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 337751 : planner(Query *parse, const char *query_string, int cursorOptions,
334 : ParamListInfo boundParams, ExplainState *es)
335 : {
336 : PlannedStmt *result;
337 :
338 337751 : if (planner_hook)
339 50519 : result = (*planner_hook) (parse, query_string, cursorOptions,
340 : boundParams, es);
341 : else
342 287232 : result = standard_planner(parse, query_string, cursorOptions,
343 : boundParams, es);
344 :
345 333765 : pgstat_report_plan_id(result->planId, false);
346 :
347 333765 : return result;
348 : }
349 :
350 : PlannedStmt *
351 337751 : 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 337751 : glob = makeNode(PlannerGlobal);
372 :
373 337751 : glob->boundParams = boundParams;
374 337751 : glob->subplans = NIL;
375 337751 : glob->subpaths = NIL;
376 337751 : glob->subroots = NIL;
377 337751 : glob->rewindPlanIDs = NULL;
378 337751 : glob->finalrtable = NIL;
379 337751 : glob->allRelids = NULL;
380 337751 : glob->prunableRelids = NULL;
381 337751 : glob->finalrteperminfos = NIL;
382 337751 : glob->finalrowmarks = NIL;
383 337751 : glob->resultRelations = NIL;
384 337751 : glob->appendRelations = NIL;
385 337751 : glob->partPruneInfos = NIL;
386 337751 : glob->relationOids = NIL;
387 337751 : glob->invalItems = NIL;
388 337751 : glob->paramExecTypes = NIL;
389 337751 : glob->lastPHId = 0;
390 337751 : glob->lastRowMarkId = 0;
391 337751 : glob->lastPlanNodeId = 0;
392 337751 : glob->transientPlan = false;
393 337751 : glob->dependsOnRole = false;
394 337751 : glob->partition_directory = NULL;
395 337751 : 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 337751 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
419 319564 : IsUnderPostmaster &&
420 319564 : parse->commandType == CMD_SELECT &&
421 258143 : !parse->hasModifyingCTE &&
422 258020 : max_parallel_workers_per_gather > 0 &&
423 257328 : !IsParallelWorker())
424 : {
425 : /* all the cheap tests pass, so scan the query tree */
426 257288 : glob->maxParallelHazard = max_parallel_hazard(parse);
427 257288 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
428 : }
429 : else
430 : {
431 : /* skip the query tree scan, just assume it's unsafe */
432 80463 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
433 80463 : 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 560093 : glob->parallelModeNeeded = glob->parallelModeOK &&
454 222342 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
455 :
456 : /* Determine what fraction of the plan is likely to be scanned */
457 337751 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
458 : {
459 : /*
460 : * We have no real idea how many tuples the user will ultimately FETCH
461 : * from a cursor, but it is often the case that he doesn't want 'em
462 : * all, or would prefer a fast-start plan anyway so that he can
463 : * process some of the tuples sooner. Use a GUC parameter to decide
464 : * what fraction to optimize for.
465 : */
466 3237 : tuple_fraction = cursor_tuple_fraction;
467 :
468 : /*
469 : * We document cursor_tuple_fraction as simply being a fraction, which
470 : * means the edge cases 0 and 1 have to be treated specially here. We
471 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
472 : */
473 3237 : if (tuple_fraction >= 1.0)
474 0 : tuple_fraction = 0.0;
475 3237 : else if (tuple_fraction <= 0.0)
476 0 : tuple_fraction = 1e-10;
477 : }
478 : else
479 : {
480 : /* Default assumption is we need all the tuples */
481 334514 : 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 337751 : glob->default_pgs_mask = PGS_APPEND | PGS_MERGE_APPEND | PGS_FOREIGNJOIN |
499 : PGS_GATHER | PGS_CONSIDER_NONPARTIAL;
500 337751 : if (enable_tidscan)
501 337751 : glob->default_pgs_mask |= PGS_TIDSCAN;
502 337751 : if (enable_seqscan)
503 320624 : glob->default_pgs_mask |= PGS_SEQSCAN;
504 337751 : if (enable_indexscan)
505 335504 : glob->default_pgs_mask |= PGS_INDEXSCAN | PGS_INDEXONLYSCAN;
506 337751 : if (enable_indexonlyscan)
507 336352 : glob->default_pgs_mask |= PGS_CONSIDER_INDEXONLY;
508 337751 : if (enable_bitmapscan)
509 328296 : glob->default_pgs_mask |= PGS_BITMAPSCAN;
510 337751 : if (enable_mergejoin)
511 : {
512 335832 : glob->default_pgs_mask |= PGS_MERGEJOIN_PLAIN;
513 335832 : if (enable_material)
514 335765 : glob->default_pgs_mask |= PGS_MERGEJOIN_MATERIALIZE;
515 : }
516 337751 : if (enable_nestloop)
517 : {
518 337447 : glob->default_pgs_mask |= PGS_NESTLOOP_PLAIN;
519 337447 : if (enable_material)
520 337147 : glob->default_pgs_mask |= PGS_NESTLOOP_MATERIALIZE;
521 337447 : if (enable_memoize)
522 337340 : glob->default_pgs_mask |= PGS_NESTLOOP_MEMOIZE;
523 : }
524 337751 : if (enable_hashjoin)
525 335512 : glob->default_pgs_mask |= PGS_HASHJOIN;
526 337751 : if (enable_gathermerge)
527 337751 : glob->default_pgs_mask |= PGS_GATHER_MERGE;
528 337751 : if (enable_partitionwise_join)
529 2065 : glob->default_pgs_mask |= PGS_CONSIDER_PARTITIONWISE;
530 :
531 : /* Allow plugins to take control after we've initialized "glob" */
532 337751 : if (planner_setup_hook)
533 88505 : (*planner_setup_hook) (glob, parse, query_string, cursorOptions,
534 : &tuple_fraction, es);
535 :
536 : /* primary planning entry point (may recurse for subqueries) */
537 336984 : 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 334035 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
542 334035 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
543 :
544 334035 : 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 333765 : 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 333765 : 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 333765 : if (glob->paramExecTypes != NIL)
630 : {
631 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
632 148254 : forboth(lp, glob->subplans, lr, glob->subroots)
633 : {
634 33957 : Plan *subplan = (Plan *) lfirst(lp);
635 33957 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
636 :
637 33957 : SS_finalize_plan(subroot, subplan);
638 : }
639 114297 : 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 333765 : 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 367722 : forboth(lp, glob->subplans, lr, glob->subroots)
652 : {
653 33957 : Plan *subplan = (Plan *) lfirst(lp);
654 33957 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
655 :
656 33957 : lfirst(lp) = set_plan_references(subroot, subplan);
657 : }
658 :
659 : /* build the PlannedStmt result */
660 333765 : result = makeNode(PlannedStmt);
661 :
662 333765 : result->commandType = parse->commandType;
663 333765 : result->queryId = parse->queryId;
664 333765 : result->planOrigin = PLAN_STMT_STANDARD;
665 333765 : result->hasReturning = (parse->returningList != NIL);
666 333765 : result->hasModifyingCTE = parse->hasModifyingCTE;
667 333765 : result->canSetTag = parse->canSetTag;
668 333765 : result->transientPlan = glob->transientPlan;
669 333765 : result->dependsOnRole = glob->dependsOnRole;
670 333765 : result->parallelModeNeeded = glob->parallelModeNeeded;
671 333765 : result->planTree = top_plan;
672 333765 : result->partPruneInfos = glob->partPruneInfos;
673 333765 : result->rtable = glob->finalrtable;
674 667530 : result->unprunableRelids = bms_difference(glob->allRelids,
675 333765 : glob->prunableRelids);
676 333765 : result->permInfos = glob->finalrteperminfos;
677 333765 : result->subrtinfos = glob->subrtinfos;
678 333765 : result->appendRelations = glob->appendRelations;
679 333765 : result->subplans = glob->subplans;
680 333765 : result->rewindPlanIDs = glob->rewindPlanIDs;
681 333765 : 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 402905 : foreach(lc, glob->resultRelations)
688 69140 : result->resultRelationRelids = bms_add_member(result->resultRelationRelids,
689 : lfirst_int(lc));
690 344563 : foreach(lc, glob->finalrowmarks)
691 10798 : result->rowMarkRelids = bms_add_member(result->rowMarkRelids,
692 10798 : ((PlanRowMark *) lfirst(lc))->rti);
693 :
694 333765 : result->relationOids = glob->relationOids;
695 333765 : result->invalItems = glob->invalItems;
696 333765 : result->paramExecTypes = glob->paramExecTypes;
697 : /* utilityStmt should be null, but we might as well copy it */
698 333765 : result->utilityStmt = parse->utilityStmt;
699 333765 : result->elidedNodes = glob->elidedNodes;
700 333765 : result->stmt_location = parse->stmt_location;
701 333765 : result->stmt_len = parse->stmt_len;
702 :
703 333765 : result->jitFlags = PGJIT_NONE;
704 333765 : 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 333765 : if (planner_shutdown_hook)
730 86968 : (*planner_shutdown_hook) (glob, parse, query_string, result);
731 :
732 333765 : if (glob->partition_directory != NULL)
733 9168 : DestroyPartitionDirectory(glob->partition_directory);
734 :
735 333765 : 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 401402 : 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 401402 : root = makeNode(PlannerInfo);
792 401402 : root->parse = parse;
793 401402 : root->glob = glob;
794 401402 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
795 401402 : root->plan_name = plan_name;
796 401402 : if (alternative_root != NULL)
797 1331 : root->alternative_plan_name = alternative_root->plan_name;
798 : else
799 400071 : root->alternative_plan_name = plan_name;
800 401402 : root->parent_root = parent_root;
801 401402 : root->plan_params = NIL;
802 401402 : root->outer_params = NULL;
803 401402 : root->planner_cxt = CurrentMemoryContext;
804 401402 : root->init_plans = NIL;
805 401402 : root->cte_plan_ids = NIL;
806 401402 : root->multiexpr_params = NIL;
807 401402 : root->join_domains = NIL;
808 401402 : root->eq_classes = NIL;
809 401402 : root->ec_merging_done = false;
810 401402 : root->last_rinfo_serial = 0;
811 401402 : root->all_result_relids =
812 401402 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
813 401402 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
814 401402 : root->append_rel_list = NIL;
815 401402 : root->row_identity_vars = NIL;
816 401402 : root->rowMarks = NIL;
817 401402 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
818 401402 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
819 401402 : root->processed_groupClause = NIL;
820 401402 : root->processed_distinctClause = NIL;
821 401402 : root->processed_tlist = NIL;
822 401402 : root->update_colnos = NIL;
823 401402 : root->grouping_map = NULL;
824 401402 : root->minmax_aggs = NIL;
825 401402 : root->qual_security_level = 0;
826 401402 : root->hasPseudoConstantQuals = false;
827 401402 : root->hasAlternativeSubPlans = false;
828 401402 : root->placeholdersFrozen = false;
829 401402 : root->hasRecursion = hasRecursion;
830 401402 : root->assumeReplanning = false;
831 401402 : if (hasRecursion)
832 692 : root->wt_param_id = assign_special_exec_param(root);
833 : else
834 400710 : root->wt_param_id = -1;
835 401402 : 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 401402 : 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 401402 : if (parse->cteList)
849 2210 : SS_process_ctes(root);
850 :
851 : /*
852 : * If it's a MERGE command, transform the joinlist as appropriate.
853 : */
854 401398 : 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 401398 : 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 401398 : 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 401398 : if (parse->hasSubLinks)
881 31385 : 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 401398 : 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 401394 : 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 401394 : if (parse->setOperations)
904 5500 : flatten_simple_union_all(root);
905 :
906 : /*
907 : * Survey the rangetable to see what kinds of entries are present. We can
908 : * skip some later processing if relevant SQL features are not used; for
909 : * example if there are no JOIN RTEs we can avoid the expense of doing
910 : * flatten_join_alias_vars(). This must be done after we have finished
911 : * adding rangetable entries, of course. (Note: actually, processing of
912 : * inherited or partitioned rels can cause RTEs for their child tables to
913 : * get added later; but those must all be RTE_RELATION entries, so they
914 : * don't invalidate the conclusions drawn here.)
915 : */
916 401394 : root->hasJoinRTEs = false;
917 401394 : root->hasLateralRTEs = false;
918 401394 : root->group_rtindex = 0;
919 401394 : hasOuterJoins = false;
920 401394 : hasResultRTEs = false;
921 1110443 : foreach(l, parse->rtable)
922 : {
923 709049 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
924 :
925 709049 : switch (rte->rtekind)
926 : {
927 74442 : case RTE_JOIN:
928 74442 : root->hasJoinRTEs = true;
929 74442 : if (IS_OUTER_JOIN(rte->jointype))
930 37252 : hasOuterJoins = true;
931 74442 : break;
932 148002 : case RTE_RESULT:
933 148002 : hasResultRTEs = true;
934 148002 : break;
935 4274 : case RTE_GROUP:
936 : Assert(parse->hasGroupRTE);
937 4274 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
938 4274 : break;
939 482331 : default:
940 : /* No work here for other RTE types */
941 482331 : break;
942 : }
943 :
944 709049 : if (rte->lateral)
945 8263 : 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 709049 : 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 401394 : if (parse->resultRelation)
963 : {
964 65469 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
965 :
966 65469 : if (!rte->inh)
967 63153 : root->leaf_result_relids =
968 63153 : 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 1109655 : foreach(l, parse->rtable)
986 : {
987 708527 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
988 :
989 708527 : if (rte->perminfoindex != 0 &&
990 383808 : rte->relkind == RELKIND_VIEW)
991 : {
992 : RTEPermissionInfo *perminfo;
993 : bool result;
994 :
995 16524 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
996 16524 : result = ExecCheckOneRelPerms(perminfo);
997 16524 : 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 401128 : 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 401128 : 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 398495 : parse->targetList = (List *)
1023 401128 : preprocess_expression(root, (Node *) parse->targetList,
1024 : EXPRKIND_TARGET);
1025 :
1026 398495 : newWithCheckOptions = NIL;
1027 401063 : 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 398495 : parse->withCheckOptions = newWithCheckOptions;
1037 :
1038 398495 : parse->returningList = (List *)
1039 398495 : preprocess_expression(root, (Node *) parse->returningList,
1040 : EXPRKIND_TARGET);
1041 :
1042 398495 : preprocess_qual_conditions(root, (Node *) parse->jointree);
1043 :
1044 398495 : parse->havingQual = preprocess_expression(root, parse->havingQual,
1045 : EXPRKIND_QUAL);
1046 :
1047 400927 : foreach(l, parse->windowClause)
1048 : {
1049 2432 : WindowClause *wc = lfirst_node(WindowClause, l);
1050 :
1051 : /* partitionClause/orderClause are sort/group expressions */
1052 2432 : wc->startOffset = preprocess_expression(root, wc->startOffset,
1053 : EXPRKIND_LIMIT);
1054 2432 : wc->endOffset = preprocess_expression(root, wc->endOffset,
1055 : EXPRKIND_LIMIT);
1056 : }
1057 :
1058 398495 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
1059 : EXPRKIND_LIMIT);
1060 398495 : parse->limitCount = preprocess_expression(root, parse->limitCount,
1061 : EXPRKIND_LIMIT);
1062 :
1063 398495 : if (parse->onConflict)
1064 : {
1065 3578 : parse->onConflict->arbiterElems = (List *)
1066 1789 : preprocess_expression(root,
1067 1789 : (Node *) parse->onConflict->arbiterElems,
1068 : EXPRKIND_ARBITER_ELEM);
1069 3578 : parse->onConflict->arbiterWhere =
1070 1789 : preprocess_expression(root,
1071 1789 : parse->onConflict->arbiterWhere,
1072 : EXPRKIND_QUAL);
1073 3578 : parse->onConflict->onConflictSet = (List *)
1074 1789 : preprocess_expression(root,
1075 1789 : (Node *) parse->onConflict->onConflictSet,
1076 : EXPRKIND_TARGET);
1077 1789 : parse->onConflict->onConflictWhere =
1078 1789 : preprocess_expression(root,
1079 1789 : parse->onConflict->onConflictWhere,
1080 : EXPRKIND_QUAL);
1081 : /* exclRelTlist contains only Vars, so no preprocessing needed */
1082 : }
1083 :
1084 400761 : foreach(l, parse->mergeActionList)
1085 : {
1086 2266 : MergeAction *action = (MergeAction *) lfirst(l);
1087 :
1088 2266 : action->targetList = (List *)
1089 2266 : preprocess_expression(root,
1090 2266 : (Node *) action->targetList,
1091 : EXPRKIND_TARGET);
1092 2266 : action->qual =
1093 2266 : preprocess_expression(root,
1094 : (Node *) action->qual,
1095 : EXPRKIND_QUAL);
1096 : }
1097 :
1098 398495 : parse->mergeJoinCondition =
1099 398495 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1100 :
1101 398495 : root->append_rel_list = (List *)
1102 398495 : preprocess_expression(root, (Node *) root->append_rel_list,
1103 : EXPRKIND_APPINFO);
1104 :
1105 : /* Also need to preprocess expressions within RTEs */
1106 1103722 : foreach(l, parse->rtable)
1107 : {
1108 705227 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1109 : int kind;
1110 : ListCell *lcsq;
1111 :
1112 705227 : if (rte->rtekind == RTE_RELATION)
1113 : {
1114 370383 : if (rte->tablesample)
1115 180 : rte->tablesample = (TableSampleClause *)
1116 180 : preprocess_expression(root,
1117 180 : (Node *) rte->tablesample,
1118 : EXPRKIND_TABLESAMPLE);
1119 : }
1120 334844 : else if (rte->rtekind == RTE_SUBQUERY)
1121 : {
1122 : /*
1123 : * We don't want to do all preprocessing yet on the subquery's
1124 : * expressions, since that will happen when we plan it. But if it
1125 : * contains any join aliases of our level, those have to get
1126 : * expanded now, because planning of the subquery won't do it.
1127 : * That's only possible if the subquery is LATERAL.
1128 : */
1129 64045 : if (rte->lateral && root->hasJoinRTEs)
1130 1595 : rte->subquery = (Query *)
1131 1595 : flatten_join_alias_vars(root, root->parse,
1132 1595 : (Node *) rte->subquery);
1133 : }
1134 270799 : else if (rte->rtekind == RTE_FUNCTION)
1135 : {
1136 : /* Preprocess the function expression(s) fully */
1137 35447 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1138 35447 : rte->functions = (List *)
1139 35447 : preprocess_expression(root, (Node *) rte->functions, kind);
1140 : }
1141 235352 : else if (rte->rtekind == RTE_TABLEFUNC)
1142 : {
1143 : /* Preprocess the function expression(s) fully */
1144 519 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1145 519 : rte->tablefunc = (TableFunc *)
1146 519 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1147 : }
1148 234833 : else if (rte->rtekind == RTE_VALUES)
1149 : {
1150 : /* Preprocess the values lists fully */
1151 6882 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1152 6882 : rte->values_lists = (List *)
1153 6882 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1154 : }
1155 227951 : else if (rte->rtekind == RTE_GROUP)
1156 : {
1157 : /* Preprocess the groupexprs list fully */
1158 4274 : rte->groupexprs = (List *)
1159 4274 : preprocess_expression(root, (Node *) rte->groupexprs,
1160 : EXPRKIND_GROUPEXPR);
1161 : }
1162 :
1163 : /*
1164 : * Process each element of the securityQuals list as if it were a
1165 : * separate qual expression (as indeed it is). We need to do it this
1166 : * way to get proper canonicalization of AND/OR structure. Note that
1167 : * this converts each element into an implicit-AND sublist.
1168 : */
1169 708032 : foreach(lcsq, rte->securityQuals)
1170 : {
1171 2805 : lfirst(lcsq) = preprocess_expression(root,
1172 2805 : (Node *) lfirst(lcsq),
1173 : EXPRKIND_QUAL);
1174 : }
1175 : }
1176 :
1177 : /*
1178 : * Now that we are done preprocessing expressions, and in particular done
1179 : * flattening join alias variables, get rid of the joinaliasvars lists.
1180 : * They no longer match what expressions in the rest of the tree look
1181 : * like, because we have not preprocessed expressions in those lists (and
1182 : * do not want to; for example, expanding a SubLink there would result in
1183 : * a useless unreferenced subplan). Leaving them in place simply creates
1184 : * a hazard for later scans of the tree. We could try to prevent that by
1185 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1186 : * but that doesn't sound very reliable.
1187 : */
1188 398495 : if (root->hasJoinRTEs)
1189 : {
1190 259067 : foreach(l, parse->rtable)
1191 : {
1192 212310 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1193 :
1194 212310 : rte->joinaliasvars = NIL;
1195 : }
1196 : }
1197 :
1198 : /*
1199 : * Before we flatten GROUP Vars, check which HAVING clauses have collation
1200 : * conflicts. When GROUP BY uses a nondeterministic collation, values
1201 : * that are "equal" for grouping may be distinguishable under a different
1202 : * collation. If such a HAVING clause were moved to WHERE, it would
1203 : * filter individual rows before grouping, potentially eliminating some
1204 : * members of a group and thereby changing aggregate results.
1205 : *
1206 : * We do this check before flatten_group_exprs because we can easily
1207 : * identify grouping expressions by checking whether a Var references
1208 : * RTE_GROUP, and such Vars directly carry the GROUP BY collation as their
1209 : * varcollid. After flattening, these Vars are replaced by the underlying
1210 : * expressions, and we would have to match expressions in the HAVING
1211 : * clause back to grouping expressions, which is much more complex.
1212 : */
1213 398495 : if (parse->hasGroupRTE)
1214 : havingCollationConflicts =
1215 4274 : find_having_collation_conflicts(parse, root->group_rtindex);
1216 : else
1217 394221 : havingCollationConflicts = NULL;
1218 :
1219 : /*
1220 : * Replace any Vars in the subquery's targetlist and havingQual that
1221 : * reference GROUP outputs with the underlying grouping expressions.
1222 : *
1223 : * Note that we need to perform this replacement after we've preprocessed
1224 : * the grouping expressions. This is to ensure that there is only one
1225 : * instance of SubPlan for each SubLink contained within the grouping
1226 : * expressions.
1227 : */
1228 398495 : if (parse->hasGroupRTE)
1229 : {
1230 4274 : parse->targetList = (List *)
1231 4274 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1232 4274 : parse->havingQual =
1233 4274 : flatten_group_exprs(root, root->parse, parse->havingQual);
1234 : }
1235 :
1236 : /* Constant-folding might have removed all set-returning functions */
1237 398495 : if (parse->hasTargetSRFs)
1238 10375 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1239 :
1240 : /*
1241 : * If we have grouping sets, expand the groupingSets tree of this query to
1242 : * a flat list of grouping sets. We need to do this before optimizing
1243 : * HAVING, since we can't easily tell if there's an empty grouping set
1244 : * until we have this representation.
1245 : */
1246 398495 : if (parse->groupingSets)
1247 : {
1248 905 : parse->groupingSets =
1249 905 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1250 : }
1251 :
1252 : /*
1253 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1254 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1255 : * volatile functions (since a HAVING clause is supposed to be executed
1256 : * only once per group). We also can't do this if there are any grouping
1257 : * sets and the clause references any columns that are nullable by the
1258 : * grouping sets; the nulled values of those columns are not available
1259 : * before the grouping step. (The test on groupClause might seem wrong,
1260 : * but it's okay: it's just an optimization to avoid running pull_varnos
1261 : * when there cannot be any Vars in the HAVING clause.)
1262 : *
1263 : * We also cannot do this if the HAVING clause uses a different collation
1264 : * than the GROUP BY for any grouping expression whose GROUP BY collation
1265 : * is nondeterministic. This is detected before flatten_group_exprs (see
1266 : * find_having_collation_conflicts above) and recorded in the
1267 : * havingCollationConflicts bitmapset. The bitmapset indexes remain valid
1268 : * here because flatten_group_exprs uses expression_tree_mutator, which
1269 : * preserves the list length and ordering of havingQual.
1270 : *
1271 : * Also, it may be that the clause is so expensive to execute that we're
1272 : * better off doing it only once per group, despite the loss of
1273 : * selectivity. This is hard to estimate short of doing the entire
1274 : * planning process twice, so we use a heuristic: clauses containing
1275 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1276 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1277 : * instead of after.
1278 : *
1279 : * If the query has no empty grouping set then we can simply move such a
1280 : * clause into WHERE; any group that fails the clause will not be in the
1281 : * output because none of its tuples will reach the grouping or
1282 : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1283 : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1284 : * clause must be degenerate (variable-free), so we can copy it into WHERE
1285 : * so that query_planner() can use it in a gating Result node. (This could
1286 : * be done better, but it seems not worth optimizing.)
1287 : *
1288 : * Note that a HAVING clause may contain expressions that are not fully
1289 : * preprocessed. This can happen if these expressions are part of
1290 : * grouping items. In such cases, they are replaced with GROUP Vars in
1291 : * the parser and then replaced back after we're done with expression
1292 : * preprocessing on havingQual. This is not an issue if the clause
1293 : * remains in HAVING, because these expressions will be matched to lower
1294 : * target items in setrefs.c. However, if the clause is moved or copied
1295 : * into WHERE, we need to ensure that these expressions are fully
1296 : * preprocessed.
1297 : *
1298 : * Note that both havingQual and parse->jointree->quals are in
1299 : * implicitly-ANDed-list form at this point, even though they are declared
1300 : * as Node *.
1301 : */
1302 398495 : newHaving = NIL;
1303 398495 : havingIdx = 0;
1304 399735 : foreach(l, (List *) parse->havingQual)
1305 : {
1306 1240 : Node *havingclause = (Node *) lfirst(l);
1307 :
1308 1801 : if (contain_agg_clause(havingclause) ||
1309 1122 : contain_volatile_functions(havingclause) ||
1310 1122 : contain_subplans(havingclause) ||
1311 561 : bms_is_member(havingIdx, havingCollationConflicts) ||
1312 601 : (parse->groupClause && parse->groupingSets &&
1313 100 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1314 : {
1315 : /* keep it in HAVING */
1316 799 : newHaving = lappend(newHaving, havingclause);
1317 : }
1318 441 : else if (parse->groupClause &&
1319 411 : (parse->groupingSets == NIL ||
1320 40 : (List *) linitial(parse->groupingSets) != NIL))
1321 401 : {
1322 : /* There is GROUP BY, but no empty grouping set */
1323 : Node *whereclause;
1324 :
1325 : /* Preprocess the HAVING clause fully */
1326 401 : whereclause = preprocess_expression(root, havingclause,
1327 : EXPRKIND_QUAL);
1328 : /* ... and move it to WHERE */
1329 401 : parse->jointree->quals = (Node *)
1330 401 : list_concat((List *) parse->jointree->quals,
1331 : (List *) whereclause);
1332 : }
1333 : else
1334 : {
1335 : /* There is an empty grouping set (perhaps implicitly) */
1336 : Node *whereclause;
1337 :
1338 : /* Preprocess the HAVING clause fully */
1339 40 : whereclause = preprocess_expression(root, copyObject(havingclause),
1340 : EXPRKIND_QUAL);
1341 : /* ... and put a copy in WHERE */
1342 80 : parse->jointree->quals = (Node *)
1343 40 : list_concat((List *) parse->jointree->quals,
1344 : (List *) whereclause);
1345 : /* ... and also keep it in HAVING */
1346 40 : newHaving = lappend(newHaving, havingclause);
1347 : }
1348 :
1349 1240 : havingIdx++;
1350 : }
1351 398495 : parse->havingQual = (Node *) newHaving;
1352 :
1353 : /*
1354 : * If we have any outer joins, try to reduce them to plain inner joins.
1355 : * This step is most easily done after we've done expression
1356 : * preprocessing.
1357 : */
1358 398495 : if (hasOuterJoins)
1359 25825 : reduce_outer_joins(root);
1360 :
1361 : /*
1362 : * If we have any RTE_RESULT relations, see if they can be deleted from
1363 : * the jointree. We also rely on this processing to flatten single-child
1364 : * FromExprs underneath outer joins. This step is most effectively done
1365 : * after we've done expression preprocessing and outer join reduction.
1366 : */
1367 398495 : if (hasResultRTEs || hasOuterJoins)
1368 170000 : remove_useless_result_rtes(root);
1369 :
1370 : /*
1371 : * Do the main planning.
1372 : */
1373 398495 : grouping_planner(root, tuple_fraction, setops);
1374 :
1375 : /*
1376 : * Capture the set of outer-level param IDs we have access to, for use in
1377 : * extParam/allParam calculations later.
1378 : */
1379 398449 : SS_identify_outer_params(root);
1380 :
1381 : /*
1382 : * If any initPlans were created in this query level, adjust the surviving
1383 : * Paths' costs and parallel-safety flags to account for them. The
1384 : * initPlans won't actually get attached to the plan tree till
1385 : * create_plan() runs, but we must include their effects now.
1386 : */
1387 398449 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1388 398449 : SS_charge_for_initplans(root, final_rel);
1389 :
1390 : /*
1391 : * Make sure we've identified the cheapest Path for the final rel. (By
1392 : * doing this here not in grouping_planner, we include initPlan costs in
1393 : * the decision, though it's unlikely that will change anything.)
1394 : */
1395 398449 : set_cheapest(final_rel);
1396 :
1397 398449 : return root;
1398 : }
1399 :
1400 : /*
1401 : * preprocess_expression
1402 : * Do subquery_planner's preprocessing work for an expression,
1403 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1404 : * conditions), a HAVING clause, or a few other things.
1405 : */
1406 : static Node *
1407 3358590 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1408 : {
1409 : /*
1410 : * Fall out quickly if expression is empty. This occurs often enough to
1411 : * be worth checking. Note that null->null is the correct conversion for
1412 : * implicit-AND result format, too.
1413 : */
1414 3358590 : if (expr == NULL)
1415 2640867 : return NULL;
1416 :
1417 : /*
1418 : * If the query has any join RTEs, replace join alias variables with
1419 : * base-relation variables. We must do this first, since any expressions
1420 : * we may extract from the joinaliasvars lists have not been preprocessed.
1421 : * For example, if we did this after sublink processing, sublinks expanded
1422 : * out from join aliases would not get processed. But we can skip this in
1423 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1424 : * they can't contain any Vars of the current query level.
1425 : */
1426 717723 : if (root->hasJoinRTEs &&
1427 332882 : !(kind == EXPRKIND_RTFUNC ||
1428 166292 : kind == EXPRKIND_VALUES ||
1429 : kind == EXPRKIND_TABLESAMPLE ||
1430 : kind == EXPRKIND_TABLEFUNC))
1431 166277 : expr = flatten_join_alias_vars(root, root->parse, expr);
1432 :
1433 : /*
1434 : * Simplify constant expressions. For function RTEs, this was already
1435 : * done by preprocess_function_rtes. (But note we must do it again for
1436 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1437 : * un-simplified subexpressions inserted by flattening of subqueries or
1438 : * join alias variables.)
1439 : *
1440 : * Note: an essential effect of this is to convert named-argument function
1441 : * calls to positional notation and insert the current actual values of
1442 : * any default arguments for functions. To ensure that happens, we *must*
1443 : * process all expressions here. Previous PG versions sometimes skipped
1444 : * const-simplification if it didn't seem worth the trouble, but we can't
1445 : * do that anymore.
1446 : *
1447 : * Note: this also flattens nested AND and OR expressions into N-argument
1448 : * form. All processing of a qual expression after this point must be
1449 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1450 : * with AND directly under AND, nor OR directly under OR.
1451 : */
1452 717723 : if (kind != EXPRKIND_RTFUNC)
1453 686920 : expr = eval_const_expressions(root, expr);
1454 :
1455 : /*
1456 : * If it's a qual or havingQual, canonicalize it.
1457 : */
1458 715090 : if (kind == EXPRKIND_QUAL)
1459 : {
1460 264951 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1461 :
1462 : #ifdef OPTIMIZER_DEBUG
1463 : printf("After canonicalize_qual()\n");
1464 : pprint(expr);
1465 : #endif
1466 : }
1467 :
1468 : /*
1469 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1470 : * hashfuncid of any that might execute more quickly by using hash lookups
1471 : * instead of a linear search.
1472 : */
1473 715090 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1474 : {
1475 656320 : convert_saop_to_hashed_saop(expr);
1476 : }
1477 :
1478 : /* Expand SubLinks to SubPlans */
1479 715090 : if (root->parse->hasSubLinks)
1480 92128 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1481 :
1482 : /*
1483 : * XXX do not insert anything here unless you have grokked the comments in
1484 : * SS_replace_correlation_vars ...
1485 : */
1486 :
1487 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1488 715090 : if (root->query_level > 1)
1489 149701 : expr = SS_replace_correlation_vars(root, expr);
1490 :
1491 : /*
1492 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1493 : * don't want to do this before eval_const_expressions, since the latter
1494 : * would be unable to simplify a top-level AND correctly. Also,
1495 : * SS_process_sublinks expects explicit-AND format.)
1496 : */
1497 715090 : if (kind == EXPRKIND_QUAL)
1498 264951 : expr = (Node *) make_ands_implicit((Expr *) expr);
1499 :
1500 715090 : return expr;
1501 : }
1502 :
1503 : /*
1504 : * preprocess_qual_conditions
1505 : * Recursively scan the query's jointree and do subquery_planner's
1506 : * preprocessing work on each qual condition found therein.
1507 : */
1508 : static void
1509 1011806 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1510 : {
1511 1011806 : if (jtnode == NULL)
1512 0 : return;
1513 1011806 : if (IsA(jtnode, RangeTblRef))
1514 : {
1515 : /* nothing to do here */
1516 : }
1517 496749 : else if (IsA(jtnode, FromExpr))
1518 : {
1519 413473 : FromExpr *f = (FromExpr *) jtnode;
1520 : ListCell *l;
1521 :
1522 860232 : foreach(l, f->fromlist)
1523 446759 : preprocess_qual_conditions(root, lfirst(l));
1524 :
1525 413473 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1526 : }
1527 83276 : else if (IsA(jtnode, JoinExpr))
1528 : {
1529 83276 : JoinExpr *j = (JoinExpr *) jtnode;
1530 :
1531 83276 : preprocess_qual_conditions(root, j->larg);
1532 83276 : preprocess_qual_conditions(root, j->rarg);
1533 :
1534 83276 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1535 : }
1536 : else
1537 0 : elog(ERROR, "unrecognized node type: %d",
1538 : (int) nodeTag(jtnode));
1539 : }
1540 :
1541 : /*
1542 : * find_having_collation_conflicts
1543 : * Identify HAVING clauses that must not be moved to WHERE due to collation
1544 : * mismatches with GROUP BY.
1545 : *
1546 : * This must be called before flatten_group_exprs, while the HAVING clause
1547 : * still contains GROUP Vars (Vars referencing RTE_GROUP). These GROUP Vars
1548 : * carry the GROUP BY collation as their varcollid. A GROUP Var with a
1549 : * nondeterministic varcollid conflicts whenever some collation-aware ancestor
1550 : * on its path applies a different inputcollid: that operator would distinguish
1551 : * values which the GROUP BY considers equal, so the clause is unsafe to push
1552 : * to WHERE.
1553 : *
1554 : * Returns a Bitmapset of zero-based indexes into the havingQual list for
1555 : * clauses that have collation conflicts and must stay in HAVING.
1556 : */
1557 : static Bitmapset *
1558 4274 : find_having_collation_conflicts(Query *parse, Index group_rtindex)
1559 : {
1560 4274 : Bitmapset *result = NULL;
1561 : having_collation_ctx ctx;
1562 : int idx;
1563 :
1564 4274 : if (parse->havingQual == NULL)
1565 3466 : return NULL;
1566 :
1567 808 : ctx.group_rtindex = group_rtindex;
1568 808 : ctx.ancestor_collids = NIL;
1569 :
1570 808 : idx = 0;
1571 2563 : foreach_ptr(Node, clause, (List *) parse->havingQual)
1572 : {
1573 947 : if (having_collation_conflict_walker(clause, &ctx))
1574 60 : result = bms_add_member(result, idx);
1575 947 : idx++;
1576 : Assert(ctx.ancestor_collids == NIL);
1577 : }
1578 :
1579 808 : return result;
1580 : }
1581 :
1582 : /*
1583 : * Walker function for find_having_collation_conflicts.
1584 : *
1585 : * Walk the clause top-down, maintaining a stack of inputcollids contributed
1586 : * by collation-aware ancestors. At each GROUP Var with a nondeterministic
1587 : * varcollid, the clause has a conflict if any ancestor's inputcollid differs
1588 : * from the GROUP Var's varcollid. Most collation-aware nodes expose their
1589 : * inputcollid through exprInputCollation(); RowCompareExpr is the exception,
1590 : * as it carries one inputcollid per column in inputcollids[], so we descend
1591 : * into its (largs[i], rargs[i]) pairs explicitly with the matching collation
1592 : * pushed onto the stack.
1593 : */
1594 : static bool
1595 4166 : having_collation_conflict_walker(Node *node, having_collation_ctx *ctx)
1596 : {
1597 : Oid this_collid;
1598 : bool result;
1599 :
1600 4166 : if (node == NULL)
1601 443 : return false;
1602 :
1603 3723 : if (IsA(node, Var))
1604 : {
1605 898 : Var *var = (Var *) node;
1606 :
1607 : /* We should not see any upper-level Vars here */
1608 : Assert(var->varlevelsup == 0);
1609 :
1610 898 : if (var->varno == ctx->group_rtindex &&
1611 514 : OidIsValid(var->varcollid) &&
1612 331 : !get_collation_isdeterministic(var->varcollid))
1613 : {
1614 230 : foreach_oid(collid, ctx->ancestor_collids)
1615 : {
1616 130 : if (collid != var->varcollid)
1617 60 : return true;
1618 : }
1619 : }
1620 838 : return false;
1621 : }
1622 :
1623 2825 : if (IsA(node, RowCompareExpr))
1624 : {
1625 10 : RowCompareExpr *rcexpr = (RowCompareExpr *) node;
1626 : ListCell *lc_l;
1627 : ListCell *lc_r;
1628 : ListCell *lc_c;
1629 :
1630 : /*
1631 : * Each column of a row comparison is compared under its own
1632 : * inputcollids[i]. Walk each (largs[i], rargs[i]) pair with that
1633 : * collation pushed, so a Var in column i is checked against the
1634 : * collation that actually applies to it.
1635 : */
1636 10 : forthree(lc_l, rcexpr->largs,
1637 : lc_r, rcexpr->rargs,
1638 : lc_c, rcexpr->inputcollids)
1639 : {
1640 10 : Oid collid = lfirst_oid(lc_c);
1641 : bool found;
1642 :
1643 10 : if (OidIsValid(collid))
1644 10 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1645 : collid);
1646 :
1647 10 : found = having_collation_conflict_walker((Node *) lfirst(lc_l),
1648 10 : ctx) ||
1649 0 : having_collation_conflict_walker((Node *) lfirst(lc_r),
1650 : ctx);
1651 :
1652 10 : if (OidIsValid(collid))
1653 10 : ctx->ancestor_collids =
1654 10 : list_delete_last(ctx->ancestor_collids);
1655 :
1656 10 : if (found)
1657 10 : return true;
1658 : }
1659 0 : return false;
1660 : }
1661 :
1662 2815 : this_collid = exprInputCollation(node);
1663 2815 : if (OidIsValid(this_collid))
1664 356 : ctx->ancestor_collids = lappend_oid(ctx->ancestor_collids,
1665 : this_collid);
1666 :
1667 2815 : result = expression_tree_walker(node, having_collation_conflict_walker,
1668 : ctx);
1669 :
1670 2815 : if (OidIsValid(this_collid))
1671 356 : ctx->ancestor_collids = list_delete_last(ctx->ancestor_collids);
1672 :
1673 2815 : return result;
1674 : }
1675 :
1676 : /*
1677 : * preprocess_phv_expression
1678 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1679 : *
1680 : * If a LATERAL subquery references an output of another subquery, and that
1681 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1682 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1683 : * and later pull it back up during find_lateral_references, which runs after
1684 : * subquery_planner has preprocessed all the expressions that were in the
1685 : * current query level to start with. So we need to preprocess it then.
1686 : */
1687 : Expr *
1688 75 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1689 : {
1690 75 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1691 : }
1692 :
1693 : /*--------------------
1694 : * grouping_planner
1695 : * Perform planning steps related to grouping, aggregation, etc.
1696 : *
1697 : * This function adds all required top-level processing to the scan/join
1698 : * Path(s) produced by query_planner.
1699 : *
1700 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1701 : * tuple_fraction is interpreted as follows:
1702 : * 0: expect all tuples to be retrieved (normal case)
1703 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1704 : * from the plan to be retrieved
1705 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1706 : * expected to be retrieved (ie, a LIMIT specification).
1707 : * setops is used for set operation subqueries to provide the subquery with
1708 : * the context in which it's being used so that Paths correctly sorted for the
1709 : * set operation can be generated. NULL when not planning a set operation
1710 : * child, or when a child of a set op that isn't interested in sorted input.
1711 : *
1712 : * Returns nothing; the useful output is in the Paths we attach to the
1713 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1714 : * root->processed_tlist contains the final processed targetlist.
1715 : *
1716 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1717 : * to leave this to the caller.
1718 : *--------------------
1719 : */
1720 : static void
1721 398495 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1722 : SetOperationStmt *setops)
1723 : {
1724 398495 : Query *parse = root->parse;
1725 398495 : int64 offset_est = 0;
1726 398495 : int64 count_est = 0;
1727 398495 : double limit_tuples = -1.0;
1728 398495 : bool have_postponed_srfs = false;
1729 : PathTarget *final_target;
1730 : List *final_targets;
1731 : List *final_targets_contain_srfs;
1732 : bool final_target_parallel_safe;
1733 : RelOptInfo *current_rel;
1734 : RelOptInfo *final_rel;
1735 : FinalPathExtraData extra;
1736 : ListCell *lc;
1737 :
1738 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1739 398495 : if (parse->limitCount || parse->limitOffset)
1740 : {
1741 3714 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1742 : &offset_est, &count_est);
1743 :
1744 : /*
1745 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1746 : * estimate the effects of using a bounded sort.
1747 : */
1748 3714 : if (count_est > 0 && offset_est >= 0)
1749 3252 : limit_tuples = (double) count_est + (double) offset_est;
1750 : }
1751 :
1752 : /* Make tuple_fraction accessible to lower-level routines */
1753 398495 : root->tuple_fraction = tuple_fraction;
1754 :
1755 398495 : if (parse->setOperations)
1756 : {
1757 : /*
1758 : * Construct Paths for set operations. The results will not need any
1759 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1760 : * special work for recursive unions is the responsibility of
1761 : * plan_set_operations.
1762 : */
1763 4956 : current_rel = plan_set_operations(root);
1764 :
1765 : /*
1766 : * We should not need to call preprocess_targetlist, since we must be
1767 : * in a SELECT query node. Instead, use the processed_tlist returned
1768 : * by plan_set_operations (since this tells whether it returned any
1769 : * resjunk columns!), and transfer any sort key information from the
1770 : * original tlist.
1771 : */
1772 : Assert(parse->commandType == CMD_SELECT);
1773 :
1774 : /* for safety, copy processed_tlist instead of modifying in-place */
1775 4952 : root->processed_tlist =
1776 4952 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1777 : parse->targetList);
1778 :
1779 : /* Also extract the PathTarget form of the setop result tlist */
1780 4952 : final_target = current_rel->cheapest_total_path->pathtarget;
1781 :
1782 : /* And check whether it's parallel safe */
1783 : final_target_parallel_safe =
1784 4952 : is_parallel_safe(root, (Node *) final_target->exprs);
1785 :
1786 : /* The setop result tlist couldn't contain any SRFs */
1787 : Assert(!parse->hasTargetSRFs);
1788 4952 : final_targets = final_targets_contain_srfs = NIL;
1789 :
1790 : /*
1791 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1792 : * checked already, but let's make sure).
1793 : */
1794 4952 : if (parse->rowMarks)
1795 0 : ereport(ERROR,
1796 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1797 : /*------
1798 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1799 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1800 : LCS_asString(linitial_node(RowMarkClause,
1801 : parse->rowMarks)->strength))));
1802 :
1803 : /*
1804 : * Calculate pathkeys that represent result ordering requirements
1805 : */
1806 : Assert(parse->distinctClause == NIL);
1807 4952 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1808 : parse->sortClause,
1809 : root->processed_tlist);
1810 : }
1811 : else
1812 : {
1813 : /* No set operations, do regular planning */
1814 : PathTarget *sort_input_target;
1815 : List *sort_input_targets;
1816 : List *sort_input_targets_contain_srfs;
1817 : bool sort_input_target_parallel_safe;
1818 : PathTarget *grouping_target;
1819 : List *grouping_targets;
1820 : List *grouping_targets_contain_srfs;
1821 : bool grouping_target_parallel_safe;
1822 : PathTarget *scanjoin_target;
1823 : List *scanjoin_targets;
1824 : List *scanjoin_targets_contain_srfs;
1825 : bool scanjoin_target_parallel_safe;
1826 : bool scanjoin_target_same_exprs;
1827 : bool have_grouping;
1828 393539 : WindowFuncLists *wflists = NULL;
1829 393539 : List *activeWindows = NIL;
1830 393539 : grouping_sets_data *gset_data = NULL;
1831 : standard_qp_extra qp_extra;
1832 :
1833 : /* A recursive query should always have setOperations */
1834 : Assert(!root->hasRecursion);
1835 :
1836 : /* Preprocess grouping sets and GROUP BY clause, if any */
1837 393539 : if (parse->groupingSets)
1838 : {
1839 905 : gset_data = preprocess_grouping_sets(root);
1840 : }
1841 392634 : else if (parse->groupClause)
1842 : {
1843 : /* Preprocess regular GROUP BY clause, if any */
1844 3434 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1845 : }
1846 :
1847 : /*
1848 : * Preprocess targetlist. Note that much of the remaining planning
1849 : * work will be done with the PathTarget representation of tlists, but
1850 : * we must also maintain the full representation of the final tlist so
1851 : * that we can transfer its decoration (resnames etc) to the topmost
1852 : * tlist of the finished Plan. This is kept in processed_tlist.
1853 : */
1854 393535 : preprocess_targetlist(root);
1855 :
1856 : /*
1857 : * Mark all the aggregates with resolved aggtranstypes, and detect
1858 : * aggregates that are duplicates or can share transition state. We
1859 : * must do this before slicing and dicing the tlist into various
1860 : * pathtargets, else some copies of the Aggref nodes might escape
1861 : * being marked.
1862 : */
1863 393535 : if (parse->hasAggs)
1864 : {
1865 33296 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1866 33296 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1867 : }
1868 :
1869 : /*
1870 : * Locate any window functions in the tlist. (We don't need to look
1871 : * anywhere else, since expressions used in ORDER BY will be in there
1872 : * too.) Note that they could all have been eliminated by constant
1873 : * folding, in which case we don't need to do any more work.
1874 : */
1875 393535 : if (parse->hasWindowFuncs)
1876 : {
1877 2222 : wflists = find_window_functions((Node *) root->processed_tlist,
1878 2222 : list_length(parse->windowClause));
1879 2222 : if (wflists->numWindowFuncs > 0)
1880 : {
1881 : /*
1882 : * See if any modifications can be made to each WindowClause
1883 : * to allow the executor to execute the WindowFuncs more
1884 : * quickly.
1885 : */
1886 2217 : optimize_window_clauses(root, wflists);
1887 :
1888 : /* Extract the list of windows actually in use. */
1889 2217 : activeWindows = select_active_windows(root, wflists);
1890 :
1891 : /* Make sure they all have names, for EXPLAIN's use. */
1892 2217 : name_active_windows(activeWindows);
1893 : }
1894 : else
1895 5 : parse->hasWindowFuncs = false;
1896 : }
1897 :
1898 : /*
1899 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1900 : * adding logic between here and the query_planner() call. Anything
1901 : * that is needed in MIN/MAX-optimizable cases will have to be
1902 : * duplicated in planagg.c.
1903 : */
1904 393535 : if (parse->hasAggs)
1905 33296 : preprocess_minmax_aggregates(root);
1906 :
1907 : /*
1908 : * Figure out whether there's a hard limit on the number of rows that
1909 : * query_planner's result subplan needs to return. Even if we know a
1910 : * hard limit overall, it doesn't apply if the query has any
1911 : * grouping/aggregation operations, or SRFs in the tlist.
1912 : */
1913 393535 : if (parse->groupClause ||
1914 389265 : parse->groupingSets ||
1915 389200 : parse->distinctClause ||
1916 387263 : parse->hasAggs ||
1917 357639 : parse->hasWindowFuncs ||
1918 355538 : parse->hasTargetSRFs ||
1919 345497 : root->hasHavingQual)
1920 48058 : root->limit_tuples = -1.0;
1921 : else
1922 345477 : root->limit_tuples = limit_tuples;
1923 :
1924 : /* Set up data needed by standard_qp_callback */
1925 393535 : qp_extra.activeWindows = activeWindows;
1926 393535 : qp_extra.gset_data = gset_data;
1927 :
1928 : /*
1929 : * If we're a subquery for a set operation, store the SetOperationStmt
1930 : * in qp_extra.
1931 : */
1932 393535 : qp_extra.setop = setops;
1933 :
1934 : /*
1935 : * Generate the best unsorted and presorted paths for the scan/join
1936 : * portion of this Query, ie the processing represented by the
1937 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1938 : * We also generate (in standard_qp_callback) pathkey representations
1939 : * of the query's sort clause, distinct clause, etc.
1940 : */
1941 393535 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1942 :
1943 : /*
1944 : * Convert the query's result tlist into PathTarget format.
1945 : *
1946 : * Note: this cannot be done before query_planner() has performed
1947 : * appendrel expansion, because that might add resjunk entries to
1948 : * root->processed_tlist. Waiting till afterwards is also helpful
1949 : * because the target width estimates can use per-Var width numbers
1950 : * that were obtained within query_planner().
1951 : */
1952 393501 : final_target = create_pathtarget(root, root->processed_tlist);
1953 : final_target_parallel_safe =
1954 393501 : is_parallel_safe(root, (Node *) final_target->exprs);
1955 :
1956 : /*
1957 : * If ORDER BY was given, consider whether we should use a post-sort
1958 : * projection, and compute the adjusted target for preceding steps if
1959 : * so.
1960 : */
1961 393501 : if (parse->sortClause)
1962 : {
1963 58879 : sort_input_target = make_sort_input_target(root,
1964 : final_target,
1965 : &have_postponed_srfs);
1966 : sort_input_target_parallel_safe =
1967 58879 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1968 : }
1969 : else
1970 : {
1971 334622 : sort_input_target = final_target;
1972 334622 : sort_input_target_parallel_safe = final_target_parallel_safe;
1973 : }
1974 :
1975 : /*
1976 : * If we have window functions to deal with, the output from any
1977 : * grouping step needs to be what the window functions want;
1978 : * otherwise, it should be sort_input_target.
1979 : */
1980 393501 : if (activeWindows)
1981 : {
1982 2217 : grouping_target = make_window_input_target(root,
1983 : final_target,
1984 : activeWindows);
1985 : grouping_target_parallel_safe =
1986 2217 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1987 : }
1988 : else
1989 : {
1990 391284 : grouping_target = sort_input_target;
1991 391284 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1992 : }
1993 :
1994 : /*
1995 : * If we have grouping or aggregation to do, the topmost scan/join
1996 : * plan node must emit what the grouping step wants; otherwise, it
1997 : * should emit grouping_target.
1998 : */
1999 389231 : have_grouping = (parse->groupClause || parse->groupingSets ||
2000 782732 : parse->hasAggs || root->hasHavingQual);
2001 393501 : if (have_grouping)
2002 : {
2003 34010 : scanjoin_target = make_group_input_target(root, final_target);
2004 : scanjoin_target_parallel_safe =
2005 34010 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2006 : }
2007 : else
2008 : {
2009 359491 : scanjoin_target = grouping_target;
2010 359491 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2011 : }
2012 :
2013 : /*
2014 : * If there are any SRFs in the targetlist, we must separate each of
2015 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2016 : * each of the named targets with a SRF-free version, and remember the
2017 : * list of additional projection steps we need to add afterwards.
2018 : */
2019 393501 : if (parse->hasTargetSRFs)
2020 : {
2021 : /* final_target doesn't recompute any SRFs in sort_input_target */
2022 10375 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2023 : &final_targets,
2024 : &final_targets_contain_srfs);
2025 10375 : final_target = linitial_node(PathTarget, final_targets);
2026 : Assert(!linitial_int(final_targets_contain_srfs));
2027 : /* likewise for sort_input_target vs. grouping_target */
2028 10375 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2029 : &sort_input_targets,
2030 : &sort_input_targets_contain_srfs);
2031 10375 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2032 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2033 : /* likewise for grouping_target vs. scanjoin_target */
2034 10375 : split_pathtarget_at_srfs_grouping(root,
2035 : grouping_target, scanjoin_target,
2036 : &grouping_targets,
2037 : &grouping_targets_contain_srfs);
2038 10375 : grouping_target = linitial_node(PathTarget, grouping_targets);
2039 : Assert(!linitial_int(grouping_targets_contain_srfs));
2040 : /* scanjoin_target will not have any SRFs precomputed for it */
2041 10375 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2042 : &scanjoin_targets,
2043 : &scanjoin_targets_contain_srfs);
2044 10375 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2045 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2046 : }
2047 : else
2048 : {
2049 : /* initialize lists; for most of these, dummy values are OK */
2050 383126 : final_targets = final_targets_contain_srfs = NIL;
2051 383126 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2052 383126 : grouping_targets = grouping_targets_contain_srfs = NIL;
2053 383126 : scanjoin_targets = list_make1(scanjoin_target);
2054 383126 : scanjoin_targets_contain_srfs = NIL;
2055 : }
2056 :
2057 : /* Apply scan/join target. */
2058 393501 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2059 393501 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2060 393501 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2061 : scanjoin_targets_contain_srfs,
2062 : scanjoin_target_parallel_safe,
2063 : scanjoin_target_same_exprs);
2064 :
2065 : /*
2066 : * Save the various upper-rel PathTargets we just computed into
2067 : * root->upper_targets[]. The core code doesn't use this, but it
2068 : * provides a convenient place for extensions to get at the info. For
2069 : * consistency, we save all the intermediate targets, even though some
2070 : * of the corresponding upperrels might not be needed for this query.
2071 : */
2072 393501 : root->upper_targets[UPPERREL_FINAL] = final_target;
2073 393501 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2074 393501 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2075 393501 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
2076 393501 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2077 393501 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2078 :
2079 : /*
2080 : * If we have grouping and/or aggregation, consider ways to implement
2081 : * that. We build a new upperrel representing the output of this
2082 : * phase.
2083 : */
2084 393501 : if (have_grouping)
2085 : {
2086 34010 : current_rel = create_grouping_paths(root,
2087 : current_rel,
2088 : grouping_target,
2089 : grouping_target_parallel_safe,
2090 : gset_data);
2091 : /* Fix things up if grouping_target contains SRFs */
2092 34006 : if (parse->hasTargetSRFs)
2093 299 : adjust_paths_for_srfs(root, current_rel,
2094 : grouping_targets,
2095 : grouping_targets_contain_srfs);
2096 : }
2097 :
2098 : /*
2099 : * If we have window functions, consider ways to implement those. We
2100 : * build a new upperrel representing the output of this phase.
2101 : */
2102 393497 : if (activeWindows)
2103 : {
2104 2217 : current_rel = create_window_paths(root,
2105 : current_rel,
2106 : grouping_target,
2107 : sort_input_target,
2108 : sort_input_target_parallel_safe,
2109 : wflists,
2110 : activeWindows);
2111 : /* Fix things up if sort_input_target contains SRFs */
2112 2217 : if (parse->hasTargetSRFs)
2113 10 : adjust_paths_for_srfs(root, current_rel,
2114 : sort_input_targets,
2115 : sort_input_targets_contain_srfs);
2116 : }
2117 :
2118 : /*
2119 : * If there is a DISTINCT clause, consider ways to implement that. We
2120 : * build a new upperrel representing the output of this phase.
2121 : */
2122 393497 : if (parse->distinctClause)
2123 : {
2124 1964 : current_rel = create_distinct_paths(root,
2125 : current_rel,
2126 : sort_input_target);
2127 : }
2128 : } /* end of if (setOperations) */
2129 :
2130 : /*
2131 : * If ORDER BY was given, consider ways to implement that, and generate a
2132 : * new upperrel containing only paths that emit the correct ordering and
2133 : * project the correct final_target. We can apply the original
2134 : * limit_tuples limit in sort costing here, but only if there are no
2135 : * postponed SRFs.
2136 : */
2137 398449 : if (parse->sortClause)
2138 : {
2139 62146 : current_rel = create_ordered_paths(root,
2140 : current_rel,
2141 : final_target,
2142 : final_target_parallel_safe,
2143 : have_postponed_srfs ? -1.0 :
2144 : limit_tuples);
2145 : /* Fix things up if final_target contains SRFs */
2146 62146 : if (parse->hasTargetSRFs)
2147 182 : adjust_paths_for_srfs(root, current_rel,
2148 : final_targets,
2149 : final_targets_contain_srfs);
2150 : }
2151 :
2152 : /*
2153 : * Now we are prepared to build the final-output upperrel.
2154 : */
2155 398449 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2156 :
2157 : /*
2158 : * If the input rel is marked consider_parallel and there's nothing that's
2159 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2160 : * consider_parallel as well. Note that if the query has rowMarks or is
2161 : * not a SELECT, consider_parallel will be false for every relation in the
2162 : * query.
2163 : */
2164 537761 : if (current_rel->consider_parallel &&
2165 278604 : is_parallel_safe(root, parse->limitOffset) &&
2166 139292 : is_parallel_safe(root, parse->limitCount))
2167 139287 : final_rel->consider_parallel = true;
2168 :
2169 : /*
2170 : * If the current_rel belongs to a single FDW, so does the final_rel.
2171 : */
2172 398449 : final_rel->serverid = current_rel->serverid;
2173 398449 : final_rel->userid = current_rel->userid;
2174 398449 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2175 398449 : final_rel->fdwroutine = current_rel->fdwroutine;
2176 :
2177 : /*
2178 : * Generate paths for the final_rel. Insert all surviving paths, with
2179 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2180 : */
2181 813415 : foreach(lc, current_rel->pathlist)
2182 : {
2183 414966 : Path *path = (Path *) lfirst(lc);
2184 :
2185 : /*
2186 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2187 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2188 : * here. If there are only non-locking rowmarks, they should be
2189 : * handled by the ModifyTable node instead. However, root->rowMarks
2190 : * is what goes into the LockRows node.)
2191 : */
2192 414966 : if (parse->rowMarks)
2193 : {
2194 6811 : path = (Path *) create_lockrows_path(root, final_rel, path,
2195 : root->rowMarks,
2196 : assign_special_exec_param(root));
2197 : }
2198 :
2199 : /*
2200 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2201 : */
2202 414966 : if (limit_needed(parse))
2203 : {
2204 4189 : path = (Path *) create_limit_path(root, final_rel, path,
2205 : parse->limitOffset,
2206 : parse->limitCount,
2207 : parse->limitOption,
2208 : offset_est, count_est);
2209 : }
2210 :
2211 : /*
2212 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
2213 : */
2214 414966 : if (parse->commandType != CMD_SELECT)
2215 : {
2216 : Index rootRelation;
2217 65115 : List *resultRelations = NIL;
2218 65115 : List *updateColnosLists = NIL;
2219 65115 : List *withCheckOptionLists = NIL;
2220 65115 : List *returningLists = NIL;
2221 65115 : List *mergeActionLists = NIL;
2222 65115 : List *mergeJoinConditions = NIL;
2223 : List *rowMarks;
2224 :
2225 65115 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
2226 : {
2227 : /* Inherited UPDATE/DELETE/MERGE */
2228 2288 : RelOptInfo *top_result_rel = find_base_rel(root,
2229 : parse->resultRelation);
2230 2288 : int resultRelation = -1;
2231 :
2232 : /* Pass the root result rel forward to the executor. */
2233 2288 : rootRelation = parse->resultRelation;
2234 :
2235 : /* Add only leaf children to ModifyTable. */
2236 6718 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
2237 6718 : resultRelation)) >= 0)
2238 : {
2239 4430 : RelOptInfo *this_result_rel = find_base_rel(root,
2240 : resultRelation);
2241 :
2242 : /*
2243 : * Also exclude any leaf rels that have turned dummy since
2244 : * being added to the list, for example, by being excluded
2245 : * by constraint exclusion.
2246 : */
2247 4430 : if (IS_DUMMY_REL(this_result_rel))
2248 153 : continue;
2249 :
2250 : /* Build per-target-rel lists needed by ModifyTable */
2251 4277 : resultRelations = lappend_int(resultRelations,
2252 : resultRelation);
2253 4277 : if (parse->commandType == CMD_UPDATE)
2254 : {
2255 2933 : List *update_colnos = root->update_colnos;
2256 :
2257 2933 : if (this_result_rel != top_result_rel)
2258 : update_colnos =
2259 2933 : adjust_inherited_attnums_multilevel(root,
2260 : update_colnos,
2261 : this_result_rel->relid,
2262 : top_result_rel->relid);
2263 2933 : updateColnosLists = lappend(updateColnosLists,
2264 : update_colnos);
2265 : }
2266 4277 : if (parse->withCheckOptions)
2267 : {
2268 416 : List *withCheckOptions = parse->withCheckOptions;
2269 :
2270 416 : if (this_result_rel != top_result_rel)
2271 : withCheckOptions = (List *)
2272 416 : adjust_appendrel_attrs_multilevel(root,
2273 : (Node *) withCheckOptions,
2274 : this_result_rel,
2275 : top_result_rel);
2276 416 : withCheckOptionLists = lappend(withCheckOptionLists,
2277 : withCheckOptions);
2278 : }
2279 4277 : if (parse->returningList)
2280 : {
2281 645 : List *returningList = parse->returningList;
2282 :
2283 645 : if (this_result_rel != top_result_rel)
2284 : returningList = (List *)
2285 645 : adjust_appendrel_attrs_multilevel(root,
2286 : (Node *) returningList,
2287 : this_result_rel,
2288 : top_result_rel);
2289 645 : returningLists = lappend(returningLists,
2290 : returningList);
2291 : }
2292 4277 : if (parse->mergeActionList)
2293 : {
2294 : ListCell *l;
2295 425 : List *mergeActionList = NIL;
2296 :
2297 : /*
2298 : * Copy MergeActions and translate stuff that
2299 : * references attribute numbers.
2300 : */
2301 1322 : foreach(l, parse->mergeActionList)
2302 : {
2303 897 : MergeAction *action = lfirst(l),
2304 897 : *leaf_action = copyObject(action);
2305 :
2306 897 : leaf_action->qual =
2307 897 : adjust_appendrel_attrs_multilevel(root,
2308 : (Node *) action->qual,
2309 : this_result_rel,
2310 : top_result_rel);
2311 897 : leaf_action->targetList = (List *)
2312 897 : adjust_appendrel_attrs_multilevel(root,
2313 897 : (Node *) action->targetList,
2314 : this_result_rel,
2315 : top_result_rel);
2316 897 : if (leaf_action->commandType == CMD_UPDATE)
2317 491 : leaf_action->updateColnos =
2318 491 : adjust_inherited_attnums_multilevel(root,
2319 : action->updateColnos,
2320 : this_result_rel->relid,
2321 : top_result_rel->relid);
2322 897 : mergeActionList = lappend(mergeActionList,
2323 : leaf_action);
2324 : }
2325 :
2326 425 : mergeActionLists = lappend(mergeActionLists,
2327 : mergeActionList);
2328 : }
2329 4277 : if (parse->commandType == CMD_MERGE)
2330 : {
2331 425 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2332 :
2333 425 : if (this_result_rel != top_result_rel)
2334 : mergeJoinCondition =
2335 425 : adjust_appendrel_attrs_multilevel(root,
2336 : mergeJoinCondition,
2337 : this_result_rel,
2338 : top_result_rel);
2339 425 : mergeJoinConditions = lappend(mergeJoinConditions,
2340 : mergeJoinCondition);
2341 : }
2342 : }
2343 :
2344 2288 : if (resultRelations == NIL)
2345 : {
2346 : /*
2347 : * We managed to exclude every child rel, so generate a
2348 : * dummy one-relation plan using info for the top target
2349 : * rel (even though that may not be a leaf target).
2350 : * Although it's clear that no data will be updated or
2351 : * deleted, we still need to have a ModifyTable node so
2352 : * that any statement triggers will be executed. (This
2353 : * could be cleaner if we fixed nodeModifyTable.c to allow
2354 : * zero target relations, but that probably wouldn't be a
2355 : * net win.)
2356 : */
2357 28 : resultRelations = list_make1_int(parse->resultRelation);
2358 28 : if (parse->commandType == CMD_UPDATE)
2359 26 : updateColnosLists = list_make1(root->update_colnos);
2360 28 : if (parse->withCheckOptions)
2361 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2362 28 : if (parse->returningList)
2363 15 : returningLists = list_make1(parse->returningList);
2364 28 : if (parse->mergeActionList)
2365 1 : mergeActionLists = list_make1(parse->mergeActionList);
2366 28 : if (parse->commandType == CMD_MERGE)
2367 1 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2368 : }
2369 : }
2370 : else
2371 : {
2372 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2373 62827 : rootRelation = 0; /* there's no separate root rel */
2374 62827 : resultRelations = list_make1_int(parse->resultRelation);
2375 62827 : if (parse->commandType == CMD_UPDATE)
2376 8814 : updateColnosLists = list_make1(root->update_colnos);
2377 62827 : if (parse->withCheckOptions)
2378 889 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2379 62827 : if (parse->returningList)
2380 2204 : returningLists = list_make1(parse->returningList);
2381 62827 : if (parse->mergeActionList)
2382 1286 : mergeActionLists = list_make1(parse->mergeActionList);
2383 62827 : if (parse->commandType == CMD_MERGE)
2384 1286 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2385 : }
2386 :
2387 : /*
2388 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2389 : * will have dealt with fetching non-locked marked rows, else we
2390 : * need to have ModifyTable do that.
2391 : */
2392 65115 : if (parse->rowMarks)
2393 0 : rowMarks = NIL;
2394 : else
2395 65115 : rowMarks = root->rowMarks;
2396 :
2397 : path = (Path *)
2398 65115 : create_modifytable_path(root, final_rel,
2399 : path,
2400 : parse->commandType,
2401 65115 : parse->canSetTag,
2402 65115 : parse->resultRelation,
2403 : rootRelation,
2404 : resultRelations,
2405 : updateColnosLists,
2406 : withCheckOptionLists,
2407 : returningLists,
2408 : rowMarks,
2409 : parse->onConflict,
2410 : mergeActionLists,
2411 : mergeJoinConditions,
2412 : parse->forPortionOf,
2413 : assign_special_exec_param(root));
2414 : }
2415 :
2416 : /* And shove it into final_rel */
2417 414966 : add_path(final_rel, path);
2418 : }
2419 :
2420 : /*
2421 : * Generate partial paths for final_rel, too, if outer query levels might
2422 : * be able to make use of them.
2423 : */
2424 398449 : if (final_rel->consider_parallel && root->query_level > 1 &&
2425 24855 : !limit_needed(parse))
2426 : {
2427 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2428 24789 : foreach(lc, current_rel->partial_pathlist)
2429 : {
2430 112 : Path *partial_path = (Path *) lfirst(lc);
2431 :
2432 112 : add_partial_path(final_rel, partial_path);
2433 : }
2434 : }
2435 :
2436 398449 : extra.limit_needed = limit_needed(parse);
2437 398449 : extra.limit_tuples = limit_tuples;
2438 398449 : extra.count_est = count_est;
2439 398449 : extra.offset_est = offset_est;
2440 :
2441 : /*
2442 : * If there is an FDW that's responsible for all baserels of the query,
2443 : * let it consider adding ForeignPaths.
2444 : */
2445 398449 : if (final_rel->fdwroutine &&
2446 657 : final_rel->fdwroutine->GetForeignUpperPaths)
2447 620 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2448 : current_rel, final_rel,
2449 : &extra);
2450 :
2451 : /* Let extensions possibly add some more paths */
2452 398449 : if (create_upper_paths_hook)
2453 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2454 : current_rel, final_rel, &extra);
2455 :
2456 : /* Note: currently, we leave it to callers to do set_cheapest() */
2457 398449 : }
2458 :
2459 : /*
2460 : * Do preprocessing for groupingSets clause and related data.
2461 : *
2462 : * We expect that parse->groupingSets has already been expanded into a flat
2463 : * list of grouping sets (that is, just integer Lists of ressortgroupref
2464 : * numbers) by expand_grouping_sets(). This function handles the preliminary
2465 : * steps of organizing the grouping sets into lists of rollups, and preparing
2466 : * annotations which will later be filled in with size estimates.
2467 : */
2468 : static grouping_sets_data *
2469 905 : preprocess_grouping_sets(PlannerInfo *root)
2470 : {
2471 905 : Query *parse = root->parse;
2472 : List *sets;
2473 905 : int maxref = 0;
2474 : ListCell *lc_set;
2475 905 : grouping_sets_data *gd = palloc0_object(grouping_sets_data);
2476 :
2477 : /*
2478 : * We don't currently make any attempt to optimize the groupClause when
2479 : * there are grouping sets, so just duplicate it in processed_groupClause.
2480 : */
2481 905 : root->processed_groupClause = parse->groupClause;
2482 :
2483 : /* Detect unhashable and unsortable grouping expressions */
2484 905 : gd->any_hashable = false;
2485 905 : gd->unhashable_refs = NULL;
2486 905 : gd->unsortable_refs = NULL;
2487 905 : gd->unsortable_sets = NIL;
2488 :
2489 905 : if (parse->groupClause)
2490 : {
2491 : ListCell *lc;
2492 :
2493 2651 : foreach(lc, parse->groupClause)
2494 : {
2495 1811 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2496 1811 : Index ref = gc->tleSortGroupRef;
2497 :
2498 1811 : if (ref > maxref)
2499 1771 : maxref = ref;
2500 :
2501 1811 : if (!gc->hashable)
2502 24 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2503 :
2504 1811 : if (!OidIsValid(gc->sortop))
2505 33 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2506 : }
2507 : }
2508 :
2509 : /* Allocate workspace array for remapping */
2510 905 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2511 :
2512 : /*
2513 : * If we have any unsortable sets, we must extract them before trying to
2514 : * prepare rollups. Unsortable sets don't go through
2515 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2516 : * here.
2517 : */
2518 905 : if (!bms_is_empty(gd->unsortable_refs))
2519 : {
2520 33 : List *sortable_sets = NIL;
2521 : ListCell *lc;
2522 :
2523 99 : foreach(lc, parse->groupingSets)
2524 : {
2525 70 : List *gset = (List *) lfirst(lc);
2526 :
2527 70 : if (bms_overlap_list(gd->unsortable_refs, gset))
2528 : {
2529 38 : GroupingSetData *gs = makeNode(GroupingSetData);
2530 :
2531 38 : gs->set = gset;
2532 38 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2533 :
2534 : /*
2535 : * We must enforce here that an unsortable set is hashable;
2536 : * later code assumes this. Parse analysis only checks that
2537 : * every individual column is either hashable or sortable.
2538 : *
2539 : * Note that passing this test doesn't guarantee we can
2540 : * generate a plan; there might be other showstoppers.
2541 : */
2542 38 : if (bms_overlap_list(gd->unhashable_refs, gset))
2543 4 : ereport(ERROR,
2544 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2545 : errmsg("could not implement GROUP BY"),
2546 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2547 : }
2548 : else
2549 32 : sortable_sets = lappend(sortable_sets, gset);
2550 : }
2551 :
2552 29 : if (sortable_sets)
2553 24 : sets = extract_rollup_sets(sortable_sets);
2554 : else
2555 5 : sets = NIL;
2556 : }
2557 : else
2558 872 : sets = extract_rollup_sets(parse->groupingSets);
2559 :
2560 2376 : foreach(lc_set, sets)
2561 : {
2562 1475 : List *current_sets = (List *) lfirst(lc_set);
2563 1475 : RollupData *rollup = makeNode(RollupData);
2564 : GroupingSetData *gs;
2565 :
2566 : /*
2567 : * Reorder the current list of grouping sets into correct prefix
2568 : * order. If only one aggregation pass is needed, try to make the
2569 : * list match the ORDER BY clause; if more than one pass is needed, we
2570 : * don't bother with that.
2571 : *
2572 : * Note that this reorders the sets from smallest-member-first to
2573 : * largest-member-first, and applies the GroupingSetData annotations,
2574 : * though the data will be filled in later.
2575 : */
2576 1475 : current_sets = reorder_grouping_sets(current_sets,
2577 1475 : (list_length(sets) == 1
2578 : ? parse->sortClause
2579 : : NIL));
2580 :
2581 : /*
2582 : * Get the initial (and therefore largest) grouping set.
2583 : */
2584 1475 : gs = linitial_node(GroupingSetData, current_sets);
2585 :
2586 : /*
2587 : * Order the groupClause appropriately. If the first grouping set is
2588 : * empty, then the groupClause must also be empty; otherwise we have
2589 : * to force the groupClause to match that grouping set's order.
2590 : *
2591 : * (The first grouping set can be empty even though parse->groupClause
2592 : * is not empty only if all non-empty grouping sets are unsortable.
2593 : * The groupClauses for hashed grouping sets are built later on.)
2594 : */
2595 1475 : if (gs->set)
2596 1410 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2597 : else
2598 65 : rollup->groupClause = NIL;
2599 :
2600 : /*
2601 : * Is it hashable? We pretend empty sets are hashable even though we
2602 : * actually force them not to be hashed later. But don't bother if
2603 : * there's nothing but empty sets (since in that case we can't hash
2604 : * anything).
2605 : */
2606 1475 : if (gs->set &&
2607 1410 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2608 : {
2609 1390 : rollup->hashable = true;
2610 1390 : gd->any_hashable = true;
2611 : }
2612 :
2613 : /*
2614 : * Now that we've pinned down an order for the groupClause for this
2615 : * list of grouping sets, we need to remap the entries in the grouping
2616 : * sets from sortgrouprefs to plain indices (0-based) into the
2617 : * groupClause for this collection of grouping sets. We keep the
2618 : * original form for later use, though.
2619 : */
2620 1475 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2621 : current_sets,
2622 : gd->tleref_to_colnum_map);
2623 1475 : rollup->gsets_data = current_sets;
2624 :
2625 1475 : gd->rollups = lappend(gd->rollups, rollup);
2626 : }
2627 :
2628 901 : if (gd->unsortable_sets)
2629 : {
2630 : /*
2631 : * We have not yet pinned down a groupclause for this, but we will
2632 : * need index-based lists for estimation purposes. Construct
2633 : * hash_sets_idx based on the entire original groupclause for now.
2634 : */
2635 29 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2636 : gd->unsortable_sets,
2637 : gd->tleref_to_colnum_map);
2638 29 : gd->any_hashable = true;
2639 : }
2640 :
2641 901 : return gd;
2642 : }
2643 :
2644 : /*
2645 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2646 : * (without annotation) mapped to indexes into the given groupclause.
2647 : */
2648 : static List *
2649 4147 : remap_to_groupclause_idx(List *groupClause,
2650 : List *gsets,
2651 : int *tleref_to_colnum_map)
2652 : {
2653 4147 : int ref = 0;
2654 4147 : List *result = NIL;
2655 : ListCell *lc;
2656 :
2657 9861 : foreach(lc, groupClause)
2658 : {
2659 5714 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2660 :
2661 5714 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2662 : }
2663 :
2664 9444 : foreach(lc, gsets)
2665 : {
2666 5297 : List *set = NIL;
2667 : ListCell *lc2;
2668 5297 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2669 :
2670 11720 : foreach(lc2, gs->set)
2671 : {
2672 6423 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2673 : }
2674 :
2675 5297 : result = lappend(result, set);
2676 : }
2677 :
2678 4147 : return result;
2679 : }
2680 :
2681 :
2682 : /*
2683 : * preprocess_rowmarks - set up PlanRowMarks if needed
2684 : */
2685 : static void
2686 401128 : preprocess_rowmarks(PlannerInfo *root)
2687 : {
2688 401128 : Query *parse = root->parse;
2689 : Bitmapset *rels;
2690 : List *prowmarks;
2691 : ListCell *l;
2692 : int i;
2693 :
2694 401128 : if (parse->rowMarks)
2695 : {
2696 : /*
2697 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2698 : * grouping, since grouping renders a reference to individual tuple
2699 : * CTIDs invalid. This is also checked at parse time, but that's
2700 : * insufficient because of rule substitution, query pullup, etc.
2701 : */
2702 6564 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2703 : parse->rowMarks)->strength);
2704 : }
2705 : else
2706 : {
2707 : /*
2708 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2709 : * UPDATE/SHARE.
2710 : */
2711 394564 : if (parse->commandType != CMD_UPDATE &&
2712 384155 : parse->commandType != CMD_DELETE &&
2713 380436 : parse->commandType != CMD_MERGE)
2714 378956 : return;
2715 : }
2716 :
2717 : /*
2718 : * We need to have rowmarks for all base relations except the target. We
2719 : * make a bitmapset of all base rels and then remove the items we don't
2720 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2721 : */
2722 22172 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2723 22172 : if (parse->resultRelation)
2724 15608 : rels = bms_del_member(rels, parse->resultRelation);
2725 :
2726 : /*
2727 : * Convert RowMarkClauses to PlanRowMark representation.
2728 : */
2729 22172 : prowmarks = NIL;
2730 28871 : foreach(l, parse->rowMarks)
2731 : {
2732 6699 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2733 6699 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2734 : PlanRowMark *newrc;
2735 :
2736 : /*
2737 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2738 : * applied to an update/delete target rel. If that ever becomes
2739 : * possible, we should drop the target from the PlanRowMark list.
2740 : */
2741 : Assert(rc->rti != parse->resultRelation);
2742 :
2743 : /*
2744 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2745 : * can't support true locking. Subqueries that got flattened into the
2746 : * main query should be ignored completely. Any that didn't will get
2747 : * ROW_MARK_COPY items in the next loop.
2748 : */
2749 6699 : if (rte->rtekind != RTE_RELATION)
2750 46 : continue;
2751 :
2752 6653 : rels = bms_del_member(rels, rc->rti);
2753 :
2754 6653 : newrc = makeNode(PlanRowMark);
2755 6653 : newrc->rti = newrc->prti = rc->rti;
2756 6653 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2757 6653 : newrc->markType = select_rowmark_type(rte, rc->strength);
2758 6653 : newrc->allMarkTypes = (1 << newrc->markType);
2759 6653 : newrc->strength = rc->strength;
2760 6653 : newrc->waitPolicy = rc->waitPolicy;
2761 6653 : newrc->isParent = false;
2762 :
2763 6653 : prowmarks = lappend(prowmarks, newrc);
2764 : }
2765 :
2766 : /*
2767 : * Now, add rowmarks for any non-target, non-locked base relations.
2768 : */
2769 22172 : i = 0;
2770 52951 : foreach(l, parse->rtable)
2771 : {
2772 30779 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2773 : PlanRowMark *newrc;
2774 :
2775 30779 : i++;
2776 30779 : if (!bms_is_member(i, rels))
2777 27831 : continue;
2778 :
2779 2948 : newrc = makeNode(PlanRowMark);
2780 2948 : newrc->rti = newrc->prti = i;
2781 2948 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2782 2948 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2783 2948 : newrc->allMarkTypes = (1 << newrc->markType);
2784 2948 : newrc->strength = LCS_NONE;
2785 2948 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2786 2948 : newrc->isParent = false;
2787 :
2788 2948 : prowmarks = lappend(prowmarks, newrc);
2789 : }
2790 :
2791 22172 : root->rowMarks = prowmarks;
2792 : }
2793 :
2794 : /*
2795 : * Select RowMarkType to use for a given table
2796 : */
2797 : RowMarkType
2798 11495 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2799 : {
2800 11495 : if (rte->rtekind != RTE_RELATION)
2801 : {
2802 : /* If it's not a table at all, use ROW_MARK_COPY */
2803 1109 : return ROW_MARK_COPY;
2804 : }
2805 10386 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2806 : {
2807 : /* Let the FDW select the rowmark type, if it wants to */
2808 114 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2809 :
2810 114 : if (fdwroutine->GetForeignRowMarkType != NULL)
2811 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2812 : /* Otherwise, use ROW_MARK_COPY by default */
2813 114 : return ROW_MARK_COPY;
2814 : }
2815 : else
2816 : {
2817 : /* Regular table, apply the appropriate lock type */
2818 10272 : switch (strength)
2819 : {
2820 2048 : case LCS_NONE:
2821 :
2822 : /*
2823 : * We don't need a tuple lock, only the ability to re-fetch
2824 : * the row.
2825 : */
2826 2048 : return ROW_MARK_REFERENCE;
2827 : break;
2828 7055 : case LCS_FORKEYSHARE:
2829 7055 : return ROW_MARK_KEYSHARE;
2830 : break;
2831 201 : case LCS_FORSHARE:
2832 201 : return ROW_MARK_SHARE;
2833 : break;
2834 41 : case LCS_FORNOKEYUPDATE:
2835 41 : return ROW_MARK_NOKEYEXCLUSIVE;
2836 : break;
2837 927 : case LCS_FORUPDATE:
2838 927 : return ROW_MARK_EXCLUSIVE;
2839 : break;
2840 : }
2841 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2842 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2843 : }
2844 : }
2845 :
2846 : /*
2847 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2848 : *
2849 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2850 : * results back in *count_est and *offset_est. These variables are set to
2851 : * 0 if the corresponding clause is not present, and -1 if it's present
2852 : * but we couldn't estimate the value for it. (The "0" convention is OK
2853 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2854 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2855 : * usual practice of never estimating less than one row.) These values will
2856 : * be passed to create_limit_path, which see if you change this code.
2857 : *
2858 : * The return value is the suitably adjusted tuple_fraction to use for
2859 : * planning the query. This adjustment is not overridable, since it reflects
2860 : * plan actions that grouping_planner() will certainly take, not assumptions
2861 : * about context.
2862 : */
2863 : static double
2864 3714 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2865 : int64 *offset_est, int64 *count_est)
2866 : {
2867 3714 : Query *parse = root->parse;
2868 : Node *est;
2869 : double limit_fraction;
2870 :
2871 : /* Should not be called unless LIMIT or OFFSET */
2872 : Assert(parse->limitCount || parse->limitOffset);
2873 :
2874 : /*
2875 : * Try to obtain the clause values. We use estimate_expression_value
2876 : * primarily because it can sometimes do something useful with Params.
2877 : */
2878 3714 : if (parse->limitCount)
2879 : {
2880 3272 : est = estimate_expression_value(root, parse->limitCount);
2881 3272 : if (est && IsA(est, Const))
2882 : {
2883 3267 : if (((Const *) est)->constisnull)
2884 : {
2885 : /* NULL indicates LIMIT ALL, ie, no limit */
2886 0 : *count_est = 0; /* treat as not present */
2887 : }
2888 : else
2889 : {
2890 3267 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2891 3267 : if (*count_est <= 0)
2892 190 : *count_est = 1; /* force to at least 1 */
2893 : }
2894 : }
2895 : else
2896 5 : *count_est = -1; /* can't estimate */
2897 : }
2898 : else
2899 442 : *count_est = 0; /* not present */
2900 :
2901 3714 : if (parse->limitOffset)
2902 : {
2903 652 : est = estimate_expression_value(root, parse->limitOffset);
2904 652 : if (est && IsA(est, Const))
2905 : {
2906 632 : if (((Const *) est)->constisnull)
2907 : {
2908 : /* Treat NULL as no offset; the executor will too */
2909 0 : *offset_est = 0; /* treat as not present */
2910 : }
2911 : else
2912 : {
2913 632 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2914 632 : if (*offset_est < 0)
2915 0 : *offset_est = 0; /* treat as not present */
2916 : }
2917 : }
2918 : else
2919 20 : *offset_est = -1; /* can't estimate */
2920 : }
2921 : else
2922 3062 : *offset_est = 0; /* not present */
2923 :
2924 3714 : if (*count_est != 0)
2925 : {
2926 : /*
2927 : * A LIMIT clause limits the absolute number of tuples returned.
2928 : * However, if it's not a constant LIMIT then we have to guess; for
2929 : * lack of a better idea, assume 10% of the plan's result is wanted.
2930 : */
2931 3272 : if (*count_est < 0 || *offset_est < 0)
2932 : {
2933 : /* LIMIT or OFFSET is an expression ... punt ... */
2934 20 : limit_fraction = 0.10;
2935 : }
2936 : else
2937 : {
2938 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2939 3252 : limit_fraction = (double) *count_est + (double) *offset_est;
2940 : }
2941 :
2942 : /*
2943 : * If we have absolute limits from both caller and LIMIT, use the
2944 : * smaller value; likewise if they are both fractional. If one is
2945 : * fractional and the other absolute, we can't easily determine which
2946 : * is smaller, but we use the heuristic that the absolute will usually
2947 : * be smaller.
2948 : */
2949 3272 : if (tuple_fraction >= 1.0)
2950 : {
2951 5 : if (limit_fraction >= 1.0)
2952 : {
2953 : /* both absolute */
2954 5 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2955 : }
2956 : else
2957 : {
2958 : /* caller absolute, limit fractional; use caller's value */
2959 : }
2960 : }
2961 3267 : else if (tuple_fraction > 0.0)
2962 : {
2963 82 : if (limit_fraction >= 1.0)
2964 : {
2965 : /* caller fractional, limit absolute; use limit */
2966 82 : tuple_fraction = limit_fraction;
2967 : }
2968 : else
2969 : {
2970 : /* both fractional */
2971 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2972 : }
2973 : }
2974 : else
2975 : {
2976 : /* no info from caller, just use limit */
2977 3185 : tuple_fraction = limit_fraction;
2978 : }
2979 : }
2980 442 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2981 : {
2982 : /*
2983 : * We have an OFFSET but no LIMIT. This acts entirely differently
2984 : * from the LIMIT case: here, we need to increase rather than decrease
2985 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2986 : * tuples to be fetched instead of fewer. This only matters if we got
2987 : * a tuple_fraction > 0, however.
2988 : *
2989 : * As above, use 10% if OFFSET is present but unestimatable.
2990 : */
2991 12 : if (*offset_est < 0)
2992 0 : limit_fraction = 0.10;
2993 : else
2994 12 : limit_fraction = (double) *offset_est;
2995 :
2996 : /*
2997 : * If we have absolute counts from both caller and OFFSET, add them
2998 : * together; likewise if they are both fractional. If one is
2999 : * fractional and the other absolute, we want to take the larger, and
3000 : * we heuristically assume that's the fractional one.
3001 : */
3002 12 : if (tuple_fraction >= 1.0)
3003 : {
3004 0 : if (limit_fraction >= 1.0)
3005 : {
3006 : /* both absolute, so add them together */
3007 0 : tuple_fraction += limit_fraction;
3008 : }
3009 : else
3010 : {
3011 : /* caller absolute, limit fractional; use limit */
3012 0 : tuple_fraction = limit_fraction;
3013 : }
3014 : }
3015 : else
3016 : {
3017 12 : if (limit_fraction >= 1.0)
3018 : {
3019 : /* caller fractional, limit absolute; use caller's value */
3020 : }
3021 : else
3022 : {
3023 : /* both fractional, so add them together */
3024 0 : tuple_fraction += limit_fraction;
3025 0 : if (tuple_fraction >= 1.0)
3026 0 : tuple_fraction = 0.0; /* assume fetch all */
3027 : }
3028 : }
3029 : }
3030 :
3031 3714 : return tuple_fraction;
3032 : }
3033 :
3034 : /*
3035 : * limit_needed - do we actually need a Limit plan node?
3036 : *
3037 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
3038 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
3039 : * locution for an optimization fence. (Because other places in the planner
3040 : * merely check whether parse->limitOffset isn't NULL, it will still work as
3041 : * an optimization fence --- we're just suppressing unnecessary run-time
3042 : * overhead.)
3043 : *
3044 : * This might look like it could be merged into preprocess_limit, but there's
3045 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3046 : * in preprocess_limit it's good enough to consider estimated values.
3047 : */
3048 : bool
3049 854646 : limit_needed(Query *parse)
3050 : {
3051 : Node *node;
3052 :
3053 854646 : node = parse->limitCount;
3054 854646 : if (node)
3055 : {
3056 7779 : if (IsA(node, Const))
3057 : {
3058 : /* NULL indicates LIMIT ALL, ie, no limit */
3059 7558 : if (!((Const *) node)->constisnull)
3060 7558 : return true; /* LIMIT with a constant value */
3061 : }
3062 : else
3063 221 : return true; /* non-constant LIMIT */
3064 : }
3065 :
3066 846867 : node = parse->limitOffset;
3067 846867 : if (node)
3068 : {
3069 1283 : if (IsA(node, Const))
3070 : {
3071 : /* Treat NULL as no offset; the executor would too */
3072 1025 : if (!((Const *) node)->constisnull)
3073 : {
3074 1025 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3075 :
3076 1025 : if (offset != 0)
3077 107 : return true; /* OFFSET with a nonzero value */
3078 : }
3079 : }
3080 : else
3081 258 : return true; /* non-constant OFFSET */
3082 : }
3083 :
3084 846502 : return false; /* don't need a Limit plan node */
3085 : }
3086 :
3087 : /*
3088 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3089 : *
3090 : * The idea here is to adjust the ordering of the GROUP BY elements
3091 : * (which in itself is semantically insignificant) to match ORDER BY,
3092 : * thereby allowing a single sort operation to both implement the ORDER BY
3093 : * requirement and set up for a Unique step that implements GROUP BY.
3094 : * We also consider partial match between GROUP BY and ORDER BY elements,
3095 : * which could allow to implement ORDER BY using the incremental sort.
3096 : *
3097 : * We also consider other orderings of the GROUP BY elements, which could
3098 : * match the sort ordering of other possible plans (eg an indexscan) and
3099 : * thereby reduce cost. This is implemented during the generation of grouping
3100 : * paths. See get_useful_group_keys_orderings() for details.
3101 : *
3102 : * Note: we need no comparable processing of the distinctClause because
3103 : * the parser already enforced that that matches ORDER BY.
3104 : *
3105 : * Note: we return a fresh List, but its elements are the same
3106 : * SortGroupClauses appearing in parse->groupClause. This is important
3107 : * because later processing may modify the processed_groupClause list.
3108 : *
3109 : * For grouping sets, the order of items is instead forced to agree with that
3110 : * of the grouping set (and items not in the grouping set are skipped). The
3111 : * work of sorting the order of grouping set elements to match the ORDER BY if
3112 : * possible is done elsewhere.
3113 : */
3114 : static List *
3115 7487 : preprocess_groupclause(PlannerInfo *root, List *force)
3116 : {
3117 7487 : Query *parse = root->parse;
3118 7487 : List *new_groupclause = NIL;
3119 : ListCell *sl;
3120 : ListCell *gl;
3121 :
3122 : /* For grouping sets, we need to force the ordering */
3123 7487 : if (force)
3124 : {
3125 9704 : foreach(sl, force)
3126 : {
3127 5651 : Index ref = lfirst_int(sl);
3128 5651 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3129 :
3130 5651 : new_groupclause = lappend(new_groupclause, cl);
3131 : }
3132 :
3133 4053 : return new_groupclause;
3134 : }
3135 :
3136 : /* If no ORDER BY, nothing useful to do here */
3137 3434 : if (parse->sortClause == NIL)
3138 1954 : return list_copy(parse->groupClause);
3139 :
3140 : /*
3141 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3142 : * items, but only as far as we can make a matching prefix.
3143 : *
3144 : * This code assumes that the sortClause contains no duplicate items.
3145 : */
3146 2902 : foreach(sl, parse->sortClause)
3147 : {
3148 1938 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3149 :
3150 2825 : foreach(gl, parse->groupClause)
3151 : {
3152 2309 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3153 :
3154 2309 : if (equal(gc, sc))
3155 : {
3156 1422 : new_groupclause = lappend(new_groupclause, gc);
3157 1422 : break;
3158 : }
3159 : }
3160 1938 : if (gl == NULL)
3161 516 : break; /* no match, so stop scanning */
3162 : }
3163 :
3164 :
3165 : /* If no match at all, no point in reordering GROUP BY */
3166 1480 : if (new_groupclause == NIL)
3167 228 : return list_copy(parse->groupClause);
3168 :
3169 : /*
3170 : * Add any remaining GROUP BY items to the new list. We don't require a
3171 : * complete match, because even partial match allows ORDER BY to be
3172 : * implemented using incremental sort. Also, give up if there are any
3173 : * non-sortable GROUP BY items, since then there's no hope anyway.
3174 : */
3175 2801 : foreach(gl, parse->groupClause)
3176 : {
3177 1549 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3178 :
3179 1549 : if (list_member_ptr(new_groupclause, gc))
3180 1422 : continue; /* it matched an ORDER BY item */
3181 127 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
3182 0 : return list_copy(parse->groupClause);
3183 127 : new_groupclause = lappend(new_groupclause, gc);
3184 : }
3185 :
3186 : /* Success --- install the rearranged GROUP BY list */
3187 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3188 1252 : return new_groupclause;
3189 : }
3190 :
3191 : /*
3192 : * Extract lists of grouping sets that can be implemented using a single
3193 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3194 : *
3195 : * Input must be sorted with smallest sets first. Result has each sublist
3196 : * sorted with smallest sets first.
3197 : *
3198 : * We want to produce the absolute minimum possible number of lists here to
3199 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3200 : * of finding the minimal partition of a partially-ordered set into chains
3201 : * (which is what we need, taking the list of grouping sets as a poset ordered
3202 : * by set inclusion) can be mapped to the problem of finding the maximum
3203 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3204 : * time with a worst case of no worse than O(n^2.5) and usually much
3205 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3206 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3207 : * half a second on my modest system even with optimization off and assertions
3208 : * on.)
3209 : */
3210 : static List *
3211 896 : extract_rollup_sets(List *groupingSets)
3212 : {
3213 896 : int num_sets_raw = list_length(groupingSets);
3214 896 : int num_empty = 0;
3215 896 : int num_sets = 0; /* distinct sets */
3216 896 : int num_chains = 0;
3217 896 : List *result = NIL;
3218 : List **results;
3219 : List **orig_sets;
3220 : Bitmapset **set_masks;
3221 : int *chains;
3222 : short **adjacency;
3223 : short *adjacency_buf;
3224 : BipartiteMatchState *state;
3225 : int i;
3226 : int j;
3227 : int j_size;
3228 896 : ListCell *lc1 = list_head(groupingSets);
3229 : ListCell *lc;
3230 :
3231 : /*
3232 : * Start by stripping out empty sets. The algorithm doesn't require this,
3233 : * but the planner currently needs all empty sets to be returned in the
3234 : * first list, so we strip them here and add them back after.
3235 : */
3236 1486 : while (lc1 && lfirst(lc1) == NIL)
3237 : {
3238 590 : ++num_empty;
3239 590 : lc1 = lnext(groupingSets, lc1);
3240 : }
3241 :
3242 : /* bail out now if it turns out that all we had were empty sets. */
3243 896 : if (!lc1)
3244 65 : return list_make1(groupingSets);
3245 :
3246 : /*----------
3247 : * We don't strictly need to remove duplicate sets here, but if we don't,
3248 : * they tend to become scattered through the result, which is a bit
3249 : * confusing (and irritating if we ever decide to optimize them out).
3250 : * So we remove them here and add them back after.
3251 : *
3252 : * For each non-duplicate set, we fill in the following:
3253 : *
3254 : * orig_sets[i] = list of the original set lists
3255 : * set_masks[i] = bitmapset for testing inclusion
3256 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3257 : *
3258 : * chains[i] will be the result group this set is assigned to.
3259 : *
3260 : * We index all of these from 1 rather than 0 because it is convenient
3261 : * to leave 0 free for the NIL node in the graph algorithm.
3262 : *----------
3263 : */
3264 831 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3265 831 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3266 831 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3267 831 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3268 :
3269 831 : j_size = 0;
3270 831 : j = 0;
3271 831 : i = 1;
3272 :
3273 2861 : for_each_cell(lc, groupingSets, lc1)
3274 : {
3275 2030 : List *candidate = (List *) lfirst(lc);
3276 2030 : Bitmapset *candidate_set = NULL;
3277 : ListCell *lc2;
3278 2030 : int dup_of = 0;
3279 :
3280 4827 : foreach(lc2, candidate)
3281 : {
3282 2797 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3283 : }
3284 :
3285 : /* we can only be a dup if we're the same length as a previous set */
3286 2030 : if (j_size == list_length(candidate))
3287 : {
3288 : int k;
3289 :
3290 1828 : for (k = j; k < i; ++k)
3291 : {
3292 1169 : if (bms_equal(set_masks[k], candidate_set))
3293 : {
3294 129 : dup_of = k;
3295 129 : break;
3296 : }
3297 : }
3298 : }
3299 1242 : else if (j_size < list_length(candidate))
3300 : {
3301 1242 : j_size = list_length(candidate);
3302 1242 : j = i;
3303 : }
3304 :
3305 2030 : if (dup_of > 0)
3306 : {
3307 129 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3308 129 : bms_free(candidate_set);
3309 : }
3310 : else
3311 : {
3312 : int k;
3313 1901 : int n_adj = 0;
3314 :
3315 1901 : orig_sets[i] = list_make1(candidate);
3316 1901 : set_masks[i] = candidate_set;
3317 :
3318 : /* fill in adjacency list; no need to compare equal-size sets */
3319 :
3320 2973 : for (k = j - 1; k > 0; --k)
3321 : {
3322 1072 : if (bms_is_subset(set_masks[k], candidate_set))
3323 937 : adjacency_buf[++n_adj] = k;
3324 : }
3325 :
3326 1901 : if (n_adj > 0)
3327 : {
3328 513 : adjacency_buf[0] = n_adj;
3329 513 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3330 513 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3331 : }
3332 : else
3333 1388 : adjacency[i] = NULL;
3334 :
3335 1901 : ++i;
3336 : }
3337 : }
3338 :
3339 831 : num_sets = i - 1;
3340 :
3341 : /*
3342 : * Apply the graph matching algorithm to do the work.
3343 : */
3344 831 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3345 :
3346 : /*
3347 : * Now, the state->pair* fields have the info we need to assign sets to
3348 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3349 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3350 : * it in one pass)
3351 : */
3352 831 : chains = palloc0((num_sets + 1) * sizeof(int));
3353 :
3354 2732 : for (i = 1; i <= num_sets; ++i)
3355 : {
3356 1901 : int u = state->pair_vu[i];
3357 1901 : int v = state->pair_uv[i];
3358 :
3359 1901 : if (u > 0 && u < i)
3360 0 : chains[i] = chains[u];
3361 1901 : else if (v > 0 && v < i)
3362 491 : chains[i] = chains[v];
3363 : else
3364 1410 : chains[i] = ++num_chains;
3365 : }
3366 :
3367 : /* build result lists. */
3368 831 : results = palloc0((num_chains + 1) * sizeof(List *));
3369 :
3370 2732 : for (i = 1; i <= num_sets; ++i)
3371 : {
3372 1901 : int c = chains[i];
3373 :
3374 : Assert(c > 0);
3375 :
3376 1901 : results[c] = list_concat(results[c], orig_sets[i]);
3377 : }
3378 :
3379 : /* push any empty sets back on the first list. */
3380 1306 : while (num_empty-- > 0)
3381 475 : results[1] = lcons(NIL, results[1]);
3382 :
3383 : /* make result list */
3384 2241 : for (i = 1; i <= num_chains; ++i)
3385 1410 : result = lappend(result, results[i]);
3386 :
3387 : /*
3388 : * Free all the things.
3389 : *
3390 : * (This is over-fussy for small sets but for large sets we could have
3391 : * tied up a nontrivial amount of memory.)
3392 : */
3393 831 : BipartiteMatchFree(state);
3394 831 : pfree(results);
3395 831 : pfree(chains);
3396 2732 : for (i = 1; i <= num_sets; ++i)
3397 1901 : if (adjacency[i])
3398 513 : pfree(adjacency[i]);
3399 831 : pfree(adjacency);
3400 831 : pfree(adjacency_buf);
3401 831 : pfree(orig_sets);
3402 2732 : for (i = 1; i <= num_sets; ++i)
3403 1901 : bms_free(set_masks[i]);
3404 831 : pfree(set_masks);
3405 :
3406 831 : return result;
3407 : }
3408 :
3409 : /*
3410 : * Reorder the elements of a list of grouping sets such that they have correct
3411 : * prefix relationships. Also inserts the GroupingSetData annotations.
3412 : *
3413 : * The input must be ordered with smallest sets first; the result is returned
3414 : * with largest sets first. Note that the result shares no list substructure
3415 : * with the input, so it's safe for the caller to modify it later.
3416 : *
3417 : * If we're passed in a sortclause, we follow its order of columns to the
3418 : * extent possible, to minimize the chance that we add unnecessary sorts.
3419 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3420 : * gets implemented in one pass.)
3421 : */
3422 : static List *
3423 1475 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3424 : {
3425 : ListCell *lc;
3426 1475 : List *previous = NIL;
3427 1475 : List *result = NIL;
3428 :
3429 4095 : foreach(lc, groupingSets)
3430 : {
3431 2620 : List *candidate = (List *) lfirst(lc);
3432 2620 : List *new_elems = list_difference_int(candidate, previous);
3433 2620 : GroupingSetData *gs = makeNode(GroupingSetData);
3434 :
3435 2764 : while (list_length(sortclause) > list_length(previous) &&
3436 : new_elems != NIL)
3437 : {
3438 244 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3439 244 : int ref = sc->tleSortGroupRef;
3440 :
3441 244 : if (list_member_int(new_elems, ref))
3442 : {
3443 144 : previous = lappend_int(previous, ref);
3444 144 : new_elems = list_delete_int(new_elems, ref);
3445 : }
3446 : else
3447 : {
3448 : /* diverged from the sortclause; give up on it */
3449 100 : sortclause = NIL;
3450 100 : break;
3451 : }
3452 : }
3453 :
3454 2620 : previous = list_concat(previous, new_elems);
3455 :
3456 2620 : gs->set = list_copy(previous);
3457 2620 : result = lcons(gs, result);
3458 : }
3459 :
3460 1475 : list_free(previous);
3461 :
3462 1475 : return result;
3463 : }
3464 :
3465 : /*
3466 : * has_volatile_pathkey
3467 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3468 : * containing a volatile function. Otherwise returns false.
3469 : */
3470 : static bool
3471 2205 : has_volatile_pathkey(List *keys)
3472 : {
3473 : ListCell *lc;
3474 :
3475 4532 : foreach(lc, keys)
3476 : {
3477 2340 : PathKey *pathkey = lfirst_node(PathKey, lc);
3478 :
3479 2340 : if (pathkey->pk_eclass->ec_has_volatile)
3480 13 : return true;
3481 : }
3482 :
3483 2192 : return false;
3484 : }
3485 :
3486 : /*
3487 : * adjust_group_pathkeys_for_groupagg
3488 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3489 : * pre-ordered input for ordered aggregates.
3490 : *
3491 : * We define "best" as the pathkeys that suit the largest number of
3492 : * aggregate functions. We find these by looking at the first ORDER BY /
3493 : * DISTINCT aggregate and take the pathkeys for that before searching for
3494 : * other aggregates that require the same or a more strict variation of the
3495 : * same pathkeys. We then repeat that process for any remaining aggregates
3496 : * with different pathkeys and if we find another set of pathkeys that suits a
3497 : * larger number of aggregates then we select those pathkeys instead.
3498 : *
3499 : * When the best pathkeys are found we also mark each Aggref that can use
3500 : * those pathkeys as aggpresorted = true.
3501 : *
3502 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3503 : * volatile functions, we never make use of these pathkeys. We want to ensure
3504 : * that sorts using volatile functions are done independently in each Aggref
3505 : * rather than once at the query level. If we were to allow this then Aggrefs
3506 : * with compatible sort orders would all transition their rows in the same
3507 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3508 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3509 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3510 : * left to perform their sorts individually. To avoid this inconsistent
3511 : * behavior which could make Aggref results depend on what other Aggrefs the
3512 : * query contains, we always force Aggrefs with volatile functions to perform
3513 : * their own sorts.
3514 : */
3515 : static void
3516 1869 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3517 : {
3518 1869 : List *grouppathkeys = root->group_pathkeys;
3519 : List *bestpathkeys;
3520 : Bitmapset *bestaggs;
3521 : Bitmapset *unprocessed_aggs;
3522 : ListCell *lc;
3523 : int i;
3524 :
3525 : /* Shouldn't be here if there are grouping sets */
3526 : Assert(root->parse->groupingSets == NIL);
3527 : /* Shouldn't be here unless there are some ordered aggregates */
3528 : Assert(root->numOrderedAggs > 0);
3529 :
3530 : /* Do nothing if disabled */
3531 1869 : if (!enable_presorted_aggregate)
3532 5 : return;
3533 :
3534 : /*
3535 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3536 : * the indexes of all AggInfos to be processed below.
3537 : */
3538 1864 : unprocessed_aggs = NULL;
3539 4294 : foreach(lc, root->agginfos)
3540 : {
3541 2430 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3542 2430 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3543 :
3544 2430 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3545 214 : continue;
3546 :
3547 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3548 2216 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3549 248 : continue;
3550 :
3551 : /* Additional safety checks are needed if there's a FILTER clause */
3552 1968 : if (aggref->aggfilter != NULL)
3553 : {
3554 : ListCell *lc2;
3555 43 : bool allow_presort = true;
3556 :
3557 : /*
3558 : * When the Aggref has a FILTER clause, it's possible that the
3559 : * filter removes rows that cannot be sorted because the
3560 : * expression to sort by results in an error during its
3561 : * evaluation. This is a problem for presorting as that happens
3562 : * before the FILTER, whereas without presorting, the Aggregate
3563 : * node will apply the FILTER *before* sorting. So that we never
3564 : * try to sort anything that might error, here we aim to skip over
3565 : * any Aggrefs with arguments with expressions which, when
3566 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3567 : * may be more cases that should be allowed, but more thought
3568 : * needs to be given. Err on the side of caution.
3569 : */
3570 83 : foreach(lc2, aggref->args)
3571 : {
3572 58 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3573 58 : Expr *expr = tle->expr;
3574 :
3575 68 : while (IsA(expr, RelabelType))
3576 10 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3577 :
3578 : /* Common case, Vars and Consts are ok */
3579 58 : if (IsA(expr, Var) || IsA(expr, Const))
3580 40 : continue;
3581 :
3582 : /* Unsupported. Don't try to presort for this Aggref */
3583 18 : allow_presort = false;
3584 18 : break;
3585 : }
3586 :
3587 : /* Skip unsupported Aggrefs */
3588 43 : if (!allow_presort)
3589 18 : continue;
3590 : }
3591 :
3592 1950 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3593 : foreach_current_index(lc));
3594 : }
3595 :
3596 : /*
3597 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3598 : * for the given set of aggregates.
3599 : *
3600 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3601 : * this during the first loop using the pathkeys for the very first
3602 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3603 : * a more strict set of compatible pathkeys. Once the outer loop is
3604 : * complete, we mark off all the aggregates with compatible pathkeys then
3605 : * remove those from the unprocessed_aggs and repeat the process to try to
3606 : * find another set of pathkeys that are suitable for a larger number of
3607 : * aggregates. The outer loop will stop when there are not enough
3608 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3609 : * to suit a larger number of aggregates.
3610 : */
3611 1864 : bestpathkeys = NIL;
3612 1864 : bestaggs = NULL;
3613 3679 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3614 : {
3615 1815 : Bitmapset *aggindexes = NULL;
3616 1815 : List *currpathkeys = NIL;
3617 :
3618 1815 : i = -1;
3619 4020 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3620 : {
3621 2205 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3622 2205 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3623 : List *sortlist;
3624 : List *pathkeys;
3625 :
3626 2205 : if (aggref->aggdistinct != NIL)
3627 597 : sortlist = aggref->aggdistinct;
3628 : else
3629 1608 : sortlist = aggref->aggorder;
3630 :
3631 2205 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3632 : aggref->args);
3633 :
3634 : /*
3635 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3636 : * or DISTINCT clause.
3637 : */
3638 2205 : if (has_volatile_pathkey(pathkeys))
3639 : {
3640 13 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3641 13 : continue;
3642 : }
3643 :
3644 : /*
3645 : * When not set yet, take the pathkeys from the first unprocessed
3646 : * aggregate.
3647 : */
3648 2192 : if (currpathkeys == NIL)
3649 : {
3650 1812 : currpathkeys = pathkeys;
3651 :
3652 : /* include the GROUP BY pathkeys, if they exist */
3653 1812 : if (grouppathkeys != NIL)
3654 226 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3655 : currpathkeys);
3656 :
3657 : /* record that we found pathkeys for this aggregate */
3658 1812 : aggindexes = bms_add_member(aggindexes, i);
3659 : }
3660 : else
3661 : {
3662 : /* now look for a stronger set of matching pathkeys */
3663 :
3664 : /* include the GROUP BY pathkeys, if they exist */
3665 380 : if (grouppathkeys != NIL)
3666 240 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3667 : pathkeys);
3668 :
3669 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3670 380 : switch (compare_pathkeys(currpathkeys, pathkeys))
3671 : {
3672 10 : case PATHKEYS_BETTER2:
3673 : /* 'pathkeys' are stronger, use these ones instead */
3674 10 : currpathkeys = pathkeys;
3675 : pg_fallthrough;
3676 :
3677 55 : case PATHKEYS_BETTER1:
3678 : /* 'pathkeys' are less strict */
3679 : pg_fallthrough;
3680 :
3681 : case PATHKEYS_EQUAL:
3682 : /* mark this aggregate as covered by 'currpathkeys' */
3683 55 : aggindexes = bms_add_member(aggindexes, i);
3684 55 : break;
3685 :
3686 325 : case PATHKEYS_DIFFERENT:
3687 325 : break;
3688 : }
3689 : }
3690 : }
3691 :
3692 : /* remove the aggregates that we've just processed */
3693 1815 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3694 :
3695 : /*
3696 : * If this pass included more aggregates than the previous best then
3697 : * use these ones as the best set.
3698 : */
3699 1815 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3700 : {
3701 1727 : bestaggs = aggindexes;
3702 1727 : bestpathkeys = currpathkeys;
3703 : }
3704 : }
3705 :
3706 : /*
3707 : * If we found any ordered aggregates, update root->group_pathkeys to add
3708 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3709 : * the original GROUP BY pathkeys already.
3710 : */
3711 1864 : if (bestpathkeys != NIL)
3712 1677 : root->group_pathkeys = bestpathkeys;
3713 :
3714 : /*
3715 : * Now that we've found the best set of aggregates we can set the
3716 : * presorted flag to indicate to the executor that it needn't bother
3717 : * performing a sort for these Aggrefs. We're able to do this now as
3718 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3719 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3720 : * of ordered aggregates.
3721 : */
3722 1864 : i = -1;
3723 3621 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3724 : {
3725 1757 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3726 :
3727 3529 : foreach(lc, agginfo->aggrefs)
3728 : {
3729 1772 : Aggref *aggref = lfirst_node(Aggref, lc);
3730 :
3731 1772 : aggref->aggpresorted = true;
3732 : }
3733 : }
3734 : }
3735 :
3736 : /*
3737 : * Compute query_pathkeys and other pathkeys during plan generation
3738 : */
3739 : static void
3740 393524 : standard_qp_callback(PlannerInfo *root, void *extra)
3741 : {
3742 393524 : Query *parse = root->parse;
3743 393524 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3744 393524 : List *tlist = root->processed_tlist;
3745 393524 : List *activeWindows = qp_extra->activeWindows;
3746 :
3747 : /*
3748 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3749 : * aggregate requirements.
3750 : */
3751 393524 : if (qp_extra->gset_data)
3752 : {
3753 : /*
3754 : * With grouping sets, just use the first RollupData's groupClause. We
3755 : * don't make any effort to optimize grouping clauses when there are
3756 : * grouping sets, nor can we combine aggregate ordering keys with
3757 : * grouping.
3758 : */
3759 901 : List *rollups = qp_extra->gset_data->rollups;
3760 901 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3761 :
3762 901 : if (grouping_is_sortable(groupClause))
3763 : {
3764 : bool sortable;
3765 :
3766 : /*
3767 : * The groupClause is logically below the grouping step. So if
3768 : * there is an RTE entry for the grouping step, we need to remove
3769 : * its RT index from the sort expressions before we make PathKeys
3770 : * for them.
3771 : */
3772 901 : root->group_pathkeys =
3773 901 : make_pathkeys_for_sortclauses_extended(root,
3774 : &groupClause,
3775 : tlist,
3776 : false,
3777 901 : parse->hasGroupRTE,
3778 : &sortable,
3779 : false);
3780 : Assert(sortable);
3781 901 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3782 : }
3783 : else
3784 : {
3785 0 : root->group_pathkeys = NIL;
3786 0 : root->num_groupby_pathkeys = 0;
3787 : }
3788 : }
3789 392623 : else if (parse->groupClause || root->numOrderedAggs > 0)
3790 5115 : {
3791 : /*
3792 : * With a plain GROUP BY list, we can remove any grouping items that
3793 : * are proven redundant by EquivalenceClass processing. For example,
3794 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3795 : * especially common cases, but they're nearly free to detect. Note
3796 : * that we remove redundant items from processed_groupClause but not
3797 : * the original parse->groupClause.
3798 : */
3799 : bool sortable;
3800 :
3801 : /*
3802 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3803 : * EquivalenceClass'es if it's not set yet.
3804 : */
3805 5115 : root->group_pathkeys =
3806 5115 : make_pathkeys_for_sortclauses_extended(root,
3807 : &root->processed_groupClause,
3808 : tlist,
3809 : true,
3810 : false,
3811 : &sortable,
3812 : true);
3813 5115 : if (!sortable)
3814 : {
3815 : /* Can't sort; no point in considering aggregate ordering either */
3816 0 : root->group_pathkeys = NIL;
3817 0 : root->num_groupby_pathkeys = 0;
3818 : }
3819 : else
3820 : {
3821 5115 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3822 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3823 5115 : if (root->numOrderedAggs > 0)
3824 1869 : adjust_group_pathkeys_for_groupagg(root);
3825 : }
3826 : }
3827 : else
3828 : {
3829 387508 : root->group_pathkeys = NIL;
3830 387508 : root->num_groupby_pathkeys = 0;
3831 : }
3832 :
3833 : /* We consider only the first (bottom) window in pathkeys logic */
3834 393524 : if (activeWindows != NIL)
3835 : {
3836 2217 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3837 :
3838 2217 : root->window_pathkeys = make_pathkeys_for_window(root,
3839 : wc,
3840 : tlist);
3841 : }
3842 : else
3843 391307 : root->window_pathkeys = NIL;
3844 :
3845 : /*
3846 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3847 : * redundant by EquivalenceClass processing. The non-redundant list is
3848 : * kept in root->processed_distinctClause, leaving the original
3849 : * parse->distinctClause alone.
3850 : */
3851 393524 : if (parse->distinctClause)
3852 : {
3853 : bool sortable;
3854 :
3855 : /* Make a copy since pathkey processing can modify the list */
3856 1964 : root->processed_distinctClause = list_copy(parse->distinctClause);
3857 1964 : root->distinct_pathkeys =
3858 1964 : make_pathkeys_for_sortclauses_extended(root,
3859 : &root->processed_distinctClause,
3860 : tlist,
3861 : true,
3862 : false,
3863 : &sortable,
3864 : false);
3865 1964 : if (!sortable)
3866 5 : root->distinct_pathkeys = NIL;
3867 : }
3868 : else
3869 391560 : root->distinct_pathkeys = NIL;
3870 :
3871 393524 : root->sort_pathkeys =
3872 393524 : make_pathkeys_for_sortclauses(root,
3873 : parse->sortClause,
3874 : tlist);
3875 :
3876 : /* setting setop_pathkeys might be useful to the union planner */
3877 393524 : if (qp_extra->setop != NULL)
3878 : {
3879 : List *groupClauses;
3880 : bool sortable;
3881 :
3882 10483 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3883 :
3884 10483 : root->setop_pathkeys =
3885 10483 : make_pathkeys_for_sortclauses_extended(root,
3886 : &groupClauses,
3887 : tlist,
3888 : false,
3889 : false,
3890 : &sortable,
3891 : false);
3892 10483 : if (!sortable)
3893 120 : root->setop_pathkeys = NIL;
3894 : }
3895 : else
3896 383041 : root->setop_pathkeys = NIL;
3897 :
3898 : /*
3899 : * Figure out whether we want a sorted result from query_planner.
3900 : *
3901 : * If we have a sortable GROUP BY clause, then we want a result sorted
3902 : * properly for grouping. Otherwise, if we have window functions to
3903 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3904 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3905 : * we try to produce output that's sufficiently well sorted for the
3906 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3907 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3908 : * for a set operation which can benefit from presorted results and have a
3909 : * sortable targetlist, we want to sort by the target list.
3910 : *
3911 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3912 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3913 : * that might just leave us failing to exploit an available sort order at
3914 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3915 : * much easier, since we know that the parser ensured that one is a
3916 : * superset of the other.
3917 : */
3918 393524 : if (root->group_pathkeys)
3919 5629 : root->query_pathkeys = root->group_pathkeys;
3920 387895 : else if (root->window_pathkeys)
3921 1772 : root->query_pathkeys = root->window_pathkeys;
3922 772246 : else if (list_length(root->distinct_pathkeys) >
3923 386123 : list_length(root->sort_pathkeys))
3924 1594 : root->query_pathkeys = root->distinct_pathkeys;
3925 384529 : else if (root->sort_pathkeys)
3926 56659 : root->query_pathkeys = root->sort_pathkeys;
3927 327870 : else if (root->setop_pathkeys != NIL)
3928 9495 : root->query_pathkeys = root->setop_pathkeys;
3929 : else
3930 318375 : root->query_pathkeys = NIL;
3931 393524 : }
3932 :
3933 : /*
3934 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3935 : *
3936 : * path_rows: number of output rows from scan/join step
3937 : * gd: grouping sets data including list of grouping sets and their clauses
3938 : * target_list: target list containing group clause references
3939 : *
3940 : * If doing grouping sets, we also annotate the gsets data with the estimates
3941 : * for each set and each individual rollup list, with a view to later
3942 : * determining whether some combination of them could be hashed instead.
3943 : */
3944 : static double
3945 41006 : get_number_of_groups(PlannerInfo *root,
3946 : double path_rows,
3947 : grouping_sets_data *gd,
3948 : List *target_list)
3949 : {
3950 41006 : Query *parse = root->parse;
3951 : double dNumGroups;
3952 :
3953 41006 : if (parse->groupClause)
3954 : {
3955 : List *groupExprs;
3956 :
3957 8681 : if (parse->groupingSets)
3958 : {
3959 : /* Add up the estimates for each grouping set */
3960 : ListCell *lc;
3961 :
3962 : Assert(gd); /* keep Coverity happy */
3963 :
3964 836 : dNumGroups = 0;
3965 :
3966 2246 : foreach(lc, gd->rollups)
3967 : {
3968 1410 : RollupData *rollup = lfirst_node(RollupData, lc);
3969 : ListCell *lc2;
3970 : ListCell *lc3;
3971 :
3972 1410 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3973 : target_list);
3974 :
3975 1410 : rollup->numGroups = 0.0;
3976 :
3977 3915 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3978 : {
3979 2505 : List *gset = (List *) lfirst(lc2);
3980 2505 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3981 2505 : double numGroups = estimate_num_groups(root,
3982 : groupExprs,
3983 : path_rows,
3984 : &gset,
3985 : NULL);
3986 :
3987 2505 : gs->numGroups = numGroups;
3988 2505 : rollup->numGroups += numGroups;
3989 : }
3990 :
3991 1410 : dNumGroups += rollup->numGroups;
3992 : }
3993 :
3994 836 : if (gd->hash_sets_idx)
3995 : {
3996 : ListCell *lc2;
3997 :
3998 29 : gd->dNumHashGroups = 0;
3999 :
4000 29 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
4001 : target_list);
4002 :
4003 63 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
4004 : {
4005 34 : List *gset = (List *) lfirst(lc);
4006 34 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
4007 34 : double numGroups = estimate_num_groups(root,
4008 : groupExprs,
4009 : path_rows,
4010 : &gset,
4011 : NULL);
4012 :
4013 34 : gs->numGroups = numGroups;
4014 34 : gd->dNumHashGroups += numGroups;
4015 : }
4016 :
4017 29 : dNumGroups += gd->dNumHashGroups;
4018 : }
4019 : }
4020 : else
4021 : {
4022 : /* Plain GROUP BY -- estimate based on optimized groupClause */
4023 7845 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
4024 : target_list);
4025 :
4026 7845 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
4027 : NULL, NULL);
4028 : }
4029 : }
4030 32325 : else if (parse->groupingSets)
4031 : {
4032 : /* Empty grouping sets ... one result row for each one */
4033 50 : dNumGroups = list_length(parse->groupingSets);
4034 : }
4035 32275 : else if (parse->hasAggs || root->hasHavingQual)
4036 : {
4037 : /* Plain aggregation, one result row */
4038 32275 : dNumGroups = 1;
4039 : }
4040 : else
4041 : {
4042 : /* Not grouping */
4043 0 : dNumGroups = 1;
4044 : }
4045 :
4046 41006 : return dNumGroups;
4047 : }
4048 :
4049 : /*
4050 : * create_grouping_paths
4051 : *
4052 : * Build a new upperrel containing Paths for grouping and/or aggregation.
4053 : * Along the way, we also build an upperrel for Paths which are partially
4054 : * grouped and/or aggregated. A partially grouped and/or aggregated path
4055 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
4056 : * the only partially grouped paths we build are also partial paths; that
4057 : * is, they need a Gather and then a FinalizeAggregate.
4058 : *
4059 : * input_rel: contains the source-data Paths
4060 : * target: the pathtarget for the result Paths to compute
4061 : * gd: grouping sets data including list of grouping sets and their clauses
4062 : *
4063 : * Note: all Paths in input_rel are expected to return the target computed
4064 : * by make_group_input_target.
4065 : */
4066 : static RelOptInfo *
4067 34010 : create_grouping_paths(PlannerInfo *root,
4068 : RelOptInfo *input_rel,
4069 : PathTarget *target,
4070 : bool target_parallel_safe,
4071 : grouping_sets_data *gd)
4072 : {
4073 34010 : Query *parse = root->parse;
4074 : RelOptInfo *grouped_rel;
4075 : RelOptInfo *partially_grouped_rel;
4076 : AggClauseCosts agg_costs;
4077 :
4078 204060 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
4079 34010 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
4080 :
4081 : /*
4082 : * Create grouping relation to hold fully aggregated grouping and/or
4083 : * aggregation paths.
4084 : */
4085 34010 : grouped_rel = make_grouping_rel(root, input_rel, target,
4086 : target_parallel_safe, parse->havingQual);
4087 :
4088 : /*
4089 : * Create either paths for a degenerate grouping or paths for ordinary
4090 : * grouping, as appropriate.
4091 : */
4092 34010 : if (is_degenerate_grouping(root))
4093 35 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
4094 : else
4095 : {
4096 33975 : int flags = 0;
4097 : GroupPathExtraData extra;
4098 :
4099 : /*
4100 : * Determine whether it's possible to perform sort-based
4101 : * implementations of grouping. (Note that if processed_groupClause
4102 : * is empty, grouping_is_sortable() is trivially true, and all the
4103 : * pathkeys_contained_in() tests will succeed too, so that we'll
4104 : * consider every surviving input path.)
4105 : *
4106 : * If we have grouping sets, we might be able to sort some but not all
4107 : * of them; in this case, we need can_sort to be true as long as we
4108 : * must consider any sorted-input plan.
4109 : */
4110 33975 : if ((gd && gd->rollups != NIL)
4111 33094 : || grouping_is_sortable(root->processed_groupClause))
4112 33970 : flags |= GROUPING_CAN_USE_SORT;
4113 :
4114 : /*
4115 : * Determine whether we should consider hash-based implementations of
4116 : * grouping.
4117 : *
4118 : * Hashed aggregation only applies if we're grouping. If we have
4119 : * grouping sets, some groups might be hashable but others not; in
4120 : * this case we set can_hash true as long as there is nothing globally
4121 : * preventing us from hashing (and we should therefore consider plans
4122 : * with hashes).
4123 : *
4124 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
4125 : * BY aggregates. (Doing so would imply storing *all* the input
4126 : * values in the hash table, and/or running many sorts in parallel,
4127 : * either of which seems like a certain loser.) We similarly don't
4128 : * support ordered-set aggregates in hashed aggregation, but that case
4129 : * is also included in the numOrderedAggs count.
4130 : *
4131 : * Note: grouping_is_hashable() is much more expensive to check than
4132 : * the other gating conditions, so we want to do it last.
4133 : */
4134 33975 : if ((parse->groupClause != NIL &&
4135 7516 : root->numOrderedAggs == 0 &&
4136 3246 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
4137 4050 : flags |= GROUPING_CAN_USE_HASH;
4138 :
4139 : /*
4140 : * Determine whether partial aggregation is possible.
4141 : */
4142 33975 : if (can_partial_agg(root))
4143 29869 : flags |= GROUPING_CAN_PARTIAL_AGG;
4144 :
4145 33975 : extra.flags = flags;
4146 33975 : extra.target_parallel_safe = target_parallel_safe;
4147 33975 : extra.havingQual = parse->havingQual;
4148 33975 : extra.targetList = parse->targetList;
4149 33975 : extra.partial_costs_set = false;
4150 :
4151 : /*
4152 : * Determine whether partitionwise aggregation is in theory possible.
4153 : * It can be disabled by the user, and for now, we don't try to
4154 : * support grouping sets. create_ordinary_grouping_paths() will check
4155 : * additional conditions, such as whether input_rel is partitioned.
4156 : */
4157 33975 : if (enable_partitionwise_aggregate && !parse->groupingSets)
4158 580 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
4159 : else
4160 33395 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
4161 :
4162 33975 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
4163 : &agg_costs, gd, &extra,
4164 : &partially_grouped_rel);
4165 : }
4166 :
4167 34006 : set_cheapest(grouped_rel);
4168 34006 : return grouped_rel;
4169 : }
4170 :
4171 : /*
4172 : * make_grouping_rel
4173 : *
4174 : * Create a new grouping rel and set basic properties.
4175 : *
4176 : * input_rel represents the underlying scan/join relation.
4177 : * target is the output expected from the grouping relation.
4178 : */
4179 : static RelOptInfo *
4180 35805 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
4181 : PathTarget *target, bool target_parallel_safe,
4182 : Node *havingQual)
4183 : {
4184 : RelOptInfo *grouped_rel;
4185 :
4186 35805 : if (IS_OTHER_REL(input_rel))
4187 : {
4188 1795 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
4189 : input_rel->relids);
4190 1795 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
4191 : }
4192 : else
4193 : {
4194 : /*
4195 : * By tradition, the relids set for the main grouping relation is
4196 : * NULL. (This could be changed, but might require adjustments
4197 : * elsewhere.)
4198 : */
4199 34010 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
4200 : }
4201 :
4202 : /* Set target. */
4203 35805 : grouped_rel->reltarget = target;
4204 :
4205 : /*
4206 : * If the input relation is not parallel-safe, then the grouped relation
4207 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4208 : * target list and HAVING quals are parallel-safe.
4209 : */
4210 58640 : if (input_rel->consider_parallel && target_parallel_safe &&
4211 22835 : is_parallel_safe(root, havingQual))
4212 22810 : grouped_rel->consider_parallel = true;
4213 :
4214 : /* Assume that the same path generation strategies are allowed */
4215 35805 : grouped_rel->pgs_mask = input_rel->pgs_mask;
4216 :
4217 : /*
4218 : * If the input rel belongs to a single FDW, so does the grouped rel.
4219 : */
4220 35805 : grouped_rel->serverid = input_rel->serverid;
4221 35805 : grouped_rel->userid = input_rel->userid;
4222 35805 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
4223 35805 : grouped_rel->fdwroutine = input_rel->fdwroutine;
4224 :
4225 35805 : return grouped_rel;
4226 : }
4227 :
4228 : /*
4229 : * is_degenerate_grouping
4230 : *
4231 : * A degenerate grouping is one in which the query has a HAVING qual and/or
4232 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
4233 : * grouping sets are all empty).
4234 : */
4235 : static bool
4236 34010 : is_degenerate_grouping(PlannerInfo *root)
4237 : {
4238 34010 : Query *parse = root->parse;
4239 :
4240 32899 : return (root->hasHavingQual || parse->groupingSets) &&
4241 66909 : !parse->hasAggs && parse->groupClause == NIL;
4242 : }
4243 :
4244 : /*
4245 : * create_degenerate_grouping_paths
4246 : *
4247 : * When the grouping is degenerate (see is_degenerate_grouping), we are
4248 : * supposed to emit either zero or one row for each grouping set depending on
4249 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
4250 : * either HAVING or the targetlist, so we actually do not need the FROM table
4251 : * at all! We can just throw away the plan-so-far and generate a Result node.
4252 : * This is a sufficiently unusual corner case that it's not worth contorting
4253 : * the structure of this module to avoid having to generate the earlier paths
4254 : * in the first place.
4255 : */
4256 : static void
4257 35 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4258 : RelOptInfo *grouped_rel)
4259 : {
4260 35 : Query *parse = root->parse;
4261 : int nrows;
4262 : Path *path;
4263 :
4264 35 : nrows = list_length(parse->groupingSets);
4265 35 : if (nrows > 1)
4266 : {
4267 : /*
4268 : * Doesn't seem worthwhile writing code to cons up a generate_series
4269 : * or a values scan to emit multiple rows. Instead just make N clones
4270 : * and append them. (With a volatile HAVING clause, this means you
4271 : * might get between 0 and N output rows. Offhand I think that's
4272 : * desired.)
4273 : */
4274 10 : AppendPathInput append = {0};
4275 :
4276 30 : while (--nrows >= 0)
4277 : {
4278 : path = (Path *)
4279 20 : create_group_result_path(root, grouped_rel,
4280 20 : grouped_rel->reltarget,
4281 20 : (List *) parse->havingQual);
4282 20 : append.subpaths = lappend(append.subpaths, path);
4283 : }
4284 : path = (Path *)
4285 10 : create_append_path(root,
4286 : grouped_rel,
4287 : append,
4288 : NIL,
4289 : NULL,
4290 : 0,
4291 : false,
4292 : -1);
4293 : }
4294 : else
4295 : {
4296 : /* No grouping sets, or just one, so one output row */
4297 : path = (Path *)
4298 25 : create_group_result_path(root, grouped_rel,
4299 25 : grouped_rel->reltarget,
4300 25 : (List *) parse->havingQual);
4301 : }
4302 :
4303 35 : add_path(grouped_rel, path);
4304 35 : }
4305 :
4306 : /*
4307 : * create_ordinary_grouping_paths
4308 : *
4309 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4310 : *
4311 : * We need to consider sorted and hashed aggregation in the same function,
4312 : * because otherwise (1) it would be harder to throw an appropriate error
4313 : * message if neither way works, and (2) we should not allow hashtable size
4314 : * considerations to dissuade us from using hashing if sorting is not possible.
4315 : *
4316 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4317 : * function creates, or to NULL if it doesn't create one.
4318 : */
4319 : static void
4320 35770 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4321 : RelOptInfo *grouped_rel,
4322 : const AggClauseCosts *agg_costs,
4323 : grouping_sets_data *gd,
4324 : GroupPathExtraData *extra,
4325 : RelOptInfo **partially_grouped_rel_p)
4326 : {
4327 35770 : RelOptInfo *partially_grouped_rel = NULL;
4328 35770 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4329 :
4330 : /*
4331 : * If this is the topmost grouping relation or if the parent relation is
4332 : * doing some form of partitionwise aggregation, then we may be able to do
4333 : * it at this level also. However, if the input relation is not
4334 : * partitioned, partitionwise aggregate is impossible.
4335 : */
4336 35770 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4337 2375 : IS_PARTITIONED_REL(input_rel))
4338 : {
4339 : /*
4340 : * If this is the topmost relation or if the parent relation is doing
4341 : * full partitionwise aggregation, then we can do full partitionwise
4342 : * aggregation provided that the GROUP BY clause contains all of the
4343 : * partitioning columns at this level and the collation used by GROUP
4344 : * BY matches the partitioning collation. Otherwise, we can do at
4345 : * most partial partitionwise aggregation. But if partial aggregation
4346 : * is not supported in general then we can't use it for partitionwise
4347 : * aggregation either.
4348 : *
4349 : * Check parse->groupClause not processed_groupClause, because it's
4350 : * okay if some of the partitioning columns were proved redundant.
4351 : */
4352 1360 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4353 640 : group_by_has_partkey(input_rel, extra->targetList,
4354 640 : root->parse->groupClause))
4355 404 : patype = PARTITIONWISE_AGGREGATE_FULL;
4356 316 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4357 281 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4358 : else
4359 35 : patype = PARTITIONWISE_AGGREGATE_NONE;
4360 : }
4361 :
4362 : /*
4363 : * Before generating paths for grouped_rel, we first generate any possible
4364 : * partially grouped paths; that way, later code can easily consider both
4365 : * parallel and non-parallel approaches to grouping.
4366 : */
4367 35770 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4368 : {
4369 : bool force_rel_creation;
4370 :
4371 : /*
4372 : * If we're doing partitionwise aggregation at this level, force
4373 : * creation of a partially_grouped_rel so we can add partitionwise
4374 : * paths to it.
4375 : */
4376 31604 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4377 :
4378 : partially_grouped_rel =
4379 31604 : create_partial_grouping_paths(root,
4380 : grouped_rel,
4381 : input_rel,
4382 : gd,
4383 : extra,
4384 : force_rel_creation);
4385 : }
4386 :
4387 : /* Set out parameter. */
4388 35770 : *partially_grouped_rel_p = partially_grouped_rel;
4389 :
4390 : /* Apply partitionwise aggregation technique, if possible. */
4391 35770 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4392 685 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4393 : partially_grouped_rel, agg_costs,
4394 : gd, patype, extra);
4395 :
4396 : /* If we are doing partial aggregation only, return. */
4397 35770 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4398 : {
4399 : Assert(partially_grouped_rel);
4400 :
4401 713 : if (partially_grouped_rel->pathlist)
4402 713 : set_cheapest(partially_grouped_rel);
4403 :
4404 713 : return;
4405 : }
4406 :
4407 : /* Gather any partially grouped partial paths. */
4408 35057 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4409 2324 : gather_grouping_paths(root, partially_grouped_rel);
4410 :
4411 : /* Now choose the best path(s) for partially_grouped_rel. */
4412 35057 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4413 2492 : set_cheapest(partially_grouped_rel);
4414 :
4415 : /* Build final grouping paths */
4416 35057 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4417 : partially_grouped_rel, agg_costs, gd,
4418 : extra);
4419 :
4420 : /* Give a helpful error if we failed to find any implementation */
4421 35057 : if (grouped_rel->pathlist == NIL)
4422 4 : ereport(ERROR,
4423 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4424 : errmsg("could not implement GROUP BY"),
4425 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4426 :
4427 : /*
4428 : * If there is an FDW that's responsible for all baserels of the query,
4429 : * let it consider adding ForeignPaths.
4430 : */
4431 35053 : if (grouped_rel->fdwroutine &&
4432 170 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4433 169 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4434 : input_rel, grouped_rel,
4435 : extra);
4436 :
4437 : /* Let extensions possibly add some more paths */
4438 35053 : if (create_upper_paths_hook)
4439 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4440 : input_rel, grouped_rel,
4441 : extra);
4442 : }
4443 :
4444 : /*
4445 : * For a given input path, consider the possible ways of doing grouping sets on
4446 : * it, by combinations of hashing and sorting. This can be called multiple
4447 : * times, so it's important that it not scribble on input. No result is
4448 : * returned, but any generated paths are added to grouped_rel.
4449 : */
4450 : static void
4451 1782 : consider_groupingsets_paths(PlannerInfo *root,
4452 : RelOptInfo *grouped_rel,
4453 : Path *path,
4454 : bool is_sorted,
4455 : bool can_hash,
4456 : grouping_sets_data *gd,
4457 : const AggClauseCosts *agg_costs,
4458 : double dNumGroups)
4459 : {
4460 1782 : Query *parse = root->parse;
4461 1782 : Size hash_mem_limit = get_hash_memory_limit();
4462 :
4463 : /*
4464 : * If we're not being offered sorted input, then only consider plans that
4465 : * can be done entirely by hashing.
4466 : *
4467 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4468 : * the input is actually sorted despite not being advertised as such, we
4469 : * prefer to make use of that in order to use less memory.
4470 : *
4471 : * If none of the grouping sets are sortable, then ignore the hash_mem
4472 : * limit and generate a path anyway, since otherwise we'll just fail.
4473 : */
4474 1782 : if (!is_sorted)
4475 : {
4476 807 : List *new_rollups = NIL;
4477 807 : RollupData *unhashed_rollup = NULL;
4478 : List *sets_data;
4479 807 : List *empty_sets_data = NIL;
4480 807 : List *empty_sets = NIL;
4481 : ListCell *lc;
4482 807 : ListCell *l_start = list_head(gd->rollups);
4483 807 : AggStrategy strat = AGG_HASHED;
4484 : double hashsize;
4485 807 : double exclude_groups = 0.0;
4486 :
4487 : Assert(can_hash);
4488 :
4489 : /*
4490 : * If the input is coincidentally sorted usefully (which can happen
4491 : * even if is_sorted is false, since that only means that our caller
4492 : * has set up the sorting for us), then save some hashtable space by
4493 : * making use of that. But we need to watch out for degenerate cases:
4494 : *
4495 : * 1) If there are any empty grouping sets, then group_pathkeys might
4496 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4497 : * there will be a rollup containing only empty groups, and the
4498 : * pathkeys_contained_in test is vacuously true; this is ok.
4499 : *
4500 : * XXX: the above relies on the fact that group_pathkeys is generated
4501 : * from the first rollup. If we add the ability to consider multiple
4502 : * sort orders for grouping input, this assumption might fail.
4503 : *
4504 : * 2) If there are no empty sets and only unsortable sets, then the
4505 : * rollups list will be empty (and thus l_start == NULL), and
4506 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4507 : * pathkeys_contained_in test doesn't cause us to crash.
4508 : */
4509 1609 : if (l_start != NULL &&
4510 802 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4511 : {
4512 40 : unhashed_rollup = lfirst_node(RollupData, l_start);
4513 40 : exclude_groups = unhashed_rollup->numGroups;
4514 40 : l_start = lnext(gd->rollups, l_start);
4515 : }
4516 :
4517 807 : hashsize = estimate_hashagg_tablesize(root,
4518 : path,
4519 : agg_costs,
4520 : dNumGroups - exclude_groups);
4521 :
4522 : /*
4523 : * gd->rollups is empty if we have only unsortable columns to work
4524 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4525 : * sorted-input case to generate usable mixed paths.
4526 : */
4527 807 : if (hashsize > hash_mem_limit && gd->rollups)
4528 15 : return; /* nope, won't fit */
4529 :
4530 : /*
4531 : * We need to burst the existing rollups list into individual grouping
4532 : * sets and recompute a groupClause for each set.
4533 : */
4534 792 : sets_data = list_copy(gd->unsortable_sets);
4535 :
4536 2013 : for_each_cell(lc, gd->rollups, l_start)
4537 : {
4538 1241 : RollupData *rollup = lfirst_node(RollupData, lc);
4539 :
4540 : /*
4541 : * If we find an unhashable rollup that's not been skipped by the
4542 : * "actually sorted" check above, we can't cope; we'd need sorted
4543 : * input (with a different sort order) but we can't get that here.
4544 : * So bail out; we'll get a valid path from the is_sorted case
4545 : * instead.
4546 : *
4547 : * The mere presence of empty grouping sets doesn't make a rollup
4548 : * unhashable (see preprocess_grouping_sets), we handle those
4549 : * specially below.
4550 : */
4551 1241 : if (!rollup->hashable)
4552 20 : return;
4553 :
4554 1221 : sets_data = list_concat(sets_data, rollup->gsets_data);
4555 : }
4556 3053 : foreach(lc, sets_data)
4557 : {
4558 2281 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4559 2281 : List *gset = gs->set;
4560 : RollupData *rollup;
4561 :
4562 2281 : if (gset == NIL)
4563 : {
4564 : /* Empty grouping sets can't be hashed. */
4565 445 : empty_sets_data = lappend(empty_sets_data, gs);
4566 445 : empty_sets = lappend(empty_sets, NIL);
4567 : }
4568 : else
4569 : {
4570 1836 : rollup = makeNode(RollupData);
4571 :
4572 1836 : rollup->groupClause = preprocess_groupclause(root, gset);
4573 1836 : rollup->gsets_data = list_make1(gs);
4574 1836 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4575 : rollup->gsets_data,
4576 : gd->tleref_to_colnum_map);
4577 1836 : rollup->numGroups = gs->numGroups;
4578 1836 : rollup->hashable = true;
4579 1836 : rollup->is_hashed = true;
4580 1836 : new_rollups = lappend(new_rollups, rollup);
4581 : }
4582 : }
4583 :
4584 : /*
4585 : * If we didn't find anything nonempty to hash, then bail. We'll
4586 : * generate a path from the is_sorted case.
4587 : */
4588 772 : if (new_rollups == NIL)
4589 0 : return;
4590 :
4591 : /*
4592 : * If there were empty grouping sets they should have been in the
4593 : * first rollup.
4594 : */
4595 : Assert(!unhashed_rollup || !empty_sets);
4596 :
4597 772 : if (unhashed_rollup)
4598 : {
4599 40 : new_rollups = lappend(new_rollups, unhashed_rollup);
4600 40 : strat = AGG_MIXED;
4601 : }
4602 732 : else if (empty_sets)
4603 : {
4604 405 : RollupData *rollup = makeNode(RollupData);
4605 :
4606 405 : rollup->groupClause = NIL;
4607 405 : rollup->gsets_data = empty_sets_data;
4608 405 : rollup->gsets = empty_sets;
4609 405 : rollup->numGroups = list_length(empty_sets);
4610 405 : rollup->hashable = false;
4611 405 : rollup->is_hashed = false;
4612 405 : new_rollups = lappend(new_rollups, rollup);
4613 405 : strat = AGG_MIXED;
4614 : }
4615 :
4616 772 : add_path(grouped_rel, (Path *)
4617 772 : create_groupingsets_path(root,
4618 : grouped_rel,
4619 : path,
4620 772 : (List *) parse->havingQual,
4621 : strat,
4622 : new_rollups,
4623 : agg_costs));
4624 772 : return;
4625 : }
4626 :
4627 : /*
4628 : * If we have sorted input but nothing we can do with it, bail.
4629 : */
4630 975 : if (gd->rollups == NIL)
4631 0 : return;
4632 :
4633 : /*
4634 : * Given sorted input, we try and make two paths: one sorted and one mixed
4635 : * sort/hash. (We need to try both because hashagg might be disabled, or
4636 : * some columns might not be sortable.)
4637 : *
4638 : * can_hash is passed in as false if some obstacle elsewhere (such as
4639 : * ordered aggs) means that we shouldn't consider hashing at all.
4640 : */
4641 975 : if (can_hash && gd->any_hashable)
4642 : {
4643 896 : List *rollups = NIL;
4644 896 : List *hash_sets = list_copy(gd->unsortable_sets);
4645 896 : double availspace = hash_mem_limit;
4646 : ListCell *lc;
4647 :
4648 : /*
4649 : * Account first for space needed for groups we can't sort at all.
4650 : */
4651 896 : availspace -= estimate_hashagg_tablesize(root,
4652 : path,
4653 : agg_costs,
4654 : gd->dNumHashGroups);
4655 :
4656 896 : if (availspace > 0 && list_length(gd->rollups) > 1)
4657 : {
4658 : double scale;
4659 501 : int num_rollups = list_length(gd->rollups);
4660 : int k_capacity;
4661 501 : int *k_weights = palloc(num_rollups * sizeof(int));
4662 501 : Bitmapset *hash_items = NULL;
4663 : int i;
4664 :
4665 : /*
4666 : * We treat this as a knapsack problem: the knapsack capacity
4667 : * represents hash_mem, the item weights are the estimated memory
4668 : * usage of the hashtables needed to implement a single rollup,
4669 : * and we really ought to use the cost saving as the item value;
4670 : * however, currently the costs assigned to sort nodes don't
4671 : * reflect the comparison costs well, and so we treat all items as
4672 : * of equal value (each rollup we hash instead saves us one sort).
4673 : *
4674 : * To use the discrete knapsack, we need to scale the values to a
4675 : * reasonably small bounded range. We choose to allow a 5% error
4676 : * margin; we have no more than 4096 rollups in the worst possible
4677 : * case, which with a 5% error margin will require a bit over 42MB
4678 : * of workspace. (Anyone wanting to plan queries that complex had
4679 : * better have the memory for it. In more reasonable cases, with
4680 : * no more than a couple of dozen rollups, the memory usage will
4681 : * be negligible.)
4682 : *
4683 : * k_capacity is naturally bounded, but we clamp the values for
4684 : * scale and weight (below) to avoid overflows or underflows (or
4685 : * uselessly trying to use a scale factor less than 1 byte).
4686 : */
4687 501 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4688 501 : k_capacity = (int) floor(availspace / scale);
4689 :
4690 : /*
4691 : * We leave the first rollup out of consideration since it's the
4692 : * one that matches the input sort order. We assign indexes "i"
4693 : * to only those entries considered for hashing; the second loop,
4694 : * below, must use the same condition.
4695 : */
4696 501 : i = 0;
4697 1222 : for_each_from(lc, gd->rollups, 1)
4698 : {
4699 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4700 :
4701 721 : if (rollup->hashable)
4702 : {
4703 721 : double sz = estimate_hashagg_tablesize(root,
4704 : path,
4705 : agg_costs,
4706 : rollup->numGroups);
4707 :
4708 : /*
4709 : * If sz is enormous, but hash_mem (and hence scale) is
4710 : * small, avoid integer overflow here.
4711 : */
4712 721 : k_weights[i] = (int) Min(floor(sz / scale),
4713 : k_capacity + 1.0);
4714 721 : ++i;
4715 : }
4716 : }
4717 :
4718 : /*
4719 : * Apply knapsack algorithm; compute the set of items which
4720 : * maximizes the value stored (in this case the number of sorts
4721 : * saved) while keeping the total size (approximately) within
4722 : * capacity.
4723 : */
4724 501 : if (i > 0)
4725 501 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4726 :
4727 501 : if (!bms_is_empty(hash_items))
4728 : {
4729 501 : rollups = list_make1(linitial(gd->rollups));
4730 :
4731 501 : i = 0;
4732 1222 : for_each_from(lc, gd->rollups, 1)
4733 : {
4734 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4735 :
4736 721 : if (rollup->hashable)
4737 : {
4738 721 : if (bms_is_member(i, hash_items))
4739 691 : hash_sets = list_concat(hash_sets,
4740 691 : rollup->gsets_data);
4741 : else
4742 30 : rollups = lappend(rollups, rollup);
4743 721 : ++i;
4744 : }
4745 : else
4746 0 : rollups = lappend(rollups, rollup);
4747 : }
4748 : }
4749 : }
4750 :
4751 896 : if (!rollups && hash_sets)
4752 20 : rollups = list_copy(gd->rollups);
4753 :
4754 1703 : foreach(lc, hash_sets)
4755 : {
4756 807 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4757 807 : RollupData *rollup = makeNode(RollupData);
4758 :
4759 : Assert(gs->set != NIL);
4760 :
4761 807 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4762 807 : rollup->gsets_data = list_make1(gs);
4763 807 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4764 : rollup->gsets_data,
4765 : gd->tleref_to_colnum_map);
4766 807 : rollup->numGroups = gs->numGroups;
4767 807 : rollup->hashable = true;
4768 807 : rollup->is_hashed = true;
4769 807 : rollups = lcons(rollup, rollups);
4770 : }
4771 :
4772 896 : if (rollups)
4773 : {
4774 521 : add_path(grouped_rel, (Path *)
4775 521 : create_groupingsets_path(root,
4776 : grouped_rel,
4777 : path,
4778 521 : (List *) parse->havingQual,
4779 : AGG_MIXED,
4780 : rollups,
4781 : agg_costs));
4782 : }
4783 : }
4784 :
4785 : /*
4786 : * Now try the simple sorted case.
4787 : */
4788 975 : if (!gd->unsortable_sets)
4789 951 : add_path(grouped_rel, (Path *)
4790 951 : create_groupingsets_path(root,
4791 : grouped_rel,
4792 : path,
4793 951 : (List *) parse->havingQual,
4794 : AGG_SORTED,
4795 : gd->rollups,
4796 : agg_costs));
4797 : }
4798 :
4799 : /*
4800 : * create_window_paths
4801 : *
4802 : * Build a new upperrel containing Paths for window-function evaluation.
4803 : *
4804 : * input_rel: contains the source-data Paths
4805 : * input_target: result of make_window_input_target
4806 : * output_target: what the topmost WindowAggPath should return
4807 : * wflists: result of find_window_functions
4808 : * activeWindows: result of select_active_windows
4809 : *
4810 : * Note: all Paths in input_rel are expected to return input_target.
4811 : */
4812 : static RelOptInfo *
4813 2217 : create_window_paths(PlannerInfo *root,
4814 : RelOptInfo *input_rel,
4815 : PathTarget *input_target,
4816 : PathTarget *output_target,
4817 : bool output_target_parallel_safe,
4818 : WindowFuncLists *wflists,
4819 : List *activeWindows)
4820 : {
4821 : RelOptInfo *window_rel;
4822 : ListCell *lc;
4823 :
4824 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4825 2217 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4826 :
4827 : /*
4828 : * If the input relation is not parallel-safe, then the window relation
4829 : * can't be parallel-safe, either. Otherwise, we need to examine the
4830 : * target list and active windows for non-parallel-safe constructs.
4831 : */
4832 2217 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4833 0 : is_parallel_safe(root, (Node *) activeWindows))
4834 0 : window_rel->consider_parallel = true;
4835 :
4836 : /*
4837 : * If the input rel belongs to a single FDW, so does the window rel.
4838 : */
4839 2217 : window_rel->serverid = input_rel->serverid;
4840 2217 : window_rel->userid = input_rel->userid;
4841 2217 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4842 2217 : window_rel->fdwroutine = input_rel->fdwroutine;
4843 :
4844 : /*
4845 : * Consider computing window functions starting from the existing
4846 : * cheapest-total path (which will likely require a sort) as well as any
4847 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4848 : */
4849 4696 : foreach(lc, input_rel->pathlist)
4850 : {
4851 2479 : Path *path = (Path *) lfirst(lc);
4852 : int presorted_keys;
4853 :
4854 2741 : if (path == input_rel->cheapest_total_path ||
4855 262 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4856 116 : &presorted_keys) ||
4857 116 : presorted_keys > 0)
4858 2384 : create_one_window_path(root,
4859 : window_rel,
4860 : path,
4861 : input_target,
4862 : output_target,
4863 : wflists,
4864 : activeWindows);
4865 : }
4866 :
4867 : /*
4868 : * If there is an FDW that's responsible for all baserels of the query,
4869 : * let it consider adding ForeignPaths.
4870 : */
4871 2217 : if (window_rel->fdwroutine &&
4872 6 : window_rel->fdwroutine->GetForeignUpperPaths)
4873 6 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4874 : input_rel, window_rel,
4875 : NULL);
4876 :
4877 : /* Let extensions possibly add some more paths */
4878 2217 : if (create_upper_paths_hook)
4879 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4880 : input_rel, window_rel, NULL);
4881 :
4882 : /* Now choose the best path(s) */
4883 2217 : set_cheapest(window_rel);
4884 :
4885 2217 : return window_rel;
4886 : }
4887 :
4888 : /*
4889 : * Stack window-function implementation steps atop the given Path, and
4890 : * add the result to window_rel.
4891 : *
4892 : * window_rel: upperrel to contain result
4893 : * path: input Path to use (must return input_target)
4894 : * input_target: result of make_window_input_target
4895 : * output_target: what the topmost WindowAggPath should return
4896 : * wflists: result of find_window_functions
4897 : * activeWindows: result of select_active_windows
4898 : */
4899 : static void
4900 2384 : create_one_window_path(PlannerInfo *root,
4901 : RelOptInfo *window_rel,
4902 : Path *path,
4903 : PathTarget *input_target,
4904 : PathTarget *output_target,
4905 : WindowFuncLists *wflists,
4906 : List *activeWindows)
4907 : {
4908 : PathTarget *window_target;
4909 : ListCell *l;
4910 2384 : List *topqual = NIL;
4911 :
4912 : /*
4913 : * Since each window clause could require a different sort order, we stack
4914 : * up a WindowAgg node for each clause, with sort steps between them as
4915 : * needed. (We assume that select_active_windows chose a good order for
4916 : * executing the clauses in.)
4917 : *
4918 : * input_target should contain all Vars and Aggs needed for the result.
4919 : * (In some cases we wouldn't need to propagate all of these all the way
4920 : * to the top, since they might only be needed as inputs to WindowFuncs.
4921 : * It's probably not worth trying to optimize that though.) It must also
4922 : * contain all window partitioning and sorting expressions, to ensure
4923 : * they're computed only once at the bottom of the stack (that's critical
4924 : * for volatile functions). As we climb up the stack, we'll add outputs
4925 : * for the WindowFuncs computed at each level.
4926 : */
4927 2384 : window_target = input_target;
4928 :
4929 4923 : foreach(l, activeWindows)
4930 : {
4931 2539 : WindowClause *wc = lfirst_node(WindowClause, l);
4932 : List *window_pathkeys;
4933 2539 : List *runcondition = NIL;
4934 : int presorted_keys;
4935 : bool is_sorted;
4936 : bool topwindow;
4937 : ListCell *lc2;
4938 :
4939 2539 : window_pathkeys = make_pathkeys_for_window(root,
4940 : wc,
4941 : root->processed_tlist);
4942 :
4943 2539 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4944 : path->pathkeys,
4945 : &presorted_keys);
4946 :
4947 : /* Sort if necessary */
4948 2539 : if (!is_sorted)
4949 : {
4950 : /*
4951 : * No presorted keys or incremental sort disabled, just perform a
4952 : * complete sort.
4953 : */
4954 1826 : if (presorted_keys == 0 || !enable_incremental_sort)
4955 1775 : path = (Path *) create_sort_path(root, window_rel,
4956 : path,
4957 : window_pathkeys,
4958 : -1.0);
4959 : else
4960 : {
4961 : /*
4962 : * Since we have presorted keys and incremental sort is
4963 : * enabled, just use incremental sort.
4964 : */
4965 51 : path = (Path *) create_incremental_sort_path(root,
4966 : window_rel,
4967 : path,
4968 : window_pathkeys,
4969 : presorted_keys,
4970 : -1.0);
4971 : }
4972 : }
4973 :
4974 2539 : if (lnext(activeWindows, l))
4975 : {
4976 : /*
4977 : * Add the current WindowFuncs to the output target for this
4978 : * intermediate WindowAggPath. We must copy window_target to
4979 : * avoid changing the previous path's target.
4980 : *
4981 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4982 : * we do need to account for the increase in tlist width.
4983 : */
4984 155 : int64 tuple_width = window_target->width;
4985 :
4986 155 : window_target = copy_pathtarget(window_target);
4987 370 : foreach(lc2, wflists->windowFuncs[wc->winref])
4988 : {
4989 215 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4990 :
4991 215 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4992 215 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4993 : }
4994 155 : window_target->width = clamp_width_est(tuple_width);
4995 : }
4996 : else
4997 : {
4998 : /* Install the goal target in the topmost WindowAgg */
4999 2384 : window_target = output_target;
5000 : }
5001 :
5002 : /* mark the final item in the list as the top-level window */
5003 2539 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
5004 :
5005 : /*
5006 : * Collect the WindowFuncRunConditions from each WindowFunc and
5007 : * convert them into OpExprs
5008 : */
5009 5811 : foreach(lc2, wflists->windowFuncs[wc->winref])
5010 : {
5011 : ListCell *lc3;
5012 3272 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5013 :
5014 3422 : foreach(lc3, wfunc->runCondition)
5015 : {
5016 150 : WindowFuncRunCondition *wfuncrc =
5017 : lfirst_node(WindowFuncRunCondition, lc3);
5018 : Expr *opexpr;
5019 : Expr *leftop;
5020 : Expr *rightop;
5021 :
5022 150 : if (wfuncrc->wfunc_left)
5023 : {
5024 135 : leftop = (Expr *) copyObject(wfunc);
5025 135 : rightop = copyObject(wfuncrc->arg);
5026 : }
5027 : else
5028 : {
5029 15 : leftop = copyObject(wfuncrc->arg);
5030 15 : rightop = (Expr *) copyObject(wfunc);
5031 : }
5032 :
5033 150 : opexpr = make_opclause(wfuncrc->opno,
5034 : BOOLOID,
5035 : false,
5036 : leftop,
5037 : rightop,
5038 : InvalidOid,
5039 : wfuncrc->inputcollid);
5040 :
5041 150 : runcondition = lappend(runcondition, opexpr);
5042 :
5043 150 : if (!topwindow)
5044 20 : topqual = lappend(topqual, opexpr);
5045 : }
5046 : }
5047 :
5048 : path = (Path *)
5049 2539 : create_windowagg_path(root, window_rel, path, window_target,
5050 2539 : wflists->windowFuncs[wc->winref],
5051 : runcondition, wc,
5052 : topwindow ? topqual : NIL, topwindow);
5053 : }
5054 :
5055 2384 : add_path(window_rel, path);
5056 2384 : }
5057 :
5058 : /*
5059 : * create_distinct_paths
5060 : *
5061 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
5062 : *
5063 : * input_rel: contains the source-data Paths
5064 : * target: the pathtarget for the result Paths to compute
5065 : *
5066 : * Note: input paths should already compute the desired pathtarget, since
5067 : * Sort/Unique won't project anything.
5068 : */
5069 : static RelOptInfo *
5070 1964 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5071 : PathTarget *target)
5072 : {
5073 : RelOptInfo *distinct_rel;
5074 :
5075 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
5076 1964 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
5077 :
5078 : /*
5079 : * We don't compute anything at this level, so distinct_rel will be
5080 : * parallel-safe if the input rel is parallel-safe. In particular, if
5081 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
5082 : * output those expressions, and will not be parallel-safe unless those
5083 : * expressions are parallel-safe.
5084 : */
5085 1964 : distinct_rel->consider_parallel = input_rel->consider_parallel;
5086 :
5087 : /*
5088 : * If the input rel belongs to a single FDW, so does the distinct_rel.
5089 : */
5090 1964 : distinct_rel->serverid = input_rel->serverid;
5091 1964 : distinct_rel->userid = input_rel->userid;
5092 1964 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5093 1964 : distinct_rel->fdwroutine = input_rel->fdwroutine;
5094 :
5095 : /* build distinct paths based on input_rel's pathlist */
5096 1964 : create_final_distinct_paths(root, input_rel, distinct_rel);
5097 :
5098 : /* now build distinct paths based on input_rel's partial_pathlist */
5099 1964 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
5100 :
5101 : /* Give a helpful error if we failed to create any paths */
5102 1964 : if (distinct_rel->pathlist == NIL)
5103 0 : ereport(ERROR,
5104 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5105 : errmsg("could not implement DISTINCT"),
5106 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
5107 :
5108 : /*
5109 : * If there is an FDW that's responsible for all baserels of the query,
5110 : * let it consider adding ForeignPaths.
5111 : */
5112 1964 : if (distinct_rel->fdwroutine &&
5113 8 : distinct_rel->fdwroutine->GetForeignUpperPaths)
5114 8 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5115 : UPPERREL_DISTINCT,
5116 : input_rel,
5117 : distinct_rel,
5118 : NULL);
5119 :
5120 : /* Let extensions possibly add some more paths */
5121 1964 : if (create_upper_paths_hook)
5122 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
5123 : distinct_rel, NULL);
5124 :
5125 : /* Now choose the best path(s) */
5126 1964 : set_cheapest(distinct_rel);
5127 :
5128 1964 : return distinct_rel;
5129 : }
5130 :
5131 : /*
5132 : * create_partial_distinct_paths
5133 : *
5134 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
5135 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
5136 : * paths on top and add a final unique/aggregate path to remove any duplicate
5137 : * produced from combining rows from parallel workers.
5138 : */
5139 : static void
5140 1964 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5141 : RelOptInfo *final_distinct_rel,
5142 : PathTarget *target)
5143 : {
5144 : RelOptInfo *partial_distinct_rel;
5145 : Query *parse;
5146 : List *distinctExprs;
5147 : double numDistinctRows;
5148 : Path *cheapest_partial_path;
5149 : ListCell *lc;
5150 :
5151 : /* nothing to do when there are no partial paths in the input rel */
5152 1964 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
5153 1874 : return;
5154 :
5155 90 : parse = root->parse;
5156 :
5157 : /* can't do parallel DISTINCT ON */
5158 90 : if (parse->hasDistinctOn)
5159 0 : return;
5160 :
5161 90 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
5162 : NULL);
5163 90 : partial_distinct_rel->reltarget = target;
5164 90 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
5165 :
5166 : /*
5167 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
5168 : */
5169 90 : partial_distinct_rel->serverid = input_rel->serverid;
5170 90 : partial_distinct_rel->userid = input_rel->userid;
5171 90 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5172 90 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
5173 :
5174 90 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5175 :
5176 90 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5177 : parse->targetList);
5178 :
5179 : /* estimate how many distinct rows we'll get from each worker */
5180 90 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5181 : cheapest_partial_path->rows,
5182 : NULL, NULL);
5183 :
5184 : /*
5185 : * Try sorting the cheapest path and incrementally sorting any paths with
5186 : * presorted keys and put a unique paths atop of those. We'll also
5187 : * attempt to reorder the required pathkeys to match the input path's
5188 : * pathkeys as much as possible, in hopes of avoiding a possible need to
5189 : * re-sort.
5190 : */
5191 90 : if (grouping_is_sortable(root->processed_distinctClause))
5192 : {
5193 193 : foreach(lc, input_rel->partial_pathlist)
5194 : {
5195 103 : Path *input_path = (Path *) lfirst(lc);
5196 : Path *sorted_path;
5197 103 : List *useful_pathkeys_list = NIL;
5198 :
5199 : useful_pathkeys_list =
5200 103 : get_useful_pathkeys_for_distinct(root,
5201 : root->distinct_pathkeys,
5202 : input_path->pathkeys);
5203 : Assert(list_length(useful_pathkeys_list) > 0);
5204 :
5205 319 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5206 : {
5207 113 : sorted_path = make_ordered_path(root,
5208 : partial_distinct_rel,
5209 : input_path,
5210 : cheapest_partial_path,
5211 : useful_pathkeys,
5212 : -1.0);
5213 :
5214 113 : if (sorted_path == NULL)
5215 8 : continue;
5216 :
5217 : /*
5218 : * An empty distinct_pathkeys means all tuples have the same
5219 : * value for the DISTINCT clause. See
5220 : * create_final_distinct_paths()
5221 : */
5222 105 : if (root->distinct_pathkeys == NIL)
5223 : {
5224 : Node *limitCount;
5225 :
5226 5 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5227 : sizeof(int64),
5228 : Int64GetDatum(1), false,
5229 : true);
5230 :
5231 : /*
5232 : * Apply a LimitPath onto the partial path to restrict the
5233 : * tuples from each worker to 1.
5234 : * create_final_distinct_paths will need to apply an
5235 : * additional LimitPath to restrict this to a single row
5236 : * after the Gather node. If the query already has a
5237 : * LIMIT clause, then we could end up with three Limit
5238 : * nodes in the final plan. Consolidating the top two of
5239 : * these could be done, but does not seem worth troubling
5240 : * over.
5241 : */
5242 5 : add_partial_path(partial_distinct_rel, (Path *)
5243 5 : create_limit_path(root, partial_distinct_rel,
5244 : sorted_path,
5245 : NULL,
5246 : limitCount,
5247 : LIMIT_OPTION_COUNT,
5248 : 0, 1));
5249 : }
5250 : else
5251 : {
5252 100 : add_partial_path(partial_distinct_rel, (Path *)
5253 100 : create_unique_path(root, partial_distinct_rel,
5254 : sorted_path,
5255 100 : list_length(root->distinct_pathkeys),
5256 : numDistinctRows));
5257 : }
5258 : }
5259 : }
5260 : }
5261 :
5262 : /*
5263 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5264 : * we're not on the hook to ensure we do our best to create at least one
5265 : * path here, we treat enable_hashagg as a hard off-switch rather than the
5266 : * slightly softer variant in create_final_distinct_paths.
5267 : */
5268 90 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5269 : {
5270 65 : add_partial_path(partial_distinct_rel, (Path *)
5271 65 : create_agg_path(root,
5272 : partial_distinct_rel,
5273 : cheapest_partial_path,
5274 : cheapest_partial_path->pathtarget,
5275 : AGG_HASHED,
5276 : AGGSPLIT_SIMPLE,
5277 : root->processed_distinctClause,
5278 : NIL,
5279 : NULL,
5280 : numDistinctRows));
5281 : }
5282 :
5283 : /*
5284 : * If there is an FDW that's responsible for all baserels of the query,
5285 : * let it consider adding ForeignPaths.
5286 : */
5287 90 : if (partial_distinct_rel->fdwroutine &&
5288 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5289 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5290 : UPPERREL_PARTIAL_DISTINCT,
5291 : input_rel,
5292 : partial_distinct_rel,
5293 : NULL);
5294 :
5295 : /* Let extensions possibly add some more partial paths */
5296 90 : if (create_upper_paths_hook)
5297 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5298 : input_rel, partial_distinct_rel, NULL);
5299 :
5300 90 : if (partial_distinct_rel->partial_pathlist != NIL)
5301 : {
5302 90 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5303 90 : set_cheapest(partial_distinct_rel);
5304 :
5305 : /*
5306 : * Finally, create paths to distinctify the final result. This step
5307 : * is needed to remove any duplicates due to combining rows from
5308 : * parallel workers.
5309 : */
5310 90 : create_final_distinct_paths(root, partial_distinct_rel,
5311 : final_distinct_rel);
5312 : }
5313 : }
5314 :
5315 : /*
5316 : * create_final_distinct_paths
5317 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5318 : *
5319 : * input_rel: contains the source-data paths
5320 : * distinct_rel: destination relation for storing created paths
5321 : */
5322 : static RelOptInfo *
5323 2054 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5324 : RelOptInfo *distinct_rel)
5325 : {
5326 2054 : Query *parse = root->parse;
5327 2054 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5328 : double numDistinctRows;
5329 : bool allow_hash;
5330 :
5331 : /* Estimate number of distinct rows there will be */
5332 2054 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5333 1995 : root->hasHavingQual)
5334 : {
5335 : /*
5336 : * If there was grouping or aggregation, use the number of input rows
5337 : * as the estimated number of DISTINCT rows (ie, assume the input is
5338 : * already mostly unique).
5339 : */
5340 59 : numDistinctRows = cheapest_input_path->rows;
5341 : }
5342 : else
5343 : {
5344 : /*
5345 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5346 : */
5347 : List *distinctExprs;
5348 :
5349 1995 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5350 : parse->targetList);
5351 1995 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5352 : cheapest_input_path->rows,
5353 : NULL, NULL);
5354 : }
5355 :
5356 : /*
5357 : * Consider sort-based implementations of DISTINCT, if possible.
5358 : */
5359 2054 : if (grouping_is_sortable(root->processed_distinctClause))
5360 : {
5361 : /*
5362 : * Firstly, if we have any adequately-presorted paths, just stick a
5363 : * Unique node on those. We also, consider doing an explicit sort of
5364 : * the cheapest input path and Unique'ing that. If any paths have
5365 : * presorted keys then we'll create an incremental sort atop of those
5366 : * before adding a unique node on the top. We'll also attempt to
5367 : * reorder the required pathkeys to match the input path's pathkeys as
5368 : * much as possible, in hopes of avoiding a possible need to re-sort.
5369 : *
5370 : * When we have DISTINCT ON, we must sort by the more rigorous of
5371 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5372 : * Also, if we do have to do an explicit sort, we might as well use
5373 : * the more rigorous ordering to avoid a second sort later. (Note
5374 : * that the parser will have ensured that one clause is a prefix of
5375 : * the other.)
5376 : */
5377 : List *needed_pathkeys;
5378 : ListCell *lc;
5379 2049 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5380 :
5381 2236 : if (parse->hasDistinctOn &&
5382 187 : list_length(root->distinct_pathkeys) <
5383 187 : list_length(root->sort_pathkeys))
5384 52 : needed_pathkeys = root->sort_pathkeys;
5385 : else
5386 1997 : needed_pathkeys = root->distinct_pathkeys;
5387 :
5388 5477 : foreach(lc, input_rel->pathlist)
5389 : {
5390 3428 : Path *input_path = (Path *) lfirst(lc);
5391 : Path *sorted_path;
5392 3428 : List *useful_pathkeys_list = NIL;
5393 :
5394 : useful_pathkeys_list =
5395 3428 : get_useful_pathkeys_for_distinct(root,
5396 : needed_pathkeys,
5397 : input_path->pathkeys);
5398 : Assert(list_length(useful_pathkeys_list) > 0);
5399 :
5400 10784 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5401 : {
5402 3928 : sorted_path = make_ordered_path(root,
5403 : distinct_rel,
5404 : input_path,
5405 : cheapest_input_path,
5406 : useful_pathkeys,
5407 : limittuples);
5408 :
5409 3928 : if (sorted_path == NULL)
5410 531 : continue;
5411 :
5412 : /*
5413 : * distinct_pathkeys may have become empty if all of the
5414 : * pathkeys were determined to be redundant. If all of the
5415 : * pathkeys are redundant then each DISTINCT target must only
5416 : * allow a single value, therefore all resulting tuples must
5417 : * be identical (or at least indistinguishable by an equality
5418 : * check). We can uniquify these tuples simply by just taking
5419 : * the first tuple. All we do here is add a path to do "LIMIT
5420 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5421 : * still have a non-NIL sort_pathkeys list, so we must still
5422 : * only do this with paths which are correctly sorted by
5423 : * sort_pathkeys.
5424 : */
5425 3397 : if (root->distinct_pathkeys == NIL)
5426 : {
5427 : Node *limitCount;
5428 :
5429 96 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5430 : sizeof(int64),
5431 : Int64GetDatum(1), false,
5432 : true);
5433 :
5434 : /*
5435 : * If the query already has a LIMIT clause, then we could
5436 : * end up with a duplicate LimitPath in the final plan.
5437 : * That does not seem worth troubling over too much.
5438 : */
5439 96 : add_path(distinct_rel, (Path *)
5440 96 : create_limit_path(root, distinct_rel, sorted_path,
5441 : NULL, limitCount,
5442 : LIMIT_OPTION_COUNT, 0, 1));
5443 : }
5444 : else
5445 : {
5446 3301 : add_path(distinct_rel, (Path *)
5447 3301 : create_unique_path(root, distinct_rel,
5448 : sorted_path,
5449 3301 : list_length(root->distinct_pathkeys),
5450 : numDistinctRows));
5451 : }
5452 : }
5453 : }
5454 : }
5455 :
5456 : /*
5457 : * Consider hash-based implementations of DISTINCT, if possible.
5458 : *
5459 : * If we were not able to make any other types of path, we *must* hash or
5460 : * die trying. If we do have other choices, there are two things that
5461 : * should prevent selection of hashing: if the query uses DISTINCT ON
5462 : * (because it won't really have the expected behavior if we hash), or if
5463 : * enable_hashagg is off.
5464 : *
5465 : * Note: grouping_is_hashable() is much more expensive to check than the
5466 : * other gating conditions, so we want to do it last.
5467 : */
5468 2054 : if (distinct_rel->pathlist == NIL)
5469 5 : allow_hash = true; /* we have no alternatives */
5470 2049 : else if (parse->hasDistinctOn || !enable_hashagg)
5471 312 : allow_hash = false; /* policy-based decision not to hash */
5472 : else
5473 1737 : allow_hash = true; /* default */
5474 :
5475 2054 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5476 : {
5477 : /* Generate hashed aggregate path --- no sort needed */
5478 1742 : add_path(distinct_rel, (Path *)
5479 1742 : create_agg_path(root,
5480 : distinct_rel,
5481 : cheapest_input_path,
5482 : cheapest_input_path->pathtarget,
5483 : AGG_HASHED,
5484 : AGGSPLIT_SIMPLE,
5485 : root->processed_distinctClause,
5486 : NIL,
5487 : NULL,
5488 : numDistinctRows));
5489 : }
5490 :
5491 2054 : return distinct_rel;
5492 : }
5493 :
5494 : /*
5495 : * get_useful_pathkeys_for_distinct
5496 : * Get useful orderings of pathkeys for distinctClause by reordering
5497 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5498 : *
5499 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5500 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5501 : */
5502 : static List *
5503 3531 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5504 : List *path_pathkeys)
5505 : {
5506 3531 : List *useful_pathkeys_list = NIL;
5507 3531 : List *useful_pathkeys = NIL;
5508 :
5509 : /* always include the given 'needed_pathkeys' */
5510 3531 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5511 : needed_pathkeys);
5512 :
5513 3531 : if (!enable_distinct_reordering)
5514 0 : return useful_pathkeys_list;
5515 :
5516 : /*
5517 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5518 : * that match 'needed_pathkeys', but only up to the longest matching
5519 : * prefix.
5520 : *
5521 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5522 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5523 : * the desired behavior.
5524 : */
5525 8805 : foreach_node(PathKey, pathkey, path_pathkeys)
5526 : {
5527 : /*
5528 : * The PathKey nodes are canonical, so they can be checked for
5529 : * equality by simple pointer comparison.
5530 : */
5531 1765 : if (!list_member_ptr(needed_pathkeys, pathkey))
5532 7 : break;
5533 1758 : if (root->parse->hasDistinctOn &&
5534 140 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5535 15 : break;
5536 :
5537 1743 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5538 : }
5539 :
5540 : /* If no match at all, no point in reordering needed_pathkeys */
5541 3531 : if (useful_pathkeys == NIL)
5542 2016 : return useful_pathkeys_list;
5543 :
5544 : /*
5545 : * If not full match, the resulting pathkey list is not useful without
5546 : * incremental sort.
5547 : */
5548 1515 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5549 924 : !enable_incremental_sort)
5550 48 : return useful_pathkeys_list;
5551 :
5552 : /* Append the remaining PathKey nodes in needed_pathkeys */
5553 1467 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5554 : needed_pathkeys);
5555 :
5556 : /*
5557 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5558 : * just drop it.
5559 : */
5560 1467 : if (compare_pathkeys(needed_pathkeys,
5561 : useful_pathkeys) == PATHKEYS_EQUAL)
5562 957 : return useful_pathkeys_list;
5563 :
5564 510 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5565 : useful_pathkeys);
5566 :
5567 510 : return useful_pathkeys_list;
5568 : }
5569 :
5570 : /*
5571 : * create_ordered_paths
5572 : *
5573 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5574 : *
5575 : * All paths in the result must satisfy the ORDER BY ordering.
5576 : * The only new paths we need consider are an explicit full sort
5577 : * and incremental sort on the cheapest-total existing path.
5578 : *
5579 : * input_rel: contains the source-data Paths
5580 : * target: the output tlist the result Paths must emit
5581 : * limit_tuples: estimated bound on the number of output tuples,
5582 : * or -1 if no LIMIT or couldn't estimate
5583 : *
5584 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5585 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5586 : */
5587 : static RelOptInfo *
5588 62146 : create_ordered_paths(PlannerInfo *root,
5589 : RelOptInfo *input_rel,
5590 : PathTarget *target,
5591 : bool target_parallel_safe,
5592 : double limit_tuples)
5593 : {
5594 62146 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5595 : RelOptInfo *ordered_rel;
5596 : ListCell *lc;
5597 :
5598 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5599 62146 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5600 :
5601 : /*
5602 : * If the input relation is not parallel-safe, then the ordered relation
5603 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5604 : * target list is parallel-safe.
5605 : */
5606 62146 : if (input_rel->consider_parallel && target_parallel_safe)
5607 43506 : ordered_rel->consider_parallel = true;
5608 :
5609 : /* Assume that the same path generation strategies are allowed. */
5610 62146 : ordered_rel->pgs_mask = input_rel->pgs_mask;
5611 :
5612 : /*
5613 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5614 : */
5615 62146 : ordered_rel->serverid = input_rel->serverid;
5616 62146 : ordered_rel->userid = input_rel->userid;
5617 62146 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5618 62146 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5619 :
5620 158752 : foreach(lc, input_rel->pathlist)
5621 : {
5622 96606 : Path *input_path = (Path *) lfirst(lc);
5623 : Path *sorted_path;
5624 : bool is_sorted;
5625 : int presorted_keys;
5626 :
5627 96606 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5628 : input_path->pathkeys, &presorted_keys);
5629 :
5630 96606 : if (is_sorted)
5631 37203 : sorted_path = input_path;
5632 : else
5633 : {
5634 : /*
5635 : * Try at least sorting the cheapest path and also try
5636 : * incrementally sorting any path which is partially sorted
5637 : * already (no need to deal with paths which have presorted keys
5638 : * when incremental sort is disabled unless it's the cheapest
5639 : * input path).
5640 : */
5641 59403 : if (input_path != cheapest_input_path &&
5642 4844 : (presorted_keys == 0 || !enable_incremental_sort))
5643 1612 : continue;
5644 :
5645 : /*
5646 : * We've no need to consider both a sort and incremental sort.
5647 : * We'll just do a sort if there are no presorted keys and an
5648 : * incremental sort when there are presorted keys.
5649 : */
5650 57791 : if (presorted_keys == 0 || !enable_incremental_sort)
5651 54106 : sorted_path = (Path *) create_sort_path(root,
5652 : ordered_rel,
5653 : input_path,
5654 : root->sort_pathkeys,
5655 : limit_tuples);
5656 : else
5657 3685 : sorted_path = (Path *) create_incremental_sort_path(root,
5658 : ordered_rel,
5659 : input_path,
5660 : root->sort_pathkeys,
5661 : presorted_keys,
5662 : limit_tuples);
5663 : }
5664 :
5665 : /*
5666 : * If the pathtarget of the result path has different expressions from
5667 : * the target to be applied, a projection step is needed.
5668 : */
5669 94994 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5670 288 : sorted_path = apply_projection_to_path(root, ordered_rel,
5671 : sorted_path, target);
5672 :
5673 94994 : add_path(ordered_rel, sorted_path);
5674 : }
5675 :
5676 : /*
5677 : * generate_gather_paths() will have already generated a simple Gather
5678 : * path for the best parallel path, if any, and the loop above will have
5679 : * considered sorting it. Similarly, generate_gather_paths() will also
5680 : * have generated order-preserving Gather Merge plans which can be used
5681 : * without sorting if they happen to match the sort_pathkeys, and the loop
5682 : * above will have handled those as well. However, there's one more
5683 : * possibility: it may make sense to sort the cheapest partial path or
5684 : * incrementally sort any partial path that is partially sorted according
5685 : * to the required output order and then use Gather Merge.
5686 : */
5687 62146 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5688 43386 : input_rel->partial_pathlist != NIL)
5689 : {
5690 : Path *cheapest_partial_path;
5691 :
5692 2411 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5693 :
5694 5377 : foreach(lc, input_rel->partial_pathlist)
5695 : {
5696 2966 : Path *input_path = (Path *) lfirst(lc);
5697 : Path *sorted_path;
5698 : bool is_sorted;
5699 : int presorted_keys;
5700 : double total_groups;
5701 :
5702 2966 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5703 : input_path->pathkeys,
5704 : &presorted_keys);
5705 :
5706 2966 : if (is_sorted)
5707 476 : continue;
5708 :
5709 : /*
5710 : * Try at least sorting the cheapest path and also try
5711 : * incrementally sorting any path which is partially sorted
5712 : * already (no need to deal with paths which have presorted keys
5713 : * when incremental sort is disabled unless it's the cheapest
5714 : * partial path).
5715 : */
5716 2490 : if (input_path != cheapest_partial_path &&
5717 101 : (presorted_keys == 0 || !enable_incremental_sort))
5718 0 : continue;
5719 :
5720 : /*
5721 : * We've no need to consider both a sort and incremental sort.
5722 : * We'll just do a sort if there are no presorted keys and an
5723 : * incremental sort when there are presorted keys.
5724 : */
5725 2490 : if (presorted_keys == 0 || !enable_incremental_sort)
5726 2374 : sorted_path = (Path *) create_sort_path(root,
5727 : ordered_rel,
5728 : input_path,
5729 : root->sort_pathkeys,
5730 : limit_tuples);
5731 : else
5732 116 : sorted_path = (Path *) create_incremental_sort_path(root,
5733 : ordered_rel,
5734 : input_path,
5735 : root->sort_pathkeys,
5736 : presorted_keys,
5737 : limit_tuples);
5738 2490 : total_groups = compute_gather_rows(sorted_path);
5739 : sorted_path = (Path *)
5740 2490 : create_gather_merge_path(root, ordered_rel,
5741 : sorted_path,
5742 : sorted_path->pathtarget,
5743 : root->sort_pathkeys, NULL,
5744 : &total_groups);
5745 :
5746 : /*
5747 : * If the pathtarget of the result path has different expressions
5748 : * from the target to be applied, a projection step is needed.
5749 : */
5750 2490 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5751 5 : sorted_path = apply_projection_to_path(root, ordered_rel,
5752 : sorted_path, target);
5753 :
5754 2490 : add_path(ordered_rel, sorted_path);
5755 : }
5756 : }
5757 :
5758 : /*
5759 : * If there is an FDW that's responsible for all baserels of the query,
5760 : * let it consider adding ForeignPaths.
5761 : */
5762 62146 : if (ordered_rel->fdwroutine &&
5763 197 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5764 189 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5765 : input_rel, ordered_rel,
5766 : NULL);
5767 :
5768 : /* Let extensions possibly add some more paths */
5769 62146 : if (create_upper_paths_hook)
5770 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5771 : input_rel, ordered_rel, NULL);
5772 :
5773 : /*
5774 : * No need to bother with set_cheapest here; grouping_planner does not
5775 : * need us to do it.
5776 : */
5777 : Assert(ordered_rel->pathlist != NIL);
5778 :
5779 62146 : return ordered_rel;
5780 : }
5781 :
5782 :
5783 : /*
5784 : * make_group_input_target
5785 : * Generate appropriate PathTarget for initial input to grouping nodes.
5786 : *
5787 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5788 : * the query's final targetlist; for example, it certainly can't emit any
5789 : * aggregate function calls. This routine generates the correct target
5790 : * for the scan/join subplan.
5791 : *
5792 : * The query target list passed from the parser already contains entries
5793 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5794 : * for variables used only in HAVING clauses; so we need to add those
5795 : * variables to the subplan target list. Also, we flatten all expressions
5796 : * except GROUP BY items into their component variables; other expressions
5797 : * will be computed by the upper plan nodes rather than by the subplan.
5798 : * For example, given a query like
5799 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5800 : * we want to pass this targetlist to the subplan:
5801 : * a+b,c,d
5802 : * where the a+b target will be used by the Sort/Group steps, and the
5803 : * other targets will be used for computing the final results.
5804 : *
5805 : * 'final_target' is the query's final target list (in PathTarget form)
5806 : *
5807 : * The result is the PathTarget to be computed by the Paths returned from
5808 : * query_planner().
5809 : */
5810 : static PathTarget *
5811 34010 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5812 : {
5813 34010 : Query *parse = root->parse;
5814 : PathTarget *input_target;
5815 : List *non_group_cols;
5816 : List *non_group_vars;
5817 : int i;
5818 : ListCell *lc;
5819 :
5820 : /*
5821 : * We must build a target containing all grouping columns, plus any other
5822 : * Vars mentioned in the query's targetlist and HAVING qual.
5823 : */
5824 34010 : input_target = create_empty_pathtarget();
5825 34010 : non_group_cols = NIL;
5826 :
5827 34010 : i = 0;
5828 85562 : foreach(lc, final_target->exprs)
5829 : {
5830 51552 : Expr *expr = (Expr *) lfirst(lc);
5831 51552 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5832 :
5833 59660 : if (sgref && root->processed_groupClause &&
5834 8108 : get_sortgroupref_clause_noerr(sgref,
5835 : root->processed_groupClause) != NULL)
5836 : {
5837 : /*
5838 : * It's a grouping column, so add it to the input target as-is.
5839 : *
5840 : * Note that the target is logically below the grouping step. So
5841 : * with grouping sets we need to remove the RT index of the
5842 : * grouping step if there is any from the target expression.
5843 : */
5844 6592 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5845 : {
5846 : Assert(root->group_rtindex > 0);
5847 : expr = (Expr *)
5848 1803 : remove_nulling_relids((Node *) expr,
5849 1803 : bms_make_singleton(root->group_rtindex),
5850 : NULL);
5851 : }
5852 6592 : add_column_to_pathtarget(input_target, expr, sgref);
5853 : }
5854 : else
5855 : {
5856 : /*
5857 : * Non-grouping column, so just remember the expression for later
5858 : * call to pull_var_clause.
5859 : */
5860 44960 : non_group_cols = lappend(non_group_cols, expr);
5861 : }
5862 :
5863 51552 : i++;
5864 : }
5865 :
5866 : /*
5867 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5868 : */
5869 34010 : if (parse->havingQual)
5870 820 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5871 :
5872 : /*
5873 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5874 : * add them to the input target if not already present. (A Var used
5875 : * directly as a GROUP BY item will be present already.) Note this
5876 : * includes Vars used in resjunk items, so we are covering the needs of
5877 : * ORDER BY and window specifications. Vars used within Aggrefs and
5878 : * WindowFuncs will be pulled out here, too.
5879 : *
5880 : * Note that the target is logically below the grouping step. So with
5881 : * grouping sets we need to remove the RT index of the grouping step if
5882 : * there is any from the non-group Vars.
5883 : */
5884 34010 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5885 : PVC_RECURSE_AGGREGATES |
5886 : PVC_RECURSE_WINDOWFUNCS |
5887 : PVC_INCLUDE_PLACEHOLDERS);
5888 34010 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5889 : {
5890 : Assert(root->group_rtindex > 0);
5891 : non_group_vars = (List *)
5892 836 : remove_nulling_relids((Node *) non_group_vars,
5893 836 : bms_make_singleton(root->group_rtindex),
5894 : NULL);
5895 : }
5896 34010 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5897 :
5898 : /* clean up cruft */
5899 34010 : list_free(non_group_vars);
5900 34010 : list_free(non_group_cols);
5901 :
5902 : /* XXX this causes some redundant cost calculation ... */
5903 34010 : return set_pathtarget_cost_width(root, input_target);
5904 : }
5905 :
5906 : /*
5907 : * make_partial_grouping_target
5908 : * Generate appropriate PathTarget for output of partial aggregate
5909 : * (or partial grouping, if there are no aggregates) nodes.
5910 : *
5911 : * A partial aggregation node needs to emit all the same aggregates that
5912 : * a regular aggregation node would, plus any aggregates used in HAVING;
5913 : * except that the Aggref nodes should be marked as partial aggregates.
5914 : *
5915 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5916 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5917 : * these would be Vars that are grouped by or used in grouping expressions.)
5918 : *
5919 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5920 : * havingQual represents the HAVING clause.
5921 : */
5922 : static PathTarget *
5923 3205 : make_partial_grouping_target(PlannerInfo *root,
5924 : PathTarget *grouping_target,
5925 : Node *havingQual)
5926 : {
5927 : PathTarget *partial_target;
5928 : List *non_group_cols;
5929 : List *non_group_exprs;
5930 : int i;
5931 : ListCell *lc;
5932 :
5933 3205 : partial_target = create_empty_pathtarget();
5934 3205 : non_group_cols = NIL;
5935 :
5936 3205 : i = 0;
5937 10561 : foreach(lc, grouping_target->exprs)
5938 : {
5939 7356 : Expr *expr = (Expr *) lfirst(lc);
5940 7356 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5941 :
5942 11288 : if (sgref && root->processed_groupClause &&
5943 3932 : get_sortgroupref_clause_noerr(sgref,
5944 : root->processed_groupClause) != NULL)
5945 : {
5946 : /*
5947 : * It's a grouping column, so add it to the partial_target as-is.
5948 : * (This allows the upper agg step to repeat the grouping calcs.)
5949 : */
5950 2330 : add_column_to_pathtarget(partial_target, expr, sgref);
5951 : }
5952 : else
5953 : {
5954 : /*
5955 : * Non-grouping column, so just remember the expression for later
5956 : * call to pull_var_clause.
5957 : */
5958 5026 : non_group_cols = lappend(non_group_cols, expr);
5959 : }
5960 :
5961 7356 : i++;
5962 : }
5963 :
5964 : /*
5965 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5966 : */
5967 3205 : if (havingQual)
5968 729 : non_group_cols = lappend(non_group_cols, havingQual);
5969 :
5970 : /*
5971 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5972 : * non-group cols (plus HAVING), and add them to the partial_target if not
5973 : * already present. (An expression used directly as a GROUP BY item will
5974 : * be present already.) Note this includes Vars used in resjunk items, so
5975 : * we are covering the needs of ORDER BY and window specifications.
5976 : */
5977 3205 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5978 : PVC_INCLUDE_AGGREGATES |
5979 : PVC_RECURSE_WINDOWFUNCS |
5980 : PVC_INCLUDE_PLACEHOLDERS);
5981 :
5982 3205 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5983 :
5984 : /*
5985 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5986 : * are at the top level of the target list, so we can just scan the list
5987 : * rather than recursing through the expression trees.
5988 : */
5989 11075 : foreach(lc, partial_target->exprs)
5990 : {
5991 7870 : Aggref *aggref = (Aggref *) lfirst(lc);
5992 :
5993 7870 : if (IsA(aggref, Aggref))
5994 : {
5995 : Aggref *newaggref;
5996 :
5997 : /*
5998 : * We shouldn't need to copy the substructure of the Aggref node,
5999 : * but flat-copy the node itself to avoid damaging other trees.
6000 : */
6001 5515 : newaggref = makeNode(Aggref);
6002 5515 : memcpy(newaggref, aggref, sizeof(Aggref));
6003 :
6004 : /* For now, assume serialization is required */
6005 5515 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
6006 :
6007 5515 : lfirst(lc) = newaggref;
6008 : }
6009 : }
6010 :
6011 : /* clean up cruft */
6012 3205 : list_free(non_group_exprs);
6013 3205 : list_free(non_group_cols);
6014 :
6015 : /* XXX this causes some redundant cost calculation ... */
6016 3205 : return set_pathtarget_cost_width(root, partial_target);
6017 : }
6018 :
6019 : /*
6020 : * mark_partial_aggref
6021 : * Adjust an Aggref to make it represent a partial-aggregation step.
6022 : *
6023 : * The Aggref node is modified in-place; caller must do any copying required.
6024 : */
6025 : void
6026 15327 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
6027 : {
6028 : /* aggtranstype should be computed by this point */
6029 : Assert(OidIsValid(agg->aggtranstype));
6030 : /* ... but aggsplit should still be as the parser left it */
6031 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
6032 :
6033 : /* Mark the Aggref with the intended partial-aggregation mode */
6034 15327 : agg->aggsplit = aggsplit;
6035 :
6036 : /*
6037 : * Adjust result type if needed. Normally, a partial aggregate returns
6038 : * the aggregate's transition type; but if that's INTERNAL and we're
6039 : * serializing, it returns BYTEA instead.
6040 : */
6041 15327 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
6042 : {
6043 13423 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
6044 261 : agg->aggtype = BYTEAOID;
6045 : else
6046 13162 : agg->aggtype = agg->aggtranstype;
6047 : }
6048 15327 : }
6049 :
6050 : /*
6051 : * postprocess_setop_tlist
6052 : * Fix up targetlist returned by plan_set_operations().
6053 : *
6054 : * We need to transpose sort key info from the orig_tlist into new_tlist.
6055 : * NOTE: this would not be good enough if we supported resjunk sort keys
6056 : * for results of set operations --- then, we'd need to project a whole
6057 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
6058 : * find any resjunk columns in orig_tlist.
6059 : */
6060 : static List *
6061 4952 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
6062 : {
6063 : ListCell *l;
6064 4952 : ListCell *orig_tlist_item = list_head(orig_tlist);
6065 :
6066 19219 : foreach(l, new_tlist)
6067 : {
6068 14267 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
6069 : TargetEntry *orig_tle;
6070 :
6071 : /* ignore resjunk columns in setop result */
6072 14267 : if (new_tle->resjunk)
6073 0 : continue;
6074 :
6075 : Assert(orig_tlist_item != NULL);
6076 14267 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
6077 14267 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
6078 14267 : if (orig_tle->resjunk) /* should not happen */
6079 0 : elog(ERROR, "resjunk output columns are not implemented");
6080 : Assert(new_tle->resno == orig_tle->resno);
6081 14267 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
6082 : }
6083 4952 : if (orig_tlist_item != NULL)
6084 0 : elog(ERROR, "resjunk output columns are not implemented");
6085 4952 : return new_tlist;
6086 : }
6087 :
6088 : /*
6089 : * optimize_window_clauses
6090 : * Call each WindowFunc's prosupport function to see if we're able to
6091 : * make any adjustments to any of the WindowClause's so that the executor
6092 : * can execute the window functions in a more optimal way.
6093 : *
6094 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
6095 : * may allow more things to be done here in the future.
6096 : */
6097 : static void
6098 2217 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
6099 : {
6100 2217 : List *windowClause = root->parse->windowClause;
6101 : ListCell *lc;
6102 :
6103 4639 : foreach(lc, windowClause)
6104 : {
6105 2422 : WindowClause *wc = lfirst_node(WindowClause, lc);
6106 : ListCell *lc2;
6107 2422 : int optimizedFrameOptions = 0;
6108 :
6109 : Assert(wc->winref <= wflists->maxWinRef);
6110 :
6111 : /* skip any WindowClauses that have no WindowFuncs */
6112 2422 : if (wflists->windowFuncs[wc->winref] == NIL)
6113 20 : continue;
6114 :
6115 2945 : foreach(lc2, wflists->windowFuncs[wc->winref])
6116 : {
6117 : SupportRequestOptimizeWindowClause req;
6118 : SupportRequestOptimizeWindowClause *res;
6119 2437 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
6120 : Oid prosupport;
6121 :
6122 2437 : prosupport = get_func_support(wfunc->winfnoid);
6123 :
6124 : /* Check if there's a support function for 'wfunc' */
6125 2437 : if (!OidIsValid(prosupport))
6126 1894 : break; /* can't optimize this WindowClause */
6127 :
6128 730 : req.type = T_SupportRequestOptimizeWindowClause;
6129 730 : req.window_clause = wc;
6130 730 : req.window_func = wfunc;
6131 730 : req.frameOptions = wc->frameOptions;
6132 :
6133 : /* call the support function */
6134 : res = (SupportRequestOptimizeWindowClause *)
6135 730 : DatumGetPointer(OidFunctionCall1(prosupport,
6136 : PointerGetDatum(&req)));
6137 :
6138 : /*
6139 : * Skip to next WindowClause if the support function does not
6140 : * support this request type.
6141 : */
6142 730 : if (res == NULL)
6143 187 : break;
6144 :
6145 : /*
6146 : * Save these frameOptions for the first WindowFunc for this
6147 : * WindowClause.
6148 : */
6149 543 : if (foreach_current_index(lc2) == 0)
6150 523 : optimizedFrameOptions = res->frameOptions;
6151 :
6152 : /*
6153 : * On subsequent WindowFuncs, if the frameOptions are not the same
6154 : * then we're unable to optimize the frameOptions for this
6155 : * WindowClause.
6156 : */
6157 20 : else if (optimizedFrameOptions != res->frameOptions)
6158 0 : break; /* skip to the next WindowClause, if any */
6159 : }
6160 :
6161 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
6162 2402 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
6163 : {
6164 : ListCell *lc3;
6165 :
6166 : /* apply the new frame options */
6167 508 : wc->frameOptions = optimizedFrameOptions;
6168 :
6169 : /*
6170 : * We now check to see if changing the frameOptions has caused
6171 : * this WindowClause to be a duplicate of some other WindowClause.
6172 : * This can only happen if we have multiple WindowClauses, so
6173 : * don't bother if there's only 1.
6174 : */
6175 508 : if (list_length(windowClause) == 1)
6176 433 : continue;
6177 :
6178 : /*
6179 : * Do the duplicate check and reuse the existing WindowClause if
6180 : * we find a duplicate.
6181 : */
6182 190 : foreach(lc3, windowClause)
6183 : {
6184 145 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
6185 :
6186 : /* skip over the WindowClause we're currently editing */
6187 145 : if (existing_wc == wc)
6188 45 : continue;
6189 :
6190 : /*
6191 : * Perform the same duplicate check that is done in
6192 : * transformWindowFuncCall.
6193 : */
6194 200 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
6195 100 : equal(wc->orderClause, existing_wc->orderClause) &&
6196 100 : wc->frameOptions == existing_wc->frameOptions &&
6197 60 : equal(wc->startOffset, existing_wc->startOffset) &&
6198 30 : equal(wc->endOffset, existing_wc->endOffset))
6199 : {
6200 : ListCell *lc4;
6201 :
6202 : /*
6203 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
6204 : * This required adjusting each WindowFunc's winref and
6205 : * moving the WindowFuncs in 'wc' to the list of
6206 : * WindowFuncs in 'existing_wc'.
6207 : */
6208 65 : foreach(lc4, wflists->windowFuncs[wc->winref])
6209 : {
6210 35 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
6211 :
6212 35 : wfunc->winref = existing_wc->winref;
6213 : }
6214 :
6215 : /* move list items */
6216 60 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
6217 30 : wflists->windowFuncs[wc->winref]);
6218 30 : wflists->windowFuncs[wc->winref] = NIL;
6219 :
6220 : /*
6221 : * transformWindowFuncCall() should have made sure there
6222 : * are no other duplicates, so we needn't bother looking
6223 : * any further.
6224 : */
6225 30 : break;
6226 : }
6227 : }
6228 : }
6229 : }
6230 2217 : }
6231 :
6232 : /*
6233 : * select_active_windows
6234 : * Create a list of the "active" window clauses (ie, those referenced
6235 : * by non-deleted WindowFuncs) in the order they are to be executed.
6236 : */
6237 : static List *
6238 2217 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
6239 : {
6240 2217 : List *windowClause = root->parse->windowClause;
6241 2217 : List *result = NIL;
6242 : ListCell *lc;
6243 2217 : int nActive = 0;
6244 2217 : WindowClauseSortData *actives = palloc_array(WindowClauseSortData,
6245 : list_length(windowClause));
6246 :
6247 : /* First, construct an array of the active windows */
6248 4639 : foreach(lc, windowClause)
6249 : {
6250 2422 : WindowClause *wc = lfirst_node(WindowClause, lc);
6251 :
6252 : /* It's only active if wflists shows some related WindowFuncs */
6253 : Assert(wc->winref <= wflists->maxWinRef);
6254 2422 : if (wflists->windowFuncs[wc->winref] == NIL)
6255 50 : continue;
6256 :
6257 2372 : actives[nActive].wc = wc; /* original clause */
6258 :
6259 : /*
6260 : * For sorting, we want the list of partition keys followed by the
6261 : * list of sort keys. But pathkeys construction will remove duplicates
6262 : * between the two, so we can as well (even though we can't detect all
6263 : * of the duplicates, since some may come from ECs - that might mean
6264 : * we miss optimization chances here). We must, however, ensure that
6265 : * the order of entries is preserved with respect to the ones we do
6266 : * keep.
6267 : *
6268 : * partitionClause and orderClause had their own duplicates removed in
6269 : * parse analysis, so we're only concerned here with removing
6270 : * orderClause entries that also appear in partitionClause.
6271 : */
6272 4744 : actives[nActive].uniqueOrder =
6273 2372 : list_concat_unique(list_copy(wc->partitionClause),
6274 2372 : wc->orderClause);
6275 2372 : nActive++;
6276 : }
6277 :
6278 : /*
6279 : * Sort active windows by their partitioning/ordering clauses, ignoring
6280 : * any framing clauses, so that the windows that need the same sorting are
6281 : * adjacent in the list. When we come to generate paths, this will avoid
6282 : * inserting additional Sort nodes.
6283 : *
6284 : * This is how we implement a specific requirement from the SQL standard,
6285 : * which says that when two or more windows are order-equivalent (i.e.
6286 : * have matching partition and order clauses, even if their names or
6287 : * framing clauses differ), then all peer rows must be presented in the
6288 : * same order in all of them. If we allowed multiple sort nodes for such
6289 : * cases, we'd risk having the peer rows end up in different orders in
6290 : * equivalent windows due to sort instability. (See General Rule 4 of
6291 : * <window clause> in SQL2008 - SQL2016.)
6292 : *
6293 : * Additionally, if the entire list of clauses of one window is a prefix
6294 : * of another, put first the window with stronger sorting requirements.
6295 : * This way we will first sort for stronger window, and won't have to sort
6296 : * again for the weaker one.
6297 : */
6298 2217 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6299 :
6300 : /* build ordered list of the original WindowClause nodes */
6301 4589 : for (int i = 0; i < nActive; i++)
6302 2372 : result = lappend(result, actives[i].wc);
6303 :
6304 2217 : pfree(actives);
6305 :
6306 2217 : return result;
6307 : }
6308 :
6309 : /*
6310 : * name_active_windows
6311 : * Ensure all active windows have unique names.
6312 : *
6313 : * The parser will have checked that user-assigned window names are unique
6314 : * within the Query. Here we assign made-up names to any unnamed
6315 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6316 : * at parse time, because it'd mess up decompilation of views.)
6317 : *
6318 : * activeWindows: result of select_active_windows
6319 : */
6320 : static void
6321 2217 : name_active_windows(List *activeWindows)
6322 : {
6323 2217 : int next_n = 1;
6324 : char newname[16];
6325 : ListCell *lc;
6326 :
6327 4589 : foreach(lc, activeWindows)
6328 : {
6329 2372 : WindowClause *wc = lfirst_node(WindowClause, lc);
6330 :
6331 : /* Nothing to do if it has a name already. */
6332 2372 : if (wc->name)
6333 480 : continue;
6334 :
6335 : /* Select a name not currently present in the list. */
6336 : for (;;)
6337 5 : {
6338 : ListCell *lc2;
6339 :
6340 1897 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6341 4084 : foreach(lc2, activeWindows)
6342 : {
6343 2192 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6344 :
6345 2192 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6346 5 : break; /* matched */
6347 : }
6348 1897 : if (lc2 == NULL)
6349 1892 : break; /* reached the end with no match */
6350 : }
6351 1892 : wc->name = pstrdup(newname);
6352 : }
6353 2217 : }
6354 :
6355 : /*
6356 : * common_prefix_cmp
6357 : * QSort comparison function for WindowClauseSortData
6358 : *
6359 : * Sort the windows by the required sorting clauses. First, compare the sort
6360 : * clauses themselves. Second, if one window's clauses are a prefix of another
6361 : * one's clauses, put the window with more sort clauses first.
6362 : *
6363 : * We purposefully sort by the highest tleSortGroupRef first. Since
6364 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6365 : * and because here we sort the lowest tleSortGroupRefs last, if a
6366 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6367 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6368 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6369 : * reducing the total number of sorts required for the query.
6370 : */
6371 : static int
6372 170 : common_prefix_cmp(const void *a, const void *b)
6373 : {
6374 170 : const WindowClauseSortData *wcsa = a;
6375 170 : const WindowClauseSortData *wcsb = b;
6376 : ListCell *item_a;
6377 : ListCell *item_b;
6378 :
6379 305 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6380 : {
6381 220 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6382 220 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6383 :
6384 220 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6385 85 : return -1;
6386 210 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6387 55 : return 1;
6388 155 : else if (sca->sortop > scb->sortop)
6389 0 : return -1;
6390 155 : else if (sca->sortop < scb->sortop)
6391 20 : return 1;
6392 135 : else if (sca->nulls_first && !scb->nulls_first)
6393 0 : return -1;
6394 135 : else if (!sca->nulls_first && scb->nulls_first)
6395 0 : return 1;
6396 : /* no need to compare eqop, since it is fully determined by sortop */
6397 : }
6398 :
6399 85 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6400 5 : return -1;
6401 80 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6402 25 : return 1;
6403 :
6404 55 : return 0;
6405 : }
6406 :
6407 : /*
6408 : * make_window_input_target
6409 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6410 : *
6411 : * When the query has window functions, this function computes the desired
6412 : * target to be computed by the node just below the first WindowAgg.
6413 : * This tlist must contain all values needed to evaluate the window functions,
6414 : * compute the final target list, and perform any required final sort step.
6415 : * If multiple WindowAggs are needed, each intermediate one adds its window
6416 : * function results onto this base tlist; only the topmost WindowAgg computes
6417 : * the actual desired target list.
6418 : *
6419 : * This function is much like make_group_input_target, though not quite enough
6420 : * like it to share code. As in that function, we flatten most expressions
6421 : * into their component variables. But we do not want to flatten window
6422 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6423 : * evaluations of them, which would be bad (possibly even resulting in
6424 : * inconsistent answers, if they contain volatile functions).
6425 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6426 : * make_group_input_target, because we may no longer have access to the
6427 : * individual Vars in them.
6428 : *
6429 : * Another key difference from make_group_input_target is that we don't
6430 : * flatten Aggref expressions, since those are to be computed below the
6431 : * window functions and just referenced like Vars above that.
6432 : *
6433 : * 'final_target' is the query's final target list (in PathTarget form)
6434 : * 'activeWindows' is the list of active windows previously identified by
6435 : * select_active_windows.
6436 : *
6437 : * The result is the PathTarget to be computed by the plan node immediately
6438 : * below the first WindowAgg node.
6439 : */
6440 : static PathTarget *
6441 2217 : make_window_input_target(PlannerInfo *root,
6442 : PathTarget *final_target,
6443 : List *activeWindows)
6444 : {
6445 : PathTarget *input_target;
6446 : Bitmapset *sgrefs;
6447 : List *flattenable_cols;
6448 : List *flattenable_vars;
6449 : int i;
6450 : ListCell *lc;
6451 :
6452 : Assert(root->parse->hasWindowFuncs);
6453 :
6454 : /*
6455 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6456 : * into a bitmapset for convenient reference below.
6457 : */
6458 2217 : sgrefs = NULL;
6459 4589 : foreach(lc, activeWindows)
6460 : {
6461 2372 : WindowClause *wc = lfirst_node(WindowClause, lc);
6462 : ListCell *lc2;
6463 :
6464 2994 : foreach(lc2, wc->partitionClause)
6465 : {
6466 622 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6467 :
6468 622 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6469 : }
6470 4277 : foreach(lc2, wc->orderClause)
6471 : {
6472 1905 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6473 :
6474 1905 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6475 : }
6476 : }
6477 :
6478 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6479 2373 : foreach(lc, root->processed_groupClause)
6480 : {
6481 156 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6482 :
6483 156 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6484 : }
6485 :
6486 : /*
6487 : * Construct a target containing all the non-flattenable targetlist items,
6488 : * and save aside the others for a moment.
6489 : */
6490 2217 : input_target = create_empty_pathtarget();
6491 2217 : flattenable_cols = NIL;
6492 :
6493 2217 : i = 0;
6494 9324 : foreach(lc, final_target->exprs)
6495 : {
6496 7107 : Expr *expr = (Expr *) lfirst(lc);
6497 7107 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6498 :
6499 : /*
6500 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6501 : * that such items can't contain window functions, so it's okay to
6502 : * compute them below the WindowAgg nodes.)
6503 : */
6504 7107 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6505 : {
6506 : /*
6507 : * Don't want to deconstruct this value, so add it to the input
6508 : * target as-is.
6509 : */
6510 2389 : add_column_to_pathtarget(input_target, expr, sgref);
6511 : }
6512 : else
6513 : {
6514 : /*
6515 : * Column is to be flattened, so just remember the expression for
6516 : * later call to pull_var_clause.
6517 : */
6518 4718 : flattenable_cols = lappend(flattenable_cols, expr);
6519 : }
6520 :
6521 7107 : i++;
6522 : }
6523 :
6524 : /*
6525 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6526 : * add them to the input target if not already present. (Some might be
6527 : * there already because they're used directly as window/group clauses.)
6528 : *
6529 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6530 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6531 : * at higher levels. On the other hand, we should recurse into
6532 : * WindowFuncs to make sure their input expressions are available.
6533 : */
6534 2217 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6535 : PVC_INCLUDE_AGGREGATES |
6536 : PVC_RECURSE_WINDOWFUNCS |
6537 : PVC_INCLUDE_PLACEHOLDERS);
6538 2217 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6539 :
6540 : /* clean up cruft */
6541 2217 : list_free(flattenable_vars);
6542 2217 : list_free(flattenable_cols);
6543 :
6544 : /* XXX this causes some redundant cost calculation ... */
6545 2217 : return set_pathtarget_cost_width(root, input_target);
6546 : }
6547 :
6548 : /*
6549 : * make_pathkeys_for_window
6550 : * Create a pathkeys list describing the required input ordering
6551 : * for the given WindowClause.
6552 : *
6553 : * Modifies wc's partitionClause to remove any clauses which are deemed
6554 : * redundant by the pathkey logic.
6555 : *
6556 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6557 : * In the future we might try to implement windowing using hashing, in which
6558 : * case the ordering could be relaxed, but for now we always sort.
6559 : */
6560 : static List *
6561 4756 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6562 : List *tlist)
6563 : {
6564 4756 : List *window_pathkeys = NIL;
6565 :
6566 : /* Throw error if can't sort */
6567 4756 : if (!grouping_is_sortable(wc->partitionClause))
6568 0 : ereport(ERROR,
6569 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6570 : errmsg("could not implement window PARTITION BY"),
6571 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6572 4756 : if (!grouping_is_sortable(wc->orderClause))
6573 0 : ereport(ERROR,
6574 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6575 : errmsg("could not implement window ORDER BY"),
6576 : errdetail("Window ordering columns must be of sortable datatypes.")));
6577 :
6578 : /*
6579 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6580 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6581 : */
6582 4756 : if (wc->partitionClause != NIL)
6583 : {
6584 : bool sortable;
6585 :
6586 1088 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6587 : &wc->partitionClause,
6588 : tlist,
6589 : true,
6590 : false,
6591 : &sortable,
6592 : false);
6593 :
6594 : Assert(sortable);
6595 : }
6596 :
6597 : /*
6598 : * In principle, we could also consider removing redundant ORDER BY items
6599 : * too as doing so does not alter the result of peer row checks done by
6600 : * the executor. However, we must *not* remove the ordering column for
6601 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6602 : * if it's known to be equal to some partitioning column.
6603 : */
6604 4756 : if (wc->orderClause != NIL)
6605 : {
6606 : List *orderby_pathkeys;
6607 :
6608 3724 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6609 : wc->orderClause,
6610 : tlist);
6611 :
6612 : /* Okay, make the combined pathkeys */
6613 3724 : if (window_pathkeys != NIL)
6614 781 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6615 : else
6616 2943 : window_pathkeys = orderby_pathkeys;
6617 : }
6618 :
6619 4756 : return window_pathkeys;
6620 : }
6621 :
6622 : /*
6623 : * make_sort_input_target
6624 : * Generate appropriate PathTarget for initial input to Sort step.
6625 : *
6626 : * If the query has ORDER BY, this function chooses the target to be computed
6627 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6628 : * project) steps. This might or might not be identical to the query's final
6629 : * output target.
6630 : *
6631 : * The main argument for keeping the sort-input tlist the same as the final
6632 : * is that we avoid a separate projection node (which will be needed if
6633 : * they're different, because Sort can't project). However, there are also
6634 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6635 : * a consistent order of evaluation for any volatile functions in the tlist,
6636 : * and if there's also a LIMIT, we can stop the query without ever computing
6637 : * tlist functions for later rows, which is beneficial for both volatile and
6638 : * expensive functions.
6639 : *
6640 : * Our current policy is to postpone volatile expressions till after the sort
6641 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6642 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6643 : * postpone set-returning expressions, because running them beforehand would
6644 : * bloat the sort dataset, and because it might cause unexpected output order
6645 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6646 : * in the tlist should be evaluated at the same plan step, so that they can
6647 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6648 : * mustn't postpone any SRFs. (Note that in principle that policy should
6649 : * probably get applied to the group/window input targetlists too, but we
6650 : * have not done that historically.) Lastly, expensive expressions are
6651 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6652 : * partial evaluation of the query is possible (if neither is true, we expect
6653 : * to have to evaluate the expressions for every row anyway), or if there are
6654 : * any volatile or set-returning expressions (since once we've put in a
6655 : * projection at all, it won't cost any more to postpone more stuff).
6656 : *
6657 : * Another issue that could potentially be considered here is that
6658 : * evaluating tlist expressions could result in data that's either wider
6659 : * or narrower than the input Vars, thus changing the volume of data that
6660 : * has to go through the Sort. However, we usually have only a very bad
6661 : * idea of the output width of any expression more complex than a Var,
6662 : * so for now it seems too risky to try to optimize on that basis.
6663 : *
6664 : * Note that if we do produce a modified sort-input target, and then the
6665 : * query ends up not using an explicit Sort, no particular harm is done:
6666 : * we'll initially use the modified target for the preceding path nodes,
6667 : * but then change them to the final target with apply_projection_to_path.
6668 : * Moreover, in such a case the guarantees about evaluation order of
6669 : * volatile functions still hold, since the rows are sorted already.
6670 : *
6671 : * This function has some things in common with make_group_input_target and
6672 : * make_window_input_target, though the detailed rules for what to do are
6673 : * different. We never flatten/postpone any grouping or ordering columns;
6674 : * those are needed before the sort. If we do flatten a particular
6675 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6676 : * computed earlier.
6677 : *
6678 : * 'final_target' is the query's final target list (in PathTarget form)
6679 : * 'have_postponed_srfs' is an output argument, see below
6680 : *
6681 : * The result is the PathTarget to be computed by the plan node immediately
6682 : * below the Sort step (and the Distinct step, if any). This will be
6683 : * exactly final_target if we decide a projection step wouldn't be helpful.
6684 : *
6685 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6686 : * any set-returning functions to after the Sort.
6687 : */
6688 : static PathTarget *
6689 58879 : make_sort_input_target(PlannerInfo *root,
6690 : PathTarget *final_target,
6691 : bool *have_postponed_srfs)
6692 : {
6693 58879 : Query *parse = root->parse;
6694 : PathTarget *input_target;
6695 : int ncols;
6696 : bool *col_is_srf;
6697 : bool *postpone_col;
6698 : bool have_srf;
6699 : bool have_volatile;
6700 : bool have_expensive;
6701 : bool have_srf_sortcols;
6702 : bool postpone_srfs;
6703 : List *postponable_cols;
6704 : List *postponable_vars;
6705 : int i;
6706 : ListCell *lc;
6707 :
6708 : /* Shouldn't get here unless query has ORDER BY */
6709 : Assert(parse->sortClause);
6710 :
6711 58879 : *have_postponed_srfs = false; /* default result */
6712 :
6713 : /* Inspect tlist and collect per-column information */
6714 58879 : ncols = list_length(final_target->exprs);
6715 58879 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6716 58879 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6717 58879 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6718 :
6719 58879 : i = 0;
6720 342655 : foreach(lc, final_target->exprs)
6721 : {
6722 283776 : Expr *expr = (Expr *) lfirst(lc);
6723 :
6724 : /*
6725 : * If the column has a sortgroupref, assume it has to be evaluated
6726 : * before sorting. Generally such columns would be ORDER BY, GROUP
6727 : * BY, etc targets. One exception is columns that were removed from
6728 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6729 : * only be Vars anyway. There don't seem to be any cases where it
6730 : * would be worth the trouble to double-check.
6731 : */
6732 283776 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6733 : {
6734 : /*
6735 : * Check for SRF or volatile functions. Check the SRF case first
6736 : * because we must know whether we have any postponed SRFs.
6737 : */
6738 200596 : if (parse->hasTargetSRFs &&
6739 180 : expression_returns_set((Node *) expr))
6740 : {
6741 : /* We'll decide below whether these are postponable */
6742 80 : col_is_srf[i] = true;
6743 80 : have_srf = true;
6744 : }
6745 200336 : else if (contain_volatile_functions((Node *) expr))
6746 : {
6747 : /* Unconditionally postpone */
6748 185 : postpone_col[i] = true;
6749 185 : have_volatile = true;
6750 : }
6751 : else
6752 : {
6753 : /*
6754 : * Else check the cost. XXX it's annoying to have to do this
6755 : * when set_pathtarget_cost_width() just did it. Refactor to
6756 : * allow sharing the work?
6757 : */
6758 : QualCost cost;
6759 :
6760 200151 : cost_qual_eval_node(&cost, (Node *) expr, root);
6761 :
6762 : /*
6763 : * We arbitrarily define "expensive" as "more than 10X
6764 : * cpu_operator_cost". Note this will take in any PL function
6765 : * with default cost.
6766 : */
6767 200151 : if (cost.per_tuple > 10 * cpu_operator_cost)
6768 : {
6769 14339 : postpone_col[i] = true;
6770 14339 : have_expensive = true;
6771 : }
6772 : }
6773 : }
6774 : else
6775 : {
6776 : /* For sortgroupref cols, just check if any contain SRFs */
6777 83360 : if (!have_srf_sortcols &&
6778 83617 : parse->hasTargetSRFs &&
6779 277 : expression_returns_set((Node *) expr))
6780 122 : have_srf_sortcols = true;
6781 : }
6782 :
6783 283776 : i++;
6784 : }
6785 :
6786 : /*
6787 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6788 : */
6789 58879 : postpone_srfs = (have_srf && !have_srf_sortcols);
6790 :
6791 : /*
6792 : * If we don't need a post-sort projection, just return final_target.
6793 : */
6794 58879 : if (!(postpone_srfs || have_volatile ||
6795 58646 : (have_expensive &&
6796 8591 : (parse->limitCount || root->tuple_fraction > 0))))
6797 58616 : return final_target;
6798 :
6799 : /*
6800 : * Report whether the post-sort projection will contain set-returning
6801 : * functions. This is important because it affects whether the Sort can
6802 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6803 : * to return.
6804 : */
6805 263 : *have_postponed_srfs = postpone_srfs;
6806 :
6807 : /*
6808 : * Construct the sort-input target, taking all non-postponable columns and
6809 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6810 : * the postponable ones.
6811 : */
6812 263 : input_target = create_empty_pathtarget();
6813 263 : postponable_cols = NIL;
6814 :
6815 263 : i = 0;
6816 1855 : foreach(lc, final_target->exprs)
6817 : {
6818 1592 : Expr *expr = (Expr *) lfirst(lc);
6819 :
6820 1592 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6821 315 : postponable_cols = lappend(postponable_cols, expr);
6822 : else
6823 1277 : add_column_to_pathtarget(input_target, expr,
6824 1277 : get_pathtarget_sortgroupref(final_target, i));
6825 :
6826 1592 : i++;
6827 : }
6828 :
6829 : /*
6830 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6831 : * postponable columns, and add them to the sort-input target if not
6832 : * already present. (Some might be there already.) We mustn't
6833 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6834 : * would be unable to recompute them.
6835 : */
6836 263 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6837 : PVC_INCLUDE_AGGREGATES |
6838 : PVC_INCLUDE_WINDOWFUNCS |
6839 : PVC_INCLUDE_PLACEHOLDERS);
6840 263 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6841 :
6842 : /* clean up cruft */
6843 263 : list_free(postponable_vars);
6844 263 : list_free(postponable_cols);
6845 :
6846 : /* XXX this represents even more redundant cost calculation ... */
6847 263 : return set_pathtarget_cost_width(root, input_target);
6848 : }
6849 :
6850 : /*
6851 : * get_cheapest_fractional_path
6852 : * Find the cheapest path for retrieving a specified fraction of all
6853 : * the tuples expected to be returned by the given relation.
6854 : *
6855 : * Do not consider parameterized paths. If the caller needs a path for upper
6856 : * rel, it can't have parameterized paths. If the caller needs an append
6857 : * subpath, it could become limited by the treatment of similar
6858 : * parameterization of all the subpaths.
6859 : *
6860 : * We interpret tuple_fraction the same way as grouping_planner.
6861 : *
6862 : * We assume set_cheapest() has been run on the given rel.
6863 : */
6864 : Path *
6865 365893 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6866 : {
6867 365893 : Path *best_path = rel->cheapest_total_path;
6868 : ListCell *l;
6869 :
6870 : /* If all tuples will be retrieved, just return the cheapest-total path */
6871 365893 : if (tuple_fraction <= 0.0)
6872 358817 : return best_path;
6873 :
6874 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6875 7076 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6876 3073 : tuple_fraction /= best_path->rows;
6877 :
6878 18111 : foreach(l, rel->pathlist)
6879 : {
6880 11035 : Path *path = (Path *) lfirst(l);
6881 :
6882 11035 : if (path->param_info)
6883 166 : continue;
6884 :
6885 14662 : if (path == rel->cheapest_total_path ||
6886 3793 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6887 10547 : continue;
6888 :
6889 322 : best_path = path;
6890 : }
6891 :
6892 7076 : return best_path;
6893 : }
6894 :
6895 : /*
6896 : * adjust_paths_for_srfs
6897 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6898 : *
6899 : * The executor can only handle set-returning functions that appear at the
6900 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6901 : * that are not at top level, we need to split up the evaluation into multiple
6902 : * plan levels in which each level satisfies this constraint. This function
6903 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6904 : * output tlist to insert appropriate projection steps.
6905 : *
6906 : * The given targets and targets_contain_srfs lists are from
6907 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6908 : * target in targets.
6909 : */
6910 : static void
6911 10866 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6912 : List *targets, List *targets_contain_srfs)
6913 : {
6914 : ListCell *lc;
6915 :
6916 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6917 : Assert(!linitial_int(targets_contain_srfs));
6918 :
6919 : /* If no SRFs appear at this plan level, nothing to do */
6920 10866 : if (list_length(targets) == 1)
6921 491 : return;
6922 :
6923 : /*
6924 : * Stack SRF-evaluation nodes atop each path for the rel.
6925 : *
6926 : * In principle we should re-run set_cheapest() here to identify the
6927 : * cheapest path, but it seems unlikely that adding the same tlist eval
6928 : * costs to all the paths would change that, so we don't bother. Instead,
6929 : * just assume that the cheapest-startup and cheapest-total paths remain
6930 : * so. (There should be no parameterized paths anymore, so we needn't
6931 : * worry about updating cheapest_parameterized_paths.)
6932 : */
6933 20781 : foreach(lc, rel->pathlist)
6934 : {
6935 10406 : Path *subpath = (Path *) lfirst(lc);
6936 10406 : Path *newpath = subpath;
6937 : ListCell *lc1,
6938 : *lc2;
6939 :
6940 : Assert(subpath->param_info == NULL);
6941 32004 : forboth(lc1, targets, lc2, targets_contain_srfs)
6942 : {
6943 21598 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6944 21598 : bool contains_srfs = (bool) lfirst_int(lc2);
6945 :
6946 : /* If this level doesn't contain SRFs, do regular projection */
6947 21598 : if (contains_srfs)
6948 10456 : newpath = (Path *) create_set_projection_path(root,
6949 : rel,
6950 : newpath,
6951 : thistarget);
6952 : else
6953 11142 : newpath = (Path *) apply_projection_to_path(root,
6954 : rel,
6955 : newpath,
6956 : thistarget);
6957 : }
6958 10406 : lfirst(lc) = newpath;
6959 10406 : if (subpath == rel->cheapest_startup_path)
6960 239 : rel->cheapest_startup_path = newpath;
6961 10406 : if (subpath == rel->cheapest_total_path)
6962 239 : rel->cheapest_total_path = newpath;
6963 : }
6964 :
6965 : /* Likewise for partial paths, if any */
6966 10390 : foreach(lc, rel->partial_pathlist)
6967 : {
6968 15 : Path *subpath = (Path *) lfirst(lc);
6969 15 : Path *newpath = subpath;
6970 : ListCell *lc1,
6971 : *lc2;
6972 :
6973 : Assert(subpath->param_info == NULL);
6974 60 : forboth(lc1, targets, lc2, targets_contain_srfs)
6975 : {
6976 45 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6977 45 : bool contains_srfs = (bool) lfirst_int(lc2);
6978 :
6979 : /* If this level doesn't contain SRFs, do regular projection */
6980 45 : if (contains_srfs)
6981 15 : newpath = (Path *) create_set_projection_path(root,
6982 : rel,
6983 : newpath,
6984 : thistarget);
6985 : else
6986 : {
6987 : /* avoid apply_projection_to_path, in case of multiple refs */
6988 30 : newpath = (Path *) create_projection_path(root,
6989 : rel,
6990 : newpath,
6991 : thistarget);
6992 : }
6993 : }
6994 15 : lfirst(lc) = newpath;
6995 : }
6996 : }
6997 :
6998 : /*
6999 : * expression_planner
7000 : * Perform planner's transformations on a standalone expression.
7001 : *
7002 : * Various utility commands need to evaluate expressions that are not part
7003 : * of a plannable query. They can do so using the executor's regular
7004 : * expression-execution machinery, but first the expression has to be fed
7005 : * through here to transform it from parser output to something executable.
7006 : *
7007 : * Currently, we disallow sublinks in standalone expressions, so there's no
7008 : * real "planning" involved here. (That might not always be true though.)
7009 : * What we must do is run eval_const_expressions to ensure that any function
7010 : * calls are converted to positional notation and function default arguments
7011 : * get inserted. The fact that constant subexpressions get simplified is a
7012 : * side-effect that is useful when the expression will get evaluated more than
7013 : * once. Also, we must fix operator function IDs.
7014 : *
7015 : * This does not return any information about dependencies of the expression.
7016 : * Hence callers should use the results only for the duration of the current
7017 : * query. Callers that would like to cache the results for longer should use
7018 : * expression_planner_with_deps, probably via the plancache.
7019 : *
7020 : * Note: this must not make any damaging changes to the passed-in expression
7021 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
7022 : * we first do an expression_tree_mutator-based walk, what is returned will
7023 : * be a new node tree.) The result is constructed in the current memory
7024 : * context; beware that this can leak a lot of additional stuff there, too.
7025 : */
7026 : Expr *
7027 138472 : expression_planner(Expr *expr)
7028 : {
7029 : Node *result;
7030 :
7031 : /*
7032 : * Convert named-argument function calls, insert default arguments and
7033 : * simplify constant subexprs
7034 : */
7035 138472 : result = eval_const_expressions(NULL, (Node *) expr);
7036 :
7037 : /* Fill in opfuncid values if missing */
7038 138452 : fix_opfuncids(result);
7039 :
7040 138452 : return (Expr *) result;
7041 : }
7042 :
7043 : /*
7044 : * expression_planner_with_deps
7045 : * Perform planner's transformations on a standalone expression,
7046 : * returning expression dependency information along with the result.
7047 : *
7048 : * This is identical to expression_planner() except that it also returns
7049 : * information about possible dependencies of the expression, ie identities of
7050 : * objects whose definitions affect the result. As in a PlannedStmt, these
7051 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
7052 : */
7053 : Expr *
7054 266 : expression_planner_with_deps(Expr *expr,
7055 : List **relationOids,
7056 : List **invalItems)
7057 : {
7058 : Node *result;
7059 : PlannerGlobal glob;
7060 : PlannerInfo root;
7061 :
7062 : /* Make up dummy planner state so we can use setrefs machinery */
7063 7714 : MemSet(&glob, 0, sizeof(glob));
7064 266 : glob.type = T_PlannerGlobal;
7065 266 : glob.relationOids = NIL;
7066 266 : glob.invalItems = NIL;
7067 :
7068 25004 : MemSet(&root, 0, sizeof(root));
7069 266 : root.type = T_PlannerInfo;
7070 266 : root.glob = &glob;
7071 :
7072 : /*
7073 : * Convert named-argument function calls, insert default arguments and
7074 : * simplify constant subexprs. Collect identities of inlined functions
7075 : * and elided domains, too.
7076 : */
7077 266 : result = eval_const_expressions(&root, (Node *) expr);
7078 :
7079 : /* Fill in opfuncid values if missing */
7080 266 : fix_opfuncids(result);
7081 :
7082 : /*
7083 : * Now walk the finished expression to find anything else we ought to
7084 : * record as an expression dependency.
7085 : */
7086 266 : (void) extract_query_dependencies_walker(result, &root);
7087 :
7088 266 : *relationOids = glob.relationOids;
7089 266 : *invalItems = glob.invalItems;
7090 :
7091 266 : return (Expr *) result;
7092 : }
7093 :
7094 :
7095 : /*
7096 : * plan_cluster_use_sort
7097 : * Use the planner to decide how CLUSTER should implement sorting
7098 : *
7099 : * tableOid is the OID of a table to be clustered on its index indexOid
7100 : * (which is already known to be a btree index). Decide whether it's
7101 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
7102 : * Return true to use sorting, false to use an indexscan.
7103 : *
7104 : * Note: caller had better already hold some type of lock on the table.
7105 : */
7106 : bool
7107 145 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
7108 : {
7109 : PlannerInfo *root;
7110 : Query *query;
7111 : PlannerGlobal *glob;
7112 : RangeTblEntry *rte;
7113 : RelOptInfo *rel;
7114 : IndexOptInfo *indexInfo;
7115 : QualCost indexExprCost;
7116 : Cost comparisonCost;
7117 : Path *seqScanPath;
7118 : Path seqScanAndSortPath;
7119 : IndexPath *indexScanPath;
7120 : ListCell *lc;
7121 :
7122 : /* We can short-circuit the cost comparison if indexscans are disabled */
7123 145 : if (!enable_indexscan)
7124 20 : return true; /* use sort */
7125 :
7126 : /* Set up mostly-dummy planner state */
7127 125 : query = makeNode(Query);
7128 125 : query->commandType = CMD_SELECT;
7129 :
7130 125 : glob = makeNode(PlannerGlobal);
7131 :
7132 125 : root = makeNode(PlannerInfo);
7133 125 : root->parse = query;
7134 125 : root->glob = glob;
7135 125 : root->query_level = 1;
7136 125 : root->planner_cxt = CurrentMemoryContext;
7137 125 : root->wt_param_id = -1;
7138 125 : root->join_domains = list_make1(makeNode(JoinDomain));
7139 :
7140 : /* Build a minimal RTE for the rel */
7141 125 : rte = makeNode(RangeTblEntry);
7142 125 : rte->rtekind = RTE_RELATION;
7143 125 : rte->relid = tableOid;
7144 125 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7145 125 : rte->rellockmode = AccessShareLock;
7146 125 : rte->lateral = false;
7147 125 : rte->inh = false;
7148 125 : rte->inFromCl = true;
7149 125 : query->rtable = list_make1(rte);
7150 125 : addRTEPermissionInfo(&query->rteperminfos, rte);
7151 :
7152 : /* Set up RTE/RelOptInfo arrays */
7153 125 : setup_simple_rel_arrays(root);
7154 :
7155 : /* Build RelOptInfo */
7156 125 : rel = build_simple_rel(root, 1, NULL);
7157 :
7158 : /* Locate IndexOptInfo for the target index */
7159 125 : indexInfo = NULL;
7160 155 : foreach(lc, rel->indexlist)
7161 : {
7162 155 : indexInfo = lfirst_node(IndexOptInfo, lc);
7163 155 : if (indexInfo->indexoid == indexOid)
7164 125 : break;
7165 : }
7166 :
7167 : /*
7168 : * It's possible that get_relation_info did not generate an IndexOptInfo
7169 : * for the desired index; this could happen if it's not yet reached its
7170 : * indcheckxmin usability horizon, or if it's a system index and we're
7171 : * ignoring system indexes. In such cases we should tell CLUSTER to not
7172 : * trust the index contents but use seqscan-and-sort.
7173 : */
7174 125 : if (lc == NULL) /* not in the list? */
7175 0 : return true; /* use sort */
7176 :
7177 : /*
7178 : * Rather than doing all the pushups that would be needed to use
7179 : * set_baserel_size_estimates, just do a quick hack for rows and width.
7180 : */
7181 125 : rel->rows = rel->tuples;
7182 125 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
7183 :
7184 125 : root->total_table_pages = rel->pages;
7185 :
7186 : /*
7187 : * Determine eval cost of the index expressions, if any. We need to
7188 : * charge twice that amount for each tuple comparison that happens during
7189 : * the sort, since tuplesort.c will have to re-evaluate the index
7190 : * expressions each time. (XXX that's pretty inefficient...)
7191 : */
7192 125 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
7193 125 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
7194 :
7195 : /* Estimate the cost of seq scan + sort */
7196 125 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
7197 125 : cost_sort(&seqScanAndSortPath, root, NIL,
7198 : seqScanPath->disabled_nodes,
7199 125 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
7200 : comparisonCost, maintenance_work_mem, -1.0);
7201 :
7202 : /* Estimate the cost of index scan */
7203 125 : indexScanPath = create_index_path(root, indexInfo,
7204 : NIL, NIL, NIL, NIL,
7205 : ForwardScanDirection, false,
7206 : NULL, 1.0, false);
7207 :
7208 125 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
7209 : }
7210 :
7211 : /*
7212 : * plan_create_index_workers
7213 : * Use the planner to decide how many parallel worker processes
7214 : * CREATE INDEX should request for use
7215 : *
7216 : * tableOid is the table on which the index is to be built. indexOid is the
7217 : * OID of an index to be created or reindexed (which must be an index with
7218 : * support for parallel builds - currently btree, GIN, or BRIN).
7219 : *
7220 : * Return value is the number of parallel worker processes to request. It
7221 : * may be unsafe to proceed if this is 0. Note that this does not include the
7222 : * leader participating as a worker (value is always a number of parallel
7223 : * worker processes).
7224 : *
7225 : * Note: caller had better already hold some type of lock on the table and
7226 : * index.
7227 : */
7228 : int
7229 23038 : plan_create_index_workers(Oid tableOid, Oid indexOid)
7230 : {
7231 : PlannerInfo *root;
7232 : Query *query;
7233 : PlannerGlobal *glob;
7234 : RangeTblEntry *rte;
7235 : Relation heap;
7236 : Relation index;
7237 : RelOptInfo *rel;
7238 : int parallel_workers;
7239 : BlockNumber heap_blocks;
7240 : double reltuples;
7241 : double allvisfrac;
7242 :
7243 : /*
7244 : * We don't allow performing parallel operation in standalone backend or
7245 : * when parallelism is disabled.
7246 : */
7247 23038 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
7248 301 : return 0;
7249 :
7250 : /* Set up largely-dummy planner state */
7251 22737 : query = makeNode(Query);
7252 22737 : query->commandType = CMD_SELECT;
7253 :
7254 22737 : glob = makeNode(PlannerGlobal);
7255 :
7256 22737 : root = makeNode(PlannerInfo);
7257 22737 : root->parse = query;
7258 22737 : root->glob = glob;
7259 22737 : root->query_level = 1;
7260 22737 : root->planner_cxt = CurrentMemoryContext;
7261 22737 : root->wt_param_id = -1;
7262 22737 : root->join_domains = list_make1(makeNode(JoinDomain));
7263 :
7264 : /*
7265 : * Build a minimal RTE.
7266 : *
7267 : * Mark the RTE with inh = true. This is a kludge to prevent
7268 : * get_relation_info() from fetching index info, which is necessary
7269 : * because it does not expect that any IndexOptInfo is currently
7270 : * undergoing REINDEX.
7271 : */
7272 22737 : rte = makeNode(RangeTblEntry);
7273 22737 : rte->rtekind = RTE_RELATION;
7274 22737 : rte->relid = tableOid;
7275 22737 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7276 22737 : rte->rellockmode = AccessShareLock;
7277 22737 : rte->lateral = false;
7278 22737 : rte->inh = true;
7279 22737 : rte->inFromCl = true;
7280 22737 : query->rtable = list_make1(rte);
7281 22737 : addRTEPermissionInfo(&query->rteperminfos, rte);
7282 :
7283 : /* Set up RTE/RelOptInfo arrays */
7284 22737 : setup_simple_rel_arrays(root);
7285 :
7286 : /* Build RelOptInfo */
7287 22737 : rel = build_simple_rel(root, 1, NULL);
7288 :
7289 : /* Rels are assumed already locked by the caller */
7290 22737 : heap = table_open(tableOid, NoLock);
7291 22737 : index = index_open(indexOid, NoLock);
7292 :
7293 : /*
7294 : * Determine if it's safe to proceed.
7295 : *
7296 : * Currently, parallel workers can't access the leader's temporary tables.
7297 : * Furthermore, any index predicate or index expressions must be parallel
7298 : * safe.
7299 : */
7300 22737 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7301 21352 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7302 21262 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7303 : {
7304 1475 : parallel_workers = 0;
7305 1475 : goto done;
7306 : }
7307 :
7308 : /*
7309 : * If parallel_workers storage parameter is set for the table, accept that
7310 : * as the number of parallel worker processes to launch (though still cap
7311 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7312 : * consider any other factor when parallel_workers is set. (e.g., memory
7313 : * use by workers.)
7314 : */
7315 21262 : if (rel->rel_parallel_workers != -1)
7316 : {
7317 59 : parallel_workers = Min(rel->rel_parallel_workers,
7318 : max_parallel_maintenance_workers);
7319 59 : goto done;
7320 : }
7321 :
7322 : /*
7323 : * Estimate heap relation size ourselves, since rel->pages cannot be
7324 : * trusted (heap RTE was marked as inheritance parent)
7325 : */
7326 21203 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7327 :
7328 : /*
7329 : * Determine number of workers to scan the heap relation using generic
7330 : * model
7331 : */
7332 21203 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7333 : max_parallel_maintenance_workers);
7334 :
7335 : /*
7336 : * Cap workers based on available maintenance_work_mem as needed.
7337 : *
7338 : * Note that each tuplesort participant receives an even share of the
7339 : * total maintenance_work_mem budget. Aim to leave participants
7340 : * (including the leader as a participant) with no less than 32MB of
7341 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7342 : * immediately past the threshold of being capable of launching a single
7343 : * parallel worker to sort.
7344 : */
7345 21312 : while (parallel_workers > 0 &&
7346 219 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7347 109 : parallel_workers--;
7348 :
7349 21203 : done:
7350 22737 : index_close(index, NoLock);
7351 22737 : table_close(heap, NoLock);
7352 :
7353 22737 : return parallel_workers;
7354 : }
7355 :
7356 : /*
7357 : * add_paths_to_grouping_rel
7358 : *
7359 : * Add non-partial paths to grouping relation.
7360 : */
7361 : static void
7362 35057 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7363 : RelOptInfo *grouped_rel,
7364 : RelOptInfo *partially_grouped_rel,
7365 : const AggClauseCosts *agg_costs,
7366 : grouping_sets_data *gd,
7367 : GroupPathExtraData *extra)
7368 : {
7369 35057 : Query *parse = root->parse;
7370 35057 : Path *cheapest_path = input_rel->cheapest_total_path;
7371 35057 : Path *cheapest_partially_grouped_path = NULL;
7372 : ListCell *lc;
7373 35057 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7374 35057 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7375 35057 : List *havingQual = (List *) extra->havingQual;
7376 35057 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7377 35057 : double dNumGroups = 0;
7378 35057 : double dNumFinalGroups = 0;
7379 :
7380 : /*
7381 : * Estimate number of groups for non-split aggregation.
7382 : */
7383 35057 : dNumGroups = get_number_of_groups(root,
7384 : cheapest_path->rows,
7385 : gd,
7386 : extra->targetList);
7387 :
7388 35057 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7389 : {
7390 2492 : cheapest_partially_grouped_path =
7391 : partially_grouped_rel->cheapest_total_path;
7392 :
7393 : /*
7394 : * Estimate number of groups for final phase of partial aggregation.
7395 : */
7396 : dNumFinalGroups =
7397 2492 : get_number_of_groups(root,
7398 : cheapest_partially_grouped_path->rows,
7399 : gd,
7400 : extra->targetList);
7401 : }
7402 :
7403 35057 : if (can_sort)
7404 : {
7405 : /*
7406 : * Use any available suitably-sorted path as input, and also consider
7407 : * sorting the cheapest-total path and incremental sort on any paths
7408 : * with presorted keys.
7409 : */
7410 72686 : foreach(lc, input_rel->pathlist)
7411 : {
7412 : ListCell *lc2;
7413 37634 : Path *path = (Path *) lfirst(lc);
7414 37634 : Path *path_save = path;
7415 37634 : List *pathkey_orderings = NIL;
7416 :
7417 : /* generate alternative group orderings that might be useful */
7418 37634 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7419 :
7420 : Assert(list_length(pathkey_orderings) > 0);
7421 :
7422 75385 : foreach(lc2, pathkey_orderings)
7423 : {
7424 37751 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7425 :
7426 : /* restore the path (we replace it in the loop) */
7427 37751 : path = path_save;
7428 :
7429 37751 : path = make_ordered_path(root,
7430 : grouped_rel,
7431 : path,
7432 : cheapest_path,
7433 : info->pathkeys,
7434 : -1.0);
7435 37751 : if (path == NULL)
7436 306 : continue;
7437 :
7438 : /* Now decide what to stick atop it */
7439 37445 : if (parse->groupingSets)
7440 : {
7441 975 : consider_groupingsets_paths(root, grouped_rel,
7442 : path, true, can_hash,
7443 : gd, agg_costs, dNumGroups);
7444 : }
7445 36470 : else if (parse->hasAggs)
7446 : {
7447 : /*
7448 : * We have aggregation, possibly with plain GROUP BY. Make
7449 : * an AggPath.
7450 : */
7451 35798 : add_path(grouped_rel, (Path *)
7452 35798 : create_agg_path(root,
7453 : grouped_rel,
7454 : path,
7455 35798 : grouped_rel->reltarget,
7456 35798 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7457 : AGGSPLIT_SIMPLE,
7458 : info->clauses,
7459 : havingQual,
7460 : agg_costs,
7461 : dNumGroups));
7462 : }
7463 672 : else if (parse->groupClause)
7464 : {
7465 : /*
7466 : * We have GROUP BY without aggregation or grouping sets.
7467 : * Make a GroupPath.
7468 : */
7469 672 : add_path(grouped_rel, (Path *)
7470 672 : create_group_path(root,
7471 : grouped_rel,
7472 : path,
7473 : info->clauses,
7474 : havingQual,
7475 : dNumGroups));
7476 : }
7477 : else
7478 : {
7479 : /* Other cases should have been handled above */
7480 : Assert(false);
7481 : }
7482 : }
7483 : }
7484 :
7485 : /*
7486 : * Instead of operating directly on the input relation, we can
7487 : * consider finalizing a partially aggregated path.
7488 : */
7489 35052 : if (partially_grouped_rel != NULL)
7490 : {
7491 6420 : foreach(lc, partially_grouped_rel->pathlist)
7492 : {
7493 : ListCell *lc2;
7494 3928 : Path *path = (Path *) lfirst(lc);
7495 3928 : Path *path_save = path;
7496 3928 : List *pathkey_orderings = NIL;
7497 :
7498 : /* generate alternative group orderings that might be useful */
7499 3928 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7500 :
7501 : Assert(list_length(pathkey_orderings) > 0);
7502 :
7503 : /* process all potentially interesting grouping reorderings */
7504 7856 : foreach(lc2, pathkey_orderings)
7505 : {
7506 3928 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7507 :
7508 : /* restore the path (we replace it in the loop) */
7509 3928 : path = path_save;
7510 :
7511 3928 : path = make_ordered_path(root,
7512 : grouped_rel,
7513 : path,
7514 : cheapest_partially_grouped_path,
7515 : info->pathkeys,
7516 : -1.0);
7517 :
7518 3928 : if (path == NULL)
7519 172 : continue;
7520 :
7521 3756 : if (parse->hasAggs)
7522 3554 : add_path(grouped_rel, (Path *)
7523 3554 : create_agg_path(root,
7524 : grouped_rel,
7525 : path,
7526 3554 : grouped_rel->reltarget,
7527 3554 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7528 : AGGSPLIT_FINAL_DESERIAL,
7529 : info->clauses,
7530 : havingQual,
7531 : agg_final_costs,
7532 : dNumFinalGroups));
7533 : else
7534 202 : add_path(grouped_rel, (Path *)
7535 202 : create_group_path(root,
7536 : grouped_rel,
7537 : path,
7538 : info->clauses,
7539 : havingQual,
7540 : dNumFinalGroups));
7541 :
7542 : }
7543 : }
7544 : }
7545 : }
7546 :
7547 35057 : if (can_hash)
7548 : {
7549 5072 : if (parse->groupingSets)
7550 : {
7551 : /*
7552 : * Try for a hash-only groupingsets path over unsorted input.
7553 : */
7554 807 : consider_groupingsets_paths(root, grouped_rel,
7555 : cheapest_path, false, true,
7556 : gd, agg_costs, dNumGroups);
7557 : }
7558 : else
7559 : {
7560 : /*
7561 : * Generate a HashAgg Path. We just need an Agg over the
7562 : * cheapest-total input path, since input order won't matter.
7563 : */
7564 4265 : add_path(grouped_rel, (Path *)
7565 4265 : create_agg_path(root, grouped_rel,
7566 : cheapest_path,
7567 4265 : grouped_rel->reltarget,
7568 : AGG_HASHED,
7569 : AGGSPLIT_SIMPLE,
7570 : root->processed_groupClause,
7571 : havingQual,
7572 : agg_costs,
7573 : dNumGroups));
7574 : }
7575 :
7576 : /*
7577 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7578 : * grouped path, assuming there is one
7579 : */
7580 5072 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7581 : {
7582 1182 : add_path(grouped_rel, (Path *)
7583 1182 : create_agg_path(root,
7584 : grouped_rel,
7585 : cheapest_partially_grouped_path,
7586 1182 : grouped_rel->reltarget,
7587 : AGG_HASHED,
7588 : AGGSPLIT_FINAL_DESERIAL,
7589 : root->processed_groupClause,
7590 : havingQual,
7591 : agg_final_costs,
7592 : dNumFinalGroups));
7593 : }
7594 : }
7595 :
7596 : /*
7597 : * When partitionwise aggregate is used, we might have fully aggregated
7598 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7599 : * consider a path for grouped_rel consisting of a Parallel Append of
7600 : * non-partial paths from each child.
7601 : */
7602 35057 : if (grouped_rel->partial_pathlist != NIL)
7603 265 : gather_grouping_paths(root, grouped_rel);
7604 35057 : }
7605 :
7606 : /*
7607 : * create_partial_grouping_paths
7608 : *
7609 : * Create a new upper relation representing the result of partial aggregation
7610 : * and populate it with appropriate paths. Note that we don't finalize the
7611 : * lists of paths here, so the caller can add additional partial or non-partial
7612 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7613 : * the returned upper relation.
7614 : *
7615 : * All paths for this new upper relation -- both partial and non-partial --
7616 : * have been partially aggregated but require a subsequent FinalizeAggregate
7617 : * step.
7618 : *
7619 : * NB: This function is allowed to return NULL if it determines that there is
7620 : * no real need to create a new RelOptInfo.
7621 : */
7622 : static RelOptInfo *
7623 31604 : create_partial_grouping_paths(PlannerInfo *root,
7624 : RelOptInfo *grouped_rel,
7625 : RelOptInfo *input_rel,
7626 : grouping_sets_data *gd,
7627 : GroupPathExtraData *extra,
7628 : bool force_rel_creation)
7629 : {
7630 31604 : Query *parse = root->parse;
7631 : RelOptInfo *partially_grouped_rel;
7632 31604 : RelOptInfo *eager_agg_rel = NULL;
7633 31604 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7634 31604 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7635 31604 : Path *cheapest_partial_path = NULL;
7636 31604 : Path *cheapest_total_path = NULL;
7637 31604 : double dNumPartialGroups = 0;
7638 31604 : double dNumPartialPartialGroups = 0;
7639 : ListCell *lc;
7640 31604 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7641 31604 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7642 :
7643 : /*
7644 : * Check whether any partially aggregated paths have been generated
7645 : * through eager aggregation.
7646 : */
7647 31604 : if (input_rel->grouped_rel &&
7648 777 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7649 777 : input_rel->grouped_rel->pathlist != NIL)
7650 727 : eager_agg_rel = input_rel->grouped_rel;
7651 :
7652 : /*
7653 : * Consider whether we should generate partially aggregated non-partial
7654 : * paths. We can only do this if we have a non-partial path, and only if
7655 : * the parent of the input rel is performing partial partitionwise
7656 : * aggregation. (Note that extra->patype is the type of partitionwise
7657 : * aggregation being used at the parent level, not this level.)
7658 : */
7659 31604 : if (input_rel->pathlist != NIL &&
7660 31604 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7661 713 : cheapest_total_path = input_rel->cheapest_total_path;
7662 :
7663 : /*
7664 : * If parallelism is possible for grouped_rel, then we should consider
7665 : * generating partially-grouped partial paths. However, if the input rel
7666 : * has no partial paths, then we can't.
7667 : */
7668 31604 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7669 2744 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7670 :
7671 : /*
7672 : * If we can't partially aggregate partial paths, and we can't partially
7673 : * aggregate non-partial paths, and no partially aggregated paths were
7674 : * generated by eager aggregation, then don't bother creating the new
7675 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7676 : */
7677 31604 : if (cheapest_total_path == NULL &&
7678 28567 : cheapest_partial_path == NULL &&
7679 28480 : eager_agg_rel == NULL &&
7680 28480 : !force_rel_creation)
7681 28399 : return NULL;
7682 :
7683 : /*
7684 : * Build a new upper relation to represent the result of partially
7685 : * aggregating the rows from the input relation.
7686 : */
7687 3205 : partially_grouped_rel = fetch_upper_rel(root,
7688 : UPPERREL_PARTIAL_GROUP_AGG,
7689 : grouped_rel->relids);
7690 3205 : partially_grouped_rel->consider_parallel =
7691 3205 : grouped_rel->consider_parallel;
7692 3205 : partially_grouped_rel->pgs_mask = grouped_rel->pgs_mask;
7693 3205 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7694 3205 : partially_grouped_rel->serverid = grouped_rel->serverid;
7695 3205 : partially_grouped_rel->userid = grouped_rel->userid;
7696 3205 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7697 3205 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7698 :
7699 : /*
7700 : * Build target list for partial aggregate paths. These paths cannot just
7701 : * emit the same tlist as regular aggregate paths, because (1) we must
7702 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7703 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7704 : */
7705 3205 : partially_grouped_rel->reltarget =
7706 3205 : make_partial_grouping_target(root, grouped_rel->reltarget,
7707 : extra->havingQual);
7708 :
7709 3205 : if (!extra->partial_costs_set)
7710 : {
7711 : /*
7712 : * Collect statistics about aggregates for estimating costs of
7713 : * performing aggregation in parallel.
7714 : */
7715 11532 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7716 11532 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7717 1922 : if (parse->hasAggs)
7718 : {
7719 : /* partial phase */
7720 1813 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7721 : agg_partial_costs);
7722 :
7723 : /* final phase */
7724 1813 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7725 : agg_final_costs);
7726 : }
7727 :
7728 1922 : extra->partial_costs_set = true;
7729 : }
7730 :
7731 : /* Estimate number of partial groups. */
7732 3205 : if (cheapest_total_path != NULL)
7733 : dNumPartialGroups =
7734 713 : get_number_of_groups(root,
7735 : cheapest_total_path->rows,
7736 : gd,
7737 : extra->targetList);
7738 3205 : if (cheapest_partial_path != NULL)
7739 : dNumPartialPartialGroups =
7740 2744 : get_number_of_groups(root,
7741 : cheapest_partial_path->rows,
7742 : gd,
7743 : extra->targetList);
7744 :
7745 3205 : if (can_sort && cheapest_total_path != NULL)
7746 : {
7747 : /* This should have been checked previously */
7748 : Assert(parse->hasAggs || parse->groupClause);
7749 :
7750 : /*
7751 : * Use any available suitably-sorted path as input, and also consider
7752 : * sorting the cheapest partial path.
7753 : */
7754 1426 : foreach(lc, input_rel->pathlist)
7755 : {
7756 : ListCell *lc2;
7757 713 : Path *path = (Path *) lfirst(lc);
7758 713 : Path *path_save = path;
7759 713 : List *pathkey_orderings = NIL;
7760 :
7761 : /* generate alternative group orderings that might be useful */
7762 713 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7763 :
7764 : Assert(list_length(pathkey_orderings) > 0);
7765 :
7766 : /* process all potentially interesting grouping reorderings */
7767 1426 : foreach(lc2, pathkey_orderings)
7768 : {
7769 713 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7770 :
7771 : /* restore the path (we replace it in the loop) */
7772 713 : path = path_save;
7773 :
7774 713 : path = make_ordered_path(root,
7775 : partially_grouped_rel,
7776 : path,
7777 : cheapest_total_path,
7778 : info->pathkeys,
7779 : -1.0);
7780 :
7781 713 : if (path == NULL)
7782 0 : continue;
7783 :
7784 713 : if (parse->hasAggs)
7785 653 : add_path(partially_grouped_rel, (Path *)
7786 653 : create_agg_path(root,
7787 : partially_grouped_rel,
7788 : path,
7789 653 : partially_grouped_rel->reltarget,
7790 653 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7791 : AGGSPLIT_INITIAL_SERIAL,
7792 : info->clauses,
7793 : NIL,
7794 : agg_partial_costs,
7795 : dNumPartialGroups));
7796 : else
7797 60 : add_path(partially_grouped_rel, (Path *)
7798 60 : create_group_path(root,
7799 : partially_grouped_rel,
7800 : path,
7801 : info->clauses,
7802 : NIL,
7803 : dNumPartialGroups));
7804 : }
7805 : }
7806 : }
7807 :
7808 3205 : if (can_sort && cheapest_partial_path != NULL)
7809 : {
7810 : /* Similar to above logic, but for partial paths. */
7811 5863 : foreach(lc, input_rel->partial_pathlist)
7812 : {
7813 : ListCell *lc2;
7814 3119 : Path *path = (Path *) lfirst(lc);
7815 3119 : Path *path_save = path;
7816 3119 : List *pathkey_orderings = NIL;
7817 :
7818 : /* generate alternative group orderings that might be useful */
7819 3119 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7820 :
7821 : Assert(list_length(pathkey_orderings) > 0);
7822 :
7823 : /* process all potentially interesting grouping reorderings */
7824 6238 : foreach(lc2, pathkey_orderings)
7825 : {
7826 3119 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7827 :
7828 :
7829 : /* restore the path (we replace it in the loop) */
7830 3119 : path = path_save;
7831 :
7832 3119 : path = make_ordered_path(root,
7833 : partially_grouped_rel,
7834 : path,
7835 : cheapest_partial_path,
7836 : info->pathkeys,
7837 : -1.0);
7838 :
7839 3119 : if (path == NULL)
7840 5 : continue;
7841 :
7842 3114 : if (parse->hasAggs)
7843 3015 : add_partial_path(partially_grouped_rel, (Path *)
7844 3015 : create_agg_path(root,
7845 : partially_grouped_rel,
7846 : path,
7847 3015 : partially_grouped_rel->reltarget,
7848 3015 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7849 : AGGSPLIT_INITIAL_SERIAL,
7850 : info->clauses,
7851 : NIL,
7852 : agg_partial_costs,
7853 : dNumPartialPartialGroups));
7854 : else
7855 99 : add_partial_path(partially_grouped_rel, (Path *)
7856 99 : create_group_path(root,
7857 : partially_grouped_rel,
7858 : path,
7859 : info->clauses,
7860 : NIL,
7861 : dNumPartialPartialGroups));
7862 : }
7863 : }
7864 : }
7865 :
7866 : /*
7867 : * Add a partially-grouped HashAgg Path where possible
7868 : */
7869 3205 : if (can_hash && cheapest_total_path != NULL)
7870 : {
7871 : /* Checked above */
7872 : Assert(parse->hasAggs || parse->groupClause);
7873 :
7874 713 : add_path(partially_grouped_rel, (Path *)
7875 713 : create_agg_path(root,
7876 : partially_grouped_rel,
7877 : cheapest_total_path,
7878 713 : partially_grouped_rel->reltarget,
7879 : AGG_HASHED,
7880 : AGGSPLIT_INITIAL_SERIAL,
7881 : root->processed_groupClause,
7882 : NIL,
7883 : agg_partial_costs,
7884 : dNumPartialGroups));
7885 : }
7886 :
7887 : /*
7888 : * Now add a partially-grouped HashAgg partial Path where possible
7889 : */
7890 3205 : if (can_hash && cheapest_partial_path != NULL)
7891 : {
7892 1434 : add_partial_path(partially_grouped_rel, (Path *)
7893 1434 : create_agg_path(root,
7894 : partially_grouped_rel,
7895 : cheapest_partial_path,
7896 1434 : partially_grouped_rel->reltarget,
7897 : AGG_HASHED,
7898 : AGGSPLIT_INITIAL_SERIAL,
7899 : root->processed_groupClause,
7900 : NIL,
7901 : agg_partial_costs,
7902 : dNumPartialPartialGroups));
7903 : }
7904 :
7905 : /*
7906 : * Add any partially aggregated paths generated by eager aggregation to
7907 : * the new upper relation after applying projection steps as needed.
7908 : */
7909 3205 : if (eager_agg_rel)
7910 : {
7911 : /* Add the paths */
7912 1894 : foreach(lc, eager_agg_rel->pathlist)
7913 : {
7914 1167 : Path *path = (Path *) lfirst(lc);
7915 :
7916 : /* Shouldn't have any parameterized paths anymore */
7917 : Assert(path->param_info == NULL);
7918 :
7919 1167 : path = (Path *) create_projection_path(root,
7920 : partially_grouped_rel,
7921 : path,
7922 1167 : partially_grouped_rel->reltarget);
7923 :
7924 1167 : add_path(partially_grouped_rel, path);
7925 : }
7926 :
7927 : /*
7928 : * Likewise add the partial paths, but only if parallelism is possible
7929 : * for partially_grouped_rel.
7930 : */
7931 727 : if (partially_grouped_rel->consider_parallel)
7932 : {
7933 1690 : foreach(lc, eager_agg_rel->partial_pathlist)
7934 : {
7935 1010 : Path *path = (Path *) lfirst(lc);
7936 :
7937 : /* Shouldn't have any parameterized paths anymore */
7938 : Assert(path->param_info == NULL);
7939 :
7940 1010 : path = (Path *) create_projection_path(root,
7941 : partially_grouped_rel,
7942 : path,
7943 1010 : partially_grouped_rel->reltarget);
7944 :
7945 1010 : add_partial_path(partially_grouped_rel, path);
7946 : }
7947 : }
7948 : }
7949 :
7950 : /*
7951 : * If there is an FDW that's responsible for all baserels of the query,
7952 : * let it consider adding partially grouped ForeignPaths.
7953 : */
7954 3205 : if (partially_grouped_rel->fdwroutine &&
7955 3 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7956 : {
7957 3 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7958 :
7959 3 : fdwroutine->GetForeignUpperPaths(root,
7960 : UPPERREL_PARTIAL_GROUP_AGG,
7961 : input_rel, partially_grouped_rel,
7962 : extra);
7963 : }
7964 :
7965 3205 : return partially_grouped_rel;
7966 : }
7967 :
7968 : /*
7969 : * make_ordered_path
7970 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7971 : * return NULL if it doesn't make sense to generate an ordered path in
7972 : * this case.
7973 : */
7974 : static Path *
7975 49552 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7976 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7977 : {
7978 : bool is_sorted;
7979 : int presorted_keys;
7980 :
7981 49552 : is_sorted = pathkeys_count_contained_in(pathkeys,
7982 : path->pathkeys,
7983 : &presorted_keys);
7984 :
7985 49552 : if (!is_sorted)
7986 : {
7987 : /*
7988 : * Try at least sorting the cheapest path and also try incrementally
7989 : * sorting any path which is partially sorted already (no need to deal
7990 : * with paths which have presorted keys when incremental sort is
7991 : * disabled unless it's the cheapest input path).
7992 : */
7993 13115 : if (path != cheapest_path &&
7994 2157 : (presorted_keys == 0 || !enable_incremental_sort))
7995 1022 : return NULL;
7996 :
7997 : /*
7998 : * We've no need to consider both a sort and incremental sort. We'll
7999 : * just do a sort if there are no presorted keys and an incremental
8000 : * sort when there are presorted keys.
8001 : */
8002 12093 : if (presorted_keys == 0 || !enable_incremental_sort)
8003 10812 : path = (Path *) create_sort_path(root,
8004 : rel,
8005 : path,
8006 : pathkeys,
8007 : limit_tuples);
8008 : else
8009 1281 : path = (Path *) create_incremental_sort_path(root,
8010 : rel,
8011 : path,
8012 : pathkeys,
8013 : presorted_keys,
8014 : limit_tuples);
8015 : }
8016 :
8017 48530 : return path;
8018 : }
8019 :
8020 : /*
8021 : * Generate Gather and Gather Merge paths for a grouping relation or partial
8022 : * grouping relation.
8023 : *
8024 : * generate_useful_gather_paths does most of the work, but we also consider a
8025 : * special case: we could try sorting the data by the group_pathkeys and then
8026 : * applying Gather Merge.
8027 : *
8028 : * NB: This function shouldn't be used for anything other than a grouped or
8029 : * partially grouped relation not only because of the fact that it explicitly
8030 : * references group_pathkeys but we pass "true" as the third argument to
8031 : * generate_useful_gather_paths().
8032 : */
8033 : static void
8034 2589 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
8035 : {
8036 : ListCell *lc;
8037 : Path *cheapest_partial_path;
8038 : List *groupby_pathkeys;
8039 :
8040 : /*
8041 : * This occurs after any partial aggregation has taken place, so trim off
8042 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
8043 : */
8044 2589 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
8045 15 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
8046 : root->num_groupby_pathkeys);
8047 : else
8048 2574 : groupby_pathkeys = root->group_pathkeys;
8049 :
8050 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
8051 2589 : generate_useful_gather_paths(root, rel, true);
8052 :
8053 2589 : cheapest_partial_path = linitial(rel->partial_pathlist);
8054 :
8055 : /* XXX Shouldn't this also consider the group-key-reordering? */
8056 6150 : foreach(lc, rel->partial_pathlist)
8057 : {
8058 3561 : Path *path = (Path *) lfirst(lc);
8059 : bool is_sorted;
8060 : int presorted_keys;
8061 : double total_groups;
8062 :
8063 3561 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
8064 : path->pathkeys,
8065 : &presorted_keys);
8066 :
8067 3561 : if (is_sorted)
8068 2328 : continue;
8069 :
8070 : /*
8071 : * Try at least sorting the cheapest path and also try incrementally
8072 : * sorting any path which is partially sorted already (no need to deal
8073 : * with paths which have presorted keys when incremental sort is
8074 : * disabled unless it's the cheapest input path).
8075 : */
8076 1233 : if (path != cheapest_partial_path &&
8077 0 : (presorted_keys == 0 || !enable_incremental_sort))
8078 0 : continue;
8079 :
8080 : /*
8081 : * We've no need to consider both a sort and incremental sort. We'll
8082 : * just do a sort if there are no presorted keys and an incremental
8083 : * sort when there are presorted keys.
8084 : */
8085 1233 : if (presorted_keys == 0 || !enable_incremental_sort)
8086 1233 : path = (Path *) create_sort_path(root, rel, path,
8087 : groupby_pathkeys,
8088 : -1.0);
8089 : else
8090 0 : path = (Path *) create_incremental_sort_path(root,
8091 : rel,
8092 : path,
8093 : groupby_pathkeys,
8094 : presorted_keys,
8095 : -1.0);
8096 1233 : total_groups = compute_gather_rows(path);
8097 : path = (Path *)
8098 1233 : create_gather_merge_path(root,
8099 : rel,
8100 : path,
8101 1233 : rel->reltarget,
8102 : groupby_pathkeys,
8103 : NULL,
8104 : &total_groups);
8105 :
8106 1233 : add_path(rel, path);
8107 : }
8108 2589 : }
8109 :
8110 : /*
8111 : * can_partial_agg
8112 : *
8113 : * Determines whether or not partial grouping and/or aggregation is possible.
8114 : * Returns true when possible, false otherwise.
8115 : */
8116 : static bool
8117 33975 : can_partial_agg(PlannerInfo *root)
8118 : {
8119 33975 : Query *parse = root->parse;
8120 :
8121 33975 : if (!parse->hasAggs && parse->groupClause == NIL)
8122 : {
8123 : /*
8124 : * We don't know how to do parallel aggregation unless we have either
8125 : * some aggregates or a grouping clause.
8126 : */
8127 0 : return false;
8128 : }
8129 33975 : else if (parse->groupingSets)
8130 : {
8131 : /* We don't know how to do grouping sets in parallel. */
8132 886 : return false;
8133 : }
8134 33089 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
8135 : {
8136 : /* Insufficient support for partial mode. */
8137 3220 : return false;
8138 : }
8139 :
8140 : /* Everything looks good. */
8141 29869 : return true;
8142 : }
8143 :
8144 : /*
8145 : * apply_scanjoin_target_to_paths
8146 : *
8147 : * Adjust the final scan/join relation, and recursively all of its children,
8148 : * to generate the final scan/join target. It would be more correct to model
8149 : * this as a separate planning step with a new RelOptInfo at the toplevel and
8150 : * for each child relation, but doing it this way is noticeably cheaper.
8151 : * Maybe that problem can be solved at some point, but for now we do this.
8152 : *
8153 : * If tlist_same_exprs is true, then the scan/join target to be applied has
8154 : * the same expressions as the existing reltarget, so we need only insert the
8155 : * appropriate sortgroupref information. By avoiding the creation of
8156 : * projection paths we save effort both immediately and at plan creation time.
8157 : */
8158 : static void
8159 413534 : apply_scanjoin_target_to_paths(PlannerInfo *root,
8160 : RelOptInfo *rel,
8161 : List *scanjoin_targets,
8162 : List *scanjoin_targets_contain_srfs,
8163 : bool scanjoin_target_parallel_safe,
8164 : bool tlist_same_exprs)
8165 : {
8166 413534 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
8167 : PathTarget *scanjoin_target;
8168 : ListCell *lc;
8169 :
8170 : /* This recurses, so be paranoid. */
8171 413534 : check_stack_depth();
8172 :
8173 : /*
8174 : * If the rel only has Append and MergeAppend paths, we want to drop its
8175 : * existing paths and generate new ones. This function would still be
8176 : * correct if we kept the existing paths: we'd modify them to generate the
8177 : * correct target above the partitioning Append, and then they'd compete
8178 : * on cost with paths generating the target below the Append. However, in
8179 : * our current cost model the latter way is always the same or cheaper
8180 : * cost, so modifying the existing paths would just be useless work.
8181 : * Moreover, when the cost is the same, varying roundoff errors might
8182 : * sometimes allow an existing path to be picked, resulting in undesirable
8183 : * cross-platform plan variations. So we drop old paths and thereby force
8184 : * the work to be done below the Append.
8185 : *
8186 : * However, there are several cases when this optimization is not safe. If
8187 : * the rel isn't partitioned, then none of the paths will be Append or
8188 : * MergeAppend paths, so we should definitely not do this. If it is
8189 : * partitioned but is a joinrel, it may have Append and MergeAppend paths,
8190 : * but it can also have join paths that we can't afford to discard.
8191 : *
8192 : * Some care is needed, because we have to allow
8193 : * generate_useful_gather_paths to see the old partial paths in the next
8194 : * stanza. Hence, zap the main pathlist here, then allow
8195 : * generate_useful_gather_paths to add path(s) to the main list, and
8196 : * finally zap the partial pathlist.
8197 : */
8198 413534 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8199 8716 : rel->pathlist = NIL;
8200 :
8201 : /*
8202 : * If the scan/join target is not parallel-safe, partial paths cannot
8203 : * generate it.
8204 : */
8205 413534 : if (!scanjoin_target_parallel_safe)
8206 : {
8207 : /*
8208 : * Since we can't generate the final scan/join target in parallel
8209 : * workers, this is our last opportunity to use any partial paths that
8210 : * exist; so build Gather path(s) that use them and emit whatever the
8211 : * current reltarget is. We don't do this in the case where the
8212 : * target is parallel-safe, since we will be able to generate superior
8213 : * paths by doing it after the final scan/join target has been
8214 : * applied.
8215 : */
8216 53690 : generate_useful_gather_paths(root, rel, false);
8217 :
8218 : /* Can't use parallel query above this level. */
8219 53690 : rel->partial_pathlist = NIL;
8220 53690 : rel->consider_parallel = false;
8221 : }
8222 :
8223 : /* Finish dropping old paths for a partitioned rel, per comment above */
8224 413534 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
8225 8716 : rel->partial_pathlist = NIL;
8226 :
8227 : /* Extract SRF-free scan/join target. */
8228 413534 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
8229 :
8230 : /*
8231 : * Apply the SRF-free scan/join target to each existing path.
8232 : *
8233 : * If the tlist exprs are the same, we can just inject the sortgroupref
8234 : * information into the existing pathtargets. Otherwise, replace each
8235 : * path with a projection path that generates the SRF-free scan/join
8236 : * target. This can't change the ordering of paths within rel->pathlist,
8237 : * so we just modify the list in place.
8238 : */
8239 863754 : foreach(lc, rel->pathlist)
8240 : {
8241 450220 : Path *subpath = (Path *) lfirst(lc);
8242 :
8243 : /* Shouldn't have any parameterized paths anymore */
8244 : Assert(subpath->param_info == NULL);
8245 :
8246 450220 : if (tlist_same_exprs)
8247 162820 : subpath->pathtarget->sortgrouprefs =
8248 162820 : scanjoin_target->sortgrouprefs;
8249 : else
8250 : {
8251 : Path *newpath;
8252 :
8253 287400 : newpath = (Path *) create_projection_path(root, rel, subpath,
8254 : scanjoin_target);
8255 287400 : lfirst(lc) = newpath;
8256 : }
8257 : }
8258 :
8259 : /* Likewise adjust the targets for any partial paths. */
8260 433784 : foreach(lc, rel->partial_pathlist)
8261 : {
8262 20250 : Path *subpath = (Path *) lfirst(lc);
8263 :
8264 : /* Shouldn't have any parameterized paths anymore */
8265 : Assert(subpath->param_info == NULL);
8266 :
8267 20250 : if (tlist_same_exprs)
8268 16177 : subpath->pathtarget->sortgrouprefs =
8269 16177 : scanjoin_target->sortgrouprefs;
8270 : else
8271 : {
8272 : Path *newpath;
8273 :
8274 4073 : newpath = (Path *) create_projection_path(root, rel, subpath,
8275 : scanjoin_target);
8276 4073 : lfirst(lc) = newpath;
8277 : }
8278 : }
8279 :
8280 : /*
8281 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8282 : * atop each existing path. (Note that this function doesn't look at the
8283 : * cheapest-path fields, which is a good thing because they're bogus right
8284 : * now.)
8285 : */
8286 413534 : if (root->parse->hasTargetSRFs)
8287 10375 : adjust_paths_for_srfs(root, rel,
8288 : scanjoin_targets,
8289 : scanjoin_targets_contain_srfs);
8290 :
8291 : /*
8292 : * Update the rel's target to be the final (with SRFs) scan/join target.
8293 : * This now matches the actual output of all the paths, and we might get
8294 : * confused in createplan.c if they don't agree. We must do this now so
8295 : * that any append paths made in the next part will use the correct
8296 : * pathtarget (cf. create_append_path).
8297 : *
8298 : * Note that this is also necessary if GetForeignUpperPaths() gets called
8299 : * on the final scan/join relation or on any of its children, since the
8300 : * FDW might look at the rel's target to create ForeignPaths.
8301 : */
8302 413534 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8303 :
8304 : /*
8305 : * If the relation is partitioned, recursively apply the scan/join target
8306 : * to all partitions, and generate brand-new Append paths in which the
8307 : * scan/join target is computed below the Append rather than above it.
8308 : * Since Append is not projection-capable, that might save a separate
8309 : * Result node, and it also is important for partitionwise aggregate.
8310 : */
8311 413534 : if (rel_is_partitioned)
8312 : {
8313 10019 : List *live_children = NIL;
8314 : int i;
8315 :
8316 : /* Adjust each partition. */
8317 10019 : i = -1;
8318 30092 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8319 : {
8320 20073 : RelOptInfo *child_rel = rel->part_rels[i];
8321 : AppendRelInfo **appinfos;
8322 : int nappinfos;
8323 20073 : List *child_scanjoin_targets = NIL;
8324 :
8325 : Assert(child_rel != NULL);
8326 :
8327 : /* Dummy children can be ignored. */
8328 20073 : if (IS_DUMMY_REL(child_rel))
8329 40 : continue;
8330 :
8331 : /* Translate scan/join targets for this child. */
8332 20033 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8333 : &nappinfos);
8334 40066 : foreach(lc, scanjoin_targets)
8335 : {
8336 20033 : PathTarget *target = lfirst_node(PathTarget, lc);
8337 :
8338 20033 : target = copy_pathtarget(target);
8339 20033 : target->exprs = (List *)
8340 20033 : adjust_appendrel_attrs(root,
8341 20033 : (Node *) target->exprs,
8342 : nappinfos, appinfos);
8343 20033 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8344 : target);
8345 : }
8346 20033 : pfree(appinfos);
8347 :
8348 : /* Recursion does the real work. */
8349 20033 : apply_scanjoin_target_to_paths(root, child_rel,
8350 : child_scanjoin_targets,
8351 : scanjoin_targets_contain_srfs,
8352 : scanjoin_target_parallel_safe,
8353 : tlist_same_exprs);
8354 :
8355 : /* Save non-dummy children for Append paths. */
8356 20033 : if (!IS_DUMMY_REL(child_rel))
8357 20033 : live_children = lappend(live_children, child_rel);
8358 : }
8359 :
8360 : /* Build new paths for this relation by appending child paths. */
8361 10019 : add_paths_to_append_rel(root, rel, live_children);
8362 : }
8363 :
8364 : /*
8365 : * Consider generating Gather or Gather Merge paths. We must only do this
8366 : * if the relation is parallel safe, and we don't do it for child rels to
8367 : * avoid creating multiple Gather nodes within the same plan. We must do
8368 : * this after all paths have been generated and before set_cheapest, since
8369 : * one of the generated paths may turn out to be the cheapest one.
8370 : */
8371 413534 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
8372 139993 : generate_useful_gather_paths(root, rel, false);
8373 :
8374 : /*
8375 : * Reassess which paths are the cheapest, now that we've potentially added
8376 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8377 : * this relation.
8378 : */
8379 413534 : set_cheapest(rel);
8380 413534 : }
8381 :
8382 : /*
8383 : * create_partitionwise_grouping_paths
8384 : *
8385 : * If the partition keys of input relation are part of the GROUP BY clause, all
8386 : * the rows belonging to a given group come from a single partition. This
8387 : * allows aggregation/grouping over a partitioned relation to be broken down
8388 : * into aggregation/grouping on each partition. This should be no worse, and
8389 : * often better, than the normal approach.
8390 : *
8391 : * However, if the GROUP BY clause does not contain all the partition keys,
8392 : * rows from a given group may be spread across multiple partitions. In that
8393 : * case, we perform partial aggregation for each group, append the results,
8394 : * and then finalize aggregation. This is less certain to win than the
8395 : * previous case. It may win if the PartialAggregate stage greatly reduces
8396 : * the number of groups, because fewer rows will pass through the Append node.
8397 : * It may lose if we have lots of small groups.
8398 : */
8399 : static void
8400 685 : create_partitionwise_grouping_paths(PlannerInfo *root,
8401 : RelOptInfo *input_rel,
8402 : RelOptInfo *grouped_rel,
8403 : RelOptInfo *partially_grouped_rel,
8404 : const AggClauseCosts *agg_costs,
8405 : grouping_sets_data *gd,
8406 : PartitionwiseAggregateType patype,
8407 : GroupPathExtraData *extra)
8408 : {
8409 685 : List *grouped_live_children = NIL;
8410 685 : List *partially_grouped_live_children = NIL;
8411 685 : PathTarget *target = grouped_rel->reltarget;
8412 685 : bool partial_grouping_valid = true;
8413 : int i;
8414 :
8415 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8416 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8417 : partially_grouped_rel != NULL);
8418 :
8419 : /* Add paths for partitionwise aggregation/grouping. */
8420 685 : i = -1;
8421 2480 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8422 : {
8423 1795 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8424 : PathTarget *child_target;
8425 : AppendRelInfo **appinfos;
8426 : int nappinfos;
8427 : GroupPathExtraData child_extra;
8428 : RelOptInfo *child_grouped_rel;
8429 : RelOptInfo *child_partially_grouped_rel;
8430 :
8431 : Assert(child_input_rel != NULL);
8432 :
8433 : /* Dummy children can be ignored. */
8434 1795 : if (IS_DUMMY_REL(child_input_rel))
8435 0 : continue;
8436 :
8437 1795 : child_target = copy_pathtarget(target);
8438 :
8439 : /*
8440 : * Copy the given "extra" structure as is and then override the
8441 : * members specific to this child.
8442 : */
8443 1795 : memcpy(&child_extra, extra, sizeof(child_extra));
8444 :
8445 1795 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8446 : &nappinfos);
8447 :
8448 1795 : child_target->exprs = (List *)
8449 1795 : adjust_appendrel_attrs(root,
8450 1795 : (Node *) target->exprs,
8451 : nappinfos, appinfos);
8452 :
8453 : /* Translate havingQual and targetList. */
8454 1795 : child_extra.havingQual = (Node *)
8455 : adjust_appendrel_attrs(root,
8456 : extra->havingQual,
8457 : nappinfos, appinfos);
8458 1795 : child_extra.targetList = (List *)
8459 1795 : adjust_appendrel_attrs(root,
8460 1795 : (Node *) extra->targetList,
8461 : nappinfos, appinfos);
8462 :
8463 : /*
8464 : * extra->patype was the value computed for our parent rel; patype is
8465 : * the value for this relation. For the child, our value is its
8466 : * parent rel's value.
8467 : */
8468 1795 : child_extra.patype = patype;
8469 :
8470 : /*
8471 : * Create grouping relation to hold fully aggregated grouping and/or
8472 : * aggregation paths for the child.
8473 : */
8474 1795 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8475 : child_target,
8476 1795 : extra->target_parallel_safe,
8477 : child_extra.havingQual);
8478 :
8479 : /* Create grouping paths for this child relation. */
8480 1795 : create_ordinary_grouping_paths(root, child_input_rel,
8481 : child_grouped_rel,
8482 : agg_costs, gd, &child_extra,
8483 : &child_partially_grouped_rel);
8484 :
8485 1795 : if (child_partially_grouped_rel)
8486 : {
8487 : partially_grouped_live_children =
8488 1283 : lappend(partially_grouped_live_children,
8489 : child_partially_grouped_rel);
8490 : }
8491 : else
8492 512 : partial_grouping_valid = false;
8493 :
8494 1795 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8495 : {
8496 1082 : set_cheapest(child_grouped_rel);
8497 1082 : grouped_live_children = lappend(grouped_live_children,
8498 : child_grouped_rel);
8499 : }
8500 :
8501 1795 : pfree(appinfos);
8502 : }
8503 :
8504 : /*
8505 : * Try to create append paths for partially grouped children. For full
8506 : * partitionwise aggregation, we might have paths in the partial_pathlist
8507 : * if parallel aggregation is possible. For partial partitionwise
8508 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8509 : *
8510 : * NB: We must have a partially grouped path for every child in order to
8511 : * generate a partially grouped path for this relation.
8512 : */
8513 685 : if (partially_grouped_rel && partial_grouping_valid)
8514 : {
8515 : Assert(partially_grouped_live_children != NIL);
8516 :
8517 501 : add_paths_to_append_rel(root, partially_grouped_rel,
8518 : partially_grouped_live_children);
8519 : }
8520 :
8521 : /* If possible, create append paths for fully grouped children. */
8522 685 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8523 : {
8524 : Assert(grouped_live_children != NIL);
8525 :
8526 404 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8527 : }
8528 685 : }
8529 :
8530 : /*
8531 : * group_by_has_partkey
8532 : *
8533 : * Returns true if all the partition keys of the given relation are part of
8534 : * the GROUP BY clauses, including having matching collation, false otherwise.
8535 : */
8536 : static bool
8537 640 : group_by_has_partkey(RelOptInfo *input_rel,
8538 : List *targetList,
8539 : List *groupClause)
8540 : {
8541 640 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8542 640 : int cnt = 0;
8543 : int partnatts;
8544 :
8545 : /* Input relation should be partitioned. */
8546 : Assert(input_rel->part_scheme);
8547 :
8548 : /* Rule out early, if there are no partition keys present. */
8549 640 : if (!input_rel->partexprs)
8550 0 : return false;
8551 :
8552 640 : partnatts = input_rel->part_scheme->partnatts;
8553 :
8554 1074 : for (cnt = 0; cnt < partnatts; cnt++)
8555 : {
8556 670 : List *partexprs = input_rel->partexprs[cnt];
8557 : ListCell *lc;
8558 670 : bool found = false;
8559 :
8560 1001 : foreach(lc, partexprs)
8561 : {
8562 : ListCell *lg;
8563 775 : Expr *partexpr = lfirst(lc);
8564 775 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8565 :
8566 1206 : foreach(lg, groupexprs)
8567 : {
8568 875 : Expr *groupexpr = lfirst(lg);
8569 875 : Oid groupcoll = exprCollation((Node *) groupexpr);
8570 :
8571 : /*
8572 : * Note: we can assume there is at most one RelabelType node;
8573 : * eval_const_expressions() will have simplified if more than
8574 : * one.
8575 : */
8576 875 : if (IsA(groupexpr, RelabelType))
8577 20 : groupexpr = ((RelabelType *) groupexpr)->arg;
8578 :
8579 875 : if (equal(groupexpr, partexpr))
8580 : {
8581 : /*
8582 : * Reject a match if the grouping collation does not match
8583 : * the partitioning collation.
8584 : */
8585 444 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8586 : partcoll != groupcoll)
8587 10 : return false;
8588 :
8589 434 : found = true;
8590 434 : break;
8591 : }
8592 : }
8593 :
8594 765 : if (found)
8595 434 : break;
8596 : }
8597 :
8598 : /*
8599 : * If none of the partition key expressions match with any of the
8600 : * GROUP BY expression, return false.
8601 : */
8602 660 : if (!found)
8603 226 : return false;
8604 : }
8605 :
8606 404 : return true;
8607 : }
8608 :
8609 : /*
8610 : * generate_setop_child_grouplist
8611 : * Build a SortGroupClause list defining the sort/grouping properties
8612 : * of the child of a set operation.
8613 : *
8614 : * This is similar to generate_setop_grouplist() but differs as the setop
8615 : * child query's targetlist entries may already have a tleSortGroupRef
8616 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8617 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8618 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8619 : * any of the columns in the targetlist don't match to the setop's colTypes
8620 : * then we return an empty list. This may leave some TLEs with unreferenced
8621 : * ressortgroupref markings, but that's harmless.
8622 : */
8623 : static List *
8624 10483 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8625 : {
8626 10483 : List *grouplist = copyObject(op->groupClauses);
8627 : ListCell *lg;
8628 : ListCell *lt;
8629 : ListCell *ct;
8630 :
8631 10483 : lg = list_head(grouplist);
8632 10483 : ct = list_head(op->colTypes);
8633 40763 : foreach(lt, targetlist)
8634 : {
8635 30525 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8636 : SortGroupClause *sgc;
8637 : Oid coltype;
8638 :
8639 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8640 30525 : if (tle->resjunk)
8641 0 : continue;
8642 :
8643 : /*
8644 : * We expect every non-resjunk target to have a SortGroupClause and
8645 : * colTypes.
8646 : */
8647 : Assert(lg != NULL);
8648 : Assert(ct != NULL);
8649 30525 : sgc = (SortGroupClause *) lfirst(lg);
8650 30525 : coltype = lfirst_oid(ct);
8651 :
8652 : /* reject if target type isn't the same as the setop target type */
8653 30525 : if (coltype != exprType((Node *) tle->expr))
8654 245 : return NIL;
8655 :
8656 30280 : lg = lnext(grouplist, lg);
8657 30280 : ct = lnext(op->colTypes, ct);
8658 :
8659 : /* assign a tleSortGroupRef, or reuse the existing one */
8660 30280 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8661 : }
8662 :
8663 : Assert(lg == NULL);
8664 : Assert(ct == NULL);
8665 :
8666 10238 : return grouplist;
8667 : }
8668 :
8669 : /*
8670 : * create_unique_paths
8671 : * Build a new RelOptInfo containing Paths that represent elimination of
8672 : * distinct rows from the input data. Distinct-ness is defined according to
8673 : * the needs of the semijoin represented by sjinfo. If it is not possible
8674 : * to identify how to make the data unique, NULL is returned.
8675 : *
8676 : * If used at all, this is likely to be called repeatedly on the same rel,
8677 : * so we cache the result.
8678 : */
8679 : RelOptInfo *
8680 6924 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8681 : {
8682 : RelOptInfo *unique_rel;
8683 6924 : List *sortPathkeys = NIL;
8684 6924 : List *groupClause = NIL;
8685 : MemoryContext oldcontext;
8686 :
8687 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8688 : Assert(sjinfo->jointype == JOIN_SEMI);
8689 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8690 :
8691 : /* If result already cached, return it */
8692 6924 : if (rel->unique_rel)
8693 1196 : return rel->unique_rel;
8694 :
8695 : /* If it's not possible to unique-ify, return NULL */
8696 5728 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8697 104 : return NULL;
8698 :
8699 : /*
8700 : * Punt if this is a child relation and we failed to build a unique-ified
8701 : * relation for its parent. This can happen if all the RHS columns were
8702 : * found to be equated to constants when unique-ifying the parent table,
8703 : * leaving no columns to unique-ify.
8704 : */
8705 5624 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8706 10 : return NULL;
8707 :
8708 : /*
8709 : * When called during GEQO join planning, we are in a short-lived memory
8710 : * context. We must make sure that the unique rel and any subsidiary data
8711 : * structures created for a baserel survive the GEQO cycle, else the
8712 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8713 : * are creating those for a joinrel during GEQO, we don't want them to
8714 : * clutter the main planning context. Upshot is that the best solution is
8715 : * to explicitly allocate memory in the same context the given RelOptInfo
8716 : * is in.
8717 : */
8718 5614 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8719 :
8720 5614 : unique_rel = makeNode(RelOptInfo);
8721 5614 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8722 :
8723 : /*
8724 : * clear path info
8725 : */
8726 5614 : unique_rel->pathlist = NIL;
8727 5614 : unique_rel->ppilist = NIL;
8728 5614 : unique_rel->partial_pathlist = NIL;
8729 5614 : unique_rel->cheapest_startup_path = NULL;
8730 5614 : unique_rel->cheapest_total_path = NULL;
8731 5614 : unique_rel->cheapest_parameterized_paths = NIL;
8732 :
8733 : /*
8734 : * Build the target list for the unique rel. We also build the pathkeys
8735 : * that represent the ordering requirements for the sort-based
8736 : * implementation, and the list of SortGroupClause nodes that represent
8737 : * the columns to be grouped on for the hash-based implementation.
8738 : *
8739 : * For a child rel, we can construct these fields from those of its
8740 : * parent.
8741 : */
8742 5614 : if (IS_OTHER_REL(rel))
8743 360 : {
8744 : PathTarget *child_unique_target;
8745 : PathTarget *parent_unique_target;
8746 :
8747 360 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8748 :
8749 360 : child_unique_target = copy_pathtarget(parent_unique_target);
8750 :
8751 : /* Translate the target expressions */
8752 360 : child_unique_target->exprs = (List *)
8753 360 : adjust_appendrel_attrs_multilevel(root,
8754 360 : (Node *) parent_unique_target->exprs,
8755 : rel,
8756 360 : rel->top_parent);
8757 :
8758 360 : unique_rel->reltarget = child_unique_target;
8759 :
8760 360 : sortPathkeys = rel->top_parent->unique_pathkeys;
8761 360 : groupClause = rel->top_parent->unique_groupclause;
8762 : }
8763 : else
8764 : {
8765 : List *newtlist;
8766 : int nextresno;
8767 5254 : List *sortList = NIL;
8768 : ListCell *lc1;
8769 : ListCell *lc2;
8770 :
8771 : /*
8772 : * The values we are supposed to unique-ify may be expressions in the
8773 : * variables of the input rel's targetlist. We have to add any such
8774 : * expressions to the unique rel's targetlist.
8775 : *
8776 : * To complicate matters, some of the values to be unique-ified may be
8777 : * known redundant by the EquivalenceClass machinery (e.g., because
8778 : * they have been equated to constants). There is no need to compare
8779 : * such values during unique-ification, and indeed we had better not
8780 : * try because the Vars involved may not have propagated as high as
8781 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8782 : * detect such cases, which is a tad inefficient but it doesn't seem
8783 : * worth building specialized infrastructure for this.
8784 : */
8785 5254 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8786 5254 : nextresno = list_length(newtlist) + 1;
8787 :
8788 10709 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8789 : {
8790 5455 : Expr *uniqexpr = lfirst(lc1);
8791 5455 : Oid in_oper = lfirst_oid(lc2);
8792 : Oid sortop;
8793 : TargetEntry *tle;
8794 5455 : bool made_tle = false;
8795 :
8796 5455 : tle = tlist_member(uniqexpr, newtlist);
8797 5455 : if (!tle)
8798 : {
8799 2752 : tle = makeTargetEntry(uniqexpr,
8800 : nextresno,
8801 : NULL,
8802 : false);
8803 2752 : newtlist = lappend(newtlist, tle);
8804 2752 : nextresno++;
8805 2752 : made_tle = true;
8806 : }
8807 :
8808 : /*
8809 : * Try to build an ORDER BY list to sort the input compatibly. We
8810 : * do this for each sortable clause even when the clauses are not
8811 : * all sortable, so that we can detect clauses that are redundant
8812 : * according to the pathkey machinery.
8813 : */
8814 5455 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8815 5455 : if (OidIsValid(sortop))
8816 : {
8817 : Oid eqop;
8818 : SortGroupClause *sortcl;
8819 :
8820 : /*
8821 : * The Unique node will need equality operators. Normally
8822 : * these are the same as the IN clause operators, but if those
8823 : * are cross-type operators then the equality operators are
8824 : * the ones for the IN clause operators' RHS datatype.
8825 : */
8826 5455 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8827 5455 : if (!OidIsValid(eqop)) /* shouldn't happen */
8828 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8829 : sortop);
8830 :
8831 5455 : sortcl = makeNode(SortGroupClause);
8832 5455 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8833 5455 : sortcl->eqop = eqop;
8834 5455 : sortcl->sortop = sortop;
8835 5455 : sortcl->reverse_sort = false;
8836 5455 : sortcl->nulls_first = false;
8837 5455 : sortcl->hashable = false; /* no need to make this accurate */
8838 5455 : sortList = lappend(sortList, sortcl);
8839 :
8840 : /*
8841 : * At each step, convert the SortGroupClause list to pathkey
8842 : * form. If the just-added SortGroupClause is redundant, the
8843 : * result will be shorter than the SortGroupClause list.
8844 : */
8845 5455 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8846 : newtlist);
8847 5455 : if (list_length(sortPathkeys) != list_length(sortList))
8848 : {
8849 : /* Drop the redundant SortGroupClause */
8850 1710 : sortList = list_delete_last(sortList);
8851 : Assert(list_length(sortPathkeys) == list_length(sortList));
8852 : /* Undo tlist addition, if we made one */
8853 1710 : if (made_tle)
8854 : {
8855 10 : newtlist = list_delete_last(newtlist);
8856 10 : nextresno--;
8857 : }
8858 : /* We need not consider this clause for hashing, either */
8859 1710 : continue;
8860 : }
8861 : }
8862 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8863 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8864 : in_oper);
8865 :
8866 3745 : if (sjinfo->semi_can_hash)
8867 : {
8868 : /* Create a GROUP BY list for the Agg node to use */
8869 : Oid eq_oper;
8870 : SortGroupClause *groupcl;
8871 :
8872 : /*
8873 : * Get the hashable equality operators for the Agg node to
8874 : * use. Normally these are the same as the IN clause
8875 : * operators, but if those are cross-type operators then the
8876 : * equality operators are the ones for the IN clause
8877 : * operators' RHS datatype.
8878 : */
8879 3745 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8880 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8881 : in_oper);
8882 :
8883 3745 : groupcl = makeNode(SortGroupClause);
8884 3745 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8885 3745 : groupcl->eqop = eq_oper;
8886 3745 : groupcl->sortop = sortop;
8887 3745 : groupcl->reverse_sort = false;
8888 3745 : groupcl->nulls_first = false;
8889 3745 : groupcl->hashable = true;
8890 3745 : groupClause = lappend(groupClause, groupcl);
8891 : }
8892 : }
8893 :
8894 : /*
8895 : * Done building the sortPathkeys and groupClause. But the
8896 : * sortPathkeys are bogus if not all the clauses were sortable.
8897 : */
8898 5254 : if (!sjinfo->semi_can_btree)
8899 0 : sortPathkeys = NIL;
8900 :
8901 : /*
8902 : * It can happen that all the RHS columns are equated to constants.
8903 : * We'd have to do something special to unique-ify in that case, and
8904 : * it's such an unlikely-in-the-real-world case that it's not worth
8905 : * the effort. So just punt if we found no columns to unique-ify.
8906 : */
8907 5254 : if (sortPathkeys == NIL && groupClause == NIL)
8908 : {
8909 1625 : MemoryContextSwitchTo(oldcontext);
8910 1625 : return NULL;
8911 : }
8912 :
8913 : /* Convert the required targetlist back to PathTarget form */
8914 3629 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8915 : }
8916 :
8917 : /* build unique paths based on input rel's pathlist */
8918 3989 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8919 : sjinfo, unique_rel);
8920 :
8921 : /* build unique paths based on input rel's partial_pathlist */
8922 3989 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8923 : sjinfo, unique_rel);
8924 :
8925 : /* Now choose the best path(s) */
8926 3989 : set_cheapest(unique_rel);
8927 :
8928 : /*
8929 : * There shouldn't be any partial paths for the unique relation;
8930 : * otherwise, we won't be able to properly guarantee uniqueness.
8931 : */
8932 : Assert(unique_rel->partial_pathlist == NIL);
8933 :
8934 : /* Cache the result */
8935 3989 : rel->unique_rel = unique_rel;
8936 3989 : rel->unique_pathkeys = sortPathkeys;
8937 3989 : rel->unique_groupclause = groupClause;
8938 :
8939 3989 : MemoryContextSwitchTo(oldcontext);
8940 :
8941 3989 : return unique_rel;
8942 : }
8943 :
8944 : /*
8945 : * create_final_unique_paths
8946 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8947 : */
8948 : static void
8949 7121 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8950 : List *sortPathkeys, List *groupClause,
8951 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8952 : {
8953 7121 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8954 :
8955 : /* Estimate number of output rows */
8956 7121 : unique_rel->rows = estimate_num_groups(root,
8957 : sjinfo->semi_rhs_exprs,
8958 : cheapest_input_path->rows,
8959 : NULL,
8960 : NULL);
8961 :
8962 : /* Consider sort-based implementations, if possible. */
8963 7121 : if (sjinfo->semi_can_btree)
8964 : {
8965 : ListCell *lc;
8966 :
8967 : /*
8968 : * Use any available suitably-sorted path as input, and also consider
8969 : * sorting the cheapest-total path and incremental sort on any paths
8970 : * with presorted keys.
8971 : *
8972 : * To save planning time, we ignore parameterized input paths unless
8973 : * they are the cheapest-total path.
8974 : */
8975 15520 : foreach(lc, input_rel->pathlist)
8976 : {
8977 8399 : Path *input_path = (Path *) lfirst(lc);
8978 : Path *path;
8979 : bool is_sorted;
8980 : int presorted_keys;
8981 :
8982 : /*
8983 : * Ignore parameterized paths that are not the cheapest-total
8984 : * path.
8985 : */
8986 8399 : if (input_path->param_info &&
8987 : input_path != cheapest_input_path)
8988 740 : continue;
8989 :
8990 7695 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8991 : input_path->pathkeys,
8992 : &presorted_keys);
8993 :
8994 : /*
8995 : * Ignore paths that are not suitably or partially sorted, unless
8996 : * they are the cheapest total path (no need to deal with paths
8997 : * which have presorted keys when incremental sort is disabled).
8998 : */
8999 7695 : if (!is_sorted && input_path != cheapest_input_path &&
9000 76 : (presorted_keys == 0 || !enable_incremental_sort))
9001 36 : continue;
9002 :
9003 : /*
9004 : * Make a separate ProjectionPath in case we need a Result node.
9005 : */
9006 7659 : path = (Path *) create_projection_path(root,
9007 : unique_rel,
9008 : input_path,
9009 7659 : unique_rel->reltarget);
9010 :
9011 7659 : if (!is_sorted)
9012 : {
9013 : /*
9014 : * We've no need to consider both a sort and incremental sort.
9015 : * We'll just do a sort if there are no presorted keys and an
9016 : * incremental sort when there are presorted keys.
9017 : */
9018 3972 : if (presorted_keys == 0 || !enable_incremental_sort)
9019 3932 : path = (Path *) create_sort_path(root,
9020 : unique_rel,
9021 : path,
9022 : sortPathkeys,
9023 : -1.0);
9024 : else
9025 40 : path = (Path *) create_incremental_sort_path(root,
9026 : unique_rel,
9027 : path,
9028 : sortPathkeys,
9029 : presorted_keys,
9030 : -1.0);
9031 : }
9032 :
9033 7659 : path = (Path *) create_unique_path(root, unique_rel, path,
9034 : list_length(sortPathkeys),
9035 : unique_rel->rows);
9036 :
9037 7659 : add_path(unique_rel, path);
9038 : }
9039 : }
9040 :
9041 : /* Consider hash-based implementation, if possible. */
9042 7121 : if (sjinfo->semi_can_hash)
9043 : {
9044 : Path *path;
9045 :
9046 : /*
9047 : * Make a separate ProjectionPath in case we need a Result node.
9048 : */
9049 7121 : path = (Path *) create_projection_path(root,
9050 : unique_rel,
9051 : cheapest_input_path,
9052 7121 : unique_rel->reltarget);
9053 :
9054 7121 : path = (Path *) create_agg_path(root,
9055 : unique_rel,
9056 : path,
9057 : cheapest_input_path->pathtarget,
9058 : AGG_HASHED,
9059 : AGGSPLIT_SIMPLE,
9060 : groupClause,
9061 : NIL,
9062 : NULL,
9063 : unique_rel->rows);
9064 :
9065 7121 : add_path(unique_rel, path);
9066 : }
9067 7121 : }
9068 :
9069 : /*
9070 : * create_partial_unique_paths
9071 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
9072 : */
9073 : static void
9074 3989 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
9075 : List *sortPathkeys, List *groupClause,
9076 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
9077 : {
9078 : RelOptInfo *partial_unique_rel;
9079 : Path *cheapest_partial_path;
9080 :
9081 : /* nothing to do when there are no partial paths in the input rel */
9082 3989 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
9083 857 : return;
9084 :
9085 : /*
9086 : * nothing to do if there's anything in the targetlist that's
9087 : * parallel-restricted.
9088 : */
9089 3132 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
9090 0 : return;
9091 :
9092 3132 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
9093 :
9094 3132 : partial_unique_rel = makeNode(RelOptInfo);
9095 3132 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
9096 :
9097 : /*
9098 : * clear path info
9099 : */
9100 3132 : partial_unique_rel->pathlist = NIL;
9101 3132 : partial_unique_rel->ppilist = NIL;
9102 3132 : partial_unique_rel->partial_pathlist = NIL;
9103 3132 : partial_unique_rel->cheapest_startup_path = NULL;
9104 3132 : partial_unique_rel->cheapest_total_path = NULL;
9105 3132 : partial_unique_rel->cheapest_parameterized_paths = NIL;
9106 :
9107 : /* Estimate number of output rows */
9108 3132 : partial_unique_rel->rows = estimate_num_groups(root,
9109 : sjinfo->semi_rhs_exprs,
9110 : cheapest_partial_path->rows,
9111 : NULL,
9112 : NULL);
9113 3132 : partial_unique_rel->reltarget = unique_rel->reltarget;
9114 :
9115 : /* Consider sort-based implementations, if possible. */
9116 3132 : if (sjinfo->semi_can_btree)
9117 : {
9118 : ListCell *lc;
9119 :
9120 : /*
9121 : * Use any available suitably-sorted path as input, and also consider
9122 : * sorting the cheapest partial path and incremental sort on any paths
9123 : * with presorted keys.
9124 : */
9125 6518 : foreach(lc, input_rel->partial_pathlist)
9126 : {
9127 3386 : Path *input_path = (Path *) lfirst(lc);
9128 : Path *path;
9129 : bool is_sorted;
9130 : int presorted_keys;
9131 :
9132 3386 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
9133 : input_path->pathkeys,
9134 : &presorted_keys);
9135 :
9136 : /*
9137 : * Ignore paths that are not suitably or partially sorted, unless
9138 : * they are the cheapest partial path (no need to deal with paths
9139 : * which have presorted keys when incremental sort is disabled).
9140 : */
9141 3386 : if (!is_sorted && input_path != cheapest_partial_path &&
9142 0 : (presorted_keys == 0 || !enable_incremental_sort))
9143 0 : continue;
9144 :
9145 : /*
9146 : * Make a separate ProjectionPath in case we need a Result node.
9147 : */
9148 3386 : path = (Path *) create_projection_path(root,
9149 : partial_unique_rel,
9150 : input_path,
9151 3386 : partial_unique_rel->reltarget);
9152 :
9153 3386 : if (!is_sorted)
9154 : {
9155 : /*
9156 : * We've no need to consider both a sort and incremental sort.
9157 : * We'll just do a sort if there are no presorted keys and an
9158 : * incremental sort when there are presorted keys.
9159 : */
9160 3092 : if (presorted_keys == 0 || !enable_incremental_sort)
9161 3092 : path = (Path *) create_sort_path(root,
9162 : partial_unique_rel,
9163 : path,
9164 : sortPathkeys,
9165 : -1.0);
9166 : else
9167 0 : path = (Path *) create_incremental_sort_path(root,
9168 : partial_unique_rel,
9169 : path,
9170 : sortPathkeys,
9171 : presorted_keys,
9172 : -1.0);
9173 : }
9174 :
9175 3386 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
9176 : list_length(sortPathkeys),
9177 : partial_unique_rel->rows);
9178 :
9179 3386 : add_partial_path(partial_unique_rel, path);
9180 : }
9181 : }
9182 :
9183 : /* Consider hash-based implementation, if possible. */
9184 3132 : if (sjinfo->semi_can_hash)
9185 : {
9186 : Path *path;
9187 :
9188 : /*
9189 : * Make a separate ProjectionPath in case we need a Result node.
9190 : */
9191 3132 : path = (Path *) create_projection_path(root,
9192 : partial_unique_rel,
9193 : cheapest_partial_path,
9194 3132 : partial_unique_rel->reltarget);
9195 :
9196 3132 : path = (Path *) create_agg_path(root,
9197 : partial_unique_rel,
9198 : path,
9199 : cheapest_partial_path->pathtarget,
9200 : AGG_HASHED,
9201 : AGGSPLIT_SIMPLE,
9202 : groupClause,
9203 : NIL,
9204 : NULL,
9205 : partial_unique_rel->rows);
9206 :
9207 3132 : add_partial_path(partial_unique_rel, path);
9208 : }
9209 :
9210 3132 : if (partial_unique_rel->partial_pathlist != NIL)
9211 : {
9212 3132 : generate_useful_gather_paths(root, partial_unique_rel, true);
9213 3132 : set_cheapest(partial_unique_rel);
9214 :
9215 : /*
9216 : * Finally, create paths to unique-ify the final result. This step is
9217 : * needed to remove any duplicates due to combining rows from parallel
9218 : * workers.
9219 : */
9220 3132 : create_final_unique_paths(root, partial_unique_rel,
9221 : sortPathkeys, groupClause,
9222 : sjinfo, unique_rel);
9223 : }
9224 : }
9225 :
9226 : /*
9227 : * Choose a unique name for some subroot.
9228 : *
9229 : * Modifies glob->subplanNames to track names already used.
9230 : */
9231 : char *
9232 65353 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
9233 : {
9234 : unsigned n;
9235 :
9236 : /*
9237 : * If a numeric suffix is not required, then search the list of
9238 : * previously-assigned names for a match. If none is found, then we can
9239 : * use the provided name without modification.
9240 : */
9241 65353 : if (!always_number)
9242 : {
9243 19321 : bool found = false;
9244 :
9245 46554 : foreach_ptr(char, subplan_name, glob->subplanNames)
9246 : {
9247 12263 : if (strcmp(subplan_name, name) == 0)
9248 : {
9249 4351 : found = true;
9250 4351 : break;
9251 : }
9252 : }
9253 :
9254 19321 : if (!found)
9255 : {
9256 : /* pstrdup here is just to avoid cast-away-const */
9257 14970 : char *chosen_name = pstrdup(name);
9258 :
9259 14970 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
9260 14970 : return chosen_name;
9261 : }
9262 : }
9263 :
9264 : /*
9265 : * If a numeric suffix is required or if the un-suffixed name is already
9266 : * in use, then loop until we find a positive integer that produces a
9267 : * novel name.
9268 : */
9269 50383 : for (n = 1; true; ++n)
9270 43342 : {
9271 93725 : char *proposed_name = psprintf("%s_%u", name, n);
9272 93725 : bool found = false;
9273 :
9274 349309 : foreach_ptr(char, subplan_name, glob->subplanNames)
9275 : {
9276 205201 : if (strcmp(subplan_name, proposed_name) == 0)
9277 : {
9278 43342 : found = true;
9279 43342 : break;
9280 : }
9281 : }
9282 :
9283 93725 : if (!found)
9284 : {
9285 50383 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9286 50383 : return proposed_name;
9287 : }
9288 :
9289 43342 : pfree(proposed_name);
9290 : }
9291 : }
|
