Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * planner.c
4 : * The query optimizer external interface.
5 : *
6 : * Portions Copyright (c) 1996-2025, 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 when grouping_planner() plans upper rels */
77 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
78 :
79 :
80 : /* Expression kind codes for preprocess_expression */
81 : #define EXPRKIND_QUAL 0
82 : #define EXPRKIND_TARGET 1
83 : #define EXPRKIND_RTFUNC 2
84 : #define EXPRKIND_RTFUNC_LATERAL 3
85 : #define EXPRKIND_VALUES 4
86 : #define EXPRKIND_VALUES_LATERAL 5
87 : #define EXPRKIND_LIMIT 6
88 : #define EXPRKIND_APPINFO 7
89 : #define EXPRKIND_PHV 8
90 : #define EXPRKIND_TABLESAMPLE 9
91 : #define EXPRKIND_ARBITER_ELEM 10
92 : #define EXPRKIND_TABLEFUNC 11
93 : #define EXPRKIND_TABLEFUNC_LATERAL 12
94 : #define EXPRKIND_GROUPEXPR 13
95 :
96 : /*
97 : * Data specific to grouping sets
98 : */
99 : typedef struct
100 : {
101 : List *rollups;
102 : List *hash_sets_idx;
103 : double dNumHashGroups;
104 : bool any_hashable;
105 : Bitmapset *unsortable_refs;
106 : Bitmapset *unhashable_refs;
107 : List *unsortable_sets;
108 : int *tleref_to_colnum_map;
109 : } grouping_sets_data;
110 :
111 : /*
112 : * Temporary structure for use during WindowClause reordering in order to be
113 : * able to sort WindowClauses on partitioning/ordering prefix.
114 : */
115 : typedef struct
116 : {
117 : WindowClause *wc;
118 : List *uniqueOrder; /* A List of unique ordering/partitioning
119 : * clauses per Window */
120 : } WindowClauseSortData;
121 :
122 : /* Passthrough data for standard_qp_callback */
123 : typedef struct
124 : {
125 : List *activeWindows; /* active windows, if any */
126 : grouping_sets_data *gset_data; /* grouping sets data, if any */
127 : SetOperationStmt *setop; /* parent set operation or NULL if not a
128 : * subquery belonging to a set operation */
129 : } standard_qp_extra;
130 :
131 : /* Local functions */
132 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
133 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
134 : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
135 : SetOperationStmt *setops);
136 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
137 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
138 : int *tleref_to_colnum_map);
139 : static void preprocess_rowmarks(PlannerInfo *root);
140 : static double preprocess_limit(PlannerInfo *root,
141 : double tuple_fraction,
142 : int64 *offset_est, int64 *count_est);
143 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
144 : static List *extract_rollup_sets(List *groupingSets);
145 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
146 : static void standard_qp_callback(PlannerInfo *root, void *extra);
147 : static double get_number_of_groups(PlannerInfo *root,
148 : double path_rows,
149 : grouping_sets_data *gd,
150 : List *target_list);
151 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
152 : RelOptInfo *input_rel,
153 : PathTarget *target,
154 : bool target_parallel_safe,
155 : grouping_sets_data *gd);
156 : static bool is_degenerate_grouping(PlannerInfo *root);
157 : static void create_degenerate_grouping_paths(PlannerInfo *root,
158 : RelOptInfo *input_rel,
159 : RelOptInfo *grouped_rel);
160 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
161 : PathTarget *target, bool target_parallel_safe,
162 : Node *havingQual);
163 : static void create_ordinary_grouping_paths(PlannerInfo *root,
164 : RelOptInfo *input_rel,
165 : RelOptInfo *grouped_rel,
166 : const AggClauseCosts *agg_costs,
167 : grouping_sets_data *gd,
168 : GroupPathExtraData *extra,
169 : RelOptInfo **partially_grouped_rel_p);
170 : static void consider_groupingsets_paths(PlannerInfo *root,
171 : RelOptInfo *grouped_rel,
172 : Path *path,
173 : bool is_sorted,
174 : bool can_hash,
175 : grouping_sets_data *gd,
176 : const AggClauseCosts *agg_costs,
177 : double dNumGroups);
178 : static RelOptInfo *create_window_paths(PlannerInfo *root,
179 : RelOptInfo *input_rel,
180 : PathTarget *input_target,
181 : PathTarget *output_target,
182 : bool output_target_parallel_safe,
183 : WindowFuncLists *wflists,
184 : List *activeWindows);
185 : static void create_one_window_path(PlannerInfo *root,
186 : RelOptInfo *window_rel,
187 : Path *path,
188 : PathTarget *input_target,
189 : PathTarget *output_target,
190 : WindowFuncLists *wflists,
191 : List *activeWindows);
192 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
193 : RelOptInfo *input_rel,
194 : PathTarget *target);
195 : static void create_partial_distinct_paths(PlannerInfo *root,
196 : RelOptInfo *input_rel,
197 : RelOptInfo *final_distinct_rel,
198 : PathTarget *target);
199 : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
200 : RelOptInfo *input_rel,
201 : RelOptInfo *distinct_rel);
202 : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
203 : List *needed_pathkeys,
204 : List *path_pathkeys);
205 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
206 : RelOptInfo *input_rel,
207 : PathTarget *target,
208 : bool target_parallel_safe,
209 : double limit_tuples);
210 : static PathTarget *make_group_input_target(PlannerInfo *root,
211 : PathTarget *final_target);
212 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
213 : PathTarget *grouping_target,
214 : Node *havingQual);
215 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
216 : static void optimize_window_clauses(PlannerInfo *root,
217 : WindowFuncLists *wflists);
218 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
219 : static void name_active_windows(List *activeWindows);
220 : static PathTarget *make_window_input_target(PlannerInfo *root,
221 : PathTarget *final_target,
222 : List *activeWindows);
223 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
224 : List *tlist);
225 : static PathTarget *make_sort_input_target(PlannerInfo *root,
226 : PathTarget *final_target,
227 : bool *have_postponed_srfs);
228 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
229 : List *targets, List *targets_contain_srfs);
230 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
231 : RelOptInfo *grouped_rel,
232 : RelOptInfo *partially_grouped_rel,
233 : const AggClauseCosts *agg_costs,
234 : grouping_sets_data *gd,
235 : double dNumGroups,
236 : GroupPathExtraData *extra);
237 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
238 : RelOptInfo *grouped_rel,
239 : RelOptInfo *input_rel,
240 : grouping_sets_data *gd,
241 : GroupPathExtraData *extra,
242 : bool force_rel_creation);
243 : static Path *make_ordered_path(PlannerInfo *root,
244 : RelOptInfo *rel,
245 : Path *path,
246 : Path *cheapest_path,
247 : List *pathkeys,
248 : double limit_tuples);
249 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
250 : static bool can_partial_agg(PlannerInfo *root);
251 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
252 : RelOptInfo *rel,
253 : List *scanjoin_targets,
254 : List *scanjoin_targets_contain_srfs,
255 : bool scanjoin_target_parallel_safe,
256 : bool tlist_same_exprs);
257 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
258 : RelOptInfo *input_rel,
259 : RelOptInfo *grouped_rel,
260 : RelOptInfo *partially_grouped_rel,
261 : const AggClauseCosts *agg_costs,
262 : grouping_sets_data *gd,
263 : PartitionwiseAggregateType patype,
264 : GroupPathExtraData *extra);
265 : static bool group_by_has_partkey(RelOptInfo *input_rel,
266 : List *targetList,
267 : List *groupClause);
268 : static int common_prefix_cmp(const void *a, const void *b);
269 : static List *generate_setop_child_grouplist(SetOperationStmt *op,
270 : List *targetlist);
271 : static void create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
272 : List *sortPathkeys, List *groupClause,
273 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
274 : static void create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
275 : List *sortPathkeys, List *groupClause,
276 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
277 :
278 :
279 : /*****************************************************************************
280 : *
281 : * Query optimizer entry point
282 : *
283 : * To support loadable plugins that monitor or modify planner behavior,
284 : * we provide a hook variable that lets a plugin get control before and
285 : * after the standard planning process. The plugin would normally call
286 : * standard_planner().
287 : *
288 : * Note to plugin authors: standard_planner() scribbles on its Query input,
289 : * so you'd better copy that data structure if you want to plan more than once.
290 : *
291 : *****************************************************************************/
292 : PlannedStmt *
293 450048 : planner(Query *parse, const char *query_string, int cursorOptions,
294 : ParamListInfo boundParams)
295 : {
296 : PlannedStmt *result;
297 :
298 450048 : if (planner_hook)
299 94502 : result = (*planner_hook) (parse, query_string, cursorOptions, boundParams);
300 : else
301 355546 : result = standard_planner(parse, query_string, cursorOptions, boundParams);
302 :
303 445310 : pgstat_report_plan_id(result->planId, false);
304 :
305 445310 : return result;
306 : }
307 :
308 : PlannedStmt *
309 450048 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
310 : ParamListInfo boundParams)
311 : {
312 : PlannedStmt *result;
313 : PlannerGlobal *glob;
314 : double tuple_fraction;
315 : PlannerInfo *root;
316 : RelOptInfo *final_rel;
317 : Path *best_path;
318 : Plan *top_plan;
319 : ListCell *lp,
320 : *lr;
321 :
322 : /*
323 : * Set up global state for this planner invocation. This data is needed
324 : * across all levels of sub-Query that might exist in the given command,
325 : * so we keep it in a separate struct that's linked to by each per-Query
326 : * PlannerInfo.
327 : */
328 450048 : glob = makeNode(PlannerGlobal);
329 :
330 450048 : glob->boundParams = boundParams;
331 450048 : glob->subplans = NIL;
332 450048 : glob->subpaths = NIL;
333 450048 : glob->subroots = NIL;
334 450048 : glob->rewindPlanIDs = NULL;
335 450048 : glob->finalrtable = NIL;
336 450048 : glob->allRelids = NULL;
337 450048 : glob->prunableRelids = NULL;
338 450048 : glob->finalrteperminfos = NIL;
339 450048 : glob->finalrowmarks = NIL;
340 450048 : glob->resultRelations = NIL;
341 450048 : glob->appendRelations = NIL;
342 450048 : glob->partPruneInfos = NIL;
343 450048 : glob->relationOids = NIL;
344 450048 : glob->invalItems = NIL;
345 450048 : glob->paramExecTypes = NIL;
346 450048 : glob->lastPHId = 0;
347 450048 : glob->lastRowMarkId = 0;
348 450048 : glob->lastPlanNodeId = 0;
349 450048 : glob->transientPlan = false;
350 450048 : glob->dependsOnRole = false;
351 450048 : glob->partition_directory = NULL;
352 450048 : glob->rel_notnullatts_hash = NULL;
353 :
354 : /*
355 : * Assess whether it's feasible to use parallel mode for this query. We
356 : * can't do this in a standalone backend, or if the command will try to
357 : * modify any data, or if this is a cursor operation, or if GUCs are set
358 : * to values that don't permit parallelism, or if parallel-unsafe
359 : * functions are present in the query tree.
360 : *
361 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
362 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
363 : * the command is writing into a completely new table which workers won't
364 : * be able to see. If the workers could see the table, the fact that
365 : * group locking would cause them to ignore the leader's heavyweight GIN
366 : * page locks would make this unsafe. We'll have to fix that somehow if
367 : * we want to allow parallel inserts in general; updates and deletes have
368 : * additional problems especially around combo CIDs.)
369 : *
370 : * For now, we don't try to use parallel mode if we're running inside a
371 : * parallel worker. We might eventually be able to relax this
372 : * restriction, but for now it seems best not to have parallel workers
373 : * trying to create their own parallel workers.
374 : */
375 450048 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
376 421866 : IsUnderPostmaster &&
377 421866 : parse->commandType == CMD_SELECT &&
378 337582 : !parse->hasModifyingCTE &&
379 337440 : max_parallel_workers_per_gather > 0 &&
380 336836 : !IsParallelWorker())
381 : {
382 : /* all the cheap tests pass, so scan the query tree */
383 336788 : glob->maxParallelHazard = max_parallel_hazard(parse);
384 336788 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
385 : }
386 : else
387 : {
388 : /* skip the query tree scan, just assume it's unsafe */
389 113260 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
390 113260 : glob->parallelModeOK = false;
391 : }
392 :
393 : /*
394 : * glob->parallelModeNeeded is normally set to false here and changed to
395 : * true during plan creation if a Gather or Gather Merge plan is actually
396 : * created (cf. create_gather_plan, create_gather_merge_plan).
397 : *
398 : * However, if debug_parallel_query = on or debug_parallel_query =
399 : * regress, then we impose parallel mode whenever it's safe to do so, even
400 : * if the final plan doesn't use parallelism. It's not safe to do so if
401 : * the query contains anything parallel-unsafe; parallelModeOK will be
402 : * false in that case. Note that parallelModeOK can't change after this
403 : * point. Otherwise, everything in the query is either parallel-safe or
404 : * parallel-restricted, and in either case it should be OK to impose
405 : * parallel-mode restrictions. If that ends up breaking something, then
406 : * either some function the user included in the query is incorrectly
407 : * labeled as parallel-safe or parallel-restricted when in reality it's
408 : * parallel-unsafe, or else the query planner itself has a bug.
409 : */
410 738428 : glob->parallelModeNeeded = glob->parallelModeOK &&
411 288380 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
412 :
413 : /* Determine what fraction of the plan is likely to be scanned */
414 450048 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
415 : {
416 : /*
417 : * We have no real idea how many tuples the user will ultimately FETCH
418 : * from a cursor, but it is often the case that he doesn't want 'em
419 : * all, or would prefer a fast-start plan anyway so that he can
420 : * process some of the tuples sooner. Use a GUC parameter to decide
421 : * what fraction to optimize for.
422 : */
423 4702 : tuple_fraction = cursor_tuple_fraction;
424 :
425 : /*
426 : * We document cursor_tuple_fraction as simply being a fraction, which
427 : * means the edge cases 0 and 1 have to be treated specially here. We
428 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
429 : */
430 4702 : if (tuple_fraction >= 1.0)
431 0 : tuple_fraction = 0.0;
432 4702 : else if (tuple_fraction <= 0.0)
433 0 : tuple_fraction = 1e-10;
434 : }
435 : else
436 : {
437 : /* Default assumption is we need all the tuples */
438 445346 : tuple_fraction = 0.0;
439 : }
440 :
441 : /* primary planning entry point (may recurse for subqueries) */
442 450048 : root = subquery_planner(glob, parse, NULL, false, tuple_fraction, NULL);
443 :
444 : /* Select best Path and turn it into a Plan */
445 445706 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
446 445706 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
447 :
448 445706 : top_plan = create_plan(root, best_path);
449 :
450 : /*
451 : * If creating a plan for a scrollable cursor, make sure it can run
452 : * backwards on demand. Add a Material node at the top at need.
453 : */
454 445310 : if (cursorOptions & CURSOR_OPT_SCROLL)
455 : {
456 266 : if (!ExecSupportsBackwardScan(top_plan))
457 32 : top_plan = materialize_finished_plan(top_plan);
458 : }
459 :
460 : /*
461 : * Optionally add a Gather node for testing purposes, provided this is
462 : * actually a safe thing to do.
463 : *
464 : * We can add Gather even when top_plan has parallel-safe initPlans, but
465 : * then we have to move the initPlans to the Gather node because of
466 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
467 : * regression tests when debug_parallel_query = regress, because initPlans
468 : * that would normally appear on the top_plan move to the Gather, causing
469 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
470 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
471 : */
472 445310 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
473 194 : top_plan->parallel_safe &&
474 128 : (top_plan->initPlan == NIL ||
475 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
476 : {
477 128 : Gather *gather = makeNode(Gather);
478 : Cost initplan_cost;
479 : bool unsafe_initplans;
480 :
481 128 : gather->plan.targetlist = top_plan->targetlist;
482 128 : gather->plan.qual = NIL;
483 128 : gather->plan.lefttree = top_plan;
484 128 : gather->plan.righttree = NULL;
485 128 : gather->num_workers = 1;
486 128 : gather->single_copy = true;
487 128 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
488 :
489 : /* Transfer any initPlans to the new top node */
490 128 : gather->plan.initPlan = top_plan->initPlan;
491 128 : top_plan->initPlan = NIL;
492 :
493 : /*
494 : * Since this Gather has no parallel-aware descendants to signal to,
495 : * we don't need a rescan Param.
496 : */
497 128 : gather->rescan_param = -1;
498 :
499 : /*
500 : * Ideally we'd use cost_gather here, but setting up dummy path data
501 : * to satisfy it doesn't seem much cleaner than knowing what it does.
502 : */
503 128 : gather->plan.startup_cost = top_plan->startup_cost +
504 : parallel_setup_cost;
505 128 : gather->plan.total_cost = top_plan->total_cost +
506 128 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
507 128 : gather->plan.plan_rows = top_plan->plan_rows;
508 128 : gather->plan.plan_width = top_plan->plan_width;
509 128 : gather->plan.parallel_aware = false;
510 128 : gather->plan.parallel_safe = false;
511 :
512 : /*
513 : * Delete the initplans' cost from top_plan. We needn't add it to the
514 : * Gather node, since the above coding already included it there.
515 : */
516 128 : SS_compute_initplan_cost(gather->plan.initPlan,
517 : &initplan_cost, &unsafe_initplans);
518 128 : top_plan->startup_cost -= initplan_cost;
519 128 : top_plan->total_cost -= initplan_cost;
520 :
521 : /* use parallel mode for parallel plans. */
522 128 : root->glob->parallelModeNeeded = true;
523 :
524 128 : top_plan = &gather->plan;
525 : }
526 :
527 : /*
528 : * If any Params were generated, run through the plan tree and compute
529 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
530 : * set_plan_references' tree traversal, but for now it has to be separate
531 : * because we need to visit subplans before not after main plan.
532 : */
533 445310 : if (glob->paramExecTypes != NIL)
534 : {
535 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
536 197782 : forboth(lp, glob->subplans, lr, glob->subroots)
537 : {
538 43306 : Plan *subplan = (Plan *) lfirst(lp);
539 43306 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
540 :
541 43306 : SS_finalize_plan(subroot, subplan);
542 : }
543 154476 : SS_finalize_plan(root, top_plan);
544 : }
545 :
546 : /* final cleanup of the plan */
547 : Assert(glob->finalrtable == NIL);
548 : Assert(glob->finalrteperminfos == NIL);
549 : Assert(glob->finalrowmarks == NIL);
550 : Assert(glob->resultRelations == NIL);
551 : Assert(glob->appendRelations == NIL);
552 445310 : top_plan = set_plan_references(root, top_plan);
553 : /* ... and the subplans (both regular subplans and initplans) */
554 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
555 488616 : forboth(lp, glob->subplans, lr, glob->subroots)
556 : {
557 43306 : Plan *subplan = (Plan *) lfirst(lp);
558 43306 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
559 :
560 43306 : lfirst(lp) = set_plan_references(subroot, subplan);
561 : }
562 :
563 : /* build the PlannedStmt result */
564 445310 : result = makeNode(PlannedStmt);
565 :
566 445310 : result->commandType = parse->commandType;
567 445310 : result->queryId = parse->queryId;
568 445310 : result->planOrigin = PLAN_STMT_STANDARD;
569 445310 : result->hasReturning = (parse->returningList != NIL);
570 445310 : result->hasModifyingCTE = parse->hasModifyingCTE;
571 445310 : result->canSetTag = parse->canSetTag;
572 445310 : result->transientPlan = glob->transientPlan;
573 445310 : result->dependsOnRole = glob->dependsOnRole;
574 445310 : result->parallelModeNeeded = glob->parallelModeNeeded;
575 445310 : result->planTree = top_plan;
576 445310 : result->partPruneInfos = glob->partPruneInfos;
577 445310 : result->rtable = glob->finalrtable;
578 890620 : result->unprunableRelids = bms_difference(glob->allRelids,
579 445310 : glob->prunableRelids);
580 445310 : result->permInfos = glob->finalrteperminfos;
581 445310 : result->resultRelations = glob->resultRelations;
582 445310 : result->appendRelations = glob->appendRelations;
583 445310 : result->subplans = glob->subplans;
584 445310 : result->rewindPlanIDs = glob->rewindPlanIDs;
585 445310 : result->rowMarks = glob->finalrowmarks;
586 445310 : result->relationOids = glob->relationOids;
587 445310 : result->invalItems = glob->invalItems;
588 445310 : result->paramExecTypes = glob->paramExecTypes;
589 : /* utilityStmt should be null, but we might as well copy it */
590 445310 : result->utilityStmt = parse->utilityStmt;
591 445310 : result->stmt_location = parse->stmt_location;
592 445310 : result->stmt_len = parse->stmt_len;
593 :
594 445310 : result->jitFlags = PGJIT_NONE;
595 445310 : if (jit_enabled && jit_above_cost >= 0 &&
596 444526 : top_plan->total_cost > jit_above_cost)
597 : {
598 946 : result->jitFlags |= PGJIT_PERFORM;
599 :
600 : /*
601 : * Decide how much effort should be put into generating better code.
602 : */
603 946 : if (jit_optimize_above_cost >= 0 &&
604 946 : top_plan->total_cost > jit_optimize_above_cost)
605 432 : result->jitFlags |= PGJIT_OPT3;
606 946 : if (jit_inline_above_cost >= 0 &&
607 946 : top_plan->total_cost > jit_inline_above_cost)
608 432 : result->jitFlags |= PGJIT_INLINE;
609 :
610 : /*
611 : * Decide which operations should be JITed.
612 : */
613 946 : if (jit_expressions)
614 946 : result->jitFlags |= PGJIT_EXPR;
615 946 : if (jit_tuple_deforming)
616 946 : result->jitFlags |= PGJIT_DEFORM;
617 : }
618 :
619 445310 : if (glob->partition_directory != NULL)
620 11582 : DestroyPartitionDirectory(glob->partition_directory);
621 :
622 445310 : return result;
623 : }
624 :
625 :
626 : /*--------------------
627 : * subquery_planner
628 : * Invokes the planner on a subquery. We recurse to here for each
629 : * sub-SELECT found in the query tree.
630 : *
631 : * glob is the global state for the current planner run.
632 : * parse is the querytree produced by the parser & rewriter.
633 : * parent_root is the immediate parent Query's info (NULL at the top level).
634 : * hasRecursion is true if this is a recursive WITH query.
635 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
636 : * tuple_fraction is interpreted as explained for grouping_planner, below.
637 : * setops is used for set operation subqueries to provide the subquery with
638 : * the context in which it's being used so that Paths correctly sorted for the
639 : * set operation can be generated. NULL when not planning a set operation
640 : * child, or when a child of a set op that isn't interested in sorted input.
641 : *
642 : * Basically, this routine does the stuff that should only be done once
643 : * per Query object. It then calls grouping_planner. At one time,
644 : * grouping_planner could be invoked recursively on the same Query object;
645 : * that's not currently true, but we keep the separation between the two
646 : * routines anyway, in case we need it again someday.
647 : *
648 : * subquery_planner will be called recursively to handle sub-Query nodes
649 : * found within the query's expressions and rangetable.
650 : *
651 : * Returns the PlannerInfo struct ("root") that contains all data generated
652 : * while planning the subquery. In particular, the Path(s) attached to
653 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
654 : * cheapest way(s) to implement the query. The top level will select the
655 : * best Path and pass it through createplan.c to produce a finished Plan.
656 : *--------------------
657 : */
658 : PlannerInfo *
659 520588 : subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root,
660 : bool hasRecursion, double tuple_fraction,
661 : SetOperationStmt *setops)
662 : {
663 : PlannerInfo *root;
664 : List *newWithCheckOptions;
665 : List *newHaving;
666 : bool hasOuterJoins;
667 : bool hasResultRTEs;
668 : RelOptInfo *final_rel;
669 : ListCell *l;
670 :
671 : /* Create a PlannerInfo data structure for this subquery */
672 520588 : root = makeNode(PlannerInfo);
673 520588 : root->parse = parse;
674 520588 : root->glob = glob;
675 520588 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
676 520588 : root->parent_root = parent_root;
677 520588 : root->plan_params = NIL;
678 520588 : root->outer_params = NULL;
679 520588 : root->planner_cxt = CurrentMemoryContext;
680 520588 : root->init_plans = NIL;
681 520588 : root->cte_plan_ids = NIL;
682 520588 : root->multiexpr_params = NIL;
683 520588 : root->join_domains = NIL;
684 520588 : root->eq_classes = NIL;
685 520588 : root->ec_merging_done = false;
686 520588 : root->last_rinfo_serial = 0;
687 520588 : root->all_result_relids =
688 520588 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
689 520588 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
690 520588 : root->append_rel_list = NIL;
691 520588 : root->row_identity_vars = NIL;
692 520588 : root->rowMarks = NIL;
693 520588 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
694 520588 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
695 520588 : root->processed_groupClause = NIL;
696 520588 : root->processed_distinctClause = NIL;
697 520588 : root->processed_tlist = NIL;
698 520588 : root->update_colnos = NIL;
699 520588 : root->grouping_map = NULL;
700 520588 : root->minmax_aggs = NIL;
701 520588 : root->qual_security_level = 0;
702 520588 : root->hasPseudoConstantQuals = false;
703 520588 : root->hasAlternativeSubPlans = false;
704 520588 : root->placeholdersFrozen = false;
705 520588 : root->hasRecursion = hasRecursion;
706 520588 : if (hasRecursion)
707 926 : root->wt_param_id = assign_special_exec_param(root);
708 : else
709 519662 : root->wt_param_id = -1;
710 520588 : root->non_recursive_path = NULL;
711 520588 : root->partColsUpdated = false;
712 :
713 : /*
714 : * Create the top-level join domain. This won't have valid contents until
715 : * deconstruct_jointree fills it in, but the node needs to exist before
716 : * that so we can build EquivalenceClasses referencing it.
717 : */
718 520588 : root->join_domains = list_make1(makeNode(JoinDomain));
719 :
720 : /*
721 : * If there is a WITH list, process each WITH query and either convert it
722 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
723 : */
724 520588 : if (parse->cteList)
725 2856 : SS_process_ctes(root);
726 :
727 : /*
728 : * If it's a MERGE command, transform the joinlist as appropriate.
729 : */
730 520582 : transform_MERGE_to_join(parse);
731 :
732 : /*
733 : * Scan the rangetable for relation RTEs and retrieve the necessary
734 : * catalog information for each relation. Using this information, clear
735 : * the inh flag for any relation that has no children, collect not-null
736 : * attribute numbers for any relation that has column not-null
737 : * constraints, and expand virtual generated columns for any relation that
738 : * contains them. Note that this step does not descend into sublinks and
739 : * subqueries; if we pull up any sublinks or subqueries below, their
740 : * relation RTEs are processed just before pulling them up.
741 : */
742 520582 : parse = root->parse = preprocess_relation_rtes(root);
743 :
744 : /*
745 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
746 : * that we don't need so many special cases to deal with that situation.
747 : */
748 520582 : replace_empty_jointree(parse);
749 :
750 : /*
751 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
752 : * to transform them into joins. Note that this step does not descend
753 : * into subqueries; if we pull up any subqueries below, their SubLinks are
754 : * processed just before pulling them up.
755 : */
756 520582 : if (parse->hasSubLinks)
757 36050 : pull_up_sublinks(root);
758 :
759 : /*
760 : * Scan the rangetable for function RTEs, do const-simplification on them,
761 : * and then inline them if possible (producing subqueries that might get
762 : * pulled up next). Recursion issues here are handled in the same way as
763 : * for SubLinks.
764 : */
765 520582 : preprocess_function_rtes(root);
766 :
767 : /*
768 : * Check to see if any subqueries in the jointree can be merged into this
769 : * query.
770 : */
771 520576 : pull_up_subqueries(root);
772 :
773 : /*
774 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
775 : * do this now because it requires applying pull_up_subqueries to the leaf
776 : * queries of the UNION ALL, which weren't touched above because they
777 : * weren't referenced by the jointree (they will be after we do this).
778 : */
779 520570 : if (parse->setOperations)
780 6452 : flatten_simple_union_all(root);
781 :
782 : /*
783 : * Survey the rangetable to see what kinds of entries are present. We can
784 : * skip some later processing if relevant SQL features are not used; for
785 : * example if there are no JOIN RTEs we can avoid the expense of doing
786 : * flatten_join_alias_vars(). This must be done after we have finished
787 : * adding rangetable entries, of course. (Note: actually, processing of
788 : * inherited or partitioned rels can cause RTEs for their child tables to
789 : * get added later; but those must all be RTE_RELATION entries, so they
790 : * don't invalidate the conclusions drawn here.)
791 : */
792 520570 : root->hasJoinRTEs = false;
793 520570 : root->hasLateralRTEs = false;
794 520570 : root->group_rtindex = 0;
795 520570 : hasOuterJoins = false;
796 520570 : hasResultRTEs = false;
797 1410334 : foreach(l, parse->rtable)
798 : {
799 889764 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
800 :
801 889764 : switch (rte->rtekind)
802 : {
803 85706 : case RTE_JOIN:
804 85706 : root->hasJoinRTEs = true;
805 85706 : if (IS_OUTER_JOIN(rte->jointype))
806 47994 : hasOuterJoins = true;
807 85706 : break;
808 203134 : case RTE_RESULT:
809 203134 : hasResultRTEs = true;
810 203134 : break;
811 4466 : case RTE_GROUP:
812 : Assert(parse->hasGroupRTE);
813 4466 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
814 4466 : break;
815 596458 : default:
816 : /* No work here for other RTE types */
817 596458 : break;
818 : }
819 :
820 889764 : if (rte->lateral)
821 10636 : root->hasLateralRTEs = true;
822 :
823 : /*
824 : * We can also determine the maximum security level required for any
825 : * securityQuals now. Addition of inheritance-child RTEs won't affect
826 : * this, because child tables don't have their own securityQuals; see
827 : * expand_single_inheritance_child().
828 : */
829 889764 : if (rte->securityQuals)
830 2604 : root->qual_security_level = Max(root->qual_security_level,
831 : list_length(rte->securityQuals));
832 : }
833 :
834 : /*
835 : * If we have now verified that the query target relation is
836 : * non-inheriting, mark it as a leaf target.
837 : */
838 520570 : if (parse->resultRelation)
839 : {
840 91154 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
841 :
842 91154 : if (!rte->inh)
843 88280 : root->leaf_result_relids =
844 88280 : bms_make_singleton(parse->resultRelation);
845 : }
846 :
847 : /*
848 : * This would be a convenient time to check access permissions for all
849 : * relations mentioned in the query, since it would be better to fail now,
850 : * before doing any detailed planning. However, for historical reasons,
851 : * we leave this to be done at executor startup.
852 : *
853 : * Note, however, that we do need to check access permissions for any view
854 : * relations mentioned in the query, in order to prevent information being
855 : * leaked by selectivity estimation functions, which only check view owner
856 : * permissions on underlying tables (see all_rows_selectable() and its
857 : * callers). This is a little ugly, because it means that access
858 : * permissions for views will be checked twice, which is another reason
859 : * why it would be better to do all the ACL checks here.
860 : */
861 1409214 : foreach(l, parse->rtable)
862 : {
863 889030 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
864 :
865 889030 : if (rte->perminfoindex != 0 &&
866 477108 : rte->relkind == RELKIND_VIEW)
867 : {
868 : RTEPermissionInfo *perminfo;
869 : bool result;
870 :
871 20104 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
872 20104 : result = ExecCheckOneRelPerms(perminfo);
873 20104 : if (!result)
874 386 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
875 386 : get_rel_name(perminfo->relid));
876 : }
877 : }
878 :
879 : /*
880 : * Preprocess RowMark information. We need to do this after subquery
881 : * pullup, so that all base relations are present.
882 : */
883 520184 : preprocess_rowmarks(root);
884 :
885 : /*
886 : * Set hasHavingQual to remember if HAVING clause is present. Needed
887 : * because preprocess_expression will reduce a constant-true condition to
888 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
889 : */
890 520184 : root->hasHavingQual = (parse->havingQual != NULL);
891 :
892 : /*
893 : * Do expression preprocessing on targetlist and quals, as well as other
894 : * random expressions in the querytree. Note that we do not need to
895 : * handle sort/group expressions explicitly, because they are actually
896 : * part of the targetlist.
897 : */
898 516312 : parse->targetList = (List *)
899 520184 : preprocess_expression(root, (Node *) parse->targetList,
900 : EXPRKIND_TARGET);
901 :
902 516312 : newWithCheckOptions = NIL;
903 518800 : foreach(l, parse->withCheckOptions)
904 : {
905 2488 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
906 :
907 2488 : wco->qual = preprocess_expression(root, wco->qual,
908 : EXPRKIND_QUAL);
909 2488 : if (wco->qual != NULL)
910 2088 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
911 : }
912 516312 : parse->withCheckOptions = newWithCheckOptions;
913 :
914 516312 : parse->returningList = (List *)
915 516312 : preprocess_expression(root, (Node *) parse->returningList,
916 : EXPRKIND_TARGET);
917 :
918 516312 : preprocess_qual_conditions(root, (Node *) parse->jointree);
919 :
920 516312 : parse->havingQual = preprocess_expression(root, parse->havingQual,
921 : EXPRKIND_QUAL);
922 :
923 518930 : foreach(l, parse->windowClause)
924 : {
925 2618 : WindowClause *wc = lfirst_node(WindowClause, l);
926 :
927 : /* partitionClause/orderClause are sort/group expressions */
928 2618 : wc->startOffset = preprocess_expression(root, wc->startOffset,
929 : EXPRKIND_LIMIT);
930 2618 : wc->endOffset = preprocess_expression(root, wc->endOffset,
931 : EXPRKIND_LIMIT);
932 : }
933 :
934 516312 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
935 : EXPRKIND_LIMIT);
936 516312 : parse->limitCount = preprocess_expression(root, parse->limitCount,
937 : EXPRKIND_LIMIT);
938 :
939 516312 : if (parse->onConflict)
940 : {
941 3592 : parse->onConflict->arbiterElems = (List *)
942 1796 : preprocess_expression(root,
943 1796 : (Node *) parse->onConflict->arbiterElems,
944 : EXPRKIND_ARBITER_ELEM);
945 3592 : parse->onConflict->arbiterWhere =
946 1796 : preprocess_expression(root,
947 1796 : parse->onConflict->arbiterWhere,
948 : EXPRKIND_QUAL);
949 3592 : parse->onConflict->onConflictSet = (List *)
950 1796 : preprocess_expression(root,
951 1796 : (Node *) parse->onConflict->onConflictSet,
952 : EXPRKIND_TARGET);
953 1796 : parse->onConflict->onConflictWhere =
954 1796 : preprocess_expression(root,
955 1796 : parse->onConflict->onConflictWhere,
956 : EXPRKIND_QUAL);
957 : /* exclRelTlist contains only Vars, so no preprocessing needed */
958 : }
959 :
960 519036 : foreach(l, parse->mergeActionList)
961 : {
962 2724 : MergeAction *action = (MergeAction *) lfirst(l);
963 :
964 2724 : action->targetList = (List *)
965 2724 : preprocess_expression(root,
966 2724 : (Node *) action->targetList,
967 : EXPRKIND_TARGET);
968 2724 : action->qual =
969 2724 : preprocess_expression(root,
970 : (Node *) action->qual,
971 : EXPRKIND_QUAL);
972 : }
973 :
974 516312 : parse->mergeJoinCondition =
975 516312 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
976 :
977 516312 : root->append_rel_list = (List *)
978 516312 : preprocess_expression(root, (Node *) root->append_rel_list,
979 : EXPRKIND_APPINFO);
980 :
981 : /* Also need to preprocess expressions within RTEs */
982 1400488 : foreach(l, parse->rtable)
983 : {
984 884176 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
985 : int kind;
986 : ListCell *lcsq;
987 :
988 884176 : if (rte->rtekind == RTE_RELATION)
989 : {
990 461186 : if (rte->tablesample)
991 228 : rte->tablesample = (TableSampleClause *)
992 228 : preprocess_expression(root,
993 228 : (Node *) rte->tablesample,
994 : EXPRKIND_TABLESAMPLE);
995 : }
996 422990 : else if (rte->rtekind == RTE_SUBQUERY)
997 : {
998 : /*
999 : * We don't want to do all preprocessing yet on the subquery's
1000 : * expressions, since that will happen when we plan it. But if it
1001 : * contains any join aliases of our level, those have to get
1002 : * expanded now, because planning of the subquery won't do it.
1003 : * That's only possible if the subquery is LATERAL.
1004 : */
1005 68388 : if (rte->lateral && root->hasJoinRTEs)
1006 1432 : rte->subquery = (Query *)
1007 1432 : flatten_join_alias_vars(root, root->parse,
1008 1432 : (Node *) rte->subquery);
1009 : }
1010 354602 : else if (rte->rtekind == RTE_FUNCTION)
1011 : {
1012 : /* Preprocess the function expression(s) fully */
1013 50946 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1014 50946 : rte->functions = (List *)
1015 50946 : preprocess_expression(root, (Node *) rte->functions, kind);
1016 : }
1017 303656 : else if (rte->rtekind == RTE_TABLEFUNC)
1018 : {
1019 : /* Preprocess the function expression(s) fully */
1020 626 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1021 626 : rte->tablefunc = (TableFunc *)
1022 626 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1023 : }
1024 303030 : else if (rte->rtekind == RTE_VALUES)
1025 : {
1026 : /* Preprocess the values lists fully */
1027 8234 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1028 8234 : rte->values_lists = (List *)
1029 8234 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1030 : }
1031 294796 : else if (rte->rtekind == RTE_GROUP)
1032 : {
1033 : /* Preprocess the groupexprs list fully */
1034 4466 : rte->groupexprs = (List *)
1035 4466 : preprocess_expression(root, (Node *) rte->groupexprs,
1036 : EXPRKIND_GROUPEXPR);
1037 : }
1038 :
1039 : /*
1040 : * Process each element of the securityQuals list as if it were a
1041 : * separate qual expression (as indeed it is). We need to do it this
1042 : * way to get proper canonicalization of AND/OR structure. Note that
1043 : * this converts each element into an implicit-AND sublist.
1044 : */
1045 887120 : foreach(lcsq, rte->securityQuals)
1046 : {
1047 2944 : lfirst(lcsq) = preprocess_expression(root,
1048 2944 : (Node *) lfirst(lcsq),
1049 : EXPRKIND_QUAL);
1050 : }
1051 : }
1052 :
1053 : /*
1054 : * Now that we are done preprocessing expressions, and in particular done
1055 : * flattening join alias variables, get rid of the joinaliasvars lists.
1056 : * They no longer match what expressions in the rest of the tree look
1057 : * like, because we have not preprocessed expressions in those lists (and
1058 : * do not want to; for example, expanding a SubLink there would result in
1059 : * a useless unreferenced subplan). Leaving them in place simply creates
1060 : * a hazard for later scans of the tree. We could try to prevent that by
1061 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1062 : * but that doesn't sound very reliable.
1063 : */
1064 516312 : if (root->hasJoinRTEs)
1065 : {
1066 296208 : foreach(l, parse->rtable)
1067 : {
1068 244356 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1069 :
1070 244356 : rte->joinaliasvars = NIL;
1071 : }
1072 : }
1073 :
1074 : /*
1075 : * Replace any Vars in the subquery's targetlist and havingQual that
1076 : * reference GROUP outputs with the underlying grouping expressions.
1077 : *
1078 : * Note that we need to perform this replacement after we've preprocessed
1079 : * the grouping expressions. This is to ensure that there is only one
1080 : * instance of SubPlan for each SubLink contained within the grouping
1081 : * expressions.
1082 : */
1083 516312 : if (parse->hasGroupRTE)
1084 : {
1085 4466 : parse->targetList = (List *)
1086 4466 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1087 4466 : parse->havingQual =
1088 4466 : flatten_group_exprs(root, root->parse, parse->havingQual);
1089 : }
1090 :
1091 : /* Constant-folding might have removed all set-returning functions */
1092 516312 : if (parse->hasTargetSRFs)
1093 11674 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1094 :
1095 : /*
1096 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1097 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1098 : * volatile functions (since a HAVING clause is supposed to be executed
1099 : * only once per group). We also can't do this if there are any nonempty
1100 : * grouping sets and the clause references any columns that are nullable
1101 : * by the grouping sets; moving such a clause into WHERE would potentially
1102 : * change the results. (If there are only empty grouping sets, then the
1103 : * HAVING clause must be degenerate as discussed below.)
1104 : *
1105 : * Also, it may be that the clause is so expensive to execute that we're
1106 : * better off doing it only once per group, despite the loss of
1107 : * selectivity. This is hard to estimate short of doing the entire
1108 : * planning process twice, so we use a heuristic: clauses containing
1109 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1110 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1111 : * instead of after.
1112 : *
1113 : * If the query has explicit grouping then we can simply move such a
1114 : * clause into WHERE; any group that fails the clause will not be in the
1115 : * output because none of its tuples will reach the grouping or
1116 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
1117 : * HAVING clause, which we put in WHERE so that query_planner() can use it
1118 : * in a gating Result node, but also keep in HAVING to ensure that we
1119 : * don't emit a bogus aggregated row. (This could be done better, but it
1120 : * seems not worth optimizing.)
1121 : *
1122 : * Note that a HAVING clause may contain expressions that are not fully
1123 : * preprocessed. This can happen if these expressions are part of
1124 : * grouping items. In such cases, they are replaced with GROUP Vars in
1125 : * the parser and then replaced back after we've done with expression
1126 : * preprocessing on havingQual. This is not an issue if the clause
1127 : * remains in HAVING, because these expressions will be matched to lower
1128 : * target items in setrefs.c. However, if the clause is moved or copied
1129 : * into WHERE, we need to ensure that these expressions are fully
1130 : * preprocessed.
1131 : *
1132 : * Note that both havingQual and parse->jointree->quals are in
1133 : * implicitly-ANDed-list form at this point, even though they are declared
1134 : * as Node *.
1135 : */
1136 516312 : newHaving = NIL;
1137 517470 : foreach(l, (List *) parse->havingQual)
1138 : {
1139 1158 : Node *havingclause = (Node *) lfirst(l);
1140 :
1141 1494 : if (contain_agg_clause(havingclause) ||
1142 672 : contain_volatile_functions(havingclause) ||
1143 336 : contain_subplans(havingclause) ||
1144 420 : (parse->groupClause && parse->groupingSets &&
1145 84 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1146 : {
1147 : /* keep it in HAVING */
1148 894 : newHaving = lappend(newHaving, havingclause);
1149 : }
1150 264 : else if (parse->groupClause)
1151 : {
1152 : Node *whereclause;
1153 :
1154 : /* Preprocess the HAVING clause fully */
1155 246 : whereclause = preprocess_expression(root, havingclause,
1156 : EXPRKIND_QUAL);
1157 : /* ... and move it to WHERE */
1158 246 : parse->jointree->quals = (Node *)
1159 246 : list_concat((List *) parse->jointree->quals,
1160 : (List *) whereclause);
1161 : }
1162 : else
1163 : {
1164 : Node *whereclause;
1165 :
1166 : /* Preprocess the HAVING clause fully */
1167 18 : whereclause = preprocess_expression(root, copyObject(havingclause),
1168 : EXPRKIND_QUAL);
1169 : /* ... and put a copy in WHERE */
1170 36 : parse->jointree->quals = (Node *)
1171 18 : list_concat((List *) parse->jointree->quals,
1172 : (List *) whereclause);
1173 : /* ... and also keep it in HAVING */
1174 18 : newHaving = lappend(newHaving, havingclause);
1175 : }
1176 : }
1177 516312 : parse->havingQual = (Node *) newHaving;
1178 :
1179 : /*
1180 : * If we have any outer joins, try to reduce them to plain inner joins.
1181 : * This step is most easily done after we've done expression
1182 : * preprocessing.
1183 : */
1184 516312 : if (hasOuterJoins)
1185 33614 : reduce_outer_joins(root);
1186 :
1187 : /*
1188 : * If we have any RTE_RESULT relations, see if they can be deleted from
1189 : * the jointree. We also rely on this processing to flatten single-child
1190 : * FromExprs underneath outer joins. This step is most effectively done
1191 : * after we've done expression preprocessing and outer join reduction.
1192 : */
1193 516312 : if (hasResultRTEs || hasOuterJoins)
1194 231358 : remove_useless_result_rtes(root);
1195 :
1196 : /*
1197 : * Do the main planning.
1198 : */
1199 516312 : grouping_planner(root, tuple_fraction, setops);
1200 :
1201 : /*
1202 : * Capture the set of outer-level param IDs we have access to, for use in
1203 : * extParam/allParam calculations later.
1204 : */
1205 516240 : SS_identify_outer_params(root);
1206 :
1207 : /*
1208 : * If any initPlans were created in this query level, adjust the surviving
1209 : * Paths' costs and parallel-safety flags to account for them. The
1210 : * initPlans won't actually get attached to the plan tree till
1211 : * create_plan() runs, but we must include their effects now.
1212 : */
1213 516240 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1214 516240 : SS_charge_for_initplans(root, final_rel);
1215 :
1216 : /*
1217 : * Make sure we've identified the cheapest Path for the final rel. (By
1218 : * doing this here not in grouping_planner, we include initPlan costs in
1219 : * the decision, though it's unlikely that will change anything.)
1220 : */
1221 516240 : set_cheapest(final_rel);
1222 :
1223 516240 : return root;
1224 : }
1225 :
1226 : /*
1227 : * preprocess_expression
1228 : * Do subquery_planner's preprocessing work for an expression,
1229 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1230 : * conditions), a HAVING clause, or a few other things.
1231 : */
1232 : static Node *
1233 4330430 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1234 : {
1235 : /*
1236 : * Fall out quickly if expression is empty. This occurs often enough to
1237 : * be worth checking. Note that null->null is the correct conversion for
1238 : * implicit-AND result format, too.
1239 : */
1240 4330430 : if (expr == NULL)
1241 3424402 : return NULL;
1242 :
1243 : /*
1244 : * If the query has any join RTEs, replace join alias variables with
1245 : * base-relation variables. We must do this first, since any expressions
1246 : * we may extract from the joinaliasvars lists have not been preprocessed.
1247 : * For example, if we did this after sublink processing, sublinks expanded
1248 : * out from join aliases would not get processed. But we can skip this in
1249 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1250 : * they can't contain any Vars of the current query level.
1251 : */
1252 906028 : if (root->hasJoinRTEs &&
1253 374878 : !(kind == EXPRKIND_RTFUNC ||
1254 187266 : kind == EXPRKIND_VALUES ||
1255 : kind == EXPRKIND_TABLESAMPLE ||
1256 : kind == EXPRKIND_TABLEFUNC))
1257 187248 : expr = flatten_join_alias_vars(root, root->parse, expr);
1258 :
1259 : /*
1260 : * Simplify constant expressions. For function RTEs, this was already
1261 : * done by preprocess_function_rtes. (But note we must do it again for
1262 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1263 : * un-simplified subexpressions inserted by flattening of subqueries or
1264 : * join alias variables.)
1265 : *
1266 : * Note: an essential effect of this is to convert named-argument function
1267 : * calls to positional notation and insert the current actual values of
1268 : * any default arguments for functions. To ensure that happens, we *must*
1269 : * process all expressions here. Previous PG versions sometimes skipped
1270 : * const-simplification if it didn't seem worth the trouble, but we can't
1271 : * do that anymore.
1272 : *
1273 : * Note: this also flattens nested AND and OR expressions into N-argument
1274 : * form. All processing of a qual expression after this point must be
1275 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1276 : * with AND directly under AND, nor OR directly under OR.
1277 : */
1278 906028 : if (kind != EXPRKIND_RTFUNC)
1279 863362 : expr = eval_const_expressions(root, expr);
1280 :
1281 : /*
1282 : * If it's a qual or havingQual, canonicalize it.
1283 : */
1284 902156 : if (kind == EXPRKIND_QUAL)
1285 : {
1286 320930 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1287 :
1288 : #ifdef OPTIMIZER_DEBUG
1289 : printf("After canonicalize_qual()\n");
1290 : pprint(expr);
1291 : #endif
1292 : }
1293 :
1294 : /*
1295 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1296 : * hashfuncid of any that might execute more quickly by using hash lookups
1297 : * instead of a linear search.
1298 : */
1299 902156 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1300 : {
1301 824402 : convert_saop_to_hashed_saop(expr);
1302 : }
1303 :
1304 : /* Expand SubLinks to SubPlans */
1305 902156 : if (root->parse->hasSubLinks)
1306 104116 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1307 :
1308 : /*
1309 : * XXX do not insert anything here unless you have grokked the comments in
1310 : * SS_replace_correlation_vars ...
1311 : */
1312 :
1313 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1314 902156 : if (root->query_level > 1)
1315 154704 : expr = SS_replace_correlation_vars(root, expr);
1316 :
1317 : /*
1318 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1319 : * don't want to do this before eval_const_expressions, since the latter
1320 : * would be unable to simplify a top-level AND correctly. Also,
1321 : * SS_process_sublinks expects explicit-AND format.)
1322 : */
1323 902156 : if (kind == EXPRKIND_QUAL)
1324 320930 : expr = (Node *) make_ands_implicit((Expr *) expr);
1325 :
1326 902156 : return expr;
1327 : }
1328 :
1329 : /*
1330 : * preprocess_qual_conditions
1331 : * Recursively scan the query's jointree and do subquery_planner's
1332 : * preprocessing work on each qual condition found therein.
1333 : */
1334 : static void
1335 1277672 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1336 : {
1337 1277672 : if (jtnode == NULL)
1338 0 : return;
1339 1277672 : if (IsA(jtnode, RangeTblRef))
1340 : {
1341 : /* nothing to do here */
1342 : }
1343 624220 : else if (IsA(jtnode, FromExpr))
1344 : {
1345 530928 : FromExpr *f = (FromExpr *) jtnode;
1346 : ListCell *l;
1347 :
1348 1105704 : foreach(l, f->fromlist)
1349 574776 : preprocess_qual_conditions(root, lfirst(l));
1350 :
1351 530928 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1352 : }
1353 93292 : else if (IsA(jtnode, JoinExpr))
1354 : {
1355 93292 : JoinExpr *j = (JoinExpr *) jtnode;
1356 :
1357 93292 : preprocess_qual_conditions(root, j->larg);
1358 93292 : preprocess_qual_conditions(root, j->rarg);
1359 :
1360 93292 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1361 : }
1362 : else
1363 0 : elog(ERROR, "unrecognized node type: %d",
1364 : (int) nodeTag(jtnode));
1365 : }
1366 :
1367 : /*
1368 : * preprocess_phv_expression
1369 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1370 : *
1371 : * If a LATERAL subquery references an output of another subquery, and that
1372 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1373 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1374 : * and later pull it back up during find_lateral_references, which runs after
1375 : * subquery_planner has preprocessed all the expressions that were in the
1376 : * current query level to start with. So we need to preprocess it then.
1377 : */
1378 : Expr *
1379 90 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1380 : {
1381 90 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1382 : }
1383 :
1384 : /*--------------------
1385 : * grouping_planner
1386 : * Perform planning steps related to grouping, aggregation, etc.
1387 : *
1388 : * This function adds all required top-level processing to the scan/join
1389 : * Path(s) produced by query_planner.
1390 : *
1391 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1392 : * tuple_fraction is interpreted as follows:
1393 : * 0: expect all tuples to be retrieved (normal case)
1394 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1395 : * from the plan to be retrieved
1396 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1397 : * expected to be retrieved (ie, a LIMIT specification).
1398 : * setops is used for set operation subqueries to provide the subquery with
1399 : * the context in which it's being used so that Paths correctly sorted for the
1400 : * set operation can be generated. NULL when not planning a set operation
1401 : * child, or when a child of a set op that isn't interested in sorted input.
1402 : *
1403 : * Returns nothing; the useful output is in the Paths we attach to the
1404 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1405 : * root->processed_tlist contains the final processed targetlist.
1406 : *
1407 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1408 : * to leave this to the caller.
1409 : *--------------------
1410 : */
1411 : static void
1412 516312 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1413 : SetOperationStmt *setops)
1414 : {
1415 516312 : Query *parse = root->parse;
1416 516312 : int64 offset_est = 0;
1417 516312 : int64 count_est = 0;
1418 516312 : double limit_tuples = -1.0;
1419 516312 : bool have_postponed_srfs = false;
1420 : PathTarget *final_target;
1421 : List *final_targets;
1422 : List *final_targets_contain_srfs;
1423 : bool final_target_parallel_safe;
1424 : RelOptInfo *current_rel;
1425 : RelOptInfo *final_rel;
1426 : FinalPathExtraData extra;
1427 : ListCell *lc;
1428 :
1429 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1430 516312 : if (parse->limitCount || parse->limitOffset)
1431 : {
1432 5006 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1433 : &offset_est, &count_est);
1434 :
1435 : /*
1436 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1437 : * estimate the effects of using a bounded sort.
1438 : */
1439 5006 : if (count_est > 0 && offset_est >= 0)
1440 4472 : limit_tuples = (double) count_est + (double) offset_est;
1441 : }
1442 :
1443 : /* Make tuple_fraction accessible to lower-level routines */
1444 516312 : root->tuple_fraction = tuple_fraction;
1445 :
1446 516312 : if (parse->setOperations)
1447 : {
1448 : /*
1449 : * Construct Paths for set operations. The results will not need any
1450 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1451 : * special work for recursive unions is the responsibility of
1452 : * plan_set_operations.
1453 : */
1454 5982 : current_rel = plan_set_operations(root);
1455 :
1456 : /*
1457 : * We should not need to call preprocess_targetlist, since we must be
1458 : * in a SELECT query node. Instead, use the processed_tlist returned
1459 : * by plan_set_operations (since this tells whether it returned any
1460 : * resjunk columns!), and transfer any sort key information from the
1461 : * original tlist.
1462 : */
1463 : Assert(parse->commandType == CMD_SELECT);
1464 :
1465 : /* for safety, copy processed_tlist instead of modifying in-place */
1466 5976 : root->processed_tlist =
1467 5976 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1468 : parse->targetList);
1469 :
1470 : /* Also extract the PathTarget form of the setop result tlist */
1471 5976 : final_target = current_rel->cheapest_total_path->pathtarget;
1472 :
1473 : /* And check whether it's parallel safe */
1474 : final_target_parallel_safe =
1475 5976 : is_parallel_safe(root, (Node *) final_target->exprs);
1476 :
1477 : /* The setop result tlist couldn't contain any SRFs */
1478 : Assert(!parse->hasTargetSRFs);
1479 5976 : final_targets = final_targets_contain_srfs = NIL;
1480 :
1481 : /*
1482 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1483 : * checked already, but let's make sure).
1484 : */
1485 5976 : if (parse->rowMarks)
1486 0 : ereport(ERROR,
1487 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1488 : /*------
1489 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1490 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1491 : LCS_asString(linitial_node(RowMarkClause,
1492 : parse->rowMarks)->strength))));
1493 :
1494 : /*
1495 : * Calculate pathkeys that represent result ordering requirements
1496 : */
1497 : Assert(parse->distinctClause == NIL);
1498 5976 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1499 : parse->sortClause,
1500 : root->processed_tlist);
1501 : }
1502 : else
1503 : {
1504 : /* No set operations, do regular planning */
1505 : PathTarget *sort_input_target;
1506 : List *sort_input_targets;
1507 : List *sort_input_targets_contain_srfs;
1508 : bool sort_input_target_parallel_safe;
1509 : PathTarget *grouping_target;
1510 : List *grouping_targets;
1511 : List *grouping_targets_contain_srfs;
1512 : bool grouping_target_parallel_safe;
1513 : PathTarget *scanjoin_target;
1514 : List *scanjoin_targets;
1515 : List *scanjoin_targets_contain_srfs;
1516 : bool scanjoin_target_parallel_safe;
1517 : bool scanjoin_target_same_exprs;
1518 : bool have_grouping;
1519 510330 : WindowFuncLists *wflists = NULL;
1520 510330 : List *activeWindows = NIL;
1521 510330 : grouping_sets_data *gset_data = NULL;
1522 : standard_qp_extra qp_extra;
1523 :
1524 : /* A recursive query should always have setOperations */
1525 : Assert(!root->hasRecursion);
1526 :
1527 : /* Preprocess grouping sets and GROUP BY clause, if any */
1528 510330 : if (parse->groupingSets)
1529 : {
1530 878 : gset_data = preprocess_grouping_sets(root);
1531 : }
1532 509452 : else if (parse->groupClause)
1533 : {
1534 : /* Preprocess regular GROUP BY clause, if any */
1535 3630 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1536 : }
1537 :
1538 : /*
1539 : * Preprocess targetlist. Note that much of the remaining planning
1540 : * work will be done with the PathTarget representation of tlists, but
1541 : * we must also maintain the full representation of the final tlist so
1542 : * that we can transfer its decoration (resnames etc) to the topmost
1543 : * tlist of the finished Plan. This is kept in processed_tlist.
1544 : */
1545 510324 : preprocess_targetlist(root);
1546 :
1547 : /*
1548 : * Mark all the aggregates with resolved aggtranstypes, and detect
1549 : * aggregates that are duplicates or can share transition state. We
1550 : * must do this before slicing and dicing the tlist into various
1551 : * pathtargets, else some copies of the Aggref nodes might escape
1552 : * being marked.
1553 : */
1554 510324 : if (parse->hasAggs)
1555 : {
1556 38152 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1557 38152 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1558 : }
1559 :
1560 : /*
1561 : * Locate any window functions in the tlist. (We don't need to look
1562 : * anywhere else, since expressions used in ORDER BY will be in there
1563 : * too.) Note that they could all have been eliminated by constant
1564 : * folding, in which case we don't need to do any more work.
1565 : */
1566 510324 : if (parse->hasWindowFuncs)
1567 : {
1568 2384 : wflists = find_window_functions((Node *) root->processed_tlist,
1569 2384 : list_length(parse->windowClause));
1570 2384 : if (wflists->numWindowFuncs > 0)
1571 : {
1572 : /*
1573 : * See if any modifications can be made to each WindowClause
1574 : * to allow the executor to execute the WindowFuncs more
1575 : * quickly.
1576 : */
1577 2378 : optimize_window_clauses(root, wflists);
1578 :
1579 : /* Extract the list of windows actually in use. */
1580 2378 : activeWindows = select_active_windows(root, wflists);
1581 :
1582 : /* Make sure they all have names, for EXPLAIN's use. */
1583 2378 : name_active_windows(activeWindows);
1584 : }
1585 : else
1586 6 : parse->hasWindowFuncs = false;
1587 : }
1588 :
1589 : /*
1590 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1591 : * adding logic between here and the query_planner() call. Anything
1592 : * that is needed in MIN/MAX-optimizable cases will have to be
1593 : * duplicated in planagg.c.
1594 : */
1595 510324 : if (parse->hasAggs)
1596 38152 : preprocess_minmax_aggregates(root);
1597 :
1598 : /*
1599 : * Figure out whether there's a hard limit on the number of rows that
1600 : * query_planner's result subplan needs to return. Even if we know a
1601 : * hard limit overall, it doesn't apply if the query has any
1602 : * grouping/aggregation operations, or SRFs in the tlist.
1603 : */
1604 510324 : if (parse->groupClause ||
1605 505864 : parse->groupingSets ||
1606 505822 : parse->distinctClause ||
1607 503236 : parse->hasAggs ||
1608 468974 : parse->hasWindowFuncs ||
1609 466734 : parse->hasTargetSRFs ||
1610 455528 : root->hasHavingQual)
1611 54814 : root->limit_tuples = -1.0;
1612 : else
1613 455510 : root->limit_tuples = limit_tuples;
1614 :
1615 : /* Set up data needed by standard_qp_callback */
1616 510324 : qp_extra.activeWindows = activeWindows;
1617 510324 : qp_extra.gset_data = gset_data;
1618 :
1619 : /*
1620 : * If we're a subquery for a set operation, store the SetOperationStmt
1621 : * in qp_extra.
1622 : */
1623 510324 : qp_extra.setop = setops;
1624 :
1625 : /*
1626 : * Generate the best unsorted and presorted paths for the scan/join
1627 : * portion of this Query, ie the processing represented by the
1628 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1629 : * We also generate (in standard_qp_callback) pathkey representations
1630 : * of the query's sort clause, distinct clause, etc.
1631 : */
1632 510324 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1633 :
1634 : /*
1635 : * Convert the query's result tlist into PathTarget format.
1636 : *
1637 : * Note: this cannot be done before query_planner() has performed
1638 : * appendrel expansion, because that might add resjunk entries to
1639 : * root->processed_tlist. Waiting till afterwards is also helpful
1640 : * because the target width estimates can use per-Var width numbers
1641 : * that were obtained within query_planner().
1642 : */
1643 510270 : final_target = create_pathtarget(root, root->processed_tlist);
1644 : final_target_parallel_safe =
1645 510270 : is_parallel_safe(root, (Node *) final_target->exprs);
1646 :
1647 : /*
1648 : * If ORDER BY was given, consider whether we should use a post-sort
1649 : * projection, and compute the adjusted target for preceding steps if
1650 : * so.
1651 : */
1652 510270 : if (parse->sortClause)
1653 : {
1654 69806 : sort_input_target = make_sort_input_target(root,
1655 : final_target,
1656 : &have_postponed_srfs);
1657 : sort_input_target_parallel_safe =
1658 69806 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1659 : }
1660 : else
1661 : {
1662 440464 : sort_input_target = final_target;
1663 440464 : sort_input_target_parallel_safe = final_target_parallel_safe;
1664 : }
1665 :
1666 : /*
1667 : * If we have window functions to deal with, the output from any
1668 : * grouping step needs to be what the window functions want;
1669 : * otherwise, it should be sort_input_target.
1670 : */
1671 510270 : if (activeWindows)
1672 : {
1673 2378 : grouping_target = make_window_input_target(root,
1674 : final_target,
1675 : activeWindows);
1676 : grouping_target_parallel_safe =
1677 2378 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1678 : }
1679 : else
1680 : {
1681 507892 : grouping_target = sort_input_target;
1682 507892 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1683 : }
1684 :
1685 : /*
1686 : * If we have grouping or aggregation to do, the topmost scan/join
1687 : * plan node must emit what the grouping step wants; otherwise, it
1688 : * should emit grouping_target.
1689 : */
1690 505810 : have_grouping = (parse->groupClause || parse->groupingSets ||
1691 1016080 : parse->hasAggs || root->hasHavingQual);
1692 510270 : if (have_grouping)
1693 : {
1694 38820 : scanjoin_target = make_group_input_target(root, final_target);
1695 : scanjoin_target_parallel_safe =
1696 38820 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1697 : }
1698 : else
1699 : {
1700 471450 : scanjoin_target = grouping_target;
1701 471450 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1702 : }
1703 :
1704 : /*
1705 : * If there are any SRFs in the targetlist, we must separate each of
1706 : * these PathTargets into SRF-computing and SRF-free targets. Replace
1707 : * each of the named targets with a SRF-free version, and remember the
1708 : * list of additional projection steps we need to add afterwards.
1709 : */
1710 510270 : if (parse->hasTargetSRFs)
1711 : {
1712 : /* final_target doesn't recompute any SRFs in sort_input_target */
1713 11674 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1714 : &final_targets,
1715 : &final_targets_contain_srfs);
1716 11674 : final_target = linitial_node(PathTarget, final_targets);
1717 : Assert(!linitial_int(final_targets_contain_srfs));
1718 : /* likewise for sort_input_target vs. grouping_target */
1719 11674 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1720 : &sort_input_targets,
1721 : &sort_input_targets_contain_srfs);
1722 11674 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
1723 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1724 : /* likewise for grouping_target vs. scanjoin_target */
1725 11674 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
1726 : &grouping_targets,
1727 : &grouping_targets_contain_srfs);
1728 11674 : grouping_target = linitial_node(PathTarget, grouping_targets);
1729 : Assert(!linitial_int(grouping_targets_contain_srfs));
1730 : /* scanjoin_target will not have any SRFs precomputed for it */
1731 11674 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1732 : &scanjoin_targets,
1733 : &scanjoin_targets_contain_srfs);
1734 11674 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1735 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1736 : }
1737 : else
1738 : {
1739 : /* initialize lists; for most of these, dummy values are OK */
1740 498596 : final_targets = final_targets_contain_srfs = NIL;
1741 498596 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1742 498596 : grouping_targets = grouping_targets_contain_srfs = NIL;
1743 498596 : scanjoin_targets = list_make1(scanjoin_target);
1744 498596 : scanjoin_targets_contain_srfs = NIL;
1745 : }
1746 :
1747 : /* Apply scan/join target. */
1748 510270 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1749 510270 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1750 510270 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1751 : scanjoin_targets_contain_srfs,
1752 : scanjoin_target_parallel_safe,
1753 : scanjoin_target_same_exprs);
1754 :
1755 : /*
1756 : * Save the various upper-rel PathTargets we just computed into
1757 : * root->upper_targets[]. The core code doesn't use this, but it
1758 : * provides a convenient place for extensions to get at the info. For
1759 : * consistency, we save all the intermediate targets, even though some
1760 : * of the corresponding upperrels might not be needed for this query.
1761 : */
1762 510270 : root->upper_targets[UPPERREL_FINAL] = final_target;
1763 510270 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1764 510270 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1765 510270 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1766 510270 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1767 510270 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1768 :
1769 : /*
1770 : * If we have grouping and/or aggregation, consider ways to implement
1771 : * that. We build a new upperrel representing the output of this
1772 : * phase.
1773 : */
1774 510270 : if (have_grouping)
1775 : {
1776 38820 : current_rel = create_grouping_paths(root,
1777 : current_rel,
1778 : grouping_target,
1779 : grouping_target_parallel_safe,
1780 : gset_data);
1781 : /* Fix things up if grouping_target contains SRFs */
1782 38814 : if (parse->hasTargetSRFs)
1783 426 : adjust_paths_for_srfs(root, current_rel,
1784 : grouping_targets,
1785 : grouping_targets_contain_srfs);
1786 : }
1787 :
1788 : /*
1789 : * If we have window functions, consider ways to implement those. We
1790 : * build a new upperrel representing the output of this phase.
1791 : */
1792 510264 : if (activeWindows)
1793 : {
1794 2378 : current_rel = create_window_paths(root,
1795 : current_rel,
1796 : grouping_target,
1797 : sort_input_target,
1798 : sort_input_target_parallel_safe,
1799 : wflists,
1800 : activeWindows);
1801 : /* Fix things up if sort_input_target contains SRFs */
1802 2378 : if (parse->hasTargetSRFs)
1803 12 : adjust_paths_for_srfs(root, current_rel,
1804 : sort_input_targets,
1805 : sort_input_targets_contain_srfs);
1806 : }
1807 :
1808 : /*
1809 : * If there is a DISTINCT clause, consider ways to implement that. We
1810 : * build a new upperrel representing the output of this phase.
1811 : */
1812 510264 : if (parse->distinctClause)
1813 : {
1814 2620 : current_rel = create_distinct_paths(root,
1815 : current_rel,
1816 : sort_input_target);
1817 : }
1818 : } /* end of if (setOperations) */
1819 :
1820 : /*
1821 : * If ORDER BY was given, consider ways to implement that, and generate a
1822 : * new upperrel containing only paths that emit the correct ordering and
1823 : * project the correct final_target. We can apply the original
1824 : * limit_tuples limit in sort costing here, but only if there are no
1825 : * postponed SRFs.
1826 : */
1827 516240 : if (parse->sortClause)
1828 : {
1829 73652 : current_rel = create_ordered_paths(root,
1830 : current_rel,
1831 : final_target,
1832 : final_target_parallel_safe,
1833 : have_postponed_srfs ? -1.0 :
1834 : limit_tuples);
1835 : /* Fix things up if final_target contains SRFs */
1836 73652 : if (parse->hasTargetSRFs)
1837 196 : adjust_paths_for_srfs(root, current_rel,
1838 : final_targets,
1839 : final_targets_contain_srfs);
1840 : }
1841 :
1842 : /*
1843 : * Now we are prepared to build the final-output upperrel.
1844 : */
1845 516240 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1846 :
1847 : /*
1848 : * If the input rel is marked consider_parallel and there's nothing that's
1849 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1850 : * consider_parallel as well. Note that if the query has rowMarks or is
1851 : * not a SELECT, consider_parallel will be false for every relation in the
1852 : * query.
1853 : */
1854 685300 : if (current_rel->consider_parallel &&
1855 338096 : is_parallel_safe(root, parse->limitOffset) &&
1856 169036 : is_parallel_safe(root, parse->limitCount))
1857 169030 : final_rel->consider_parallel = true;
1858 :
1859 : /*
1860 : * If the current_rel belongs to a single FDW, so does the final_rel.
1861 : */
1862 516240 : final_rel->serverid = current_rel->serverid;
1863 516240 : final_rel->userid = current_rel->userid;
1864 516240 : final_rel->useridiscurrent = current_rel->useridiscurrent;
1865 516240 : final_rel->fdwroutine = current_rel->fdwroutine;
1866 :
1867 : /*
1868 : * Generate paths for the final_rel. Insert all surviving paths, with
1869 : * LockRows, Limit, and/or ModifyTable steps added if needed.
1870 : */
1871 1051850 : foreach(lc, current_rel->pathlist)
1872 : {
1873 535610 : Path *path = (Path *) lfirst(lc);
1874 :
1875 : /*
1876 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1877 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
1878 : * here. If there are only non-locking rowmarks, they should be
1879 : * handled by the ModifyTable node instead. However, root->rowMarks
1880 : * is what goes into the LockRows node.)
1881 : */
1882 535610 : if (parse->rowMarks)
1883 : {
1884 8246 : path = (Path *) create_lockrows_path(root, final_rel, path,
1885 : root->rowMarks,
1886 : assign_special_exec_param(root));
1887 : }
1888 :
1889 : /*
1890 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1891 : */
1892 535610 : if (limit_needed(parse))
1893 : {
1894 5932 : path = (Path *) create_limit_path(root, final_rel, path,
1895 : parse->limitOffset,
1896 : parse->limitCount,
1897 : parse->limitOption,
1898 : offset_est, count_est);
1899 : }
1900 :
1901 : /*
1902 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
1903 : */
1904 535610 : if (parse->commandType != CMD_SELECT)
1905 : {
1906 : Index rootRelation;
1907 90624 : List *resultRelations = NIL;
1908 90624 : List *updateColnosLists = NIL;
1909 90624 : List *withCheckOptionLists = NIL;
1910 90624 : List *returningLists = NIL;
1911 90624 : List *mergeActionLists = NIL;
1912 90624 : List *mergeJoinConditions = NIL;
1913 : List *rowMarks;
1914 :
1915 90624 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
1916 : {
1917 : /* Inherited UPDATE/DELETE/MERGE */
1918 2844 : RelOptInfo *top_result_rel = find_base_rel(root,
1919 : parse->resultRelation);
1920 2844 : int resultRelation = -1;
1921 :
1922 : /* Pass the root result rel forward to the executor. */
1923 2844 : rootRelation = parse->resultRelation;
1924 :
1925 : /* Add only leaf children to ModifyTable. */
1926 8306 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
1927 8306 : resultRelation)) >= 0)
1928 : {
1929 5462 : RelOptInfo *this_result_rel = find_base_rel(root,
1930 : resultRelation);
1931 :
1932 : /*
1933 : * Also exclude any leaf rels that have turned dummy since
1934 : * being added to the list, for example, by being excluded
1935 : * by constraint exclusion.
1936 : */
1937 5462 : if (IS_DUMMY_REL(this_result_rel))
1938 174 : continue;
1939 :
1940 : /* Build per-target-rel lists needed by ModifyTable */
1941 5288 : resultRelations = lappend_int(resultRelations,
1942 : resultRelation);
1943 5288 : if (parse->commandType == CMD_UPDATE)
1944 : {
1945 3636 : List *update_colnos = root->update_colnos;
1946 :
1947 3636 : if (this_result_rel != top_result_rel)
1948 : update_colnos =
1949 3636 : adjust_inherited_attnums_multilevel(root,
1950 : update_colnos,
1951 : this_result_rel->relid,
1952 : top_result_rel->relid);
1953 3636 : updateColnosLists = lappend(updateColnosLists,
1954 : update_colnos);
1955 : }
1956 5288 : if (parse->withCheckOptions)
1957 : {
1958 504 : List *withCheckOptions = parse->withCheckOptions;
1959 :
1960 504 : if (this_result_rel != top_result_rel)
1961 : withCheckOptions = (List *)
1962 504 : adjust_appendrel_attrs_multilevel(root,
1963 : (Node *) withCheckOptions,
1964 : this_result_rel,
1965 : top_result_rel);
1966 504 : withCheckOptionLists = lappend(withCheckOptionLists,
1967 : withCheckOptions);
1968 : }
1969 5288 : if (parse->returningList)
1970 : {
1971 840 : List *returningList = parse->returningList;
1972 :
1973 840 : if (this_result_rel != top_result_rel)
1974 : returningList = (List *)
1975 840 : adjust_appendrel_attrs_multilevel(root,
1976 : (Node *) returningList,
1977 : this_result_rel,
1978 : top_result_rel);
1979 840 : returningLists = lappend(returningLists,
1980 : returningList);
1981 : }
1982 5288 : if (parse->mergeActionList)
1983 : {
1984 : ListCell *l;
1985 522 : List *mergeActionList = NIL;
1986 :
1987 : /*
1988 : * Copy MergeActions and translate stuff that
1989 : * references attribute numbers.
1990 : */
1991 1614 : foreach(l, parse->mergeActionList)
1992 : {
1993 1092 : MergeAction *action = lfirst(l),
1994 1092 : *leaf_action = copyObject(action);
1995 :
1996 1092 : leaf_action->qual =
1997 1092 : adjust_appendrel_attrs_multilevel(root,
1998 : (Node *) action->qual,
1999 : this_result_rel,
2000 : top_result_rel);
2001 1092 : leaf_action->targetList = (List *)
2002 1092 : adjust_appendrel_attrs_multilevel(root,
2003 1092 : (Node *) action->targetList,
2004 : this_result_rel,
2005 : top_result_rel);
2006 1092 : if (leaf_action->commandType == CMD_UPDATE)
2007 592 : leaf_action->updateColnos =
2008 592 : adjust_inherited_attnums_multilevel(root,
2009 : action->updateColnos,
2010 : this_result_rel->relid,
2011 : top_result_rel->relid);
2012 1092 : mergeActionList = lappend(mergeActionList,
2013 : leaf_action);
2014 : }
2015 :
2016 522 : mergeActionLists = lappend(mergeActionLists,
2017 : mergeActionList);
2018 : }
2019 5288 : if (parse->commandType == CMD_MERGE)
2020 : {
2021 522 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2022 :
2023 522 : if (this_result_rel != top_result_rel)
2024 : mergeJoinCondition =
2025 522 : adjust_appendrel_attrs_multilevel(root,
2026 : mergeJoinCondition,
2027 : this_result_rel,
2028 : top_result_rel);
2029 522 : mergeJoinConditions = lappend(mergeJoinConditions,
2030 : mergeJoinCondition);
2031 : }
2032 : }
2033 :
2034 2844 : if (resultRelations == NIL)
2035 : {
2036 : /*
2037 : * We managed to exclude every child rel, so generate a
2038 : * dummy one-relation plan using info for the top target
2039 : * rel (even though that may not be a leaf target).
2040 : * Although it's clear that no data will be updated or
2041 : * deleted, we still need to have a ModifyTable node so
2042 : * that any statement triggers will be executed. (This
2043 : * could be cleaner if we fixed nodeModifyTable.c to allow
2044 : * zero target relations, but that probably wouldn't be a
2045 : * net win.)
2046 : */
2047 30 : resultRelations = list_make1_int(parse->resultRelation);
2048 30 : if (parse->commandType == CMD_UPDATE)
2049 30 : updateColnosLists = list_make1(root->update_colnos);
2050 30 : if (parse->withCheckOptions)
2051 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2052 30 : if (parse->returningList)
2053 18 : returningLists = list_make1(parse->returningList);
2054 30 : if (parse->mergeActionList)
2055 0 : mergeActionLists = list_make1(parse->mergeActionList);
2056 30 : if (parse->commandType == CMD_MERGE)
2057 0 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2058 : }
2059 : }
2060 : else
2061 : {
2062 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2063 87780 : rootRelation = 0; /* there's no separate root rel */
2064 87780 : resultRelations = list_make1_int(parse->resultRelation);
2065 87780 : if (parse->commandType == CMD_UPDATE)
2066 11928 : updateColnosLists = list_make1(root->update_colnos);
2067 87780 : if (parse->withCheckOptions)
2068 926 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2069 87780 : if (parse->returningList)
2070 2432 : returningLists = list_make1(parse->returningList);
2071 87780 : if (parse->mergeActionList)
2072 1512 : mergeActionLists = list_make1(parse->mergeActionList);
2073 87780 : if (parse->commandType == CMD_MERGE)
2074 1512 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2075 : }
2076 :
2077 : /*
2078 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2079 : * will have dealt with fetching non-locked marked rows, else we
2080 : * need to have ModifyTable do that.
2081 : */
2082 90624 : if (parse->rowMarks)
2083 0 : rowMarks = NIL;
2084 : else
2085 90624 : rowMarks = root->rowMarks;
2086 :
2087 : path = (Path *)
2088 90624 : create_modifytable_path(root, final_rel,
2089 : path,
2090 : parse->commandType,
2091 90624 : parse->canSetTag,
2092 90624 : parse->resultRelation,
2093 : rootRelation,
2094 90624 : root->partColsUpdated,
2095 : resultRelations,
2096 : updateColnosLists,
2097 : withCheckOptionLists,
2098 : returningLists,
2099 : rowMarks,
2100 : parse->onConflict,
2101 : mergeActionLists,
2102 : mergeJoinConditions,
2103 : assign_special_exec_param(root));
2104 : }
2105 :
2106 : /* And shove it into final_rel */
2107 535610 : add_path(final_rel, path);
2108 : }
2109 :
2110 : /*
2111 : * Generate partial paths for final_rel, too, if outer query levels might
2112 : * be able to make use of them.
2113 : */
2114 516240 : if (final_rel->consider_parallel && root->query_level > 1 &&
2115 24272 : !limit_needed(parse))
2116 : {
2117 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2118 24194 : foreach(lc, current_rel->partial_pathlist)
2119 : {
2120 108 : Path *partial_path = (Path *) lfirst(lc);
2121 :
2122 108 : add_partial_path(final_rel, partial_path);
2123 : }
2124 : }
2125 :
2126 516240 : extra.limit_needed = limit_needed(parse);
2127 516240 : extra.limit_tuples = limit_tuples;
2128 516240 : extra.count_est = count_est;
2129 516240 : extra.offset_est = offset_est;
2130 :
2131 : /*
2132 : * If there is an FDW that's responsible for all baserels of the query,
2133 : * let it consider adding ForeignPaths.
2134 : */
2135 516240 : if (final_rel->fdwroutine &&
2136 1256 : final_rel->fdwroutine->GetForeignUpperPaths)
2137 1188 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2138 : current_rel, final_rel,
2139 : &extra);
2140 :
2141 : /* Let extensions possibly add some more paths */
2142 516240 : if (create_upper_paths_hook)
2143 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2144 : current_rel, final_rel, &extra);
2145 :
2146 : /* Note: currently, we leave it to callers to do set_cheapest() */
2147 516240 : }
2148 :
2149 : /*
2150 : * Do preprocessing for groupingSets clause and related data. This handles the
2151 : * preliminary steps of expanding the grouping sets, organizing them into lists
2152 : * of rollups, and preparing annotations which will later be filled in with
2153 : * size estimates.
2154 : */
2155 : static grouping_sets_data *
2156 878 : preprocess_grouping_sets(PlannerInfo *root)
2157 : {
2158 878 : Query *parse = root->parse;
2159 : List *sets;
2160 878 : int maxref = 0;
2161 : ListCell *lc_set;
2162 878 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2163 :
2164 878 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
2165 :
2166 878 : gd->any_hashable = false;
2167 878 : gd->unhashable_refs = NULL;
2168 878 : gd->unsortable_refs = NULL;
2169 878 : gd->unsortable_sets = NIL;
2170 :
2171 : /*
2172 : * We don't currently make any attempt to optimize the groupClause when
2173 : * there are grouping sets, so just duplicate it in processed_groupClause.
2174 : */
2175 878 : root->processed_groupClause = parse->groupClause;
2176 :
2177 878 : if (parse->groupClause)
2178 : {
2179 : ListCell *lc;
2180 :
2181 2684 : foreach(lc, parse->groupClause)
2182 : {
2183 1848 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2184 1848 : Index ref = gc->tleSortGroupRef;
2185 :
2186 1848 : if (ref > maxref)
2187 1812 : maxref = ref;
2188 :
2189 1848 : if (!gc->hashable)
2190 30 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2191 :
2192 1848 : if (!OidIsValid(gc->sortop))
2193 42 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2194 : }
2195 : }
2196 :
2197 : /* Allocate workspace array for remapping */
2198 878 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2199 :
2200 : /*
2201 : * If we have any unsortable sets, we must extract them before trying to
2202 : * prepare rollups. Unsortable sets don't go through
2203 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2204 : * here.
2205 : */
2206 878 : if (!bms_is_empty(gd->unsortable_refs))
2207 : {
2208 42 : List *sortable_sets = NIL;
2209 : ListCell *lc;
2210 :
2211 126 : foreach(lc, parse->groupingSets)
2212 : {
2213 90 : List *gset = (List *) lfirst(lc);
2214 :
2215 90 : if (bms_overlap_list(gd->unsortable_refs, gset))
2216 : {
2217 48 : GroupingSetData *gs = makeNode(GroupingSetData);
2218 :
2219 48 : gs->set = gset;
2220 48 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2221 :
2222 : /*
2223 : * We must enforce here that an unsortable set is hashable;
2224 : * later code assumes this. Parse analysis only checks that
2225 : * every individual column is either hashable or sortable.
2226 : *
2227 : * Note that passing this test doesn't guarantee we can
2228 : * generate a plan; there might be other showstoppers.
2229 : */
2230 48 : if (bms_overlap_list(gd->unhashable_refs, gset))
2231 6 : ereport(ERROR,
2232 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2233 : errmsg("could not implement GROUP BY"),
2234 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2235 : }
2236 : else
2237 42 : sortable_sets = lappend(sortable_sets, gset);
2238 : }
2239 :
2240 36 : if (sortable_sets)
2241 30 : sets = extract_rollup_sets(sortable_sets);
2242 : else
2243 6 : sets = NIL;
2244 : }
2245 : else
2246 836 : sets = extract_rollup_sets(parse->groupingSets);
2247 :
2248 2310 : foreach(lc_set, sets)
2249 : {
2250 1438 : List *current_sets = (List *) lfirst(lc_set);
2251 1438 : RollupData *rollup = makeNode(RollupData);
2252 : GroupingSetData *gs;
2253 :
2254 : /*
2255 : * Reorder the current list of grouping sets into correct prefix
2256 : * order. If only one aggregation pass is needed, try to make the
2257 : * list match the ORDER BY clause; if more than one pass is needed, we
2258 : * don't bother with that.
2259 : *
2260 : * Note that this reorders the sets from smallest-member-first to
2261 : * largest-member-first, and applies the GroupingSetData annotations,
2262 : * though the data will be filled in later.
2263 : */
2264 1438 : current_sets = reorder_grouping_sets(current_sets,
2265 1438 : (list_length(sets) == 1
2266 : ? parse->sortClause
2267 : : NIL));
2268 :
2269 : /*
2270 : * Get the initial (and therefore largest) grouping set.
2271 : */
2272 1438 : gs = linitial_node(GroupingSetData, current_sets);
2273 :
2274 : /*
2275 : * Order the groupClause appropriately. If the first grouping set is
2276 : * empty, then the groupClause must also be empty; otherwise we have
2277 : * to force the groupClause to match that grouping set's order.
2278 : *
2279 : * (The first grouping set can be empty even though parse->groupClause
2280 : * is not empty only if all non-empty grouping sets are unsortable.
2281 : * The groupClauses for hashed grouping sets are built later on.)
2282 : */
2283 1438 : if (gs->set)
2284 1396 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2285 : else
2286 42 : rollup->groupClause = NIL;
2287 :
2288 : /*
2289 : * Is it hashable? We pretend empty sets are hashable even though we
2290 : * actually force them not to be hashed later. But don't bother if
2291 : * there's nothing but empty sets (since in that case we can't hash
2292 : * anything).
2293 : */
2294 1438 : if (gs->set &&
2295 1396 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2296 : {
2297 1372 : rollup->hashable = true;
2298 1372 : gd->any_hashable = true;
2299 : }
2300 :
2301 : /*
2302 : * Now that we've pinned down an order for the groupClause for this
2303 : * list of grouping sets, we need to remap the entries in the grouping
2304 : * sets from sortgrouprefs to plain indices (0-based) into the
2305 : * groupClause for this collection of grouping sets. We keep the
2306 : * original form for later use, though.
2307 : */
2308 1438 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2309 : current_sets,
2310 : gd->tleref_to_colnum_map);
2311 1438 : rollup->gsets_data = current_sets;
2312 :
2313 1438 : gd->rollups = lappend(gd->rollups, rollup);
2314 : }
2315 :
2316 872 : if (gd->unsortable_sets)
2317 : {
2318 : /*
2319 : * We have not yet pinned down a groupclause for this, but we will
2320 : * need index-based lists for estimation purposes. Construct
2321 : * hash_sets_idx based on the entire original groupclause for now.
2322 : */
2323 36 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2324 : gd->unsortable_sets,
2325 : gd->tleref_to_colnum_map);
2326 36 : gd->any_hashable = true;
2327 : }
2328 :
2329 872 : return gd;
2330 : }
2331 :
2332 : /*
2333 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2334 : * (without annotation) mapped to indexes into the given groupclause.
2335 : */
2336 : static List *
2337 4200 : remap_to_groupclause_idx(List *groupClause,
2338 : List *gsets,
2339 : int *tleref_to_colnum_map)
2340 : {
2341 4200 : int ref = 0;
2342 4200 : List *result = NIL;
2343 : ListCell *lc;
2344 :
2345 10264 : foreach(lc, groupClause)
2346 : {
2347 6064 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2348 :
2349 6064 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2350 : }
2351 :
2352 9702 : foreach(lc, gsets)
2353 : {
2354 5502 : List *set = NIL;
2355 : ListCell *lc2;
2356 5502 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2357 :
2358 12404 : foreach(lc2, gs->set)
2359 : {
2360 6902 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2361 : }
2362 :
2363 5502 : result = lappend(result, set);
2364 : }
2365 :
2366 4200 : return result;
2367 : }
2368 :
2369 :
2370 : /*
2371 : * preprocess_rowmarks - set up PlanRowMarks if needed
2372 : */
2373 : static void
2374 520184 : preprocess_rowmarks(PlannerInfo *root)
2375 : {
2376 520184 : Query *parse = root->parse;
2377 : Bitmapset *rels;
2378 : List *prowmarks;
2379 : ListCell *l;
2380 : int i;
2381 :
2382 520184 : if (parse->rowMarks)
2383 : {
2384 : /*
2385 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2386 : * grouping, since grouping renders a reference to individual tuple
2387 : * CTIDs invalid. This is also checked at parse time, but that's
2388 : * insufficient because of rule substitution, query pullup, etc.
2389 : */
2390 7758 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2391 : parse->rowMarks)->strength);
2392 : }
2393 : else
2394 : {
2395 : /*
2396 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2397 : * UPDATE/SHARE.
2398 : */
2399 512426 : if (parse->commandType != CMD_UPDATE &&
2400 498506 : parse->commandType != CMD_DELETE &&
2401 494188 : parse->commandType != CMD_MERGE)
2402 492436 : return;
2403 : }
2404 :
2405 : /*
2406 : * We need to have rowmarks for all base relations except the target. We
2407 : * make a bitmapset of all base rels and then remove the items we don't
2408 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2409 : */
2410 27748 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2411 27748 : if (parse->resultRelation)
2412 19990 : rels = bms_del_member(rels, parse->resultRelation);
2413 :
2414 : /*
2415 : * Convert RowMarkClauses to PlanRowMark representation.
2416 : */
2417 27748 : prowmarks = NIL;
2418 35736 : foreach(l, parse->rowMarks)
2419 : {
2420 7988 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2421 7988 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2422 : PlanRowMark *newrc;
2423 :
2424 : /*
2425 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2426 : * applied to an update/delete target rel. If that ever becomes
2427 : * possible, we should drop the target from the PlanRowMark list.
2428 : */
2429 : Assert(rc->rti != parse->resultRelation);
2430 :
2431 : /*
2432 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2433 : * can't support true locking. Subqueries that got flattened into the
2434 : * main query should be ignored completely. Any that didn't will get
2435 : * ROW_MARK_COPY items in the next loop.
2436 : */
2437 7988 : if (rte->rtekind != RTE_RELATION)
2438 60 : continue;
2439 :
2440 7928 : rels = bms_del_member(rels, rc->rti);
2441 :
2442 7928 : newrc = makeNode(PlanRowMark);
2443 7928 : newrc->rti = newrc->prti = rc->rti;
2444 7928 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2445 7928 : newrc->markType = select_rowmark_type(rte, rc->strength);
2446 7928 : newrc->allMarkTypes = (1 << newrc->markType);
2447 7928 : newrc->strength = rc->strength;
2448 7928 : newrc->waitPolicy = rc->waitPolicy;
2449 7928 : newrc->isParent = false;
2450 :
2451 7928 : prowmarks = lappend(prowmarks, newrc);
2452 : }
2453 :
2454 : /*
2455 : * Now, add rowmarks for any non-target, non-locked base relations.
2456 : */
2457 27748 : i = 0;
2458 66090 : foreach(l, parse->rtable)
2459 : {
2460 38342 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2461 : PlanRowMark *newrc;
2462 :
2463 38342 : i++;
2464 38342 : if (!bms_is_member(i, rels))
2465 34700 : continue;
2466 :
2467 3642 : newrc = makeNode(PlanRowMark);
2468 3642 : newrc->rti = newrc->prti = i;
2469 3642 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2470 3642 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2471 3642 : newrc->allMarkTypes = (1 << newrc->markType);
2472 3642 : newrc->strength = LCS_NONE;
2473 3642 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2474 3642 : newrc->isParent = false;
2475 :
2476 3642 : prowmarks = lappend(prowmarks, newrc);
2477 : }
2478 :
2479 27748 : root->rowMarks = prowmarks;
2480 : }
2481 :
2482 : /*
2483 : * Select RowMarkType to use for a given table
2484 : */
2485 : RowMarkType
2486 13902 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2487 : {
2488 13902 : if (rte->rtekind != RTE_RELATION)
2489 : {
2490 : /* If it's not a table at all, use ROW_MARK_COPY */
2491 1392 : return ROW_MARK_COPY;
2492 : }
2493 12510 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2494 : {
2495 : /* Let the FDW select the rowmark type, if it wants to */
2496 212 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2497 :
2498 212 : if (fdwroutine->GetForeignRowMarkType != NULL)
2499 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2500 : /* Otherwise, use ROW_MARK_COPY by default */
2501 212 : return ROW_MARK_COPY;
2502 : }
2503 : else
2504 : {
2505 : /* Regular table, apply the appropriate lock type */
2506 12298 : switch (strength)
2507 : {
2508 2480 : case LCS_NONE:
2509 :
2510 : /*
2511 : * We don't need a tuple lock, only the ability to re-fetch
2512 : * the row.
2513 : */
2514 2480 : return ROW_MARK_REFERENCE;
2515 : break;
2516 7938 : case LCS_FORKEYSHARE:
2517 7938 : return ROW_MARK_KEYSHARE;
2518 : break;
2519 300 : case LCS_FORSHARE:
2520 300 : return ROW_MARK_SHARE;
2521 : break;
2522 72 : case LCS_FORNOKEYUPDATE:
2523 72 : return ROW_MARK_NOKEYEXCLUSIVE;
2524 : break;
2525 1508 : case LCS_FORUPDATE:
2526 1508 : return ROW_MARK_EXCLUSIVE;
2527 : break;
2528 : }
2529 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2530 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2531 : }
2532 : }
2533 :
2534 : /*
2535 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2536 : *
2537 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2538 : * results back in *count_est and *offset_est. These variables are set to
2539 : * 0 if the corresponding clause is not present, and -1 if it's present
2540 : * but we couldn't estimate the value for it. (The "0" convention is OK
2541 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2542 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2543 : * usual practice of never estimating less than one row.) These values will
2544 : * be passed to create_limit_path, which see if you change this code.
2545 : *
2546 : * The return value is the suitably adjusted tuple_fraction to use for
2547 : * planning the query. This adjustment is not overridable, since it reflects
2548 : * plan actions that grouping_planner() will certainly take, not assumptions
2549 : * about context.
2550 : */
2551 : static double
2552 5006 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2553 : int64 *offset_est, int64 *count_est)
2554 : {
2555 5006 : Query *parse = root->parse;
2556 : Node *est;
2557 : double limit_fraction;
2558 :
2559 : /* Should not be called unless LIMIT or OFFSET */
2560 : Assert(parse->limitCount || parse->limitOffset);
2561 :
2562 : /*
2563 : * Try to obtain the clause values. We use estimate_expression_value
2564 : * primarily because it can sometimes do something useful with Params.
2565 : */
2566 5006 : if (parse->limitCount)
2567 : {
2568 4496 : est = estimate_expression_value(root, parse->limitCount);
2569 4496 : if (est && IsA(est, Const))
2570 : {
2571 4490 : if (((Const *) est)->constisnull)
2572 : {
2573 : /* NULL indicates LIMIT ALL, ie, no limit */
2574 0 : *count_est = 0; /* treat as not present */
2575 : }
2576 : else
2577 : {
2578 4490 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2579 4490 : if (*count_est <= 0)
2580 150 : *count_est = 1; /* force to at least 1 */
2581 : }
2582 : }
2583 : else
2584 6 : *count_est = -1; /* can't estimate */
2585 : }
2586 : else
2587 510 : *count_est = 0; /* not present */
2588 :
2589 5006 : if (parse->limitOffset)
2590 : {
2591 882 : est = estimate_expression_value(root, parse->limitOffset);
2592 882 : if (est && IsA(est, Const))
2593 : {
2594 858 : if (((Const *) est)->constisnull)
2595 : {
2596 : /* Treat NULL as no offset; the executor will too */
2597 0 : *offset_est = 0; /* treat as not present */
2598 : }
2599 : else
2600 : {
2601 858 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2602 858 : if (*offset_est < 0)
2603 0 : *offset_est = 0; /* treat as not present */
2604 : }
2605 : }
2606 : else
2607 24 : *offset_est = -1; /* can't estimate */
2608 : }
2609 : else
2610 4124 : *offset_est = 0; /* not present */
2611 :
2612 5006 : if (*count_est != 0)
2613 : {
2614 : /*
2615 : * A LIMIT clause limits the absolute number of tuples returned.
2616 : * However, if it's not a constant LIMIT then we have to guess; for
2617 : * lack of a better idea, assume 10% of the plan's result is wanted.
2618 : */
2619 4496 : if (*count_est < 0 || *offset_est < 0)
2620 : {
2621 : /* LIMIT or OFFSET is an expression ... punt ... */
2622 24 : limit_fraction = 0.10;
2623 : }
2624 : else
2625 : {
2626 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2627 4472 : limit_fraction = (double) *count_est + (double) *offset_est;
2628 : }
2629 :
2630 : /*
2631 : * If we have absolute limits from both caller and LIMIT, use the
2632 : * smaller value; likewise if they are both fractional. If one is
2633 : * fractional and the other absolute, we can't easily determine which
2634 : * is smaller, but we use the heuristic that the absolute will usually
2635 : * be smaller.
2636 : */
2637 4496 : if (tuple_fraction >= 1.0)
2638 : {
2639 6 : if (limit_fraction >= 1.0)
2640 : {
2641 : /* both absolute */
2642 6 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2643 : }
2644 : else
2645 : {
2646 : /* caller absolute, limit fractional; use caller's value */
2647 : }
2648 : }
2649 4490 : else if (tuple_fraction > 0.0)
2650 : {
2651 148 : if (limit_fraction >= 1.0)
2652 : {
2653 : /* caller fractional, limit absolute; use limit */
2654 148 : tuple_fraction = limit_fraction;
2655 : }
2656 : else
2657 : {
2658 : /* both fractional */
2659 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2660 : }
2661 : }
2662 : else
2663 : {
2664 : /* no info from caller, just use limit */
2665 4342 : tuple_fraction = limit_fraction;
2666 : }
2667 : }
2668 510 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2669 : {
2670 : /*
2671 : * We have an OFFSET but no LIMIT. This acts entirely differently
2672 : * from the LIMIT case: here, we need to increase rather than decrease
2673 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2674 : * tuples to be fetched instead of fewer. This only matters if we got
2675 : * a tuple_fraction > 0, however.
2676 : *
2677 : * As above, use 10% if OFFSET is present but unestimatable.
2678 : */
2679 12 : if (*offset_est < 0)
2680 0 : limit_fraction = 0.10;
2681 : else
2682 12 : limit_fraction = (double) *offset_est;
2683 :
2684 : /*
2685 : * If we have absolute counts from both caller and OFFSET, add them
2686 : * together; likewise if they are both fractional. If one is
2687 : * fractional and the other absolute, we want to take the larger, and
2688 : * we heuristically assume that's the fractional one.
2689 : */
2690 12 : if (tuple_fraction >= 1.0)
2691 : {
2692 0 : if (limit_fraction >= 1.0)
2693 : {
2694 : /* both absolute, so add them together */
2695 0 : tuple_fraction += limit_fraction;
2696 : }
2697 : else
2698 : {
2699 : /* caller absolute, limit fractional; use limit */
2700 0 : tuple_fraction = limit_fraction;
2701 : }
2702 : }
2703 : else
2704 : {
2705 12 : if (limit_fraction >= 1.0)
2706 : {
2707 : /* caller fractional, limit absolute; use caller's value */
2708 : }
2709 : else
2710 : {
2711 : /* both fractional, so add them together */
2712 0 : tuple_fraction += limit_fraction;
2713 0 : if (tuple_fraction >= 1.0)
2714 0 : tuple_fraction = 0.0; /* assume fetch all */
2715 : }
2716 : }
2717 : }
2718 :
2719 5006 : return tuple_fraction;
2720 : }
2721 :
2722 : /*
2723 : * limit_needed - do we actually need a Limit plan node?
2724 : *
2725 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2726 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2727 : * locution for an optimization fence. (Because other places in the planner
2728 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2729 : * an optimization fence --- we're just suppressing unnecessary run-time
2730 : * overhead.)
2731 : *
2732 : * This might look like it could be merged into preprocess_limit, but there's
2733 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2734 : * in preprocess_limit it's good enough to consider estimated values.
2735 : */
2736 : bool
2737 1089090 : limit_needed(Query *parse)
2738 : {
2739 : Node *node;
2740 :
2741 1089090 : node = parse->limitCount;
2742 1089090 : if (node)
2743 : {
2744 10762 : if (IsA(node, Const))
2745 : {
2746 : /* NULL indicates LIMIT ALL, ie, no limit */
2747 10526 : if (!((Const *) node)->constisnull)
2748 10526 : return true; /* LIMIT with a constant value */
2749 : }
2750 : else
2751 236 : return true; /* non-constant LIMIT */
2752 : }
2753 :
2754 1078328 : node = parse->limitOffset;
2755 1078328 : if (node)
2756 : {
2757 1498 : if (IsA(node, Const))
2758 : {
2759 : /* Treat NULL as no offset; the executor would too */
2760 1190 : if (!((Const *) node)->constisnull)
2761 : {
2762 1190 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2763 :
2764 1190 : if (offset != 0)
2765 110 : return true; /* OFFSET with a nonzero value */
2766 : }
2767 : }
2768 : else
2769 308 : return true; /* non-constant OFFSET */
2770 : }
2771 :
2772 1077910 : return false; /* don't need a Limit plan node */
2773 : }
2774 :
2775 : /*
2776 : * preprocess_groupclause - do preparatory work on GROUP BY clause
2777 : *
2778 : * The idea here is to adjust the ordering of the GROUP BY elements
2779 : * (which in itself is semantically insignificant) to match ORDER BY,
2780 : * thereby allowing a single sort operation to both implement the ORDER BY
2781 : * requirement and set up for a Unique step that implements GROUP BY.
2782 : * We also consider partial match between GROUP BY and ORDER BY elements,
2783 : * which could allow to implement ORDER BY using the incremental sort.
2784 : *
2785 : * We also consider other orderings of the GROUP BY elements, which could
2786 : * match the sort ordering of other possible plans (eg an indexscan) and
2787 : * thereby reduce cost. This is implemented during the generation of grouping
2788 : * paths. See get_useful_group_keys_orderings() for details.
2789 : *
2790 : * Note: we need no comparable processing of the distinctClause because
2791 : * the parser already enforced that that matches ORDER BY.
2792 : *
2793 : * Note: we return a fresh List, but its elements are the same
2794 : * SortGroupClauses appearing in parse->groupClause. This is important
2795 : * because later processing may modify the processed_groupClause list.
2796 : *
2797 : * For grouping sets, the order of items is instead forced to agree with that
2798 : * of the grouping set (and items not in the grouping set are skipped). The
2799 : * work of sorting the order of grouping set elements to match the ORDER BY if
2800 : * possible is done elsewhere.
2801 : */
2802 : static List *
2803 7752 : preprocess_groupclause(PlannerInfo *root, List *force)
2804 : {
2805 7752 : Query *parse = root->parse;
2806 7752 : List *new_groupclause = NIL;
2807 : ListCell *sl;
2808 : ListCell *gl;
2809 :
2810 : /* For grouping sets, we need to force the ordering */
2811 7752 : if (force)
2812 : {
2813 10108 : foreach(sl, force)
2814 : {
2815 5986 : Index ref = lfirst_int(sl);
2816 5986 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2817 :
2818 5986 : new_groupclause = lappend(new_groupclause, cl);
2819 : }
2820 :
2821 4122 : return new_groupclause;
2822 : }
2823 :
2824 : /* If no ORDER BY, nothing useful to do here */
2825 3630 : if (parse->sortClause == NIL)
2826 2052 : return list_copy(parse->groupClause);
2827 :
2828 : /*
2829 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2830 : * items, but only as far as we can make a matching prefix.
2831 : *
2832 : * This code assumes that the sortClause contains no duplicate items.
2833 : */
2834 3062 : foreach(sl, parse->sortClause)
2835 : {
2836 2130 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2837 :
2838 3226 : foreach(gl, parse->groupClause)
2839 : {
2840 2580 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2841 :
2842 2580 : if (equal(gc, sc))
2843 : {
2844 1484 : new_groupclause = lappend(new_groupclause, gc);
2845 1484 : break;
2846 : }
2847 : }
2848 2130 : if (gl == NULL)
2849 646 : break; /* no match, so stop scanning */
2850 : }
2851 :
2852 :
2853 : /* If no match at all, no point in reordering GROUP BY */
2854 1578 : if (new_groupclause == NIL)
2855 298 : return list_copy(parse->groupClause);
2856 :
2857 : /*
2858 : * Add any remaining GROUP BY items to the new list. We don't require a
2859 : * complete match, because even partial match allows ORDER BY to be
2860 : * implemented using incremental sort. Also, give up if there are any
2861 : * non-sortable GROUP BY items, since then there's no hope anyway.
2862 : */
2863 2930 : foreach(gl, parse->groupClause)
2864 : {
2865 1650 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2866 :
2867 1650 : if (list_member_ptr(new_groupclause, gc))
2868 1484 : continue; /* it matched an ORDER BY item */
2869 166 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2870 0 : return list_copy(parse->groupClause);
2871 166 : new_groupclause = lappend(new_groupclause, gc);
2872 : }
2873 :
2874 : /* Success --- install the rearranged GROUP BY list */
2875 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2876 1280 : return new_groupclause;
2877 : }
2878 :
2879 : /*
2880 : * Extract lists of grouping sets that can be implemented using a single
2881 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2882 : *
2883 : * Input must be sorted with smallest sets first. Result has each sublist
2884 : * sorted with smallest sets first.
2885 : *
2886 : * We want to produce the absolute minimum possible number of lists here to
2887 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2888 : * of finding the minimal partition of a partially-ordered set into chains
2889 : * (which is what we need, taking the list of grouping sets as a poset ordered
2890 : * by set inclusion) can be mapped to the problem of finding the maximum
2891 : * cardinality matching on a bipartite graph, which is solvable in polynomial
2892 : * time with a worst case of no worse than O(n^2.5) and usually much
2893 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
2894 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
2895 : * half a second on my modest system even with optimization off and assertions
2896 : * on.)
2897 : */
2898 : static List *
2899 866 : extract_rollup_sets(List *groupingSets)
2900 : {
2901 866 : int num_sets_raw = list_length(groupingSets);
2902 866 : int num_empty = 0;
2903 866 : int num_sets = 0; /* distinct sets */
2904 866 : int num_chains = 0;
2905 866 : List *result = NIL;
2906 : List **results;
2907 : List **orig_sets;
2908 : Bitmapset **set_masks;
2909 : int *chains;
2910 : short **adjacency;
2911 : short *adjacency_buf;
2912 : BipartiteMatchState *state;
2913 : int i;
2914 : int j;
2915 : int j_size;
2916 866 : ListCell *lc1 = list_head(groupingSets);
2917 : ListCell *lc;
2918 :
2919 : /*
2920 : * Start by stripping out empty sets. The algorithm doesn't require this,
2921 : * but the planner currently needs all empty sets to be returned in the
2922 : * first list, so we strip them here and add them back after.
2923 : */
2924 1476 : while (lc1 && lfirst(lc1) == NIL)
2925 : {
2926 610 : ++num_empty;
2927 610 : lc1 = lnext(groupingSets, lc1);
2928 : }
2929 :
2930 : /* bail out now if it turns out that all we had were empty sets. */
2931 866 : if (!lc1)
2932 42 : return list_make1(groupingSets);
2933 :
2934 : /*----------
2935 : * We don't strictly need to remove duplicate sets here, but if we don't,
2936 : * they tend to become scattered through the result, which is a bit
2937 : * confusing (and irritating if we ever decide to optimize them out).
2938 : * So we remove them here and add them back after.
2939 : *
2940 : * For each non-duplicate set, we fill in the following:
2941 : *
2942 : * orig_sets[i] = list of the original set lists
2943 : * set_masks[i] = bitmapset for testing inclusion
2944 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2945 : *
2946 : * chains[i] will be the result group this set is assigned to.
2947 : *
2948 : * We index all of these from 1 rather than 0 because it is convenient
2949 : * to leave 0 free for the NIL node in the graph algorithm.
2950 : *----------
2951 : */
2952 824 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
2953 824 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
2954 824 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
2955 824 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
2956 :
2957 824 : j_size = 0;
2958 824 : j = 0;
2959 824 : i = 1;
2960 :
2961 2948 : for_each_cell(lc, groupingSets, lc1)
2962 : {
2963 2124 : List *candidate = (List *) lfirst(lc);
2964 2124 : Bitmapset *candidate_set = NULL;
2965 : ListCell *lc2;
2966 2124 : int dup_of = 0;
2967 :
2968 5142 : foreach(lc2, candidate)
2969 : {
2970 3018 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
2971 : }
2972 :
2973 : /* we can only be a dup if we're the same length as a previous set */
2974 2124 : if (j_size == list_length(candidate))
2975 : {
2976 : int k;
2977 :
2978 1904 : for (k = j; k < i; ++k)
2979 : {
2980 1236 : if (bms_equal(set_masks[k], candidate_set))
2981 : {
2982 158 : dup_of = k;
2983 158 : break;
2984 : }
2985 : }
2986 : }
2987 1298 : else if (j_size < list_length(candidate))
2988 : {
2989 1298 : j_size = list_length(candidate);
2990 1298 : j = i;
2991 : }
2992 :
2993 2124 : if (dup_of > 0)
2994 : {
2995 158 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
2996 158 : bms_free(candidate_set);
2997 : }
2998 : else
2999 : {
3000 : int k;
3001 1966 : int n_adj = 0;
3002 :
3003 1966 : orig_sets[i] = list_make1(candidate);
3004 1966 : set_masks[i] = candidate_set;
3005 :
3006 : /* fill in adjacency list; no need to compare equal-size sets */
3007 :
3008 3238 : for (k = j - 1; k > 0; --k)
3009 : {
3010 1272 : if (bms_is_subset(set_masks[k], candidate_set))
3011 1110 : adjacency_buf[++n_adj] = k;
3012 : }
3013 :
3014 1966 : if (n_adj > 0)
3015 : {
3016 598 : adjacency_buf[0] = n_adj;
3017 598 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3018 598 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3019 : }
3020 : else
3021 1368 : adjacency[i] = NULL;
3022 :
3023 1966 : ++i;
3024 : }
3025 : }
3026 :
3027 824 : num_sets = i - 1;
3028 :
3029 : /*
3030 : * Apply the graph matching algorithm to do the work.
3031 : */
3032 824 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3033 :
3034 : /*
3035 : * Now, the state->pair* fields have the info we need to assign sets to
3036 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3037 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3038 : * it in one pass)
3039 : */
3040 824 : chains = palloc0((num_sets + 1) * sizeof(int));
3041 :
3042 2790 : for (i = 1; i <= num_sets; ++i)
3043 : {
3044 1966 : int u = state->pair_vu[i];
3045 1966 : int v = state->pair_uv[i];
3046 :
3047 1966 : if (u > 0 && u < i)
3048 0 : chains[i] = chains[u];
3049 1966 : else if (v > 0 && v < i)
3050 570 : chains[i] = chains[v];
3051 : else
3052 1396 : chains[i] = ++num_chains;
3053 : }
3054 :
3055 : /* build result lists. */
3056 824 : results = palloc0((num_chains + 1) * sizeof(List *));
3057 :
3058 2790 : for (i = 1; i <= num_sets; ++i)
3059 : {
3060 1966 : int c = chains[i];
3061 :
3062 : Assert(c > 0);
3063 :
3064 1966 : results[c] = list_concat(results[c], orig_sets[i]);
3065 : }
3066 :
3067 : /* push any empty sets back on the first list. */
3068 1344 : while (num_empty-- > 0)
3069 520 : results[1] = lcons(NIL, results[1]);
3070 :
3071 : /* make result list */
3072 2220 : for (i = 1; i <= num_chains; ++i)
3073 1396 : result = lappend(result, results[i]);
3074 :
3075 : /*
3076 : * Free all the things.
3077 : *
3078 : * (This is over-fussy for small sets but for large sets we could have
3079 : * tied up a nontrivial amount of memory.)
3080 : */
3081 824 : BipartiteMatchFree(state);
3082 824 : pfree(results);
3083 824 : pfree(chains);
3084 2790 : for (i = 1; i <= num_sets; ++i)
3085 1966 : if (adjacency[i])
3086 598 : pfree(adjacency[i]);
3087 824 : pfree(adjacency);
3088 824 : pfree(adjacency_buf);
3089 824 : pfree(orig_sets);
3090 2790 : for (i = 1; i <= num_sets; ++i)
3091 1966 : bms_free(set_masks[i]);
3092 824 : pfree(set_masks);
3093 :
3094 824 : return result;
3095 : }
3096 :
3097 : /*
3098 : * Reorder the elements of a list of grouping sets such that they have correct
3099 : * prefix relationships. Also inserts the GroupingSetData annotations.
3100 : *
3101 : * The input must be ordered with smallest sets first; the result is returned
3102 : * with largest sets first. Note that the result shares no list substructure
3103 : * with the input, so it's safe for the caller to modify it later.
3104 : *
3105 : * If we're passed in a sortclause, we follow its order of columns to the
3106 : * extent possible, to minimize the chance that we add unnecessary sorts.
3107 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3108 : * gets implemented in one pass.)
3109 : */
3110 : static List *
3111 1438 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3112 : {
3113 : ListCell *lc;
3114 1438 : List *previous = NIL;
3115 1438 : List *result = NIL;
3116 :
3117 4172 : foreach(lc, groupingSets)
3118 : {
3119 2734 : List *candidate = (List *) lfirst(lc);
3120 2734 : List *new_elems = list_difference_int(candidate, previous);
3121 2734 : GroupingSetData *gs = makeNode(GroupingSetData);
3122 :
3123 2898 : while (list_length(sortclause) > list_length(previous) &&
3124 : new_elems != NIL)
3125 : {
3126 272 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3127 272 : int ref = sc->tleSortGroupRef;
3128 :
3129 272 : if (list_member_int(new_elems, ref))
3130 : {
3131 164 : previous = lappend_int(previous, ref);
3132 164 : new_elems = list_delete_int(new_elems, ref);
3133 : }
3134 : else
3135 : {
3136 : /* diverged from the sortclause; give up on it */
3137 108 : sortclause = NIL;
3138 108 : break;
3139 : }
3140 : }
3141 :
3142 2734 : previous = list_concat(previous, new_elems);
3143 :
3144 2734 : gs->set = list_copy(previous);
3145 2734 : result = lcons(gs, result);
3146 : }
3147 :
3148 1438 : list_free(previous);
3149 :
3150 1438 : return result;
3151 : }
3152 :
3153 : /*
3154 : * has_volatile_pathkey
3155 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3156 : * containing a volatile function. Otherwise returns false.
3157 : */
3158 : static bool
3159 2792 : has_volatile_pathkey(List *keys)
3160 : {
3161 : ListCell *lc;
3162 :
3163 5728 : foreach(lc, keys)
3164 : {
3165 2954 : PathKey *pathkey = lfirst_node(PathKey, lc);
3166 :
3167 2954 : if (pathkey->pk_eclass->ec_has_volatile)
3168 18 : return true;
3169 : }
3170 :
3171 2774 : return false;
3172 : }
3173 :
3174 : /*
3175 : * adjust_group_pathkeys_for_groupagg
3176 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3177 : * pre-ordered input for ordered aggregates.
3178 : *
3179 : * We define "best" as the pathkeys that suit the largest number of
3180 : * aggregate functions. We find these by looking at the first ORDER BY /
3181 : * DISTINCT aggregate and take the pathkeys for that before searching for
3182 : * other aggregates that require the same or a more strict variation of the
3183 : * same pathkeys. We then repeat that process for any remaining aggregates
3184 : * with different pathkeys and if we find another set of pathkeys that suits a
3185 : * larger number of aggregates then we select those pathkeys instead.
3186 : *
3187 : * When the best pathkeys are found we also mark each Aggref that can use
3188 : * those pathkeys as aggpresorted = true.
3189 : *
3190 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3191 : * volatile functions, we never make use of these pathkeys. We want to ensure
3192 : * that sorts using volatile functions are done independently in each Aggref
3193 : * rather than once at the query level. If we were to allow this then Aggrefs
3194 : * with compatible sort orders would all transition their rows in the same
3195 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3196 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3197 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3198 : * left to perform their sorts individually. To avoid this inconsistent
3199 : * behavior which could make Aggref results depend on what other Aggrefs the
3200 : * query contains, we always force Aggrefs with volatile functions to perform
3201 : * their own sorts.
3202 : */
3203 : static void
3204 2396 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3205 : {
3206 2396 : List *grouppathkeys = root->group_pathkeys;
3207 : List *bestpathkeys;
3208 : Bitmapset *bestaggs;
3209 : Bitmapset *unprocessed_aggs;
3210 : ListCell *lc;
3211 : int i;
3212 :
3213 : /* Shouldn't be here if there are grouping sets */
3214 : Assert(root->parse->groupingSets == NIL);
3215 : /* Shouldn't be here unless there are some ordered aggregates */
3216 : Assert(root->numOrderedAggs > 0);
3217 :
3218 : /* Do nothing if disabled */
3219 2396 : if (!enable_presorted_aggregate)
3220 6 : return;
3221 :
3222 : /*
3223 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3224 : * the indexes of all AggInfos to be processed below.
3225 : */
3226 2390 : unprocessed_aggs = NULL;
3227 5464 : foreach(lc, root->agginfos)
3228 : {
3229 3074 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3230 3074 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3231 :
3232 3074 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3233 264 : continue;
3234 :
3235 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3236 2810 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3237 300 : continue;
3238 :
3239 : /* Additional safety checks are needed if there's a FILTER clause */
3240 2510 : if (aggref->aggfilter != NULL)
3241 : {
3242 : ListCell *lc2;
3243 54 : bool allow_presort = true;
3244 :
3245 : /*
3246 : * When the Aggref has a FILTER clause, it's possible that the
3247 : * filter removes rows that cannot be sorted because the
3248 : * expression to sort by results in an error during its
3249 : * evaluation. This is a problem for presorting as that happens
3250 : * before the FILTER, whereas without presorting, the Aggregate
3251 : * node will apply the FILTER *before* sorting. So that we never
3252 : * try to sort anything that might error, here we aim to skip over
3253 : * any Aggrefs with arguments with expressions which, when
3254 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3255 : * may be more cases that should be allowed, but more thought
3256 : * needs to be given. Err on the side of caution.
3257 : */
3258 102 : foreach(lc2, aggref->args)
3259 : {
3260 72 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3261 72 : Expr *expr = tle->expr;
3262 :
3263 84 : while (IsA(expr, RelabelType))
3264 12 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3265 :
3266 : /* Common case, Vars and Consts are ok */
3267 72 : if (IsA(expr, Var) || IsA(expr, Const))
3268 48 : continue;
3269 :
3270 : /* Unsupported. Don't try to presort for this Aggref */
3271 24 : allow_presort = false;
3272 24 : break;
3273 : }
3274 :
3275 : /* Skip unsupported Aggrefs */
3276 54 : if (!allow_presort)
3277 24 : continue;
3278 : }
3279 :
3280 2486 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3281 : foreach_current_index(lc));
3282 : }
3283 :
3284 : /*
3285 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3286 : * for the given set of aggregates.
3287 : *
3288 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3289 : * this during the first loop using the pathkeys for the very first
3290 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3291 : * a more strict set of compatible pathkeys. Once the outer loop is
3292 : * complete, we mark off all the aggregates with compatible pathkeys then
3293 : * remove those from the unprocessed_aggs and repeat the process to try to
3294 : * find another set of pathkeys that are suitable for a larger number of
3295 : * aggregates. The outer loop will stop when there are not enough
3296 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3297 : * to suit a larger number of aggregates.
3298 : */
3299 2390 : bestpathkeys = NIL;
3300 2390 : bestaggs = NULL;
3301 4714 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3302 : {
3303 2324 : Bitmapset *aggindexes = NULL;
3304 2324 : List *currpathkeys = NIL;
3305 :
3306 2324 : i = -1;
3307 5116 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3308 : {
3309 2792 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3310 2792 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3311 : List *sortlist;
3312 : List *pathkeys;
3313 :
3314 2792 : if (aggref->aggdistinct != NIL)
3315 718 : sortlist = aggref->aggdistinct;
3316 : else
3317 2074 : sortlist = aggref->aggorder;
3318 :
3319 2792 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3320 : aggref->args);
3321 :
3322 : /*
3323 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3324 : * or DISTINCT clause.
3325 : */
3326 2792 : if (has_volatile_pathkey(pathkeys))
3327 : {
3328 18 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3329 18 : continue;
3330 : }
3331 :
3332 : /*
3333 : * When not set yet, take the pathkeys from the first unprocessed
3334 : * aggregate.
3335 : */
3336 2774 : if (currpathkeys == NIL)
3337 : {
3338 2318 : currpathkeys = pathkeys;
3339 :
3340 : /* include the GROUP BY pathkeys, if they exist */
3341 2318 : if (grouppathkeys != NIL)
3342 276 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3343 : currpathkeys);
3344 :
3345 : /* record that we found pathkeys for this aggregate */
3346 2318 : aggindexes = bms_add_member(aggindexes, i);
3347 : }
3348 : else
3349 : {
3350 : /* now look for a stronger set of matching pathkeys */
3351 :
3352 : /* include the GROUP BY pathkeys, if they exist */
3353 456 : if (grouppathkeys != NIL)
3354 288 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3355 : pathkeys);
3356 :
3357 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3358 456 : switch (compare_pathkeys(currpathkeys, pathkeys))
3359 : {
3360 12 : case PATHKEYS_BETTER2:
3361 : /* 'pathkeys' are stronger, use these ones instead */
3362 12 : currpathkeys = pathkeys;
3363 : /* FALLTHROUGH */
3364 :
3365 66 : case PATHKEYS_BETTER1:
3366 : /* 'pathkeys' are less strict */
3367 : /* FALLTHROUGH */
3368 :
3369 : case PATHKEYS_EQUAL:
3370 : /* mark this aggregate as covered by 'currpathkeys' */
3371 66 : aggindexes = bms_add_member(aggindexes, i);
3372 66 : break;
3373 :
3374 390 : case PATHKEYS_DIFFERENT:
3375 390 : break;
3376 : }
3377 : }
3378 : }
3379 :
3380 : /* remove the aggregates that we've just processed */
3381 2324 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3382 :
3383 : /*
3384 : * If this pass included more aggregates than the previous best then
3385 : * use these ones as the best set.
3386 : */
3387 2324 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3388 : {
3389 2216 : bestaggs = aggindexes;
3390 2216 : bestpathkeys = currpathkeys;
3391 : }
3392 : }
3393 :
3394 : /*
3395 : * If we found any ordered aggregates, update root->group_pathkeys to add
3396 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3397 : * the original GROUP BY pathkeys already.
3398 : */
3399 2390 : if (bestpathkeys != NIL)
3400 2156 : root->group_pathkeys = bestpathkeys;
3401 :
3402 : /*
3403 : * Now that we've found the best set of aggregates we can set the
3404 : * presorted flag to indicate to the executor that it needn't bother
3405 : * performing a sort for these Aggrefs. We're able to do this now as
3406 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3407 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3408 : * of ordered aggregates.
3409 : */
3410 2390 : i = -1;
3411 4642 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3412 : {
3413 2252 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3414 :
3415 4522 : foreach(lc, agginfo->aggrefs)
3416 : {
3417 2270 : Aggref *aggref = lfirst_node(Aggref, lc);
3418 :
3419 2270 : aggref->aggpresorted = true;
3420 : }
3421 : }
3422 : }
3423 :
3424 : /*
3425 : * Compute query_pathkeys and other pathkeys during plan generation
3426 : */
3427 : static void
3428 510306 : standard_qp_callback(PlannerInfo *root, void *extra)
3429 : {
3430 510306 : Query *parse = root->parse;
3431 510306 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3432 510306 : List *tlist = root->processed_tlist;
3433 510306 : List *activeWindows = qp_extra->activeWindows;
3434 :
3435 : /*
3436 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3437 : * aggregate requirements.
3438 : */
3439 510306 : if (qp_extra->gset_data)
3440 : {
3441 : /*
3442 : * With grouping sets, just use the first RollupData's groupClause. We
3443 : * don't make any effort to optimize grouping clauses when there are
3444 : * grouping sets, nor can we combine aggregate ordering keys with
3445 : * grouping.
3446 : */
3447 872 : List *rollups = qp_extra->gset_data->rollups;
3448 872 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3449 :
3450 872 : if (grouping_is_sortable(groupClause))
3451 : {
3452 : bool sortable;
3453 :
3454 : /*
3455 : * The groupClause is logically below the grouping step. So if
3456 : * there is an RTE entry for the grouping step, we need to remove
3457 : * its RT index from the sort expressions before we make PathKeys
3458 : * for them.
3459 : */
3460 872 : root->group_pathkeys =
3461 872 : make_pathkeys_for_sortclauses_extended(root,
3462 : &groupClause,
3463 : tlist,
3464 : false,
3465 872 : parse->hasGroupRTE,
3466 : &sortable,
3467 : false);
3468 : Assert(sortable);
3469 872 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3470 : }
3471 : else
3472 : {
3473 0 : root->group_pathkeys = NIL;
3474 0 : root->num_groupby_pathkeys = 0;
3475 : }
3476 : }
3477 509434 : else if (parse->groupClause || root->numOrderedAggs > 0)
3478 5782 : {
3479 : /*
3480 : * With a plain GROUP BY list, we can remove any grouping items that
3481 : * are proven redundant by EquivalenceClass processing. For example,
3482 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3483 : * especially common cases, but they're nearly free to detect. Note
3484 : * that we remove redundant items from processed_groupClause but not
3485 : * the original parse->groupClause.
3486 : */
3487 : bool sortable;
3488 :
3489 : /*
3490 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3491 : * EquivalenceClass'es if it's not set yet.
3492 : */
3493 5782 : root->group_pathkeys =
3494 5782 : make_pathkeys_for_sortclauses_extended(root,
3495 : &root->processed_groupClause,
3496 : tlist,
3497 : true,
3498 : false,
3499 : &sortable,
3500 : true);
3501 5782 : if (!sortable)
3502 : {
3503 : /* Can't sort; no point in considering aggregate ordering either */
3504 0 : root->group_pathkeys = NIL;
3505 0 : root->num_groupby_pathkeys = 0;
3506 : }
3507 : else
3508 : {
3509 5782 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3510 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3511 5782 : if (root->numOrderedAggs > 0)
3512 2396 : adjust_group_pathkeys_for_groupagg(root);
3513 : }
3514 : }
3515 : else
3516 : {
3517 503652 : root->group_pathkeys = NIL;
3518 503652 : root->num_groupby_pathkeys = 0;
3519 : }
3520 :
3521 : /* We consider only the first (bottom) window in pathkeys logic */
3522 510306 : if (activeWindows != NIL)
3523 : {
3524 2378 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3525 :
3526 2378 : root->window_pathkeys = make_pathkeys_for_window(root,
3527 : wc,
3528 : tlist);
3529 : }
3530 : else
3531 507928 : root->window_pathkeys = NIL;
3532 :
3533 : /*
3534 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3535 : * redundant by EquivalenceClass processing. The non-redundant list is
3536 : * kept in root->processed_distinctClause, leaving the original
3537 : * parse->distinctClause alone.
3538 : */
3539 510306 : if (parse->distinctClause)
3540 : {
3541 : bool sortable;
3542 :
3543 : /* Make a copy since pathkey processing can modify the list */
3544 2620 : root->processed_distinctClause = list_copy(parse->distinctClause);
3545 2620 : root->distinct_pathkeys =
3546 2620 : make_pathkeys_for_sortclauses_extended(root,
3547 : &root->processed_distinctClause,
3548 : tlist,
3549 : true,
3550 : false,
3551 : &sortable,
3552 : false);
3553 2620 : if (!sortable)
3554 6 : root->distinct_pathkeys = NIL;
3555 : }
3556 : else
3557 507686 : root->distinct_pathkeys = NIL;
3558 :
3559 510306 : root->sort_pathkeys =
3560 510306 : make_pathkeys_for_sortclauses(root,
3561 : parse->sortClause,
3562 : tlist);
3563 :
3564 : /* setting setop_pathkeys might be useful to the union planner */
3565 510306 : if (qp_extra->setop != NULL)
3566 : {
3567 : List *groupClauses;
3568 : bool sortable;
3569 :
3570 12274 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3571 :
3572 12274 : root->setop_pathkeys =
3573 12274 : make_pathkeys_for_sortclauses_extended(root,
3574 : &groupClauses,
3575 : tlist,
3576 : false,
3577 : false,
3578 : &sortable,
3579 : false);
3580 12274 : if (!sortable)
3581 204 : root->setop_pathkeys = NIL;
3582 : }
3583 : else
3584 498032 : root->setop_pathkeys = NIL;
3585 :
3586 : /*
3587 : * Figure out whether we want a sorted result from query_planner.
3588 : *
3589 : * If we have a sortable GROUP BY clause, then we want a result sorted
3590 : * properly for grouping. Otherwise, if we have window functions to
3591 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3592 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3593 : * we try to produce output that's sufficiently well sorted for the
3594 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3595 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3596 : * for a set operation which can benefit from presorted results and have a
3597 : * sortable targetlist, we want to sort by the target list.
3598 : *
3599 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3600 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3601 : * that might just leave us failing to exploit an available sort order at
3602 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3603 : * much easier, since we know that the parser ensured that one is a
3604 : * superset of the other.
3605 : */
3606 510306 : if (root->group_pathkeys)
3607 6298 : root->query_pathkeys = root->group_pathkeys;
3608 504008 : else if (root->window_pathkeys)
3609 2032 : root->query_pathkeys = root->window_pathkeys;
3610 1003952 : else if (list_length(root->distinct_pathkeys) >
3611 501976 : list_length(root->sort_pathkeys))
3612 2158 : root->query_pathkeys = root->distinct_pathkeys;
3613 499818 : else if (root->sort_pathkeys)
3614 67304 : root->query_pathkeys = root->sort_pathkeys;
3615 432514 : else if (root->setop_pathkeys != NIL)
3616 10878 : root->query_pathkeys = root->setop_pathkeys;
3617 : else
3618 421636 : root->query_pathkeys = NIL;
3619 510306 : }
3620 :
3621 : /*
3622 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3623 : *
3624 : * path_rows: number of output rows from scan/join step
3625 : * gd: grouping sets data including list of grouping sets and their clauses
3626 : * target_list: target list containing group clause references
3627 : *
3628 : * If doing grouping sets, we also annotate the gsets data with the estimates
3629 : * for each set and each individual rollup list, with a view to later
3630 : * determining whether some combination of them could be hashed instead.
3631 : */
3632 : static double
3633 42080 : get_number_of_groups(PlannerInfo *root,
3634 : double path_rows,
3635 : grouping_sets_data *gd,
3636 : List *target_list)
3637 : {
3638 42080 : Query *parse = root->parse;
3639 : double dNumGroups;
3640 :
3641 42080 : if (parse->groupClause)
3642 : {
3643 : List *groupExprs;
3644 :
3645 6940 : if (parse->groupingSets)
3646 : {
3647 : /* Add up the estimates for each grouping set */
3648 : ListCell *lc;
3649 :
3650 : Assert(gd); /* keep Coverity happy */
3651 :
3652 830 : dNumGroups = 0;
3653 :
3654 2226 : foreach(lc, gd->rollups)
3655 : {
3656 1396 : RollupData *rollup = lfirst_node(RollupData, lc);
3657 : ListCell *lc2;
3658 : ListCell *lc3;
3659 :
3660 1396 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3661 : target_list);
3662 :
3663 1396 : rollup->numGroups = 0.0;
3664 :
3665 4040 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3666 : {
3667 2644 : List *gset = (List *) lfirst(lc2);
3668 2644 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3669 2644 : double numGroups = estimate_num_groups(root,
3670 : groupExprs,
3671 : path_rows,
3672 : &gset,
3673 : NULL);
3674 :
3675 2644 : gs->numGroups = numGroups;
3676 2644 : rollup->numGroups += numGroups;
3677 : }
3678 :
3679 1396 : dNumGroups += rollup->numGroups;
3680 : }
3681 :
3682 830 : if (gd->hash_sets_idx)
3683 : {
3684 : ListCell *lc2;
3685 :
3686 36 : gd->dNumHashGroups = 0;
3687 :
3688 36 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3689 : target_list);
3690 :
3691 78 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3692 : {
3693 42 : List *gset = (List *) lfirst(lc);
3694 42 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3695 42 : double numGroups = estimate_num_groups(root,
3696 : groupExprs,
3697 : path_rows,
3698 : &gset,
3699 : NULL);
3700 :
3701 42 : gs->numGroups = numGroups;
3702 42 : gd->dNumHashGroups += numGroups;
3703 : }
3704 :
3705 36 : dNumGroups += gd->dNumHashGroups;
3706 : }
3707 : }
3708 : else
3709 : {
3710 : /* Plain GROUP BY -- estimate based on optimized groupClause */
3711 6110 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3712 : target_list);
3713 :
3714 6110 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3715 : NULL, NULL);
3716 : }
3717 : }
3718 35140 : else if (parse->groupingSets)
3719 : {
3720 : /* Empty grouping sets ... one result row for each one */
3721 42 : dNumGroups = list_length(parse->groupingSets);
3722 : }
3723 35098 : else if (parse->hasAggs || root->hasHavingQual)
3724 : {
3725 : /* Plain aggregation, one result row */
3726 35098 : dNumGroups = 1;
3727 : }
3728 : else
3729 : {
3730 : /* Not grouping */
3731 0 : dNumGroups = 1;
3732 : }
3733 :
3734 42080 : return dNumGroups;
3735 : }
3736 :
3737 : /*
3738 : * create_grouping_paths
3739 : *
3740 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3741 : * Along the way, we also build an upperrel for Paths which are partially
3742 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3743 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3744 : * the only partially grouped paths we build are also partial paths; that
3745 : * is, they need a Gather and then a FinalizeAggregate.
3746 : *
3747 : * input_rel: contains the source-data Paths
3748 : * target: the pathtarget for the result Paths to compute
3749 : * gd: grouping sets data including list of grouping sets and their clauses
3750 : *
3751 : * Note: all Paths in input_rel are expected to return the target computed
3752 : * by make_group_input_target.
3753 : */
3754 : static RelOptInfo *
3755 38820 : create_grouping_paths(PlannerInfo *root,
3756 : RelOptInfo *input_rel,
3757 : PathTarget *target,
3758 : bool target_parallel_safe,
3759 : grouping_sets_data *gd)
3760 : {
3761 38820 : Query *parse = root->parse;
3762 : RelOptInfo *grouped_rel;
3763 : RelOptInfo *partially_grouped_rel;
3764 : AggClauseCosts agg_costs;
3765 :
3766 232920 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3767 38820 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3768 :
3769 : /*
3770 : * Create grouping relation to hold fully aggregated grouping and/or
3771 : * aggregation paths.
3772 : */
3773 38820 : grouped_rel = make_grouping_rel(root, input_rel, target,
3774 : target_parallel_safe, parse->havingQual);
3775 :
3776 : /*
3777 : * Create either paths for a degenerate grouping or paths for ordinary
3778 : * grouping, as appropriate.
3779 : */
3780 38820 : if (is_degenerate_grouping(root))
3781 18 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3782 : else
3783 : {
3784 38802 : int flags = 0;
3785 : GroupPathExtraData extra;
3786 :
3787 : /*
3788 : * Determine whether it's possible to perform sort-based
3789 : * implementations of grouping. (Note that if processed_groupClause
3790 : * is empty, grouping_is_sortable() is trivially true, and all the
3791 : * pathkeys_contained_in() tests will succeed too, so that we'll
3792 : * consider every surviving input path.)
3793 : *
3794 : * If we have grouping sets, we might be able to sort some but not all
3795 : * of them; in this case, we need can_sort to be true as long as we
3796 : * must consider any sorted-input plan.
3797 : */
3798 38802 : if ((gd && gd->rollups != NIL)
3799 37936 : || grouping_is_sortable(root->processed_groupClause))
3800 38796 : flags |= GROUPING_CAN_USE_SORT;
3801 :
3802 : /*
3803 : * Determine whether we should consider hash-based implementations of
3804 : * grouping.
3805 : *
3806 : * Hashed aggregation only applies if we're grouping. If we have
3807 : * grouping sets, some groups might be hashable but others not; in
3808 : * this case we set can_hash true as long as there is nothing globally
3809 : * preventing us from hashing (and we should therefore consider plans
3810 : * with hashes).
3811 : *
3812 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3813 : * BY aggregates. (Doing so would imply storing *all* the input
3814 : * values in the hash table, and/or running many sorts in parallel,
3815 : * either of which seems like a certain loser.) We similarly don't
3816 : * support ordered-set aggregates in hashed aggregation, but that case
3817 : * is also included in the numOrderedAggs count.
3818 : *
3819 : * Note: grouping_is_hashable() is much more expensive to check than
3820 : * the other gating conditions, so we want to do it last.
3821 : */
3822 38802 : if ((parse->groupClause != NIL &&
3823 8640 : root->numOrderedAggs == 0 &&
3824 4180 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3825 4176 : flags |= GROUPING_CAN_USE_HASH;
3826 :
3827 : /*
3828 : * Determine whether partial aggregation is possible.
3829 : */
3830 38802 : if (can_partial_agg(root))
3831 34148 : flags |= GROUPING_CAN_PARTIAL_AGG;
3832 :
3833 38802 : extra.flags = flags;
3834 38802 : extra.target_parallel_safe = target_parallel_safe;
3835 38802 : extra.havingQual = parse->havingQual;
3836 38802 : extra.targetList = parse->targetList;
3837 38802 : extra.partial_costs_set = false;
3838 :
3839 : /*
3840 : * Determine whether partitionwise aggregation is in theory possible.
3841 : * It can be disabled by the user, and for now, we don't try to
3842 : * support grouping sets. create_ordinary_grouping_paths() will check
3843 : * additional conditions, such as whether input_rel is partitioned.
3844 : */
3845 38802 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3846 556 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3847 : else
3848 38246 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3849 :
3850 38802 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3851 : &agg_costs, gd, &extra,
3852 : &partially_grouped_rel);
3853 : }
3854 :
3855 38814 : set_cheapest(grouped_rel);
3856 38814 : return grouped_rel;
3857 : }
3858 :
3859 : /*
3860 : * make_grouping_rel
3861 : *
3862 : * Create a new grouping rel and set basic properties.
3863 : *
3864 : * input_rel represents the underlying scan/join relation.
3865 : * target is the output expected from the grouping relation.
3866 : */
3867 : static RelOptInfo *
3868 40314 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3869 : PathTarget *target, bool target_parallel_safe,
3870 : Node *havingQual)
3871 : {
3872 : RelOptInfo *grouped_rel;
3873 :
3874 40314 : if (IS_OTHER_REL(input_rel))
3875 : {
3876 1494 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3877 : input_rel->relids);
3878 1494 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3879 : }
3880 : else
3881 : {
3882 : /*
3883 : * By tradition, the relids set for the main grouping relation is
3884 : * NULL. (This could be changed, but might require adjustments
3885 : * elsewhere.)
3886 : */
3887 38820 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3888 : }
3889 :
3890 : /* Set target. */
3891 40314 : grouped_rel->reltarget = target;
3892 :
3893 : /*
3894 : * If the input relation is not parallel-safe, then the grouped relation
3895 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3896 : * target list and HAVING quals are parallel-safe.
3897 : */
3898 68480 : if (input_rel->consider_parallel && target_parallel_safe &&
3899 28166 : is_parallel_safe(root, (Node *) havingQual))
3900 28148 : grouped_rel->consider_parallel = true;
3901 :
3902 : /*
3903 : * If the input rel belongs to a single FDW, so does the grouped rel.
3904 : */
3905 40314 : grouped_rel->serverid = input_rel->serverid;
3906 40314 : grouped_rel->userid = input_rel->userid;
3907 40314 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3908 40314 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3909 :
3910 40314 : return grouped_rel;
3911 : }
3912 :
3913 : /*
3914 : * is_degenerate_grouping
3915 : *
3916 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3917 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3918 : * grouping sets are all empty).
3919 : */
3920 : static bool
3921 38820 : is_degenerate_grouping(PlannerInfo *root)
3922 : {
3923 38820 : Query *parse = root->parse;
3924 :
3925 37800 : return (root->hasHavingQual || parse->groupingSets) &&
3926 76620 : !parse->hasAggs && parse->groupClause == NIL;
3927 : }
3928 :
3929 : /*
3930 : * create_degenerate_grouping_paths
3931 : *
3932 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3933 : * supposed to emit either zero or one row for each grouping set depending on
3934 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3935 : * either HAVING or the targetlist, so we actually do not need the FROM table
3936 : * at all! We can just throw away the plan-so-far and generate a Result node.
3937 : * This is a sufficiently unusual corner case that it's not worth contorting
3938 : * the structure of this module to avoid having to generate the earlier paths
3939 : * in the first place.
3940 : */
3941 : static void
3942 18 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3943 : RelOptInfo *grouped_rel)
3944 : {
3945 18 : Query *parse = root->parse;
3946 : int nrows;
3947 : Path *path;
3948 :
3949 18 : nrows = list_length(parse->groupingSets);
3950 18 : if (nrows > 1)
3951 : {
3952 : /*
3953 : * Doesn't seem worthwhile writing code to cons up a generate_series
3954 : * or a values scan to emit multiple rows. Instead just make N clones
3955 : * and append them. (With a volatile HAVING clause, this means you
3956 : * might get between 0 and N output rows. Offhand I think that's
3957 : * desired.)
3958 : */
3959 0 : List *paths = NIL;
3960 :
3961 0 : while (--nrows >= 0)
3962 : {
3963 : path = (Path *)
3964 0 : create_group_result_path(root, grouped_rel,
3965 0 : grouped_rel->reltarget,
3966 0 : (List *) parse->havingQual);
3967 0 : paths = lappend(paths, path);
3968 : }
3969 : path = (Path *)
3970 0 : create_append_path(root,
3971 : grouped_rel,
3972 : paths,
3973 : NIL,
3974 : NIL,
3975 : NULL,
3976 : 0,
3977 : false,
3978 : -1);
3979 : }
3980 : else
3981 : {
3982 : /* No grouping sets, or just one, so one output row */
3983 : path = (Path *)
3984 18 : create_group_result_path(root, grouped_rel,
3985 18 : grouped_rel->reltarget,
3986 18 : (List *) parse->havingQual);
3987 : }
3988 :
3989 18 : add_path(grouped_rel, path);
3990 18 : }
3991 :
3992 : /*
3993 : * create_ordinary_grouping_paths
3994 : *
3995 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
3996 : *
3997 : * We need to consider sorted and hashed aggregation in the same function,
3998 : * because otherwise (1) it would be harder to throw an appropriate error
3999 : * message if neither way works, and (2) we should not allow hashtable size
4000 : * considerations to dissuade us from using hashing if sorting is not possible.
4001 : *
4002 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4003 : * function creates, or to NULL if it doesn't create one.
4004 : */
4005 : static void
4006 40296 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4007 : RelOptInfo *grouped_rel,
4008 : const AggClauseCosts *agg_costs,
4009 : grouping_sets_data *gd,
4010 : GroupPathExtraData *extra,
4011 : RelOptInfo **partially_grouped_rel_p)
4012 : {
4013 40296 : Path *cheapest_path = input_rel->cheapest_total_path;
4014 40296 : RelOptInfo *partially_grouped_rel = NULL;
4015 : double dNumGroups;
4016 40296 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4017 :
4018 : /*
4019 : * If this is the topmost grouping relation or if the parent relation is
4020 : * doing some form of partitionwise aggregation, then we may be able to do
4021 : * it at this level also. However, if the input relation is not
4022 : * partitioned, partitionwise aggregate is impossible.
4023 : */
4024 40296 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4025 2050 : IS_PARTITIONED_REL(input_rel))
4026 : {
4027 : /*
4028 : * If this is the topmost relation or if the parent relation is doing
4029 : * full partitionwise aggregation, then we can do full partitionwise
4030 : * aggregation provided that the GROUP BY clause contains all of the
4031 : * partitioning columns at this level and the collation used by GROUP
4032 : * BY matches the partitioning collation. Otherwise, we can do at
4033 : * most partial partitionwise aggregation. But if partial aggregation
4034 : * is not supported in general then we can't use it for partitionwise
4035 : * aggregation either.
4036 : *
4037 : * Check parse->groupClause not processed_groupClause, because it's
4038 : * okay if some of the partitioning columns were proved redundant.
4039 : */
4040 1160 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4041 556 : group_by_has_partkey(input_rel, extra->targetList,
4042 556 : root->parse->groupClause))
4043 320 : patype = PARTITIONWISE_AGGREGATE_FULL;
4044 284 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4045 242 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4046 : else
4047 42 : patype = PARTITIONWISE_AGGREGATE_NONE;
4048 : }
4049 :
4050 : /*
4051 : * Before generating paths for grouped_rel, we first generate any possible
4052 : * partially grouped paths; that way, later code can easily consider both
4053 : * parallel and non-parallel approaches to grouping.
4054 : */
4055 40296 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4056 : {
4057 : bool force_rel_creation;
4058 :
4059 : /*
4060 : * If we're doing partitionwise aggregation at this level, force
4061 : * creation of a partially_grouped_rel so we can add partitionwise
4062 : * paths to it.
4063 : */
4064 35570 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4065 :
4066 : partially_grouped_rel =
4067 35570 : create_partial_grouping_paths(root,
4068 : grouped_rel,
4069 : input_rel,
4070 : gd,
4071 : extra,
4072 : force_rel_creation);
4073 : }
4074 :
4075 : /* Set out parameter. */
4076 40296 : *partially_grouped_rel_p = partially_grouped_rel;
4077 :
4078 : /* Apply partitionwise aggregation technique, if possible. */
4079 40296 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4080 562 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4081 : partially_grouped_rel, agg_costs,
4082 : gd, patype, extra);
4083 :
4084 : /* If we are doing partial aggregation only, return. */
4085 40296 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4086 : {
4087 : Assert(partially_grouped_rel);
4088 :
4089 618 : if (partially_grouped_rel->pathlist)
4090 618 : set_cheapest(partially_grouped_rel);
4091 :
4092 618 : return;
4093 : }
4094 :
4095 : /* Gather any partially grouped partial paths. */
4096 39678 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4097 : {
4098 1484 : gather_grouping_paths(root, partially_grouped_rel);
4099 1484 : set_cheapest(partially_grouped_rel);
4100 : }
4101 :
4102 : /*
4103 : * Estimate number of groups.
4104 : */
4105 39678 : dNumGroups = get_number_of_groups(root,
4106 : cheapest_path->rows,
4107 : gd,
4108 : extra->targetList);
4109 :
4110 : /* Build final grouping paths */
4111 39678 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4112 : partially_grouped_rel, agg_costs, gd,
4113 : dNumGroups, extra);
4114 :
4115 : /* Give a helpful error if we failed to find any implementation */
4116 39678 : if (grouped_rel->pathlist == NIL)
4117 6 : ereport(ERROR,
4118 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4119 : errmsg("could not implement GROUP BY"),
4120 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4121 :
4122 : /*
4123 : * If there is an FDW that's responsible for all baserels of the query,
4124 : * let it consider adding ForeignPaths.
4125 : */
4126 39672 : if (grouped_rel->fdwroutine &&
4127 336 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4128 336 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4129 : input_rel, grouped_rel,
4130 : extra);
4131 :
4132 : /* Let extensions possibly add some more paths */
4133 39672 : if (create_upper_paths_hook)
4134 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4135 : input_rel, grouped_rel,
4136 : extra);
4137 : }
4138 :
4139 : /*
4140 : * For a given input path, consider the possible ways of doing grouping sets on
4141 : * it, by combinations of hashing and sorting. This can be called multiple
4142 : * times, so it's important that it not scribble on input. No result is
4143 : * returned, but any generated paths are added to grouped_rel.
4144 : */
4145 : static void
4146 1732 : consider_groupingsets_paths(PlannerInfo *root,
4147 : RelOptInfo *grouped_rel,
4148 : Path *path,
4149 : bool is_sorted,
4150 : bool can_hash,
4151 : grouping_sets_data *gd,
4152 : const AggClauseCosts *agg_costs,
4153 : double dNumGroups)
4154 : {
4155 1732 : Query *parse = root->parse;
4156 1732 : Size hash_mem_limit = get_hash_memory_limit();
4157 :
4158 : /*
4159 : * If we're not being offered sorted input, then only consider plans that
4160 : * can be done entirely by hashing.
4161 : *
4162 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4163 : * the input is actually sorted despite not being advertised as such, we
4164 : * prefer to make use of that in order to use less memory.
4165 : *
4166 : * If none of the grouping sets are sortable, then ignore the hash_mem
4167 : * limit and generate a path anyway, since otherwise we'll just fail.
4168 : */
4169 1732 : if (!is_sorted)
4170 : {
4171 794 : List *new_rollups = NIL;
4172 794 : RollupData *unhashed_rollup = NULL;
4173 : List *sets_data;
4174 794 : List *empty_sets_data = NIL;
4175 794 : List *empty_sets = NIL;
4176 : ListCell *lc;
4177 794 : ListCell *l_start = list_head(gd->rollups);
4178 794 : AggStrategy strat = AGG_HASHED;
4179 : double hashsize;
4180 794 : double exclude_groups = 0.0;
4181 :
4182 : Assert(can_hash);
4183 :
4184 : /*
4185 : * If the input is coincidentally sorted usefully (which can happen
4186 : * even if is_sorted is false, since that only means that our caller
4187 : * has set up the sorting for us), then save some hashtable space by
4188 : * making use of that. But we need to watch out for degenerate cases:
4189 : *
4190 : * 1) If there are any empty grouping sets, then group_pathkeys might
4191 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4192 : * there will be a rollup containing only empty groups, and the
4193 : * pathkeys_contained_in test is vacuously true; this is ok.
4194 : *
4195 : * XXX: the above relies on the fact that group_pathkeys is generated
4196 : * from the first rollup. If we add the ability to consider multiple
4197 : * sort orders for grouping input, this assumption might fail.
4198 : *
4199 : * 2) If there are no empty sets and only unsortable sets, then the
4200 : * rollups list will be empty (and thus l_start == NULL), and
4201 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4202 : * pathkeys_contained_in test doesn't cause us to crash.
4203 : */
4204 1582 : if (l_start != NULL &&
4205 788 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4206 : {
4207 12 : unhashed_rollup = lfirst_node(RollupData, l_start);
4208 12 : exclude_groups = unhashed_rollup->numGroups;
4209 12 : l_start = lnext(gd->rollups, l_start);
4210 : }
4211 :
4212 794 : hashsize = estimate_hashagg_tablesize(root,
4213 : path,
4214 : agg_costs,
4215 : dNumGroups - exclude_groups);
4216 :
4217 : /*
4218 : * gd->rollups is empty if we have only unsortable columns to work
4219 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4220 : * sorted-input case to generate usable mixed paths.
4221 : */
4222 794 : if (hashsize > hash_mem_limit && gd->rollups)
4223 18 : return; /* nope, won't fit */
4224 :
4225 : /*
4226 : * We need to burst the existing rollups list into individual grouping
4227 : * sets and recompute a groupClause for each set.
4228 : */
4229 776 : sets_data = list_copy(gd->unsortable_sets);
4230 :
4231 1980 : for_each_cell(lc, gd->rollups, l_start)
4232 : {
4233 1228 : RollupData *rollup = lfirst_node(RollupData, lc);
4234 :
4235 : /*
4236 : * If we find an unhashable rollup that's not been skipped by the
4237 : * "actually sorted" check above, we can't cope; we'd need sorted
4238 : * input (with a different sort order) but we can't get that here.
4239 : * So bail out; we'll get a valid path from the is_sorted case
4240 : * instead.
4241 : *
4242 : * The mere presence of empty grouping sets doesn't make a rollup
4243 : * unhashable (see preprocess_grouping_sets), we handle those
4244 : * specially below.
4245 : */
4246 1228 : if (!rollup->hashable)
4247 24 : return;
4248 :
4249 1204 : sets_data = list_concat(sets_data, rollup->gsets_data);
4250 : }
4251 3162 : foreach(lc, sets_data)
4252 : {
4253 2410 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4254 2410 : List *gset = gs->set;
4255 : RollupData *rollup;
4256 :
4257 2410 : if (gset == NIL)
4258 : {
4259 : /* Empty grouping sets can't be hashed. */
4260 484 : empty_sets_data = lappend(empty_sets_data, gs);
4261 484 : empty_sets = lappend(empty_sets, NIL);
4262 : }
4263 : else
4264 : {
4265 1926 : rollup = makeNode(RollupData);
4266 :
4267 1926 : rollup->groupClause = preprocess_groupclause(root, gset);
4268 1926 : rollup->gsets_data = list_make1(gs);
4269 1926 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4270 : rollup->gsets_data,
4271 : gd->tleref_to_colnum_map);
4272 1926 : rollup->numGroups = gs->numGroups;
4273 1926 : rollup->hashable = true;
4274 1926 : rollup->is_hashed = true;
4275 1926 : new_rollups = lappend(new_rollups, rollup);
4276 : }
4277 : }
4278 :
4279 : /*
4280 : * If we didn't find anything nonempty to hash, then bail. We'll
4281 : * generate a path from the is_sorted case.
4282 : */
4283 752 : if (new_rollups == NIL)
4284 0 : return;
4285 :
4286 : /*
4287 : * If there were empty grouping sets they should have been in the
4288 : * first rollup.
4289 : */
4290 : Assert(!unhashed_rollup || !empty_sets);
4291 :
4292 752 : if (unhashed_rollup)
4293 : {
4294 12 : new_rollups = lappend(new_rollups, unhashed_rollup);
4295 12 : strat = AGG_MIXED;
4296 : }
4297 740 : else if (empty_sets)
4298 : {
4299 436 : RollupData *rollup = makeNode(RollupData);
4300 :
4301 436 : rollup->groupClause = NIL;
4302 436 : rollup->gsets_data = empty_sets_data;
4303 436 : rollup->gsets = empty_sets;
4304 436 : rollup->numGroups = list_length(empty_sets);
4305 436 : rollup->hashable = false;
4306 436 : rollup->is_hashed = false;
4307 436 : new_rollups = lappend(new_rollups, rollup);
4308 436 : strat = AGG_MIXED;
4309 : }
4310 :
4311 752 : add_path(grouped_rel, (Path *)
4312 752 : create_groupingsets_path(root,
4313 : grouped_rel,
4314 : path,
4315 752 : (List *) parse->havingQual,
4316 : strat,
4317 : new_rollups,
4318 : agg_costs));
4319 752 : return;
4320 : }
4321 :
4322 : /*
4323 : * If we have sorted input but nothing we can do with it, bail.
4324 : */
4325 938 : if (gd->rollups == NIL)
4326 0 : return;
4327 :
4328 : /*
4329 : * Given sorted input, we try and make two paths: one sorted and one mixed
4330 : * sort/hash. (We need to try both because hashagg might be disabled, or
4331 : * some columns might not be sortable.)
4332 : *
4333 : * can_hash is passed in as false if some obstacle elsewhere (such as
4334 : * ordered aggs) means that we shouldn't consider hashing at all.
4335 : */
4336 938 : if (can_hash && gd->any_hashable)
4337 : {
4338 860 : List *rollups = NIL;
4339 860 : List *hash_sets = list_copy(gd->unsortable_sets);
4340 860 : double availspace = hash_mem_limit;
4341 : ListCell *lc;
4342 :
4343 : /*
4344 : * Account first for space needed for groups we can't sort at all.
4345 : */
4346 860 : availspace -= estimate_hashagg_tablesize(root,
4347 : path,
4348 : agg_costs,
4349 : gd->dNumHashGroups);
4350 :
4351 860 : if (availspace > 0 && list_length(gd->rollups) > 1)
4352 : {
4353 : double scale;
4354 444 : int num_rollups = list_length(gd->rollups);
4355 : int k_capacity;
4356 444 : int *k_weights = palloc(num_rollups * sizeof(int));
4357 444 : Bitmapset *hash_items = NULL;
4358 : int i;
4359 :
4360 : /*
4361 : * We treat this as a knapsack problem: the knapsack capacity
4362 : * represents hash_mem, the item weights are the estimated memory
4363 : * usage of the hashtables needed to implement a single rollup,
4364 : * and we really ought to use the cost saving as the item value;
4365 : * however, currently the costs assigned to sort nodes don't
4366 : * reflect the comparison costs well, and so we treat all items as
4367 : * of equal value (each rollup we hash instead saves us one sort).
4368 : *
4369 : * To use the discrete knapsack, we need to scale the values to a
4370 : * reasonably small bounded range. We choose to allow a 5% error
4371 : * margin; we have no more than 4096 rollups in the worst possible
4372 : * case, which with a 5% error margin will require a bit over 42MB
4373 : * of workspace. (Anyone wanting to plan queries that complex had
4374 : * better have the memory for it. In more reasonable cases, with
4375 : * no more than a couple of dozen rollups, the memory usage will
4376 : * be negligible.)
4377 : *
4378 : * k_capacity is naturally bounded, but we clamp the values for
4379 : * scale and weight (below) to avoid overflows or underflows (or
4380 : * uselessly trying to use a scale factor less than 1 byte).
4381 : */
4382 444 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4383 444 : k_capacity = (int) floor(availspace / scale);
4384 :
4385 : /*
4386 : * We leave the first rollup out of consideration since it's the
4387 : * one that matches the input sort order. We assign indexes "i"
4388 : * to only those entries considered for hashing; the second loop,
4389 : * below, must use the same condition.
4390 : */
4391 444 : i = 0;
4392 1140 : for_each_from(lc, gd->rollups, 1)
4393 : {
4394 696 : RollupData *rollup = lfirst_node(RollupData, lc);
4395 :
4396 696 : if (rollup->hashable)
4397 : {
4398 696 : double sz = estimate_hashagg_tablesize(root,
4399 : path,
4400 : agg_costs,
4401 : rollup->numGroups);
4402 :
4403 : /*
4404 : * If sz is enormous, but hash_mem (and hence scale) is
4405 : * small, avoid integer overflow here.
4406 : */
4407 696 : k_weights[i] = (int) Min(floor(sz / scale),
4408 : k_capacity + 1.0);
4409 696 : ++i;
4410 : }
4411 : }
4412 :
4413 : /*
4414 : * Apply knapsack algorithm; compute the set of items which
4415 : * maximizes the value stored (in this case the number of sorts
4416 : * saved) while keeping the total size (approximately) within
4417 : * capacity.
4418 : */
4419 444 : if (i > 0)
4420 444 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4421 :
4422 444 : if (!bms_is_empty(hash_items))
4423 : {
4424 444 : rollups = list_make1(linitial(gd->rollups));
4425 :
4426 444 : i = 0;
4427 1140 : for_each_from(lc, gd->rollups, 1)
4428 : {
4429 696 : RollupData *rollup = lfirst_node(RollupData, lc);
4430 :
4431 696 : if (rollup->hashable)
4432 : {
4433 696 : if (bms_is_member(i, hash_items))
4434 660 : hash_sets = list_concat(hash_sets,
4435 660 : rollup->gsets_data);
4436 : else
4437 36 : rollups = lappend(rollups, rollup);
4438 696 : ++i;
4439 : }
4440 : else
4441 0 : rollups = lappend(rollups, rollup);
4442 : }
4443 : }
4444 : }
4445 :
4446 860 : if (!rollups && hash_sets)
4447 24 : rollups = list_copy(gd->rollups);
4448 :
4449 1660 : foreach(lc, hash_sets)
4450 : {
4451 800 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4452 800 : RollupData *rollup = makeNode(RollupData);
4453 :
4454 : Assert(gs->set != NIL);
4455 :
4456 800 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4457 800 : rollup->gsets_data = list_make1(gs);
4458 800 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4459 : rollup->gsets_data,
4460 : gd->tleref_to_colnum_map);
4461 800 : rollup->numGroups = gs->numGroups;
4462 800 : rollup->hashable = true;
4463 800 : rollup->is_hashed = true;
4464 800 : rollups = lcons(rollup, rollups);
4465 : }
4466 :
4467 860 : if (rollups)
4468 : {
4469 468 : add_path(grouped_rel, (Path *)
4470 468 : create_groupingsets_path(root,
4471 : grouped_rel,
4472 : path,
4473 468 : (List *) parse->havingQual,
4474 : AGG_MIXED,
4475 : rollups,
4476 : agg_costs));
4477 : }
4478 : }
4479 :
4480 : /*
4481 : * Now try the simple sorted case.
4482 : */
4483 938 : if (!gd->unsortable_sets)
4484 908 : add_path(grouped_rel, (Path *)
4485 908 : create_groupingsets_path(root,
4486 : grouped_rel,
4487 : path,
4488 908 : (List *) parse->havingQual,
4489 : AGG_SORTED,
4490 : gd->rollups,
4491 : agg_costs));
4492 : }
4493 :
4494 : /*
4495 : * create_window_paths
4496 : *
4497 : * Build a new upperrel containing Paths for window-function evaluation.
4498 : *
4499 : * input_rel: contains the source-data Paths
4500 : * input_target: result of make_window_input_target
4501 : * output_target: what the topmost WindowAggPath should return
4502 : * wflists: result of find_window_functions
4503 : * activeWindows: result of select_active_windows
4504 : *
4505 : * Note: all Paths in input_rel are expected to return input_target.
4506 : */
4507 : static RelOptInfo *
4508 2378 : create_window_paths(PlannerInfo *root,
4509 : RelOptInfo *input_rel,
4510 : PathTarget *input_target,
4511 : PathTarget *output_target,
4512 : bool output_target_parallel_safe,
4513 : WindowFuncLists *wflists,
4514 : List *activeWindows)
4515 : {
4516 : RelOptInfo *window_rel;
4517 : ListCell *lc;
4518 :
4519 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4520 2378 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4521 :
4522 : /*
4523 : * If the input relation is not parallel-safe, then the window relation
4524 : * can't be parallel-safe, either. Otherwise, we need to examine the
4525 : * target list and active windows for non-parallel-safe constructs.
4526 : */
4527 2378 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4528 0 : is_parallel_safe(root, (Node *) activeWindows))
4529 0 : window_rel->consider_parallel = true;
4530 :
4531 : /*
4532 : * If the input rel belongs to a single FDW, so does the window rel.
4533 : */
4534 2378 : window_rel->serverid = input_rel->serverid;
4535 2378 : window_rel->userid = input_rel->userid;
4536 2378 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4537 2378 : window_rel->fdwroutine = input_rel->fdwroutine;
4538 :
4539 : /*
4540 : * Consider computing window functions starting from the existing
4541 : * cheapest-total path (which will likely require a sort) as well as any
4542 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4543 : */
4544 5078 : foreach(lc, input_rel->pathlist)
4545 : {
4546 2700 : Path *path = (Path *) lfirst(lc);
4547 : int presorted_keys;
4548 :
4549 3022 : if (path == input_rel->cheapest_total_path ||
4550 322 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4551 140 : &presorted_keys) ||
4552 140 : presorted_keys > 0)
4553 2586 : create_one_window_path(root,
4554 : window_rel,
4555 : path,
4556 : input_target,
4557 : output_target,
4558 : wflists,
4559 : activeWindows);
4560 : }
4561 :
4562 : /*
4563 : * If there is an FDW that's responsible for all baserels of the query,
4564 : * let it consider adding ForeignPaths.
4565 : */
4566 2378 : if (window_rel->fdwroutine &&
4567 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4568 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4569 : input_rel, window_rel,
4570 : NULL);
4571 :
4572 : /* Let extensions possibly add some more paths */
4573 2378 : if (create_upper_paths_hook)
4574 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4575 : input_rel, window_rel, NULL);
4576 :
4577 : /* Now choose the best path(s) */
4578 2378 : set_cheapest(window_rel);
4579 :
4580 2378 : return window_rel;
4581 : }
4582 :
4583 : /*
4584 : * Stack window-function implementation steps atop the given Path, and
4585 : * add the result to window_rel.
4586 : *
4587 : * window_rel: upperrel to contain result
4588 : * path: input Path to use (must return input_target)
4589 : * input_target: result of make_window_input_target
4590 : * output_target: what the topmost WindowAggPath should return
4591 : * wflists: result of find_window_functions
4592 : * activeWindows: result of select_active_windows
4593 : */
4594 : static void
4595 2586 : create_one_window_path(PlannerInfo *root,
4596 : RelOptInfo *window_rel,
4597 : Path *path,
4598 : PathTarget *input_target,
4599 : PathTarget *output_target,
4600 : WindowFuncLists *wflists,
4601 : List *activeWindows)
4602 : {
4603 : PathTarget *window_target;
4604 : ListCell *l;
4605 2586 : List *topqual = NIL;
4606 :
4607 : /*
4608 : * Since each window clause could require a different sort order, we stack
4609 : * up a WindowAgg node for each clause, with sort steps between them as
4610 : * needed. (We assume that select_active_windows chose a good order for
4611 : * executing the clauses in.)
4612 : *
4613 : * input_target should contain all Vars and Aggs needed for the result.
4614 : * (In some cases we wouldn't need to propagate all of these all the way
4615 : * to the top, since they might only be needed as inputs to WindowFuncs.
4616 : * It's probably not worth trying to optimize that though.) It must also
4617 : * contain all window partitioning and sorting expressions, to ensure
4618 : * they're computed only once at the bottom of the stack (that's critical
4619 : * for volatile functions). As we climb up the stack, we'll add outputs
4620 : * for the WindowFuncs computed at each level.
4621 : */
4622 2586 : window_target = input_target;
4623 :
4624 5340 : foreach(l, activeWindows)
4625 : {
4626 2754 : WindowClause *wc = lfirst_node(WindowClause, l);
4627 : List *window_pathkeys;
4628 2754 : List *runcondition = NIL;
4629 : int presorted_keys;
4630 : bool is_sorted;
4631 : bool topwindow;
4632 : ListCell *lc2;
4633 :
4634 2754 : window_pathkeys = make_pathkeys_for_window(root,
4635 : wc,
4636 : root->processed_tlist);
4637 :
4638 2754 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4639 : path->pathkeys,
4640 : &presorted_keys);
4641 :
4642 : /* Sort if necessary */
4643 2754 : if (!is_sorted)
4644 : {
4645 : /*
4646 : * No presorted keys or incremental sort disabled, just perform a
4647 : * complete sort.
4648 : */
4649 2106 : if (presorted_keys == 0 || !enable_incremental_sort)
4650 2044 : path = (Path *) create_sort_path(root, window_rel,
4651 : path,
4652 : window_pathkeys,
4653 : -1.0);
4654 : else
4655 : {
4656 : /*
4657 : * Since we have presorted keys and incremental sort is
4658 : * enabled, just use incremental sort.
4659 : */
4660 62 : path = (Path *) create_incremental_sort_path(root,
4661 : window_rel,
4662 : path,
4663 : window_pathkeys,
4664 : presorted_keys,
4665 : -1.0);
4666 : }
4667 : }
4668 :
4669 2754 : if (lnext(activeWindows, l))
4670 : {
4671 : /*
4672 : * Add the current WindowFuncs to the output target for this
4673 : * intermediate WindowAggPath. We must copy window_target to
4674 : * avoid changing the previous path's target.
4675 : *
4676 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4677 : * we do need to account for the increase in tlist width.
4678 : */
4679 168 : int64 tuple_width = window_target->width;
4680 :
4681 168 : window_target = copy_pathtarget(window_target);
4682 384 : foreach(lc2, wflists->windowFuncs[wc->winref])
4683 : {
4684 216 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4685 :
4686 216 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4687 216 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4688 : }
4689 168 : window_target->width = clamp_width_est(tuple_width);
4690 : }
4691 : else
4692 : {
4693 : /* Install the goal target in the topmost WindowAgg */
4694 2586 : window_target = output_target;
4695 : }
4696 :
4697 : /* mark the final item in the list as the top-level window */
4698 2754 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4699 :
4700 : /*
4701 : * Collect the WindowFuncRunConditions from each WindowFunc and
4702 : * convert them into OpExprs
4703 : */
4704 6246 : foreach(lc2, wflists->windowFuncs[wc->winref])
4705 : {
4706 : ListCell *lc3;
4707 3492 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4708 :
4709 3672 : foreach(lc3, wfunc->runCondition)
4710 : {
4711 180 : WindowFuncRunCondition *wfuncrc =
4712 : lfirst_node(WindowFuncRunCondition, lc3);
4713 : Expr *opexpr;
4714 : Expr *leftop;
4715 : Expr *rightop;
4716 :
4717 180 : if (wfuncrc->wfunc_left)
4718 : {
4719 162 : leftop = (Expr *) copyObject(wfunc);
4720 162 : rightop = copyObject(wfuncrc->arg);
4721 : }
4722 : else
4723 : {
4724 18 : leftop = copyObject(wfuncrc->arg);
4725 18 : rightop = (Expr *) copyObject(wfunc);
4726 : }
4727 :
4728 180 : opexpr = make_opclause(wfuncrc->opno,
4729 : BOOLOID,
4730 : false,
4731 : leftop,
4732 : rightop,
4733 : InvalidOid,
4734 : wfuncrc->inputcollid);
4735 :
4736 180 : runcondition = lappend(runcondition, opexpr);
4737 :
4738 180 : if (!topwindow)
4739 24 : topqual = lappend(topqual, opexpr);
4740 : }
4741 : }
4742 :
4743 : path = (Path *)
4744 2754 : create_windowagg_path(root, window_rel, path, window_target,
4745 2754 : wflists->windowFuncs[wc->winref],
4746 : runcondition, wc,
4747 : topwindow ? topqual : NIL, topwindow);
4748 : }
4749 :
4750 2586 : add_path(window_rel, path);
4751 2586 : }
4752 :
4753 : /*
4754 : * create_distinct_paths
4755 : *
4756 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4757 : *
4758 : * input_rel: contains the source-data Paths
4759 : * target: the pathtarget for the result Paths to compute
4760 : *
4761 : * Note: input paths should already compute the desired pathtarget, since
4762 : * Sort/Unique won't project anything.
4763 : */
4764 : static RelOptInfo *
4765 2620 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4766 : PathTarget *target)
4767 : {
4768 : RelOptInfo *distinct_rel;
4769 :
4770 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4771 2620 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4772 :
4773 : /*
4774 : * We don't compute anything at this level, so distinct_rel will be
4775 : * parallel-safe if the input rel is parallel-safe. In particular, if
4776 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4777 : * output those expressions, and will not be parallel-safe unless those
4778 : * expressions are parallel-safe.
4779 : */
4780 2620 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4781 :
4782 : /*
4783 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4784 : */
4785 2620 : distinct_rel->serverid = input_rel->serverid;
4786 2620 : distinct_rel->userid = input_rel->userid;
4787 2620 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4788 2620 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4789 :
4790 : /* build distinct paths based on input_rel's pathlist */
4791 2620 : create_final_distinct_paths(root, input_rel, distinct_rel);
4792 :
4793 : /* now build distinct paths based on input_rel's partial_pathlist */
4794 2620 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4795 :
4796 : /* Give a helpful error if we failed to create any paths */
4797 2620 : if (distinct_rel->pathlist == NIL)
4798 0 : ereport(ERROR,
4799 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4800 : errmsg("could not implement DISTINCT"),
4801 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4802 :
4803 : /*
4804 : * If there is an FDW that's responsible for all baserels of the query,
4805 : * let it consider adding ForeignPaths.
4806 : */
4807 2620 : if (distinct_rel->fdwroutine &&
4808 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4809 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4810 : UPPERREL_DISTINCT,
4811 : input_rel,
4812 : distinct_rel,
4813 : NULL);
4814 :
4815 : /* Let extensions possibly add some more paths */
4816 2620 : if (create_upper_paths_hook)
4817 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4818 : distinct_rel, NULL);
4819 :
4820 : /* Now choose the best path(s) */
4821 2620 : set_cheapest(distinct_rel);
4822 :
4823 2620 : return distinct_rel;
4824 : }
4825 :
4826 : /*
4827 : * create_partial_distinct_paths
4828 : *
4829 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4830 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4831 : * paths on top and add a final unique/aggregate path to remove any duplicate
4832 : * produced from combining rows from parallel workers.
4833 : */
4834 : static void
4835 2620 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4836 : RelOptInfo *final_distinct_rel,
4837 : PathTarget *target)
4838 : {
4839 : RelOptInfo *partial_distinct_rel;
4840 : Query *parse;
4841 : List *distinctExprs;
4842 : double numDistinctRows;
4843 : Path *cheapest_partial_path;
4844 : ListCell *lc;
4845 :
4846 : /* nothing to do when there are no partial paths in the input rel */
4847 2620 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4848 2512 : return;
4849 :
4850 108 : parse = root->parse;
4851 :
4852 : /* can't do parallel DISTINCT ON */
4853 108 : if (parse->hasDistinctOn)
4854 0 : return;
4855 :
4856 108 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4857 : NULL);
4858 108 : partial_distinct_rel->reltarget = target;
4859 108 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4860 :
4861 : /*
4862 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4863 : */
4864 108 : partial_distinct_rel->serverid = input_rel->serverid;
4865 108 : partial_distinct_rel->userid = input_rel->userid;
4866 108 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4867 108 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4868 :
4869 108 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4870 :
4871 108 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4872 : parse->targetList);
4873 :
4874 : /* estimate how many distinct rows we'll get from each worker */
4875 108 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4876 : cheapest_partial_path->rows,
4877 : NULL, NULL);
4878 :
4879 : /*
4880 : * Try sorting the cheapest path and incrementally sorting any paths with
4881 : * presorted keys and put a unique paths atop of those. We'll also
4882 : * attempt to reorder the required pathkeys to match the input path's
4883 : * pathkeys as much as possible, in hopes of avoiding a possible need to
4884 : * re-sort.
4885 : */
4886 108 : if (grouping_is_sortable(root->processed_distinctClause))
4887 : {
4888 234 : foreach(lc, input_rel->partial_pathlist)
4889 : {
4890 126 : Path *input_path = (Path *) lfirst(lc);
4891 : Path *sorted_path;
4892 126 : List *useful_pathkeys_list = NIL;
4893 :
4894 : useful_pathkeys_list =
4895 126 : get_useful_pathkeys_for_distinct(root,
4896 : root->distinct_pathkeys,
4897 : input_path->pathkeys);
4898 : Assert(list_length(useful_pathkeys_list) > 0);
4899 :
4900 390 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
4901 : {
4902 138 : sorted_path = make_ordered_path(root,
4903 : partial_distinct_rel,
4904 : input_path,
4905 : cheapest_partial_path,
4906 : useful_pathkeys,
4907 : -1.0);
4908 :
4909 138 : if (sorted_path == NULL)
4910 12 : continue;
4911 :
4912 : /*
4913 : * An empty distinct_pathkeys means all tuples have the same
4914 : * value for the DISTINCT clause. See
4915 : * create_final_distinct_paths()
4916 : */
4917 126 : if (root->distinct_pathkeys == NIL)
4918 : {
4919 : Node *limitCount;
4920 :
4921 6 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
4922 : sizeof(int64),
4923 : Int64GetDatum(1), false,
4924 : true);
4925 :
4926 : /*
4927 : * Apply a LimitPath onto the partial path to restrict the
4928 : * tuples from each worker to 1.
4929 : * create_final_distinct_paths will need to apply an
4930 : * additional LimitPath to restrict this to a single row
4931 : * after the Gather node. If the query already has a
4932 : * LIMIT clause, then we could end up with three Limit
4933 : * nodes in the final plan. Consolidating the top two of
4934 : * these could be done, but does not seem worth troubling
4935 : * over.
4936 : */
4937 6 : add_partial_path(partial_distinct_rel, (Path *)
4938 6 : create_limit_path(root, partial_distinct_rel,
4939 : sorted_path,
4940 : NULL,
4941 : limitCount,
4942 : LIMIT_OPTION_COUNT,
4943 : 0, 1));
4944 : }
4945 : else
4946 : {
4947 120 : add_partial_path(partial_distinct_rel, (Path *)
4948 120 : create_unique_path(root, partial_distinct_rel,
4949 : sorted_path,
4950 120 : list_length(root->distinct_pathkeys),
4951 : numDistinctRows));
4952 : }
4953 : }
4954 : }
4955 : }
4956 :
4957 : /*
4958 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
4959 : * we're not on the hook to ensure we do our best to create at least one
4960 : * path here, we treat enable_hashagg as a hard off-switch rather than the
4961 : * slightly softer variant in create_final_distinct_paths.
4962 : */
4963 108 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
4964 : {
4965 78 : add_partial_path(partial_distinct_rel, (Path *)
4966 78 : create_agg_path(root,
4967 : partial_distinct_rel,
4968 : cheapest_partial_path,
4969 : cheapest_partial_path->pathtarget,
4970 : AGG_HASHED,
4971 : AGGSPLIT_SIMPLE,
4972 : root->processed_distinctClause,
4973 : NIL,
4974 : NULL,
4975 : numDistinctRows));
4976 : }
4977 :
4978 : /*
4979 : * If there is an FDW that's responsible for all baserels of the query,
4980 : * let it consider adding ForeignPaths.
4981 : */
4982 108 : if (partial_distinct_rel->fdwroutine &&
4983 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
4984 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4985 : UPPERREL_PARTIAL_DISTINCT,
4986 : input_rel,
4987 : partial_distinct_rel,
4988 : NULL);
4989 :
4990 : /* Let extensions possibly add some more partial paths */
4991 108 : if (create_upper_paths_hook)
4992 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
4993 : input_rel, partial_distinct_rel, NULL);
4994 :
4995 108 : if (partial_distinct_rel->partial_pathlist != NIL)
4996 : {
4997 108 : generate_useful_gather_paths(root, partial_distinct_rel, true);
4998 108 : set_cheapest(partial_distinct_rel);
4999 :
5000 : /*
5001 : * Finally, create paths to distinctify the final result. This step
5002 : * is needed to remove any duplicates due to combining rows from
5003 : * parallel workers.
5004 : */
5005 108 : create_final_distinct_paths(root, partial_distinct_rel,
5006 : final_distinct_rel);
5007 : }
5008 : }
5009 :
5010 : /*
5011 : * create_final_distinct_paths
5012 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5013 : *
5014 : * input_rel: contains the source-data paths
5015 : * distinct_rel: destination relation for storing created paths
5016 : */
5017 : static RelOptInfo *
5018 2728 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5019 : RelOptInfo *distinct_rel)
5020 : {
5021 2728 : Query *parse = root->parse;
5022 2728 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5023 : double numDistinctRows;
5024 : bool allow_hash;
5025 :
5026 : /* Estimate number of distinct rows there will be */
5027 2728 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5028 2654 : root->hasHavingQual)
5029 : {
5030 : /*
5031 : * If there was grouping or aggregation, use the number of input rows
5032 : * as the estimated number of DISTINCT rows (ie, assume the input is
5033 : * already mostly unique).
5034 : */
5035 74 : numDistinctRows = cheapest_input_path->rows;
5036 : }
5037 : else
5038 : {
5039 : /*
5040 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5041 : */
5042 : List *distinctExprs;
5043 :
5044 2654 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5045 : parse->targetList);
5046 2654 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5047 : cheapest_input_path->rows,
5048 : NULL, NULL);
5049 : }
5050 :
5051 : /*
5052 : * Consider sort-based implementations of DISTINCT, if possible.
5053 : */
5054 2728 : if (grouping_is_sortable(root->processed_distinctClause))
5055 : {
5056 : /*
5057 : * Firstly, if we have any adequately-presorted paths, just stick a
5058 : * Unique node on those. We also, consider doing an explicit sort of
5059 : * the cheapest input path and Unique'ing that. If any paths have
5060 : * presorted keys then we'll create an incremental sort atop of those
5061 : * before adding a unique node on the top. We'll also attempt to
5062 : * reorder the required pathkeys to match the input path's pathkeys as
5063 : * much as possible, in hopes of avoiding a possible need to re-sort.
5064 : *
5065 : * When we have DISTINCT ON, we must sort by the more rigorous of
5066 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5067 : * Also, if we do have to do an explicit sort, we might as well use
5068 : * the more rigorous ordering to avoid a second sort later. (Note
5069 : * that the parser will have ensured that one clause is a prefix of
5070 : * the other.)
5071 : */
5072 : List *needed_pathkeys;
5073 : ListCell *lc;
5074 2722 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5075 :
5076 2970 : if (parse->hasDistinctOn &&
5077 248 : list_length(root->distinct_pathkeys) <
5078 248 : list_length(root->sort_pathkeys))
5079 54 : needed_pathkeys = root->sort_pathkeys;
5080 : else
5081 2668 : needed_pathkeys = root->distinct_pathkeys;
5082 :
5083 6954 : foreach(lc, input_rel->pathlist)
5084 : {
5085 4232 : Path *input_path = (Path *) lfirst(lc);
5086 : Path *sorted_path;
5087 4232 : List *useful_pathkeys_list = NIL;
5088 :
5089 : useful_pathkeys_list =
5090 4232 : get_useful_pathkeys_for_distinct(root,
5091 : needed_pathkeys,
5092 : input_path->pathkeys);
5093 : Assert(list_length(useful_pathkeys_list) > 0);
5094 :
5095 13188 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5096 : {
5097 4724 : sorted_path = make_ordered_path(root,
5098 : distinct_rel,
5099 : input_path,
5100 : cheapest_input_path,
5101 : useful_pathkeys,
5102 : limittuples);
5103 :
5104 4724 : if (sorted_path == NULL)
5105 550 : continue;
5106 :
5107 : /*
5108 : * distinct_pathkeys may have become empty if all of the
5109 : * pathkeys were determined to be redundant. If all of the
5110 : * pathkeys are redundant then each DISTINCT target must only
5111 : * allow a single value, therefore all resulting tuples must
5112 : * be identical (or at least indistinguishable by an equality
5113 : * check). We can uniquify these tuples simply by just taking
5114 : * the first tuple. All we do here is add a path to do "LIMIT
5115 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5116 : * still have a non-NIL sort_pathkeys list, so we must still
5117 : * only do this with paths which are correctly sorted by
5118 : * sort_pathkeys.
5119 : */
5120 4174 : if (root->distinct_pathkeys == NIL)
5121 : {
5122 : Node *limitCount;
5123 :
5124 116 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5125 : sizeof(int64),
5126 : Int64GetDatum(1), false,
5127 : true);
5128 :
5129 : /*
5130 : * If the query already has a LIMIT clause, then we could
5131 : * end up with a duplicate LimitPath in the final plan.
5132 : * That does not seem worth troubling over too much.
5133 : */
5134 116 : add_path(distinct_rel, (Path *)
5135 116 : create_limit_path(root, distinct_rel, sorted_path,
5136 : NULL, limitCount,
5137 : LIMIT_OPTION_COUNT, 0, 1));
5138 : }
5139 : else
5140 : {
5141 4058 : add_path(distinct_rel, (Path *)
5142 4058 : create_unique_path(root, distinct_rel,
5143 : sorted_path,
5144 4058 : list_length(root->distinct_pathkeys),
5145 : numDistinctRows));
5146 : }
5147 : }
5148 : }
5149 : }
5150 :
5151 : /*
5152 : * Consider hash-based implementations of DISTINCT, if possible.
5153 : *
5154 : * If we were not able to make any other types of path, we *must* hash or
5155 : * die trying. If we do have other choices, there are two things that
5156 : * should prevent selection of hashing: if the query uses DISTINCT ON
5157 : * (because it won't really have the expected behavior if we hash), or if
5158 : * enable_hashagg is off.
5159 : *
5160 : * Note: grouping_is_hashable() is much more expensive to check than the
5161 : * other gating conditions, so we want to do it last.
5162 : */
5163 2728 : if (distinct_rel->pathlist == NIL)
5164 6 : allow_hash = true; /* we have no alternatives */
5165 2722 : else if (parse->hasDistinctOn || !enable_hashagg)
5166 398 : allow_hash = false; /* policy-based decision not to hash */
5167 : else
5168 2324 : allow_hash = true; /* default */
5169 :
5170 2728 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5171 : {
5172 : /* Generate hashed aggregate path --- no sort needed */
5173 2330 : add_path(distinct_rel, (Path *)
5174 2330 : create_agg_path(root,
5175 : distinct_rel,
5176 : cheapest_input_path,
5177 : cheapest_input_path->pathtarget,
5178 : AGG_HASHED,
5179 : AGGSPLIT_SIMPLE,
5180 : root->processed_distinctClause,
5181 : NIL,
5182 : NULL,
5183 : numDistinctRows));
5184 : }
5185 :
5186 2728 : return distinct_rel;
5187 : }
5188 :
5189 : /*
5190 : * get_useful_pathkeys_for_distinct
5191 : * Get useful orderings of pathkeys for distinctClause by reordering
5192 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5193 : *
5194 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5195 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5196 : */
5197 : static List *
5198 4358 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5199 : List *path_pathkeys)
5200 : {
5201 4358 : List *useful_pathkeys_list = NIL;
5202 4358 : List *useful_pathkeys = NIL;
5203 :
5204 : /* always include the given 'needed_pathkeys' */
5205 4358 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5206 : needed_pathkeys);
5207 :
5208 4358 : if (!enable_distinct_reordering)
5209 0 : return useful_pathkeys_list;
5210 :
5211 : /*
5212 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5213 : * that match 'needed_pathkeys', but only up to the longest matching
5214 : * prefix.
5215 : *
5216 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5217 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5218 : * the desired behavior.
5219 : */
5220 10630 : foreach_node(PathKey, pathkey, path_pathkeys)
5221 : {
5222 : /*
5223 : * The PathKey nodes are canonical, so they can be checked for
5224 : * equality by simple pointer comparison.
5225 : */
5226 1942 : if (!list_member_ptr(needed_pathkeys, pathkey))
5227 10 : break;
5228 1932 : if (root->parse->hasDistinctOn &&
5229 200 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5230 18 : break;
5231 :
5232 1914 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5233 : }
5234 :
5235 : /* If no match at all, no point in reordering needed_pathkeys */
5236 4358 : if (useful_pathkeys == NIL)
5237 2708 : return useful_pathkeys_list;
5238 :
5239 : /*
5240 : * If not full match, the resulting pathkey list is not useful without
5241 : * incremental sort.
5242 : */
5243 1650 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5244 904 : !enable_incremental_sort)
5245 60 : return useful_pathkeys_list;
5246 :
5247 : /* Append the remaining PathKey nodes in needed_pathkeys */
5248 1590 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5249 : needed_pathkeys);
5250 :
5251 : /*
5252 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5253 : * just drop it.
5254 : */
5255 1590 : if (compare_pathkeys(needed_pathkeys,
5256 : useful_pathkeys) == PATHKEYS_EQUAL)
5257 1086 : return useful_pathkeys_list;
5258 :
5259 504 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5260 : useful_pathkeys);
5261 :
5262 504 : return useful_pathkeys_list;
5263 : }
5264 :
5265 : /*
5266 : * create_ordered_paths
5267 : *
5268 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5269 : *
5270 : * All paths in the result must satisfy the ORDER BY ordering.
5271 : * The only new paths we need consider are an explicit full sort
5272 : * and incremental sort on the cheapest-total existing path.
5273 : *
5274 : * input_rel: contains the source-data Paths
5275 : * target: the output tlist the result Paths must emit
5276 : * limit_tuples: estimated bound on the number of output tuples,
5277 : * or -1 if no LIMIT or couldn't estimate
5278 : *
5279 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5280 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5281 : */
5282 : static RelOptInfo *
5283 73652 : create_ordered_paths(PlannerInfo *root,
5284 : RelOptInfo *input_rel,
5285 : PathTarget *target,
5286 : bool target_parallel_safe,
5287 : double limit_tuples)
5288 : {
5289 73652 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5290 : RelOptInfo *ordered_rel;
5291 : ListCell *lc;
5292 :
5293 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5294 73652 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5295 :
5296 : /*
5297 : * If the input relation is not parallel-safe, then the ordered relation
5298 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5299 : * target list is parallel-safe.
5300 : */
5301 73652 : if (input_rel->consider_parallel && target_parallel_safe)
5302 51222 : ordered_rel->consider_parallel = true;
5303 :
5304 : /*
5305 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5306 : */
5307 73652 : ordered_rel->serverid = input_rel->serverid;
5308 73652 : ordered_rel->userid = input_rel->userid;
5309 73652 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5310 73652 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5311 :
5312 185516 : foreach(lc, input_rel->pathlist)
5313 : {
5314 111864 : Path *input_path = (Path *) lfirst(lc);
5315 : Path *sorted_path;
5316 : bool is_sorted;
5317 : int presorted_keys;
5318 :
5319 111864 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5320 : input_path->pathkeys, &presorted_keys);
5321 :
5322 111864 : if (is_sorted)
5323 41640 : sorted_path = input_path;
5324 : else
5325 : {
5326 : /*
5327 : * Try at least sorting the cheapest path and also try
5328 : * incrementally sorting any path which is partially sorted
5329 : * already (no need to deal with paths which have presorted keys
5330 : * when incremental sort is disabled unless it's the cheapest
5331 : * input path).
5332 : */
5333 70224 : if (input_path != cheapest_input_path &&
5334 5918 : (presorted_keys == 0 || !enable_incremental_sort))
5335 1886 : continue;
5336 :
5337 : /*
5338 : * We've no need to consider both a sort and incremental sort.
5339 : * We'll just do a sort if there are no presorted keys and an
5340 : * incremental sort when there are presorted keys.
5341 : */
5342 68338 : if (presorted_keys == 0 || !enable_incremental_sort)
5343 63682 : sorted_path = (Path *) create_sort_path(root,
5344 : ordered_rel,
5345 : input_path,
5346 : root->sort_pathkeys,
5347 : limit_tuples);
5348 : else
5349 4656 : sorted_path = (Path *) create_incremental_sort_path(root,
5350 : ordered_rel,
5351 : input_path,
5352 : root->sort_pathkeys,
5353 : presorted_keys,
5354 : limit_tuples);
5355 : }
5356 :
5357 : /*
5358 : * If the pathtarget of the result path has different expressions from
5359 : * the target to be applied, a projection step is needed.
5360 : */
5361 109978 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5362 294 : sorted_path = apply_projection_to_path(root, ordered_rel,
5363 : sorted_path, target);
5364 :
5365 109978 : add_path(ordered_rel, sorted_path);
5366 : }
5367 :
5368 : /*
5369 : * generate_gather_paths() will have already generated a simple Gather
5370 : * path for the best parallel path, if any, and the loop above will have
5371 : * considered sorting it. Similarly, generate_gather_paths() will also
5372 : * have generated order-preserving Gather Merge plans which can be used
5373 : * without sorting if they happen to match the sort_pathkeys, and the loop
5374 : * above will have handled those as well. However, there's one more
5375 : * possibility: it may make sense to sort the cheapest partial path or
5376 : * incrementally sort any partial path that is partially sorted according
5377 : * to the required output order and then use Gather Merge.
5378 : */
5379 73652 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5380 51018 : input_rel->partial_pathlist != NIL)
5381 : {
5382 : Path *cheapest_partial_path;
5383 :
5384 2234 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5385 :
5386 4674 : foreach(lc, input_rel->partial_pathlist)
5387 : {
5388 2440 : Path *input_path = (Path *) lfirst(lc);
5389 : Path *sorted_path;
5390 : bool is_sorted;
5391 : int presorted_keys;
5392 : double total_groups;
5393 :
5394 2440 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5395 : input_path->pathkeys,
5396 : &presorted_keys);
5397 :
5398 2440 : if (is_sorted)
5399 182 : continue;
5400 :
5401 : /*
5402 : * Try at least sorting the cheapest path and also try
5403 : * incrementally sorting any path which is partially sorted
5404 : * already (no need to deal with paths which have presorted keys
5405 : * when incremental sort is disabled unless it's the cheapest
5406 : * partial path).
5407 : */
5408 2258 : if (input_path != cheapest_partial_path &&
5409 42 : (presorted_keys == 0 || !enable_incremental_sort))
5410 0 : continue;
5411 :
5412 : /*
5413 : * We've no need to consider both a sort and incremental sort.
5414 : * We'll just do a sort if there are no presorted keys and an
5415 : * incremental sort when there are presorted keys.
5416 : */
5417 2258 : if (presorted_keys == 0 || !enable_incremental_sort)
5418 2198 : sorted_path = (Path *) create_sort_path(root,
5419 : ordered_rel,
5420 : input_path,
5421 : root->sort_pathkeys,
5422 : limit_tuples);
5423 : else
5424 60 : sorted_path = (Path *) create_incremental_sort_path(root,
5425 : ordered_rel,
5426 : input_path,
5427 : root->sort_pathkeys,
5428 : presorted_keys,
5429 : limit_tuples);
5430 2258 : total_groups = compute_gather_rows(sorted_path);
5431 : sorted_path = (Path *)
5432 2258 : create_gather_merge_path(root, ordered_rel,
5433 : sorted_path,
5434 : sorted_path->pathtarget,
5435 : root->sort_pathkeys, NULL,
5436 : &total_groups);
5437 :
5438 : /*
5439 : * If the pathtarget of the result path has different expressions
5440 : * from the target to be applied, a projection step is needed.
5441 : */
5442 2258 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5443 6 : sorted_path = apply_projection_to_path(root, ordered_rel,
5444 : sorted_path, target);
5445 :
5446 2258 : add_path(ordered_rel, sorted_path);
5447 : }
5448 : }
5449 :
5450 : /*
5451 : * If there is an FDW that's responsible for all baserels of the query,
5452 : * let it consider adding ForeignPaths.
5453 : */
5454 73652 : if (ordered_rel->fdwroutine &&
5455 384 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5456 370 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5457 : input_rel, ordered_rel,
5458 : NULL);
5459 :
5460 : /* Let extensions possibly add some more paths */
5461 73652 : if (create_upper_paths_hook)
5462 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5463 : input_rel, ordered_rel, NULL);
5464 :
5465 : /*
5466 : * No need to bother with set_cheapest here; grouping_planner does not
5467 : * need us to do it.
5468 : */
5469 : Assert(ordered_rel->pathlist != NIL);
5470 :
5471 73652 : return ordered_rel;
5472 : }
5473 :
5474 :
5475 : /*
5476 : * make_group_input_target
5477 : * Generate appropriate PathTarget for initial input to grouping nodes.
5478 : *
5479 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5480 : * the query's final targetlist; for example, it certainly can't emit any
5481 : * aggregate function calls. This routine generates the correct target
5482 : * for the scan/join subplan.
5483 : *
5484 : * The query target list passed from the parser already contains entries
5485 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5486 : * for variables used only in HAVING clauses; so we need to add those
5487 : * variables to the subplan target list. Also, we flatten all expressions
5488 : * except GROUP BY items into their component variables; other expressions
5489 : * will be computed by the upper plan nodes rather than by the subplan.
5490 : * For example, given a query like
5491 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5492 : * we want to pass this targetlist to the subplan:
5493 : * a+b,c,d
5494 : * where the a+b target will be used by the Sort/Group steps, and the
5495 : * other targets will be used for computing the final results.
5496 : *
5497 : * 'final_target' is the query's final target list (in PathTarget form)
5498 : *
5499 : * The result is the PathTarget to be computed by the Paths returned from
5500 : * query_planner().
5501 : */
5502 : static PathTarget *
5503 38820 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5504 : {
5505 38820 : Query *parse = root->parse;
5506 : PathTarget *input_target;
5507 : List *non_group_cols;
5508 : List *non_group_vars;
5509 : int i;
5510 : ListCell *lc;
5511 :
5512 : /*
5513 : * We must build a target containing all grouping columns, plus any other
5514 : * Vars mentioned in the query's targetlist and HAVING qual.
5515 : */
5516 38820 : input_target = create_empty_pathtarget();
5517 38820 : non_group_cols = NIL;
5518 :
5519 38820 : i = 0;
5520 94668 : foreach(lc, final_target->exprs)
5521 : {
5522 55848 : Expr *expr = (Expr *) lfirst(lc);
5523 55848 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5524 :
5525 64610 : if (sgref && root->processed_groupClause &&
5526 8762 : get_sortgroupref_clause_noerr(sgref,
5527 : root->processed_groupClause) != NULL)
5528 : {
5529 : /*
5530 : * It's a grouping column, so add it to the input target as-is.
5531 : *
5532 : * Note that the target is logically below the grouping step. So
5533 : * with grouping sets we need to remove the RT index of the
5534 : * grouping step if there is any from the target expression.
5535 : */
5536 7018 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5537 : {
5538 : Assert(root->group_rtindex > 0);
5539 : expr = (Expr *)
5540 1836 : remove_nulling_relids((Node *) expr,
5541 1836 : bms_make_singleton(root->group_rtindex),
5542 : NULL);
5543 : }
5544 7018 : add_column_to_pathtarget(input_target, expr, sgref);
5545 : }
5546 : else
5547 : {
5548 : /*
5549 : * Non-grouping column, so just remember the expression for later
5550 : * call to pull_var_clause.
5551 : */
5552 48830 : non_group_cols = lappend(non_group_cols, expr);
5553 : }
5554 :
5555 55848 : i++;
5556 : }
5557 :
5558 : /*
5559 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5560 : */
5561 38820 : if (parse->havingQual)
5562 882 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5563 :
5564 : /*
5565 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5566 : * add them to the input target if not already present. (A Var used
5567 : * directly as a GROUP BY item will be present already.) Note this
5568 : * includes Vars used in resjunk items, so we are covering the needs of
5569 : * ORDER BY and window specifications. Vars used within Aggrefs and
5570 : * WindowFuncs will be pulled out here, too.
5571 : *
5572 : * Note that the target is logically below the grouping step. So with
5573 : * grouping sets we need to remove the RT index of the grouping step if
5574 : * there is any from the non-group Vars.
5575 : */
5576 38820 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5577 : PVC_RECURSE_AGGREGATES |
5578 : PVC_RECURSE_WINDOWFUNCS |
5579 : PVC_INCLUDE_PLACEHOLDERS);
5580 38820 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5581 : {
5582 : Assert(root->group_rtindex > 0);
5583 : non_group_vars = (List *)
5584 830 : remove_nulling_relids((Node *) non_group_vars,
5585 830 : bms_make_singleton(root->group_rtindex),
5586 : NULL);
5587 : }
5588 38820 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5589 :
5590 : /* clean up cruft */
5591 38820 : list_free(non_group_vars);
5592 38820 : list_free(non_group_cols);
5593 :
5594 : /* XXX this causes some redundant cost calculation ... */
5595 38820 : return set_pathtarget_cost_width(root, input_target);
5596 : }
5597 :
5598 : /*
5599 : * make_partial_grouping_target
5600 : * Generate appropriate PathTarget for output of partial aggregate
5601 : * (or partial grouping, if there are no aggregates) nodes.
5602 : *
5603 : * A partial aggregation node needs to emit all the same aggregates that
5604 : * a regular aggregation node would, plus any aggregates used in HAVING;
5605 : * except that the Aggref nodes should be marked as partial aggregates.
5606 : *
5607 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5608 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5609 : * these would be Vars that are grouped by or used in grouping expressions.)
5610 : *
5611 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5612 : * havingQual represents the HAVING clause.
5613 : */
5614 : static PathTarget *
5615 2200 : make_partial_grouping_target(PlannerInfo *root,
5616 : PathTarget *grouping_target,
5617 : Node *havingQual)
5618 : {
5619 : PathTarget *partial_target;
5620 : List *non_group_cols;
5621 : List *non_group_exprs;
5622 : int i;
5623 : ListCell *lc;
5624 :
5625 2200 : partial_target = create_empty_pathtarget();
5626 2200 : non_group_cols = NIL;
5627 :
5628 2200 : i = 0;
5629 7818 : foreach(lc, grouping_target->exprs)
5630 : {
5631 5618 : Expr *expr = (Expr *) lfirst(lc);
5632 5618 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5633 :
5634 9438 : if (sgref && root->processed_groupClause &&
5635 3820 : get_sortgroupref_clause_noerr(sgref,
5636 : root->processed_groupClause) != NULL)
5637 : {
5638 : /*
5639 : * It's a grouping column, so add it to the partial_target as-is.
5640 : * (This allows the upper agg step to repeat the grouping calcs.)
5641 : */
5642 1906 : add_column_to_pathtarget(partial_target, expr, sgref);
5643 : }
5644 : else
5645 : {
5646 : /*
5647 : * Non-grouping column, so just remember the expression for later
5648 : * call to pull_var_clause.
5649 : */
5650 3712 : non_group_cols = lappend(non_group_cols, expr);
5651 : }
5652 :
5653 5618 : i++;
5654 : }
5655 :
5656 : /*
5657 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5658 : */
5659 2200 : if (havingQual)
5660 824 : non_group_cols = lappend(non_group_cols, havingQual);
5661 :
5662 : /*
5663 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5664 : * non-group cols (plus HAVING), and add them to the partial_target if not
5665 : * already present. (An expression used directly as a GROUP BY item will
5666 : * be present already.) Note this includes Vars used in resjunk items, so
5667 : * we are covering the needs of ORDER BY and window specifications.
5668 : */
5669 2200 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5670 : PVC_INCLUDE_AGGREGATES |
5671 : PVC_RECURSE_WINDOWFUNCS |
5672 : PVC_INCLUDE_PLACEHOLDERS);
5673 :
5674 2200 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5675 :
5676 : /*
5677 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5678 : * are at the top level of the target list, so we can just scan the list
5679 : * rather than recursing through the expression trees.
5680 : */
5681 8390 : foreach(lc, partial_target->exprs)
5682 : {
5683 6190 : Aggref *aggref = (Aggref *) lfirst(lc);
5684 :
5685 6190 : if (IsA(aggref, Aggref))
5686 : {
5687 : Aggref *newaggref;
5688 :
5689 : /*
5690 : * We shouldn't need to copy the substructure of the Aggref node,
5691 : * but flat-copy the node itself to avoid damaging other trees.
5692 : */
5693 4254 : newaggref = makeNode(Aggref);
5694 4254 : memcpy(newaggref, aggref, sizeof(Aggref));
5695 :
5696 : /* For now, assume serialization is required */
5697 4254 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5698 :
5699 4254 : lfirst(lc) = newaggref;
5700 : }
5701 : }
5702 :
5703 : /* clean up cruft */
5704 2200 : list_free(non_group_exprs);
5705 2200 : list_free(non_group_cols);
5706 :
5707 : /* XXX this causes some redundant cost calculation ... */
5708 2200 : return set_pathtarget_cost_width(root, partial_target);
5709 : }
5710 :
5711 : /*
5712 : * mark_partial_aggref
5713 : * Adjust an Aggref to make it represent a partial-aggregation step.
5714 : *
5715 : * The Aggref node is modified in-place; caller must do any copying required.
5716 : */
5717 : void
5718 7066 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5719 : {
5720 : /* aggtranstype should be computed by this point */
5721 : Assert(OidIsValid(agg->aggtranstype));
5722 : /* ... but aggsplit should still be as the parser left it */
5723 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5724 :
5725 : /* Mark the Aggref with the intended partial-aggregation mode */
5726 7066 : agg->aggsplit = aggsplit;
5727 :
5728 : /*
5729 : * Adjust result type if needed. Normally, a partial aggregate returns
5730 : * the aggregate's transition type; but if that's INTERNAL and we're
5731 : * serializing, it returns BYTEA instead.
5732 : */
5733 7066 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5734 : {
5735 5660 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5736 242 : agg->aggtype = BYTEAOID;
5737 : else
5738 5418 : agg->aggtype = agg->aggtranstype;
5739 : }
5740 7066 : }
5741 :
5742 : /*
5743 : * postprocess_setop_tlist
5744 : * Fix up targetlist returned by plan_set_operations().
5745 : *
5746 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5747 : * NOTE: this would not be good enough if we supported resjunk sort keys
5748 : * for results of set operations --- then, we'd need to project a whole
5749 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5750 : * find any resjunk columns in orig_tlist.
5751 : */
5752 : static List *
5753 5976 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5754 : {
5755 : ListCell *l;
5756 5976 : ListCell *orig_tlist_item = list_head(orig_tlist);
5757 :
5758 23118 : foreach(l, new_tlist)
5759 : {
5760 17142 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5761 : TargetEntry *orig_tle;
5762 :
5763 : /* ignore resjunk columns in setop result */
5764 17142 : if (new_tle->resjunk)
5765 0 : continue;
5766 :
5767 : Assert(orig_tlist_item != NULL);
5768 17142 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5769 17142 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5770 17142 : if (orig_tle->resjunk) /* should not happen */
5771 0 : elog(ERROR, "resjunk output columns are not implemented");
5772 : Assert(new_tle->resno == orig_tle->resno);
5773 17142 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5774 : }
5775 5976 : if (orig_tlist_item != NULL)
5776 0 : elog(ERROR, "resjunk output columns are not implemented");
5777 5976 : return new_tlist;
5778 : }
5779 :
5780 : /*
5781 : * optimize_window_clauses
5782 : * Call each WindowFunc's prosupport function to see if we're able to
5783 : * make any adjustments to any of the WindowClause's so that the executor
5784 : * can execute the window functions in a more optimal way.
5785 : *
5786 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5787 : * may allow more things to be done here in the future.
5788 : */
5789 : static void
5790 2378 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5791 : {
5792 2378 : List *windowClause = root->parse->windowClause;
5793 : ListCell *lc;
5794 :
5795 4984 : foreach(lc, windowClause)
5796 : {
5797 2606 : WindowClause *wc = lfirst_node(WindowClause, lc);
5798 : ListCell *lc2;
5799 2606 : int optimizedFrameOptions = 0;
5800 :
5801 : Assert(wc->winref <= wflists->maxWinRef);
5802 :
5803 : /* skip any WindowClauses that have no WindowFuncs */
5804 2606 : if (wflists->windowFuncs[wc->winref] == NIL)
5805 24 : continue;
5806 :
5807 3122 : foreach(lc2, wflists->windowFuncs[wc->winref])
5808 : {
5809 : SupportRequestOptimizeWindowClause req;
5810 : SupportRequestOptimizeWindowClause *res;
5811 2624 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5812 : Oid prosupport;
5813 :
5814 2624 : prosupport = get_func_support(wfunc->winfnoid);
5815 :
5816 : /* Check if there's a support function for 'wfunc' */
5817 2624 : if (!OidIsValid(prosupport))
5818 2084 : break; /* can't optimize this WindowClause */
5819 :
5820 760 : req.type = T_SupportRequestOptimizeWindowClause;
5821 760 : req.window_clause = wc;
5822 760 : req.window_func = wfunc;
5823 760 : req.frameOptions = wc->frameOptions;
5824 :
5825 : /* call the support function */
5826 : res = (SupportRequestOptimizeWindowClause *)
5827 760 : DatumGetPointer(OidFunctionCall1(prosupport,
5828 : PointerGetDatum(&req)));
5829 :
5830 : /*
5831 : * Skip to next WindowClause if the support function does not
5832 : * support this request type.
5833 : */
5834 760 : if (res == NULL)
5835 220 : break;
5836 :
5837 : /*
5838 : * Save these frameOptions for the first WindowFunc for this
5839 : * WindowClause.
5840 : */
5841 540 : if (foreach_current_index(lc2) == 0)
5842 516 : optimizedFrameOptions = res->frameOptions;
5843 :
5844 : /*
5845 : * On subsequent WindowFuncs, if the frameOptions are not the same
5846 : * then we're unable to optimize the frameOptions for this
5847 : * WindowClause.
5848 : */
5849 24 : else if (optimizedFrameOptions != res->frameOptions)
5850 0 : break; /* skip to the next WindowClause, if any */
5851 : }
5852 :
5853 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
5854 2582 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5855 : {
5856 : ListCell *lc3;
5857 :
5858 : /* apply the new frame options */
5859 498 : wc->frameOptions = optimizedFrameOptions;
5860 :
5861 : /*
5862 : * We now check to see if changing the frameOptions has caused
5863 : * this WindowClause to be a duplicate of some other WindowClause.
5864 : * This can only happen if we have multiple WindowClauses, so
5865 : * don't bother if there's only 1.
5866 : */
5867 498 : if (list_length(windowClause) == 1)
5868 408 : continue;
5869 :
5870 : /*
5871 : * Do the duplicate check and reuse the existing WindowClause if
5872 : * we find a duplicate.
5873 : */
5874 228 : foreach(lc3, windowClause)
5875 : {
5876 174 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5877 :
5878 : /* skip over the WindowClause we're currently editing */
5879 174 : if (existing_wc == wc)
5880 54 : continue;
5881 :
5882 : /*
5883 : * Perform the same duplicate check that is done in
5884 : * transformWindowFuncCall.
5885 : */
5886 240 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
5887 120 : equal(wc->orderClause, existing_wc->orderClause) &&
5888 120 : wc->frameOptions == existing_wc->frameOptions &&
5889 72 : equal(wc->startOffset, existing_wc->startOffset) &&
5890 36 : equal(wc->endOffset, existing_wc->endOffset))
5891 : {
5892 : ListCell *lc4;
5893 :
5894 : /*
5895 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
5896 : * This required adjusting each WindowFunc's winref and
5897 : * moving the WindowFuncs in 'wc' to the list of
5898 : * WindowFuncs in 'existing_wc'.
5899 : */
5900 78 : foreach(lc4, wflists->windowFuncs[wc->winref])
5901 : {
5902 42 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
5903 :
5904 42 : wfunc->winref = existing_wc->winref;
5905 : }
5906 :
5907 : /* move list items */
5908 72 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
5909 36 : wflists->windowFuncs[wc->winref]);
5910 36 : wflists->windowFuncs[wc->winref] = NIL;
5911 :
5912 : /*
5913 : * transformWindowFuncCall() should have made sure there
5914 : * are no other duplicates, so we needn't bother looking
5915 : * any further.
5916 : */
5917 36 : break;
5918 : }
5919 : }
5920 : }
5921 : }
5922 2378 : }
5923 :
5924 : /*
5925 : * select_active_windows
5926 : * Create a list of the "active" window clauses (ie, those referenced
5927 : * by non-deleted WindowFuncs) in the order they are to be executed.
5928 : */
5929 : static List *
5930 2378 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5931 : {
5932 2378 : List *windowClause = root->parse->windowClause;
5933 2378 : List *result = NIL;
5934 : ListCell *lc;
5935 2378 : int nActive = 0;
5936 2378 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5937 2378 : * list_length(windowClause));
5938 :
5939 : /* First, construct an array of the active windows */
5940 4984 : foreach(lc, windowClause)
5941 : {
5942 2606 : WindowClause *wc = lfirst_node(WindowClause, lc);
5943 :
5944 : /* It's only active if wflists shows some related WindowFuncs */
5945 : Assert(wc->winref <= wflists->maxWinRef);
5946 2606 : if (wflists->windowFuncs[wc->winref] == NIL)
5947 60 : continue;
5948 :
5949 2546 : actives[nActive].wc = wc; /* original clause */
5950 :
5951 : /*
5952 : * For sorting, we want the list of partition keys followed by the
5953 : * list of sort keys. But pathkeys construction will remove duplicates
5954 : * between the two, so we can as well (even though we can't detect all
5955 : * of the duplicates, since some may come from ECs - that might mean
5956 : * we miss optimization chances here). We must, however, ensure that
5957 : * the order of entries is preserved with respect to the ones we do
5958 : * keep.
5959 : *
5960 : * partitionClause and orderClause had their own duplicates removed in
5961 : * parse analysis, so we're only concerned here with removing
5962 : * orderClause entries that also appear in partitionClause.
5963 : */
5964 5092 : actives[nActive].uniqueOrder =
5965 2546 : list_concat_unique(list_copy(wc->partitionClause),
5966 2546 : wc->orderClause);
5967 2546 : nActive++;
5968 : }
5969 :
5970 : /*
5971 : * Sort active windows by their partitioning/ordering clauses, ignoring
5972 : * any framing clauses, so that the windows that need the same sorting are
5973 : * adjacent in the list. When we come to generate paths, this will avoid
5974 : * inserting additional Sort nodes.
5975 : *
5976 : * This is how we implement a specific requirement from the SQL standard,
5977 : * which says that when two or more windows are order-equivalent (i.e.
5978 : * have matching partition and order clauses, even if their names or
5979 : * framing clauses differ), then all peer rows must be presented in the
5980 : * same order in all of them. If we allowed multiple sort nodes for such
5981 : * cases, we'd risk having the peer rows end up in different orders in
5982 : * equivalent windows due to sort instability. (See General Rule 4 of
5983 : * <window clause> in SQL2008 - SQL2016.)
5984 : *
5985 : * Additionally, if the entire list of clauses of one window is a prefix
5986 : * of another, put first the window with stronger sorting requirements.
5987 : * This way we will first sort for stronger window, and won't have to sort
5988 : * again for the weaker one.
5989 : */
5990 2378 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5991 :
5992 : /* build ordered list of the original WindowClause nodes */
5993 4924 : for (int i = 0; i < nActive; i++)
5994 2546 : result = lappend(result, actives[i].wc);
5995 :
5996 2378 : pfree(actives);
5997 :
5998 2378 : return result;
5999 : }
6000 :
6001 : /*
6002 : * name_active_windows
6003 : * Ensure all active windows have unique names.
6004 : *
6005 : * The parser will have checked that user-assigned window names are unique
6006 : * within the Query. Here we assign made-up names to any unnamed
6007 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6008 : * at parse time, because it'd mess up decompilation of views.)
6009 : *
6010 : * activeWindows: result of select_active_windows
6011 : */
6012 : static void
6013 2378 : name_active_windows(List *activeWindows)
6014 : {
6015 2378 : int next_n = 1;
6016 : char newname[16];
6017 : ListCell *lc;
6018 :
6019 4924 : foreach(lc, activeWindows)
6020 : {
6021 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
6022 :
6023 : /* Nothing to do if it has a name already. */
6024 2546 : if (wc->name)
6025 498 : continue;
6026 :
6027 : /* Select a name not currently present in the list. */
6028 : for (;;)
6029 6 : {
6030 : ListCell *lc2;
6031 :
6032 2054 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6033 4456 : foreach(lc2, activeWindows)
6034 : {
6035 2408 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6036 :
6037 2408 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6038 6 : break; /* matched */
6039 : }
6040 2054 : if (lc2 == NULL)
6041 2048 : break; /* reached the end with no match */
6042 : }
6043 2048 : wc->name = pstrdup(newname);
6044 : }
6045 2378 : }
6046 :
6047 : /*
6048 : * common_prefix_cmp
6049 : * QSort comparison function for WindowClauseSortData
6050 : *
6051 : * Sort the windows by the required sorting clauses. First, compare the sort
6052 : * clauses themselves. Second, if one window's clauses are a prefix of another
6053 : * one's clauses, put the window with more sort clauses first.
6054 : *
6055 : * We purposefully sort by the highest tleSortGroupRef first. Since
6056 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6057 : * and because here we sort the lowest tleSortGroupRefs last, if a
6058 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6059 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6060 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6061 : * reducing the total number of sorts required for the query.
6062 : */
6063 : static int
6064 186 : common_prefix_cmp(const void *a, const void *b)
6065 : {
6066 186 : const WindowClauseSortData *wcsa = a;
6067 186 : const WindowClauseSortData *wcsb = b;
6068 : ListCell *item_a;
6069 : ListCell *item_b;
6070 :
6071 330 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6072 : {
6073 246 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6074 246 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6075 :
6076 246 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6077 102 : return -1;
6078 234 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6079 66 : return 1;
6080 168 : else if (sca->sortop > scb->sortop)
6081 0 : return -1;
6082 168 : else if (sca->sortop < scb->sortop)
6083 24 : return 1;
6084 144 : else if (sca->nulls_first && !scb->nulls_first)
6085 0 : return -1;
6086 144 : else if (!sca->nulls_first && scb->nulls_first)
6087 0 : return 1;
6088 : /* no need to compare eqop, since it is fully determined by sortop */
6089 : }
6090 :
6091 84 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6092 6 : return -1;
6093 78 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6094 30 : return 1;
6095 :
6096 48 : return 0;
6097 : }
6098 :
6099 : /*
6100 : * make_window_input_target
6101 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6102 : *
6103 : * When the query has window functions, this function computes the desired
6104 : * target to be computed by the node just below the first WindowAgg.
6105 : * This tlist must contain all values needed to evaluate the window functions,
6106 : * compute the final target list, and perform any required final sort step.
6107 : * If multiple WindowAggs are needed, each intermediate one adds its window
6108 : * function results onto this base tlist; only the topmost WindowAgg computes
6109 : * the actual desired target list.
6110 : *
6111 : * This function is much like make_group_input_target, though not quite enough
6112 : * like it to share code. As in that function, we flatten most expressions
6113 : * into their component variables. But we do not want to flatten window
6114 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6115 : * evaluations of them, which would be bad (possibly even resulting in
6116 : * inconsistent answers, if they contain volatile functions).
6117 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6118 : * make_group_input_target, because we may no longer have access to the
6119 : * individual Vars in them.
6120 : *
6121 : * Another key difference from make_group_input_target is that we don't
6122 : * flatten Aggref expressions, since those are to be computed below the
6123 : * window functions and just referenced like Vars above that.
6124 : *
6125 : * 'final_target' is the query's final target list (in PathTarget form)
6126 : * 'activeWindows' is the list of active windows previously identified by
6127 : * select_active_windows.
6128 : *
6129 : * The result is the PathTarget to be computed by the plan node immediately
6130 : * below the first WindowAgg node.
6131 : */
6132 : static PathTarget *
6133 2378 : make_window_input_target(PlannerInfo *root,
6134 : PathTarget *final_target,
6135 : List *activeWindows)
6136 : {
6137 : PathTarget *input_target;
6138 : Bitmapset *sgrefs;
6139 : List *flattenable_cols;
6140 : List *flattenable_vars;
6141 : int i;
6142 : ListCell *lc;
6143 :
6144 : Assert(root->parse->hasWindowFuncs);
6145 :
6146 : /*
6147 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6148 : * into a bitmapset for convenient reference below.
6149 : */
6150 2378 : sgrefs = NULL;
6151 4924 : foreach(lc, activeWindows)
6152 : {
6153 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
6154 : ListCell *lc2;
6155 :
6156 3290 : foreach(lc2, wc->partitionClause)
6157 : {
6158 744 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6159 :
6160 744 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6161 : }
6162 4722 : foreach(lc2, wc->orderClause)
6163 : {
6164 2176 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6165 :
6166 2176 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6167 : }
6168 : }
6169 :
6170 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6171 2564 : foreach(lc, root->processed_groupClause)
6172 : {
6173 186 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6174 :
6175 186 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6176 : }
6177 :
6178 : /*
6179 : * Construct a target containing all the non-flattenable targetlist items,
6180 : * and save aside the others for a moment.
6181 : */
6182 2378 : input_target = create_empty_pathtarget();
6183 2378 : flattenable_cols = NIL;
6184 :
6185 2378 : i = 0;
6186 10238 : foreach(lc, final_target->exprs)
6187 : {
6188 7860 : Expr *expr = (Expr *) lfirst(lc);
6189 7860 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6190 :
6191 : /*
6192 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6193 : * that such items can't contain window functions, so it's okay to
6194 : * compute them below the WindowAgg nodes.)
6195 : */
6196 7860 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6197 : {
6198 : /*
6199 : * Don't want to deconstruct this value, so add it to the input
6200 : * target as-is.
6201 : */
6202 2774 : add_column_to_pathtarget(input_target, expr, sgref);
6203 : }
6204 : else
6205 : {
6206 : /*
6207 : * Column is to be flattened, so just remember the expression for
6208 : * later call to pull_var_clause.
6209 : */
6210 5086 : flattenable_cols = lappend(flattenable_cols, expr);
6211 : }
6212 :
6213 7860 : i++;
6214 : }
6215 :
6216 : /*
6217 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6218 : * add them to the input target if not already present. (Some might be
6219 : * there already because they're used directly as window/group clauses.)
6220 : *
6221 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6222 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6223 : * at higher levels. On the other hand, we should recurse into
6224 : * WindowFuncs to make sure their input expressions are available.
6225 : */
6226 2378 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6227 : PVC_INCLUDE_AGGREGATES |
6228 : PVC_RECURSE_WINDOWFUNCS |
6229 : PVC_INCLUDE_PLACEHOLDERS);
6230 2378 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6231 :
6232 : /* clean up cruft */
6233 2378 : list_free(flattenable_vars);
6234 2378 : list_free(flattenable_cols);
6235 :
6236 : /* XXX this causes some redundant cost calculation ... */
6237 2378 : return set_pathtarget_cost_width(root, input_target);
6238 : }
6239 :
6240 : /*
6241 : * make_pathkeys_for_window
6242 : * Create a pathkeys list describing the required input ordering
6243 : * for the given WindowClause.
6244 : *
6245 : * Modifies wc's partitionClause to remove any clauses which are deemed
6246 : * redundant by the pathkey logic.
6247 : *
6248 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6249 : * In the future we might try to implement windowing using hashing, in which
6250 : * case the ordering could be relaxed, but for now we always sort.
6251 : */
6252 : static List *
6253 5132 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6254 : List *tlist)
6255 : {
6256 5132 : List *window_pathkeys = NIL;
6257 :
6258 : /* Throw error if can't sort */
6259 5132 : if (!grouping_is_sortable(wc->partitionClause))
6260 0 : ereport(ERROR,
6261 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6262 : errmsg("could not implement window PARTITION BY"),
6263 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6264 5132 : if (!grouping_is_sortable(wc->orderClause))
6265 0 : ereport(ERROR,
6266 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6267 : errmsg("could not implement window ORDER BY"),
6268 : errdetail("Window ordering columns must be of sortable datatypes.")));
6269 :
6270 : /*
6271 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6272 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6273 : */
6274 5132 : if (wc->partitionClause != NIL)
6275 : {
6276 : bool sortable;
6277 :
6278 1290 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6279 : &wc->partitionClause,
6280 : tlist,
6281 : true,
6282 : false,
6283 : &sortable,
6284 : false);
6285 :
6286 : Assert(sortable);
6287 : }
6288 :
6289 : /*
6290 : * In principle, we could also consider removing redundant ORDER BY items
6291 : * too as doing so does not alter the result of peer row checks done by
6292 : * the executor. However, we must *not* remove the ordering column for
6293 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6294 : * if it's known to be equal to some partitioning column.
6295 : */
6296 5132 : if (wc->orderClause != NIL)
6297 : {
6298 : List *orderby_pathkeys;
6299 :
6300 4276 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6301 : wc->orderClause,
6302 : tlist);
6303 :
6304 : /* Okay, make the combined pathkeys */
6305 4276 : if (window_pathkeys != NIL)
6306 934 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6307 : else
6308 3342 : window_pathkeys = orderby_pathkeys;
6309 : }
6310 :
6311 5132 : return window_pathkeys;
6312 : }
6313 :
6314 : /*
6315 : * make_sort_input_target
6316 : * Generate appropriate PathTarget for initial input to Sort step.
6317 : *
6318 : * If the query has ORDER BY, this function chooses the target to be computed
6319 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6320 : * project) steps. This might or might not be identical to the query's final
6321 : * output target.
6322 : *
6323 : * The main argument for keeping the sort-input tlist the same as the final
6324 : * is that we avoid a separate projection node (which will be needed if
6325 : * they're different, because Sort can't project). However, there are also
6326 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6327 : * a consistent order of evaluation for any volatile functions in the tlist,
6328 : * and if there's also a LIMIT, we can stop the query without ever computing
6329 : * tlist functions for later rows, which is beneficial for both volatile and
6330 : * expensive functions.
6331 : *
6332 : * Our current policy is to postpone volatile expressions till after the sort
6333 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6334 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6335 : * postpone set-returning expressions, because running them beforehand would
6336 : * bloat the sort dataset, and because it might cause unexpected output order
6337 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6338 : * in the tlist should be evaluated at the same plan step, so that they can
6339 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6340 : * mustn't postpone any SRFs. (Note that in principle that policy should
6341 : * probably get applied to the group/window input targetlists too, but we
6342 : * have not done that historically.) Lastly, expensive expressions are
6343 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6344 : * partial evaluation of the query is possible (if neither is true, we expect
6345 : * to have to evaluate the expressions for every row anyway), or if there are
6346 : * any volatile or set-returning expressions (since once we've put in a
6347 : * projection at all, it won't cost any more to postpone more stuff).
6348 : *
6349 : * Another issue that could potentially be considered here is that
6350 : * evaluating tlist expressions could result in data that's either wider
6351 : * or narrower than the input Vars, thus changing the volume of data that
6352 : * has to go through the Sort. However, we usually have only a very bad
6353 : * idea of the output width of any expression more complex than a Var,
6354 : * so for now it seems too risky to try to optimize on that basis.
6355 : *
6356 : * Note that if we do produce a modified sort-input target, and then the
6357 : * query ends up not using an explicit Sort, no particular harm is done:
6358 : * we'll initially use the modified target for the preceding path nodes,
6359 : * but then change them to the final target with apply_projection_to_path.
6360 : * Moreover, in such a case the guarantees about evaluation order of
6361 : * volatile functions still hold, since the rows are sorted already.
6362 : *
6363 : * This function has some things in common with make_group_input_target and
6364 : * make_window_input_target, though the detailed rules for what to do are
6365 : * different. We never flatten/postpone any grouping or ordering columns;
6366 : * those are needed before the sort. If we do flatten a particular
6367 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6368 : * computed earlier.
6369 : *
6370 : * 'final_target' is the query's final target list (in PathTarget form)
6371 : * 'have_postponed_srfs' is an output argument, see below
6372 : *
6373 : * The result is the PathTarget to be computed by the plan node immediately
6374 : * below the Sort step (and the Distinct step, if any). This will be
6375 : * exactly final_target if we decide a projection step wouldn't be helpful.
6376 : *
6377 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6378 : * any set-returning functions to after the Sort.
6379 : */
6380 : static PathTarget *
6381 69806 : make_sort_input_target(PlannerInfo *root,
6382 : PathTarget *final_target,
6383 : bool *have_postponed_srfs)
6384 : {
6385 69806 : Query *parse = root->parse;
6386 : PathTarget *input_target;
6387 : int ncols;
6388 : bool *col_is_srf;
6389 : bool *postpone_col;
6390 : bool have_srf;
6391 : bool have_volatile;
6392 : bool have_expensive;
6393 : bool have_srf_sortcols;
6394 : bool postpone_srfs;
6395 : List *postponable_cols;
6396 : List *postponable_vars;
6397 : int i;
6398 : ListCell *lc;
6399 :
6400 : /* Shouldn't get here unless query has ORDER BY */
6401 : Assert(parse->sortClause);
6402 :
6403 69806 : *have_postponed_srfs = false; /* default result */
6404 :
6405 : /* Inspect tlist and collect per-column information */
6406 69806 : ncols = list_length(final_target->exprs);
6407 69806 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6408 69806 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6409 69806 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6410 :
6411 69806 : i = 0;
6412 421176 : foreach(lc, final_target->exprs)
6413 : {
6414 351370 : Expr *expr = (Expr *) lfirst(lc);
6415 :
6416 : /*
6417 : * If the column has a sortgroupref, assume it has to be evaluated
6418 : * before sorting. Generally such columns would be ORDER BY, GROUP
6419 : * BY, etc targets. One exception is columns that were removed from
6420 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6421 : * only be Vars anyway. There don't seem to be any cases where it
6422 : * would be worth the trouble to double-check.
6423 : */
6424 351370 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6425 : {
6426 : /*
6427 : * Check for SRF or volatile functions. Check the SRF case first
6428 : * because we must know whether we have any postponed SRFs.
6429 : */
6430 253108 : if (parse->hasTargetSRFs &&
6431 216 : expression_returns_set((Node *) expr))
6432 : {
6433 : /* We'll decide below whether these are postponable */
6434 96 : col_is_srf[i] = true;
6435 96 : have_srf = true;
6436 : }
6437 252796 : else if (contain_volatile_functions((Node *) expr))
6438 : {
6439 : /* Unconditionally postpone */
6440 148 : postpone_col[i] = true;
6441 148 : have_volatile = true;
6442 : }
6443 : else
6444 : {
6445 : /*
6446 : * Else check the cost. XXX it's annoying to have to do this
6447 : * when set_pathtarget_cost_width() just did it. Refactor to
6448 : * allow sharing the work?
6449 : */
6450 : QualCost cost;
6451 :
6452 252648 : cost_qual_eval_node(&cost, (Node *) expr, root);
6453 :
6454 : /*
6455 : * We arbitrarily define "expensive" as "more than 10X
6456 : * cpu_operator_cost". Note this will take in any PL function
6457 : * with default cost.
6458 : */
6459 252648 : if (cost.per_tuple > 10 * cpu_operator_cost)
6460 : {
6461 16522 : postpone_col[i] = true;
6462 16522 : have_expensive = true;
6463 : }
6464 : }
6465 : }
6466 : else
6467 : {
6468 : /* For sortgroupref cols, just check if any contain SRFs */
6469 98478 : if (!have_srf_sortcols &&
6470 98788 : parse->hasTargetSRFs &&
6471 310 : expression_returns_set((Node *) expr))
6472 124 : have_srf_sortcols = true;
6473 : }
6474 :
6475 351370 : i++;
6476 : }
6477 :
6478 : /*
6479 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6480 : */
6481 69806 : postpone_srfs = (have_srf && !have_srf_sortcols);
6482 :
6483 : /*
6484 : * If we don't need a post-sort projection, just return final_target.
6485 : */
6486 69806 : if (!(postpone_srfs || have_volatile ||
6487 69602 : (have_expensive &&
6488 9726 : (parse->limitCount || root->tuple_fraction > 0))))
6489 69566 : return final_target;
6490 :
6491 : /*
6492 : * Report whether the post-sort projection will contain set-returning
6493 : * functions. This is important because it affects whether the Sort can
6494 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6495 : * to return.
6496 : */
6497 240 : *have_postponed_srfs = postpone_srfs;
6498 :
6499 : /*
6500 : * Construct the sort-input target, taking all non-postponable columns and
6501 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6502 : * the postponable ones.
6503 : */
6504 240 : input_target = create_empty_pathtarget();
6505 240 : postponable_cols = NIL;
6506 :
6507 240 : i = 0;
6508 1990 : foreach(lc, final_target->exprs)
6509 : {
6510 1750 : Expr *expr = (Expr *) lfirst(lc);
6511 :
6512 1750 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6513 298 : postponable_cols = lappend(postponable_cols, expr);
6514 : else
6515 1452 : add_column_to_pathtarget(input_target, expr,
6516 1452 : get_pathtarget_sortgroupref(final_target, i));
6517 :
6518 1750 : i++;
6519 : }
6520 :
6521 : /*
6522 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6523 : * postponable columns, and add them to the sort-input target if not
6524 : * already present. (Some might be there already.) We mustn't
6525 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6526 : * would be unable to recompute them.
6527 : */
6528 240 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6529 : PVC_INCLUDE_AGGREGATES |
6530 : PVC_INCLUDE_WINDOWFUNCS |
6531 : PVC_INCLUDE_PLACEHOLDERS);
6532 240 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6533 :
6534 : /* clean up cruft */
6535 240 : list_free(postponable_vars);
6536 240 : list_free(postponable_cols);
6537 :
6538 : /* XXX this represents even more redundant cost calculation ... */
6539 240 : return set_pathtarget_cost_width(root, input_target);
6540 : }
6541 :
6542 : /*
6543 : * get_cheapest_fractional_path
6544 : * Find the cheapest path for retrieving a specified fraction of all
6545 : * the tuples expected to be returned by the given relation.
6546 : *
6547 : * Do not consider parameterized paths. If the caller needs a path for upper
6548 : * rel, it can't have parameterized paths. If the caller needs an append
6549 : * subpath, it could become limited by the treatment of similar
6550 : * parameterization of all the subpaths.
6551 : *
6552 : * We interpret tuple_fraction the same way as grouping_planner.
6553 : *
6554 : * We assume set_cheapest() has been run on the given rel.
6555 : */
6556 : Path *
6557 486072 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6558 : {
6559 486072 : Path *best_path = rel->cheapest_total_path;
6560 : ListCell *l;
6561 :
6562 : /* If all tuples will be retrieved, just return the cheapest-total path */
6563 486072 : if (tuple_fraction <= 0.0)
6564 476702 : return best_path;
6565 :
6566 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6567 9370 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6568 3796 : tuple_fraction /= best_path->rows;
6569 :
6570 24556 : foreach(l, rel->pathlist)
6571 : {
6572 15186 : Path *path = (Path *) lfirst(l);
6573 :
6574 15186 : if (path->param_info)
6575 200 : continue;
6576 :
6577 20602 : if (path == rel->cheapest_total_path ||
6578 5616 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6579 14470 : continue;
6580 :
6581 516 : best_path = path;
6582 : }
6583 :
6584 9370 : return best_path;
6585 : }
6586 :
6587 : /*
6588 : * adjust_paths_for_srfs
6589 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6590 : *
6591 : * The executor can only handle set-returning functions that appear at the
6592 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6593 : * that are not at top level, we need to split up the evaluation into multiple
6594 : * plan levels in which each level satisfies this constraint. This function
6595 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6596 : * output tlist to insert appropriate projection steps.
6597 : *
6598 : * The given targets and targets_contain_srfs lists are from
6599 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6600 : * target in targets.
6601 : */
6602 : static void
6603 12308 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6604 : List *targets, List *targets_contain_srfs)
6605 : {
6606 : ListCell *lc;
6607 :
6608 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6609 : Assert(!linitial_int(targets_contain_srfs));
6610 :
6611 : /* If no SRFs appear at this plan level, nothing to do */
6612 12308 : if (list_length(targets) == 1)
6613 634 : return;
6614 :
6615 : /*
6616 : * Stack SRF-evaluation nodes atop each path for the rel.
6617 : *
6618 : * In principle we should re-run set_cheapest() here to identify the
6619 : * cheapest path, but it seems unlikely that adding the same tlist eval
6620 : * costs to all the paths would change that, so we don't bother. Instead,
6621 : * just assume that the cheapest-startup and cheapest-total paths remain
6622 : * so. (There should be no parameterized paths anymore, so we needn't
6623 : * worry about updating cheapest_parameterized_paths.)
6624 : */
6625 23374 : foreach(lc, rel->pathlist)
6626 : {
6627 11700 : Path *subpath = (Path *) lfirst(lc);
6628 11700 : Path *newpath = subpath;
6629 : ListCell *lc1,
6630 : *lc2;
6631 :
6632 : Assert(subpath->param_info == NULL);
6633 36284 : forboth(lc1, targets, lc2, targets_contain_srfs)
6634 : {
6635 24584 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6636 24584 : bool contains_srfs = (bool) lfirst_int(lc2);
6637 :
6638 : /* If this level doesn't contain SRFs, do regular projection */
6639 24584 : if (contains_srfs)
6640 11760 : newpath = (Path *) create_set_projection_path(root,
6641 : rel,
6642 : newpath,
6643 : thistarget);
6644 : else
6645 12824 : newpath = (Path *) apply_projection_to_path(root,
6646 : rel,
6647 : newpath,
6648 : thistarget);
6649 : }
6650 11700 : lfirst(lc) = newpath;
6651 11700 : if (subpath == rel->cheapest_startup_path)
6652 378 : rel->cheapest_startup_path = newpath;
6653 11700 : if (subpath == rel->cheapest_total_path)
6654 378 : rel->cheapest_total_path = newpath;
6655 : }
6656 :
6657 : /* Likewise for partial paths, if any */
6658 11680 : foreach(lc, rel->partial_pathlist)
6659 : {
6660 6 : Path *subpath = (Path *) lfirst(lc);
6661 6 : Path *newpath = subpath;
6662 : ListCell *lc1,
6663 : *lc2;
6664 :
6665 : Assert(subpath->param_info == NULL);
6666 24 : forboth(lc1, targets, lc2, targets_contain_srfs)
6667 : {
6668 18 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6669 18 : bool contains_srfs = (bool) lfirst_int(lc2);
6670 :
6671 : /* If this level doesn't contain SRFs, do regular projection */
6672 18 : if (contains_srfs)
6673 6 : newpath = (Path *) create_set_projection_path(root,
6674 : rel,
6675 : newpath,
6676 : thistarget);
6677 : else
6678 : {
6679 : /* avoid apply_projection_to_path, in case of multiple refs */
6680 12 : newpath = (Path *) create_projection_path(root,
6681 : rel,
6682 : newpath,
6683 : thistarget);
6684 : }
6685 : }
6686 6 : lfirst(lc) = newpath;
6687 : }
6688 : }
6689 :
6690 : /*
6691 : * expression_planner
6692 : * Perform planner's transformations on a standalone expression.
6693 : *
6694 : * Various utility commands need to evaluate expressions that are not part
6695 : * of a plannable query. They can do so using the executor's regular
6696 : * expression-execution machinery, but first the expression has to be fed
6697 : * through here to transform it from parser output to something executable.
6698 : *
6699 : * Currently, we disallow sublinks in standalone expressions, so there's no
6700 : * real "planning" involved here. (That might not always be true though.)
6701 : * What we must do is run eval_const_expressions to ensure that any function
6702 : * calls are converted to positional notation and function default arguments
6703 : * get inserted. The fact that constant subexpressions get simplified is a
6704 : * side-effect that is useful when the expression will get evaluated more than
6705 : * once. Also, we must fix operator function IDs.
6706 : *
6707 : * This does not return any information about dependencies of the expression.
6708 : * Hence callers should use the results only for the duration of the current
6709 : * query. Callers that would like to cache the results for longer should use
6710 : * expression_planner_with_deps, probably via the plancache.
6711 : *
6712 : * Note: this must not make any damaging changes to the passed-in expression
6713 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6714 : * we first do an expression_tree_mutator-based walk, what is returned will
6715 : * be a new node tree.) The result is constructed in the current memory
6716 : * context; beware that this can leak a lot of additional stuff there, too.
6717 : */
6718 : Expr *
6719 240492 : expression_planner(Expr *expr)
6720 : {
6721 : Node *result;
6722 :
6723 : /*
6724 : * Convert named-argument function calls, insert default arguments and
6725 : * simplify constant subexprs
6726 : */
6727 240492 : result = eval_const_expressions(NULL, (Node *) expr);
6728 :
6729 : /* Fill in opfuncid values if missing */
6730 240474 : fix_opfuncids(result);
6731 :
6732 240474 : return (Expr *) result;
6733 : }
6734 :
6735 : /*
6736 : * expression_planner_with_deps
6737 : * Perform planner's transformations on a standalone expression,
6738 : * returning expression dependency information along with the result.
6739 : *
6740 : * This is identical to expression_planner() except that it also returns
6741 : * information about possible dependencies of the expression, ie identities of
6742 : * objects whose definitions affect the result. As in a PlannedStmt, these
6743 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6744 : */
6745 : Expr *
6746 356 : expression_planner_with_deps(Expr *expr,
6747 : List **relationOids,
6748 : List **invalItems)
6749 : {
6750 : Node *result;
6751 : PlannerGlobal glob;
6752 : PlannerInfo root;
6753 :
6754 : /* Make up dummy planner state so we can use setrefs machinery */
6755 8188 : MemSet(&glob, 0, sizeof(glob));
6756 356 : glob.type = T_PlannerGlobal;
6757 356 : glob.relationOids = NIL;
6758 356 : glob.invalItems = NIL;
6759 :
6760 31684 : MemSet(&root, 0, sizeof(root));
6761 356 : root.type = T_PlannerInfo;
6762 356 : root.glob = &glob;
6763 :
6764 : /*
6765 : * Convert named-argument function calls, insert default arguments and
6766 : * simplify constant subexprs. Collect identities of inlined functions
6767 : * and elided domains, too.
6768 : */
6769 356 : result = eval_const_expressions(&root, (Node *) expr);
6770 :
6771 : /* Fill in opfuncid values if missing */
6772 356 : fix_opfuncids(result);
6773 :
6774 : /*
6775 : * Now walk the finished expression to find anything else we ought to
6776 : * record as an expression dependency.
6777 : */
6778 356 : (void) extract_query_dependencies_walker(result, &root);
6779 :
6780 356 : *relationOids = glob.relationOids;
6781 356 : *invalItems = glob.invalItems;
6782 :
6783 356 : return (Expr *) result;
6784 : }
6785 :
6786 :
6787 : /*
6788 : * plan_cluster_use_sort
6789 : * Use the planner to decide how CLUSTER should implement sorting
6790 : *
6791 : * tableOid is the OID of a table to be clustered on its index indexOid
6792 : * (which is already known to be a btree index). Decide whether it's
6793 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6794 : * Return true to use sorting, false to use an indexscan.
6795 : *
6796 : * Note: caller had better already hold some type of lock on the table.
6797 : */
6798 : bool
6799 192 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6800 : {
6801 : PlannerInfo *root;
6802 : Query *query;
6803 : PlannerGlobal *glob;
6804 : RangeTblEntry *rte;
6805 : RelOptInfo *rel;
6806 : IndexOptInfo *indexInfo;
6807 : QualCost indexExprCost;
6808 : Cost comparisonCost;
6809 : Path *seqScanPath;
6810 : Path seqScanAndSortPath;
6811 : IndexPath *indexScanPath;
6812 : ListCell *lc;
6813 :
6814 : /* We can short-circuit the cost comparison if indexscans are disabled */
6815 192 : if (!enable_indexscan)
6816 30 : return true; /* use sort */
6817 :
6818 : /* Set up mostly-dummy planner state */
6819 162 : query = makeNode(Query);
6820 162 : query->commandType = CMD_SELECT;
6821 :
6822 162 : glob = makeNode(PlannerGlobal);
6823 :
6824 162 : root = makeNode(PlannerInfo);
6825 162 : root->parse = query;
6826 162 : root->glob = glob;
6827 162 : root->query_level = 1;
6828 162 : root->planner_cxt = CurrentMemoryContext;
6829 162 : root->wt_param_id = -1;
6830 162 : root->join_domains = list_make1(makeNode(JoinDomain));
6831 :
6832 : /* Build a minimal RTE for the rel */
6833 162 : rte = makeNode(RangeTblEntry);
6834 162 : rte->rtekind = RTE_RELATION;
6835 162 : rte->relid = tableOid;
6836 162 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6837 162 : rte->rellockmode = AccessShareLock;
6838 162 : rte->lateral = false;
6839 162 : rte->inh = false;
6840 162 : rte->inFromCl = true;
6841 162 : query->rtable = list_make1(rte);
6842 162 : addRTEPermissionInfo(&query->rteperminfos, rte);
6843 :
6844 : /* Set up RTE/RelOptInfo arrays */
6845 162 : setup_simple_rel_arrays(root);
6846 :
6847 : /* Build RelOptInfo */
6848 162 : rel = build_simple_rel(root, 1, NULL);
6849 :
6850 : /* Locate IndexOptInfo for the target index */
6851 162 : indexInfo = NULL;
6852 200 : foreach(lc, rel->indexlist)
6853 : {
6854 200 : indexInfo = lfirst_node(IndexOptInfo, lc);
6855 200 : if (indexInfo->indexoid == indexOid)
6856 162 : break;
6857 : }
6858 :
6859 : /*
6860 : * It's possible that get_relation_info did not generate an IndexOptInfo
6861 : * for the desired index; this could happen if it's not yet reached its
6862 : * indcheckxmin usability horizon, or if it's a system index and we're
6863 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6864 : * trust the index contents but use seqscan-and-sort.
6865 : */
6866 162 : if (lc == NULL) /* not in the list? */
6867 0 : return true; /* use sort */
6868 :
6869 : /*
6870 : * Rather than doing all the pushups that would be needed to use
6871 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6872 : */
6873 162 : rel->rows = rel->tuples;
6874 162 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6875 :
6876 162 : root->total_table_pages = rel->pages;
6877 :
6878 : /*
6879 : * Determine eval cost of the index expressions, if any. We need to
6880 : * charge twice that amount for each tuple comparison that happens during
6881 : * the sort, since tuplesort.c will have to re-evaluate the index
6882 : * expressions each time. (XXX that's pretty inefficient...)
6883 : */
6884 162 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6885 162 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6886 :
6887 : /* Estimate the cost of seq scan + sort */
6888 162 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6889 162 : cost_sort(&seqScanAndSortPath, root, NIL,
6890 : seqScanPath->disabled_nodes,
6891 162 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6892 : comparisonCost, maintenance_work_mem, -1.0);
6893 :
6894 : /* Estimate the cost of index scan */
6895 162 : indexScanPath = create_index_path(root, indexInfo,
6896 : NIL, NIL, NIL, NIL,
6897 : ForwardScanDirection, false,
6898 : NULL, 1.0, false);
6899 :
6900 162 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6901 : }
6902 :
6903 : /*
6904 : * plan_create_index_workers
6905 : * Use the planner to decide how many parallel worker processes
6906 : * CREATE INDEX should request for use
6907 : *
6908 : * tableOid is the table on which the index is to be built. indexOid is the
6909 : * OID of an index to be created or reindexed (which must be an index with
6910 : * support for parallel builds - currently btree, GIN, or BRIN).
6911 : *
6912 : * Return value is the number of parallel worker processes to request. It
6913 : * may be unsafe to proceed if this is 0. Note that this does not include the
6914 : * leader participating as a worker (value is always a number of parallel
6915 : * worker processes).
6916 : *
6917 : * Note: caller had better already hold some type of lock on the table and
6918 : * index.
6919 : */
6920 : int
6921 34924 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6922 : {
6923 : PlannerInfo *root;
6924 : Query *query;
6925 : PlannerGlobal *glob;
6926 : RangeTblEntry *rte;
6927 : Relation heap;
6928 : Relation index;
6929 : RelOptInfo *rel;
6930 : int parallel_workers;
6931 : BlockNumber heap_blocks;
6932 : double reltuples;
6933 : double allvisfrac;
6934 :
6935 : /*
6936 : * We don't allow performing parallel operation in standalone backend or
6937 : * when parallelism is disabled.
6938 : */
6939 34924 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
6940 506 : return 0;
6941 :
6942 : /* Set up largely-dummy planner state */
6943 34418 : query = makeNode(Query);
6944 34418 : query->commandType = CMD_SELECT;
6945 :
6946 34418 : glob = makeNode(PlannerGlobal);
6947 :
6948 34418 : root = makeNode(PlannerInfo);
6949 34418 : root->parse = query;
6950 34418 : root->glob = glob;
6951 34418 : root->query_level = 1;
6952 34418 : root->planner_cxt = CurrentMemoryContext;
6953 34418 : root->wt_param_id = -1;
6954 34418 : root->join_domains = list_make1(makeNode(JoinDomain));
6955 :
6956 : /*
6957 : * Build a minimal RTE.
6958 : *
6959 : * Mark the RTE with inh = true. This is a kludge to prevent
6960 : * get_relation_info() from fetching index info, which is necessary
6961 : * because it does not expect that any IndexOptInfo is currently
6962 : * undergoing REINDEX.
6963 : */
6964 34418 : rte = makeNode(RangeTblEntry);
6965 34418 : rte->rtekind = RTE_RELATION;
6966 34418 : rte->relid = tableOid;
6967 34418 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6968 34418 : rte->rellockmode = AccessShareLock;
6969 34418 : rte->lateral = false;
6970 34418 : rte->inh = true;
6971 34418 : rte->inFromCl = true;
6972 34418 : query->rtable = list_make1(rte);
6973 34418 : addRTEPermissionInfo(&query->rteperminfos, rte);
6974 :
6975 : /* Set up RTE/RelOptInfo arrays */
6976 34418 : setup_simple_rel_arrays(root);
6977 :
6978 : /* Build RelOptInfo */
6979 34418 : rel = build_simple_rel(root, 1, NULL);
6980 :
6981 : /* Rels are assumed already locked by the caller */
6982 34418 : heap = table_open(tableOid, NoLock);
6983 34418 : index = index_open(indexOid, NoLock);
6984 :
6985 : /*
6986 : * Determine if it's safe to proceed.
6987 : *
6988 : * Currently, parallel workers can't access the leader's temporary tables.
6989 : * Furthermore, any index predicate or index expressions must be parallel
6990 : * safe.
6991 : */
6992 34418 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6993 32364 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6994 32244 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6995 : {
6996 2174 : parallel_workers = 0;
6997 2174 : goto done;
6998 : }
6999 :
7000 : /*
7001 : * If parallel_workers storage parameter is set for the table, accept that
7002 : * as the number of parallel worker processes to launch (though still cap
7003 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7004 : * consider any other factor when parallel_workers is set. (e.g., memory
7005 : * use by workers.)
7006 : */
7007 32244 : if (rel->rel_parallel_workers != -1)
7008 : {
7009 20 : parallel_workers = Min(rel->rel_parallel_workers,
7010 : max_parallel_maintenance_workers);
7011 20 : goto done;
7012 : }
7013 :
7014 : /*
7015 : * Estimate heap relation size ourselves, since rel->pages cannot be
7016 : * trusted (heap RTE was marked as inheritance parent)
7017 : */
7018 32224 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7019 :
7020 : /*
7021 : * Determine number of workers to scan the heap relation using generic
7022 : * model
7023 : */
7024 32224 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7025 : max_parallel_maintenance_workers);
7026 :
7027 : /*
7028 : * Cap workers based on available maintenance_work_mem as needed.
7029 : *
7030 : * Note that each tuplesort participant receives an even share of the
7031 : * total maintenance_work_mem budget. Aim to leave participants
7032 : * (including the leader as a participant) with no less than 32MB of
7033 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7034 : * immediately past the threshold of being capable of launching a single
7035 : * parallel worker to sort.
7036 : */
7037 32380 : while (parallel_workers > 0 &&
7038 314 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7039 156 : parallel_workers--;
7040 :
7041 32224 : done:
7042 34418 : index_close(index, NoLock);
7043 34418 : table_close(heap, NoLock);
7044 :
7045 34418 : return parallel_workers;
7046 : }
7047 :
7048 : /*
7049 : * add_paths_to_grouping_rel
7050 : *
7051 : * Add non-partial paths to grouping relation.
7052 : */
7053 : static void
7054 39678 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7055 : RelOptInfo *grouped_rel,
7056 : RelOptInfo *partially_grouped_rel,
7057 : const AggClauseCosts *agg_costs,
7058 : grouping_sets_data *gd, double dNumGroups,
7059 : GroupPathExtraData *extra)
7060 : {
7061 39678 : Query *parse = root->parse;
7062 39678 : Path *cheapest_path = input_rel->cheapest_total_path;
7063 : ListCell *lc;
7064 39678 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7065 39678 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7066 39678 : List *havingQual = (List *) extra->havingQual;
7067 39678 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7068 :
7069 39678 : if (can_sort)
7070 : {
7071 : /*
7072 : * Use any available suitably-sorted path as input, and also consider
7073 : * sorting the cheapest-total path and incremental sort on any paths
7074 : * with presorted keys.
7075 : */
7076 82020 : foreach(lc, input_rel->pathlist)
7077 : {
7078 : ListCell *lc2;
7079 42348 : Path *path = (Path *) lfirst(lc);
7080 42348 : Path *path_save = path;
7081 42348 : List *pathkey_orderings = NIL;
7082 :
7083 : /* generate alternative group orderings that might be useful */
7084 42348 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7085 :
7086 : Assert(list_length(pathkey_orderings) > 0);
7087 :
7088 84840 : foreach(lc2, pathkey_orderings)
7089 : {
7090 42492 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7091 :
7092 : /* restore the path (we replace it in the loop) */
7093 42492 : path = path_save;
7094 :
7095 42492 : path = make_ordered_path(root,
7096 : grouped_rel,
7097 : path,
7098 : cheapest_path,
7099 : info->pathkeys,
7100 : -1.0);
7101 42492 : if (path == NULL)
7102 368 : continue;
7103 :
7104 : /* Now decide what to stick atop it */
7105 42124 : if (parse->groupingSets)
7106 : {
7107 938 : consider_groupingsets_paths(root, grouped_rel,
7108 : path, true, can_hash,
7109 : gd, agg_costs, dNumGroups);
7110 : }
7111 41186 : else if (parse->hasAggs)
7112 : {
7113 : /*
7114 : * We have aggregation, possibly with plain GROUP BY. Make
7115 : * an AggPath.
7116 : */
7117 40414 : add_path(grouped_rel, (Path *)
7118 40414 : create_agg_path(root,
7119 : grouped_rel,
7120 : path,
7121 40414 : grouped_rel->reltarget,
7122 40414 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7123 : AGGSPLIT_SIMPLE,
7124 : info->clauses,
7125 : havingQual,
7126 : agg_costs,
7127 : dNumGroups));
7128 : }
7129 772 : else if (parse->groupClause)
7130 : {
7131 : /*
7132 : * We have GROUP BY without aggregation or grouping sets.
7133 : * Make a GroupPath.
7134 : */
7135 772 : add_path(grouped_rel, (Path *)
7136 772 : create_group_path(root,
7137 : grouped_rel,
7138 : path,
7139 : info->clauses,
7140 : havingQual,
7141 : dNumGroups));
7142 : }
7143 : else
7144 : {
7145 : /* Other cases should have been handled above */
7146 : Assert(false);
7147 : }
7148 : }
7149 : }
7150 :
7151 : /*
7152 : * Instead of operating directly on the input relation, we can
7153 : * consider finalizing a partially aggregated path.
7154 : */
7155 39672 : if (partially_grouped_rel != NULL)
7156 : {
7157 3994 : foreach(lc, partially_grouped_rel->pathlist)
7158 : {
7159 : ListCell *lc2;
7160 2412 : Path *path = (Path *) lfirst(lc);
7161 2412 : Path *path_save = path;
7162 2412 : List *pathkey_orderings = NIL;
7163 :
7164 : /* generate alternative group orderings that might be useful */
7165 2412 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7166 :
7167 : Assert(list_length(pathkey_orderings) > 0);
7168 :
7169 : /* process all potentially interesting grouping reorderings */
7170 4824 : foreach(lc2, pathkey_orderings)
7171 : {
7172 2412 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7173 :
7174 : /* restore the path (we replace it in the loop) */
7175 2412 : path = path_save;
7176 :
7177 2412 : path = make_ordered_path(root,
7178 : grouped_rel,
7179 : path,
7180 2412 : partially_grouped_rel->cheapest_total_path,
7181 : info->pathkeys,
7182 : -1.0);
7183 :
7184 2412 : if (path == NULL)
7185 108 : continue;
7186 :
7187 2304 : if (parse->hasAggs)
7188 2056 : add_path(grouped_rel, (Path *)
7189 2056 : create_agg_path(root,
7190 : grouped_rel,
7191 : path,
7192 2056 : grouped_rel->reltarget,
7193 2056 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7194 : AGGSPLIT_FINAL_DESERIAL,
7195 : info->clauses,
7196 : havingQual,
7197 : agg_final_costs,
7198 : dNumGroups));
7199 : else
7200 248 : add_path(grouped_rel, (Path *)
7201 248 : create_group_path(root,
7202 : grouped_rel,
7203 : path,
7204 : info->clauses,
7205 : havingQual,
7206 : dNumGroups));
7207 :
7208 : }
7209 : }
7210 : }
7211 : }
7212 :
7213 39678 : if (can_hash)
7214 : {
7215 4980 : if (parse->groupingSets)
7216 : {
7217 : /*
7218 : * Try for a hash-only groupingsets path over unsorted input.
7219 : */
7220 794 : consider_groupingsets_paths(root, grouped_rel,
7221 : cheapest_path, false, true,
7222 : gd, agg_costs, dNumGroups);
7223 : }
7224 : else
7225 : {
7226 : /*
7227 : * Generate a HashAgg Path. We just need an Agg over the
7228 : * cheapest-total input path, since input order won't matter.
7229 : */
7230 4186 : add_path(grouped_rel, (Path *)
7231 4186 : create_agg_path(root, grouped_rel,
7232 : cheapest_path,
7233 4186 : grouped_rel->reltarget,
7234 : AGG_HASHED,
7235 : AGGSPLIT_SIMPLE,
7236 : root->processed_groupClause,
7237 : havingQual,
7238 : agg_costs,
7239 : dNumGroups));
7240 : }
7241 :
7242 : /*
7243 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7244 : * grouped path, assuming there is one
7245 : */
7246 4980 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7247 : {
7248 784 : Path *path = partially_grouped_rel->cheapest_total_path;
7249 :
7250 784 : add_path(grouped_rel, (Path *)
7251 784 : create_agg_path(root,
7252 : grouped_rel,
7253 : path,
7254 784 : grouped_rel->reltarget,
7255 : AGG_HASHED,
7256 : AGGSPLIT_FINAL_DESERIAL,
7257 : root->processed_groupClause,
7258 : havingQual,
7259 : agg_final_costs,
7260 : dNumGroups));
7261 : }
7262 : }
7263 :
7264 : /*
7265 : * When partitionwise aggregate is used, we might have fully aggregated
7266 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7267 : * consider a path for grouped_rel consisting of a Parallel Append of
7268 : * non-partial paths from each child.
7269 : */
7270 39678 : if (grouped_rel->partial_pathlist != NIL)
7271 162 : gather_grouping_paths(root, grouped_rel);
7272 39678 : }
7273 :
7274 : /*
7275 : * create_partial_grouping_paths
7276 : *
7277 : * Create a new upper relation representing the result of partial aggregation
7278 : * and populate it with appropriate paths. Note that we don't finalize the
7279 : * lists of paths here, so the caller can add additional partial or non-partial
7280 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7281 : * the returned upper relation.
7282 : *
7283 : * All paths for this new upper relation -- both partial and non-partial --
7284 : * have been partially aggregated but require a subsequent FinalizeAggregate
7285 : * step.
7286 : *
7287 : * NB: This function is allowed to return NULL if it determines that there is
7288 : * no real need to create a new RelOptInfo.
7289 : */
7290 : static RelOptInfo *
7291 35570 : create_partial_grouping_paths(PlannerInfo *root,
7292 : RelOptInfo *grouped_rel,
7293 : RelOptInfo *input_rel,
7294 : grouping_sets_data *gd,
7295 : GroupPathExtraData *extra,
7296 : bool force_rel_creation)
7297 : {
7298 35570 : Query *parse = root->parse;
7299 : RelOptInfo *partially_grouped_rel;
7300 35570 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7301 35570 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7302 35570 : Path *cheapest_partial_path = NULL;
7303 35570 : Path *cheapest_total_path = NULL;
7304 35570 : double dNumPartialGroups = 0;
7305 35570 : double dNumPartialPartialGroups = 0;
7306 : ListCell *lc;
7307 35570 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7308 35570 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7309 :
7310 : /*
7311 : * Consider whether we should generate partially aggregated non-partial
7312 : * paths. We can only do this if we have a non-partial path, and only if
7313 : * the parent of the input rel is performing partial partitionwise
7314 : * aggregation. (Note that extra->patype is the type of partitionwise
7315 : * aggregation being used at the parent level, not this level.)
7316 : */
7317 35570 : if (input_rel->pathlist != NIL &&
7318 35570 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7319 618 : cheapest_total_path = input_rel->cheapest_total_path;
7320 :
7321 : /*
7322 : * If parallelism is possible for grouped_rel, then we should consider
7323 : * generating partially-grouped partial paths. However, if the input rel
7324 : * has no partial paths, then we can't.
7325 : */
7326 35570 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7327 1784 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7328 :
7329 : /*
7330 : * If we can't partially aggregate partial paths, and we can't partially
7331 : * aggregate non-partial paths, then don't bother creating the new
7332 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7333 : */
7334 35570 : if (cheapest_total_path == NULL &&
7335 33468 : cheapest_partial_path == NULL &&
7336 33468 : !force_rel_creation)
7337 33370 : return NULL;
7338 :
7339 : /*
7340 : * Build a new upper relation to represent the result of partially
7341 : * aggregating the rows from the input relation.
7342 : */
7343 2200 : partially_grouped_rel = fetch_upper_rel(root,
7344 : UPPERREL_PARTIAL_GROUP_AGG,
7345 : grouped_rel->relids);
7346 2200 : partially_grouped_rel->consider_parallel =
7347 2200 : grouped_rel->consider_parallel;
7348 2200 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7349 2200 : partially_grouped_rel->serverid = grouped_rel->serverid;
7350 2200 : partially_grouped_rel->userid = grouped_rel->userid;
7351 2200 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7352 2200 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7353 :
7354 : /*
7355 : * Build target list for partial aggregate paths. These paths cannot just
7356 : * emit the same tlist as regular aggregate paths, because (1) we must
7357 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7358 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7359 : */
7360 2200 : partially_grouped_rel->reltarget =
7361 2200 : make_partial_grouping_target(root, grouped_rel->reltarget,
7362 : extra->havingQual);
7363 :
7364 2200 : if (!extra->partial_costs_set)
7365 : {
7366 : /*
7367 : * Collect statistics about aggregates for estimating costs of
7368 : * performing aggregation in parallel.
7369 : */
7370 7764 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7371 7764 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7372 1294 : if (parse->hasAggs)
7373 : {
7374 : /* partial phase */
7375 1160 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7376 : agg_partial_costs);
7377 :
7378 : /* final phase */
7379 1160 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7380 : agg_final_costs);
7381 : }
7382 :
7383 1294 : extra->partial_costs_set = true;
7384 : }
7385 :
7386 : /* Estimate number of partial groups. */
7387 2200 : if (cheapest_total_path != NULL)
7388 : dNumPartialGroups =
7389 618 : get_number_of_groups(root,
7390 : cheapest_total_path->rows,
7391 : gd,
7392 : extra->targetList);
7393 2200 : if (cheapest_partial_path != NULL)
7394 : dNumPartialPartialGroups =
7395 1784 : get_number_of_groups(root,
7396 : cheapest_partial_path->rows,
7397 : gd,
7398 : extra->targetList);
7399 :
7400 2200 : if (can_sort && cheapest_total_path != NULL)
7401 : {
7402 : /* This should have been checked previously */
7403 : Assert(parse->hasAggs || parse->groupClause);
7404 :
7405 : /*
7406 : * Use any available suitably-sorted path as input, and also consider
7407 : * sorting the cheapest partial path.
7408 : */
7409 1236 : foreach(lc, input_rel->pathlist)
7410 : {
7411 : ListCell *lc2;
7412 618 : Path *path = (Path *) lfirst(lc);
7413 618 : Path *path_save = path;
7414 618 : List *pathkey_orderings = NIL;
7415 :
7416 : /* generate alternative group orderings that might be useful */
7417 618 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7418 :
7419 : Assert(list_length(pathkey_orderings) > 0);
7420 :
7421 : /* process all potentially interesting grouping reorderings */
7422 1236 : foreach(lc2, pathkey_orderings)
7423 : {
7424 618 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7425 :
7426 : /* restore the path (we replace it in the loop) */
7427 618 : path = path_save;
7428 :
7429 618 : path = make_ordered_path(root,
7430 : partially_grouped_rel,
7431 : path,
7432 : cheapest_total_path,
7433 : info->pathkeys,
7434 : -1.0);
7435 :
7436 618 : if (path == NULL)
7437 0 : continue;
7438 :
7439 618 : if (parse->hasAggs)
7440 546 : add_path(partially_grouped_rel, (Path *)
7441 546 : create_agg_path(root,
7442 : partially_grouped_rel,
7443 : path,
7444 546 : partially_grouped_rel->reltarget,
7445 546 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7446 : AGGSPLIT_INITIAL_SERIAL,
7447 : info->clauses,
7448 : NIL,
7449 : agg_partial_costs,
7450 : dNumPartialGroups));
7451 : else
7452 72 : add_path(partially_grouped_rel, (Path *)
7453 72 : create_group_path(root,
7454 : partially_grouped_rel,
7455 : path,
7456 : info->clauses,
7457 : NIL,
7458 : dNumPartialGroups));
7459 : }
7460 : }
7461 : }
7462 :
7463 2200 : if (can_sort && cheapest_partial_path != NULL)
7464 : {
7465 : /* Similar to above logic, but for partial paths. */
7466 3580 : foreach(lc, input_rel->partial_pathlist)
7467 : {
7468 : ListCell *lc2;
7469 1796 : Path *path = (Path *) lfirst(lc);
7470 1796 : Path *path_save = path;
7471 1796 : List *pathkey_orderings = NIL;
7472 :
7473 : /* generate alternative group orderings that might be useful */
7474 1796 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7475 :
7476 : Assert(list_length(pathkey_orderings) > 0);
7477 :
7478 : /* process all potentially interesting grouping reorderings */
7479 3592 : foreach(lc2, pathkey_orderings)
7480 : {
7481 1796 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7482 :
7483 :
7484 : /* restore the path (we replace it in the loop) */
7485 1796 : path = path_save;
7486 :
7487 1796 : path = make_ordered_path(root,
7488 : partially_grouped_rel,
7489 : path,
7490 : cheapest_partial_path,
7491 : info->pathkeys,
7492 : -1.0);
7493 :
7494 1796 : if (path == NULL)
7495 6 : continue;
7496 :
7497 1790 : if (parse->hasAggs)
7498 1668 : add_partial_path(partially_grouped_rel, (Path *)
7499 1668 : create_agg_path(root,
7500 : partially_grouped_rel,
7501 : path,
7502 1668 : partially_grouped_rel->reltarget,
7503 1668 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7504 : AGGSPLIT_INITIAL_SERIAL,
7505 : info->clauses,
7506 : NIL,
7507 : agg_partial_costs,
7508 : dNumPartialPartialGroups));
7509 : else
7510 122 : add_partial_path(partially_grouped_rel, (Path *)
7511 122 : create_group_path(root,
7512 : partially_grouped_rel,
7513 : path,
7514 : info->clauses,
7515 : NIL,
7516 : dNumPartialPartialGroups));
7517 : }
7518 : }
7519 : }
7520 :
7521 : /*
7522 : * Add a partially-grouped HashAgg Path where possible
7523 : */
7524 2200 : if (can_hash && cheapest_total_path != NULL)
7525 : {
7526 : /* Checked above */
7527 : Assert(parse->hasAggs || parse->groupClause);
7528 :
7529 618 : add_path(partially_grouped_rel, (Path *)
7530 618 : create_agg_path(root,
7531 : partially_grouped_rel,
7532 : cheapest_total_path,
7533 618 : partially_grouped_rel->reltarget,
7534 : AGG_HASHED,
7535 : AGGSPLIT_INITIAL_SERIAL,
7536 : root->processed_groupClause,
7537 : NIL,
7538 : agg_partial_costs,
7539 : dNumPartialGroups));
7540 : }
7541 :
7542 : /*
7543 : * Now add a partially-grouped HashAgg partial Path where possible
7544 : */
7545 2200 : if (can_hash && cheapest_partial_path != NULL)
7546 : {
7547 986 : add_partial_path(partially_grouped_rel, (Path *)
7548 986 : create_agg_path(root,
7549 : partially_grouped_rel,
7550 : cheapest_partial_path,
7551 986 : partially_grouped_rel->reltarget,
7552 : AGG_HASHED,
7553 : AGGSPLIT_INITIAL_SERIAL,
7554 : root->processed_groupClause,
7555 : NIL,
7556 : agg_partial_costs,
7557 : dNumPartialPartialGroups));
7558 : }
7559 :
7560 : /*
7561 : * If there is an FDW that's responsible for all baserels of the query,
7562 : * let it consider adding partially grouped ForeignPaths.
7563 : */
7564 2200 : if (partially_grouped_rel->fdwroutine &&
7565 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7566 : {
7567 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7568 :
7569 6 : fdwroutine->GetForeignUpperPaths(root,
7570 : UPPERREL_PARTIAL_GROUP_AGG,
7571 : input_rel, partially_grouped_rel,
7572 : extra);
7573 : }
7574 :
7575 2200 : return partially_grouped_rel;
7576 : }
7577 :
7578 : /*
7579 : * make_ordered_path
7580 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7581 : * return NULL if it doesn't make sense to generate an ordered path in
7582 : * this case.
7583 : */
7584 : static Path *
7585 52180 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7586 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7587 : {
7588 : bool is_sorted;
7589 : int presorted_keys;
7590 :
7591 52180 : is_sorted = pathkeys_count_contained_in(pathkeys,
7592 : path->pathkeys,
7593 : &presorted_keys);
7594 :
7595 52180 : if (!is_sorted)
7596 : {
7597 : /*
7598 : * Try at least sorting the cheapest path and also try incrementally
7599 : * sorting any path which is partially sorted already (no need to deal
7600 : * with paths which have presorted keys when incremental sort is
7601 : * disabled unless it's the cheapest input path).
7602 : */
7603 12858 : if (path != cheapest_path &&
7604 2052 : (presorted_keys == 0 || !enable_incremental_sort))
7605 1044 : return NULL;
7606 :
7607 : /*
7608 : * We've no need to consider both a sort and incremental sort. We'll
7609 : * just do a sort if there are no presorted keys and an incremental
7610 : * sort when there are presorted keys.
7611 : */
7612 11814 : if (presorted_keys == 0 || !enable_incremental_sort)
7613 10638 : path = (Path *) create_sort_path(root,
7614 : rel,
7615 : path,
7616 : pathkeys,
7617 : limit_tuples);
7618 : else
7619 1176 : path = (Path *) create_incremental_sort_path(root,
7620 : rel,
7621 : path,
7622 : pathkeys,
7623 : presorted_keys,
7624 : limit_tuples);
7625 : }
7626 :
7627 51136 : return path;
7628 : }
7629 :
7630 : /*
7631 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7632 : * grouping relation.
7633 : *
7634 : * generate_useful_gather_paths does most of the work, but we also consider a
7635 : * special case: we could try sorting the data by the group_pathkeys and then
7636 : * applying Gather Merge.
7637 : *
7638 : * NB: This function shouldn't be used for anything other than a grouped or
7639 : * partially grouped relation not only because of the fact that it explicitly
7640 : * references group_pathkeys but we pass "true" as the third argument to
7641 : * generate_useful_gather_paths().
7642 : */
7643 : static void
7644 1646 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7645 : {
7646 : ListCell *lc;
7647 : Path *cheapest_partial_path;
7648 : List *groupby_pathkeys;
7649 :
7650 : /*
7651 : * This occurs after any partial aggregation has taken place, so trim off
7652 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7653 : */
7654 1646 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7655 18 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7656 : root->num_groupby_pathkeys);
7657 : else
7658 1628 : groupby_pathkeys = root->group_pathkeys;
7659 :
7660 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7661 1646 : generate_useful_gather_paths(root, rel, true);
7662 :
7663 1646 : cheapest_partial_path = linitial(rel->partial_pathlist);
7664 :
7665 : /* XXX Shouldn't this also consider the group-key-reordering? */
7666 3898 : foreach(lc, rel->partial_pathlist)
7667 : {
7668 2252 : Path *path = (Path *) lfirst(lc);
7669 : bool is_sorted;
7670 : int presorted_keys;
7671 : double total_groups;
7672 :
7673 2252 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7674 : path->pathkeys,
7675 : &presorted_keys);
7676 :
7677 2252 : if (is_sorted)
7678 1472 : continue;
7679 :
7680 : /*
7681 : * Try at least sorting the cheapest path and also try incrementally
7682 : * sorting any path which is partially sorted already (no need to deal
7683 : * with paths which have presorted keys when incremental sort is
7684 : * disabled unless it's the cheapest input path).
7685 : */
7686 780 : if (path != cheapest_partial_path &&
7687 0 : (presorted_keys == 0 || !enable_incremental_sort))
7688 0 : continue;
7689 :
7690 : /*
7691 : * We've no need to consider both a sort and incremental sort. We'll
7692 : * just do a sort if there are no presorted keys and an incremental
7693 : * sort when there are presorted keys.
7694 : */
7695 780 : if (presorted_keys == 0 || !enable_incremental_sort)
7696 780 : path = (Path *) create_sort_path(root, rel, path,
7697 : groupby_pathkeys,
7698 : -1.0);
7699 : else
7700 0 : path = (Path *) create_incremental_sort_path(root,
7701 : rel,
7702 : path,
7703 : groupby_pathkeys,
7704 : presorted_keys,
7705 : -1.0);
7706 780 : total_groups = compute_gather_rows(path);
7707 : path = (Path *)
7708 780 : create_gather_merge_path(root,
7709 : rel,
7710 : path,
7711 780 : rel->reltarget,
7712 : groupby_pathkeys,
7713 : NULL,
7714 : &total_groups);
7715 :
7716 780 : add_path(rel, path);
7717 : }
7718 1646 : }
7719 :
7720 : /*
7721 : * can_partial_agg
7722 : *
7723 : * Determines whether or not partial grouping and/or aggregation is possible.
7724 : * Returns true when possible, false otherwise.
7725 : */
7726 : static bool
7727 38802 : can_partial_agg(PlannerInfo *root)
7728 : {
7729 38802 : Query *parse = root->parse;
7730 :
7731 38802 : if (!parse->hasAggs && parse->groupClause == NIL)
7732 : {
7733 : /*
7734 : * We don't know how to do parallel aggregation unless we have either
7735 : * some aggregates or a grouping clause.
7736 : */
7737 0 : return false;
7738 : }
7739 38802 : else if (parse->groupingSets)
7740 : {
7741 : /* We don't know how to do grouping sets in parallel. */
7742 872 : return false;
7743 : }
7744 37930 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7745 : {
7746 : /* Insufficient support for partial mode. */
7747 3782 : return false;
7748 : }
7749 :
7750 : /* Everything looks good. */
7751 34148 : return true;
7752 : }
7753 :
7754 : /*
7755 : * apply_scanjoin_target_to_paths
7756 : *
7757 : * Adjust the final scan/join relation, and recursively all of its children,
7758 : * to generate the final scan/join target. It would be more correct to model
7759 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7760 : * for each child relation, but doing it this way is noticeably cheaper.
7761 : * Maybe that problem can be solved at some point, but for now we do this.
7762 : *
7763 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7764 : * the same expressions as the existing reltarget, so we need only insert the
7765 : * appropriate sortgroupref information. By avoiding the creation of
7766 : * projection paths we save effort both immediately and at plan creation time.
7767 : */
7768 : static void
7769 533210 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7770 : RelOptInfo *rel,
7771 : List *scanjoin_targets,
7772 : List *scanjoin_targets_contain_srfs,
7773 : bool scanjoin_target_parallel_safe,
7774 : bool tlist_same_exprs)
7775 : {
7776 533210 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7777 : PathTarget *scanjoin_target;
7778 : ListCell *lc;
7779 :
7780 : /* This recurses, so be paranoid. */
7781 533210 : check_stack_depth();
7782 :
7783 : /*
7784 : * If the rel is partitioned, we want to drop its existing paths and
7785 : * generate new ones. This function would still be correct if we kept the
7786 : * existing paths: we'd modify them to generate the correct target above
7787 : * the partitioning Append, and then they'd compete on cost with paths
7788 : * generating the target below the Append. However, in our current cost
7789 : * model the latter way is always the same or cheaper cost, so modifying
7790 : * the existing paths would just be useless work. Moreover, when the cost
7791 : * is the same, varying roundoff errors might sometimes allow an existing
7792 : * path to be picked, resulting in undesirable cross-platform plan
7793 : * variations. So we drop old paths and thereby force the work to be done
7794 : * below the Append, except in the case of a non-parallel-safe target.
7795 : *
7796 : * Some care is needed, because we have to allow
7797 : * generate_useful_gather_paths to see the old partial paths in the next
7798 : * stanza. Hence, zap the main pathlist here, then allow
7799 : * generate_useful_gather_paths to add path(s) to the main list, and
7800 : * finally zap the partial pathlist.
7801 : */
7802 533210 : if (rel_is_partitioned)
7803 12602 : rel->pathlist = NIL;
7804 :
7805 : /*
7806 : * If the scan/join target is not parallel-safe, partial paths cannot
7807 : * generate it.
7808 : */
7809 533210 : if (!scanjoin_target_parallel_safe)
7810 : {
7811 : /*
7812 : * Since we can't generate the final scan/join target in parallel
7813 : * workers, this is our last opportunity to use any partial paths that
7814 : * exist; so build Gather path(s) that use them and emit whatever the
7815 : * current reltarget is. We don't do this in the case where the
7816 : * target is parallel-safe, since we will be able to generate superior
7817 : * paths by doing it after the final scan/join target has been
7818 : * applied.
7819 : */
7820 78460 : generate_useful_gather_paths(root, rel, false);
7821 :
7822 : /* Can't use parallel query above this level. */
7823 78460 : rel->partial_pathlist = NIL;
7824 78460 : rel->consider_parallel = false;
7825 : }
7826 :
7827 : /* Finish dropping old paths for a partitioned rel, per comment above */
7828 533210 : if (rel_is_partitioned)
7829 12602 : rel->partial_pathlist = NIL;
7830 :
7831 : /* Extract SRF-free scan/join target. */
7832 533210 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7833 :
7834 : /*
7835 : * Apply the SRF-free scan/join target to each existing path.
7836 : *
7837 : * If the tlist exprs are the same, we can just inject the sortgroupref
7838 : * information into the existing pathtargets. Otherwise, replace each
7839 : * path with a projection path that generates the SRF-free scan/join
7840 : * target. This can't change the ordering of paths within rel->pathlist,
7841 : * so we just modify the list in place.
7842 : */
7843 1105184 : foreach(lc, rel->pathlist)
7844 : {
7845 571974 : Path *subpath = (Path *) lfirst(lc);
7846 :
7847 : /* Shouldn't have any parameterized paths anymore */
7848 : Assert(subpath->param_info == NULL);
7849 :
7850 571974 : if (tlist_same_exprs)
7851 201230 : subpath->pathtarget->sortgrouprefs =
7852 201230 : scanjoin_target->sortgrouprefs;
7853 : else
7854 : {
7855 : Path *newpath;
7856 :
7857 370744 : newpath = (Path *) create_projection_path(root, rel, subpath,
7858 : scanjoin_target);
7859 370744 : lfirst(lc) = newpath;
7860 : }
7861 : }
7862 :
7863 : /* Likewise adjust the targets for any partial paths. */
7864 552878 : foreach(lc, rel->partial_pathlist)
7865 : {
7866 19668 : Path *subpath = (Path *) lfirst(lc);
7867 :
7868 : /* Shouldn't have any parameterized paths anymore */
7869 : Assert(subpath->param_info == NULL);
7870 :
7871 19668 : if (tlist_same_exprs)
7872 16028 : subpath->pathtarget->sortgrouprefs =
7873 16028 : scanjoin_target->sortgrouprefs;
7874 : else
7875 : {
7876 : Path *newpath;
7877 :
7878 3640 : newpath = (Path *) create_projection_path(root, rel, subpath,
7879 : scanjoin_target);
7880 3640 : lfirst(lc) = newpath;
7881 : }
7882 : }
7883 :
7884 : /*
7885 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7886 : * atop each existing path. (Note that this function doesn't look at the
7887 : * cheapest-path fields, which is a good thing because they're bogus right
7888 : * now.)
7889 : */
7890 533210 : if (root->parse->hasTargetSRFs)
7891 11674 : adjust_paths_for_srfs(root, rel,
7892 : scanjoin_targets,
7893 : scanjoin_targets_contain_srfs);
7894 :
7895 : /*
7896 : * Update the rel's target to be the final (with SRFs) scan/join target.
7897 : * This now matches the actual output of all the paths, and we might get
7898 : * confused in createplan.c if they don't agree. We must do this now so
7899 : * that any append paths made in the next part will use the correct
7900 : * pathtarget (cf. create_append_path).
7901 : *
7902 : * Note that this is also necessary if GetForeignUpperPaths() gets called
7903 : * on the final scan/join relation or on any of its children, since the
7904 : * FDW might look at the rel's target to create ForeignPaths.
7905 : */
7906 533210 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7907 :
7908 : /*
7909 : * If the relation is partitioned, recursively apply the scan/join target
7910 : * to all partitions, and generate brand-new Append paths in which the
7911 : * scan/join target is computed below the Append rather than above it.
7912 : * Since Append is not projection-capable, that might save a separate
7913 : * Result node, and it also is important for partitionwise aggregate.
7914 : */
7915 533210 : if (rel_is_partitioned)
7916 : {
7917 12602 : List *live_children = NIL;
7918 : int i;
7919 :
7920 : /* Adjust each partition. */
7921 12602 : i = -1;
7922 35584 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
7923 : {
7924 22982 : RelOptInfo *child_rel = rel->part_rels[i];
7925 : AppendRelInfo **appinfos;
7926 : int nappinfos;
7927 22982 : List *child_scanjoin_targets = NIL;
7928 :
7929 : Assert(child_rel != NULL);
7930 :
7931 : /* Dummy children can be ignored. */
7932 22982 : if (IS_DUMMY_REL(child_rel))
7933 42 : continue;
7934 :
7935 : /* Translate scan/join targets for this child. */
7936 22940 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7937 : &nappinfos);
7938 45880 : foreach(lc, scanjoin_targets)
7939 : {
7940 22940 : PathTarget *target = lfirst_node(PathTarget, lc);
7941 :
7942 22940 : target = copy_pathtarget(target);
7943 22940 : target->exprs = (List *)
7944 22940 : adjust_appendrel_attrs(root,
7945 22940 : (Node *) target->exprs,
7946 : nappinfos, appinfos);
7947 22940 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7948 : target);
7949 : }
7950 22940 : pfree(appinfos);
7951 :
7952 : /* Recursion does the real work. */
7953 22940 : apply_scanjoin_target_to_paths(root, child_rel,
7954 : child_scanjoin_targets,
7955 : scanjoin_targets_contain_srfs,
7956 : scanjoin_target_parallel_safe,
7957 : tlist_same_exprs);
7958 :
7959 : /* Save non-dummy children for Append paths. */
7960 22940 : if (!IS_DUMMY_REL(child_rel))
7961 22940 : live_children = lappend(live_children, child_rel);
7962 : }
7963 :
7964 : /* Build new paths for this relation by appending child paths. */
7965 12602 : add_paths_to_append_rel(root, rel, live_children);
7966 : }
7967 :
7968 : /*
7969 : * Consider generating Gather or Gather Merge paths. We must only do this
7970 : * if the relation is parallel safe, and we don't do it for child rels to
7971 : * avoid creating multiple Gather nodes within the same plan. We must do
7972 : * this after all paths have been generated and before set_cheapest, since
7973 : * one of the generated paths may turn out to be the cheapest one.
7974 : */
7975 533210 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7976 169672 : generate_useful_gather_paths(root, rel, false);
7977 :
7978 : /*
7979 : * Reassess which paths are the cheapest, now that we've potentially added
7980 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7981 : * this relation.
7982 : */
7983 533210 : set_cheapest(rel);
7984 533210 : }
7985 :
7986 : /*
7987 : * create_partitionwise_grouping_paths
7988 : *
7989 : * If the partition keys of input relation are part of the GROUP BY clause, all
7990 : * the rows belonging to a given group come from a single partition. This
7991 : * allows aggregation/grouping over a partitioned relation to be broken down
7992 : * into aggregation/grouping on each partition. This should be no worse, and
7993 : * often better, than the normal approach.
7994 : *
7995 : * However, if the GROUP BY clause does not contain all the partition keys,
7996 : * rows from a given group may be spread across multiple partitions. In that
7997 : * case, we perform partial aggregation for each group, append the results,
7998 : * and then finalize aggregation. This is less certain to win than the
7999 : * previous case. It may win if the PartialAggregate stage greatly reduces
8000 : * the number of groups, because fewer rows will pass through the Append node.
8001 : * It may lose if we have lots of small groups.
8002 : */
8003 : static void
8004 562 : create_partitionwise_grouping_paths(PlannerInfo *root,
8005 : RelOptInfo *input_rel,
8006 : RelOptInfo *grouped_rel,
8007 : RelOptInfo *partially_grouped_rel,
8008 : const AggClauseCosts *agg_costs,
8009 : grouping_sets_data *gd,
8010 : PartitionwiseAggregateType patype,
8011 : GroupPathExtraData *extra)
8012 : {
8013 562 : List *grouped_live_children = NIL;
8014 562 : List *partially_grouped_live_children = NIL;
8015 562 : PathTarget *target = grouped_rel->reltarget;
8016 562 : bool partial_grouping_valid = true;
8017 : int i;
8018 :
8019 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8020 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8021 : partially_grouped_rel != NULL);
8022 :
8023 : /* Add paths for partitionwise aggregation/grouping. */
8024 562 : i = -1;
8025 2056 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8026 : {
8027 1494 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8028 : PathTarget *child_target;
8029 : AppendRelInfo **appinfos;
8030 : int nappinfos;
8031 : GroupPathExtraData child_extra;
8032 : RelOptInfo *child_grouped_rel;
8033 : RelOptInfo *child_partially_grouped_rel;
8034 :
8035 : Assert(child_input_rel != NULL);
8036 :
8037 : /* Dummy children can be ignored. */
8038 1494 : if (IS_DUMMY_REL(child_input_rel))
8039 0 : continue;
8040 :
8041 1494 : child_target = copy_pathtarget(target);
8042 :
8043 : /*
8044 : * Copy the given "extra" structure as is and then override the
8045 : * members specific to this child.
8046 : */
8047 1494 : memcpy(&child_extra, extra, sizeof(child_extra));
8048 :
8049 1494 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8050 : &nappinfos);
8051 :
8052 1494 : child_target->exprs = (List *)
8053 1494 : adjust_appendrel_attrs(root,
8054 1494 : (Node *) target->exprs,
8055 : nappinfos, appinfos);
8056 :
8057 : /* Translate havingQual and targetList. */
8058 1494 : child_extra.havingQual = (Node *)
8059 : adjust_appendrel_attrs(root,
8060 : extra->havingQual,
8061 : nappinfos, appinfos);
8062 1494 : child_extra.targetList = (List *)
8063 1494 : adjust_appendrel_attrs(root,
8064 1494 : (Node *) extra->targetList,
8065 : nappinfos, appinfos);
8066 :
8067 : /*
8068 : * extra->patype was the value computed for our parent rel; patype is
8069 : * the value for this relation. For the child, our value is its
8070 : * parent rel's value.
8071 : */
8072 1494 : child_extra.patype = patype;
8073 :
8074 : /*
8075 : * Create grouping relation to hold fully aggregated grouping and/or
8076 : * aggregation paths for the child.
8077 : */
8078 1494 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8079 : child_target,
8080 1494 : extra->target_parallel_safe,
8081 : child_extra.havingQual);
8082 :
8083 : /* Create grouping paths for this child relation. */
8084 1494 : create_ordinary_grouping_paths(root, child_input_rel,
8085 : child_grouped_rel,
8086 : agg_costs, gd, &child_extra,
8087 : &child_partially_grouped_rel);
8088 :
8089 1494 : if (child_partially_grouped_rel)
8090 : {
8091 : partially_grouped_live_children =
8092 906 : lappend(partially_grouped_live_children,
8093 : child_partially_grouped_rel);
8094 : }
8095 : else
8096 588 : partial_grouping_valid = false;
8097 :
8098 1494 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8099 : {
8100 876 : set_cheapest(child_grouped_rel);
8101 876 : grouped_live_children = lappend(grouped_live_children,
8102 : child_grouped_rel);
8103 : }
8104 :
8105 1494 : pfree(appinfos);
8106 : }
8107 :
8108 : /*
8109 : * Try to create append paths for partially grouped children. For full
8110 : * partitionwise aggregation, we might have paths in the partial_pathlist
8111 : * if parallel aggregation is possible. For partial partitionwise
8112 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8113 : *
8114 : * NB: We must have a partially grouped path for every child in order to
8115 : * generate a partially grouped path for this relation.
8116 : */
8117 562 : if (partially_grouped_rel && partial_grouping_valid)
8118 : {
8119 : Assert(partially_grouped_live_children != NIL);
8120 :
8121 350 : add_paths_to_append_rel(root, partially_grouped_rel,
8122 : partially_grouped_live_children);
8123 :
8124 : /*
8125 : * We need call set_cheapest, since the finalization step will use the
8126 : * cheapest path from the rel.
8127 : */
8128 350 : if (partially_grouped_rel->pathlist)
8129 350 : set_cheapest(partially_grouped_rel);
8130 : }
8131 :
8132 : /* If possible, create append paths for fully grouped children. */
8133 562 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8134 : {
8135 : Assert(grouped_live_children != NIL);
8136 :
8137 320 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8138 : }
8139 562 : }
8140 :
8141 : /*
8142 : * group_by_has_partkey
8143 : *
8144 : * Returns true if all the partition keys of the given relation are part of
8145 : * the GROUP BY clauses, including having matching collation, false otherwise.
8146 : */
8147 : static bool
8148 556 : group_by_has_partkey(RelOptInfo *input_rel,
8149 : List *targetList,
8150 : List *groupClause)
8151 : {
8152 556 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8153 556 : int cnt = 0;
8154 : int partnatts;
8155 :
8156 : /* Input relation should be partitioned. */
8157 : Assert(input_rel->part_scheme);
8158 :
8159 : /* Rule out early, if there are no partition keys present. */
8160 556 : if (!input_rel->partexprs)
8161 0 : return false;
8162 :
8163 556 : partnatts = input_rel->part_scheme->partnatts;
8164 :
8165 912 : for (cnt = 0; cnt < partnatts; cnt++)
8166 : {
8167 592 : List *partexprs = input_rel->partexprs[cnt];
8168 : ListCell *lc;
8169 592 : bool found = false;
8170 :
8171 810 : foreach(lc, partexprs)
8172 : {
8173 : ListCell *lg;
8174 586 : Expr *partexpr = lfirst(lc);
8175 586 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8176 :
8177 924 : foreach(lg, groupexprs)
8178 : {
8179 706 : Expr *groupexpr = lfirst(lg);
8180 706 : Oid groupcoll = exprCollation((Node *) groupexpr);
8181 :
8182 : /*
8183 : * Note: we can assume there is at most one RelabelType node;
8184 : * eval_const_expressions() will have simplified if more than
8185 : * one.
8186 : */
8187 706 : if (IsA(groupexpr, RelabelType))
8188 24 : groupexpr = ((RelabelType *) groupexpr)->arg;
8189 :
8190 706 : if (equal(groupexpr, partexpr))
8191 : {
8192 : /*
8193 : * Reject a match if the grouping collation does not match
8194 : * the partitioning collation.
8195 : */
8196 368 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8197 : partcoll != groupcoll)
8198 12 : return false;
8199 :
8200 356 : found = true;
8201 356 : break;
8202 : }
8203 : }
8204 :
8205 574 : if (found)
8206 356 : break;
8207 : }
8208 :
8209 : /*
8210 : * If none of the partition key expressions match with any of the
8211 : * GROUP BY expression, return false.
8212 : */
8213 580 : if (!found)
8214 224 : return false;
8215 : }
8216 :
8217 320 : return true;
8218 : }
8219 :
8220 : /*
8221 : * generate_setop_child_grouplist
8222 : * Build a SortGroupClause list defining the sort/grouping properties
8223 : * of the child of a set operation.
8224 : *
8225 : * This is similar to generate_setop_grouplist() but differs as the setop
8226 : * child query's targetlist entries may already have a tleSortGroupRef
8227 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8228 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8229 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8230 : * any of the columns in the targetlist don't match to the setop's colTypes
8231 : * then we return an empty list. This may leave some TLEs with unreferenced
8232 : * ressortgroupref markings, but that's harmless.
8233 : */
8234 : static List *
8235 12274 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8236 : {
8237 12274 : List *grouplist = copyObject(op->groupClauses);
8238 : ListCell *lg;
8239 : ListCell *lt;
8240 : ListCell *ct;
8241 :
8242 12274 : lg = list_head(grouplist);
8243 12274 : ct = list_head(op->colTypes);
8244 47734 : foreach(lt, targetlist)
8245 : {
8246 35874 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8247 : SortGroupClause *sgc;
8248 : Oid coltype;
8249 :
8250 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8251 35874 : if (tle->resjunk)
8252 0 : continue;
8253 :
8254 : /*
8255 : * We expect every non-resjunk target to have a SortGroupClause and
8256 : * colTypes.
8257 : */
8258 : Assert(lg != NULL);
8259 : Assert(ct != NULL);
8260 35874 : sgc = (SortGroupClause *) lfirst(lg);
8261 35874 : coltype = lfirst_oid(ct);
8262 :
8263 : /* reject if target type isn't the same as the setop target type */
8264 35874 : if (coltype != exprType((Node *) tle->expr))
8265 414 : return NIL;
8266 :
8267 35460 : lg = lnext(grouplist, lg);
8268 35460 : ct = lnext(op->colTypes, ct);
8269 :
8270 : /* assign a tleSortGroupRef, or reuse the existing one */
8271 35460 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8272 : }
8273 :
8274 : Assert(lg == NULL);
8275 : Assert(ct == NULL);
8276 :
8277 11860 : return grouplist;
8278 : }
8279 :
8280 : /*
8281 : * create_unique_paths
8282 : * Build a new RelOptInfo containing Paths that represent elimination of
8283 : * distinct rows from the input data. Distinct-ness is defined according to
8284 : * the needs of the semijoin represented by sjinfo. If it is not possible
8285 : * to identify how to make the data unique, NULL is returned.
8286 : *
8287 : * If used at all, this is likely to be called repeatedly on the same rel,
8288 : * so we cache the result.
8289 : */
8290 : RelOptInfo *
8291 8788 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8292 : {
8293 : RelOptInfo *unique_rel;
8294 8788 : List *sortPathkeys = NIL;
8295 8788 : List *groupClause = NIL;
8296 : MemoryContext oldcontext;
8297 :
8298 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8299 : Assert(sjinfo->jointype == JOIN_SEMI);
8300 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8301 :
8302 : /* If result already cached, return it */
8303 8788 : if (rel->unique_rel)
8304 1820 : return rel->unique_rel;
8305 :
8306 : /* If it's not possible to unique-ify, return NULL */
8307 6968 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8308 132 : return NULL;
8309 :
8310 : /*
8311 : * Punt if this is a child relation and we failed to build a unique-ified
8312 : * relation for its parent. This can happen if all the RHS columns were
8313 : * found to be equated to constants when unique-ifying the parent table,
8314 : * leaving no columns to unique-ify.
8315 : */
8316 6836 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8317 12 : return NULL;
8318 :
8319 : /*
8320 : * When called during GEQO join planning, we are in a short-lived memory
8321 : * context. We must make sure that the unique rel and any subsidiary data
8322 : * structures created for a baserel survive the GEQO cycle, else the
8323 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8324 : * are creating those for a joinrel during GEQO, we don't want them to
8325 : * clutter the main planning context. Upshot is that the best solution is
8326 : * to explicitly allocate memory in the same context the given RelOptInfo
8327 : * is in.
8328 : */
8329 6824 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8330 :
8331 6824 : unique_rel = makeNode(RelOptInfo);
8332 6824 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8333 :
8334 : /*
8335 : * clear path info
8336 : */
8337 6824 : unique_rel->pathlist = NIL;
8338 6824 : unique_rel->ppilist = NIL;
8339 6824 : unique_rel->partial_pathlist = NIL;
8340 6824 : unique_rel->cheapest_startup_path = NULL;
8341 6824 : unique_rel->cheapest_total_path = NULL;
8342 6824 : unique_rel->cheapest_parameterized_paths = NIL;
8343 :
8344 : /*
8345 : * Build the target list for the unique rel. We also build the pathkeys
8346 : * that represent the ordering requirements for the sort-based
8347 : * implementation, and the list of SortGroupClause nodes that represent
8348 : * the columns to be grouped on for the hash-based implementation.
8349 : *
8350 : * For a child rel, we can construct these fields from those of its
8351 : * parent.
8352 : */
8353 6824 : if (IS_OTHER_REL(rel))
8354 432 : {
8355 : PathTarget *child_unique_target;
8356 : PathTarget *parent_unique_target;
8357 :
8358 432 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8359 :
8360 432 : child_unique_target = copy_pathtarget(parent_unique_target);
8361 :
8362 : /* Translate the target expressions */
8363 432 : child_unique_target->exprs = (List *)
8364 432 : adjust_appendrel_attrs_multilevel(root,
8365 432 : (Node *) parent_unique_target->exprs,
8366 : rel,
8367 432 : rel->top_parent);
8368 :
8369 432 : unique_rel->reltarget = child_unique_target;
8370 :
8371 432 : sortPathkeys = rel->top_parent->unique_pathkeys;
8372 432 : groupClause = rel->top_parent->unique_groupclause;
8373 : }
8374 : else
8375 : {
8376 : List *newtlist;
8377 : int nextresno;
8378 6392 : List *sortList = NIL;
8379 : ListCell *lc1;
8380 : ListCell *lc2;
8381 :
8382 : /*
8383 : * The values we are supposed to unique-ify may be expressions in the
8384 : * variables of the input rel's targetlist. We have to add any such
8385 : * expressions to the unique rel's targetlist.
8386 : *
8387 : * To complicate matters, some of the values to be unique-ified may be
8388 : * known redundant by the EquivalenceClass machinery (e.g., because
8389 : * they have been equated to constants). There is no need to compare
8390 : * such values during unique-ification, and indeed we had better not
8391 : * try because the Vars involved may not have propagated as high as
8392 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8393 : * detect such cases, which is a tad inefficient but it doesn't seem
8394 : * worth building specialized infrastructure for this.
8395 : */
8396 6392 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8397 6392 : nextresno = list_length(newtlist) + 1;
8398 :
8399 13018 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8400 : {
8401 6626 : Expr *uniqexpr = lfirst(lc1);
8402 6626 : Oid in_oper = lfirst_oid(lc2);
8403 : Oid sortop;
8404 : TargetEntry *tle;
8405 6626 : bool made_tle = false;
8406 :
8407 6626 : tle = tlist_member(uniqexpr, newtlist);
8408 6626 : if (!tle)
8409 : {
8410 3196 : tle = makeTargetEntry((Expr *) uniqexpr,
8411 : nextresno,
8412 : NULL,
8413 : false);
8414 3196 : newtlist = lappend(newtlist, tle);
8415 3196 : nextresno++;
8416 3196 : made_tle = true;
8417 : }
8418 :
8419 : /*
8420 : * Try to build an ORDER BY list to sort the input compatibly. We
8421 : * do this for each sortable clause even when the clauses are not
8422 : * all sortable, so that we can detect clauses that are redundant
8423 : * according to the pathkey machinery.
8424 : */
8425 6626 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8426 6626 : if (OidIsValid(sortop))
8427 : {
8428 : Oid eqop;
8429 : SortGroupClause *sortcl;
8430 :
8431 : /*
8432 : * The Unique node will need equality operators. Normally
8433 : * these are the same as the IN clause operators, but if those
8434 : * are cross-type operators then the equality operators are
8435 : * the ones for the IN clause operators' RHS datatype.
8436 : */
8437 6626 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8438 6626 : if (!OidIsValid(eqop)) /* shouldn't happen */
8439 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8440 : sortop);
8441 :
8442 6626 : sortcl = makeNode(SortGroupClause);
8443 6626 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8444 6626 : sortcl->eqop = eqop;
8445 6626 : sortcl->sortop = sortop;
8446 6626 : sortcl->reverse_sort = false;
8447 6626 : sortcl->nulls_first = false;
8448 6626 : sortcl->hashable = false; /* no need to make this accurate */
8449 6626 : sortList = lappend(sortList, sortcl);
8450 :
8451 : /*
8452 : * At each step, convert the SortGroupClause list to pathkey
8453 : * form. If the just-added SortGroupClause is redundant, the
8454 : * result will be shorter than the SortGroupClause list.
8455 : */
8456 6626 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8457 : newtlist);
8458 6626 : if (list_length(sortPathkeys) != list_length(sortList))
8459 : {
8460 : /* Drop the redundant SortGroupClause */
8461 2052 : sortList = list_delete_last(sortList);
8462 : Assert(list_length(sortPathkeys) == list_length(sortList));
8463 : /* Undo tlist addition, if we made one */
8464 2052 : if (made_tle)
8465 : {
8466 12 : newtlist = list_delete_last(newtlist);
8467 12 : nextresno--;
8468 : }
8469 : /* We need not consider this clause for hashing, either */
8470 2052 : continue;
8471 : }
8472 : }
8473 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8474 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8475 : in_oper);
8476 :
8477 4574 : if (sjinfo->semi_can_hash)
8478 : {
8479 : /* Create a GROUP BY list for the Agg node to use */
8480 : Oid eq_oper;
8481 : SortGroupClause *groupcl;
8482 :
8483 : /*
8484 : * Get the hashable equality operators for the Agg node to
8485 : * use. Normally these are the same as the IN clause
8486 : * operators, but if those are cross-type operators then the
8487 : * equality operators are the ones for the IN clause
8488 : * operators' RHS datatype.
8489 : */
8490 4574 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8491 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8492 : in_oper);
8493 :
8494 4574 : groupcl = makeNode(SortGroupClause);
8495 4574 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8496 4574 : groupcl->eqop = eq_oper;
8497 4574 : groupcl->sortop = sortop;
8498 4574 : groupcl->reverse_sort = false;
8499 4574 : groupcl->nulls_first = false;
8500 4574 : groupcl->hashable = true;
8501 4574 : groupClause = lappend(groupClause, groupcl);
8502 : }
8503 : }
8504 :
8505 : /*
8506 : * Done building the sortPathkeys and groupClause. But the
8507 : * sortPathkeys are bogus if not all the clauses were sortable.
8508 : */
8509 6392 : if (!sjinfo->semi_can_btree)
8510 0 : sortPathkeys = NIL;
8511 :
8512 : /*
8513 : * It can happen that all the RHS columns are equated to constants.
8514 : * We'd have to do something special to unique-ify in that case, and
8515 : * it's such an unlikely-in-the-real-world case that it's not worth
8516 : * the effort. So just punt if we found no columns to unique-ify.
8517 : */
8518 6392 : if (sortPathkeys == NIL && groupClause == NIL)
8519 : {
8520 1950 : MemoryContextSwitchTo(oldcontext);
8521 1950 : return NULL;
8522 : }
8523 :
8524 : /* Convert the required targetlist back to PathTarget form */
8525 4442 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8526 : }
8527 :
8528 : /* build unique paths based on input rel's pathlist */
8529 4874 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8530 : sjinfo, unique_rel);
8531 :
8532 : /* build unique paths based on input rel's partial_pathlist */
8533 4874 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8534 : sjinfo, unique_rel);
8535 :
8536 : /* Now choose the best path(s) */
8537 4874 : set_cheapest(unique_rel);
8538 :
8539 : /*
8540 : * There shouldn't be any partial paths for the unique relation;
8541 : * otherwise, we won't be able to properly guarantee uniqueness.
8542 : */
8543 : Assert(unique_rel->partial_pathlist == NIL);
8544 :
8545 : /* Cache the result */
8546 4874 : rel->unique_rel = unique_rel;
8547 4874 : rel->unique_pathkeys = sortPathkeys;
8548 4874 : rel->unique_groupclause = groupClause;
8549 :
8550 4874 : MemoryContextSwitchTo(oldcontext);
8551 :
8552 4874 : return unique_rel;
8553 : }
8554 :
8555 : /*
8556 : * create_final_unique_paths
8557 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8558 : */
8559 : static void
8560 8520 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8561 : List *sortPathkeys, List *groupClause,
8562 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8563 : {
8564 8520 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8565 :
8566 : /* Estimate number of output rows */
8567 8520 : unique_rel->rows = estimate_num_groups(root,
8568 : sjinfo->semi_rhs_exprs,
8569 : cheapest_input_path->rows,
8570 : NULL,
8571 : NULL);
8572 :
8573 : /* Consider sort-based implementations, if possible. */
8574 8520 : if (sjinfo->semi_can_btree)
8575 : {
8576 : ListCell *lc;
8577 :
8578 : /*
8579 : * Use any available suitably-sorted path as input, and also consider
8580 : * sorting the cheapest-total path and incremental sort on any paths
8581 : * with presorted keys.
8582 : *
8583 : * To save planning time, we ignore parameterized input paths unless
8584 : * they are the cheapest-total path.
8585 : */
8586 18596 : foreach(lc, input_rel->pathlist)
8587 : {
8588 10076 : Path *input_path = (Path *) lfirst(lc);
8589 : Path *path;
8590 : bool is_sorted;
8591 : int presorted_keys;
8592 :
8593 : /*
8594 : * Ignore parameterized paths that are not the cheapest-total
8595 : * path.
8596 : */
8597 10076 : if (input_path->param_info &&
8598 : input_path != cheapest_input_path)
8599 916 : continue;
8600 :
8601 9204 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8602 : input_path->pathkeys,
8603 : &presorted_keys);
8604 :
8605 : /*
8606 : * Ignore paths that are not suitably or partially sorted, unless
8607 : * they are the cheapest total path (no need to deal with paths
8608 : * which have presorted keys when incremental sort is disabled).
8609 : */
8610 9204 : if (!is_sorted && input_path != cheapest_input_path &&
8611 92 : (presorted_keys == 0 || !enable_incremental_sort))
8612 44 : continue;
8613 :
8614 : /*
8615 : * Make a separate ProjectionPath in case we need a Result node.
8616 : */
8617 9160 : path = (Path *) create_projection_path(root,
8618 : unique_rel,
8619 : input_path,
8620 9160 : unique_rel->reltarget);
8621 :
8622 9160 : if (!is_sorted)
8623 : {
8624 : /*
8625 : * We've no need to consider both a sort and incremental sort.
8626 : * We'll just do a sort if there are no presorted keys and an
8627 : * incremental sort when there are presorted keys.
8628 : */
8629 4854 : if (presorted_keys == 0 || !enable_incremental_sort)
8630 4806 : path = (Path *) create_sort_path(root,
8631 : unique_rel,
8632 : path,
8633 : sortPathkeys,
8634 : -1.0);
8635 : else
8636 48 : path = (Path *) create_incremental_sort_path(root,
8637 : unique_rel,
8638 : path,
8639 : sortPathkeys,
8640 : presorted_keys,
8641 : -1.0);
8642 : }
8643 :
8644 9160 : path = (Path *) create_unique_path(root, unique_rel, path,
8645 : list_length(sortPathkeys),
8646 : unique_rel->rows);
8647 :
8648 9160 : add_path(unique_rel, path);
8649 : }
8650 : }
8651 :
8652 : /* Consider hash-based implementation, if possible. */
8653 8520 : if (sjinfo->semi_can_hash)
8654 : {
8655 : Path *path;
8656 :
8657 : /*
8658 : * Make a separate ProjectionPath in case we need a Result node.
8659 : */
8660 8520 : path = (Path *) create_projection_path(root,
8661 : unique_rel,
8662 : cheapest_input_path,
8663 8520 : unique_rel->reltarget);
8664 :
8665 8520 : path = (Path *) create_agg_path(root,
8666 : unique_rel,
8667 : path,
8668 : cheapest_input_path->pathtarget,
8669 : AGG_HASHED,
8670 : AGGSPLIT_SIMPLE,
8671 : groupClause,
8672 : NIL,
8673 : NULL,
8674 : unique_rel->rows);
8675 :
8676 8520 : add_path(unique_rel, path);
8677 : }
8678 8520 : }
8679 :
8680 : /*
8681 : * create_partial_unique_paths
8682 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
8683 : */
8684 : static void
8685 4874 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8686 : List *sortPathkeys, List *groupClause,
8687 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8688 : {
8689 : RelOptInfo *partial_unique_rel;
8690 : Path *cheapest_partial_path;
8691 :
8692 : /* nothing to do when there are no partial paths in the input rel */
8693 4874 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
8694 1228 : return;
8695 :
8696 : /*
8697 : * nothing to do if there's anything in the targetlist that's
8698 : * parallel-restricted.
8699 : */
8700 3646 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
8701 0 : return;
8702 :
8703 3646 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
8704 :
8705 3646 : partial_unique_rel = makeNode(RelOptInfo);
8706 3646 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
8707 :
8708 : /*
8709 : * clear path info
8710 : */
8711 3646 : partial_unique_rel->pathlist = NIL;
8712 3646 : partial_unique_rel->ppilist = NIL;
8713 3646 : partial_unique_rel->partial_pathlist = NIL;
8714 3646 : partial_unique_rel->cheapest_startup_path = NULL;
8715 3646 : partial_unique_rel->cheapest_total_path = NULL;
8716 3646 : partial_unique_rel->cheapest_parameterized_paths = NIL;
8717 :
8718 : /* Estimate number of output rows */
8719 3646 : partial_unique_rel->rows = estimate_num_groups(root,
8720 : sjinfo->semi_rhs_exprs,
8721 : cheapest_partial_path->rows,
8722 : NULL,
8723 : NULL);
8724 3646 : partial_unique_rel->reltarget = unique_rel->reltarget;
8725 :
8726 : /* Consider sort-based implementations, if possible. */
8727 3646 : if (sjinfo->semi_can_btree)
8728 : {
8729 : ListCell *lc;
8730 :
8731 : /*
8732 : * Use any available suitably-sorted path as input, and also consider
8733 : * sorting the cheapest partial path and incremental sort on any paths
8734 : * with presorted keys.
8735 : */
8736 7604 : foreach(lc, input_rel->partial_pathlist)
8737 : {
8738 3958 : Path *input_path = (Path *) lfirst(lc);
8739 : Path *path;
8740 : bool is_sorted;
8741 : int presorted_keys;
8742 :
8743 3958 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8744 : input_path->pathkeys,
8745 : &presorted_keys);
8746 :
8747 : /*
8748 : * Ignore paths that are not suitably or partially sorted, unless
8749 : * they are the cheapest partial path (no need to deal with paths
8750 : * which have presorted keys when incremental sort is disabled).
8751 : */
8752 3958 : if (!is_sorted && input_path != cheapest_partial_path &&
8753 0 : (presorted_keys == 0 || !enable_incremental_sort))
8754 0 : continue;
8755 :
8756 : /*
8757 : * Make a separate ProjectionPath in case we need a Result node.
8758 : */
8759 3958 : path = (Path *) create_projection_path(root,
8760 : partial_unique_rel,
8761 : input_path,
8762 3958 : partial_unique_rel->reltarget);
8763 :
8764 3958 : if (!is_sorted)
8765 : {
8766 : /*
8767 : * We've no need to consider both a sort and incremental sort.
8768 : * We'll just do a sort if there are no presorted keys and an
8769 : * incremental sort when there are presorted keys.
8770 : */
8771 3598 : if (presorted_keys == 0 || !enable_incremental_sort)
8772 3598 : path = (Path *) create_sort_path(root,
8773 : partial_unique_rel,
8774 : path,
8775 : sortPathkeys,
8776 : -1.0);
8777 : else
8778 0 : path = (Path *) create_incremental_sort_path(root,
8779 : partial_unique_rel,
8780 : path,
8781 : sortPathkeys,
8782 : presorted_keys,
8783 : -1.0);
8784 : }
8785 :
8786 3958 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
8787 : list_length(sortPathkeys),
8788 : partial_unique_rel->rows);
8789 :
8790 3958 : add_partial_path(partial_unique_rel, path);
8791 : }
8792 : }
8793 :
8794 : /* Consider hash-based implementation, if possible. */
8795 3646 : if (sjinfo->semi_can_hash)
8796 : {
8797 : Path *path;
8798 :
8799 : /*
8800 : * Make a separate ProjectionPath in case we need a Result node.
8801 : */
8802 3646 : path = (Path *) create_projection_path(root,
8803 : partial_unique_rel,
8804 : cheapest_partial_path,
8805 3646 : partial_unique_rel->reltarget);
8806 :
8807 3646 : path = (Path *) create_agg_path(root,
8808 : partial_unique_rel,
8809 : path,
8810 : cheapest_partial_path->pathtarget,
8811 : AGG_HASHED,
8812 : AGGSPLIT_SIMPLE,
8813 : groupClause,
8814 : NIL,
8815 : NULL,
8816 : partial_unique_rel->rows);
8817 :
8818 3646 : add_partial_path(partial_unique_rel, path);
8819 : }
8820 :
8821 3646 : if (partial_unique_rel->partial_pathlist != NIL)
8822 : {
8823 3646 : generate_useful_gather_paths(root, partial_unique_rel, true);
8824 3646 : set_cheapest(partial_unique_rel);
8825 :
8826 : /*
8827 : * Finally, create paths to unique-ify the final result. This step is
8828 : * needed to remove any duplicates due to combining rows from parallel
8829 : * workers.
8830 : */
8831 3646 : create_final_unique_paths(root, partial_unique_rel,
8832 : sortPathkeys, groupClause,
8833 : sjinfo, unique_rel);
8834 : }
8835 : }
|
