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 333923 : planner(Query *parse, const char *query_string, int cursorOptions,
334 : ParamListInfo boundParams, ExplainState *es)
335 : {
336 : PlannedStmt *result;
337 :
338 333923 : if (planner_hook)
339 50758 : result = (*planner_hook) (parse, query_string, cursorOptions,
340 : boundParams, es);
341 : else
342 283165 : result = standard_planner(parse, query_string, cursorOptions,
343 : boundParams, es);
344 :
345 329892 : pgstat_report_plan_id(result->planId, false);
346 :
347 329892 : return result;
348 : }
349 :
350 : PlannedStmt *
351 333923 : 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 333923 : glob = makeNode(PlannerGlobal);
372 :
373 333923 : glob->boundParams = boundParams;
374 333923 : glob->subplans = NIL;
375 333923 : glob->subpaths = NIL;
376 333923 : glob->subroots = NIL;
377 333923 : glob->rewindPlanIDs = NULL;
378 333923 : glob->finalrtable = NIL;
379 333923 : glob->allRelids = NULL;
380 333923 : glob->prunableRelids = NULL;
381 333923 : glob->finalrteperminfos = NIL;
382 333923 : glob->finalrowmarks = NIL;
383 333923 : glob->resultRelations = NIL;
384 333923 : glob->appendRelations = NIL;
385 333923 : glob->partPruneInfos = NIL;
386 333923 : glob->relationOids = NIL;
387 333923 : glob->invalItems = NIL;
388 333923 : glob->paramExecTypes = NIL;
389 333923 : glob->lastPHId = 0;
390 333923 : glob->lastRowMarkId = 0;
391 333923 : glob->lastPlanNodeId = 0;
392 333923 : glob->transientPlan = false;
393 333923 : glob->dependsOnRole = false;
394 333923 : glob->partition_directory = NULL;
395 333923 : 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 333923 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
419 315648 : IsUnderPostmaster &&
420 315648 : parse->commandType == CMD_SELECT &&
421 254656 : !parse->hasModifyingCTE &&
422 254533 : max_parallel_workers_per_gather > 0 &&
423 253850 : !IsParallelWorker())
424 : {
425 : /* all the cheap tests pass, so scan the query tree */
426 253810 : glob->maxParallelHazard = max_parallel_hazard(parse);
427 253810 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
428 : }
429 : else
430 : {
431 : /* skip the query tree scan, just assume it's unsafe */
432 80113 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
433 80113 : 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 553217 : glob->parallelModeNeeded = glob->parallelModeOK &&
454 219294 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
455 :
456 : /* Determine what fraction of the plan is likely to be scanned */
457 333923 : 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 3238 : 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 3238 : if (tuple_fraction >= 1.0)
474 0 : tuple_fraction = 0.0;
475 3238 : 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 330685 : 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 333923 : glob->default_pgs_mask = PGS_APPEND | PGS_MERGE_APPEND | PGS_FOREIGNJOIN |
499 : PGS_GATHER | PGS_CONSIDER_NONPARTIAL;
500 333923 : if (enable_tidscan)
501 333923 : glob->default_pgs_mask |= PGS_TIDSCAN;
502 333923 : if (enable_seqscan)
503 316721 : glob->default_pgs_mask |= PGS_SEQSCAN;
504 333923 : if (enable_indexscan)
505 331676 : glob->default_pgs_mask |= PGS_INDEXSCAN | PGS_INDEXONLYSCAN;
506 333923 : if (enable_indexonlyscan)
507 332509 : glob->default_pgs_mask |= PGS_CONSIDER_INDEXONLY;
508 333923 : if (enable_bitmapscan)
509 324380 : glob->default_pgs_mask |= PGS_BITMAPSCAN;
510 333923 : if (enable_mergejoin)
511 : {
512 331943 : glob->default_pgs_mask |= PGS_MERGEJOIN_PLAIN;
513 331943 : if (enable_material)
514 331876 : glob->default_pgs_mask |= PGS_MERGEJOIN_MATERIALIZE;
515 : }
516 333923 : if (enable_nestloop)
517 : {
518 333619 : glob->default_pgs_mask |= PGS_NESTLOOP_PLAIN;
519 333619 : if (enable_material)
520 333283 : glob->default_pgs_mask |= PGS_NESTLOOP_MATERIALIZE;
521 333619 : if (enable_memoize)
522 333481 : glob->default_pgs_mask |= PGS_NESTLOOP_MEMOIZE;
523 : }
524 333923 : if (enable_hashjoin)
525 331624 : glob->default_pgs_mask |= PGS_HASHJOIN;
526 333923 : if (enable_gathermerge)
527 333923 : glob->default_pgs_mask |= PGS_GATHER_MERGE;
528 333923 : if (enable_partitionwise_join)
529 2083 : glob->default_pgs_mask |= PGS_CONSIDER_PARTITIONWISE;
530 :
531 : /* Allow plugins to take control after we've initialized "glob" */
532 333923 : if (planner_setup_hook)
533 89174 : (*planner_setup_hook) (glob, parse, query_string, cursorOptions,
534 : &tuple_fraction, es);
535 :
536 : /* primary planning entry point (may recurse for subqueries) */
537 333148 : 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 330162 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
542 330162 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
543 :
544 330162 : 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 329892 : 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 329892 : 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 329892 : if (glob->paramExecTypes != NIL)
630 : {
631 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
632 144499 : forboth(lp, glob->subplans, lr, glob->subroots)
633 : {
634 32250 : Plan *subplan = (Plan *) lfirst(lp);
635 32250 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
636 :
637 32250 : SS_finalize_plan(subroot, subplan);
638 : }
639 112249 : 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 329892 : 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 362142 : forboth(lp, glob->subplans, lr, glob->subroots)
652 : {
653 32250 : Plan *subplan = (Plan *) lfirst(lp);
654 32250 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
655 :
656 32250 : lfirst(lp) = set_plan_references(subroot, subplan);
657 : }
658 :
659 : /* build the PlannedStmt result */
660 329892 : result = makeNode(PlannedStmt);
661 :
662 329892 : result->commandType = parse->commandType;
663 329892 : result->queryId = parse->queryId;
664 329892 : result->planOrigin = PLAN_STMT_STANDARD;
665 329892 : result->hasReturning = (parse->returningList != NIL);
666 329892 : result->hasModifyingCTE = parse->hasModifyingCTE;
667 329892 : result->canSetTag = parse->canSetTag;
668 329892 : result->transientPlan = glob->transientPlan;
669 329892 : result->dependsOnRole = glob->dependsOnRole;
670 329892 : result->parallelModeNeeded = glob->parallelModeNeeded;
671 329892 : result->planTree = top_plan;
672 329892 : result->partPruneInfos = glob->partPruneInfos;
673 329892 : result->rtable = glob->finalrtable;
674 659784 : result->unprunableRelids = bms_difference(glob->allRelids,
675 329892 : glob->prunableRelids);
676 329892 : result->permInfos = glob->finalrteperminfos;
677 329892 : result->subrtinfos = glob->subrtinfos;
678 329892 : result->appendRelations = glob->appendRelations;
679 329892 : result->subplans = glob->subplans;
680 329892 : result->rewindPlanIDs = glob->rewindPlanIDs;
681 329892 : 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 398637 : foreach(lc, glob->resultRelations)
688 68745 : result->resultRelationRelids = bms_add_member(result->resultRelationRelids,
689 : lfirst_int(lc));
690 340691 : foreach(lc, glob->finalrowmarks)
691 10799 : result->rowMarkRelids = bms_add_member(result->rowMarkRelids,
692 10799 : ((PlanRowMark *) lfirst(lc))->rti);
693 :
694 329892 : result->relationOids = glob->relationOids;
695 329892 : result->invalItems = glob->invalItems;
696 329892 : result->paramExecTypes = glob->paramExecTypes;
697 : /* utilityStmt should be null, but we might as well copy it */
698 329892 : result->utilityStmt = parse->utilityStmt;
699 329892 : result->elidedNodes = glob->elidedNodes;
700 329892 : result->stmt_location = parse->stmt_location;
701 329892 : result->stmt_len = parse->stmt_len;
702 :
703 329892 : result->jitFlags = PGJIT_NONE;
704 329892 : 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 329892 : if (planner_shutdown_hook)
730 87621 : (*planner_shutdown_hook) (glob, parse, query_string, result);
731 :
732 329892 : if (glob->partition_directory != NULL)
733 9200 : DestroyPartitionDirectory(glob->partition_directory);
734 :
735 329892 : 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 394709 : 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 394709 : root = makeNode(PlannerInfo);
792 394709 : root->parse = parse;
793 394709 : root->glob = glob;
794 394709 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
795 394709 : root->plan_name = plan_name;
796 394709 : if (alternative_root != NULL)
797 1262 : root->alternative_plan_name = alternative_root->plan_name;
798 : else
799 393447 : root->alternative_plan_name = plan_name;
800 394709 : root->parent_root = parent_root;
801 394709 : root->plan_params = NIL;
802 394709 : root->outer_params = NULL;
803 394709 : root->planner_cxt = CurrentMemoryContext;
804 394709 : root->init_plans = NIL;
805 394709 : root->cte_plan_ids = NIL;
806 394709 : root->multiexpr_params = NIL;
807 394709 : root->join_domains = NIL;
808 394709 : root->eq_classes = NIL;
809 394709 : root->ec_merging_done = false;
810 394709 : root->last_rinfo_serial = 0;
811 394709 : root->all_result_relids =
812 394709 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
813 394709 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
814 394709 : root->append_rel_list = NIL;
815 394709 : root->row_identity_vars = NIL;
816 394709 : root->rowMarks = NIL;
817 394709 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
818 394709 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
819 394709 : root->processed_groupClause = NIL;
820 394709 : root->processed_distinctClause = NIL;
821 394709 : root->processed_tlist = NIL;
822 394709 : root->update_colnos = NIL;
823 394709 : root->grouping_map = NULL;
824 394709 : root->minmax_aggs = NIL;
825 394709 : root->qual_security_level = 0;
826 394709 : root->hasPseudoConstantQuals = false;
827 394709 : root->hasAlternativeSubPlans = false;
828 394709 : root->placeholdersFrozen = false;
829 394709 : root->hasRecursion = hasRecursion;
830 394709 : root->assumeReplanning = false;
831 394709 : if (hasRecursion)
832 637 : root->wt_param_id = assign_special_exec_param(root);
833 : else
834 394072 : root->wt_param_id = -1;
835 394709 : 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 394709 : 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 394709 : if (parse->cteList)
849 2166 : SS_process_ctes(root);
850 :
851 : /*
852 : * If it's a MERGE command, transform the joinlist as appropriate.
853 : */
854 394705 : 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 394705 : 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 394705 : 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 394705 : if (parse->hasSubLinks)
881 30170 : 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 394705 : 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 394701 : 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 394701 : if (parse->setOperations)
904 5375 : 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 394701 : root->hasJoinRTEs = false;
917 394701 : root->hasLateralRTEs = false;
918 394701 : root->group_rtindex = 0;
919 394701 : hasOuterJoins = false;
920 394701 : hasResultRTEs = false;
921 1088085 : foreach(l, parse->rtable)
922 : {
923 693384 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
924 :
925 693384 : switch (rte->rtekind)
926 : {
927 71192 : case RTE_JOIN:
928 71192 : root->hasJoinRTEs = true;
929 71192 : if (IS_OUTER_JOIN(rte->jointype))
930 36256 : hasOuterJoins = true;
931 71192 : break;
932 147851 : case RTE_RESULT:
933 147851 : hasResultRTEs = true;
934 147851 : break;
935 4348 : case RTE_GROUP:
936 : Assert(parse->hasGroupRTE);
937 4348 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
938 4348 : break;
939 469993 : default:
940 : /* No work here for other RTE types */
941 469993 : break;
942 : }
943 :
944 693384 : if (rte->lateral)
945 8130 : 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 693384 : 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 394701 : if (parse->resultRelation)
963 : {
964 65036 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
965 :
966 65036 : if (!rte->inh)
967 62688 : root->leaf_result_relids =
968 62688 : 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 1087297 : foreach(l, parse->rtable)
986 : {
987 692862 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
988 :
989 692862 : if (rte->perminfoindex != 0 &&
990 373405 : rte->relkind == RELKIND_VIEW)
991 : {
992 : RTEPermissionInfo *perminfo;
993 : bool result;
994 :
995 16297 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
996 16297 : result = ExecCheckOneRelPerms(perminfo);
997 16297 : 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 394435 : 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 394435 : 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 391778 : parse->targetList = (List *)
1023 394435 : preprocess_expression(root, (Node *) parse->targetList,
1024 : EXPRKIND_TARGET);
1025 :
1026 391778 : newWithCheckOptions = NIL;
1027 394346 : 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 391778 : parse->withCheckOptions = newWithCheckOptions;
1037 :
1038 391778 : parse->returningList = (List *)
1039 391778 : preprocess_expression(root, (Node *) parse->returningList,
1040 : EXPRKIND_TARGET);
1041 :
1042 391778 : preprocess_qual_conditions(root, (Node *) parse->jointree);
1043 :
1044 391774 : parse->havingQual = preprocess_expression(root, parse->havingQual,
1045 : EXPRKIND_QUAL);
1046 :
1047 394221 : foreach(l, parse->windowClause)
1048 : {
1049 2447 : WindowClause *wc = lfirst_node(WindowClause, l);
1050 :
1051 : /* partitionClause/orderClause are sort/group expressions */
1052 2447 : wc->startOffset = preprocess_expression(root, wc->startOffset,
1053 : EXPRKIND_LIMIT);
1054 2447 : wc->endOffset = preprocess_expression(root, wc->endOffset,
1055 : EXPRKIND_LIMIT);
1056 : }
1057 :
1058 391774 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
1059 : EXPRKIND_LIMIT);
1060 391774 : parse->limitCount = preprocess_expression(root, parse->limitCount,
1061 : EXPRKIND_LIMIT);
1062 :
1063 391774 : if (parse->onConflict)
1064 : {
1065 3580 : parse->onConflict->arbiterElems = (List *)
1066 1790 : preprocess_expression(root,
1067 1790 : (Node *) parse->onConflict->arbiterElems,
1068 : EXPRKIND_ARBITER_ELEM);
1069 3580 : parse->onConflict->arbiterWhere =
1070 1790 : preprocess_expression(root,
1071 1790 : parse->onConflict->arbiterWhere,
1072 : EXPRKIND_QUAL);
1073 3580 : parse->onConflict->onConflictSet = (List *)
1074 1790 : preprocess_expression(root,
1075 1790 : (Node *) parse->onConflict->onConflictSet,
1076 : EXPRKIND_TARGET);
1077 1790 : parse->onConflict->onConflictWhere =
1078 1790 : preprocess_expression(root,
1079 1790 : parse->onConflict->onConflictWhere,
1080 : EXPRKIND_QUAL);
1081 : /* exclRelTlist contains only Vars, so no preprocessing needed */
1082 : }
1083 :
1084 391774 : if (parse->forPortionOf)
1085 : {
1086 2072 : parse->forPortionOf->targetRange =
1087 1036 : preprocess_expression(root,
1088 1036 : parse->forPortionOf->targetRange,
1089 : EXPRKIND_TARGET);
1090 1036 : if (contain_volatile_functions(parse->forPortionOf->targetRange))
1091 8 : ereport(ERROR,
1092 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1093 : errmsg("FOR PORTION OF bounds cannot contain volatile functions")));
1094 : }
1095 :
1096 394033 : foreach(l, parse->mergeActionList)
1097 : {
1098 2267 : MergeAction *action = (MergeAction *) lfirst(l);
1099 :
1100 2267 : action->targetList = (List *)
1101 2267 : preprocess_expression(root,
1102 2267 : (Node *) action->targetList,
1103 : EXPRKIND_TARGET);
1104 2267 : action->qual =
1105 2267 : preprocess_expression(root,
1106 : (Node *) action->qual,
1107 : EXPRKIND_QUAL);
1108 : }
1109 :
1110 391766 : parse->mergeJoinCondition =
1111 391766 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1112 :
1113 391766 : root->append_rel_list = (List *)
1114 391766 : preprocess_expression(root, (Node *) root->append_rel_list,
1115 : EXPRKIND_APPINFO);
1116 :
1117 : /* Also need to preprocess expressions within RTEs */
1118 1081292 : foreach(l, parse->rtable)
1119 : {
1120 689526 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1121 : int kind;
1122 : ListCell *lcsq;
1123 :
1124 689526 : if (rte->rtekind == RTE_RELATION)
1125 : {
1126 360186 : if (rte->tablesample)
1127 180 : rte->tablesample = (TableSampleClause *)
1128 180 : preprocess_expression(root,
1129 180 : (Node *) rte->tablesample,
1130 : EXPRKIND_TABLESAMPLE);
1131 : }
1132 329340 : else if (rte->rtekind == RTE_SUBQUERY)
1133 : {
1134 : /*
1135 : * We don't want to do all preprocessing yet on the subquery's
1136 : * expressions, since that will happen when we plan it. But if it
1137 : * contains any join aliases of our level, those have to get
1138 : * expanded now, because planning of the subquery won't do it.
1139 : * That's only possible if the subquery is LATERAL.
1140 : */
1141 62695 : if (rte->lateral && root->hasJoinRTEs)
1142 1623 : rte->subquery = (Query *)
1143 1623 : flatten_join_alias_vars(root, root->parse,
1144 1623 : (Node *) rte->subquery);
1145 : }
1146 266645 : else if (rte->rtekind == RTE_FUNCTION)
1147 : {
1148 : /* Preprocess the function expression(s) fully */
1149 34665 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1150 34665 : rte->functions = (List *)
1151 34665 : preprocess_expression(root, (Node *) rte->functions, kind);
1152 : }
1153 231980 : else if (rte->rtekind == RTE_TABLEFUNC)
1154 : {
1155 : /* Preprocess the function expression(s) fully */
1156 519 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1157 519 : rte->tablefunc = (TableFunc *)
1158 519 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1159 : }
1160 231461 : else if (rte->rtekind == RTE_VALUES)
1161 : {
1162 : /* Preprocess the values lists fully */
1163 6972 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1164 6972 : rte->values_lists = (List *)
1165 6972 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1166 : }
1167 224489 : else if (rte->rtekind == RTE_GROUP)
1168 : {
1169 : /* Preprocess the groupexprs list fully */
1170 4348 : rte->groupexprs = (List *)
1171 4348 : preprocess_expression(root, (Node *) rte->groupexprs,
1172 : EXPRKIND_GROUPEXPR);
1173 : }
1174 :
1175 : /*
1176 : * Process each element of the securityQuals list as if it were a
1177 : * separate qual expression (as indeed it is). We need to do it this
1178 : * way to get proper canonicalization of AND/OR structure. Note that
1179 : * this converts each element into an implicit-AND sublist.
1180 : */
1181 692331 : foreach(lcsq, rte->securityQuals)
1182 : {
1183 2805 : lfirst(lcsq) = preprocess_expression(root,
1184 2805 : (Node *) lfirst(lcsq),
1185 : EXPRKIND_QUAL);
1186 : }
1187 : }
1188 :
1189 : /*
1190 : * Now that we are done preprocessing expressions, and in particular done
1191 : * flattening join alias variables, get rid of the joinaliasvars lists.
1192 : * They no longer match what expressions in the rest of the tree look
1193 : * like, because we have not preprocessed expressions in those lists (and
1194 : * do not want to; for example, expanding a SubLink there would result in
1195 : * a useless unreferenced subplan). Leaving them in place simply creates
1196 : * a hazard for later scans of the tree. We could try to prevent that by
1197 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1198 : * but that doesn't sound very reliable.
1199 : */
1200 391766 : if (root->hasJoinRTEs)
1201 : {
1202 248331 : foreach(l, parse->rtable)
1203 : {
1204 203541 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1205 :
1206 203541 : rte->joinaliasvars = NIL;
1207 : }
1208 : }
1209 :
1210 : /*
1211 : * Before we flatten GROUP Vars, check which HAVING clauses have collation
1212 : * conflicts. When GROUP BY uses a nondeterministic collation, values
1213 : * that are "equal" for grouping may be distinguishable under a different
1214 : * collation. If such a HAVING clause were moved to WHERE, it would
1215 : * filter individual rows before grouping, potentially eliminating some
1216 : * members of a group and thereby changing aggregate results.
1217 : *
1218 : * We do this check before flatten_group_exprs because we can easily
1219 : * identify grouping expressions by checking whether a Var references
1220 : * RTE_GROUP, and such Vars directly carry the GROUP BY collation as their
1221 : * varcollid. After flattening, these Vars are replaced by the underlying
1222 : * expressions, and we would have to match expressions in the HAVING
1223 : * clause back to grouping expressions, which is much more complex.
1224 : */
1225 391766 : if (parse->hasGroupRTE)
1226 : havingCollationConflicts =
1227 4348 : find_having_collation_conflicts(parse, root->group_rtindex);
1228 : else
1229 387418 : havingCollationConflicts = NULL;
1230 :
1231 : /*
1232 : * Replace any Vars in the subquery's targetlist and havingQual that
1233 : * reference GROUP outputs with the underlying grouping expressions.
1234 : *
1235 : * Note that we need to perform this replacement after we've preprocessed
1236 : * the grouping expressions. This is to ensure that there is only one
1237 : * instance of SubPlan for each SubLink contained within the grouping
1238 : * expressions.
1239 : */
1240 391766 : if (parse->hasGroupRTE)
1241 : {
1242 4348 : parse->targetList = (List *)
1243 4348 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1244 4348 : parse->havingQual =
1245 4348 : flatten_group_exprs(root, root->parse, parse->havingQual);
1246 : }
1247 :
1248 : /* Constant-folding might have removed all set-returning functions */
1249 391766 : if (parse->hasTargetSRFs)
1250 10174 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1251 :
1252 : /*
1253 : * If we have grouping sets, expand the groupingSets tree of this query to
1254 : * a flat list of grouping sets. We need to do this before optimizing
1255 : * HAVING, since we can't easily tell if there's an empty grouping set
1256 : * until we have this representation.
1257 : */
1258 391766 : if (parse->groupingSets)
1259 : {
1260 915 : parse->groupingSets =
1261 915 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1262 : }
1263 :
1264 : /*
1265 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1266 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1267 : * volatile functions (since a HAVING clause is supposed to be executed
1268 : * only once per group). We also can't do this if there are any grouping
1269 : * sets and the clause references any columns that are nullable by the
1270 : * grouping sets; the nulled values of those columns are not available
1271 : * before the grouping step. (The test on groupClause might seem wrong,
1272 : * but it's okay: it's just an optimization to avoid running pull_varnos
1273 : * when there cannot be any Vars in the HAVING clause.)
1274 : *
1275 : * We also cannot do this if the HAVING clause uses a different collation
1276 : * than the GROUP BY for any grouping expression whose GROUP BY collation
1277 : * is nondeterministic. This is detected before flatten_group_exprs (see
1278 : * find_having_collation_conflicts above) and recorded in the
1279 : * havingCollationConflicts bitmapset. The bitmapset indexes remain valid
1280 : * here because flatten_group_exprs uses expression_tree_mutator, which
1281 : * preserves the list length and ordering of havingQual.
1282 : *
1283 : * Also, it may be that the clause is so expensive to execute that we're
1284 : * better off doing it only once per group, despite the loss of
1285 : * selectivity. This is hard to estimate short of doing the entire
1286 : * planning process twice, so we use a heuristic: clauses containing
1287 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1288 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1289 : * instead of after.
1290 : *
1291 : * If the query has no empty grouping set then we can simply move such a
1292 : * clause into WHERE; any group that fails the clause will not be in the
1293 : * output because none of its tuples will reach the grouping or
1294 : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1295 : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1296 : * clause must be degenerate (variable-free), so we can copy it into WHERE
1297 : * so that query_planner() can use it in a gating Result node. (This could
1298 : * be done better, but it seems not worth optimizing.)
1299 : *
1300 : * Note that a HAVING clause may contain expressions that are not fully
1301 : * preprocessed. This can happen if these expressions are part of
1302 : * grouping items. In such cases, they are replaced with GROUP Vars in
1303 : * the parser and then replaced back after we're done with expression
1304 : * preprocessing on havingQual. This is not an issue if the clause
1305 : * remains in HAVING, because these expressions will be matched to lower
1306 : * target items in setrefs.c. However, if the clause is moved or copied
1307 : * into WHERE, we need to ensure that these expressions are fully
1308 : * preprocessed.
1309 : *
1310 : * Note that both havingQual and parse->jointree->quals are in
1311 : * implicitly-ANDed-list form at this point, even though they are declared
1312 : * as Node *.
1313 : */
1314 391766 : newHaving = NIL;
1315 391766 : havingIdx = 0;
1316 393036 : foreach(l, (List *) parse->havingQual)
1317 : {
1318 1270 : Node *havingclause = (Node *) lfirst(l);
1319 :
1320 1861 : if (contain_agg_clause(havingclause) ||
1321 1182 : contain_volatile_functions(havingclause) ||
1322 1182 : contain_subplans(havingclause) ||
1323 591 : bms_is_member(havingIdx, havingCollationConflicts) ||
1324 611 : (parse->groupClause && parse->groupingSets &&
1325 100 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1326 : {
1327 : /* keep it in HAVING */
1328 819 : newHaving = lappend(newHaving, havingclause);
1329 : }
1330 451 : else if (parse->groupClause &&
1331 421 : (parse->groupingSets == NIL ||
1332 40 : (List *) linitial(parse->groupingSets) != NIL))
1333 411 : {
1334 : /* There is GROUP BY, but no empty grouping set */
1335 : Node *whereclause;
1336 :
1337 : /* Preprocess the HAVING clause fully */
1338 411 : whereclause = preprocess_expression(root, havingclause,
1339 : EXPRKIND_QUAL);
1340 : /* ... and move it to WHERE */
1341 411 : parse->jointree->quals = (Node *)
1342 411 : list_concat((List *) parse->jointree->quals,
1343 : (List *) whereclause);
1344 : }
1345 : else
1346 : {
1347 : /* There is an empty grouping set (perhaps implicitly) */
1348 : Node *whereclause;
1349 :
1350 : /* Preprocess the HAVING clause fully */
1351 40 : whereclause = preprocess_expression(root, copyObject(havingclause),
1352 : EXPRKIND_QUAL);
1353 : /* ... and put a copy in WHERE */
1354 80 : parse->jointree->quals = (Node *)
1355 40 : list_concat((List *) parse->jointree->quals,
1356 : (List *) whereclause);
1357 : /* ... and also keep it in HAVING */
1358 40 : newHaving = lappend(newHaving, havingclause);
1359 : }
1360 :
1361 1270 : havingIdx++;
1362 : }
1363 391766 : parse->havingQual = (Node *) newHaving;
1364 :
1365 : /*
1366 : * If we have any outer joins, try to reduce them to plain inner joins.
1367 : * This step is most easily done after we've done expression
1368 : * preprocessing.
1369 : */
1370 391766 : if (hasOuterJoins)
1371 25199 : reduce_outer_joins(root);
1372 :
1373 : /*
1374 : * If we have any RTE_RESULT relations, see if they can be deleted from
1375 : * the jointree. We also rely on this processing to flatten single-child
1376 : * FromExprs underneath outer joins. This step is most effectively done
1377 : * after we've done expression preprocessing and outer join reduction.
1378 : */
1379 391766 : if (hasResultRTEs || hasOuterJoins)
1380 169198 : remove_useless_result_rtes(root);
1381 :
1382 : /*
1383 : * Do the main planning.
1384 : */
1385 391766 : grouping_planner(root, tuple_fraction, setops);
1386 :
1387 : /*
1388 : * Capture the set of outer-level param IDs we have access to, for use in
1389 : * extParam/allParam calculations later.
1390 : */
1391 391719 : SS_identify_outer_params(root);
1392 :
1393 : /*
1394 : * If any initPlans were created in this query level, adjust the surviving
1395 : * Paths' costs and parallel-safety flags to account for them. The
1396 : * initPlans won't actually get attached to the plan tree till
1397 : * create_plan() runs, but we must include their effects now.
1398 : */
1399 391719 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1400 391719 : SS_charge_for_initplans(root, final_rel);
1401 :
1402 : /*
1403 : * Make sure we've identified the cheapest Path for the final rel. (By
1404 : * doing this here not in grouping_planner, we include initPlan costs in
1405 : * the decision, though it's unlikely that will change anything.)
1406 : */
1407 391719 : set_cheapest(final_rel);
1408 :
1409 391719 : return root;
1410 : }
1411 :
1412 : /*
1413 : * preprocess_expression
1414 : * Do subquery_planner's preprocessing work for an expression,
1415 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1416 : * conditions), a HAVING clause, or a few other things.
1417 : */
1418 : static Node *
1419 3301829 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1420 : {
1421 : /*
1422 : * Fall out quickly if expression is empty. This occurs often enough to
1423 : * be worth checking. Note that null->null is the correct conversion for
1424 : * implicit-AND result format, too.
1425 : */
1426 3301829 : if (expr == NULL)
1427 2597814 : return NULL;
1428 :
1429 : /*
1430 : * If the query has any join RTEs, replace join alias variables with
1431 : * base-relation variables. We must do this first, since any expressions
1432 : * we may extract from the joinaliasvars lists have not been preprocessed.
1433 : * For example, if we did this after sublink processing, sublinks expanded
1434 : * out from join aliases would not get processed. But we can skip this in
1435 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1436 : * they can't contain any Vars of the current query level.
1437 : */
1438 704015 : if (root->hasJoinRTEs &&
1439 319408 : !(kind == EXPRKIND_RTFUNC ||
1440 159555 : kind == EXPRKIND_VALUES ||
1441 : kind == EXPRKIND_TABLESAMPLE ||
1442 : kind == EXPRKIND_TABLEFUNC))
1443 159540 : expr = flatten_join_alias_vars(root, root->parse, expr);
1444 :
1445 : /*
1446 : * Simplify constant expressions. For function RTEs, this was already
1447 : * done by preprocess_function_rtes. (But note we must do it again for
1448 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1449 : * un-simplified subexpressions inserted by flattening of subqueries or
1450 : * join alias variables.)
1451 : *
1452 : * Note: an essential effect of this is to convert named-argument function
1453 : * calls to positional notation and insert the current actual values of
1454 : * any default arguments for functions. To ensure that happens, we *must*
1455 : * process all expressions here. Previous PG versions sometimes skipped
1456 : * const-simplification if it didn't seem worth the trouble, but we can't
1457 : * do that anymore.
1458 : *
1459 : * Note: this also flattens nested AND and OR expressions into N-argument
1460 : * form. All processing of a qual expression after this point must be
1461 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1462 : * with AND directly under AND, nor OR directly under OR.
1463 : */
1464 704015 : if (kind != EXPRKIND_RTFUNC)
1465 673818 : expr = eval_const_expressions(root, expr);
1466 :
1467 : /*
1468 : * If it's a qual or havingQual, canonicalize it.
1469 : */
1470 701354 : if (kind == EXPRKIND_QUAL)
1471 : {
1472 257440 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1473 :
1474 : #ifdef OPTIMIZER_DEBUG
1475 : printf("After canonicalize_qual()\n");
1476 : pprint(expr);
1477 : #endif
1478 : }
1479 :
1480 : /*
1481 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1482 : * hashfuncid of any that might execute more quickly by using hash lookups
1483 : * instead of a linear search.
1484 : */
1485 701354 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1486 : {
1487 643262 : convert_saop_to_hashed_saop(expr);
1488 : }
1489 :
1490 : /* Expand SubLinks to SubPlans */
1491 701354 : if (root->parse->hasSubLinks)
1492 88677 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1493 :
1494 : /*
1495 : * XXX do not insert anything here unless you have grokked the comments in
1496 : * SS_replace_correlation_vars ...
1497 : */
1498 :
1499 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1500 701354 : if (root->query_level > 1)
1501 142714 : expr = SS_replace_correlation_vars(root, expr);
1502 :
1503 : /*
1504 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1505 : * don't want to do this before eval_const_expressions, since the latter
1506 : * would be unable to simplify a top-level AND correctly. Also,
1507 : * SS_process_sublinks expects explicit-AND format.)
1508 : */
1509 701354 : if (kind == EXPRKIND_QUAL)
1510 257440 : expr = (Node *) make_ands_implicit((Expr *) expr);
1511 :
1512 701354 : return expr;
1513 : }
1514 :
1515 : /*
1516 : * preprocess_qual_conditions
1517 : * Recursively scan the query's jointree and do subquery_planner's
1518 : * preprocessing work on each qual condition found therein.
1519 : */
1520 : static void
1521 991022 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1522 : {
1523 991022 : if (jtnode == NULL)
1524 0 : return;
1525 991022 : if (IsA(jtnode, RangeTblRef))
1526 : {
1527 : /* nothing to do here */
1528 : }
1529 486555 : else if (IsA(jtnode, FromExpr))
1530 : {
1531 406647 : FromExpr *f = (FromExpr *) jtnode;
1532 : ListCell *l;
1533 :
1534 846075 : foreach(l, f->fromlist)
1535 439428 : preprocess_qual_conditions(root, lfirst(l));
1536 :
1537 406647 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1538 : }
1539 79908 : else if (IsA(jtnode, JoinExpr))
1540 : {
1541 79908 : JoinExpr *j = (JoinExpr *) jtnode;
1542 :
1543 79908 : preprocess_qual_conditions(root, j->larg);
1544 79908 : preprocess_qual_conditions(root, j->rarg);
1545 :
1546 79908 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1547 : }
1548 : else
1549 0 : elog(ERROR, "unrecognized node type: %d",
1550 : (int) nodeTag(jtnode));
1551 : }
1552 :
1553 : /*
1554 : * find_having_collation_conflicts
1555 : * Identify HAVING clauses that must not be moved to WHERE due to collation
1556 : * mismatches with GROUP BY.
1557 : *
1558 : * This must be called before flatten_group_exprs, while the HAVING clause
1559 : * still contains GROUP Vars (Vars referencing RTE_GROUP). These GROUP Vars
1560 : * carry the GROUP BY collation as their varcollid. A GROUP Var with a
1561 : * nondeterministic varcollid conflicts whenever some collation-aware ancestor
1562 : * on its path applies a different inputcollid: that operator would distinguish
1563 : * values which the GROUP BY considers equal, so the clause is unsafe to push
1564 : * to WHERE.
1565 : *
1566 : * Returns a Bitmapset of zero-based indexes into the havingQual list for
1567 : * clauses that have collation conflicts and must stay in HAVING.
1568 : */
1569 : static Bitmapset *
1570 4348 : find_having_collation_conflicts(Query *parse, Index group_rtindex)
1571 : {
1572 4348 : Bitmapset *result = NULL;
1573 : having_collation_ctx ctx;
1574 : int idx;
1575 :
1576 4348 : if (parse->havingQual == NULL)
1577 3510 : return NULL;
1578 :
1579 838 : ctx.group_rtindex = group_rtindex;
1580 838 : ctx.ancestor_collids = NIL;
1581 :
1582 838 : idx = 0;
1583 2653 : foreach_ptr(Node, clause, (List *) parse->havingQual)
1584 : {
1585 977 : if (having_collation_conflict_walker(clause, &ctx))
1586 80 : result = bms_add_member(result, idx);
1587 977 : idx++;
1588 : Assert(ctx.ancestor_collids == NIL);
1589 : }
1590 :
1591 838 : return result;
1592 : }
1593 :
1594 : /*
1595 : * Walker function for find_having_collation_conflicts.
1596 : *
1597 : * Walk the clause top-down, maintaining a stack of inputcollids contributed
1598 : * by collation-aware ancestors. At each GROUP Var with a nondeterministic
1599 : * varcollid, the clause has a conflict if any ancestor's inputcollid differs
1600 : * from the GROUP Var's varcollid. Most collation-aware nodes expose their
1601 : * inputcollid through exprInputCollation(). Two structural exceptions need
1602 : * special handling:
1603 : *
1604 : * - RowCompareExpr carries one inputcollid per column in inputcollids[], so we
1605 : * descend into its (largs[i], rargs[i]) pairs explicitly with the matching
1606 : * collation pushed onto the stack.
1607 : *
1608 : * - A simple CASE (CaseExpr with a non-NULL arg) holds the arg outside the
1609 : * WHEN's OpExpr, even though the WHEN's OpExpr is the place where the
1610 : * comparison's inputcollid lives. Parse analysis builds each WHEN as
1611 : * "OpExpr(CaseTestExpr op val)" -- the CaseTestExpr is a placeholder for
1612 : * the arg. Before walking cexpr->arg we therefore push every WHEN's
1613 : * inputcollid onto the ancestor stack, so a GROUP Var at the arg is
1614 : * checked against the same collations the WHEN comparisons would apply.
1615 : * The WHEN bodies and defresult are then walked under the unchanged stack
1616 : * so their own collation contexts are picked up by the default path.
1617 : */
1618 : static bool
1619 4306 : having_collation_conflict_walker(Node *node, having_collation_ctx *ctx)
1620 : {
1621 : Oid this_collid;
1622 : bool result;
1623 :
1624 4306 : if (node == NULL)
1625 453 : return false;
1626 :
1627 3853 : if (IsA(node, Var))
1628 : {
1629 928 : Var *var = (Var *) node;
1630 :
1631 : /* We should not see any upper-level Vars here */
1632 : Assert(var->varlevelsup == 0);
1633 :
1634 928 : if (var->varno == ctx->group_rtindex &&
1635 544 : OidIsValid(var->varcollid) &&
1636 361 : !get_collation_isdeterministic(var->varcollid))
1637 : {
1638 280 : foreach_oid(collid, ctx->ancestor_collids)
1639 : {
1640 160 : if (collid != var->varcollid)
1641 80 : return true;
1642 : }
1643 : }
1644 848 : return false;
1645 : }
1646 :
1647 2925 : if (IsA(node, RowCompareExpr))
1648 : {
1649 10 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
1650 : ListCell *lc_l;
1651 : ListCell *lc_r;
1652 : ListCell *lc_c;
1653 :
1654 : /*
1655 : * Each column of a row comparison is compared under its own
1656 : * inputcollids[i]. Walk each (largs[i], rargs[i]) pair with that
1657 : * collation pushed, so a Var in column i is checked against the
1658 : * collation that actually applies to it.
1659 : */
1660 10 : forthree(lc_l, rcexpr->largs,
1661 : lc_r, rcexpr->rargs,
1662 : lc_c, rcexpr->inputcollids)
1663 : {
1664 10 : Oid collid = lfirst_oid(lc_c);
1665 : bool found;
1666 :
1667 10 : if (OidIsValid(collid))
1668 10 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1669 : collid);
1670 :
1671 10 : found = having_collation_conflict_walker((Node *) lfirst(lc_l),
1672 10 : ctx) ||
1673 0 : having_collation_conflict_walker((Node *) lfirst(lc_r),
1674 : ctx);
1675 :
1676 10 : if (OidIsValid(collid))
1677 10 : ctx->ancestor_collids =
1678 10 : list_delete_last(ctx->ancestor_collids);
1679 :
1680 10 : if (found)
1681 10 : return true;
1682 : }
1683 0 : return false;
1684 : }
1685 :
1686 2915 : if (IsA(node, CaseExpr) && ((CaseExpr *) node)->arg != NULL)
1687 : {
1688 30 : CaseExpr *cexpr = (CaseExpr *) node;
1689 30 : int saved_len = list_length(ctx->ancestor_collids);
1690 : bool found;
1691 :
1692 : /*
1693 : * Push every WHEN's inputcollid before walking cexpr->arg, since each
1694 : * WHEN implicitly compares the arg under that inputcollid.
1695 : */
1696 90 : foreach_node(CaseWhen, cw, cexpr->args)
1697 : {
1698 30 : Oid collid = exprInputCollation((Node *) cw->expr);
1699 :
1700 30 : if (OidIsValid(collid))
1701 30 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1702 : collid);
1703 : }
1704 :
1705 30 : found = having_collation_conflict_walker((Node *) cexpr->arg, ctx);
1706 :
1707 30 : ctx->ancestor_collids = list_truncate(ctx->ancestor_collids,
1708 : saved_len);
1709 :
1710 30 : if (found)
1711 20 : return true;
1712 :
1713 : /*
1714 : * Walk the WHEN bodies and defresult under the unchanged ancestor
1715 : * stack; any inputcollids inside them are picked up by the default
1716 : * path.
1717 : */
1718 30 : foreach_node(CaseWhen, cw, cexpr->args)
1719 : {
1720 20 : if (having_collation_conflict_walker((Node *) cw->expr, ctx) ||
1721 10 : having_collation_conflict_walker((Node *) cw->result, ctx))
1722 0 : return true;
1723 : }
1724 10 : return having_collation_conflict_walker((Node *) cexpr->defresult,
1725 : ctx);
1726 : }
1727 :
1728 2885 : this_collid = exprInputCollation(node);
1729 2885 : if (OidIsValid(this_collid))
1730 366 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1731 : this_collid);
1732 :
1733 2885 : result = expression_tree_walker(node, having_collation_conflict_walker,
1734 : ctx);
1735 :
1736 2885 : if (OidIsValid(this_collid))
1737 366 : ctx->ancestor_collids = list_delete_last(ctx->ancestor_collids);
1738 :
1739 2885 : return result;
1740 : }
1741 :
1742 : /*
1743 : * preprocess_phv_expression
1744 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1745 : *
1746 : * If a LATERAL subquery references an output of another subquery, and that
1747 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1748 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1749 : * and later pull it back up during find_lateral_references, which runs after
1750 : * subquery_planner has preprocessed all the expressions that were in the
1751 : * current query level to start with. So we need to preprocess it then.
1752 : */
1753 : Expr *
1754 75 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1755 : {
1756 75 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1757 : }
1758 :
1759 : /*--------------------
1760 : * grouping_planner
1761 : * Perform planning steps related to grouping, aggregation, etc.
1762 : *
1763 : * This function adds all required top-level processing to the scan/join
1764 : * Path(s) produced by query_planner.
1765 : *
1766 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1767 : * tuple_fraction is interpreted as follows:
1768 : * 0: expect all tuples to be retrieved (normal case)
1769 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1770 : * from the plan to be retrieved
1771 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1772 : * expected to be retrieved (ie, a LIMIT specification).
1773 : * setops is used for set operation subqueries to provide the subquery with
1774 : * the context in which it's being used so that Paths correctly sorted for the
1775 : * set operation can be generated. NULL when not planning a set operation
1776 : * child, or when a child of a set op that isn't interested in sorted input.
1777 : *
1778 : * Returns nothing; the useful output is in the Paths we attach to the
1779 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1780 : * root->processed_tlist contains the final processed targetlist.
1781 : *
1782 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1783 : * to leave this to the caller.
1784 : *--------------------
1785 : */
1786 : static void
1787 391766 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1788 : SetOperationStmt *setops)
1789 : {
1790 391766 : Query *parse = root->parse;
1791 391766 : int64 offset_est = 0;
1792 391766 : int64 count_est = 0;
1793 391766 : double limit_tuples = -1.0;
1794 391766 : bool have_postponed_srfs = false;
1795 : PathTarget *final_target;
1796 : List *final_targets;
1797 : List *final_targets_contain_srfs;
1798 : bool final_target_parallel_safe;
1799 : RelOptInfo *current_rel;
1800 : RelOptInfo *final_rel;
1801 : FinalPathExtraData extra;
1802 : ListCell *lc;
1803 :
1804 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1805 391766 : if (parse->limitCount || parse->limitOffset)
1806 : {
1807 3717 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1808 : &offset_est, &count_est);
1809 :
1810 : /*
1811 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1812 : * estimate the effects of using a bounded sort.
1813 : */
1814 3717 : if (count_est > 0 && offset_est >= 0)
1815 3255 : limit_tuples = (double) count_est + (double) offset_est;
1816 : }
1817 :
1818 : /* Make tuple_fraction accessible to lower-level routines */
1819 391766 : root->tuple_fraction = tuple_fraction;
1820 :
1821 391766 : if (parse->setOperations)
1822 : {
1823 : /*
1824 : * Construct Paths for set operations. The results will not need any
1825 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1826 : * special work for recursive unions is the responsibility of
1827 : * plan_set_operations.
1828 : */
1829 4831 : current_rel = plan_set_operations(root);
1830 :
1831 : /*
1832 : * We should not need to call preprocess_targetlist, since we must be
1833 : * in a SELECT query node. Instead, use the processed_tlist returned
1834 : * by plan_set_operations (since this tells whether it returned any
1835 : * resjunk columns!), and transfer any sort key information from the
1836 : * original tlist.
1837 : */
1838 : Assert(parse->commandType == CMD_SELECT);
1839 :
1840 : /* for safety, copy processed_tlist instead of modifying in-place */
1841 4827 : root->processed_tlist =
1842 4827 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1843 : parse->targetList);
1844 :
1845 : /* Also extract the PathTarget form of the setop result tlist */
1846 4827 : final_target = current_rel->cheapest_total_path->pathtarget;
1847 :
1848 : /* And check whether it's parallel safe */
1849 : final_target_parallel_safe =
1850 4827 : is_parallel_safe(root, (Node *) final_target->exprs);
1851 :
1852 : /* The setop result tlist couldn't contain any SRFs */
1853 : Assert(!parse->hasTargetSRFs);
1854 4827 : final_targets = final_targets_contain_srfs = NIL;
1855 :
1856 : /*
1857 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1858 : * checked already, but let's make sure).
1859 : */
1860 4827 : if (parse->rowMarks)
1861 0 : ereport(ERROR,
1862 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1863 : /*------
1864 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1865 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1866 : LCS_asString(linitial_node(RowMarkClause,
1867 : parse->rowMarks)->strength))));
1868 :
1869 : /*
1870 : * Calculate pathkeys that represent result ordering requirements
1871 : */
1872 : Assert(parse->distinctClause == NIL);
1873 4827 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1874 : parse->sortClause,
1875 : root->processed_tlist);
1876 : }
1877 : else
1878 : {
1879 : /* No set operations, do regular planning */
1880 : PathTarget *sort_input_target;
1881 : List *sort_input_targets;
1882 : List *sort_input_targets_contain_srfs;
1883 : bool sort_input_target_parallel_safe;
1884 : PathTarget *grouping_target;
1885 : List *grouping_targets;
1886 : List *grouping_targets_contain_srfs;
1887 : bool grouping_target_parallel_safe;
1888 : PathTarget *scanjoin_target;
1889 : List *scanjoin_targets;
1890 : List *scanjoin_targets_contain_srfs;
1891 : bool scanjoin_target_parallel_safe;
1892 : bool scanjoin_target_same_exprs;
1893 : bool have_grouping;
1894 386935 : WindowFuncLists *wflists = NULL;
1895 386935 : List *activeWindows = NIL;
1896 386935 : grouping_sets_data *gset_data = NULL;
1897 : standard_qp_extra qp_extra;
1898 :
1899 : /* A recursive query should always have setOperations */
1900 : Assert(!root->hasRecursion);
1901 :
1902 : /* Preprocess grouping sets and GROUP BY clause, if any */
1903 386935 : if (parse->groupingSets)
1904 : {
1905 915 : gset_data = preprocess_grouping_sets(root);
1906 : }
1907 386020 : else if (parse->groupClause)
1908 : {
1909 : /* Preprocess regular GROUP BY clause, if any */
1910 3508 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1911 : }
1912 :
1913 : /*
1914 : * Preprocess targetlist. Note that much of the remaining planning
1915 : * work will be done with the PathTarget representation of tlists, but
1916 : * we must also maintain the full representation of the final tlist so
1917 : * that we can transfer its decoration (resnames etc) to the topmost
1918 : * tlist of the finished Plan. This is kept in processed_tlist.
1919 : */
1920 386931 : preprocess_targetlist(root);
1921 :
1922 : /*
1923 : * Mark all the aggregates with resolved aggtranstypes, and detect
1924 : * aggregates that are duplicates or can share transition state. We
1925 : * must do this before slicing and dicing the tlist into various
1926 : * pathtargets, else some copies of the Aggref nodes might escape
1927 : * being marked.
1928 : */
1929 386931 : if (parse->hasAggs)
1930 : {
1931 33198 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1932 33198 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1933 : }
1934 :
1935 : /*
1936 : * Locate any window functions in the tlist. (We don't need to look
1937 : * anywhere else, since expressions used in ORDER BY will be in there
1938 : * too.) Note that they could all have been eliminated by constant
1939 : * folding, in which case we don't need to do any more work.
1940 : */
1941 386931 : if (parse->hasWindowFuncs)
1942 : {
1943 2237 : wflists = find_window_functions((Node *) root->processed_tlist,
1944 2237 : list_length(parse->windowClause));
1945 2237 : if (wflists->numWindowFuncs > 0)
1946 : {
1947 : /*
1948 : * See if any modifications can be made to each WindowClause
1949 : * to allow the executor to execute the WindowFuncs more
1950 : * quickly.
1951 : */
1952 2232 : optimize_window_clauses(root, wflists);
1953 :
1954 : /* Extract the list of windows actually in use. */
1955 2232 : activeWindows = select_active_windows(root, wflists);
1956 :
1957 : /* Make sure they all have names, for EXPLAIN's use. */
1958 2232 : name_active_windows(activeWindows);
1959 : }
1960 : else
1961 5 : parse->hasWindowFuncs = false;
1962 : }
1963 :
1964 : /*
1965 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1966 : * adding logic between here and the query_planner() call. Anything
1967 : * that is needed in MIN/MAX-optimizable cases will have to be
1968 : * duplicated in planagg.c.
1969 : */
1970 386931 : if (parse->hasAggs)
1971 33198 : preprocess_minmax_aggregates(root);
1972 :
1973 : /*
1974 : * Figure out whether there's a hard limit on the number of rows that
1975 : * query_planner's result subplan needs to return. Even if we know a
1976 : * hard limit overall, it doesn't apply if the query has any
1977 : * grouping/aggregation operations, or SRFs in the tlist.
1978 : */
1979 386931 : if (parse->groupClause ||
1980 382587 : parse->groupingSets ||
1981 382512 : parse->distinctClause ||
1982 380611 : parse->hasAggs ||
1983 351139 : parse->hasWindowFuncs ||
1984 349023 : parse->hasTargetSRFs ||
1985 339183 : root->hasHavingQual)
1986 47768 : root->limit_tuples = -1.0;
1987 : else
1988 339163 : root->limit_tuples = limit_tuples;
1989 :
1990 : /* Set up data needed by standard_qp_callback */
1991 386931 : qp_extra.activeWindows = activeWindows;
1992 386931 : qp_extra.gset_data = gset_data;
1993 :
1994 : /*
1995 : * If we're a subquery for a set operation, store the SetOperationStmt
1996 : * in qp_extra.
1997 : */
1998 386931 : qp_extra.setop = setops;
1999 :
2000 : /*
2001 : * Generate the best unsorted and presorted paths for the scan/join
2002 : * portion of this Query, ie the processing represented by the
2003 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
2004 : * We also generate (in standard_qp_callback) pathkey representations
2005 : * of the query's sort clause, distinct clause, etc.
2006 : */
2007 386931 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
2008 :
2009 : /*
2010 : * Convert the query's result tlist into PathTarget format.
2011 : *
2012 : * Note: this cannot be done before query_planner() has performed
2013 : * appendrel expansion, because that might add resjunk entries to
2014 : * root->processed_tlist. Waiting till afterwards is also helpful
2015 : * because the target width estimates can use per-Var width numbers
2016 : * that were obtained within query_planner().
2017 : */
2018 386896 : final_target = create_pathtarget(root, root->processed_tlist);
2019 : final_target_parallel_safe =
2020 386896 : is_parallel_safe(root, (Node *) final_target->exprs);
2021 :
2022 : /*
2023 : * If ORDER BY was given, consider whether we should use a post-sort
2024 : * projection, and compute the adjusted target for preceding steps if
2025 : * so.
2026 : */
2027 386896 : if (parse->sortClause)
2028 : {
2029 57681 : sort_input_target = make_sort_input_target(root,
2030 : final_target,
2031 : &have_postponed_srfs);
2032 : sort_input_target_parallel_safe =
2033 57681 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
2034 : }
2035 : else
2036 : {
2037 329215 : sort_input_target = final_target;
2038 329215 : sort_input_target_parallel_safe = final_target_parallel_safe;
2039 : }
2040 :
2041 : /*
2042 : * If we have window functions to deal with, the output from any
2043 : * grouping step needs to be what the window functions want;
2044 : * otherwise, it should be sort_input_target.
2045 : */
2046 386896 : if (activeWindows)
2047 : {
2048 2232 : grouping_target = make_window_input_target(root,
2049 : final_target,
2050 : activeWindows);
2051 : grouping_target_parallel_safe =
2052 2232 : is_parallel_safe(root, (Node *) grouping_target->exprs);
2053 : }
2054 : else
2055 : {
2056 384664 : grouping_target = sort_input_target;
2057 384664 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
2058 : }
2059 :
2060 : /*
2061 : * If we have grouping or aggregation to do, the topmost scan/join
2062 : * plan node must emit what the grouping step wants; otherwise, it
2063 : * should emit grouping_target.
2064 : */
2065 382552 : have_grouping = (parse->groupClause || parse->groupingSets ||
2066 769448 : parse->hasAggs || root->hasHavingQual);
2067 386896 : if (have_grouping)
2068 : {
2069 33942 : scanjoin_target = make_group_input_target(root, final_target);
2070 : scanjoin_target_parallel_safe =
2071 33942 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2072 : }
2073 : else
2074 : {
2075 352954 : scanjoin_target = grouping_target;
2076 352954 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2077 : }
2078 :
2079 : /*
2080 : * If there are any SRFs in the targetlist, we must separate each of
2081 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2082 : * each of the named targets with a SRF-free version, and remember the
2083 : * list of additional projection steps we need to add afterwards.
2084 : */
2085 386896 : if (parse->hasTargetSRFs)
2086 : {
2087 : /* final_target doesn't recompute any SRFs in sort_input_target */
2088 10174 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2089 : &final_targets,
2090 : &final_targets_contain_srfs);
2091 10174 : final_target = linitial_node(PathTarget, final_targets);
2092 : Assert(!linitial_int(final_targets_contain_srfs));
2093 : /* likewise for sort_input_target vs. grouping_target */
2094 10174 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2095 : &sort_input_targets,
2096 : &sort_input_targets_contain_srfs);
2097 10174 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2098 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2099 : /* likewise for grouping_target vs. scanjoin_target */
2100 10174 : split_pathtarget_at_srfs_grouping(root,
2101 : grouping_target, scanjoin_target,
2102 : &grouping_targets,
2103 : &grouping_targets_contain_srfs);
2104 10174 : grouping_target = linitial_node(PathTarget, grouping_targets);
2105 : Assert(!linitial_int(grouping_targets_contain_srfs));
2106 : /* scanjoin_target will not have any SRFs precomputed for it */
2107 10174 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2108 : &scanjoin_targets,
2109 : &scanjoin_targets_contain_srfs);
2110 10174 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2111 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2112 : }
2113 : else
2114 : {
2115 : /* initialize lists; for most of these, dummy values are OK */
2116 376722 : final_targets = final_targets_contain_srfs = NIL;
2117 376722 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2118 376722 : grouping_targets = grouping_targets_contain_srfs = NIL;
2119 376722 : scanjoin_targets = list_make1(scanjoin_target);
2120 376722 : scanjoin_targets_contain_srfs = NIL;
2121 : }
2122 :
2123 : /* Apply scan/join target. */
2124 386896 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2125 386896 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2126 386896 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2127 : scanjoin_targets_contain_srfs,
2128 : scanjoin_target_parallel_safe,
2129 : scanjoin_target_same_exprs);
2130 :
2131 : /*
2132 : * Save the various upper-rel PathTargets we just computed into
2133 : * root->upper_targets[]. The core code doesn't use this, but it
2134 : * provides a convenient place for extensions to get at the info. For
2135 : * consistency, we save all the intermediate targets, even though some
2136 : * of the corresponding upperrels might not be needed for this query.
2137 : */
2138 386896 : root->upper_targets[UPPERREL_FINAL] = final_target;
2139 386896 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2140 386896 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2141 386896 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
2142 386896 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2143 386896 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2144 :
2145 : /*
2146 : * If we have grouping and/or aggregation, consider ways to implement
2147 : * that. We build a new upperrel representing the output of this
2148 : * phase.
2149 : */
2150 386896 : if (have_grouping)
2151 : {
2152 33942 : current_rel = create_grouping_paths(root,
2153 : current_rel,
2154 : grouping_target,
2155 : grouping_target_parallel_safe,
2156 : gset_data);
2157 : /* Fix things up if grouping_target contains SRFs */
2158 33938 : if (parse->hasTargetSRFs)
2159 299 : adjust_paths_for_srfs(root, current_rel,
2160 : grouping_targets,
2161 : grouping_targets_contain_srfs);
2162 : }
2163 :
2164 : /*
2165 : * If we have window functions, consider ways to implement those. We
2166 : * build a new upperrel representing the output of this phase.
2167 : */
2168 386892 : if (activeWindows)
2169 : {
2170 2232 : current_rel = create_window_paths(root,
2171 : current_rel,
2172 : grouping_target,
2173 : sort_input_target,
2174 : sort_input_target_parallel_safe,
2175 : wflists,
2176 : activeWindows);
2177 : /* Fix things up if sort_input_target contains SRFs */
2178 2232 : if (parse->hasTargetSRFs)
2179 10 : adjust_paths_for_srfs(root, current_rel,
2180 : sort_input_targets,
2181 : sort_input_targets_contain_srfs);
2182 : }
2183 :
2184 : /*
2185 : * If there is a DISTINCT clause, consider ways to implement that. We
2186 : * build a new upperrel representing the output of this phase.
2187 : */
2188 386892 : if (parse->distinctClause)
2189 : {
2190 1928 : current_rel = create_distinct_paths(root,
2191 : current_rel,
2192 : sort_input_target);
2193 : }
2194 : } /* end of if (setOperations) */
2195 :
2196 : /*
2197 : * If ORDER BY was given, consider ways to implement that, and generate a
2198 : * new upperrel containing only paths that emit the correct ordering and
2199 : * project the correct final_target. We can apply the original
2200 : * limit_tuples limit in sort costing here, but only if there are no
2201 : * postponed SRFs.
2202 : */
2203 391719 : if (parse->sortClause)
2204 : {
2205 60879 : current_rel = create_ordered_paths(root,
2206 : current_rel,
2207 : final_target,
2208 : final_target_parallel_safe,
2209 : have_postponed_srfs ? -1.0 :
2210 : limit_tuples);
2211 : /* Fix things up if final_target contains SRFs */
2212 60879 : if (parse->hasTargetSRFs)
2213 182 : adjust_paths_for_srfs(root, current_rel,
2214 : final_targets,
2215 : final_targets_contain_srfs);
2216 : }
2217 :
2218 : /*
2219 : * Now we are prepared to build the final-output upperrel.
2220 : */
2221 391719 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2222 :
2223 : /*
2224 : * If the input rel is marked consider_parallel and there's nothing that's
2225 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2226 : * consider_parallel as well. Note that if the query has rowMarks or is
2227 : * not a SELECT, consider_parallel will be false for every relation in the
2228 : * query.
2229 : */
2230 526577 : if (current_rel->consider_parallel &&
2231 269696 : is_parallel_safe(root, parse->limitOffset) &&
2232 134838 : is_parallel_safe(root, parse->limitCount))
2233 134833 : final_rel->consider_parallel = true;
2234 :
2235 : /*
2236 : * If the current_rel belongs to a single FDW, so does the final_rel.
2237 : */
2238 391719 : final_rel->serverid = current_rel->serverid;
2239 391719 : final_rel->userid = current_rel->userid;
2240 391719 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2241 391719 : final_rel->fdwroutine = current_rel->fdwroutine;
2242 :
2243 : /*
2244 : * Generate paths for the final_rel. Insert all surviving paths, with
2245 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2246 : */
2247 799867 : foreach(lc, current_rel->pathlist)
2248 : {
2249 408148 : Path *path = (Path *) lfirst(lc);
2250 :
2251 : /*
2252 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2253 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2254 : * here. If there are only non-locking rowmarks, they should be
2255 : * handled by the ModifyTable node instead. However, root->rowMarks
2256 : * is what goes into the LockRows node.)
2257 : */
2258 408148 : if (parse->rowMarks)
2259 : {
2260 6812 : path = (Path *) create_lockrows_path(root, final_rel, path,
2261 : root->rowMarks,
2262 : assign_special_exec_param(root));
2263 : }
2264 :
2265 : /*
2266 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2267 : */
2268 408148 : if (limit_needed(parse))
2269 : {
2270 4203 : path = (Path *) create_limit_path(root, final_rel, path,
2271 : parse->limitOffset,
2272 : parse->limitCount,
2273 : parse->limitOption,
2274 : offset_est, count_est);
2275 : }
2276 :
2277 : /*
2278 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
2279 : */
2280 408148 : if (parse->commandType != CMD_SELECT)
2281 : {
2282 : Index rootRelation;
2283 64666 : List *resultRelations = NIL;
2284 64666 : List *updateColnosLists = NIL;
2285 64666 : List *withCheckOptionLists = NIL;
2286 64666 : List *returningLists = NIL;
2287 64666 : List *mergeActionLists = NIL;
2288 64666 : List *mergeJoinConditions = NIL;
2289 : List *rowMarks;
2290 :
2291 64666 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
2292 : {
2293 : /* Inherited UPDATE/DELETE/MERGE */
2294 2320 : RelOptInfo *top_result_rel = find_base_rel(root,
2295 : parse->resultRelation);
2296 2320 : int resultRelation = -1;
2297 :
2298 : /* Pass the root result rel forward to the executor. */
2299 2320 : rootRelation = parse->resultRelation;
2300 :
2301 : /* Add only leaf children to ModifyTable. */
2302 6804 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
2303 6804 : resultRelation)) >= 0)
2304 : {
2305 4484 : RelOptInfo *this_result_rel = find_base_rel(root,
2306 : resultRelation);
2307 :
2308 : /*
2309 : * Also exclude any leaf rels that have turned dummy since
2310 : * being added to the list, for example, by being excluded
2311 : * by constraint exclusion.
2312 : */
2313 4484 : if (IS_DUMMY_REL(this_result_rel))
2314 153 : continue;
2315 :
2316 : /* Build per-target-rel lists needed by ModifyTable */
2317 4331 : resultRelations = lappend_int(resultRelations,
2318 : resultRelation);
2319 4331 : if (parse->commandType == CMD_UPDATE)
2320 : {
2321 2962 : List *update_colnos = root->update_colnos;
2322 :
2323 2962 : if (this_result_rel != top_result_rel)
2324 : update_colnos =
2325 2962 : adjust_inherited_attnums_multilevel(root,
2326 : update_colnos,
2327 : this_result_rel->relid,
2328 : top_result_rel->relid);
2329 2962 : updateColnosLists = lappend(updateColnosLists,
2330 : update_colnos);
2331 : }
2332 4331 : if (parse->withCheckOptions)
2333 : {
2334 416 : List *withCheckOptions = parse->withCheckOptions;
2335 :
2336 416 : if (this_result_rel != top_result_rel)
2337 : withCheckOptions = (List *)
2338 416 : adjust_appendrel_attrs_multilevel(root,
2339 : (Node *) withCheckOptions,
2340 : this_result_rel,
2341 : top_result_rel);
2342 416 : withCheckOptionLists = lappend(withCheckOptionLists,
2343 : withCheckOptions);
2344 : }
2345 4331 : if (parse->returningList)
2346 : {
2347 659 : List *returningList = parse->returningList;
2348 :
2349 659 : if (this_result_rel != top_result_rel)
2350 : returningList = (List *)
2351 659 : adjust_appendrel_attrs_multilevel(root,
2352 : (Node *) returningList,
2353 : this_result_rel,
2354 : top_result_rel);
2355 659 : returningLists = lappend(returningLists,
2356 : returningList);
2357 : }
2358 4331 : if (parse->mergeActionList)
2359 : {
2360 : ListCell *l;
2361 425 : List *mergeActionList = NIL;
2362 :
2363 : /*
2364 : * Copy MergeActions and translate stuff that
2365 : * references attribute numbers.
2366 : */
2367 1322 : foreach(l, parse->mergeActionList)
2368 : {
2369 897 : MergeAction *action = lfirst(l),
2370 897 : *leaf_action = copyObject(action);
2371 :
2372 897 : leaf_action->qual =
2373 897 : adjust_appendrel_attrs_multilevel(root,
2374 : (Node *) action->qual,
2375 : this_result_rel,
2376 : top_result_rel);
2377 897 : leaf_action->targetList = (List *)
2378 897 : adjust_appendrel_attrs_multilevel(root,
2379 897 : (Node *) action->targetList,
2380 : this_result_rel,
2381 : top_result_rel);
2382 897 : if (leaf_action->commandType == CMD_UPDATE)
2383 491 : leaf_action->updateColnos =
2384 491 : adjust_inherited_attnums_multilevel(root,
2385 : action->updateColnos,
2386 : this_result_rel->relid,
2387 : top_result_rel->relid);
2388 897 : mergeActionList = lappend(mergeActionList,
2389 : leaf_action);
2390 : }
2391 :
2392 425 : mergeActionLists = lappend(mergeActionLists,
2393 : mergeActionList);
2394 : }
2395 4331 : if (parse->commandType == CMD_MERGE)
2396 : {
2397 425 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2398 :
2399 425 : if (this_result_rel != top_result_rel)
2400 : mergeJoinCondition =
2401 425 : adjust_appendrel_attrs_multilevel(root,
2402 : mergeJoinCondition,
2403 : this_result_rel,
2404 : top_result_rel);
2405 425 : mergeJoinConditions = lappend(mergeJoinConditions,
2406 : mergeJoinCondition);
2407 : }
2408 : }
2409 :
2410 2320 : if (resultRelations == NIL)
2411 : {
2412 : /*
2413 : * We managed to exclude every child rel, so generate a
2414 : * dummy one-relation plan using info for the top target
2415 : * rel (even though that may not be a leaf target).
2416 : * Although it's clear that no data will be updated or
2417 : * deleted, we still need to have a ModifyTable node so
2418 : * that any statement triggers will be executed. (This
2419 : * could be cleaner if we fixed nodeModifyTable.c to allow
2420 : * zero target relations, but that probably wouldn't be a
2421 : * net win.)
2422 : */
2423 28 : resultRelations = list_make1_int(parse->resultRelation);
2424 28 : if (parse->commandType == CMD_UPDATE)
2425 26 : updateColnosLists = list_make1(root->update_colnos);
2426 28 : if (parse->withCheckOptions)
2427 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2428 28 : if (parse->returningList)
2429 15 : returningLists = list_make1(parse->returningList);
2430 28 : if (parse->mergeActionList)
2431 1 : mergeActionLists = list_make1(parse->mergeActionList);
2432 28 : if (parse->commandType == CMD_MERGE)
2433 1 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2434 : }
2435 : }
2436 : else
2437 : {
2438 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2439 62346 : rootRelation = 0; /* there's no separate root rel */
2440 62346 : resultRelations = list_make1_int(parse->resultRelation);
2441 62346 : if (parse->commandType == CMD_UPDATE)
2442 8842 : updateColnosLists = list_make1(root->update_colnos);
2443 62346 : if (parse->withCheckOptions)
2444 889 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2445 62346 : if (parse->returningList)
2446 2209 : returningLists = list_make1(parse->returningList);
2447 62346 : if (parse->mergeActionList)
2448 1287 : mergeActionLists = list_make1(parse->mergeActionList);
2449 62346 : if (parse->commandType == CMD_MERGE)
2450 1287 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2451 : }
2452 :
2453 : /*
2454 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2455 : * will have dealt with fetching non-locked marked rows, else we
2456 : * need to have ModifyTable do that.
2457 : */
2458 64666 : if (parse->rowMarks)
2459 0 : rowMarks = NIL;
2460 : else
2461 64666 : rowMarks = root->rowMarks;
2462 :
2463 : path = (Path *)
2464 64666 : create_modifytable_path(root, final_rel,
2465 : path,
2466 : parse->commandType,
2467 64666 : parse->canSetTag,
2468 64666 : parse->resultRelation,
2469 : rootRelation,
2470 : resultRelations,
2471 : updateColnosLists,
2472 : withCheckOptionLists,
2473 : returningLists,
2474 : rowMarks,
2475 : parse->onConflict,
2476 : mergeActionLists,
2477 : mergeJoinConditions,
2478 : parse->forPortionOf,
2479 : assign_special_exec_param(root));
2480 : }
2481 :
2482 : /* And shove it into final_rel */
2483 408148 : add_path(final_rel, path);
2484 : }
2485 :
2486 : /*
2487 : * Generate partial paths for final_rel, too, if outer query levels might
2488 : * be able to make use of them.
2489 : */
2490 391719 : if (final_rel->consider_parallel && root->query_level > 1 &&
2491 23068 : !limit_needed(parse))
2492 : {
2493 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2494 23002 : foreach(lc, current_rel->partial_pathlist)
2495 : {
2496 112 : Path *partial_path = (Path *) lfirst(lc);
2497 :
2498 112 : add_partial_path(final_rel, partial_path);
2499 : }
2500 : }
2501 :
2502 391719 : extra.limit_needed = limit_needed(parse);
2503 391719 : extra.limit_tuples = limit_tuples;
2504 391719 : extra.count_est = count_est;
2505 391719 : extra.offset_est = offset_est;
2506 :
2507 : /*
2508 : * If there is an FDW that's responsible for all baserels of the query,
2509 : * let it consider adding ForeignPaths.
2510 : */
2511 391719 : if (final_rel->fdwroutine &&
2512 658 : final_rel->fdwroutine->GetForeignUpperPaths)
2513 621 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2514 : current_rel, final_rel,
2515 : &extra);
2516 :
2517 : /* Let extensions possibly add some more paths */
2518 391719 : if (create_upper_paths_hook)
2519 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2520 : current_rel, final_rel, &extra);
2521 :
2522 : /* Note: currently, we leave it to callers to do set_cheapest() */
2523 391719 : }
2524 :
2525 : /*
2526 : * Do preprocessing for groupingSets clause and related data.
2527 : *
2528 : * We expect that parse->groupingSets has already been expanded into a flat
2529 : * list of grouping sets (that is, just integer Lists of ressortgroupref
2530 : * numbers) by expand_grouping_sets(). This function handles the preliminary
2531 : * steps of organizing the grouping sets into lists of rollups, and preparing
2532 : * annotations which will later be filled in with size estimates.
2533 : */
2534 : static grouping_sets_data *
2535 915 : preprocess_grouping_sets(PlannerInfo *root)
2536 : {
2537 915 : Query *parse = root->parse;
2538 : List *sets;
2539 915 : int maxref = 0;
2540 : ListCell *lc_set;
2541 915 : grouping_sets_data *gd = palloc0_object(grouping_sets_data);
2542 :
2543 : /*
2544 : * We don't currently make any attempt to optimize the groupClause when
2545 : * there are grouping sets, so just duplicate it in processed_groupClause.
2546 : */
2547 915 : root->processed_groupClause = parse->groupClause;
2548 :
2549 : /* Detect unhashable and unsortable grouping expressions */
2550 915 : gd->any_hashable = false;
2551 915 : gd->unhashable_refs = NULL;
2552 915 : gd->unsortable_refs = NULL;
2553 915 : gd->unsortable_sets = NIL;
2554 :
2555 915 : if (parse->groupClause)
2556 : {
2557 : ListCell *lc;
2558 :
2559 2651 : foreach(lc, parse->groupClause)
2560 : {
2561 1811 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2562 1811 : Index ref = gc->tleSortGroupRef;
2563 :
2564 1811 : if (ref > maxref)
2565 1771 : maxref = ref;
2566 :
2567 1811 : if (!gc->hashable)
2568 24 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2569 :
2570 1811 : if (!OidIsValid(gc->sortop))
2571 33 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2572 : }
2573 : }
2574 :
2575 : /* Allocate workspace array for remapping */
2576 915 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2577 :
2578 : /*
2579 : * If we have any unsortable sets, we must extract them before trying to
2580 : * prepare rollups. Unsortable sets don't go through
2581 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2582 : * here.
2583 : */
2584 915 : if (!bms_is_empty(gd->unsortable_refs))
2585 : {
2586 33 : List *sortable_sets = NIL;
2587 : ListCell *lc;
2588 :
2589 99 : foreach(lc, parse->groupingSets)
2590 : {
2591 70 : List *gset = (List *) lfirst(lc);
2592 :
2593 70 : if (bms_overlap_list(gd->unsortable_refs, gset))
2594 : {
2595 38 : GroupingSetData *gs = makeNode(GroupingSetData);
2596 :
2597 38 : gs->set = gset;
2598 38 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2599 :
2600 : /*
2601 : * We must enforce here that an unsortable set is hashable;
2602 : * later code assumes this. Parse analysis only checks that
2603 : * every individual column is either hashable or sortable.
2604 : *
2605 : * Note that passing this test doesn't guarantee we can
2606 : * generate a plan; there might be other showstoppers.
2607 : */
2608 38 : if (bms_overlap_list(gd->unhashable_refs, gset))
2609 4 : ereport(ERROR,
2610 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2611 : errmsg("could not implement GROUP BY"),
2612 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2613 : }
2614 : else
2615 32 : sortable_sets = lappend(sortable_sets, gset);
2616 : }
2617 :
2618 29 : if (sortable_sets)
2619 24 : sets = extract_rollup_sets(sortable_sets);
2620 : else
2621 5 : sets = NIL;
2622 : }
2623 : else
2624 882 : sets = extract_rollup_sets(parse->groupingSets);
2625 :
2626 2396 : foreach(lc_set, sets)
2627 : {
2628 1485 : List *current_sets = (List *) lfirst(lc_set);
2629 1485 : RollupData *rollup = makeNode(RollupData);
2630 : GroupingSetData *gs;
2631 :
2632 : /*
2633 : * Reorder the current list of grouping sets into correct prefix
2634 : * order. If only one aggregation pass is needed, try to make the
2635 : * list match the ORDER BY clause; if more than one pass is needed, we
2636 : * don't bother with that.
2637 : *
2638 : * Note that this reorders the sets from smallest-member-first to
2639 : * largest-member-first, and applies the GroupingSetData annotations,
2640 : * though the data will be filled in later.
2641 : */
2642 1485 : current_sets = reorder_grouping_sets(current_sets,
2643 1485 : (list_length(sets) == 1
2644 : ? parse->sortClause
2645 : : NIL));
2646 :
2647 : /*
2648 : * Get the initial (and therefore largest) grouping set.
2649 : */
2650 1485 : gs = linitial_node(GroupingSetData, current_sets);
2651 :
2652 : /*
2653 : * Order the groupClause appropriately. If the first grouping set is
2654 : * empty, then the groupClause must also be empty; otherwise we have
2655 : * to force the groupClause to match that grouping set's order.
2656 : *
2657 : * (The first grouping set can be empty even though parse->groupClause
2658 : * is not empty only if all non-empty grouping sets are unsortable.
2659 : * The groupClauses for hashed grouping sets are built later on.)
2660 : */
2661 1485 : if (gs->set)
2662 1410 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2663 : else
2664 75 : rollup->groupClause = NIL;
2665 :
2666 : /*
2667 : * Is it hashable? We pretend empty sets are hashable even though we
2668 : * actually force them not to be hashed later. But don't bother if
2669 : * there's nothing but empty sets (since in that case we can't hash
2670 : * anything).
2671 : */
2672 1485 : if (gs->set &&
2673 1410 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2674 : {
2675 1390 : rollup->hashable = true;
2676 1390 : gd->any_hashable = true;
2677 : }
2678 :
2679 : /*
2680 : * Now that we've pinned down an order for the groupClause for this
2681 : * list of grouping sets, we need to remap the entries in the grouping
2682 : * sets from sortgrouprefs to plain indices (0-based) into the
2683 : * groupClause for this collection of grouping sets. We keep the
2684 : * original form for later use, though.
2685 : */
2686 1485 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2687 : current_sets,
2688 : gd->tleref_to_colnum_map);
2689 1485 : rollup->gsets_data = current_sets;
2690 :
2691 1485 : gd->rollups = lappend(gd->rollups, rollup);
2692 : }
2693 :
2694 911 : if (gd->unsortable_sets)
2695 : {
2696 : /*
2697 : * We have not yet pinned down a groupclause for this, but we will
2698 : * need index-based lists for estimation purposes. Construct
2699 : * hash_sets_idx based on the entire original groupclause for now.
2700 : */
2701 29 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2702 : gd->unsortable_sets,
2703 : gd->tleref_to_colnum_map);
2704 29 : gd->any_hashable = true;
2705 : }
2706 :
2707 911 : return gd;
2708 : }
2709 :
2710 : /*
2711 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2712 : * (without annotation) mapped to indexes into the given groupclause.
2713 : */
2714 : static List *
2715 4157 : remap_to_groupclause_idx(List *groupClause,
2716 : List *gsets,
2717 : int *tleref_to_colnum_map)
2718 : {
2719 4157 : int ref = 0;
2720 4157 : List *result = NIL;
2721 : ListCell *lc;
2722 :
2723 9871 : foreach(lc, groupClause)
2724 : {
2725 5714 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2726 :
2727 5714 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2728 : }
2729 :
2730 9464 : foreach(lc, gsets)
2731 : {
2732 5307 : List *set = NIL;
2733 : ListCell *lc2;
2734 5307 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2735 :
2736 11730 : foreach(lc2, gs->set)
2737 : {
2738 6423 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2739 : }
2740 :
2741 5307 : result = lappend(result, set);
2742 : }
2743 :
2744 4157 : return result;
2745 : }
2746 :
2747 :
2748 : /*
2749 : * preprocess_rowmarks - set up PlanRowMarks if needed
2750 : */
2751 : static void
2752 394435 : preprocess_rowmarks(PlannerInfo *root)
2753 : {
2754 394435 : Query *parse = root->parse;
2755 : Bitmapset *rels;
2756 : List *prowmarks;
2757 : ListCell *l;
2758 : int i;
2759 :
2760 394435 : if (parse->rowMarks)
2761 : {
2762 : /*
2763 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2764 : * grouping, since grouping renders a reference to individual tuple
2765 : * CTIDs invalid. This is also checked at parse time, but that's
2766 : * insufficient because of rule substitution, query pullup, etc.
2767 : */
2768 6565 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2769 : parse->rowMarks)->strength);
2770 : }
2771 : else
2772 : {
2773 : /*
2774 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2775 : * UPDATE/SHARE.
2776 : */
2777 387870 : if (parse->commandType != CMD_UPDATE &&
2778 377408 : parse->commandType != CMD_DELETE &&
2779 373666 : parse->commandType != CMD_MERGE)
2780 372185 : return;
2781 : }
2782 :
2783 : /*
2784 : * We need to have rowmarks for all base relations except the target. We
2785 : * make a bitmapset of all base rels and then remove the items we don't
2786 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2787 : */
2788 22250 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2789 22250 : if (parse->resultRelation)
2790 15685 : rels = bms_del_member(rels, parse->resultRelation);
2791 :
2792 : /*
2793 : * Convert RowMarkClauses to PlanRowMark representation.
2794 : */
2795 22250 : prowmarks = NIL;
2796 28950 : foreach(l, parse->rowMarks)
2797 : {
2798 6700 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2799 6700 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2800 : PlanRowMark *newrc;
2801 :
2802 : /*
2803 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2804 : * applied to an update/delete target rel. If that ever becomes
2805 : * possible, we should drop the target from the PlanRowMark list.
2806 : */
2807 : Assert(rc->rti != parse->resultRelation);
2808 :
2809 : /*
2810 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2811 : * can't support true locking. Subqueries that got flattened into the
2812 : * main query should be ignored completely. Any that didn't will get
2813 : * ROW_MARK_COPY items in the next loop.
2814 : */
2815 6700 : if (rte->rtekind != RTE_RELATION)
2816 46 : continue;
2817 :
2818 6654 : rels = bms_del_member(rels, rc->rti);
2819 :
2820 6654 : newrc = makeNode(PlanRowMark);
2821 6654 : newrc->rti = newrc->prti = rc->rti;
2822 6654 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2823 6654 : newrc->markType = select_rowmark_type(rte, rc->strength);
2824 6654 : newrc->allMarkTypes = (1 << newrc->markType);
2825 6654 : newrc->strength = rc->strength;
2826 6654 : newrc->waitPolicy = rc->waitPolicy;
2827 6654 : newrc->isParent = false;
2828 :
2829 6654 : prowmarks = lappend(prowmarks, newrc);
2830 : }
2831 :
2832 : /*
2833 : * Now, add rowmarks for any non-target, non-locked base relations.
2834 : */
2835 22250 : i = 0;
2836 53110 : foreach(l, parse->rtable)
2837 : {
2838 30860 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2839 : PlanRowMark *newrc;
2840 :
2841 30860 : i++;
2842 30860 : if (!bms_is_member(i, rels))
2843 27911 : continue;
2844 :
2845 2949 : newrc = makeNode(PlanRowMark);
2846 2949 : newrc->rti = newrc->prti = i;
2847 2949 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2848 2949 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2849 2949 : newrc->allMarkTypes = (1 << newrc->markType);
2850 2949 : newrc->strength = LCS_NONE;
2851 2949 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2852 2949 : newrc->isParent = false;
2853 :
2854 2949 : prowmarks = lappend(prowmarks, newrc);
2855 : }
2856 :
2857 22250 : root->rowMarks = prowmarks;
2858 : }
2859 :
2860 : /*
2861 : * Select RowMarkType to use for a given table
2862 : */
2863 : RowMarkType
2864 11497 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2865 : {
2866 11497 : if (rte->rtekind != RTE_RELATION)
2867 : {
2868 : /* If it's not a table at all, use ROW_MARK_COPY */
2869 1110 : return ROW_MARK_COPY;
2870 : }
2871 10387 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2872 : {
2873 : /* Let the FDW select the rowmark type, if it wants to */
2874 114 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2875 :
2876 114 : if (fdwroutine->GetForeignRowMarkType != NULL)
2877 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2878 : /* Otherwise, use ROW_MARK_COPY by default */
2879 114 : return ROW_MARK_COPY;
2880 : }
2881 : else
2882 : {
2883 : /* Regular table, apply the appropriate lock type */
2884 10273 : switch (strength)
2885 : {
2886 2048 : case LCS_NONE:
2887 :
2888 : /*
2889 : * We don't need a tuple lock, only the ability to re-fetch
2890 : * the row.
2891 : */
2892 2048 : return ROW_MARK_REFERENCE;
2893 : break;
2894 7055 : case LCS_FORKEYSHARE:
2895 7055 : return ROW_MARK_KEYSHARE;
2896 : break;
2897 201 : case LCS_FORSHARE:
2898 201 : return ROW_MARK_SHARE;
2899 : break;
2900 41 : case LCS_FORNOKEYUPDATE:
2901 41 : return ROW_MARK_NOKEYEXCLUSIVE;
2902 : break;
2903 928 : case LCS_FORUPDATE:
2904 928 : return ROW_MARK_EXCLUSIVE;
2905 : break;
2906 : }
2907 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2908 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2909 : }
2910 : }
2911 :
2912 : /*
2913 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2914 : *
2915 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2916 : * results back in *count_est and *offset_est. These variables are set to
2917 : * 0 if the corresponding clause is not present, and -1 if it's present
2918 : * but we couldn't estimate the value for it. (The "0" convention is OK
2919 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2920 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2921 : * usual practice of never estimating less than one row.) These values will
2922 : * be passed to create_limit_path, which see if you change this code.
2923 : *
2924 : * The return value is the suitably adjusted tuple_fraction to use for
2925 : * planning the query. This adjustment is not overridable, since it reflects
2926 : * plan actions that grouping_planner() will certainly take, not assumptions
2927 : * about context.
2928 : */
2929 : static double
2930 3717 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2931 : int64 *offset_est, int64 *count_est)
2932 : {
2933 3717 : Query *parse = root->parse;
2934 : Node *est;
2935 : double limit_fraction;
2936 :
2937 : /* Should not be called unless LIMIT or OFFSET */
2938 : Assert(parse->limitCount || parse->limitOffset);
2939 :
2940 : /*
2941 : * Try to obtain the clause values. We use estimate_expression_value
2942 : * primarily because it can sometimes do something useful with Params.
2943 : */
2944 3717 : if (parse->limitCount)
2945 : {
2946 3275 : est = estimate_expression_value(root, parse->limitCount);
2947 3275 : if (est && IsA(est, Const))
2948 : {
2949 3270 : if (((Const *) est)->constisnull)
2950 : {
2951 : /* NULL indicates LIMIT ALL, ie, no limit */
2952 0 : *count_est = 0; /* treat as not present */
2953 : }
2954 : else
2955 : {
2956 3270 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2957 3270 : if (*count_est <= 0)
2958 190 : *count_est = 1; /* force to at least 1 */
2959 : }
2960 : }
2961 : else
2962 5 : *count_est = -1; /* can't estimate */
2963 : }
2964 : else
2965 442 : *count_est = 0; /* not present */
2966 :
2967 3717 : if (parse->limitOffset)
2968 : {
2969 652 : est = estimate_expression_value(root, parse->limitOffset);
2970 652 : if (est && IsA(est, Const))
2971 : {
2972 632 : if (((Const *) est)->constisnull)
2973 : {
2974 : /* Treat NULL as no offset; the executor will too */
2975 0 : *offset_est = 0; /* treat as not present */
2976 : }
2977 : else
2978 : {
2979 632 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2980 632 : if (*offset_est < 0)
2981 0 : *offset_est = 0; /* treat as not present */
2982 : }
2983 : }
2984 : else
2985 20 : *offset_est = -1; /* can't estimate */
2986 : }
2987 : else
2988 3065 : *offset_est = 0; /* not present */
2989 :
2990 3717 : if (*count_est != 0)
2991 : {
2992 : /*
2993 : * A LIMIT clause limits the absolute number of tuples returned.
2994 : * However, if it's not a constant LIMIT then we have to guess; for
2995 : * lack of a better idea, assume 10% of the plan's result is wanted.
2996 : */
2997 3275 : if (*count_est < 0 || *offset_est < 0)
2998 : {
2999 : /* LIMIT or OFFSET is an expression ... punt ... */
3000 20 : limit_fraction = 0.10;
3001 : }
3002 : else
3003 : {
3004 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
3005 3255 : limit_fraction = (double) *count_est + (double) *offset_est;
3006 : }
3007 :
3008 : /*
3009 : * If we have absolute limits from both caller and LIMIT, use the
3010 : * smaller value; likewise if they are both fractional. If one is
3011 : * fractional and the other absolute, we can't easily determine which
3012 : * is smaller, but we use the heuristic that the absolute will usually
3013 : * be smaller.
3014 : */
3015 3275 : if (tuple_fraction >= 1.0)
3016 : {
3017 5 : if (limit_fraction >= 1.0)
3018 : {
3019 : /* both absolute */
3020 5 : tuple_fraction = Min(tuple_fraction, limit_fraction);
3021 : }
3022 : else
3023 : {
3024 : /* caller absolute, limit fractional; use caller's value */
3025 : }
3026 : }
3027 3270 : else if (tuple_fraction > 0.0)
3028 : {
3029 82 : if (limit_fraction >= 1.0)
3030 : {
3031 : /* caller fractional, limit absolute; use limit */
3032 82 : tuple_fraction = limit_fraction;
3033 : }
3034 : else
3035 : {
3036 : /* both fractional */
3037 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
3038 : }
3039 : }
3040 : else
3041 : {
3042 : /* no info from caller, just use limit */
3043 3188 : tuple_fraction = limit_fraction;
3044 : }
3045 : }
3046 442 : else if (*offset_est != 0 && tuple_fraction > 0.0)
3047 : {
3048 : /*
3049 : * We have an OFFSET but no LIMIT. This acts entirely differently
3050 : * from the LIMIT case: here, we need to increase rather than decrease
3051 : * the caller's tuple_fraction, because the OFFSET acts to cause more
3052 : * tuples to be fetched instead of fewer. This only matters if we got
3053 : * a tuple_fraction > 0, however.
3054 : *
3055 : * As above, use 10% if OFFSET is present but unestimatable.
3056 : */
3057 12 : if (*offset_est < 0)
3058 0 : limit_fraction = 0.10;
3059 : else
3060 12 : limit_fraction = (double) *offset_est;
3061 :
3062 : /*
3063 : * If we have absolute counts from both caller and OFFSET, add them
3064 : * together; likewise if they are both fractional. If one is
3065 : * fractional and the other absolute, we want to take the larger, and
3066 : * we heuristically assume that's the fractional one.
3067 : */
3068 12 : if (tuple_fraction >= 1.0)
3069 : {
3070 0 : if (limit_fraction >= 1.0)
3071 : {
3072 : /* both absolute, so add them together */
3073 0 : tuple_fraction += limit_fraction;
3074 : }
3075 : else
3076 : {
3077 : /* caller absolute, limit fractional; use limit */
3078 0 : tuple_fraction = limit_fraction;
3079 : }
3080 : }
3081 : else
3082 : {
3083 12 : if (limit_fraction >= 1.0)
3084 : {
3085 : /* caller fractional, limit absolute; use caller's value */
3086 : }
3087 : else
3088 : {
3089 : /* both fractional, so add them together */
3090 0 : tuple_fraction += limit_fraction;
3091 0 : if (tuple_fraction >= 1.0)
3092 0 : tuple_fraction = 0.0; /* assume fetch all */
3093 : }
3094 : }
3095 : }
3096 :
3097 3717 : return tuple_fraction;
3098 : }
3099 :
3100 : /*
3101 : * limit_needed - do we actually need a Limit plan node?
3102 : *
3103 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
3104 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
3105 : * locution for an optimization fence. (Because other places in the planner
3106 : * merely check whether parse->limitOffset isn't NULL, it will still work as
3107 : * an optimization fence --- we're just suppressing unnecessary run-time
3108 : * overhead.)
3109 : *
3110 : * This might look like it could be merged into preprocess_limit, but there's
3111 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3112 : * in preprocess_limit it's good enough to consider estimated values.
3113 : */
3114 : bool
3115 838453 : limit_needed(Query *parse)
3116 : {
3117 : Node *node;
3118 :
3119 838453 : node = parse->limitCount;
3120 838453 : if (node)
3121 : {
3122 7796 : if (IsA(node, Const))
3123 : {
3124 : /* NULL indicates LIMIT ALL, ie, no limit */
3125 7575 : if (!((Const *) node)->constisnull)
3126 7575 : return true; /* LIMIT with a constant value */
3127 : }
3128 : else
3129 221 : return true; /* non-constant LIMIT */
3130 : }
3131 :
3132 830657 : node = parse->limitOffset;
3133 830657 : if (node)
3134 : {
3135 1283 : if (IsA(node, Const))
3136 : {
3137 : /* Treat NULL as no offset; the executor would too */
3138 1025 : if (!((Const *) node)->constisnull)
3139 : {
3140 1025 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3141 :
3142 1025 : if (offset != 0)
3143 107 : return true; /* OFFSET with a nonzero value */
3144 : }
3145 : }
3146 : else
3147 258 : return true; /* non-constant OFFSET */
3148 : }
3149 :
3150 830292 : return false; /* don't need a Limit plan node */
3151 : }
3152 :
3153 : /*
3154 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3155 : *
3156 : * The idea here is to adjust the ordering of the GROUP BY elements
3157 : * (which in itself is semantically insignificant) to match ORDER BY,
3158 : * thereby allowing a single sort operation to both implement the ORDER BY
3159 : * requirement and set up for a Unique step that implements GROUP BY.
3160 : * We also consider partial match between GROUP BY and ORDER BY elements,
3161 : * which could allow to implement ORDER BY using the incremental sort.
3162 : *
3163 : * We also consider other orderings of the GROUP BY elements, which could
3164 : * match the sort ordering of other possible plans (eg an indexscan) and
3165 : * thereby reduce cost. This is implemented during the generation of grouping
3166 : * paths. See get_useful_group_keys_orderings() for details.
3167 : *
3168 : * Note: we need no comparable processing of the distinctClause because
3169 : * the parser already enforced that that matches ORDER BY.
3170 : *
3171 : * Note: we return a fresh List, but its elements are the same
3172 : * SortGroupClauses appearing in parse->groupClause. This is important
3173 : * because later processing may modify the processed_groupClause list.
3174 : *
3175 : * For grouping sets, the order of items is instead forced to agree with that
3176 : * of the grouping set (and items not in the grouping set are skipped). The
3177 : * work of sorting the order of grouping set elements to match the ORDER BY if
3178 : * possible is done elsewhere.
3179 : */
3180 : static List *
3181 7561 : preprocess_groupclause(PlannerInfo *root, List *force)
3182 : {
3183 7561 : Query *parse = root->parse;
3184 7561 : List *new_groupclause = NIL;
3185 : ListCell *sl;
3186 : ListCell *gl;
3187 :
3188 : /* For grouping sets, we need to force the ordering */
3189 7561 : if (force)
3190 : {
3191 9704 : foreach(sl, force)
3192 : {
3193 5651 : Index ref = lfirst_int(sl);
3194 5651 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3195 :
3196 5651 : new_groupclause = lappend(new_groupclause, cl);
3197 : }
3198 :
3199 4053 : return new_groupclause;
3200 : }
3201 :
3202 : /* If no ORDER BY, nothing useful to do here */
3203 3508 : if (parse->sortClause == NIL)
3204 1984 : return list_copy(parse->groupClause);
3205 :
3206 : /*
3207 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3208 : * items, but only as far as we can make a matching prefix.
3209 : *
3210 : * This code assumes that the sortClause contains no duplicate items.
3211 : */
3212 3000 : foreach(sl, parse->sortClause)
3213 : {
3214 1992 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3215 :
3216 2899 : foreach(gl, parse->groupClause)
3217 : {
3218 2383 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3219 :
3220 2383 : if (equal(gc, sc))
3221 : {
3222 1476 : new_groupclause = lappend(new_groupclause, gc);
3223 1476 : break;
3224 : }
3225 : }
3226 1992 : if (gl == NULL)
3227 516 : break; /* no match, so stop scanning */
3228 : }
3229 :
3230 :
3231 : /* If no match at all, no point in reordering GROUP BY */
3232 1524 : if (new_groupclause == NIL)
3233 228 : return list_copy(parse->groupClause);
3234 :
3235 : /*
3236 : * Add any remaining GROUP BY items to the new list. We don't require a
3237 : * complete match, because even partial match allows ORDER BY to be
3238 : * implemented using incremental sort. Also, give up if there are any
3239 : * non-sortable GROUP BY items, since then there's no hope anyway.
3240 : */
3241 2909 : foreach(gl, parse->groupClause)
3242 : {
3243 1613 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3244 :
3245 1613 : if (list_member_ptr(new_groupclause, gc))
3246 1476 : continue; /* it matched an ORDER BY item */
3247 137 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
3248 0 : return list_copy(parse->groupClause);
3249 137 : new_groupclause = lappend(new_groupclause, gc);
3250 : }
3251 :
3252 : /* Success --- install the rearranged GROUP BY list */
3253 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3254 1296 : return new_groupclause;
3255 : }
3256 :
3257 : /*
3258 : * Extract lists of grouping sets that can be implemented using a single
3259 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3260 : *
3261 : * Input must be sorted with smallest sets first. Result has each sublist
3262 : * sorted with smallest sets first.
3263 : *
3264 : * We want to produce the absolute minimum possible number of lists here to
3265 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3266 : * of finding the minimal partition of a partially-ordered set into chains
3267 : * (which is what we need, taking the list of grouping sets as a poset ordered
3268 : * by set inclusion) can be mapped to the problem of finding the maximum
3269 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3270 : * time with a worst case of no worse than O(n^2.5) and usually much
3271 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3272 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3273 : * half a second on my modest system even with optimization off and assertions
3274 : * on.)
3275 : */
3276 : static List *
3277 906 : extract_rollup_sets(List *groupingSets)
3278 : {
3279 906 : int num_sets_raw = list_length(groupingSets);
3280 906 : int num_empty = 0;
3281 906 : int num_sets = 0; /* distinct sets */
3282 906 : int num_chains = 0;
3283 906 : List *result = NIL;
3284 : List **results;
3285 : List **orig_sets;
3286 : Bitmapset **set_masks;
3287 : int *chains;
3288 : short **adjacency;
3289 : short *adjacency_buf;
3290 : BipartiteMatchState *state;
3291 : int i;
3292 : int j;
3293 : int j_size;
3294 906 : ListCell *lc1 = list_head(groupingSets);
3295 : ListCell *lc;
3296 :
3297 : /*
3298 : * Start by stripping out empty sets. The algorithm doesn't require this,
3299 : * but the planner currently needs all empty sets to be returned in the
3300 : * first list, so we strip them here and add them back after.
3301 : */
3302 1506 : while (lc1 && lfirst(lc1) == NIL)
3303 : {
3304 600 : ++num_empty;
3305 600 : lc1 = lnext(groupingSets, lc1);
3306 : }
3307 :
3308 : /* bail out now if it turns out that all we had were empty sets. */
3309 906 : if (!lc1)
3310 75 : return list_make1(groupingSets);
3311 :
3312 : /*----------
3313 : * We don't strictly need to remove duplicate sets here, but if we don't,
3314 : * they tend to become scattered through the result, which is a bit
3315 : * confusing (and irritating if we ever decide to optimize them out).
3316 : * So we remove them here and add them back after.
3317 : *
3318 : * For each non-duplicate set, we fill in the following:
3319 : *
3320 : * orig_sets[i] = list of the original set lists
3321 : * set_masks[i] = bitmapset for testing inclusion
3322 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3323 : *
3324 : * chains[i] will be the result group this set is assigned to.
3325 : *
3326 : * We index all of these from 1 rather than 0 because it is convenient
3327 : * to leave 0 free for the NIL node in the graph algorithm.
3328 : *----------
3329 : */
3330 831 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3331 831 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3332 831 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3333 831 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3334 :
3335 831 : j_size = 0;
3336 831 : j = 0;
3337 831 : i = 1;
3338 :
3339 2861 : for_each_cell(lc, groupingSets, lc1)
3340 : {
3341 2030 : List *candidate = (List *) lfirst(lc);
3342 2030 : Bitmapset *candidate_set = NULL;
3343 : ListCell *lc2;
3344 2030 : int dup_of = 0;
3345 :
3346 4827 : foreach(lc2, candidate)
3347 : {
3348 2797 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3349 : }
3350 :
3351 : /* we can only be a dup if we're the same length as a previous set */
3352 2030 : if (j_size == list_length(candidate))
3353 : {
3354 : int k;
3355 :
3356 1828 : for (k = j; k < i; ++k)
3357 : {
3358 1169 : if (bms_equal(set_masks[k], candidate_set))
3359 : {
3360 129 : dup_of = k;
3361 129 : break;
3362 : }
3363 : }
3364 : }
3365 1242 : else if (j_size < list_length(candidate))
3366 : {
3367 1242 : j_size = list_length(candidate);
3368 1242 : j = i;
3369 : }
3370 :
3371 2030 : if (dup_of > 0)
3372 : {
3373 129 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3374 129 : bms_free(candidate_set);
3375 : }
3376 : else
3377 : {
3378 : int k;
3379 1901 : int n_adj = 0;
3380 :
3381 1901 : orig_sets[i] = list_make1(candidate);
3382 1901 : set_masks[i] = candidate_set;
3383 :
3384 : /* fill in adjacency list; no need to compare equal-size sets */
3385 :
3386 2973 : for (k = j - 1; k > 0; --k)
3387 : {
3388 1072 : if (bms_is_subset(set_masks[k], candidate_set))
3389 937 : adjacency_buf[++n_adj] = k;
3390 : }
3391 :
3392 1901 : if (n_adj > 0)
3393 : {
3394 513 : adjacency_buf[0] = n_adj;
3395 513 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3396 513 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3397 : }
3398 : else
3399 1388 : adjacency[i] = NULL;
3400 :
3401 1901 : ++i;
3402 : }
3403 : }
3404 :
3405 831 : num_sets = i - 1;
3406 :
3407 : /*
3408 : * Apply the graph matching algorithm to do the work.
3409 : */
3410 831 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3411 :
3412 : /*
3413 : * Now, the state->pair* fields have the info we need to assign sets to
3414 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3415 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3416 : * it in one pass)
3417 : */
3418 831 : chains = palloc0((num_sets + 1) * sizeof(int));
3419 :
3420 2732 : for (i = 1; i <= num_sets; ++i)
3421 : {
3422 1901 : int u = state->pair_vu[i];
3423 1901 : int v = state->pair_uv[i];
3424 :
3425 1901 : if (u > 0 && u < i)
3426 0 : chains[i] = chains[u];
3427 1901 : else if (v > 0 && v < i)
3428 491 : chains[i] = chains[v];
3429 : else
3430 1410 : chains[i] = ++num_chains;
3431 : }
3432 :
3433 : /* build result lists. */
3434 831 : results = palloc0((num_chains + 1) * sizeof(List *));
3435 :
3436 2732 : for (i = 1; i <= num_sets; ++i)
3437 : {
3438 1901 : int c = chains[i];
3439 :
3440 : Assert(c > 0);
3441 :
3442 1901 : results[c] = list_concat(results[c], orig_sets[i]);
3443 : }
3444 :
3445 : /* push any empty sets back on the first list. */
3446 1306 : while (num_empty-- > 0)
3447 475 : results[1] = lcons(NIL, results[1]);
3448 :
3449 : /* make result list */
3450 2241 : for (i = 1; i <= num_chains; ++i)
3451 1410 : result = lappend(result, results[i]);
3452 :
3453 : /*
3454 : * Free all the things.
3455 : *
3456 : * (This is over-fussy for small sets but for large sets we could have
3457 : * tied up a nontrivial amount of memory.)
3458 : */
3459 831 : BipartiteMatchFree(state);
3460 831 : pfree(results);
3461 831 : pfree(chains);
3462 2732 : for (i = 1; i <= num_sets; ++i)
3463 1901 : if (adjacency[i])
3464 513 : pfree(adjacency[i]);
3465 831 : pfree(adjacency);
3466 831 : pfree(adjacency_buf);
3467 831 : pfree(orig_sets);
3468 2732 : for (i = 1; i <= num_sets; ++i)
3469 1901 : bms_free(set_masks[i]);
3470 831 : pfree(set_masks);
3471 :
3472 831 : return result;
3473 : }
3474 :
3475 : /*
3476 : * Reorder the elements of a list of grouping sets such that they have correct
3477 : * prefix relationships. Also inserts the GroupingSetData annotations.
3478 : *
3479 : * The input must be ordered with smallest sets first; the result is returned
3480 : * with largest sets first. Note that the result shares no list substructure
3481 : * with the input, so it's safe for the caller to modify it later.
3482 : *
3483 : * If we're passed in a sortclause, we follow its order of columns to the
3484 : * extent possible, to minimize the chance that we add unnecessary sorts.
3485 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3486 : * gets implemented in one pass.)
3487 : */
3488 : static List *
3489 1485 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3490 : {
3491 : ListCell *lc;
3492 1485 : List *previous = NIL;
3493 1485 : List *result = NIL;
3494 :
3495 4115 : foreach(lc, groupingSets)
3496 : {
3497 2630 : List *candidate = (List *) lfirst(lc);
3498 2630 : List *new_elems = list_difference_int(candidate, previous);
3499 2630 : GroupingSetData *gs = makeNode(GroupingSetData);
3500 :
3501 2774 : while (list_length(sortclause) > list_length(previous) &&
3502 : new_elems != NIL)
3503 : {
3504 244 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3505 244 : int ref = sc->tleSortGroupRef;
3506 :
3507 244 : if (list_member_int(new_elems, ref))
3508 : {
3509 144 : previous = lappend_int(previous, ref);
3510 144 : new_elems = list_delete_int(new_elems, ref);
3511 : }
3512 : else
3513 : {
3514 : /* diverged from the sortclause; give up on it */
3515 100 : sortclause = NIL;
3516 100 : break;
3517 : }
3518 : }
3519 :
3520 2630 : previous = list_concat(previous, new_elems);
3521 :
3522 2630 : gs->set = list_copy(previous);
3523 2630 : result = lcons(gs, result);
3524 : }
3525 :
3526 1485 : list_free(previous);
3527 :
3528 1485 : return result;
3529 : }
3530 :
3531 : /*
3532 : * has_volatile_pathkey
3533 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3534 : * containing a volatile function. Otherwise returns false.
3535 : */
3536 : static bool
3537 2000 : has_volatile_pathkey(List *keys)
3538 : {
3539 : ListCell *lc;
3540 :
3541 4122 : foreach(lc, keys)
3542 : {
3543 2135 : PathKey *pathkey = lfirst_node(PathKey, lc);
3544 :
3545 2135 : if (pathkey->pk_eclass->ec_has_volatile)
3546 13 : return true;
3547 : }
3548 :
3549 1987 : return false;
3550 : }
3551 :
3552 : /*
3553 : * adjust_group_pathkeys_for_groupagg
3554 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3555 : * pre-ordered input for ordered aggregates.
3556 : *
3557 : * We define "best" as the pathkeys that suit the largest number of
3558 : * aggregate functions. We find these by looking at the first ORDER BY /
3559 : * DISTINCT aggregate and take the pathkeys for that before searching for
3560 : * other aggregates that require the same or a more strict variation of the
3561 : * same pathkeys. We then repeat that process for any remaining aggregates
3562 : * with different pathkeys and if we find another set of pathkeys that suits a
3563 : * larger number of aggregates then we select those pathkeys instead.
3564 : *
3565 : * When the best pathkeys are found we also mark each Aggref that can use
3566 : * those pathkeys as aggpresorted = true.
3567 : *
3568 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3569 : * volatile functions, we never make use of these pathkeys. We want to ensure
3570 : * that sorts using volatile functions are done independently in each Aggref
3571 : * rather than once at the query level. If we were to allow this then Aggrefs
3572 : * with compatible sort orders would all transition their rows in the same
3573 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3574 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3575 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3576 : * left to perform their sorts individually. To avoid this inconsistent
3577 : * behavior which could make Aggref results depend on what other Aggrefs the
3578 : * query contains, we always force Aggrefs with volatile functions to perform
3579 : * their own sorts.
3580 : */
3581 : static void
3582 1664 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3583 : {
3584 1664 : List *grouppathkeys = root->group_pathkeys;
3585 : List *bestpathkeys;
3586 : Bitmapset *bestaggs;
3587 : Bitmapset *unprocessed_aggs;
3588 : ListCell *lc;
3589 : int i;
3590 :
3591 : /* Shouldn't be here if there are grouping sets */
3592 : Assert(root->parse->groupingSets == NIL);
3593 : /* Shouldn't be here unless there are some ordered aggregates */
3594 : Assert(root->numOrderedAggs > 0);
3595 :
3596 : /* Do nothing if disabled */
3597 1664 : if (!enable_presorted_aggregate)
3598 5 : return;
3599 :
3600 : /*
3601 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3602 : * the indexes of all AggInfos to be processed below.
3603 : */
3604 1659 : unprocessed_aggs = NULL;
3605 3884 : foreach(lc, root->agginfos)
3606 : {
3607 2225 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3608 2225 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3609 :
3610 2225 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3611 214 : continue;
3612 :
3613 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3614 2011 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3615 248 : continue;
3616 :
3617 : /* Additional safety checks are needed if there's a FILTER clause */
3618 1763 : if (aggref->aggfilter != NULL)
3619 : {
3620 : ListCell *lc2;
3621 43 : bool allow_presort = true;
3622 :
3623 : /*
3624 : * When the Aggref has a FILTER clause, it's possible that the
3625 : * filter removes rows that cannot be sorted because the
3626 : * expression to sort by results in an error during its
3627 : * evaluation. This is a problem for presorting as that happens
3628 : * before the FILTER, whereas without presorting, the Aggregate
3629 : * node will apply the FILTER *before* sorting. So that we never
3630 : * try to sort anything that might error, here we aim to skip over
3631 : * any Aggrefs with arguments with expressions which, when
3632 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3633 : * may be more cases that should be allowed, but more thought
3634 : * needs to be given. Err on the side of caution.
3635 : */
3636 83 : foreach(lc2, aggref->args)
3637 : {
3638 58 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3639 58 : Expr *expr = tle->expr;
3640 :
3641 68 : while (IsA(expr, RelabelType))
3642 10 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3643 :
3644 : /* Common case, Vars and Consts are ok */
3645 58 : if (IsA(expr, Var) || IsA(expr, Const))
3646 40 : continue;
3647 :
3648 : /* Unsupported. Don't try to presort for this Aggref */
3649 18 : allow_presort = false;
3650 18 : break;
3651 : }
3652 :
3653 : /* Skip unsupported Aggrefs */
3654 43 : if (!allow_presort)
3655 18 : continue;
3656 : }
3657 :
3658 1745 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3659 : foreach_current_index(lc));
3660 : }
3661 :
3662 : /*
3663 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3664 : * for the given set of aggregates.
3665 : *
3666 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3667 : * this during the first loop using the pathkeys for the very first
3668 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3669 : * a more strict set of compatible pathkeys. Once the outer loop is
3670 : * complete, we mark off all the aggregates with compatible pathkeys then
3671 : * remove those from the unprocessed_aggs and repeat the process to try to
3672 : * find another set of pathkeys that are suitable for a larger number of
3673 : * aggregates. The outer loop will stop when there are not enough
3674 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3675 : * to suit a larger number of aggregates.
3676 : */
3677 1659 : bestpathkeys = NIL;
3678 1659 : bestaggs = NULL;
3679 3269 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3680 : {
3681 1610 : Bitmapset *aggindexes = NULL;
3682 1610 : List *currpathkeys = NIL;
3683 :
3684 1610 : i = -1;
3685 3610 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3686 : {
3687 2000 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3688 2000 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3689 : List *sortlist;
3690 : List *pathkeys;
3691 :
3692 2000 : if (aggref->aggdistinct != NIL)
3693 599 : sortlist = aggref->aggdistinct;
3694 : else
3695 1401 : sortlist = aggref->aggorder;
3696 :
3697 2000 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3698 : aggref->args);
3699 :
3700 : /*
3701 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3702 : * or DISTINCT clause.
3703 : */
3704 2000 : if (has_volatile_pathkey(pathkeys))
3705 : {
3706 13 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3707 13 : continue;
3708 : }
3709 :
3710 : /*
3711 : * When not set yet, take the pathkeys from the first unprocessed
3712 : * aggregate.
3713 : */
3714 1987 : if (currpathkeys == NIL)
3715 : {
3716 1607 : currpathkeys = pathkeys;
3717 :
3718 : /* include the GROUP BY pathkeys, if they exist */
3719 1607 : if (grouppathkeys != NIL)
3720 226 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3721 : currpathkeys);
3722 :
3723 : /* record that we found pathkeys for this aggregate */
3724 1607 : aggindexes = bms_add_member(aggindexes, i);
3725 : }
3726 : else
3727 : {
3728 : /* now look for a stronger set of matching pathkeys */
3729 :
3730 : /* include the GROUP BY pathkeys, if they exist */
3731 380 : if (grouppathkeys != NIL)
3732 240 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3733 : pathkeys);
3734 :
3735 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3736 380 : switch (compare_pathkeys(currpathkeys, pathkeys))
3737 : {
3738 10 : case PATHKEYS_BETTER2:
3739 : /* 'pathkeys' are stronger, use these ones instead */
3740 10 : currpathkeys = pathkeys;
3741 : pg_fallthrough;
3742 :
3743 55 : case PATHKEYS_BETTER1:
3744 : /* 'pathkeys' are less strict */
3745 : pg_fallthrough;
3746 :
3747 : case PATHKEYS_EQUAL:
3748 : /* mark this aggregate as covered by 'currpathkeys' */
3749 55 : aggindexes = bms_add_member(aggindexes, i);
3750 55 : break;
3751 :
3752 325 : case PATHKEYS_DIFFERENT:
3753 325 : break;
3754 : }
3755 : }
3756 : }
3757 :
3758 : /* remove the aggregates that we've just processed */
3759 1610 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3760 :
3761 : /*
3762 : * If this pass included more aggregates than the previous best then
3763 : * use these ones as the best set.
3764 : */
3765 1610 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3766 : {
3767 1522 : bestaggs = aggindexes;
3768 1522 : bestpathkeys = currpathkeys;
3769 : }
3770 : }
3771 :
3772 : /*
3773 : * If we found any ordered aggregates, update root->group_pathkeys to add
3774 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3775 : * the original GROUP BY pathkeys already.
3776 : */
3777 1659 : if (bestpathkeys != NIL)
3778 1472 : root->group_pathkeys = bestpathkeys;
3779 :
3780 : /*
3781 : * Now that we've found the best set of aggregates we can set the
3782 : * presorted flag to indicate to the executor that it needn't bother
3783 : * performing a sort for these Aggrefs. We're able to do this now as
3784 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3785 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3786 : * of ordered aggregates.
3787 : */
3788 1659 : i = -1;
3789 3211 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3790 : {
3791 1552 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3792 :
3793 3119 : foreach(lc, agginfo->aggrefs)
3794 : {
3795 1567 : Aggref *aggref = lfirst_node(Aggref, lc);
3796 :
3797 1567 : aggref->aggpresorted = true;
3798 : }
3799 : }
3800 : }
3801 :
3802 : /*
3803 : * Compute query_pathkeys and other pathkeys during plan generation
3804 : */
3805 : static void
3806 386919 : standard_qp_callback(PlannerInfo *root, void *extra)
3807 : {
3808 386919 : Query *parse = root->parse;
3809 386919 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3810 386919 : List *tlist = root->processed_tlist;
3811 386919 : List *activeWindows = qp_extra->activeWindows;
3812 :
3813 : /*
3814 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3815 : * aggregate requirements.
3816 : */
3817 386919 : if (qp_extra->gset_data)
3818 : {
3819 : /*
3820 : * With grouping sets, just use the first RollupData's groupClause. We
3821 : * don't make any effort to optimize grouping clauses when there are
3822 : * grouping sets, nor can we combine aggregate ordering keys with
3823 : * grouping.
3824 : */
3825 911 : List *rollups = qp_extra->gset_data->rollups;
3826 911 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3827 :
3828 911 : if (grouping_is_sortable(groupClause))
3829 : {
3830 : bool sortable;
3831 :
3832 : /*
3833 : * The groupClause is logically below the grouping step. So if
3834 : * there is an RTE entry for the grouping step, we need to remove
3835 : * its RT index from the sort expressions before we make PathKeys
3836 : * for them.
3837 : */
3838 911 : root->group_pathkeys =
3839 911 : make_pathkeys_for_sortclauses_extended(root,
3840 : &groupClause,
3841 : tlist,
3842 : false,
3843 911 : parse->hasGroupRTE,
3844 : &sortable,
3845 : false);
3846 : Assert(sortable);
3847 911 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3848 : }
3849 : else
3850 : {
3851 0 : root->group_pathkeys = NIL;
3852 0 : root->num_groupby_pathkeys = 0;
3853 : }
3854 : }
3855 386008 : else if (parse->groupClause || root->numOrderedAggs > 0)
3856 4984 : {
3857 : /*
3858 : * With a plain GROUP BY list, we can remove any grouping items that
3859 : * are proven redundant by EquivalenceClass processing. For example,
3860 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3861 : * especially common cases, but they're nearly free to detect. Note
3862 : * that we remove redundant items from processed_groupClause but not
3863 : * the original parse->groupClause.
3864 : */
3865 : bool sortable;
3866 :
3867 : /*
3868 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3869 : * EquivalenceClass'es if it's not set yet.
3870 : */
3871 4984 : root->group_pathkeys =
3872 4984 : make_pathkeys_for_sortclauses_extended(root,
3873 : &root->processed_groupClause,
3874 : tlist,
3875 : true,
3876 : false,
3877 : &sortable,
3878 : true);
3879 4984 : if (!sortable)
3880 : {
3881 : /* Can't sort; no point in considering aggregate ordering either */
3882 0 : root->group_pathkeys = NIL;
3883 0 : root->num_groupby_pathkeys = 0;
3884 : }
3885 : else
3886 : {
3887 4984 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3888 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3889 4984 : if (root->numOrderedAggs > 0)
3890 1664 : adjust_group_pathkeys_for_groupagg(root);
3891 : }
3892 : }
3893 : else
3894 : {
3895 381024 : root->group_pathkeys = NIL;
3896 381024 : root->num_groupby_pathkeys = 0;
3897 : }
3898 :
3899 : /* We consider only the first (bottom) window in pathkeys logic */
3900 386919 : if (activeWindows != NIL)
3901 : {
3902 2232 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3903 :
3904 2232 : root->window_pathkeys = make_pathkeys_for_window(root,
3905 : wc,
3906 : tlist);
3907 : }
3908 : else
3909 384687 : root->window_pathkeys = NIL;
3910 :
3911 : /*
3912 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3913 : * redundant by EquivalenceClass processing. The non-redundant list is
3914 : * kept in root->processed_distinctClause, leaving the original
3915 : * parse->distinctClause alone.
3916 : */
3917 386919 : if (parse->distinctClause)
3918 : {
3919 : bool sortable;
3920 :
3921 : /* Make a copy since pathkey processing can modify the list */
3922 1928 : root->processed_distinctClause = list_copy(parse->distinctClause);
3923 1928 : root->distinct_pathkeys =
3924 1928 : make_pathkeys_for_sortclauses_extended(root,
3925 : &root->processed_distinctClause,
3926 : tlist,
3927 : true,
3928 : false,
3929 : &sortable,
3930 : false);
3931 1928 : if (!sortable)
3932 5 : root->distinct_pathkeys = NIL;
3933 : }
3934 : else
3935 384991 : root->distinct_pathkeys = NIL;
3936 :
3937 386919 : root->sort_pathkeys =
3938 386919 : make_pathkeys_for_sortclauses(root,
3939 : parse->sortClause,
3940 : tlist);
3941 :
3942 : /* setting setop_pathkeys might be useful to the union planner */
3943 386919 : if (qp_extra->setop != NULL)
3944 : {
3945 : List *groupClauses;
3946 : bool sortable;
3947 :
3948 10478 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3949 :
3950 10478 : root->setop_pathkeys =
3951 10478 : make_pathkeys_for_sortclauses_extended(root,
3952 : &groupClauses,
3953 : tlist,
3954 : false,
3955 : false,
3956 : &sortable,
3957 : false);
3958 10478 : if (!sortable)
3959 120 : root->setop_pathkeys = NIL;
3960 : }
3961 : else
3962 376441 : root->setop_pathkeys = NIL;
3963 :
3964 : /*
3965 : * Figure out whether we want a sorted result from query_planner.
3966 : *
3967 : * If we have a sortable GROUP BY clause, then we want a result sorted
3968 : * properly for grouping. Otherwise, if we have window functions to
3969 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3970 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3971 : * we try to produce output that's sufficiently well sorted for the
3972 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3973 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3974 : * for a set operation which can benefit from presorted results and have a
3975 : * sortable targetlist, we want to sort by the target list.
3976 : *
3977 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3978 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3979 : * that might just leave us failing to exploit an available sort order at
3980 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3981 : * much easier, since we know that the parser ensured that one is a
3982 : * superset of the other.
3983 : */
3984 386919 : if (root->group_pathkeys)
3985 5498 : root->query_pathkeys = root->group_pathkeys;
3986 381421 : else if (root->window_pathkeys)
3987 1787 : root->query_pathkeys = root->window_pathkeys;
3988 759268 : else if (list_length(root->distinct_pathkeys) >
3989 379634 : list_length(root->sort_pathkeys))
3990 1558 : root->query_pathkeys = root->distinct_pathkeys;
3991 378076 : else if (root->sort_pathkeys)
3992 55417 : root->query_pathkeys = root->sort_pathkeys;
3993 322659 : else if (root->setop_pathkeys != NIL)
3994 9494 : root->query_pathkeys = root->setop_pathkeys;
3995 : else
3996 313165 : root->query_pathkeys = NIL;
3997 386919 : }
3998 :
3999 : /*
4000 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
4001 : *
4002 : * path_rows: number of output rows from scan/join step
4003 : * gd: grouping sets data including list of grouping sets and their clauses
4004 : * target_list: target list containing group clause references
4005 : *
4006 : * If doing grouping sets, we also annotate the gsets data with the estimates
4007 : * for each set and each individual rollup list, with a view to later
4008 : * determining whether some combination of them could be hashed instead.
4009 : */
4010 : static double
4011 40940 : get_number_of_groups(PlannerInfo *root,
4012 : double path_rows,
4013 : grouping_sets_data *gd,
4014 : List *target_list)
4015 : {
4016 40940 : Query *parse = root->parse;
4017 : double dNumGroups;
4018 :
4019 40940 : if (parse->groupClause)
4020 : {
4021 : List *groupExprs;
4022 :
4023 8765 : if (parse->groupingSets)
4024 : {
4025 : /* Add up the estimates for each grouping set */
4026 : ListCell *lc;
4027 :
4028 : Assert(gd); /* keep Coverity happy */
4029 :
4030 836 : dNumGroups = 0;
4031 :
4032 2246 : foreach(lc, gd->rollups)
4033 : {
4034 1410 : RollupData *rollup = lfirst_node(RollupData, lc);
4035 : ListCell *lc2;
4036 : ListCell *lc3;
4037 :
4038 1410 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
4039 : target_list);
4040 :
4041 1410 : rollup->numGroups = 0.0;
4042 :
4043 3915 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
4044 : {
4045 2505 : List *gset = (List *) lfirst(lc2);
4046 2505 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
4047 2505 : double numGroups = estimate_num_groups(root,
4048 : groupExprs,
4049 : path_rows,
4050 : &gset,
4051 : NULL);
4052 :
4053 2505 : gs->numGroups = numGroups;
4054 2505 : rollup->numGroups += numGroups;
4055 : }
4056 :
4057 1410 : dNumGroups += rollup->numGroups;
4058 : }
4059 :
4060 836 : if (gd->hash_sets_idx)
4061 : {
4062 : ListCell *lc2;
4063 :
4064 29 : gd->dNumHashGroups = 0;
4065 :
4066 29 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
4067 : target_list);
4068 :
4069 63 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
4070 : {
4071 34 : List *gset = (List *) lfirst(lc);
4072 34 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
4073 34 : double numGroups = estimate_num_groups(root,
4074 : groupExprs,
4075 : path_rows,
4076 : &gset,
4077 : NULL);
4078 :
4079 34 : gs->numGroups = numGroups;
4080 34 : gd->dNumHashGroups += numGroups;
4081 : }
4082 :
4083 29 : dNumGroups += gd->dNumHashGroups;
4084 : }
4085 : }
4086 : else
4087 : {
4088 : /* Plain GROUP BY -- estimate based on optimized groupClause */
4089 7929 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
4090 : target_list);
4091 :
4092 7929 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
4093 : NULL, NULL);
4094 : }
4095 : }
4096 32175 : else if (parse->groupingSets)
4097 : {
4098 : /* Empty grouping sets ... one result row for each one */
4099 50 : dNumGroups = list_length(parse->groupingSets);
4100 : }
4101 32125 : else if (parse->hasAggs || root->hasHavingQual)
4102 : {
4103 : /* Plain aggregation, one result row */
4104 32125 : dNumGroups = 1;
4105 : }
4106 : else
4107 : {
4108 : /* Not grouping */
4109 0 : dNumGroups = 1;
4110 : }
4111 :
4112 40940 : return dNumGroups;
4113 : }
4114 :
4115 : /*
4116 : * create_grouping_paths
4117 : *
4118 : * Build a new upperrel containing Paths for grouping and/or aggregation.
4119 : * Along the way, we also build an upperrel for Paths which are partially
4120 : * grouped and/or aggregated. A partially grouped and/or aggregated path
4121 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
4122 : * the only partially grouped paths we build are also partial paths; that
4123 : * is, they need a Gather and then a FinalizeAggregate.
4124 : *
4125 : * input_rel: contains the source-data Paths
4126 : * target: the pathtarget for the result Paths to compute
4127 : * gd: grouping sets data including list of grouping sets and their clauses
4128 : *
4129 : * Note: all Paths in input_rel are expected to return the target computed
4130 : * by make_group_input_target.
4131 : */
4132 : static RelOptInfo *
4133 33942 : create_grouping_paths(PlannerInfo *root,
4134 : RelOptInfo *input_rel,
4135 : PathTarget *target,
4136 : bool target_parallel_safe,
4137 : grouping_sets_data *gd)
4138 : {
4139 33942 : Query *parse = root->parse;
4140 : RelOptInfo *grouped_rel;
4141 : RelOptInfo *partially_grouped_rel;
4142 : AggClauseCosts agg_costs;
4143 :
4144 203652 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
4145 33942 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
4146 :
4147 : /*
4148 : * Create grouping relation to hold fully aggregated grouping and/or
4149 : * aggregation paths.
4150 : */
4151 33942 : grouped_rel = make_grouping_rel(root, input_rel, target,
4152 : target_parallel_safe, parse->havingQual);
4153 :
4154 : /*
4155 : * Create either paths for a degenerate grouping or paths for ordinary
4156 : * grouping, as appropriate.
4157 : */
4158 33942 : if (is_degenerate_grouping(root))
4159 45 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
4160 : else
4161 : {
4162 33897 : int flags = 0;
4163 : GroupPathExtraData extra;
4164 :
4165 : /*
4166 : * Determine whether it's possible to perform sort-based
4167 : * implementations of grouping. (Note that if processed_groupClause
4168 : * is empty, grouping_is_sortable() is trivially true, and all the
4169 : * pathkeys_contained_in() tests will succeed too, so that we'll
4170 : * consider every surviving input path.)
4171 : *
4172 : * If we have grouping sets, we might be able to sort some but not all
4173 : * of them; in this case, we need can_sort to be true as long as we
4174 : * must consider any sorted-input plan.
4175 : */
4176 33897 : if ((gd && gd->rollups != NIL)
4177 33016 : || grouping_is_sortable(root->processed_groupClause))
4178 33892 : flags |= GROUPING_CAN_USE_SORT;
4179 :
4180 : /*
4181 : * Determine whether we should consider hash-based implementations of
4182 : * grouping.
4183 : *
4184 : * Hashed aggregation only applies if we're grouping. If we have
4185 : * grouping sets, some groups might be hashable but others not; in
4186 : * this case we set can_hash true as long as there is nothing globally
4187 : * preventing us from hashing (and we should therefore consider plans
4188 : * with hashes).
4189 : *
4190 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
4191 : * BY aggregates. (Doing so would imply storing *all* the input
4192 : * values in the hash table, and/or running many sorts in parallel,
4193 : * either of which seems like a certain loser.) We similarly don't
4194 : * support ordered-set aggregates in hashed aggregation, but that case
4195 : * is also included in the numOrderedAggs count.
4196 : *
4197 : * Note: grouping_is_hashable() is much more expensive to check than
4198 : * the other gating conditions, so we want to do it last.
4199 : */
4200 33897 : if ((parse->groupClause != NIL &&
4201 7664 : root->numOrderedAggs == 0 &&
4202 3320 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
4203 4124 : flags |= GROUPING_CAN_USE_HASH;
4204 :
4205 : /*
4206 : * Determine whether partial aggregation is possible.
4207 : */
4208 33897 : if (can_partial_agg(root))
4209 29974 : flags |= GROUPING_CAN_PARTIAL_AGG;
4210 :
4211 33897 : extra.flags = flags;
4212 33897 : extra.target_parallel_safe = target_parallel_safe;
4213 33897 : extra.havingQual = parse->havingQual;
4214 33897 : extra.targetList = parse->targetList;
4215 33897 : extra.partial_costs_set = false;
4216 :
4217 : /*
4218 : * Determine whether partitionwise aggregation is in theory possible.
4219 : * It can be disabled by the user, and for now, we don't try to
4220 : * support grouping sets. create_ordinary_grouping_paths() will check
4221 : * additional conditions, such as whether input_rel is partitioned.
4222 : */
4223 33897 : if (enable_partitionwise_aggregate && !parse->groupingSets)
4224 580 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
4225 : else
4226 33317 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
4227 :
4228 33897 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
4229 : &agg_costs, gd, &extra,
4230 : &partially_grouped_rel);
4231 : }
4232 :
4233 33938 : set_cheapest(grouped_rel);
4234 33938 : return grouped_rel;
4235 : }
4236 :
4237 : /*
4238 : * make_grouping_rel
4239 : *
4240 : * Create a new grouping rel and set basic properties.
4241 : *
4242 : * input_rel represents the underlying scan/join relation.
4243 : * target is the output expected from the grouping relation.
4244 : */
4245 : static RelOptInfo *
4246 35737 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
4247 : PathTarget *target, bool target_parallel_safe,
4248 : Node *havingQual)
4249 : {
4250 : RelOptInfo *grouped_rel;
4251 :
4252 35737 : if (IS_OTHER_REL(input_rel))
4253 : {
4254 1795 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
4255 : input_rel->relids);
4256 1795 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
4257 : }
4258 : else
4259 : {
4260 : /*
4261 : * By tradition, the relids set for the main grouping relation is
4262 : * NULL. (This could be changed, but might require adjustments
4263 : * elsewhere.)
4264 : */
4265 33942 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
4266 : }
4267 :
4268 : /* Set target. */
4269 35737 : grouped_rel->reltarget = target;
4270 :
4271 : /*
4272 : * If the input relation is not parallel-safe, then the grouped relation
4273 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4274 : * target list and HAVING quals are parallel-safe.
4275 : */
4276 58713 : if (input_rel->consider_parallel && target_parallel_safe &&
4277 22976 : is_parallel_safe(root, havingQual))
4278 22951 : grouped_rel->consider_parallel = true;
4279 :
4280 : /* Assume that the same path generation strategies are allowed */
4281 35737 : grouped_rel->pgs_mask = input_rel->pgs_mask;
4282 :
4283 : /*
4284 : * If the input rel belongs to a single FDW, so does the grouped rel.
4285 : */
4286 35737 : grouped_rel->serverid = input_rel->serverid;
4287 35737 : grouped_rel->userid = input_rel->userid;
4288 35737 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
4289 35737 : grouped_rel->fdwroutine = input_rel->fdwroutine;
4290 :
4291 35737 : return grouped_rel;
4292 : }
4293 :
4294 : /*
4295 : * is_degenerate_grouping
4296 : *
4297 : * A degenerate grouping is one in which the query has a HAVING qual and/or
4298 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
4299 : * grouping sets are all empty).
4300 : */
4301 : static bool
4302 33942 : is_degenerate_grouping(PlannerInfo *root)
4303 : {
4304 33942 : Query *parse = root->parse;
4305 :
4306 32801 : return (root->hasHavingQual || parse->groupingSets) &&
4307 66743 : !parse->hasAggs && parse->groupClause == NIL;
4308 : }
4309 :
4310 : /*
4311 : * create_degenerate_grouping_paths
4312 : *
4313 : * When the grouping is degenerate (see is_degenerate_grouping), we are
4314 : * supposed to emit either zero or one row for each grouping set depending on
4315 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
4316 : * either HAVING or the targetlist, so we actually do not need the FROM table
4317 : * at all! We can just throw away the plan-so-far and generate a Result node.
4318 : * This is a sufficiently unusual corner case that it's not worth contorting
4319 : * the structure of this module to avoid having to generate the earlier paths
4320 : * in the first place.
4321 : */
4322 : static void
4323 45 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4324 : RelOptInfo *grouped_rel)
4325 : {
4326 45 : Query *parse = root->parse;
4327 : int nrows;
4328 : Path *path;
4329 :
4330 45 : nrows = list_length(parse->groupingSets);
4331 45 : if (nrows > 1)
4332 : {
4333 : /*
4334 : * Doesn't seem worthwhile writing code to cons up a generate_series
4335 : * or a values scan to emit multiple rows. Instead just make N clones
4336 : * and append them. (With a volatile HAVING clause, this means you
4337 : * might get between 0 and N output rows. Offhand I think that's
4338 : * desired.)
4339 : */
4340 10 : AppendPathInput append = {0};
4341 :
4342 30 : while (--nrows >= 0)
4343 : {
4344 : path = (Path *)
4345 20 : create_group_result_path(root, grouped_rel,
4346 20 : grouped_rel->reltarget,
4347 20 : (List *) parse->havingQual);
4348 20 : append.subpaths = lappend(append.subpaths, path);
4349 : }
4350 : path = (Path *)
4351 10 : create_append_path(root,
4352 : grouped_rel,
4353 : append,
4354 : NIL,
4355 : NULL,
4356 : 0,
4357 : false,
4358 : -1);
4359 : }
4360 : else
4361 : {
4362 : /* No grouping sets, or just one, so one output row */
4363 : path = (Path *)
4364 35 : create_group_result_path(root, grouped_rel,
4365 35 : grouped_rel->reltarget,
4366 35 : (List *) parse->havingQual);
4367 : }
4368 :
4369 45 : add_path(grouped_rel, path);
4370 45 : }
4371 :
4372 : /*
4373 : * create_ordinary_grouping_paths
4374 : *
4375 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4376 : *
4377 : * We need to consider sorted and hashed aggregation in the same function,
4378 : * because otherwise (1) it would be harder to throw an appropriate error
4379 : * message if neither way works, and (2) we should not allow hashtable size
4380 : * considerations to dissuade us from using hashing if sorting is not possible.
4381 : *
4382 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4383 : * function creates, or to NULL if it doesn't create one.
4384 : */
4385 : static void
4386 35692 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4387 : RelOptInfo *grouped_rel,
4388 : const AggClauseCosts *agg_costs,
4389 : grouping_sets_data *gd,
4390 : GroupPathExtraData *extra,
4391 : RelOptInfo **partially_grouped_rel_p)
4392 : {
4393 35692 : RelOptInfo *partially_grouped_rel = NULL;
4394 35692 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4395 :
4396 : /*
4397 : * If this is the topmost grouping relation or if the parent relation is
4398 : * doing some form of partitionwise aggregation, then we may be able to do
4399 : * it at this level also. However, if the input relation is not
4400 : * partitioned, partitionwise aggregate is impossible.
4401 : */
4402 35692 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4403 2375 : IS_PARTITIONED_REL(input_rel))
4404 : {
4405 : /*
4406 : * If this is the topmost relation or if the parent relation is doing
4407 : * full partitionwise aggregation, then we can do full partitionwise
4408 : * aggregation provided that the GROUP BY clause contains all of the
4409 : * partitioning columns at this level and the collation used by GROUP
4410 : * BY matches the partitioning collation. Otherwise, we can do at
4411 : * most partial partitionwise aggregation. But if partial aggregation
4412 : * is not supported in general then we can't use it for partitionwise
4413 : * aggregation either.
4414 : *
4415 : * Check parse->groupClause not processed_groupClause, because it's
4416 : * okay if some of the partitioning columns were proved redundant.
4417 : */
4418 1360 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4419 640 : group_by_has_partkey(input_rel, extra->targetList,
4420 640 : root->parse->groupClause))
4421 404 : patype = PARTITIONWISE_AGGREGATE_FULL;
4422 316 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4423 281 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4424 : else
4425 35 : patype = PARTITIONWISE_AGGREGATE_NONE;
4426 : }
4427 :
4428 : /*
4429 : * Before generating paths for grouped_rel, we first generate any possible
4430 : * partially grouped paths; that way, later code can easily consider both
4431 : * parallel and non-parallel approaches to grouping.
4432 : */
4433 35692 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4434 : {
4435 : bool force_rel_creation;
4436 :
4437 : /*
4438 : * If we're doing partitionwise aggregation at this level, force
4439 : * creation of a partially_grouped_rel so we can add partitionwise
4440 : * paths to it.
4441 : */
4442 31709 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4443 :
4444 : partially_grouped_rel =
4445 31709 : create_partial_grouping_paths(root,
4446 : grouped_rel,
4447 : input_rel,
4448 : gd,
4449 : extra,
4450 : force_rel_creation);
4451 : }
4452 :
4453 : /* Set out parameter. */
4454 35692 : *partially_grouped_rel_p = partially_grouped_rel;
4455 :
4456 : /* Apply partitionwise aggregation technique, if possible. */
4457 35692 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4458 685 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4459 : partially_grouped_rel, agg_costs,
4460 : gd, patype, extra);
4461 :
4462 : /* If we are doing partial aggregation only, return. */
4463 35692 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4464 : {
4465 : Assert(partially_grouped_rel);
4466 :
4467 713 : if (partially_grouped_rel->pathlist)
4468 713 : set_cheapest(partially_grouped_rel);
4469 :
4470 713 : return;
4471 : }
4472 :
4473 : /* Gather any partially grouped partial paths. */
4474 34979 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4475 2325 : gather_grouping_paths(root, partially_grouped_rel);
4476 :
4477 : /* Now choose the best path(s) for partially_grouped_rel. */
4478 34979 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4479 2503 : set_cheapest(partially_grouped_rel);
4480 :
4481 : /* Build final grouping paths */
4482 34979 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4483 : partially_grouped_rel, agg_costs, gd,
4484 : extra);
4485 :
4486 : /* Give a helpful error if we failed to find any implementation */
4487 34979 : if (grouped_rel->pathlist == NIL)
4488 4 : ereport(ERROR,
4489 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4490 : errmsg("could not implement GROUP BY"),
4491 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4492 :
4493 : /*
4494 : * If there is an FDW that's responsible for all baserels of the query,
4495 : * let it consider adding ForeignPaths.
4496 : */
4497 34975 : if (grouped_rel->fdwroutine &&
4498 171 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4499 170 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4500 : input_rel, grouped_rel,
4501 : extra);
4502 :
4503 : /* Let extensions possibly add some more paths */
4504 34975 : if (create_upper_paths_hook)
4505 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4506 : input_rel, grouped_rel,
4507 : extra);
4508 : }
4509 :
4510 : /*
4511 : * For a given input path, consider the possible ways of doing grouping sets on
4512 : * it, by combinations of hashing and sorting. This can be called multiple
4513 : * times, so it's important that it not scribble on input. No result is
4514 : * returned, but any generated paths are added to grouped_rel.
4515 : */
4516 : static void
4517 1782 : consider_groupingsets_paths(PlannerInfo *root,
4518 : RelOptInfo *grouped_rel,
4519 : Path *path,
4520 : bool is_sorted,
4521 : bool can_hash,
4522 : grouping_sets_data *gd,
4523 : const AggClauseCosts *agg_costs,
4524 : double dNumGroups)
4525 : {
4526 1782 : Query *parse = root->parse;
4527 1782 : Size hash_mem_limit = get_hash_memory_limit();
4528 :
4529 : /*
4530 : * If we're not being offered sorted input, then only consider plans that
4531 : * can be done entirely by hashing.
4532 : *
4533 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4534 : * the input is actually sorted despite not being advertised as such, we
4535 : * prefer to make use of that in order to use less memory.
4536 : *
4537 : * If none of the grouping sets are sortable, then ignore the hash_mem
4538 : * limit and generate a path anyway, since otherwise we'll just fail.
4539 : */
4540 1782 : if (!is_sorted)
4541 : {
4542 807 : List *new_rollups = NIL;
4543 807 : RollupData *unhashed_rollup = NULL;
4544 : List *sets_data;
4545 807 : List *empty_sets_data = NIL;
4546 807 : List *empty_sets = NIL;
4547 : ListCell *lc;
4548 807 : ListCell *l_start = list_head(gd->rollups);
4549 807 : AggStrategy strat = AGG_HASHED;
4550 : double hashsize;
4551 807 : double exclude_groups = 0.0;
4552 :
4553 : Assert(can_hash);
4554 :
4555 : /*
4556 : * If the input is coincidentally sorted usefully (which can happen
4557 : * even if is_sorted is false, since that only means that our caller
4558 : * has set up the sorting for us), then save some hashtable space by
4559 : * making use of that. But we need to watch out for degenerate cases:
4560 : *
4561 : * 1) If there are any empty grouping sets, then group_pathkeys might
4562 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4563 : * there will be a rollup containing only empty groups, and the
4564 : * pathkeys_contained_in test is vacuously true; this is ok.
4565 : *
4566 : * XXX: the above relies on the fact that group_pathkeys is generated
4567 : * from the first rollup. If we add the ability to consider multiple
4568 : * sort orders for grouping input, this assumption might fail.
4569 : *
4570 : * 2) If there are no empty sets and only unsortable sets, then the
4571 : * rollups list will be empty (and thus l_start == NULL), and
4572 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4573 : * pathkeys_contained_in test doesn't cause us to crash.
4574 : */
4575 1609 : if (l_start != NULL &&
4576 802 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4577 : {
4578 40 : unhashed_rollup = lfirst_node(RollupData, l_start);
4579 40 : exclude_groups = unhashed_rollup->numGroups;
4580 40 : l_start = lnext(gd->rollups, l_start);
4581 : }
4582 :
4583 807 : hashsize = estimate_hashagg_tablesize(root,
4584 : path,
4585 : agg_costs,
4586 : dNumGroups - exclude_groups);
4587 :
4588 : /*
4589 : * gd->rollups is empty if we have only unsortable columns to work
4590 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4591 : * sorted-input case to generate usable mixed paths.
4592 : */
4593 807 : if (hashsize > hash_mem_limit && gd->rollups)
4594 15 : return; /* nope, won't fit */
4595 :
4596 : /*
4597 : * We need to burst the existing rollups list into individual grouping
4598 : * sets and recompute a groupClause for each set.
4599 : */
4600 792 : sets_data = list_copy(gd->unsortable_sets);
4601 :
4602 2013 : for_each_cell(lc, gd->rollups, l_start)
4603 : {
4604 1241 : RollupData *rollup = lfirst_node(RollupData, lc);
4605 :
4606 : /*
4607 : * If we find an unhashable rollup that's not been skipped by the
4608 : * "actually sorted" check above, we can't cope; we'd need sorted
4609 : * input (with a different sort order) but we can't get that here.
4610 : * So bail out; we'll get a valid path from the is_sorted case
4611 : * instead.
4612 : *
4613 : * The mere presence of empty grouping sets doesn't make a rollup
4614 : * unhashable (see preprocess_grouping_sets), we handle those
4615 : * specially below.
4616 : */
4617 1241 : if (!rollup->hashable)
4618 20 : return;
4619 :
4620 1221 : sets_data = list_concat(sets_data, rollup->gsets_data);
4621 : }
4622 3053 : foreach(lc, sets_data)
4623 : {
4624 2281 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4625 2281 : List *gset = gs->set;
4626 : RollupData *rollup;
4627 :
4628 2281 : if (gset == NIL)
4629 : {
4630 : /* Empty grouping sets can't be hashed. */
4631 445 : empty_sets_data = lappend(empty_sets_data, gs);
4632 445 : empty_sets = lappend(empty_sets, NIL);
4633 : }
4634 : else
4635 : {
4636 1836 : rollup = makeNode(RollupData);
4637 :
4638 1836 : rollup->groupClause = preprocess_groupclause(root, gset);
4639 1836 : rollup->gsets_data = list_make1(gs);
4640 1836 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4641 : rollup->gsets_data,
4642 : gd->tleref_to_colnum_map);
4643 1836 : rollup->numGroups = gs->numGroups;
4644 1836 : rollup->hashable = true;
4645 1836 : rollup->is_hashed = true;
4646 1836 : new_rollups = lappend(new_rollups, rollup);
4647 : }
4648 : }
4649 :
4650 : /*
4651 : * If we didn't find anything nonempty to hash, then bail. We'll
4652 : * generate a path from the is_sorted case.
4653 : */
4654 772 : if (new_rollups == NIL)
4655 0 : return;
4656 :
4657 : /*
4658 : * If there were empty grouping sets they should have been in the
4659 : * first rollup.
4660 : */
4661 : Assert(!unhashed_rollup || !empty_sets);
4662 :
4663 772 : if (unhashed_rollup)
4664 : {
4665 40 : new_rollups = lappend(new_rollups, unhashed_rollup);
4666 40 : strat = AGG_MIXED;
4667 : }
4668 732 : else if (empty_sets)
4669 : {
4670 405 : RollupData *rollup = makeNode(RollupData);
4671 :
4672 405 : rollup->groupClause = NIL;
4673 405 : rollup->gsets_data = empty_sets_data;
4674 405 : rollup->gsets = empty_sets;
4675 405 : rollup->numGroups = list_length(empty_sets);
4676 405 : rollup->hashable = false;
4677 405 : rollup->is_hashed = false;
4678 405 : new_rollups = lappend(new_rollups, rollup);
4679 405 : strat = AGG_MIXED;
4680 : }
4681 :
4682 772 : add_path(grouped_rel, (Path *)
4683 772 : create_groupingsets_path(root,
4684 : grouped_rel,
4685 : path,
4686 772 : (List *) parse->havingQual,
4687 : strat,
4688 : new_rollups,
4689 : agg_costs));
4690 772 : return;
4691 : }
4692 :
4693 : /*
4694 : * If we have sorted input but nothing we can do with it, bail.
4695 : */
4696 975 : if (gd->rollups == NIL)
4697 0 : return;
4698 :
4699 : /*
4700 : * Given sorted input, we try and make two paths: one sorted and one mixed
4701 : * sort/hash. (We need to try both because hashagg might be disabled, or
4702 : * some columns might not be sortable.)
4703 : *
4704 : * can_hash is passed in as false if some obstacle elsewhere (such as
4705 : * ordered aggs) means that we shouldn't consider hashing at all.
4706 : */
4707 975 : if (can_hash && gd->any_hashable)
4708 : {
4709 896 : List *rollups = NIL;
4710 896 : List *hash_sets = list_copy(gd->unsortable_sets);
4711 896 : double availspace = hash_mem_limit;
4712 : ListCell *lc;
4713 :
4714 : /*
4715 : * Account first for space needed for groups we can't sort at all.
4716 : */
4717 896 : availspace -= estimate_hashagg_tablesize(root,
4718 : path,
4719 : agg_costs,
4720 : gd->dNumHashGroups);
4721 :
4722 896 : if (availspace > 0 && list_length(gd->rollups) > 1)
4723 : {
4724 : double scale;
4725 501 : int num_rollups = list_length(gd->rollups);
4726 : int k_capacity;
4727 501 : int *k_weights = palloc(num_rollups * sizeof(int));
4728 501 : Bitmapset *hash_items = NULL;
4729 : int i;
4730 :
4731 : /*
4732 : * We treat this as a knapsack problem: the knapsack capacity
4733 : * represents hash_mem, the item weights are the estimated memory
4734 : * usage of the hashtables needed to implement a single rollup,
4735 : * and we really ought to use the cost saving as the item value;
4736 : * however, currently the costs assigned to sort nodes don't
4737 : * reflect the comparison costs well, and so we treat all items as
4738 : * of equal value (each rollup we hash instead saves us one sort).
4739 : *
4740 : * To use the discrete knapsack, we need to scale the values to a
4741 : * reasonably small bounded range. We choose to allow a 5% error
4742 : * margin; we have no more than 4096 rollups in the worst possible
4743 : * case, which with a 5% error margin will require a bit over 42MB
4744 : * of workspace. (Anyone wanting to plan queries that complex had
4745 : * better have the memory for it. In more reasonable cases, with
4746 : * no more than a couple of dozen rollups, the memory usage will
4747 : * be negligible.)
4748 : *
4749 : * k_capacity is naturally bounded, but we clamp the values for
4750 : * scale and weight (below) to avoid overflows or underflows (or
4751 : * uselessly trying to use a scale factor less than 1 byte).
4752 : */
4753 501 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4754 501 : k_capacity = (int) floor(availspace / scale);
4755 :
4756 : /*
4757 : * We leave the first rollup out of consideration since it's the
4758 : * one that matches the input sort order. We assign indexes "i"
4759 : * to only those entries considered for hashing; the second loop,
4760 : * below, must use the same condition.
4761 : */
4762 501 : i = 0;
4763 1222 : for_each_from(lc, gd->rollups, 1)
4764 : {
4765 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4766 :
4767 721 : if (rollup->hashable)
4768 : {
4769 721 : double sz = estimate_hashagg_tablesize(root,
4770 : path,
4771 : agg_costs,
4772 : rollup->numGroups);
4773 :
4774 : /*
4775 : * If sz is enormous, but hash_mem (and hence scale) is
4776 : * small, avoid integer overflow here.
4777 : */
4778 721 : k_weights[i] = (int) Min(floor(sz / scale),
4779 : k_capacity + 1.0);
4780 721 : ++i;
4781 : }
4782 : }
4783 :
4784 : /*
4785 : * Apply knapsack algorithm; compute the set of items which
4786 : * maximizes the value stored (in this case the number of sorts
4787 : * saved) while keeping the total size (approximately) within
4788 : * capacity.
4789 : */
4790 501 : if (i > 0)
4791 501 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4792 :
4793 501 : if (!bms_is_empty(hash_items))
4794 : {
4795 501 : rollups = list_make1(linitial(gd->rollups));
4796 :
4797 501 : i = 0;
4798 1222 : for_each_from(lc, gd->rollups, 1)
4799 : {
4800 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4801 :
4802 721 : if (rollup->hashable)
4803 : {
4804 721 : if (bms_is_member(i, hash_items))
4805 691 : hash_sets = list_concat(hash_sets,
4806 691 : rollup->gsets_data);
4807 : else
4808 30 : rollups = lappend(rollups, rollup);
4809 721 : ++i;
4810 : }
4811 : else
4812 0 : rollups = lappend(rollups, rollup);
4813 : }
4814 : }
4815 : }
4816 :
4817 896 : if (!rollups && hash_sets)
4818 20 : rollups = list_copy(gd->rollups);
4819 :
4820 1703 : foreach(lc, hash_sets)
4821 : {
4822 807 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4823 807 : RollupData *rollup = makeNode(RollupData);
4824 :
4825 : Assert(gs->set != NIL);
4826 :
4827 807 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4828 807 : rollup->gsets_data = list_make1(gs);
4829 807 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4830 : rollup->gsets_data,
4831 : gd->tleref_to_colnum_map);
4832 807 : rollup->numGroups = gs->numGroups;
4833 807 : rollup->hashable = true;
4834 807 : rollup->is_hashed = true;
4835 807 : rollups = lcons(rollup, rollups);
4836 : }
4837 :
4838 896 : if (rollups)
4839 : {
4840 521 : add_path(grouped_rel, (Path *)
4841 521 : create_groupingsets_path(root,
4842 : grouped_rel,
4843 : path,
4844 521 : (List *) parse->havingQual,
4845 : AGG_MIXED,
4846 : rollups,
4847 : agg_costs));
4848 : }
4849 : }
4850 :
4851 : /*
4852 : * Now try the simple sorted case.
4853 : */
4854 975 : if (!gd->unsortable_sets)
4855 951 : add_path(grouped_rel, (Path *)
4856 951 : create_groupingsets_path(root,
4857 : grouped_rel,
4858 : path,
4859 951 : (List *) parse->havingQual,
4860 : AGG_SORTED,
4861 : gd->rollups,
4862 : agg_costs));
4863 : }
4864 :
4865 : /*
4866 : * create_window_paths
4867 : *
4868 : * Build a new upperrel containing Paths for window-function evaluation.
4869 : *
4870 : * input_rel: contains the source-data Paths
4871 : * input_target: result of make_window_input_target
4872 : * output_target: what the topmost WindowAggPath should return
4873 : * wflists: result of find_window_functions
4874 : * activeWindows: result of select_active_windows
4875 : *
4876 : * Note: all Paths in input_rel are expected to return input_target.
4877 : */
4878 : static RelOptInfo *
4879 2232 : create_window_paths(PlannerInfo *root,
4880 : RelOptInfo *input_rel,
4881 : PathTarget *input_target,
4882 : PathTarget *output_target,
4883 : bool output_target_parallel_safe,
4884 : WindowFuncLists *wflists,
4885 : List *activeWindows)
4886 : {
4887 : RelOptInfo *window_rel;
4888 : ListCell *lc;
4889 :
4890 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4891 2232 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4892 :
4893 : /*
4894 : * If the input relation is not parallel-safe, then the window relation
4895 : * can't be parallel-safe, either. Otherwise, we need to examine the
4896 : * target list and active windows for non-parallel-safe constructs.
4897 : */
4898 2232 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4899 0 : is_parallel_safe(root, (Node *) activeWindows))
4900 0 : window_rel->consider_parallel = true;
4901 :
4902 : /*
4903 : * If the input rel belongs to a single FDW, so does the window rel.
4904 : */
4905 2232 : window_rel->serverid = input_rel->serverid;
4906 2232 : window_rel->userid = input_rel->userid;
4907 2232 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4908 2232 : window_rel->fdwroutine = input_rel->fdwroutine;
4909 :
4910 : /*
4911 : * Consider computing window functions starting from the existing
4912 : * cheapest-total path (which will likely require a sort) as well as any
4913 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4914 : */
4915 4726 : foreach(lc, input_rel->pathlist)
4916 : {
4917 2494 : Path *path = (Path *) lfirst(lc);
4918 : int presorted_keys;
4919 :
4920 2756 : if (path == input_rel->cheapest_total_path ||
4921 262 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4922 116 : &presorted_keys) ||
4923 116 : presorted_keys > 0)
4924 2399 : create_one_window_path(root,
4925 : window_rel,
4926 : path,
4927 : input_target,
4928 : output_target,
4929 : wflists,
4930 : activeWindows);
4931 : }
4932 :
4933 : /*
4934 : * If there is an FDW that's responsible for all baserels of the query,
4935 : * let it consider adding ForeignPaths.
4936 : */
4937 2232 : if (window_rel->fdwroutine &&
4938 6 : window_rel->fdwroutine->GetForeignUpperPaths)
4939 6 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4940 : input_rel, window_rel,
4941 : NULL);
4942 :
4943 : /* Let extensions possibly add some more paths */
4944 2232 : if (create_upper_paths_hook)
4945 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4946 : input_rel, window_rel, NULL);
4947 :
4948 : /* Now choose the best path(s) */
4949 2232 : set_cheapest(window_rel);
4950 :
4951 2232 : return window_rel;
4952 : }
4953 :
4954 : /*
4955 : * Stack window-function implementation steps atop the given Path, and
4956 : * add the result to window_rel.
4957 : *
4958 : * window_rel: upperrel to contain result
4959 : * path: input Path to use (must return input_target)
4960 : * input_target: result of make_window_input_target
4961 : * output_target: what the topmost WindowAggPath should return
4962 : * wflists: result of find_window_functions
4963 : * activeWindows: result of select_active_windows
4964 : */
4965 : static void
4966 2399 : create_one_window_path(PlannerInfo *root,
4967 : RelOptInfo *window_rel,
4968 : Path *path,
4969 : PathTarget *input_target,
4970 : PathTarget *output_target,
4971 : WindowFuncLists *wflists,
4972 : List *activeWindows)
4973 : {
4974 : PathTarget *window_target;
4975 : ListCell *l;
4976 2399 : List *topqual = NIL;
4977 :
4978 : /*
4979 : * Since each window clause could require a different sort order, we stack
4980 : * up a WindowAgg node for each clause, with sort steps between them as
4981 : * needed. (We assume that select_active_windows chose a good order for
4982 : * executing the clauses in.)
4983 : *
4984 : * input_target should contain all Vars and Aggs needed for the result.
4985 : * (In some cases we wouldn't need to propagate all of these all the way
4986 : * to the top, since they might only be needed as inputs to WindowFuncs.
4987 : * It's probably not worth trying to optimize that though.) It must also
4988 : * contain all window partitioning and sorting expressions, to ensure
4989 : * they're computed only once at the bottom of the stack (that's critical
4990 : * for volatile functions). As we climb up the stack, we'll add outputs
4991 : * for the WindowFuncs computed at each level.
4992 : */
4993 2399 : window_target = input_target;
4994 :
4995 4953 : foreach(l, activeWindows)
4996 : {
4997 2554 : WindowClause *wc = lfirst_node(WindowClause, l);
4998 : List *window_pathkeys;
4999 2554 : List *runcondition = NIL;
5000 : int presorted_keys;
5001 : bool is_sorted;
5002 : bool topwindow;
5003 : ListCell *lc2;
5004 :
5005 2554 : window_pathkeys = make_pathkeys_for_window(root,
5006 : wc,
5007 : root->processed_tlist);
5008 :
5009 2554 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
5010 : path->pathkeys,
5011 : &presorted_keys);
5012 :
5013 : /* Sort if necessary */
5014 2554 : if (!is_sorted)
5015 : {
5016 : /*
5017 : * No presorted keys or incremental sort disabled, just perform a
5018 : * complete sort.
5019 : */
5020 1841 : if (presorted_keys == 0 || !enable_incremental_sort)
5021 1790 : path = (Path *) create_sort_path(root, window_rel,
5022 : path,
5023 : window_pathkeys,
5024 : -1.0);
5025 : else
5026 : {
5027 : /*
5028 : * Since we have presorted keys and incremental sort is
5029 : * enabled, just use incremental sort.
5030 : */
5031 51 : path = (Path *) create_incremental_sort_path(root,
5032 : window_rel,
5033 : path,
5034 : window_pathkeys,
5035 : presorted_keys,
5036 : -1.0);
5037 : }
5038 : }
5039 :
5040 2554 : if (lnext(activeWindows, l))
5041 : {
5042 : /*
5043 : * Add the current WindowFuncs to the output target for this
5044 : * intermediate WindowAggPath. We must copy window_target to
5045 : * avoid changing the previous path's target.
5046 : *
5047 : * Note: a WindowFunc adds nothing to the target's eval costs; but
5048 : * we do need to account for the increase in tlist width.
5049 : */
5050 155 : int64 tuple_width = window_target->width;
5051 :
5052 155 : window_target = copy_pathtarget(window_target);
5053 370 : foreach(lc2, wflists->windowFuncs[wc->winref])
5054 : {
5055 215 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5056 :
5057 215 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
5058 215 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
5059 : }
5060 155 : window_target->width = clamp_width_est(tuple_width);
5061 : }
5062 : else
5063 : {
5064 : /* Install the goal target in the topmost WindowAgg */
5065 2399 : window_target = output_target;
5066 : }
5067 :
5068 : /* mark the final item in the list as the top-level window */
5069 2554 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
5070 :
5071 : /*
5072 : * Collect the WindowFuncRunConditions from each WindowFunc and
5073 : * convert them into OpExprs
5074 : */
5075 5841 : foreach(lc2, wflists->windowFuncs[wc->winref])
5076 : {
5077 : ListCell *lc3;
5078 3287 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5079 :
5080 3437 : foreach(lc3, wfunc->runCondition)
5081 : {
5082 150 : WindowFuncRunCondition *wfuncrc =
5083 : lfirst_node(WindowFuncRunCondition, lc3);
5084 : Expr *opexpr;
5085 : Expr *leftop;
5086 : Expr *rightop;
5087 :
5088 150 : if (wfuncrc->wfunc_left)
5089 : {
5090 135 : leftop = (Expr *) copyObject(wfunc);
5091 135 : rightop = copyObject(wfuncrc->arg);
5092 : }
5093 : else
5094 : {
5095 15 : leftop = copyObject(wfuncrc->arg);
5096 15 : rightop = (Expr *) copyObject(wfunc);
5097 : }
5098 :
5099 150 : opexpr = make_opclause(wfuncrc->opno,
5100 : BOOLOID,
5101 : false,
5102 : leftop,
5103 : rightop,
5104 : InvalidOid,
5105 : wfuncrc->inputcollid);
5106 :
5107 150 : runcondition = lappend(runcondition, opexpr);
5108 :
5109 150 : if (!topwindow)
5110 20 : topqual = lappend(topqual, opexpr);
5111 : }
5112 : }
5113 :
5114 : path = (Path *)
5115 2554 : create_windowagg_path(root, window_rel, path, window_target,
5116 2554 : wflists->windowFuncs[wc->winref],
5117 : runcondition, wc,
5118 : topwindow ? topqual : NIL, topwindow);
5119 : }
5120 :
5121 2399 : add_path(window_rel, path);
5122 2399 : }
5123 :
5124 : /*
5125 : * create_distinct_paths
5126 : *
5127 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
5128 : *
5129 : * input_rel: contains the source-data Paths
5130 : * target: the pathtarget for the result Paths to compute
5131 : *
5132 : * Note: input paths should already compute the desired pathtarget, since
5133 : * Sort/Unique won't project anything.
5134 : */
5135 : static RelOptInfo *
5136 1928 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5137 : PathTarget *target)
5138 : {
5139 : RelOptInfo *distinct_rel;
5140 :
5141 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
5142 1928 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
5143 :
5144 : /*
5145 : * We don't compute anything at this level, so distinct_rel will be
5146 : * parallel-safe if the input rel is parallel-safe. In particular, if
5147 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
5148 : * output those expressions, and will not be parallel-safe unless those
5149 : * expressions are parallel-safe.
5150 : */
5151 1928 : distinct_rel->consider_parallel = input_rel->consider_parallel;
5152 :
5153 : /*
5154 : * If the input rel belongs to a single FDW, so does the distinct_rel.
5155 : */
5156 1928 : distinct_rel->serverid = input_rel->serverid;
5157 1928 : distinct_rel->userid = input_rel->userid;
5158 1928 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5159 1928 : distinct_rel->fdwroutine = input_rel->fdwroutine;
5160 :
5161 : /* build distinct paths based on input_rel's pathlist */
5162 1928 : create_final_distinct_paths(root, input_rel, distinct_rel);
5163 :
5164 : /* now build distinct paths based on input_rel's partial_pathlist */
5165 1928 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
5166 :
5167 : /* Give a helpful error if we failed to create any paths */
5168 1928 : if (distinct_rel->pathlist == NIL)
5169 0 : ereport(ERROR,
5170 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5171 : errmsg("could not implement DISTINCT"),
5172 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
5173 :
5174 : /*
5175 : * If there is an FDW that's responsible for all baserels of the query,
5176 : * let it consider adding ForeignPaths.
5177 : */
5178 1928 : if (distinct_rel->fdwroutine &&
5179 8 : distinct_rel->fdwroutine->GetForeignUpperPaths)
5180 8 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5181 : UPPERREL_DISTINCT,
5182 : input_rel,
5183 : distinct_rel,
5184 : NULL);
5185 :
5186 : /* Let extensions possibly add some more paths */
5187 1928 : if (create_upper_paths_hook)
5188 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
5189 : distinct_rel, NULL);
5190 :
5191 : /* Now choose the best path(s) */
5192 1928 : set_cheapest(distinct_rel);
5193 :
5194 1928 : return distinct_rel;
5195 : }
5196 :
5197 : /*
5198 : * create_partial_distinct_paths
5199 : *
5200 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
5201 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
5202 : * paths on top and add a final unique/aggregate path to remove any duplicate
5203 : * produced from combining rows from parallel workers.
5204 : */
5205 : static void
5206 1928 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5207 : RelOptInfo *final_distinct_rel,
5208 : PathTarget *target)
5209 : {
5210 : RelOptInfo *partial_distinct_rel;
5211 : Query *parse;
5212 : List *distinctExprs;
5213 : double numDistinctRows;
5214 : Path *cheapest_partial_path;
5215 : ListCell *lc;
5216 :
5217 : /* nothing to do when there are no partial paths in the input rel */
5218 1928 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
5219 1838 : return;
5220 :
5221 90 : parse = root->parse;
5222 :
5223 : /* can't do parallel DISTINCT ON */
5224 90 : if (parse->hasDistinctOn)
5225 0 : return;
5226 :
5227 90 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
5228 : NULL);
5229 90 : partial_distinct_rel->reltarget = target;
5230 90 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
5231 :
5232 : /*
5233 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
5234 : */
5235 90 : partial_distinct_rel->serverid = input_rel->serverid;
5236 90 : partial_distinct_rel->userid = input_rel->userid;
5237 90 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5238 90 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
5239 :
5240 90 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5241 :
5242 90 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5243 : parse->targetList);
5244 :
5245 : /* estimate how many distinct rows we'll get from each worker */
5246 90 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5247 : cheapest_partial_path->rows,
5248 : NULL, NULL);
5249 :
5250 : /*
5251 : * Try sorting the cheapest path and incrementally sorting any paths with
5252 : * presorted keys and put a unique paths atop of those. We'll also
5253 : * attempt to reorder the required pathkeys to match the input path's
5254 : * pathkeys as much as possible, in hopes of avoiding a possible need to
5255 : * re-sort.
5256 : */
5257 90 : if (grouping_is_sortable(root->processed_distinctClause))
5258 : {
5259 193 : foreach(lc, input_rel->partial_pathlist)
5260 : {
5261 103 : Path *input_path = (Path *) lfirst(lc);
5262 : Path *sorted_path;
5263 103 : List *useful_pathkeys_list = NIL;
5264 :
5265 : useful_pathkeys_list =
5266 103 : get_useful_pathkeys_for_distinct(root,
5267 : root->distinct_pathkeys,
5268 : input_path->pathkeys);
5269 : Assert(list_length(useful_pathkeys_list) > 0);
5270 :
5271 319 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5272 : {
5273 113 : sorted_path = make_ordered_path(root,
5274 : partial_distinct_rel,
5275 : input_path,
5276 : cheapest_partial_path,
5277 : useful_pathkeys,
5278 : -1.0);
5279 :
5280 113 : if (sorted_path == NULL)
5281 8 : continue;
5282 :
5283 : /*
5284 : * An empty distinct_pathkeys means all tuples have the same
5285 : * value for the DISTINCT clause. See
5286 : * create_final_distinct_paths()
5287 : */
5288 105 : if (root->distinct_pathkeys == NIL)
5289 : {
5290 : Node *limitCount;
5291 :
5292 5 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5293 : sizeof(int64),
5294 : Int64GetDatum(1), false,
5295 : true);
5296 :
5297 : /*
5298 : * Apply a LimitPath onto the partial path to restrict the
5299 : * tuples from each worker to 1.
5300 : * create_final_distinct_paths will need to apply an
5301 : * additional LimitPath to restrict this to a single row
5302 : * after the Gather node. If the query already has a
5303 : * LIMIT clause, then we could end up with three Limit
5304 : * nodes in the final plan. Consolidating the top two of
5305 : * these could be done, but does not seem worth troubling
5306 : * over.
5307 : */
5308 5 : add_partial_path(partial_distinct_rel, (Path *)
5309 5 : create_limit_path(root, partial_distinct_rel,
5310 : sorted_path,
5311 : NULL,
5312 : limitCount,
5313 : LIMIT_OPTION_COUNT,
5314 : 0, 1));
5315 : }
5316 : else
5317 : {
5318 100 : add_partial_path(partial_distinct_rel, (Path *)
5319 100 : create_unique_path(root, partial_distinct_rel,
5320 : sorted_path,
5321 100 : list_length(root->distinct_pathkeys),
5322 : numDistinctRows));
5323 : }
5324 : }
5325 : }
5326 : }
5327 :
5328 : /*
5329 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5330 : * we're not on the hook to ensure we do our best to create at least one
5331 : * path here, we treat enable_hashagg as a hard off-switch rather than the
5332 : * slightly softer variant in create_final_distinct_paths.
5333 : */
5334 90 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5335 : {
5336 65 : add_partial_path(partial_distinct_rel, (Path *)
5337 65 : create_agg_path(root,
5338 : partial_distinct_rel,
5339 : cheapest_partial_path,
5340 : cheapest_partial_path->pathtarget,
5341 : AGG_HASHED,
5342 : AGGSPLIT_SIMPLE,
5343 : root->processed_distinctClause,
5344 : NIL,
5345 : NULL,
5346 : numDistinctRows));
5347 : }
5348 :
5349 : /*
5350 : * If there is an FDW that's responsible for all baserels of the query,
5351 : * let it consider adding ForeignPaths.
5352 : */
5353 90 : if (partial_distinct_rel->fdwroutine &&
5354 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5355 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5356 : UPPERREL_PARTIAL_DISTINCT,
5357 : input_rel,
5358 : partial_distinct_rel,
5359 : NULL);
5360 :
5361 : /* Let extensions possibly add some more partial paths */
5362 90 : if (create_upper_paths_hook)
5363 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5364 : input_rel, partial_distinct_rel, NULL);
5365 :
5366 90 : if (partial_distinct_rel->partial_pathlist != NIL)
5367 : {
5368 90 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5369 90 : set_cheapest(partial_distinct_rel);
5370 :
5371 : /*
5372 : * Finally, create paths to distinctify the final result. This step
5373 : * is needed to remove any duplicates due to combining rows from
5374 : * parallel workers.
5375 : */
5376 90 : create_final_distinct_paths(root, partial_distinct_rel,
5377 : final_distinct_rel);
5378 : }
5379 : }
5380 :
5381 : /*
5382 : * create_final_distinct_paths
5383 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5384 : *
5385 : * input_rel: contains the source-data paths
5386 : * distinct_rel: destination relation for storing created paths
5387 : */
5388 : static RelOptInfo *
5389 2018 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5390 : RelOptInfo *distinct_rel)
5391 : {
5392 2018 : Query *parse = root->parse;
5393 2018 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5394 : double numDistinctRows;
5395 : bool allow_hash;
5396 :
5397 : /* Estimate number of distinct rows there will be */
5398 2018 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5399 1959 : root->hasHavingQual)
5400 : {
5401 : /*
5402 : * If there was grouping or aggregation, use the number of input rows
5403 : * as the estimated number of DISTINCT rows (ie, assume the input is
5404 : * already mostly unique).
5405 : */
5406 59 : numDistinctRows = cheapest_input_path->rows;
5407 : }
5408 : else
5409 : {
5410 : /*
5411 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5412 : */
5413 : List *distinctExprs;
5414 :
5415 1959 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5416 : parse->targetList);
5417 1959 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5418 : cheapest_input_path->rows,
5419 : NULL, NULL);
5420 : }
5421 :
5422 : /*
5423 : * Consider sort-based implementations of DISTINCT, if possible.
5424 : */
5425 2018 : if (grouping_is_sortable(root->processed_distinctClause))
5426 : {
5427 : /*
5428 : * Firstly, if we have any adequately-presorted paths, just stick a
5429 : * Unique node on those. We also, consider doing an explicit sort of
5430 : * the cheapest input path and Unique'ing that. If any paths have
5431 : * presorted keys then we'll create an incremental sort atop of those
5432 : * before adding a unique node on the top. We'll also attempt to
5433 : * reorder the required pathkeys to match the input path's pathkeys as
5434 : * much as possible, in hopes of avoiding a possible need to re-sort.
5435 : *
5436 : * When we have DISTINCT ON, we must sort by the more rigorous of
5437 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5438 : * Also, if we do have to do an explicit sort, we might as well use
5439 : * the more rigorous ordering to avoid a second sort later. (Note
5440 : * that the parser will have ensured that one clause is a prefix of
5441 : * the other.)
5442 : */
5443 : List *needed_pathkeys;
5444 : ListCell *lc;
5445 2013 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5446 :
5447 2200 : if (parse->hasDistinctOn &&
5448 187 : list_length(root->distinct_pathkeys) <
5449 187 : list_length(root->sort_pathkeys))
5450 53 : needed_pathkeys = root->sort_pathkeys;
5451 : else
5452 1960 : needed_pathkeys = root->distinct_pathkeys;
5453 :
5454 5335 : foreach(lc, input_rel->pathlist)
5455 : {
5456 3322 : Path *input_path = (Path *) lfirst(lc);
5457 : Path *sorted_path;
5458 3322 : List *useful_pathkeys_list = NIL;
5459 :
5460 : useful_pathkeys_list =
5461 3322 : get_useful_pathkeys_for_distinct(root,
5462 : needed_pathkeys,
5463 : input_path->pathkeys);
5464 : Assert(list_length(useful_pathkeys_list) > 0);
5465 :
5466 10466 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5467 : {
5468 3822 : sorted_path = make_ordered_path(root,
5469 : distinct_rel,
5470 : input_path,
5471 : cheapest_input_path,
5472 : useful_pathkeys,
5473 : limittuples);
5474 :
5475 3822 : if (sorted_path == NULL)
5476 531 : continue;
5477 :
5478 : /*
5479 : * distinct_pathkeys may have become empty if all of the
5480 : * pathkeys were determined to be redundant. If all of the
5481 : * pathkeys are redundant then each DISTINCT target must only
5482 : * allow a single value, therefore all resulting tuples must
5483 : * be identical (or at least indistinguishable by an equality
5484 : * check). We can uniquify these tuples simply by just taking
5485 : * the first tuple. All we do here is add a path to do "LIMIT
5486 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5487 : * still have a non-NIL sort_pathkeys list, so we must still
5488 : * only do this with paths which are correctly sorted by
5489 : * sort_pathkeys.
5490 : */
5491 3291 : if (root->distinct_pathkeys == NIL)
5492 : {
5493 : Node *limitCount;
5494 :
5495 96 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5496 : sizeof(int64),
5497 : Int64GetDatum(1), false,
5498 : true);
5499 :
5500 : /*
5501 : * If the query already has a LIMIT clause, then we could
5502 : * end up with a duplicate LimitPath in the final plan.
5503 : * That does not seem worth troubling over too much.
5504 : */
5505 96 : add_path(distinct_rel, (Path *)
5506 96 : create_limit_path(root, distinct_rel, sorted_path,
5507 : NULL, limitCount,
5508 : LIMIT_OPTION_COUNT, 0, 1));
5509 : }
5510 : else
5511 : {
5512 3195 : add_path(distinct_rel, (Path *)
5513 3195 : create_unique_path(root, distinct_rel,
5514 : sorted_path,
5515 3195 : list_length(root->distinct_pathkeys),
5516 : numDistinctRows));
5517 : }
5518 : }
5519 : }
5520 : }
5521 :
5522 : /*
5523 : * Consider hash-based implementations of DISTINCT, if possible.
5524 : *
5525 : * If we were not able to make any other types of path, we *must* hash or
5526 : * die trying. If we do have other choices, there are two things that
5527 : * should prevent selection of hashing: if the query uses DISTINCT ON
5528 : * (because it won't really have the expected behavior if we hash), or if
5529 : * enable_hashagg is off.
5530 : *
5531 : * Note: grouping_is_hashable() is much more expensive to check than the
5532 : * other gating conditions, so we want to do it last.
5533 : */
5534 2018 : if (distinct_rel->pathlist == NIL)
5535 5 : allow_hash = true; /* we have no alternatives */
5536 2013 : else if (parse->hasDistinctOn || !enable_hashagg)
5537 312 : allow_hash = false; /* policy-based decision not to hash */
5538 : else
5539 1701 : allow_hash = true; /* default */
5540 :
5541 2018 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5542 : {
5543 : /* Generate hashed aggregate path --- no sort needed */
5544 1706 : add_path(distinct_rel, (Path *)
5545 1706 : create_agg_path(root,
5546 : distinct_rel,
5547 : cheapest_input_path,
5548 : cheapest_input_path->pathtarget,
5549 : AGG_HASHED,
5550 : AGGSPLIT_SIMPLE,
5551 : root->processed_distinctClause,
5552 : NIL,
5553 : NULL,
5554 : numDistinctRows));
5555 : }
5556 :
5557 2018 : return distinct_rel;
5558 : }
5559 :
5560 : /*
5561 : * get_useful_pathkeys_for_distinct
5562 : * Get useful orderings of pathkeys for distinctClause by reordering
5563 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5564 : *
5565 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5566 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5567 : */
5568 : static List *
5569 3425 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5570 : List *path_pathkeys)
5571 : {
5572 3425 : List *useful_pathkeys_list = NIL;
5573 3425 : List *useful_pathkeys = NIL;
5574 :
5575 : /* always include the given 'needed_pathkeys' */
5576 3425 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5577 : needed_pathkeys);
5578 :
5579 3425 : if (!enable_distinct_reordering)
5580 0 : return useful_pathkeys_list;
5581 :
5582 : /*
5583 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5584 : * that match 'needed_pathkeys', but only up to the longest matching
5585 : * prefix.
5586 : *
5587 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5588 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5589 : * the desired behavior.
5590 : */
5591 8553 : foreach_node(PathKey, pathkey, path_pathkeys)
5592 : {
5593 : /*
5594 : * The PathKey nodes are canonical, so they can be checked for
5595 : * equality by simple pointer comparison.
5596 : */
5597 1725 : if (!list_member_ptr(needed_pathkeys, pathkey))
5598 7 : break;
5599 1718 : if (root->parse->hasDistinctOn &&
5600 139 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5601 15 : break;
5602 :
5603 1703 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5604 : }
5605 :
5606 : /* If no match at all, no point in reordering needed_pathkeys */
5607 3425 : if (useful_pathkeys == NIL)
5608 1970 : return useful_pathkeys_list;
5609 :
5610 : /*
5611 : * If not full match, the resulting pathkey list is not useful without
5612 : * incremental sort.
5613 : */
5614 1455 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5615 924 : !enable_incremental_sort)
5616 48 : return useful_pathkeys_list;
5617 :
5618 : /* Append the remaining PathKey nodes in needed_pathkeys */
5619 1407 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5620 : needed_pathkeys);
5621 :
5622 : /*
5623 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5624 : * just drop it.
5625 : */
5626 1407 : if (compare_pathkeys(needed_pathkeys,
5627 : useful_pathkeys) == PATHKEYS_EQUAL)
5628 897 : return useful_pathkeys_list;
5629 :
5630 510 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5631 : useful_pathkeys);
5632 :
5633 510 : return useful_pathkeys_list;
5634 : }
5635 :
5636 : /*
5637 : * create_ordered_paths
5638 : *
5639 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5640 : *
5641 : * All paths in the result must satisfy the ORDER BY ordering.
5642 : * The only new paths we need consider are an explicit full sort
5643 : * and incremental sort on the cheapest-total existing path.
5644 : *
5645 : * input_rel: contains the source-data Paths
5646 : * target: the output tlist the result Paths must emit
5647 : * limit_tuples: estimated bound on the number of output tuples,
5648 : * or -1 if no LIMIT or couldn't estimate
5649 : *
5650 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5651 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5652 : */
5653 : static RelOptInfo *
5654 60879 : create_ordered_paths(PlannerInfo *root,
5655 : RelOptInfo *input_rel,
5656 : PathTarget *target,
5657 : bool target_parallel_safe,
5658 : double limit_tuples)
5659 : {
5660 60879 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5661 : RelOptInfo *ordered_rel;
5662 : ListCell *lc;
5663 :
5664 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5665 60879 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5666 :
5667 : /*
5668 : * If the input relation is not parallel-safe, then the ordered relation
5669 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5670 : * target list is parallel-safe.
5671 : */
5672 60879 : if (input_rel->consider_parallel && target_parallel_safe)
5673 42762 : ordered_rel->consider_parallel = true;
5674 :
5675 : /* Assume that the same path generation strategies are allowed. */
5676 60879 : ordered_rel->pgs_mask = input_rel->pgs_mask;
5677 :
5678 : /*
5679 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5680 : */
5681 60879 : ordered_rel->serverid = input_rel->serverid;
5682 60879 : ordered_rel->userid = input_rel->userid;
5683 60879 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5684 60879 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5685 :
5686 155163 : foreach(lc, input_rel->pathlist)
5687 : {
5688 94284 : Path *input_path = (Path *) lfirst(lc);
5689 : Path *sorted_path;
5690 : bool is_sorted;
5691 : int presorted_keys;
5692 :
5693 94284 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5694 : input_path->pathkeys, &presorted_keys);
5695 :
5696 94284 : if (is_sorted)
5697 36364 : sorted_path = input_path;
5698 : else
5699 : {
5700 : /*
5701 : * Try at least sorting the cheapest path and also try
5702 : * incrementally sorting any path which is partially sorted
5703 : * already (no need to deal with paths which have presorted keys
5704 : * when incremental sort is disabled unless it's the cheapest
5705 : * input path).
5706 : */
5707 57920 : if (input_path != cheapest_input_path &&
5708 4575 : (presorted_keys == 0 || !enable_incremental_sort))
5709 1603 : continue;
5710 :
5711 : /*
5712 : * We've no need to consider both a sort and incremental sort.
5713 : * We'll just do a sort if there are no presorted keys and an
5714 : * incremental sort when there are presorted keys.
5715 : */
5716 56317 : if (presorted_keys == 0 || !enable_incremental_sort)
5717 52897 : sorted_path = (Path *) create_sort_path(root,
5718 : ordered_rel,
5719 : input_path,
5720 : root->sort_pathkeys,
5721 : limit_tuples);
5722 : else
5723 3420 : sorted_path = (Path *) create_incremental_sort_path(root,
5724 : ordered_rel,
5725 : input_path,
5726 : root->sort_pathkeys,
5727 : presorted_keys,
5728 : limit_tuples);
5729 : }
5730 :
5731 : /*
5732 : * If the pathtarget of the result path has different expressions from
5733 : * the target to be applied, a projection step is needed.
5734 : */
5735 92681 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5736 298 : sorted_path = apply_projection_to_path(root, ordered_rel,
5737 : sorted_path, target);
5738 :
5739 92681 : add_path(ordered_rel, sorted_path);
5740 : }
5741 :
5742 : /*
5743 : * generate_gather_paths() will have already generated a simple Gather
5744 : * path for the best parallel path, if any, and the loop above will have
5745 : * considered sorting it. Similarly, generate_gather_paths() will also
5746 : * have generated order-preserving Gather Merge plans which can be used
5747 : * without sorting if they happen to match the sort_pathkeys, and the loop
5748 : * above will have handled those as well. However, there's one more
5749 : * possibility: it may make sense to sort the cheapest partial path or
5750 : * incrementally sort any partial path that is partially sorted according
5751 : * to the required output order and then use Gather Merge.
5752 : */
5753 60879 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5754 42642 : input_rel->partial_pathlist != NIL)
5755 : {
5756 : Path *cheapest_partial_path;
5757 :
5758 2367 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5759 :
5760 5289 : foreach(lc, input_rel->partial_pathlist)
5761 : {
5762 2922 : Path *input_path = (Path *) lfirst(lc);
5763 : Path *sorted_path;
5764 : bool is_sorted;
5765 : int presorted_keys;
5766 : double total_groups;
5767 :
5768 2922 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5769 : input_path->pathkeys,
5770 : &presorted_keys);
5771 :
5772 2922 : if (is_sorted)
5773 476 : continue;
5774 :
5775 : /*
5776 : * Try at least sorting the cheapest path and also try
5777 : * incrementally sorting any path which is partially sorted
5778 : * already (no need to deal with paths which have presorted keys
5779 : * when incremental sort is disabled unless it's the cheapest
5780 : * partial path).
5781 : */
5782 2446 : if (input_path != cheapest_partial_path &&
5783 101 : (presorted_keys == 0 || !enable_incremental_sort))
5784 0 : continue;
5785 :
5786 : /*
5787 : * We've no need to consider both a sort and incremental sort.
5788 : * We'll just do a sort if there are no presorted keys and an
5789 : * incremental sort when there are presorted keys.
5790 : */
5791 2446 : if (presorted_keys == 0 || !enable_incremental_sort)
5792 2330 : sorted_path = (Path *) create_sort_path(root,
5793 : ordered_rel,
5794 : input_path,
5795 : root->sort_pathkeys,
5796 : limit_tuples);
5797 : else
5798 116 : sorted_path = (Path *) create_incremental_sort_path(root,
5799 : ordered_rel,
5800 : input_path,
5801 : root->sort_pathkeys,
5802 : presorted_keys,
5803 : limit_tuples);
5804 2446 : total_groups = compute_gather_rows(sorted_path);
5805 : sorted_path = (Path *)
5806 2446 : create_gather_merge_path(root, ordered_rel,
5807 : sorted_path,
5808 : sorted_path->pathtarget,
5809 : root->sort_pathkeys, NULL,
5810 : &total_groups);
5811 :
5812 : /*
5813 : * If the pathtarget of the result path has different expressions
5814 : * from the target to be applied, a projection step is needed.
5815 : */
5816 2446 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5817 5 : sorted_path = apply_projection_to_path(root, ordered_rel,
5818 : sorted_path, target);
5819 :
5820 2446 : add_path(ordered_rel, sorted_path);
5821 : }
5822 : }
5823 :
5824 : /*
5825 : * If there is an FDW that's responsible for all baserels of the query,
5826 : * let it consider adding ForeignPaths.
5827 : */
5828 60879 : if (ordered_rel->fdwroutine &&
5829 197 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5830 189 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5831 : input_rel, ordered_rel,
5832 : NULL);
5833 :
5834 : /* Let extensions possibly add some more paths */
5835 60879 : if (create_upper_paths_hook)
5836 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5837 : input_rel, ordered_rel, NULL);
5838 :
5839 : /*
5840 : * No need to bother with set_cheapest here; grouping_planner does not
5841 : * need us to do it.
5842 : */
5843 : Assert(ordered_rel->pathlist != NIL);
5844 :
5845 60879 : return ordered_rel;
5846 : }
5847 :
5848 :
5849 : /*
5850 : * make_group_input_target
5851 : * Generate appropriate PathTarget for initial input to grouping nodes.
5852 : *
5853 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5854 : * the query's final targetlist; for example, it certainly can't emit any
5855 : * aggregate function calls. This routine generates the correct target
5856 : * for the scan/join subplan.
5857 : *
5858 : * The query target list passed from the parser already contains entries
5859 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5860 : * for variables used only in HAVING clauses; so we need to add those
5861 : * variables to the subplan target list. Also, we flatten all expressions
5862 : * except GROUP BY items into their component variables; other expressions
5863 : * will be computed by the upper plan nodes rather than by the subplan.
5864 : * For example, given a query like
5865 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5866 : * we want to pass this targetlist to the subplan:
5867 : * a+b,c,d
5868 : * where the a+b target will be used by the Sort/Group steps, and the
5869 : * other targets will be used for computing the final results.
5870 : *
5871 : * 'final_target' is the query's final target list (in PathTarget form)
5872 : *
5873 : * The result is the PathTarget to be computed by the Paths returned from
5874 : * query_planner().
5875 : */
5876 : static PathTarget *
5877 33942 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5878 : {
5879 33942 : Query *parse = root->parse;
5880 : PathTarget *input_target;
5881 : List *non_group_cols;
5882 : List *non_group_vars;
5883 : int i;
5884 : ListCell *lc;
5885 :
5886 : /*
5887 : * We must build a target containing all grouping columns, plus any other
5888 : * Vars mentioned in the query's targetlist and HAVING qual.
5889 : */
5890 33942 : input_target = create_empty_pathtarget();
5891 33942 : non_group_cols = NIL;
5892 :
5893 33942 : i = 0;
5894 85575 : foreach(lc, final_target->exprs)
5895 : {
5896 51633 : Expr *expr = (Expr *) lfirst(lc);
5897 51633 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5898 :
5899 59835 : if (sgref && root->processed_groupClause &&
5900 8202 : get_sortgroupref_clause_noerr(sgref,
5901 : root->processed_groupClause) != NULL)
5902 : {
5903 : /*
5904 : * It's a grouping column, so add it to the input target as-is.
5905 : *
5906 : * Note that the target is logically below the grouping step. So
5907 : * with grouping sets we need to remove the RT index of the
5908 : * grouping step if there is any from the target expression.
5909 : */
5910 6686 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5911 : {
5912 : Assert(root->group_rtindex > 0);
5913 : expr = (Expr *)
5914 1803 : remove_nulling_relids((Node *) expr,
5915 1803 : bms_make_singleton(root->group_rtindex),
5916 : NULL);
5917 : }
5918 6686 : add_column_to_pathtarget(input_target, expr, sgref);
5919 : }
5920 : else
5921 : {
5922 : /*
5923 : * Non-grouping column, so just remember the expression for later
5924 : * call to pull_var_clause.
5925 : */
5926 44947 : non_group_cols = lappend(non_group_cols, expr);
5927 : }
5928 :
5929 51633 : i++;
5930 : }
5931 :
5932 : /*
5933 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5934 : */
5935 33942 : if (parse->havingQual)
5936 840 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5937 :
5938 : /*
5939 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5940 : * add them to the input target if not already present. (A Var used
5941 : * directly as a GROUP BY item will be present already.) Note this
5942 : * includes Vars used in resjunk items, so we are covering the needs of
5943 : * ORDER BY and window specifications. Vars used within Aggrefs and
5944 : * WindowFuncs will be pulled out here, too.
5945 : *
5946 : * Note that the target is logically below the grouping step. So with
5947 : * grouping sets we need to remove the RT index of the grouping step if
5948 : * there is any from the non-group Vars.
5949 : */
5950 33942 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5951 : PVC_RECURSE_AGGREGATES |
5952 : PVC_RECURSE_WINDOWFUNCS |
5953 : PVC_INCLUDE_PLACEHOLDERS);
5954 33942 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5955 : {
5956 : Assert(root->group_rtindex > 0);
5957 : non_group_vars = (List *)
5958 836 : remove_nulling_relids((Node *) non_group_vars,
5959 836 : bms_make_singleton(root->group_rtindex),
5960 : NULL);
5961 : }
5962 33942 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5963 :
5964 : /* clean up cruft */
5965 33942 : list_free(non_group_vars);
5966 33942 : list_free(non_group_cols);
5967 :
5968 : /* XXX this causes some redundant cost calculation ... */
5969 33942 : return set_pathtarget_cost_width(root, input_target);
5970 : }
5971 :
5972 : /*
5973 : * make_partial_grouping_target
5974 : * Generate appropriate PathTarget for output of partial aggregate
5975 : * (or partial grouping, if there are no aggregates) nodes.
5976 : *
5977 : * A partial aggregation node needs to emit all the same aggregates that
5978 : * a regular aggregation node would, plus any aggregates used in HAVING;
5979 : * except that the Aggref nodes should be marked as partial aggregates.
5980 : *
5981 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5982 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5983 : * these would be Vars that are grouped by or used in grouping expressions.)
5984 : *
5985 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5986 : * havingQual represents the HAVING clause.
5987 : */
5988 : static PathTarget *
5989 3216 : make_partial_grouping_target(PlannerInfo *root,
5990 : PathTarget *grouping_target,
5991 : Node *havingQual)
5992 : {
5993 : PathTarget *partial_target;
5994 : List *non_group_cols;
5995 : List *non_group_exprs;
5996 : int i;
5997 : ListCell *lc;
5998 :
5999 3216 : partial_target = create_empty_pathtarget();
6000 3216 : non_group_cols = NIL;
6001 :
6002 3216 : i = 0;
6003 10593 : foreach(lc, grouping_target->exprs)
6004 : {
6005 7377 : Expr *expr = (Expr *) lfirst(lc);
6006 7377 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
6007 :
6008 11319 : if (sgref && root->processed_groupClause &&
6009 3942 : get_sortgroupref_clause_noerr(sgref,
6010 : root->processed_groupClause) != NULL)
6011 : {
6012 : /*
6013 : * It's a grouping column, so add it to the partial_target as-is.
6014 : * (This allows the upper agg step to repeat the grouping calcs.)
6015 : */
6016 2340 : add_column_to_pathtarget(partial_target, expr, sgref);
6017 : }
6018 : else
6019 : {
6020 : /*
6021 : * Non-grouping column, so just remember the expression for later
6022 : * call to pull_var_clause.
6023 : */
6024 5037 : non_group_cols = lappend(non_group_cols, expr);
6025 : }
6026 :
6027 7377 : i++;
6028 : }
6029 :
6030 : /*
6031 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
6032 : */
6033 3216 : if (havingQual)
6034 729 : non_group_cols = lappend(non_group_cols, havingQual);
6035 :
6036 : /*
6037 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
6038 : * non-group cols (plus HAVING), and add them to the partial_target if not
6039 : * already present. (An expression used directly as a GROUP BY item will
6040 : * be present already.) Note this includes Vars used in resjunk items, so
6041 : * we are covering the needs of ORDER BY and window specifications.
6042 : */
6043 3216 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
6044 : PVC_INCLUDE_AGGREGATES |
6045 : PVC_RECURSE_WINDOWFUNCS |
6046 : PVC_INCLUDE_PLACEHOLDERS);
6047 :
6048 3216 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
6049 :
6050 : /*
6051 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
6052 : * are at the top level of the target list, so we can just scan the list
6053 : * rather than recursing through the expression trees.
6054 : */
6055 11107 : foreach(lc, partial_target->exprs)
6056 : {
6057 7891 : Aggref *aggref = (Aggref *) lfirst(lc);
6058 :
6059 7891 : if (IsA(aggref, Aggref))
6060 : {
6061 : Aggref *newaggref;
6062 :
6063 : /*
6064 : * We shouldn't need to copy the substructure of the Aggref node,
6065 : * but flat-copy the node itself to avoid damaging other trees.
6066 : */
6067 5526 : newaggref = makeNode(Aggref);
6068 5526 : memcpy(newaggref, aggref, sizeof(Aggref));
6069 :
6070 : /* For now, assume serialization is required */
6071 5526 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
6072 :
6073 5526 : lfirst(lc) = newaggref;
6074 : }
6075 : }
6076 :
6077 : /* clean up cruft */
6078 3216 : list_free(non_group_exprs);
6079 3216 : list_free(non_group_cols);
6080 :
6081 : /* XXX this causes some redundant cost calculation ... */
6082 3216 : return set_pathtarget_cost_width(root, partial_target);
6083 : }
6084 :
6085 : /*
6086 : * mark_partial_aggref
6087 : * Adjust an Aggref to make it represent a partial-aggregation step.
6088 : *
6089 : * The Aggref node is modified in-place; caller must do any copying required.
6090 : */
6091 : void
6092 15395 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
6093 : {
6094 : /* aggtranstype should be computed by this point */
6095 : Assert(OidIsValid(agg->aggtranstype));
6096 : /* ... but aggsplit should still be as the parser left it */
6097 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
6098 :
6099 : /* Mark the Aggref with the intended partial-aggregation mode */
6100 15395 : agg->aggsplit = aggsplit;
6101 :
6102 : /*
6103 : * Adjust result type if needed. Normally, a partial aggregate returns
6104 : * the aggregate's transition type; but if that's INTERNAL and we're
6105 : * serializing, it returns BYTEA instead.
6106 : */
6107 15395 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
6108 : {
6109 13480 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
6110 261 : agg->aggtype = BYTEAOID;
6111 : else
6112 13219 : agg->aggtype = agg->aggtranstype;
6113 : }
6114 15395 : }
6115 :
6116 : /*
6117 : * postprocess_setop_tlist
6118 : * Fix up targetlist returned by plan_set_operations().
6119 : *
6120 : * We need to transpose sort key info from the orig_tlist into new_tlist.
6121 : * NOTE: this would not be good enough if we supported resjunk sort keys
6122 : * for results of set operations --- then, we'd need to project a whole
6123 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
6124 : * find any resjunk columns in orig_tlist.
6125 : */
6126 : static List *
6127 4827 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
6128 : {
6129 : ListCell *l;
6130 4827 : ListCell *orig_tlist_item = list_head(orig_tlist);
6131 :
6132 18622 : foreach(l, new_tlist)
6133 : {
6134 13795 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
6135 : TargetEntry *orig_tle;
6136 :
6137 : /* ignore resjunk columns in setop result */
6138 13795 : if (new_tle->resjunk)
6139 0 : continue;
6140 :
6141 : Assert(orig_tlist_item != NULL);
6142 13795 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
6143 13795 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
6144 13795 : if (orig_tle->resjunk) /* should not happen */
6145 0 : elog(ERROR, "resjunk output columns are not implemented");
6146 : Assert(new_tle->resno == orig_tle->resno);
6147 13795 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
6148 : }
6149 4827 : if (orig_tlist_item != NULL)
6150 0 : elog(ERROR, "resjunk output columns are not implemented");
6151 4827 : return new_tlist;
6152 : }
6153 :
6154 : /*
6155 : * optimize_window_clauses
6156 : * Call each WindowFunc's prosupport function to see if we're able to
6157 : * make any adjustments to any of the WindowClause's so that the executor
6158 : * can execute the window functions in a more optimal way.
6159 : *
6160 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
6161 : * may allow more things to be done here in the future.
6162 : */
6163 : static void
6164 2232 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
6165 : {
6166 2232 : List *windowClause = root->parse->windowClause;
6167 : ListCell *lc;
6168 :
6169 4669 : foreach(lc, windowClause)
6170 : {
6171 2437 : WindowClause *wc = lfirst_node(WindowClause, lc);
6172 : ListCell *lc2;
6173 2437 : int optimizedFrameOptions = 0;
6174 :
6175 : Assert(wc->winref <= wflists->maxWinRef);
6176 :
6177 : /* skip any WindowClauses that have no WindowFuncs */
6178 2437 : if (wflists->windowFuncs[wc->winref] == NIL)
6179 20 : continue;
6180 :
6181 2960 : foreach(lc2, wflists->windowFuncs[wc->winref])
6182 : {
6183 : SupportRequestOptimizeWindowClause req;
6184 : SupportRequestOptimizeWindowClause *res;
6185 2452 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
6186 : Oid prosupport;
6187 :
6188 2452 : prosupport = get_func_support(wfunc->winfnoid);
6189 :
6190 : /* Check if there's a support function for 'wfunc' */
6191 2452 : if (!OidIsValid(prosupport))
6192 1909 : break; /* can't optimize this WindowClause */
6193 :
6194 730 : req.type = T_SupportRequestOptimizeWindowClause;
6195 730 : req.window_clause = wc;
6196 730 : req.window_func = wfunc;
6197 730 : req.frameOptions = wc->frameOptions;
6198 :
6199 : /* call the support function */
6200 : res = (SupportRequestOptimizeWindowClause *)
6201 730 : DatumGetPointer(OidFunctionCall1(prosupport,
6202 : PointerGetDatum(&req)));
6203 :
6204 : /*
6205 : * Skip to next WindowClause if the support function does not
6206 : * support this request type.
6207 : */
6208 730 : if (res == NULL)
6209 187 : break;
6210 :
6211 : /*
6212 : * Save these frameOptions for the first WindowFunc for this
6213 : * WindowClause.
6214 : */
6215 543 : if (foreach_current_index(lc2) == 0)
6216 523 : optimizedFrameOptions = res->frameOptions;
6217 :
6218 : /*
6219 : * On subsequent WindowFuncs, if the frameOptions are not the same
6220 : * then we're unable to optimize the frameOptions for this
6221 : * WindowClause.
6222 : */
6223 20 : else if (optimizedFrameOptions != res->frameOptions)
6224 0 : break; /* skip to the next WindowClause, if any */
6225 : }
6226 :
6227 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
6228 2417 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
6229 : {
6230 : ListCell *lc3;
6231 :
6232 : /* apply the new frame options */
6233 508 : wc->frameOptions = optimizedFrameOptions;
6234 :
6235 : /*
6236 : * We now check to see if changing the frameOptions has caused
6237 : * this WindowClause to be a duplicate of some other WindowClause.
6238 : * This can only happen if we have multiple WindowClauses, so
6239 : * don't bother if there's only 1.
6240 : */
6241 508 : if (list_length(windowClause) == 1)
6242 433 : continue;
6243 :
6244 : /*
6245 : * Do the duplicate check and reuse the existing WindowClause if
6246 : * we find a duplicate.
6247 : */
6248 190 : foreach(lc3, windowClause)
6249 : {
6250 145 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
6251 :
6252 : /* skip over the WindowClause we're currently editing */
6253 145 : if (existing_wc == wc)
6254 45 : continue;
6255 :
6256 : /*
6257 : * Perform the same duplicate check that is done in
6258 : * transformWindowFuncCall.
6259 : */
6260 200 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
6261 100 : equal(wc->orderClause, existing_wc->orderClause) &&
6262 100 : wc->frameOptions == existing_wc->frameOptions &&
6263 60 : equal(wc->startOffset, existing_wc->startOffset) &&
6264 30 : equal(wc->endOffset, existing_wc->endOffset))
6265 : {
6266 : ListCell *lc4;
6267 :
6268 : /*
6269 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
6270 : * This required adjusting each WindowFunc's winref and
6271 : * moving the WindowFuncs in 'wc' to the list of
6272 : * WindowFuncs in 'existing_wc'.
6273 : */
6274 65 : foreach(lc4, wflists->windowFuncs[wc->winref])
6275 : {
6276 35 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
6277 :
6278 35 : wfunc->winref = existing_wc->winref;
6279 : }
6280 :
6281 : /* move list items */
6282 60 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
6283 30 : wflists->windowFuncs[wc->winref]);
6284 30 : wflists->windowFuncs[wc->winref] = NIL;
6285 :
6286 : /*
6287 : * transformWindowFuncCall() should have made sure there
6288 : * are no other duplicates, so we needn't bother looking
6289 : * any further.
6290 : */
6291 30 : break;
6292 : }
6293 : }
6294 : }
6295 : }
6296 2232 : }
6297 :
6298 : /*
6299 : * select_active_windows
6300 : * Create a list of the "active" window clauses (ie, those referenced
6301 : * by non-deleted WindowFuncs) in the order they are to be executed.
6302 : */
6303 : static List *
6304 2232 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
6305 : {
6306 2232 : List *windowClause = root->parse->windowClause;
6307 2232 : List *result = NIL;
6308 : ListCell *lc;
6309 2232 : int nActive = 0;
6310 2232 : WindowClauseSortData *actives = palloc_array(WindowClauseSortData,
6311 : list_length(windowClause));
6312 :
6313 : /* First, construct an array of the active windows */
6314 4669 : foreach(lc, windowClause)
6315 : {
6316 2437 : WindowClause *wc = lfirst_node(WindowClause, lc);
6317 :
6318 : /* It's only active if wflists shows some related WindowFuncs */
6319 : Assert(wc->winref <= wflists->maxWinRef);
6320 2437 : if (wflists->windowFuncs[wc->winref] == NIL)
6321 50 : continue;
6322 :
6323 2387 : actives[nActive].wc = wc; /* original clause */
6324 :
6325 : /*
6326 : * For sorting, we want the list of partition keys followed by the
6327 : * list of sort keys. But pathkeys construction will remove duplicates
6328 : * between the two, so we can as well (even though we can't detect all
6329 : * of the duplicates, since some may come from ECs - that might mean
6330 : * we miss optimization chances here). We must, however, ensure that
6331 : * the order of entries is preserved with respect to the ones we do
6332 : * keep.
6333 : *
6334 : * partitionClause and orderClause had their own duplicates removed in
6335 : * parse analysis, so we're only concerned here with removing
6336 : * orderClause entries that also appear in partitionClause.
6337 : */
6338 4774 : actives[nActive].uniqueOrder =
6339 2387 : list_concat_unique(list_copy(wc->partitionClause),
6340 2387 : wc->orderClause);
6341 2387 : nActive++;
6342 : }
6343 :
6344 : /*
6345 : * Sort active windows by their partitioning/ordering clauses, ignoring
6346 : * any framing clauses, so that the windows that need the same sorting are
6347 : * adjacent in the list. When we come to generate paths, this will avoid
6348 : * inserting additional Sort nodes.
6349 : *
6350 : * This is how we implement a specific requirement from the SQL standard,
6351 : * which says that when two or more windows are order-equivalent (i.e.
6352 : * have matching partition and order clauses, even if their names or
6353 : * framing clauses differ), then all peer rows must be presented in the
6354 : * same order in all of them. If we allowed multiple sort nodes for such
6355 : * cases, we'd risk having the peer rows end up in different orders in
6356 : * equivalent windows due to sort instability. (See General Rule 4 of
6357 : * <window clause> in SQL2008 - SQL2016.)
6358 : *
6359 : * Additionally, if the entire list of clauses of one window is a prefix
6360 : * of another, put first the window with stronger sorting requirements.
6361 : * This way we will first sort for stronger window, and won't have to sort
6362 : * again for the weaker one.
6363 : */
6364 2232 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6365 :
6366 : /* build ordered list of the original WindowClause nodes */
6367 4619 : for (int i = 0; i < nActive; i++)
6368 2387 : result = lappend(result, actives[i].wc);
6369 :
6370 2232 : pfree(actives);
6371 :
6372 2232 : return result;
6373 : }
6374 :
6375 : /*
6376 : * name_active_windows
6377 : * Ensure all active windows have unique names.
6378 : *
6379 : * The parser will have checked that user-assigned window names are unique
6380 : * within the Query. Here we assign made-up names to any unnamed
6381 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6382 : * at parse time, because it'd mess up decompilation of views.)
6383 : *
6384 : * activeWindows: result of select_active_windows
6385 : */
6386 : static void
6387 2232 : name_active_windows(List *activeWindows)
6388 : {
6389 2232 : int next_n = 1;
6390 : char newname[16];
6391 : ListCell *lc;
6392 :
6393 4619 : foreach(lc, activeWindows)
6394 : {
6395 2387 : WindowClause *wc = lfirst_node(WindowClause, lc);
6396 :
6397 : /* Nothing to do if it has a name already. */
6398 2387 : if (wc->name)
6399 480 : continue;
6400 :
6401 : /* Select a name not currently present in the list. */
6402 : for (;;)
6403 5 : {
6404 : ListCell *lc2;
6405 :
6406 1912 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6407 4114 : foreach(lc2, activeWindows)
6408 : {
6409 2207 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6410 :
6411 2207 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6412 5 : break; /* matched */
6413 : }
6414 1912 : if (lc2 == NULL)
6415 1907 : break; /* reached the end with no match */
6416 : }
6417 1907 : wc->name = pstrdup(newname);
6418 : }
6419 2232 : }
6420 :
6421 : /*
6422 : * common_prefix_cmp
6423 : * QSort comparison function for WindowClauseSortData
6424 : *
6425 : * Sort the windows by the required sorting clauses. First, compare the sort
6426 : * clauses themselves. Second, if one window's clauses are a prefix of another
6427 : * one's clauses, put the window with more sort clauses first.
6428 : *
6429 : * We purposefully sort by the highest tleSortGroupRef first. Since
6430 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6431 : * and because here we sort the lowest tleSortGroupRefs last, if a
6432 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6433 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6434 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6435 : * reducing the total number of sorts required for the query.
6436 : */
6437 : static int
6438 170 : common_prefix_cmp(const void *a, const void *b)
6439 : {
6440 170 : const WindowClauseSortData *wcsa = a;
6441 170 : const WindowClauseSortData *wcsb = b;
6442 : ListCell *item_a;
6443 : ListCell *item_b;
6444 :
6445 305 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6446 : {
6447 220 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6448 220 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6449 :
6450 220 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6451 85 : return -1;
6452 210 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6453 55 : return 1;
6454 155 : else if (sca->sortop > scb->sortop)
6455 0 : return -1;
6456 155 : else if (sca->sortop < scb->sortop)
6457 20 : return 1;
6458 135 : else if (sca->nulls_first && !scb->nulls_first)
6459 0 : return -1;
6460 135 : else if (!sca->nulls_first && scb->nulls_first)
6461 0 : return 1;
6462 : /* no need to compare eqop, since it is fully determined by sortop */
6463 : }
6464 :
6465 85 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6466 5 : return -1;
6467 80 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6468 25 : return 1;
6469 :
6470 55 : return 0;
6471 : }
6472 :
6473 : /*
6474 : * make_window_input_target
6475 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6476 : *
6477 : * When the query has window functions, this function computes the desired
6478 : * target to be computed by the node just below the first WindowAgg.
6479 : * This tlist must contain all values needed to evaluate the window functions,
6480 : * compute the final target list, and perform any required final sort step.
6481 : * If multiple WindowAggs are needed, each intermediate one adds its window
6482 : * function results onto this base tlist; only the topmost WindowAgg computes
6483 : * the actual desired target list.
6484 : *
6485 : * This function is much like make_group_input_target, though not quite enough
6486 : * like it to share code. As in that function, we flatten most expressions
6487 : * into their component variables. But we do not want to flatten window
6488 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6489 : * evaluations of them, which would be bad (possibly even resulting in
6490 : * inconsistent answers, if they contain volatile functions).
6491 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6492 : * make_group_input_target, because we may no longer have access to the
6493 : * individual Vars in them.
6494 : *
6495 : * Another key difference from make_group_input_target is that we don't
6496 : * flatten Aggref expressions, since those are to be computed below the
6497 : * window functions and just referenced like Vars above that.
6498 : *
6499 : * 'final_target' is the query's final target list (in PathTarget form)
6500 : * 'activeWindows' is the list of active windows previously identified by
6501 : * select_active_windows.
6502 : *
6503 : * The result is the PathTarget to be computed by the plan node immediately
6504 : * below the first WindowAgg node.
6505 : */
6506 : static PathTarget *
6507 2232 : make_window_input_target(PlannerInfo *root,
6508 : PathTarget *final_target,
6509 : List *activeWindows)
6510 : {
6511 : PathTarget *input_target;
6512 : Bitmapset *sgrefs;
6513 : List *flattenable_cols;
6514 : List *flattenable_vars;
6515 : int i;
6516 : ListCell *lc;
6517 :
6518 : Assert(root->parse->hasWindowFuncs);
6519 :
6520 : /*
6521 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6522 : * into a bitmapset for convenient reference below.
6523 : */
6524 2232 : sgrefs = NULL;
6525 4619 : foreach(lc, activeWindows)
6526 : {
6527 2387 : WindowClause *wc = lfirst_node(WindowClause, lc);
6528 : ListCell *lc2;
6529 :
6530 3009 : foreach(lc2, wc->partitionClause)
6531 : {
6532 622 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6533 :
6534 622 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6535 : }
6536 4307 : foreach(lc2, wc->orderClause)
6537 : {
6538 1920 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6539 :
6540 1920 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6541 : }
6542 : }
6543 :
6544 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6545 2388 : foreach(lc, root->processed_groupClause)
6546 : {
6547 156 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6548 :
6549 156 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6550 : }
6551 :
6552 : /*
6553 : * Construct a target containing all the non-flattenable targetlist items,
6554 : * and save aside the others for a moment.
6555 : */
6556 2232 : input_target = create_empty_pathtarget();
6557 2232 : flattenable_cols = NIL;
6558 :
6559 2232 : i = 0;
6560 9369 : foreach(lc, final_target->exprs)
6561 : {
6562 7137 : Expr *expr = (Expr *) lfirst(lc);
6563 7137 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6564 :
6565 : /*
6566 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6567 : * that such items can't contain window functions, so it's okay to
6568 : * compute them below the WindowAgg nodes.)
6569 : */
6570 7137 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6571 : {
6572 : /*
6573 : * Don't want to deconstruct this value, so add it to the input
6574 : * target as-is.
6575 : */
6576 2404 : add_column_to_pathtarget(input_target, expr, sgref);
6577 : }
6578 : else
6579 : {
6580 : /*
6581 : * Column is to be flattened, so just remember the expression for
6582 : * later call to pull_var_clause.
6583 : */
6584 4733 : flattenable_cols = lappend(flattenable_cols, expr);
6585 : }
6586 :
6587 7137 : i++;
6588 : }
6589 :
6590 : /*
6591 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6592 : * add them to the input target if not already present. (Some might be
6593 : * there already because they're used directly as window/group clauses.)
6594 : *
6595 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6596 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6597 : * at higher levels. On the other hand, we should recurse into
6598 : * WindowFuncs to make sure their input expressions are available.
6599 : */
6600 2232 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6601 : PVC_INCLUDE_AGGREGATES |
6602 : PVC_RECURSE_WINDOWFUNCS |
6603 : PVC_INCLUDE_PLACEHOLDERS);
6604 2232 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6605 :
6606 : /* clean up cruft */
6607 2232 : list_free(flattenable_vars);
6608 2232 : list_free(flattenable_cols);
6609 :
6610 : /* XXX this causes some redundant cost calculation ... */
6611 2232 : return set_pathtarget_cost_width(root, input_target);
6612 : }
6613 :
6614 : /*
6615 : * make_pathkeys_for_window
6616 : * Create a pathkeys list describing the required input ordering
6617 : * for the given WindowClause.
6618 : *
6619 : * Modifies wc's partitionClause to remove any clauses which are deemed
6620 : * redundant by the pathkey logic.
6621 : *
6622 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6623 : * In the future we might try to implement windowing using hashing, in which
6624 : * case the ordering could be relaxed, but for now we always sort.
6625 : */
6626 : static List *
6627 4786 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6628 : List *tlist)
6629 : {
6630 4786 : List *window_pathkeys = NIL;
6631 :
6632 : /* Throw error if can't sort */
6633 4786 : if (!grouping_is_sortable(wc->partitionClause))
6634 0 : ereport(ERROR,
6635 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6636 : errmsg("could not implement window PARTITION BY"),
6637 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6638 4786 : if (!grouping_is_sortable(wc->orderClause))
6639 0 : ereport(ERROR,
6640 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6641 : errmsg("could not implement window ORDER BY"),
6642 : errdetail("Window ordering columns must be of sortable datatypes.")));
6643 :
6644 : /*
6645 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6646 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6647 : */
6648 4786 : if (wc->partitionClause != NIL)
6649 : {
6650 : bool sortable;
6651 :
6652 1088 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6653 : &wc->partitionClause,
6654 : tlist,
6655 : true,
6656 : false,
6657 : &sortable,
6658 : false);
6659 :
6660 : Assert(sortable);
6661 : }
6662 :
6663 : /*
6664 : * In principle, we could also consider removing redundant ORDER BY items
6665 : * too as doing so does not alter the result of peer row checks done by
6666 : * the executor. However, we must *not* remove the ordering column for
6667 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6668 : * if it's known to be equal to some partitioning column.
6669 : */
6670 4786 : if (wc->orderClause != NIL)
6671 : {
6672 : List *orderby_pathkeys;
6673 :
6674 3754 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6675 : wc->orderClause,
6676 : tlist);
6677 :
6678 : /* Okay, make the combined pathkeys */
6679 3754 : if (window_pathkeys != NIL)
6680 781 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6681 : else
6682 2973 : window_pathkeys = orderby_pathkeys;
6683 : }
6684 :
6685 4786 : return window_pathkeys;
6686 : }
6687 :
6688 : /*
6689 : * make_sort_input_target
6690 : * Generate appropriate PathTarget for initial input to Sort step.
6691 : *
6692 : * If the query has ORDER BY, this function chooses the target to be computed
6693 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6694 : * project) steps. This might or might not be identical to the query's final
6695 : * output target.
6696 : *
6697 : * The main argument for keeping the sort-input tlist the same as the final
6698 : * is that we avoid a separate projection node (which will be needed if
6699 : * they're different, because Sort can't project). However, there are also
6700 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6701 : * a consistent order of evaluation for any volatile functions in the tlist,
6702 : * and if there's also a LIMIT, we can stop the query without ever computing
6703 : * tlist functions for later rows, which is beneficial for both volatile and
6704 : * expensive functions.
6705 : *
6706 : * Our current policy is to postpone volatile expressions till after the sort
6707 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6708 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6709 : * postpone set-returning expressions, because running them beforehand would
6710 : * bloat the sort dataset, and because it might cause unexpected output order
6711 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6712 : * in the tlist should be evaluated at the same plan step, so that they can
6713 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6714 : * mustn't postpone any SRFs. (Note that in principle that policy should
6715 : * probably get applied to the group/window input targetlists too, but we
6716 : * have not done that historically.) Lastly, expensive expressions are
6717 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6718 : * partial evaluation of the query is possible (if neither is true, we expect
6719 : * to have to evaluate the expressions for every row anyway), or if there are
6720 : * any volatile or set-returning expressions (since once we've put in a
6721 : * projection at all, it won't cost any more to postpone more stuff).
6722 : *
6723 : * Another issue that could potentially be considered here is that
6724 : * evaluating tlist expressions could result in data that's either wider
6725 : * or narrower than the input Vars, thus changing the volume of data that
6726 : * has to go through the Sort. However, we usually have only a very bad
6727 : * idea of the output width of any expression more complex than a Var,
6728 : * so for now it seems too risky to try to optimize on that basis.
6729 : *
6730 : * Note that if we do produce a modified sort-input target, and then the
6731 : * query ends up not using an explicit Sort, no particular harm is done:
6732 : * we'll initially use the modified target for the preceding path nodes,
6733 : * but then change them to the final target with apply_projection_to_path.
6734 : * Moreover, in such a case the guarantees about evaluation order of
6735 : * volatile functions still hold, since the rows are sorted already.
6736 : *
6737 : * This function has some things in common with make_group_input_target and
6738 : * make_window_input_target, though the detailed rules for what to do are
6739 : * different. We never flatten/postpone any grouping or ordering columns;
6740 : * those are needed before the sort. If we do flatten a particular
6741 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6742 : * computed earlier.
6743 : *
6744 : * 'final_target' is the query's final target list (in PathTarget form)
6745 : * 'have_postponed_srfs' is an output argument, see below
6746 : *
6747 : * The result is the PathTarget to be computed by the plan node immediately
6748 : * below the Sort step (and the Distinct step, if any). This will be
6749 : * exactly final_target if we decide a projection step wouldn't be helpful.
6750 : *
6751 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6752 : * any set-returning functions to after the Sort.
6753 : */
6754 : static PathTarget *
6755 57681 : make_sort_input_target(PlannerInfo *root,
6756 : PathTarget *final_target,
6757 : bool *have_postponed_srfs)
6758 : {
6759 57681 : Query *parse = root->parse;
6760 : PathTarget *input_target;
6761 : int ncols;
6762 : bool *col_is_srf;
6763 : bool *postpone_col;
6764 : bool have_srf;
6765 : bool have_volatile;
6766 : bool have_expensive;
6767 : bool have_srf_sortcols;
6768 : bool postpone_srfs;
6769 : List *postponable_cols;
6770 : List *postponable_vars;
6771 : int i;
6772 : ListCell *lc;
6773 :
6774 : /* Shouldn't get here unless query has ORDER BY */
6775 : Assert(parse->sortClause);
6776 :
6777 57681 : *have_postponed_srfs = false; /* default result */
6778 :
6779 : /* Inspect tlist and collect per-column information */
6780 57681 : ncols = list_length(final_target->exprs);
6781 57681 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6782 57681 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6783 57681 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6784 :
6785 57681 : i = 0;
6786 329219 : foreach(lc, final_target->exprs)
6787 : {
6788 271538 : Expr *expr = (Expr *) lfirst(lc);
6789 :
6790 : /*
6791 : * If the column has a sortgroupref, assume it has to be evaluated
6792 : * before sorting. Generally such columns would be ORDER BY, GROUP
6793 : * BY, etc targets. One exception is columns that were removed from
6794 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6795 : * only be Vars anyway. There don't seem to be any cases where it
6796 : * would be worth the trouble to double-check.
6797 : */
6798 271538 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6799 : {
6800 : /*
6801 : * Check for SRF or volatile functions. Check the SRF case first
6802 : * because we must know whether we have any postponed SRFs.
6803 : */
6804 190347 : if (parse->hasTargetSRFs &&
6805 180 : expression_returns_set((Node *) expr))
6806 : {
6807 : /* We'll decide below whether these are postponable */
6808 80 : col_is_srf[i] = true;
6809 80 : have_srf = true;
6810 : }
6811 190087 : else if (contain_volatile_functions((Node *) expr))
6812 : {
6813 : /* Unconditionally postpone */
6814 190 : postpone_col[i] = true;
6815 190 : have_volatile = true;
6816 : }
6817 : else
6818 : {
6819 : /*
6820 : * Else check the cost. XXX it's annoying to have to do this
6821 : * when set_pathtarget_cost_width() just did it. Refactor to
6822 : * allow sharing the work?
6823 : */
6824 : QualCost cost;
6825 :
6826 189897 : cost_qual_eval_node(&cost, (Node *) expr, root);
6827 :
6828 : /*
6829 : * We arbitrarily define "expensive" as "more than 10X
6830 : * cpu_operator_cost". Note this will take in any PL function
6831 : * with default cost.
6832 : */
6833 189897 : if (cost.per_tuple > 10 * cpu_operator_cost)
6834 : {
6835 13923 : postpone_col[i] = true;
6836 13923 : have_expensive = true;
6837 : }
6838 : }
6839 : }
6840 : else
6841 : {
6842 : /* For sortgroupref cols, just check if any contain SRFs */
6843 81371 : if (!have_srf_sortcols &&
6844 81628 : parse->hasTargetSRFs &&
6845 277 : expression_returns_set((Node *) expr))
6846 122 : have_srf_sortcols = true;
6847 : }
6848 :
6849 271538 : i++;
6850 : }
6851 :
6852 : /*
6853 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6854 : */
6855 57681 : postpone_srfs = (have_srf && !have_srf_sortcols);
6856 :
6857 : /*
6858 : * If we don't need a post-sort projection, just return final_target.
6859 : */
6860 57681 : if (!(postpone_srfs || have_volatile ||
6861 57443 : (have_expensive &&
6862 8382 : (parse->limitCount || root->tuple_fraction > 0))))
6863 57413 : return final_target;
6864 :
6865 : /*
6866 : * Report whether the post-sort projection will contain set-returning
6867 : * functions. This is important because it affects whether the Sort can
6868 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6869 : * to return.
6870 : */
6871 268 : *have_postponed_srfs = postpone_srfs;
6872 :
6873 : /*
6874 : * Construct the sort-input target, taking all non-postponable columns and
6875 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6876 : * the postponable ones.
6877 : */
6878 268 : input_target = create_empty_pathtarget();
6879 268 : postponable_cols = NIL;
6880 :
6881 268 : i = 0;
6882 1875 : foreach(lc, final_target->exprs)
6883 : {
6884 1607 : Expr *expr = (Expr *) lfirst(lc);
6885 :
6886 1607 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6887 320 : postponable_cols = lappend(postponable_cols, expr);
6888 : else
6889 1287 : add_column_to_pathtarget(input_target, expr,
6890 1287 : get_pathtarget_sortgroupref(final_target, i));
6891 :
6892 1607 : i++;
6893 : }
6894 :
6895 : /*
6896 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6897 : * postponable columns, and add them to the sort-input target if not
6898 : * already present. (Some might be there already.) We mustn't
6899 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6900 : * would be unable to recompute them.
6901 : */
6902 268 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6903 : PVC_INCLUDE_AGGREGATES |
6904 : PVC_INCLUDE_WINDOWFUNCS |
6905 : PVC_INCLUDE_PLACEHOLDERS);
6906 268 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6907 :
6908 : /* clean up cruft */
6909 268 : list_free(postponable_vars);
6910 268 : list_free(postponable_cols);
6911 :
6912 : /* XXX this represents even more redundant cost calculation ... */
6913 268 : return set_pathtarget_cost_width(root, input_target);
6914 : }
6915 :
6916 : /*
6917 : * get_cheapest_fractional_path
6918 : * Find the cheapest path for retrieving a specified fraction of all
6919 : * the tuples expected to be returned by the given relation.
6920 : *
6921 : * Do not consider parameterized paths. If the caller needs a path for upper
6922 : * rel, it can't have parameterized paths. If the caller needs an append
6923 : * subpath, it could become limited by the treatment of similar
6924 : * parameterization of all the subpaths.
6925 : *
6926 : * We interpret tuple_fraction the same way as grouping_planner.
6927 : *
6928 : * We assume set_cheapest() has been run on the given rel.
6929 : */
6930 : Path *
6931 361242 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6932 : {
6933 361242 : Path *best_path = rel->cheapest_total_path;
6934 : ListCell *l;
6935 :
6936 : /* If all tuples will be retrieved, just return the cheapest-total path */
6937 361242 : if (tuple_fraction <= 0.0)
6938 354305 : return best_path;
6939 :
6940 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6941 6937 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6942 2933 : tuple_fraction /= best_path->rows;
6943 :
6944 17742 : foreach(l, rel->pathlist)
6945 : {
6946 10805 : Path *path = (Path *) lfirst(l);
6947 :
6948 10805 : if (path->param_info)
6949 166 : continue;
6950 :
6951 14341 : if (path == rel->cheapest_total_path ||
6952 3702 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6953 10317 : continue;
6954 :
6955 322 : best_path = path;
6956 : }
6957 :
6958 6937 : return best_path;
6959 : }
6960 :
6961 : /*
6962 : * adjust_paths_for_srfs
6963 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6964 : *
6965 : * The executor can only handle set-returning functions that appear at the
6966 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6967 : * that are not at top level, we need to split up the evaluation into multiple
6968 : * plan levels in which each level satisfies this constraint. This function
6969 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6970 : * output tlist to insert appropriate projection steps.
6971 : *
6972 : * The given targets and targets_contain_srfs lists are from
6973 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6974 : * target in targets.
6975 : */
6976 : static void
6977 10665 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6978 : List *targets, List *targets_contain_srfs)
6979 : {
6980 : ListCell *lc;
6981 :
6982 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6983 : Assert(!linitial_int(targets_contain_srfs));
6984 :
6985 : /* If no SRFs appear at this plan level, nothing to do */
6986 10665 : if (list_length(targets) == 1)
6987 491 : return;
6988 :
6989 : /*
6990 : * Stack SRF-evaluation nodes atop each path for the rel.
6991 : *
6992 : * In principle we should re-run set_cheapest() here to identify the
6993 : * cheapest path, but it seems unlikely that adding the same tlist eval
6994 : * costs to all the paths would change that, so we don't bother. Instead,
6995 : * just assume that the cheapest-startup and cheapest-total paths remain
6996 : * so. (There should be no parameterized paths anymore, so we needn't
6997 : * worry about updating cheapest_parameterized_paths.)
6998 : */
6999 20380 : foreach(lc, rel->pathlist)
7000 : {
7001 10206 : Path *subpath = (Path *) lfirst(lc);
7002 10206 : Path *newpath = subpath;
7003 : ListCell *lc1,
7004 : *lc2;
7005 :
7006 : Assert(subpath->param_info == NULL);
7007 31405 : forboth(lc1, targets, lc2, targets_contain_srfs)
7008 : {
7009 21199 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
7010 21199 : bool contains_srfs = (bool) lfirst_int(lc2);
7011 :
7012 : /* If this level doesn't contain SRFs, do regular projection */
7013 21199 : if (contains_srfs)
7014 10256 : newpath = (Path *) create_set_projection_path(root,
7015 : rel,
7016 : newpath,
7017 : thistarget);
7018 : else
7019 10943 : newpath = (Path *) apply_projection_to_path(root,
7020 : rel,
7021 : newpath,
7022 : thistarget);
7023 : }
7024 10206 : lfirst(lc) = newpath;
7025 10206 : if (subpath == rel->cheapest_startup_path)
7026 239 : rel->cheapest_startup_path = newpath;
7027 10206 : if (subpath == rel->cheapest_total_path)
7028 239 : rel->cheapest_total_path = newpath;
7029 : }
7030 :
7031 : /* Likewise for partial paths, if any */
7032 10189 : foreach(lc, rel->partial_pathlist)
7033 : {
7034 15 : Path *subpath = (Path *) lfirst(lc);
7035 15 : Path *newpath = subpath;
7036 : ListCell *lc1,
7037 : *lc2;
7038 :
7039 : Assert(subpath->param_info == NULL);
7040 60 : forboth(lc1, targets, lc2, targets_contain_srfs)
7041 : {
7042 45 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
7043 45 : bool contains_srfs = (bool) lfirst_int(lc2);
7044 :
7045 : /* If this level doesn't contain SRFs, do regular projection */
7046 45 : if (contains_srfs)
7047 15 : newpath = (Path *) create_set_projection_path(root,
7048 : rel,
7049 : newpath,
7050 : thistarget);
7051 : else
7052 : {
7053 : /* avoid apply_projection_to_path, in case of multiple refs */
7054 30 : newpath = (Path *) create_projection_path(root,
7055 : rel,
7056 : newpath,
7057 : thistarget);
7058 : }
7059 : }
7060 15 : lfirst(lc) = newpath;
7061 : }
7062 : }
7063 :
7064 : /*
7065 : * expression_planner
7066 : * Perform planner's transformations on a standalone expression.
7067 : *
7068 : * Various utility commands need to evaluate expressions that are not part
7069 : * of a plannable query. They can do so using the executor's regular
7070 : * expression-execution machinery, but first the expression has to be fed
7071 : * through here to transform it from parser output to something executable.
7072 : *
7073 : * Currently, we disallow sublinks in standalone expressions, so there's no
7074 : * real "planning" involved here. (That might not always be true though.)
7075 : * What we must do is run eval_const_expressions to ensure that any function
7076 : * calls are converted to positional notation and function default arguments
7077 : * get inserted. The fact that constant subexpressions get simplified is a
7078 : * side-effect that is useful when the expression will get evaluated more than
7079 : * once. Also, we must fix operator function IDs.
7080 : *
7081 : * This does not return any information about dependencies of the expression.
7082 : * Hence callers should use the results only for the duration of the current
7083 : * query. Callers that would like to cache the results for longer should use
7084 : * expression_planner_with_deps, probably via the plancache.
7085 : *
7086 : * Note: this must not make any damaging changes to the passed-in expression
7087 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
7088 : * we first do an expression_tree_mutator-based walk, what is returned will
7089 : * be a new node tree.) The result is constructed in the current memory
7090 : * context; beware that this can leak a lot of additional stuff there, too.
7091 : */
7092 : Expr *
7093 138071 : expression_planner(Expr *expr)
7094 : {
7095 : Node *result;
7096 :
7097 : /*
7098 : * Convert named-argument function calls, insert default arguments and
7099 : * simplify constant subexprs
7100 : */
7101 138071 : result = eval_const_expressions(NULL, (Node *) expr);
7102 :
7103 : /* Fill in opfuncid values if missing */
7104 138059 : fix_opfuncids(result);
7105 :
7106 138059 : return (Expr *) result;
7107 : }
7108 :
7109 : /*
7110 : * expression_planner_with_deps
7111 : * Perform planner's transformations on a standalone expression,
7112 : * returning expression dependency information along with the result.
7113 : *
7114 : * This is identical to expression_planner() except that it also returns
7115 : * information about possible dependencies of the expression, ie identities of
7116 : * objects whose definitions affect the result. As in a PlannedStmt, these
7117 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
7118 : */
7119 : Expr *
7120 265 : expression_planner_with_deps(Expr *expr,
7121 : List **relationOids,
7122 : List **invalItems)
7123 : {
7124 : Node *result;
7125 : PlannerGlobal glob;
7126 : PlannerInfo root;
7127 :
7128 : /* Make up dummy planner state so we can use setrefs machinery */
7129 7685 : MemSet(&glob, 0, sizeof(glob));
7130 265 : glob.type = T_PlannerGlobal;
7131 265 : glob.relationOids = NIL;
7132 265 : glob.invalItems = NIL;
7133 :
7134 24910 : MemSet(&root, 0, sizeof(root));
7135 265 : root.type = T_PlannerInfo;
7136 265 : root.glob = &glob;
7137 :
7138 : /*
7139 : * Convert named-argument function calls, insert default arguments and
7140 : * simplify constant subexprs. Collect identities of inlined functions
7141 : * and elided domains, too.
7142 : */
7143 265 : result = eval_const_expressions(&root, (Node *) expr);
7144 :
7145 : /* Fill in opfuncid values if missing */
7146 265 : fix_opfuncids(result);
7147 :
7148 : /*
7149 : * Now walk the finished expression to find anything else we ought to
7150 : * record as an expression dependency.
7151 : */
7152 265 : (void) extract_query_dependencies_walker(result, &root);
7153 :
7154 265 : *relationOids = glob.relationOids;
7155 265 : *invalItems = glob.invalItems;
7156 :
7157 265 : return (Expr *) result;
7158 : }
7159 :
7160 :
7161 : /*
7162 : * plan_cluster_use_sort
7163 : * Use the planner to decide how CLUSTER should implement sorting
7164 : *
7165 : * tableOid is the OID of a table to be clustered on its index indexOid
7166 : * (which is already known to be a btree index). Decide whether it's
7167 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
7168 : * Return true to use sorting, false to use an indexscan.
7169 : *
7170 : * Note: caller had better already hold some type of lock on the table.
7171 : */
7172 : bool
7173 146 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
7174 : {
7175 : PlannerInfo *root;
7176 : Query *query;
7177 : PlannerGlobal *glob;
7178 : RangeTblEntry *rte;
7179 : RelOptInfo *rel;
7180 : IndexOptInfo *indexInfo;
7181 : QualCost indexExprCost;
7182 : Cost comparisonCost;
7183 : Path *seqScanPath;
7184 : Path seqScanAndSortPath;
7185 : IndexPath *indexScanPath;
7186 : ListCell *lc;
7187 :
7188 : /* We can short-circuit the cost comparison if indexscans are disabled */
7189 146 : if (!enable_indexscan)
7190 20 : return true; /* use sort */
7191 :
7192 : /* Set up mostly-dummy planner state */
7193 126 : query = makeNode(Query);
7194 126 : query->commandType = CMD_SELECT;
7195 :
7196 126 : glob = makeNode(PlannerGlobal);
7197 :
7198 126 : root = makeNode(PlannerInfo);
7199 126 : root->parse = query;
7200 126 : root->glob = glob;
7201 126 : root->query_level = 1;
7202 126 : root->planner_cxt = CurrentMemoryContext;
7203 126 : root->wt_param_id = -1;
7204 126 : root->join_domains = list_make1(makeNode(JoinDomain));
7205 :
7206 : /* Build a minimal RTE for the rel */
7207 126 : rte = makeNode(RangeTblEntry);
7208 126 : rte->rtekind = RTE_RELATION;
7209 126 : rte->relid = tableOid;
7210 126 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7211 126 : rte->rellockmode = AccessShareLock;
7212 126 : rte->lateral = false;
7213 126 : rte->inh = false;
7214 126 : rte->inFromCl = true;
7215 126 : query->rtable = list_make1(rte);
7216 126 : addRTEPermissionInfo(&query->rteperminfos, rte);
7217 :
7218 : /* Set up RTE/RelOptInfo arrays */
7219 126 : setup_simple_rel_arrays(root);
7220 :
7221 : /* Build RelOptInfo */
7222 126 : rel = build_simple_rel(root, 1, NULL);
7223 :
7224 : /* Locate IndexOptInfo for the target index */
7225 126 : indexInfo = NULL;
7226 156 : foreach(lc, rel->indexlist)
7227 : {
7228 156 : indexInfo = lfirst_node(IndexOptInfo, lc);
7229 156 : if (indexInfo->indexoid == indexOid)
7230 126 : break;
7231 : }
7232 :
7233 : /*
7234 : * It's possible that get_relation_info did not generate an IndexOptInfo
7235 : * for the desired index; this could happen if it's not yet reached its
7236 : * indcheckxmin usability horizon, or if it's a system index and we're
7237 : * ignoring system indexes. In such cases we should tell CLUSTER to not
7238 : * trust the index contents but use seqscan-and-sort.
7239 : */
7240 126 : if (lc == NULL) /* not in the list? */
7241 0 : return true; /* use sort */
7242 :
7243 : /*
7244 : * Rather than doing all the pushups that would be needed to use
7245 : * set_baserel_size_estimates, just do a quick hack for rows and width.
7246 : */
7247 126 : rel->rows = rel->tuples;
7248 126 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
7249 :
7250 126 : root->total_table_pages = rel->pages;
7251 :
7252 : /*
7253 : * Determine eval cost of the index expressions, if any. We need to
7254 : * charge twice that amount for each tuple comparison that happens during
7255 : * the sort, since tuplesort.c will have to re-evaluate the index
7256 : * expressions each time. (XXX that's pretty inefficient...)
7257 : */
7258 126 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
7259 126 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
7260 :
7261 : /* Estimate the cost of seq scan + sort */
7262 126 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
7263 126 : cost_sort(&seqScanAndSortPath, root, NIL,
7264 : seqScanPath->disabled_nodes,
7265 126 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
7266 : comparisonCost, maintenance_work_mem, -1.0);
7267 :
7268 : /* Estimate the cost of index scan */
7269 126 : indexScanPath = create_index_path(root, indexInfo,
7270 : NIL, NIL, NIL, NIL,
7271 : ForwardScanDirection, false,
7272 : NULL, 1.0, false);
7273 :
7274 126 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
7275 : }
7276 :
7277 : /*
7278 : * plan_create_index_workers
7279 : * Use the planner to decide how many parallel worker processes
7280 : * CREATE INDEX should request for use
7281 : *
7282 : * tableOid is the table on which the index is to be built. indexOid is the
7283 : * OID of an index to be created or reindexed (which must be an index with
7284 : * support for parallel builds - currently btree, GIN, or BRIN).
7285 : *
7286 : * Return value is the number of parallel worker processes to request. It
7287 : * may be unsafe to proceed if this is 0. Note that this does not include the
7288 : * leader participating as a worker (value is always a number of parallel
7289 : * worker processes).
7290 : *
7291 : * Note: caller had better already hold some type of lock on the table and
7292 : * index.
7293 : */
7294 : int
7295 23248 : plan_create_index_workers(Oid tableOid, Oid indexOid)
7296 : {
7297 : PlannerInfo *root;
7298 : Query *query;
7299 : PlannerGlobal *glob;
7300 : RangeTblEntry *rte;
7301 : Relation heap;
7302 : Relation index;
7303 : RelOptInfo *rel;
7304 : int parallel_workers;
7305 : BlockNumber heap_blocks;
7306 : double reltuples;
7307 : double allvisfrac;
7308 :
7309 : /*
7310 : * We don't allow performing parallel operation in standalone backend or
7311 : * when parallelism is disabled.
7312 : */
7313 23248 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
7314 301 : return 0;
7315 :
7316 : /* Set up largely-dummy planner state */
7317 22947 : query = makeNode(Query);
7318 22947 : query->commandType = CMD_SELECT;
7319 :
7320 22947 : glob = makeNode(PlannerGlobal);
7321 :
7322 22947 : root = makeNode(PlannerInfo);
7323 22947 : root->parse = query;
7324 22947 : root->glob = glob;
7325 22947 : root->query_level = 1;
7326 22947 : root->planner_cxt = CurrentMemoryContext;
7327 22947 : root->wt_param_id = -1;
7328 22947 : root->join_domains = list_make1(makeNode(JoinDomain));
7329 :
7330 : /*
7331 : * Build a minimal RTE.
7332 : *
7333 : * Mark the RTE with inh = true. This is a kludge to prevent
7334 : * get_relation_info() from fetching index info, which is necessary
7335 : * because it does not expect that any IndexOptInfo is currently
7336 : * undergoing REINDEX.
7337 : */
7338 22947 : rte = makeNode(RangeTblEntry);
7339 22947 : rte->rtekind = RTE_RELATION;
7340 22947 : rte->relid = tableOid;
7341 22947 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7342 22947 : rte->rellockmode = AccessShareLock;
7343 22947 : rte->lateral = false;
7344 22947 : rte->inh = true;
7345 22947 : rte->inFromCl = true;
7346 22947 : query->rtable = list_make1(rte);
7347 22947 : addRTEPermissionInfo(&query->rteperminfos, rte);
7348 :
7349 : /* Set up RTE/RelOptInfo arrays */
7350 22947 : setup_simple_rel_arrays(root);
7351 :
7352 : /* Build RelOptInfo */
7353 22947 : rel = build_simple_rel(root, 1, NULL);
7354 :
7355 : /* Rels are assumed already locked by the caller */
7356 22947 : heap = table_open(tableOid, NoLock);
7357 22947 : index = index_open(indexOid, NoLock);
7358 :
7359 : /*
7360 : * Determine if it's safe to proceed.
7361 : *
7362 : * Currently, parallel workers can't access the leader's temporary tables.
7363 : * Furthermore, any index predicate or index expressions must be parallel
7364 : * safe.
7365 : */
7366 22947 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7367 21556 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7368 21466 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7369 : {
7370 1481 : parallel_workers = 0;
7371 1481 : goto done;
7372 : }
7373 :
7374 : /*
7375 : * If parallel_workers storage parameter is set for the table, accept that
7376 : * as the number of parallel worker processes to launch (though still cap
7377 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7378 : * consider any other factor when parallel_workers is set. (e.g., memory
7379 : * use by workers.)
7380 : */
7381 21466 : if (rel->rel_parallel_workers != -1)
7382 : {
7383 59 : parallel_workers = Min(rel->rel_parallel_workers,
7384 : max_parallel_maintenance_workers);
7385 59 : goto done;
7386 : }
7387 :
7388 : /*
7389 : * Estimate heap relation size ourselves, since rel->pages cannot be
7390 : * trusted (heap RTE was marked as inheritance parent)
7391 : */
7392 21407 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7393 :
7394 : /*
7395 : * Determine number of workers to scan the heap relation using generic
7396 : * model
7397 : */
7398 21407 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7399 : max_parallel_maintenance_workers);
7400 :
7401 : /*
7402 : * Cap workers based on available maintenance_work_mem as needed.
7403 : *
7404 : * Note that each tuplesort participant receives an even share of the
7405 : * total maintenance_work_mem budget. Aim to leave participants
7406 : * (including the leader as a participant) with no less than 32MB of
7407 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7408 : * immediately past the threshold of being capable of launching a single
7409 : * parallel worker to sort.
7410 : */
7411 21516 : while (parallel_workers > 0 &&
7412 219 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7413 109 : parallel_workers--;
7414 :
7415 21407 : done:
7416 22947 : index_close(index, NoLock);
7417 22947 : table_close(heap, NoLock);
7418 :
7419 22947 : return parallel_workers;
7420 : }
7421 :
7422 : /*
7423 : * add_paths_to_grouping_rel
7424 : *
7425 : * Add non-partial paths to grouping relation.
7426 : */
7427 : static void
7428 34979 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7429 : RelOptInfo *grouped_rel,
7430 : RelOptInfo *partially_grouped_rel,
7431 : const AggClauseCosts *agg_costs,
7432 : grouping_sets_data *gd,
7433 : GroupPathExtraData *extra)
7434 : {
7435 34979 : Query *parse = root->parse;
7436 34979 : Path *cheapest_path = input_rel->cheapest_total_path;
7437 34979 : Path *cheapest_partially_grouped_path = NULL;
7438 : ListCell *lc;
7439 34979 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7440 34979 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7441 34979 : List *havingQual = (List *) extra->havingQual;
7442 34979 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7443 34979 : double dNumGroups = 0;
7444 34979 : double dNumFinalGroups = 0;
7445 :
7446 : /*
7447 : * Estimate number of groups for non-split aggregation.
7448 : */
7449 34979 : dNumGroups = get_number_of_groups(root,
7450 : cheapest_path->rows,
7451 : gd,
7452 : extra->targetList);
7453 :
7454 34979 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7455 : {
7456 2503 : cheapest_partially_grouped_path =
7457 : partially_grouped_rel->cheapest_total_path;
7458 :
7459 : /*
7460 : * Estimate number of groups for final phase of partial aggregation.
7461 : */
7462 : dNumFinalGroups =
7463 2503 : get_number_of_groups(root,
7464 : cheapest_partially_grouped_path->rows,
7465 : gd,
7466 : extra->targetList);
7467 : }
7468 :
7469 34979 : if (can_sort)
7470 : {
7471 : /*
7472 : * Use any available suitably-sorted path as input, and also consider
7473 : * sorting the cheapest-total path and incremental sort on any paths
7474 : * with presorted keys.
7475 : */
7476 72541 : foreach(lc, input_rel->pathlist)
7477 : {
7478 : ListCell *lc2;
7479 37567 : Path *path = (Path *) lfirst(lc);
7480 37567 : Path *path_save = path;
7481 37567 : List *pathkey_orderings = NIL;
7482 :
7483 : /* generate alternative group orderings that might be useful */
7484 37567 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7485 :
7486 : Assert(list_length(pathkey_orderings) > 0);
7487 :
7488 75251 : foreach(lc2, pathkey_orderings)
7489 : {
7490 37684 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7491 :
7492 : /* restore the path (we replace it in the loop) */
7493 37684 : path = path_save;
7494 :
7495 37684 : path = make_ordered_path(root,
7496 : grouped_rel,
7497 : path,
7498 : cheapest_path,
7499 : info->pathkeys,
7500 : -1.0);
7501 37684 : if (path == NULL)
7502 306 : continue;
7503 :
7504 : /* Now decide what to stick atop it */
7505 37378 : if (parse->groupingSets)
7506 : {
7507 975 : consider_groupingsets_paths(root, grouped_rel,
7508 : path, true, can_hash,
7509 : gd, agg_costs, dNumGroups);
7510 : }
7511 36403 : else if (parse->hasAggs)
7512 : {
7513 : /*
7514 : * We have aggregation, possibly with plain GROUP BY. Make
7515 : * an AggPath.
7516 : */
7517 35711 : add_path(grouped_rel, (Path *)
7518 35711 : create_agg_path(root,
7519 : grouped_rel,
7520 : path,
7521 35711 : grouped_rel->reltarget,
7522 35711 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7523 : AGGSPLIT_SIMPLE,
7524 : info->clauses,
7525 : havingQual,
7526 : agg_costs,
7527 : dNumGroups));
7528 : }
7529 692 : else if (parse->groupClause)
7530 : {
7531 : /*
7532 : * We have GROUP BY without aggregation or grouping sets.
7533 : * Make a GroupPath.
7534 : */
7535 692 : add_path(grouped_rel, (Path *)
7536 692 : create_group_path(root,
7537 : grouped_rel,
7538 : path,
7539 : info->clauses,
7540 : havingQual,
7541 : dNumGroups));
7542 : }
7543 : else
7544 : {
7545 : /* Other cases should have been handled above */
7546 : Assert(false);
7547 : }
7548 : }
7549 : }
7550 :
7551 : /*
7552 : * Instead of operating directly on the input relation, we can
7553 : * consider finalizing a partially aggregated path.
7554 : */
7555 34974 : if (partially_grouped_rel != NULL)
7556 : {
7557 6452 : foreach(lc, partially_grouped_rel->pathlist)
7558 : {
7559 : ListCell *lc2;
7560 3949 : Path *path = (Path *) lfirst(lc);
7561 3949 : Path *path_save = path;
7562 3949 : List *pathkey_orderings = NIL;
7563 :
7564 : /* generate alternative group orderings that might be useful */
7565 3949 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7566 :
7567 : Assert(list_length(pathkey_orderings) > 0);
7568 :
7569 : /* process all potentially interesting grouping reorderings */
7570 7898 : foreach(lc2, pathkey_orderings)
7571 : {
7572 3949 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7573 :
7574 : /* restore the path (we replace it in the loop) */
7575 3949 : path = path_save;
7576 :
7577 3949 : path = make_ordered_path(root,
7578 : grouped_rel,
7579 : path,
7580 : cheapest_partially_grouped_path,
7581 : info->pathkeys,
7582 : -1.0);
7583 :
7584 3949 : if (path == NULL)
7585 172 : continue;
7586 :
7587 3777 : if (parse->hasAggs)
7588 3575 : add_path(grouped_rel, (Path *)
7589 3575 : create_agg_path(root,
7590 : grouped_rel,
7591 : path,
7592 3575 : grouped_rel->reltarget,
7593 3575 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7594 : AGGSPLIT_FINAL_DESERIAL,
7595 : info->clauses,
7596 : havingQual,
7597 : agg_final_costs,
7598 : dNumFinalGroups));
7599 : else
7600 202 : add_path(grouped_rel, (Path *)
7601 202 : create_group_path(root,
7602 : grouped_rel,
7603 : path,
7604 : info->clauses,
7605 : havingQual,
7606 : dNumFinalGroups));
7607 :
7608 : }
7609 : }
7610 : }
7611 : }
7612 :
7613 34979 : if (can_hash)
7614 : {
7615 5146 : if (parse->groupingSets)
7616 : {
7617 : /*
7618 : * Try for a hash-only groupingsets path over unsorted input.
7619 : */
7620 807 : consider_groupingsets_paths(root, grouped_rel,
7621 : cheapest_path, false, true,
7622 : gd, agg_costs, dNumGroups);
7623 : }
7624 : else
7625 : {
7626 : /*
7627 : * Generate a HashAgg Path. We just need an Agg over the
7628 : * cheapest-total input path, since input order won't matter.
7629 : */
7630 4339 : add_path(grouped_rel, (Path *)
7631 4339 : create_agg_path(root, grouped_rel,
7632 : cheapest_path,
7633 4339 : grouped_rel->reltarget,
7634 : AGG_HASHED,
7635 : AGGSPLIT_SIMPLE,
7636 : root->processed_groupClause,
7637 : havingQual,
7638 : agg_costs,
7639 : dNumGroups));
7640 : }
7641 :
7642 : /*
7643 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7644 : * grouped path, assuming there is one
7645 : */
7646 5146 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7647 : {
7648 1192 : add_path(grouped_rel, (Path *)
7649 1192 : create_agg_path(root,
7650 : grouped_rel,
7651 : cheapest_partially_grouped_path,
7652 1192 : grouped_rel->reltarget,
7653 : AGG_HASHED,
7654 : AGGSPLIT_FINAL_DESERIAL,
7655 : root->processed_groupClause,
7656 : havingQual,
7657 : agg_final_costs,
7658 : dNumFinalGroups));
7659 : }
7660 : }
7661 :
7662 : /*
7663 : * When partitionwise aggregate is used, we might have fully aggregated
7664 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7665 : * consider a path for grouped_rel consisting of a Parallel Append of
7666 : * non-partial paths from each child.
7667 : */
7668 34979 : if (grouped_rel->partial_pathlist != NIL)
7669 265 : gather_grouping_paths(root, grouped_rel);
7670 34979 : }
7671 :
7672 : /*
7673 : * create_partial_grouping_paths
7674 : *
7675 : * Create a new upper relation representing the result of partial aggregation
7676 : * and populate it with appropriate paths. Note that we don't finalize the
7677 : * lists of paths here, so the caller can add additional partial or non-partial
7678 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7679 : * the returned upper relation.
7680 : *
7681 : * All paths for this new upper relation -- both partial and non-partial --
7682 : * have been partially aggregated but require a subsequent FinalizeAggregate
7683 : * step.
7684 : *
7685 : * NB: This function is allowed to return NULL if it determines that there is
7686 : * no real need to create a new RelOptInfo.
7687 : */
7688 : static RelOptInfo *
7689 31709 : create_partial_grouping_paths(PlannerInfo *root,
7690 : RelOptInfo *grouped_rel,
7691 : RelOptInfo *input_rel,
7692 : grouping_sets_data *gd,
7693 : GroupPathExtraData *extra,
7694 : bool force_rel_creation)
7695 : {
7696 31709 : Query *parse = root->parse;
7697 : RelOptInfo *partially_grouped_rel;
7698 31709 : RelOptInfo *eager_agg_rel = NULL;
7699 31709 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7700 31709 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7701 31709 : Path *cheapest_partial_path = NULL;
7702 31709 : Path *cheapest_total_path = NULL;
7703 31709 : double dNumPartialGroups = 0;
7704 31709 : double dNumPartialPartialGroups = 0;
7705 : ListCell *lc;
7706 31709 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7707 31709 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7708 :
7709 : /*
7710 : * Check whether any partially aggregated paths have been generated
7711 : * through eager aggregation.
7712 : */
7713 31709 : if (input_rel->grouped_rel &&
7714 797 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7715 797 : input_rel->grouped_rel->pathlist != NIL)
7716 737 : eager_agg_rel = input_rel->grouped_rel;
7717 :
7718 : /*
7719 : * Consider whether we should generate partially aggregated non-partial
7720 : * paths. We can only do this if we have a non-partial path, and only if
7721 : * the parent of the input rel is performing partial partitionwise
7722 : * aggregation. (Note that extra->patype is the type of partitionwise
7723 : * aggregation being used at the parent level, not this level.)
7724 : */
7725 31709 : if (input_rel->pathlist != NIL &&
7726 31709 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7727 713 : cheapest_total_path = input_rel->cheapest_total_path;
7728 :
7729 : /*
7730 : * If parallelism is possible for grouped_rel, then we should consider
7731 : * generating partially-grouped partial paths. However, if the input rel
7732 : * has no partial paths, then we can't.
7733 : */
7734 31709 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7735 2745 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7736 :
7737 : /*
7738 : * If we can't partially aggregate partial paths, and we can't partially
7739 : * aggregate non-partial paths, and no partially aggregated paths were
7740 : * generated by eager aggregation, then don't bother creating the new
7741 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7742 : */
7743 31709 : if (cheapest_total_path == NULL &&
7744 28671 : cheapest_partial_path == NULL &&
7745 28574 : eager_agg_rel == NULL &&
7746 28574 : !force_rel_creation)
7747 28493 : return NULL;
7748 :
7749 : /*
7750 : * Build a new upper relation to represent the result of partially
7751 : * aggregating the rows from the input relation.
7752 : */
7753 3216 : partially_grouped_rel = fetch_upper_rel(root,
7754 : UPPERREL_PARTIAL_GROUP_AGG,
7755 : grouped_rel->relids);
7756 3216 : partially_grouped_rel->consider_parallel =
7757 3216 : grouped_rel->consider_parallel;
7758 3216 : partially_grouped_rel->pgs_mask = grouped_rel->pgs_mask;
7759 3216 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7760 3216 : partially_grouped_rel->serverid = grouped_rel->serverid;
7761 3216 : partially_grouped_rel->userid = grouped_rel->userid;
7762 3216 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7763 3216 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7764 :
7765 : /*
7766 : * Build target list for partial aggregate paths. These paths cannot just
7767 : * emit the same tlist as regular aggregate paths, because (1) we must
7768 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7769 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7770 : */
7771 3216 : partially_grouped_rel->reltarget =
7772 3216 : make_partial_grouping_target(root, grouped_rel->reltarget,
7773 : extra->havingQual);
7774 :
7775 3216 : if (!extra->partial_costs_set)
7776 : {
7777 : /*
7778 : * Collect statistics about aggregates for estimating costs of
7779 : * performing aggregation in parallel.
7780 : */
7781 11598 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7782 11598 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7783 1933 : if (parse->hasAggs)
7784 : {
7785 : /* partial phase */
7786 1824 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7787 : agg_partial_costs);
7788 :
7789 : /* final phase */
7790 1824 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7791 : agg_final_costs);
7792 : }
7793 :
7794 1933 : extra->partial_costs_set = true;
7795 : }
7796 :
7797 : /* Estimate number of partial groups. */
7798 3216 : if (cheapest_total_path != NULL)
7799 : dNumPartialGroups =
7800 713 : get_number_of_groups(root,
7801 : cheapest_total_path->rows,
7802 : gd,
7803 : extra->targetList);
7804 3216 : if (cheapest_partial_path != NULL)
7805 : dNumPartialPartialGroups =
7806 2745 : get_number_of_groups(root,
7807 : cheapest_partial_path->rows,
7808 : gd,
7809 : extra->targetList);
7810 :
7811 3216 : if (can_sort && cheapest_total_path != NULL)
7812 : {
7813 : /* This should have been checked previously */
7814 : Assert(parse->hasAggs || parse->groupClause);
7815 :
7816 : /*
7817 : * Use any available suitably-sorted path as input, and also consider
7818 : * sorting the cheapest partial path.
7819 : */
7820 1426 : foreach(lc, input_rel->pathlist)
7821 : {
7822 : ListCell *lc2;
7823 713 : Path *path = (Path *) lfirst(lc);
7824 713 : Path *path_save = path;
7825 713 : List *pathkey_orderings = NIL;
7826 :
7827 : /* generate alternative group orderings that might be useful */
7828 713 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7829 :
7830 : Assert(list_length(pathkey_orderings) > 0);
7831 :
7832 : /* process all potentially interesting grouping reorderings */
7833 1426 : foreach(lc2, pathkey_orderings)
7834 : {
7835 713 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7836 :
7837 : /* restore the path (we replace it in the loop) */
7838 713 : path = path_save;
7839 :
7840 713 : path = make_ordered_path(root,
7841 : partially_grouped_rel,
7842 : path,
7843 : cheapest_total_path,
7844 : info->pathkeys,
7845 : -1.0);
7846 :
7847 713 : if (path == NULL)
7848 0 : continue;
7849 :
7850 713 : if (parse->hasAggs)
7851 653 : add_path(partially_grouped_rel, (Path *)
7852 653 : create_agg_path(root,
7853 : partially_grouped_rel,
7854 : path,
7855 653 : partially_grouped_rel->reltarget,
7856 653 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7857 : AGGSPLIT_INITIAL_SERIAL,
7858 : info->clauses,
7859 : NIL,
7860 : agg_partial_costs,
7861 : dNumPartialGroups));
7862 : else
7863 60 : add_path(partially_grouped_rel, (Path *)
7864 60 : create_group_path(root,
7865 : partially_grouped_rel,
7866 : path,
7867 : info->clauses,
7868 : NIL,
7869 : dNumPartialGroups));
7870 : }
7871 : }
7872 : }
7873 :
7874 3216 : if (can_sort && cheapest_partial_path != NULL)
7875 : {
7876 : /* Similar to above logic, but for partial paths. */
7877 5865 : foreach(lc, input_rel->partial_pathlist)
7878 : {
7879 : ListCell *lc2;
7880 3120 : Path *path = (Path *) lfirst(lc);
7881 3120 : Path *path_save = path;
7882 3120 : List *pathkey_orderings = NIL;
7883 :
7884 : /* generate alternative group orderings that might be useful */
7885 3120 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7886 :
7887 : Assert(list_length(pathkey_orderings) > 0);
7888 :
7889 : /* process all potentially interesting grouping reorderings */
7890 6240 : foreach(lc2, pathkey_orderings)
7891 : {
7892 3120 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7893 :
7894 :
7895 : /* restore the path (we replace it in the loop) */
7896 3120 : path = path_save;
7897 :
7898 3120 : path = make_ordered_path(root,
7899 : partially_grouped_rel,
7900 : path,
7901 : cheapest_partial_path,
7902 : info->pathkeys,
7903 : -1.0);
7904 :
7905 3120 : if (path == NULL)
7906 5 : continue;
7907 :
7908 3115 : if (parse->hasAggs)
7909 3016 : add_partial_path(partially_grouped_rel, (Path *)
7910 3016 : create_agg_path(root,
7911 : partially_grouped_rel,
7912 : path,
7913 3016 : partially_grouped_rel->reltarget,
7914 3016 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7915 : AGGSPLIT_INITIAL_SERIAL,
7916 : info->clauses,
7917 : NIL,
7918 : agg_partial_costs,
7919 : dNumPartialPartialGroups));
7920 : else
7921 99 : add_partial_path(partially_grouped_rel, (Path *)
7922 99 : create_group_path(root,
7923 : partially_grouped_rel,
7924 : path,
7925 : info->clauses,
7926 : NIL,
7927 : dNumPartialPartialGroups));
7928 : }
7929 : }
7930 : }
7931 :
7932 : /*
7933 : * Add a partially-grouped HashAgg Path where possible
7934 : */
7935 3216 : if (can_hash && cheapest_total_path != NULL)
7936 : {
7937 : /* Checked above */
7938 : Assert(parse->hasAggs || parse->groupClause);
7939 :
7940 713 : add_path(partially_grouped_rel, (Path *)
7941 713 : create_agg_path(root,
7942 : partially_grouped_rel,
7943 : cheapest_total_path,
7944 713 : partially_grouped_rel->reltarget,
7945 : AGG_HASHED,
7946 : AGGSPLIT_INITIAL_SERIAL,
7947 : root->processed_groupClause,
7948 : NIL,
7949 : agg_partial_costs,
7950 : dNumPartialGroups));
7951 : }
7952 :
7953 : /*
7954 : * Now add a partially-grouped HashAgg partial Path where possible
7955 : */
7956 3216 : if (can_hash && cheapest_partial_path != NULL)
7957 : {
7958 1434 : add_partial_path(partially_grouped_rel, (Path *)
7959 1434 : create_agg_path(root,
7960 : partially_grouped_rel,
7961 : cheapest_partial_path,
7962 1434 : partially_grouped_rel->reltarget,
7963 : AGG_HASHED,
7964 : AGGSPLIT_INITIAL_SERIAL,
7965 : root->processed_groupClause,
7966 : NIL,
7967 : agg_partial_costs,
7968 : dNumPartialPartialGroups));
7969 : }
7970 :
7971 : /*
7972 : * Add any partially aggregated paths generated by eager aggregation to
7973 : * the new upper relation after applying projection steps as needed.
7974 : */
7975 3216 : if (eager_agg_rel)
7976 : {
7977 : /* Add the paths */
7978 1924 : foreach(lc, eager_agg_rel->pathlist)
7979 : {
7980 1187 : Path *path = (Path *) lfirst(lc);
7981 :
7982 : /* Shouldn't have any parameterized paths anymore */
7983 : Assert(path->param_info == NULL);
7984 :
7985 1187 : path = (Path *) create_projection_path(root,
7986 : partially_grouped_rel,
7987 : path,
7988 1187 : partially_grouped_rel->reltarget);
7989 :
7990 1187 : add_path(partially_grouped_rel, path);
7991 : }
7992 :
7993 : /*
7994 : * Likewise add the partial paths, but only if parallelism is possible
7995 : * for partially_grouped_rel.
7996 : */
7997 737 : if (partially_grouped_rel->consider_parallel)
7998 : {
7999 1700 : foreach(lc, eager_agg_rel->partial_pathlist)
8000 : {
8001 1010 : Path *path = (Path *) lfirst(lc);
8002 :
8003 : /* Shouldn't have any parameterized paths anymore */
8004 : Assert(path->param_info == NULL);
8005 :
8006 1010 : path = (Path *) create_projection_path(root,
8007 : partially_grouped_rel,
8008 : path,
8009 1010 : partially_grouped_rel->reltarget);
8010 :
8011 1010 : add_partial_path(partially_grouped_rel, path);
8012 : }
8013 : }
8014 : }
8015 :
8016 : /*
8017 : * If there is an FDW that's responsible for all baserels of the query,
8018 : * let it consider adding partially grouped ForeignPaths.
8019 : */
8020 3216 : if (partially_grouped_rel->fdwroutine &&
8021 3 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
8022 : {
8023 3 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
8024 :
8025 3 : fdwroutine->GetForeignUpperPaths(root,
8026 : UPPERREL_PARTIAL_GROUP_AGG,
8027 : input_rel, partially_grouped_rel,
8028 : extra);
8029 : }
8030 :
8031 3216 : return partially_grouped_rel;
8032 : }
8033 :
8034 : /*
8035 : * make_ordered_path
8036 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
8037 : * return NULL if it doesn't make sense to generate an ordered path in
8038 : * this case.
8039 : */
8040 : static Path *
8041 49401 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
8042 : Path *cheapest_path, List *pathkeys, double limit_tuples)
8043 : {
8044 : bool is_sorted;
8045 : int presorted_keys;
8046 :
8047 49401 : is_sorted = pathkeys_count_contained_in(pathkeys,
8048 : path->pathkeys,
8049 : &presorted_keys);
8050 :
8051 49401 : if (!is_sorted)
8052 : {
8053 : /*
8054 : * Try at least sorting the cheapest path and also try incrementally
8055 : * sorting any path which is partially sorted already (no need to deal
8056 : * with paths which have presorted keys when incremental sort is
8057 : * disabled unless it's the cheapest input path).
8058 : */
8059 13155 : if (path != cheapest_path &&
8060 2157 : (presorted_keys == 0 || !enable_incremental_sort))
8061 1022 : return NULL;
8062 :
8063 : /*
8064 : * We've no need to consider both a sort and incremental sort. We'll
8065 : * just do a sort if there are no presorted keys and an incremental
8066 : * sort when there are presorted keys.
8067 : */
8068 12133 : if (presorted_keys == 0 || !enable_incremental_sort)
8069 10852 : path = (Path *) create_sort_path(root,
8070 : rel,
8071 : path,
8072 : pathkeys,
8073 : limit_tuples);
8074 : else
8075 1281 : path = (Path *) create_incremental_sort_path(root,
8076 : rel,
8077 : path,
8078 : pathkeys,
8079 : presorted_keys,
8080 : limit_tuples);
8081 : }
8082 :
8083 48379 : return path;
8084 : }
8085 :
8086 : /*
8087 : * Generate Gather and Gather Merge paths for a grouping relation or partial
8088 : * grouping relation.
8089 : *
8090 : * generate_useful_gather_paths does most of the work, but we also consider a
8091 : * special case: we could try sorting the data by the group_pathkeys and then
8092 : * applying Gather Merge.
8093 : *
8094 : * NB: This function shouldn't be used for anything other than a grouped or
8095 : * partially grouped relation not only because of the fact that it explicitly
8096 : * references group_pathkeys but we pass "true" as the third argument to
8097 : * generate_useful_gather_paths().
8098 : */
8099 : static void
8100 2590 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
8101 : {
8102 : ListCell *lc;
8103 : Path *cheapest_partial_path;
8104 : List *groupby_pathkeys;
8105 :
8106 : /*
8107 : * This occurs after any partial aggregation has taken place, so trim off
8108 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
8109 : */
8110 2590 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
8111 15 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
8112 : root->num_groupby_pathkeys);
8113 : else
8114 2575 : groupby_pathkeys = root->group_pathkeys;
8115 :
8116 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
8117 2590 : generate_useful_gather_paths(root, rel, true);
8118 :
8119 2590 : cheapest_partial_path = linitial(rel->partial_pathlist);
8120 :
8121 : /* XXX Shouldn't this also consider the group-key-reordering? */
8122 6152 : foreach(lc, rel->partial_pathlist)
8123 : {
8124 3562 : Path *path = (Path *) lfirst(lc);
8125 : bool is_sorted;
8126 : int presorted_keys;
8127 : double total_groups;
8128 :
8129 3562 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
8130 : path->pathkeys,
8131 : &presorted_keys);
8132 :
8133 3562 : if (is_sorted)
8134 2329 : continue;
8135 :
8136 : /*
8137 : * Try at least sorting the cheapest path and also try incrementally
8138 : * sorting any path which is partially sorted already (no need to deal
8139 : * with paths which have presorted keys when incremental sort is
8140 : * disabled unless it's the cheapest input path).
8141 : */
8142 1233 : if (path != cheapest_partial_path &&
8143 0 : (presorted_keys == 0 || !enable_incremental_sort))
8144 0 : continue;
8145 :
8146 : /*
8147 : * We've no need to consider both a sort and incremental sort. We'll
8148 : * just do a sort if there are no presorted keys and an incremental
8149 : * sort when there are presorted keys.
8150 : */
8151 1233 : if (presorted_keys == 0 || !enable_incremental_sort)
8152 1233 : path = (Path *) create_sort_path(root, rel, path,
8153 : groupby_pathkeys,
8154 : -1.0);
8155 : else
8156 0 : path = (Path *) create_incremental_sort_path(root,
8157 : rel,
8158 : path,
8159 : groupby_pathkeys,
8160 : presorted_keys,
8161 : -1.0);
8162 1233 : total_groups = compute_gather_rows(path);
8163 : path = (Path *)
8164 1233 : create_gather_merge_path(root,
8165 : rel,
8166 : path,
8167 1233 : rel->reltarget,
8168 : groupby_pathkeys,
8169 : NULL,
8170 : &total_groups);
8171 :
8172 1233 : add_path(rel, path);
8173 : }
8174 2590 : }
8175 :
8176 : /*
8177 : * can_partial_agg
8178 : *
8179 : * Determines whether or not partial grouping and/or aggregation is possible.
8180 : * Returns true when possible, false otherwise.
8181 : */
8182 : static bool
8183 33897 : can_partial_agg(PlannerInfo *root)
8184 : {
8185 33897 : Query *parse = root->parse;
8186 :
8187 33897 : if (!parse->hasAggs && parse->groupClause == NIL)
8188 : {
8189 : /*
8190 : * We don't know how to do parallel aggregation unless we have either
8191 : * some aggregates or a grouping clause.
8192 : */
8193 0 : return false;
8194 : }
8195 33897 : else if (parse->groupingSets)
8196 : {
8197 : /* We don't know how to do grouping sets in parallel. */
8198 886 : return false;
8199 : }
8200 33011 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
8201 : {
8202 : /* Insufficient support for partial mode. */
8203 3037 : return false;
8204 : }
8205 :
8206 : /* Everything looks good. */
8207 29974 : return true;
8208 : }
8209 :
8210 : /*
8211 : * apply_scanjoin_target_to_paths
8212 : *
8213 : * Adjust the final scan/join relation, and recursively all of its children,
8214 : * to generate the final scan/join target. It would be more correct to model
8215 : * this as a separate planning step with a new RelOptInfo at the toplevel and
8216 : * for each child relation, but doing it this way is noticeably cheaper.
8217 : * Maybe that problem can be solved at some point, but for now we do this.
8218 : *
8219 : * If tlist_same_exprs is true, then the scan/join target to be applied has
8220 : * the same expressions as the existing reltarget, so we need only insert the
8221 : * appropriate sortgroupref information. By avoiding the creation of
8222 : * projection paths we save effort both immediately and at plan creation time.
8223 : */
8224 : static void
8225 407783 : apply_scanjoin_target_to_paths(PlannerInfo *root,
8226 : RelOptInfo *rel,
8227 : List *scanjoin_targets,
8228 : List *scanjoin_targets_contain_srfs,
8229 : bool scanjoin_target_parallel_safe,
8230 : bool tlist_same_exprs)
8231 : {
8232 407783 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
8233 : PathTarget *scanjoin_target;
8234 : ListCell *lc;
8235 :
8236 : /* This recurses, so be paranoid. */
8237 407783 : check_stack_depth();
8238 :
8239 : /*
8240 : * If the rel only has Append and MergeAppend paths, we want to drop its
8241 : * existing paths and generate new ones. This function would still be
8242 : * correct if we kept the existing paths: we'd modify them to generate the
8243 : * correct target above the partitioning Append, and then they'd compete
8244 : * on cost with paths generating the target below the Append. However, in
8245 : * our current cost model the latter way is always the same or cheaper
8246 : * cost, so modifying the existing paths would just be useless work.
8247 : * Moreover, when the cost is the same, varying roundoff errors might
8248 : * sometimes allow an existing path to be picked, resulting in undesirable
8249 : * cross-platform plan variations. So we drop old paths and thereby force
8250 : * the work to be done below the Append.
8251 : *
8252 : * However, there are several cases when this optimization is not safe. If
8253 : * the rel isn't partitioned, then none of the paths will be Append or
8254 : * MergeAppend paths, so we should definitely not do this. If it is
8255 : * partitioned but is a joinrel, it may have Append and MergeAppend paths,
8256 : * but it can also have join paths that we can't afford to discard.
8257 : *
8258 : * Some care is needed, because we have to allow
8259 : * generate_useful_gather_paths to see the old partial paths in the next
8260 : * stanza. Hence, zap the main pathlist here, then allow
8261 : * generate_useful_gather_paths to add path(s) to the main list, and
8262 : * finally zap the partial pathlist.
8263 : */
8264 407783 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8265 8746 : rel->pathlist = NIL;
8266 :
8267 : /*
8268 : * If the scan/join target is not parallel-safe, partial paths cannot
8269 : * generate it.
8270 : */
8271 407783 : if (!scanjoin_target_parallel_safe)
8272 : {
8273 : /*
8274 : * Since we can't generate the final scan/join target in parallel
8275 : * workers, this is our last opportunity to use any partial paths that
8276 : * exist; so build Gather path(s) that use them and emit whatever the
8277 : * current reltarget is. We don't do this in the case where the
8278 : * target is parallel-safe, since we will be able to generate superior
8279 : * paths by doing it after the final scan/join target has been
8280 : * applied.
8281 : */
8282 52636 : generate_useful_gather_paths(root, rel, false);
8283 :
8284 : /* Can't use parallel query above this level. */
8285 52636 : rel->partial_pathlist = NIL;
8286 52636 : rel->consider_parallel = false;
8287 : }
8288 :
8289 : /* Finish dropping old paths for a partitioned rel, per comment above */
8290 407783 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8291 8746 : rel->partial_pathlist = NIL;
8292 :
8293 : /* Extract SRF-free scan/join target. */
8294 407783 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
8295 :
8296 : /*
8297 : * Apply the SRF-free scan/join target to each existing path.
8298 : *
8299 : * If the tlist exprs are the same, we can just inject the sortgroupref
8300 : * information into the existing pathtargets. Otherwise, replace each
8301 : * path with a projection path that generates the SRF-free scan/join
8302 : * target. This can't change the ordering of paths within rel->pathlist,
8303 : * so we just modify the list in place.
8304 : */
8305 851041 : foreach(lc, rel->pathlist)
8306 : {
8307 443258 : Path *subpath = (Path *) lfirst(lc);
8308 :
8309 : /* Shouldn't have any parameterized paths anymore */
8310 : Assert(subpath->param_info == NULL);
8311 :
8312 443258 : if (tlist_same_exprs)
8313 159478 : subpath->pathtarget->sortgrouprefs =
8314 159478 : scanjoin_target->sortgrouprefs;
8315 : else
8316 : {
8317 : Path *newpath;
8318 :
8319 283780 : newpath = (Path *) create_projection_path(root, rel, subpath,
8320 : scanjoin_target);
8321 283780 : lfirst(lc) = newpath;
8322 : }
8323 : }
8324 :
8325 : /* Likewise adjust the targets for any partial paths. */
8326 428022 : foreach(lc, rel->partial_pathlist)
8327 : {
8328 20239 : Path *subpath = (Path *) lfirst(lc);
8329 :
8330 : /* Shouldn't have any parameterized paths anymore */
8331 : Assert(subpath->param_info == NULL);
8332 :
8333 20239 : if (tlist_same_exprs)
8334 16106 : subpath->pathtarget->sortgrouprefs =
8335 16106 : scanjoin_target->sortgrouprefs;
8336 : else
8337 : {
8338 : Path *newpath;
8339 :
8340 4133 : newpath = (Path *) create_projection_path(root, rel, subpath,
8341 : scanjoin_target);
8342 4133 : lfirst(lc) = newpath;
8343 : }
8344 : }
8345 :
8346 : /*
8347 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8348 : * atop each existing path. (Note that this function doesn't look at the
8349 : * cheapest-path fields, which is a good thing because they're bogus right
8350 : * now.)
8351 : */
8352 407783 : if (root->parse->hasTargetSRFs)
8353 10174 : adjust_paths_for_srfs(root, rel,
8354 : scanjoin_targets,
8355 : scanjoin_targets_contain_srfs);
8356 :
8357 : /*
8358 : * Update the rel's target to be the final (with SRFs) scan/join target.
8359 : * This now matches the actual output of all the paths, and we might get
8360 : * confused in createplan.c if they don't agree. We must do this now so
8361 : * that any append paths made in the next part will use the correct
8362 : * pathtarget (cf. create_append_path).
8363 : *
8364 : * Note that this is also necessary if GetForeignUpperPaths() gets called
8365 : * on the final scan/join relation or on any of its children, since the
8366 : * FDW might look at the rel's target to create ForeignPaths.
8367 : */
8368 407783 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8369 :
8370 : /*
8371 : * If the relation is partitioned, recursively apply the scan/join target
8372 : * to all partitions, and generate brand-new Append paths in which the
8373 : * scan/join target is computed below the Append rather than above it.
8374 : * Since Append is not projection-capable, that might save a separate
8375 : * Result node, and it also is important for partitionwise aggregate.
8376 : */
8377 407783 : if (rel_is_partitioned)
8378 : {
8379 10049 : List *live_children = NIL;
8380 : int i;
8381 :
8382 : /* Adjust each partition. */
8383 10049 : i = -1;
8384 30172 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8385 : {
8386 20123 : RelOptInfo *child_rel = rel->part_rels[i];
8387 : AppendRelInfo **appinfos;
8388 : int nappinfos;
8389 20123 : List *child_scanjoin_targets = NIL;
8390 :
8391 : Assert(child_rel != NULL);
8392 :
8393 : /* Dummy children can be ignored. */
8394 20123 : if (IS_DUMMY_REL(child_rel))
8395 40 : continue;
8396 :
8397 : /* Translate scan/join targets for this child. */
8398 20083 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8399 : &nappinfos);
8400 40166 : foreach(lc, scanjoin_targets)
8401 : {
8402 20083 : PathTarget *target = lfirst_node(PathTarget, lc);
8403 :
8404 20083 : target = copy_pathtarget(target);
8405 20083 : target->exprs = (List *)
8406 20083 : adjust_appendrel_attrs(root,
8407 20083 : (Node *) target->exprs,
8408 : nappinfos, appinfos);
8409 20083 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8410 : target);
8411 : }
8412 20083 : pfree(appinfos);
8413 :
8414 : /* Recursion does the real work. */
8415 20083 : apply_scanjoin_target_to_paths(root, child_rel,
8416 : child_scanjoin_targets,
8417 : scanjoin_targets_contain_srfs,
8418 : scanjoin_target_parallel_safe,
8419 : tlist_same_exprs);
8420 :
8421 : /* Save non-dummy children for Append paths. */
8422 20083 : if (!IS_DUMMY_REL(child_rel))
8423 20083 : live_children = lappend(live_children, child_rel);
8424 : }
8425 :
8426 : /* Build new paths for this relation by appending child paths. */
8427 10049 : add_paths_to_append_rel(root, rel, live_children);
8428 : }
8429 :
8430 : /*
8431 : * Consider generating Gather or Gather Merge paths. We must only do this
8432 : * if the relation is parallel safe, and we don't do it for child rels to
8433 : * avoid creating multiple Gather nodes within the same plan. We must do
8434 : * this after all paths have been generated and before set_cheapest, since
8435 : * one of the generated paths may turn out to be the cheapest one.
8436 : */
8437 407783 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
8438 135608 : generate_useful_gather_paths(root, rel, false);
8439 :
8440 : /*
8441 : * Reassess which paths are the cheapest, now that we've potentially added
8442 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8443 : * this relation.
8444 : */
8445 407783 : set_cheapest(rel);
8446 407783 : }
8447 :
8448 : /*
8449 : * create_partitionwise_grouping_paths
8450 : *
8451 : * If the partition keys of input relation are part of the GROUP BY clause, all
8452 : * the rows belonging to a given group come from a single partition. This
8453 : * allows aggregation/grouping over a partitioned relation to be broken down
8454 : * into aggregation/grouping on each partition. This should be no worse, and
8455 : * often better, than the normal approach.
8456 : *
8457 : * However, if the GROUP BY clause does not contain all the partition keys,
8458 : * rows from a given group may be spread across multiple partitions. In that
8459 : * case, we perform partial aggregation for each group, append the results,
8460 : * and then finalize aggregation. This is less certain to win than the
8461 : * previous case. It may win if the PartialAggregate stage greatly reduces
8462 : * the number of groups, because fewer rows will pass through the Append node.
8463 : * It may lose if we have lots of small groups.
8464 : */
8465 : static void
8466 685 : create_partitionwise_grouping_paths(PlannerInfo *root,
8467 : RelOptInfo *input_rel,
8468 : RelOptInfo *grouped_rel,
8469 : RelOptInfo *partially_grouped_rel,
8470 : const AggClauseCosts *agg_costs,
8471 : grouping_sets_data *gd,
8472 : PartitionwiseAggregateType patype,
8473 : GroupPathExtraData *extra)
8474 : {
8475 685 : List *grouped_live_children = NIL;
8476 685 : List *partially_grouped_live_children = NIL;
8477 685 : PathTarget *target = grouped_rel->reltarget;
8478 685 : bool partial_grouping_valid = true;
8479 : int i;
8480 :
8481 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8482 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8483 : partially_grouped_rel != NULL);
8484 :
8485 : /* Add paths for partitionwise aggregation/grouping. */
8486 685 : i = -1;
8487 2480 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8488 : {
8489 1795 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8490 : PathTarget *child_target;
8491 : AppendRelInfo **appinfos;
8492 : int nappinfos;
8493 : GroupPathExtraData child_extra;
8494 : RelOptInfo *child_grouped_rel;
8495 : RelOptInfo *child_partially_grouped_rel;
8496 :
8497 : Assert(child_input_rel != NULL);
8498 :
8499 : /* Dummy children can be ignored. */
8500 1795 : if (IS_DUMMY_REL(child_input_rel))
8501 0 : continue;
8502 :
8503 1795 : child_target = copy_pathtarget(target);
8504 :
8505 : /*
8506 : * Copy the given "extra" structure as is and then override the
8507 : * members specific to this child.
8508 : */
8509 1795 : memcpy(&child_extra, extra, sizeof(child_extra));
8510 :
8511 1795 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8512 : &nappinfos);
8513 :
8514 1795 : child_target->exprs = (List *)
8515 1795 : adjust_appendrel_attrs(root,
8516 1795 : (Node *) target->exprs,
8517 : nappinfos, appinfos);
8518 :
8519 : /* Translate havingQual and targetList. */
8520 1795 : child_extra.havingQual = (Node *)
8521 : adjust_appendrel_attrs(root,
8522 : extra->havingQual,
8523 : nappinfos, appinfos);
8524 1795 : child_extra.targetList = (List *)
8525 1795 : adjust_appendrel_attrs(root,
8526 1795 : (Node *) extra->targetList,
8527 : nappinfos, appinfos);
8528 :
8529 : /*
8530 : * extra->patype was the value computed for our parent rel; patype is
8531 : * the value for this relation. For the child, our value is its
8532 : * parent rel's value.
8533 : */
8534 1795 : child_extra.patype = patype;
8535 :
8536 : /*
8537 : * Create grouping relation to hold fully aggregated grouping and/or
8538 : * aggregation paths for the child.
8539 : */
8540 1795 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8541 : child_target,
8542 1795 : extra->target_parallel_safe,
8543 : child_extra.havingQual);
8544 :
8545 : /* Create grouping paths for this child relation. */
8546 1795 : create_ordinary_grouping_paths(root, child_input_rel,
8547 : child_grouped_rel,
8548 : agg_costs, gd, &child_extra,
8549 : &child_partially_grouped_rel);
8550 :
8551 1795 : if (child_partially_grouped_rel)
8552 : {
8553 : partially_grouped_live_children =
8554 1283 : lappend(partially_grouped_live_children,
8555 : child_partially_grouped_rel);
8556 : }
8557 : else
8558 512 : partial_grouping_valid = false;
8559 :
8560 1795 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8561 : {
8562 1082 : set_cheapest(child_grouped_rel);
8563 1082 : grouped_live_children = lappend(grouped_live_children,
8564 : child_grouped_rel);
8565 : }
8566 :
8567 1795 : pfree(appinfos);
8568 : }
8569 :
8570 : /*
8571 : * Try to create append paths for partially grouped children. For full
8572 : * partitionwise aggregation, we might have paths in the partial_pathlist
8573 : * if parallel aggregation is possible. For partial partitionwise
8574 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8575 : *
8576 : * NB: We must have a partially grouped path for every child in order to
8577 : * generate a partially grouped path for this relation.
8578 : */
8579 685 : if (partially_grouped_rel && partial_grouping_valid)
8580 : {
8581 : Assert(partially_grouped_live_children != NIL);
8582 :
8583 501 : add_paths_to_append_rel(root, partially_grouped_rel,
8584 : partially_grouped_live_children);
8585 : }
8586 :
8587 : /* If possible, create append paths for fully grouped children. */
8588 685 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8589 : {
8590 : Assert(grouped_live_children != NIL);
8591 :
8592 404 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8593 : }
8594 685 : }
8595 :
8596 : /*
8597 : * group_by_has_partkey
8598 : *
8599 : * Returns true if all the partition keys of the given relation are part of
8600 : * the GROUP BY clauses, including having matching collation, false otherwise.
8601 : */
8602 : static bool
8603 640 : group_by_has_partkey(RelOptInfo *input_rel,
8604 : List *targetList,
8605 : List *groupClause)
8606 : {
8607 640 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8608 640 : int cnt = 0;
8609 : int partnatts;
8610 :
8611 : /* Input relation should be partitioned. */
8612 : Assert(input_rel->part_scheme);
8613 :
8614 : /* Rule out early, if there are no partition keys present. */
8615 640 : if (!input_rel->partexprs)
8616 0 : return false;
8617 :
8618 640 : partnatts = input_rel->part_scheme->partnatts;
8619 :
8620 1074 : for (cnt = 0; cnt < partnatts; cnt++)
8621 : {
8622 670 : List *partexprs = input_rel->partexprs[cnt];
8623 : ListCell *lc;
8624 670 : bool found = false;
8625 :
8626 1001 : foreach(lc, partexprs)
8627 : {
8628 : ListCell *lg;
8629 775 : Expr *partexpr = lfirst(lc);
8630 775 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8631 :
8632 1206 : foreach(lg, groupexprs)
8633 : {
8634 875 : Expr *groupexpr = lfirst(lg);
8635 875 : Oid groupcoll = exprCollation((Node *) groupexpr);
8636 :
8637 : /*
8638 : * Note: we can assume there is at most one RelabelType node;
8639 : * eval_const_expressions() will have simplified if more than
8640 : * one.
8641 : */
8642 875 : if (IsA(groupexpr, RelabelType))
8643 20 : groupexpr = ((RelabelType *) groupexpr)->arg;
8644 :
8645 875 : if (equal(groupexpr, partexpr))
8646 : {
8647 : /*
8648 : * Reject a match if the grouping collation does not match
8649 : * the partitioning collation.
8650 : */
8651 444 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8652 : partcoll != groupcoll)
8653 10 : return false;
8654 :
8655 434 : found = true;
8656 434 : break;
8657 : }
8658 : }
8659 :
8660 765 : if (found)
8661 434 : break;
8662 : }
8663 :
8664 : /*
8665 : * If none of the partition key expressions match with any of the
8666 : * GROUP BY expression, return false.
8667 : */
8668 660 : if (!found)
8669 226 : return false;
8670 : }
8671 :
8672 404 : return true;
8673 : }
8674 :
8675 : /*
8676 : * generate_setop_child_grouplist
8677 : * Build a SortGroupClause list defining the sort/grouping properties
8678 : * of the child of a set operation.
8679 : *
8680 : * This is similar to generate_setop_grouplist() but differs as the setop
8681 : * child query's targetlist entries may already have a tleSortGroupRef
8682 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8683 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8684 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8685 : * any of the columns in the targetlist don't match to the setop's colTypes
8686 : * then we return an empty list. This may leave some TLEs with unreferenced
8687 : * ressortgroupref markings, but that's harmless.
8688 : */
8689 : static List *
8690 10478 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8691 : {
8692 10478 : List *grouplist = copyObject(op->groupClauses);
8693 : ListCell *lg;
8694 : ListCell *lt;
8695 : ListCell *ct;
8696 :
8697 10478 : lg = list_head(grouplist);
8698 10478 : ct = list_head(op->colTypes);
8699 40733 : foreach(lt, targetlist)
8700 : {
8701 30496 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8702 : SortGroupClause *sgc;
8703 : Oid coltype;
8704 :
8705 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8706 30496 : if (tle->resjunk)
8707 0 : continue;
8708 :
8709 : /*
8710 : * We expect every non-resjunk target to have a SortGroupClause and
8711 : * colTypes.
8712 : */
8713 : Assert(lg != NULL);
8714 : Assert(ct != NULL);
8715 30496 : sgc = (SortGroupClause *) lfirst(lg);
8716 30496 : coltype = lfirst_oid(ct);
8717 :
8718 : /* reject if target type isn't the same as the setop target type */
8719 30496 : if (coltype != exprType((Node *) tle->expr))
8720 241 : return NIL;
8721 :
8722 30255 : lg = lnext(grouplist, lg);
8723 30255 : ct = lnext(op->colTypes, ct);
8724 :
8725 : /* assign a tleSortGroupRef, or reuse the existing one */
8726 30255 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8727 : }
8728 :
8729 : Assert(lg == NULL);
8730 : Assert(ct == NULL);
8731 :
8732 10237 : return grouplist;
8733 : }
8734 :
8735 : /*
8736 : * create_unique_paths
8737 : * Build a new RelOptInfo containing Paths that represent elimination of
8738 : * distinct rows from the input data. Distinct-ness is defined according to
8739 : * the needs of the semijoin represented by sjinfo. If it is not possible
8740 : * to identify how to make the data unique, NULL is returned.
8741 : *
8742 : * If used at all, this is likely to be called repeatedly on the same rel,
8743 : * so we cache the result.
8744 : */
8745 : RelOptInfo *
8746 6944 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8747 : {
8748 : RelOptInfo *unique_rel;
8749 6944 : List *sortPathkeys = NIL;
8750 6944 : List *groupClause = NIL;
8751 : MemoryContext oldcontext;
8752 :
8753 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8754 : Assert(sjinfo->jointype == JOIN_SEMI);
8755 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8756 :
8757 : /* If result already cached, return it */
8758 6944 : if (rel->unique_rel)
8759 1206 : return rel->unique_rel;
8760 :
8761 : /* If it's not possible to unique-ify, return NULL */
8762 5738 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8763 104 : return NULL;
8764 :
8765 : /*
8766 : * Punt if this is a child relation and we failed to build a unique-ified
8767 : * relation for its parent. This can happen if all the RHS columns were
8768 : * found to be equated to constants when unique-ifying the parent table,
8769 : * leaving no columns to unique-ify.
8770 : */
8771 5634 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8772 10 : return NULL;
8773 :
8774 : /*
8775 : * When called during GEQO join planning, we are in a short-lived memory
8776 : * context. We must make sure that the unique rel and any subsidiary data
8777 : * structures created for a baserel survive the GEQO cycle, else the
8778 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8779 : * are creating those for a joinrel during GEQO, we don't want them to
8780 : * clutter the main planning context. Upshot is that the best solution is
8781 : * to explicitly allocate memory in the same context the given RelOptInfo
8782 : * is in.
8783 : */
8784 5624 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8785 :
8786 5624 : unique_rel = makeNode(RelOptInfo);
8787 5624 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8788 :
8789 : /*
8790 : * clear path info
8791 : */
8792 5624 : unique_rel->pathlist = NIL;
8793 5624 : unique_rel->ppilist = NIL;
8794 5624 : unique_rel->partial_pathlist = NIL;
8795 5624 : unique_rel->cheapest_startup_path = NULL;
8796 5624 : unique_rel->cheapest_total_path = NULL;
8797 5624 : unique_rel->cheapest_parameterized_paths = NIL;
8798 :
8799 : /*
8800 : * Build the target list for the unique rel. We also build the pathkeys
8801 : * that represent the ordering requirements for the sort-based
8802 : * implementation, and the list of SortGroupClause nodes that represent
8803 : * the columns to be grouped on for the hash-based implementation.
8804 : *
8805 : * For a child rel, we can construct these fields from those of its
8806 : * parent.
8807 : */
8808 5624 : if (IS_OTHER_REL(rel))
8809 360 : {
8810 : PathTarget *child_unique_target;
8811 : PathTarget *parent_unique_target;
8812 :
8813 360 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8814 :
8815 360 : child_unique_target = copy_pathtarget(parent_unique_target);
8816 :
8817 : /* Translate the target expressions */
8818 360 : child_unique_target->exprs = (List *)
8819 360 : adjust_appendrel_attrs_multilevel(root,
8820 360 : (Node *) parent_unique_target->exprs,
8821 : rel,
8822 360 : rel->top_parent);
8823 :
8824 360 : unique_rel->reltarget = child_unique_target;
8825 :
8826 360 : sortPathkeys = rel->top_parent->unique_pathkeys;
8827 360 : groupClause = rel->top_parent->unique_groupclause;
8828 : }
8829 : else
8830 : {
8831 : List *newtlist;
8832 : int nextresno;
8833 5264 : List *sortList = NIL;
8834 : ListCell *lc1;
8835 : ListCell *lc2;
8836 :
8837 : /*
8838 : * The values we are supposed to unique-ify may be expressions in the
8839 : * variables of the input rel's targetlist. We have to add any such
8840 : * expressions to the unique rel's targetlist.
8841 : *
8842 : * To complicate matters, some of the values to be unique-ified may be
8843 : * known redundant by the EquivalenceClass machinery (e.g., because
8844 : * they have been equated to constants). There is no need to compare
8845 : * such values during unique-ification, and indeed we had better not
8846 : * try because the Vars involved may not have propagated as high as
8847 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8848 : * detect such cases, which is a tad inefficient but it doesn't seem
8849 : * worth building specialized infrastructure for this.
8850 : */
8851 5264 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8852 5264 : nextresno = list_length(newtlist) + 1;
8853 :
8854 10729 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8855 : {
8856 5465 : Expr *uniqexpr = lfirst(lc1);
8857 5465 : Oid in_oper = lfirst_oid(lc2);
8858 : Oid sortop;
8859 : TargetEntry *tle;
8860 5465 : bool made_tle = false;
8861 :
8862 5465 : tle = tlist_member(uniqexpr, newtlist);
8863 5465 : if (!tle)
8864 : {
8865 2752 : tle = makeTargetEntry(uniqexpr,
8866 : nextresno,
8867 : NULL,
8868 : false);
8869 2752 : newtlist = lappend(newtlist, tle);
8870 2752 : nextresno++;
8871 2752 : made_tle = true;
8872 : }
8873 :
8874 : /*
8875 : * Try to build an ORDER BY list to sort the input compatibly. We
8876 : * do this for each sortable clause even when the clauses are not
8877 : * all sortable, so that we can detect clauses that are redundant
8878 : * according to the pathkey machinery.
8879 : */
8880 5465 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8881 5465 : if (OidIsValid(sortop))
8882 : {
8883 : Oid eqop;
8884 : SortGroupClause *sortcl;
8885 :
8886 : /*
8887 : * The Unique node will need equality operators. Normally
8888 : * these are the same as the IN clause operators, but if those
8889 : * are cross-type operators then the equality operators are
8890 : * the ones for the IN clause operators' RHS datatype.
8891 : */
8892 5465 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8893 5465 : if (!OidIsValid(eqop)) /* shouldn't happen */
8894 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8895 : sortop);
8896 :
8897 5465 : sortcl = makeNode(SortGroupClause);
8898 5465 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8899 5465 : sortcl->eqop = eqop;
8900 5465 : sortcl->sortop = sortop;
8901 5465 : sortcl->reverse_sort = false;
8902 5465 : sortcl->nulls_first = false;
8903 5465 : sortcl->hashable = false; /* no need to make this accurate */
8904 5465 : sortList = lappend(sortList, sortcl);
8905 :
8906 : /*
8907 : * At each step, convert the SortGroupClause list to pathkey
8908 : * form. If the just-added SortGroupClause is redundant, the
8909 : * result will be shorter than the SortGroupClause list.
8910 : */
8911 5465 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8912 : newtlist);
8913 5465 : if (list_length(sortPathkeys) != list_length(sortList))
8914 : {
8915 : /* Drop the redundant SortGroupClause */
8916 1710 : sortList = list_delete_last(sortList);
8917 : Assert(list_length(sortPathkeys) == list_length(sortList));
8918 : /* Undo tlist addition, if we made one */
8919 1710 : if (made_tle)
8920 : {
8921 10 : newtlist = list_delete_last(newtlist);
8922 10 : nextresno--;
8923 : }
8924 : /* We need not consider this clause for hashing, either */
8925 1710 : continue;
8926 : }
8927 : }
8928 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8929 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8930 : in_oper);
8931 :
8932 3755 : if (sjinfo->semi_can_hash)
8933 : {
8934 : /* Create a GROUP BY list for the Agg node to use */
8935 : Oid eq_oper;
8936 : SortGroupClause *groupcl;
8937 :
8938 : /*
8939 : * Get the hashable equality operators for the Agg node to
8940 : * use. Normally these are the same as the IN clause
8941 : * operators, but if those are cross-type operators then the
8942 : * equality operators are the ones for the IN clause
8943 : * operators' RHS datatype.
8944 : */
8945 3755 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8946 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8947 : in_oper);
8948 :
8949 3755 : groupcl = makeNode(SortGroupClause);
8950 3755 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8951 3755 : groupcl->eqop = eq_oper;
8952 3755 : groupcl->sortop = sortop;
8953 3755 : groupcl->reverse_sort = false;
8954 3755 : groupcl->nulls_first = false;
8955 3755 : groupcl->hashable = true;
8956 3755 : groupClause = lappend(groupClause, groupcl);
8957 : }
8958 : }
8959 :
8960 : /*
8961 : * Done building the sortPathkeys and groupClause. But the
8962 : * sortPathkeys are bogus if not all the clauses were sortable.
8963 : */
8964 5264 : if (!sjinfo->semi_can_btree)
8965 0 : sortPathkeys = NIL;
8966 :
8967 : /*
8968 : * It can happen that all the RHS columns are equated to constants.
8969 : * We'd have to do something special to unique-ify in that case, and
8970 : * it's such an unlikely-in-the-real-world case that it's not worth
8971 : * the effort. So just punt if we found no columns to unique-ify.
8972 : */
8973 5264 : if (sortPathkeys == NIL && groupClause == NIL)
8974 : {
8975 1625 : MemoryContextSwitchTo(oldcontext);
8976 1625 : return NULL;
8977 : }
8978 :
8979 : /* Convert the required targetlist back to PathTarget form */
8980 3639 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8981 : }
8982 :
8983 : /* build unique paths based on input rel's pathlist */
8984 3999 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8985 : sjinfo, unique_rel);
8986 :
8987 : /* build unique paths based on input rel's partial_pathlist */
8988 3999 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8989 : sjinfo, unique_rel);
8990 :
8991 : /* Now choose the best path(s) */
8992 3999 : set_cheapest(unique_rel);
8993 :
8994 : /*
8995 : * There shouldn't be any partial paths for the unique relation;
8996 : * otherwise, we won't be able to properly guarantee uniqueness.
8997 : */
8998 : Assert(unique_rel->partial_pathlist == NIL);
8999 :
9000 : /* Cache the result */
9001 3999 : rel->unique_rel = unique_rel;
9002 3999 : rel->unique_pathkeys = sortPathkeys;
9003 3999 : rel->unique_groupclause = groupClause;
9004 :
9005 3999 : MemoryContextSwitchTo(oldcontext);
9006 :
9007 3999 : return unique_rel;
9008 : }
9009 :
9010 : /*
9011 : * create_final_unique_paths
9012 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
9013 : */
9014 : static void
9015 7131 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
9016 : List *sortPathkeys, List *groupClause,
9017 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
9018 : {
9019 7131 : Path *cheapest_input_path = input_rel->cheapest_total_path;
9020 :
9021 : /* Estimate number of output rows */
9022 7131 : unique_rel->rows = estimate_num_groups(root,
9023 : sjinfo->semi_rhs_exprs,
9024 : cheapest_input_path->rows,
9025 : NULL,
9026 : NULL);
9027 :
9028 : /* Consider sort-based implementations, if possible. */
9029 7131 : if (sjinfo->semi_can_btree)
9030 : {
9031 : ListCell *lc;
9032 :
9033 : /*
9034 : * Use any available suitably-sorted path as input, and also consider
9035 : * sorting the cheapest-total path and incremental sort on any paths
9036 : * with presorted keys.
9037 : *
9038 : * To save planning time, we ignore parameterized input paths unless
9039 : * they are the cheapest-total path.
9040 : */
9041 15540 : foreach(lc, input_rel->pathlist)
9042 : {
9043 8409 : Path *input_path = (Path *) lfirst(lc);
9044 : Path *path;
9045 : bool is_sorted;
9046 : int presorted_keys;
9047 :
9048 : /*
9049 : * Ignore parameterized paths that are not the cheapest-total
9050 : * path.
9051 : */
9052 8409 : if (input_path->param_info &&
9053 : input_path != cheapest_input_path)
9054 740 : continue;
9055 :
9056 7705 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
9057 : input_path->pathkeys,
9058 : &presorted_keys);
9059 :
9060 : /*
9061 : * Ignore paths that are not suitably or partially sorted, unless
9062 : * they are the cheapest total path (no need to deal with paths
9063 : * which have presorted keys when incremental sort is disabled).
9064 : */
9065 7705 : if (!is_sorted && input_path != cheapest_input_path &&
9066 76 : (presorted_keys == 0 || !enable_incremental_sort))
9067 36 : continue;
9068 :
9069 : /*
9070 : * Make a separate ProjectionPath in case we need a Result node.
9071 : */
9072 7669 : path = (Path *) create_projection_path(root,
9073 : unique_rel,
9074 : input_path,
9075 7669 : unique_rel->reltarget);
9076 :
9077 7669 : if (!is_sorted)
9078 : {
9079 : /*
9080 : * We've no need to consider both a sort and incremental sort.
9081 : * We'll just do a sort if there are no presorted keys and an
9082 : * incremental sort when there are presorted keys.
9083 : */
9084 3982 : if (presorted_keys == 0 || !enable_incremental_sort)
9085 3942 : path = (Path *) create_sort_path(root,
9086 : unique_rel,
9087 : path,
9088 : sortPathkeys,
9089 : -1.0);
9090 : else
9091 40 : path = (Path *) create_incremental_sort_path(root,
9092 : unique_rel,
9093 : path,
9094 : sortPathkeys,
9095 : presorted_keys,
9096 : -1.0);
9097 : }
9098 :
9099 7669 : path = (Path *) create_unique_path(root, unique_rel, path,
9100 : list_length(sortPathkeys),
9101 : unique_rel->rows);
9102 :
9103 7669 : add_path(unique_rel, path);
9104 : }
9105 : }
9106 :
9107 : /* Consider hash-based implementation, if possible. */
9108 7131 : if (sjinfo->semi_can_hash)
9109 : {
9110 : Path *path;
9111 :
9112 : /*
9113 : * Make a separate ProjectionPath in case we need a Result node.
9114 : */
9115 7131 : path = (Path *) create_projection_path(root,
9116 : unique_rel,
9117 : cheapest_input_path,
9118 7131 : unique_rel->reltarget);
9119 :
9120 7131 : path = (Path *) create_agg_path(root,
9121 : unique_rel,
9122 : path,
9123 : cheapest_input_path->pathtarget,
9124 : AGG_HASHED,
9125 : AGGSPLIT_SIMPLE,
9126 : groupClause,
9127 : NIL,
9128 : NULL,
9129 : unique_rel->rows);
9130 :
9131 7131 : add_path(unique_rel, path);
9132 : }
9133 7131 : }
9134 :
9135 : /*
9136 : * create_partial_unique_paths
9137 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
9138 : */
9139 : static void
9140 3999 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
9141 : List *sortPathkeys, List *groupClause,
9142 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
9143 : {
9144 : RelOptInfo *partial_unique_rel;
9145 : Path *cheapest_partial_path;
9146 :
9147 : /* nothing to do when there are no partial paths in the input rel */
9148 3999 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
9149 867 : return;
9150 :
9151 : /*
9152 : * nothing to do if there's anything in the targetlist that's
9153 : * parallel-restricted.
9154 : */
9155 3132 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
9156 0 : return;
9157 :
9158 3132 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
9159 :
9160 3132 : partial_unique_rel = makeNode(RelOptInfo);
9161 3132 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
9162 :
9163 : /*
9164 : * clear path info
9165 : */
9166 3132 : partial_unique_rel->pathlist = NIL;
9167 3132 : partial_unique_rel->ppilist = NIL;
9168 3132 : partial_unique_rel->partial_pathlist = NIL;
9169 3132 : partial_unique_rel->cheapest_startup_path = NULL;
9170 3132 : partial_unique_rel->cheapest_total_path = NULL;
9171 3132 : partial_unique_rel->cheapest_parameterized_paths = NIL;
9172 :
9173 : /* Estimate number of output rows */
9174 3132 : partial_unique_rel->rows = estimate_num_groups(root,
9175 : sjinfo->semi_rhs_exprs,
9176 : cheapest_partial_path->rows,
9177 : NULL,
9178 : NULL);
9179 3132 : partial_unique_rel->reltarget = unique_rel->reltarget;
9180 :
9181 : /* Consider sort-based implementations, if possible. */
9182 3132 : if (sjinfo->semi_can_btree)
9183 : {
9184 : ListCell *lc;
9185 :
9186 : /*
9187 : * Use any available suitably-sorted path as input, and also consider
9188 : * sorting the cheapest partial path and incremental sort on any paths
9189 : * with presorted keys.
9190 : */
9191 6518 : foreach(lc, input_rel->partial_pathlist)
9192 : {
9193 3386 : Path *input_path = (Path *) lfirst(lc);
9194 : Path *path;
9195 : bool is_sorted;
9196 : int presorted_keys;
9197 :
9198 3386 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
9199 : input_path->pathkeys,
9200 : &presorted_keys);
9201 :
9202 : /*
9203 : * Ignore paths that are not suitably or partially sorted, unless
9204 : * they are the cheapest partial path (no need to deal with paths
9205 : * which have presorted keys when incremental sort is disabled).
9206 : */
9207 3386 : if (!is_sorted && input_path != cheapest_partial_path &&
9208 0 : (presorted_keys == 0 || !enable_incremental_sort))
9209 0 : continue;
9210 :
9211 : /*
9212 : * Make a separate ProjectionPath in case we need a Result node.
9213 : */
9214 3386 : path = (Path *) create_projection_path(root,
9215 : partial_unique_rel,
9216 : input_path,
9217 3386 : partial_unique_rel->reltarget);
9218 :
9219 3386 : if (!is_sorted)
9220 : {
9221 : /*
9222 : * We've no need to consider both a sort and incremental sort.
9223 : * We'll just do a sort if there are no presorted keys and an
9224 : * incremental sort when there are presorted keys.
9225 : */
9226 3092 : if (presorted_keys == 0 || !enable_incremental_sort)
9227 3092 : path = (Path *) create_sort_path(root,
9228 : partial_unique_rel,
9229 : path,
9230 : sortPathkeys,
9231 : -1.0);
9232 : else
9233 0 : path = (Path *) create_incremental_sort_path(root,
9234 : partial_unique_rel,
9235 : path,
9236 : sortPathkeys,
9237 : presorted_keys,
9238 : -1.0);
9239 : }
9240 :
9241 3386 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
9242 : list_length(sortPathkeys),
9243 : partial_unique_rel->rows);
9244 :
9245 3386 : add_partial_path(partial_unique_rel, path);
9246 : }
9247 : }
9248 :
9249 : /* Consider hash-based implementation, if possible. */
9250 3132 : if (sjinfo->semi_can_hash)
9251 : {
9252 : Path *path;
9253 :
9254 : /*
9255 : * Make a separate ProjectionPath in case we need a Result node.
9256 : */
9257 3132 : path = (Path *) create_projection_path(root,
9258 : partial_unique_rel,
9259 : cheapest_partial_path,
9260 3132 : partial_unique_rel->reltarget);
9261 :
9262 3132 : path = (Path *) create_agg_path(root,
9263 : partial_unique_rel,
9264 : path,
9265 : cheapest_partial_path->pathtarget,
9266 : AGG_HASHED,
9267 : AGGSPLIT_SIMPLE,
9268 : groupClause,
9269 : NIL,
9270 : NULL,
9271 : partial_unique_rel->rows);
9272 :
9273 3132 : add_partial_path(partial_unique_rel, path);
9274 : }
9275 :
9276 3132 : if (partial_unique_rel->partial_pathlist != NIL)
9277 : {
9278 3132 : generate_useful_gather_paths(root, partial_unique_rel, true);
9279 3132 : set_cheapest(partial_unique_rel);
9280 :
9281 : /*
9282 : * Finally, create paths to unique-ify the final result. This step is
9283 : * needed to remove any duplicates due to combining rows from parallel
9284 : * workers.
9285 : */
9286 3132 : create_final_unique_paths(root, partial_unique_rel,
9287 : sortPathkeys, groupClause,
9288 : sjinfo, unique_rel);
9289 : }
9290 : }
9291 :
9292 : /*
9293 : * Choose a unique name for some subroot.
9294 : *
9295 : * Modifies glob->subplanNames to track names already used.
9296 : */
9297 : char *
9298 62496 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
9299 : {
9300 : unsigned n;
9301 :
9302 : /*
9303 : * If a numeric suffix is not required, then search the list of
9304 : * previously-assigned names for a match. If none is found, then we can
9305 : * use the provided name without modification.
9306 : */
9307 62496 : if (!always_number)
9308 : {
9309 18506 : bool found = false;
9310 :
9311 45076 : foreach_ptr(char, subplan_name, glob->subplanNames)
9312 : {
9313 11592 : if (strcmp(subplan_name, name) == 0)
9314 : {
9315 3528 : found = true;
9316 3528 : break;
9317 : }
9318 : }
9319 :
9320 18506 : if (!found)
9321 : {
9322 : /* pstrdup here is just to avoid cast-away-const */
9323 14978 : char *chosen_name = pstrdup(name);
9324 :
9325 14978 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
9326 14978 : return chosen_name;
9327 : }
9328 : }
9329 :
9330 : /*
9331 : * If a numeric suffix is required or if the un-suffixed name is already
9332 : * in use, then loop until we find a positive integer that produces a
9333 : * novel name.
9334 : */
9335 47518 : for (n = 1; true; ++n)
9336 37634 : {
9337 85152 : char *proposed_name = psprintf("%s_%u", name, n);
9338 85152 : bool found = false;
9339 :
9340 304725 : foreach_ptr(char, subplan_name, glob->subplanNames)
9341 : {
9342 172055 : if (strcmp(subplan_name, proposed_name) == 0)
9343 : {
9344 37634 : found = true;
9345 37634 : break;
9346 : }
9347 : }
9348 :
9349 85152 : if (!found)
9350 : {
9351 47518 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9352 47518 : return proposed_name;
9353 : }
9354 :
9355 37634 : pfree(proposed_name);
9356 : }
9357 : }
|
