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 : /*
138 : * Context for the find_having_collation_conflicts walker.
139 : *
140 : * ancestor_collids is a stack of inputcollids contributed by collation-aware
141 : * ancestors of the current node. Entries are pushed before recursing into a
142 : * node's children and popped afterwards, so the stack reflects exactly the
143 : * inputcollids on the current root-to-node path.
144 : */
145 : typedef struct
146 : {
147 : Index group_rtindex;
148 : List *ancestor_collids;
149 : } having_collation_ctx;
150 :
151 : /* Local functions */
152 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
153 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
154 : static Bitmapset *find_having_collation_conflicts(Query *parse,
155 : Index group_rtindex);
156 : static bool having_collation_conflict_walker(Node *node,
157 : having_collation_ctx *ctx);
158 : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
159 : SetOperationStmt *setops);
160 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
161 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
162 : int *tleref_to_colnum_map);
163 : static void preprocess_rowmarks(PlannerInfo *root);
164 : static double preprocess_limit(PlannerInfo *root,
165 : double tuple_fraction,
166 : int64 *offset_est, int64 *count_est);
167 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
168 : static List *extract_rollup_sets(List *groupingSets);
169 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
170 : static void standard_qp_callback(PlannerInfo *root, void *extra);
171 : static double get_number_of_groups(PlannerInfo *root,
172 : double path_rows,
173 : grouping_sets_data *gd,
174 : List *target_list);
175 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
176 : RelOptInfo *input_rel,
177 : PathTarget *target,
178 : bool target_parallel_safe,
179 : grouping_sets_data *gd);
180 : static bool is_degenerate_grouping(PlannerInfo *root);
181 : static void create_degenerate_grouping_paths(PlannerInfo *root,
182 : RelOptInfo *input_rel,
183 : RelOptInfo *grouped_rel);
184 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
185 : PathTarget *target, bool target_parallel_safe,
186 : Node *havingQual);
187 : static void create_ordinary_grouping_paths(PlannerInfo *root,
188 : RelOptInfo *input_rel,
189 : RelOptInfo *grouped_rel,
190 : const AggClauseCosts *agg_costs,
191 : grouping_sets_data *gd,
192 : GroupPathExtraData *extra,
193 : RelOptInfo **partially_grouped_rel_p);
194 : static void consider_groupingsets_paths(PlannerInfo *root,
195 : RelOptInfo *grouped_rel,
196 : Path *path,
197 : bool is_sorted,
198 : bool can_hash,
199 : grouping_sets_data *gd,
200 : const AggClauseCosts *agg_costs,
201 : double dNumGroups);
202 : static RelOptInfo *create_window_paths(PlannerInfo *root,
203 : RelOptInfo *input_rel,
204 : PathTarget *input_target,
205 : PathTarget *output_target,
206 : bool output_target_parallel_safe,
207 : WindowFuncLists *wflists,
208 : List *activeWindows);
209 : static void create_one_window_path(PlannerInfo *root,
210 : RelOptInfo *window_rel,
211 : Path *path,
212 : PathTarget *input_target,
213 : PathTarget *output_target,
214 : WindowFuncLists *wflists,
215 : List *activeWindows);
216 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
217 : RelOptInfo *input_rel,
218 : PathTarget *target);
219 : static void create_partial_distinct_paths(PlannerInfo *root,
220 : RelOptInfo *input_rel,
221 : RelOptInfo *final_distinct_rel,
222 : PathTarget *target);
223 : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
224 : RelOptInfo *input_rel,
225 : RelOptInfo *distinct_rel);
226 : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
227 : List *needed_pathkeys,
228 : List *path_pathkeys);
229 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
230 : RelOptInfo *input_rel,
231 : PathTarget *target,
232 : bool target_parallel_safe,
233 : double limit_tuples);
234 : static PathTarget *make_group_input_target(PlannerInfo *root,
235 : PathTarget *final_target);
236 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
237 : PathTarget *grouping_target,
238 : Node *havingQual);
239 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
240 : static void optimize_window_clauses(PlannerInfo *root,
241 : WindowFuncLists *wflists);
242 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
243 : static void name_active_windows(List *activeWindows);
244 : static PathTarget *make_window_input_target(PlannerInfo *root,
245 : PathTarget *final_target,
246 : List *activeWindows);
247 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
248 : List *tlist);
249 : static PathTarget *make_sort_input_target(PlannerInfo *root,
250 : PathTarget *final_target,
251 : bool *have_postponed_srfs);
252 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
253 : List *targets, List *targets_contain_srfs);
254 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
255 : RelOptInfo *grouped_rel,
256 : RelOptInfo *partially_grouped_rel,
257 : const AggClauseCosts *agg_costs,
258 : grouping_sets_data *gd,
259 : GroupPathExtraData *extra);
260 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
261 : RelOptInfo *grouped_rel,
262 : RelOptInfo *input_rel,
263 : grouping_sets_data *gd,
264 : GroupPathExtraData *extra,
265 : bool force_rel_creation);
266 : static Path *make_ordered_path(PlannerInfo *root,
267 : RelOptInfo *rel,
268 : Path *path,
269 : Path *cheapest_path,
270 : List *pathkeys,
271 : double limit_tuples);
272 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
273 : static bool can_partial_agg(PlannerInfo *root);
274 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
275 : RelOptInfo *rel,
276 : List *scanjoin_targets,
277 : List *scanjoin_targets_contain_srfs,
278 : bool scanjoin_target_parallel_safe,
279 : bool tlist_same_exprs);
280 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
281 : RelOptInfo *input_rel,
282 : RelOptInfo *grouped_rel,
283 : RelOptInfo *partially_grouped_rel,
284 : const AggClauseCosts *agg_costs,
285 : grouping_sets_data *gd,
286 : PartitionwiseAggregateType patype,
287 : GroupPathExtraData *extra);
288 : static bool group_by_has_partkey(RelOptInfo *input_rel,
289 : List *targetList,
290 : List *groupClause);
291 : static int common_prefix_cmp(const void *a, const void *b);
292 : static List *generate_setop_child_grouplist(SetOperationStmt *op,
293 : List *targetlist);
294 : static void create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
295 : List *sortPathkeys, List *groupClause,
296 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
297 : static void create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
298 : List *sortPathkeys, List *groupClause,
299 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
300 :
301 :
302 : /*****************************************************************************
303 : *
304 : * Query optimizer entry point
305 : *
306 : * Inputs:
307 : * parse: an analyzed-and-rewritten query tree for an optimizable statement
308 : * query_string: source text for the query tree (used for error reports)
309 : * cursorOptions: bitmask of CURSOR_OPT_XXX flags, see parsenodes.h
310 : * boundParams: passed-in parameter values, or NULL if none
311 : * es: ExplainState if being called from EXPLAIN, else NULL
312 : *
313 : * The result is a PlannedStmt tree.
314 : *
315 : * PARAM_EXTERN Param nodes within the parse tree can be replaced by Consts
316 : * using values from boundParams, if those values are marked PARAM_FLAG_CONST.
317 : * Parameter values not so marked are still relied on for estimation purposes.
318 : *
319 : * The ExplainState pointer is not currently used by the core planner, but it
320 : * is passed through to some planner hooks so that they can report information
321 : * back to EXPLAIN extension hooks.
322 : *
323 : * To support loadable plugins that monitor or modify planner behavior,
324 : * we provide a hook variable that lets a plugin get control before and
325 : * after the standard planning process. The plugin would normally call
326 : * standard_planner().
327 : *
328 : * Note to plugin authors: standard_planner() scribbles on its Query input,
329 : * so you'd better copy that data structure if you want to plan more than once.
330 : *
331 : *****************************************************************************/
332 : PlannedStmt *
333 337553 : planner(Query *parse, const char *query_string, int cursorOptions,
334 : ParamListInfo boundParams, ExplainState *es)
335 : {
336 : PlannedStmt *result;
337 :
338 337553 : if (planner_hook)
339 50386 : result = (*planner_hook) (parse, query_string, cursorOptions,
340 : boundParams, es);
341 : else
342 287167 : result = standard_planner(parse, query_string, cursorOptions,
343 : boundParams, es);
344 :
345 333567 : pgstat_report_plan_id(result->planId, false);
346 :
347 333567 : return result;
348 : }
349 :
350 : PlannedStmt *
351 337553 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
352 : ParamListInfo boundParams, ExplainState *es)
353 : {
354 : PlannedStmt *result;
355 : PlannerGlobal *glob;
356 : double tuple_fraction;
357 : PlannerInfo *root;
358 : RelOptInfo *final_rel;
359 : Path *best_path;
360 : Plan *top_plan;
361 : ListCell *lp,
362 : *lr,
363 : *lc;
364 :
365 : /*
366 : * Set up global state for this planner invocation. This data is needed
367 : * across all levels of sub-Query that might exist in the given command,
368 : * so we keep it in a separate struct that's linked to by each per-Query
369 : * PlannerInfo.
370 : */
371 337553 : glob = makeNode(PlannerGlobal);
372 :
373 337553 : glob->boundParams = boundParams;
374 337553 : glob->subplans = NIL;
375 337553 : glob->subpaths = NIL;
376 337553 : glob->subroots = NIL;
377 337553 : glob->rewindPlanIDs = NULL;
378 337553 : glob->finalrtable = NIL;
379 337553 : glob->allRelids = NULL;
380 337553 : glob->prunableRelids = NULL;
381 337553 : glob->finalrteperminfos = NIL;
382 337553 : glob->finalrowmarks = NIL;
383 337553 : glob->resultRelations = NIL;
384 337553 : glob->appendRelations = NIL;
385 337553 : glob->partPruneInfos = NIL;
386 337553 : glob->relationOids = NIL;
387 337553 : glob->invalItems = NIL;
388 337553 : glob->paramExecTypes = NIL;
389 337553 : glob->lastPHId = 0;
390 337553 : glob->lastRowMarkId = 0;
391 337553 : glob->lastPlanNodeId = 0;
392 337553 : glob->transientPlan = false;
393 337553 : glob->dependsOnRole = false;
394 337553 : glob->partition_directory = NULL;
395 337553 : glob->rel_notnullatts_hash = NULL;
396 :
397 : /*
398 : * Assess whether it's feasible to use parallel mode for this query. We
399 : * can't do this in a standalone backend, or if the command will try to
400 : * modify any data, or if this is a cursor operation, or if GUCs are set
401 : * to values that don't permit parallelism, or if parallel-unsafe
402 : * functions are present in the query tree.
403 : *
404 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
405 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
406 : * the command is writing into a completely new table which workers won't
407 : * be able to see. If the workers could see the table, the fact that
408 : * group locking would cause them to ignore the leader's heavyweight GIN
409 : * page locks would make this unsafe. We'll have to fix that somehow if
410 : * we want to allow parallel inserts in general; updates and deletes have
411 : * additional problems especially around combo CIDs.)
412 : *
413 : * For now, we don't try to use parallel mode if we're running inside a
414 : * parallel worker. We might eventually be able to relax this
415 : * restriction, but for now it seems best not to have parallel workers
416 : * trying to create their own parallel workers.
417 : */
418 337553 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
419 319514 : IsUnderPostmaster &&
420 319514 : parse->commandType == CMD_SELECT &&
421 258021 : !parse->hasModifyingCTE &&
422 257898 : max_parallel_workers_per_gather > 0 &&
423 257205 : !IsParallelWorker())
424 : {
425 : /* all the cheap tests pass, so scan the query tree */
426 257165 : glob->maxParallelHazard = max_parallel_hazard(parse);
427 257165 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
428 : }
429 : else
430 : {
431 : /* skip the query tree scan, just assume it's unsafe */
432 80388 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
433 80388 : glob->parallelModeOK = false;
434 : }
435 :
436 : /*
437 : * glob->parallelModeNeeded is normally set to false here and changed to
438 : * true during plan creation if a Gather or Gather Merge plan is actually
439 : * created (cf. create_gather_plan, create_gather_merge_plan).
440 : *
441 : * However, if debug_parallel_query = on or debug_parallel_query =
442 : * regress, then we impose parallel mode whenever it's safe to do so, even
443 : * if the final plan doesn't use parallelism. It's not safe to do so if
444 : * the query contains anything parallel-unsafe; parallelModeOK will be
445 : * false in that case. Note that parallelModeOK can't change after this
446 : * point. Otherwise, everything in the query is either parallel-safe or
447 : * parallel-restricted, and in either case it should be OK to impose
448 : * parallel-mode restrictions. If that ends up breaking something, then
449 : * either some function the user included in the query is incorrectly
450 : * labeled as parallel-safe or parallel-restricted when in reality it's
451 : * parallel-unsafe, or else the query planner itself has a bug.
452 : */
453 559704 : glob->parallelModeNeeded = glob->parallelModeOK &&
454 222151 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
455 :
456 : /* Determine what fraction of the plan is likely to be scanned */
457 337553 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
458 : {
459 : /*
460 : * We have no real idea how many tuples the user will ultimately FETCH
461 : * from a cursor, but it is often the case that he doesn't want 'em
462 : * all, or would prefer a fast-start plan anyway so that he can
463 : * process some of the tuples sooner. Use a GUC parameter to decide
464 : * what fraction to optimize for.
465 : */
466 3237 : tuple_fraction = cursor_tuple_fraction;
467 :
468 : /*
469 : * We document cursor_tuple_fraction as simply being a fraction, which
470 : * means the edge cases 0 and 1 have to be treated specially here. We
471 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
472 : */
473 3237 : if (tuple_fraction >= 1.0)
474 0 : tuple_fraction = 0.0;
475 3237 : else if (tuple_fraction <= 0.0)
476 0 : tuple_fraction = 1e-10;
477 : }
478 : else
479 : {
480 : /* Default assumption is we need all the tuples */
481 334316 : tuple_fraction = 0.0;
482 : }
483 :
484 : /*
485 : * Compute the initial path generation strategy mask.
486 : *
487 : * Some strategies, such as PGS_FOREIGNJOIN, have no corresponding enable_*
488 : * GUC, and so the corresponding bits are always set in the default
489 : * strategy mask.
490 : *
491 : * It may seem surprising that enable_indexscan sets both PGS_INDEXSCAN
492 : * and PGS_INDEXONLYSCAN. However, the historical behavior of this GUC
493 : * corresponds to this exactly: enable_indexscan=off disables both
494 : * index-scan and index-only scan paths, whereas enable_indexonlyscan=off
495 : * converts the index-only scan paths that we would have considered into
496 : * index scan paths.
497 : */
498 337553 : glob->default_pgs_mask = PGS_APPEND | PGS_MERGE_APPEND | PGS_FOREIGNJOIN |
499 : PGS_GATHER | PGS_CONSIDER_NONPARTIAL;
500 337553 : if (enable_tidscan)
501 337553 : glob->default_pgs_mask |= PGS_TIDSCAN;
502 337553 : if (enable_seqscan)
503 320558 : glob->default_pgs_mask |= PGS_SEQSCAN;
504 337553 : if (enable_indexscan)
505 335306 : glob->default_pgs_mask |= PGS_INDEXSCAN | PGS_INDEXONLYSCAN;
506 337553 : if (enable_indexonlyscan)
507 336169 : glob->default_pgs_mask |= PGS_CONSIDER_INDEXONLY;
508 337553 : if (enable_bitmapscan)
509 328323 : glob->default_pgs_mask |= PGS_BITMAPSCAN;
510 337553 : if (enable_mergejoin)
511 : {
512 335616 : glob->default_pgs_mask |= PGS_MERGEJOIN_PLAIN;
513 335616 : if (enable_material)
514 335549 : glob->default_pgs_mask |= PGS_MERGEJOIN_MATERIALIZE;
515 : }
516 337553 : if (enable_nestloop)
517 : {
518 337249 : glob->default_pgs_mask |= PGS_NESTLOOP_PLAIN;
519 337249 : if (enable_material)
520 336931 : glob->default_pgs_mask |= PGS_NESTLOOP_MATERIALIZE;
521 337249 : if (enable_memoize)
522 337142 : glob->default_pgs_mask |= PGS_NESTLOOP_MEMOIZE;
523 : }
524 337553 : if (enable_hashjoin)
525 335281 : glob->default_pgs_mask |= PGS_HASHJOIN;
526 337553 : if (enable_gathermerge)
527 337553 : glob->default_pgs_mask |= PGS_GATHER_MERGE;
528 337553 : if (enable_partitionwise_join)
529 2065 : glob->default_pgs_mask |= PGS_CONSIDER_PARTITIONWISE;
530 :
531 : /* Allow plugins to take control after we've initialized "glob" */
532 337553 : if (planner_setup_hook)
533 88383 : (*planner_setup_hook) (glob, parse, query_string, cursorOptions,
534 : &tuple_fraction, es);
535 :
536 : /* primary planning entry point (may recurse for subqueries) */
537 336786 : root = subquery_planner(glob, parse, NULL, NULL, NULL, false,
538 : tuple_fraction, NULL);
539 :
540 : /* Select best Path and turn it into a Plan */
541 333837 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
542 333837 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
543 :
544 333837 : top_plan = create_plan(root, best_path);
545 :
546 : /*
547 : * If creating a plan for a scrollable cursor, make sure it can run
548 : * backwards on demand. Add a Material node at the top at need.
549 : */
550 333567 : if (cursorOptions & CURSOR_OPT_SCROLL)
551 : {
552 221 : if (!ExecSupportsBackwardScan(top_plan))
553 26 : top_plan = materialize_finished_plan(top_plan);
554 : }
555 :
556 : /*
557 : * Optionally add a Gather node for testing purposes, provided this is
558 : * actually a safe thing to do.
559 : *
560 : * We can add Gather even when top_plan has parallel-safe initPlans, but
561 : * then we have to move the initPlans to the Gather node because of
562 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
563 : * regression tests when debug_parallel_query = regress, because initPlans
564 : * that would normally appear on the top_plan move to the Gather, causing
565 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
566 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
567 : */
568 333567 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
569 157 : top_plan->parallel_safe &&
570 104 : (top_plan->initPlan == NIL ||
571 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
572 : {
573 104 : Gather *gather = makeNode(Gather);
574 : Cost initplan_cost;
575 : bool unsafe_initplans;
576 :
577 104 : gather->plan.targetlist = top_plan->targetlist;
578 104 : gather->plan.qual = NIL;
579 104 : gather->plan.lefttree = top_plan;
580 104 : gather->plan.righttree = NULL;
581 104 : gather->num_workers = 1;
582 104 : gather->single_copy = true;
583 104 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
584 :
585 : /* Transfer any initPlans to the new top node */
586 104 : gather->plan.initPlan = top_plan->initPlan;
587 104 : top_plan->initPlan = NIL;
588 :
589 : /*
590 : * Since this Gather has no parallel-aware descendants to signal to,
591 : * we don't need a rescan Param.
592 : */
593 104 : gather->rescan_param = -1;
594 :
595 : /*
596 : * Ideally we'd use cost_gather here, but setting up dummy path data
597 : * to satisfy it doesn't seem much cleaner than knowing what it does.
598 : */
599 104 : gather->plan.startup_cost = top_plan->startup_cost +
600 : parallel_setup_cost;
601 104 : gather->plan.total_cost = top_plan->total_cost +
602 104 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
603 104 : gather->plan.plan_rows = top_plan->plan_rows;
604 104 : gather->plan.plan_width = top_plan->plan_width;
605 104 : gather->plan.parallel_aware = false;
606 104 : gather->plan.parallel_safe = false;
607 :
608 : /*
609 : * Delete the initplans' cost from top_plan. We needn't add it to the
610 : * Gather node, since the above coding already included it there.
611 : */
612 104 : SS_compute_initplan_cost(gather->plan.initPlan,
613 : &initplan_cost, &unsafe_initplans);
614 104 : top_plan->startup_cost -= initplan_cost;
615 104 : top_plan->total_cost -= initplan_cost;
616 :
617 : /* use parallel mode for parallel plans. */
618 104 : root->glob->parallelModeNeeded = true;
619 :
620 104 : top_plan = &gather->plan;
621 : }
622 :
623 : /*
624 : * If any Params were generated, run through the plan tree and compute
625 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
626 : * set_plan_references' tree traversal, but for now it has to be separate
627 : * because we need to visit subplans before not after main plan.
628 : */
629 333567 : if (glob->paramExecTypes != NIL)
630 : {
631 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
632 148375 : forboth(lp, glob->subplans, lr, glob->subroots)
633 : {
634 33961 : Plan *subplan = (Plan *) lfirst(lp);
635 33961 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
636 :
637 33961 : SS_finalize_plan(subroot, subplan);
638 : }
639 114414 : SS_finalize_plan(root, top_plan);
640 : }
641 :
642 : /* final cleanup of the plan */
643 : Assert(glob->finalrtable == NIL);
644 : Assert(glob->finalrteperminfos == NIL);
645 : Assert(glob->finalrowmarks == NIL);
646 : Assert(glob->resultRelations == NIL);
647 : Assert(glob->appendRelations == NIL);
648 333567 : top_plan = set_plan_references(root, top_plan);
649 : /* ... and the subplans (both regular subplans and initplans) */
650 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
651 367528 : forboth(lp, glob->subplans, lr, glob->subroots)
652 : {
653 33961 : Plan *subplan = (Plan *) lfirst(lp);
654 33961 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
655 :
656 33961 : lfirst(lp) = set_plan_references(subroot, subplan);
657 : }
658 :
659 : /* build the PlannedStmt result */
660 333567 : result = makeNode(PlannedStmt);
661 :
662 333567 : result->commandType = parse->commandType;
663 333567 : result->queryId = parse->queryId;
664 333567 : result->planOrigin = PLAN_STMT_STANDARD;
665 333567 : result->hasReturning = (parse->returningList != NIL);
666 333567 : result->hasModifyingCTE = parse->hasModifyingCTE;
667 333567 : result->canSetTag = parse->canSetTag;
668 333567 : result->transientPlan = glob->transientPlan;
669 333567 : result->dependsOnRole = glob->dependsOnRole;
670 333567 : result->parallelModeNeeded = glob->parallelModeNeeded;
671 333567 : result->planTree = top_plan;
672 333567 : result->partPruneInfos = glob->partPruneInfos;
673 333567 : result->rtable = glob->finalrtable;
674 667134 : result->unprunableRelids = bms_difference(glob->allRelids,
675 333567 : glob->prunableRelids);
676 333567 : result->permInfos = glob->finalrteperminfos;
677 333567 : result->subrtinfos = glob->subrtinfos;
678 333567 : result->appendRelations = glob->appendRelations;
679 333567 : result->subplans = glob->subplans;
680 333567 : result->rewindPlanIDs = glob->rewindPlanIDs;
681 333567 : result->rowMarks = glob->finalrowmarks;
682 :
683 : /*
684 : * Compute resultRelationRelids and rowMarkRelids from resultRelations and
685 : * rowMarks. These can be used for cheap membership checks.
686 : */
687 402780 : foreach(lc, glob->resultRelations)
688 69213 : result->resultRelationRelids = bms_add_member(result->resultRelationRelids,
689 : lfirst_int(lc));
690 344381 : foreach(lc, glob->finalrowmarks)
691 10814 : result->rowMarkRelids = bms_add_member(result->rowMarkRelids,
692 10814 : ((PlanRowMark *) lfirst(lc))->rti);
693 :
694 333567 : result->relationOids = glob->relationOids;
695 333567 : result->invalItems = glob->invalItems;
696 333567 : result->paramExecTypes = glob->paramExecTypes;
697 : /* utilityStmt should be null, but we might as well copy it */
698 333567 : result->utilityStmt = parse->utilityStmt;
699 333567 : result->elidedNodes = glob->elidedNodes;
700 333567 : result->stmt_location = parse->stmt_location;
701 333567 : result->stmt_len = parse->stmt_len;
702 :
703 333567 : result->jitFlags = PGJIT_NONE;
704 333567 : if (jit_enabled && jit_above_cost >= 0 &&
705 0 : top_plan->total_cost > jit_above_cost)
706 : {
707 0 : result->jitFlags |= PGJIT_PERFORM;
708 :
709 : /*
710 : * Decide how much effort should be put into generating better code.
711 : */
712 0 : if (jit_optimize_above_cost >= 0 &&
713 0 : top_plan->total_cost > jit_optimize_above_cost)
714 0 : result->jitFlags |= PGJIT_OPT3;
715 0 : if (jit_inline_above_cost >= 0 &&
716 0 : top_plan->total_cost > jit_inline_above_cost)
717 0 : result->jitFlags |= PGJIT_INLINE;
718 :
719 : /*
720 : * Decide which operations should be JITed.
721 : */
722 0 : if (jit_expressions)
723 0 : result->jitFlags |= PGJIT_EXPR;
724 0 : if (jit_tuple_deforming)
725 0 : result->jitFlags |= PGJIT_DEFORM;
726 : }
727 :
728 : /* Allow plugins to take control before we discard "glob" */
729 333567 : if (planner_shutdown_hook)
730 86846 : (*planner_shutdown_hook) (glob, parse, query_string, result);
731 :
732 333567 : if (glob->partition_directory != NULL)
733 9168 : DestroyPartitionDirectory(glob->partition_directory);
734 :
735 333567 : return result;
736 : }
737 :
738 :
739 : /*--------------------
740 : * subquery_planner
741 : * Invokes the planner on a subquery. We recurse to here for each
742 : * sub-SELECT found in the query tree.
743 : *
744 : * glob is the global state for the current planner run.
745 : * parse is the querytree produced by the parser & rewriter.
746 : * plan_name is the name to assign to this subplan (NULL at the top level).
747 : * parent_root is the immediate parent Query's info (NULL at the top level).
748 : * alternative_root is a previously created PlannerInfo for which this query
749 : * level is an alternative implementation, or else NULL.
750 : * hasRecursion is true if this is a recursive WITH query.
751 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
752 : * tuple_fraction is interpreted as explained for grouping_planner, below.
753 : * setops is used for set operation subqueries to provide the subquery with
754 : * the context in which it's being used so that Paths correctly sorted for the
755 : * set operation can be generated. NULL when not planning a set operation
756 : * child, or when a child of a set op that isn't interested in sorted input.
757 : *
758 : * Basically, this routine does the stuff that should only be done once
759 : * per Query object. It then calls grouping_planner. At one time,
760 : * grouping_planner could be invoked recursively on the same Query object;
761 : * that's not currently true, but we keep the separation between the two
762 : * routines anyway, in case we need it again someday.
763 : *
764 : * subquery_planner will be called recursively to handle sub-Query nodes
765 : * found within the query's expressions and rangetable.
766 : *
767 : * Returns the PlannerInfo struct ("root") that contains all data generated
768 : * while planning the subquery. In particular, the Path(s) attached to
769 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
770 : * cheapest way(s) to implement the query. The top level will select the
771 : * best Path and pass it through createplan.c to produce a finished Plan.
772 : *--------------------
773 : */
774 : PlannerInfo *
775 401212 : subquery_planner(PlannerGlobal *glob, Query *parse, char *plan_name,
776 : PlannerInfo *parent_root, PlannerInfo *alternative_root,
777 : bool hasRecursion, double tuple_fraction,
778 : SetOperationStmt *setops)
779 : {
780 : PlannerInfo *root;
781 : List *newWithCheckOptions;
782 : List *newHaving;
783 : Bitmapset *havingCollationConflicts;
784 : int havingIdx;
785 : bool hasOuterJoins;
786 : bool hasResultRTEs;
787 : RelOptInfo *final_rel;
788 : ListCell *l;
789 :
790 : /* Create a PlannerInfo data structure for this subquery */
791 401212 : root = makeNode(PlannerInfo);
792 401212 : root->parse = parse;
793 401212 : root->glob = glob;
794 401212 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
795 401212 : root->plan_name = plan_name;
796 401212 : if (alternative_root != NULL)
797 1331 : root->alternative_plan_name = alternative_root->plan_name;
798 : else
799 399881 : root->alternative_plan_name = plan_name;
800 401212 : root->parent_root = parent_root;
801 401212 : root->plan_params = NIL;
802 401212 : root->outer_params = NULL;
803 401212 : root->planner_cxt = CurrentMemoryContext;
804 401212 : root->init_plans = NIL;
805 401212 : root->cte_plan_ids = NIL;
806 401212 : root->multiexpr_params = NIL;
807 401212 : root->join_domains = NIL;
808 401212 : root->eq_classes = NIL;
809 401212 : root->ec_merging_done = false;
810 401212 : root->last_rinfo_serial = 0;
811 401212 : root->all_result_relids =
812 401212 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
813 401212 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
814 401212 : root->append_rel_list = NIL;
815 401212 : root->row_identity_vars = NIL;
816 401212 : root->rowMarks = NIL;
817 401212 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
818 401212 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
819 401212 : root->processed_groupClause = NIL;
820 401212 : root->processed_distinctClause = NIL;
821 401212 : root->processed_tlist = NIL;
822 401212 : root->update_colnos = NIL;
823 401212 : root->grouping_map = NULL;
824 401212 : root->minmax_aggs = NIL;
825 401212 : root->qual_security_level = 0;
826 401212 : root->hasPseudoConstantQuals = false;
827 401212 : root->hasAlternativeSubPlans = false;
828 401212 : root->placeholdersFrozen = false;
829 401212 : root->hasRecursion = hasRecursion;
830 401212 : root->assumeReplanning = false;
831 401212 : if (hasRecursion)
832 692 : root->wt_param_id = assign_special_exec_param(root);
833 : else
834 400520 : root->wt_param_id = -1;
835 401212 : root->non_recursive_path = NULL;
836 :
837 : /*
838 : * Create the top-level join domain. This won't have valid contents until
839 : * deconstruct_jointree fills it in, but the node needs to exist before
840 : * that so we can build EquivalenceClasses referencing it.
841 : */
842 401212 : root->join_domains = list_make1(makeNode(JoinDomain));
843 :
844 : /*
845 : * If there is a WITH list, process each WITH query and either convert it
846 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
847 : */
848 401212 : if (parse->cteList)
849 2211 : SS_process_ctes(root);
850 :
851 : /*
852 : * If it's a MERGE command, transform the joinlist as appropriate.
853 : */
854 401208 : transform_MERGE_to_join(parse);
855 :
856 : /*
857 : * Scan the rangetable for relation RTEs and retrieve the necessary
858 : * catalog information for each relation. Using this information, clear
859 : * the inh flag for any relation that has no children, collect not-null
860 : * attribute numbers for any relation that has column not-null
861 : * constraints, and expand virtual generated columns for any relation that
862 : * contains them. Note that this step does not descend into sublinks and
863 : * subqueries; if we pull up any sublinks or subqueries below, their
864 : * relation RTEs are processed just before pulling them up.
865 : */
866 401208 : parse = root->parse = preprocess_relation_rtes(root);
867 :
868 : /*
869 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
870 : * that we don't need so many special cases to deal with that situation.
871 : */
872 401208 : replace_empty_jointree(parse);
873 :
874 : /*
875 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
876 : * to transform them into joins. Note that this step does not descend
877 : * into subqueries; if we pull up any subqueries below, their SubLinks are
878 : * processed just before pulling them up.
879 : */
880 401208 : if (parse->hasSubLinks)
881 31388 : pull_up_sublinks(root);
882 :
883 : /*
884 : * Scan the rangetable for function RTEs, do const-simplification on them,
885 : * and then inline them if possible (producing subqueries that might get
886 : * pulled up next). Recursion issues here are handled in the same way as
887 : * for SubLinks.
888 : */
889 401208 : preprocess_function_rtes(root);
890 :
891 : /*
892 : * Check to see if any subqueries in the jointree can be merged into this
893 : * query.
894 : */
895 401204 : pull_up_subqueries(root);
896 :
897 : /*
898 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
899 : * do this now because it requires applying pull_up_subqueries to the leaf
900 : * queries of the UNION ALL, which weren't touched above because they
901 : * weren't referenced by the jointree (they will be after we do this).
902 : */
903 401204 : if (parse->setOperations)
904 5500 : flatten_simple_union_all(root);
905 :
906 : /*
907 : * Survey the rangetable to see what kinds of entries are present. We can
908 : * skip some later processing if relevant SQL features are not used; for
909 : * example if there are no JOIN RTEs we can avoid the expense of doing
910 : * flatten_join_alias_vars(). This must be done after we have finished
911 : * adding rangetable entries, of course. (Note: actually, processing of
912 : * inherited or partitioned rels can cause RTEs for their child tables to
913 : * get added later; but those must all be RTE_RELATION entries, so they
914 : * don't invalidate the conclusions drawn here.)
915 : */
916 401204 : root->hasJoinRTEs = false;
917 401204 : root->hasLateralRTEs = false;
918 401204 : root->group_rtindex = 0;
919 401204 : hasOuterJoins = false;
920 401204 : hasResultRTEs = false;
921 1110359 : foreach(l, parse->rtable)
922 : {
923 709155 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
924 :
925 709155 : switch (rte->rtekind)
926 : {
927 74466 : case RTE_JOIN:
928 74466 : root->hasJoinRTEs = true;
929 74466 : if (IS_OUTER_JOIN(rte->jointype))
930 37270 : hasOuterJoins = true;
931 74466 : break;
932 147568 : case RTE_RESULT:
933 147568 : hasResultRTEs = true;
934 147568 : break;
935 4321 : case RTE_GROUP:
936 : Assert(parse->hasGroupRTE);
937 4321 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
938 4321 : break;
939 482800 : default:
940 : /* No work here for other RTE types */
941 482800 : break;
942 : }
943 :
944 709155 : if (rte->lateral)
945 8329 : root->hasLateralRTEs = true;
946 :
947 : /*
948 : * We can also determine the maximum security level required for any
949 : * securityQuals now. Addition of inheritance-child RTEs won't affect
950 : * this, because child tables don't have their own securityQuals; see
951 : * expand_single_inheritance_child().
952 : */
953 709155 : if (rte->securityQuals)
954 2455 : root->qual_security_level = Max(root->qual_security_level,
955 : list_length(rte->securityQuals));
956 : }
957 :
958 : /*
959 : * If we have now verified that the query target relation is
960 : * non-inheriting, mark it as a leaf target.
961 : */
962 401204 : if (parse->resultRelation)
963 : {
964 65542 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
965 :
966 65542 : if (!rte->inh)
967 63226 : root->leaf_result_relids =
968 63226 : bms_make_singleton(parse->resultRelation);
969 : }
970 :
971 : /*
972 : * This would be a convenient time to check access permissions for all
973 : * relations mentioned in the query, since it would be better to fail now,
974 : * before doing any detailed planning. However, for historical reasons,
975 : * we leave this to be done at executor startup.
976 : *
977 : * Note, however, that we do need to check access permissions for any view
978 : * relations mentioned in the query, in order to prevent information being
979 : * leaked by selectivity estimation functions, which only check view owner
980 : * permissions on underlying tables (see all_rows_selectable() and its
981 : * callers). This is a little ugly, because it means that access
982 : * permissions for views will be checked twice, which is another reason
983 : * why it would be better to do all the ACL checks here.
984 : */
985 1109571 : foreach(l, parse->rtable)
986 : {
987 708633 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
988 :
989 708633 : if (rte->perminfoindex != 0 &&
990 384154 : rte->relkind == RELKIND_VIEW)
991 : {
992 : RTEPermissionInfo *perminfo;
993 : bool result;
994 :
995 16532 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
996 16532 : result = ExecCheckOneRelPerms(perminfo);
997 16532 : if (!result)
998 266 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
999 266 : get_rel_name(perminfo->relid));
1000 : }
1001 : }
1002 :
1003 : /*
1004 : * Preprocess RowMark information. We need to do this after subquery
1005 : * pullup, so that all base relations are present.
1006 : */
1007 400938 : preprocess_rowmarks(root);
1008 :
1009 : /*
1010 : * Set hasHavingQual to remember if HAVING clause is present. Needed
1011 : * because preprocess_expression will reduce a constant-true condition to
1012 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
1013 : */
1014 400938 : root->hasHavingQual = (parse->havingQual != NULL);
1015 :
1016 : /*
1017 : * Do expression preprocessing on targetlist and quals, as well as other
1018 : * random expressions in the querytree. Note that we do not need to
1019 : * handle sort/group expressions explicitly, because they are actually
1020 : * part of the targetlist.
1021 : */
1022 398305 : parse->targetList = (List *)
1023 400938 : preprocess_expression(root, (Node *) parse->targetList,
1024 : EXPRKIND_TARGET);
1025 :
1026 398305 : newWithCheckOptions = NIL;
1027 400873 : foreach(l, parse->withCheckOptions)
1028 : {
1029 2568 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
1030 :
1031 2568 : wco->qual = preprocess_expression(root, wco->qual,
1032 : EXPRKIND_QUAL);
1033 2568 : if (wco->qual != NULL)
1034 2236 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
1035 : }
1036 398305 : parse->withCheckOptions = newWithCheckOptions;
1037 :
1038 398305 : parse->returningList = (List *)
1039 398305 : preprocess_expression(root, (Node *) parse->returningList,
1040 : EXPRKIND_TARGET);
1041 :
1042 398305 : preprocess_qual_conditions(root, (Node *) parse->jointree);
1043 :
1044 398305 : parse->havingQual = preprocess_expression(root, parse->havingQual,
1045 : EXPRKIND_QUAL);
1046 :
1047 400737 : foreach(l, parse->windowClause)
1048 : {
1049 2432 : WindowClause *wc = lfirst_node(WindowClause, l);
1050 :
1051 : /* partitionClause/orderClause are sort/group expressions */
1052 2432 : wc->startOffset = preprocess_expression(root, wc->startOffset,
1053 : EXPRKIND_LIMIT);
1054 2432 : wc->endOffset = preprocess_expression(root, wc->endOffset,
1055 : EXPRKIND_LIMIT);
1056 : }
1057 :
1058 398305 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
1059 : EXPRKIND_LIMIT);
1060 398305 : parse->limitCount = preprocess_expression(root, parse->limitCount,
1061 : EXPRKIND_LIMIT);
1062 :
1063 398305 : if (parse->onConflict)
1064 : {
1065 3578 : parse->onConflict->arbiterElems = (List *)
1066 1789 : preprocess_expression(root,
1067 1789 : (Node *) parse->onConflict->arbiterElems,
1068 : EXPRKIND_ARBITER_ELEM);
1069 3578 : parse->onConflict->arbiterWhere =
1070 1789 : preprocess_expression(root,
1071 1789 : parse->onConflict->arbiterWhere,
1072 : EXPRKIND_QUAL);
1073 3578 : parse->onConflict->onConflictSet = (List *)
1074 1789 : preprocess_expression(root,
1075 1789 : (Node *) parse->onConflict->onConflictSet,
1076 : EXPRKIND_TARGET);
1077 1789 : parse->onConflict->onConflictWhere =
1078 1789 : preprocess_expression(root,
1079 1789 : parse->onConflict->onConflictWhere,
1080 : EXPRKIND_QUAL);
1081 : /* exclRelTlist contains only Vars, so no preprocessing needed */
1082 : }
1083 :
1084 400571 : foreach(l, parse->mergeActionList)
1085 : {
1086 2266 : MergeAction *action = (MergeAction *) lfirst(l);
1087 :
1088 2266 : action->targetList = (List *)
1089 2266 : preprocess_expression(root,
1090 2266 : (Node *) action->targetList,
1091 : EXPRKIND_TARGET);
1092 2266 : action->qual =
1093 2266 : preprocess_expression(root,
1094 : (Node *) action->qual,
1095 : EXPRKIND_QUAL);
1096 : }
1097 :
1098 398305 : parse->mergeJoinCondition =
1099 398305 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1100 :
1101 398305 : root->append_rel_list = (List *)
1102 398305 : preprocess_expression(root, (Node *) root->append_rel_list,
1103 : EXPRKIND_APPINFO);
1104 :
1105 : /* Also need to preprocess expressions within RTEs */
1106 1103638 : foreach(l, parse->rtable)
1107 : {
1108 705333 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1109 : int kind;
1110 : ListCell *lcsq;
1111 :
1112 705333 : if (rte->rtekind == RTE_RELATION)
1113 : {
1114 370721 : if (rte->tablesample)
1115 180 : rte->tablesample = (TableSampleClause *)
1116 180 : preprocess_expression(root,
1117 180 : (Node *) rte->tablesample,
1118 : EXPRKIND_TABLESAMPLE);
1119 : }
1120 334612 : else if (rte->rtekind == RTE_SUBQUERY)
1121 : {
1122 : /*
1123 : * We don't want to do all preprocessing yet on the subquery's
1124 : * expressions, since that will happen when we plan it. But if it
1125 : * contains any join aliases of our level, those have to get
1126 : * expanded now, because planning of the subquery won't do it.
1127 : * That's only possible if the subquery is LATERAL.
1128 : */
1129 64068 : if (rte->lateral && root->hasJoinRTEs)
1130 1595 : rte->subquery = (Query *)
1131 1595 : flatten_join_alias_vars(root, root->parse,
1132 1595 : (Node *) rte->subquery);
1133 : }
1134 270544 : else if (rte->rtekind == RTE_FUNCTION)
1135 : {
1136 : /* Preprocess the function expression(s) fully */
1137 35541 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1138 35541 : rte->functions = (List *)
1139 35541 : preprocess_expression(root, (Node *) rte->functions, kind);
1140 : }
1141 235003 : else if (rte->rtekind == RTE_TABLEFUNC)
1142 : {
1143 : /* Preprocess the function expression(s) fully */
1144 519 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1145 519 : rte->tablefunc = (TableFunc *)
1146 519 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1147 : }
1148 234484 : else if (rte->rtekind == RTE_VALUES)
1149 : {
1150 : /* Preprocess the values lists fully */
1151 6891 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1152 6891 : rte->values_lists = (List *)
1153 6891 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1154 : }
1155 227593 : else if (rte->rtekind == RTE_GROUP)
1156 : {
1157 : /* Preprocess the groupexprs list fully */
1158 4321 : rte->groupexprs = (List *)
1159 4321 : preprocess_expression(root, (Node *) rte->groupexprs,
1160 : EXPRKIND_GROUPEXPR);
1161 : }
1162 :
1163 : /*
1164 : * Process each element of the securityQuals list as if it were a
1165 : * separate qual expression (as indeed it is). We need to do it this
1166 : * way to get proper canonicalization of AND/OR structure. Note that
1167 : * this converts each element into an implicit-AND sublist.
1168 : */
1169 708138 : foreach(lcsq, rte->securityQuals)
1170 : {
1171 2805 : lfirst(lcsq) = preprocess_expression(root,
1172 2805 : (Node *) lfirst(lcsq),
1173 : EXPRKIND_QUAL);
1174 : }
1175 : }
1176 :
1177 : /*
1178 : * Now that we are done preprocessing expressions, and in particular done
1179 : * flattening join alias variables, get rid of the joinaliasvars lists.
1180 : * They no longer match what expressions in the rest of the tree look
1181 : * like, because we have not preprocessed expressions in those lists (and
1182 : * do not want to; for example, expanding a SubLink there would result in
1183 : * a useless unreferenced subplan). Leaving them in place simply creates
1184 : * a hazard for later scans of the tree. We could try to prevent that by
1185 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1186 : * but that doesn't sound very reliable.
1187 : */
1188 398305 : if (root->hasJoinRTEs)
1189 : {
1190 259158 : foreach(l, parse->rtable)
1191 : {
1192 212384 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1193 :
1194 212384 : rte->joinaliasvars = NIL;
1195 : }
1196 : }
1197 :
1198 : /*
1199 : * Before we flatten GROUP Vars, check which HAVING clauses have collation
1200 : * conflicts. When GROUP BY uses a nondeterministic collation, values
1201 : * that are "equal" for grouping may be distinguishable under a different
1202 : * collation. If such a HAVING clause were moved to WHERE, it would
1203 : * filter individual rows before grouping, potentially eliminating some
1204 : * members of a group and thereby changing aggregate results.
1205 : *
1206 : * We do this check before flatten_group_exprs because we can easily
1207 : * identify grouping expressions by checking whether a Var references
1208 : * RTE_GROUP, and such Vars directly carry the GROUP BY collation as their
1209 : * varcollid. After flattening, these Vars are replaced by the underlying
1210 : * expressions, and we would have to match expressions in the HAVING
1211 : * clause back to grouping expressions, which is much more complex.
1212 : */
1213 398305 : if (parse->hasGroupRTE)
1214 : havingCollationConflicts =
1215 4321 : find_having_collation_conflicts(parse, root->group_rtindex);
1216 : else
1217 393984 : havingCollationConflicts = NULL;
1218 :
1219 : /*
1220 : * Replace any Vars in the subquery's targetlist and havingQual that
1221 : * reference GROUP outputs with the underlying grouping expressions.
1222 : *
1223 : * Note that we need to perform this replacement after we've preprocessed
1224 : * the grouping expressions. This is to ensure that there is only one
1225 : * instance of SubPlan for each SubLink contained within the grouping
1226 : * expressions.
1227 : */
1228 398305 : if (parse->hasGroupRTE)
1229 : {
1230 4321 : parse->targetList = (List *)
1231 4321 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1232 4321 : parse->havingQual =
1233 4321 : flatten_group_exprs(root, root->parse, parse->havingQual);
1234 : }
1235 :
1236 : /* Constant-folding might have removed all set-returning functions */
1237 398305 : if (parse->hasTargetSRFs)
1238 10366 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1239 :
1240 : /*
1241 : * If we have grouping sets, expand the groupingSets tree of this query to
1242 : * a flat list of grouping sets. We need to do this before optimizing
1243 : * HAVING, since we can't easily tell if there's an empty grouping set
1244 : * until we have this representation.
1245 : */
1246 398305 : if (parse->groupingSets)
1247 : {
1248 905 : parse->groupingSets =
1249 905 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1250 : }
1251 :
1252 : /*
1253 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1254 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1255 : * volatile functions (since a HAVING clause is supposed to be executed
1256 : * only once per group). We also can't do this if there are any grouping
1257 : * sets and the clause references any columns that are nullable by the
1258 : * grouping sets; the nulled values of those columns are not available
1259 : * before the grouping step. (The test on groupClause might seem wrong,
1260 : * but it's okay: it's just an optimization to avoid running pull_varnos
1261 : * when there cannot be any Vars in the HAVING clause.)
1262 : *
1263 : * We also cannot do this if the HAVING clause uses a different collation
1264 : * than the GROUP BY for any grouping expression whose GROUP BY collation
1265 : * is nondeterministic. This is detected before flatten_group_exprs (see
1266 : * find_having_collation_conflicts above) and recorded in the
1267 : * havingCollationConflicts bitmapset. The bitmapset indexes remain valid
1268 : * here because flatten_group_exprs uses expression_tree_mutator, which
1269 : * preserves the list length and ordering of havingQual.
1270 : *
1271 : * Also, it may be that the clause is so expensive to execute that we're
1272 : * better off doing it only once per group, despite the loss of
1273 : * selectivity. This is hard to estimate short of doing the entire
1274 : * planning process twice, so we use a heuristic: clauses containing
1275 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1276 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1277 : * instead of after.
1278 : *
1279 : * If the query has no empty grouping set then we can simply move such a
1280 : * clause into WHERE; any group that fails the clause will not be in the
1281 : * output because none of its tuples will reach the grouping or
1282 : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1283 : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1284 : * clause must be degenerate (variable-free), so we can copy it into WHERE
1285 : * so that query_planner() can use it in a gating Result node. (This could
1286 : * be done better, but it seems not worth optimizing.)
1287 : *
1288 : * Note that a HAVING clause may contain expressions that are not fully
1289 : * preprocessed. This can happen if these expressions are part of
1290 : * grouping items. In such cases, they are replaced with GROUP Vars in
1291 : * the parser and then replaced back after we're done with expression
1292 : * preprocessing on havingQual. This is not an issue if the clause
1293 : * remains in HAVING, because these expressions will be matched to lower
1294 : * target items in setrefs.c. However, if the clause is moved or copied
1295 : * into WHERE, we need to ensure that these expressions are fully
1296 : * preprocessed.
1297 : *
1298 : * Note that both havingQual and parse->jointree->quals are in
1299 : * implicitly-ANDed-list form at this point, even though they are declared
1300 : * as Node *.
1301 : */
1302 398305 : newHaving = NIL;
1303 398305 : havingIdx = 0;
1304 399577 : foreach(l, (List *) parse->havingQual)
1305 : {
1306 1272 : Node *havingclause = (Node *) lfirst(l);
1307 :
1308 1863 : if (contain_agg_clause(havingclause) ||
1309 1182 : contain_volatile_functions(havingclause) ||
1310 1182 : contain_subplans(havingclause) ||
1311 591 : bms_is_member(havingIdx, havingCollationConflicts) ||
1312 611 : (parse->groupClause && parse->groupingSets &&
1313 100 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1314 : {
1315 : /* keep it in HAVING */
1316 821 : newHaving = lappend(newHaving, havingclause);
1317 : }
1318 451 : else if (parse->groupClause &&
1319 421 : (parse->groupingSets == NIL ||
1320 40 : (List *) linitial(parse->groupingSets) != NIL))
1321 411 : {
1322 : /* There is GROUP BY, but no empty grouping set */
1323 : Node *whereclause;
1324 :
1325 : /* Preprocess the HAVING clause fully */
1326 411 : whereclause = preprocess_expression(root, havingclause,
1327 : EXPRKIND_QUAL);
1328 : /* ... and move it to WHERE */
1329 411 : parse->jointree->quals = (Node *)
1330 411 : list_concat((List *) parse->jointree->quals,
1331 : (List *) whereclause);
1332 : }
1333 : else
1334 : {
1335 : /* There is an empty grouping set (perhaps implicitly) */
1336 : Node *whereclause;
1337 :
1338 : /* Preprocess the HAVING clause fully */
1339 40 : whereclause = preprocess_expression(root, copyObject(havingclause),
1340 : EXPRKIND_QUAL);
1341 : /* ... and put a copy in WHERE */
1342 80 : parse->jointree->quals = (Node *)
1343 40 : list_concat((List *) parse->jointree->quals,
1344 : (List *) whereclause);
1345 : /* ... and also keep it in HAVING */
1346 40 : newHaving = lappend(newHaving, havingclause);
1347 : }
1348 :
1349 1272 : havingIdx++;
1350 : }
1351 398305 : parse->havingQual = (Node *) newHaving;
1352 :
1353 : /*
1354 : * If we have any outer joins, try to reduce them to plain inner joins.
1355 : * This step is most easily done after we've done expression
1356 : * preprocessing.
1357 : */
1358 398305 : if (hasOuterJoins)
1359 25838 : reduce_outer_joins(root);
1360 :
1361 : /*
1362 : * If we have any RTE_RESULT relations, see if they can be deleted from
1363 : * the jointree. We also rely on this processing to flatten single-child
1364 : * FromExprs underneath outer joins. This step is most effectively done
1365 : * after we've done expression preprocessing and outer join reduction.
1366 : */
1367 398305 : if (hasResultRTEs || hasOuterJoins)
1368 169579 : remove_useless_result_rtes(root);
1369 :
1370 : /*
1371 : * Do the main planning.
1372 : */
1373 398305 : grouping_planner(root, tuple_fraction, setops);
1374 :
1375 : /*
1376 : * Capture the set of outer-level param IDs we have access to, for use in
1377 : * extParam/allParam calculations later.
1378 : */
1379 398259 : SS_identify_outer_params(root);
1380 :
1381 : /*
1382 : * If any initPlans were created in this query level, adjust the surviving
1383 : * Paths' costs and parallel-safety flags to account for them. The
1384 : * initPlans won't actually get attached to the plan tree till
1385 : * create_plan() runs, but we must include their effects now.
1386 : */
1387 398259 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1388 398259 : SS_charge_for_initplans(root, final_rel);
1389 :
1390 : /*
1391 : * Make sure we've identified the cheapest Path for the final rel. (By
1392 : * doing this here not in grouping_planner, we include initPlan costs in
1393 : * the decision, though it's unlikely that will change anything.)
1394 : */
1395 398259 : set_cheapest(final_rel);
1396 :
1397 398259 : return root;
1398 : }
1399 :
1400 : /*
1401 : * preprocess_expression
1402 : * Do subquery_planner's preprocessing work for an expression,
1403 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1404 : * conditions), a HAVING clause, or a few other things.
1405 : */
1406 : static Node *
1407 3357258 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1408 : {
1409 : /*
1410 : * Fall out quickly if expression is empty. This occurs often enough to
1411 : * be worth checking. Note that null->null is the correct conversion for
1412 : * implicit-AND result format, too.
1413 : */
1414 3357258 : if (expr == NULL)
1415 2639341 : return NULL;
1416 :
1417 : /*
1418 : * If the query has any join RTEs, replace join alias variables with
1419 : * base-relation variables. We must do this first, since any expressions
1420 : * we may extract from the joinaliasvars lists have not been preprocessed.
1421 : * For example, if we did this after sublink processing, sublinks expanded
1422 : * out from join aliases would not get processed. But we can skip this in
1423 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1424 : * they can't contain any Vars of the current query level.
1425 : */
1426 717917 : if (root->hasJoinRTEs &&
1427 332999 : !(kind == EXPRKIND_RTFUNC ||
1428 166351 : kind == EXPRKIND_VALUES ||
1429 : kind == EXPRKIND_TABLESAMPLE ||
1430 : kind == EXPRKIND_TABLEFUNC))
1431 166336 : expr = flatten_join_alias_vars(root, root->parse, expr);
1432 :
1433 : /*
1434 : * Simplify constant expressions. For function RTEs, this was already
1435 : * done by preprocess_function_rtes. (But note we must do it again for
1436 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1437 : * un-simplified subexpressions inserted by flattening of subqueries or
1438 : * join alias variables.)
1439 : *
1440 : * Note: an essential effect of this is to convert named-argument function
1441 : * calls to positional notation and insert the current actual values of
1442 : * any default arguments for functions. To ensure that happens, we *must*
1443 : * process all expressions here. Previous PG versions sometimes skipped
1444 : * const-simplification if it didn't seem worth the trouble, but we can't
1445 : * do that anymore.
1446 : *
1447 : * Note: this also flattens nested AND and OR expressions into N-argument
1448 : * form. All processing of a qual expression after this point must be
1449 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1450 : * with AND directly under AND, nor OR directly under OR.
1451 : */
1452 717917 : if (kind != EXPRKIND_RTFUNC)
1453 687086 : expr = eval_const_expressions(root, expr);
1454 :
1455 : /*
1456 : * If it's a qual or havingQual, canonicalize it.
1457 : */
1458 715284 : if (kind == EXPRKIND_QUAL)
1459 : {
1460 265180 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1461 :
1462 : #ifdef OPTIMIZER_DEBUG
1463 : printf("After canonicalize_qual()\n");
1464 : pprint(expr);
1465 : #endif
1466 : }
1467 :
1468 : /*
1469 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1470 : * hashfuncid of any that might execute more quickly by using hash lookups
1471 : * instead of a linear search.
1472 : */
1473 715284 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1474 : {
1475 656366 : convert_saop_to_hashed_saop(expr);
1476 : }
1477 :
1478 : /* Expand SubLinks to SubPlans */
1479 715284 : if (root->parse->hasSubLinks)
1480 92122 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1481 :
1482 : /*
1483 : * XXX do not insert anything here unless you have grokked the comments in
1484 : * SS_replace_correlation_vars ...
1485 : */
1486 :
1487 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1488 715284 : if (root->query_level > 1)
1489 149712 : expr = SS_replace_correlation_vars(root, expr);
1490 :
1491 : /*
1492 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1493 : * don't want to do this before eval_const_expressions, since the latter
1494 : * would be unable to simplify a top-level AND correctly. Also,
1495 : * SS_process_sublinks expects explicit-AND format.)
1496 : */
1497 715284 : if (kind == EXPRKIND_QUAL)
1498 265180 : expr = (Node *) make_ands_implicit((Expr *) expr);
1499 :
1500 715284 : return expr;
1501 : }
1502 :
1503 : /*
1504 : * preprocess_qual_conditions
1505 : * Recursively scan the query's jointree and do subquery_planner's
1506 : * preprocessing work on each qual condition found therein.
1507 : */
1508 : static void
1509 1011595 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1510 : {
1511 1011595 : if (jtnode == NULL)
1512 0 : return;
1513 1011595 : if (IsA(jtnode, RangeTblRef))
1514 : {
1515 : /* nothing to do here */
1516 : }
1517 496587 : else if (IsA(jtnode, FromExpr))
1518 : {
1519 413287 : FromExpr *f = (FromExpr *) jtnode;
1520 : ListCell *l;
1521 :
1522 859977 : foreach(l, f->fromlist)
1523 446690 : preprocess_qual_conditions(root, lfirst(l));
1524 :
1525 413287 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1526 : }
1527 83300 : else if (IsA(jtnode, JoinExpr))
1528 : {
1529 83300 : JoinExpr *j = (JoinExpr *) jtnode;
1530 :
1531 83300 : preprocess_qual_conditions(root, j->larg);
1532 83300 : preprocess_qual_conditions(root, j->rarg);
1533 :
1534 83300 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1535 : }
1536 : else
1537 0 : elog(ERROR, "unrecognized node type: %d",
1538 : (int) nodeTag(jtnode));
1539 : }
1540 :
1541 : /*
1542 : * find_having_collation_conflicts
1543 : * Identify HAVING clauses that must not be moved to WHERE due to collation
1544 : * mismatches with GROUP BY.
1545 : *
1546 : * This must be called before flatten_group_exprs, while the HAVING clause
1547 : * still contains GROUP Vars (Vars referencing RTE_GROUP). These GROUP Vars
1548 : * carry the GROUP BY collation as their varcollid. A GROUP Var with a
1549 : * nondeterministic varcollid conflicts whenever some collation-aware ancestor
1550 : * on its path applies a different inputcollid: that operator would distinguish
1551 : * values which the GROUP BY considers equal, so the clause is unsafe to push
1552 : * to WHERE.
1553 : *
1554 : * Returns a Bitmapset of zero-based indexes into the havingQual list for
1555 : * clauses that have collation conflicts and must stay in HAVING.
1556 : */
1557 : static Bitmapset *
1558 4321 : find_having_collation_conflicts(Query *parse, Index group_rtindex)
1559 : {
1560 4321 : Bitmapset *result = NULL;
1561 : having_collation_ctx ctx;
1562 : int idx;
1563 :
1564 4321 : if (parse->havingQual == NULL)
1565 3483 : return NULL;
1566 :
1567 838 : ctx.group_rtindex = group_rtindex;
1568 838 : ctx.ancestor_collids = NIL;
1569 :
1570 838 : idx = 0;
1571 2653 : foreach_ptr(Node, clause, (List *) parse->havingQual)
1572 : {
1573 977 : if (having_collation_conflict_walker(clause, &ctx))
1574 80 : result = bms_add_member(result, idx);
1575 977 : idx++;
1576 : Assert(ctx.ancestor_collids == NIL);
1577 : }
1578 :
1579 838 : return result;
1580 : }
1581 :
1582 : /*
1583 : * Walker function for find_having_collation_conflicts.
1584 : *
1585 : * Walk the clause top-down, maintaining a stack of inputcollids contributed
1586 : * by collation-aware ancestors. At each GROUP Var with a nondeterministic
1587 : * varcollid, the clause has a conflict if any ancestor's inputcollid differs
1588 : * from the GROUP Var's varcollid. Most collation-aware nodes expose their
1589 : * inputcollid through exprInputCollation(). Two structural exceptions need
1590 : * special handling:
1591 : *
1592 : * - RowCompareExpr carries one inputcollid per column in inputcollids[], so we
1593 : * descend into its (largs[i], rargs[i]) pairs explicitly with the matching
1594 : * collation pushed onto the stack.
1595 : *
1596 : * - A simple CASE (CaseExpr with a non-NULL arg) holds the arg outside the
1597 : * WHEN's OpExpr, even though the WHEN's OpExpr is the place where the
1598 : * comparison's inputcollid lives. Parse analysis builds each WHEN as
1599 : * "OpExpr(CaseTestExpr op val)" -- the CaseTestExpr is a placeholder for
1600 : * the arg. Before walking cexpr->arg we therefore push every WHEN's
1601 : * inputcollid onto the ancestor stack, so a GROUP Var at the arg is
1602 : * checked against the same collations the WHEN comparisons would apply.
1603 : * The WHEN bodies and defresult are then walked under the unchanged stack
1604 : * so their own collation contexts are picked up by the default path.
1605 : */
1606 : static bool
1607 4306 : having_collation_conflict_walker(Node *node, having_collation_ctx *ctx)
1608 : {
1609 : Oid this_collid;
1610 : bool result;
1611 :
1612 4306 : if (node == NULL)
1613 453 : return false;
1614 :
1615 3853 : if (IsA(node, Var))
1616 : {
1617 928 : Var *var = (Var *) node;
1618 :
1619 : /* We should not see any upper-level Vars here */
1620 : Assert(var->varlevelsup == 0);
1621 :
1622 928 : if (var->varno == ctx->group_rtindex &&
1623 544 : OidIsValid(var->varcollid) &&
1624 361 : !get_collation_isdeterministic(var->varcollid))
1625 : {
1626 280 : foreach_oid(collid, ctx->ancestor_collids)
1627 : {
1628 160 : if (collid != var->varcollid)
1629 80 : return true;
1630 : }
1631 : }
1632 848 : return false;
1633 : }
1634 :
1635 2925 : if (IsA(node, RowCompareExpr))
1636 : {
1637 10 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
1638 : ListCell *lc_l;
1639 : ListCell *lc_r;
1640 : ListCell *lc_c;
1641 :
1642 : /*
1643 : * Each column of a row comparison is compared under its own
1644 : * inputcollids[i]. Walk each (largs[i], rargs[i]) pair with that
1645 : * collation pushed, so a Var in column i is checked against the
1646 : * collation that actually applies to it.
1647 : */
1648 10 : forthree(lc_l, rcexpr->largs,
1649 : lc_r, rcexpr->rargs,
1650 : lc_c, rcexpr->inputcollids)
1651 : {
1652 10 : Oid collid = lfirst_oid(lc_c);
1653 : bool found;
1654 :
1655 10 : if (OidIsValid(collid))
1656 10 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1657 : collid);
1658 :
1659 10 : found = having_collation_conflict_walker((Node *) lfirst(lc_l),
1660 10 : ctx) ||
1661 0 : having_collation_conflict_walker((Node *) lfirst(lc_r),
1662 : ctx);
1663 :
1664 10 : if (OidIsValid(collid))
1665 10 : ctx->ancestor_collids =
1666 10 : list_delete_last(ctx->ancestor_collids);
1667 :
1668 10 : if (found)
1669 10 : return true;
1670 : }
1671 0 : return false;
1672 : }
1673 :
1674 2915 : if (IsA(node, CaseExpr) && ((CaseExpr *) node)->arg != NULL)
1675 : {
1676 30 : CaseExpr *cexpr = (CaseExpr *) node;
1677 30 : int saved_len = list_length(ctx->ancestor_collids);
1678 : bool found;
1679 :
1680 : /*
1681 : * Push every WHEN's inputcollid before walking cexpr->arg, since each
1682 : * WHEN implicitly compares the arg under that inputcollid.
1683 : */
1684 90 : foreach_node(CaseWhen, cw, cexpr->args)
1685 : {
1686 30 : Oid collid = exprInputCollation((Node *) cw->expr);
1687 :
1688 30 : if (OidIsValid(collid))
1689 30 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1690 : collid);
1691 : }
1692 :
1693 30 : found = having_collation_conflict_walker((Node *) cexpr->arg, ctx);
1694 :
1695 30 : ctx->ancestor_collids = list_truncate(ctx->ancestor_collids,
1696 : saved_len);
1697 :
1698 30 : if (found)
1699 20 : return true;
1700 :
1701 : /*
1702 : * Walk the WHEN bodies and defresult under the unchanged ancestor
1703 : * stack; any inputcollids inside them are picked up by the default
1704 : * path.
1705 : */
1706 30 : foreach_node(CaseWhen, cw, cexpr->args)
1707 : {
1708 20 : if (having_collation_conflict_walker((Node *) cw->expr, ctx) ||
1709 10 : having_collation_conflict_walker((Node *) cw->result, ctx))
1710 0 : return true;
1711 : }
1712 10 : return having_collation_conflict_walker((Node *) cexpr->defresult,
1713 : ctx);
1714 : }
1715 :
1716 2885 : this_collid = exprInputCollation(node);
1717 2885 : if (OidIsValid(this_collid))
1718 366 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1719 : this_collid);
1720 :
1721 2885 : result = expression_tree_walker(node, having_collation_conflict_walker,
1722 : ctx);
1723 :
1724 2885 : if (OidIsValid(this_collid))
1725 366 : ctx->ancestor_collids = list_delete_last(ctx->ancestor_collids);
1726 :
1727 2885 : return result;
1728 : }
1729 :
1730 : /*
1731 : * preprocess_phv_expression
1732 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1733 : *
1734 : * If a LATERAL subquery references an output of another subquery, and that
1735 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1736 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1737 : * and later pull it back up during find_lateral_references, which runs after
1738 : * subquery_planner has preprocessed all the expressions that were in the
1739 : * current query level to start with. So we need to preprocess it then.
1740 : */
1741 : Expr *
1742 75 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1743 : {
1744 75 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1745 : }
1746 :
1747 : /*--------------------
1748 : * grouping_planner
1749 : * Perform planning steps related to grouping, aggregation, etc.
1750 : *
1751 : * This function adds all required top-level processing to the scan/join
1752 : * Path(s) produced by query_planner.
1753 : *
1754 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1755 : * tuple_fraction is interpreted as follows:
1756 : * 0: expect all tuples to be retrieved (normal case)
1757 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1758 : * from the plan to be retrieved
1759 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1760 : * expected to be retrieved (ie, a LIMIT specification).
1761 : * setops is used for set operation subqueries to provide the subquery with
1762 : * the context in which it's being used so that Paths correctly sorted for the
1763 : * set operation can be generated. NULL when not planning a set operation
1764 : * child, or when a child of a set op that isn't interested in sorted input.
1765 : *
1766 : * Returns nothing; the useful output is in the Paths we attach to the
1767 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1768 : * root->processed_tlist contains the final processed targetlist.
1769 : *
1770 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1771 : * to leave this to the caller.
1772 : *--------------------
1773 : */
1774 : static void
1775 398305 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1776 : SetOperationStmt *setops)
1777 : {
1778 398305 : Query *parse = root->parse;
1779 398305 : int64 offset_est = 0;
1780 398305 : int64 count_est = 0;
1781 398305 : double limit_tuples = -1.0;
1782 398305 : bool have_postponed_srfs = false;
1783 : PathTarget *final_target;
1784 : List *final_targets;
1785 : List *final_targets_contain_srfs;
1786 : bool final_target_parallel_safe;
1787 : RelOptInfo *current_rel;
1788 : RelOptInfo *final_rel;
1789 : FinalPathExtraData extra;
1790 : ListCell *lc;
1791 :
1792 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1793 398305 : if (parse->limitCount || parse->limitOffset)
1794 : {
1795 3712 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1796 : &offset_est, &count_est);
1797 :
1798 : /*
1799 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1800 : * estimate the effects of using a bounded sort.
1801 : */
1802 3712 : if (count_est > 0 && offset_est >= 0)
1803 3250 : limit_tuples = (double) count_est + (double) offset_est;
1804 : }
1805 :
1806 : /* Make tuple_fraction accessible to lower-level routines */
1807 398305 : root->tuple_fraction = tuple_fraction;
1808 :
1809 398305 : if (parse->setOperations)
1810 : {
1811 : /*
1812 : * Construct Paths for set operations. The results will not need any
1813 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1814 : * special work for recursive unions is the responsibility of
1815 : * plan_set_operations.
1816 : */
1817 4956 : current_rel = plan_set_operations(root);
1818 :
1819 : /*
1820 : * We should not need to call preprocess_targetlist, since we must be
1821 : * in a SELECT query node. Instead, use the processed_tlist returned
1822 : * by plan_set_operations (since this tells whether it returned any
1823 : * resjunk columns!), and transfer any sort key information from the
1824 : * original tlist.
1825 : */
1826 : Assert(parse->commandType == CMD_SELECT);
1827 :
1828 : /* for safety, copy processed_tlist instead of modifying in-place */
1829 4952 : root->processed_tlist =
1830 4952 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1831 : parse->targetList);
1832 :
1833 : /* Also extract the PathTarget form of the setop result tlist */
1834 4952 : final_target = current_rel->cheapest_total_path->pathtarget;
1835 :
1836 : /* And check whether it's parallel safe */
1837 : final_target_parallel_safe =
1838 4952 : is_parallel_safe(root, (Node *) final_target->exprs);
1839 :
1840 : /* The setop result tlist couldn't contain any SRFs */
1841 : Assert(!parse->hasTargetSRFs);
1842 4952 : final_targets = final_targets_contain_srfs = NIL;
1843 :
1844 : /*
1845 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1846 : * checked already, but let's make sure).
1847 : */
1848 4952 : if (parse->rowMarks)
1849 0 : ereport(ERROR,
1850 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1851 : /*------
1852 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1853 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1854 : LCS_asString(linitial_node(RowMarkClause,
1855 : parse->rowMarks)->strength))));
1856 :
1857 : /*
1858 : * Calculate pathkeys that represent result ordering requirements
1859 : */
1860 : Assert(parse->distinctClause == NIL);
1861 4952 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1862 : parse->sortClause,
1863 : root->processed_tlist);
1864 : }
1865 : else
1866 : {
1867 : /* No set operations, do regular planning */
1868 : PathTarget *sort_input_target;
1869 : List *sort_input_targets;
1870 : List *sort_input_targets_contain_srfs;
1871 : bool sort_input_target_parallel_safe;
1872 : PathTarget *grouping_target;
1873 : List *grouping_targets;
1874 : List *grouping_targets_contain_srfs;
1875 : bool grouping_target_parallel_safe;
1876 : PathTarget *scanjoin_target;
1877 : List *scanjoin_targets;
1878 : List *scanjoin_targets_contain_srfs;
1879 : bool scanjoin_target_parallel_safe;
1880 : bool scanjoin_target_same_exprs;
1881 : bool have_grouping;
1882 393349 : WindowFuncLists *wflists = NULL;
1883 393349 : List *activeWindows = NIL;
1884 393349 : grouping_sets_data *gset_data = NULL;
1885 : standard_qp_extra qp_extra;
1886 :
1887 : /* A recursive query should always have setOperations */
1888 : Assert(!root->hasRecursion);
1889 :
1890 : /* Preprocess grouping sets and GROUP BY clause, if any */
1891 393349 : if (parse->groupingSets)
1892 : {
1893 905 : gset_data = preprocess_grouping_sets(root);
1894 : }
1895 392444 : else if (parse->groupClause)
1896 : {
1897 : /* Preprocess regular GROUP BY clause, if any */
1898 3481 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1899 : }
1900 :
1901 : /*
1902 : * Preprocess targetlist. Note that much of the remaining planning
1903 : * work will be done with the PathTarget representation of tlists, but
1904 : * we must also maintain the full representation of the final tlist so
1905 : * that we can transfer its decoration (resnames etc) to the topmost
1906 : * tlist of the finished Plan. This is kept in processed_tlist.
1907 : */
1908 393345 : preprocess_targetlist(root);
1909 :
1910 : /*
1911 : * Mark all the aggregates with resolved aggtranstypes, and detect
1912 : * aggregates that are duplicates or can share transition state. We
1913 : * must do this before slicing and dicing the tlist into various
1914 : * pathtargets, else some copies of the Aggref nodes might escape
1915 : * being marked.
1916 : */
1917 393345 : if (parse->hasAggs)
1918 : {
1919 33363 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1920 33363 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1921 : }
1922 :
1923 : /*
1924 : * Locate any window functions in the tlist. (We don't need to look
1925 : * anywhere else, since expressions used in ORDER BY will be in there
1926 : * too.) Note that they could all have been eliminated by constant
1927 : * folding, in which case we don't need to do any more work.
1928 : */
1929 393345 : if (parse->hasWindowFuncs)
1930 : {
1931 2222 : wflists = find_window_functions((Node *) root->processed_tlist,
1932 2222 : list_length(parse->windowClause));
1933 2222 : if (wflists->numWindowFuncs > 0)
1934 : {
1935 : /*
1936 : * See if any modifications can be made to each WindowClause
1937 : * to allow the executor to execute the WindowFuncs more
1938 : * quickly.
1939 : */
1940 2217 : optimize_window_clauses(root, wflists);
1941 :
1942 : /* Extract the list of windows actually in use. */
1943 2217 : activeWindows = select_active_windows(root, wflists);
1944 :
1945 : /* Make sure they all have names, for EXPLAIN's use. */
1946 2217 : name_active_windows(activeWindows);
1947 : }
1948 : else
1949 5 : parse->hasWindowFuncs = false;
1950 : }
1951 :
1952 : /*
1953 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1954 : * adding logic between here and the query_planner() call. Anything
1955 : * that is needed in MIN/MAX-optimizable cases will have to be
1956 : * duplicated in planagg.c.
1957 : */
1958 393345 : if (parse->hasAggs)
1959 33363 : preprocess_minmax_aggregates(root);
1960 :
1961 : /*
1962 : * Figure out whether there's a hard limit on the number of rows that
1963 : * query_planner's result subplan needs to return. Even if we know a
1964 : * hard limit overall, it doesn't apply if the query has any
1965 : * grouping/aggregation operations, or SRFs in the tlist.
1966 : */
1967 393345 : if (parse->groupClause ||
1968 389028 : parse->groupingSets ||
1969 388963 : parse->distinctClause ||
1970 387016 : parse->hasAggs ||
1971 357352 : parse->hasWindowFuncs ||
1972 355251 : parse->hasTargetSRFs ||
1973 345219 : root->hasHavingQual)
1974 48146 : root->limit_tuples = -1.0;
1975 : else
1976 345199 : root->limit_tuples = limit_tuples;
1977 :
1978 : /* Set up data needed by standard_qp_callback */
1979 393345 : qp_extra.activeWindows = activeWindows;
1980 393345 : qp_extra.gset_data = gset_data;
1981 :
1982 : /*
1983 : * If we're a subquery for a set operation, store the SetOperationStmt
1984 : * in qp_extra.
1985 : */
1986 393345 : qp_extra.setop = setops;
1987 :
1988 : /*
1989 : * Generate the best unsorted and presorted paths for the scan/join
1990 : * portion of this Query, ie the processing represented by the
1991 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1992 : * We also generate (in standard_qp_callback) pathkey representations
1993 : * of the query's sort clause, distinct clause, etc.
1994 : */
1995 393345 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1996 :
1997 : /*
1998 : * Convert the query's result tlist into PathTarget format.
1999 : *
2000 : * Note: this cannot be done before query_planner() has performed
2001 : * appendrel expansion, because that might add resjunk entries to
2002 : * root->processed_tlist. Waiting till afterwards is also helpful
2003 : * because the target width estimates can use per-Var width numbers
2004 : * that were obtained within query_planner().
2005 : */
2006 393311 : final_target = create_pathtarget(root, root->processed_tlist);
2007 : final_target_parallel_safe =
2008 393311 : is_parallel_safe(root, (Node *) final_target->exprs);
2009 :
2010 : /*
2011 : * If ORDER BY was given, consider whether we should use a post-sort
2012 : * projection, and compute the adjusted target for preceding steps if
2013 : * so.
2014 : */
2015 393311 : if (parse->sortClause)
2016 : {
2017 58900 : sort_input_target = make_sort_input_target(root,
2018 : final_target,
2019 : &have_postponed_srfs);
2020 : sort_input_target_parallel_safe =
2021 58900 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
2022 : }
2023 : else
2024 : {
2025 334411 : sort_input_target = final_target;
2026 334411 : sort_input_target_parallel_safe = final_target_parallel_safe;
2027 : }
2028 :
2029 : /*
2030 : * If we have window functions to deal with, the output from any
2031 : * grouping step needs to be what the window functions want;
2032 : * otherwise, it should be sort_input_target.
2033 : */
2034 393311 : if (activeWindows)
2035 : {
2036 2217 : grouping_target = make_window_input_target(root,
2037 : final_target,
2038 : activeWindows);
2039 : grouping_target_parallel_safe =
2040 2217 : is_parallel_safe(root, (Node *) grouping_target->exprs);
2041 : }
2042 : else
2043 : {
2044 391094 : grouping_target = sort_input_target;
2045 391094 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
2046 : }
2047 :
2048 : /*
2049 : * If we have grouping or aggregation to do, the topmost scan/join
2050 : * plan node must emit what the grouping step wants; otherwise, it
2051 : * should emit grouping_target.
2052 : */
2053 388994 : have_grouping = (parse->groupClause || parse->groupingSets ||
2054 782305 : parse->hasAggs || root->hasHavingQual);
2055 393311 : if (have_grouping)
2056 : {
2057 34097 : scanjoin_target = make_group_input_target(root, final_target);
2058 : scanjoin_target_parallel_safe =
2059 34097 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2060 : }
2061 : else
2062 : {
2063 359214 : scanjoin_target = grouping_target;
2064 359214 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2065 : }
2066 :
2067 : /*
2068 : * If there are any SRFs in the targetlist, we must separate each of
2069 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2070 : * each of the named targets with a SRF-free version, and remember the
2071 : * list of additional projection steps we need to add afterwards.
2072 : */
2073 393311 : if (parse->hasTargetSRFs)
2074 : {
2075 : /* final_target doesn't recompute any SRFs in sort_input_target */
2076 10366 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2077 : &final_targets,
2078 : &final_targets_contain_srfs);
2079 10366 : final_target = linitial_node(PathTarget, final_targets);
2080 : Assert(!linitial_int(final_targets_contain_srfs));
2081 : /* likewise for sort_input_target vs. grouping_target */
2082 10366 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2083 : &sort_input_targets,
2084 : &sort_input_targets_contain_srfs);
2085 10366 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2086 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2087 : /* likewise for grouping_target vs. scanjoin_target */
2088 10366 : split_pathtarget_at_srfs_grouping(root,
2089 : grouping_target, scanjoin_target,
2090 : &grouping_targets,
2091 : &grouping_targets_contain_srfs);
2092 10366 : grouping_target = linitial_node(PathTarget, grouping_targets);
2093 : Assert(!linitial_int(grouping_targets_contain_srfs));
2094 : /* scanjoin_target will not have any SRFs precomputed for it */
2095 10366 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2096 : &scanjoin_targets,
2097 : &scanjoin_targets_contain_srfs);
2098 10366 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2099 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2100 : }
2101 : else
2102 : {
2103 : /* initialize lists; for most of these, dummy values are OK */
2104 382945 : final_targets = final_targets_contain_srfs = NIL;
2105 382945 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2106 382945 : grouping_targets = grouping_targets_contain_srfs = NIL;
2107 382945 : scanjoin_targets = list_make1(scanjoin_target);
2108 382945 : scanjoin_targets_contain_srfs = NIL;
2109 : }
2110 :
2111 : /* Apply scan/join target. */
2112 393311 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2113 393311 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2114 393311 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2115 : scanjoin_targets_contain_srfs,
2116 : scanjoin_target_parallel_safe,
2117 : scanjoin_target_same_exprs);
2118 :
2119 : /*
2120 : * Save the various upper-rel PathTargets we just computed into
2121 : * root->upper_targets[]. The core code doesn't use this, but it
2122 : * provides a convenient place for extensions to get at the info. For
2123 : * consistency, we save all the intermediate targets, even though some
2124 : * of the corresponding upperrels might not be needed for this query.
2125 : */
2126 393311 : root->upper_targets[UPPERREL_FINAL] = final_target;
2127 393311 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2128 393311 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2129 393311 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
2130 393311 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2131 393311 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2132 :
2133 : /*
2134 : * If we have grouping and/or aggregation, consider ways to implement
2135 : * that. We build a new upperrel representing the output of this
2136 : * phase.
2137 : */
2138 393311 : if (have_grouping)
2139 : {
2140 34097 : current_rel = create_grouping_paths(root,
2141 : current_rel,
2142 : grouping_target,
2143 : grouping_target_parallel_safe,
2144 : gset_data);
2145 : /* Fix things up if grouping_target contains SRFs */
2146 34093 : if (parse->hasTargetSRFs)
2147 299 : adjust_paths_for_srfs(root, current_rel,
2148 : grouping_targets,
2149 : grouping_targets_contain_srfs);
2150 : }
2151 :
2152 : /*
2153 : * If we have window functions, consider ways to implement those. We
2154 : * build a new upperrel representing the output of this phase.
2155 : */
2156 393307 : if (activeWindows)
2157 : {
2158 2217 : current_rel = create_window_paths(root,
2159 : current_rel,
2160 : grouping_target,
2161 : sort_input_target,
2162 : sort_input_target_parallel_safe,
2163 : wflists,
2164 : activeWindows);
2165 : /* Fix things up if sort_input_target contains SRFs */
2166 2217 : if (parse->hasTargetSRFs)
2167 10 : adjust_paths_for_srfs(root, current_rel,
2168 : sort_input_targets,
2169 : sort_input_targets_contain_srfs);
2170 : }
2171 :
2172 : /*
2173 : * If there is a DISTINCT clause, consider ways to implement that. We
2174 : * build a new upperrel representing the output of this phase.
2175 : */
2176 393307 : if (parse->distinctClause)
2177 : {
2178 1974 : current_rel = create_distinct_paths(root,
2179 : current_rel,
2180 : sort_input_target);
2181 : }
2182 : } /* end of if (setOperations) */
2183 :
2184 : /*
2185 : * If ORDER BY was given, consider ways to implement that, and generate a
2186 : * new upperrel containing only paths that emit the correct ordering and
2187 : * project the correct final_target. We can apply the original
2188 : * limit_tuples limit in sort costing here, but only if there are no
2189 : * postponed SRFs.
2190 : */
2191 398259 : if (parse->sortClause)
2192 : {
2193 62167 : current_rel = create_ordered_paths(root,
2194 : current_rel,
2195 : final_target,
2196 : final_target_parallel_safe,
2197 : have_postponed_srfs ? -1.0 :
2198 : limit_tuples);
2199 : /* Fix things up if final_target contains SRFs */
2200 62167 : if (parse->hasTargetSRFs)
2201 182 : adjust_paths_for_srfs(root, current_rel,
2202 : final_targets,
2203 : final_targets_contain_srfs);
2204 : }
2205 :
2206 : /*
2207 : * Now we are prepared to build the final-output upperrel.
2208 : */
2209 398259 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2210 :
2211 : /*
2212 : * If the input rel is marked consider_parallel and there's nothing that's
2213 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2214 : * consider_parallel as well. Note that if the query has rowMarks or is
2215 : * not a SELECT, consider_parallel will be false for every relation in the
2216 : * query.
2217 : */
2218 537659 : if (current_rel->consider_parallel &&
2219 278780 : is_parallel_safe(root, parse->limitOffset) &&
2220 139380 : is_parallel_safe(root, parse->limitCount))
2221 139375 : final_rel->consider_parallel = true;
2222 :
2223 : /*
2224 : * If the current_rel belongs to a single FDW, so does the final_rel.
2225 : */
2226 398259 : final_rel->serverid = current_rel->serverid;
2227 398259 : final_rel->userid = current_rel->userid;
2228 398259 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2229 398259 : final_rel->fdwroutine = current_rel->fdwroutine;
2230 :
2231 : /*
2232 : * Generate paths for the final_rel. Insert all surviving paths, with
2233 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2234 : */
2235 813086 : foreach(lc, current_rel->pathlist)
2236 : {
2237 414827 : Path *path = (Path *) lfirst(lc);
2238 :
2239 : /*
2240 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2241 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2242 : * here. If there are only non-locking rowmarks, they should be
2243 : * handled by the ModifyTable node instead. However, root->rowMarks
2244 : * is what goes into the LockRows node.)
2245 : */
2246 414827 : if (parse->rowMarks)
2247 : {
2248 6826 : path = (Path *) create_lockrows_path(root, final_rel, path,
2249 : root->rowMarks,
2250 : assign_special_exec_param(root));
2251 : }
2252 :
2253 : /*
2254 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2255 : */
2256 414827 : if (limit_needed(parse))
2257 : {
2258 4197 : path = (Path *) create_limit_path(root, final_rel, path,
2259 : parse->limitOffset,
2260 : parse->limitCount,
2261 : parse->limitOption,
2262 : offset_est, count_est);
2263 : }
2264 :
2265 : /*
2266 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
2267 : */
2268 414827 : if (parse->commandType != CMD_SELECT)
2269 : {
2270 : Index rootRelation;
2271 65188 : List *resultRelations = NIL;
2272 65188 : List *updateColnosLists = NIL;
2273 65188 : List *withCheckOptionLists = NIL;
2274 65188 : List *returningLists = NIL;
2275 65188 : List *mergeActionLists = NIL;
2276 65188 : List *mergeJoinConditions = NIL;
2277 : List *rowMarks;
2278 :
2279 65188 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
2280 : {
2281 : /* Inherited UPDATE/DELETE/MERGE */
2282 2288 : RelOptInfo *top_result_rel = find_base_rel(root,
2283 : parse->resultRelation);
2284 2288 : int resultRelation = -1;
2285 :
2286 : /* Pass the root result rel forward to the executor. */
2287 2288 : rootRelation = parse->resultRelation;
2288 :
2289 : /* Add only leaf children to ModifyTable. */
2290 6718 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
2291 6718 : resultRelation)) >= 0)
2292 : {
2293 4430 : RelOptInfo *this_result_rel = find_base_rel(root,
2294 : resultRelation);
2295 :
2296 : /*
2297 : * Also exclude any leaf rels that have turned dummy since
2298 : * being added to the list, for example, by being excluded
2299 : * by constraint exclusion.
2300 : */
2301 4430 : if (IS_DUMMY_REL(this_result_rel))
2302 153 : continue;
2303 :
2304 : /* Build per-target-rel lists needed by ModifyTable */
2305 4277 : resultRelations = lappend_int(resultRelations,
2306 : resultRelation);
2307 4277 : if (parse->commandType == CMD_UPDATE)
2308 : {
2309 2933 : List *update_colnos = root->update_colnos;
2310 :
2311 2933 : if (this_result_rel != top_result_rel)
2312 : update_colnos =
2313 2933 : adjust_inherited_attnums_multilevel(root,
2314 : update_colnos,
2315 : this_result_rel->relid,
2316 : top_result_rel->relid);
2317 2933 : updateColnosLists = lappend(updateColnosLists,
2318 : update_colnos);
2319 : }
2320 4277 : if (parse->withCheckOptions)
2321 : {
2322 416 : List *withCheckOptions = parse->withCheckOptions;
2323 :
2324 416 : if (this_result_rel != top_result_rel)
2325 : withCheckOptions = (List *)
2326 416 : adjust_appendrel_attrs_multilevel(root,
2327 : (Node *) withCheckOptions,
2328 : this_result_rel,
2329 : top_result_rel);
2330 416 : withCheckOptionLists = lappend(withCheckOptionLists,
2331 : withCheckOptions);
2332 : }
2333 4277 : if (parse->returningList)
2334 : {
2335 645 : List *returningList = parse->returningList;
2336 :
2337 645 : if (this_result_rel != top_result_rel)
2338 : returningList = (List *)
2339 645 : adjust_appendrel_attrs_multilevel(root,
2340 : (Node *) returningList,
2341 : this_result_rel,
2342 : top_result_rel);
2343 645 : returningLists = lappend(returningLists,
2344 : returningList);
2345 : }
2346 4277 : if (parse->mergeActionList)
2347 : {
2348 : ListCell *l;
2349 425 : List *mergeActionList = NIL;
2350 :
2351 : /*
2352 : * Copy MergeActions and translate stuff that
2353 : * references attribute numbers.
2354 : */
2355 1322 : foreach(l, parse->mergeActionList)
2356 : {
2357 897 : MergeAction *action = lfirst(l),
2358 897 : *leaf_action = copyObject(action);
2359 :
2360 897 : leaf_action->qual =
2361 897 : adjust_appendrel_attrs_multilevel(root,
2362 : (Node *) action->qual,
2363 : this_result_rel,
2364 : top_result_rel);
2365 897 : leaf_action->targetList = (List *)
2366 897 : adjust_appendrel_attrs_multilevel(root,
2367 897 : (Node *) action->targetList,
2368 : this_result_rel,
2369 : top_result_rel);
2370 897 : if (leaf_action->commandType == CMD_UPDATE)
2371 491 : leaf_action->updateColnos =
2372 491 : adjust_inherited_attnums_multilevel(root,
2373 : action->updateColnos,
2374 : this_result_rel->relid,
2375 : top_result_rel->relid);
2376 897 : mergeActionList = lappend(mergeActionList,
2377 : leaf_action);
2378 : }
2379 :
2380 425 : mergeActionLists = lappend(mergeActionLists,
2381 : mergeActionList);
2382 : }
2383 4277 : if (parse->commandType == CMD_MERGE)
2384 : {
2385 425 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2386 :
2387 425 : if (this_result_rel != top_result_rel)
2388 : mergeJoinCondition =
2389 425 : adjust_appendrel_attrs_multilevel(root,
2390 : mergeJoinCondition,
2391 : this_result_rel,
2392 : top_result_rel);
2393 425 : mergeJoinConditions = lappend(mergeJoinConditions,
2394 : mergeJoinCondition);
2395 : }
2396 : }
2397 :
2398 2288 : if (resultRelations == NIL)
2399 : {
2400 : /*
2401 : * We managed to exclude every child rel, so generate a
2402 : * dummy one-relation plan using info for the top target
2403 : * rel (even though that may not be a leaf target).
2404 : * Although it's clear that no data will be updated or
2405 : * deleted, we still need to have a ModifyTable node so
2406 : * that any statement triggers will be executed. (This
2407 : * could be cleaner if we fixed nodeModifyTable.c to allow
2408 : * zero target relations, but that probably wouldn't be a
2409 : * net win.)
2410 : */
2411 28 : resultRelations = list_make1_int(parse->resultRelation);
2412 28 : if (parse->commandType == CMD_UPDATE)
2413 26 : updateColnosLists = list_make1(root->update_colnos);
2414 28 : if (parse->withCheckOptions)
2415 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2416 28 : if (parse->returningList)
2417 15 : returningLists = list_make1(parse->returningList);
2418 28 : if (parse->mergeActionList)
2419 1 : mergeActionLists = list_make1(parse->mergeActionList);
2420 28 : if (parse->commandType == CMD_MERGE)
2421 1 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2422 : }
2423 : }
2424 : else
2425 : {
2426 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2427 62900 : rootRelation = 0; /* there's no separate root rel */
2428 62900 : resultRelations = list_make1_int(parse->resultRelation);
2429 62900 : if (parse->commandType == CMD_UPDATE)
2430 8821 : updateColnosLists = list_make1(root->update_colnos);
2431 62900 : if (parse->withCheckOptions)
2432 889 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2433 62900 : if (parse->returningList)
2434 2207 : returningLists = list_make1(parse->returningList);
2435 62900 : if (parse->mergeActionList)
2436 1286 : mergeActionLists = list_make1(parse->mergeActionList);
2437 62900 : if (parse->commandType == CMD_MERGE)
2438 1286 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2439 : }
2440 :
2441 : /*
2442 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2443 : * will have dealt with fetching non-locked marked rows, else we
2444 : * need to have ModifyTable do that.
2445 : */
2446 65188 : if (parse->rowMarks)
2447 0 : rowMarks = NIL;
2448 : else
2449 65188 : rowMarks = root->rowMarks;
2450 :
2451 : path = (Path *)
2452 65188 : create_modifytable_path(root, final_rel,
2453 : path,
2454 : parse->commandType,
2455 65188 : parse->canSetTag,
2456 65188 : parse->resultRelation,
2457 : rootRelation,
2458 : resultRelations,
2459 : updateColnosLists,
2460 : withCheckOptionLists,
2461 : returningLists,
2462 : rowMarks,
2463 : parse->onConflict,
2464 : mergeActionLists,
2465 : mergeJoinConditions,
2466 : parse->forPortionOf,
2467 : assign_special_exec_param(root));
2468 : }
2469 :
2470 : /* And shove it into final_rel */
2471 414827 : add_path(final_rel, path);
2472 : }
2473 :
2474 : /*
2475 : * Generate partial paths for final_rel, too, if outer query levels might
2476 : * be able to make use of them.
2477 : */
2478 398259 : if (final_rel->consider_parallel && root->query_level > 1 &&
2479 24850 : !limit_needed(parse))
2480 : {
2481 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2482 24784 : foreach(lc, current_rel->partial_pathlist)
2483 : {
2484 112 : Path *partial_path = (Path *) lfirst(lc);
2485 :
2486 112 : add_partial_path(final_rel, partial_path);
2487 : }
2488 : }
2489 :
2490 398259 : extra.limit_needed = limit_needed(parse);
2491 398259 : extra.limit_tuples = limit_tuples;
2492 398259 : extra.count_est = count_est;
2493 398259 : extra.offset_est = offset_est;
2494 :
2495 : /*
2496 : * If there is an FDW that's responsible for all baserels of the query,
2497 : * let it consider adding ForeignPaths.
2498 : */
2499 398259 : if (final_rel->fdwroutine &&
2500 657 : final_rel->fdwroutine->GetForeignUpperPaths)
2501 620 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2502 : current_rel, final_rel,
2503 : &extra);
2504 :
2505 : /* Let extensions possibly add some more paths */
2506 398259 : if (create_upper_paths_hook)
2507 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2508 : current_rel, final_rel, &extra);
2509 :
2510 : /* Note: currently, we leave it to callers to do set_cheapest() */
2511 398259 : }
2512 :
2513 : /*
2514 : * Do preprocessing for groupingSets clause and related data.
2515 : *
2516 : * We expect that parse->groupingSets has already been expanded into a flat
2517 : * list of grouping sets (that is, just integer Lists of ressortgroupref
2518 : * numbers) by expand_grouping_sets(). This function handles the preliminary
2519 : * steps of organizing the grouping sets into lists of rollups, and preparing
2520 : * annotations which will later be filled in with size estimates.
2521 : */
2522 : static grouping_sets_data *
2523 905 : preprocess_grouping_sets(PlannerInfo *root)
2524 : {
2525 905 : Query *parse = root->parse;
2526 : List *sets;
2527 905 : int maxref = 0;
2528 : ListCell *lc_set;
2529 905 : grouping_sets_data *gd = palloc0_object(grouping_sets_data);
2530 :
2531 : /*
2532 : * We don't currently make any attempt to optimize the groupClause when
2533 : * there are grouping sets, so just duplicate it in processed_groupClause.
2534 : */
2535 905 : root->processed_groupClause = parse->groupClause;
2536 :
2537 : /* Detect unhashable and unsortable grouping expressions */
2538 905 : gd->any_hashable = false;
2539 905 : gd->unhashable_refs = NULL;
2540 905 : gd->unsortable_refs = NULL;
2541 905 : gd->unsortable_sets = NIL;
2542 :
2543 905 : if (parse->groupClause)
2544 : {
2545 : ListCell *lc;
2546 :
2547 2651 : foreach(lc, parse->groupClause)
2548 : {
2549 1811 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2550 1811 : Index ref = gc->tleSortGroupRef;
2551 :
2552 1811 : if (ref > maxref)
2553 1771 : maxref = ref;
2554 :
2555 1811 : if (!gc->hashable)
2556 24 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2557 :
2558 1811 : if (!OidIsValid(gc->sortop))
2559 33 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2560 : }
2561 : }
2562 :
2563 : /* Allocate workspace array for remapping */
2564 905 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2565 :
2566 : /*
2567 : * If we have any unsortable sets, we must extract them before trying to
2568 : * prepare rollups. Unsortable sets don't go through
2569 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2570 : * here.
2571 : */
2572 905 : if (!bms_is_empty(gd->unsortable_refs))
2573 : {
2574 33 : List *sortable_sets = NIL;
2575 : ListCell *lc;
2576 :
2577 99 : foreach(lc, parse->groupingSets)
2578 : {
2579 70 : List *gset = (List *) lfirst(lc);
2580 :
2581 70 : if (bms_overlap_list(gd->unsortable_refs, gset))
2582 : {
2583 38 : GroupingSetData *gs = makeNode(GroupingSetData);
2584 :
2585 38 : gs->set = gset;
2586 38 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2587 :
2588 : /*
2589 : * We must enforce here that an unsortable set is hashable;
2590 : * later code assumes this. Parse analysis only checks that
2591 : * every individual column is either hashable or sortable.
2592 : *
2593 : * Note that passing this test doesn't guarantee we can
2594 : * generate a plan; there might be other showstoppers.
2595 : */
2596 38 : if (bms_overlap_list(gd->unhashable_refs, gset))
2597 4 : ereport(ERROR,
2598 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2599 : errmsg("could not implement GROUP BY"),
2600 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2601 : }
2602 : else
2603 32 : sortable_sets = lappend(sortable_sets, gset);
2604 : }
2605 :
2606 29 : if (sortable_sets)
2607 24 : sets = extract_rollup_sets(sortable_sets);
2608 : else
2609 5 : sets = NIL;
2610 : }
2611 : else
2612 872 : sets = extract_rollup_sets(parse->groupingSets);
2613 :
2614 2376 : foreach(lc_set, sets)
2615 : {
2616 1475 : List *current_sets = (List *) lfirst(lc_set);
2617 1475 : RollupData *rollup = makeNode(RollupData);
2618 : GroupingSetData *gs;
2619 :
2620 : /*
2621 : * Reorder the current list of grouping sets into correct prefix
2622 : * order. If only one aggregation pass is needed, try to make the
2623 : * list match the ORDER BY clause; if more than one pass is needed, we
2624 : * don't bother with that.
2625 : *
2626 : * Note that this reorders the sets from smallest-member-first to
2627 : * largest-member-first, and applies the GroupingSetData annotations,
2628 : * though the data will be filled in later.
2629 : */
2630 1475 : current_sets = reorder_grouping_sets(current_sets,
2631 1475 : (list_length(sets) == 1
2632 : ? parse->sortClause
2633 : : NIL));
2634 :
2635 : /*
2636 : * Get the initial (and therefore largest) grouping set.
2637 : */
2638 1475 : gs = linitial_node(GroupingSetData, current_sets);
2639 :
2640 : /*
2641 : * Order the groupClause appropriately. If the first grouping set is
2642 : * empty, then the groupClause must also be empty; otherwise we have
2643 : * to force the groupClause to match that grouping set's order.
2644 : *
2645 : * (The first grouping set can be empty even though parse->groupClause
2646 : * is not empty only if all non-empty grouping sets are unsortable.
2647 : * The groupClauses for hashed grouping sets are built later on.)
2648 : */
2649 1475 : if (gs->set)
2650 1410 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2651 : else
2652 65 : rollup->groupClause = NIL;
2653 :
2654 : /*
2655 : * Is it hashable? We pretend empty sets are hashable even though we
2656 : * actually force them not to be hashed later. But don't bother if
2657 : * there's nothing but empty sets (since in that case we can't hash
2658 : * anything).
2659 : */
2660 1475 : if (gs->set &&
2661 1410 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2662 : {
2663 1390 : rollup->hashable = true;
2664 1390 : gd->any_hashable = true;
2665 : }
2666 :
2667 : /*
2668 : * Now that we've pinned down an order for the groupClause for this
2669 : * list of grouping sets, we need to remap the entries in the grouping
2670 : * sets from sortgrouprefs to plain indices (0-based) into the
2671 : * groupClause for this collection of grouping sets. We keep the
2672 : * original form for later use, though.
2673 : */
2674 1475 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2675 : current_sets,
2676 : gd->tleref_to_colnum_map);
2677 1475 : rollup->gsets_data = current_sets;
2678 :
2679 1475 : gd->rollups = lappend(gd->rollups, rollup);
2680 : }
2681 :
2682 901 : if (gd->unsortable_sets)
2683 : {
2684 : /*
2685 : * We have not yet pinned down a groupclause for this, but we will
2686 : * need index-based lists for estimation purposes. Construct
2687 : * hash_sets_idx based on the entire original groupclause for now.
2688 : */
2689 29 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2690 : gd->unsortable_sets,
2691 : gd->tleref_to_colnum_map);
2692 29 : gd->any_hashable = true;
2693 : }
2694 :
2695 901 : return gd;
2696 : }
2697 :
2698 : /*
2699 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2700 : * (without annotation) mapped to indexes into the given groupclause.
2701 : */
2702 : static List *
2703 4147 : remap_to_groupclause_idx(List *groupClause,
2704 : List *gsets,
2705 : int *tleref_to_colnum_map)
2706 : {
2707 4147 : int ref = 0;
2708 4147 : List *result = NIL;
2709 : ListCell *lc;
2710 :
2711 9861 : foreach(lc, groupClause)
2712 : {
2713 5714 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2714 :
2715 5714 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2716 : }
2717 :
2718 9444 : foreach(lc, gsets)
2719 : {
2720 5297 : List *set = NIL;
2721 : ListCell *lc2;
2722 5297 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2723 :
2724 11720 : foreach(lc2, gs->set)
2725 : {
2726 6423 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2727 : }
2728 :
2729 5297 : result = lappend(result, set);
2730 : }
2731 :
2732 4147 : return result;
2733 : }
2734 :
2735 :
2736 : /*
2737 : * preprocess_rowmarks - set up PlanRowMarks if needed
2738 : */
2739 : static void
2740 400938 : preprocess_rowmarks(PlannerInfo *root)
2741 : {
2742 400938 : Query *parse = root->parse;
2743 : Bitmapset *rels;
2744 : List *prowmarks;
2745 : ListCell *l;
2746 : int i;
2747 :
2748 400938 : if (parse->rowMarks)
2749 : {
2750 : /*
2751 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2752 : * grouping, since grouping renders a reference to individual tuple
2753 : * CTIDs invalid. This is also checked at parse time, but that's
2754 : * insufficient because of rule substitution, query pullup, etc.
2755 : */
2756 6579 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2757 : parse->rowMarks)->strength);
2758 : }
2759 : else
2760 : {
2761 : /*
2762 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2763 : * UPDATE/SHARE.
2764 : */
2765 394359 : if (parse->commandType != CMD_UPDATE &&
2766 383943 : parse->commandType != CMD_DELETE &&
2767 380223 : parse->commandType != CMD_MERGE)
2768 378743 : return;
2769 : }
2770 :
2771 : /*
2772 : * We need to have rowmarks for all base relations except the target. We
2773 : * make a bitmapset of all base rels and then remove the items we don't
2774 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2775 : */
2776 22195 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2777 22195 : if (parse->resultRelation)
2778 15616 : rels = bms_del_member(rels, parse->resultRelation);
2779 :
2780 : /*
2781 : * Convert RowMarkClauses to PlanRowMark representation.
2782 : */
2783 22195 : prowmarks = NIL;
2784 28909 : foreach(l, parse->rowMarks)
2785 : {
2786 6714 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2787 6714 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2788 : PlanRowMark *newrc;
2789 :
2790 : /*
2791 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2792 : * applied to an update/delete target rel. If that ever becomes
2793 : * possible, we should drop the target from the PlanRowMark list.
2794 : */
2795 : Assert(rc->rti != parse->resultRelation);
2796 :
2797 : /*
2798 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2799 : * can't support true locking. Subqueries that got flattened into the
2800 : * main query should be ignored completely. Any that didn't will get
2801 : * ROW_MARK_COPY items in the next loop.
2802 : */
2803 6714 : if (rte->rtekind != RTE_RELATION)
2804 46 : continue;
2805 :
2806 6668 : rels = bms_del_member(rels, rc->rti);
2807 :
2808 6668 : newrc = makeNode(PlanRowMark);
2809 6668 : newrc->rti = newrc->prti = rc->rti;
2810 6668 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2811 6668 : newrc->markType = select_rowmark_type(rte, rc->strength);
2812 6668 : newrc->allMarkTypes = (1 << newrc->markType);
2813 6668 : newrc->strength = rc->strength;
2814 6668 : newrc->waitPolicy = rc->waitPolicy;
2815 6668 : newrc->isParent = false;
2816 :
2817 6668 : prowmarks = lappend(prowmarks, newrc);
2818 : }
2819 :
2820 : /*
2821 : * Now, add rowmarks for any non-target, non-locked base relations.
2822 : */
2823 22195 : i = 0;
2824 52998 : foreach(l, parse->rtable)
2825 : {
2826 30803 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2827 : PlanRowMark *newrc;
2828 :
2829 30803 : i++;
2830 30803 : if (!bms_is_member(i, rels))
2831 27854 : continue;
2832 :
2833 2949 : newrc = makeNode(PlanRowMark);
2834 2949 : newrc->rti = newrc->prti = i;
2835 2949 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2836 2949 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2837 2949 : newrc->allMarkTypes = (1 << newrc->markType);
2838 2949 : newrc->strength = LCS_NONE;
2839 2949 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2840 2949 : newrc->isParent = false;
2841 :
2842 2949 : prowmarks = lappend(prowmarks, newrc);
2843 : }
2844 :
2845 22195 : root->rowMarks = prowmarks;
2846 : }
2847 :
2848 : /*
2849 : * Select RowMarkType to use for a given table
2850 : */
2851 : RowMarkType
2852 11511 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2853 : {
2854 11511 : if (rte->rtekind != RTE_RELATION)
2855 : {
2856 : /* If it's not a table at all, use ROW_MARK_COPY */
2857 1110 : return ROW_MARK_COPY;
2858 : }
2859 10401 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2860 : {
2861 : /* Let the FDW select the rowmark type, if it wants to */
2862 114 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2863 :
2864 114 : if (fdwroutine->GetForeignRowMarkType != NULL)
2865 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2866 : /* Otherwise, use ROW_MARK_COPY by default */
2867 114 : return ROW_MARK_COPY;
2868 : }
2869 : else
2870 : {
2871 : /* Regular table, apply the appropriate lock type */
2872 10287 : switch (strength)
2873 : {
2874 2048 : case LCS_NONE:
2875 :
2876 : /*
2877 : * We don't need a tuple lock, only the ability to re-fetch
2878 : * the row.
2879 : */
2880 2048 : return ROW_MARK_REFERENCE;
2881 : break;
2882 7070 : case LCS_FORKEYSHARE:
2883 7070 : return ROW_MARK_KEYSHARE;
2884 : break;
2885 201 : case LCS_FORSHARE:
2886 201 : return ROW_MARK_SHARE;
2887 : break;
2888 41 : case LCS_FORNOKEYUPDATE:
2889 41 : return ROW_MARK_NOKEYEXCLUSIVE;
2890 : break;
2891 927 : case LCS_FORUPDATE:
2892 927 : return ROW_MARK_EXCLUSIVE;
2893 : break;
2894 : }
2895 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2896 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2897 : }
2898 : }
2899 :
2900 : /*
2901 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2902 : *
2903 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2904 : * results back in *count_est and *offset_est. These variables are set to
2905 : * 0 if the corresponding clause is not present, and -1 if it's present
2906 : * but we couldn't estimate the value for it. (The "0" convention is OK
2907 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2908 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2909 : * usual practice of never estimating less than one row.) These values will
2910 : * be passed to create_limit_path, which see if you change this code.
2911 : *
2912 : * The return value is the suitably adjusted tuple_fraction to use for
2913 : * planning the query. This adjustment is not overridable, since it reflects
2914 : * plan actions that grouping_planner() will certainly take, not assumptions
2915 : * about context.
2916 : */
2917 : static double
2918 3712 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2919 : int64 *offset_est, int64 *count_est)
2920 : {
2921 3712 : Query *parse = root->parse;
2922 : Node *est;
2923 : double limit_fraction;
2924 :
2925 : /* Should not be called unless LIMIT or OFFSET */
2926 : Assert(parse->limitCount || parse->limitOffset);
2927 :
2928 : /*
2929 : * Try to obtain the clause values. We use estimate_expression_value
2930 : * primarily because it can sometimes do something useful with Params.
2931 : */
2932 3712 : if (parse->limitCount)
2933 : {
2934 3270 : est = estimate_expression_value(root, parse->limitCount);
2935 3270 : if (est && IsA(est, Const))
2936 : {
2937 3265 : if (((Const *) est)->constisnull)
2938 : {
2939 : /* NULL indicates LIMIT ALL, ie, no limit */
2940 0 : *count_est = 0; /* treat as not present */
2941 : }
2942 : else
2943 : {
2944 3265 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2945 3265 : if (*count_est <= 0)
2946 190 : *count_est = 1; /* force to at least 1 */
2947 : }
2948 : }
2949 : else
2950 5 : *count_est = -1; /* can't estimate */
2951 : }
2952 : else
2953 442 : *count_est = 0; /* not present */
2954 :
2955 3712 : if (parse->limitOffset)
2956 : {
2957 652 : est = estimate_expression_value(root, parse->limitOffset);
2958 652 : if (est && IsA(est, Const))
2959 : {
2960 632 : if (((Const *) est)->constisnull)
2961 : {
2962 : /* Treat NULL as no offset; the executor will too */
2963 0 : *offset_est = 0; /* treat as not present */
2964 : }
2965 : else
2966 : {
2967 632 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2968 632 : if (*offset_est < 0)
2969 0 : *offset_est = 0; /* treat as not present */
2970 : }
2971 : }
2972 : else
2973 20 : *offset_est = -1; /* can't estimate */
2974 : }
2975 : else
2976 3060 : *offset_est = 0; /* not present */
2977 :
2978 3712 : if (*count_est != 0)
2979 : {
2980 : /*
2981 : * A LIMIT clause limits the absolute number of tuples returned.
2982 : * However, if it's not a constant LIMIT then we have to guess; for
2983 : * lack of a better idea, assume 10% of the plan's result is wanted.
2984 : */
2985 3270 : if (*count_est < 0 || *offset_est < 0)
2986 : {
2987 : /* LIMIT or OFFSET is an expression ... punt ... */
2988 20 : limit_fraction = 0.10;
2989 : }
2990 : else
2991 : {
2992 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2993 3250 : limit_fraction = (double) *count_est + (double) *offset_est;
2994 : }
2995 :
2996 : /*
2997 : * If we have absolute limits from both caller and LIMIT, use the
2998 : * smaller value; likewise if they are both fractional. If one is
2999 : * fractional and the other absolute, we can't easily determine which
3000 : * is smaller, but we use the heuristic that the absolute will usually
3001 : * be smaller.
3002 : */
3003 3270 : if (tuple_fraction >= 1.0)
3004 : {
3005 5 : if (limit_fraction >= 1.0)
3006 : {
3007 : /* both absolute */
3008 5 : tuple_fraction = Min(tuple_fraction, limit_fraction);
3009 : }
3010 : else
3011 : {
3012 : /* caller absolute, limit fractional; use caller's value */
3013 : }
3014 : }
3015 3265 : else if (tuple_fraction > 0.0)
3016 : {
3017 82 : if (limit_fraction >= 1.0)
3018 : {
3019 : /* caller fractional, limit absolute; use limit */
3020 82 : tuple_fraction = limit_fraction;
3021 : }
3022 : else
3023 : {
3024 : /* both fractional */
3025 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
3026 : }
3027 : }
3028 : else
3029 : {
3030 : /* no info from caller, just use limit */
3031 3183 : tuple_fraction = limit_fraction;
3032 : }
3033 : }
3034 442 : else if (*offset_est != 0 && tuple_fraction > 0.0)
3035 : {
3036 : /*
3037 : * We have an OFFSET but no LIMIT. This acts entirely differently
3038 : * from the LIMIT case: here, we need to increase rather than decrease
3039 : * the caller's tuple_fraction, because the OFFSET acts to cause more
3040 : * tuples to be fetched instead of fewer. This only matters if we got
3041 : * a tuple_fraction > 0, however.
3042 : *
3043 : * As above, use 10% if OFFSET is present but unestimatable.
3044 : */
3045 12 : if (*offset_est < 0)
3046 0 : limit_fraction = 0.10;
3047 : else
3048 12 : limit_fraction = (double) *offset_est;
3049 :
3050 : /*
3051 : * If we have absolute counts from both caller and OFFSET, add them
3052 : * together; likewise if they are both fractional. If one is
3053 : * fractional and the other absolute, we want to take the larger, and
3054 : * we heuristically assume that's the fractional one.
3055 : */
3056 12 : if (tuple_fraction >= 1.0)
3057 : {
3058 0 : if (limit_fraction >= 1.0)
3059 : {
3060 : /* both absolute, so add them together */
3061 0 : tuple_fraction += limit_fraction;
3062 : }
3063 : else
3064 : {
3065 : /* caller absolute, limit fractional; use limit */
3066 0 : tuple_fraction = limit_fraction;
3067 : }
3068 : }
3069 : else
3070 : {
3071 12 : if (limit_fraction >= 1.0)
3072 : {
3073 : /* caller fractional, limit absolute; use caller's value */
3074 : }
3075 : else
3076 : {
3077 : /* both fractional, so add them together */
3078 0 : tuple_fraction += limit_fraction;
3079 0 : if (tuple_fraction >= 1.0)
3080 0 : tuple_fraction = 0.0; /* assume fetch all */
3081 : }
3082 : }
3083 : }
3084 :
3085 3712 : return tuple_fraction;
3086 : }
3087 :
3088 : /*
3089 : * limit_needed - do we actually need a Limit plan node?
3090 : *
3091 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
3092 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
3093 : * locution for an optimization fence. (Because other places in the planner
3094 : * merely check whether parse->limitOffset isn't NULL, it will still work as
3095 : * an optimization fence --- we're just suppressing unnecessary run-time
3096 : * overhead.)
3097 : *
3098 : * This might look like it could be merged into preprocess_limit, but there's
3099 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3100 : * in preprocess_limit it's good enough to consider estimated values.
3101 : */
3102 : bool
3103 854312 : limit_needed(Query *parse)
3104 : {
3105 : Node *node;
3106 :
3107 854312 : node = parse->limitCount;
3108 854312 : if (node)
3109 : {
3110 7785 : if (IsA(node, Const))
3111 : {
3112 : /* NULL indicates LIMIT ALL, ie, no limit */
3113 7564 : if (!((Const *) node)->constisnull)
3114 7564 : return true; /* LIMIT with a constant value */
3115 : }
3116 : else
3117 221 : return true; /* non-constant LIMIT */
3118 : }
3119 :
3120 846527 : node = parse->limitOffset;
3121 846527 : if (node)
3122 : {
3123 1283 : if (IsA(node, Const))
3124 : {
3125 : /* Treat NULL as no offset; the executor would too */
3126 1025 : if (!((Const *) node)->constisnull)
3127 : {
3128 1025 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3129 :
3130 1025 : if (offset != 0)
3131 107 : return true; /* OFFSET with a nonzero value */
3132 : }
3133 : }
3134 : else
3135 258 : return true; /* non-constant OFFSET */
3136 : }
3137 :
3138 846162 : return false; /* don't need a Limit plan node */
3139 : }
3140 :
3141 : /*
3142 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3143 : *
3144 : * The idea here is to adjust the ordering of the GROUP BY elements
3145 : * (which in itself is semantically insignificant) to match ORDER BY,
3146 : * thereby allowing a single sort operation to both implement the ORDER BY
3147 : * requirement and set up for a Unique step that implements GROUP BY.
3148 : * We also consider partial match between GROUP BY and ORDER BY elements,
3149 : * which could allow to implement ORDER BY using the incremental sort.
3150 : *
3151 : * We also consider other orderings of the GROUP BY elements, which could
3152 : * match the sort ordering of other possible plans (eg an indexscan) and
3153 : * thereby reduce cost. This is implemented during the generation of grouping
3154 : * paths. See get_useful_group_keys_orderings() for details.
3155 : *
3156 : * Note: we need no comparable processing of the distinctClause because
3157 : * the parser already enforced that that matches ORDER BY.
3158 : *
3159 : * Note: we return a fresh List, but its elements are the same
3160 : * SortGroupClauses appearing in parse->groupClause. This is important
3161 : * because later processing may modify the processed_groupClause list.
3162 : *
3163 : * For grouping sets, the order of items is instead forced to agree with that
3164 : * of the grouping set (and items not in the grouping set are skipped). The
3165 : * work of sorting the order of grouping set elements to match the ORDER BY if
3166 : * possible is done elsewhere.
3167 : */
3168 : static List *
3169 7534 : preprocess_groupclause(PlannerInfo *root, List *force)
3170 : {
3171 7534 : Query *parse = root->parse;
3172 7534 : List *new_groupclause = NIL;
3173 : ListCell *sl;
3174 : ListCell *gl;
3175 :
3176 : /* For grouping sets, we need to force the ordering */
3177 7534 : if (force)
3178 : {
3179 9704 : foreach(sl, force)
3180 : {
3181 5651 : Index ref = lfirst_int(sl);
3182 5651 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3183 :
3184 5651 : new_groupclause = lappend(new_groupclause, cl);
3185 : }
3186 :
3187 4053 : return new_groupclause;
3188 : }
3189 :
3190 : /* If no ORDER BY, nothing useful to do here */
3191 3481 : if (parse->sortClause == NIL)
3192 1984 : return list_copy(parse->groupClause);
3193 :
3194 : /*
3195 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3196 : * items, but only as far as we can make a matching prefix.
3197 : *
3198 : * This code assumes that the sortClause contains no duplicate items.
3199 : */
3200 2946 : foreach(sl, parse->sortClause)
3201 : {
3202 1965 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3203 :
3204 2872 : foreach(gl, parse->groupClause)
3205 : {
3206 2356 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3207 :
3208 2356 : if (equal(gc, sc))
3209 : {
3210 1449 : new_groupclause = lappend(new_groupclause, gc);
3211 1449 : break;
3212 : }
3213 : }
3214 1965 : if (gl == NULL)
3215 516 : break; /* no match, so stop scanning */
3216 : }
3217 :
3218 :
3219 : /* If no match at all, no point in reordering GROUP BY */
3220 1497 : if (new_groupclause == NIL)
3221 228 : return list_copy(parse->groupClause);
3222 :
3223 : /*
3224 : * Add any remaining GROUP BY items to the new list. We don't require a
3225 : * complete match, because even partial match allows ORDER BY to be
3226 : * implemented using incremental sort. Also, give up if there are any
3227 : * non-sortable GROUP BY items, since then there's no hope anyway.
3228 : */
3229 2855 : foreach(gl, parse->groupClause)
3230 : {
3231 1586 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3232 :
3233 1586 : if (list_member_ptr(new_groupclause, gc))
3234 1449 : continue; /* it matched an ORDER BY item */
3235 137 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
3236 0 : return list_copy(parse->groupClause);
3237 137 : new_groupclause = lappend(new_groupclause, gc);
3238 : }
3239 :
3240 : /* Success --- install the rearranged GROUP BY list */
3241 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3242 1269 : return new_groupclause;
3243 : }
3244 :
3245 : /*
3246 : * Extract lists of grouping sets that can be implemented using a single
3247 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3248 : *
3249 : * Input must be sorted with smallest sets first. Result has each sublist
3250 : * sorted with smallest sets first.
3251 : *
3252 : * We want to produce the absolute minimum possible number of lists here to
3253 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3254 : * of finding the minimal partition of a partially-ordered set into chains
3255 : * (which is what we need, taking the list of grouping sets as a poset ordered
3256 : * by set inclusion) can be mapped to the problem of finding the maximum
3257 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3258 : * time with a worst case of no worse than O(n^2.5) and usually much
3259 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3260 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3261 : * half a second on my modest system even with optimization off and assertions
3262 : * on.)
3263 : */
3264 : static List *
3265 896 : extract_rollup_sets(List *groupingSets)
3266 : {
3267 896 : int num_sets_raw = list_length(groupingSets);
3268 896 : int num_empty = 0;
3269 896 : int num_sets = 0; /* distinct sets */
3270 896 : int num_chains = 0;
3271 896 : List *result = NIL;
3272 : List **results;
3273 : List **orig_sets;
3274 : Bitmapset **set_masks;
3275 : int *chains;
3276 : short **adjacency;
3277 : short *adjacency_buf;
3278 : BipartiteMatchState *state;
3279 : int i;
3280 : int j;
3281 : int j_size;
3282 896 : ListCell *lc1 = list_head(groupingSets);
3283 : ListCell *lc;
3284 :
3285 : /*
3286 : * Start by stripping out empty sets. The algorithm doesn't require this,
3287 : * but the planner currently needs all empty sets to be returned in the
3288 : * first list, so we strip them here and add them back after.
3289 : */
3290 1486 : while (lc1 && lfirst(lc1) == NIL)
3291 : {
3292 590 : ++num_empty;
3293 590 : lc1 = lnext(groupingSets, lc1);
3294 : }
3295 :
3296 : /* bail out now if it turns out that all we had were empty sets. */
3297 896 : if (!lc1)
3298 65 : return list_make1(groupingSets);
3299 :
3300 : /*----------
3301 : * We don't strictly need to remove duplicate sets here, but if we don't,
3302 : * they tend to become scattered through the result, which is a bit
3303 : * confusing (and irritating if we ever decide to optimize them out).
3304 : * So we remove them here and add them back after.
3305 : *
3306 : * For each non-duplicate set, we fill in the following:
3307 : *
3308 : * orig_sets[i] = list of the original set lists
3309 : * set_masks[i] = bitmapset for testing inclusion
3310 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3311 : *
3312 : * chains[i] will be the result group this set is assigned to.
3313 : *
3314 : * We index all of these from 1 rather than 0 because it is convenient
3315 : * to leave 0 free for the NIL node in the graph algorithm.
3316 : *----------
3317 : */
3318 831 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3319 831 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3320 831 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3321 831 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3322 :
3323 831 : j_size = 0;
3324 831 : j = 0;
3325 831 : i = 1;
3326 :
3327 2861 : for_each_cell(lc, groupingSets, lc1)
3328 : {
3329 2030 : List *candidate = (List *) lfirst(lc);
3330 2030 : Bitmapset *candidate_set = NULL;
3331 : ListCell *lc2;
3332 2030 : int dup_of = 0;
3333 :
3334 4827 : foreach(lc2, candidate)
3335 : {
3336 2797 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3337 : }
3338 :
3339 : /* we can only be a dup if we're the same length as a previous set */
3340 2030 : if (j_size == list_length(candidate))
3341 : {
3342 : int k;
3343 :
3344 1828 : for (k = j; k < i; ++k)
3345 : {
3346 1169 : if (bms_equal(set_masks[k], candidate_set))
3347 : {
3348 129 : dup_of = k;
3349 129 : break;
3350 : }
3351 : }
3352 : }
3353 1242 : else if (j_size < list_length(candidate))
3354 : {
3355 1242 : j_size = list_length(candidate);
3356 1242 : j = i;
3357 : }
3358 :
3359 2030 : if (dup_of > 0)
3360 : {
3361 129 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3362 129 : bms_free(candidate_set);
3363 : }
3364 : else
3365 : {
3366 : int k;
3367 1901 : int n_adj = 0;
3368 :
3369 1901 : orig_sets[i] = list_make1(candidate);
3370 1901 : set_masks[i] = candidate_set;
3371 :
3372 : /* fill in adjacency list; no need to compare equal-size sets */
3373 :
3374 2973 : for (k = j - 1; k > 0; --k)
3375 : {
3376 1072 : if (bms_is_subset(set_masks[k], candidate_set))
3377 937 : adjacency_buf[++n_adj] = k;
3378 : }
3379 :
3380 1901 : if (n_adj > 0)
3381 : {
3382 513 : adjacency_buf[0] = n_adj;
3383 513 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3384 513 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3385 : }
3386 : else
3387 1388 : adjacency[i] = NULL;
3388 :
3389 1901 : ++i;
3390 : }
3391 : }
3392 :
3393 831 : num_sets = i - 1;
3394 :
3395 : /*
3396 : * Apply the graph matching algorithm to do the work.
3397 : */
3398 831 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3399 :
3400 : /*
3401 : * Now, the state->pair* fields have the info we need to assign sets to
3402 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3403 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3404 : * it in one pass)
3405 : */
3406 831 : chains = palloc0((num_sets + 1) * sizeof(int));
3407 :
3408 2732 : for (i = 1; i <= num_sets; ++i)
3409 : {
3410 1901 : int u = state->pair_vu[i];
3411 1901 : int v = state->pair_uv[i];
3412 :
3413 1901 : if (u > 0 && u < i)
3414 0 : chains[i] = chains[u];
3415 1901 : else if (v > 0 && v < i)
3416 491 : chains[i] = chains[v];
3417 : else
3418 1410 : chains[i] = ++num_chains;
3419 : }
3420 :
3421 : /* build result lists. */
3422 831 : results = palloc0((num_chains + 1) * sizeof(List *));
3423 :
3424 2732 : for (i = 1; i <= num_sets; ++i)
3425 : {
3426 1901 : int c = chains[i];
3427 :
3428 : Assert(c > 0);
3429 :
3430 1901 : results[c] = list_concat(results[c], orig_sets[i]);
3431 : }
3432 :
3433 : /* push any empty sets back on the first list. */
3434 1306 : while (num_empty-- > 0)
3435 475 : results[1] = lcons(NIL, results[1]);
3436 :
3437 : /* make result list */
3438 2241 : for (i = 1; i <= num_chains; ++i)
3439 1410 : result = lappend(result, results[i]);
3440 :
3441 : /*
3442 : * Free all the things.
3443 : *
3444 : * (This is over-fussy for small sets but for large sets we could have
3445 : * tied up a nontrivial amount of memory.)
3446 : */
3447 831 : BipartiteMatchFree(state);
3448 831 : pfree(results);
3449 831 : pfree(chains);
3450 2732 : for (i = 1; i <= num_sets; ++i)
3451 1901 : if (adjacency[i])
3452 513 : pfree(adjacency[i]);
3453 831 : pfree(adjacency);
3454 831 : pfree(adjacency_buf);
3455 831 : pfree(orig_sets);
3456 2732 : for (i = 1; i <= num_sets; ++i)
3457 1901 : bms_free(set_masks[i]);
3458 831 : pfree(set_masks);
3459 :
3460 831 : return result;
3461 : }
3462 :
3463 : /*
3464 : * Reorder the elements of a list of grouping sets such that they have correct
3465 : * prefix relationships. Also inserts the GroupingSetData annotations.
3466 : *
3467 : * The input must be ordered with smallest sets first; the result is returned
3468 : * with largest sets first. Note that the result shares no list substructure
3469 : * with the input, so it's safe for the caller to modify it later.
3470 : *
3471 : * If we're passed in a sortclause, we follow its order of columns to the
3472 : * extent possible, to minimize the chance that we add unnecessary sorts.
3473 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3474 : * gets implemented in one pass.)
3475 : */
3476 : static List *
3477 1475 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3478 : {
3479 : ListCell *lc;
3480 1475 : List *previous = NIL;
3481 1475 : List *result = NIL;
3482 :
3483 4095 : foreach(lc, groupingSets)
3484 : {
3485 2620 : List *candidate = (List *) lfirst(lc);
3486 2620 : List *new_elems = list_difference_int(candidate, previous);
3487 2620 : GroupingSetData *gs = makeNode(GroupingSetData);
3488 :
3489 2764 : while (list_length(sortclause) > list_length(previous) &&
3490 : new_elems != NIL)
3491 : {
3492 244 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3493 244 : int ref = sc->tleSortGroupRef;
3494 :
3495 244 : if (list_member_int(new_elems, ref))
3496 : {
3497 144 : previous = lappend_int(previous, ref);
3498 144 : new_elems = list_delete_int(new_elems, ref);
3499 : }
3500 : else
3501 : {
3502 : /* diverged from the sortclause; give up on it */
3503 100 : sortclause = NIL;
3504 100 : break;
3505 : }
3506 : }
3507 :
3508 2620 : previous = list_concat(previous, new_elems);
3509 :
3510 2620 : gs->set = list_copy(previous);
3511 2620 : result = lcons(gs, result);
3512 : }
3513 :
3514 1475 : list_free(previous);
3515 :
3516 1475 : return result;
3517 : }
3518 :
3519 : /*
3520 : * has_volatile_pathkey
3521 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3522 : * containing a volatile function. Otherwise returns false.
3523 : */
3524 : static bool
3525 2209 : has_volatile_pathkey(List *keys)
3526 : {
3527 : ListCell *lc;
3528 :
3529 4540 : foreach(lc, keys)
3530 : {
3531 2344 : PathKey *pathkey = lfirst_node(PathKey, lc);
3532 :
3533 2344 : if (pathkey->pk_eclass->ec_has_volatile)
3534 13 : return true;
3535 : }
3536 :
3537 2196 : return false;
3538 : }
3539 :
3540 : /*
3541 : * adjust_group_pathkeys_for_groupagg
3542 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3543 : * pre-ordered input for ordered aggregates.
3544 : *
3545 : * We define "best" as the pathkeys that suit the largest number of
3546 : * aggregate functions. We find these by looking at the first ORDER BY /
3547 : * DISTINCT aggregate and take the pathkeys for that before searching for
3548 : * other aggregates that require the same or a more strict variation of the
3549 : * same pathkeys. We then repeat that process for any remaining aggregates
3550 : * with different pathkeys and if we find another set of pathkeys that suits a
3551 : * larger number of aggregates then we select those pathkeys instead.
3552 : *
3553 : * When the best pathkeys are found we also mark each Aggref that can use
3554 : * those pathkeys as aggpresorted = true.
3555 : *
3556 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3557 : * volatile functions, we never make use of these pathkeys. We want to ensure
3558 : * that sorts using volatile functions are done independently in each Aggref
3559 : * rather than once at the query level. If we were to allow this then Aggrefs
3560 : * with compatible sort orders would all transition their rows in the same
3561 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3562 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3563 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3564 : * left to perform their sorts individually. To avoid this inconsistent
3565 : * behavior which could make Aggref results depend on what other Aggrefs the
3566 : * query contains, we always force Aggrefs with volatile functions to perform
3567 : * their own sorts.
3568 : */
3569 : static void
3570 1873 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3571 : {
3572 1873 : List *grouppathkeys = root->group_pathkeys;
3573 : List *bestpathkeys;
3574 : Bitmapset *bestaggs;
3575 : Bitmapset *unprocessed_aggs;
3576 : ListCell *lc;
3577 : int i;
3578 :
3579 : /* Shouldn't be here if there are grouping sets */
3580 : Assert(root->parse->groupingSets == NIL);
3581 : /* Shouldn't be here unless there are some ordered aggregates */
3582 : Assert(root->numOrderedAggs > 0);
3583 :
3584 : /* Do nothing if disabled */
3585 1873 : if (!enable_presorted_aggregate)
3586 5 : return;
3587 :
3588 : /*
3589 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3590 : * the indexes of all AggInfos to be processed below.
3591 : */
3592 1868 : unprocessed_aggs = NULL;
3593 4302 : foreach(lc, root->agginfos)
3594 : {
3595 2434 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3596 2434 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3597 :
3598 2434 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3599 214 : continue;
3600 :
3601 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3602 2220 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3603 248 : continue;
3604 :
3605 : /* Additional safety checks are needed if there's a FILTER clause */
3606 1972 : if (aggref->aggfilter != NULL)
3607 : {
3608 : ListCell *lc2;
3609 43 : bool allow_presort = true;
3610 :
3611 : /*
3612 : * When the Aggref has a FILTER clause, it's possible that the
3613 : * filter removes rows that cannot be sorted because the
3614 : * expression to sort by results in an error during its
3615 : * evaluation. This is a problem for presorting as that happens
3616 : * before the FILTER, whereas without presorting, the Aggregate
3617 : * node will apply the FILTER *before* sorting. So that we never
3618 : * try to sort anything that might error, here we aim to skip over
3619 : * any Aggrefs with arguments with expressions which, when
3620 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3621 : * may be more cases that should be allowed, but more thought
3622 : * needs to be given. Err on the side of caution.
3623 : */
3624 83 : foreach(lc2, aggref->args)
3625 : {
3626 58 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3627 58 : Expr *expr = tle->expr;
3628 :
3629 68 : while (IsA(expr, RelabelType))
3630 10 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3631 :
3632 : /* Common case, Vars and Consts are ok */
3633 58 : if (IsA(expr, Var) || IsA(expr, Const))
3634 40 : continue;
3635 :
3636 : /* Unsupported. Don't try to presort for this Aggref */
3637 18 : allow_presort = false;
3638 18 : break;
3639 : }
3640 :
3641 : /* Skip unsupported Aggrefs */
3642 43 : if (!allow_presort)
3643 18 : continue;
3644 : }
3645 :
3646 1954 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3647 : foreach_current_index(lc));
3648 : }
3649 :
3650 : /*
3651 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3652 : * for the given set of aggregates.
3653 : *
3654 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3655 : * this during the first loop using the pathkeys for the very first
3656 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3657 : * a more strict set of compatible pathkeys. Once the outer loop is
3658 : * complete, we mark off all the aggregates with compatible pathkeys then
3659 : * remove those from the unprocessed_aggs and repeat the process to try to
3660 : * find another set of pathkeys that are suitable for a larger number of
3661 : * aggregates. The outer loop will stop when there are not enough
3662 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3663 : * to suit a larger number of aggregates.
3664 : */
3665 1868 : bestpathkeys = NIL;
3666 1868 : bestaggs = NULL;
3667 3687 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3668 : {
3669 1819 : Bitmapset *aggindexes = NULL;
3670 1819 : List *currpathkeys = NIL;
3671 :
3672 1819 : i = -1;
3673 4028 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3674 : {
3675 2209 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3676 2209 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3677 : List *sortlist;
3678 : List *pathkeys;
3679 :
3680 2209 : if (aggref->aggdistinct != NIL)
3681 597 : sortlist = aggref->aggdistinct;
3682 : else
3683 1612 : sortlist = aggref->aggorder;
3684 :
3685 2209 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3686 : aggref->args);
3687 :
3688 : /*
3689 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3690 : * or DISTINCT clause.
3691 : */
3692 2209 : if (has_volatile_pathkey(pathkeys))
3693 : {
3694 13 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3695 13 : continue;
3696 : }
3697 :
3698 : /*
3699 : * When not set yet, take the pathkeys from the first unprocessed
3700 : * aggregate.
3701 : */
3702 2196 : if (currpathkeys == NIL)
3703 : {
3704 1816 : currpathkeys = pathkeys;
3705 :
3706 : /* include the GROUP BY pathkeys, if they exist */
3707 1816 : if (grouppathkeys != NIL)
3708 226 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3709 : currpathkeys);
3710 :
3711 : /* record that we found pathkeys for this aggregate */
3712 1816 : aggindexes = bms_add_member(aggindexes, i);
3713 : }
3714 : else
3715 : {
3716 : /* now look for a stronger set of matching pathkeys */
3717 :
3718 : /* include the GROUP BY pathkeys, if they exist */
3719 380 : if (grouppathkeys != NIL)
3720 240 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3721 : pathkeys);
3722 :
3723 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3724 380 : switch (compare_pathkeys(currpathkeys, pathkeys))
3725 : {
3726 10 : case PATHKEYS_BETTER2:
3727 : /* 'pathkeys' are stronger, use these ones instead */
3728 10 : currpathkeys = pathkeys;
3729 : pg_fallthrough;
3730 :
3731 55 : case PATHKEYS_BETTER1:
3732 : /* 'pathkeys' are less strict */
3733 : pg_fallthrough;
3734 :
3735 : case PATHKEYS_EQUAL:
3736 : /* mark this aggregate as covered by 'currpathkeys' */
3737 55 : aggindexes = bms_add_member(aggindexes, i);
3738 55 : break;
3739 :
3740 325 : case PATHKEYS_DIFFERENT:
3741 325 : break;
3742 : }
3743 : }
3744 : }
3745 :
3746 : /* remove the aggregates that we've just processed */
3747 1819 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3748 :
3749 : /*
3750 : * If this pass included more aggregates than the previous best then
3751 : * use these ones as the best set.
3752 : */
3753 1819 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3754 : {
3755 1731 : bestaggs = aggindexes;
3756 1731 : bestpathkeys = currpathkeys;
3757 : }
3758 : }
3759 :
3760 : /*
3761 : * If we found any ordered aggregates, update root->group_pathkeys to add
3762 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3763 : * the original GROUP BY pathkeys already.
3764 : */
3765 1868 : if (bestpathkeys != NIL)
3766 1681 : root->group_pathkeys = bestpathkeys;
3767 :
3768 : /*
3769 : * Now that we've found the best set of aggregates we can set the
3770 : * presorted flag to indicate to the executor that it needn't bother
3771 : * performing a sort for these Aggrefs. We're able to do this now as
3772 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3773 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3774 : * of ordered aggregates.
3775 : */
3776 1868 : i = -1;
3777 3629 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3778 : {
3779 1761 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3780 :
3781 3537 : foreach(lc, agginfo->aggrefs)
3782 : {
3783 1776 : Aggref *aggref = lfirst_node(Aggref, lc);
3784 :
3785 1776 : aggref->aggpresorted = true;
3786 : }
3787 : }
3788 : }
3789 :
3790 : /*
3791 : * Compute query_pathkeys and other pathkeys during plan generation
3792 : */
3793 : static void
3794 393334 : standard_qp_callback(PlannerInfo *root, void *extra)
3795 : {
3796 393334 : Query *parse = root->parse;
3797 393334 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3798 393334 : List *tlist = root->processed_tlist;
3799 393334 : List *activeWindows = qp_extra->activeWindows;
3800 :
3801 : /*
3802 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3803 : * aggregate requirements.
3804 : */
3805 393334 : if (qp_extra->gset_data)
3806 : {
3807 : /*
3808 : * With grouping sets, just use the first RollupData's groupClause. We
3809 : * don't make any effort to optimize grouping clauses when there are
3810 : * grouping sets, nor can we combine aggregate ordering keys with
3811 : * grouping.
3812 : */
3813 901 : List *rollups = qp_extra->gset_data->rollups;
3814 901 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3815 :
3816 901 : if (grouping_is_sortable(groupClause))
3817 : {
3818 : bool sortable;
3819 :
3820 : /*
3821 : * The groupClause is logically below the grouping step. So if
3822 : * there is an RTE entry for the grouping step, we need to remove
3823 : * its RT index from the sort expressions before we make PathKeys
3824 : * for them.
3825 : */
3826 901 : root->group_pathkeys =
3827 901 : make_pathkeys_for_sortclauses_extended(root,
3828 : &groupClause,
3829 : tlist,
3830 : false,
3831 901 : parse->hasGroupRTE,
3832 : &sortable,
3833 : false);
3834 : Assert(sortable);
3835 901 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3836 : }
3837 : else
3838 : {
3839 0 : root->group_pathkeys = NIL;
3840 0 : root->num_groupby_pathkeys = 0;
3841 : }
3842 : }
3843 392433 : else if (parse->groupClause || root->numOrderedAggs > 0)
3844 5166 : {
3845 : /*
3846 : * With a plain GROUP BY list, we can remove any grouping items that
3847 : * are proven redundant by EquivalenceClass processing. For example,
3848 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3849 : * especially common cases, but they're nearly free to detect. Note
3850 : * that we remove redundant items from processed_groupClause but not
3851 : * the original parse->groupClause.
3852 : */
3853 : bool sortable;
3854 :
3855 : /*
3856 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3857 : * EquivalenceClass'es if it's not set yet.
3858 : */
3859 5166 : root->group_pathkeys =
3860 5166 : make_pathkeys_for_sortclauses_extended(root,
3861 : &root->processed_groupClause,
3862 : tlist,
3863 : true,
3864 : false,
3865 : &sortable,
3866 : true);
3867 5166 : if (!sortable)
3868 : {
3869 : /* Can't sort; no point in considering aggregate ordering either */
3870 0 : root->group_pathkeys = NIL;
3871 0 : root->num_groupby_pathkeys = 0;
3872 : }
3873 : else
3874 : {
3875 5166 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3876 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3877 5166 : if (root->numOrderedAggs > 0)
3878 1873 : adjust_group_pathkeys_for_groupagg(root);
3879 : }
3880 : }
3881 : else
3882 : {
3883 387267 : root->group_pathkeys = NIL;
3884 387267 : root->num_groupby_pathkeys = 0;
3885 : }
3886 :
3887 : /* We consider only the first (bottom) window in pathkeys logic */
3888 393334 : if (activeWindows != NIL)
3889 : {
3890 2217 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3891 :
3892 2217 : root->window_pathkeys = make_pathkeys_for_window(root,
3893 : wc,
3894 : tlist);
3895 : }
3896 : else
3897 391117 : root->window_pathkeys = NIL;
3898 :
3899 : /*
3900 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3901 : * redundant by EquivalenceClass processing. The non-redundant list is
3902 : * kept in root->processed_distinctClause, leaving the original
3903 : * parse->distinctClause alone.
3904 : */
3905 393334 : if (parse->distinctClause)
3906 : {
3907 : bool sortable;
3908 :
3909 : /* Make a copy since pathkey processing can modify the list */
3910 1974 : root->processed_distinctClause = list_copy(parse->distinctClause);
3911 1974 : root->distinct_pathkeys =
3912 1974 : make_pathkeys_for_sortclauses_extended(root,
3913 : &root->processed_distinctClause,
3914 : tlist,
3915 : true,
3916 : false,
3917 : &sortable,
3918 : false);
3919 1974 : if (!sortable)
3920 5 : root->distinct_pathkeys = NIL;
3921 : }
3922 : else
3923 391360 : root->distinct_pathkeys = NIL;
3924 :
3925 393334 : root->sort_pathkeys =
3926 393334 : make_pathkeys_for_sortclauses(root,
3927 : parse->sortClause,
3928 : tlist);
3929 :
3930 : /* setting setop_pathkeys might be useful to the union planner */
3931 393334 : if (qp_extra->setop != NULL)
3932 : {
3933 : List *groupClauses;
3934 : bool sortable;
3935 :
3936 10483 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3937 :
3938 10483 : root->setop_pathkeys =
3939 10483 : make_pathkeys_for_sortclauses_extended(root,
3940 : &groupClauses,
3941 : tlist,
3942 : false,
3943 : false,
3944 : &sortable,
3945 : false);
3946 10483 : if (!sortable)
3947 120 : root->setop_pathkeys = NIL;
3948 : }
3949 : else
3950 382851 : root->setop_pathkeys = NIL;
3951 :
3952 : /*
3953 : * Figure out whether we want a sorted result from query_planner.
3954 : *
3955 : * If we have a sortable GROUP BY clause, then we want a result sorted
3956 : * properly for grouping. Otherwise, if we have window functions to
3957 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3958 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3959 : * we try to produce output that's sufficiently well sorted for the
3960 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3961 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3962 : * for a set operation which can benefit from presorted results and have a
3963 : * sortable targetlist, we want to sort by the target list.
3964 : *
3965 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3966 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3967 : * that might just leave us failing to exploit an available sort order at
3968 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3969 : * much easier, since we know that the parser ensured that one is a
3970 : * superset of the other.
3971 : */
3972 393334 : if (root->group_pathkeys)
3973 5680 : root->query_pathkeys = root->group_pathkeys;
3974 387654 : else if (root->window_pathkeys)
3975 1772 : root->query_pathkeys = root->window_pathkeys;
3976 771764 : else if (list_length(root->distinct_pathkeys) >
3977 385882 : list_length(root->sort_pathkeys))
3978 1604 : root->query_pathkeys = root->distinct_pathkeys;
3979 384278 : else if (root->sort_pathkeys)
3980 56663 : root->query_pathkeys = root->sort_pathkeys;
3981 327615 : else if (root->setop_pathkeys != NIL)
3982 9495 : root->query_pathkeys = root->setop_pathkeys;
3983 : else
3984 318120 : root->query_pathkeys = NIL;
3985 393334 : }
3986 :
3987 : /*
3988 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3989 : *
3990 : * path_rows: number of output rows from scan/join step
3991 : * gd: grouping sets data including list of grouping sets and their clauses
3992 : * target_list: target list containing group clause references
3993 : *
3994 : * If doing grouping sets, we also annotate the gsets data with the estimates
3995 : * for each set and each individual rollup list, with a view to later
3996 : * determining whether some combination of them could be hashed instead.
3997 : */
3998 : static double
3999 41095 : get_number_of_groups(PlannerInfo *root,
4000 : double path_rows,
4001 : grouping_sets_data *gd,
4002 : List *target_list)
4003 : {
4004 41095 : Query *parse = root->parse;
4005 : double dNumGroups;
4006 :
4007 41095 : if (parse->groupClause)
4008 : {
4009 : List *groupExprs;
4010 :
4011 8728 : if (parse->groupingSets)
4012 : {
4013 : /* Add up the estimates for each grouping set */
4014 : ListCell *lc;
4015 :
4016 : Assert(gd); /* keep Coverity happy */
4017 :
4018 836 : dNumGroups = 0;
4019 :
4020 2246 : foreach(lc, gd->rollups)
4021 : {
4022 1410 : RollupData *rollup = lfirst_node(RollupData, lc);
4023 : ListCell *lc2;
4024 : ListCell *lc3;
4025 :
4026 1410 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
4027 : target_list);
4028 :
4029 1410 : rollup->numGroups = 0.0;
4030 :
4031 3915 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
4032 : {
4033 2505 : List *gset = (List *) lfirst(lc2);
4034 2505 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
4035 2505 : double numGroups = estimate_num_groups(root,
4036 : groupExprs,
4037 : path_rows,
4038 : &gset,
4039 : NULL);
4040 :
4041 2505 : gs->numGroups = numGroups;
4042 2505 : rollup->numGroups += numGroups;
4043 : }
4044 :
4045 1410 : dNumGroups += rollup->numGroups;
4046 : }
4047 :
4048 836 : if (gd->hash_sets_idx)
4049 : {
4050 : ListCell *lc2;
4051 :
4052 29 : gd->dNumHashGroups = 0;
4053 :
4054 29 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
4055 : target_list);
4056 :
4057 63 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
4058 : {
4059 34 : List *gset = (List *) lfirst(lc);
4060 34 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
4061 34 : double numGroups = estimate_num_groups(root,
4062 : groupExprs,
4063 : path_rows,
4064 : &gset,
4065 : NULL);
4066 :
4067 34 : gs->numGroups = numGroups;
4068 34 : gd->dNumHashGroups += numGroups;
4069 : }
4070 :
4071 29 : dNumGroups += gd->dNumHashGroups;
4072 : }
4073 : }
4074 : else
4075 : {
4076 : /* Plain GROUP BY -- estimate based on optimized groupClause */
4077 7892 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
4078 : target_list);
4079 :
4080 7892 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
4081 : NULL, NULL);
4082 : }
4083 : }
4084 32367 : else if (parse->groupingSets)
4085 : {
4086 : /* Empty grouping sets ... one result row for each one */
4087 50 : dNumGroups = list_length(parse->groupingSets);
4088 : }
4089 32317 : else if (parse->hasAggs || root->hasHavingQual)
4090 : {
4091 : /* Plain aggregation, one result row */
4092 32317 : dNumGroups = 1;
4093 : }
4094 : else
4095 : {
4096 : /* Not grouping */
4097 0 : dNumGroups = 1;
4098 : }
4099 :
4100 41095 : return dNumGroups;
4101 : }
4102 :
4103 : /*
4104 : * create_grouping_paths
4105 : *
4106 : * Build a new upperrel containing Paths for grouping and/or aggregation.
4107 : * Along the way, we also build an upperrel for Paths which are partially
4108 : * grouped and/or aggregated. A partially grouped and/or aggregated path
4109 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
4110 : * the only partially grouped paths we build are also partial paths; that
4111 : * is, they need a Gather and then a FinalizeAggregate.
4112 : *
4113 : * input_rel: contains the source-data Paths
4114 : * target: the pathtarget for the result Paths to compute
4115 : * gd: grouping sets data including list of grouping sets and their clauses
4116 : *
4117 : * Note: all Paths in input_rel are expected to return the target computed
4118 : * by make_group_input_target.
4119 : */
4120 : static RelOptInfo *
4121 34097 : create_grouping_paths(PlannerInfo *root,
4122 : RelOptInfo *input_rel,
4123 : PathTarget *target,
4124 : bool target_parallel_safe,
4125 : grouping_sets_data *gd)
4126 : {
4127 34097 : Query *parse = root->parse;
4128 : RelOptInfo *grouped_rel;
4129 : RelOptInfo *partially_grouped_rel;
4130 : AggClauseCosts agg_costs;
4131 :
4132 204582 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
4133 34097 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
4134 :
4135 : /*
4136 : * Create grouping relation to hold fully aggregated grouping and/or
4137 : * aggregation paths.
4138 : */
4139 34097 : grouped_rel = make_grouping_rel(root, input_rel, target,
4140 : target_parallel_safe, parse->havingQual);
4141 :
4142 : /*
4143 : * Create either paths for a degenerate grouping or paths for ordinary
4144 : * grouping, as appropriate.
4145 : */
4146 34097 : if (is_degenerate_grouping(root))
4147 35 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
4148 : else
4149 : {
4150 34062 : int flags = 0;
4151 : GroupPathExtraData extra;
4152 :
4153 : /*
4154 : * Determine whether it's possible to perform sort-based
4155 : * implementations of grouping. (Note that if processed_groupClause
4156 : * is empty, grouping_is_sortable() is trivially true, and all the
4157 : * pathkeys_contained_in() tests will succeed too, so that we'll
4158 : * consider every surviving input path.)
4159 : *
4160 : * If we have grouping sets, we might be able to sort some but not all
4161 : * of them; in this case, we need can_sort to be true as long as we
4162 : * must consider any sorted-input plan.
4163 : */
4164 34062 : if ((gd && gd->rollups != NIL)
4165 33181 : || grouping_is_sortable(root->processed_groupClause))
4166 34057 : flags |= GROUPING_CAN_USE_SORT;
4167 :
4168 : /*
4169 : * Determine whether we should consider hash-based implementations of
4170 : * grouping.
4171 : *
4172 : * Hashed aggregation only applies if we're grouping. If we have
4173 : * grouping sets, some groups might be hashable but others not; in
4174 : * this case we set can_hash true as long as there is nothing globally
4175 : * preventing us from hashing (and we should therefore consider plans
4176 : * with hashes).
4177 : *
4178 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
4179 : * BY aggregates. (Doing so would imply storing *all* the input
4180 : * values in the hash table, and/or running many sorts in parallel,
4181 : * either of which seems like a certain loser.) We similarly don't
4182 : * support ordered-set aggregates in hashed aggregation, but that case
4183 : * is also included in the numOrderedAggs count.
4184 : *
4185 : * Note: grouping_is_hashable() is much more expensive to check than
4186 : * the other gating conditions, so we want to do it last.
4187 : */
4188 34062 : if ((parse->groupClause != NIL &&
4189 7610 : root->numOrderedAggs == 0 &&
4190 3293 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
4191 4097 : flags |= GROUPING_CAN_USE_HASH;
4192 :
4193 : /*
4194 : * Determine whether partial aggregation is possible.
4195 : */
4196 34062 : if (can_partial_agg(root))
4197 29938 : flags |= GROUPING_CAN_PARTIAL_AGG;
4198 :
4199 34062 : extra.flags = flags;
4200 34062 : extra.target_parallel_safe = target_parallel_safe;
4201 34062 : extra.havingQual = parse->havingQual;
4202 34062 : extra.targetList = parse->targetList;
4203 34062 : extra.partial_costs_set = false;
4204 :
4205 : /*
4206 : * Determine whether partitionwise aggregation is in theory possible.
4207 : * It can be disabled by the user, and for now, we don't try to
4208 : * support grouping sets. create_ordinary_grouping_paths() will check
4209 : * additional conditions, such as whether input_rel is partitioned.
4210 : */
4211 34062 : if (enable_partitionwise_aggregate && !parse->groupingSets)
4212 580 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
4213 : else
4214 33482 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
4215 :
4216 34062 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
4217 : &agg_costs, gd, &extra,
4218 : &partially_grouped_rel);
4219 : }
4220 :
4221 34093 : set_cheapest(grouped_rel);
4222 34093 : return grouped_rel;
4223 : }
4224 :
4225 : /*
4226 : * make_grouping_rel
4227 : *
4228 : * Create a new grouping rel and set basic properties.
4229 : *
4230 : * input_rel represents the underlying scan/join relation.
4231 : * target is the output expected from the grouping relation.
4232 : */
4233 : static RelOptInfo *
4234 35892 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
4235 : PathTarget *target, bool target_parallel_safe,
4236 : Node *havingQual)
4237 : {
4238 : RelOptInfo *grouped_rel;
4239 :
4240 35892 : if (IS_OTHER_REL(input_rel))
4241 : {
4242 1795 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
4243 : input_rel->relids);
4244 1795 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
4245 : }
4246 : else
4247 : {
4248 : /*
4249 : * By tradition, the relids set for the main grouping relation is
4250 : * NULL. (This could be changed, but might require adjustments
4251 : * elsewhere.)
4252 : */
4253 34097 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
4254 : }
4255 :
4256 : /* Set target. */
4257 35892 : grouped_rel->reltarget = target;
4258 :
4259 : /*
4260 : * If the input relation is not parallel-safe, then the grouped relation
4261 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4262 : * target list and HAVING quals are parallel-safe.
4263 : */
4264 58792 : if (input_rel->consider_parallel && target_parallel_safe &&
4265 22900 : is_parallel_safe(root, havingQual))
4266 22875 : grouped_rel->consider_parallel = true;
4267 :
4268 : /* Assume that the same path generation strategies are allowed */
4269 35892 : grouped_rel->pgs_mask = input_rel->pgs_mask;
4270 :
4271 : /*
4272 : * If the input rel belongs to a single FDW, so does the grouped rel.
4273 : */
4274 35892 : grouped_rel->serverid = input_rel->serverid;
4275 35892 : grouped_rel->userid = input_rel->userid;
4276 35892 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
4277 35892 : grouped_rel->fdwroutine = input_rel->fdwroutine;
4278 :
4279 35892 : return grouped_rel;
4280 : }
4281 :
4282 : /*
4283 : * is_degenerate_grouping
4284 : *
4285 : * A degenerate grouping is one in which the query has a HAVING qual and/or
4286 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
4287 : * grouping sets are all empty).
4288 : */
4289 : static bool
4290 34097 : is_degenerate_grouping(PlannerInfo *root)
4291 : {
4292 34097 : Query *parse = root->parse;
4293 :
4294 32954 : return (root->hasHavingQual || parse->groupingSets) &&
4295 67051 : !parse->hasAggs && parse->groupClause == NIL;
4296 : }
4297 :
4298 : /*
4299 : * create_degenerate_grouping_paths
4300 : *
4301 : * When the grouping is degenerate (see is_degenerate_grouping), we are
4302 : * supposed to emit either zero or one row for each grouping set depending on
4303 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
4304 : * either HAVING or the targetlist, so we actually do not need the FROM table
4305 : * at all! We can just throw away the plan-so-far and generate a Result node.
4306 : * This is a sufficiently unusual corner case that it's not worth contorting
4307 : * the structure of this module to avoid having to generate the earlier paths
4308 : * in the first place.
4309 : */
4310 : static void
4311 35 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4312 : RelOptInfo *grouped_rel)
4313 : {
4314 35 : Query *parse = root->parse;
4315 : int nrows;
4316 : Path *path;
4317 :
4318 35 : nrows = list_length(parse->groupingSets);
4319 35 : if (nrows > 1)
4320 : {
4321 : /*
4322 : * Doesn't seem worthwhile writing code to cons up a generate_series
4323 : * or a values scan to emit multiple rows. Instead just make N clones
4324 : * and append them. (With a volatile HAVING clause, this means you
4325 : * might get between 0 and N output rows. Offhand I think that's
4326 : * desired.)
4327 : */
4328 10 : AppendPathInput append = {0};
4329 :
4330 30 : while (--nrows >= 0)
4331 : {
4332 : path = (Path *)
4333 20 : create_group_result_path(root, grouped_rel,
4334 20 : grouped_rel->reltarget,
4335 20 : (List *) parse->havingQual);
4336 20 : append.subpaths = lappend(append.subpaths, path);
4337 : }
4338 : path = (Path *)
4339 10 : create_append_path(root,
4340 : grouped_rel,
4341 : append,
4342 : NIL,
4343 : NULL,
4344 : 0,
4345 : false,
4346 : -1);
4347 : }
4348 : else
4349 : {
4350 : /* No grouping sets, or just one, so one output row */
4351 : path = (Path *)
4352 25 : create_group_result_path(root, grouped_rel,
4353 25 : grouped_rel->reltarget,
4354 25 : (List *) parse->havingQual);
4355 : }
4356 :
4357 35 : add_path(grouped_rel, path);
4358 35 : }
4359 :
4360 : /*
4361 : * create_ordinary_grouping_paths
4362 : *
4363 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4364 : *
4365 : * We need to consider sorted and hashed aggregation in the same function,
4366 : * because otherwise (1) it would be harder to throw an appropriate error
4367 : * message if neither way works, and (2) we should not allow hashtable size
4368 : * considerations to dissuade us from using hashing if sorting is not possible.
4369 : *
4370 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4371 : * function creates, or to NULL if it doesn't create one.
4372 : */
4373 : static void
4374 35857 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4375 : RelOptInfo *grouped_rel,
4376 : const AggClauseCosts *agg_costs,
4377 : grouping_sets_data *gd,
4378 : GroupPathExtraData *extra,
4379 : RelOptInfo **partially_grouped_rel_p)
4380 : {
4381 35857 : RelOptInfo *partially_grouped_rel = NULL;
4382 35857 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4383 :
4384 : /*
4385 : * If this is the topmost grouping relation or if the parent relation is
4386 : * doing some form of partitionwise aggregation, then we may be able to do
4387 : * it at this level also. However, if the input relation is not
4388 : * partitioned, partitionwise aggregate is impossible.
4389 : */
4390 35857 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4391 2375 : IS_PARTITIONED_REL(input_rel))
4392 : {
4393 : /*
4394 : * If this is the topmost relation or if the parent relation is doing
4395 : * full partitionwise aggregation, then we can do full partitionwise
4396 : * aggregation provided that the GROUP BY clause contains all of the
4397 : * partitioning columns at this level and the collation used by GROUP
4398 : * BY matches the partitioning collation. Otherwise, we can do at
4399 : * most partial partitionwise aggregation. But if partial aggregation
4400 : * is not supported in general then we can't use it for partitionwise
4401 : * aggregation either.
4402 : *
4403 : * Check parse->groupClause not processed_groupClause, because it's
4404 : * okay if some of the partitioning columns were proved redundant.
4405 : */
4406 1360 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4407 640 : group_by_has_partkey(input_rel, extra->targetList,
4408 640 : root->parse->groupClause))
4409 404 : patype = PARTITIONWISE_AGGREGATE_FULL;
4410 316 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4411 281 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4412 : else
4413 35 : patype = PARTITIONWISE_AGGREGATE_NONE;
4414 : }
4415 :
4416 : /*
4417 : * Before generating paths for grouped_rel, we first generate any possible
4418 : * partially grouped paths; that way, later code can easily consider both
4419 : * parallel and non-parallel approaches to grouping.
4420 : */
4421 35857 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4422 : {
4423 : bool force_rel_creation;
4424 :
4425 : /*
4426 : * If we're doing partitionwise aggregation at this level, force
4427 : * creation of a partially_grouped_rel so we can add partitionwise
4428 : * paths to it.
4429 : */
4430 31673 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4431 :
4432 : partially_grouped_rel =
4433 31673 : create_partial_grouping_paths(root,
4434 : grouped_rel,
4435 : input_rel,
4436 : gd,
4437 : extra,
4438 : force_rel_creation);
4439 : }
4440 :
4441 : /* Set out parameter. */
4442 35857 : *partially_grouped_rel_p = partially_grouped_rel;
4443 :
4444 : /* Apply partitionwise aggregation technique, if possible. */
4445 35857 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4446 685 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4447 : partially_grouped_rel, agg_costs,
4448 : gd, patype, extra);
4449 :
4450 : /* If we are doing partial aggregation only, return. */
4451 35857 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4452 : {
4453 : Assert(partially_grouped_rel);
4454 :
4455 713 : if (partially_grouped_rel->pathlist)
4456 713 : set_cheapest(partially_grouped_rel);
4457 :
4458 713 : return;
4459 : }
4460 :
4461 : /* Gather any partially grouped partial paths. */
4462 35144 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4463 2325 : gather_grouping_paths(root, partially_grouped_rel);
4464 :
4465 : /* Now choose the best path(s) for partially_grouped_rel. */
4466 35144 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4467 2493 : set_cheapest(partially_grouped_rel);
4468 :
4469 : /* Build final grouping paths */
4470 35144 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4471 : partially_grouped_rel, agg_costs, gd,
4472 : extra);
4473 :
4474 : /* Give a helpful error if we failed to find any implementation */
4475 35144 : if (grouped_rel->pathlist == NIL)
4476 4 : ereport(ERROR,
4477 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4478 : errmsg("could not implement GROUP BY"),
4479 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4480 :
4481 : /*
4482 : * If there is an FDW that's responsible for all baserels of the query,
4483 : * let it consider adding ForeignPaths.
4484 : */
4485 35140 : if (grouped_rel->fdwroutine &&
4486 170 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4487 169 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4488 : input_rel, grouped_rel,
4489 : extra);
4490 :
4491 : /* Let extensions possibly add some more paths */
4492 35140 : if (create_upper_paths_hook)
4493 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4494 : input_rel, grouped_rel,
4495 : extra);
4496 : }
4497 :
4498 : /*
4499 : * For a given input path, consider the possible ways of doing grouping sets on
4500 : * it, by combinations of hashing and sorting. This can be called multiple
4501 : * times, so it's important that it not scribble on input. No result is
4502 : * returned, but any generated paths are added to grouped_rel.
4503 : */
4504 : static void
4505 1782 : consider_groupingsets_paths(PlannerInfo *root,
4506 : RelOptInfo *grouped_rel,
4507 : Path *path,
4508 : bool is_sorted,
4509 : bool can_hash,
4510 : grouping_sets_data *gd,
4511 : const AggClauseCosts *agg_costs,
4512 : double dNumGroups)
4513 : {
4514 1782 : Query *parse = root->parse;
4515 1782 : Size hash_mem_limit = get_hash_memory_limit();
4516 :
4517 : /*
4518 : * If we're not being offered sorted input, then only consider plans that
4519 : * can be done entirely by hashing.
4520 : *
4521 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4522 : * the input is actually sorted despite not being advertised as such, we
4523 : * prefer to make use of that in order to use less memory.
4524 : *
4525 : * If none of the grouping sets are sortable, then ignore the hash_mem
4526 : * limit and generate a path anyway, since otherwise we'll just fail.
4527 : */
4528 1782 : if (!is_sorted)
4529 : {
4530 807 : List *new_rollups = NIL;
4531 807 : RollupData *unhashed_rollup = NULL;
4532 : List *sets_data;
4533 807 : List *empty_sets_data = NIL;
4534 807 : List *empty_sets = NIL;
4535 : ListCell *lc;
4536 807 : ListCell *l_start = list_head(gd->rollups);
4537 807 : AggStrategy strat = AGG_HASHED;
4538 : double hashsize;
4539 807 : double exclude_groups = 0.0;
4540 :
4541 : Assert(can_hash);
4542 :
4543 : /*
4544 : * If the input is coincidentally sorted usefully (which can happen
4545 : * even if is_sorted is false, since that only means that our caller
4546 : * has set up the sorting for us), then save some hashtable space by
4547 : * making use of that. But we need to watch out for degenerate cases:
4548 : *
4549 : * 1) If there are any empty grouping sets, then group_pathkeys might
4550 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4551 : * there will be a rollup containing only empty groups, and the
4552 : * pathkeys_contained_in test is vacuously true; this is ok.
4553 : *
4554 : * XXX: the above relies on the fact that group_pathkeys is generated
4555 : * from the first rollup. If we add the ability to consider multiple
4556 : * sort orders for grouping input, this assumption might fail.
4557 : *
4558 : * 2) If there are no empty sets and only unsortable sets, then the
4559 : * rollups list will be empty (and thus l_start == NULL), and
4560 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4561 : * pathkeys_contained_in test doesn't cause us to crash.
4562 : */
4563 1609 : if (l_start != NULL &&
4564 802 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4565 : {
4566 40 : unhashed_rollup = lfirst_node(RollupData, l_start);
4567 40 : exclude_groups = unhashed_rollup->numGroups;
4568 40 : l_start = lnext(gd->rollups, l_start);
4569 : }
4570 :
4571 807 : hashsize = estimate_hashagg_tablesize(root,
4572 : path,
4573 : agg_costs,
4574 : dNumGroups - exclude_groups);
4575 :
4576 : /*
4577 : * gd->rollups is empty if we have only unsortable columns to work
4578 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4579 : * sorted-input case to generate usable mixed paths.
4580 : */
4581 807 : if (hashsize > hash_mem_limit && gd->rollups)
4582 15 : return; /* nope, won't fit */
4583 :
4584 : /*
4585 : * We need to burst the existing rollups list into individual grouping
4586 : * sets and recompute a groupClause for each set.
4587 : */
4588 792 : sets_data = list_copy(gd->unsortable_sets);
4589 :
4590 2013 : for_each_cell(lc, gd->rollups, l_start)
4591 : {
4592 1241 : RollupData *rollup = lfirst_node(RollupData, lc);
4593 :
4594 : /*
4595 : * If we find an unhashable rollup that's not been skipped by the
4596 : * "actually sorted" check above, we can't cope; we'd need sorted
4597 : * input (with a different sort order) but we can't get that here.
4598 : * So bail out; we'll get a valid path from the is_sorted case
4599 : * instead.
4600 : *
4601 : * The mere presence of empty grouping sets doesn't make a rollup
4602 : * unhashable (see preprocess_grouping_sets), we handle those
4603 : * specially below.
4604 : */
4605 1241 : if (!rollup->hashable)
4606 20 : return;
4607 :
4608 1221 : sets_data = list_concat(sets_data, rollup->gsets_data);
4609 : }
4610 3053 : foreach(lc, sets_data)
4611 : {
4612 2281 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4613 2281 : List *gset = gs->set;
4614 : RollupData *rollup;
4615 :
4616 2281 : if (gset == NIL)
4617 : {
4618 : /* Empty grouping sets can't be hashed. */
4619 445 : empty_sets_data = lappend(empty_sets_data, gs);
4620 445 : empty_sets = lappend(empty_sets, NIL);
4621 : }
4622 : else
4623 : {
4624 1836 : rollup = makeNode(RollupData);
4625 :
4626 1836 : rollup->groupClause = preprocess_groupclause(root, gset);
4627 1836 : rollup->gsets_data = list_make1(gs);
4628 1836 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4629 : rollup->gsets_data,
4630 : gd->tleref_to_colnum_map);
4631 1836 : rollup->numGroups = gs->numGroups;
4632 1836 : rollup->hashable = true;
4633 1836 : rollup->is_hashed = true;
4634 1836 : new_rollups = lappend(new_rollups, rollup);
4635 : }
4636 : }
4637 :
4638 : /*
4639 : * If we didn't find anything nonempty to hash, then bail. We'll
4640 : * generate a path from the is_sorted case.
4641 : */
4642 772 : if (new_rollups == NIL)
4643 0 : return;
4644 :
4645 : /*
4646 : * If there were empty grouping sets they should have been in the
4647 : * first rollup.
4648 : */
4649 : Assert(!unhashed_rollup || !empty_sets);
4650 :
4651 772 : if (unhashed_rollup)
4652 : {
4653 40 : new_rollups = lappend(new_rollups, unhashed_rollup);
4654 40 : strat = AGG_MIXED;
4655 : }
4656 732 : else if (empty_sets)
4657 : {
4658 405 : RollupData *rollup = makeNode(RollupData);
4659 :
4660 405 : rollup->groupClause = NIL;
4661 405 : rollup->gsets_data = empty_sets_data;
4662 405 : rollup->gsets = empty_sets;
4663 405 : rollup->numGroups = list_length(empty_sets);
4664 405 : rollup->hashable = false;
4665 405 : rollup->is_hashed = false;
4666 405 : new_rollups = lappend(new_rollups, rollup);
4667 405 : strat = AGG_MIXED;
4668 : }
4669 :
4670 772 : add_path(grouped_rel, (Path *)
4671 772 : create_groupingsets_path(root,
4672 : grouped_rel,
4673 : path,
4674 772 : (List *) parse->havingQual,
4675 : strat,
4676 : new_rollups,
4677 : agg_costs));
4678 772 : return;
4679 : }
4680 :
4681 : /*
4682 : * If we have sorted input but nothing we can do with it, bail.
4683 : */
4684 975 : if (gd->rollups == NIL)
4685 0 : return;
4686 :
4687 : /*
4688 : * Given sorted input, we try and make two paths: one sorted and one mixed
4689 : * sort/hash. (We need to try both because hashagg might be disabled, or
4690 : * some columns might not be sortable.)
4691 : *
4692 : * can_hash is passed in as false if some obstacle elsewhere (such as
4693 : * ordered aggs) means that we shouldn't consider hashing at all.
4694 : */
4695 975 : if (can_hash && gd->any_hashable)
4696 : {
4697 896 : List *rollups = NIL;
4698 896 : List *hash_sets = list_copy(gd->unsortable_sets);
4699 896 : double availspace = hash_mem_limit;
4700 : ListCell *lc;
4701 :
4702 : /*
4703 : * Account first for space needed for groups we can't sort at all.
4704 : */
4705 896 : availspace -= estimate_hashagg_tablesize(root,
4706 : path,
4707 : agg_costs,
4708 : gd->dNumHashGroups);
4709 :
4710 896 : if (availspace > 0 && list_length(gd->rollups) > 1)
4711 : {
4712 : double scale;
4713 501 : int num_rollups = list_length(gd->rollups);
4714 : int k_capacity;
4715 501 : int *k_weights = palloc(num_rollups * sizeof(int));
4716 501 : Bitmapset *hash_items = NULL;
4717 : int i;
4718 :
4719 : /*
4720 : * We treat this as a knapsack problem: the knapsack capacity
4721 : * represents hash_mem, the item weights are the estimated memory
4722 : * usage of the hashtables needed to implement a single rollup,
4723 : * and we really ought to use the cost saving as the item value;
4724 : * however, currently the costs assigned to sort nodes don't
4725 : * reflect the comparison costs well, and so we treat all items as
4726 : * of equal value (each rollup we hash instead saves us one sort).
4727 : *
4728 : * To use the discrete knapsack, we need to scale the values to a
4729 : * reasonably small bounded range. We choose to allow a 5% error
4730 : * margin; we have no more than 4096 rollups in the worst possible
4731 : * case, which with a 5% error margin will require a bit over 42MB
4732 : * of workspace. (Anyone wanting to plan queries that complex had
4733 : * better have the memory for it. In more reasonable cases, with
4734 : * no more than a couple of dozen rollups, the memory usage will
4735 : * be negligible.)
4736 : *
4737 : * k_capacity is naturally bounded, but we clamp the values for
4738 : * scale and weight (below) to avoid overflows or underflows (or
4739 : * uselessly trying to use a scale factor less than 1 byte).
4740 : */
4741 501 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4742 501 : k_capacity = (int) floor(availspace / scale);
4743 :
4744 : /*
4745 : * We leave the first rollup out of consideration since it's the
4746 : * one that matches the input sort order. We assign indexes "i"
4747 : * to only those entries considered for hashing; the second loop,
4748 : * below, must use the same condition.
4749 : */
4750 501 : i = 0;
4751 1222 : for_each_from(lc, gd->rollups, 1)
4752 : {
4753 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4754 :
4755 721 : if (rollup->hashable)
4756 : {
4757 721 : double sz = estimate_hashagg_tablesize(root,
4758 : path,
4759 : agg_costs,
4760 : rollup->numGroups);
4761 :
4762 : /*
4763 : * If sz is enormous, but hash_mem (and hence scale) is
4764 : * small, avoid integer overflow here.
4765 : */
4766 721 : k_weights[i] = (int) Min(floor(sz / scale),
4767 : k_capacity + 1.0);
4768 721 : ++i;
4769 : }
4770 : }
4771 :
4772 : /*
4773 : * Apply knapsack algorithm; compute the set of items which
4774 : * maximizes the value stored (in this case the number of sorts
4775 : * saved) while keeping the total size (approximately) within
4776 : * capacity.
4777 : */
4778 501 : if (i > 0)
4779 501 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4780 :
4781 501 : if (!bms_is_empty(hash_items))
4782 : {
4783 501 : rollups = list_make1(linitial(gd->rollups));
4784 :
4785 501 : i = 0;
4786 1222 : for_each_from(lc, gd->rollups, 1)
4787 : {
4788 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4789 :
4790 721 : if (rollup->hashable)
4791 : {
4792 721 : if (bms_is_member(i, hash_items))
4793 691 : hash_sets = list_concat(hash_sets,
4794 691 : rollup->gsets_data);
4795 : else
4796 30 : rollups = lappend(rollups, rollup);
4797 721 : ++i;
4798 : }
4799 : else
4800 0 : rollups = lappend(rollups, rollup);
4801 : }
4802 : }
4803 : }
4804 :
4805 896 : if (!rollups && hash_sets)
4806 20 : rollups = list_copy(gd->rollups);
4807 :
4808 1703 : foreach(lc, hash_sets)
4809 : {
4810 807 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4811 807 : RollupData *rollup = makeNode(RollupData);
4812 :
4813 : Assert(gs->set != NIL);
4814 :
4815 807 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4816 807 : rollup->gsets_data = list_make1(gs);
4817 807 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4818 : rollup->gsets_data,
4819 : gd->tleref_to_colnum_map);
4820 807 : rollup->numGroups = gs->numGroups;
4821 807 : rollup->hashable = true;
4822 807 : rollup->is_hashed = true;
4823 807 : rollups = lcons(rollup, rollups);
4824 : }
4825 :
4826 896 : if (rollups)
4827 : {
4828 521 : add_path(grouped_rel, (Path *)
4829 521 : create_groupingsets_path(root,
4830 : grouped_rel,
4831 : path,
4832 521 : (List *) parse->havingQual,
4833 : AGG_MIXED,
4834 : rollups,
4835 : agg_costs));
4836 : }
4837 : }
4838 :
4839 : /*
4840 : * Now try the simple sorted case.
4841 : */
4842 975 : if (!gd->unsortable_sets)
4843 951 : add_path(grouped_rel, (Path *)
4844 951 : create_groupingsets_path(root,
4845 : grouped_rel,
4846 : path,
4847 951 : (List *) parse->havingQual,
4848 : AGG_SORTED,
4849 : gd->rollups,
4850 : agg_costs));
4851 : }
4852 :
4853 : /*
4854 : * create_window_paths
4855 : *
4856 : * Build a new upperrel containing Paths for window-function evaluation.
4857 : *
4858 : * input_rel: contains the source-data Paths
4859 : * input_target: result of make_window_input_target
4860 : * output_target: what the topmost WindowAggPath should return
4861 : * wflists: result of find_window_functions
4862 : * activeWindows: result of select_active_windows
4863 : *
4864 : * Note: all Paths in input_rel are expected to return input_target.
4865 : */
4866 : static RelOptInfo *
4867 2217 : create_window_paths(PlannerInfo *root,
4868 : RelOptInfo *input_rel,
4869 : PathTarget *input_target,
4870 : PathTarget *output_target,
4871 : bool output_target_parallel_safe,
4872 : WindowFuncLists *wflists,
4873 : List *activeWindows)
4874 : {
4875 : RelOptInfo *window_rel;
4876 : ListCell *lc;
4877 :
4878 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4879 2217 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4880 :
4881 : /*
4882 : * If the input relation is not parallel-safe, then the window relation
4883 : * can't be parallel-safe, either. Otherwise, we need to examine the
4884 : * target list and active windows for non-parallel-safe constructs.
4885 : */
4886 2217 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4887 0 : is_parallel_safe(root, (Node *) activeWindows))
4888 0 : window_rel->consider_parallel = true;
4889 :
4890 : /*
4891 : * If the input rel belongs to a single FDW, so does the window rel.
4892 : */
4893 2217 : window_rel->serverid = input_rel->serverid;
4894 2217 : window_rel->userid = input_rel->userid;
4895 2217 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4896 2217 : window_rel->fdwroutine = input_rel->fdwroutine;
4897 :
4898 : /*
4899 : * Consider computing window functions starting from the existing
4900 : * cheapest-total path (which will likely require a sort) as well as any
4901 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4902 : */
4903 4696 : foreach(lc, input_rel->pathlist)
4904 : {
4905 2479 : Path *path = (Path *) lfirst(lc);
4906 : int presorted_keys;
4907 :
4908 2741 : if (path == input_rel->cheapest_total_path ||
4909 262 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4910 116 : &presorted_keys) ||
4911 116 : presorted_keys > 0)
4912 2384 : create_one_window_path(root,
4913 : window_rel,
4914 : path,
4915 : input_target,
4916 : output_target,
4917 : wflists,
4918 : activeWindows);
4919 : }
4920 :
4921 : /*
4922 : * If there is an FDW that's responsible for all baserels of the query,
4923 : * let it consider adding ForeignPaths.
4924 : */
4925 2217 : if (window_rel->fdwroutine &&
4926 6 : window_rel->fdwroutine->GetForeignUpperPaths)
4927 6 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4928 : input_rel, window_rel,
4929 : NULL);
4930 :
4931 : /* Let extensions possibly add some more paths */
4932 2217 : if (create_upper_paths_hook)
4933 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4934 : input_rel, window_rel, NULL);
4935 :
4936 : /* Now choose the best path(s) */
4937 2217 : set_cheapest(window_rel);
4938 :
4939 2217 : return window_rel;
4940 : }
4941 :
4942 : /*
4943 : * Stack window-function implementation steps atop the given Path, and
4944 : * add the result to window_rel.
4945 : *
4946 : * window_rel: upperrel to contain result
4947 : * path: input Path to use (must return input_target)
4948 : * input_target: result of make_window_input_target
4949 : * output_target: what the topmost WindowAggPath should return
4950 : * wflists: result of find_window_functions
4951 : * activeWindows: result of select_active_windows
4952 : */
4953 : static void
4954 2384 : create_one_window_path(PlannerInfo *root,
4955 : RelOptInfo *window_rel,
4956 : Path *path,
4957 : PathTarget *input_target,
4958 : PathTarget *output_target,
4959 : WindowFuncLists *wflists,
4960 : List *activeWindows)
4961 : {
4962 : PathTarget *window_target;
4963 : ListCell *l;
4964 2384 : List *topqual = NIL;
4965 :
4966 : /*
4967 : * Since each window clause could require a different sort order, we stack
4968 : * up a WindowAgg node for each clause, with sort steps between them as
4969 : * needed. (We assume that select_active_windows chose a good order for
4970 : * executing the clauses in.)
4971 : *
4972 : * input_target should contain all Vars and Aggs needed for the result.
4973 : * (In some cases we wouldn't need to propagate all of these all the way
4974 : * to the top, since they might only be needed as inputs to WindowFuncs.
4975 : * It's probably not worth trying to optimize that though.) It must also
4976 : * contain all window partitioning and sorting expressions, to ensure
4977 : * they're computed only once at the bottom of the stack (that's critical
4978 : * for volatile functions). As we climb up the stack, we'll add outputs
4979 : * for the WindowFuncs computed at each level.
4980 : */
4981 2384 : window_target = input_target;
4982 :
4983 4923 : foreach(l, activeWindows)
4984 : {
4985 2539 : WindowClause *wc = lfirst_node(WindowClause, l);
4986 : List *window_pathkeys;
4987 2539 : List *runcondition = NIL;
4988 : int presorted_keys;
4989 : bool is_sorted;
4990 : bool topwindow;
4991 : ListCell *lc2;
4992 :
4993 2539 : window_pathkeys = make_pathkeys_for_window(root,
4994 : wc,
4995 : root->processed_tlist);
4996 :
4997 2539 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4998 : path->pathkeys,
4999 : &presorted_keys);
5000 :
5001 : /* Sort if necessary */
5002 2539 : if (!is_sorted)
5003 : {
5004 : /*
5005 : * No presorted keys or incremental sort disabled, just perform a
5006 : * complete sort.
5007 : */
5008 1826 : if (presorted_keys == 0 || !enable_incremental_sort)
5009 1775 : path = (Path *) create_sort_path(root, window_rel,
5010 : path,
5011 : window_pathkeys,
5012 : -1.0);
5013 : else
5014 : {
5015 : /*
5016 : * Since we have presorted keys and incremental sort is
5017 : * enabled, just use incremental sort.
5018 : */
5019 51 : path = (Path *) create_incremental_sort_path(root,
5020 : window_rel,
5021 : path,
5022 : window_pathkeys,
5023 : presorted_keys,
5024 : -1.0);
5025 : }
5026 : }
5027 :
5028 2539 : if (lnext(activeWindows, l))
5029 : {
5030 : /*
5031 : * Add the current WindowFuncs to the output target for this
5032 : * intermediate WindowAggPath. We must copy window_target to
5033 : * avoid changing the previous path's target.
5034 : *
5035 : * Note: a WindowFunc adds nothing to the target's eval costs; but
5036 : * we do need to account for the increase in tlist width.
5037 : */
5038 155 : int64 tuple_width = window_target->width;
5039 :
5040 155 : window_target = copy_pathtarget(window_target);
5041 370 : foreach(lc2, wflists->windowFuncs[wc->winref])
5042 : {
5043 215 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5044 :
5045 215 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
5046 215 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
5047 : }
5048 155 : window_target->width = clamp_width_est(tuple_width);
5049 : }
5050 : else
5051 : {
5052 : /* Install the goal target in the topmost WindowAgg */
5053 2384 : window_target = output_target;
5054 : }
5055 :
5056 : /* mark the final item in the list as the top-level window */
5057 2539 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
5058 :
5059 : /*
5060 : * Collect the WindowFuncRunConditions from each WindowFunc and
5061 : * convert them into OpExprs
5062 : */
5063 5811 : foreach(lc2, wflists->windowFuncs[wc->winref])
5064 : {
5065 : ListCell *lc3;
5066 3272 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5067 :
5068 3422 : foreach(lc3, wfunc->runCondition)
5069 : {
5070 150 : WindowFuncRunCondition *wfuncrc =
5071 : lfirst_node(WindowFuncRunCondition, lc3);
5072 : Expr *opexpr;
5073 : Expr *leftop;
5074 : Expr *rightop;
5075 :
5076 150 : if (wfuncrc->wfunc_left)
5077 : {
5078 135 : leftop = (Expr *) copyObject(wfunc);
5079 135 : rightop = copyObject(wfuncrc->arg);
5080 : }
5081 : else
5082 : {
5083 15 : leftop = copyObject(wfuncrc->arg);
5084 15 : rightop = (Expr *) copyObject(wfunc);
5085 : }
5086 :
5087 150 : opexpr = make_opclause(wfuncrc->opno,
5088 : BOOLOID,
5089 : false,
5090 : leftop,
5091 : rightop,
5092 : InvalidOid,
5093 : wfuncrc->inputcollid);
5094 :
5095 150 : runcondition = lappend(runcondition, opexpr);
5096 :
5097 150 : if (!topwindow)
5098 20 : topqual = lappend(topqual, opexpr);
5099 : }
5100 : }
5101 :
5102 : path = (Path *)
5103 2539 : create_windowagg_path(root, window_rel, path, window_target,
5104 2539 : wflists->windowFuncs[wc->winref],
5105 : runcondition, wc,
5106 : topwindow ? topqual : NIL, topwindow);
5107 : }
5108 :
5109 2384 : add_path(window_rel, path);
5110 2384 : }
5111 :
5112 : /*
5113 : * create_distinct_paths
5114 : *
5115 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
5116 : *
5117 : * input_rel: contains the source-data Paths
5118 : * target: the pathtarget for the result Paths to compute
5119 : *
5120 : * Note: input paths should already compute the desired pathtarget, since
5121 : * Sort/Unique won't project anything.
5122 : */
5123 : static RelOptInfo *
5124 1974 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5125 : PathTarget *target)
5126 : {
5127 : RelOptInfo *distinct_rel;
5128 :
5129 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
5130 1974 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
5131 :
5132 : /*
5133 : * We don't compute anything at this level, so distinct_rel will be
5134 : * parallel-safe if the input rel is parallel-safe. In particular, if
5135 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
5136 : * output those expressions, and will not be parallel-safe unless those
5137 : * expressions are parallel-safe.
5138 : */
5139 1974 : distinct_rel->consider_parallel = input_rel->consider_parallel;
5140 :
5141 : /*
5142 : * If the input rel belongs to a single FDW, so does the distinct_rel.
5143 : */
5144 1974 : distinct_rel->serverid = input_rel->serverid;
5145 1974 : distinct_rel->userid = input_rel->userid;
5146 1974 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5147 1974 : distinct_rel->fdwroutine = input_rel->fdwroutine;
5148 :
5149 : /* build distinct paths based on input_rel's pathlist */
5150 1974 : create_final_distinct_paths(root, input_rel, distinct_rel);
5151 :
5152 : /* now build distinct paths based on input_rel's partial_pathlist */
5153 1974 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
5154 :
5155 : /* Give a helpful error if we failed to create any paths */
5156 1974 : if (distinct_rel->pathlist == NIL)
5157 0 : ereport(ERROR,
5158 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5159 : errmsg("could not implement DISTINCT"),
5160 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
5161 :
5162 : /*
5163 : * If there is an FDW that's responsible for all baserels of the query,
5164 : * let it consider adding ForeignPaths.
5165 : */
5166 1974 : if (distinct_rel->fdwroutine &&
5167 8 : distinct_rel->fdwroutine->GetForeignUpperPaths)
5168 8 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5169 : UPPERREL_DISTINCT,
5170 : input_rel,
5171 : distinct_rel,
5172 : NULL);
5173 :
5174 : /* Let extensions possibly add some more paths */
5175 1974 : if (create_upper_paths_hook)
5176 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
5177 : distinct_rel, NULL);
5178 :
5179 : /* Now choose the best path(s) */
5180 1974 : set_cheapest(distinct_rel);
5181 :
5182 1974 : return distinct_rel;
5183 : }
5184 :
5185 : /*
5186 : * create_partial_distinct_paths
5187 : *
5188 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
5189 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
5190 : * paths on top and add a final unique/aggregate path to remove any duplicate
5191 : * produced from combining rows from parallel workers.
5192 : */
5193 : static void
5194 1974 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5195 : RelOptInfo *final_distinct_rel,
5196 : PathTarget *target)
5197 : {
5198 : RelOptInfo *partial_distinct_rel;
5199 : Query *parse;
5200 : List *distinctExprs;
5201 : double numDistinctRows;
5202 : Path *cheapest_partial_path;
5203 : ListCell *lc;
5204 :
5205 : /* nothing to do when there are no partial paths in the input rel */
5206 1974 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
5207 1884 : return;
5208 :
5209 90 : parse = root->parse;
5210 :
5211 : /* can't do parallel DISTINCT ON */
5212 90 : if (parse->hasDistinctOn)
5213 0 : return;
5214 :
5215 90 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
5216 : NULL);
5217 90 : partial_distinct_rel->reltarget = target;
5218 90 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
5219 :
5220 : /*
5221 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
5222 : */
5223 90 : partial_distinct_rel->serverid = input_rel->serverid;
5224 90 : partial_distinct_rel->userid = input_rel->userid;
5225 90 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5226 90 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
5227 :
5228 90 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5229 :
5230 90 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5231 : parse->targetList);
5232 :
5233 : /* estimate how many distinct rows we'll get from each worker */
5234 90 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5235 : cheapest_partial_path->rows,
5236 : NULL, NULL);
5237 :
5238 : /*
5239 : * Try sorting the cheapest path and incrementally sorting any paths with
5240 : * presorted keys and put a unique paths atop of those. We'll also
5241 : * attempt to reorder the required pathkeys to match the input path's
5242 : * pathkeys as much as possible, in hopes of avoiding a possible need to
5243 : * re-sort.
5244 : */
5245 90 : if (grouping_is_sortable(root->processed_distinctClause))
5246 : {
5247 193 : foreach(lc, input_rel->partial_pathlist)
5248 : {
5249 103 : Path *input_path = (Path *) lfirst(lc);
5250 : Path *sorted_path;
5251 103 : List *useful_pathkeys_list = NIL;
5252 :
5253 : useful_pathkeys_list =
5254 103 : get_useful_pathkeys_for_distinct(root,
5255 : root->distinct_pathkeys,
5256 : input_path->pathkeys);
5257 : Assert(list_length(useful_pathkeys_list) > 0);
5258 :
5259 319 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5260 : {
5261 113 : sorted_path = make_ordered_path(root,
5262 : partial_distinct_rel,
5263 : input_path,
5264 : cheapest_partial_path,
5265 : useful_pathkeys,
5266 : -1.0);
5267 :
5268 113 : if (sorted_path == NULL)
5269 8 : continue;
5270 :
5271 : /*
5272 : * An empty distinct_pathkeys means all tuples have the same
5273 : * value for the DISTINCT clause. See
5274 : * create_final_distinct_paths()
5275 : */
5276 105 : if (root->distinct_pathkeys == NIL)
5277 : {
5278 : Node *limitCount;
5279 :
5280 5 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5281 : sizeof(int64),
5282 : Int64GetDatum(1), false,
5283 : true);
5284 :
5285 : /*
5286 : * Apply a LimitPath onto the partial path to restrict the
5287 : * tuples from each worker to 1.
5288 : * create_final_distinct_paths will need to apply an
5289 : * additional LimitPath to restrict this to a single row
5290 : * after the Gather node. If the query already has a
5291 : * LIMIT clause, then we could end up with three Limit
5292 : * nodes in the final plan. Consolidating the top two of
5293 : * these could be done, but does not seem worth troubling
5294 : * over.
5295 : */
5296 5 : add_partial_path(partial_distinct_rel, (Path *)
5297 5 : create_limit_path(root, partial_distinct_rel,
5298 : sorted_path,
5299 : NULL,
5300 : limitCount,
5301 : LIMIT_OPTION_COUNT,
5302 : 0, 1));
5303 : }
5304 : else
5305 : {
5306 100 : add_partial_path(partial_distinct_rel, (Path *)
5307 100 : create_unique_path(root, partial_distinct_rel,
5308 : sorted_path,
5309 100 : list_length(root->distinct_pathkeys),
5310 : numDistinctRows));
5311 : }
5312 : }
5313 : }
5314 : }
5315 :
5316 : /*
5317 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5318 : * we're not on the hook to ensure we do our best to create at least one
5319 : * path here, we treat enable_hashagg as a hard off-switch rather than the
5320 : * slightly softer variant in create_final_distinct_paths.
5321 : */
5322 90 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5323 : {
5324 65 : add_partial_path(partial_distinct_rel, (Path *)
5325 65 : create_agg_path(root,
5326 : partial_distinct_rel,
5327 : cheapest_partial_path,
5328 : cheapest_partial_path->pathtarget,
5329 : AGG_HASHED,
5330 : AGGSPLIT_SIMPLE,
5331 : root->processed_distinctClause,
5332 : NIL,
5333 : NULL,
5334 : numDistinctRows));
5335 : }
5336 :
5337 : /*
5338 : * If there is an FDW that's responsible for all baserels of the query,
5339 : * let it consider adding ForeignPaths.
5340 : */
5341 90 : if (partial_distinct_rel->fdwroutine &&
5342 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5343 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5344 : UPPERREL_PARTIAL_DISTINCT,
5345 : input_rel,
5346 : partial_distinct_rel,
5347 : NULL);
5348 :
5349 : /* Let extensions possibly add some more partial paths */
5350 90 : if (create_upper_paths_hook)
5351 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5352 : input_rel, partial_distinct_rel, NULL);
5353 :
5354 90 : if (partial_distinct_rel->partial_pathlist != NIL)
5355 : {
5356 90 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5357 90 : set_cheapest(partial_distinct_rel);
5358 :
5359 : /*
5360 : * Finally, create paths to distinctify the final result. This step
5361 : * is needed to remove any duplicates due to combining rows from
5362 : * parallel workers.
5363 : */
5364 90 : create_final_distinct_paths(root, partial_distinct_rel,
5365 : final_distinct_rel);
5366 : }
5367 : }
5368 :
5369 : /*
5370 : * create_final_distinct_paths
5371 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5372 : *
5373 : * input_rel: contains the source-data paths
5374 : * distinct_rel: destination relation for storing created paths
5375 : */
5376 : static RelOptInfo *
5377 2064 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5378 : RelOptInfo *distinct_rel)
5379 : {
5380 2064 : Query *parse = root->parse;
5381 2064 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5382 : double numDistinctRows;
5383 : bool allow_hash;
5384 :
5385 : /* Estimate number of distinct rows there will be */
5386 2064 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5387 2005 : root->hasHavingQual)
5388 : {
5389 : /*
5390 : * If there was grouping or aggregation, use the number of input rows
5391 : * as the estimated number of DISTINCT rows (ie, assume the input is
5392 : * already mostly unique).
5393 : */
5394 59 : numDistinctRows = cheapest_input_path->rows;
5395 : }
5396 : else
5397 : {
5398 : /*
5399 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5400 : */
5401 : List *distinctExprs;
5402 :
5403 2005 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5404 : parse->targetList);
5405 2005 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5406 : cheapest_input_path->rows,
5407 : NULL, NULL);
5408 : }
5409 :
5410 : /*
5411 : * Consider sort-based implementations of DISTINCT, if possible.
5412 : */
5413 2064 : if (grouping_is_sortable(root->processed_distinctClause))
5414 : {
5415 : /*
5416 : * Firstly, if we have any adequately-presorted paths, just stick a
5417 : * Unique node on those. We also, consider doing an explicit sort of
5418 : * the cheapest input path and Unique'ing that. If any paths have
5419 : * presorted keys then we'll create an incremental sort atop of those
5420 : * before adding a unique node on the top. We'll also attempt to
5421 : * reorder the required pathkeys to match the input path's pathkeys as
5422 : * much as possible, in hopes of avoiding a possible need to re-sort.
5423 : *
5424 : * When we have DISTINCT ON, we must sort by the more rigorous of
5425 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5426 : * Also, if we do have to do an explicit sort, we might as well use
5427 : * the more rigorous ordering to avoid a second sort later. (Note
5428 : * that the parser will have ensured that one clause is a prefix of
5429 : * the other.)
5430 : */
5431 : List *needed_pathkeys;
5432 : ListCell *lc;
5433 2059 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5434 :
5435 2246 : if (parse->hasDistinctOn &&
5436 187 : list_length(root->distinct_pathkeys) <
5437 187 : list_length(root->sort_pathkeys))
5438 52 : needed_pathkeys = root->sort_pathkeys;
5439 : else
5440 2007 : needed_pathkeys = root->distinct_pathkeys;
5441 :
5442 5497 : foreach(lc, input_rel->pathlist)
5443 : {
5444 3438 : Path *input_path = (Path *) lfirst(lc);
5445 : Path *sorted_path;
5446 3438 : List *useful_pathkeys_list = NIL;
5447 :
5448 : useful_pathkeys_list =
5449 3438 : get_useful_pathkeys_for_distinct(root,
5450 : needed_pathkeys,
5451 : input_path->pathkeys);
5452 : Assert(list_length(useful_pathkeys_list) > 0);
5453 :
5454 10814 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5455 : {
5456 3938 : sorted_path = make_ordered_path(root,
5457 : distinct_rel,
5458 : input_path,
5459 : cheapest_input_path,
5460 : useful_pathkeys,
5461 : limittuples);
5462 :
5463 3938 : if (sorted_path == NULL)
5464 531 : continue;
5465 :
5466 : /*
5467 : * distinct_pathkeys may have become empty if all of the
5468 : * pathkeys were determined to be redundant. If all of the
5469 : * pathkeys are redundant then each DISTINCT target must only
5470 : * allow a single value, therefore all resulting tuples must
5471 : * be identical (or at least indistinguishable by an equality
5472 : * check). We can uniquify these tuples simply by just taking
5473 : * the first tuple. All we do here is add a path to do "LIMIT
5474 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5475 : * still have a non-NIL sort_pathkeys list, so we must still
5476 : * only do this with paths which are correctly sorted by
5477 : * sort_pathkeys.
5478 : */
5479 3407 : if (root->distinct_pathkeys == NIL)
5480 : {
5481 : Node *limitCount;
5482 :
5483 96 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5484 : sizeof(int64),
5485 : Int64GetDatum(1), false,
5486 : true);
5487 :
5488 : /*
5489 : * If the query already has a LIMIT clause, then we could
5490 : * end up with a duplicate LimitPath in the final plan.
5491 : * That does not seem worth troubling over too much.
5492 : */
5493 96 : add_path(distinct_rel, (Path *)
5494 96 : create_limit_path(root, distinct_rel, sorted_path,
5495 : NULL, limitCount,
5496 : LIMIT_OPTION_COUNT, 0, 1));
5497 : }
5498 : else
5499 : {
5500 3311 : add_path(distinct_rel, (Path *)
5501 3311 : create_unique_path(root, distinct_rel,
5502 : sorted_path,
5503 3311 : list_length(root->distinct_pathkeys),
5504 : numDistinctRows));
5505 : }
5506 : }
5507 : }
5508 : }
5509 :
5510 : /*
5511 : * Consider hash-based implementations of DISTINCT, if possible.
5512 : *
5513 : * If we were not able to make any other types of path, we *must* hash or
5514 : * die trying. If we do have other choices, there are two things that
5515 : * should prevent selection of hashing: if the query uses DISTINCT ON
5516 : * (because it won't really have the expected behavior if we hash), or if
5517 : * enable_hashagg is off.
5518 : *
5519 : * Note: grouping_is_hashable() is much more expensive to check than the
5520 : * other gating conditions, so we want to do it last.
5521 : */
5522 2064 : if (distinct_rel->pathlist == NIL)
5523 5 : allow_hash = true; /* we have no alternatives */
5524 2059 : else if (parse->hasDistinctOn || !enable_hashagg)
5525 312 : allow_hash = false; /* policy-based decision not to hash */
5526 : else
5527 1747 : allow_hash = true; /* default */
5528 :
5529 2064 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5530 : {
5531 : /* Generate hashed aggregate path --- no sort needed */
5532 1752 : add_path(distinct_rel, (Path *)
5533 1752 : create_agg_path(root,
5534 : distinct_rel,
5535 : cheapest_input_path,
5536 : cheapest_input_path->pathtarget,
5537 : AGG_HASHED,
5538 : AGGSPLIT_SIMPLE,
5539 : root->processed_distinctClause,
5540 : NIL,
5541 : NULL,
5542 : numDistinctRows));
5543 : }
5544 :
5545 2064 : return distinct_rel;
5546 : }
5547 :
5548 : /*
5549 : * get_useful_pathkeys_for_distinct
5550 : * Get useful orderings of pathkeys for distinctClause by reordering
5551 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5552 : *
5553 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5554 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5555 : */
5556 : static List *
5557 3541 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5558 : List *path_pathkeys)
5559 : {
5560 3541 : List *useful_pathkeys_list = NIL;
5561 3541 : List *useful_pathkeys = NIL;
5562 :
5563 : /* always include the given 'needed_pathkeys' */
5564 3541 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5565 : needed_pathkeys);
5566 :
5567 3541 : if (!enable_distinct_reordering)
5568 0 : return useful_pathkeys_list;
5569 :
5570 : /*
5571 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5572 : * that match 'needed_pathkeys', but only up to the longest matching
5573 : * prefix.
5574 : *
5575 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5576 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5577 : * the desired behavior.
5578 : */
5579 8825 : foreach_node(PathKey, pathkey, path_pathkeys)
5580 : {
5581 : /*
5582 : * The PathKey nodes are canonical, so they can be checked for
5583 : * equality by simple pointer comparison.
5584 : */
5585 1765 : if (!list_member_ptr(needed_pathkeys, pathkey))
5586 7 : break;
5587 1758 : if (root->parse->hasDistinctOn &&
5588 140 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5589 15 : break;
5590 :
5591 1743 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5592 : }
5593 :
5594 : /* If no match at all, no point in reordering needed_pathkeys */
5595 3541 : if (useful_pathkeys == NIL)
5596 2026 : return useful_pathkeys_list;
5597 :
5598 : /*
5599 : * If not full match, the resulting pathkey list is not useful without
5600 : * incremental sort.
5601 : */
5602 1515 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5603 924 : !enable_incremental_sort)
5604 48 : return useful_pathkeys_list;
5605 :
5606 : /* Append the remaining PathKey nodes in needed_pathkeys */
5607 1467 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5608 : needed_pathkeys);
5609 :
5610 : /*
5611 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5612 : * just drop it.
5613 : */
5614 1467 : if (compare_pathkeys(needed_pathkeys,
5615 : useful_pathkeys) == PATHKEYS_EQUAL)
5616 957 : return useful_pathkeys_list;
5617 :
5618 510 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5619 : useful_pathkeys);
5620 :
5621 510 : return useful_pathkeys_list;
5622 : }
5623 :
5624 : /*
5625 : * create_ordered_paths
5626 : *
5627 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5628 : *
5629 : * All paths in the result must satisfy the ORDER BY ordering.
5630 : * The only new paths we need consider are an explicit full sort
5631 : * and incremental sort on the cheapest-total existing path.
5632 : *
5633 : * input_rel: contains the source-data Paths
5634 : * target: the output tlist the result Paths must emit
5635 : * limit_tuples: estimated bound on the number of output tuples,
5636 : * or -1 if no LIMIT or couldn't estimate
5637 : *
5638 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5639 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5640 : */
5641 : static RelOptInfo *
5642 62167 : create_ordered_paths(PlannerInfo *root,
5643 : RelOptInfo *input_rel,
5644 : PathTarget *target,
5645 : bool target_parallel_safe,
5646 : double limit_tuples)
5647 : {
5648 62167 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5649 : RelOptInfo *ordered_rel;
5650 : ListCell *lc;
5651 :
5652 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5653 62167 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5654 :
5655 : /*
5656 : * If the input relation is not parallel-safe, then the ordered relation
5657 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5658 : * target list is parallel-safe.
5659 : */
5660 62167 : if (input_rel->consider_parallel && target_parallel_safe)
5661 43517 : ordered_rel->consider_parallel = true;
5662 :
5663 : /* Assume that the same path generation strategies are allowed. */
5664 62167 : ordered_rel->pgs_mask = input_rel->pgs_mask;
5665 :
5666 : /*
5667 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5668 : */
5669 62167 : ordered_rel->serverid = input_rel->serverid;
5670 62167 : ordered_rel->userid = input_rel->userid;
5671 62167 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5672 62167 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5673 :
5674 158824 : foreach(lc, input_rel->pathlist)
5675 : {
5676 96657 : Path *input_path = (Path *) lfirst(lc);
5677 : Path *sorted_path;
5678 : bool is_sorted;
5679 : int presorted_keys;
5680 :
5681 96657 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5682 : input_path->pathkeys, &presorted_keys);
5683 :
5684 96657 : if (is_sorted)
5685 37225 : sorted_path = input_path;
5686 : else
5687 : {
5688 : /*
5689 : * Try at least sorting the cheapest path and also try
5690 : * incrementally sorting any path which is partially sorted
5691 : * already (no need to deal with paths which have presorted keys
5692 : * when incremental sort is disabled unless it's the cheapest
5693 : * input path).
5694 : */
5695 59432 : if (input_path != cheapest_input_path &&
5696 4842 : (presorted_keys == 0 || !enable_incremental_sort))
5697 1608 : continue;
5698 :
5699 : /*
5700 : * We've no need to consider both a sort and incremental sort.
5701 : * We'll just do a sort if there are no presorted keys and an
5702 : * incremental sort when there are presorted keys.
5703 : */
5704 57824 : if (presorted_keys == 0 || !enable_incremental_sort)
5705 54139 : sorted_path = (Path *) create_sort_path(root,
5706 : ordered_rel,
5707 : input_path,
5708 : root->sort_pathkeys,
5709 : limit_tuples);
5710 : else
5711 3685 : sorted_path = (Path *) create_incremental_sort_path(root,
5712 : ordered_rel,
5713 : input_path,
5714 : root->sort_pathkeys,
5715 : presorted_keys,
5716 : limit_tuples);
5717 : }
5718 :
5719 : /*
5720 : * If the pathtarget of the result path has different expressions from
5721 : * the target to be applied, a projection step is needed.
5722 : */
5723 95049 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5724 288 : sorted_path = apply_projection_to_path(root, ordered_rel,
5725 : sorted_path, target);
5726 :
5727 95049 : add_path(ordered_rel, sorted_path);
5728 : }
5729 :
5730 : /*
5731 : * generate_gather_paths() will have already generated a simple Gather
5732 : * path for the best parallel path, if any, and the loop above will have
5733 : * considered sorting it. Similarly, generate_gather_paths() will also
5734 : * have generated order-preserving Gather Merge plans which can be used
5735 : * without sorting if they happen to match the sort_pathkeys, and the loop
5736 : * above will have handled those as well. However, there's one more
5737 : * possibility: it may make sense to sort the cheapest partial path or
5738 : * incrementally sort any partial path that is partially sorted according
5739 : * to the required output order and then use Gather Merge.
5740 : */
5741 62167 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5742 43397 : input_rel->partial_pathlist != NIL)
5743 : {
5744 : Path *cheapest_partial_path;
5745 :
5746 2411 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5747 :
5748 5377 : foreach(lc, input_rel->partial_pathlist)
5749 : {
5750 2966 : Path *input_path = (Path *) lfirst(lc);
5751 : Path *sorted_path;
5752 : bool is_sorted;
5753 : int presorted_keys;
5754 : double total_groups;
5755 :
5756 2966 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5757 : input_path->pathkeys,
5758 : &presorted_keys);
5759 :
5760 2966 : if (is_sorted)
5761 476 : continue;
5762 :
5763 : /*
5764 : * Try at least sorting the cheapest path and also try
5765 : * incrementally sorting any path which is partially sorted
5766 : * already (no need to deal with paths which have presorted keys
5767 : * when incremental sort is disabled unless it's the cheapest
5768 : * partial path).
5769 : */
5770 2490 : if (input_path != cheapest_partial_path &&
5771 101 : (presorted_keys == 0 || !enable_incremental_sort))
5772 0 : continue;
5773 :
5774 : /*
5775 : * We've no need to consider both a sort and incremental sort.
5776 : * We'll just do a sort if there are no presorted keys and an
5777 : * incremental sort when there are presorted keys.
5778 : */
5779 2490 : if (presorted_keys == 0 || !enable_incremental_sort)
5780 2374 : sorted_path = (Path *) create_sort_path(root,
5781 : ordered_rel,
5782 : input_path,
5783 : root->sort_pathkeys,
5784 : limit_tuples);
5785 : else
5786 116 : sorted_path = (Path *) create_incremental_sort_path(root,
5787 : ordered_rel,
5788 : input_path,
5789 : root->sort_pathkeys,
5790 : presorted_keys,
5791 : limit_tuples);
5792 2490 : total_groups = compute_gather_rows(sorted_path);
5793 : sorted_path = (Path *)
5794 2490 : create_gather_merge_path(root, ordered_rel,
5795 : sorted_path,
5796 : sorted_path->pathtarget,
5797 : root->sort_pathkeys, NULL,
5798 : &total_groups);
5799 :
5800 : /*
5801 : * If the pathtarget of the result path has different expressions
5802 : * from the target to be applied, a projection step is needed.
5803 : */
5804 2490 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5805 5 : sorted_path = apply_projection_to_path(root, ordered_rel,
5806 : sorted_path, target);
5807 :
5808 2490 : add_path(ordered_rel, sorted_path);
5809 : }
5810 : }
5811 :
5812 : /*
5813 : * If there is an FDW that's responsible for all baserels of the query,
5814 : * let it consider adding ForeignPaths.
5815 : */
5816 62167 : if (ordered_rel->fdwroutine &&
5817 197 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5818 189 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5819 : input_rel, ordered_rel,
5820 : NULL);
5821 :
5822 : /* Let extensions possibly add some more paths */
5823 62167 : if (create_upper_paths_hook)
5824 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5825 : input_rel, ordered_rel, NULL);
5826 :
5827 : /*
5828 : * No need to bother with set_cheapest here; grouping_planner does not
5829 : * need us to do it.
5830 : */
5831 : Assert(ordered_rel->pathlist != NIL);
5832 :
5833 62167 : return ordered_rel;
5834 : }
5835 :
5836 :
5837 : /*
5838 : * make_group_input_target
5839 : * Generate appropriate PathTarget for initial input to grouping nodes.
5840 : *
5841 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5842 : * the query's final targetlist; for example, it certainly can't emit any
5843 : * aggregate function calls. This routine generates the correct target
5844 : * for the scan/join subplan.
5845 : *
5846 : * The query target list passed from the parser already contains entries
5847 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5848 : * for variables used only in HAVING clauses; so we need to add those
5849 : * variables to the subplan target list. Also, we flatten all expressions
5850 : * except GROUP BY items into their component variables; other expressions
5851 : * will be computed by the upper plan nodes rather than by the subplan.
5852 : * For example, given a query like
5853 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5854 : * we want to pass this targetlist to the subplan:
5855 : * a+b,c,d
5856 : * where the a+b target will be used by the Sort/Group steps, and the
5857 : * other targets will be used for computing the final results.
5858 : *
5859 : * 'final_target' is the query's final target list (in PathTarget form)
5860 : *
5861 : * The result is the PathTarget to be computed by the Paths returned from
5862 : * query_planner().
5863 : */
5864 : static PathTarget *
5865 34097 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5866 : {
5867 34097 : Query *parse = root->parse;
5868 : PathTarget *input_target;
5869 : List *non_group_cols;
5870 : List *non_group_vars;
5871 : int i;
5872 : ListCell *lc;
5873 :
5874 : /*
5875 : * We must build a target containing all grouping columns, plus any other
5876 : * Vars mentioned in the query's targetlist and HAVING qual.
5877 : */
5878 34097 : input_target = create_empty_pathtarget();
5879 34097 : non_group_cols = NIL;
5880 :
5881 34097 : i = 0;
5882 85780 : foreach(lc, final_target->exprs)
5883 : {
5884 51683 : Expr *expr = (Expr *) lfirst(lc);
5885 51683 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5886 :
5887 59858 : if (sgref && root->processed_groupClause &&
5888 8175 : get_sortgroupref_clause_noerr(sgref,
5889 : root->processed_groupClause) != NULL)
5890 : {
5891 : /*
5892 : * It's a grouping column, so add it to the input target as-is.
5893 : *
5894 : * Note that the target is logically below the grouping step. So
5895 : * with grouping sets we need to remove the RT index of the
5896 : * grouping step if there is any from the target expression.
5897 : */
5898 6659 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5899 : {
5900 : Assert(root->group_rtindex > 0);
5901 : expr = (Expr *)
5902 1803 : remove_nulling_relids((Node *) expr,
5903 1803 : bms_make_singleton(root->group_rtindex),
5904 : NULL);
5905 : }
5906 6659 : add_column_to_pathtarget(input_target, expr, sgref);
5907 : }
5908 : else
5909 : {
5910 : /*
5911 : * Non-grouping column, so just remember the expression for later
5912 : * call to pull_var_clause.
5913 : */
5914 45024 : non_group_cols = lappend(non_group_cols, expr);
5915 : }
5916 :
5917 51683 : i++;
5918 : }
5919 :
5920 : /*
5921 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5922 : */
5923 34097 : if (parse->havingQual)
5924 842 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5925 :
5926 : /*
5927 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5928 : * add them to the input target if not already present. (A Var used
5929 : * directly as a GROUP BY item will be present already.) Note this
5930 : * includes Vars used in resjunk items, so we are covering the needs of
5931 : * ORDER BY and window specifications. Vars used within Aggrefs and
5932 : * WindowFuncs will be pulled out here, too.
5933 : *
5934 : * Note that the target is logically below the grouping step. So with
5935 : * grouping sets we need to remove the RT index of the grouping step if
5936 : * there is any from the non-group Vars.
5937 : */
5938 34097 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5939 : PVC_RECURSE_AGGREGATES |
5940 : PVC_RECURSE_WINDOWFUNCS |
5941 : PVC_INCLUDE_PLACEHOLDERS);
5942 34097 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5943 : {
5944 : Assert(root->group_rtindex > 0);
5945 : non_group_vars = (List *)
5946 836 : remove_nulling_relids((Node *) non_group_vars,
5947 836 : bms_make_singleton(root->group_rtindex),
5948 : NULL);
5949 : }
5950 34097 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5951 :
5952 : /* clean up cruft */
5953 34097 : list_free(non_group_vars);
5954 34097 : list_free(non_group_cols);
5955 :
5956 : /* XXX this causes some redundant cost calculation ... */
5957 34097 : return set_pathtarget_cost_width(root, input_target);
5958 : }
5959 :
5960 : /*
5961 : * make_partial_grouping_target
5962 : * Generate appropriate PathTarget for output of partial aggregate
5963 : * (or partial grouping, if there are no aggregates) nodes.
5964 : *
5965 : * A partial aggregation node needs to emit all the same aggregates that
5966 : * a regular aggregation node would, plus any aggregates used in HAVING;
5967 : * except that the Aggref nodes should be marked as partial aggregates.
5968 : *
5969 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5970 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5971 : * these would be Vars that are grouped by or used in grouping expressions.)
5972 : *
5973 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5974 : * havingQual represents the HAVING clause.
5975 : */
5976 : static PathTarget *
5977 3206 : make_partial_grouping_target(PlannerInfo *root,
5978 : PathTarget *grouping_target,
5979 : Node *havingQual)
5980 : {
5981 : PathTarget *partial_target;
5982 : List *non_group_cols;
5983 : List *non_group_exprs;
5984 : int i;
5985 : ListCell *lc;
5986 :
5987 3206 : partial_target = create_empty_pathtarget();
5988 3206 : non_group_cols = NIL;
5989 :
5990 3206 : i = 0;
5991 10563 : foreach(lc, grouping_target->exprs)
5992 : {
5993 7357 : Expr *expr = (Expr *) lfirst(lc);
5994 7357 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5995 :
5996 11289 : if (sgref && root->processed_groupClause &&
5997 3932 : get_sortgroupref_clause_noerr(sgref,
5998 : root->processed_groupClause) != NULL)
5999 : {
6000 : /*
6001 : * It's a grouping column, so add it to the partial_target as-is.
6002 : * (This allows the upper agg step to repeat the grouping calcs.)
6003 : */
6004 2330 : add_column_to_pathtarget(partial_target, expr, sgref);
6005 : }
6006 : else
6007 : {
6008 : /*
6009 : * Non-grouping column, so just remember the expression for later
6010 : * call to pull_var_clause.
6011 : */
6012 5027 : non_group_cols = lappend(non_group_cols, expr);
6013 : }
6014 :
6015 7357 : i++;
6016 : }
6017 :
6018 : /*
6019 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
6020 : */
6021 3206 : if (havingQual)
6022 729 : non_group_cols = lappend(non_group_cols, havingQual);
6023 :
6024 : /*
6025 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
6026 : * non-group cols (plus HAVING), and add them to the partial_target if not
6027 : * already present. (An expression used directly as a GROUP BY item will
6028 : * be present already.) Note this includes Vars used in resjunk items, so
6029 : * we are covering the needs of ORDER BY and window specifications.
6030 : */
6031 3206 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
6032 : PVC_INCLUDE_AGGREGATES |
6033 : PVC_RECURSE_WINDOWFUNCS |
6034 : PVC_INCLUDE_PLACEHOLDERS);
6035 :
6036 3206 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
6037 :
6038 : /*
6039 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
6040 : * are at the top level of the target list, so we can just scan the list
6041 : * rather than recursing through the expression trees.
6042 : */
6043 11077 : foreach(lc, partial_target->exprs)
6044 : {
6045 7871 : Aggref *aggref = (Aggref *) lfirst(lc);
6046 :
6047 7871 : if (IsA(aggref, Aggref))
6048 : {
6049 : Aggref *newaggref;
6050 :
6051 : /*
6052 : * We shouldn't need to copy the substructure of the Aggref node,
6053 : * but flat-copy the node itself to avoid damaging other trees.
6054 : */
6055 5516 : newaggref = makeNode(Aggref);
6056 5516 : memcpy(newaggref, aggref, sizeof(Aggref));
6057 :
6058 : /* For now, assume serialization is required */
6059 5516 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
6060 :
6061 5516 : lfirst(lc) = newaggref;
6062 : }
6063 : }
6064 :
6065 : /* clean up cruft */
6066 3206 : list_free(non_group_exprs);
6067 3206 : list_free(non_group_cols);
6068 :
6069 : /* XXX this causes some redundant cost calculation ... */
6070 3206 : return set_pathtarget_cost_width(root, partial_target);
6071 : }
6072 :
6073 : /*
6074 : * mark_partial_aggref
6075 : * Adjust an Aggref to make it represent a partial-aggregation step.
6076 : *
6077 : * The Aggref node is modified in-place; caller must do any copying required.
6078 : */
6079 : void
6080 15330 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
6081 : {
6082 : /* aggtranstype should be computed by this point */
6083 : Assert(OidIsValid(agg->aggtranstype));
6084 : /* ... but aggsplit should still be as the parser left it */
6085 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
6086 :
6087 : /* Mark the Aggref with the intended partial-aggregation mode */
6088 15330 : agg->aggsplit = aggsplit;
6089 :
6090 : /*
6091 : * Adjust result type if needed. Normally, a partial aggregate returns
6092 : * the aggregate's transition type; but if that's INTERNAL and we're
6093 : * serializing, it returns BYTEA instead.
6094 : */
6095 15330 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
6096 : {
6097 13425 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
6098 261 : agg->aggtype = BYTEAOID;
6099 : else
6100 13164 : agg->aggtype = agg->aggtranstype;
6101 : }
6102 15330 : }
6103 :
6104 : /*
6105 : * postprocess_setop_tlist
6106 : * Fix up targetlist returned by plan_set_operations().
6107 : *
6108 : * We need to transpose sort key info from the orig_tlist into new_tlist.
6109 : * NOTE: this would not be good enough if we supported resjunk sort keys
6110 : * for results of set operations --- then, we'd need to project a whole
6111 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
6112 : * find any resjunk columns in orig_tlist.
6113 : */
6114 : static List *
6115 4952 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
6116 : {
6117 : ListCell *l;
6118 4952 : ListCell *orig_tlist_item = list_head(orig_tlist);
6119 :
6120 19219 : foreach(l, new_tlist)
6121 : {
6122 14267 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
6123 : TargetEntry *orig_tle;
6124 :
6125 : /* ignore resjunk columns in setop result */
6126 14267 : if (new_tle->resjunk)
6127 0 : continue;
6128 :
6129 : Assert(orig_tlist_item != NULL);
6130 14267 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
6131 14267 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
6132 14267 : if (orig_tle->resjunk) /* should not happen */
6133 0 : elog(ERROR, "resjunk output columns are not implemented");
6134 : Assert(new_tle->resno == orig_tle->resno);
6135 14267 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
6136 : }
6137 4952 : if (orig_tlist_item != NULL)
6138 0 : elog(ERROR, "resjunk output columns are not implemented");
6139 4952 : return new_tlist;
6140 : }
6141 :
6142 : /*
6143 : * optimize_window_clauses
6144 : * Call each WindowFunc's prosupport function to see if we're able to
6145 : * make any adjustments to any of the WindowClause's so that the executor
6146 : * can execute the window functions in a more optimal way.
6147 : *
6148 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
6149 : * may allow more things to be done here in the future.
6150 : */
6151 : static void
6152 2217 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
6153 : {
6154 2217 : List *windowClause = root->parse->windowClause;
6155 : ListCell *lc;
6156 :
6157 4639 : foreach(lc, windowClause)
6158 : {
6159 2422 : WindowClause *wc = lfirst_node(WindowClause, lc);
6160 : ListCell *lc2;
6161 2422 : int optimizedFrameOptions = 0;
6162 :
6163 : Assert(wc->winref <= wflists->maxWinRef);
6164 :
6165 : /* skip any WindowClauses that have no WindowFuncs */
6166 2422 : if (wflists->windowFuncs[wc->winref] == NIL)
6167 20 : continue;
6168 :
6169 2945 : foreach(lc2, wflists->windowFuncs[wc->winref])
6170 : {
6171 : SupportRequestOptimizeWindowClause req;
6172 : SupportRequestOptimizeWindowClause *res;
6173 2437 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
6174 : Oid prosupport;
6175 :
6176 2437 : prosupport = get_func_support(wfunc->winfnoid);
6177 :
6178 : /* Check if there's a support function for 'wfunc' */
6179 2437 : if (!OidIsValid(prosupport))
6180 1894 : break; /* can't optimize this WindowClause */
6181 :
6182 730 : req.type = T_SupportRequestOptimizeWindowClause;
6183 730 : req.window_clause = wc;
6184 730 : req.window_func = wfunc;
6185 730 : req.frameOptions = wc->frameOptions;
6186 :
6187 : /* call the support function */
6188 : res = (SupportRequestOptimizeWindowClause *)
6189 730 : DatumGetPointer(OidFunctionCall1(prosupport,
6190 : PointerGetDatum(&req)));
6191 :
6192 : /*
6193 : * Skip to next WindowClause if the support function does not
6194 : * support this request type.
6195 : */
6196 730 : if (res == NULL)
6197 187 : break;
6198 :
6199 : /*
6200 : * Save these frameOptions for the first WindowFunc for this
6201 : * WindowClause.
6202 : */
6203 543 : if (foreach_current_index(lc2) == 0)
6204 523 : optimizedFrameOptions = res->frameOptions;
6205 :
6206 : /*
6207 : * On subsequent WindowFuncs, if the frameOptions are not the same
6208 : * then we're unable to optimize the frameOptions for this
6209 : * WindowClause.
6210 : */
6211 20 : else if (optimizedFrameOptions != res->frameOptions)
6212 0 : break; /* skip to the next WindowClause, if any */
6213 : }
6214 :
6215 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
6216 2402 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
6217 : {
6218 : ListCell *lc3;
6219 :
6220 : /* apply the new frame options */
6221 508 : wc->frameOptions = optimizedFrameOptions;
6222 :
6223 : /*
6224 : * We now check to see if changing the frameOptions has caused
6225 : * this WindowClause to be a duplicate of some other WindowClause.
6226 : * This can only happen if we have multiple WindowClauses, so
6227 : * don't bother if there's only 1.
6228 : */
6229 508 : if (list_length(windowClause) == 1)
6230 433 : continue;
6231 :
6232 : /*
6233 : * Do the duplicate check and reuse the existing WindowClause if
6234 : * we find a duplicate.
6235 : */
6236 190 : foreach(lc3, windowClause)
6237 : {
6238 145 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
6239 :
6240 : /* skip over the WindowClause we're currently editing */
6241 145 : if (existing_wc == wc)
6242 45 : continue;
6243 :
6244 : /*
6245 : * Perform the same duplicate check that is done in
6246 : * transformWindowFuncCall.
6247 : */
6248 200 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
6249 100 : equal(wc->orderClause, existing_wc->orderClause) &&
6250 100 : wc->frameOptions == existing_wc->frameOptions &&
6251 60 : equal(wc->startOffset, existing_wc->startOffset) &&
6252 30 : equal(wc->endOffset, existing_wc->endOffset))
6253 : {
6254 : ListCell *lc4;
6255 :
6256 : /*
6257 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
6258 : * This required adjusting each WindowFunc's winref and
6259 : * moving the WindowFuncs in 'wc' to the list of
6260 : * WindowFuncs in 'existing_wc'.
6261 : */
6262 65 : foreach(lc4, wflists->windowFuncs[wc->winref])
6263 : {
6264 35 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
6265 :
6266 35 : wfunc->winref = existing_wc->winref;
6267 : }
6268 :
6269 : /* move list items */
6270 60 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
6271 30 : wflists->windowFuncs[wc->winref]);
6272 30 : wflists->windowFuncs[wc->winref] = NIL;
6273 :
6274 : /*
6275 : * transformWindowFuncCall() should have made sure there
6276 : * are no other duplicates, so we needn't bother looking
6277 : * any further.
6278 : */
6279 30 : break;
6280 : }
6281 : }
6282 : }
6283 : }
6284 2217 : }
6285 :
6286 : /*
6287 : * select_active_windows
6288 : * Create a list of the "active" window clauses (ie, those referenced
6289 : * by non-deleted WindowFuncs) in the order they are to be executed.
6290 : */
6291 : static List *
6292 2217 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
6293 : {
6294 2217 : List *windowClause = root->parse->windowClause;
6295 2217 : List *result = NIL;
6296 : ListCell *lc;
6297 2217 : int nActive = 0;
6298 2217 : WindowClauseSortData *actives = palloc_array(WindowClauseSortData,
6299 : list_length(windowClause));
6300 :
6301 : /* First, construct an array of the active windows */
6302 4639 : foreach(lc, windowClause)
6303 : {
6304 2422 : WindowClause *wc = lfirst_node(WindowClause, lc);
6305 :
6306 : /* It's only active if wflists shows some related WindowFuncs */
6307 : Assert(wc->winref <= wflists->maxWinRef);
6308 2422 : if (wflists->windowFuncs[wc->winref] == NIL)
6309 50 : continue;
6310 :
6311 2372 : actives[nActive].wc = wc; /* original clause */
6312 :
6313 : /*
6314 : * For sorting, we want the list of partition keys followed by the
6315 : * list of sort keys. But pathkeys construction will remove duplicates
6316 : * between the two, so we can as well (even though we can't detect all
6317 : * of the duplicates, since some may come from ECs - that might mean
6318 : * we miss optimization chances here). We must, however, ensure that
6319 : * the order of entries is preserved with respect to the ones we do
6320 : * keep.
6321 : *
6322 : * partitionClause and orderClause had their own duplicates removed in
6323 : * parse analysis, so we're only concerned here with removing
6324 : * orderClause entries that also appear in partitionClause.
6325 : */
6326 4744 : actives[nActive].uniqueOrder =
6327 2372 : list_concat_unique(list_copy(wc->partitionClause),
6328 2372 : wc->orderClause);
6329 2372 : nActive++;
6330 : }
6331 :
6332 : /*
6333 : * Sort active windows by their partitioning/ordering clauses, ignoring
6334 : * any framing clauses, so that the windows that need the same sorting are
6335 : * adjacent in the list. When we come to generate paths, this will avoid
6336 : * inserting additional Sort nodes.
6337 : *
6338 : * This is how we implement a specific requirement from the SQL standard,
6339 : * which says that when two or more windows are order-equivalent (i.e.
6340 : * have matching partition and order clauses, even if their names or
6341 : * framing clauses differ), then all peer rows must be presented in the
6342 : * same order in all of them. If we allowed multiple sort nodes for such
6343 : * cases, we'd risk having the peer rows end up in different orders in
6344 : * equivalent windows due to sort instability. (See General Rule 4 of
6345 : * <window clause> in SQL2008 - SQL2016.)
6346 : *
6347 : * Additionally, if the entire list of clauses of one window is a prefix
6348 : * of another, put first the window with stronger sorting requirements.
6349 : * This way we will first sort for stronger window, and won't have to sort
6350 : * again for the weaker one.
6351 : */
6352 2217 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6353 :
6354 : /* build ordered list of the original WindowClause nodes */
6355 4589 : for (int i = 0; i < nActive; i++)
6356 2372 : result = lappend(result, actives[i].wc);
6357 :
6358 2217 : pfree(actives);
6359 :
6360 2217 : return result;
6361 : }
6362 :
6363 : /*
6364 : * name_active_windows
6365 : * Ensure all active windows have unique names.
6366 : *
6367 : * The parser will have checked that user-assigned window names are unique
6368 : * within the Query. Here we assign made-up names to any unnamed
6369 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6370 : * at parse time, because it'd mess up decompilation of views.)
6371 : *
6372 : * activeWindows: result of select_active_windows
6373 : */
6374 : static void
6375 2217 : name_active_windows(List *activeWindows)
6376 : {
6377 2217 : int next_n = 1;
6378 : char newname[16];
6379 : ListCell *lc;
6380 :
6381 4589 : foreach(lc, activeWindows)
6382 : {
6383 2372 : WindowClause *wc = lfirst_node(WindowClause, lc);
6384 :
6385 : /* Nothing to do if it has a name already. */
6386 2372 : if (wc->name)
6387 480 : continue;
6388 :
6389 : /* Select a name not currently present in the list. */
6390 : for (;;)
6391 5 : {
6392 : ListCell *lc2;
6393 :
6394 1897 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6395 4084 : foreach(lc2, activeWindows)
6396 : {
6397 2192 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6398 :
6399 2192 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6400 5 : break; /* matched */
6401 : }
6402 1897 : if (lc2 == NULL)
6403 1892 : break; /* reached the end with no match */
6404 : }
6405 1892 : wc->name = pstrdup(newname);
6406 : }
6407 2217 : }
6408 :
6409 : /*
6410 : * common_prefix_cmp
6411 : * QSort comparison function for WindowClauseSortData
6412 : *
6413 : * Sort the windows by the required sorting clauses. First, compare the sort
6414 : * clauses themselves. Second, if one window's clauses are a prefix of another
6415 : * one's clauses, put the window with more sort clauses first.
6416 : *
6417 : * We purposefully sort by the highest tleSortGroupRef first. Since
6418 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6419 : * and because here we sort the lowest tleSortGroupRefs last, if a
6420 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6421 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6422 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6423 : * reducing the total number of sorts required for the query.
6424 : */
6425 : static int
6426 170 : common_prefix_cmp(const void *a, const void *b)
6427 : {
6428 170 : const WindowClauseSortData *wcsa = a;
6429 170 : const WindowClauseSortData *wcsb = b;
6430 : ListCell *item_a;
6431 : ListCell *item_b;
6432 :
6433 305 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6434 : {
6435 220 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6436 220 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6437 :
6438 220 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6439 85 : return -1;
6440 210 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6441 55 : return 1;
6442 155 : else if (sca->sortop > scb->sortop)
6443 0 : return -1;
6444 155 : else if (sca->sortop < scb->sortop)
6445 20 : return 1;
6446 135 : else if (sca->nulls_first && !scb->nulls_first)
6447 0 : return -1;
6448 135 : else if (!sca->nulls_first && scb->nulls_first)
6449 0 : return 1;
6450 : /* no need to compare eqop, since it is fully determined by sortop */
6451 : }
6452 :
6453 85 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6454 5 : return -1;
6455 80 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6456 25 : return 1;
6457 :
6458 55 : return 0;
6459 : }
6460 :
6461 : /*
6462 : * make_window_input_target
6463 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6464 : *
6465 : * When the query has window functions, this function computes the desired
6466 : * target to be computed by the node just below the first WindowAgg.
6467 : * This tlist must contain all values needed to evaluate the window functions,
6468 : * compute the final target list, and perform any required final sort step.
6469 : * If multiple WindowAggs are needed, each intermediate one adds its window
6470 : * function results onto this base tlist; only the topmost WindowAgg computes
6471 : * the actual desired target list.
6472 : *
6473 : * This function is much like make_group_input_target, though not quite enough
6474 : * like it to share code. As in that function, we flatten most expressions
6475 : * into their component variables. But we do not want to flatten window
6476 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6477 : * evaluations of them, which would be bad (possibly even resulting in
6478 : * inconsistent answers, if they contain volatile functions).
6479 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6480 : * make_group_input_target, because we may no longer have access to the
6481 : * individual Vars in them.
6482 : *
6483 : * Another key difference from make_group_input_target is that we don't
6484 : * flatten Aggref expressions, since those are to be computed below the
6485 : * window functions and just referenced like Vars above that.
6486 : *
6487 : * 'final_target' is the query's final target list (in PathTarget form)
6488 : * 'activeWindows' is the list of active windows previously identified by
6489 : * select_active_windows.
6490 : *
6491 : * The result is the PathTarget to be computed by the plan node immediately
6492 : * below the first WindowAgg node.
6493 : */
6494 : static PathTarget *
6495 2217 : make_window_input_target(PlannerInfo *root,
6496 : PathTarget *final_target,
6497 : List *activeWindows)
6498 : {
6499 : PathTarget *input_target;
6500 : Bitmapset *sgrefs;
6501 : List *flattenable_cols;
6502 : List *flattenable_vars;
6503 : int i;
6504 : ListCell *lc;
6505 :
6506 : Assert(root->parse->hasWindowFuncs);
6507 :
6508 : /*
6509 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6510 : * into a bitmapset for convenient reference below.
6511 : */
6512 2217 : sgrefs = NULL;
6513 4589 : foreach(lc, activeWindows)
6514 : {
6515 2372 : WindowClause *wc = lfirst_node(WindowClause, lc);
6516 : ListCell *lc2;
6517 :
6518 2994 : foreach(lc2, wc->partitionClause)
6519 : {
6520 622 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6521 :
6522 622 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6523 : }
6524 4277 : foreach(lc2, wc->orderClause)
6525 : {
6526 1905 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6527 :
6528 1905 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6529 : }
6530 : }
6531 :
6532 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6533 2373 : foreach(lc, root->processed_groupClause)
6534 : {
6535 156 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6536 :
6537 156 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6538 : }
6539 :
6540 : /*
6541 : * Construct a target containing all the non-flattenable targetlist items,
6542 : * and save aside the others for a moment.
6543 : */
6544 2217 : input_target = create_empty_pathtarget();
6545 2217 : flattenable_cols = NIL;
6546 :
6547 2217 : i = 0;
6548 9324 : foreach(lc, final_target->exprs)
6549 : {
6550 7107 : Expr *expr = (Expr *) lfirst(lc);
6551 7107 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6552 :
6553 : /*
6554 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6555 : * that such items can't contain window functions, so it's okay to
6556 : * compute them below the WindowAgg nodes.)
6557 : */
6558 7107 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6559 : {
6560 : /*
6561 : * Don't want to deconstruct this value, so add it to the input
6562 : * target as-is.
6563 : */
6564 2389 : add_column_to_pathtarget(input_target, expr, sgref);
6565 : }
6566 : else
6567 : {
6568 : /*
6569 : * Column is to be flattened, so just remember the expression for
6570 : * later call to pull_var_clause.
6571 : */
6572 4718 : flattenable_cols = lappend(flattenable_cols, expr);
6573 : }
6574 :
6575 7107 : i++;
6576 : }
6577 :
6578 : /*
6579 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6580 : * add them to the input target if not already present. (Some might be
6581 : * there already because they're used directly as window/group clauses.)
6582 : *
6583 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6584 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6585 : * at higher levels. On the other hand, we should recurse into
6586 : * WindowFuncs to make sure their input expressions are available.
6587 : */
6588 2217 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6589 : PVC_INCLUDE_AGGREGATES |
6590 : PVC_RECURSE_WINDOWFUNCS |
6591 : PVC_INCLUDE_PLACEHOLDERS);
6592 2217 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6593 :
6594 : /* clean up cruft */
6595 2217 : list_free(flattenable_vars);
6596 2217 : list_free(flattenable_cols);
6597 :
6598 : /* XXX this causes some redundant cost calculation ... */
6599 2217 : return set_pathtarget_cost_width(root, input_target);
6600 : }
6601 :
6602 : /*
6603 : * make_pathkeys_for_window
6604 : * Create a pathkeys list describing the required input ordering
6605 : * for the given WindowClause.
6606 : *
6607 : * Modifies wc's partitionClause to remove any clauses which are deemed
6608 : * redundant by the pathkey logic.
6609 : *
6610 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6611 : * In the future we might try to implement windowing using hashing, in which
6612 : * case the ordering could be relaxed, but for now we always sort.
6613 : */
6614 : static List *
6615 4756 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6616 : List *tlist)
6617 : {
6618 4756 : List *window_pathkeys = NIL;
6619 :
6620 : /* Throw error if can't sort */
6621 4756 : if (!grouping_is_sortable(wc->partitionClause))
6622 0 : ereport(ERROR,
6623 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6624 : errmsg("could not implement window PARTITION BY"),
6625 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6626 4756 : if (!grouping_is_sortable(wc->orderClause))
6627 0 : ereport(ERROR,
6628 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6629 : errmsg("could not implement window ORDER BY"),
6630 : errdetail("Window ordering columns must be of sortable datatypes.")));
6631 :
6632 : /*
6633 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6634 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6635 : */
6636 4756 : if (wc->partitionClause != NIL)
6637 : {
6638 : bool sortable;
6639 :
6640 1088 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6641 : &wc->partitionClause,
6642 : tlist,
6643 : true,
6644 : false,
6645 : &sortable,
6646 : false);
6647 :
6648 : Assert(sortable);
6649 : }
6650 :
6651 : /*
6652 : * In principle, we could also consider removing redundant ORDER BY items
6653 : * too as doing so does not alter the result of peer row checks done by
6654 : * the executor. However, we must *not* remove the ordering column for
6655 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6656 : * if it's known to be equal to some partitioning column.
6657 : */
6658 4756 : if (wc->orderClause != NIL)
6659 : {
6660 : List *orderby_pathkeys;
6661 :
6662 3724 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6663 : wc->orderClause,
6664 : tlist);
6665 :
6666 : /* Okay, make the combined pathkeys */
6667 3724 : if (window_pathkeys != NIL)
6668 781 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6669 : else
6670 2943 : window_pathkeys = orderby_pathkeys;
6671 : }
6672 :
6673 4756 : return window_pathkeys;
6674 : }
6675 :
6676 : /*
6677 : * make_sort_input_target
6678 : * Generate appropriate PathTarget for initial input to Sort step.
6679 : *
6680 : * If the query has ORDER BY, this function chooses the target to be computed
6681 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6682 : * project) steps. This might or might not be identical to the query's final
6683 : * output target.
6684 : *
6685 : * The main argument for keeping the sort-input tlist the same as the final
6686 : * is that we avoid a separate projection node (which will be needed if
6687 : * they're different, because Sort can't project). However, there are also
6688 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6689 : * a consistent order of evaluation for any volatile functions in the tlist,
6690 : * and if there's also a LIMIT, we can stop the query without ever computing
6691 : * tlist functions for later rows, which is beneficial for both volatile and
6692 : * expensive functions.
6693 : *
6694 : * Our current policy is to postpone volatile expressions till after the sort
6695 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6696 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6697 : * postpone set-returning expressions, because running them beforehand would
6698 : * bloat the sort dataset, and because it might cause unexpected output order
6699 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6700 : * in the tlist should be evaluated at the same plan step, so that they can
6701 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6702 : * mustn't postpone any SRFs. (Note that in principle that policy should
6703 : * probably get applied to the group/window input targetlists too, but we
6704 : * have not done that historically.) Lastly, expensive expressions are
6705 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6706 : * partial evaluation of the query is possible (if neither is true, we expect
6707 : * to have to evaluate the expressions for every row anyway), or if there are
6708 : * any volatile or set-returning expressions (since once we've put in a
6709 : * projection at all, it won't cost any more to postpone more stuff).
6710 : *
6711 : * Another issue that could potentially be considered here is that
6712 : * evaluating tlist expressions could result in data that's either wider
6713 : * or narrower than the input Vars, thus changing the volume of data that
6714 : * has to go through the Sort. However, we usually have only a very bad
6715 : * idea of the output width of any expression more complex than a Var,
6716 : * so for now it seems too risky to try to optimize on that basis.
6717 : *
6718 : * Note that if we do produce a modified sort-input target, and then the
6719 : * query ends up not using an explicit Sort, no particular harm is done:
6720 : * we'll initially use the modified target for the preceding path nodes,
6721 : * but then change them to the final target with apply_projection_to_path.
6722 : * Moreover, in such a case the guarantees about evaluation order of
6723 : * volatile functions still hold, since the rows are sorted already.
6724 : *
6725 : * This function has some things in common with make_group_input_target and
6726 : * make_window_input_target, though the detailed rules for what to do are
6727 : * different. We never flatten/postpone any grouping or ordering columns;
6728 : * those are needed before the sort. If we do flatten a particular
6729 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6730 : * computed earlier.
6731 : *
6732 : * 'final_target' is the query's final target list (in PathTarget form)
6733 : * 'have_postponed_srfs' is an output argument, see below
6734 : *
6735 : * The result is the PathTarget to be computed by the plan node immediately
6736 : * below the Sort step (and the Distinct step, if any). This will be
6737 : * exactly final_target if we decide a projection step wouldn't be helpful.
6738 : *
6739 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6740 : * any set-returning functions to after the Sort.
6741 : */
6742 : static PathTarget *
6743 58900 : make_sort_input_target(PlannerInfo *root,
6744 : PathTarget *final_target,
6745 : bool *have_postponed_srfs)
6746 : {
6747 58900 : Query *parse = root->parse;
6748 : PathTarget *input_target;
6749 : int ncols;
6750 : bool *col_is_srf;
6751 : bool *postpone_col;
6752 : bool have_srf;
6753 : bool have_volatile;
6754 : bool have_expensive;
6755 : bool have_srf_sortcols;
6756 : bool postpone_srfs;
6757 : List *postponable_cols;
6758 : List *postponable_vars;
6759 : int i;
6760 : ListCell *lc;
6761 :
6762 : /* Shouldn't get here unless query has ORDER BY */
6763 : Assert(parse->sortClause);
6764 :
6765 58900 : *have_postponed_srfs = false; /* default result */
6766 :
6767 : /* Inspect tlist and collect per-column information */
6768 58900 : ncols = list_length(final_target->exprs);
6769 58900 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6770 58900 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6771 58900 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6772 :
6773 58900 : i = 0;
6774 342736 : foreach(lc, final_target->exprs)
6775 : {
6776 283836 : Expr *expr = (Expr *) lfirst(lc);
6777 :
6778 : /*
6779 : * If the column has a sortgroupref, assume it has to be evaluated
6780 : * before sorting. Generally such columns would be ORDER BY, GROUP
6781 : * BY, etc targets. One exception is columns that were removed from
6782 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6783 : * only be Vars anyway. There don't seem to be any cases where it
6784 : * would be worth the trouble to double-check.
6785 : */
6786 283836 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6787 : {
6788 : /*
6789 : * Check for SRF or volatile functions. Check the SRF case first
6790 : * because we must know whether we have any postponed SRFs.
6791 : */
6792 200615 : if (parse->hasTargetSRFs &&
6793 180 : expression_returns_set((Node *) expr))
6794 : {
6795 : /* We'll decide below whether these are postponable */
6796 80 : col_is_srf[i] = true;
6797 80 : have_srf = true;
6798 : }
6799 200355 : else if (contain_volatile_functions((Node *) expr))
6800 : {
6801 : /* Unconditionally postpone */
6802 185 : postpone_col[i] = true;
6803 185 : have_volatile = true;
6804 : }
6805 : else
6806 : {
6807 : /*
6808 : * Else check the cost. XXX it's annoying to have to do this
6809 : * when set_pathtarget_cost_width() just did it. Refactor to
6810 : * allow sharing the work?
6811 : */
6812 : QualCost cost;
6813 :
6814 200170 : cost_qual_eval_node(&cost, (Node *) expr, root);
6815 :
6816 : /*
6817 : * We arbitrarily define "expensive" as "more than 10X
6818 : * cpu_operator_cost". Note this will take in any PL function
6819 : * with default cost.
6820 : */
6821 200170 : if (cost.per_tuple > 10 * cpu_operator_cost)
6822 : {
6823 14339 : postpone_col[i] = true;
6824 14339 : have_expensive = true;
6825 : }
6826 : }
6827 : }
6828 : else
6829 : {
6830 : /* For sortgroupref cols, just check if any contain SRFs */
6831 83401 : if (!have_srf_sortcols &&
6832 83658 : parse->hasTargetSRFs &&
6833 277 : expression_returns_set((Node *) expr))
6834 122 : have_srf_sortcols = true;
6835 : }
6836 :
6837 283836 : i++;
6838 : }
6839 :
6840 : /*
6841 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6842 : */
6843 58900 : postpone_srfs = (have_srf && !have_srf_sortcols);
6844 :
6845 : /*
6846 : * If we don't need a post-sort projection, just return final_target.
6847 : */
6848 58900 : if (!(postpone_srfs || have_volatile ||
6849 58667 : (have_expensive &&
6850 8591 : (parse->limitCount || root->tuple_fraction > 0))))
6851 58637 : return final_target;
6852 :
6853 : /*
6854 : * Report whether the post-sort projection will contain set-returning
6855 : * functions. This is important because it affects whether the Sort can
6856 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6857 : * to return.
6858 : */
6859 263 : *have_postponed_srfs = postpone_srfs;
6860 :
6861 : /*
6862 : * Construct the sort-input target, taking all non-postponable columns and
6863 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6864 : * the postponable ones.
6865 : */
6866 263 : input_target = create_empty_pathtarget();
6867 263 : postponable_cols = NIL;
6868 :
6869 263 : i = 0;
6870 1855 : foreach(lc, final_target->exprs)
6871 : {
6872 1592 : Expr *expr = (Expr *) lfirst(lc);
6873 :
6874 1592 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6875 315 : postponable_cols = lappend(postponable_cols, expr);
6876 : else
6877 1277 : add_column_to_pathtarget(input_target, expr,
6878 1277 : get_pathtarget_sortgroupref(final_target, i));
6879 :
6880 1592 : i++;
6881 : }
6882 :
6883 : /*
6884 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6885 : * postponable columns, and add them to the sort-input target if not
6886 : * already present. (Some might be there already.) We mustn't
6887 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6888 : * would be unable to recompute them.
6889 : */
6890 263 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6891 : PVC_INCLUDE_AGGREGATES |
6892 : PVC_INCLUDE_WINDOWFUNCS |
6893 : PVC_INCLUDE_PLACEHOLDERS);
6894 263 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6895 :
6896 : /* clean up cruft */
6897 263 : list_free(postponable_vars);
6898 263 : list_free(postponable_cols);
6899 :
6900 : /* XXX this represents even more redundant cost calculation ... */
6901 263 : return set_pathtarget_cost_width(root, input_target);
6902 : }
6903 :
6904 : /*
6905 : * get_cheapest_fractional_path
6906 : * Find the cheapest path for retrieving a specified fraction of all
6907 : * the tuples expected to be returned by the given relation.
6908 : *
6909 : * Do not consider parameterized paths. If the caller needs a path for upper
6910 : * rel, it can't have parameterized paths. If the caller needs an append
6911 : * subpath, it could become limited by the treatment of similar
6912 : * parameterization of all the subpaths.
6913 : *
6914 : * We interpret tuple_fraction the same way as grouping_planner.
6915 : *
6916 : * We assume set_cheapest() has been run on the given rel.
6917 : */
6918 : Path *
6919 365698 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6920 : {
6921 365698 : Path *best_path = rel->cheapest_total_path;
6922 : ListCell *l;
6923 :
6924 : /* If all tuples will be retrieved, just return the cheapest-total path */
6925 365698 : if (tuple_fraction <= 0.0)
6926 358627 : return best_path;
6927 :
6928 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6929 7071 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6930 3068 : tuple_fraction /= best_path->rows;
6931 :
6932 18101 : foreach(l, rel->pathlist)
6933 : {
6934 11030 : Path *path = (Path *) lfirst(l);
6935 :
6936 11030 : if (path->param_info)
6937 166 : continue;
6938 :
6939 14657 : if (path == rel->cheapest_total_path ||
6940 3793 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6941 10542 : continue;
6942 :
6943 322 : best_path = path;
6944 : }
6945 :
6946 7071 : return best_path;
6947 : }
6948 :
6949 : /*
6950 : * adjust_paths_for_srfs
6951 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6952 : *
6953 : * The executor can only handle set-returning functions that appear at the
6954 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6955 : * that are not at top level, we need to split up the evaluation into multiple
6956 : * plan levels in which each level satisfies this constraint. This function
6957 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6958 : * output tlist to insert appropriate projection steps.
6959 : *
6960 : * The given targets and targets_contain_srfs lists are from
6961 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6962 : * target in targets.
6963 : */
6964 : static void
6965 10857 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6966 : List *targets, List *targets_contain_srfs)
6967 : {
6968 : ListCell *lc;
6969 :
6970 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6971 : Assert(!linitial_int(targets_contain_srfs));
6972 :
6973 : /* If no SRFs appear at this plan level, nothing to do */
6974 10857 : if (list_length(targets) == 1)
6975 491 : return;
6976 :
6977 : /*
6978 : * Stack SRF-evaluation nodes atop each path for the rel.
6979 : *
6980 : * In principle we should re-run set_cheapest() here to identify the
6981 : * cheapest path, but it seems unlikely that adding the same tlist eval
6982 : * costs to all the paths would change that, so we don't bother. Instead,
6983 : * just assume that the cheapest-startup and cheapest-total paths remain
6984 : * so. (There should be no parameterized paths anymore, so we needn't
6985 : * worry about updating cheapest_parameterized_paths.)
6986 : */
6987 20764 : foreach(lc, rel->pathlist)
6988 : {
6989 10398 : Path *subpath = (Path *) lfirst(lc);
6990 10398 : Path *newpath = subpath;
6991 : ListCell *lc1,
6992 : *lc2;
6993 :
6994 : Assert(subpath->param_info == NULL);
6995 31981 : forboth(lc1, targets, lc2, targets_contain_srfs)
6996 : {
6997 21583 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6998 21583 : bool contains_srfs = (bool) lfirst_int(lc2);
6999 :
7000 : /* If this level doesn't contain SRFs, do regular projection */
7001 21583 : if (contains_srfs)
7002 10448 : newpath = (Path *) create_set_projection_path(root,
7003 : rel,
7004 : newpath,
7005 : thistarget);
7006 : else
7007 11135 : newpath = (Path *) apply_projection_to_path(root,
7008 : rel,
7009 : newpath,
7010 : thistarget);
7011 : }
7012 10398 : lfirst(lc) = newpath;
7013 10398 : if (subpath == rel->cheapest_startup_path)
7014 239 : rel->cheapest_startup_path = newpath;
7015 10398 : if (subpath == rel->cheapest_total_path)
7016 239 : rel->cheapest_total_path = newpath;
7017 : }
7018 :
7019 : /* Likewise for partial paths, if any */
7020 10381 : foreach(lc, rel->partial_pathlist)
7021 : {
7022 15 : Path *subpath = (Path *) lfirst(lc);
7023 15 : Path *newpath = subpath;
7024 : ListCell *lc1,
7025 : *lc2;
7026 :
7027 : Assert(subpath->param_info == NULL);
7028 60 : forboth(lc1, targets, lc2, targets_contain_srfs)
7029 : {
7030 45 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
7031 45 : bool contains_srfs = (bool) lfirst_int(lc2);
7032 :
7033 : /* If this level doesn't contain SRFs, do regular projection */
7034 45 : if (contains_srfs)
7035 15 : newpath = (Path *) create_set_projection_path(root,
7036 : rel,
7037 : newpath,
7038 : thistarget);
7039 : else
7040 : {
7041 : /* avoid apply_projection_to_path, in case of multiple refs */
7042 30 : newpath = (Path *) create_projection_path(root,
7043 : rel,
7044 : newpath,
7045 : thistarget);
7046 : }
7047 : }
7048 15 : lfirst(lc) = newpath;
7049 : }
7050 : }
7051 :
7052 : /*
7053 : * expression_planner
7054 : * Perform planner's transformations on a standalone expression.
7055 : *
7056 : * Various utility commands need to evaluate expressions that are not part
7057 : * of a plannable query. They can do so using the executor's regular
7058 : * expression-execution machinery, but first the expression has to be fed
7059 : * through here to transform it from parser output to something executable.
7060 : *
7061 : * Currently, we disallow sublinks in standalone expressions, so there's no
7062 : * real "planning" involved here. (That might not always be true though.)
7063 : * What we must do is run eval_const_expressions to ensure that any function
7064 : * calls are converted to positional notation and function default arguments
7065 : * get inserted. The fact that constant subexpressions get simplified is a
7066 : * side-effect that is useful when the expression will get evaluated more than
7067 : * once. Also, we must fix operator function IDs.
7068 : *
7069 : * This does not return any information about dependencies of the expression.
7070 : * Hence callers should use the results only for the duration of the current
7071 : * query. Callers that would like to cache the results for longer should use
7072 : * expression_planner_with_deps, probably via the plancache.
7073 : *
7074 : * Note: this must not make any damaging changes to the passed-in expression
7075 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
7076 : * we first do an expression_tree_mutator-based walk, what is returned will
7077 : * be a new node tree.) The result is constructed in the current memory
7078 : * context; beware that this can leak a lot of additional stuff there, too.
7079 : */
7080 : Expr *
7081 138753 : expression_planner(Expr *expr)
7082 : {
7083 : Node *result;
7084 :
7085 : /*
7086 : * Convert named-argument function calls, insert default arguments and
7087 : * simplify constant subexprs
7088 : */
7089 138753 : result = eval_const_expressions(NULL, (Node *) expr);
7090 :
7091 : /* Fill in opfuncid values if missing */
7092 138733 : fix_opfuncids(result);
7093 :
7094 138733 : return (Expr *) result;
7095 : }
7096 :
7097 : /*
7098 : * expression_planner_with_deps
7099 : * Perform planner's transformations on a standalone expression,
7100 : * returning expression dependency information along with the result.
7101 : *
7102 : * This is identical to expression_planner() except that it also returns
7103 : * information about possible dependencies of the expression, ie identities of
7104 : * objects whose definitions affect the result. As in a PlannedStmt, these
7105 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
7106 : */
7107 : Expr *
7108 256 : expression_planner_with_deps(Expr *expr,
7109 : List **relationOids,
7110 : List **invalItems)
7111 : {
7112 : Node *result;
7113 : PlannerGlobal glob;
7114 : PlannerInfo root;
7115 :
7116 : /* Make up dummy planner state so we can use setrefs machinery */
7117 7424 : MemSet(&glob, 0, sizeof(glob));
7118 256 : glob.type = T_PlannerGlobal;
7119 256 : glob.relationOids = NIL;
7120 256 : glob.invalItems = NIL;
7121 :
7122 24064 : MemSet(&root, 0, sizeof(root));
7123 256 : root.type = T_PlannerInfo;
7124 256 : root.glob = &glob;
7125 :
7126 : /*
7127 : * Convert named-argument function calls, insert default arguments and
7128 : * simplify constant subexprs. Collect identities of inlined functions
7129 : * and elided domains, too.
7130 : */
7131 256 : result = eval_const_expressions(&root, (Node *) expr);
7132 :
7133 : /* Fill in opfuncid values if missing */
7134 256 : fix_opfuncids(result);
7135 :
7136 : /*
7137 : * Now walk the finished expression to find anything else we ought to
7138 : * record as an expression dependency.
7139 : */
7140 256 : (void) extract_query_dependencies_walker(result, &root);
7141 :
7142 256 : *relationOids = glob.relationOids;
7143 256 : *invalItems = glob.invalItems;
7144 :
7145 256 : return (Expr *) result;
7146 : }
7147 :
7148 :
7149 : /*
7150 : * plan_cluster_use_sort
7151 : * Use the planner to decide how CLUSTER should implement sorting
7152 : *
7153 : * tableOid is the OID of a table to be clustered on its index indexOid
7154 : * (which is already known to be a btree index). Decide whether it's
7155 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
7156 : * Return true to use sorting, false to use an indexscan.
7157 : *
7158 : * Note: caller had better already hold some type of lock on the table.
7159 : */
7160 : bool
7161 144 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
7162 : {
7163 : PlannerInfo *root;
7164 : Query *query;
7165 : PlannerGlobal *glob;
7166 : RangeTblEntry *rte;
7167 : RelOptInfo *rel;
7168 : IndexOptInfo *indexInfo;
7169 : QualCost indexExprCost;
7170 : Cost comparisonCost;
7171 : Path *seqScanPath;
7172 : Path seqScanAndSortPath;
7173 : IndexPath *indexScanPath;
7174 : ListCell *lc;
7175 :
7176 : /* We can short-circuit the cost comparison if indexscans are disabled */
7177 144 : if (!enable_indexscan)
7178 20 : return true; /* use sort */
7179 :
7180 : /* Set up mostly-dummy planner state */
7181 124 : query = makeNode(Query);
7182 124 : query->commandType = CMD_SELECT;
7183 :
7184 124 : glob = makeNode(PlannerGlobal);
7185 :
7186 124 : root = makeNode(PlannerInfo);
7187 124 : root->parse = query;
7188 124 : root->glob = glob;
7189 124 : root->query_level = 1;
7190 124 : root->planner_cxt = CurrentMemoryContext;
7191 124 : root->wt_param_id = -1;
7192 124 : root->join_domains = list_make1(makeNode(JoinDomain));
7193 :
7194 : /* Build a minimal RTE for the rel */
7195 124 : rte = makeNode(RangeTblEntry);
7196 124 : rte->rtekind = RTE_RELATION;
7197 124 : rte->relid = tableOid;
7198 124 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7199 124 : rte->rellockmode = AccessShareLock;
7200 124 : rte->lateral = false;
7201 124 : rte->inh = false;
7202 124 : rte->inFromCl = true;
7203 124 : query->rtable = list_make1(rte);
7204 124 : addRTEPermissionInfo(&query->rteperminfos, rte);
7205 :
7206 : /* Set up RTE/RelOptInfo arrays */
7207 124 : setup_simple_rel_arrays(root);
7208 :
7209 : /* Build RelOptInfo */
7210 124 : rel = build_simple_rel(root, 1, NULL);
7211 :
7212 : /* Locate IndexOptInfo for the target index */
7213 124 : indexInfo = NULL;
7214 154 : foreach(lc, rel->indexlist)
7215 : {
7216 154 : indexInfo = lfirst_node(IndexOptInfo, lc);
7217 154 : if (indexInfo->indexoid == indexOid)
7218 124 : break;
7219 : }
7220 :
7221 : /*
7222 : * It's possible that get_relation_info did not generate an IndexOptInfo
7223 : * for the desired index; this could happen if it's not yet reached its
7224 : * indcheckxmin usability horizon, or if it's a system index and we're
7225 : * ignoring system indexes. In such cases we should tell CLUSTER to not
7226 : * trust the index contents but use seqscan-and-sort.
7227 : */
7228 124 : if (lc == NULL) /* not in the list? */
7229 0 : return true; /* use sort */
7230 :
7231 : /*
7232 : * Rather than doing all the pushups that would be needed to use
7233 : * set_baserel_size_estimates, just do a quick hack for rows and width.
7234 : */
7235 124 : rel->rows = rel->tuples;
7236 124 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
7237 :
7238 124 : root->total_table_pages = rel->pages;
7239 :
7240 : /*
7241 : * Determine eval cost of the index expressions, if any. We need to
7242 : * charge twice that amount for each tuple comparison that happens during
7243 : * the sort, since tuplesort.c will have to re-evaluate the index
7244 : * expressions each time. (XXX that's pretty inefficient...)
7245 : */
7246 124 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
7247 124 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
7248 :
7249 : /* Estimate the cost of seq scan + sort */
7250 124 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
7251 124 : cost_sort(&seqScanAndSortPath, root, NIL,
7252 : seqScanPath->disabled_nodes,
7253 124 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
7254 : comparisonCost, maintenance_work_mem, -1.0);
7255 :
7256 : /* Estimate the cost of index scan */
7257 124 : indexScanPath = create_index_path(root, indexInfo,
7258 : NIL, NIL, NIL, NIL,
7259 : ForwardScanDirection, false,
7260 : NULL, 1.0, false);
7261 :
7262 124 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
7263 : }
7264 :
7265 : /*
7266 : * plan_create_index_workers
7267 : * Use the planner to decide how many parallel worker processes
7268 : * CREATE INDEX should request for use
7269 : *
7270 : * tableOid is the table on which the index is to be built. indexOid is the
7271 : * OID of an index to be created or reindexed (which must be an index with
7272 : * support for parallel builds - currently btree, GIN, or BRIN).
7273 : *
7274 : * Return value is the number of parallel worker processes to request. It
7275 : * may be unsafe to proceed if this is 0. Note that this does not include the
7276 : * leader participating as a worker (value is always a number of parallel
7277 : * worker processes).
7278 : *
7279 : * Note: caller had better already hold some type of lock on the table and
7280 : * index.
7281 : */
7282 : int
7283 23258 : plan_create_index_workers(Oid tableOid, Oid indexOid)
7284 : {
7285 : PlannerInfo *root;
7286 : Query *query;
7287 : PlannerGlobal *glob;
7288 : RangeTblEntry *rte;
7289 : Relation heap;
7290 : Relation index;
7291 : RelOptInfo *rel;
7292 : int parallel_workers;
7293 : BlockNumber heap_blocks;
7294 : double reltuples;
7295 : double allvisfrac;
7296 :
7297 : /*
7298 : * We don't allow performing parallel operation in standalone backend or
7299 : * when parallelism is disabled.
7300 : */
7301 23258 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
7302 301 : return 0;
7303 :
7304 : /* Set up largely-dummy planner state */
7305 22957 : query = makeNode(Query);
7306 22957 : query->commandType = CMD_SELECT;
7307 :
7308 22957 : glob = makeNode(PlannerGlobal);
7309 :
7310 22957 : root = makeNode(PlannerInfo);
7311 22957 : root->parse = query;
7312 22957 : root->glob = glob;
7313 22957 : root->query_level = 1;
7314 22957 : root->planner_cxt = CurrentMemoryContext;
7315 22957 : root->wt_param_id = -1;
7316 22957 : root->join_domains = list_make1(makeNode(JoinDomain));
7317 :
7318 : /*
7319 : * Build a minimal RTE.
7320 : *
7321 : * Mark the RTE with inh = true. This is a kludge to prevent
7322 : * get_relation_info() from fetching index info, which is necessary
7323 : * because it does not expect that any IndexOptInfo is currently
7324 : * undergoing REINDEX.
7325 : */
7326 22957 : rte = makeNode(RangeTblEntry);
7327 22957 : rte->rtekind = RTE_RELATION;
7328 22957 : rte->relid = tableOid;
7329 22957 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7330 22957 : rte->rellockmode = AccessShareLock;
7331 22957 : rte->lateral = false;
7332 22957 : rte->inh = true;
7333 22957 : rte->inFromCl = true;
7334 22957 : query->rtable = list_make1(rte);
7335 22957 : addRTEPermissionInfo(&query->rteperminfos, rte);
7336 :
7337 : /* Set up RTE/RelOptInfo arrays */
7338 22957 : setup_simple_rel_arrays(root);
7339 :
7340 : /* Build RelOptInfo */
7341 22957 : rel = build_simple_rel(root, 1, NULL);
7342 :
7343 : /* Rels are assumed already locked by the caller */
7344 22957 : heap = table_open(tableOid, NoLock);
7345 22957 : index = index_open(indexOid, NoLock);
7346 :
7347 : /*
7348 : * Determine if it's safe to proceed.
7349 : *
7350 : * Currently, parallel workers can't access the leader's temporary tables.
7351 : * Furthermore, any index predicate or index expressions must be parallel
7352 : * safe.
7353 : */
7354 22957 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7355 21562 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7356 21472 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7357 : {
7358 1485 : parallel_workers = 0;
7359 1485 : goto done;
7360 : }
7361 :
7362 : /*
7363 : * If parallel_workers storage parameter is set for the table, accept that
7364 : * as the number of parallel worker processes to launch (though still cap
7365 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7366 : * consider any other factor when parallel_workers is set. (e.g., memory
7367 : * use by workers.)
7368 : */
7369 21472 : if (rel->rel_parallel_workers != -1)
7370 : {
7371 57 : parallel_workers = Min(rel->rel_parallel_workers,
7372 : max_parallel_maintenance_workers);
7373 57 : goto done;
7374 : }
7375 :
7376 : /*
7377 : * Estimate heap relation size ourselves, since rel->pages cannot be
7378 : * trusted (heap RTE was marked as inheritance parent)
7379 : */
7380 21415 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7381 :
7382 : /*
7383 : * Determine number of workers to scan the heap relation using generic
7384 : * model
7385 : */
7386 21415 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7387 : max_parallel_maintenance_workers);
7388 :
7389 : /*
7390 : * Cap workers based on available maintenance_work_mem as needed.
7391 : *
7392 : * Note that each tuplesort participant receives an even share of the
7393 : * total maintenance_work_mem budget. Aim to leave participants
7394 : * (including the leader as a participant) with no less than 32MB of
7395 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7396 : * immediately past the threshold of being capable of launching a single
7397 : * parallel worker to sort.
7398 : */
7399 21524 : while (parallel_workers > 0 &&
7400 219 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7401 109 : parallel_workers--;
7402 :
7403 21415 : done:
7404 22957 : index_close(index, NoLock);
7405 22957 : table_close(heap, NoLock);
7406 :
7407 22957 : return parallel_workers;
7408 : }
7409 :
7410 : /*
7411 : * add_paths_to_grouping_rel
7412 : *
7413 : * Add non-partial paths to grouping relation.
7414 : */
7415 : static void
7416 35144 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7417 : RelOptInfo *grouped_rel,
7418 : RelOptInfo *partially_grouped_rel,
7419 : const AggClauseCosts *agg_costs,
7420 : grouping_sets_data *gd,
7421 : GroupPathExtraData *extra)
7422 : {
7423 35144 : Query *parse = root->parse;
7424 35144 : Path *cheapest_path = input_rel->cheapest_total_path;
7425 35144 : Path *cheapest_partially_grouped_path = NULL;
7426 : ListCell *lc;
7427 35144 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7428 35144 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7429 35144 : List *havingQual = (List *) extra->havingQual;
7430 35144 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7431 35144 : double dNumGroups = 0;
7432 35144 : double dNumFinalGroups = 0;
7433 :
7434 : /*
7435 : * Estimate number of groups for non-split aggregation.
7436 : */
7437 35144 : dNumGroups = get_number_of_groups(root,
7438 : cheapest_path->rows,
7439 : gd,
7440 : extra->targetList);
7441 :
7442 35144 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7443 : {
7444 2493 : cheapest_partially_grouped_path =
7445 : partially_grouped_rel->cheapest_total_path;
7446 :
7447 : /*
7448 : * Estimate number of groups for final phase of partial aggregation.
7449 : */
7450 : dNumFinalGroups =
7451 2493 : get_number_of_groups(root,
7452 : cheapest_partially_grouped_path->rows,
7453 : gd,
7454 : extra->targetList);
7455 : }
7456 :
7457 35144 : if (can_sort)
7458 : {
7459 : /*
7460 : * Use any available suitably-sorted path as input, and also consider
7461 : * sorting the cheapest-total path and incremental sort on any paths
7462 : * with presorted keys.
7463 : */
7464 72861 : foreach(lc, input_rel->pathlist)
7465 : {
7466 : ListCell *lc2;
7467 37722 : Path *path = (Path *) lfirst(lc);
7468 37722 : Path *path_save = path;
7469 37722 : List *pathkey_orderings = NIL;
7470 :
7471 : /* generate alternative group orderings that might be useful */
7472 37722 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7473 :
7474 : Assert(list_length(pathkey_orderings) > 0);
7475 :
7476 75561 : foreach(lc2, pathkey_orderings)
7477 : {
7478 37839 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7479 :
7480 : /* restore the path (we replace it in the loop) */
7481 37839 : path = path_save;
7482 :
7483 37839 : path = make_ordered_path(root,
7484 : grouped_rel,
7485 : path,
7486 : cheapest_path,
7487 : info->pathkeys,
7488 : -1.0);
7489 37839 : if (path == NULL)
7490 306 : continue;
7491 :
7492 : /* Now decide what to stick atop it */
7493 37533 : if (parse->groupingSets)
7494 : {
7495 975 : consider_groupingsets_paths(root, grouped_rel,
7496 : path, true, can_hash,
7497 : gd, agg_costs, dNumGroups);
7498 : }
7499 36558 : else if (parse->hasAggs)
7500 : {
7501 : /*
7502 : * We have aggregation, possibly with plain GROUP BY. Make
7503 : * an AggPath.
7504 : */
7505 35866 : add_path(grouped_rel, (Path *)
7506 35866 : create_agg_path(root,
7507 : grouped_rel,
7508 : path,
7509 35866 : grouped_rel->reltarget,
7510 35866 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7511 : AGGSPLIT_SIMPLE,
7512 : info->clauses,
7513 : havingQual,
7514 : agg_costs,
7515 : dNumGroups));
7516 : }
7517 692 : else if (parse->groupClause)
7518 : {
7519 : /*
7520 : * We have GROUP BY without aggregation or grouping sets.
7521 : * Make a GroupPath.
7522 : */
7523 692 : add_path(grouped_rel, (Path *)
7524 692 : create_group_path(root,
7525 : grouped_rel,
7526 : path,
7527 : info->clauses,
7528 : havingQual,
7529 : dNumGroups));
7530 : }
7531 : else
7532 : {
7533 : /* Other cases should have been handled above */
7534 : Assert(false);
7535 : }
7536 : }
7537 : }
7538 :
7539 : /*
7540 : * Instead of operating directly on the input relation, we can
7541 : * consider finalizing a partially aggregated path.
7542 : */
7543 35139 : if (partially_grouped_rel != NULL)
7544 : {
7545 6422 : foreach(lc, partially_grouped_rel->pathlist)
7546 : {
7547 : ListCell *lc2;
7548 3929 : Path *path = (Path *) lfirst(lc);
7549 3929 : Path *path_save = path;
7550 3929 : List *pathkey_orderings = NIL;
7551 :
7552 : /* generate alternative group orderings that might be useful */
7553 3929 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7554 :
7555 : Assert(list_length(pathkey_orderings) > 0);
7556 :
7557 : /* process all potentially interesting grouping reorderings */
7558 7858 : foreach(lc2, pathkey_orderings)
7559 : {
7560 3929 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7561 :
7562 : /* restore the path (we replace it in the loop) */
7563 3929 : path = path_save;
7564 :
7565 3929 : path = make_ordered_path(root,
7566 : grouped_rel,
7567 : path,
7568 : cheapest_partially_grouped_path,
7569 : info->pathkeys,
7570 : -1.0);
7571 :
7572 3929 : if (path == NULL)
7573 172 : continue;
7574 :
7575 3757 : if (parse->hasAggs)
7576 3555 : add_path(grouped_rel, (Path *)
7577 3555 : create_agg_path(root,
7578 : grouped_rel,
7579 : path,
7580 3555 : grouped_rel->reltarget,
7581 3555 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7582 : AGGSPLIT_FINAL_DESERIAL,
7583 : info->clauses,
7584 : havingQual,
7585 : agg_final_costs,
7586 : dNumFinalGroups));
7587 : else
7588 202 : add_path(grouped_rel, (Path *)
7589 202 : create_group_path(root,
7590 : grouped_rel,
7591 : path,
7592 : info->clauses,
7593 : havingQual,
7594 : dNumFinalGroups));
7595 :
7596 : }
7597 : }
7598 : }
7599 : }
7600 :
7601 35144 : if (can_hash)
7602 : {
7603 5119 : if (parse->groupingSets)
7604 : {
7605 : /*
7606 : * Try for a hash-only groupingsets path over unsorted input.
7607 : */
7608 807 : consider_groupingsets_paths(root, grouped_rel,
7609 : cheapest_path, false, true,
7610 : gd, agg_costs, dNumGroups);
7611 : }
7612 : else
7613 : {
7614 : /*
7615 : * Generate a HashAgg Path. We just need an Agg over the
7616 : * cheapest-total input path, since input order won't matter.
7617 : */
7618 4312 : add_path(grouped_rel, (Path *)
7619 4312 : create_agg_path(root, grouped_rel,
7620 : cheapest_path,
7621 4312 : grouped_rel->reltarget,
7622 : AGG_HASHED,
7623 : AGGSPLIT_SIMPLE,
7624 : root->processed_groupClause,
7625 : havingQual,
7626 : agg_costs,
7627 : dNumGroups));
7628 : }
7629 :
7630 : /*
7631 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7632 : * grouped path, assuming there is one
7633 : */
7634 5119 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7635 : {
7636 1182 : add_path(grouped_rel, (Path *)
7637 1182 : create_agg_path(root,
7638 : grouped_rel,
7639 : cheapest_partially_grouped_path,
7640 1182 : grouped_rel->reltarget,
7641 : AGG_HASHED,
7642 : AGGSPLIT_FINAL_DESERIAL,
7643 : root->processed_groupClause,
7644 : havingQual,
7645 : agg_final_costs,
7646 : dNumFinalGroups));
7647 : }
7648 : }
7649 :
7650 : /*
7651 : * When partitionwise aggregate is used, we might have fully aggregated
7652 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7653 : * consider a path for grouped_rel consisting of a Parallel Append of
7654 : * non-partial paths from each child.
7655 : */
7656 35144 : if (grouped_rel->partial_pathlist != NIL)
7657 265 : gather_grouping_paths(root, grouped_rel);
7658 35144 : }
7659 :
7660 : /*
7661 : * create_partial_grouping_paths
7662 : *
7663 : * Create a new upper relation representing the result of partial aggregation
7664 : * and populate it with appropriate paths. Note that we don't finalize the
7665 : * lists of paths here, so the caller can add additional partial or non-partial
7666 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7667 : * the returned upper relation.
7668 : *
7669 : * All paths for this new upper relation -- both partial and non-partial --
7670 : * have been partially aggregated but require a subsequent FinalizeAggregate
7671 : * step.
7672 : *
7673 : * NB: This function is allowed to return NULL if it determines that there is
7674 : * no real need to create a new RelOptInfo.
7675 : */
7676 : static RelOptInfo *
7677 31673 : create_partial_grouping_paths(PlannerInfo *root,
7678 : RelOptInfo *grouped_rel,
7679 : RelOptInfo *input_rel,
7680 : grouping_sets_data *gd,
7681 : GroupPathExtraData *extra,
7682 : bool force_rel_creation)
7683 : {
7684 31673 : Query *parse = root->parse;
7685 : RelOptInfo *partially_grouped_rel;
7686 31673 : RelOptInfo *eager_agg_rel = NULL;
7687 31673 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7688 31673 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7689 31673 : Path *cheapest_partial_path = NULL;
7690 31673 : Path *cheapest_total_path = NULL;
7691 31673 : double dNumPartialGroups = 0;
7692 31673 : double dNumPartialPartialGroups = 0;
7693 : ListCell *lc;
7694 31673 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7695 31673 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7696 :
7697 : /*
7698 : * Check whether any partially aggregated paths have been generated
7699 : * through eager aggregation.
7700 : */
7701 31673 : if (input_rel->grouped_rel &&
7702 777 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7703 777 : input_rel->grouped_rel->pathlist != NIL)
7704 727 : eager_agg_rel = input_rel->grouped_rel;
7705 :
7706 : /*
7707 : * Consider whether we should generate partially aggregated non-partial
7708 : * paths. We can only do this if we have a non-partial path, and only if
7709 : * the parent of the input rel is performing partial partitionwise
7710 : * aggregation. (Note that extra->patype is the type of partitionwise
7711 : * aggregation being used at the parent level, not this level.)
7712 : */
7713 31673 : if (input_rel->pathlist != NIL &&
7714 31673 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7715 713 : cheapest_total_path = input_rel->cheapest_total_path;
7716 :
7717 : /*
7718 : * If parallelism is possible for grouped_rel, then we should consider
7719 : * generating partially-grouped partial paths. However, if the input rel
7720 : * has no partial paths, then we can't.
7721 : */
7722 31673 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7723 2745 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7724 :
7725 : /*
7726 : * If we can't partially aggregate partial paths, and we can't partially
7727 : * aggregate non-partial paths, and no partially aggregated paths were
7728 : * generated by eager aggregation, then don't bother creating the new
7729 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7730 : */
7731 31673 : if (cheapest_total_path == NULL &&
7732 28635 : cheapest_partial_path == NULL &&
7733 28548 : eager_agg_rel == NULL &&
7734 28548 : !force_rel_creation)
7735 28467 : return NULL;
7736 :
7737 : /*
7738 : * Build a new upper relation to represent the result of partially
7739 : * aggregating the rows from the input relation.
7740 : */
7741 3206 : partially_grouped_rel = fetch_upper_rel(root,
7742 : UPPERREL_PARTIAL_GROUP_AGG,
7743 : grouped_rel->relids);
7744 3206 : partially_grouped_rel->consider_parallel =
7745 3206 : grouped_rel->consider_parallel;
7746 3206 : partially_grouped_rel->pgs_mask = grouped_rel->pgs_mask;
7747 3206 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7748 3206 : partially_grouped_rel->serverid = grouped_rel->serverid;
7749 3206 : partially_grouped_rel->userid = grouped_rel->userid;
7750 3206 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7751 3206 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7752 :
7753 : /*
7754 : * Build target list for partial aggregate paths. These paths cannot just
7755 : * emit the same tlist as regular aggregate paths, because (1) we must
7756 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7757 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7758 : */
7759 3206 : partially_grouped_rel->reltarget =
7760 3206 : make_partial_grouping_target(root, grouped_rel->reltarget,
7761 : extra->havingQual);
7762 :
7763 3206 : if (!extra->partial_costs_set)
7764 : {
7765 : /*
7766 : * Collect statistics about aggregates for estimating costs of
7767 : * performing aggregation in parallel.
7768 : */
7769 11538 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7770 11538 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7771 1923 : if (parse->hasAggs)
7772 : {
7773 : /* partial phase */
7774 1814 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7775 : agg_partial_costs);
7776 :
7777 : /* final phase */
7778 1814 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7779 : agg_final_costs);
7780 : }
7781 :
7782 1923 : extra->partial_costs_set = true;
7783 : }
7784 :
7785 : /* Estimate number of partial groups. */
7786 3206 : if (cheapest_total_path != NULL)
7787 : dNumPartialGroups =
7788 713 : get_number_of_groups(root,
7789 : cheapest_total_path->rows,
7790 : gd,
7791 : extra->targetList);
7792 3206 : if (cheapest_partial_path != NULL)
7793 : dNumPartialPartialGroups =
7794 2745 : get_number_of_groups(root,
7795 : cheapest_partial_path->rows,
7796 : gd,
7797 : extra->targetList);
7798 :
7799 3206 : if (can_sort && cheapest_total_path != NULL)
7800 : {
7801 : /* This should have been checked previously */
7802 : Assert(parse->hasAggs || parse->groupClause);
7803 :
7804 : /*
7805 : * Use any available suitably-sorted path as input, and also consider
7806 : * sorting the cheapest partial path.
7807 : */
7808 1426 : foreach(lc, input_rel->pathlist)
7809 : {
7810 : ListCell *lc2;
7811 713 : Path *path = (Path *) lfirst(lc);
7812 713 : Path *path_save = path;
7813 713 : List *pathkey_orderings = NIL;
7814 :
7815 : /* generate alternative group orderings that might be useful */
7816 713 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7817 :
7818 : Assert(list_length(pathkey_orderings) > 0);
7819 :
7820 : /* process all potentially interesting grouping reorderings */
7821 1426 : foreach(lc2, pathkey_orderings)
7822 : {
7823 713 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7824 :
7825 : /* restore the path (we replace it in the loop) */
7826 713 : path = path_save;
7827 :
7828 713 : path = make_ordered_path(root,
7829 : partially_grouped_rel,
7830 : path,
7831 : cheapest_total_path,
7832 : info->pathkeys,
7833 : -1.0);
7834 :
7835 713 : if (path == NULL)
7836 0 : continue;
7837 :
7838 713 : if (parse->hasAggs)
7839 653 : add_path(partially_grouped_rel, (Path *)
7840 653 : create_agg_path(root,
7841 : partially_grouped_rel,
7842 : path,
7843 653 : partially_grouped_rel->reltarget,
7844 653 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7845 : AGGSPLIT_INITIAL_SERIAL,
7846 : info->clauses,
7847 : NIL,
7848 : agg_partial_costs,
7849 : dNumPartialGroups));
7850 : else
7851 60 : add_path(partially_grouped_rel, (Path *)
7852 60 : create_group_path(root,
7853 : partially_grouped_rel,
7854 : path,
7855 : info->clauses,
7856 : NIL,
7857 : dNumPartialGroups));
7858 : }
7859 : }
7860 : }
7861 :
7862 3206 : if (can_sort && cheapest_partial_path != NULL)
7863 : {
7864 : /* Similar to above logic, but for partial paths. */
7865 5865 : foreach(lc, input_rel->partial_pathlist)
7866 : {
7867 : ListCell *lc2;
7868 3120 : Path *path = (Path *) lfirst(lc);
7869 3120 : Path *path_save = path;
7870 3120 : List *pathkey_orderings = NIL;
7871 :
7872 : /* generate alternative group orderings that might be useful */
7873 3120 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7874 :
7875 : Assert(list_length(pathkey_orderings) > 0);
7876 :
7877 : /* process all potentially interesting grouping reorderings */
7878 6240 : foreach(lc2, pathkey_orderings)
7879 : {
7880 3120 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7881 :
7882 :
7883 : /* restore the path (we replace it in the loop) */
7884 3120 : path = path_save;
7885 :
7886 3120 : path = make_ordered_path(root,
7887 : partially_grouped_rel,
7888 : path,
7889 : cheapest_partial_path,
7890 : info->pathkeys,
7891 : -1.0);
7892 :
7893 3120 : if (path == NULL)
7894 5 : continue;
7895 :
7896 3115 : if (parse->hasAggs)
7897 3016 : add_partial_path(partially_grouped_rel, (Path *)
7898 3016 : create_agg_path(root,
7899 : partially_grouped_rel,
7900 : path,
7901 3016 : partially_grouped_rel->reltarget,
7902 3016 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7903 : AGGSPLIT_INITIAL_SERIAL,
7904 : info->clauses,
7905 : NIL,
7906 : agg_partial_costs,
7907 : dNumPartialPartialGroups));
7908 : else
7909 99 : add_partial_path(partially_grouped_rel, (Path *)
7910 99 : create_group_path(root,
7911 : partially_grouped_rel,
7912 : path,
7913 : info->clauses,
7914 : NIL,
7915 : dNumPartialPartialGroups));
7916 : }
7917 : }
7918 : }
7919 :
7920 : /*
7921 : * Add a partially-grouped HashAgg Path where possible
7922 : */
7923 3206 : if (can_hash && cheapest_total_path != NULL)
7924 : {
7925 : /* Checked above */
7926 : Assert(parse->hasAggs || parse->groupClause);
7927 :
7928 713 : add_path(partially_grouped_rel, (Path *)
7929 713 : create_agg_path(root,
7930 : partially_grouped_rel,
7931 : cheapest_total_path,
7932 713 : partially_grouped_rel->reltarget,
7933 : AGG_HASHED,
7934 : AGGSPLIT_INITIAL_SERIAL,
7935 : root->processed_groupClause,
7936 : NIL,
7937 : agg_partial_costs,
7938 : dNumPartialGroups));
7939 : }
7940 :
7941 : /*
7942 : * Now add a partially-grouped HashAgg partial Path where possible
7943 : */
7944 3206 : if (can_hash && cheapest_partial_path != NULL)
7945 : {
7946 1434 : add_partial_path(partially_grouped_rel, (Path *)
7947 1434 : create_agg_path(root,
7948 : partially_grouped_rel,
7949 : cheapest_partial_path,
7950 1434 : partially_grouped_rel->reltarget,
7951 : AGG_HASHED,
7952 : AGGSPLIT_INITIAL_SERIAL,
7953 : root->processed_groupClause,
7954 : NIL,
7955 : agg_partial_costs,
7956 : dNumPartialPartialGroups));
7957 : }
7958 :
7959 : /*
7960 : * Add any partially aggregated paths generated by eager aggregation to
7961 : * the new upper relation after applying projection steps as needed.
7962 : */
7963 3206 : if (eager_agg_rel)
7964 : {
7965 : /* Add the paths */
7966 1894 : foreach(lc, eager_agg_rel->pathlist)
7967 : {
7968 1167 : Path *path = (Path *) lfirst(lc);
7969 :
7970 : /* Shouldn't have any parameterized paths anymore */
7971 : Assert(path->param_info == NULL);
7972 :
7973 1167 : path = (Path *) create_projection_path(root,
7974 : partially_grouped_rel,
7975 : path,
7976 1167 : partially_grouped_rel->reltarget);
7977 :
7978 1167 : add_path(partially_grouped_rel, path);
7979 : }
7980 :
7981 : /*
7982 : * Likewise add the partial paths, but only if parallelism is possible
7983 : * for partially_grouped_rel.
7984 : */
7985 727 : if (partially_grouped_rel->consider_parallel)
7986 : {
7987 1690 : foreach(lc, eager_agg_rel->partial_pathlist)
7988 : {
7989 1010 : Path *path = (Path *) lfirst(lc);
7990 :
7991 : /* Shouldn't have any parameterized paths anymore */
7992 : Assert(path->param_info == NULL);
7993 :
7994 1010 : path = (Path *) create_projection_path(root,
7995 : partially_grouped_rel,
7996 : path,
7997 1010 : partially_grouped_rel->reltarget);
7998 :
7999 1010 : add_partial_path(partially_grouped_rel, path);
8000 : }
8001 : }
8002 : }
8003 :
8004 : /*
8005 : * If there is an FDW that's responsible for all baserels of the query,
8006 : * let it consider adding partially grouped ForeignPaths.
8007 : */
8008 3206 : if (partially_grouped_rel->fdwroutine &&
8009 3 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
8010 : {
8011 3 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
8012 :
8013 3 : fdwroutine->GetForeignUpperPaths(root,
8014 : UPPERREL_PARTIAL_GROUP_AGG,
8015 : input_rel, partially_grouped_rel,
8016 : extra);
8017 : }
8018 :
8019 3206 : return partially_grouped_rel;
8020 : }
8021 :
8022 : /*
8023 : * make_ordered_path
8024 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
8025 : * return NULL if it doesn't make sense to generate an ordered path in
8026 : * this case.
8027 : */
8028 : static Path *
8029 49652 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
8030 : Path *cheapest_path, List *pathkeys, double limit_tuples)
8031 : {
8032 : bool is_sorted;
8033 : int presorted_keys;
8034 :
8035 49652 : is_sorted = pathkeys_count_contained_in(pathkeys,
8036 : path->pathkeys,
8037 : &presorted_keys);
8038 :
8039 49652 : if (!is_sorted)
8040 : {
8041 : /*
8042 : * Try at least sorting the cheapest path and also try incrementally
8043 : * sorting any path which is partially sorted already (no need to deal
8044 : * with paths which have presorted keys when incremental sort is
8045 : * disabled unless it's the cheapest input path).
8046 : */
8047 13176 : if (path != cheapest_path &&
8048 2157 : (presorted_keys == 0 || !enable_incremental_sort))
8049 1022 : return NULL;
8050 :
8051 : /*
8052 : * We've no need to consider both a sort and incremental sort. We'll
8053 : * just do a sort if there are no presorted keys and an incremental
8054 : * sort when there are presorted keys.
8055 : */
8056 12154 : if (presorted_keys == 0 || !enable_incremental_sort)
8057 10873 : path = (Path *) create_sort_path(root,
8058 : rel,
8059 : path,
8060 : pathkeys,
8061 : limit_tuples);
8062 : else
8063 1281 : path = (Path *) create_incremental_sort_path(root,
8064 : rel,
8065 : path,
8066 : pathkeys,
8067 : presorted_keys,
8068 : limit_tuples);
8069 : }
8070 :
8071 48630 : return path;
8072 : }
8073 :
8074 : /*
8075 : * Generate Gather and Gather Merge paths for a grouping relation or partial
8076 : * grouping relation.
8077 : *
8078 : * generate_useful_gather_paths does most of the work, but we also consider a
8079 : * special case: we could try sorting the data by the group_pathkeys and then
8080 : * applying Gather Merge.
8081 : *
8082 : * NB: This function shouldn't be used for anything other than a grouped or
8083 : * partially grouped relation not only because of the fact that it explicitly
8084 : * references group_pathkeys but we pass "true" as the third argument to
8085 : * generate_useful_gather_paths().
8086 : */
8087 : static void
8088 2590 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
8089 : {
8090 : ListCell *lc;
8091 : Path *cheapest_partial_path;
8092 : List *groupby_pathkeys;
8093 :
8094 : /*
8095 : * This occurs after any partial aggregation has taken place, so trim off
8096 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
8097 : */
8098 2590 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
8099 15 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
8100 : root->num_groupby_pathkeys);
8101 : else
8102 2575 : groupby_pathkeys = root->group_pathkeys;
8103 :
8104 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
8105 2590 : generate_useful_gather_paths(root, rel, true);
8106 :
8107 2590 : cheapest_partial_path = linitial(rel->partial_pathlist);
8108 :
8109 : /* XXX Shouldn't this also consider the group-key-reordering? */
8110 6152 : foreach(lc, rel->partial_pathlist)
8111 : {
8112 3562 : Path *path = (Path *) lfirst(lc);
8113 : bool is_sorted;
8114 : int presorted_keys;
8115 : double total_groups;
8116 :
8117 3562 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
8118 : path->pathkeys,
8119 : &presorted_keys);
8120 :
8121 3562 : if (is_sorted)
8122 2329 : continue;
8123 :
8124 : /*
8125 : * Try at least sorting the cheapest path and also try incrementally
8126 : * sorting any path which is partially sorted already (no need to deal
8127 : * with paths which have presorted keys when incremental sort is
8128 : * disabled unless it's the cheapest input path).
8129 : */
8130 1233 : if (path != cheapest_partial_path &&
8131 0 : (presorted_keys == 0 || !enable_incremental_sort))
8132 0 : continue;
8133 :
8134 : /*
8135 : * We've no need to consider both a sort and incremental sort. We'll
8136 : * just do a sort if there are no presorted keys and an incremental
8137 : * sort when there are presorted keys.
8138 : */
8139 1233 : if (presorted_keys == 0 || !enable_incremental_sort)
8140 1233 : path = (Path *) create_sort_path(root, rel, path,
8141 : groupby_pathkeys,
8142 : -1.0);
8143 : else
8144 0 : path = (Path *) create_incremental_sort_path(root,
8145 : rel,
8146 : path,
8147 : groupby_pathkeys,
8148 : presorted_keys,
8149 : -1.0);
8150 1233 : total_groups = compute_gather_rows(path);
8151 : path = (Path *)
8152 1233 : create_gather_merge_path(root,
8153 : rel,
8154 : path,
8155 1233 : rel->reltarget,
8156 : groupby_pathkeys,
8157 : NULL,
8158 : &total_groups);
8159 :
8160 1233 : add_path(rel, path);
8161 : }
8162 2590 : }
8163 :
8164 : /*
8165 : * can_partial_agg
8166 : *
8167 : * Determines whether or not partial grouping and/or aggregation is possible.
8168 : * Returns true when possible, false otherwise.
8169 : */
8170 : static bool
8171 34062 : can_partial_agg(PlannerInfo *root)
8172 : {
8173 34062 : Query *parse = root->parse;
8174 :
8175 34062 : if (!parse->hasAggs && parse->groupClause == NIL)
8176 : {
8177 : /*
8178 : * We don't know how to do parallel aggregation unless we have either
8179 : * some aggregates or a grouping clause.
8180 : */
8181 0 : return false;
8182 : }
8183 34062 : else if (parse->groupingSets)
8184 : {
8185 : /* We don't know how to do grouping sets in parallel. */
8186 886 : return false;
8187 : }
8188 33176 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
8189 : {
8190 : /* Insufficient support for partial mode. */
8191 3238 : return false;
8192 : }
8193 :
8194 : /* Everything looks good. */
8195 29938 : return true;
8196 : }
8197 :
8198 : /*
8199 : * apply_scanjoin_target_to_paths
8200 : *
8201 : * Adjust the final scan/join relation, and recursively all of its children,
8202 : * to generate the final scan/join target. It would be more correct to model
8203 : * this as a separate planning step with a new RelOptInfo at the toplevel and
8204 : * for each child relation, but doing it this way is noticeably cheaper.
8205 : * Maybe that problem can be solved at some point, but for now we do this.
8206 : *
8207 : * If tlist_same_exprs is true, then the scan/join target to be applied has
8208 : * the same expressions as the existing reltarget, so we need only insert the
8209 : * appropriate sortgroupref information. By avoiding the creation of
8210 : * projection paths we save effort both immediately and at plan creation time.
8211 : */
8212 : static void
8213 413355 : apply_scanjoin_target_to_paths(PlannerInfo *root,
8214 : RelOptInfo *rel,
8215 : List *scanjoin_targets,
8216 : List *scanjoin_targets_contain_srfs,
8217 : bool scanjoin_target_parallel_safe,
8218 : bool tlist_same_exprs)
8219 : {
8220 413355 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
8221 : PathTarget *scanjoin_target;
8222 : ListCell *lc;
8223 :
8224 : /* This recurses, so be paranoid. */
8225 413355 : check_stack_depth();
8226 :
8227 : /*
8228 : * If the rel only has Append and MergeAppend paths, we want to drop its
8229 : * existing paths and generate new ones. This function would still be
8230 : * correct if we kept the existing paths: we'd modify them to generate the
8231 : * correct target above the partitioning Append, and then they'd compete
8232 : * on cost with paths generating the target below the Append. However, in
8233 : * our current cost model the latter way is always the same or cheaper
8234 : * cost, so modifying the existing paths would just be useless work.
8235 : * Moreover, when the cost is the same, varying roundoff errors might
8236 : * sometimes allow an existing path to be picked, resulting in undesirable
8237 : * cross-platform plan variations. So we drop old paths and thereby force
8238 : * the work to be done below the Append.
8239 : *
8240 : * However, there are several cases when this optimization is not safe. If
8241 : * the rel isn't partitioned, then none of the paths will be Append or
8242 : * MergeAppend paths, so we should definitely not do this. If it is
8243 : * partitioned but is a joinrel, it may have Append and MergeAppend paths,
8244 : * but it can also have join paths that we can't afford to discard.
8245 : *
8246 : * Some care is needed, because we have to allow
8247 : * generate_useful_gather_paths to see the old partial paths in the next
8248 : * stanza. Hence, zap the main pathlist here, then allow
8249 : * generate_useful_gather_paths to add path(s) to the main list, and
8250 : * finally zap the partial pathlist.
8251 : */
8252 413355 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8253 8719 : rel->pathlist = NIL;
8254 :
8255 : /*
8256 : * If the scan/join target is not parallel-safe, partial paths cannot
8257 : * generate it.
8258 : */
8259 413355 : if (!scanjoin_target_parallel_safe)
8260 : {
8261 : /*
8262 : * Since we can't generate the final scan/join target in parallel
8263 : * workers, this is our last opportunity to use any partial paths that
8264 : * exist; so build Gather path(s) that use them and emit whatever the
8265 : * current reltarget is. We don't do this in the case where the
8266 : * target is parallel-safe, since we will be able to generate superior
8267 : * paths by doing it after the final scan/join target has been
8268 : * applied.
8269 : */
8270 53769 : generate_useful_gather_paths(root, rel, false);
8271 :
8272 : /* Can't use parallel query above this level. */
8273 53769 : rel->partial_pathlist = NIL;
8274 53769 : rel->consider_parallel = false;
8275 : }
8276 :
8277 : /* Finish dropping old paths for a partitioned rel, per comment above */
8278 413355 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8279 8719 : rel->partial_pathlist = NIL;
8280 :
8281 : /* Extract SRF-free scan/join target. */
8282 413355 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
8283 :
8284 : /*
8285 : * Apply the SRF-free scan/join target to each existing path.
8286 : *
8287 : * If the tlist exprs are the same, we can just inject the sortgroupref
8288 : * information into the existing pathtargets. Otherwise, replace each
8289 : * path with a projection path that generates the SRF-free scan/join
8290 : * target. This can't change the ordering of paths within rel->pathlist,
8291 : * so we just modify the list in place.
8292 : */
8293 863406 : foreach(lc, rel->pathlist)
8294 : {
8295 450051 : Path *subpath = (Path *) lfirst(lc);
8296 :
8297 : /* Shouldn't have any parameterized paths anymore */
8298 : Assert(subpath->param_info == NULL);
8299 :
8300 450051 : if (tlist_same_exprs)
8301 162995 : subpath->pathtarget->sortgrouprefs =
8302 162995 : scanjoin_target->sortgrouprefs;
8303 : else
8304 : {
8305 : Path *newpath;
8306 :
8307 287056 : newpath = (Path *) create_projection_path(root, rel, subpath,
8308 : scanjoin_target);
8309 287056 : lfirst(lc) = newpath;
8310 : }
8311 : }
8312 :
8313 : /* Likewise adjust the targets for any partial paths. */
8314 433613 : foreach(lc, rel->partial_pathlist)
8315 : {
8316 20258 : Path *subpath = (Path *) lfirst(lc);
8317 :
8318 : /* Shouldn't have any parameterized paths anymore */
8319 : Assert(subpath->param_info == NULL);
8320 :
8321 20258 : if (tlist_same_exprs)
8322 16185 : subpath->pathtarget->sortgrouprefs =
8323 16185 : scanjoin_target->sortgrouprefs;
8324 : else
8325 : {
8326 : Path *newpath;
8327 :
8328 4073 : newpath = (Path *) create_projection_path(root, rel, subpath,
8329 : scanjoin_target);
8330 4073 : lfirst(lc) = newpath;
8331 : }
8332 : }
8333 :
8334 : /*
8335 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8336 : * atop each existing path. (Note that this function doesn't look at the
8337 : * cheapest-path fields, which is a good thing because they're bogus right
8338 : * now.)
8339 : */
8340 413355 : if (root->parse->hasTargetSRFs)
8341 10366 : adjust_paths_for_srfs(root, rel,
8342 : scanjoin_targets,
8343 : scanjoin_targets_contain_srfs);
8344 :
8345 : /*
8346 : * Update the rel's target to be the final (with SRFs) scan/join target.
8347 : * This now matches the actual output of all the paths, and we might get
8348 : * confused in createplan.c if they don't agree. We must do this now so
8349 : * that any append paths made in the next part will use the correct
8350 : * pathtarget (cf. create_append_path).
8351 : *
8352 : * Note that this is also necessary if GetForeignUpperPaths() gets called
8353 : * on the final scan/join relation or on any of its children, since the
8354 : * FDW might look at the rel's target to create ForeignPaths.
8355 : */
8356 413355 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8357 :
8358 : /*
8359 : * If the relation is partitioned, recursively apply the scan/join target
8360 : * to all partitions, and generate brand-new Append paths in which the
8361 : * scan/join target is computed below the Append rather than above it.
8362 : * Since Append is not projection-capable, that might save a separate
8363 : * Result node, and it also is important for partitionwise aggregate.
8364 : */
8365 413355 : if (rel_is_partitioned)
8366 : {
8367 10022 : List *live_children = NIL;
8368 : int i;
8369 :
8370 : /* Adjust each partition. */
8371 10022 : i = -1;
8372 30106 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8373 : {
8374 20084 : RelOptInfo *child_rel = rel->part_rels[i];
8375 : AppendRelInfo **appinfos;
8376 : int nappinfos;
8377 20084 : List *child_scanjoin_targets = NIL;
8378 :
8379 : Assert(child_rel != NULL);
8380 :
8381 : /* Dummy children can be ignored. */
8382 20084 : if (IS_DUMMY_REL(child_rel))
8383 40 : continue;
8384 :
8385 : /* Translate scan/join targets for this child. */
8386 20044 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8387 : &nappinfos);
8388 40088 : foreach(lc, scanjoin_targets)
8389 : {
8390 20044 : PathTarget *target = lfirst_node(PathTarget, lc);
8391 :
8392 20044 : target = copy_pathtarget(target);
8393 20044 : target->exprs = (List *)
8394 20044 : adjust_appendrel_attrs(root,
8395 20044 : (Node *) target->exprs,
8396 : nappinfos, appinfos);
8397 20044 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8398 : target);
8399 : }
8400 20044 : pfree(appinfos);
8401 :
8402 : /* Recursion does the real work. */
8403 20044 : apply_scanjoin_target_to_paths(root, child_rel,
8404 : child_scanjoin_targets,
8405 : scanjoin_targets_contain_srfs,
8406 : scanjoin_target_parallel_safe,
8407 : tlist_same_exprs);
8408 :
8409 : /* Save non-dummy children for Append paths. */
8410 20044 : if (!IS_DUMMY_REL(child_rel))
8411 20044 : live_children = lappend(live_children, child_rel);
8412 : }
8413 :
8414 : /* Build new paths for this relation by appending child paths. */
8415 10022 : add_paths_to_append_rel(root, rel, live_children);
8416 : }
8417 :
8418 : /*
8419 : * Consider generating Gather or Gather Merge paths. We must only do this
8420 : * if the relation is parallel safe, and we don't do it for child rels to
8421 : * avoid creating multiple Gather nodes within the same plan. We must do
8422 : * this after all paths have been generated and before set_cheapest, since
8423 : * one of the generated paths may turn out to be the cheapest one.
8424 : */
8425 413355 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
8426 140081 : generate_useful_gather_paths(root, rel, false);
8427 :
8428 : /*
8429 : * Reassess which paths are the cheapest, now that we've potentially added
8430 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8431 : * this relation.
8432 : */
8433 413355 : set_cheapest(rel);
8434 413355 : }
8435 :
8436 : /*
8437 : * create_partitionwise_grouping_paths
8438 : *
8439 : * If the partition keys of input relation are part of the GROUP BY clause, all
8440 : * the rows belonging to a given group come from a single partition. This
8441 : * allows aggregation/grouping over a partitioned relation to be broken down
8442 : * into aggregation/grouping on each partition. This should be no worse, and
8443 : * often better, than the normal approach.
8444 : *
8445 : * However, if the GROUP BY clause does not contain all the partition keys,
8446 : * rows from a given group may be spread across multiple partitions. In that
8447 : * case, we perform partial aggregation for each group, append the results,
8448 : * and then finalize aggregation. This is less certain to win than the
8449 : * previous case. It may win if the PartialAggregate stage greatly reduces
8450 : * the number of groups, because fewer rows will pass through the Append node.
8451 : * It may lose if we have lots of small groups.
8452 : */
8453 : static void
8454 685 : create_partitionwise_grouping_paths(PlannerInfo *root,
8455 : RelOptInfo *input_rel,
8456 : RelOptInfo *grouped_rel,
8457 : RelOptInfo *partially_grouped_rel,
8458 : const AggClauseCosts *agg_costs,
8459 : grouping_sets_data *gd,
8460 : PartitionwiseAggregateType patype,
8461 : GroupPathExtraData *extra)
8462 : {
8463 685 : List *grouped_live_children = NIL;
8464 685 : List *partially_grouped_live_children = NIL;
8465 685 : PathTarget *target = grouped_rel->reltarget;
8466 685 : bool partial_grouping_valid = true;
8467 : int i;
8468 :
8469 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8470 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8471 : partially_grouped_rel != NULL);
8472 :
8473 : /* Add paths for partitionwise aggregation/grouping. */
8474 685 : i = -1;
8475 2480 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8476 : {
8477 1795 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8478 : PathTarget *child_target;
8479 : AppendRelInfo **appinfos;
8480 : int nappinfos;
8481 : GroupPathExtraData child_extra;
8482 : RelOptInfo *child_grouped_rel;
8483 : RelOptInfo *child_partially_grouped_rel;
8484 :
8485 : Assert(child_input_rel != NULL);
8486 :
8487 : /* Dummy children can be ignored. */
8488 1795 : if (IS_DUMMY_REL(child_input_rel))
8489 0 : continue;
8490 :
8491 1795 : child_target = copy_pathtarget(target);
8492 :
8493 : /*
8494 : * Copy the given "extra" structure as is and then override the
8495 : * members specific to this child.
8496 : */
8497 1795 : memcpy(&child_extra, extra, sizeof(child_extra));
8498 :
8499 1795 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8500 : &nappinfos);
8501 :
8502 1795 : child_target->exprs = (List *)
8503 1795 : adjust_appendrel_attrs(root,
8504 1795 : (Node *) target->exprs,
8505 : nappinfos, appinfos);
8506 :
8507 : /* Translate havingQual and targetList. */
8508 1795 : child_extra.havingQual = (Node *)
8509 : adjust_appendrel_attrs(root,
8510 : extra->havingQual,
8511 : nappinfos, appinfos);
8512 1795 : child_extra.targetList = (List *)
8513 1795 : adjust_appendrel_attrs(root,
8514 1795 : (Node *) extra->targetList,
8515 : nappinfos, appinfos);
8516 :
8517 : /*
8518 : * extra->patype was the value computed for our parent rel; patype is
8519 : * the value for this relation. For the child, our value is its
8520 : * parent rel's value.
8521 : */
8522 1795 : child_extra.patype = patype;
8523 :
8524 : /*
8525 : * Create grouping relation to hold fully aggregated grouping and/or
8526 : * aggregation paths for the child.
8527 : */
8528 1795 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8529 : child_target,
8530 1795 : extra->target_parallel_safe,
8531 : child_extra.havingQual);
8532 :
8533 : /* Create grouping paths for this child relation. */
8534 1795 : create_ordinary_grouping_paths(root, child_input_rel,
8535 : child_grouped_rel,
8536 : agg_costs, gd, &child_extra,
8537 : &child_partially_grouped_rel);
8538 :
8539 1795 : if (child_partially_grouped_rel)
8540 : {
8541 : partially_grouped_live_children =
8542 1283 : lappend(partially_grouped_live_children,
8543 : child_partially_grouped_rel);
8544 : }
8545 : else
8546 512 : partial_grouping_valid = false;
8547 :
8548 1795 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8549 : {
8550 1082 : set_cheapest(child_grouped_rel);
8551 1082 : grouped_live_children = lappend(grouped_live_children,
8552 : child_grouped_rel);
8553 : }
8554 :
8555 1795 : pfree(appinfos);
8556 : }
8557 :
8558 : /*
8559 : * Try to create append paths for partially grouped children. For full
8560 : * partitionwise aggregation, we might have paths in the partial_pathlist
8561 : * if parallel aggregation is possible. For partial partitionwise
8562 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8563 : *
8564 : * NB: We must have a partially grouped path for every child in order to
8565 : * generate a partially grouped path for this relation.
8566 : */
8567 685 : if (partially_grouped_rel && partial_grouping_valid)
8568 : {
8569 : Assert(partially_grouped_live_children != NIL);
8570 :
8571 501 : add_paths_to_append_rel(root, partially_grouped_rel,
8572 : partially_grouped_live_children);
8573 : }
8574 :
8575 : /* If possible, create append paths for fully grouped children. */
8576 685 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8577 : {
8578 : Assert(grouped_live_children != NIL);
8579 :
8580 404 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8581 : }
8582 685 : }
8583 :
8584 : /*
8585 : * group_by_has_partkey
8586 : *
8587 : * Returns true if all the partition keys of the given relation are part of
8588 : * the GROUP BY clauses, including having matching collation, false otherwise.
8589 : */
8590 : static bool
8591 640 : group_by_has_partkey(RelOptInfo *input_rel,
8592 : List *targetList,
8593 : List *groupClause)
8594 : {
8595 640 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8596 640 : int cnt = 0;
8597 : int partnatts;
8598 :
8599 : /* Input relation should be partitioned. */
8600 : Assert(input_rel->part_scheme);
8601 :
8602 : /* Rule out early, if there are no partition keys present. */
8603 640 : if (!input_rel->partexprs)
8604 0 : return false;
8605 :
8606 640 : partnatts = input_rel->part_scheme->partnatts;
8607 :
8608 1074 : for (cnt = 0; cnt < partnatts; cnt++)
8609 : {
8610 670 : List *partexprs = input_rel->partexprs[cnt];
8611 : ListCell *lc;
8612 670 : bool found = false;
8613 :
8614 1001 : foreach(lc, partexprs)
8615 : {
8616 : ListCell *lg;
8617 775 : Expr *partexpr = lfirst(lc);
8618 775 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8619 :
8620 1206 : foreach(lg, groupexprs)
8621 : {
8622 875 : Expr *groupexpr = lfirst(lg);
8623 875 : Oid groupcoll = exprCollation((Node *) groupexpr);
8624 :
8625 : /*
8626 : * Note: we can assume there is at most one RelabelType node;
8627 : * eval_const_expressions() will have simplified if more than
8628 : * one.
8629 : */
8630 875 : if (IsA(groupexpr, RelabelType))
8631 20 : groupexpr = ((RelabelType *) groupexpr)->arg;
8632 :
8633 875 : if (equal(groupexpr, partexpr))
8634 : {
8635 : /*
8636 : * Reject a match if the grouping collation does not match
8637 : * the partitioning collation.
8638 : */
8639 444 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8640 : partcoll != groupcoll)
8641 10 : return false;
8642 :
8643 434 : found = true;
8644 434 : break;
8645 : }
8646 : }
8647 :
8648 765 : if (found)
8649 434 : break;
8650 : }
8651 :
8652 : /*
8653 : * If none of the partition key expressions match with any of the
8654 : * GROUP BY expression, return false.
8655 : */
8656 660 : if (!found)
8657 226 : return false;
8658 : }
8659 :
8660 404 : return true;
8661 : }
8662 :
8663 : /*
8664 : * generate_setop_child_grouplist
8665 : * Build a SortGroupClause list defining the sort/grouping properties
8666 : * of the child of a set operation.
8667 : *
8668 : * This is similar to generate_setop_grouplist() but differs as the setop
8669 : * child query's targetlist entries may already have a tleSortGroupRef
8670 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8671 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8672 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8673 : * any of the columns in the targetlist don't match to the setop's colTypes
8674 : * then we return an empty list. This may leave some TLEs with unreferenced
8675 : * ressortgroupref markings, but that's harmless.
8676 : */
8677 : static List *
8678 10483 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8679 : {
8680 10483 : List *grouplist = copyObject(op->groupClauses);
8681 : ListCell *lg;
8682 : ListCell *lt;
8683 : ListCell *ct;
8684 :
8685 10483 : lg = list_head(grouplist);
8686 10483 : ct = list_head(op->colTypes);
8687 40763 : foreach(lt, targetlist)
8688 : {
8689 30525 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8690 : SortGroupClause *sgc;
8691 : Oid coltype;
8692 :
8693 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8694 30525 : if (tle->resjunk)
8695 0 : continue;
8696 :
8697 : /*
8698 : * We expect every non-resjunk target to have a SortGroupClause and
8699 : * colTypes.
8700 : */
8701 : Assert(lg != NULL);
8702 : Assert(ct != NULL);
8703 30525 : sgc = (SortGroupClause *) lfirst(lg);
8704 30525 : coltype = lfirst_oid(ct);
8705 :
8706 : /* reject if target type isn't the same as the setop target type */
8707 30525 : if (coltype != exprType((Node *) tle->expr))
8708 245 : return NIL;
8709 :
8710 30280 : lg = lnext(grouplist, lg);
8711 30280 : ct = lnext(op->colTypes, ct);
8712 :
8713 : /* assign a tleSortGroupRef, or reuse the existing one */
8714 30280 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8715 : }
8716 :
8717 : Assert(lg == NULL);
8718 : Assert(ct == NULL);
8719 :
8720 10238 : return grouplist;
8721 : }
8722 :
8723 : /*
8724 : * create_unique_paths
8725 : * Build a new RelOptInfo containing Paths that represent elimination of
8726 : * distinct rows from the input data. Distinct-ness is defined according to
8727 : * the needs of the semijoin represented by sjinfo. If it is not possible
8728 : * to identify how to make the data unique, NULL is returned.
8729 : *
8730 : * If used at all, this is likely to be called repeatedly on the same rel,
8731 : * so we cache the result.
8732 : */
8733 : RelOptInfo *
8734 6924 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8735 : {
8736 : RelOptInfo *unique_rel;
8737 6924 : List *sortPathkeys = NIL;
8738 6924 : List *groupClause = NIL;
8739 : MemoryContext oldcontext;
8740 :
8741 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8742 : Assert(sjinfo->jointype == JOIN_SEMI);
8743 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8744 :
8745 : /* If result already cached, return it */
8746 6924 : if (rel->unique_rel)
8747 1196 : return rel->unique_rel;
8748 :
8749 : /* If it's not possible to unique-ify, return NULL */
8750 5728 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8751 104 : return NULL;
8752 :
8753 : /*
8754 : * Punt if this is a child relation and we failed to build a unique-ified
8755 : * relation for its parent. This can happen if all the RHS columns were
8756 : * found to be equated to constants when unique-ifying the parent table,
8757 : * leaving no columns to unique-ify.
8758 : */
8759 5624 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8760 10 : return NULL;
8761 :
8762 : /*
8763 : * When called during GEQO join planning, we are in a short-lived memory
8764 : * context. We must make sure that the unique rel and any subsidiary data
8765 : * structures created for a baserel survive the GEQO cycle, else the
8766 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8767 : * are creating those for a joinrel during GEQO, we don't want them to
8768 : * clutter the main planning context. Upshot is that the best solution is
8769 : * to explicitly allocate memory in the same context the given RelOptInfo
8770 : * is in.
8771 : */
8772 5614 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8773 :
8774 5614 : unique_rel = makeNode(RelOptInfo);
8775 5614 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8776 :
8777 : /*
8778 : * clear path info
8779 : */
8780 5614 : unique_rel->pathlist = NIL;
8781 5614 : unique_rel->ppilist = NIL;
8782 5614 : unique_rel->partial_pathlist = NIL;
8783 5614 : unique_rel->cheapest_startup_path = NULL;
8784 5614 : unique_rel->cheapest_total_path = NULL;
8785 5614 : unique_rel->cheapest_parameterized_paths = NIL;
8786 :
8787 : /*
8788 : * Build the target list for the unique rel. We also build the pathkeys
8789 : * that represent the ordering requirements for the sort-based
8790 : * implementation, and the list of SortGroupClause nodes that represent
8791 : * the columns to be grouped on for the hash-based implementation.
8792 : *
8793 : * For a child rel, we can construct these fields from those of its
8794 : * parent.
8795 : */
8796 5614 : if (IS_OTHER_REL(rel))
8797 360 : {
8798 : PathTarget *child_unique_target;
8799 : PathTarget *parent_unique_target;
8800 :
8801 360 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8802 :
8803 360 : child_unique_target = copy_pathtarget(parent_unique_target);
8804 :
8805 : /* Translate the target expressions */
8806 360 : child_unique_target->exprs = (List *)
8807 360 : adjust_appendrel_attrs_multilevel(root,
8808 360 : (Node *) parent_unique_target->exprs,
8809 : rel,
8810 360 : rel->top_parent);
8811 :
8812 360 : unique_rel->reltarget = child_unique_target;
8813 :
8814 360 : sortPathkeys = rel->top_parent->unique_pathkeys;
8815 360 : groupClause = rel->top_parent->unique_groupclause;
8816 : }
8817 : else
8818 : {
8819 : List *newtlist;
8820 : int nextresno;
8821 5254 : List *sortList = NIL;
8822 : ListCell *lc1;
8823 : ListCell *lc2;
8824 :
8825 : /*
8826 : * The values we are supposed to unique-ify may be expressions in the
8827 : * variables of the input rel's targetlist. We have to add any such
8828 : * expressions to the unique rel's targetlist.
8829 : *
8830 : * To complicate matters, some of the values to be unique-ified may be
8831 : * known redundant by the EquivalenceClass machinery (e.g., because
8832 : * they have been equated to constants). There is no need to compare
8833 : * such values during unique-ification, and indeed we had better not
8834 : * try because the Vars involved may not have propagated as high as
8835 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8836 : * detect such cases, which is a tad inefficient but it doesn't seem
8837 : * worth building specialized infrastructure for this.
8838 : */
8839 5254 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8840 5254 : nextresno = list_length(newtlist) + 1;
8841 :
8842 10709 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8843 : {
8844 5455 : Expr *uniqexpr = lfirst(lc1);
8845 5455 : Oid in_oper = lfirst_oid(lc2);
8846 : Oid sortop;
8847 : TargetEntry *tle;
8848 5455 : bool made_tle = false;
8849 :
8850 5455 : tle = tlist_member(uniqexpr, newtlist);
8851 5455 : if (!tle)
8852 : {
8853 2752 : tle = makeTargetEntry(uniqexpr,
8854 : nextresno,
8855 : NULL,
8856 : false);
8857 2752 : newtlist = lappend(newtlist, tle);
8858 2752 : nextresno++;
8859 2752 : made_tle = true;
8860 : }
8861 :
8862 : /*
8863 : * Try to build an ORDER BY list to sort the input compatibly. We
8864 : * do this for each sortable clause even when the clauses are not
8865 : * all sortable, so that we can detect clauses that are redundant
8866 : * according to the pathkey machinery.
8867 : */
8868 5455 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8869 5455 : if (OidIsValid(sortop))
8870 : {
8871 : Oid eqop;
8872 : SortGroupClause *sortcl;
8873 :
8874 : /*
8875 : * The Unique node will need equality operators. Normally
8876 : * these are the same as the IN clause operators, but if those
8877 : * are cross-type operators then the equality operators are
8878 : * the ones for the IN clause operators' RHS datatype.
8879 : */
8880 5455 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8881 5455 : if (!OidIsValid(eqop)) /* shouldn't happen */
8882 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8883 : sortop);
8884 :
8885 5455 : sortcl = makeNode(SortGroupClause);
8886 5455 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8887 5455 : sortcl->eqop = eqop;
8888 5455 : sortcl->sortop = sortop;
8889 5455 : sortcl->reverse_sort = false;
8890 5455 : sortcl->nulls_first = false;
8891 5455 : sortcl->hashable = false; /* no need to make this accurate */
8892 5455 : sortList = lappend(sortList, sortcl);
8893 :
8894 : /*
8895 : * At each step, convert the SortGroupClause list to pathkey
8896 : * form. If the just-added SortGroupClause is redundant, the
8897 : * result will be shorter than the SortGroupClause list.
8898 : */
8899 5455 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8900 : newtlist);
8901 5455 : if (list_length(sortPathkeys) != list_length(sortList))
8902 : {
8903 : /* Drop the redundant SortGroupClause */
8904 1710 : sortList = list_delete_last(sortList);
8905 : Assert(list_length(sortPathkeys) == list_length(sortList));
8906 : /* Undo tlist addition, if we made one */
8907 1710 : if (made_tle)
8908 : {
8909 10 : newtlist = list_delete_last(newtlist);
8910 10 : nextresno--;
8911 : }
8912 : /* We need not consider this clause for hashing, either */
8913 1710 : continue;
8914 : }
8915 : }
8916 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8917 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8918 : in_oper);
8919 :
8920 3745 : if (sjinfo->semi_can_hash)
8921 : {
8922 : /* Create a GROUP BY list for the Agg node to use */
8923 : Oid eq_oper;
8924 : SortGroupClause *groupcl;
8925 :
8926 : /*
8927 : * Get the hashable equality operators for the Agg node to
8928 : * use. Normally these are the same as the IN clause
8929 : * operators, but if those are cross-type operators then the
8930 : * equality operators are the ones for the IN clause
8931 : * operators' RHS datatype.
8932 : */
8933 3745 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8934 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8935 : in_oper);
8936 :
8937 3745 : groupcl = makeNode(SortGroupClause);
8938 3745 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8939 3745 : groupcl->eqop = eq_oper;
8940 3745 : groupcl->sortop = sortop;
8941 3745 : groupcl->reverse_sort = false;
8942 3745 : groupcl->nulls_first = false;
8943 3745 : groupcl->hashable = true;
8944 3745 : groupClause = lappend(groupClause, groupcl);
8945 : }
8946 : }
8947 :
8948 : /*
8949 : * Done building the sortPathkeys and groupClause. But the
8950 : * sortPathkeys are bogus if not all the clauses were sortable.
8951 : */
8952 5254 : if (!sjinfo->semi_can_btree)
8953 0 : sortPathkeys = NIL;
8954 :
8955 : /*
8956 : * It can happen that all the RHS columns are equated to constants.
8957 : * We'd have to do something special to unique-ify in that case, and
8958 : * it's such an unlikely-in-the-real-world case that it's not worth
8959 : * the effort. So just punt if we found no columns to unique-ify.
8960 : */
8961 5254 : if (sortPathkeys == NIL && groupClause == NIL)
8962 : {
8963 1625 : MemoryContextSwitchTo(oldcontext);
8964 1625 : return NULL;
8965 : }
8966 :
8967 : /* Convert the required targetlist back to PathTarget form */
8968 3629 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8969 : }
8970 :
8971 : /* build unique paths based on input rel's pathlist */
8972 3989 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8973 : sjinfo, unique_rel);
8974 :
8975 : /* build unique paths based on input rel's partial_pathlist */
8976 3989 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8977 : sjinfo, unique_rel);
8978 :
8979 : /* Now choose the best path(s) */
8980 3989 : set_cheapest(unique_rel);
8981 :
8982 : /*
8983 : * There shouldn't be any partial paths for the unique relation;
8984 : * otherwise, we won't be able to properly guarantee uniqueness.
8985 : */
8986 : Assert(unique_rel->partial_pathlist == NIL);
8987 :
8988 : /* Cache the result */
8989 3989 : rel->unique_rel = unique_rel;
8990 3989 : rel->unique_pathkeys = sortPathkeys;
8991 3989 : rel->unique_groupclause = groupClause;
8992 :
8993 3989 : MemoryContextSwitchTo(oldcontext);
8994 :
8995 3989 : return unique_rel;
8996 : }
8997 :
8998 : /*
8999 : * create_final_unique_paths
9000 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
9001 : */
9002 : static void
9003 7121 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
9004 : List *sortPathkeys, List *groupClause,
9005 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
9006 : {
9007 7121 : Path *cheapest_input_path = input_rel->cheapest_total_path;
9008 :
9009 : /* Estimate number of output rows */
9010 7121 : unique_rel->rows = estimate_num_groups(root,
9011 : sjinfo->semi_rhs_exprs,
9012 : cheapest_input_path->rows,
9013 : NULL,
9014 : NULL);
9015 :
9016 : /* Consider sort-based implementations, if possible. */
9017 7121 : if (sjinfo->semi_can_btree)
9018 : {
9019 : ListCell *lc;
9020 :
9021 : /*
9022 : * Use any available suitably-sorted path as input, and also consider
9023 : * sorting the cheapest-total path and incremental sort on any paths
9024 : * with presorted keys.
9025 : *
9026 : * To save planning time, we ignore parameterized input paths unless
9027 : * they are the cheapest-total path.
9028 : */
9029 15520 : foreach(lc, input_rel->pathlist)
9030 : {
9031 8399 : Path *input_path = (Path *) lfirst(lc);
9032 : Path *path;
9033 : bool is_sorted;
9034 : int presorted_keys;
9035 :
9036 : /*
9037 : * Ignore parameterized paths that are not the cheapest-total
9038 : * path.
9039 : */
9040 8399 : if (input_path->param_info &&
9041 : input_path != cheapest_input_path)
9042 740 : continue;
9043 :
9044 7695 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
9045 : input_path->pathkeys,
9046 : &presorted_keys);
9047 :
9048 : /*
9049 : * Ignore paths that are not suitably or partially sorted, unless
9050 : * they are the cheapest total path (no need to deal with paths
9051 : * which have presorted keys when incremental sort is disabled).
9052 : */
9053 7695 : if (!is_sorted && input_path != cheapest_input_path &&
9054 76 : (presorted_keys == 0 || !enable_incremental_sort))
9055 36 : continue;
9056 :
9057 : /*
9058 : * Make a separate ProjectionPath in case we need a Result node.
9059 : */
9060 7659 : path = (Path *) create_projection_path(root,
9061 : unique_rel,
9062 : input_path,
9063 7659 : unique_rel->reltarget);
9064 :
9065 7659 : if (!is_sorted)
9066 : {
9067 : /*
9068 : * We've no need to consider both a sort and incremental sort.
9069 : * We'll just do a sort if there are no presorted keys and an
9070 : * incremental sort when there are presorted keys.
9071 : */
9072 3972 : if (presorted_keys == 0 || !enable_incremental_sort)
9073 3932 : path = (Path *) create_sort_path(root,
9074 : unique_rel,
9075 : path,
9076 : sortPathkeys,
9077 : -1.0);
9078 : else
9079 40 : path = (Path *) create_incremental_sort_path(root,
9080 : unique_rel,
9081 : path,
9082 : sortPathkeys,
9083 : presorted_keys,
9084 : -1.0);
9085 : }
9086 :
9087 7659 : path = (Path *) create_unique_path(root, unique_rel, path,
9088 : list_length(sortPathkeys),
9089 : unique_rel->rows);
9090 :
9091 7659 : add_path(unique_rel, path);
9092 : }
9093 : }
9094 :
9095 : /* Consider hash-based implementation, if possible. */
9096 7121 : if (sjinfo->semi_can_hash)
9097 : {
9098 : Path *path;
9099 :
9100 : /*
9101 : * Make a separate ProjectionPath in case we need a Result node.
9102 : */
9103 7121 : path = (Path *) create_projection_path(root,
9104 : unique_rel,
9105 : cheapest_input_path,
9106 7121 : unique_rel->reltarget);
9107 :
9108 7121 : path = (Path *) create_agg_path(root,
9109 : unique_rel,
9110 : path,
9111 : cheapest_input_path->pathtarget,
9112 : AGG_HASHED,
9113 : AGGSPLIT_SIMPLE,
9114 : groupClause,
9115 : NIL,
9116 : NULL,
9117 : unique_rel->rows);
9118 :
9119 7121 : add_path(unique_rel, path);
9120 : }
9121 7121 : }
9122 :
9123 : /*
9124 : * create_partial_unique_paths
9125 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
9126 : */
9127 : static void
9128 3989 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
9129 : List *sortPathkeys, List *groupClause,
9130 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
9131 : {
9132 : RelOptInfo *partial_unique_rel;
9133 : Path *cheapest_partial_path;
9134 :
9135 : /* nothing to do when there are no partial paths in the input rel */
9136 3989 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
9137 857 : return;
9138 :
9139 : /*
9140 : * nothing to do if there's anything in the targetlist that's
9141 : * parallel-restricted.
9142 : */
9143 3132 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
9144 0 : return;
9145 :
9146 3132 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
9147 :
9148 3132 : partial_unique_rel = makeNode(RelOptInfo);
9149 3132 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
9150 :
9151 : /*
9152 : * clear path info
9153 : */
9154 3132 : partial_unique_rel->pathlist = NIL;
9155 3132 : partial_unique_rel->ppilist = NIL;
9156 3132 : partial_unique_rel->partial_pathlist = NIL;
9157 3132 : partial_unique_rel->cheapest_startup_path = NULL;
9158 3132 : partial_unique_rel->cheapest_total_path = NULL;
9159 3132 : partial_unique_rel->cheapest_parameterized_paths = NIL;
9160 :
9161 : /* Estimate number of output rows */
9162 3132 : partial_unique_rel->rows = estimate_num_groups(root,
9163 : sjinfo->semi_rhs_exprs,
9164 : cheapest_partial_path->rows,
9165 : NULL,
9166 : NULL);
9167 3132 : partial_unique_rel->reltarget = unique_rel->reltarget;
9168 :
9169 : /* Consider sort-based implementations, if possible. */
9170 3132 : if (sjinfo->semi_can_btree)
9171 : {
9172 : ListCell *lc;
9173 :
9174 : /*
9175 : * Use any available suitably-sorted path as input, and also consider
9176 : * sorting the cheapest partial path and incremental sort on any paths
9177 : * with presorted keys.
9178 : */
9179 6518 : foreach(lc, input_rel->partial_pathlist)
9180 : {
9181 3386 : Path *input_path = (Path *) lfirst(lc);
9182 : Path *path;
9183 : bool is_sorted;
9184 : int presorted_keys;
9185 :
9186 3386 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
9187 : input_path->pathkeys,
9188 : &presorted_keys);
9189 :
9190 : /*
9191 : * Ignore paths that are not suitably or partially sorted, unless
9192 : * they are the cheapest partial path (no need to deal with paths
9193 : * which have presorted keys when incremental sort is disabled).
9194 : */
9195 3386 : if (!is_sorted && input_path != cheapest_partial_path &&
9196 0 : (presorted_keys == 0 || !enable_incremental_sort))
9197 0 : continue;
9198 :
9199 : /*
9200 : * Make a separate ProjectionPath in case we need a Result node.
9201 : */
9202 3386 : path = (Path *) create_projection_path(root,
9203 : partial_unique_rel,
9204 : input_path,
9205 3386 : partial_unique_rel->reltarget);
9206 :
9207 3386 : if (!is_sorted)
9208 : {
9209 : /*
9210 : * We've no need to consider both a sort and incremental sort.
9211 : * We'll just do a sort if there are no presorted keys and an
9212 : * incremental sort when there are presorted keys.
9213 : */
9214 3092 : if (presorted_keys == 0 || !enable_incremental_sort)
9215 3092 : path = (Path *) create_sort_path(root,
9216 : partial_unique_rel,
9217 : path,
9218 : sortPathkeys,
9219 : -1.0);
9220 : else
9221 0 : path = (Path *) create_incremental_sort_path(root,
9222 : partial_unique_rel,
9223 : path,
9224 : sortPathkeys,
9225 : presorted_keys,
9226 : -1.0);
9227 : }
9228 :
9229 3386 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
9230 : list_length(sortPathkeys),
9231 : partial_unique_rel->rows);
9232 :
9233 3386 : add_partial_path(partial_unique_rel, path);
9234 : }
9235 : }
9236 :
9237 : /* Consider hash-based implementation, if possible. */
9238 3132 : if (sjinfo->semi_can_hash)
9239 : {
9240 : Path *path;
9241 :
9242 : /*
9243 : * Make a separate ProjectionPath in case we need a Result node.
9244 : */
9245 3132 : path = (Path *) create_projection_path(root,
9246 : partial_unique_rel,
9247 : cheapest_partial_path,
9248 3132 : partial_unique_rel->reltarget);
9249 :
9250 3132 : path = (Path *) create_agg_path(root,
9251 : partial_unique_rel,
9252 : path,
9253 : cheapest_partial_path->pathtarget,
9254 : AGG_HASHED,
9255 : AGGSPLIT_SIMPLE,
9256 : groupClause,
9257 : NIL,
9258 : NULL,
9259 : partial_unique_rel->rows);
9260 :
9261 3132 : add_partial_path(partial_unique_rel, path);
9262 : }
9263 :
9264 3132 : if (partial_unique_rel->partial_pathlist != NIL)
9265 : {
9266 3132 : generate_useful_gather_paths(root, partial_unique_rel, true);
9267 3132 : set_cheapest(partial_unique_rel);
9268 :
9269 : /*
9270 : * Finally, create paths to unique-ify the final result. This step is
9271 : * needed to remove any duplicates due to combining rows from parallel
9272 : * workers.
9273 : */
9274 3132 : create_final_unique_paths(root, partial_unique_rel,
9275 : sortPathkeys, groupClause,
9276 : sjinfo, unique_rel);
9277 : }
9278 : }
9279 :
9280 : /*
9281 : * Choose a unique name for some subroot.
9282 : *
9283 : * Modifies glob->subplanNames to track names already used.
9284 : */
9285 : char *
9286 65361 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
9287 : {
9288 : unsigned n;
9289 :
9290 : /*
9291 : * If a numeric suffix is not required, then search the list of
9292 : * previously-assigned names for a match. If none is found, then we can
9293 : * use the provided name without modification.
9294 : */
9295 65361 : if (!always_number)
9296 : {
9297 19326 : bool found = false;
9298 :
9299 46567 : foreach_ptr(char, subplan_name, glob->subplanNames)
9300 : {
9301 12266 : if (strcmp(subplan_name, name) == 0)
9302 : {
9303 4351 : found = true;
9304 4351 : break;
9305 : }
9306 : }
9307 :
9308 19326 : if (!found)
9309 : {
9310 : /* pstrdup here is just to avoid cast-away-const */
9311 14975 : char *chosen_name = pstrdup(name);
9312 :
9313 14975 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
9314 14975 : return chosen_name;
9315 : }
9316 : }
9317 :
9318 : /*
9319 : * If a numeric suffix is required or if the un-suffixed name is already
9320 : * in use, then loop until we find a positive integer that produces a
9321 : * novel name.
9322 : */
9323 50386 : for (n = 1; true; ++n)
9324 43342 : {
9325 93728 : char *proposed_name = psprintf("%s_%u", name, n);
9326 93728 : bool found = false;
9327 :
9328 349315 : foreach_ptr(char, subplan_name, glob->subplanNames)
9329 : {
9330 205201 : if (strcmp(subplan_name, proposed_name) == 0)
9331 : {
9332 43342 : found = true;
9333 43342 : break;
9334 : }
9335 : }
9336 :
9337 93728 : if (!found)
9338 : {
9339 50386 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9340 50386 : return proposed_name;
9341 : }
9342 :
9343 43342 : pfree(proposed_name);
9344 : }
9345 : }
|
