675 : static double
2676 3707 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2677 : int64 *offset_est, int64 *count_est)
2678 : {
2679 3707 : Query *parse = root->parse;
2680 : Node *est;
2681 : double limit_fraction;
2682 :
2683 : /* Should not be called unless LIMIT or OFFSET */
2684 : Assert(parse->limitCount || parse->limitOffset);
2685 :
2686 : /*
2687 : * Try to obtain the clause values. We use estimate_expression_value
2688 : * primarily because it can sometimes do something useful with Params.
2689 : */
2690 3707 : if (parse->limitCount)
2691 : {
2692 3265 : est = estimate_expression_value(root, parse->limitCount);
2693 3265 : if (est && IsA(est, Const))
2694 : {
2695 3260 : if (((Const *) est)->constisnull)
2696 : {
2697 : /* NULL indicates LIMIT ALL, ie, no limit */
2698 0 : *count_est = 0; /* treat as not present */
2699 : }
2700 : else
2701 : {
2702 3260 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2703 3260 : if (*count_est <= 0)
2704 190 : *count_est = 1; /* force to at least 1 */
2705 : }
2706 : }
2707 : else
2708 5 : *count_est = -1; /* can't estimate */
2709 : }
2710 : else
2711 442 : *count_est = 0; /* not present */
2712 :
2713 3707 : if (parse->limitOffset)
2714 : {
2715 652 : est = estimate_expression_value(root, parse->limitOffset);
2716 652 : if (est && IsA(est, Const))
2717 : {
2718 632 : if (((Const *) est)->constisnull)
2719 : {
2720 : /* Treat NULL as no offset; the executor will too */
2721 0 : *offset_est = 0; /* treat as not present */
2722 : }
2723 : else
2724 : {
2725 632 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2726 632 : if (*offset_est < 0)
2727 0 : *offset_est = 0; /* treat as not present */
2728 : }
2729 : }
2730 : else
2731 20 : *offset_est = -1; /* can't estimate */
2732 : }
2733 : else
2734 3055 : *offset_est = 0; /* not present */
2735 :
2736 3707 : if (*count_est != 0)
2737 : {
2738 : /*
2739 : * A LIMIT clause limits the absolute number of tuples returned.
2740 : * However, if it's not a constant LIMIT then we have to guess; for
2741 : * lack of a better idea, assume 10% of the plan's result is wanted.
2742 : */
2743 3265 : if (*count_est < 0 || *offset_est < 0)
2744 : {
2745 : /* LIMIT or OFFSET is an expression ... punt ... */
2746 20 : limit_fraction = 0.10;
2747 : }
2748 : else
2749 : {
2750 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2751 3245 : limit_fraction = (double) *count_est + (double) *offset_est;
2752 : }
2753 :
2754 : /*
2755 : * If we have absolute limits from both caller and LIMIT, use the
2756 : * smaller value; likewise if they are both fractional. If one is
2757 : * fractional and the other absolute, we can't easily determine which
2758 : * is smaller, but we use the heuristic that the absolute will usually
2759 : * be smaller.
2760 : */
2761 3265 : if (tuple_fraction >= 1.0)
2762 : {
2763 5 : if (limit_fraction >= 1.0)
2764 : {
2765 : /* both absolute */
2766 5 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2767 : }
2768 : else
2769 : {
2770 : /* caller absolute, limit fractional; use caller's value */
2771 : }
2772 : }
2773 3260 : else if (tuple_fraction > 0.0)
2774 : {
2775 82 : if (limit_fraction >= 1.0)
2776 : {
2777 : /* caller fractional, limit absolute; use limit */
2778 82 : tuple_fraction = limit_fraction;
2779 : }
2780 : else
2781 : {
2782 : /* both fractional */
2783 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2784 : }
2785 : }
2786 : else
2787 : {
2788 : /* no info from caller, just use limit */
2789 3178 : tuple_fraction = limit_fraction;
2790 : }
2791 : }
2792 442 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2793 : {
2794 : /*
2795 : * We have an OFFSET but no LIMIT. This acts entirely differently
2796 : * from the LIMIT case: here, we need to increase rather than decrease
2797 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2798 : * tuples to be fetched instead of fewer. This only matters if we got
2799 : * a tuple_fraction > 0, however.
2800 : *
2801 : * As above, use 10% if OFFSET is present but unestimatable.
2802 : */
2803 12 : if (*offset_est < 0)
2804 0 : limit_fraction = 0.10;
2805 : else
2806 12 : limit_fraction = (double) *offset_est;
2807 :
2808 : /*
2809 : * If we have absolute counts from both caller and OFFSET, add them
2810 : * together; likewise if they are both fractional. If one is
2811 : * fractional and the other absolute, we want to take the larger, and
2812 : * we heuristically assume that's the fractional one.
2813 : */
2814 12 : if (tuple_fraction >= 1.0)
2815 : {
2816 0 : if (limit_fraction >= 1.0)
2817 : {
2818 : /* both absolute, so add them together */
2819 0 : tuple_fraction += limit_fraction;
2820 : }
2821 : else
2822 : {
2823 : /* caller absolute, limit fractional; use limit */
2824 0 : tuple_fraction = limit_fraction;
2825 : }
2826 : }
2827 : else
2828 : {
2829 12 : if (limit_fraction >= 1.0)
2830 : {
2831 : /* caller fractional, limit absolute; use caller's value */
2832 : }
2833 : else
2834 : {
2835 : /* both fractional, so add them together */
2836 0 : tuple_fraction += limit_fraction;
2837 0 : if (tuple_fraction >= 1.0)
2838 0 : tuple_fraction = 0.0; /* assume fetch all */
2839 : }
2840 : }
2841 : }
2842 :
2843 3707 : return tuple_fraction;
2844 : }
2845 :
2846 : /*
2847 : * limit_needed - do we actually need a Limit plan node?
2848 : *
2849 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2850 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2851 : * locution for an optimization fence. (Because other places in the planner
2852 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2853 : * an optimization fence --- we're just suppressing unnecessary run-time
2854 : * overhead.)
2855 : *
2856 : * This might look like it could be merged into preprocess_limit, but there's
2857 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2858 : * in preprocess_limit it's good enough to consider estimated values.
2859 : */
2860 : bool
2861 853551 : limit_needed(Query *parse)
2862 : {
2863 : Node *node;
2864 :
2865 853551 : node = parse->limitCount;
2866 853551 : if (node)
2867 : {
2868 7757 : if (IsA(node, Const))
2869 : {
2870 : /* NULL indicates LIMIT ALL, ie, no limit */
2871 7541 : if (!((Const *) node)->constisnull)
2872 7541 : return true; /* LIMIT with a constant value */
2873 : }
2874 : else
2875 216 : return true; /* non-constant LIMIT */
2876 : }
2877 :
2878 845794 : node = parse->limitOffset;
2879 845794 : if (node)
2880 : {
2881 1283 : if (IsA(node, Const))
2882 : {
2883 : /* Treat NULL as no offset; the executor would too */
2884 1025 : if (!((Const *) node)->constisnull)
2885 : {
2886 1025 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2887 :
2888 1025 : if (offset != 0)
2889 107 : return true; /* OFFSET with a nonzero value */
2890 : }
2891 : }
2892 : else
2893 258 : return true; /* non-constant OFFSET */
2894 : }
2895 :
2896 845429 : return false; /* don't need a Limit plan node */
2897 : }
2898 :
2899 : /*
2900 : * preprocess_groupclause - do preparatory work on GROUP BY clause
2901 : *
2902 : * The idea here is to adjust the ordering of the GROUP BY elements
2903 : * (which in itself is semantically insignificant) to match ORDER BY,
2904 : * thereby allowing a single sort operation to both implement the ORDER BY
2905 : * requirement and set up for a Unique step that implements GROUP BY.
2906 : * We also consider partial match between GROUP BY and ORDER BY elements,
2907 : * which could allow to implement ORDER BY using the incremental sort.
2908 : *
2909 : * We also consider other orderings of the GROUP BY elements, which could
2910 : * match the sort ordering of other possible plans (eg an indexscan) and
2911 : * thereby reduce cost. This is implemented during the generation of grouping
2912 : * paths. See get_useful_group_keys_orderings() for details.
2913 : *
2914 : * Note: we need no comparable processing of the distinctClause because
2915 : * the parser already enforced that that matches ORDER BY.
2916 : *
2917 : * Note: we return a fresh List, but its elements are the same
2918 : * SortGroupClauses appearing in parse->groupClause. This is important
2919 : * because later processing may modify the processed_groupClause list.
2920 : *
2921 : * For grouping sets, the order of items is instead forced to agree with that
2922 : * of the grouping set (and items not in the grouping set are skipped). The
2923 : * work of sorting the order of grouping set elements to match the ORDER BY if
2924 : * possible is done elsewhere.
2925 : */
2926 : static List *
2927 7350 : preprocess_groupclause(PlannerInfo *root, List *force)
2928 : {
2929 7350 : Query *parse = root->parse;
2930 7350 : List *new_groupclause = NIL;
2931 : ListCell *sl;
2932 : ListCell *gl;
2933 :
2934 : /* For grouping sets, we need to force the ordering */
2935 7350 : if (force)
2936 : {
2937 9704 : foreach(sl, force)
2938 : {
2939 5651 : Index ref = lfirst_int(sl);
2940 5651 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2941 :
2942 5651 : new_groupclause = lappend(new_groupclause, cl);
2943 : }
2944 :
2945 4053 : return new_groupclause;
2946 : }
2947 :
2948 : /* If no ORDER BY, nothing useful to do here */
2949 3297 : if (parse->sortClause == NIL)
2950 1859 : return list_copy(parse->groupClause);
2951 :
2952 : /*
2953 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2954 : * items, but only as far as we can make a matching prefix.
2955 : *
2956 : * This code assumes that the sortClause contains no duplicate items.
2957 : */
2958 2823 : foreach(sl, parse->sortClause)
2959 : {
2960 1896 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2961 :
2962 2778 : foreach(gl, parse->groupClause)
2963 : {
2964 2267 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2965 :
2966 2267 : if (equal(gc, sc))
2967 : {
2968 1385 : new_groupclause = lappend(new_groupclause, gc);
2969 1385 : break;
2970 : }
2971 : }
2972 1896 : if (gl == NULL)
2973 511 : break; /* no match, so stop scanning */
2974 : }
2975 :
2976 :
2977 : /* If no match at all, no point in reordering GROUP BY */
2978 1438 : if (new_groupclause == NIL)
2979 223 : return list_copy(parse->groupClause);
2980 :
2981 : /*
2982 : * Add any remaining GROUP BY items to the new list. We don't require a
2983 : * complete match, because even partial match allows ORDER BY to be
2984 : * implemented using incremental sort. Also, give up if there are any
2985 : * non-sortable GROUP BY items, since then there's no hope anyway.
2986 : */
2987 2727 : foreach(gl, parse->groupClause)
2988 : {
2989 1512 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2990 :
2991 1512 : if (list_member_ptr(new_groupclause, gc))
2992 1385 : continue; /* it matched an ORDER BY item */
2993 127 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2994 0 : return list_copy(parse->groupClause);
2995 127 : new_groupclause = lappend(new_groupclause, gc);
2996 : }
2997 :
2998 : /* Success --- install the rearranged GROUP BY list */
2999 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3000 1215 : return new_groupclause;
3001 : }
3002 :
3003 : /*
3004 : * Extract lists of grouping sets that can be implemented using a single
3005 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3006 : *
3007 : * Input must be sorted with smallest sets first. Result has each sublist
3008 : * sorted with smallest sets first.
3009 : *
3010 : * We want to produce the absolute minimum possible number of lists here to
3011 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3012 : * of finding the minimal partition of a partially-ordered set into chains
3013 : * (which is what we need, taking the list of grouping sets as a poset ordered
3014 : * by set inclusion) can be mapped to the problem of finding the maximum
3015 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3016 : * time with a worst case of no worse than O(n^2.5) and usually much
3017 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3018 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3019 : * half a second on my modest system even with optimization off and assertions
3020 : * on.)
3021 : */
3022 : static List *
3023 896 : extract_rollup_sets(List *groupingSets)
3024 : {
3025 896 : int num_sets_raw = list_length(groupingSets);
3026 896 : int num_empty = 0;
3027 896 : int num_sets = 0; /* distinct sets */
3028 896 : int num_chains = 0;
3029 896 : List *result = NIL;
3030 : List **results;
3031 : List **orig_sets;
3032 : Bitmapset **set_masks;
3033 : int *chains;
3034 : short **adjacency;
3035 : short *adjacency_buf;
3036 : BipartiteMatchState *state;
3037 : int i;
3038 : int j;
3039 : int j_size;
3040 896 : ListCell *lc1 = list_head(groupingSets);
3041 : ListCell *lc;
3042 :
3043 : /*
3044 : * Start by stripping out empty sets. The algorithm doesn't require this,
3045 : * but the planner currently needs all empty sets to be returned in the
3046 : * first list, so we strip them here and add them back after.
3047 : */
3048 1486 : while (lc1 && lfirst(lc1) == NIL)
3049 : {
3050 590 : ++num_empty;
3051 590 : lc1 = lnext(groupingSets, lc1);
3052 : }
3053 :
3054 : /* bail out now if it turns out that all we had were empty sets. */
3055 896 : if (!lc1)
3056 65 : return list_make1(groupingSets);
3057 :
3058 : /*----------
3059 : * We don't strictly need to remove duplicate sets here, but if we don't,
3060 : * they tend to become scattered through the result, which is a bit
3061 : * confusing (and irritating if we ever decide to optimize them out).
3062 : * So we remove them here and add them back after.
3063 : *
3064 : * For each non-duplicate set, we fill in the following:
3065 : *
3066 : * orig_sets[i] = list of the original set lists
3067 : * set_masks[i] = bitmapset for testing inclusion
3068 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3069 : *
3070 : * chains[i] will be the result group this set is assigned to.
3071 : *
3072 : * We index all of these from 1 rather than 0 because it is convenient
3073 : * to leave 0 free for the NIL node in the graph algorithm.
3074 : *----------
3075 : */
3076 831 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3077 831 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3078 831 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3079 831 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3080 :
3081 831 : j_size = 0;
3082 831 : j = 0;
3083 831 : i = 1;
3084 :
3085 2861 : for_each_cell(lc, groupingSets, lc1)
3086 : {
3087 2030 : List *candidate = (List *) lfirst(lc);
3088 2030 : Bitmapset *candidate_set = NULL;
3089 : ListCell *lc2;
3090 2030 : int dup_of = 0;
3091 :
3092 4827 : foreach(lc2, candidate)
3093 : {
3094 2797 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3095 : }
3096 :
3097 : /* we can only be a dup if we're the same length as a previous set */
3098 2030 : if (j_size == list_length(candidate))
3099 : {
3100 : int k;
3101 :
3102 1828 : for (k = j; k < i; ++k)
3103 : {
3104 1169 : if (bms_equal(set_masks[k], candidate_set))
3105 : {
3106 129 : dup_of = k;
3107 129 : break;
3108 : }
3109 : }
3110 : }
3111 1242 : else if (j_size < list_length(candidate))
3112 : {
3113 1242 : j_size = list_length(candidate);
3114 1242 : j = i;
3115 : }
3116 :
3117 2030 : if (dup_of > 0)
3118 : {
3119 129 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3120 129 : bms_free(candidate_set);
3121 : }
3122 : else
3123 : {
3124 : int k;
3125 1901 : int n_adj = 0;
3126 :
3127 1901 : orig_sets[i] = list_make1(candidate);
3128 1901 : set_masks[i] = candidate_set;
3129 :
3130 : /* fill in adjacency list; no need to compare equal-size sets */
3131 :
3132 2973 : for (k = j - 1; k > 0; --k)
3133 : {
3134 1072 : if (bms_is_subset(set_masks[k], candidate_set))
3135 937 : adjacency_buf[++n_adj] = k;
3136 : }
3137 :
3138 1901 : if (n_adj > 0)
3139 : {
3140 513 : adjacency_buf[0] = n_adj;
3141 513 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3142 513 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3143 : }
3144 : else
3145 1388 : adjacency[i] = NULL;
3146 :
3147 1901 : ++i;
3148 : }
3149 : }
3150 :
3151 831 : num_sets = i - 1;
3152 :
3153 : /*
3154 : * Apply the graph matching algorithm to do the work.
3155 : */
3156 831 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3157 :
3158 : /*
3159 : * Now, the state->pair* fields have the info we need to assign sets to
3160 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3161 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3162 : * it in one pass)
3163 : */
3164 831 : chains = palloc0((num_sets + 1) * sizeof(int));
3165 :
3166 2732 : for (i = 1; i <= num_sets; ++i)
3167 : {
3168 1901 : int u = state->pair_vu[i];
3169 1901 : int v = state->pair_uv[i];
3170 :
3171 1901 : if (u > 0 && u < i)
3172 0 : chains[i] = chains[u];
3173 1901 : else if (v > 0 && v < i)
3174 491 : chains[i] = chains[v];
3175 : else
3176 1410 : chains[i] = ++num_chains;
3177 : }
3178 :
3179 : /* build result lists. */
3180 831 : results = palloc0((num_chains + 1) * sizeof(List *));
3181 :
3182 2732 : for (i = 1; i <= num_sets; ++i)
3183 : {
3184 1901 : int c = chains[i];
3185 :
3186 : Assert(c > 0);
3187 :
3188 1901 : results[c] = list_concat(results[c], orig_sets[i]);
3189 : }
3190 :
3191 : /* push any empty sets back on the first list. */
3192 1306 : while (num_empty-- > 0)
3193 475 : results[1] = lcons(NIL, results[1]);
3194 :
3195 : /* make result list */
3196 2241 : for (i = 1; i <= num_chains; ++i)
3197 1410 : result = lappend(result, results[i]);
3198 :
3199 : /*
3200 : * Free all the things.
3201 : *
3202 : * (This is over-fussy for small sets but for large sets we could have
3203 : * tied up a nontrivial amount of memory.)
3204 : */
3205 831 : BipartiteMatchFree(state);
3206 831 : pfree(results);
3207 831 : pfree(chains);
3208 2732 : for (i = 1; i <= num_sets; ++i)
3209 1901 : if (adjacency[i])
3210 513 : pfree(adjacency[i]);
3211 831 : pfree(adjacency);
3212 831 : pfree(adjacency_buf);
3213 831 : pfree(orig_sets);
3214 2732 : for (i = 1; i <= num_sets; ++i)
3215 1901 : bms_free(set_masks[i]);
3216 831 : pfree(set_masks);
3217 :
3218 831 : return result;
3219 : }
3220 :
3221 : /*
3222 : * Reorder the elements of a list of grouping sets such that they have correct
3223 : * prefix relationships. Also inserts the GroupingSetData annotations.
3224 : *
3225 : * The input must be ordered with smallest sets first; the result is returned
3226 : * with largest sets first. Note that the result shares no list substructure
3227 : * with the input, so it's safe for the caller to modify it later.
3228 : *
3229 : * If we're passed in a sortclause, we follow its order of columns to the
3230 : * extent possible, to minimize the chance that we add unnecessary sorts.
3231 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3232 : * gets implemented in one pass.)
3233 : */
3234 : static List *
3235 1475 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3236 : {
3237 : ListCell *lc;
3238 1475 : List *previous = NIL;
3239 1475 : List *result = NIL;
3240 :
3241 4095 : foreach(lc, groupingSets)
3242 : {
3243 2620 : List *candidate = (List *) lfirst(lc);
3244 2620 : List *new_elems = list_difference_int(candidate, previous);
3245 2620 : GroupingSetData *gs = makeNode(GroupingSetData);
3246 :
3247 2764 : while (list_length(sortclause) > list_length(previous) &&
3248 : new_elems != NIL)
3249 : {
3250 244 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3251 244 : int ref = sc->tleSortGroupRef;
3252 :
3253 244 : if (list_member_int(new_elems, ref))
3254 : {
3255 144 : previous = lappend_int(previous, ref);
3256 144 : new_elems = list_delete_int(new_elems, ref);
3257 : }
3258 : else
3259 : {
3260 : /* diverged from the sortclause; give up on it */
3261 100 : sortclause = NIL;
3262 100 : break;
3263 : }
3264 : }
3265 :
3266 2620 : previous = list_concat(previous, new_elems);
3267 :
3268 2620 : gs->set = list_copy(previous);
3269 2620 : result = lcons(gs, result);
3270 : }
3271 :
3272 1475 : list_free(previous);
3273 :
3274 1475 : return result;
3275 : }
3276 :
3277 : /*
3278 : * has_volatile_pathkey
3279 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3280 : * containing a volatile function. Otherwise returns false.
3281 : */
3282 : static bool
3283 2200 : has_volatile_pathkey(List *keys)
3284 : {
3285 : ListCell *lc;
3286 :
3287 4522 : foreach(lc, keys)
3288 : {
3289 2335 : PathKey *pathkey = lfirst_node(PathKey, lc);
3290 :
3291 2335 : if (pathkey->pk_eclass->ec_has_volatile)
3292 13 : return true;
3293 : }
3294 :
3295 2187 : return false;
3296 : }
3297 :
3298 : /*
3299 : * adjust_group_pathkeys_for_groupagg
3300 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3301 : * pre-ordered input for ordered aggregates.
3302 : *
3303 : * We define "best" as the pathkeys that suit the largest number of
3304 : * aggregate functions. We find these by looking at the first ORDER BY /
3305 : * DISTINCT aggregate and take the pathkeys for that before searching for
3306 : * other aggregates that require the same or a more strict variation of the
3307 : * same pathkeys. We then repeat that process for any remaining aggregates
3308 : * with different pathkeys and if we find another set of pathkeys that suits a
3309 : * larger number of aggregates then we select those pathkeys instead.
3310 : *
3311 : * When the best pathkeys are found we also mark each Aggref that can use
3312 : * those pathkeys as aggpresorted = true.
3313 : *
3314 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3315 : * volatile functions, we never make use of these pathkeys. We want to ensure
3316 : * that sorts using volatile functions are done independently in each Aggref
3317 : * rather than once at the query level. If we were to allow this then Aggrefs
3318 : * with compatible sort orders would all transition their rows in the same
3319 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3320 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3321 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3322 : * left to perform their sorts individually. To avoid this inconsistent
3323 : * behavior which could make Aggref results depend on what other Aggrefs the
3324 : * query contains, we always force Aggrefs with volatile functions to perform
3325 : * their own sorts.
3326 : */
3327 : static void
3328 1864 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3329 : {
3330 1864 : List *grouppathkeys = root->group_pathkeys;
3331 : List *bestpathkeys;
3332 : Bitmapset *bestaggs;
3333 : Bitmapset *unprocessed_aggs;
3334 : ListCell *lc;
3335 : int i;
3336 :
3337 : /* Shouldn't be here if there are grouping sets */
3338 : Assert(root->parse->groupingSets == NIL);
3339 : /* Shouldn't be here unless there are some ordered aggregates */
3340 : Assert(root->numOrderedAggs > 0);
3341 :
3342 : /* Do nothing if disabled */
3343 1864 : if (!enable_presorted_aggregate)
3344 5 : return;
3345 :
3346 : /*
3347 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3348 : * the indexes of all AggInfos to be processed below.
3349 : */
3350 1859 : unprocessed_aggs = NULL;
3351 4284 : foreach(lc, root->agginfos)
3352 : {
3353 2425 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3354 2425 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3355 :
3356 2425 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3357 214 : continue;
3358 :
3359 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3360 2211 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3361 248 : continue;
3362 :
3363 : /* Additional safety checks are needed if there's a FILTER clause */
3364 1963 : if (aggref->aggfilter != NULL)
3365 : {
3366 : ListCell *lc2;
3367 43 : bool allow_presort = true;
3368 :
3369 : /*
3370 : * When the Aggref has a FILTER clause, it's possible that the
3371 : * filter removes rows that cannot be sorted because the
3372 : * expression to sort by results in an error during its
3373 : * evaluation. This is a problem for presorting as that happens
3374 : * before the FILTER, whereas without presorting, the Aggregate
3375 : * node will apply the FILTER *before* sorting. So that we never
3376 : * try to sort anything that might error, here we aim to skip over
3377 : * any Aggrefs with arguments with expressions which, when
3378 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3379 : * may be more cases that should be allowed, but more thought
3380 : * needs to be given. Err on the side of caution.
3381 : */
3382 83 : foreach(lc2, aggref->args)
3383 : {
3384 58 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3385 58 : Expr *expr = tle->expr;
3386 :
3387 68 : while (IsA(expr, RelabelType))
3388 10 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3389 :
3390 : /* Common case, Vars and Consts are ok */
3391 58 : if (IsA(expr, Var) || IsA(expr, Const))
3392 40 : continue;
3393 :
3394 : /* Unsupported. Don't try to presort for this Aggref */
3395 18 : allow_presort = false;
3396 18 : break;
3397 : }
3398 :
3399 : /* Skip unsupported Aggrefs */
3400 43 : if (!allow_presort)
3401 18 : continue;
3402 : }
3403 :
3404 1945 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3405 : foreach_current_index(lc));
3406 : }
3407 :
3408 : /*
3409 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3410 : * for the given set of aggregates.
3411 : *
3412 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3413 : * this during the first loop using the pathkeys for the very first
3414 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3415 : * a more strict set of compatible pathkeys. Once the outer loop is
3416 : * complete, we mark off all the aggregates with compatible pathkeys then
3417 : * remove those from the unprocessed_aggs and repeat the process to try to
3418 : * find another set of pathkeys that are suitable for a larger number of
3419 : * aggregates. The outer loop will stop when there are not enough
3420 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3421 : * to suit a larger number of aggregates.
3422 : */
3423 1859 : bestpathkeys = NIL;
3424 1859 : bestaggs = NULL;
3425 3669 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3426 : {
3427 1810 : Bitmapset *aggindexes = NULL;
3428 1810 : List *currpathkeys = NIL;
3429 :
3430 1810 : i = -1;
3431 4010 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3432 : {
3433 2200 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3434 2200 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3435 : List *sortlist;
3436 : List *pathkeys;
3437 :
3438 2200 : if (aggref->aggdistinct != NIL)
3439 597 : sortlist = aggref->aggdistinct;
3440 : else
3441 1603 : sortlist = aggref->aggorder;
3442 :
3443 2200 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3444 : aggref->args);
3445 :
3446 : /*
3447 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3448 : * or DISTINCT clause.
3449 : */
3450 2200 : if (has_volatile_pathkey(pathkeys))
3451 : {
3452 13 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3453 13 : continue;
3454 : }
3455 :
3456 : /*
3457 : * When not set yet, take the pathkeys from the first unprocessed
3458 : * aggregate.
3459 : */
3460 2187 : if (currpathkeys == NIL)
3461 : {
3462 1807 : currpathkeys = pathkeys;
3463 :
3464 : /* include the GROUP BY pathkeys, if they exist */
3465 1807 : if (grouppathkeys != NIL)
3466 221 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3467 : currpathkeys);
3468 :
3469 : /* record that we found pathkeys for this aggregate */
3470 1807 : aggindexes = bms_add_member(aggindexes, i);
3471 : }
3472 : else
3473 : {
3474 : /* now look for a stronger set of matching pathkeys */
3475 :
3476 : /* include the GROUP BY pathkeys, if they exist */
3477 380 : if (grouppathkeys != NIL)
3478 240 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3479 : pathkeys);
3480 :
3481 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3482 380 : switch (compare_pathkeys(currpathkeys, pathkeys))
3483 : {
3484 10 : case PATHKEYS_BETTER2:
3485 : /* 'pathkeys' are stronger, use these ones instead */
3486 10 : currpathkeys = pathkeys;
3487 : pg_fallthrough;
3488 :
3489 55 : case PATHKEYS_BETTER1:
3490 : /* 'pathkeys' are less strict */
3491 : pg_fallthrough;
3492 :
3493 : case PATHKEYS_EQUAL:
3494 : /* mark this aggregate as covered by 'currpathkeys' */
3495 55 : aggindexes = bms_add_member(aggindexes, i);
3496 55 : break;
3497 :
3498 325 : case PATHKEYS_DIFFERENT:
3499 325 : break;
3500 : }
3501 : }
3502 : }
3503 :
3504 : /* remove the aggregates that we've just processed */
3505 1810 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3506 :
3507 : /*
3508 : * If this pass included more aggregates than the previous best then
3509 : * use these ones as the best set.
3510 : */
3511 1810 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3512 : {
3513 1722 : bestaggs = aggindexes;
3514 1722 : bestpathkeys = currpathkeys;
3515 : }
3516 : }
3517 :
3518 : /*
3519 : * If we found any ordered aggregates, update root->group_pathkeys to add
3520 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3521 : * the original GROUP BY pathkeys already.
3522 : */
3523 1859 : if (bestpathkeys != NIL)
3524 1672 : root->group_pathkeys = bestpathkeys;
3525 :
3526 : /*
3527 : * Now that we've found the best set of aggregates we can set the
3528 : * presorted flag to indicate to the executor that it needn't bother
3529 : * performing a sort for these Aggrefs. We're able to do this now as
3530 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3531 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3532 : * of ordered aggregates.
3533 : */
3534 1859 : i = -1;
3535 3611 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3536 : {
3537 1752 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3538 :
3539 3519 : foreach(lc, agginfo->aggrefs)
3540 : {
3541 1767 : Aggref *aggref = lfirst_node(Aggref, lc);
3542 :
3543 1767 : aggref->aggpresorted = true;
3544 : }
3545 : }
3546 : }
3547 :
3548 : /*
3549 : * Compute query_pathkeys and other pathkeys during plan generation
3550 : */
3551 : static void
3552 393126 : standard_qp_callback(PlannerInfo *root, void *extra)
3553 : {
3554 393126 : Query *parse = root->parse;
3555 393126 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3556 393126 : List *tlist = root->processed_tlist;
3557 393126 : List *activeWindows = qp_extra->activeWindows;
3558 :
3559 : /*
3560 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3561 : * aggregate requirements.
3562 : */
3563 393126 : if (qp_extra->gset_data)
3564 : {
3565 : /*
3566 : * With grouping sets, just use the first RollupData's groupClause. We
3567 : * don't make any effort to optimize grouping clauses when there are
3568 : * grouping sets, nor can we combine aggregate ordering keys with
3569 : * grouping.
3570 : */
3571 901 : List *rollups = qp_extra->gset_data->rollups;
3572 901 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3573 :
3574 901 : if (grouping_is_sortable(groupClause))
3575 : {
3576 : bool sortable;
3577 :
3578 : /*
3579 : * The groupClause is logically below the grouping step. So if
3580 : * there is an RTE entry for the grouping step, we need to remove
3581 : * its RT index from the sort expressions before we make PathKeys
3582 : * for them.
3583 : */
3584 901 : root->group_pathkeys =
3585 901 : make_pathkeys_for_sortclauses_extended(root,
3586 : &groupClause,
3587 : tlist,
3588 : false,
3589 901 : parse->hasGroupRTE,
3590 : &sortable,
3591 : false);
3592 : Assert(sortable);
3593 901 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3594 : }
3595 : else
3596 : {
3597 0 : root->group_pathkeys = NIL;
3598 0 : root->num_groupby_pathkeys = 0;
3599 : }
3600 : }
3601 392225 : else if (parse->groupClause || root->numOrderedAggs > 0)
3602 4978 : {
3603 : /*
3604 : * With a plain GROUP BY list, we can remove any grouping items that
3605 : * are proven redundant by EquivalenceClass processing. For example,
3606 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3607 : * especially common cases, but they're nearly free to detect. Note
3608 : * that we remove redundant items from processed_groupClause but not
3609 : * the original parse->groupClause.
3610 : */
3611 : bool sortable;
3612 :
3613 : /*
3614 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3615 : * EquivalenceClass'es if it's not set yet.
3616 : */
3617 4978 : root->group_pathkeys =
3618 4978 : make_pathkeys_for_sortclauses_extended(root,
3619 : &root->processed_groupClause,
3620 : tlist,
3621 : true,
3622 : false,
3623 : &sortable,
3624 : true);
3625 4978 : if (!sortable)
3626 : {
3627 : /* Can't sort; no point in considering aggregate ordering either */
3628 0 : root->group_pathkeys = NIL;
3629 0 : root->num_groupby_pathkeys = 0;
3630 : }
3631 : else
3632 : {
3633 4978 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3634 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3635 4978 : if (root->numOrderedAggs > 0)
3636 1864 : adjust_group_pathkeys_for_groupagg(root);
3637 : }
3638 : }
3639 : else
3640 : {
3641 387247 : root->group_pathkeys = NIL;
3642 387247 : root->num_groupby_pathkeys = 0;
3643 : }
3644 :
3645 : /* We consider only the first (bottom) window in pathkeys logic */
3646 393126 : if (activeWindows != NIL)
3647 : {
3648 2217 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3649 :
3650 2217 : root->window_pathkeys = make_pathkeys_for_window(root,
3651 : wc,
3652 : tlist);
3653 : }
3654 : else
3655 390909 : root->window_pathkeys = NIL;
3656 :
3657 : /*
3658 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3659 : * redundant by EquivalenceClass processing. The non-redundant list is
3660 : * kept in root->processed_distinctClause, leaving the original
3661 : * parse->distinctClause alone.
3662 : */
3663 393126 : if (parse->distinctClause)
3664 : {
3665 : bool sortable;
3666 :
3667 : /* Make a copy since pathkey processing can modify the list */
3668 1941 : root->processed_distinctClause = list_copy(parse->distinctClause);
3669 1941 : root->distinct_pathkeys =
3670 1941 : make_pathkeys_for_sortclauses_extended(root,
3671 : &root->processed_distinctClause,
3672 : tlist,
3673 : true,
3674 : false,
3675 : &sortable,
3676 : false);
3677 1941 : if (!sortable)
3678 5 : root->distinct_pathkeys = NIL;
3679 : }
3680 : else
3681 391185 : root->distinct_pathkeys = NIL;
3682 :
3683 393126 : root->sort_pathkeys =
3684 393126 : make_pathkeys_for_sortclauses(root,
3685 : parse->sortClause,
3686 : tlist);
3687 :
3688 : /* setting setop_pathkeys might be useful to the union planner */
3689 393126 : if (qp_extra->setop != NULL)
3690 : {
3691 : List *groupClauses;
3692 : bool sortable;
3693 :
3694 10463 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3695 :
3696 10463 : root->setop_pathkeys =
3697 10463 : make_pathkeys_for_sortclauses_extended(root,
3698 : &groupClauses,
3699 : tlist,
3700 : false,
3701 : false,
3702 : &sortable,
3703 : false);
3704 10463 : if (!sortable)
3705 120 : root->setop_pathkeys = NIL;
3706 : }
3707 : else
3708 382663 : root->setop_pathkeys = NIL;
3709 :
3710 : /*
3711 : * Figure out whether we want a sorted result from query_planner.
3712 : *
3713 : * If we have a sortable GROUP BY clause, then we want a result sorted
3714 : * properly for grouping. Otherwise, if we have window functions to
3715 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3716 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3717 : * we try to produce output that's sufficiently well sorted for the
3718 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3719 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3720 : * for a set operation which can benefit from presorted results and have a
3721 : * sortable targetlist, we want to sort by the target list.
3722 : *
3723 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3724 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3725 : * that might just leave us failing to exploit an available sort order at
3726 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3727 : * much easier, since we know that the parser ensured that one is a
3728 : * superset of the other.
3729 : */
3730 393126 : if (root->group_pathkeys)
3731 5522 : root->query_pathkeys = root->group_pathkeys;
3732 387604 : else if (root->window_pathkeys)
3733 1772 : root->query_pathkeys = root->window_pathkeys;
3734 771664 : else if (list_length(root->distinct_pathkeys) >
3735 385832 : list_length(root->sort_pathkeys))
3736 1571 : root->query_pathkeys = root->distinct_pathkeys;
3737 384261 : else if (root->sort_pathkeys)
3738 56474 : root->query_pathkeys = root->sort_pathkeys;
3739 327787 : else if (root->setop_pathkeys != NIL)
3740 9475 : root->query_pathkeys = root->setop_pathkeys;
3741 : else
3742 318312 : root->query_pathkeys = NIL;
3743 393126 : }
3744 :
3745 : /*
3746 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3747 : *
3748 : * path_rows: number of output rows from scan/join step
3749 : * gd: grouping sets data including list of grouping sets and their clauses
3750 : * target_list: target list containing group clause references
3751 : *
3752 : * If doing grouping sets, we also annotate the gsets data with the estimates
3753 : * for each set and each individual rollup list, with a view to later
3754 : * determining whether some combination of them could be hashed instead.
3755 : */
3756 : static double
3757 40842 : get_number_of_groups(PlannerInfo *root,
3758 : double path_rows,
3759 : grouping_sets_data *gd,
3760 : List *target_list)
3761 : {
3762 40842 : Query *parse = root->parse;
3763 : double dNumGroups;
3764 :
3765 40842 : if (parse->groupClause)
3766 : {
3767 : List *groupExprs;
3768 :
3769 8544 : if (parse->groupingSets)
3770 : {
3771 : /* Add up the estimates for each grouping set */
3772 : ListCell *lc;
3773 :
3774 : Assert(gd); /* keep Coverity happy */
3775 :
3776 836 : dNumGroups = 0;
3777 :
3778 2246 : foreach(lc, gd->rollups)
3779 : {
3780 1410 : RollupData *rollup = lfirst_node(RollupData, lc);
3781 : ListCell *lc2;
3782 : ListCell *lc3;
3783 :
3784 1410 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3785 : target_list);
3786 :
3787 1410 : rollup->numGroups = 0.0;
3788 :
3789 3915 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3790 : {
3791 2505 : List *gset = (List *) lfirst(lc2);
3792 2505 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3793 2505 : double numGroups = estimate_num_groups(root,
3794 : groupExprs,
3795 : path_rows,
3796 : &gset,
3797 : NULL);
3798 :
3799 2505 : gs->numGroups = numGroups;
3800 2505 : rollup->numGroups += numGroups;
3801 : }
3802 :
3803 1410 : dNumGroups += rollup->numGroups;
3804 : }
3805 :
3806 836 : if (gd->hash_sets_idx)
3807 : {
3808 : ListCell *lc2;
3809 :
3810 29 : gd->dNumHashGroups = 0;
3811 :
3812 29 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3813 : target_list);
3814 :
3815 63 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3816 : {
3817 34 : List *gset = (List *) lfirst(lc);
3818 34 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3819 34 : double numGroups = estimate_num_groups(root,
3820 : groupExprs,
3821 : path_rows,
3822 : &gset,
3823 : NULL);
3824 :
3825 34 : gs->numGroups = numGroups;
3826 34 : gd->dNumHashGroups += numGroups;
3827 : }
3828 :
3829 29 : dNumGroups += gd->dNumHashGroups;
3830 : }
3831 : }
3832 : else
3833 : {
3834 : /* Plain GROUP BY -- estimate based on optimized groupClause */
3835 7708 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3836 : target_list);
3837 :
3838 7708 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3839 : NULL, NULL);
3840 : }
3841 : }
3842 32298 : else if (parse->groupingSets)
3843 : {
3844 : /* Empty grouping sets ... one result row for each one */
3845 50 : dNumGroups = list_length(parse->groupingSets);
3846 : }
3847 32248 : else if (parse->hasAggs || root->hasHavingQual)
3848 : {
3849 : /* Plain aggregation, one result row */
3850 32248 : dNumGroups = 1;
3851 : }
3852 : else
3853 : {
3854 : /* Not grouping */
3855 0 : dNumGroups = 1;
3856 : }
3857 :
3858 40842 : return dNumGroups;
3859 : }
3860 :
3861 : /*
3862 : * create_grouping_paths
3863 : *
3864 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3865 : * Along the way, we also build an upperrel for Paths which are partially
3866 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3867 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3868 : * the only partially grouped paths we build are also partial paths; that
3869 : * is, they need a Gather and then a FinalizeAggregate.
3870 : *
3871 : * input_rel: contains the source-data Paths
3872 : * target: the pathtarget for the result Paths to compute
3873 : * gd: grouping sets data including list of grouping sets and their clauses
3874 : *
3875 : * Note: all Paths in input_rel are expected to return the target computed
3876 : * by make_group_input_target.
3877 : */
3878 : static RelOptInfo *
3879 33840 : create_grouping_paths(PlannerInfo *root,
3880 : RelOptInfo *input_rel,
3881 : PathTarget *target,
3882 : bool target_parallel_safe,
3883 : grouping_sets_data *gd)
3884 : {
3885 33840 : Query *parse = root->parse;
3886 : RelOptInfo *grouped_rel;
3887 : RelOptInfo *partially_grouped_rel;
3888 : AggClauseCosts agg_costs;
3889 :
3890 203040 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3891 33840 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3892 :
3893 : /*
3894 : * Create grouping relation to hold fully aggregated grouping and/or
3895 : * aggregation paths.
3896 : */
3897 33840 : grouped_rel = make_grouping_rel(root, input_rel, target,
3898 : target_parallel_safe, parse->havingQual);
3899 :
3900 : /*
3901 : * Create either paths for a degenerate grouping or paths for ordinary
3902 : * grouping, as appropriate.
3903 : */
3904 33840 : if (is_degenerate_grouping(root))
3905 35 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3906 : else
3907 : {
3908 33805 : int flags = 0;
3909 : GroupPathExtraData extra;
3910 :
3911 : /*
3912 : * Determine whether it's possible to perform sort-based
3913 : * implementations of grouping. (Note that if processed_groupClause
3914 : * is empty, grouping_is_sortable() is trivially true, and all the
3915 : * pathkeys_contained_in() tests will succeed too, so that we'll
3916 : * consider every surviving input path.)
3917 : *
3918 : * If we have grouping sets, we might be able to sort some but not all
3919 : * of them; in this case, we need can_sort to be true as long as we
3920 : * must consider any sorted-input plan.
3921 : */
3922 33805 : if ((gd && gd->rollups != NIL)
3923 32924 : || grouping_is_sortable(root->processed_groupClause))
3924 33800 : flags |= GROUPING_CAN_USE_SORT;
3925 :
3926 : /*
3927 : * Determine whether we should consider hash-based implementations of
3928 : * grouping.
3929 : *
3930 : * Hashed aggregation only applies if we're grouping. If we have
3931 : * grouping sets, some groups might be hashable but others not; in
3932 : * this case we set can_hash true as long as there is nothing globally
3933 : * preventing us from hashing (and we should therefore consider plans
3934 : * with hashes).
3935 : *
3936 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3937 : * BY aggregates. (Doing so would imply storing *all* the input
3938 : * values in the hash table, and/or running many sorts in parallel,
3939 : * either of which seems like a certain loser.) We similarly don't
3940 : * support ordered-set aggregates in hashed aggregation, but that case
3941 : * is also included in the numOrderedAggs count.
3942 : *
3943 : * Note: grouping_is_hashable() is much more expensive to check than
3944 : * the other gating conditions, so we want to do it last.
3945 : */
3946 33805 : if ((parse->groupClause != NIL &&
3947 7247 : root->numOrderedAggs == 0 &&
3948 3114 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3949 3918 : flags |= GROUPING_CAN_USE_HASH;
3950 :
3951 : /*
3952 : * Determine whether partial aggregation is possible.
3953 : */
3954 33805 : if (can_partial_agg(root))
3955 29734 : flags |= GROUPING_CAN_PARTIAL_AGG;
3956 :
3957 33805 : extra.flags = flags;
3958 33805 : extra.target_parallel_safe = target_parallel_safe;
3959 33805 : extra.havingQual = parse->havingQual;
3960 33805 : extra.targetList = parse->targetList;
3961 33805 : extra.partial_costs_set = false;
3962 :
3963 : /*
3964 : * Determine whether partitionwise aggregation is in theory possible.
3965 : * It can be disabled by the user, and for now, we don't try to
3966 : * support grouping sets. create_ordinary_grouping_paths() will check
3967 : * additional conditions, such as whether input_rel is partitioned.
3968 : */
3969 33805 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3970 580 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3971 : else
3972 33225 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3973 :
3974 33805 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3975 : &agg_costs, gd, &extra,
3976 : &partially_grouped_rel);
3977 : }
3978 :
3979 33836 : set_cheapest(grouped_rel);
3980 33836 : return grouped_rel;
3981 : }
3982 :
3983 : /*
3984 : * make_grouping_rel
3985 : *
3986 : * Create a new grouping rel and set basic properties.
3987 : *
3988 : * input_rel represents the underlying scan/join relation.
3989 : * target is the output expected from the grouping relation.
3990 : */
3991 : static RelOptInfo *
3992 35635 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3993 : PathTarget *target, bool target_parallel_safe,
3994 : Node *havingQual)
3995 : {
3996 : RelOptInfo *grouped_rel;
3997 :
3998 35635 : if (IS_OTHER_REL(input_rel))
3999 : {
4000 1795 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
4001 : input_rel->relids);
4002 1795 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
4003 : }
4004 : else
4005 : {
4006 : /*
4007 : * By tradition, the relids set for the main grouping relation is
4008 : * NULL. (This could be changed, but might require adjustments
4009 : * elsewhere.)
4010 : */
4011 33840 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
4012 : }
4013 :
4014 : /* Set target. */
4015 35635 : grouped_rel->reltarget = target;
4016 :
4017 : /*
4018 : * If the input relation is not parallel-safe, then the grouped relation
4019 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4020 : * target list and HAVING quals are parallel-safe.
4021 : */
4022 58353 : if (input_rel->consider_parallel && target_parallel_safe &&
4023 22718 : is_parallel_safe(root, havingQual))
4024 22698 : grouped_rel->consider_parallel = true;
4025 :
4026 : /* Assume that the same path generation strategies are allowed */
4027 35635 : grouped_rel->pgs_mask = input_rel->pgs_mask;
4028 :
4029 : /*
4030 : * If the input rel belongs to a single FDW, so does the grouped rel.
4031 : */
4032 35635 : grouped_rel->serverid = input_rel->serverid;
4033 35635 : grouped_rel->userid = input_rel->userid;
4034 35635 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
4035 35635 : grouped_rel->fdwroutine = input_rel->fdwroutine;
4036 :
4037 35635 : return grouped_rel;
4038 : }
4039 :
4040 : /*
4041 : * is_degenerate_grouping
4042 : *
4043 : * A degenerate grouping is one in which the query has a HAVING qual and/or
4044 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
4045 : * grouping sets are all empty).
4046 : */
4047 : static bool
4048 33840 : is_degenerate_grouping(PlannerInfo *root)
4049 : {
4050 33840 : Query *parse = root->parse;
4051 :
4052 32849 : return (root->hasHavingQual || parse->groupingSets) &&
4053 66689 : !parse->hasAggs && parse->groupClause == NIL;
4054 : }
4055 :
4056 : /*
4057 : * create_degenerate_grouping_paths
4058 : *
4059 : * When the grouping is degenerate (see is_degenerate_grouping), we are
4060 : * supposed to emit either zero or one row for each grouping set depending on
4061 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
4062 : * either HAVING or the targetlist, so we actually do not need the FROM table
4063 : * at all! We can just throw away the plan-so-far and generate a Result node.
4064 : * This is a sufficiently unusual corner case that it's not worth contorting
4065 : * the structure of this module to avoid having to generate the earlier paths
4066 : * in the first place.
4067 : */
4068 : static void
4069 35 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4070 : RelOptInfo *grouped_rel)
4071 : {
4072 35 : Query *parse = root->parse;
4073 : int nrows;
4074 : Path *path;
4075 :
4076 35 : nrows = list_length(parse->groupingSets);
4077 35 : if (nrows > 1)
4078 : {
4079 : /*
4080 : * Doesn't seem worthwhile writing code to cons up a generate_series
4081 : * or a values scan to emit multiple rows. Instead just make N clones
4082 : * and append them. (With a volatile HAVING clause, this means you
4083 : * might get between 0 and N output rows. Offhand I think that's
4084 : * desired.)
4085 : */
4086 10 : AppendPathInput append = {0};
4087 :
4088 30 : while (--nrows >= 0)
4089 : {
4090 : path = (Path *)
4091 20 : create_group_result_path(root, grouped_rel,
4092 20 : grouped_rel->reltarget,
4093 20 : (List *) parse->havingQual);
4094 20 : append.subpaths = lappend(append.subpaths, path);
4095 : }
4096 : path = (Path *)
4097 10 : create_append_path(root,
4098 : grouped_rel,
4099 : append,
4100 : NIL,
4101 : NULL,
4102 : 0,
4103 : false,
4104 : -1);
4105 : }
4106 : else
4107 : {
4108 : /* No grouping sets, or just one, so one output row */
4109 : path = (Path *)
4110 25 : create_group_result_path(root, grouped_rel,
4111 25 : grouped_rel->reltarget,
4112 25 : (List *) parse->havingQual);
4113 : }
4114 :
4115 35 : add_path(grouped_rel, path);
4116 35 : }
4117 :
4118 : /*
4119 : * create_ordinary_grouping_paths
4120 : *
4121 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4122 : *
4123 : * We need to consider sorted and hashed aggregation in the same function,
4124 : * because otherwise (1) it would be harder to throw an appropriate error
4125 : * message if neither way works, and (2) we should not allow hashtable size
4126 : * considerations to dissuade us from using hashing if sorting is not possible.
4127 : *
4128 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4129 : * function creates, or to NULL if it doesn't create one.
4130 : */
4131 : static void
4132 35600 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4133 : RelOptInfo *grouped_rel,
4134 : const AggClauseCosts *agg_costs,
4135 : grouping_sets_data *gd,
4136 : GroupPathExtraData *extra,
4137 : RelOptInfo **partially_grouped_rel_p)
4138 : {
4139 35600 : RelOptInfo *partially_grouped_rel = NULL;
4140 35600 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4141 :
4142 : /*
4143 : * If this is the topmost grouping relation or if the parent relation is
4144 : * doing some form of partitionwise aggregation, then we may be able to do
4145 : * it at this level also. However, if the input relation is not
4146 : * partitioned, partitionwise aggregate is impossible.
4147 : */
4148 35600 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4149 2375 : IS_PARTITIONED_REL(input_rel))
4150 : {
4151 : /*
4152 : * If this is the topmost relation or if the parent relation is doing
4153 : * full partitionwise aggregation, then we can do full partitionwise
4154 : * aggregation provided that the GROUP BY clause contains all of the
4155 : * partitioning columns at this level and the collation used by GROUP
4156 : * BY matches the partitioning collation. Otherwise, we can do at
4157 : * most partial partitionwise aggregation. But if partial aggregation
4158 : * is not supported in general then we can't use it for partitionwise
4159 : * aggregation either.
4160 : *
4161 : * Check parse->groupClause not processed_groupClause, because it's
4162 : * okay if some of the partitioning columns were proved redundant.
4163 : */
4164 1360 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4165 640 : group_by_has_partkey(input_rel, extra->targetList,
4166 640 : root->parse->groupClause))
4167 404 : patype = PARTITIONWISE_AGGREGATE_FULL;
4168 316 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4169 281 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4170 : else
4171 35 : patype = PARTITIONWISE_AGGREGATE_NONE;
4172 : }
4173 :
4174 : /*
4175 : * Before generating paths for grouped_rel, we first generate any possible
4176 : * partially grouped paths; that way, later code can easily consider both
4177 : * parallel and non-parallel approaches to grouping.
4178 : */
4179 35600 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4180 : {
4181 : bool force_rel_creation;
4182 :
4183 : /*
4184 : * If we're doing partitionwise aggregation at this level, force
4185 : * creation of a partially_grouped_rel so we can add partitionwise
4186 : * paths to it.
4187 : */
4188 31469 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4189 :
4190 : partially_grouped_rel =
4191 31469 : create_partial_grouping_paths(root,
4192 : grouped_rel,
4193 : input_rel,
4194 : gd,
4195 : extra,
4196 : force_rel_creation);
4197 : }
4198 :
4199 : /* Set out parameter. */
4200 35600 : *partially_grouped_rel_p = partially_grouped_rel;
4201 :
4202 : /* Apply partitionwise aggregation technique, if possible. */
4203 35600 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4204 685 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4205 : partially_grouped_rel, agg_costs,
4206 : gd, patype, extra);
4207 :
4208 : /* If we are doing partial aggregation only, return. */
4209 35600 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4210 : {
4211 : Assert(partially_grouped_rel);
4212 :
4213 713 : if (partially_grouped_rel->pathlist)
4214 713 : set_cheapest(partially_grouped_rel);
4215 :
4216 713 : return;
4217 : }
4218 :
4219 : /* Gather any partially grouped partial paths. */
4220 34887 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4221 2327 : gather_grouping_paths(root, partially_grouped_rel);
4222 :
4223 : /* Now choose the best path(s) for partially_grouped_rel. */
4224 34887 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4225 2495 : set_cheapest(partially_grouped_rel);
4226 :
4227 : /* Build final grouping paths */
4228 34887 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4229 : partially_grouped_rel, agg_costs, gd,
4230 : extra);
4231 :
4232 : /* Give a helpful error if we failed to find any implementation */
4233 34887 : if (grouped_rel->pathlist == NIL)
4234 4 : ereport(ERROR,
4235 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4236 : errmsg("could not implement GROUP BY"),
4237 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4238 :
4239 : /*
4240 : * If there is an FDW that's responsible for all baserels of the query,
4241 : * let it consider adding ForeignPaths.
4242 : */
4243 34883 : if (grouped_rel->fdwroutine &&
4244 169 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4245 168 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4246 : input_rel, grouped_rel,
4247 : extra);
4248 :
4249 : /* Let extensions possibly add some more paths */
4250 34883 : if (create_upper_paths_hook)
4251 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4252 : input_rel, grouped_rel,
4253 : extra);
4254 : }
4255 :
4256 : /*
4257 : * For a given input path, consider the possible ways of doing grouping sets on
4258 : * it, by combinations of hashing and sorting. This can be called multiple
4259 : * times, so it's important that it not scribble on input. No result is
4260 : * returned, but any generated paths are added to grouped_rel.
4261 : */
4262 : static void
4263 1782 : consider_groupingsets_paths(PlannerInfo *root,
4264 : RelOptInfo *grouped_rel,
4265 : Path *path,
4266 : bool is_sorted,
4267 : bool can_hash,
4268 : grouping_sets_data *gd,
4269 : const AggClauseCosts *agg_costs,
4270 : double dNumGroups)
4271 : {
4272 1782 : Query *parse = root->parse;
4273 1782 : Size hash_mem_limit = get_hash_memory_limit();
4274 :
4275 : /*
4276 : * If we're not being offered sorted input, then only consider plans that
4277 : * can be done entirely by hashing.
4278 : *
4279 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4280 : * the input is actually sorted despite not being advertised as such, we
4281 : * prefer to make use of that in order to use less memory.
4282 : *
4283 : * If none of the grouping sets are sortable, then ignore the hash_mem
4284 : * limit and generate a path anyway, since otherwise we'll just fail.
4285 : */
4286 1782 : if (!is_sorted)
4287 : {
4288 807 : List *new_rollups = NIL;
4289 807 : RollupData *unhashed_rollup = NULL;
4290 : List *sets_data;
4291 807 : List *empty_sets_data = NIL;
4292 807 : List *empty_sets = NIL;
4293 : ListCell *lc;
4294 807 : ListCell *l_start = list_head(gd->rollups);
4295 807 : AggStrategy strat = AGG_HASHED;
4296 : double hashsize;
4297 807 : double exclude_groups = 0.0;
4298 :
4299 : Assert(can_hash);
4300 :
4301 : /*
4302 : * If the input is coincidentally sorted usefully (which can happen
4303 : * even if is_sorted is false, since that only means that our caller
4304 : * has set up the sorting for us), then save some hashtable space by
4305 : * making use of that. But we need to watch out for degenerate cases:
4306 : *
4307 : * 1) If there are any empty grouping sets, then group_pathkeys might
4308 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4309 : * there will be a rollup containing only empty groups, and the
4310 : * pathkeys_contained_in test is vacuously true; this is ok.
4311 : *
4312 : * XXX: the above relies on the fact that group_pathkeys is generated
4313 : * from the first rollup. If we add the ability to consider multiple
4314 : * sort orders for grouping input, this assumption might fail.
4315 : *
4316 : * 2) If there are no empty sets and only unsortable sets, then the
4317 : * rollups list will be empty (and thus l_start == NULL), and
4318 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4319 : * pathkeys_contained_in test doesn't cause us to crash.
4320 : */
4321 1609 : if (l_start != NULL &&
4322 802 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4323 : {
4324 40 : unhashed_rollup = lfirst_node(RollupData, l_start);
4325 40 : exclude_groups = unhashed_rollup->numGroups;
4326 40 : l_start = lnext(gd->rollups, l_start);
4327 : }
4328 :
4329 807 : hashsize = estimate_hashagg_tablesize(root,
4330 : path,
4331 : agg_costs,
4332 : dNumGroups - exclude_groups);
4333 :
4334 : /*
4335 : * gd->rollups is empty if we have only unsortable columns to work
4336 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4337 : * sorted-input case to generate usable mixed paths.
4338 : */
4339 807 : if (hashsize > hash_mem_limit && gd->rollups)
4340 15 : return; /* nope, won't fit */
4341 :
4342 : /*
4343 : * We need to burst the existing rollups list into individual grouping
4344 : * sets and recompute a groupClause for each set.
4345 : */
4346 792 : sets_data = list_copy(gd->unsortable_sets);
4347 :
4348 2013 : for_each_cell(lc, gd->rollups, l_start)
4349 : {
4350 1241 : RollupData *rollup = lfirst_node(RollupData, lc);
4351 :
4352 : /*
4353 : * If we find an unhashable rollup that's not been skipped by the
4354 : * "actually sorted" check above, we can't cope; we'd need sorted
4355 : * input (with a different sort order) but we can't get that here.
4356 : * So bail out; we'll get a valid path from the is_sorted case
4357 : * instead.
4358 : *
4359 : * The mere presence of empty grouping sets doesn't make a rollup
4360 : * unhashable (see preprocess_grouping_sets), we handle those
4361 : * specially below.
4362 : */
4363 1241 : if (!rollup->hashable)
4364 20 : return;
4365 :
4366 1221 : sets_data = list_concat(sets_data, rollup->gsets_data);
4367 : }
4368 3053 : foreach(lc, sets_data)
4369 : {
4370 2281 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4371 2281 : List *gset = gs->set;
4372 : RollupData *rollup;
4373 :
4374 2281 : if (gset == NIL)
4375 : {
4376 : /* Empty grouping sets can't be hashed. */
4377 445 : empty_sets_data = lappend(empty_sets_data, gs);
4378 445 : empty_sets = lappend(empty_sets, NIL);
4379 : }
4380 : else
4381 : {
4382 1836 : rollup = makeNode(RollupData);
4383 :
4384 1836 : rollup->groupClause = preprocess_groupclause(root, gset);
4385 1836 : rollup->gsets_data = list_make1(gs);
4386 1836 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4387 : rollup->gsets_data,
4388 : gd->tleref_to_colnum_map);
4389 1836 : rollup->numGroups = gs->numGroups;
4390 1836 : rollup->hashable = true;
4391 1836 : rollup->is_hashed = true;
4392 1836 : new_rollups = lappend(new_rollups, rollup);
4393 : }
4394 : }
4395 :
4396 : /*
4397 : * If we didn't find anything nonempty to hash, then bail. We'll
4398 : * generate a path from the is_sorted case.
4399 : */
4400 772 : if (new_rollups == NIL)
4401 0 : return;
4402 :
4403 : /*
4404 : * If there were empty grouping sets they should have been in the
4405 : * first rollup.
4406 : */
4407 : Assert(!unhashed_rollup || !empty_sets);
4408 :
4409 772 : if (unhashed_rollup)
4410 : {
4411 40 : new_rollups = lappend(new_rollups, unhashed_rollup);
4412 40 : strat = AGG_MIXED;
4413 : }
4414 732 : else if (empty_sets)
4415 : {
4416 405 : RollupData *rollup = makeNode(RollupData);
4417 :
4418 405 : rollup->groupClause = NIL;
4419 405 : rollup->gsets_data = empty_sets_data;
4420 405 : rollup->gsets = empty_sets;
4421 405 : rollup->numGroups = list_length(empty_sets);
4422 405 : rollup->hashable = false;
4423 405 : rollup->is_hashed = false;
4424 405 : new_rollups = lappend(new_rollups, rollup);
4425 405 : strat = AGG_MIXED;
4426 : }
4427 :
4428 772 : add_path(grouped_rel, (Path *)
4429 772 : create_groupingsets_path(root,
4430 : grouped_rel,
4431 : path,
4432 772 : (List *) parse->havingQual,
4433 : strat,
4434 : new_rollups,
4435 : agg_costs));
4436 772 : return;
4437 : }
4438 :
4439 : /*
4440 : * If we have sorted input but nothing we can do with it, bail.
4441 : */
4442 975 : if (gd->rollups == NIL)
4443 0 : return;
4444 :
4445 : /*
4446 : * Given sorted input, we try and make two paths: one sorted and one mixed
4447 : * sort/hash. (We need to try both because hashagg might be disabled, or
4448 : * some columns might not be sortable.)
4449 : *
4450 : * can_hash is passed in as false if some obstacle elsewhere (such as
4451 : * ordered aggs) means that we shouldn't consider hashing at all.
4452 : */
4453 975 : if (can_hash && gd->any_hashable)
4454 : {
4455 896 : List *rollups = NIL;
4456 896 : List *hash_sets = list_copy(gd->unsortable_sets);
4457 896 : double availspace = hash_mem_limit;
4458 : ListCell *lc;
4459 :
4460 : /*
4461 : * Account first for space needed for groups we can't sort at all.
4462 : */
4463 896 : availspace -= estimate_hashagg_tablesize(root,
4464 : path,
4465 : agg_costs,
4466 : gd->dNumHashGroups);
4467 :
4468 896 : if (availspace > 0 && list_length(gd->rollups) > 1)
4469 : {
4470 : double scale;
4471 501 : int num_rollups = list_length(gd->rollups);
4472 : int k_capacity;
4473 501 : int *k_weights = palloc(num_rollups * sizeof(int));
4474 501 : Bitmapset *hash_items = NULL;
4475 : int i;
4476 :
4477 : /*
4478 : * We treat this as a knapsack problem: the knapsack capacity
4479 : * represents hash_mem, the item weights are the estimated memory
4480 : * usage of the hashtables needed to implement a single rollup,
4481 : * and we really ought to use the cost saving as the item value;
4482 : * however, currently the costs assigned to sort nodes don't
4483 : * reflect the comparison costs well, and so we treat all items as
4484 : * of equal value (each rollup we hash instead saves us one sort).
4485 : *
4486 : * To use the discrete knapsack, we need to scale the values to a
4487 : * reasonably small bounded range. We choose to allow a 5% error
4488 : * margin; we have no more than 4096 rollups in the worst possible
4489 : * case, which with a 5% error margin will require a bit over 42MB
4490 : * of workspace. (Anyone wanting to plan queries that complex had
4491 : * better have the memory for it. In more reasonable cases, with
4492 : * no more than a couple of dozen rollups, the memory usage will
4493 : * be negligible.)
4494 : *
4495 : * k_capacity is naturally bounded, but we clamp the values for
4496 : * scale and weight (below) to avoid overflows or underflows (or
4497 : * uselessly trying to use a scale factor less than 1 byte).
4498 : */
4499 501 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4500 501 : k_capacity = (int) floor(availspace / scale);
4501 :
4502 : /*
4503 : * We leave the first rollup out of consideration since it's the
4504 : * one that matches the input sort order. We assign indexes "i"
4505 : * to only those entries considered for hashing; the second loop,
4506 : * below, must use the same condition.
4507 : */
4508 501 : i = 0;
4509 1222 : for_each_from(lc, gd->rollups, 1)
4510 : {
4511 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4512 :
4513 721 : if (rollup->hashable)
4514 : {
4515 721 : double sz = estimate_hashagg_tablesize(root,
4516 : path,
4517 : agg_costs,
4518 : rollup->numGroups);
4519 :
4520 : /*
4521 : * If sz is enormous, but hash_mem (and hence scale) is
4522 : * small, avoid integer overflow here.
4523 : */
4524 721 : k_weights[i] = (int) Min(floor(sz / scale),
4525 : k_capacity + 1.0);
4526 721 : ++i;
4527 : }
4528 : }
4529 :
4530 : /*
4531 : * Apply knapsack algorithm; compute the set of items which
4532 : * maximizes the value stored (in this case the number of sorts
4533 : * saved) while keeping the total size (approximately) within
4534 : * capacity.
4535 : */
4536 501 : if (i > 0)
4537 501 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4538 :
4539 501 : if (!bms_is_empty(hash_items))
4540 : {
4541 501 : rollups = list_make1(linitial(gd->rollups));
4542 :
4543 501 : i = 0;
4544 1222 : for_each_from(lc, gd->rollups, 1)
4545 : {
4546 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4547 :
4548 721 : if (rollup->hashable)
4549 : {
4550 721 : if (bms_is_member(i, hash_items))
4551 691 : hash_sets = list_concat(hash_sets,
4552 691 : rollup->gsets_data);
4553 : else
4554 30 : rollups = lappend(rollups, rollup);
4555 721 : ++i;
4556 : }
4557 : else
4558 0 : rollups = lappend(rollups, rollup);
4559 : }
4560 : }
4561 : }
4562 :
4563 896 : if (!rollups && hash_sets)
4564 20 : rollups = list_copy(gd->rollups);
4565 :
4566 1703 : foreach(lc, hash_sets)
4567 : {
4568 807 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4569 807 : RollupData *rollup = makeNode(RollupData);
4570 :
4571 : Assert(gs->set != NIL);
4572 :
4573 807 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4574 807 : rollup->gsets_data = list_make1(gs);
4575 807 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4576 : rollup->gsets_data,
4577 : gd->tleref_to_colnum_map);
4578 807 : rollup->numGroups = gs->numGroups;
4579 807 : rollup->hashable = true;
4580 807 : rollup->is_hashed = true;
4581 807 : rollups = lcons(rollup, rollups);
4582 : }
4583 :
4584 896 : if (rollups)
4585 : {
4586 521 : add_path(grouped_rel, (Path *)
4587 521 : create_groupingsets_path(root,
4588 : grouped_rel,
4589 : path,
4590 521 : (List *) parse->havingQual,
4591 : AGG_MIXED,
4592 : rollups,
4593 : agg_costs));
4594 : }
4595 : }
4596 :
4597 : /*
4598 : * Now try the simple sorted case.
4599 : */
4600 975 : if (!gd->unsortable_sets)
4601 951 : add_path(grouped_rel, (Path *)
4602 951 : create_groupingsets_path(root,
4603 : grouped_rel,
4604 : path,
4605 951 : (List *) parse->havingQual,
4606 : AGG_SORTED,
4607 : gd->rollups,
4608 : agg_costs));
4609 : }
4610 :
4611 : /*
4612 : * create_window_paths
4613 : *
4614 : * Build a new upperrel containing Paths for window-function evaluation.
4615 : *
4616 : * input_rel: contains the source-data Paths
4617 : * input_target: result of make_window_input_target
4618 : * output_target: what the topmost WindowAggPath should return
4619 : * wflists: result of find_window_functions
4620 : * activeWindows: result of select_active_windows
4621 : *
4622 : * Note: all Paths in input_rel are expected to return input_target.
4623 : */
4624 : static RelOptInfo *
4625 2217 : create_window_paths(PlannerInfo *root,
4626 : RelOptInfo *input_rel,
4627 : PathTarget *input_target,
4628 : PathTarget *output_target,
4629 : bool output_target_parallel_safe,
4630 : WindowFuncLists *wflists,
4631 : List *activeWindows)
4632 : {
4633 : RelOptInfo *window_rel;
4634 : ListCell *lc;
4635 :
4636 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4637 2217 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4638 :
4639 : /*
4640 : * If the input relation is not parallel-safe, then the window relation
4641 : * can't be parallel-safe, either. Otherwise, we need to examine the
4642 : * target list and active windows for non-parallel-safe constructs.
4643 : */
4644 2217 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4645 0 : is_parallel_safe(root, (Node *) activeWindows))
4646 0 : window_rel->consider_parallel = true;
4647 :
4648 : /*
4649 : * If the input rel belongs to a single FDW, so does the window rel.
4650 : */
4651 2217 : window_rel->serverid = input_rel->serverid;
4652 2217 : window_rel->userid = input_rel->userid;
4653 2217 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4654 2217 : window_rel->fdwroutine = input_rel->fdwroutine;
4655 :
4656 : /*
4657 : * Consider computing window functions starting from the existing
4658 : * cheapest-total path (which will likely require a sort) as well as any
4659 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4660 : */
4661 4696 : foreach(lc, input_rel->pathlist)
4662 : {
4663 2479 : Path *path = (Path *) lfirst(lc);
4664 : int presorted_keys;
4665 :
4666 2741 : if (path == input_rel->cheapest_total_path ||
4667 262 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4668 116 : &presorted_keys) ||
4669 116 : presorted_keys > 0)
4670 2384 : create_one_window_path(root,
4671 : window_rel,
4672 : path,
4673 : input_target,
4674 : output_target,
4675 : wflists,
4676 : activeWindows);
4677 : }
4678 :
4679 : /*
4680 : * If there is an FDW that's responsible for all baserels of the query,
4681 : * let it consider adding ForeignPaths.
4682 : */
4683 2217 : if (window_rel->fdwroutine &&
4684 6 : window_rel->fdwroutine->GetForeignUpperPaths)
4685 6 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4686 : input_rel, window_rel,
4687 : NULL);
4688 :
4689 : /* Let extensions possibly add some more paths */
4690 2217 : if (create_upper_paths_hook)
4691 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4692 : input_rel, window_rel, NULL);
4693 :
4694 : /* Now choose the best path(s) */
4695 2217 : set_cheapest(window_rel);
4696 :
4697 2217 : return window_rel;
4698 : }
4699 :
4700 : /*
4701 : * Stack window-function implementation steps atop the given Path, and
4702 : * add the result to window_rel.
4703 : *
4704 : * window_rel: upperrel to contain result
4705 : * path: input Path to use (must return input_target)
4706 : * input_target: result of make_window_input_target
4707 : * output_target: what the topmost WindowAggPath should return
4708 : * wflists: result of find_window_functions
4709 : * activeWindows: result of select_active_windows
4710 : */
4711 : static void
4712 2384 : create_one_window_path(PlannerInfo *root,
4713 : RelOptInfo *window_rel,
4714 : Path *path,
4715 : PathTarget *input_target,
4716 : PathTarget *output_target,
4717 : WindowFuncLists *wflists,
4718 : List *activeWindows)
4719 : {
4720 : PathTarget *window_target;
4721 : ListCell *l;
4722 2384 : List *topqual = NIL;
4723 :
4724 : /*
4725 : * Since each window clause could require a different sort order, we stack
4726 : * up a WindowAgg node for each clause, with sort steps between them as
4727 : * needed. (We assume that select_active_windows chose a good order for
4728 : * executing the clauses in.)
4729 : *
4730 : * input_target should contain all Vars and Aggs needed for the result.
4731 : * (In some cases we wouldn't need to propagate all of these all the way
4732 : * to the top, since they might only be needed as inputs to WindowFuncs.
4733 : * It's probably not worth trying to optimize that though.) It must also
4734 : * contain all window partitioning and sorting expressions, to ensure
4735 : * they're computed only once at the bottom of the stack (that's critical
4736 : * for volatile functions). As we climb up the stack, we'll add outputs
4737 : * for the WindowFuncs computed at each level.
4738 : */
4739 2384 : window_target = input_target;
4740 :
4741 4923 : foreach(l, activeWindows)
4742 : {
4743 2539 : WindowClause *wc = lfirst_node(WindowClause, l);
4744 : List *window_pathkeys;
4745 2539 : List *runcondition = NIL;
4746 : int presorted_keys;
4747 : bool is_sorted;
4748 : bool topwindow;
4749 : ListCell *lc2;
4750 :
4751 2539 : window_pathkeys = make_pathkeys_for_window(root,
4752 : wc,
4753 : root->processed_tlist);
4754 :
4755 2539 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4756 : path->pathkeys,
4757 : &presorted_keys);
4758 :
4759 : /* Sort if necessary */
4760 2539 : if (!is_sorted)
4761 : {
4762 : /*
4763 : * No presorted keys or incremental sort disabled, just perform a
4764 : * complete sort.
4765 : */
4766 1826 : if (presorted_keys == 0 || !enable_incremental_sort)
4767 1775 : path = (Path *) create_sort_path(root, window_rel,
4768 : path,
4769 : window_pathkeys,
4770 : -1.0);
4771 : else
4772 : {
4773 : /*
4774 : * Since we have presorted keys and incremental sort is
4775 : * enabled, just use incremental sort.
4776 : */
4777 51 : path = (Path *) create_incremental_sort_path(root,
4778 : window_rel,
4779 : path,
4780 : window_pathkeys,
4781 : presorted_keys,
4782 : -1.0);
4783 : }
4784 : }
4785 :
4786 2539 : if (lnext(activeWindows, l))
4787 : {
4788 : /*
4789 : * Add the current WindowFuncs to the output target for this
4790 : * intermediate WindowAggPath. We must copy window_target to
4791 : * avoid changing the previous path's target.
4792 : *
4793 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4794 : * we do need to account for the increase in tlist width.
4795 : */
4796 155 : int64 tuple_width = window_target->width;
4797 :
4798 155 : window_target = copy_pathtarget(window_target);
4799 370 : foreach(lc2, wflists->windowFuncs[wc->winref])
4800 : {
4801 215 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4802 :
4803 215 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4804 215 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4805 : }
4806 155 : window_target->width = clamp_width_est(tuple_width);
4807 : }
4808 : else
4809 : {
4810 : /* Install the goal target in the topmost WindowAgg */
4811 2384 : window_target = output_target;
4812 : }
4813 :
4814 : /* mark the final item in the list as the top-level window */
4815 2539 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4816 :
4817 : /*
4818 : * Collect the WindowFuncRunConditions from each WindowFunc and
4819 : * convert them into OpExprs
4820 : */
4821 5811 : foreach(lc2, wflists->windowFuncs[wc->winref])
4822 : {
4823 : ListCell *lc3;
4824 3272 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4825 :
4826 3422 : foreach(lc3, wfunc->runCondition)
4827 : {
4828 150 : WindowFuncRunCondition *wfuncrc =
4829 : lfirst_node(WindowFuncRunCondition, lc3);
4830 : Expr *opexpr;
4831 : Expr *leftop;
4832 : Expr *rightop;
4833 :
4834 150 : if (wfuncrc->wfunc_left)
4835 : {
4836 135 : leftop = (Expr *) copyObject(wfunc);
4837 135 : rightop = copyObject(wfuncrc->arg);
4838 : }
4839 : else
4840 : {
4841 15 : leftop = copyObject(wfuncrc->arg);
4842 15 : rightop = (Expr *) copyObject(wfunc);
4843 : }
4844 :
4845 150 : opexpr = make_opclause(wfuncrc->opno,
4846 : BOOLOID,
4847 : false,
4848 : leftop,
4849 : rightop,
4850 : InvalidOid,
4851 : wfuncrc->inputcollid);
4852 :
4853 150 : runcondition = lappend(runcondition, opexpr);
4854 :
4855 150 : if (!topwindow)
4856 20 : topqual = lappend(topqual, opexpr);
4857 : }
4858 : }
4859 :
4860 : path = (Path *)
4861 2539 : create_windowagg_path(root, window_rel, path, window_target,
4862 2539 : wflists->windowFuncs[wc->winref],
4863 : runcondition, wc,
4864 : topwindow ? topqual : NIL, topwindow);
4865 : }
4866 :
4867 2384 : add_path(window_rel, path);
4868 2384 : }
4869 :
4870 : /*
4871 : * create_distinct_paths
4872 : *
4873 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4874 : *
4875 : * input_rel: contains the source-data Paths
4876 : * target: the pathtarget for the result Paths to compute
4877 : *
4878 : * Note: input paths should already compute the desired pathtarget, since
4879 : * Sort/Unique won't project anything.
4880 : */
4881 : static RelOptInfo *
4882 1941 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4883 : PathTarget *target)
4884 : {
4885 : RelOptInfo *distinct_rel;
4886 :
4887 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4888 1941 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4889 :
4890 : /*
4891 : * We don't compute anything at this level, so distinct_rel will be
4892 : * parallel-safe if the input rel is parallel-safe. In particular, if
4893 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4894 : * output those expressions, and will not be parallel-safe unless those
4895 : * expressions are parallel-safe.
4896 : */
4897 1941 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4898 :
4899 : /*
4900 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4901 : */
4902 1941 : distinct_rel->serverid = input_rel->serverid;
4903 1941 : distinct_rel->userid = input_rel->userid;
4904 1941 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4905 1941 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4906 :
4907 : /* build distinct paths based on input_rel's pathlist */
4908 1941 : create_final_distinct_paths(root, input_rel, distinct_rel);
4909 :
4910 : /* now build distinct paths based on input_rel's partial_pathlist */
4911 1941 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4912 :
4913 : /* Give a helpful error if we failed to create any paths */
4914 1941 : if (distinct_rel->pathlist == NIL)
4915 0 : ereport(ERROR,
4916 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4917 : errmsg("could not implement DISTINCT"),
4918 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4919 :
4920 : /*
4921 : * If there is an FDW that's responsible for all baserels of the query,
4922 : * let it consider adding ForeignPaths.
4923 : */
4924 1941 : if (distinct_rel->fdwroutine &&
4925 8 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4926 8 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4927 : UPPERREL_DISTINCT,
4928 : input_rel,
4929 : distinct_rel,
4930 : NULL);
4931 :
4932 : /* Let extensions possibly add some more paths */
4933 1941 : if (create_upper_paths_hook)
4934 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4935 : distinct_rel, NULL);
4936 :
4937 : /* Now choose the best path(s) */
4938 1941 : set_cheapest(distinct_rel);
4939 :
4940 1941 : return distinct_rel;
4941 : }
4942 :
4943 : /*
4944 : * create_partial_distinct_paths
4945 : *
4946 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4947 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4948 : * paths on top and add a final unique/aggregate path to remove any duplicate
4949 : * produced from combining rows from parallel workers.
4950 : */
4951 : static void
4952 1941 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4953 : RelOptInfo *final_distinct_rel,
4954 : PathTarget *target)
4955 : {
4956 : RelOptInfo *partial_distinct_rel;
4957 : Query *parse;
4958 : List *distinctExprs;
4959 : double numDistinctRows;
4960 : Path *cheapest_partial_path;
4961 : ListCell *lc;
4962 :
4963 : /* nothing to do when there are no partial paths in the input rel */
4964 1941 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4965 1851 : return;
4966 :
4967 90 : parse = root->parse;
4968 :
4969 : /* can't do parallel DISTINCT ON */
4970 90 : if (parse->hasDistinctOn)
4971 0 : return;
4972 :
4973 90 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4974 : NULL);
4975 90 : partial_distinct_rel->reltarget = target;
4976 90 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4977 :
4978 : /*
4979 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4980 : */
4981 90 : partial_distinct_rel->serverid = input_rel->serverid;
4982 90 : partial_distinct_rel->userid = input_rel->userid;
4983 90 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4984 90 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4985 :
4986 90 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4987 :
4988 90 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4989 : parse->targetList);
4990 :
4991 : /* estimate how many distinct rows we'll get from each worker */
4992 90 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4993 : cheapest_partial_path->rows,
4994 : NULL, NULL);
4995 :
4996 : /*
4997 : * Try sorting the cheapest path and incrementally sorting any paths with
4998 : * presorted keys and put a unique paths atop of those. We'll also
4999 : * attempt to reorder the required pathkeys to match the input path's
5000 : * pathkeys as much as possible, in hopes of avoiding a possible need to
5001 : * re-sort.
5002 : */
5003 90 : if (grouping_is_sortable(root->processed_distinctClause))
5004 : {
5005 193 : foreach(lc, input_rel->partial_pathlist)
5006 : {
5007 103 : Path *input_path = (Path *) lfirst(lc);
5008 : Path *sorted_path;
5009 103 : List *useful_pathkeys_list = NIL;
5010 :
5011 : useful_pathkeys_list =
5012 103 : get_useful_pathkeys_for_distinct(root,
5013 : root->distinct_pathkeys,
5014 : input_path->pathkeys);
5015 : Assert(list_length(useful_pathkeys_list) > 0);
5016 :
5017 319 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5018 : {
5019 113 : sorted_path = make_ordered_path(root,
5020 : partial_distinct_rel,
5021 : input_path,
5022 : cheapest_partial_path,
5023 : useful_pathkeys,
5024 : -1.0);
5025 :
5026 113 : if (sorted_path == NULL)
5027 8 : continue;
5028 :
5029 : /*
5030 : * An empty distinct_pathkeys means all tuples have the same
5031 : * value for the DISTINCT clause. See
5032 : * create_final_distinct_paths()
5033 : */
5034 105 : if (root->distinct_pathkeys == NIL)
5035 : {
5036 : Node *limitCount;
5037 :
5038 5 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5039 : sizeof(int64),
5040 : Int64GetDatum(1), false,
5041 : true);
5042 :
5043 : /*
5044 : * Apply a LimitPath onto the partial path to restrict the
5045 : * tuples from each worker to 1.
5046 : * create_final_distinct_paths will need to apply an
5047 : * additional LimitPath to restrict this to a single row
5048 : * after the Gather node. If the query already has a
5049 : * LIMIT clause, then we could end up with three Limit
5050 : * nodes in the final plan. Consolidating the top two of
5051 : * these could be done, but does not seem worth troubling
5052 : * over.
5053 : */
5054 5 : add_partial_path(partial_distinct_rel, (Path *)
5055 5 : create_limit_path(root, partial_distinct_rel,
5056 : sorted_path,
5057 : NULL,
5058 : limitCount,
5059 : LIMIT_OPTION_COUNT,
5060 : 0, 1));
5061 : }
5062 : else
5063 : {
5064 100 : add_partial_path(partial_distinct_rel, (Path *)
5065 100 : create_unique_path(root, partial_distinct_rel,
5066 : sorted_path,
5067 100 : list_length(root->distinct_pathkeys),
5068 : numDistinctRows));
5069 : }
5070 : }
5071 : }
5072 : }
5073 :
5074 : /*
5075 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5076 : * we're not on the hook to ensure we do our best to create at least one
5077 : * path here, we treat enable_hashagg as a hard off-switch rather than the
5078 : * slightly softer variant in create_final_distinct_paths.
5079 : */
5080 90 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5081 : {
5082 65 : add_partial_path(partial_distinct_rel, (Path *)
5083 65 : create_agg_path(root,
5084 : partial_distinct_rel,
5085 : cheapest_partial_path,
5086 : cheapest_partial_path->pathtarget,
5087 : AGG_HASHED,
5088 : AGGSPLIT_SIMPLE,
5089 : root->processed_distinctClause,
5090 : NIL,
5091 : NULL,
5092 : numDistinctRows));
5093 : }
5094 :
5095 : /*
5096 : * If there is an FDW that's responsible for all baserels of the query,
5097 : * let it consider adding ForeignPaths.
5098 : */
5099 90 : if (partial_distinct_rel->fdwroutine &&
5100 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5101 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5102 : UPPERREL_PARTIAL_DISTINCT,
5103 : input_rel,
5104 : partial_distinct_rel,
5105 : NULL);
5106 :
5107 : /* Let extensions possibly add some more partial paths */
5108 90 : if (create_upper_paths_hook)
5109 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5110 : input_rel, partial_distinct_rel, NULL);
5111 :
5112 90 : if (partial_distinct_rel->partial_pathlist != NIL)
5113 : {
5114 90 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5115 90 : set_cheapest(partial_distinct_rel);
5116 :
5117 : /*
5118 : * Finally, create paths to distinctify the final result. This step
5119 : * is needed to remove any duplicates due to combining rows from
5120 : * parallel workers.
5121 : */
5122 90 : create_final_distinct_paths(root, partial_distinct_rel,
5123 : final_distinct_rel);
5124 : }
5125 : }
5126 :
5127 : /*
5128 : * create_final_distinct_paths
5129 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5130 : *
5131 : * input_rel: contains the source-data paths
5132 : * distinct_rel: destination relation for storing created paths
5133 : */
5134 : static RelOptInfo *
5135 2031 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5136 : RelOptInfo *distinct_rel)
5137 : {
5138 2031 : Query *parse = root->parse;
5139 2031 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5140 : double numDistinctRows;
5141 : bool allow_hash;
5142 :
5143 : /* Estimate number of distinct rows there will be */
5144 2031 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5145 1972 : root->hasHavingQual)
5146 : {
5147 : /*
5148 : * If there was grouping or aggregation, use the number of input rows
5149 : * as the estimated number of DISTINCT rows (ie, assume the input is
5150 : * already mostly unique).
5151 : */
5152 59 : numDistinctRows = cheapest_input_path->rows;
5153 : }
5154 : else
5155 : {
5156 : /*
5157 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5158 : */
5159 : List *distinctExprs;
5160 :
5161 1972 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5162 : parse->targetList);
5163 1972 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5164 : cheapest_input_path->rows,
5165 : NULL, NULL);
5166 : }
5167 :
5168 : /*
5169 : * Consider sort-based implementations of DISTINCT, if possible.
5170 : */
5171 2031 : if (grouping_is_sortable(root->processed_distinctClause))
5172 : {
5173 : /*
5174 : * Firstly, if we have any adequately-presorted paths, just stick a
5175 : * Unique node on those. We also, consider doing an explicit sort of
5176 : * the cheapest input path and Unique'ing that. If any paths have
5177 : * presorted keys then we'll create an incremental sort atop of those
5178 : * before adding a unique node on the top. We'll also attempt to
5179 : * reorder the required pathkeys to match the input path's pathkeys as
5180 : * much as possible, in hopes of avoiding a possible need to re-sort.
5181 : *
5182 : * When we have DISTINCT ON, we must sort by the more rigorous of
5183 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5184 : * Also, if we do have to do an explicit sort, we might as well use
5185 : * the more rigorous ordering to avoid a second sort later. (Note
5186 : * that the parser will have ensured that one clause is a prefix of
5187 : * the other.)
5188 : */
5189 : List *needed_pathkeys;
5190 : ListCell *lc;
5191 2026 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5192 :
5193 2213 : if (parse->hasDistinctOn &&
5194 187 : list_length(root->distinct_pathkeys) <
5195 187 : list_length(root->sort_pathkeys))
5196 52 : needed_pathkeys = root->sort_pathkeys;
5197 : else
5198 1974 : needed_pathkeys = root->distinct_pathkeys;
5199 :
5200 5401 : foreach(lc, input_rel->pathlist)
5201 : {
5202 3375 : Path *input_path = (Path *) lfirst(lc);
5203 : Path *sorted_path;
5204 3375 : List *useful_pathkeys_list = NIL;
5205 :
5206 : useful_pathkeys_list =
5207 3375 : get_useful_pathkeys_for_distinct(root,
5208 : needed_pathkeys,
5209 : input_path->pathkeys);
5210 : Assert(list_length(useful_pathkeys_list) > 0);
5211 :
5212 10625 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5213 : {
5214 3875 : sorted_path = make_ordered_path(root,
5215 : distinct_rel,
5216 : input_path,
5217 : cheapest_input_path,
5218 : useful_pathkeys,
5219 : limittuples);
5220 :
5221 3875 : if (sorted_path == NULL)
5222 531 : continue;
5223 :
5224 : /*
5225 : * distinct_pathkeys may have become empty if all of the
5226 : * pathkeys were determined to be redundant. If all of the
5227 : * pathkeys are redundant then each DISTINCT target must only
5228 : * allow a single value, therefore all resulting tuples must
5229 : * be identical (or at least indistinguishable by an equality
5230 : * check). We can uniquify these tuples simply by just taking
5231 : * the first tuple. All we do here is add a path to do "LIMIT
5232 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5233 : * still have a non-NIL sort_pathkeys list, so we must still
5234 : * only do this with paths which are correctly sorted by
5235 : * sort_pathkeys.
5236 : */
5237 3344 : if (root->distinct_pathkeys == NIL)
5238 : {
5239 : Node *limitCount;
5240 :
5241 96 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5242 : sizeof(int64),
5243 : Int64GetDatum(1), false,
5244 : true);
5245 :
5246 : /*
5247 : * If the query already has a LIMIT clause, then we could
5248 : * end up with a duplicate LimitPath in the final plan.
5249 : * That does not seem worth troubling over too much.
5250 : */
5251 96 : add_path(distinct_rel, (Path *)
5252 96 : create_limit_path(root, distinct_rel, sorted_path,
5253 : NULL, limitCount,
5254 : LIMIT_OPTION_COUNT, 0, 1));
5255 : }
5256 : else
5257 : {
5258 3248 : add_path(distinct_rel, (Path *)
5259 3248 : create_unique_path(root, distinct_rel,
5260 : sorted_path,
5261 3248 : list_length(root->distinct_pathkeys),
5262 : numDistinctRows));
5263 : }
5264 : }
5265 : }
5266 : }
5267 :
5268 : /*
5269 : * Consider hash-based implementations of DISTINCT, if possible.
5270 : *
5271 : * If we were not able to make any other types of path, we *must* hash or
5272 : * die trying. If we do have other choices, there are two things that
5273 : * should prevent selection of hashing: if the query uses DISTINCT ON
5274 : * (because it won't really have the expected behavior if we hash), or if
5275 : * enable_hashagg is off.
5276 : *
5277 : * Note: grouping_is_hashable() is much more expensive to check than the
5278 : * other gating conditions, so we want to do it last.
5279 : */
5280 2031 : if (distinct_rel->pathlist == NIL)
5281 5 : allow_hash = true; /* we have no alternatives */
5282 2026 : else if (parse->hasDistinctOn || !enable_hashagg)
5283 312 : allow_hash = false; /* policy-based decision not to hash */
5284 : else
5285 1714 : allow_hash = true; /* default */
5286 :
5287 2031 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5288 : {
5289 : /* Generate hashed aggregate path --- no sort needed */
5290 1719 : add_path(distinct_rel, (Path *)
5291 1719 : create_agg_path(root,
5292 : distinct_rel,
5293 : cheapest_input_path,
5294 : cheapest_input_path->pathtarget,
5295 : AGG_HASHED,
5296 : AGGSPLIT_SIMPLE,
5297 : root->processed_distinctClause,
5298 : NIL,
5299 : NULL,
5300 : numDistinctRows));
5301 : }
5302 :
5303 2031 : return distinct_rel;
5304 : }
5305 :
5306 : /*
5307 : * get_useful_pathkeys_for_distinct
5308 : * Get useful orderings of pathkeys for distinctClause by reordering
5309 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5310 : *
5311 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5312 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5313 : */
5314 : static List *
5315 3478 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5316 : List *path_pathkeys)
5317 : {
5318 3478 : List *useful_pathkeys_list = NIL;
5319 3478 : List *useful_pathkeys = NIL;
5320 :
5321 : /* always include the given 'needed_pathkeys' */
5322 3478 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5323 : needed_pathkeys);
5324 :
5325 3478 : if (!enable_distinct_reordering)
5326 0 : return useful_pathkeys_list;
5327 :
5328 : /*
5329 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5330 : * that match 'needed_pathkeys', but only up to the longest matching
5331 : * prefix.
5332 : *
5333 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5334 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5335 : * the desired behavior.
5336 : */
5337 8669 : foreach_node(PathKey, pathkey, path_pathkeys)
5338 : {
5339 : /*
5340 : * The PathKey nodes are canonical, so they can be checked for
5341 : * equality by simple pointer comparison.
5342 : */
5343 1735 : if (!list_member_ptr(needed_pathkeys, pathkey))
5344 7 : break;
5345 1728 : if (root->parse->hasDistinctOn &&
5346 140 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5347 15 : break;
5348 :
5349 1713 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5350 : }
5351 :
5352 : /* If no match at all, no point in reordering needed_pathkeys */
5353 3478 : if (useful_pathkeys == NIL)
5354 1993 : return useful_pathkeys_list;
5355 :
5356 : /*
5357 : * If not full match, the resulting pathkey list is not useful without
5358 : * incremental sort.
5359 : */
5360 1485 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5361 924 : !enable_incremental_sort)
5362 48 : return useful_pathkeys_list;
5363 :
5364 : /* Append the remaining PathKey nodes in needed_pathkeys */
5365 1437 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5366 : needed_pathkeys);
5367 :
5368 : /*
5369 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5370 : * just drop it.
5371 : */
5372 1437 : if (compare_pathkeys(needed_pathkeys,
5373 : useful_pathkeys) == PATHKEYS_EQUAL)
5374 927 : return useful_pathkeys_list;
5375 :
5376 510 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5377 : useful_pathkeys);
5378 :
5379 510 : return useful_pathkeys_list;
5380 : }
5381 :
5382 : /*
5383 : * create_ordered_paths
5384 : *
5385 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5386 : *
5387 : * All paths in the result must satisfy the ORDER BY ordering.
5388 : * The only new paths we need consider are an explicit full sort
5389 : * and incremental sort on the cheapest-total existing path.
5390 : *
5391 : * input_rel: contains the source-data Paths
5392 : * target: the output tlist the result Paths must emit
5393 : * limit_tuples: estimated bound on the number of output tuples,
5394 : * or -1 if no LIMIT or couldn't estimate
5395 : *
5396 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5397 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5398 : */
5399 : static RelOptInfo *
5400 61919 : create_ordered_paths(PlannerInfo *root,
5401 : RelOptInfo *input_rel,
5402 : PathTarget *target,
5403 : bool target_parallel_safe,
5404 : double limit_tuples)
5405 : {
5406 61919 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5407 : RelOptInfo *ordered_rel;
5408 : ListCell *lc;
5409 :
5410 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5411 61919 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5412 :
5413 : /*
5414 : * If the input relation is not parallel-safe, then the ordered relation
5415 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5416 : * target list is parallel-safe.
5417 : */
5418 61919 : if (input_rel->consider_parallel && target_parallel_safe)
5419 43288 : ordered_rel->consider_parallel = true;
5420 :
5421 : /* Assume that the same path generation strategies are allowed. */
5422 61919 : ordered_rel->pgs_mask = input_rel->pgs_mask;
5423 :
5424 : /*
5425 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5426 : */
5427 61919 : ordered_rel->serverid = input_rel->serverid;
5428 61919 : ordered_rel->userid = input_rel->userid;
5429 61919 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5430 61919 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5431 :
5432 157992 : foreach(lc, input_rel->pathlist)
5433 : {
5434 96073 : Path *input_path = (Path *) lfirst(lc);
5435 : Path *sorted_path;
5436 : bool is_sorted;
5437 : int presorted_keys;
5438 :
5439 96073 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5440 : input_path->pathkeys, &presorted_keys);
5441 :
5442 96073 : if (is_sorted)
5443 37061 : sorted_path = input_path;
5444 : else
5445 : {
5446 : /*
5447 : * Try at least sorting the cheapest path and also try
5448 : * incrementally sorting any path which is partially sorted
5449 : * already (no need to deal with paths which have presorted keys
5450 : * when incremental sort is disabled unless it's the cheapest
5451 : * input path).
5452 : */
5453 59012 : if (input_path != cheapest_input_path &&
5454 4795 : (presorted_keys == 0 || !enable_incremental_sort))
5455 1613 : continue;
5456 :
5457 : /*
5458 : * We've no need to consider both a sort and incremental sort.
5459 : * We'll just do a sort if there are no presorted keys and an
5460 : * incremental sort when there are presorted keys.
5461 : */
5462 57399 : if (presorted_keys == 0 || !enable_incremental_sort)
5463 53766 : sorted_path = (Path *) create_sort_path(root,
5464 : ordered_rel,
5465 : input_path,
5466 : root->sort_pathkeys,
5467 : limit_tuples);
5468 : else
5469 3633 : sorted_path = (Path *) create_incremental_sort_path(root,
5470 : ordered_rel,
5471 : input_path,
5472 : root->sort_pathkeys,
5473 : presorted_keys,
5474 : limit_tuples);
5475 : }
5476 :
5477 : /*
5478 : * If the pathtarget of the result path has different expressions from
5479 : * the target to be applied, a projection step is needed.
5480 : */
5481 94460 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5482 288 : sorted_path = apply_projection_to_path(root, ordered_rel,
5483 : sorted_path, target);
5484 :
5485 94460 : add_path(ordered_rel, sorted_path);
5486 : }
5487 :
5488 : /*
5489 : * generate_gather_paths() will have already generated a simple Gather
5490 : * path for the best parallel path, if any, and the loop above will have
5491 : * considered sorting it. Similarly, generate_gather_paths() will also
5492 : * have generated order-preserving Gather Merge plans which can be used
5493 : * without sorting if they happen to match the sort_pathkeys, and the loop
5494 : * above will have handled those as well. However, there's one more
5495 : * possibility: it may make sense to sort the cheapest partial path or
5496 : * incrementally sort any partial path that is partially sorted according
5497 : * to the required output order and then use Gather Merge.
5498 : */
5499 61919 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5500 43168 : input_rel->partial_pathlist != NIL)
5501 : {
5502 : Path *cheapest_partial_path;
5503 :
5504 2411 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5505 :
5506 5377 : foreach(lc, input_rel->partial_pathlist)
5507 : {
5508 2966 : Path *input_path = (Path *) lfirst(lc);
5509 : Path *sorted_path;
5510 : bool is_sorted;
5511 : int presorted_keys;
5512 : double total_groups;
5513 :
5514 2966 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5515 : input_path->pathkeys,
5516 : &presorted_keys);
5517 :
5518 2966 : if (is_sorted)
5519 476 : continue;
5520 :
5521 : /*
5522 : * Try at least sorting the cheapest path and also try
5523 : * incrementally sorting any path which is partially sorted
5524 : * already (no need to deal with paths which have presorted keys
5525 : * when incremental sort is disabled unless it's the cheapest
5526 : * partial path).
5527 : */
5528 2490 : if (input_path != cheapest_partial_path &&
5529 101 : (presorted_keys == 0 || !enable_incremental_sort))
5530 0 : continue;
5531 :
5532 : /*
5533 : * We've no need to consider both a sort and incremental sort.
5534 : * We'll just do a sort if there are no presorted keys and an
5535 : * incremental sort when there are presorted keys.
5536 : */
5537 2490 : if (presorted_keys == 0 || !enable_incremental_sort)
5538 2374 : sorted_path = (Path *) create_sort_path(root,
5539 : ordered_rel,
5540 : input_path,
5541 : root->sort_pathkeys,
5542 : limit_tuples);
5543 : else
5544 116 : sorted_path = (Path *) create_incremental_sort_path(root,
5545 : ordered_rel,
5546 : input_path,
5547 : root->sort_pathkeys,
5548 : presorted_keys,
5549 : limit_tuples);
5550 2490 : total_groups = compute_gather_rows(sorted_path);
5551 : sorted_path = (Path *)
5552 2490 : create_gather_merge_path(root, ordered_rel,
5553 : sorted_path,
5554 : sorted_path->pathtarget,
5555 : root->sort_pathkeys, NULL,
5556 : &total_groups);
5557 :
5558 : /*
5559 : * If the pathtarget of the result path has different expressions
5560 : * from the target to be applied, a projection step is needed.
5561 : */
5562 2490 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5563 5 : sorted_path = apply_projection_to_path(root, ordered_rel,
5564 : sorted_path, target);
5565 :
5566 2490 : add_path(ordered_rel, sorted_path);
5567 : }
5568 : }
5569 :
5570 : /*
5571 : * If there is an FDW that's responsible for all baserels of the query,
5572 : * let it consider adding ForeignPaths.
5573 : */
5574 61919 : if (ordered_rel->fdwroutine &&
5575 195 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5576 187 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5577 : input_rel, ordered_rel,
5578 : NULL);
5579 :
5580 : /* Let extensions possibly add some more paths */
5581 61919 : if (create_upper_paths_hook)
5582 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5583 : input_rel, ordered_rel, NULL);
5584 :
5585 : /*
5586 : * No need to bother with set_cheapest here; grouping_planner does not
5587 : * need us to do it.
5588 : */
5589 : Assert(ordered_rel->pathlist != NIL);
5590 :
5591 61919 : return ordered_rel;
5592 : }
5593 :
5594 :
5595 : /*
5596 : * make_group_input_target
5597 : * Generate appropriate PathTarget for initial input to grouping nodes.
5598 : *
5599 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5600 : * the query's final targetlist; for example, it certainly can't emit any
5601 : * aggregate function calls. This routine generates the correct target
5602 : * for the scan/join subplan.
5603 : *
5604 : * The query target list passed from the parser already contains entries
5605 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5606 : * for variables used only in HAVING clauses; so we need to add those
5607 : * variables to the subplan target list. Also, we flatten all expressions
5608 : * except GROUP BY items into their component variables; other expressions
5609 : * will be computed by the upper plan nodes rather than by the subplan.
5610 : * For example, given a query like
5611 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5612 : * we want to pass this targetlist to the subplan:
5613 : * a+b,c,d
5614 : * where the a+b target will be used by the Sort/Group steps, and the
5615 : * other targets will be used for computing the final results.
5616 : *
5617 : * 'final_target' is the query's final target list (in PathTarget form)
5618 : *
5619 : * The result is the PathTarget to be computed by the Paths returned from
5620 : * query_planner().
5621 : */
5622 : static PathTarget *
5623 33840 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5624 : {
5625 33840 : Query *parse = root->parse;
5626 : PathTarget *input_target;
5627 : List *non_group_cols;
5628 : List *non_group_vars;
5629 : int i;
5630 : ListCell *lc;
5631 :
5632 : /*
5633 : * We must build a target containing all grouping columns, plus any other
5634 : * Vars mentioned in the query's targetlist and HAVING qual.
5635 : */
5636 33840 : input_target = create_empty_pathtarget();
5637 33840 : non_group_cols = NIL;
5638 :
5639 33840 : i = 0;
5640 85083 : foreach(lc, final_target->exprs)
5641 : {
5642 51243 : Expr *expr = (Expr *) lfirst(lc);
5643 51243 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5644 :
5645 59239 : if (sgref && root->processed_groupClause &&
5646 7996 : get_sortgroupref_clause_noerr(sgref,
5647 : root->processed_groupClause) != NULL)
5648 : {
5649 : /*
5650 : * It's a grouping column, so add it to the input target as-is.
5651 : *
5652 : * Note that the target is logically below the grouping step. So
5653 : * with grouping sets we need to remove the RT index of the
5654 : * grouping step if there is any from the target expression.
5655 : */
5656 6485 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5657 : {
5658 : Assert(root->group_rtindex > 0);
5659 : expr = (Expr *)
5660 1803 : remove_nulling_relids((Node *) expr,
5661 1803 : bms_make_singleton(root->group_rtindex),
5662 : NULL);
5663 : }
5664 6485 : add_column_to_pathtarget(input_target, expr, sgref);
5665 : }
5666 : else
5667 : {
5668 : /*
5669 : * Non-grouping column, so just remember the expression for later
5670 : * call to pull_var_clause.
5671 : */
5672 44758 : non_group_cols = lappend(non_group_cols, expr);
5673 : }
5674 :
5675 51243 : i++;
5676 : }
5677 :
5678 : /*
5679 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5680 : */
5681 33840 : if (parse->havingQual)
5682 755 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5683 :
5684 : /*
5685 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5686 : * add them to the input target if not already present. (A Var used
5687 : * directly as a GROUP BY item will be present already.) Note this
5688 : * includes Vars used in resjunk items, so we are covering the needs of
5689 : * ORDER BY and window specifications. Vars used within Aggrefs and
5690 : * WindowFuncs will be pulled out here, too.
5691 : *
5692 : * Note that the target is logically below the grouping step. So with
5693 : * grouping sets we need to remove the RT index of the grouping step if
5694 : * there is any from the non-group Vars.
5695 : */
5696 33840 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5697 : PVC_RECURSE_AGGREGATES |
5698 : PVC_RECURSE_WINDOWFUNCS |
5699 : PVC_INCLUDE_PLACEHOLDERS);
5700 33840 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5701 : {
5702 : Assert(root->group_rtindex > 0);
5703 : non_group_vars = (List *)
5704 836 : remove_nulling_relids((Node *) non_group_vars,
5705 836 : bms_make_singleton(root->group_rtindex),
5706 : NULL);
5707 : }
5708 33840 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5709 :
5710 : /* clean up cruft */
5711 33840 : list_free(non_group_vars);
5712 33840 : list_free(non_group_cols);
5713 :
5714 : /* XXX this causes some redundant cost calculation ... */
5715 33840 : return set_pathtarget_cost_width(root, input_target);
5716 : }
5717 :
5718 : /*
5719 : * make_partial_grouping_target
5720 : * Generate appropriate PathTarget for output of partial aggregate
5721 : * (or partial grouping, if there are no aggregates) nodes.
5722 : *
5723 : * A partial aggregation node needs to emit all the same aggregates that
5724 : * a regular aggregation node would, plus any aggregates used in HAVING;
5725 : * except that the Aggref nodes should be marked as partial aggregates.
5726 : *
5727 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5728 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5729 : * these would be Vars that are grouped by or used in grouping expressions.)
5730 : *
5731 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5732 : * havingQual represents the HAVING clause.
5733 : */
5734 : static PathTarget *
5735 3208 : make_partial_grouping_target(PlannerInfo *root,
5736 : PathTarget *grouping_target,
5737 : Node *havingQual)
5738 : {
5739 : PathTarget *partial_target;
5740 : List *non_group_cols;
5741 : List *non_group_exprs;
5742 : int i;
5743 : ListCell *lc;
5744 :
5745 3208 : partial_target = create_empty_pathtarget();
5746 3208 : non_group_cols = NIL;
5747 :
5748 3208 : i = 0;
5749 10567 : foreach(lc, grouping_target->exprs)
5750 : {
5751 7359 : Expr *expr = (Expr *) lfirst(lc);
5752 7359 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5753 :
5754 11291 : if (sgref && root->processed_groupClause &&
5755 3932 : get_sortgroupref_clause_noerr(sgref,
5756 : root->processed_groupClause) != NULL)
5757 : {
5758 : /*
5759 : * It's a grouping column, so add it to the partial_target as-is.
5760 : * (This allows the upper agg step to repeat the grouping calcs.)
5761 : */
5762 2330 : add_column_to_pathtarget(partial_target, expr, sgref);
5763 : }
5764 : else
5765 : {
5766 : /*
5767 : * Non-grouping column, so just remember the expression for later
5768 : * call to pull_var_clause.
5769 : */
5770 5029 : non_group_cols = lappend(non_group_cols, expr);
5771 : }
5772 :
5773 7359 : i++;
5774 : }
5775 :
5776 : /*
5777 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5778 : */
5779 3208 : if (havingQual)
5780 729 : non_group_cols = lappend(non_group_cols, havingQual);
5781 :
5782 : /*
5783 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5784 : * non-group cols (plus HAVING), and add them to the partial_target if not
5785 : * already present. (An expression used directly as a GROUP BY item will
5786 : * be present already.) Note this includes Vars used in resjunk items, so
5787 : * we are covering the needs of ORDER BY and window specifications.
5788 : */
5789 3208 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5790 : PVC_INCLUDE_AGGREGATES |
5791 : PVC_RECURSE_WINDOWFUNCS |
5792 : PVC_INCLUDE_PLACEHOLDERS);
5793 :
5794 3208 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5795 :
5796 : /*
5797 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5798 : * are at the top level of the target list, so we can just scan the list
5799 : * rather than recursing through the expression trees.
5800 : */
5801 11081 : foreach(lc, partial_target->exprs)
5802 : {
5803 7873 : Aggref *aggref = (Aggref *) lfirst(lc);
5804 :
5805 7873 : if (IsA(aggref, Aggref))
5806 : {
5807 : Aggref *newaggref;
5808 :
5809 : /*
5810 : * We shouldn't need to copy the substructure of the Aggref node,
5811 : * but flat-copy the node itself to avoid damaging other trees.
5812 : */
5813 5518 : newaggref = makeNode(Aggref);
5814 5518 : memcpy(newaggref, aggref, sizeof(Aggref));
5815 :
5816 : /* For now, assume serialization is required */
5817 5518 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5818 :
5819 5518 : lfirst(lc) = newaggref;
5820 : }
5821 : }
5822 :
5823 : /* clean up cruft */
5824 3208 : list_free(non_group_exprs);
5825 3208 : list_free(non_group_cols);
5826 :
5827 : /* XXX this causes some redundant cost calculation ... */
5828 3208 : return set_pathtarget_cost_width(root, partial_target);
5829 : }
5830 :
5831 : /*
5832 : * mark_partial_aggref
5833 : * Adjust an Aggref to make it represent a partial-aggregation step.
5834 : *
5835 : * The Aggref node is modified in-place; caller must do any copying required.
5836 : */
5837 : void
5838 15336 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5839 : {
5840 : /* aggtranstype should be computed by this point */
5841 : Assert(OidIsValid(agg->aggtranstype));
5842 : /* ... but aggsplit should still be as the parser left it */
5843 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5844 :
5845 : /* Mark the Aggref with the intended partial-aggregation mode */
5846 15336 : agg->aggsplit = aggsplit;
5847 :
5848 : /*
5849 : * Adjust result type if needed. Normally, a partial aggregate returns
5850 : * the aggregate's transition type; but if that's INTERNAL and we're
5851 : * serializing, it returns BYTEA instead.
5852 : */
5853 15336 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5854 : {
5855 13429 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5856 261 : agg->aggtype = BYTEAOID;
5857 : else
5858 13168 : agg->aggtype = agg->aggtranstype;
5859 : }
5860 15336 : }
5861 :
5862 : /*
5863 : * postprocess_setop_tlist
5864 : * Fix up targetlist returned by plan_set_operations().
5865 : *
5866 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5867 : * NOTE: this would not be good enough if we supported resjunk sort keys
5868 : * for results of set operations --- then, we'd need to project a whole
5869 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5870 : * find any resjunk columns in orig_tlist.
5871 : */
5872 : static List *
5873 4942 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5874 : {
5875 : ListCell *l;
5876 4942 : ListCell *orig_tlist_item = list_head(orig_tlist);
5877 :
5878 19199 : foreach(l, new_tlist)
5879 : {
5880 14257 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5881 : TargetEntry *orig_tle;
5882 :
5883 : /* ignore resjunk columns in setop result */
5884 14257 : if (new_tle->resjunk)
5885 0 : continue;
5886 :
5887 : Assert(orig_tlist_item != NULL);
5888 14257 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5889 14257 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5890 14257 : if (orig_tle->resjunk) /* should not happen */
5891 0 : elog(ERROR, "resjunk output columns are not implemented");
5892 : Assert(new_tle->resno == orig_tle->resno);
5893 14257 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5894 : }
5895 4942 : if (orig_tlist_item != NULL)
5896 0 : elog(ERROR, "resjunk output columns are not implemented");
5897 4942 : return new_tlist;
5898 : }
5899 :
5900 : /*
5901 : * optimize_window_clauses
5902 : * Call each WindowFunc's prosupport function to see if we're able to
5903 : * make any adjustments to any of the WindowClause's so that the executor
5904 : * can execute the window functions in a more optimal way.
5905 : *
5906 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5907 : * may allow more things to be done here in the future.
5908 : */
5909 : static void
5910 2217 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5911 : {
5912 2217 : List *windowClause = root->parse->windowClause;
5913 : ListCell *lc;
5914 :
5915 4639 : foreach(lc, windowClause)
5916 : {
5917 2422 : WindowClause *wc = lfirst_node(WindowClause, lc);
5918 : ListCell *lc2;
5919 2422 : int optimizedFrameOptions = 0;
5920 :
5921 : Assert(wc->winref <= wflists->maxWinRef);
5922 :
5923 : /* skip any WindowClauses that have no WindowFuncs */
5924 2422 : if (wflists->windowFuncs[wc->winref] == NIL)
5925 20 : continue;
5926 :
5927 2945 : foreach(lc2, wflists->windowFuncs[wc->winref])
5928 : {
5929 : SupportRequestOptimizeWindowClause req;
5930 : SupportRequestOptimizeWindowClause *res;
5931 2437 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5932 : Oid prosupport;
5933 :
5934 2437 : prosupport = get_func_support(wfunc->winfnoid);
5935 :
5936 : /* Check if there's a support function for 'wfunc' */
5937 2437 : if (!OidIsValid(prosupport))
5938 1894 : break; /* can't optimize this WindowClause */
5939 :
5940 730 : req.type = T_SupportRequestOptimizeWindowClause;
5941 730 : req.window_clause = wc;
5942 730 : req.window_func = wfunc;
5943 730 : req.frameOptions = wc->frameOptions;
5944 :
5945 : /* call the support function */
5946 : res = (SupportRequestOptimizeWindowClause *)
5947 730 : DatumGetPointer(OidFunctionCall1(prosupport,
5948 : PointerGetDatum(&req)));
5949 :
5950 : /*
5951 : * Skip to next WindowClause if the support function does not
5952 : * support this request type.
5953 : */
5954 730 : if (res == NULL)
5955 187 : break;
5956 :
5957 : /*
5958 : * Save these frameOptions for the first WindowFunc for this
5959 : * WindowClause.
5960 : */
5961 543 : if (foreach_current_index(lc2) == 0)
5962 523 : optimizedFrameOptions = res->frameOptions;
5963 :
5964 : /*
5965 : * On subsequent WindowFuncs, if the frameOptions are not the same
5966 : * then we're unable to optimize the frameOptions for this
5967 : * WindowClause.
5968 : */
5969 20 : else if (optimizedFrameOptions != res->frameOptions)
5970 0 : break; /* skip to the next WindowClause, if any */
5971 : }
5972 :
5973 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
5974 2402 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5975 : {
5976 : ListCell *lc3;
5977 :
5978 : /* apply the new frame options */
5979 508 : wc->frameOptions = optimizedFrameOptions;
5980 :
5981 : /*
5982 : * We now check to see if changing the frameOptions has caused
5983 : * this WindowClause to be a duplicate of some other WindowClause.
5984 : * This can only happen if we have multiple WindowClauses, so
5985 : * don't bother if there's only 1.
5986 : */
5987 508 : if (list_length(windowClause) == 1)
5988 433 : continue;
5989 :
5990 : /*
5991 : * Do the duplicate check and reuse the existing WindowClause if
5992 : * we find a duplicate.
5993 : */
5994 190 : foreach(lc3, windowClause)
5995 : {
5996 145 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5997 :
5998 : /* skip over the WindowClause we're currently editing */
5999 145 : if (existing_wc == wc)
6000 45 : continue;
6001 :
6002 : /*
6003 : * Perform the same duplicate check that is done in
6004 : * transformWindowFuncCall.
6005 : */
6006 200 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
6007 100 : equal(wc->orderClause, existing_wc->orderClause) &&
6008 100 : wc->frameOptions == existing_wc->frameOptions &&
6009 60 : equal(wc->startOffset, existing_wc->startOffset) &&
6010 30 : equal(wc->endOffset, existing_wc->endOffset))
6011 : {
6012 : ListCell *lc4;
6013 :
6014 : /*
6015 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
6016 : * This required adjusting each WindowFunc's winref and
6017 : * moving the WindowFuncs in 'wc' to the list of
6018 : * WindowFuncs in 'existing_wc'.
6019 : */
6020 65 : foreach(lc4, wflists->windowFuncs[wc->winref])
6021 : {
6022 35 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
6023 :
6024 35 : wfunc->winref = existing_wc->winref;
6025 : }
6026 :
6027 : /* move list items */
6028 60 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
6029 30 : wflists->windowFuncs[wc->winref]);
6030 30 : wflists->windowFuncs[wc->winref] = NIL;
6031 :
6032 : /*
6033 : * transformWindowFuncCall() should have made sure there
6034 : * are no other duplicates, so we needn't bother looking
6035 : * any further.
6036 : */
6037 30 : break;
6038 : }
6039 : }
6040 : }
6041 : }
6042 2217 : }
6043 :
6044 : /*
6045 : * select_active_windows
6046 : * Create a list of the "active" window clauses (ie, those referenced
6047 : * by non-deleted WindowFuncs) in the order they are to be executed.
6048 : */
6049 : static List *
6050 2217 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
6051 : {
6052 2217 : List *windowClause = root->parse->windowClause;
6053 2217 : List *result = NIL;
6054 : ListCell *lc;
6055 2217 : int nActive = 0;
6056 2217 : WindowClauseSortData *actives = palloc_array(WindowClauseSortData,
6057 : list_length(windowClause));
6058 :
6059 : /* First, construct an array of the active windows */
6060 4639 : foreach(lc, windowClause)
6061 : {
6062 2422 : WindowClause *wc = lfirst_node(WindowClause, lc);
6063 :
6064 : /* It's only active if wflists shows some related WindowFuncs */
6065 : Assert(wc->winref <= wflists->maxWinRef);
6066 2422 : if (wflists->windowFuncs[wc->winref] == NIL)
6067 50 : continue;
6068 :
6069 2372 : actives[nActive].wc = wc; /* original clause */
6070 :
6071 : /*
6072 : * For sorting, we want the list of partition keys followed by the
6073 : * list of sort keys. But pathkeys construction will remove duplicates
6074 : * between the two, so we can as well (even though we can't detect all
6075 : * of the duplicates, since some may come from ECs - that might mean
6076 : * we miss optimization chances here). We must, however, ensure that
6077 : * the order of entries is preserved with respect to the ones we do
6078 : * keep.
6079 : *
6080 : * partitionClause and orderClause had their own duplicates removed in
6081 : * parse analysis, so we're only concerned here with removing
6082 : * orderClause entries that also appear in partitionClause.
6083 : */
6084 4744 : actives[nActive].uniqueOrder =
6085 2372 : list_concat_unique(list_copy(wc->partitionClause),
6086 2372 : wc->orderClause);
6087 2372 : nActive++;
6088 : }
6089 :
6090 : /*
6091 : * Sort active windows by their partitioning/ordering clauses, ignoring
6092 : * any framing clauses, so that the windows that need the same sorting are
6093 : * adjacent in the list. When we come to generate paths, this will avoid
6094 : * inserting additional Sort nodes.
6095 : *
6096 : * This is how we implement a specific requirement from the SQL standard,
6097 : * which says that when two or more windows are order-equivalent (i.e.
6098 : * have matching partition and order clauses, even if their names or
6099 : * framing clauses differ), then all peer rows must be presented in the
6100 : * same order in all of them. If we allowed multiple sort nodes for such
6101 : * cases, we'd risk having the peer rows end up in different orders in
6102 : * equivalent windows due to sort instability. (See General Rule 4 of
6103 : * <window clause> in SQL2008 - SQL2016.)
6104 : *
6105 : * Additionally, if the entire list of clauses of one window is a prefix
6106 : * of another, put first the window with stronger sorting requirements.
6107 : * This way we will first sort for stronger window, and won't have to sort
6108 : * again for the weaker one.
6109 : */
6110 2217 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6111 :
6112 : /* build ordered list of the original WindowClause nodes */
6113 4589 : for (int i = 0; i < nActive; i++)
6114 2372 : result = lappend(result, actives[i].wc);
6115 :
6116 2217 : pfree(actives);
6117 :
6118 2217 : return result;
6119 : }
6120 :
6121 : /*
6122 : * name_active_windows
6123 : * Ensure all active windows have unique names.
6124 : *
6125 : * The parser will have checked that user-assigned window names are unique
6126 : * within the Query. Here we assign made-up names to any unnamed
6127 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6128 : * at parse time, because it'd mess up decompilation of views.)
6129 : *
6130 : * activeWindows: result of select_active_windows
6131 : */
6132 : static void
6133 2217 : name_active_windows(List *activeWindows)
6134 : {
6135 2217 : int next_n = 1;
6136 : char newname[16];
6137 : ListCell *lc;
6138 :
6139 4589 : foreach(lc, activeWindows)
6140 : {
6141 2372 : WindowClause *wc = lfirst_node(WindowClause, lc);
6142 :
6143 : /* Nothing to do if it has a name already. */
6144 2372 : if (wc->name)
6145 480 : continue;
6146 :
6147 : /* Select a name not currently present in the list. */
6148 : for (;;)
6149 5 : {
6150 : ListCell *lc2;
6151 :
6152 1897 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6153 4084 : foreach(lc2, activeWindows)
6154 : {
6155 2192 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6156 :
6157 2192 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6158 5 : break; /* matched */
6159 : }
6160 1897 : if (lc2 == NULL)
6161 1892 : break; /* reached the end with no match */
6162 : }
6163 1892 : wc->name = pstrdup(newname);
6164 : }
6165 2217 : }
6166 :
6167 : /*
6168 : * common_prefix_cmp
6169 : * QSort comparison function for WindowClauseSortData
6170 : *
6171 : * Sort the windows by the required sorting clauses. First, compare the sort
6172 : * clauses themselves. Second, if one window's clauses are a prefix of another
6173 : * one's clauses, put the window with more sort clauses first.
6174 : *
6175 : * We purposefully sort by the highest tleSortGroupRef first. Since
6176 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6177 : * and because here we sort the lowest tleSortGroupRefs last, if a
6178 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6179 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6180 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6181 : * reducing the total number of sorts required for the query.
6182 : */
6183 : static int
6184 170 : common_prefix_cmp(const void *a, const void *b)
6185 : {
6186 170 : const WindowClauseSortData *wcsa = a;
6187 170 : const WindowClauseSortData *wcsb = b;
6188 : ListCell *item_a;
6189 : ListCell *item_b;
6190 :
6191 305 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6192 : {
6193 220 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6194 220 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6195 :
6196 220 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6197 85 : return -1;
6198 210 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6199 55 : return 1;
6200 155 : else if (sca->sortop > scb->sortop)
6201 0 : return -1;
6202 155 : else if (sca->sortop < scb->sortop)
6203 20 : return 1;
6204 135 : else if (sca->nulls_first && !scb->nulls_first)
6205 0 : return -1;
6206 135 : else if (!sca->nulls_first && scb->nulls_first)
6207 0 : return 1;
6208 : /* no need to compare eqop, since it is fully determined by sortop */
6209 : }
6210 :
6211 85 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6212 5 : return -1;
6213 80 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6214 25 : return 1;
6215 :
6216 55 : return 0;
6217 : }
6218 :
6219 : /*
6220 : * make_window_input_target
6221 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6222 : *
6223 : * When the query has window functions, this function computes the desired
6224 : * target to be computed by the node just below the first WindowAgg.
6225 : * This tlist must contain all values needed to evaluate the window functions,
6226 : * compute the final target list, and perform any required final sort step.
6227 : * If multiple WindowAggs are needed, each intermediate one adds its window
6228 : * function results onto this base tlist; only the topmost WindowAgg computes
6229 : * the actual desired target list.
6230 : *
6231 : * This function is much like make_group_input_target, though not quite enough
6232 : * like it to share code. As in that function, we flatten most expressions
6233 : * into their component variables. But we do not want to flatten window
6234 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6235 : * evaluations of them, which would be bad (possibly even resulting in
6236 : * inconsistent answers, if they contain volatile functions).
6237 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6238 : * make_group_input_target, because we may no longer have access to the
6239 : * individual Vars in them.
6240 : *
6241 : * Another key difference from make_group_input_target is that we don't
6242 : * flatten Aggref expressions, since those are to be computed below the
6243 : * window functions and just referenced like Vars above that.
6244 : *
6245 : * 'final_target' is the query's final target list (in PathTarget form)
6246 : * 'activeWindows' is the list of active windows previously identified by
6247 : * select_active_windows.
6248 : *
6249 : * The result is the PathTarget to be computed by the plan node immediately
6250 : * below the first WindowAgg node.
6251 : */
6252 : static PathTarget *
6253 2217 : make_window_input_target(PlannerInfo *root,
6254 : PathTarget *final_target,
6255 : List *activeWindows)
6256 : {
6257 : PathTarget *input_target;
6258 : Bitmapset *sgrefs;
6259 : List *flattenable_cols;
6260 : List *flattenable_vars;
6261 : int i;
6262 : ListCell *lc;
6263 :
6264 : Assert(root->parse->hasWindowFuncs);
6265 :
6266 : /*
6267 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6268 : * into a bitmapset for convenient reference below.
6269 : */
6270 2217 : sgrefs = NULL;
6271 4589 : foreach(lc, activeWindows)
6272 : {
6273 2372 : WindowClause *wc = lfirst_node(WindowClause, lc);
6274 : ListCell *lc2;
6275 :
6276 2994 : foreach(lc2, wc->partitionClause)
6277 : {
6278 622 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6279 :
6280 622 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6281 : }
6282 4277 : foreach(lc2, wc->orderClause)
6283 : {
6284 1905 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6285 :
6286 1905 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6287 : }
6288 : }
6289 :
6290 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6291 2373 : foreach(lc, root->processed_groupClause)
6292 : {
6293 156 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6294 :
6295 156 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6296 : }
6297 :
6298 : /*
6299 : * Construct a target containing all the non-flattenable targetlist items,
6300 : * and save aside the others for a moment.
6301 : */
6302 2217 : input_target = create_empty_pathtarget();
6303 2217 : flattenable_cols = NIL;
6304 :
6305 2217 : i = 0;
6306 9324 : foreach(lc, final_target->exprs)
6307 : {
6308 7107 : Expr *expr = (Expr *) lfirst(lc);
6309 7107 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6310 :
6311 : /*
6312 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6313 : * that such items can't contain window functions, so it's okay to
6314 : * compute them below the WindowAgg nodes.)
6315 : */
6316 7107 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6317 : {
6318 : /*
6319 : * Don't want to deconstruct this value, so add it to the input
6320 : * target as-is.
6321 : */
6322 2389 : add_column_to_pathtarget(input_target, expr, sgref);
6323 : }
6324 : else
6325 : {
6326 : /*
6327 : * Column is to be flattened, so just remember the expression for
6328 : * later call to pull_var_clause.
6329 : */
6330 4718 : flattenable_cols = lappend(flattenable_cols, expr);
6331 : }
6332 :
6333 7107 : i++;
6334 : }
6335 :
6336 : /*
6337 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6338 : * add them to the input target if not already present. (Some might be
6339 : * there already because they're used directly as window/group clauses.)
6340 : *
6341 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6342 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6343 : * at higher levels. On the other hand, we should recurse into
6344 : * WindowFuncs to make sure their input expressions are available.
6345 : */
6346 2217 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6347 : PVC_INCLUDE_AGGREGATES |
6348 : PVC_RECURSE_WINDOWFUNCS |
6349 : PVC_INCLUDE_PLACEHOLDERS);
6350 2217 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6351 :
6352 : /* clean up cruft */
6353 2217 : list_free(flattenable_vars);
6354 2217 : list_free(flattenable_cols);
6355 :
6356 : /* XXX this causes some redundant cost calculation ... */
6357 2217 : return set_pathtarget_cost_width(root, input_target);
6358 : }
6359 :
6360 : /*
6361 : * make_pathkeys_for_window
6362 : * Create a pathkeys list describing the required input ordering
6363 : * for the given WindowClause.
6364 : *
6365 : * Modifies wc's partitionClause to remove any clauses which are deemed
6366 : * redundant by the pathkey logic.
6367 : *
6368 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6369 : * In the future we might try to implement windowing using hashing, in which
6370 : * case the ordering could be relaxed, but for now we always sort.
6371 : */
6372 : static List *
6373 4756 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6374 : List *tlist)
6375 : {
6376 4756 : List *window_pathkeys = NIL;
6377 :
6378 : /* Throw error if can't sort */
6379 4756 : if (!grouping_is_sortable(wc->partitionClause))
6380 0 : ereport(ERROR,
6381 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6382 : errmsg("could not implement window PARTITION BY"),
6383 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6384 4756 : if (!grouping_is_sortable(wc->orderClause))
6385 0 : ereport(ERROR,
6386 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6387 : errmsg("could not implement window ORDER BY"),
6388 : errdetail("Window ordering columns must be of sortable datatypes.")));
6389 :
6390 : /*
6391 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6392 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6393 : */
6394 4756 : if (wc->partitionClause != NIL)
6395 : {
6396 : bool sortable;
6397 :
6398 1088 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6399 : &wc->partitionClause,
6400 : tlist,
6401 : true,
6402 : false,
6403 : &sortable,
6404 : false);
6405 :
6406 : Assert(sortable);
6407 : }
6408 :
6409 : /*
6410 : * In principle, we could also consider removing redundant ORDER BY items
6411 : * too as doing so does not alter the result of peer row checks done by
6412 : * the executor. However, we must *not* remove the ordering column for
6413 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6414 : * if it's known to be equal to some partitioning column.
6415 : */
6416 4756 : if (wc->orderClause != NIL)
6417 : {
6418 : List *orderby_pathkeys;
6419 :
6420 3724 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6421 : wc->orderClause,
6422 : tlist);
6423 :
6424 : /* Okay, make the combined pathkeys */
6425 3724 : if (window_pathkeys != NIL)
6426 781 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6427 : else
6428 2943 : window_pathkeys = orderby_pathkeys;
6429 : }
6430 :
6431 4756 : return window_pathkeys;
6432 : }
6433 :
6434 : /*
6435 : * make_sort_input_target
6436 : * Generate appropriate PathTarget for initial input to Sort step.
6437 : *
6438 : * If the query has ORDER BY, this function chooses the target to be computed
6439 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6440 : * project) steps. This might or might not be identical to the query's final
6441 : * output target.
6442 : *
6443 : * The main argument for keeping the sort-input tlist the same as the final
6444 : * is that we avoid a separate projection node (which will be needed if
6445 : * they're different, because Sort can't project). However, there are also
6446 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6447 : * a consistent order of evaluation for any volatile functions in the tlist,
6448 : * and if there's also a LIMIT, we can stop the query without ever computing
6449 : * tlist functions for later rows, which is beneficial for both volatile and
6450 : * expensive functions.
6451 : *
6452 : * Our current policy is to postpone volatile expressions till after the sort
6453 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6454 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6455 : * postpone set-returning expressions, because running them beforehand would
6456 : * bloat the sort dataset, and because it might cause unexpected output order
6457 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6458 : * in the tlist should be evaluated at the same plan step, so that they can
6459 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6460 : * mustn't postpone any SRFs. (Note that in principle that policy should
6461 : * probably get applied to the group/window input targetlists too, but we
6462 : * have not done that historically.) Lastly, expensive expressions are
6463 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6464 : * partial evaluation of the query is possible (if neither is true, we expect
6465 : * to have to evaluate the expressions for every row anyway), or if there are
6466 : * any volatile or set-returning expressions (since once we've put in a
6467 : * projection at all, it won't cost any more to postpone more stuff).
6468 : *
6469 : * Another issue that could potentially be considered here is that
6470 : * evaluating tlist expressions could result in data that's either wider
6471 : * or narrower than the input Vars, thus changing the volume of data that
6472 : * has to go through the Sort. However, we usually have only a very bad
6473 : * idea of the output width of any expression more complex than a Var,
6474 : * so for now it seems too risky to try to optimize on that basis.
6475 : *
6476 : * Note that if we do produce a modified sort-input target, and then the
6477 : * query ends up not using an explicit Sort, no particular harm is done:
6478 : * we'll initially use the modified target for the preceding path nodes,
6479 : * but then change them to the final target with apply_projection_to_path.
6480 : * Moreover, in such a case the guarantees about evaluation order of
6481 : * volatile functions still hold, since the rows are sorted already.
6482 : *
6483 : * This function has some things in common with make_group_input_target and
6484 : * make_window_input_target, though the detailed rules for what to do are
6485 : * different. We never flatten/postpone any grouping or ordering columns;
6486 : * those are needed before the sort. If we do flatten a particular
6487 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6488 : * computed earlier.
6489 : *
6490 : * 'final_target' is the query's final target list (in PathTarget form)
6491 : * 'have_postponed_srfs' is an output argument, see below
6492 : *
6493 : * The result is the PathTarget to be computed by the plan node immediately
6494 : * below the Sort step (and the Distinct step, if any). This will be
6495 : * exactly final_target if we decide a projection step wouldn't be helpful.
6496 : *
6497 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6498 : * any set-returning functions to after the Sort.
6499 : */
6500 : static PathTarget *
6501 58652 : make_sort_input_target(PlannerInfo *root,
6502 : PathTarget *final_target,
6503 : bool *have_postponed_srfs)
6504 : {
6505 58652 : Query *parse = root->parse;
6506 : PathTarget *input_target;
6507 : int ncols;
6508 : bool *col_is_srf;
6509 : bool *postpone_col;
6510 : bool have_srf;
6511 : bool have_volatile;
6512 : bool have_expensive;
6513 : bool have_srf_sortcols;
6514 : bool postpone_srfs;
6515 : List *postponable_cols;
6516 : List *postponable_vars;
6517 : int i;
6518 : ListCell *lc;
6519 :
6520 : /* Shouldn't get here unless query has ORDER BY */
6521 : Assert(parse->sortClause);
6522 :
6523 58652 : *have_postponed_srfs = false; /* default result */
6524 :
6525 : /* Inspect tlist and collect per-column information */
6526 58652 : ncols = list_length(final_target->exprs);
6527 58652 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6528 58652 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6529 58652 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6530 :
6531 58652 : i = 0;
6532 341963 : foreach(lc, final_target->exprs)
6533 : {
6534 283311 : Expr *expr = (Expr *) lfirst(lc);
6535 :
6536 : /*
6537 : * If the column has a sortgroupref, assume it has to be evaluated
6538 : * before sorting. Generally such columns would be ORDER BY, GROUP
6539 : * BY, etc targets. One exception is columns that were removed from
6540 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6541 : * only be Vars anyway. There don't seem to be any cases where it
6542 : * would be worth the trouble to double-check.
6543 : */
6544 283311 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6545 : {
6546 : /*
6547 : * Check for SRF or volatile functions. Check the SRF case first
6548 : * because we must know whether we have any postponed SRFs.
6549 : */
6550 200415 : if (parse->hasTargetSRFs &&
6551 180 : expression_returns_set((Node *) expr))
6552 : {
6553 : /* We'll decide below whether these are postponable */
6554 80 : col_is_srf[i] = true;
6555 80 : have_srf = true;
6556 : }
6557 200155 : else if (contain_volatile_functions((Node *) expr))
6558 : {
6559 : /* Unconditionally postpone */
6560 185 : postpone_col[i] = true;
6561 185 : have_volatile = true;
6562 : }
6563 : else
6564 : {
6565 : /*
6566 : * Else check the cost. XXX it's annoying to have to do this
6567 : * when set_pathtarget_cost_width() just did it. Refactor to
6568 : * allow sharing the work?
6569 : */
6570 : QualCost cost;
6571 :
6572 199970 : cost_qual_eval_node(&cost, (Node *) expr, root);
6573 :
6574 : /*
6575 : * We arbitrarily define "expensive" as "more than 10X
6576 : * cpu_operator_cost". Note this will take in any PL function
6577 : * with default cost.
6578 : */
6579 199970 : if (cost.per_tuple > 10 * cpu_operator_cost)
6580 : {
6581 14339 : postpone_col[i] = true;
6582 14339 : have_expensive = true;
6583 : }
6584 : }
6585 : }
6586 : else
6587 : {
6588 : /* For sortgroupref cols, just check if any contain SRFs */
6589 83076 : if (!have_srf_sortcols &&
6590 83333 : parse->hasTargetSRFs &&
6591 277 : expression_returns_set((Node *) expr))
6592 122 : have_srf_sortcols = true;
6593 : }
6594 :
6595 283311 : i++;
6596 : }
6597 :
6598 : /*
6599 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6600 : */
6601 58652 : postpone_srfs = (have_srf && !have_srf_sortcols);
6602 :
6603 : /*
6604 : * If we don't need a post-sort projection, just return final_target.
6605 : */
6606 58652 : if (!(postpone_srfs || have_volatile ||
6607 58419 : (have_expensive &&
6608 8591 : (parse->limitCount || root->tuple_fraction > 0))))
6609 58389 : return final_target;
6610 :
6611 : /*
6612 : * Report whether the post-sort projection will contain set-returning
6613 : * functions. This is important because it affects whether the Sort can
6614 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6615 : * to return.
6616 : */
6617 263 : *have_postponed_srfs = postpone_srfs;
6618 :
6619 : /*
6620 : * Construct the sort-input target, taking all non-postponable columns and
6621 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6622 : * the postponable ones.
6623 : */
6624 263 : input_target = create_empty_pathtarget();
6625 263 : postponable_cols = NIL;
6626 :
6627 263 : i = 0;
6628 1855 : foreach(lc, final_target->exprs)
6629 : {
6630 1592 : Expr *expr = (Expr *) lfirst(lc);
6631 :
6632 1592 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6633 315 : postponable_cols = lappend(postponable_cols, expr);
6634 : else
6635 1277 : add_column_to_pathtarget(input_target, expr,
6636 1277 : get_pathtarget_sortgroupref(final_target, i));
6637 :
6638 1592 : i++;
6639 : }
6640 :
6641 : /*
6642 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6643 : * postponable columns, and add them to the sort-input target if not
6644 : * already present. (Some might be there already.) We mustn't
6645 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6646 : * would be unable to recompute them.
6647 : */
6648 263 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6649 : PVC_INCLUDE_AGGREGATES |
6650 : PVC_INCLUDE_WINDOWFUNCS |
6651 : PVC_INCLUDE_PLACEHOLDERS);
6652 263 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6653 :
6654 : /* clean up cruft */
6655 263 : list_free(postponable_vars);
6656 263 : list_free(postponable_cols);
6657 :
6658 : /* XXX this represents even more redundant cost calculation ... */
6659 263 : return set_pathtarget_cost_width(root, input_target);
6660 : }
6661 :
6662 : /*
6663 : * get_cheapest_fractional_path
6664 : * Find the cheapest path for retrieving a specified fraction of all
6665 : * the tuples expected to be returned by the given relation.
6666 : *
6667 : * Do not consider parameterized paths. If the caller needs a path for upper
6668 : * rel, it can't have parameterized paths. If the caller needs an append
6669 : * subpath, it could become limited by the treatment of similar
6670 : * parameterization of all the subpaths.
6671 : *
6672 : * We interpret tuple_fraction the same way as grouping_planner.
6673 : *
6674 : * We assume set_cheapest() has been run on the given rel.
6675 : */
6676 : Path *
6677 365559 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6678 : {
6679 365559 : Path *best_path = rel->cheapest_total_path;
6680 : ListCell *l;
6681 :
6682 : /* If all tuples will be retrieved, just return the cheapest-total path */
6683 365559 : if (tuple_fraction <= 0.0)
6684 358515 : return best_path;
6685 :
6686 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6687 7044 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6688 3066 : tuple_fraction /= best_path->rows;
6689 :
6690 18033 : foreach(l, rel->pathlist)
6691 : {
6692 10989 : Path *path = (Path *) lfirst(l);
6693 :
6694 10989 : if (path->param_info)
6695 166 : continue;
6696 :
6697 14602 : if (path == rel->cheapest_total_path ||
6698 3779 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6699 10501 : continue;
6700 :
6701 322 : best_path = path;
6702 : }
6703 :
6704 7044 : return best_path;
6705 : }
6706 :
6707 : /*
6708 : * adjust_paths_for_srfs
6709 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6710 : *
6711 : * The executor can only handle set-returning functions that appear at the
6712 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6713 : * that are not at top level, we need to split up the evaluation into multiple
6714 : * plan levels in which each level satisfies this constraint. This function
6715 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6716 : * output tlist to insert appropriate projection steps.
6717 : *
6718 : * The given targets and targets_contain_srfs lists are from
6719 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6720 : * target in targets.
6721 : */
6722 : static void
6723 10845 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6724 : List *targets, List *targets_contain_srfs)
6725 : {
6726 : ListCell *lc;
6727 :
6728 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6729 : Assert(!linitial_int(targets_contain_srfs));
6730 :
6731 : /* If no SRFs appear at this plan level, nothing to do */
6732 10845 : if (list_length(targets) == 1)
6733 491 : return;
6734 :
6735 : /*
6736 : * Stack SRF-evaluation nodes atop each path for the rel.
6737 : *
6738 : * In principle we should re-run set_cheapest() here to identify the
6739 : * cheapest path, but it seems unlikely that adding the same tlist eval
6740 : * costs to all the paths would change that, so we don't bother. Instead,
6741 : * just assume that the cheapest-startup and cheapest-total paths remain
6742 : * so. (There should be no parameterized paths anymore, so we needn't
6743 : * worry about updating cheapest_parameterized_paths.)
6744 : */
6745 20740 : foreach(lc, rel->pathlist)
6746 : {
6747 10386 : Path *subpath = (Path *) lfirst(lc);
6748 10386 : Path *newpath = subpath;
6749 : ListCell *lc1,
6750 : *lc2;
6751 :
6752 : Assert(subpath->param_info == NULL);
6753 31945 : forboth(lc1, targets, lc2, targets_contain_srfs)
6754 : {
6755 21559 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6756 21559 : bool contains_srfs = (bool) lfirst_int(lc2);
6757 :
6758 : /* If this level doesn't contain SRFs, do regular projection */
6759 21559 : if (contains_srfs)
6760 10436 : newpath = (Path *) create_set_projection_path(root,
6761 : rel,
6762 : newpath,
6763 : thistarget);
6764 : else
6765 11123 : newpath = (Path *) apply_projection_to_path(root,
6766 : rel,
6767 : newpath,
6768 : thistarget);
6769 : }
6770 10386 : lfirst(lc) = newpath;
6771 10386 : if (subpath == rel->cheapest_startup_path)
6772 239 : rel->cheapest_startup_path = newpath;
6773 10386 : if (subpath == rel->cheapest_total_path)
6774 239 : rel->cheapest_total_path = newpath;
6775 : }
6776 :
6777 : /* Likewise for partial paths, if any */
6778 10369 : foreach(lc, rel->partial_pathlist)
6779 : {
6780 15 : Path *subpath = (Path *) lfirst(lc);
6781 15 : Path *newpath = subpath;
6782 : ListCell *lc1,
6783 : *lc2;
6784 :
6785 : Assert(subpath->param_info == NULL);
6786 60 : forboth(lc1, targets, lc2, targets_contain_srfs)
6787 : {
6788 45 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6789 45 : bool contains_srfs = (bool) lfirst_int(lc2);
6790 :
6791 : /* If this level doesn't contain SRFs, do regular projection */
6792 45 : if (contains_srfs)
6793 15 : newpath = (Path *) create_set_projection_path(root,
6794 : rel,
6795 : newpath,
6796 : thistarget);
6797 : else
6798 : {
6799 : /* avoid apply_projection_to_path, in case of multiple refs */
6800 30 : newpath = (Path *) create_projection_path(root,
6801 : rel,
6802 : newpath,
6803 : thistarget);
6804 : }
6805 : }
6806 15 : lfirst(lc) = newpath;
6807 : }
6808 : }
6809 :
6810 : /*
6811 : * expression_planner
6812 : * Perform planner's transformations on a standalone expression.
6813 : *
6814 : * Various utility commands need to evaluate expressions that are not part
6815 : * of a plannable query. They can do so using the executor's regular
6816 : * expression-execution machinery, but first the expression has to be fed
6817 : * through here to transform it from parser output to something executable.
6818 : *
6819 : * Currently, we disallow sublinks in standalone expressions, so there's no
6820 : * real "planning" involved here. (That might not always be true though.)
6821 : * What we must do is run eval_const_expressions to ensure that any function
6822 : * calls are converted to positional notation and function default arguments
6823 : * get inserted. The fact that constant subexpressions get simplified is a
6824 : * side-effect that is useful when the expression will get evaluated more than
6825 : * once. Also, we must fix operator function IDs.
6826 : *
6827 : * This does not return any information about dependencies of the expression.
6828 : * Hence callers should use the results only for the duration of the current
6829 : * query. Callers that would like to cache the results for longer should use
6830 : * expression_planner_with_deps, probably via the plancache.
6831 : *
6832 : * Note: this must not make any damaging changes to the passed-in expression
6833 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6834 : * we first do an expression_tree_mutator-based walk, what is returned will
6835 : * be a new node tree.) The result is constructed in the current memory
6836 : * context; beware that this can leak a lot of additional stuff there, too.
6837 : */
6838 : Expr *
6839 138182 : expression_planner(Expr *expr)
6840 : {
6841 : Node *result;
6842 :
6843 : /*
6844 : * Convert named-argument function calls, insert default arguments and
6845 : * simplify constant subexprs
6846 : */
6847 138182 : result = eval_const_expressions(NULL, (Node *) expr);
6848 :
6849 : /* Fill in opfuncid values if missing */
6850 138162 : fix_opfuncids(result);
6851 :
6852 138162 : return (Expr *) result;
6853 : }
6854 :
6855 : /*
6856 : * expression_planner_with_deps
6857 : * Perform planner's transformations on a standalone expression,
6858 : * returning expression dependency information along with the result.
6859 : *
6860 : * This is identical to expression_planner() except that it also returns
6861 : * information about possible dependencies of the expression, ie identities of
6862 : * objects whose definitions affect the result. As in a PlannedStmt, these
6863 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6864 : */
6865 : Expr *
6866 261 : expression_planner_with_deps(Expr *expr,
6867 : List **relationOids,
6868 : List **invalItems)
6869 : {
6870 : Node *result;
6871 : PlannerGlobal glob;
6872 : PlannerInfo root;
6873 :
6874 : /* Make up dummy planner state so we can use setrefs machinery */
6875 7569 : MemSet(&glob, 0, sizeof(glob));
6876 261 : glob.type = T_PlannerGlobal;
6877 261 : glob.relationOids = NIL;
6878 261 : glob.invalItems = NIL;
6879 :
6880 24534 : MemSet(&root, 0, sizeof(root));
6881 261 : root.type = T_PlannerInfo;
6882 261 : root.glob = &glob;
6883 :
6884 : /*
6885 : * Convert named-argument function calls, insert default arguments and
6886 : * simplify constant subexprs. Collect identities of inlined functions
6887 : * and elided domains, too.
6888 : */
6889 261 : result = eval_const_expressions(&root, (Node *) expr);
6890 :
6891 : /* Fill in opfuncid values if missing */
6892 261 : fix_opfuncids(result);
6893 :
6894 : /*
6895 : * Now walk the finished expression to find anything else we ought to
6896 : * record as an expression dependency.
6897 : */
6898 261 : (void) extract_query_dependencies_walker(result, &root);
6899 :
6900 261 : *relationOids = glob.relationOids;
6901 261 : *invalItems = glob.invalItems;
6902 :
6903 261 : return (Expr *) result;
6904 : }
6905 :
6906 :
6907 : /*
6908 : * plan_cluster_use_sort
6909 : * Use the planner to decide how CLUSTER should implement sorting
6910 : *
6911 : * tableOid is the OID of a table to be clustered on its index indexOid
6912 : * (which is already known to be a btree index). Decide whether it's
6913 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6914 : * Return true to use sorting, false to use an indexscan.
6915 : *
6916 : * Note: caller had better already hold some type of lock on the table.
6917 : */
6918 : bool
6919 141 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6920 : {
6921 : PlannerInfo *root;
6922 : Query *query;
6923 : PlannerGlobal *glob;
6924 : RangeTblEntry *rte;
6925 : RelOptInfo *rel;
6926 : IndexOptInfo *indexInfo;
6927 : QualCost indexExprCost;
6928 : Cost comparisonCost;
6929 : Path *seqScanPath;
6930 : Path seqScanAndSortPath;
6931 : IndexPath *indexScanPath;
6932 : ListCell *lc;
6933 :
6934 : /* We can short-circuit the cost comparison if indexscans are disabled */
6935 141 : if (!enable_indexscan)
6936 20 : return true; /* use sort */
6937 :
6938 : /* Set up mostly-dummy planner state */
6939 121 : query = makeNode(Query);
6940 121 : query->commandType = CMD_SELECT;
6941 :
6942 121 : glob = makeNode(PlannerGlobal);
6943 :
6944 121 : root = makeNode(PlannerInfo);
6945 121 : root->parse = query;
6946 121 : root->glob = glob;
6947 121 : root->query_level = 1;
6948 121 : root->planner_cxt = CurrentMemoryContext;
6949 121 : root->wt_param_id = -1;
6950 121 : root->join_domains = list_make1(makeNode(JoinDomain));
6951 :
6952 : /* Build a minimal RTE for the rel */
6953 121 : rte = makeNode(RangeTblEntry);
6954 121 : rte->rtekind = RTE_RELATION;
6955 121 : rte->relid = tableOid;
6956 121 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6957 121 : rte->rellockmode = AccessShareLock;
6958 121 : rte->lateral = false;
6959 121 : rte->inh = false;
6960 121 : rte->inFromCl = true;
6961 121 : query->rtable = list_make1(rte);
6962 121 : addRTEPermissionInfo(&query->rteperminfos, rte);
6963 :
6964 : /* Set up RTE/RelOptInfo arrays */
6965 121 : setup_simple_rel_arrays(root);
6966 :
6967 : /* Build RelOptInfo */
6968 121 : rel = build_simple_rel(root, 1, NULL);
6969 :
6970 : /* Locate IndexOptInfo for the target index */
6971 121 : indexInfo = NULL;
6972 151 : foreach(lc, rel->indexlist)
6973 : {
6974 151 : indexInfo = lfirst_node(IndexOptInfo, lc);
6975 151 : if (indexInfo->indexoid == indexOid)
6976 121 : break;
6977 : }
6978 :
6979 : /*
6980 : * It's possible that get_relation_info did not generate an IndexOptInfo
6981 : * for the desired index; this could happen if it's not yet reached its
6982 : * indcheckxmin usability horizon, or if it's a system index and we're
6983 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6984 : * trust the index contents but use seqscan-and-sort.
6985 : */
6986 121 : if (lc == NULL) /* not in the list? */
6987 0 : return true; /* use sort */
6988 :
6989 : /*
6990 : * Rather than doing all the pushups that would be needed to use
6991 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6992 : */
6993 121 : rel->rows = rel->tuples;
6994 121 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6995 :
6996 121 : root->total_table_pages = rel->pages;
6997 :
6998 : /*
6999 : * Determine eval cost of the index expressions, if any. We need to
7000 : * charge twice that amount for each tuple comparison that happens during
7001 : * the sort, since tuplesort.c will have to re-evaluate the index
7002 : * expressions each time. (XXX that's pretty inefficient...)
7003 : */
7004 121 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
7005 121 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
7006 :
7007 : /* Estimate the cost of seq scan + sort */
7008 121 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
7009 121 : cost_sort(&seqScanAndSortPath, root, NIL,
7010 : seqScanPath->disabled_nodes,
7011 121 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
7012 : comparisonCost, maintenance_work_mem, -1.0);
7013 :
7014 : /* Estimate the cost of index scan */
7015 121 : indexScanPath = create_index_path(root, indexInfo,
7016 : NIL, NIL, NIL, NIL,
7017 : ForwardScanDirection, false,
7018 : NULL, 1.0, false);
7019 :
7020 121 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
7021 : }
7022 :
7023 : /*
7024 : * plan_create_index_workers
7025 : * Use the planner to decide how many parallel worker processes
7026 : * CREATE INDEX should request for use
7027 : *
7028 : * tableOid is the table on which the index is to be built. indexOid is the
7029 : * OID of an index to be created or reindexed (which must be an index with
7030 : * support for parallel builds - currently btree, GIN, or BRIN).
7031 : *
7032 : * Return value is the number of parallel worker processes to request. It
7033 : * may be unsafe to proceed if this is 0. Note that this does not include the
7034 : * leader participating as a worker (value is always a number of parallel
7035 : * worker processes).
7036 : *
7037 : * Note: caller had better already hold some type of lock on the table and
7038 : * index.
7039 : */
7040 : int
7041 22927 : plan_create_index_workers(Oid tableOid, Oid indexOid)
7042 : {
7043 : PlannerInfo *root;
7044 : Query *query;
7045 : PlannerGlobal *glob;
7046 : RangeTblEntry *rte;
7047 : Relation heap;
7048 : Relation index;
7049 : RelOptInfo *rel;
7050 : int parallel_workers;
7051 : BlockNumber heap_blocks;
7052 : double reltuples;
7053 : double allvisfrac;
7054 :
7055 : /*
7056 : * We don't allow performing parallel operation in standalone backend or
7057 : * when parallelism is disabled.
7058 : */
7059 22927 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
7060 301 : return 0;
7061 :
7062 : /* Set up largely-dummy planner state */
7063 22626 : query = makeNode(Query);
7064 22626 : query->commandType = CMD_SELECT;
7065 :
7066 22626 : glob = makeNode(PlannerGlobal);
7067 :
7068 22626 : root = makeNode(PlannerInfo);
7069 22626 : root->parse = query;
7070 22626 : root->glob = glob;
7071 22626 : root->query_level = 1;
7072 22626 : root->planner_cxt = CurrentMemoryContext;
7073 22626 : root->wt_param_id = -1;
7074 22626 : root->join_domains = list_make1(makeNode(JoinDomain));
7075 :
7076 : /*
7077 : * Build a minimal RTE.
7078 : *
7079 : * Mark the RTE with inh = true. This is a kludge to prevent
7080 : * get_relation_info() from fetching index info, which is necessary
7081 : * because it does not expect that any IndexOptInfo is currently
7082 : * undergoing REINDEX.
7083 : */
7084 22626 : rte = makeNode(RangeTblEntry);
7085 22626 : rte->rtekind = RTE_RELATION;
7086 22626 : rte->relid = tableOid;
7087 22626 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7088 22626 : rte->rellockmode = AccessShareLock;
7089 22626 : rte->lateral = false;
7090 22626 : rte->inh = true;
7091 22626 : rte->inFromCl = true;
7092 22626 : query->rtable = list_make1(rte);
7093 22626 : addRTEPermissionInfo(&query->rteperminfos, rte);
7094 :
7095 : /* Set up RTE/RelOptInfo arrays */
7096 22626 : setup_simple_rel_arrays(root);
7097 :
7098 : /* Build RelOptInfo */
7099 22626 : rel = build_simple_rel(root, 1, NULL);
7100 :
7101 : /* Rels are assumed already locked by the caller */
7102 22626 : heap = table_open(tableOid, NoLock);
7103 22626 : index = index_open(indexOid, NoLock);
7104 :
7105 : /*
7106 : * Determine if it's safe to proceed.
7107 : *
7108 : * Currently, parallel workers can't access the leader's temporary tables.
7109 : * Furthermore, any index predicate or index expressions must be parallel
7110 : * safe.
7111 : */
7112 22626 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7113 21243 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7114 21153 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7115 : {
7116 1473 : parallel_workers = 0;
7117 1473 : goto done;
7118 : }
7119 :
7120 : /*
7121 : * If parallel_workers storage parameter is set for the table, accept that
7122 : * as the number of parallel worker processes to launch (though still cap
7123 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7124 : * consider any other factor when parallel_workers is set. (e.g., memory
7125 : * use by workers.)
7126 : */
7127 21153 : if (rel->rel_parallel_workers != -1)
7128 : {
7129 57 : parallel_workers = Min(rel->rel_parallel_workers,
7130 : max_parallel_maintenance_workers);
7131 57 : goto done;
7132 : }
7133 :
7134 : /*
7135 : * Estimate heap relation size ourselves, since rel->pages cannot be
7136 : * trusted (heap RTE was marked as inheritance parent)
7137 : */
7138 21096 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7139 :
7140 : /*
7141 : * Determine number of workers to scan the heap relation using generic
7142 : * model
7143 : */
7144 21096 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7145 : max_parallel_maintenance_workers);
7146 :
7147 : /*
7148 : * Cap workers based on available maintenance_work_mem as needed.
7149 : *
7150 : * Note that each tuplesort participant receives an even share of the
7151 : * total maintenance_work_mem budget. Aim to leave participants
7152 : * (including the leader as a participant) with no less than 32MB of
7153 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7154 : * immediately past the threshold of being capable of launching a single
7155 : * parallel worker to sort.
7156 : */
7157 21205 : while (parallel_workers > 0 &&
7158 219 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7159 109 : parallel_workers--;
7160 :
7161 21096 : done:
7162 22626 : index_close(index, NoLock);
7163 22626 : table_close(heap, NoLock);
7164 :
7165 22626 : return parallel_workers;
7166 : }
7167 :
7168 : /*
7169 : * add_paths_to_grouping_rel
7170 : *
7171 : * Add non-partial paths to grouping relation.
7172 : */
7173 : static void
7174 34887 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7175 : RelOptInfo *grouped_rel,
7176 : RelOptInfo *partially_grouped_rel,
7177 : const AggClauseCosts *agg_costs,
7178 : grouping_sets_data *gd,
7179 : GroupPathExtraData *extra)
7180 : {
7181 34887 : Query *parse = root->parse;
7182 34887 : Path *cheapest_path = input_rel->cheapest_total_path;
7183 34887 : Path *cheapest_partially_grouped_path = NULL;
7184 : ListCell *lc;
7185 34887 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7186 34887 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7187 34887 : List *havingQual = (List *) extra->havingQual;
7188 34887 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7189 34887 : double dNumGroups = 0;
7190 34887 : double dNumFinalGroups = 0;
7191 :
7192 : /*
7193 : * Estimate number of groups for non-split aggregation.
7194 : */
7195 34887 : dNumGroups = get_number_of_groups(root,
7196 : cheapest_path->rows,
7197 : gd,
7198 : extra->targetList);
7199 :
7200 34887 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7201 : {
7202 2495 : cheapest_partially_grouped_path =
7203 : partially_grouped_rel->cheapest_total_path;
7204 :
7205 : /*
7206 : * Estimate number of groups for final phase of partial aggregation.
7207 : */
7208 : dNumFinalGroups =
7209 2495 : get_number_of_groups(root,
7210 : cheapest_partially_grouped_path->rows,
7211 : gd,
7212 : extra->targetList);
7213 : }
7214 :
7215 34887 : if (can_sort)
7216 : {
7217 : /*
7218 : * Use any available suitably-sorted path as input, and also consider
7219 : * sorting the cheapest-total path and incremental sort on any paths
7220 : * with presorted keys.
7221 : */
7222 72324 : foreach(lc, input_rel->pathlist)
7223 : {
7224 : ListCell *lc2;
7225 37442 : Path *path = (Path *) lfirst(lc);
7226 37442 : Path *path_save = path;
7227 37442 : List *pathkey_orderings = NIL;
7228 :
7229 : /* generate alternative group orderings that might be useful */
7230 37442 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7231 :
7232 : Assert(list_length(pathkey_orderings) > 0);
7233 :
7234 75001 : foreach(lc2, pathkey_orderings)
7235 : {
7236 37559 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7237 :
7238 : /* restore the path (we replace it in the loop) */
7239 37559 : path = path_save;
7240 :
7241 37559 : path = make_ordered_path(root,
7242 : grouped_rel,
7243 : path,
7244 : cheapest_path,
7245 : info->pathkeys,
7246 : -1.0);
7247 37559 : if (path == NULL)
7248 306 : continue;
7249 :
7250 : /* Now decide what to stick atop it */
7251 37253 : if (parse->groupingSets)
7252 : {
7253 975 : consider_groupingsets_paths(root, grouped_rel,
7254 : path, true, can_hash,
7255 : gd, agg_costs, dNumGroups);
7256 : }
7257 36278 : else if (parse->hasAggs)
7258 : {
7259 : /*
7260 : * We have aggregation, possibly with plain GROUP BY. Make
7261 : * an AggPath.
7262 : */
7263 35626 : add_path(grouped_rel, (Path *)
7264 35626 : create_agg_path(root,
7265 : grouped_rel,
7266 : path,
7267 35626 : grouped_rel->reltarget,
7268 35626 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7269 : AGGSPLIT_SIMPLE,
7270 : info->clauses,
7271 : havingQual,
7272 : agg_costs,
7273 : dNumGroups));
7274 : }
7275 652 : else if (parse->groupClause)
7276 : {
7277 : /*
7278 : * We have GROUP BY without aggregation or grouping sets.
7279 : * Make a GroupPath.
7280 : */
7281 652 : add_path(grouped_rel, (Path *)
7282 652 : create_group_path(root,
7283 : grouped_rel,
7284 : path,
7285 : info->clauses,
7286 : havingQual,
7287 : dNumGroups));
7288 : }
7289 : else
7290 : {
7291 : /* Other cases should have been handled above */
7292 : Assert(false);
7293 : }
7294 : }
7295 : }
7296 :
7297 : /*
7298 : * Instead of operating directly on the input relation, we can
7299 : * consider finalizing a partially aggregated path.
7300 : */
7301 34882 : if (partially_grouped_rel != NULL)
7302 : {
7303 6426 : foreach(lc, partially_grouped_rel->pathlist)
7304 : {
7305 : ListCell *lc2;
7306 3931 : Path *path = (Path *) lfirst(lc);
7307 3931 : Path *path_save = path;
7308 3931 : List *pathkey_orderings = NIL;
7309 :
7310 : /* generate alternative group orderings that might be useful */
7311 3931 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7312 :
7313 : Assert(list_length(pathkey_orderings) > 0);
7314 :
7315 : /* process all potentially interesting grouping reorderings */
7316 7862 : foreach(lc2, pathkey_orderings)
7317 : {
7318 3931 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7319 :
7320 : /* restore the path (we replace it in the loop) */
7321 3931 : path = path_save;
7322 :
7323 3931 : path = make_ordered_path(root,
7324 : grouped_rel,
7325 : path,
7326 : cheapest_partially_grouped_path,
7327 : info->pathkeys,
7328 : -1.0);
7329 :
7330 3931 : if (path == NULL)
7331 172 : continue;
7332 :
7333 3759 : if (parse->hasAggs)
7334 3557 : add_path(grouped_rel, (Path *)
7335 3557 : create_agg_path(root,
7336 : grouped_rel,
7337 : path,
7338 3557 : grouped_rel->reltarget,
7339 3557 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7340 : AGGSPLIT_FINAL_DESERIAL,
7341 : info->clauses,
7342 : havingQual,
7343 : agg_final_costs,
7344 : dNumFinalGroups));
7345 : else
7346 202 : add_path(grouped_rel, (Path *)
7347 202 : create_group_path(root,
7348 : grouped_rel,
7349 : path,
7350 : info->clauses,
7351 : havingQual,
7352 : dNumFinalGroups));
7353 :
7354 : }
7355 : }
7356 : }
7357 : }
7358 :
7359 34887 : if (can_hash)
7360 : {
7361 4940 : if (parse->groupingSets)
7362 : {
7363 : /*
7364 : * Try for a hash-only groupingsets path over unsorted input.
7365 : */
7366 807 : consider_groupingsets_paths(root, grouped_rel,
7367 : cheapest_path, false, true,
7368 : gd, agg_costs, dNumGroups);
7369 : }
7370 : else
7371 : {
7372 : /*
7373 : * Generate a HashAgg Path. We just need an Agg over the
7374 : * cheapest-total input path, since input order won't matter.
7375 : */
7376 4133 : add_path(grouped_rel, (Path *)
7377 4133 : create_agg_path(root, grouped_rel,
7378 : cheapest_path,
7379 4133 : grouped_rel->reltarget,
7380 : AGG_HASHED,
7381 : AGGSPLIT_SIMPLE,
7382 : root->processed_groupClause,
7383 : havingQual,
7384 : agg_costs,
7385 : dNumGroups));
7386 : }
7387 :
7388 : /*
7389 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7390 : * grouped path, assuming there is one
7391 : */
7392 4940 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7393 : {
7394 1182 : add_path(grouped_rel, (Path *)
7395 1182 : create_agg_path(root,
7396 : grouped_rel,
7397 : cheapest_partially_grouped_path,
7398 1182 : grouped_rel->reltarget,
7399 : AGG_HASHED,
7400 : AGGSPLIT_FINAL_DESERIAL,
7401 : root->processed_groupClause,
7402 : havingQual,
7403 : agg_final_costs,
7404 : dNumFinalGroups));
7405 : }
7406 : }
7407 :
7408 : /*
7409 : * When partitionwise aggregate is used, we might have fully aggregated
7410 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7411 : * consider a path for grouped_rel consisting of a Parallel Append of
7412 : * non-partial paths from each child.
7413 : */
7414 34887 : if (grouped_rel->partial_pathlist != NIL)
7415 265 : gather_grouping_paths(root, grouped_rel);
7416 34887 : }
7417 :
7418 : /*
7419 : * create_partial_grouping_paths
7420 : *
7421 : * Create a new upper relation representing the result of partial aggregation
7422 : * and populate it with appropriate paths. Note that we don't finalize the
7423 : * lists of paths here, so the caller can add additional partial or non-partial
7424 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7425 : * the returned upper relation.
7426 : *
7427 : * All paths for this new upper relation -- both partial and non-partial --
7428 : * have been partially aggregated but require a subsequent FinalizeAggregate
7429 : * step.
7430 : *
7431 : * NB: This function is allowed to return NULL if it determines that there is
7432 : * no real need to create a new RelOptInfo.
7433 : */
7434 : static RelOptInfo *
7435 31469 : create_partial_grouping_paths(PlannerInfo *root,
7436 : RelOptInfo *grouped_rel,
7437 : RelOptInfo *input_rel,
7438 : grouping_sets_data *gd,
7439 : GroupPathExtraData *extra,
7440 : bool force_rel_creation)
7441 : {
7442 31469 : Query *parse = root->parse;
7443 : RelOptInfo *partially_grouped_rel;
7444 31469 : RelOptInfo *eager_agg_rel = NULL;
7445 31469 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7446 31469 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7447 31469 : Path *cheapest_partial_path = NULL;
7448 31469 : Path *cheapest_total_path = NULL;
7449 31469 : double dNumPartialGroups = 0;
7450 31469 : double dNumPartialPartialGroups = 0;
7451 : ListCell *lc;
7452 31469 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7453 31469 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7454 :
7455 : /*
7456 : * Check whether any partially aggregated paths have been generated
7457 : * through eager aggregation.
7458 : */
7459 31469 : if (input_rel->grouped_rel &&
7460 777 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7461 777 : input_rel->grouped_rel->pathlist != NIL)
7462 727 : eager_agg_rel = input_rel->grouped_rel;
7463 :
7464 : /*
7465 : * Consider whether we should generate partially aggregated non-partial
7466 : * paths. We can only do this if we have a non-partial path, and only if
7467 : * the parent of the input rel is performing partial partitionwise
7468 : * aggregation. (Note that extra->patype is the type of partitionwise
7469 : * aggregation being used at the parent level, not this level.)
7470 : */
7471 31469 : if (input_rel->pathlist != NIL &&
7472 31469 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7473 713 : cheapest_total_path = input_rel->cheapest_total_path;
7474 :
7475 : /*
7476 : * If parallelism is possible for grouped_rel, then we should consider
7477 : * generating partially-grouped partial paths. However, if the input rel
7478 : * has no partial paths, then we can't.
7479 : */
7480 31469 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7481 2747 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7482 :
7483 : /*
7484 : * If we can't partially aggregate partial paths, and we can't partially
7485 : * aggregate non-partial paths, and no partially aggregated paths were
7486 : * generated by eager aggregation, then don't bother creating the new
7487 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7488 : */
7489 31469 : if (cheapest_total_path == NULL &&
7490 28429 : cheapest_partial_path == NULL &&
7491 28342 : eager_agg_rel == NULL &&
7492 28342 : !force_rel_creation)
7493 28261 : return NULL;
7494 :
7495 : /*
7496 : * Build a new upper relation to represent the result of partially
7497 : * aggregating the rows from the input relation.
7498 : */
7499 3208 : partially_grouped_rel = fetch_upper_rel(root,
7500 : UPPERREL_PARTIAL_GROUP_AGG,
7501 : grouped_rel->relids);
7502 3208 : partially_grouped_rel->consider_parallel =
7503 3208 : grouped_rel->consider_parallel;
7504 3208 : partially_grouped_rel->pgs_mask = grouped_rel->pgs_mask;
7505 3208 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7506 3208 : partially_grouped_rel->serverid = grouped_rel->serverid;
7507 3208 : partially_grouped_rel->userid = grouped_rel->userid;
7508 3208 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7509 3208 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7510 :
7511 : /*
7512 : * Build target list for partial aggregate paths. These paths cannot just
7513 : * emit the same tlist as regular aggregate paths, because (1) we must
7514 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7515 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7516 : */
7517 3208 : partially_grouped_rel->reltarget =
7518 3208 : make_partial_grouping_target(root, grouped_rel->reltarget,
7519 : extra->havingQual);
7520 :
7521 3208 : if (!extra->partial_costs_set)
7522 : {
7523 : /*
7524 : * Collect statistics about aggregates for estimating costs of
7525 : * performing aggregation in parallel.
7526 : */
7527 11550 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7528 11550 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7529 1925 : if (parse->hasAggs)
7530 : {
7531 : /* partial phase */
7532 1816 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7533 : agg_partial_costs);
7534 :
7535 : /* final phase */
7536 1816 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7537 : agg_final_costs);
7538 : }
7539 :
7540 1925 : extra->partial_costs_set = true;
7541 : }
7542 :
7543 : /* Estimate number of partial groups. */
7544 3208 : if (cheapest_total_path != NULL)
7545 : dNumPartialGroups =
7546 713 : get_number_of_groups(root,
7547 : cheapest_total_path->rows,
7548 : gd,
7549 : extra->targetList);
7550 3208 : if (cheapest_partial_path != NULL)
7551 : dNumPartialPartialGroups =
7552 2747 : get_number_of_groups(root,
7553 : cheapest_partial_path->rows,
7554 : gd,
7555 : extra->targetList);
7556 :
7557 3208 : if (can_sort && cheapest_total_path != NULL)
7558 : {
7559 : /* This should have been checked previously */
7560 : Assert(parse->hasAggs || parse->groupClause);
7561 :
7562 : /*
7563 : * Use any available suitably-sorted path as input, and also consider
7564 : * sorting the cheapest partial path.
7565 : */
7566 1426 : foreach(lc, input_rel->pathlist)
7567 : {
7568 : ListCell *lc2;
7569 713 : Path *path = (Path *) lfirst(lc);
7570 713 : Path *path_save = path;
7571 713 : List *pathkey_orderings = NIL;
7572 :
7573 : /* generate alternative group orderings that might be useful */
7574 713 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7575 :
7576 : Assert(list_length(pathkey_orderings) > 0);
7577 :
7578 : /* process all potentially interesting grouping reorderings */
7579 1426 : foreach(lc2, pathkey_orderings)
7580 : {
7581 713 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7582 :
7583 : /* restore the path (we replace it in the loop) */
7584 713 : path = path_save;
7585 :
7586 713 : path = make_ordered_path(root,
7587 : partially_grouped_rel,
7588 : path,
7589 : cheapest_total_path,
7590 : info->pathkeys,
7591 : -1.0);
7592 :
7593 713 : if (path == NULL)
7594 0 : continue;
7595 :
7596 713 : if (parse->hasAggs)
7597 653 : add_path(partially_grouped_rel, (Path *)
7598 653 : create_agg_path(root,
7599 : partially_grouped_rel,
7600 : path,
7601 653 : partially_grouped_rel->reltarget,
7602 653 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7603 : AGGSPLIT_INITIAL_SERIAL,
7604 : info->clauses,
7605 : NIL,
7606 : agg_partial_costs,
7607 : dNumPartialGroups));
7608 : else
7609 60 : add_path(partially_grouped_rel, (Path *)
7610 60 : create_group_path(root,
7611 : partially_grouped_rel,
7612 : path,
7613 : info->clauses,
7614 : NIL,
7615 : dNumPartialGroups));
7616 : }
7617 : }
7618 : }
7619 :
7620 3208 : if (can_sort && cheapest_partial_path != NULL)
7621 : {
7622 : /* Similar to above logic, but for partial paths. */
7623 5869 : foreach(lc, input_rel->partial_pathlist)
7624 : {
7625 : ListCell *lc2;
7626 3122 : Path *path = (Path *) lfirst(lc);
7627 3122 : Path *path_save = path;
7628 3122 : List *pathkey_orderings = NIL;
7629 :
7630 : /* generate alternative group orderings that might be useful */
7631 3122 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7632 :
7633 : Assert(list_length(pathkey_orderings) > 0);
7634 :
7635 : /* process all potentially interesting grouping reorderings */
7636 6244 : foreach(lc2, pathkey_orderings)
7637 : {
7638 3122 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7639 :
7640 :
7641 : /* restore the path (we replace it in the loop) */
7642 3122 : path = path_save;
7643 :
7644 3122 : path = make_ordered_path(root,
7645 : partially_grouped_rel,
7646 : path,
7647 : cheapest_partial_path,
7648 : info->pathkeys,
7649 : -1.0);
7650 :
7651 3122 : if (path == NULL)
7652 5 : continue;
7653 :
7654 3117 : if (parse->hasAggs)
7655 3018 : add_partial_path(partially_grouped_rel, (Path *)
7656 3018 : create_agg_path(root,
7657 : partially_grouped_rel,
7658 : path,
7659 3018 : partially_grouped_rel->reltarget,
7660 3018 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7661 : AGGSPLIT_INITIAL_SERIAL,
7662 : info->clauses,
7663 : NIL,
7664 : agg_partial_costs,
7665 : dNumPartialPartialGroups));
7666 : else
7667 99 : add_partial_path(partially_grouped_rel, (Path *)
7668 99 : create_group_path(root,
7669 : partially_grouped_rel,
7670 : path,
7671 : info->clauses,
7672 : NIL,
7673 : dNumPartialPartialGroups));
7674 : }
7675 : }
7676 : }
7677 :
7678 : /*
7679 : * Add a partially-grouped HashAgg Path where possible
7680 : */
7681 3208 : if (can_hash && cheapest_total_path != NULL)
7682 : {
7683 : /* Checked above */
7684 : Assert(parse->hasAggs || parse->groupClause);
7685 :
7686 713 : add_path(partially_grouped_rel, (Path *)
7687 713 : create_agg_path(root,
7688 : partially_grouped_rel,
7689 : cheapest_total_path,
7690 713 : partially_grouped_rel->reltarget,
7691 : AGG_HASHED,
7692 : AGGSPLIT_INITIAL_SERIAL,
7693 : root->processed_groupClause,
7694 : NIL,
7695 : agg_partial_costs,
7696 : dNumPartialGroups));
7697 : }
7698 :
7699 : /*
7700 : * Now add a partially-grouped HashAgg partial Path where possible
7701 : */
7702 3208 : if (can_hash && cheapest_partial_path != NULL)
7703 : {
7704 1434 : add_partial_path(partially_grouped_rel, (Path *)
7705 1434 : create_agg_path(root,
7706 : partially_grouped_rel,
7707 : cheapest_partial_path,
7708 1434 : partially_grouped_rel->reltarget,
7709 : AGG_HASHED,
7710 : AGGSPLIT_INITIAL_SERIAL,
7711 : root->processed_groupClause,
7712 : NIL,
7713 : agg_partial_costs,
7714 : dNumPartialPartialGroups));
7715 : }
7716 :
7717 : /*
7718 : * Add any partially aggregated paths generated by eager aggregation to
7719 : * the new upper relation after applying projection steps as needed.
7720 : */
7721 3208 : if (eager_agg_rel)
7722 : {
7723 : /* Add the paths */
7724 1894 : foreach(lc, eager_agg_rel->pathlist)
7725 : {
7726 1167 : Path *path = (Path *) lfirst(lc);
7727 :
7728 : /* Shouldn't have any parameterized paths anymore */
7729 : Assert(path->param_info == NULL);
7730 :
7731 1167 : path = (Path *) create_projection_path(root,
7732 : partially_grouped_rel,
7733 : path,
7734 1167 : partially_grouped_rel->reltarget);
7735 :
7736 1167 : add_path(partially_grouped_rel, path);
7737 : }
7738 :
7739 : /*
7740 : * Likewise add the partial paths, but only if parallelism is possible
7741 : * for partially_grouped_rel.
7742 : */
7743 727 : if (partially_grouped_rel->consider_parallel)
7744 : {
7745 1690 : foreach(lc, eager_agg_rel->partial_pathlist)
7746 : {
7747 1010 : Path *path = (Path *) lfirst(lc);
7748 :
7749 : /* Shouldn't have any parameterized paths anymore */
7750 : Assert(path->param_info == NULL);
7751 :
7752 1010 : path = (Path *) create_projection_path(root,
7753 : partially_grouped_rel,
7754 : path,
7755 1010 : partially_grouped_rel->reltarget);
7756 :
7757 1010 : add_partial_path(partially_grouped_rel, path);
7758 : }
7759 : }
7760 : }
7761 :
7762 : /*
7763 : * If there is an FDW that's responsible for all baserels of the query,
7764 : * let it consider adding partially grouped ForeignPaths.
7765 : */
7766 3208 : if (partially_grouped_rel->fdwroutine &&
7767 3 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7768 : {
7769 3 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7770 :
7771 3 : fdwroutine->GetForeignUpperPaths(root,
7772 : UPPERREL_PARTIAL_GROUP_AGG,
7773 : input_rel, partially_grouped_rel,
7774 : extra);
7775 : }
7776 :
7777 3208 : return partially_grouped_rel;
7778 : }
7779 :
7780 : /*
7781 : * make_ordered_path
7782 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7783 : * return NULL if it doesn't make sense to generate an ordered path in
7784 : * this case.
7785 : */
7786 : static Path *
7787 49313 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7788 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7789 : {
7790 : bool is_sorted;
7791 : int presorted_keys;
7792 :
7793 49313 : is_sorted = pathkeys_count_contained_in(pathkeys,
7794 : path->pathkeys,
7795 : &presorted_keys);
7796 :
7797 49313 : if (!is_sorted)
7798 : {
7799 : /*
7800 : * Try at least sorting the cheapest path and also try incrementally
7801 : * sorting any path which is partially sorted already (no need to deal
7802 : * with paths which have presorted keys when incremental sort is
7803 : * disabled unless it's the cheapest input path).
7804 : */
7805 12980 : if (path != cheapest_path &&
7806 2152 : (presorted_keys == 0 || !enable_incremental_sort))
7807 1022 : return NULL;
7808 :
7809 : /*
7810 : * We've no need to consider both a sort and incremental sort. We'll
7811 : * just do a sort if there are no presorted keys and an incremental
7812 : * sort when there are presorted keys.
7813 : */
7814 11958 : if (presorted_keys == 0 || !enable_incremental_sort)
7815 10682 : path = (Path *) create_sort_path(root,
7816 : rel,
7817 : path,
7818 : pathkeys,
7819 : limit_tuples);
7820 : else
7821 1276 : path = (Path *) create_incremental_sort_path(root,
7822 : rel,
7823 : path,
7824 : pathkeys,
7825 : presorted_keys,
7826 : limit_tuples);
7827 : }
7828 :
7829 48291 : return path;
7830 : }
7831 :
7832 : /*
7833 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7834 : * grouping relation.
7835 : *
7836 : * generate_useful_gather_paths does most of the work, but we also consider a
7837 : * special case: we could try sorting the data by the group_pathkeys and then
7838 : * applying Gather Merge.
7839 : *
7840 : * NB: This function shouldn't be used for anything other than a grouped or
7841 : * partially grouped relation not only because of the fact that it explicitly
7842 : * references group_pathkeys but we pass "true" as the third argument to
7843 : * generate_useful_gather_paths().
7844 : */
7845 : static void
7846 2592 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7847 : {
7848 : ListCell *lc;
7849 : Path *cheapest_partial_path;
7850 : List *groupby_pathkeys;
7851 :
7852 : /*
7853 : * This occurs after any partial aggregation has taken place, so trim off
7854 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7855 : */
7856 2592 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7857 15 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7858 : root->num_groupby_pathkeys);
7859 : else
7860 2577 : groupby_pathkeys = root->group_pathkeys;
7861 :
7862 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7863 2592 : generate_useful_gather_paths(root, rel, true);
7864 :
7865 2592 : cheapest_partial_path = linitial(rel->partial_pathlist);
7866 :
7867 : /* XXX Shouldn't this also consider the group-key-reordering? */
7868 6156 : foreach(lc, rel->partial_pathlist)
7869 : {
7870 3564 : Path *path = (Path *) lfirst(lc);
7871 : bool is_sorted;
7872 : int presorted_keys;
7873 : double total_groups;
7874 :
7875 3564 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7876 : path->pathkeys,
7877 : &presorted_keys);
7878 :
7879 3564 : if (is_sorted)
7880 2331 : continue;
7881 :
7882 : /*
7883 : * Try at least sorting the cheapest path and also try incrementally
7884 : * sorting any path which is partially sorted already (no need to deal
7885 : * with paths which have presorted keys when incremental sort is
7886 : * disabled unless it's the cheapest input path).
7887 : */
7888 1233 : if (path != cheapest_partial_path &&
7889 0 : (presorted_keys == 0 || !enable_incremental_sort))
7890 0 : continue;
7891 :
7892 : /*
7893 : * We've no need to consider both a sort and incremental sort. We'll
7894 : * just do a sort if there are no presorted keys and an incremental
7895 : * sort when there are presorted keys.
7896 : */
7897 1233 : if (presorted_keys == 0 || !enable_incremental_sort)
7898 1233 : path = (Path *) create_sort_path(root, rel, path,
7899 : groupby_pathkeys,
7900 : -1.0);
7901 : else
7902 0 : path = (Path *) create_incremental_sort_path(root,
7903 : rel,
7904 : path,
7905 : groupby_pathkeys,
7906 : presorted_keys,
7907 : -1.0);
7908 1233 : total_groups = compute_gather_rows(path);
7909 : path = (Path *)
7910 1233 : create_gather_merge_path(root,
7911 : rel,
7912 : path,
7913 1233 : rel->reltarget,
7914 : groupby_pathkeys,
7915 : NULL,
7916 : &total_groups);
7917 :
7918 1233 : add_path(rel, path);
7919 : }
7920 2592 : }
7921 :
7922 : /*
7923 : * can_partial_agg
7924 : *
7925 : * Determines whether or not partial grouping and/or aggregation is possible.
7926 : * Returns true when possible, false otherwise.
7927 : */
7928 : static bool
7929 33805 : can_partial_agg(PlannerInfo *root)
7930 : {
7931 33805 : Query *parse = root->parse;
7932 :
7933 33805 : if (!parse->hasAggs && parse->groupClause == NIL)
7934 : {
7935 : /*
7936 : * We don't know how to do parallel aggregation unless we have either
7937 : * some aggregates or a grouping clause.
7938 : */
7939 0 : return false;
7940 : }
7941 33805 : else if (parse->groupingSets)
7942 : {
7943 : /* We don't know how to do grouping sets in parallel. */
7944 886 : return false;
7945 : }
7946 32919 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7947 : {
7948 : /* Insufficient support for partial mode. */
7949 3185 : return false;
7950 : }
7951 :
7952 : /* Everything looks good. */
7953 29734 : return true;
7954 : }
7955 :
7956 : /*
7957 : * apply_scanjoin_target_to_paths
7958 : *
7959 : * Adjust the final scan/join relation, and recursively all of its children,
7960 : * to generate the final scan/join target. It would be more correct to model
7961 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7962 : * for each child relation, but doing it this way is noticeably cheaper.
7963 : * Maybe that problem can be solved at some point, but for now we do this.
7964 : *
7965 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7966 : * the same expressions as the existing reltarget, so we need only insert the
7967 : * appropriate sortgroupref information. By avoiding the creation of
7968 : * projection paths we save effort both immediately and at plan creation time.
7969 : */
7970 : static void
7971 413173 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7972 : RelOptInfo *rel,
7973 : List *scanjoin_targets,
7974 : List *scanjoin_targets_contain_srfs,
7975 : bool scanjoin_target_parallel_safe,
7976 : bool tlist_same_exprs)
7977 : {
7978 413173 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7979 : PathTarget *scanjoin_target;
7980 : ListCell *lc;
7981 :
7982 : /* This recurses, so be paranoid. */
7983 413173 : check_stack_depth();
7984 :
7985 : /*
7986 : * If the rel only has Append and MergeAppend paths, we want to drop its
7987 : * existing paths and generate new ones. This function would still be
7988 : * correct if we kept the existing paths: we'd modify them to generate the
7989 : * correct target above the partitioning Append, and then they'd compete
7990 : * on cost with paths generating the target below the Append. However, in
7991 : * our current cost model the latter way is always the same or cheaper
7992 : * cost, so modifying the existing paths would just be useless work.
7993 : * Moreover, when the cost is the same, varying roundoff errors might
7994 : * sometimes allow an existing path to be picked, resulting in undesirable
7995 : * cross-platform plan variations. So we drop old paths and thereby force
7996 : * the work to be done below the Append.
7997 : *
7998 : * However, there are several cases when this optimization is not safe. If
7999 : * the rel isn't partitioned, then none of the paths will be Append or
8000 : * MergeAppend paths, so we should definitely not do this. If it is
8001 : * partitioned but is a joinrel, it may have Append and MergeAppend paths,
8002 : * but it can also have join paths that we can't afford to discard.
8003 : *
8004 : * Some care is needed, because we have to allow
8005 : * generate_useful_gather_paths to see the old partial paths in the next
8006 : * stanza. Hence, zap the main pathlist here, then allow
8007 : * generate_useful_gather_paths to add path(s) to the main list, and
8008 : * finally zap the partial pathlist.
8009 : */
8010 413173 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8011 8729 : rel->pathlist = NIL;
8012 :
8013 : /*
8014 : * If the scan/join target is not parallel-safe, partial paths cannot
8015 : * generate it.
8016 : */
8017 413173 : if (!scanjoin_target_parallel_safe)
8018 : {
8019 : /*
8020 : * Since we can't generate the final scan/join target in parallel
8021 : * workers, this is our last opportunity to use any partial paths that
8022 : * exist; so build Gather path(s) that use them and emit whatever the
8023 : * current reltarget is. We don't do this in the case where the
8024 : * target is parallel-safe, since we will be able to generate superior
8025 : * paths by doing it after the final scan/join target has been
8026 : * applied.
8027 : */
8028 53693 : generate_useful_gather_paths(root, rel, false);
8029 :
8030 : /* Can't use parallel query above this level. */
8031 53693 : rel->partial_pathlist = NIL;
8032 53693 : rel->consider_parallel = false;
8033 : }
8034 :
8035 : /* Finish dropping old paths for a partitioned rel, per comment above */
8036 413173 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8037 8729 : rel->partial_pathlist = NIL;
8038 :
8039 : /* Extract SRF-free scan/join target. */
8040 413173 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
8041 :
8042 : /*
8043 : * Apply the SRF-free scan/join target to each existing path.
8044 : *
8045 : * If the tlist exprs are the same, we can just inject the sortgroupref
8046 : * information into the existing pathtargets. Otherwise, replace each
8047 : * path with a projection path that generates the SRF-free scan/join
8048 : * target. This can't change the ordering of paths within rel->pathlist,
8049 : * so we just modify the list in place.
8050 : */
8051 862917 : foreach(lc, rel->pathlist)
8052 : {
8053 449744 : Path *subpath = (Path *) lfirst(lc);
8054 :
8055 : /* Shouldn't have any parameterized paths anymore */
8056 : Assert(subpath->param_info == NULL);
8057 :
8058 449744 : if (tlist_same_exprs)
8059 162422 : subpath->pathtarget->sortgrouprefs =
8060 162422 : scanjoin_target->sortgrouprefs;
8061 : else
8062 : {
8063 : Path *newpath;
8064 :
8065 287322 : newpath = (Path *) create_projection_path(root, rel, subpath,
8066 : scanjoin_target);
8067 287322 : lfirst(lc) = newpath;
8068 : }
8069 : }
8070 :
8071 : /* Likewise adjust the targets for any partial paths. */
8072 433451 : foreach(lc, rel->partial_pathlist)
8073 : {
8074 20278 : Path *subpath = (Path *) lfirst(lc);
8075 :
8076 : /* Shouldn't have any parameterized paths anymore */
8077 : Assert(subpath->param_info == NULL);
8078 :
8079 20278 : if (tlist_same_exprs)
8080 16205 : subpath->pathtarget->sortgrouprefs =
8081 16205 : scanjoin_target->sortgrouprefs;
8082 : else
8083 : {
8084 : Path *newpath;
8085 :
8086 4073 : newpath = (Path *) create_projection_path(root, rel, subpath,
8087 : scanjoin_target);
8088 4073 : lfirst(lc) = newpath;
8089 : }
8090 : }
8091 :
8092 : /*
8093 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8094 : * atop each existing path. (Note that this function doesn't look at the
8095 : * cheapest-path fields, which is a good thing because they're bogus right
8096 : * now.)
8097 : */
8098 413173 : if (root->parse->hasTargetSRFs)
8099 10354 : adjust_paths_for_srfs(root, rel,
8100 : scanjoin_targets,
8101 : scanjoin_targets_contain_srfs);
8102 :
8103 : /*
8104 : * Update the rel's target to be the final (with SRFs) scan/join target.
8105 : * This now matches the actual output of all the paths, and we might get
8106 : * confused in createplan.c if they don't agree. We must do this now so
8107 : * that any append paths made in the next part will use the correct
8108 : * pathtarget (cf. create_append_path).
8109 : *
8110 : * Note that this is also necessary if GetForeignUpperPaths() gets called
8111 : * on the final scan/join relation or on any of its children, since the
8112 : * FDW might look at the rel's target to create ForeignPaths.
8113 : */
8114 413173 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8115 :
8116 : /*
8117 : * If the relation is partitioned, recursively apply the scan/join target
8118 : * to all partitions, and generate brand-new Append paths in which the
8119 : * scan/join target is computed below the Append rather than above it.
8120 : * Since Append is not projection-capable, that might save a separate
8121 : * Result node, and it also is important for partitionwise aggregate.
8122 : */
8123 413173 : if (rel_is_partitioned)
8124 : {
8125 10032 : List *live_children = NIL;
8126 : int i;
8127 :
8128 : /* Adjust each partition. */
8129 10032 : i = -1;
8130 30142 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8131 : {
8132 20110 : RelOptInfo *child_rel = rel->part_rels[i];
8133 : AppendRelInfo **appinfos;
8134 : int nappinfos;
8135 20110 : List *child_scanjoin_targets = NIL;
8136 :
8137 : Assert(child_rel != NULL);
8138 :
8139 : /* Dummy children can be ignored. */
8140 20110 : if (IS_DUMMY_REL(child_rel))
8141 40 : continue;
8142 :
8143 : /* Translate scan/join targets for this child. */
8144 20070 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8145 : &nappinfos);
8146 40140 : foreach(lc, scanjoin_targets)
8147 : {
8148 20070 : PathTarget *target = lfirst_node(PathTarget, lc);
8149 :
8150 20070 : target = copy_pathtarget(target);
8151 20070 : target->exprs = (List *)
8152 20070 : adjust_appendrel_attrs(root,
8153 20070 : (Node *) target->exprs,
8154 : nappinfos, appinfos);
8155 20070 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8156 : target);
8157 : }
8158 20070 : pfree(appinfos);
8159 :
8160 : /* Recursion does the real work. */
8161 20070 : apply_scanjoin_target_to_paths(root, child_rel,
8162 : child_scanjoin_targets,
8163 : scanjoin_targets_contain_srfs,
8164 : scanjoin_target_parallel_safe,
8165 : tlist_same_exprs);
8166 :
8167 : /* Save non-dummy children for Append paths. */
8168 20070 : if (!IS_DUMMY_REL(child_rel))
8169 20070 : live_children = lappend(live_children, child_rel);
8170 : }
8171 :
8172 : /* Build new paths for this relation by appending child paths. */
8173 10032 : add_paths_to_append_rel(root, rel, live_children);
8174 : }
8175 :
8176 : /*
8177 : * Consider generating Gather or Gather Merge paths. We must only do this
8178 : * if the relation is parallel safe, and we don't do it for child rels to
8179 : * avoid creating multiple Gather nodes within the same plan. We must do
8180 : * this after all paths have been generated and before set_cheapest, since
8181 : * one of the generated paths may turn out to be the cheapest one.
8182 : */
8183 413173 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
8184 139761 : generate_useful_gather_paths(root, rel, false);
8185 :
8186 : /*
8187 : * Reassess which paths are the cheapest, now that we've potentially added
8188 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8189 : * this relation.
8190 : */
8191 413173 : set_cheapest(rel);
8192 413173 : }
8193 :
8194 : /*
8195 : * create_partitionwise_grouping_paths
8196 : *
8197 : * If the partition keys of input relation are part of the GROUP BY clause, all
8198 : * the rows belonging to a given group come from a single partition. This
8199 : * allows aggregation/grouping over a partitioned relation to be broken down
8200 : * into aggregation/grouping on each partition. This should be no worse, and
8201 : * often better, than the normal approach.
8202 : *
8203 : * However, if the GROUP BY clause does not contain all the partition keys,
8204 : * rows from a given group may be spread across multiple partitions. In that
8205 : * case, we perform partial aggregation for each group, append the results,
8206 : * and then finalize aggregation. This is less certain to win than the
8207 : * previous case. It may win if the PartialAggregate stage greatly reduces
8208 : * the number of groups, because fewer rows will pass through the Append node.
8209 : * It may lose if we have lots of small groups.
8210 : */
8211 : static void
8212 685 : create_partitionwise_grouping_paths(PlannerInfo *root,
8213 : RelOptInfo *input_rel,
8214 : RelOptInfo *grouped_rel,
8215 : RelOptInfo *partially_grouped_rel,
8216 : const AggClauseCosts *agg_costs,
8217 : grouping_sets_data *gd,
8218 : PartitionwiseAggregateType patype,
8219 : GroupPathExtraData *extra)
8220 : {
8221 685 : List *grouped_live_children = NIL;
8222 685 : List *partially_grouped_live_children = NIL;
8223 685 : PathTarget *target = grouped_rel->reltarget;
8224 685 : bool partial_grouping_valid = true;
8225 : int i;
8226 :
8227 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8228 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8229 : partially_grouped_rel != NULL);
8230 :
8231 : /* Add paths for partitionwise aggregation/grouping. */
8232 685 : i = -1;
8233 2480 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8234 : {
8235 1795 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8236 : PathTarget *child_target;
8237 : AppendRelInfo **appinfos;
8238 : int nappinfos;
8239 : GroupPathExtraData child_extra;
8240 : RelOptInfo *child_grouped_rel;
8241 : RelOptInfo *child_partially_grouped_rel;
8242 :
8243 : Assert(child_input_rel != NULL);
8244 :
8245 : /* Dummy children can be ignored. */
8246 1795 : if (IS_DUMMY_REL(child_input_rel))
8247 0 : continue;
8248 :
8249 1795 : child_target = copy_pathtarget(target);
8250 :
8251 : /*
8252 : * Copy the given "extra" structure as is and then override the
8253 : * members specific to this child.
8254 : */
8255 1795 : memcpy(&child_extra, extra, sizeof(child_extra));
8256 :
8257 1795 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8258 : &nappinfos);
8259 :
8260 1795 : child_target->exprs = (List *)
8261 1795 : adjust_appendrel_attrs(root,
8262 1795 : (Node *) target->exprs,
8263 : nappinfos, appinfos);
8264 :
8265 : /* Translate havingQual and targetList. */
8266 1795 : child_extra.havingQual = (Node *)
8267 : adjust_appendrel_attrs(root,
8268 : extra->havingQual,
8269 : nappinfos, appinfos);
8270 1795 : child_extra.targetList = (List *)
8271 1795 : adjust_appendrel_attrs(root,
8272 1795 : (Node *) extra->targetList,
8273 : nappinfos, appinfos);
8274 :
8275 : /*
8276 : * extra->patype was the value computed for our parent rel; patype is
8277 : * the value for this relation. For the child, our value is its
8278 : * parent rel's value.
8279 : */
8280 1795 : child_extra.patype = patype;
8281 :
8282 : /*
8283 : * Create grouping relation to hold fully aggregated grouping and/or
8284 : * aggregation paths for the child.
8285 : */
8286 1795 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8287 : child_target,
8288 1795 : extra->target_parallel_safe,
8289 : child_extra.havingQual);
8290 :
8291 : /* Create grouping paths for this child relation. */
8292 1795 : create_ordinary_grouping_paths(root, child_input_rel,
8293 : child_grouped_rel,
8294 : agg_costs, gd, &child_extra,
8295 : &child_partially_grouped_rel);
8296 :
8297 1795 : if (child_partially_grouped_rel)
8298 : {
8299 : partially_grouped_live_children =
8300 1283 : lappend(partially_grouped_live_children,
8301 : child_partially_grouped_rel);
8302 : }
8303 : else
8304 512 : partial_grouping_valid = false;
8305 :
8306 1795 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8307 : {
8308 1082 : set_cheapest(child_grouped_rel);
8309 1082 : grouped_live_children = lappend(grouped_live_children,
8310 : child_grouped_rel);
8311 : }
8312 :
8313 1795 : pfree(appinfos);
8314 : }
8315 :
8316 : /*
8317 : * Try to create append paths for partially grouped children. For full
8318 : * partitionwise aggregation, we might have paths in the partial_pathlist
8319 : * if parallel aggregation is possible. For partial partitionwise
8320 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8321 : *
8322 : * NB: We must have a partially grouped path for every child in order to
8323 : * generate a partially grouped path for this relation.
8324 : */
8325 685 : if (partially_grouped_rel && partial_grouping_valid)
8326 : {
8327 : Assert(partially_grouped_live_children != NIL);
8328 :
8329 501 : add_paths_to_append_rel(root, partially_grouped_rel,
8330 : partially_grouped_live_children);
8331 : }
8332 :
8333 : /* If possible, create append paths for fully grouped children. */
8334 685 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8335 : {
8336 : Assert(grouped_live_children != NIL);
8337 :
8338 404 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8339 : }
8340 685 : }
8341 :
8342 : /*
8343 : * group_by_has_partkey
8344 : *
8345 : * Returns true if all the partition keys of the given relation are part of
8346 : * the GROUP BY clauses, including having matching collation, false otherwise.
8347 : */
8348 : static bool
8349 640 : group_by_has_partkey(RelOptInfo *input_rel,
8350 : List *targetList,
8351 : List *groupClause)
8352 : {
8353 640 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8354 640 : int cnt = 0;
8355 : int partnatts;
8356 :
8357 : /* Input relation should be partitioned. */
8358 : Assert(input_rel->part_scheme);
8359 :
8360 : /* Rule out early, if there are no partition keys present. */
8361 640 : if (!input_rel->partexprs)
8362 0 : return false;
8363 :
8364 640 : partnatts = input_rel->part_scheme->partnatts;
8365 :
8366 1074 : for (cnt = 0; cnt < partnatts; cnt++)
8367 : {
8368 670 : List *partexprs = input_rel->partexprs[cnt];
8369 : ListCell *lc;
8370 670 : bool found = false;
8371 :
8372 1001 : foreach(lc, partexprs)
8373 : {
8374 : ListCell *lg;
8375 775 : Expr *partexpr = lfirst(lc);
8376 775 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8377 :
8378 1206 : foreach(lg, groupexprs)
8379 : {
8380 875 : Expr *groupexpr = lfirst(lg);
8381 875 : Oid groupcoll = exprCollation((Node *) groupexpr);
8382 :
8383 : /*
8384 : * Note: we can assume there is at most one RelabelType node;
8385 : * eval_const_expressions() will have simplified if more than
8386 : * one.
8387 : */
8388 875 : if (IsA(groupexpr, RelabelType))
8389 20 : groupexpr = ((RelabelType *) groupexpr)->arg;
8390 :
8391 875 : if (equal(groupexpr, partexpr))
8392 : {
8393 : /*
8394 : * Reject a match if the grouping collation does not match
8395 : * the partitioning collation.
8396 : */
8397 444 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8398 : partcoll != groupcoll)
8399 10 : return false;
8400 :
8401 434 : found = true;
8402 434 : break;
8403 : }
8404 : }
8405 :
8406 765 : if (found)
8407 434 : break;
8408 : }
8409 :
8410 : /*
8411 : * If none of the partition key expressions match with any of the
8412 : * GROUP BY expression, return false.
8413 : */
8414 660 : if (!found)
8415 226 : return false;
8416 : }
8417 :
8418 404 : return true;
8419 : }
8420 :
8421 : /*
8422 : * generate_setop_child_grouplist
8423 : * Build a SortGroupClause list defining the sort/grouping properties
8424 : * of the child of a set operation.
8425 : *
8426 : * This is similar to generate_setop_grouplist() but differs as the setop
8427 : * child query's targetlist entries may already have a tleSortGroupRef
8428 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8429 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8430 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8431 : * any of the columns in the targetlist don't match to the setop's colTypes
8432 : * then we return an empty list. This may leave some TLEs with unreferenced
8433 : * ressortgroupref markings, but that's harmless.
8434 : */
8435 : static List *
8436 10463 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8437 : {
8438 10463 : List *grouplist = copyObject(op->groupClauses);
8439 : ListCell *lg;
8440 : ListCell *lt;
8441 : ListCell *ct;
8442 :
8443 10463 : lg = list_head(grouplist);
8444 10463 : ct = list_head(op->colTypes);
8445 40723 : foreach(lt, targetlist)
8446 : {
8447 30505 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8448 : SortGroupClause *sgc;
8449 : Oid coltype;
8450 :
8451 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8452 30505 : if (tle->resjunk)
8453 0 : continue;
8454 :
8455 : /*
8456 : * We expect every non-resjunk target to have a SortGroupClause and
8457 : * colTypes.
8458 : */
8459 : Assert(lg != NULL);
8460 : Assert(ct != NULL);
8461 30505 : sgc = (SortGroupClause *) lfirst(lg);
8462 30505 : coltype = lfirst_oid(ct);
8463 :
8464 : /* reject if target type isn't the same as the setop target type */
8465 30505 : if (coltype != exprType((Node *) tle->expr))
8466 245 : return NIL;
8467 :
8468 30260 : lg = lnext(grouplist, lg);
8469 30260 : ct = lnext(op->colTypes, ct);
8470 :
8471 : /* assign a tleSortGroupRef, or reuse the existing one */
8472 30260 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8473 : }
8474 :
8475 : Assert(lg == NULL);
8476 : Assert(ct == NULL);
8477 :
8478 10218 : return grouplist;
8479 : }
8480 :
8481 : /*
8482 : * create_unique_paths
8483 : * Build a new RelOptInfo containing Paths that represent elimination of
8484 : * distinct rows from the input data. Distinct-ness is defined according to
8485 : * the needs of the semijoin represented by sjinfo. If it is not possible
8486 : * to identify how to make the data unique, NULL is returned.
8487 : *
8488 : * If used at all, this is likely to be called repeatedly on the same rel,
8489 : * so we cache the result.
8490 : */
8491 : RelOptInfo *
8492 6904 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8493 : {
8494 : RelOptInfo *unique_rel;
8495 6904 : List *sortPathkeys = NIL;
8496 6904 : List *groupClause = NIL;
8497 : MemoryContext oldcontext;
8498 :
8499 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8500 : Assert(sjinfo->jointype == JOIN_SEMI);
8501 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8502 :
8503 : /* If result already cached, return it */
8504 6904 : if (rel->unique_rel)
8505 1196 : return rel->unique_rel;
8506 :
8507 : /* If it's not possible to unique-ify, return NULL */
8508 5708 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8509 104 : return NULL;
8510 :
8511 : /*
8512 : * Punt if this is a child relation and we failed to build a unique-ified
8513 : * relation for its parent. This can happen if all the RHS columns were
8514 : * found to be equated to constants when unique-ifying the parent table,
8515 : * leaving no columns to unique-ify.
8516 : */
8517 5604 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8518 10 : return NULL;
8519 :
8520 : /*
8521 : * When called during GEQO join planning, we are in a short-lived memory
8522 : * context. We must make sure that the unique rel and any subsidiary data
8523 : * structures created for a baserel survive the GEQO cycle, else the
8524 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8525 : * are creating those for a joinrel during GEQO, we don't want them to
8526 : * clutter the main planning context. Upshot is that the best solution is
8527 : * to explicitly allocate memory in the same context the given RelOptInfo
8528 : * is in.
8529 : */
8530 5594 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8531 :
8532 5594 : unique_rel = makeNode(RelOptInfo);
8533 5594 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8534 :
8535 : /*
8536 : * clear path info
8537 : */
8538 5594 : unique_rel->pathlist = NIL;
8539 5594 : unique_rel->ppilist = NIL;
8540 5594 : unique_rel->partial_pathlist = NIL;
8541 5594 : unique_rel->cheapest_startup_path = NULL;
8542 5594 : unique_rel->cheapest_total_path = NULL;
8543 5594 : unique_rel->cheapest_parameterized_paths = NIL;
8544 :
8545 : /*
8546 : * Build the target list for the unique rel. We also build the pathkeys
8547 : * that represent the ordering requirements for the sort-based
8548 : * implementation, and the list of SortGroupClause nodes that represent
8549 : * the columns to be grouped on for the hash-based implementation.
8550 : *
8551 : * For a child rel, we can construct these fields from those of its
8552 : * parent.
8553 : */
8554 5594 : if (IS_OTHER_REL(rel))
8555 360 : {
8556 : PathTarget *child_unique_target;
8557 : PathTarget *parent_unique_target;
8558 :
8559 360 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8560 :
8561 360 : child_unique_target = copy_pathtarget(parent_unique_target);
8562 :
8563 : /* Translate the target expressions */
8564 360 : child_unique_target->exprs = (List *)
8565 360 : adjust_appendrel_attrs_multilevel(root,
8566 360 : (Node *) parent_unique_target->exprs,
8567 : rel,
8568 360 : rel->top_parent);
8569 :
8570 360 : unique_rel->reltarget = child_unique_target;
8571 :
8572 360 : sortPathkeys = rel->top_parent->unique_pathkeys;
8573 360 : groupClause = rel->top_parent->unique_groupclause;
8574 : }
8575 : else
8576 : {
8577 : List *newtlist;
8578 : int nextresno;
8579 5234 : List *sortList = NIL;
8580 : ListCell *lc1;
8581 : ListCell *lc2;
8582 :
8583 : /*
8584 : * The values we are supposed to unique-ify may be expressions in the
8585 : * variables of the input rel's targetlist. We have to add any such
8586 : * expressions to the unique rel's targetlist.
8587 : *
8588 : * To complicate matters, some of the values to be unique-ified may be
8589 : * known redundant by the EquivalenceClass machinery (e.g., because
8590 : * they have been equated to constants). There is no need to compare
8591 : * such values during unique-ification, and indeed we had better not
8592 : * try because the Vars involved may not have propagated as high as
8593 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8594 : * detect such cases, which is a tad inefficient but it doesn't seem
8595 : * worth building specialized infrastructure for this.
8596 : */
8597 5234 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8598 5234 : nextresno = list_length(newtlist) + 1;
8599 :
8600 10669 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8601 : {
8602 5435 : Expr *uniqexpr = lfirst(lc1);
8603 5435 : Oid in_oper = lfirst_oid(lc2);
8604 : Oid sortop;
8605 : TargetEntry *tle;
8606 5435 : bool made_tle = false;
8607 :
8608 5435 : tle = tlist_member(uniqexpr, newtlist);
8609 5435 : if (!tle)
8610 : {
8611 2732 : tle = makeTargetEntry(uniqexpr,
8612 : nextresno,
8613 : NULL,
8614 : false);
8615 2732 : newtlist = lappend(newtlist, tle);
8616 2732 : nextresno++;
8617 2732 : made_tle = true;
8618 : }
8619 :
8620 : /*
8621 : * Try to build an ORDER BY list to sort the input compatibly. We
8622 : * do this for each sortable clause even when the clauses are not
8623 : * all sortable, so that we can detect clauses that are redundant
8624 : * according to the pathkey machinery.
8625 : */
8626 5435 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8627 5435 : if (OidIsValid(sortop))
8628 : {
8629 : Oid eqop;
8630 : SortGroupClause *sortcl;
8631 :
8632 : /*
8633 : * The Unique node will need equality operators. Normally
8634 : * these are the same as the IN clause operators, but if those
8635 : * are cross-type operators then the equality operators are
8636 : * the ones for the IN clause operators' RHS datatype.
8637 : */
8638 5435 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8639 5435 : if (!OidIsValid(eqop)) /* shouldn't happen */
8640 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8641 : sortop);
8642 :
8643 5435 : sortcl = makeNode(SortGroupClause);
8644 5435 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8645 5435 : sortcl->eqop = eqop;
8646 5435 : sortcl->sortop = sortop;
8647 5435 : sortcl->reverse_sort = false;
8648 5435 : sortcl->nulls_first = false;
8649 5435 : sortcl->hashable = false; /* no need to make this accurate */
8650 5435 : sortList = lappend(sortList, sortcl);
8651 :
8652 : /*
8653 : * At each step, convert the SortGroupClause list to pathkey
8654 : * form. If the just-added SortGroupClause is redundant, the
8655 : * result will be shorter than the SortGroupClause list.
8656 : */
8657 5435 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8658 : newtlist);
8659 5435 : if (list_length(sortPathkeys) != list_length(sortList))
8660 : {
8661 : /* Drop the redundant SortGroupClause */
8662 1710 : sortList = list_delete_last(sortList);
8663 : Assert(list_length(sortPathkeys) == list_length(sortList));
8664 : /* Undo tlist addition, if we made one */
8665 1710 : if (made_tle)
8666 : {
8667 10 : newtlist = list_delete_last(newtlist);
8668 10 : nextresno--;
8669 : }
8670 : /* We need not consider this clause for hashing, either */
8671 1710 : continue;
8672 : }
8673 : }
8674 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8675 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8676 : in_oper);
8677 :
8678 3725 : if (sjinfo->semi_can_hash)
8679 : {
8680 : /* Create a GROUP BY list for the Agg node to use */
8681 : Oid eq_oper;
8682 : SortGroupClause *groupcl;
8683 :
8684 : /*
8685 : * Get the hashable equality operators for the Agg node to
8686 : * use. Normally these are the same as the IN clause
8687 : * operators, but if those are cross-type operators then the
8688 : * equality operators are the ones for the IN clause
8689 : * operators' RHS datatype.
8690 : */
8691 3725 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8692 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8693 : in_oper);
8694 :
8695 3725 : groupcl = makeNode(SortGroupClause);
8696 3725 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8697 3725 : groupcl->eqop = eq_oper;
8698 3725 : groupcl->sortop = sortop;
8699 3725 : groupcl->reverse_sort = false;
8700 3725 : groupcl->nulls_first = false;
8701 3725 : groupcl->hashable = true;
8702 3725 : groupClause = lappend(groupClause, groupcl);
8703 : }
8704 : }
8705 :
8706 : /*
8707 : * Done building the sortPathkeys and groupClause. But the
8708 : * sortPathkeys are bogus if not all the clauses were sortable.
8709 : */
8710 5234 : if (!sjinfo->semi_can_btree)
8711 0 : sortPathkeys = NIL;
8712 :
8713 : /*
8714 : * It can happen that all the RHS columns are equated to constants.
8715 : * We'd have to do something special to unique-ify in that case, and
8716 : * it's such an unlikely-in-the-real-world case that it's not worth
8717 : * the effort. So just punt if we found no columns to unique-ify.
8718 : */
8719 5234 : if (sortPathkeys == NIL && groupClause == NIL)
8720 : {
8721 1625 : MemoryContextSwitchTo(oldcontext);
8722 1625 : return NULL;
8723 : }
8724 :
8725 : /* Convert the required targetlist back to PathTarget form */
8726 3609 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8727 : }
8728 :
8729 : /* build unique paths based on input rel's pathlist */
8730 3969 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8731 : sjinfo, unique_rel);
8732 :
8733 : /* build unique paths based on input rel's partial_pathlist */
8734 3969 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8735 : sjinfo, unique_rel);
8736 :
8737 : /* Now choose the best path(s) */
8738 3969 : set_cheapest(unique_rel);
8739 :
8740 : /*
8741 : * There shouldn't be any partial paths for the unique relation;
8742 : * otherwise, we won't be able to properly guarantee uniqueness.
8743 : */
8744 : Assert(unique_rel->partial_pathlist == NIL);
8745 :
8746 : /* Cache the result */
8747 3969 : rel->unique_rel = unique_rel;
8748 3969 : rel->unique_pathkeys = sortPathkeys;
8749 3969 : rel->unique_groupclause = groupClause;
8750 :
8751 3969 : MemoryContextSwitchTo(oldcontext);
8752 :
8753 3969 : return unique_rel;
8754 : }
8755 :
8756 : /*
8757 : * create_final_unique_paths
8758 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8759 : */
8760 : static void
8761 7101 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8762 : List *sortPathkeys, List *groupClause,
8763 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8764 : {
8765 7101 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8766 :
8767 : /* Estimate number of output rows */
8768 7101 : unique_rel->rows = estimate_num_groups(root,
8769 : sjinfo->semi_rhs_exprs,
8770 : cheapest_input_path->rows,
8771 : NULL,
8772 : NULL);
8773 :
8774 : /* Consider sort-based implementations, if possible. */
8775 7101 : if (sjinfo->semi_can_btree)
8776 : {
8777 : ListCell *lc;
8778 :
8779 : /*
8780 : * Use any available suitably-sorted path as input, and also consider
8781 : * sorting the cheapest-total path and incremental sort on any paths
8782 : * with presorted keys.
8783 : *
8784 : * To save planning time, we ignore parameterized input paths unless
8785 : * they are the cheapest-total path.
8786 : */
8787 15480 : foreach(lc, input_rel->pathlist)
8788 : {
8789 8379 : Path *input_path = (Path *) lfirst(lc);
8790 : Path *path;
8791 : bool is_sorted;
8792 : int presorted_keys;
8793 :
8794 : /*
8795 : * Ignore parameterized paths that are not the cheapest-total
8796 : * path.
8797 : */
8798 8379 : if (input_path->param_info &&
8799 : input_path != cheapest_input_path)
8800 740 : continue;
8801 :
8802 7675 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8803 : input_path->pathkeys,
8804 : &presorted_keys);
8805 :
8806 : /*
8807 : * Ignore paths that are not suitably or partially sorted, unless
8808 : * they are the cheapest total path (no need to deal with paths
8809 : * which have presorted keys when incremental sort is disabled).
8810 : */
8811 7675 : if (!is_sorted && input_path != cheapest_input_path &&
8812 76 : (presorted_keys == 0 || !enable_incremental_sort))
8813 36 : continue;
8814 :
8815 : /*
8816 : * Make a separate ProjectionPath in case we need a Result node.
8817 : */
8818 7639 : path = (Path *) create_projection_path(root,
8819 : unique_rel,
8820 : input_path,
8821 7639 : unique_rel->reltarget);
8822 :
8823 7639 : if (!is_sorted)
8824 : {
8825 : /*
8826 : * We've no need to consider both a sort and incremental sort.
8827 : * We'll just do a sort if there are no presorted keys and an
8828 : * incremental sort when there are presorted keys.
8829 : */
8830 3952 : if (presorted_keys == 0 || !enable_incremental_sort)
8831 3912 : path = (Path *) create_sort_path(root,
8832 : unique_rel,
8833 : path,
8834 : sortPathkeys,
8835 : -1.0);
8836 : else
8837 40 : path = (Path *) create_incremental_sort_path(root,
8838 : unique_rel,
8839 : path,
8840 : sortPathkeys,
8841 : presorted_keys,
8842 : -1.0);
8843 : }
8844 :
8845 7639 : path = (Path *) create_unique_path(root, unique_rel, path,
8846 : list_length(sortPathkeys),
8847 : unique_rel->rows);
8848 :
8849 7639 : add_path(unique_rel, path);
8850 : }
8851 : }
8852 :
8853 : /* Consider hash-based implementation, if possible. */
8854 7101 : if (sjinfo->semi_can_hash)
8855 : {
8856 : Path *path;
8857 :
8858 : /*
8859 : * Make a separate ProjectionPath in case we need a Result node.
8860 : */
8861 7101 : path = (Path *) create_projection_path(root,
8862 : unique_rel,
8863 : cheapest_input_path,
8864 7101 : unique_rel->reltarget);
8865 :
8866 7101 : path = (Path *) create_agg_path(root,
8867 : unique_rel,
8868 : path,
8869 : cheapest_input_path->pathtarget,
8870 : AGG_HASHED,
8871 : AGGSPLIT_SIMPLE,
8872 : groupClause,
8873 : NIL,
8874 : NULL,
8875 : unique_rel->rows);
8876 :
8877 7101 : add_path(unique_rel, path);
8878 : }
8879 7101 : }
8880 :
8881 : /*
8882 : * create_partial_unique_paths
8883 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
8884 : */
8885 : static void
8886 3969 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8887 : List *sortPathkeys, List *groupClause,
8888 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8889 : {
8890 : RelOptInfo *partial_unique_rel;
8891 : Path *cheapest_partial_path;
8892 :
8893 : /* nothing to do when there are no partial paths in the input rel */
8894 3969 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
8895 837 : return;
8896 :
8897 : /*
8898 : * nothing to do if there's anything in the targetlist that's
8899 : * parallel-restricted.
8900 : */
8901 3132 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
8902 0 : return;
8903 :
8904 3132 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
8905 :
8906 3132 : partial_unique_rel = makeNode(RelOptInfo);
8907 3132 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
8908 :
8909 : /*
8910 : * clear path info
8911 : */
8912 3132 : partial_unique_rel->pathlist = NIL;
8913 3132 : partial_unique_rel->ppilist = NIL;
8914 3132 : partial_unique_rel->partial_pathlist = NIL;
8915 3132 : partial_unique_rel->cheapest_startup_path = NULL;
8916 3132 : partial_unique_rel->cheapest_total_path = NULL;
8917 3132 : partial_unique_rel->cheapest_parameterized_paths = NIL;
8918 :
8919 : /* Estimate number of output rows */
8920 3132 : partial_unique_rel->rows = estimate_num_groups(root,
8921 : sjinfo->semi_rhs_exprs,
8922 : cheapest_partial_path->rows,
8923 : NULL,
8924 : NULL);
8925 3132 : partial_unique_rel->reltarget = unique_rel->reltarget;
8926 :
8927 : /* Consider sort-based implementations, if possible. */
8928 3132 : if (sjinfo->semi_can_btree)
8929 : {
8930 : ListCell *lc;
8931 :
8932 : /*
8933 : * Use any available suitably-sorted path as input, and also consider
8934 : * sorting the cheapest partial path and incremental sort on any paths
8935 : * with presorted keys.
8936 : */
8937 6518 : foreach(lc, input_rel->partial_pathlist)
8938 : {
8939 3386 : Path *input_path = (Path *) lfirst(lc);
8940 : Path *path;
8941 : bool is_sorted;
8942 : int presorted_keys;
8943 :
8944 3386 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8945 : input_path->pathkeys,
8946 : &presorted_keys);
8947 :
8948 : /*
8949 : * Ignore paths that are not suitably or partially sorted, unless
8950 : * they are the cheapest partial path (no need to deal with paths
8951 : * which have presorted keys when incremental sort is disabled).
8952 : */
8953 3386 : if (!is_sorted && input_path != cheapest_partial_path &&
8954 0 : (presorted_keys == 0 || !enable_incremental_sort))
8955 0 : continue;
8956 :
8957 : /*
8958 : * Make a separate ProjectionPath in case we need a Result node.
8959 : */
8960 3386 : path = (Path *) create_projection_path(root,
8961 : partial_unique_rel,
8962 : input_path,
8963 3386 : partial_unique_rel->reltarget);
8964 :
8965 3386 : if (!is_sorted)
8966 : {
8967 : /*
8968 : * We've no need to consider both a sort and incremental sort.
8969 : * We'll just do a sort if there are no presorted keys and an
8970 : * incremental sort when there are presorted keys.
8971 : */
8972 3092 : if (presorted_keys == 0 || !enable_incremental_sort)
8973 3092 : path = (Path *) create_sort_path(root,
8974 : partial_unique_rel,
8975 : path,
8976 : sortPathkeys,
8977 : -1.0);
8978 : else
8979 0 : path = (Path *) create_incremental_sort_path(root,
8980 : partial_unique_rel,
8981 : path,
8982 : sortPathkeys,
8983 : presorted_keys,
8984 : -1.0);
8985 : }
8986 :
8987 3386 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
8988 : list_length(sortPathkeys),
8989 : partial_unique_rel->rows);
8990 :
8991 3386 : add_partial_path(partial_unique_rel, path);
8992 : }
8993 : }
8994 :
8995 : /* Consider hash-based implementation, if possible. */
8996 3132 : if (sjinfo->semi_can_hash)
8997 : {
8998 : Path *path;
8999 :
9000 : /*
9001 : * Make a separate ProjectionPath in case we need a Result node.
9002 : */
9003 3132 : path = (Path *) create_projection_path(root,
9004 : partial_unique_rel,
9005 : cheapest_partial_path,
9006 3132 : partial_unique_rel->reltarget);
9007 :
9008 3132 : path = (Path *) create_agg_path(root,
9009 : partial_unique_rel,
9010 : path,
9011 : cheapest_partial_path->pathtarget,
9012 : AGG_HASHED,
9013 : AGGSPLIT_SIMPLE,
9014 : groupClause,
9015 : NIL,
9016 : NULL,
9017 : partial_unique_rel->rows);
9018 :
9019 3132 : add_partial_path(partial_unique_rel, path);
9020 : }
9021 :
9022 3132 : if (partial_unique_rel->partial_pathlist != NIL)
9023 : {
9024 3132 : generate_useful_gather_paths(root, partial_unique_rel, true);
9025 3132 : set_cheapest(partial_unique_rel);
9026 :
9027 : /*
9028 : * Finally, create paths to unique-ify the final result. This step is
9029 : * needed to remove any duplicates due to combining rows from parallel
9030 : * workers.
9031 : */
9032 3132 : create_final_unique_paths(root, partial_unique_rel,
9033 : sortPathkeys, groupClause,
9034 : sjinfo, unique_rel);
9035 : }
9036 : }
9037 :
9038 : /*
9039 : * Choose a unique name for some subroot.
9040 : *
9041 : * Modifies glob->subplanNames to track names already used.
9042 : */
9043 : char *
9044 65177 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
9045 : {
9046 : unsigned n;
9047 :
9048 : /*
9049 : * If a numeric suffix is not required, then search the list of
9050 : * previously-assigned names for a match. If none is found, then we can
9051 : * use the provided name without modification.
9052 : */
9053 65177 : if (!always_number)
9054 : {
9055 19257 : bool found = false;
9056 :
9057 46422 : foreach_ptr(char, subplan_name, glob->subplanNames)
9058 : {
9059 12259 : if (strcmp(subplan_name, name) == 0)
9060 : {
9061 4351 : found = true;
9062 4351 : break;
9063 : }
9064 : }
9065 :
9066 19257 : if (!found)
9067 : {
9068 : /* pstrdup here is just to avoid cast-away-const */
9069 14906 : char *chosen_name = pstrdup(name);
9070 :
9071 14906 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
9072 14906 : return chosen_name;
9073 : }
9074 : }
9075 :
9076 : /*
9077 : * If a numeric suffix is required or if the un-suffixed name is already
9078 : * in use, then loop until we find a positive integer that produces a
9079 : * novel name.
9080 : */
9081 50271 : for (n = 1; true; ++n)
9082 43312 : {
9083 93583 : char *proposed_name = psprintf("%s_%u", name, n);
9084 93583 : bool found = false;
9085 :
9086 348905 : foreach_ptr(char, subplan_name, glob->subplanNames)
9087 : {
9088 205051 : if (strcmp(subplan_name, proposed_name) == 0)
9089 : {
9090 43312 : found = true;
9091 43312 : break;
9092 : }
9093 : }
9094 :
9095 93583 : if (!found)
9096 : {
9097 50271 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9098 50271 : return proposed_name;
9099 : }
9100 :
9101 43312 : pfree(proposed_name);
9102 : }
9103 : }
|
