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 after PlannerGlobal is initialized */
77 : planner_setup_hook_type planner_setup_hook = NULL;
78 :
79 : /* Hook for plugins to get control before PlannerGlobal is discarded */
80 : planner_shutdown_hook_type planner_shutdown_hook = NULL;
81 :
82 : /* Hook for plugins to get control when grouping_planner() plans upper rels */
83 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
84 :
85 :
86 : /* Expression kind codes for preprocess_expression */
87 : #define EXPRKIND_QUAL 0
88 : #define EXPRKIND_TARGET 1
89 : #define EXPRKIND_RTFUNC 2
90 : #define EXPRKIND_RTFUNC_LATERAL 3
91 : #define EXPRKIND_VALUES 4
92 : #define EXPRKIND_VALUES_LATERAL 5
93 : #define EXPRKIND_LIMIT 6
94 : #define EXPRKIND_APPINFO 7
95 : #define EXPRKIND_PHV 8
96 : #define EXPRKIND_TABLESAMPLE 9
97 : #define EXPRKIND_ARBITER_ELEM 10
98 : #define EXPRKIND_TABLEFUNC 11
99 : #define EXPRKIND_TABLEFUNC_LATERAL 12
100 : #define EXPRKIND_GROUPEXPR 13
101 :
102 : /*
103 : * Data specific to grouping sets
104 : */
105 : typedef struct
106 : {
107 : List *rollups;
108 : List *hash_sets_idx;
109 : double dNumHashGroups;
110 : bool any_hashable;
111 : Bitmapset *unsortable_refs;
112 : Bitmapset *unhashable_refs;
113 : List *unsortable_sets;
114 : int *tleref_to_colnum_map;
115 : } grouping_sets_data;
116 :
117 : /*
118 : * Temporary structure for use during WindowClause reordering in order to be
119 : * able to sort WindowClauses on partitioning/ordering prefix.
120 : */
121 : typedef struct
122 : {
123 : WindowClause *wc;
124 : List *uniqueOrder; /* A List of unique ordering/partitioning
125 : * clauses per Window */
126 : } WindowClauseSortData;
127 :
128 : /* Passthrough data for standard_qp_callback */
129 : typedef struct
130 : {
131 : List *activeWindows; /* active windows, if any */
132 : grouping_sets_data *gset_data; /* grouping sets data, if any */
133 : SetOperationStmt *setop; /* parent set operation or NULL if not a
134 : * subquery belonging to a set operation */
135 : } standard_qp_extra;
136 :
137 : /* Local functions */
138 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
139 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
140 : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
141 : SetOperationStmt *setops);
142 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
143 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
144 : int *tleref_to_colnum_map);
145 : static void preprocess_rowmarks(PlannerInfo *root);
146 : static double preprocess_limit(PlannerInfo *root,
147 : double tuple_fraction,
148 : int64 *offset_est, int64 *count_est);
149 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
150 : static List *extract_rollup_sets(List *groupingSets);
151 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
152 : static void standard_qp_callback(PlannerInfo *root, void *extra);
153 : static double get_number_of_groups(PlannerInfo *root,
154 : double path_rows,
155 : grouping_sets_data *gd,
156 : List *target_list);
157 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
158 : RelOptInfo *input_rel,
159 : PathTarget *target,
160 : bool target_parallel_safe,
161 : grouping_sets_data *gd);
162 : static bool is_degenerate_grouping(PlannerInfo *root);
163 : static void create_degenerate_grouping_paths(PlannerInfo *root,
164 : RelOptInfo *input_rel,
165 : RelOptInfo *grouped_rel);
166 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
167 : PathTarget *target, bool target_parallel_safe,
168 : Node *havingQual);
169 : static void create_ordinary_grouping_paths(PlannerInfo *root,
170 : RelOptInfo *input_rel,
171 : RelOptInfo *grouped_rel,
172 : const AggClauseCosts *agg_costs,
173 : grouping_sets_data *gd,
174 : GroupPathExtraData *extra,
175 : RelOptInfo **partially_grouped_rel_p);
176 : static void consider_groupingsets_paths(PlannerInfo *root,
177 : RelOptInfo *grouped_rel,
178 : Path *path,
179 : bool is_sorted,
180 : bool can_hash,
181 : grouping_sets_data *gd,
182 : const AggClauseCosts *agg_costs,
183 : double dNumGroups);
184 : static RelOptInfo *create_window_paths(PlannerInfo *root,
185 : RelOptInfo *input_rel,
186 : PathTarget *input_target,
187 : PathTarget *output_target,
188 : bool output_target_parallel_safe,
189 : WindowFuncLists *wflists,
190 : List *activeWindows);
191 : static void create_one_window_path(PlannerInfo *root,
192 : RelOptInfo *window_rel,
193 : Path *path,
194 : PathTarget *input_target,
195 : PathTarget *output_target,
196 : WindowFuncLists *wflists,
197 : List *activeWindows);
198 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
199 : RelOptInfo *input_rel,
200 : PathTarget *target);
201 : static void create_partial_distinct_paths(PlannerInfo *root,
202 : RelOptInfo *input_rel,
203 : RelOptInfo *final_distinct_rel,
204 : PathTarget *target);
205 : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
206 : RelOptInfo *input_rel,
207 : RelOptInfo *distinct_rel);
208 : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
209 : List *needed_pathkeys,
210 : List *path_pathkeys);
211 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
212 : RelOptInfo *input_rel,
213 : PathTarget *target,
214 : bool target_parallel_safe,
215 : double limit_tuples);
216 : static PathTarget *make_group_input_target(PlannerInfo *root,
217 : PathTarget *final_target);
218 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
219 : PathTarget *grouping_target,
220 : Node *havingQual);
221 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
222 : static void optimize_window_clauses(PlannerInfo *root,
223 : WindowFuncLists *wflists);
224 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
225 : static void name_active_windows(List *activeWindows);
226 : static PathTarget *make_window_input_target(PlannerInfo *root,
227 : PathTarget *final_target,
228 : List *activeWindows);
229 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
230 : List *tlist);
231 : static PathTarget *make_sort_input_target(PlannerInfo *root,
232 : PathTarget *final_target,
233 : bool *have_postponed_srfs);
234 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
235 : List *targets, List *targets_contain_srfs);
236 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
237 : RelOptInfo *grouped_rel,
238 : RelOptInfo *partially_grouped_rel,
239 : const AggClauseCosts *agg_costs,
240 : grouping_sets_data *gd,
241 : GroupPathExtraData *extra);
242 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
243 : RelOptInfo *grouped_rel,
244 : RelOptInfo *input_rel,
245 : grouping_sets_data *gd,
246 : GroupPathExtraData *extra,
247 : bool force_rel_creation);
248 : static Path *make_ordered_path(PlannerInfo *root,
249 : RelOptInfo *rel,
250 : Path *path,
251 : Path *cheapest_path,
252 : List *pathkeys,
253 : double limit_tuples);
254 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
255 : static bool can_partial_agg(PlannerInfo *root);
256 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
257 : RelOptInfo *rel,
258 : List *scanjoin_targets,
259 : List *scanjoin_targets_contain_srfs,
260 : bool scanjoin_target_parallel_safe,
261 : bool tlist_same_exprs);
262 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
263 : RelOptInfo *input_rel,
264 : RelOptInfo *grouped_rel,
265 : RelOptInfo *partially_grouped_rel,
266 : const AggClauseCosts *agg_costs,
267 : grouping_sets_data *gd,
268 : PartitionwiseAggregateType patype,
269 : GroupPathExtraData *extra);
270 : static bool group_by_has_partkey(RelOptInfo *input_rel,
271 : List *targetList,
272 : List *groupClause);
273 : static int common_prefix_cmp(const void *a, const void *b);
274 : static List *generate_setop_child_grouplist(SetOperationStmt *op,
275 : List *targetlist);
276 : static void create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
277 : List *sortPathkeys, List *groupClause,
278 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
279 : static void create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
280 : List *sortPathkeys, List *groupClause,
281 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
282 :
283 :
284 : /*****************************************************************************
285 : *
286 : * Query optimizer entry point
287 : *
288 : * Inputs:
289 : * parse: an analyzed-and-rewritten query tree for an optimizable statement
290 : * query_string: source text for the query tree (used for error reports)
291 : * cursorOptions: bitmask of CURSOR_OPT_XXX flags, see parsenodes.h
292 : * boundParams: passed-in parameter values, or NULL if none
293 : * es: ExplainState if being called from EXPLAIN, else NULL
294 : *
295 : * The result is a PlannedStmt tree.
296 : *
297 : * PARAM_EXTERN Param nodes within the parse tree can be replaced by Consts
298 : * using values from boundParams, if those values are marked PARAM_FLAG_CONST.
299 : * Parameter values not so marked are still relied on for estimation purposes.
300 : *
301 : * The ExplainState pointer is not currently used by the core planner, but it
302 : * is passed through to some planner hooks so that they can report information
303 : * back to EXPLAIN extension hooks.
304 : *
305 : * To support loadable plugins that monitor or modify planner behavior,
306 : * we provide a hook variable that lets a plugin get control before and
307 : * after the standard planning process. The plugin would normally call
308 : * standard_planner().
309 : *
310 : * Note to plugin authors: standard_planner() scribbles on its Query input,
311 : * so you'd better copy that data structure if you want to plan more than once.
312 : *
313 : *****************************************************************************/
314 : PlannedStmt *
315 450226 : planner(Query *parse, const char *query_string, int cursorOptions,
316 : ParamListInfo boundParams, ExplainState *es)
317 : {
318 : PlannedStmt *result;
319 :
320 450226 : if (planner_hook)
321 94932 : result = (*planner_hook) (parse, query_string, cursorOptions,
322 : boundParams, es);
323 : else
324 355294 : result = standard_planner(parse, query_string, cursorOptions,
325 : boundParams, es);
326 :
327 445410 : pgstat_report_plan_id(result->planId, false);
328 :
329 445410 : return result;
330 : }
331 :
332 : PlannedStmt *
333 450226 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
334 : ParamListInfo boundParams, ExplainState *es)
335 : {
336 : PlannedStmt *result;
337 : PlannerGlobal *glob;
338 : double tuple_fraction;
339 : PlannerInfo *root;
340 : RelOptInfo *final_rel;
341 : Path *best_path;
342 : Plan *top_plan;
343 : ListCell *lp,
344 : *lr;
345 :
346 : /*
347 : * Set up global state for this planner invocation. This data is needed
348 : * across all levels of sub-Query that might exist in the given command,
349 : * so we keep it in a separate struct that's linked to by each per-Query
350 : * PlannerInfo.
351 : */
352 450226 : glob = makeNode(PlannerGlobal);
353 :
354 450226 : glob->boundParams = boundParams;
355 450226 : glob->subplans = NIL;
356 450226 : glob->subpaths = NIL;
357 450226 : glob->subroots = NIL;
358 450226 : glob->rewindPlanIDs = NULL;
359 450226 : glob->finalrtable = NIL;
360 450226 : glob->allRelids = NULL;
361 450226 : glob->prunableRelids = NULL;
362 450226 : glob->finalrteperminfos = NIL;
363 450226 : glob->finalrowmarks = NIL;
364 450226 : glob->resultRelations = NIL;
365 450226 : glob->appendRelations = NIL;
366 450226 : glob->partPruneInfos = NIL;
367 450226 : glob->relationOids = NIL;
368 450226 : glob->invalItems = NIL;
369 450226 : glob->paramExecTypes = NIL;
370 450226 : glob->lastPHId = 0;
371 450226 : glob->lastRowMarkId = 0;
372 450226 : glob->lastPlanNodeId = 0;
373 450226 : glob->transientPlan = false;
374 450226 : glob->dependsOnRole = false;
375 450226 : glob->partition_directory = NULL;
376 450226 : glob->rel_notnullatts_hash = NULL;
377 :
378 : /*
379 : * Assess whether it's feasible to use parallel mode for this query. We
380 : * can't do this in a standalone backend, or if the command will try to
381 : * modify any data, or if this is a cursor operation, or if GUCs are set
382 : * to values that don't permit parallelism, or if parallel-unsafe
383 : * functions are present in the query tree.
384 : *
385 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
386 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
387 : * the command is writing into a completely new table which workers won't
388 : * be able to see. If the workers could see the table, the fact that
389 : * group locking would cause them to ignore the leader's heavyweight GIN
390 : * page locks would make this unsafe. We'll have to fix that somehow if
391 : * we want to allow parallel inserts in general; updates and deletes have
392 : * additional problems especially around combo CIDs.)
393 : *
394 : * For now, we don't try to use parallel mode if we're running inside a
395 : * parallel worker. We might eventually be able to relax this
396 : * restriction, but for now it seems best not to have parallel workers
397 : * trying to create their own parallel workers.
398 : */
399 450226 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
400 422012 : IsUnderPostmaster &&
401 422012 : parse->commandType == CMD_SELECT &&
402 339512 : !parse->hasModifyingCTE &&
403 339368 : max_parallel_workers_per_gather > 0 &&
404 338754 : !IsParallelWorker())
405 : {
406 : /* all the cheap tests pass, so scan the query tree */
407 338706 : glob->maxParallelHazard = max_parallel_hazard(parse);
408 338706 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
409 : }
410 : else
411 : {
412 : /* skip the query tree scan, just assume it's unsafe */
413 111520 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
414 111520 : glob->parallelModeOK = false;
415 : }
416 :
417 : /*
418 : * glob->parallelModeNeeded is normally set to false here and changed to
419 : * true during plan creation if a Gather or Gather Merge plan is actually
420 : * created (cf. create_gather_plan, create_gather_merge_plan).
421 : *
422 : * However, if debug_parallel_query = on or debug_parallel_query =
423 : * regress, then we impose parallel mode whenever it's safe to do so, even
424 : * if the final plan doesn't use parallelism. It's not safe to do so if
425 : * the query contains anything parallel-unsafe; parallelModeOK will be
426 : * false in that case. Note that parallelModeOK can't change after this
427 : * point. Otherwise, everything in the query is either parallel-safe or
428 : * parallel-restricted, and in either case it should be OK to impose
429 : * parallel-mode restrictions. If that ends up breaking something, then
430 : * either some function the user included in the query is incorrectly
431 : * labeled as parallel-safe or parallel-restricted when in reality it's
432 : * parallel-unsafe, or else the query planner itself has a bug.
433 : */
434 739828 : glob->parallelModeNeeded = glob->parallelModeOK &&
435 289602 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
436 :
437 : /* Determine what fraction of the plan is likely to be scanned */
438 450226 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
439 : {
440 : /*
441 : * We have no real idea how many tuples the user will ultimately FETCH
442 : * from a cursor, but it is often the case that he doesn't want 'em
443 : * all, or would prefer a fast-start plan anyway so that he can
444 : * process some of the tuples sooner. Use a GUC parameter to decide
445 : * what fraction to optimize for.
446 : */
447 4700 : tuple_fraction = cursor_tuple_fraction;
448 :
449 : /*
450 : * We document cursor_tuple_fraction as simply being a fraction, which
451 : * means the edge cases 0 and 1 have to be treated specially here. We
452 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
453 : */
454 4700 : if (tuple_fraction >= 1.0)
455 0 : tuple_fraction = 0.0;
456 4700 : else if (tuple_fraction <= 0.0)
457 0 : tuple_fraction = 1e-10;
458 : }
459 : else
460 : {
461 : /* Default assumption is we need all the tuples */
462 445526 : tuple_fraction = 0.0;
463 : }
464 :
465 : /* Allow plugins to take control after we've initialized "glob" */
466 450226 : if (planner_setup_hook)
467 0 : (*planner_setup_hook) (glob, parse, query_string, &tuple_fraction, es);
468 :
469 : /* primary planning entry point (may recurse for subqueries) */
470 450226 : root = subquery_planner(glob, parse, NULL, NULL, false, tuple_fraction,
471 : NULL);
472 :
473 : /* Select best Path and turn it into a Plan */
474 445806 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
475 445806 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
476 :
477 445806 : top_plan = create_plan(root, best_path);
478 :
479 : /*
480 : * If creating a plan for a scrollable cursor, make sure it can run
481 : * backwards on demand. Add a Material node at the top at need.
482 : */
483 445410 : if (cursorOptions & CURSOR_OPT_SCROLL)
484 : {
485 266 : if (!ExecSupportsBackwardScan(top_plan))
486 32 : top_plan = materialize_finished_plan(top_plan);
487 : }
488 :
489 : /*
490 : * Optionally add a Gather node for testing purposes, provided this is
491 : * actually a safe thing to do.
492 : *
493 : * We can add Gather even when top_plan has parallel-safe initPlans, but
494 : * then we have to move the initPlans to the Gather node because of
495 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
496 : * regression tests when debug_parallel_query = regress, because initPlans
497 : * that would normally appear on the top_plan move to the Gather, causing
498 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
499 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
500 : */
501 445410 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
502 194 : top_plan->parallel_safe &&
503 128 : (top_plan->initPlan == NIL ||
504 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
505 : {
506 128 : Gather *gather = makeNode(Gather);
507 : Cost initplan_cost;
508 : bool unsafe_initplans;
509 :
510 128 : gather->plan.targetlist = top_plan->targetlist;
511 128 : gather->plan.qual = NIL;
512 128 : gather->plan.lefttree = top_plan;
513 128 : gather->plan.righttree = NULL;
514 128 : gather->num_workers = 1;
515 128 : gather->single_copy = true;
516 128 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
517 :
518 : /* Transfer any initPlans to the new top node */
519 128 : gather->plan.initPlan = top_plan->initPlan;
520 128 : top_plan->initPlan = NIL;
521 :
522 : /*
523 : * Since this Gather has no parallel-aware descendants to signal to,
524 : * we don't need a rescan Param.
525 : */
526 128 : gather->rescan_param = -1;
527 :
528 : /*
529 : * Ideally we'd use cost_gather here, but setting up dummy path data
530 : * to satisfy it doesn't seem much cleaner than knowing what it does.
531 : */
532 128 : gather->plan.startup_cost = top_plan->startup_cost +
533 : parallel_setup_cost;
534 128 : gather->plan.total_cost = top_plan->total_cost +
535 128 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
536 128 : gather->plan.plan_rows = top_plan->plan_rows;
537 128 : gather->plan.plan_width = top_plan->plan_width;
538 128 : gather->plan.parallel_aware = false;
539 128 : gather->plan.parallel_safe = false;
540 :
541 : /*
542 : * Delete the initplans' cost from top_plan. We needn't add it to the
543 : * Gather node, since the above coding already included it there.
544 : */
545 128 : SS_compute_initplan_cost(gather->plan.initPlan,
546 : &initplan_cost, &unsafe_initplans);
547 128 : top_plan->startup_cost -= initplan_cost;
548 128 : top_plan->total_cost -= initplan_cost;
549 :
550 : /* use parallel mode for parallel plans. */
551 128 : root->glob->parallelModeNeeded = true;
552 :
553 128 : top_plan = &gather->plan;
554 : }
555 :
556 : /*
557 : * If any Params were generated, run through the plan tree and compute
558 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
559 : * set_plan_references' tree traversal, but for now it has to be separate
560 : * because we need to visit subplans before not after main plan.
561 : */
562 445410 : if (glob->paramExecTypes != NIL)
563 : {
564 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
565 197148 : forboth(lp, glob->subplans, lr, glob->subroots)
566 : {
567 43510 : Plan *subplan = (Plan *) lfirst(lp);
568 43510 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
569 :
570 43510 : SS_finalize_plan(subroot, subplan);
571 : }
572 153638 : SS_finalize_plan(root, top_plan);
573 : }
574 :
575 : /* final cleanup of the plan */
576 : Assert(glob->finalrtable == NIL);
577 : Assert(glob->finalrteperminfos == NIL);
578 : Assert(glob->finalrowmarks == NIL);
579 : Assert(glob->resultRelations == NIL);
580 : Assert(glob->appendRelations == NIL);
581 445410 : top_plan = set_plan_references(root, top_plan);
582 : /* ... and the subplans (both regular subplans and initplans) */
583 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
584 488920 : forboth(lp, glob->subplans, lr, glob->subroots)
585 : {
586 43510 : Plan *subplan = (Plan *) lfirst(lp);
587 43510 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
588 :
589 43510 : lfirst(lp) = set_plan_references(subroot, subplan);
590 : }
591 :
592 : /* build the PlannedStmt result */
593 445410 : result = makeNode(PlannedStmt);
594 :
595 445410 : result->commandType = parse->commandType;
596 445410 : result->queryId = parse->queryId;
597 445410 : result->planOrigin = PLAN_STMT_STANDARD;
598 445410 : result->hasReturning = (parse->returningList != NIL);
599 445410 : result->hasModifyingCTE = parse->hasModifyingCTE;
600 445410 : result->canSetTag = parse->canSetTag;
601 445410 : result->transientPlan = glob->transientPlan;
602 445410 : result->dependsOnRole = glob->dependsOnRole;
603 445410 : result->parallelModeNeeded = glob->parallelModeNeeded;
604 445410 : result->planTree = top_plan;
605 445410 : result->partPruneInfos = glob->partPruneInfos;
606 445410 : result->rtable = glob->finalrtable;
607 890820 : result->unprunableRelids = bms_difference(glob->allRelids,
608 445410 : glob->prunableRelids);
609 445410 : result->permInfos = glob->finalrteperminfos;
610 445410 : result->resultRelations = glob->resultRelations;
611 445410 : result->appendRelations = glob->appendRelations;
612 445410 : result->subplans = glob->subplans;
613 445410 : result->rewindPlanIDs = glob->rewindPlanIDs;
614 445410 : result->rowMarks = glob->finalrowmarks;
615 445410 : result->relationOids = glob->relationOids;
616 445410 : result->invalItems = glob->invalItems;
617 445410 : result->paramExecTypes = glob->paramExecTypes;
618 : /* utilityStmt should be null, but we might as well copy it */
619 445410 : result->utilityStmt = parse->utilityStmt;
620 445410 : result->stmt_location = parse->stmt_location;
621 445410 : result->stmt_len = parse->stmt_len;
622 :
623 445410 : result->jitFlags = PGJIT_NONE;
624 445410 : if (jit_enabled && jit_above_cost >= 0 &&
625 444718 : top_plan->total_cost > jit_above_cost)
626 : {
627 954 : result->jitFlags |= PGJIT_PERFORM;
628 :
629 : /*
630 : * Decide how much effort should be put into generating better code.
631 : */
632 954 : if (jit_optimize_above_cost >= 0 &&
633 954 : top_plan->total_cost > jit_optimize_above_cost)
634 436 : result->jitFlags |= PGJIT_OPT3;
635 954 : if (jit_inline_above_cost >= 0 &&
636 954 : top_plan->total_cost > jit_inline_above_cost)
637 436 : result->jitFlags |= PGJIT_INLINE;
638 :
639 : /*
640 : * Decide which operations should be JITed.
641 : */
642 954 : if (jit_expressions)
643 954 : result->jitFlags |= PGJIT_EXPR;
644 954 : if (jit_tuple_deforming)
645 954 : result->jitFlags |= PGJIT_DEFORM;
646 : }
647 :
648 : /* Allow plugins to take control before we discard "glob" */
649 445410 : if (planner_shutdown_hook)
650 0 : (*planner_shutdown_hook) (glob, parse, query_string, result);
651 :
652 445410 : if (glob->partition_directory != NULL)
653 11780 : DestroyPartitionDirectory(glob->partition_directory);
654 :
655 445410 : return result;
656 : }
657 :
658 :
659 : /*--------------------
660 : * subquery_planner
661 : * Invokes the planner on a subquery. We recurse to here for each
662 : * sub-SELECT found in the query tree.
663 : *
664 : * glob is the global state for the current planner run.
665 : * parse is the querytree produced by the parser & rewriter.
666 : * plan_name is the name to assign to this subplan (NULL at the top level).
667 : * parent_root is the immediate parent Query's info (NULL at the top level).
668 : * hasRecursion is true if this is a recursive WITH query.
669 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
670 : * tuple_fraction is interpreted as explained for grouping_planner, below.
671 : * setops is used for set operation subqueries to provide the subquery with
672 : * the context in which it's being used so that Paths correctly sorted for the
673 : * set operation can be generated. NULL when not planning a set operation
674 : * child, or when a child of a set op that isn't interested in sorted input.
675 : *
676 : * Basically, this routine does the stuff that should only be done once
677 : * per Query object. It then calls grouping_planner. At one time,
678 : * grouping_planner could be invoked recursively on the same Query object;
679 : * that's not currently true, but we keep the separation between the two
680 : * routines anyway, in case we need it again someday.
681 : *
682 : * subquery_planner will be called recursively to handle sub-Query nodes
683 : * found within the query's expressions and rangetable.
684 : *
685 : * Returns the PlannerInfo struct ("root") that contains all data generated
686 : * while planning the subquery. In particular, the Path(s) attached to
687 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
688 : * cheapest way(s) to implement the query. The top level will select the
689 : * best Path and pass it through createplan.c to produce a finished Plan.
690 : *--------------------
691 : */
692 : PlannerInfo *
693 526750 : subquery_planner(PlannerGlobal *glob, Query *parse, char *plan_name,
694 : PlannerInfo *parent_root, bool hasRecursion,
695 : double tuple_fraction, SetOperationStmt *setops)
696 : {
697 : PlannerInfo *root;
698 : List *newWithCheckOptions;
699 : List *newHaving;
700 : bool hasOuterJoins;
701 : bool hasResultRTEs;
702 : RelOptInfo *final_rel;
703 : ListCell *l;
704 :
705 : /* Create a PlannerInfo data structure for this subquery */
706 526750 : root = makeNode(PlannerInfo);
707 526750 : root->parse = parse;
708 526750 : root->glob = glob;
709 526750 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
710 526750 : root->plan_name = plan_name;
711 526750 : root->parent_root = parent_root;
712 526750 : root->plan_params = NIL;
713 526750 : root->outer_params = NULL;
714 526750 : root->planner_cxt = CurrentMemoryContext;
715 526750 : root->init_plans = NIL;
716 526750 : root->cte_plan_ids = NIL;
717 526750 : root->multiexpr_params = NIL;
718 526750 : root->join_domains = NIL;
719 526750 : root->eq_classes = NIL;
720 526750 : root->ec_merging_done = false;
721 526750 : root->last_rinfo_serial = 0;
722 526750 : root->all_result_relids =
723 526750 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
724 526750 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
725 526750 : root->append_rel_list = NIL;
726 526750 : root->row_identity_vars = NIL;
727 526750 : root->rowMarks = NIL;
728 526750 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
729 526750 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
730 526750 : root->processed_groupClause = NIL;
731 526750 : root->processed_distinctClause = NIL;
732 526750 : root->processed_tlist = NIL;
733 526750 : root->update_colnos = NIL;
734 526750 : root->grouping_map = NULL;
735 526750 : root->minmax_aggs = NIL;
736 526750 : root->qual_security_level = 0;
737 526750 : root->hasPseudoConstantQuals = false;
738 526750 : root->hasAlternativeSubPlans = false;
739 526750 : root->placeholdersFrozen = false;
740 526750 : root->hasRecursion = hasRecursion;
741 526750 : root->assumeReplanning = false;
742 526750 : if (hasRecursion)
743 930 : root->wt_param_id = assign_special_exec_param(root);
744 : else
745 525820 : root->wt_param_id = -1;
746 526750 : root->non_recursive_path = NULL;
747 :
748 : /*
749 : * Create the top-level join domain. This won't have valid contents until
750 : * deconstruct_jointree fills it in, but the node needs to exist before
751 : * that so we can build EquivalenceClasses referencing it.
752 : */
753 526750 : root->join_domains = list_make1(makeNode(JoinDomain));
754 :
755 : /*
756 : * If there is a WITH list, process each WITH query and either convert it
757 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
758 : */
759 526750 : if (parse->cteList)
760 2874 : SS_process_ctes(root);
761 :
762 : /*
763 : * If it's a MERGE command, transform the joinlist as appropriate.
764 : */
765 526744 : transform_MERGE_to_join(parse);
766 :
767 : /*
768 : * Scan the rangetable for relation RTEs and retrieve the necessary
769 : * catalog information for each relation. Using this information, clear
770 : * the inh flag for any relation that has no children, collect not-null
771 : * attribute numbers for any relation that has column not-null
772 : * constraints, and expand virtual generated columns for any relation that
773 : * contains them. Note that this step does not descend into sublinks and
774 : * subqueries; if we pull up any sublinks or subqueries below, their
775 : * relation RTEs are processed just before pulling them up.
776 : */
777 526744 : parse = root->parse = preprocess_relation_rtes(root);
778 :
779 : /*
780 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
781 : * that we don't need so many special cases to deal with that situation.
782 : */
783 526744 : replace_empty_jointree(parse);
784 :
785 : /*
786 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
787 : * to transform them into joins. Note that this step does not descend
788 : * into subqueries; if we pull up any subqueries below, their SubLinks are
789 : * processed just before pulling them up.
790 : */
791 526744 : if (parse->hasSubLinks)
792 36282 : pull_up_sublinks(root);
793 :
794 : /*
795 : * Scan the rangetable for function RTEs, do const-simplification on them,
796 : * and then inline them if possible (producing subqueries that might get
797 : * pulled up next). Recursion issues here are handled in the same way as
798 : * for SubLinks.
799 : */
800 526744 : preprocess_function_rtes(root);
801 :
802 : /*
803 : * Check to see if any subqueries in the jointree can be merged into this
804 : * query.
805 : */
806 526738 : pull_up_subqueries(root);
807 :
808 : /*
809 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
810 : * do this now because it requires applying pull_up_subqueries to the leaf
811 : * queries of the UNION ALL, which weren't touched above because they
812 : * weren't referenced by the jointree (they will be after we do this).
813 : */
814 526732 : if (parse->setOperations)
815 6878 : flatten_simple_union_all(root);
816 :
817 : /*
818 : * Survey the rangetable to see what kinds of entries are present. We can
819 : * skip some later processing if relevant SQL features are not used; for
820 : * example if there are no JOIN RTEs we can avoid the expense of doing
821 : * flatten_join_alias_vars(). This must be done after we have finished
822 : * adding rangetable entries, of course. (Note: actually, processing of
823 : * inherited or partitioned rels can cause RTEs for their child tables to
824 : * get added later; but those must all be RTE_RELATION entries, so they
825 : * don't invalidate the conclusions drawn here.)
826 : */
827 526732 : root->hasJoinRTEs = false;
828 526732 : root->hasLateralRTEs = false;
829 526732 : root->group_rtindex = 0;
830 526732 : hasOuterJoins = false;
831 526732 : hasResultRTEs = false;
832 1441832 : foreach(l, parse->rtable)
833 : {
834 915100 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
835 :
836 915100 : switch (rte->rtekind)
837 : {
838 92882 : case RTE_JOIN:
839 92882 : root->hasJoinRTEs = true;
840 92882 : if (IS_OUTER_JOIN(rte->jointype))
841 47912 : hasOuterJoins = true;
842 92882 : break;
843 201998 : case RTE_RESULT:
844 201998 : hasResultRTEs = true;
845 201998 : break;
846 4782 : case RTE_GROUP:
847 : Assert(parse->hasGroupRTE);
848 4782 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
849 4782 : break;
850 615438 : default:
851 : /* No work here for other RTE types */
852 615438 : break;
853 : }
854 :
855 915100 : if (rte->lateral)
856 10968 : root->hasLateralRTEs = true;
857 :
858 : /*
859 : * We can also determine the maximum security level required for any
860 : * securityQuals now. Addition of inheritance-child RTEs won't affect
861 : * this, because child tables don't have their own securityQuals; see
862 : * expand_single_inheritance_child().
863 : */
864 915100 : if (rte->securityQuals)
865 2622 : root->qual_security_level = Max(root->qual_security_level,
866 : list_length(rte->securityQuals));
867 : }
868 :
869 : /*
870 : * If we have now verified that the query target relation is
871 : * non-inheriting, mark it as a leaf target.
872 : */
873 526732 : if (parse->resultRelation)
874 : {
875 89372 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
876 :
877 89372 : if (!rte->inh)
878 86474 : root->leaf_result_relids =
879 86474 : bms_make_singleton(parse->resultRelation);
880 : }
881 :
882 : /*
883 : * This would be a convenient time to check access permissions for all
884 : * relations mentioned in the query, since it would be better to fail now,
885 : * before doing any detailed planning. However, for historical reasons,
886 : * we leave this to be done at executor startup.
887 : *
888 : * Note, however, that we do need to check access permissions for any view
889 : * relations mentioned in the query, in order to prevent information being
890 : * leaked by selectivity estimation functions, which only check view owner
891 : * permissions on underlying tables (see all_rows_selectable() and its
892 : * callers). This is a little ugly, because it means that access
893 : * permissions for views will be checked twice, which is another reason
894 : * why it would be better to do all the ACL checks here.
895 : */
896 1440712 : foreach(l, parse->rtable)
897 : {
898 914366 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
899 :
900 914366 : if (rte->perminfoindex != 0 &&
901 489654 : rte->relkind == RELKIND_VIEW)
902 : {
903 : RTEPermissionInfo *perminfo;
904 : bool result;
905 :
906 20758 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
907 20758 : result = ExecCheckOneRelPerms(perminfo);
908 20758 : if (!result)
909 386 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
910 386 : get_rel_name(perminfo->relid));
911 : }
912 : }
913 :
914 : /*
915 : * Preprocess RowMark information. We need to do this after subquery
916 : * pullup, so that all base relations are present.
917 : */
918 526346 : preprocess_rowmarks(root);
919 :
920 : /*
921 : * Set hasHavingQual to remember if HAVING clause is present. Needed
922 : * because preprocess_expression will reduce a constant-true condition to
923 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
924 : */
925 526346 : root->hasHavingQual = (parse->havingQual != NULL);
926 :
927 : /*
928 : * Do expression preprocessing on targetlist and quals, as well as other
929 : * random expressions in the querytree. Note that we do not need to
930 : * handle sort/group expressions explicitly, because they are actually
931 : * part of the targetlist.
932 : */
933 522396 : parse->targetList = (List *)
934 526346 : preprocess_expression(root, (Node *) parse->targetList,
935 : EXPRKIND_TARGET);
936 :
937 522396 : newWithCheckOptions = NIL;
938 524884 : foreach(l, parse->withCheckOptions)
939 : {
940 2488 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
941 :
942 2488 : wco->qual = preprocess_expression(root, wco->qual,
943 : EXPRKIND_QUAL);
944 2488 : if (wco->qual != NULL)
945 2088 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
946 : }
947 522396 : parse->withCheckOptions = newWithCheckOptions;
948 :
949 522396 : parse->returningList = (List *)
950 522396 : preprocess_expression(root, (Node *) parse->returningList,
951 : EXPRKIND_TARGET);
952 :
953 522396 : preprocess_qual_conditions(root, (Node *) parse->jointree);
954 :
955 522396 : parse->havingQual = preprocess_expression(root, parse->havingQual,
956 : EXPRKIND_QUAL);
957 :
958 525218 : foreach(l, parse->windowClause)
959 : {
960 2822 : WindowClause *wc = lfirst_node(WindowClause, l);
961 :
962 : /* partitionClause/orderClause are sort/group expressions */
963 2822 : wc->startOffset = preprocess_expression(root, wc->startOffset,
964 : EXPRKIND_LIMIT);
965 2822 : wc->endOffset = preprocess_expression(root, wc->endOffset,
966 : EXPRKIND_LIMIT);
967 : }
968 :
969 522396 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
970 : EXPRKIND_LIMIT);
971 522396 : parse->limitCount = preprocess_expression(root, parse->limitCount,
972 : EXPRKIND_LIMIT);
973 :
974 522396 : if (parse->onConflict)
975 : {
976 3640 : parse->onConflict->arbiterElems = (List *)
977 1820 : preprocess_expression(root,
978 1820 : (Node *) parse->onConflict->arbiterElems,
979 : EXPRKIND_ARBITER_ELEM);
980 3640 : parse->onConflict->arbiterWhere =
981 1820 : preprocess_expression(root,
982 1820 : parse->onConflict->arbiterWhere,
983 : EXPRKIND_QUAL);
984 3640 : parse->onConflict->onConflictSet = (List *)
985 1820 : preprocess_expression(root,
986 1820 : (Node *) parse->onConflict->onConflictSet,
987 : EXPRKIND_TARGET);
988 1820 : parse->onConflict->onConflictWhere =
989 1820 : preprocess_expression(root,
990 1820 : parse->onConflict->onConflictWhere,
991 : EXPRKIND_QUAL);
992 : /* exclRelTlist contains only Vars, so no preprocessing needed */
993 : }
994 :
995 525192 : foreach(l, parse->mergeActionList)
996 : {
997 2796 : MergeAction *action = (MergeAction *) lfirst(l);
998 :
999 2796 : action->targetList = (List *)
1000 2796 : preprocess_expression(root,
1001 2796 : (Node *) action->targetList,
1002 : EXPRKIND_TARGET);
1003 2796 : action->qual =
1004 2796 : preprocess_expression(root,
1005 : (Node *) action->qual,
1006 : EXPRKIND_QUAL);
1007 : }
1008 :
1009 522396 : parse->mergeJoinCondition =
1010 522396 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1011 :
1012 522396 : root->append_rel_list = (List *)
1013 522396 : preprocess_expression(root, (Node *) root->append_rel_list,
1014 : EXPRKIND_APPINFO);
1015 :
1016 : /* Also need to preprocess expressions within RTEs */
1017 1431830 : foreach(l, parse->rtable)
1018 : {
1019 909434 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1020 : int kind;
1021 : ListCell *lcsq;
1022 :
1023 909434 : if (rte->rtekind == RTE_RELATION)
1024 : {
1025 473078 : if (rte->tablesample)
1026 228 : rte->tablesample = (TableSampleClause *)
1027 228 : preprocess_expression(root,
1028 228 : (Node *) rte->tablesample,
1029 : EXPRKIND_TABLESAMPLE);
1030 : }
1031 436356 : else if (rte->rtekind == RTE_SUBQUERY)
1032 : {
1033 : /*
1034 : * We don't want to do all preprocessing yet on the subquery's
1035 : * expressions, since that will happen when we plan it. But if it
1036 : * contains any join aliases of our level, those have to get
1037 : * expanded now, because planning of the subquery won't do it.
1038 : * That's only possible if the subquery is LATERAL.
1039 : */
1040 75226 : if (rte->lateral && root->hasJoinRTEs)
1041 1434 : rte->subquery = (Query *)
1042 1434 : flatten_join_alias_vars(root, root->parse,
1043 1434 : (Node *) rte->subquery);
1044 : }
1045 361130 : else if (rte->rtekind == RTE_FUNCTION)
1046 : {
1047 : /* Preprocess the function expression(s) fully */
1048 51144 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1049 51144 : rte->functions = (List *)
1050 51144 : preprocess_expression(root, (Node *) rte->functions, kind);
1051 : }
1052 309986 : else if (rte->rtekind == RTE_TABLEFUNC)
1053 : {
1054 : /* Preprocess the function expression(s) fully */
1055 626 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1056 626 : rte->tablefunc = (TableFunc *)
1057 626 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1058 : }
1059 309360 : else if (rte->rtekind == RTE_VALUES)
1060 : {
1061 : /* Preprocess the values lists fully */
1062 8272 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1063 8272 : rte->values_lists = (List *)
1064 8272 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1065 : }
1066 301088 : else if (rte->rtekind == RTE_GROUP)
1067 : {
1068 : /* Preprocess the groupexprs list fully */
1069 4782 : rte->groupexprs = (List *)
1070 4782 : preprocess_expression(root, (Node *) rte->groupexprs,
1071 : EXPRKIND_GROUPEXPR);
1072 : }
1073 :
1074 : /*
1075 : * Process each element of the securityQuals list as if it were a
1076 : * separate qual expression (as indeed it is). We need to do it this
1077 : * way to get proper canonicalization of AND/OR structure. Note that
1078 : * this converts each element into an implicit-AND sublist.
1079 : */
1080 912402 : foreach(lcsq, rte->securityQuals)
1081 : {
1082 2968 : lfirst(lcsq) = preprocess_expression(root,
1083 2968 : (Node *) lfirst(lcsq),
1084 : EXPRKIND_QUAL);
1085 : }
1086 : }
1087 :
1088 : /*
1089 : * Now that we are done preprocessing expressions, and in particular done
1090 : * flattening join alias variables, get rid of the joinaliasvars lists.
1091 : * They no longer match what expressions in the rest of the tree look
1092 : * like, because we have not preprocessed expressions in those lists (and
1093 : * do not want to; for example, expanding a SubLink there would result in
1094 : * a useless unreferenced subplan). Leaving them in place simply creates
1095 : * a hazard for later scans of the tree. We could try to prevent that by
1096 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1097 : * but that doesn't sound very reliable.
1098 : */
1099 522396 : if (root->hasJoinRTEs)
1100 : {
1101 322316 : foreach(l, parse->rtable)
1102 : {
1103 265164 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1104 :
1105 265164 : rte->joinaliasvars = NIL;
1106 : }
1107 : }
1108 :
1109 : /*
1110 : * Replace any Vars in the subquery's targetlist and havingQual that
1111 : * reference GROUP outputs with the underlying grouping expressions.
1112 : *
1113 : * Note that we need to perform this replacement after we've preprocessed
1114 : * the grouping expressions. This is to ensure that there is only one
1115 : * instance of SubPlan for each SubLink contained within the grouping
1116 : * expressions.
1117 : */
1118 522396 : if (parse->hasGroupRTE)
1119 : {
1120 4782 : parse->targetList = (List *)
1121 4782 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1122 4782 : parse->havingQual =
1123 4782 : flatten_group_exprs(root, root->parse, parse->havingQual);
1124 : }
1125 :
1126 : /* Constant-folding might have removed all set-returning functions */
1127 522396 : if (parse->hasTargetSRFs)
1128 11704 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1129 :
1130 : /*
1131 : * If we have grouping sets, expand the groupingSets tree of this query to
1132 : * a flat list of grouping sets. We need to do this before optimizing
1133 : * HAVING, since we can't easily tell if there's an empty grouping set
1134 : * until we have this representation.
1135 : */
1136 522396 : if (parse->groupingSets)
1137 : {
1138 938 : parse->groupingSets =
1139 938 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1140 : }
1141 :
1142 : /*
1143 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1144 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1145 : * volatile functions (since a HAVING clause is supposed to be executed
1146 : * only once per group). We also can't do this if there are any grouping
1147 : * sets and the clause references any columns that are nullable by the
1148 : * grouping sets; the nulled values of those columns are not available
1149 : * before the grouping step. (The test on groupClause might seem wrong,
1150 : * but it's okay: it's just an optimization to avoid running pull_varnos
1151 : * when there cannot be any Vars in the HAVING clause.)
1152 : *
1153 : * Also, it may be that the clause is so expensive to execute that we're
1154 : * better off doing it only once per group, despite the loss of
1155 : * selectivity. This is hard to estimate short of doing the entire
1156 : * planning process twice, so we use a heuristic: clauses containing
1157 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1158 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1159 : * instead of after.
1160 : *
1161 : * If the query has no empty grouping set then we can simply move such a
1162 : * clause into WHERE; any group that fails the clause will not be in the
1163 : * output because none of its tuples will reach the grouping or
1164 : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1165 : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1166 : * clause must be degenerate (variable-free), so we can copy it into WHERE
1167 : * so that query_planner() can use it in a gating Result node. (This could
1168 : * be done better, but it seems not worth optimizing.)
1169 : *
1170 : * Note that a HAVING clause may contain expressions that are not fully
1171 : * preprocessed. This can happen if these expressions are part of
1172 : * grouping items. In such cases, they are replaced with GROUP Vars in
1173 : * the parser and then replaced back after we're done with expression
1174 : * preprocessing on havingQual. This is not an issue if the clause
1175 : * remains in HAVING, because these expressions will be matched to lower
1176 : * target items in setrefs.c. However, if the clause is moved or copied
1177 : * into WHERE, we need to ensure that these expressions are fully
1178 : * preprocessed.
1179 : *
1180 : * Note that both havingQual and parse->jointree->quals are in
1181 : * implicitly-ANDed-list form at this point, even though they are declared
1182 : * as Node *.
1183 : */
1184 522396 : newHaving = NIL;
1185 523658 : foreach(l, (List *) parse->havingQual)
1186 : {
1187 1262 : Node *havingclause = (Node *) lfirst(l);
1188 :
1189 1658 : if (contain_agg_clause(havingclause) ||
1190 792 : contain_volatile_functions(havingclause) ||
1191 396 : contain_subplans(havingclause) ||
1192 516 : (parse->groupClause && parse->groupingSets &&
1193 120 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1194 : {
1195 : /* keep it in HAVING */
1196 938 : newHaving = lappend(newHaving, havingclause);
1197 : }
1198 324 : else if (parse->groupClause &&
1199 288 : (parse->groupingSets == NIL ||
1200 48 : (List *) linitial(parse->groupingSets) != NIL))
1201 276 : {
1202 : /* There is GROUP BY, but no empty grouping set */
1203 : Node *whereclause;
1204 :
1205 : /* Preprocess the HAVING clause fully */
1206 276 : whereclause = preprocess_expression(root, havingclause,
1207 : EXPRKIND_QUAL);
1208 : /* ... and move it to WHERE */
1209 276 : parse->jointree->quals = (Node *)
1210 276 : list_concat((List *) parse->jointree->quals,
1211 : (List *) whereclause);
1212 : }
1213 : else
1214 : {
1215 : /* There is an empty grouping set (perhaps implicitly) */
1216 : Node *whereclause;
1217 :
1218 : /* Preprocess the HAVING clause fully */
1219 48 : whereclause = preprocess_expression(root, copyObject(havingclause),
1220 : EXPRKIND_QUAL);
1221 : /* ... and put a copy in WHERE */
1222 96 : parse->jointree->quals = (Node *)
1223 48 : list_concat((List *) parse->jointree->quals,
1224 : (List *) whereclause);
1225 : /* ... and also keep it in HAVING */
1226 48 : newHaving = lappend(newHaving, havingclause);
1227 : }
1228 : }
1229 522396 : parse->havingQual = (Node *) newHaving;
1230 :
1231 : /*
1232 : * If we have any outer joins, try to reduce them to plain inner joins.
1233 : * This step is most easily done after we've done expression
1234 : * preprocessing.
1235 : */
1236 522396 : if (hasOuterJoins)
1237 33548 : reduce_outer_joins(root);
1238 :
1239 : /*
1240 : * If we have any RTE_RESULT relations, see if they can be deleted from
1241 : * the jointree. We also rely on this processing to flatten single-child
1242 : * FromExprs underneath outer joins. This step is most effectively done
1243 : * after we've done expression preprocessing and outer join reduction.
1244 : */
1245 522396 : if (hasResultRTEs || hasOuterJoins)
1246 230076 : remove_useless_result_rtes(root);
1247 :
1248 : /*
1249 : * Do the main planning.
1250 : */
1251 522396 : grouping_planner(root, tuple_fraction, setops);
1252 :
1253 : /*
1254 : * Capture the set of outer-level param IDs we have access to, for use in
1255 : * extParam/allParam calculations later.
1256 : */
1257 522324 : SS_identify_outer_params(root);
1258 :
1259 : /*
1260 : * If any initPlans were created in this query level, adjust the surviving
1261 : * Paths' costs and parallel-safety flags to account for them. The
1262 : * initPlans won't actually get attached to the plan tree till
1263 : * create_plan() runs, but we must include their effects now.
1264 : */
1265 522324 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1266 522324 : SS_charge_for_initplans(root, final_rel);
1267 :
1268 : /*
1269 : * Make sure we've identified the cheapest Path for the final rel. (By
1270 : * doing this here not in grouping_planner, we include initPlan costs in
1271 : * the decision, though it's unlikely that will change anything.)
1272 : */
1273 522324 : set_cheapest(final_rel);
1274 :
1275 522324 : return root;
1276 : }
1277 :
1278 : /*
1279 : * preprocess_expression
1280 : * Do subquery_planner's preprocessing work for an expression,
1281 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1282 : * conditions), a HAVING clause, or a few other things.
1283 : */
1284 : static Node *
1285 4388142 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1286 : {
1287 : /*
1288 : * Fall out quickly if expression is empty. This occurs often enough to
1289 : * be worth checking. Note that null->null is the correct conversion for
1290 : * implicit-AND result format, too.
1291 : */
1292 4388142 : if (expr == NULL)
1293 3462482 : return NULL;
1294 :
1295 : /*
1296 : * If the query has any join RTEs, replace join alias variables with
1297 : * base-relation variables. We must do this first, since any expressions
1298 : * we may extract from the joinaliasvars lists have not been preprocessed.
1299 : * For example, if we did this after sublink processing, sublinks expanded
1300 : * out from join aliases would not get processed. But we can skip this in
1301 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1302 : * they can't contain any Vars of the current query level.
1303 : */
1304 925660 : if (root->hasJoinRTEs &&
1305 408870 : !(kind == EXPRKIND_RTFUNC ||
1306 204262 : kind == EXPRKIND_VALUES ||
1307 : kind == EXPRKIND_TABLESAMPLE ||
1308 : kind == EXPRKIND_TABLEFUNC))
1309 204244 : expr = flatten_join_alias_vars(root, root->parse, expr);
1310 :
1311 : /*
1312 : * Simplify constant expressions. For function RTEs, this was already
1313 : * done by preprocess_function_rtes. (But note we must do it again for
1314 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1315 : * un-simplified subexpressions inserted by flattening of subqueries or
1316 : * join alias variables.)
1317 : *
1318 : * Note: an essential effect of this is to convert named-argument function
1319 : * calls to positional notation and insert the current actual values of
1320 : * any default arguments for functions. To ensure that happens, we *must*
1321 : * process all expressions here. Previous PG versions sometimes skipped
1322 : * const-simplification if it didn't seem worth the trouble, but we can't
1323 : * do that anymore.
1324 : *
1325 : * Note: this also flattens nested AND and OR expressions into N-argument
1326 : * form. All processing of a qual expression after this point must be
1327 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1328 : * with AND directly under AND, nor OR directly under OR.
1329 : */
1330 925660 : if (kind != EXPRKIND_RTFUNC)
1331 883126 : expr = eval_const_expressions(root, expr);
1332 :
1333 : /*
1334 : * If it's a qual or havingQual, canonicalize it.
1335 : */
1336 921710 : if (kind == EXPRKIND_QUAL)
1337 : {
1338 333072 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1339 :
1340 : #ifdef OPTIMIZER_DEBUG
1341 : printf("After canonicalize_qual()\n");
1342 : pprint(expr);
1343 : #endif
1344 : }
1345 :
1346 : /*
1347 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1348 : * hashfuncid of any that might execute more quickly by using hash lookups
1349 : * instead of a linear search.
1350 : */
1351 921710 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1352 : {
1353 842618 : convert_saop_to_hashed_saop(expr);
1354 : }
1355 :
1356 : /* Expand SubLinks to SubPlans */
1357 921710 : if (root->parse->hasSubLinks)
1358 104554 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1359 :
1360 : /*
1361 : * XXX do not insert anything here unless you have grokked the comments in
1362 : * SS_replace_correlation_vars ...
1363 : */
1364 :
1365 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1366 921710 : if (root->query_level > 1)
1367 172868 : expr = SS_replace_correlation_vars(root, expr);
1368 :
1369 : /*
1370 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1371 : * don't want to do this before eval_const_expressions, since the latter
1372 : * would be unable to simplify a top-level AND correctly. Also,
1373 : * SS_process_sublinks expects explicit-AND format.)
1374 : */
1375 921710 : if (kind == EXPRKIND_QUAL)
1376 333072 : expr = (Node *) make_ands_implicit((Expr *) expr);
1377 :
1378 921710 : return expr;
1379 : }
1380 :
1381 : /*
1382 : * preprocess_qual_conditions
1383 : * Recursively scan the query's jointree and do subquery_planner's
1384 : * preprocessing work on each qual condition found therein.
1385 : */
1386 : static void
1387 1305206 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1388 : {
1389 1305206 : if (jtnode == NULL)
1390 0 : return;
1391 1305206 : if (IsA(jtnode, RangeTblRef))
1392 : {
1393 : /* nothing to do here */
1394 : }
1395 637982 : else if (IsA(jtnode, FromExpr))
1396 : {
1397 537442 : FromExpr *f = (FromExpr *) jtnode;
1398 : ListCell *l;
1399 :
1400 1119172 : foreach(l, f->fromlist)
1401 581730 : preprocess_qual_conditions(root, lfirst(l));
1402 :
1403 537442 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1404 : }
1405 100540 : else if (IsA(jtnode, JoinExpr))
1406 : {
1407 100540 : JoinExpr *j = (JoinExpr *) jtnode;
1408 :
1409 100540 : preprocess_qual_conditions(root, j->larg);
1410 100540 : preprocess_qual_conditions(root, j->rarg);
1411 :
1412 100540 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1413 : }
1414 : else
1415 0 : elog(ERROR, "unrecognized node type: %d",
1416 : (int) nodeTag(jtnode));
1417 : }
1418 :
1419 : /*
1420 : * preprocess_phv_expression
1421 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1422 : *
1423 : * If a LATERAL subquery references an output of another subquery, and that
1424 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1425 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1426 : * and later pull it back up during find_lateral_references, which runs after
1427 : * subquery_planner has preprocessed all the expressions that were in the
1428 : * current query level to start with. So we need to preprocess it then.
1429 : */
1430 : Expr *
1431 90 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1432 : {
1433 90 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1434 : }
1435 :
1436 : /*--------------------
1437 : * grouping_planner
1438 : * Perform planning steps related to grouping, aggregation, etc.
1439 : *
1440 : * This function adds all required top-level processing to the scan/join
1441 : * Path(s) produced by query_planner.
1442 : *
1443 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1444 : * tuple_fraction is interpreted as follows:
1445 : * 0: expect all tuples to be retrieved (normal case)
1446 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1447 : * from the plan to be retrieved
1448 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1449 : * expected to be retrieved (ie, a LIMIT specification).
1450 : * setops is used for set operation subqueries to provide the subquery with
1451 : * the context in which it's being used so that Paths correctly sorted for the
1452 : * set operation can be generated. NULL when not planning a set operation
1453 : * child, or when a child of a set op that isn't interested in sorted input.
1454 : *
1455 : * Returns nothing; the useful output is in the Paths we attach to the
1456 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1457 : * root->processed_tlist contains the final processed targetlist.
1458 : *
1459 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1460 : * to leave this to the caller.
1461 : *--------------------
1462 : */
1463 : static void
1464 522396 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1465 : SetOperationStmt *setops)
1466 : {
1467 522396 : Query *parse = root->parse;
1468 522396 : int64 offset_est = 0;
1469 522396 : int64 count_est = 0;
1470 522396 : double limit_tuples = -1.0;
1471 522396 : bool have_postponed_srfs = false;
1472 : PathTarget *final_target;
1473 : List *final_targets;
1474 : List *final_targets_contain_srfs;
1475 : bool final_target_parallel_safe;
1476 : RelOptInfo *current_rel;
1477 : RelOptInfo *final_rel;
1478 : FinalPathExtraData extra;
1479 : ListCell *lc;
1480 :
1481 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1482 522396 : if (parse->limitCount || parse->limitOffset)
1483 : {
1484 5022 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1485 : &offset_est, &count_est);
1486 :
1487 : /*
1488 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1489 : * estimate the effects of using a bounded sort.
1490 : */
1491 5022 : if (count_est > 0 && offset_est >= 0)
1492 4472 : limit_tuples = (double) count_est + (double) offset_est;
1493 : }
1494 :
1495 : /* Make tuple_fraction accessible to lower-level routines */
1496 522396 : root->tuple_fraction = tuple_fraction;
1497 :
1498 522396 : if (parse->setOperations)
1499 : {
1500 : /*
1501 : * Construct Paths for set operations. The results will not need any
1502 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1503 : * special work for recursive unions is the responsibility of
1504 : * plan_set_operations.
1505 : */
1506 6102 : current_rel = plan_set_operations(root);
1507 :
1508 : /*
1509 : * We should not need to call preprocess_targetlist, since we must be
1510 : * in a SELECT query node. Instead, use the processed_tlist returned
1511 : * by plan_set_operations (since this tells whether it returned any
1512 : * resjunk columns!), and transfer any sort key information from the
1513 : * original tlist.
1514 : */
1515 : Assert(parse->commandType == CMD_SELECT);
1516 :
1517 : /* for safety, copy processed_tlist instead of modifying in-place */
1518 6096 : root->processed_tlist =
1519 6096 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1520 : parse->targetList);
1521 :
1522 : /* Also extract the PathTarget form of the setop result tlist */
1523 6096 : final_target = current_rel->cheapest_total_path->pathtarget;
1524 :
1525 : /* And check whether it's parallel safe */
1526 : final_target_parallel_safe =
1527 6096 : is_parallel_safe(root, (Node *) final_target->exprs);
1528 :
1529 : /* The setop result tlist couldn't contain any SRFs */
1530 : Assert(!parse->hasTargetSRFs);
1531 6096 : final_targets = final_targets_contain_srfs = NIL;
1532 :
1533 : /*
1534 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1535 : * checked already, but let's make sure).
1536 : */
1537 6096 : if (parse->rowMarks)
1538 0 : ereport(ERROR,
1539 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1540 : /*------
1541 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1542 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1543 : LCS_asString(linitial_node(RowMarkClause,
1544 : parse->rowMarks)->strength))));
1545 :
1546 : /*
1547 : * Calculate pathkeys that represent result ordering requirements
1548 : */
1549 : Assert(parse->distinctClause == NIL);
1550 6096 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1551 : parse->sortClause,
1552 : root->processed_tlist);
1553 : }
1554 : else
1555 : {
1556 : /* No set operations, do regular planning */
1557 : PathTarget *sort_input_target;
1558 : List *sort_input_targets;
1559 : List *sort_input_targets_contain_srfs;
1560 : bool sort_input_target_parallel_safe;
1561 : PathTarget *grouping_target;
1562 : List *grouping_targets;
1563 : List *grouping_targets_contain_srfs;
1564 : bool grouping_target_parallel_safe;
1565 : PathTarget *scanjoin_target;
1566 : List *scanjoin_targets;
1567 : List *scanjoin_targets_contain_srfs;
1568 : bool scanjoin_target_parallel_safe;
1569 : bool scanjoin_target_same_exprs;
1570 : bool have_grouping;
1571 516294 : WindowFuncLists *wflists = NULL;
1572 516294 : List *activeWindows = NIL;
1573 516294 : grouping_sets_data *gset_data = NULL;
1574 : standard_qp_extra qp_extra;
1575 :
1576 : /* A recursive query should always have setOperations */
1577 : Assert(!root->hasRecursion);
1578 :
1579 : /* Preprocess grouping sets and GROUP BY clause, if any */
1580 516294 : if (parse->groupingSets)
1581 : {
1582 938 : gset_data = preprocess_grouping_sets(root);
1583 : }
1584 515356 : else if (parse->groupClause)
1585 : {
1586 : /* Preprocess regular GROUP BY clause, if any */
1587 3910 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1588 : }
1589 :
1590 : /*
1591 : * Preprocess targetlist. Note that much of the remaining planning
1592 : * work will be done with the PathTarget representation of tlists, but
1593 : * we must also maintain the full representation of the final tlist so
1594 : * that we can transfer its decoration (resnames etc) to the topmost
1595 : * tlist of the finished Plan. This is kept in processed_tlist.
1596 : */
1597 516288 : preprocess_targetlist(root);
1598 :
1599 : /*
1600 : * Mark all the aggregates with resolved aggtranstypes, and detect
1601 : * aggregates that are duplicates or can share transition state. We
1602 : * must do this before slicing and dicing the tlist into various
1603 : * pathtargets, else some copies of the Aggref nodes might escape
1604 : * being marked.
1605 : */
1606 516288 : if (parse->hasAggs)
1607 : {
1608 38592 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1609 38592 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1610 : }
1611 :
1612 : /*
1613 : * Locate any window functions in the tlist. (We don't need to look
1614 : * anywhere else, since expressions used in ORDER BY will be in there
1615 : * too.) Note that they could all have been eliminated by constant
1616 : * folding, in which case we don't need to do any more work.
1617 : */
1618 516288 : if (parse->hasWindowFuncs)
1619 : {
1620 2570 : wflists = find_window_functions((Node *) root->processed_tlist,
1621 2570 : list_length(parse->windowClause));
1622 2570 : if (wflists->numWindowFuncs > 0)
1623 : {
1624 : /*
1625 : * See if any modifications can be made to each WindowClause
1626 : * to allow the executor to execute the WindowFuncs more
1627 : * quickly.
1628 : */
1629 2564 : optimize_window_clauses(root, wflists);
1630 :
1631 : /* Extract the list of windows actually in use. */
1632 2564 : activeWindows = select_active_windows(root, wflists);
1633 :
1634 : /* Make sure they all have names, for EXPLAIN's use. */
1635 2564 : name_active_windows(activeWindows);
1636 : }
1637 : else
1638 6 : parse->hasWindowFuncs = false;
1639 : }
1640 :
1641 : /*
1642 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1643 : * adding logic between here and the query_planner() call. Anything
1644 : * that is needed in MIN/MAX-optimizable cases will have to be
1645 : * duplicated in planagg.c.
1646 : */
1647 516288 : if (parse->hasAggs)
1648 38592 : preprocess_minmax_aggregates(root);
1649 :
1650 : /*
1651 : * Figure out whether there's a hard limit on the number of rows that
1652 : * query_planner's result subplan needs to return. Even if we know a
1653 : * hard limit overall, it doesn't apply if the query has any
1654 : * grouping/aggregation operations, or SRFs in the tlist.
1655 : */
1656 516288 : if (parse->groupClause ||
1657 511512 : parse->groupingSets ||
1658 511446 : parse->distinctClause ||
1659 508524 : parse->hasAggs ||
1660 474144 : parse->hasWindowFuncs ||
1661 471724 : parse->hasTargetSRFs ||
1662 460494 : root->hasHavingQual)
1663 55818 : root->limit_tuples = -1.0;
1664 : else
1665 460470 : root->limit_tuples = limit_tuples;
1666 :
1667 : /* Set up data needed by standard_qp_callback */
1668 516288 : qp_extra.activeWindows = activeWindows;
1669 516288 : qp_extra.gset_data = gset_data;
1670 :
1671 : /*
1672 : * If we're a subquery for a set operation, store the SetOperationStmt
1673 : * in qp_extra.
1674 : */
1675 516288 : qp_extra.setop = setops;
1676 :
1677 : /*
1678 : * Generate the best unsorted and presorted paths for the scan/join
1679 : * portion of this Query, ie the processing represented by the
1680 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1681 : * We also generate (in standard_qp_callback) pathkey representations
1682 : * of the query's sort clause, distinct clause, etc.
1683 : */
1684 516288 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1685 :
1686 : /*
1687 : * Convert the query's result tlist into PathTarget format.
1688 : *
1689 : * Note: this cannot be done before query_planner() has performed
1690 : * appendrel expansion, because that might add resjunk entries to
1691 : * root->processed_tlist. Waiting till afterwards is also helpful
1692 : * because the target width estimates can use per-Var width numbers
1693 : * that were obtained within query_planner().
1694 : */
1695 516234 : final_target = create_pathtarget(root, root->processed_tlist);
1696 : final_target_parallel_safe =
1697 516234 : is_parallel_safe(root, (Node *) final_target->exprs);
1698 :
1699 : /*
1700 : * If ORDER BY was given, consider whether we should use a post-sort
1701 : * projection, and compute the adjusted target for preceding steps if
1702 : * so.
1703 : */
1704 516234 : if (parse->sortClause)
1705 : {
1706 70148 : sort_input_target = make_sort_input_target(root,
1707 : final_target,
1708 : &have_postponed_srfs);
1709 : sort_input_target_parallel_safe =
1710 70148 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1711 : }
1712 : else
1713 : {
1714 446086 : sort_input_target = final_target;
1715 446086 : sort_input_target_parallel_safe = final_target_parallel_safe;
1716 : }
1717 :
1718 : /*
1719 : * If we have window functions to deal with, the output from any
1720 : * grouping step needs to be what the window functions want;
1721 : * otherwise, it should be sort_input_target.
1722 : */
1723 516234 : if (activeWindows)
1724 : {
1725 2564 : grouping_target = make_window_input_target(root,
1726 : final_target,
1727 : activeWindows);
1728 : grouping_target_parallel_safe =
1729 2564 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1730 : }
1731 : else
1732 : {
1733 513670 : grouping_target = sort_input_target;
1734 513670 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1735 : }
1736 :
1737 : /*
1738 : * If we have grouping or aggregation to do, the topmost scan/join
1739 : * plan node must emit what the grouping step wants; otherwise, it
1740 : * should emit grouping_target.
1741 : */
1742 511458 : have_grouping = (parse->groupClause || parse->groupingSets ||
1743 1027692 : parse->hasAggs || root->hasHavingQual);
1744 516234 : if (have_grouping)
1745 : {
1746 39284 : scanjoin_target = make_group_input_target(root, final_target);
1747 : scanjoin_target_parallel_safe =
1748 39284 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1749 : }
1750 : else
1751 : {
1752 476950 : scanjoin_target = grouping_target;
1753 476950 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1754 : }
1755 :
1756 : /*
1757 : * If there are any SRFs in the targetlist, we must separate each of
1758 : * these PathTargets into SRF-computing and SRF-free targets. Replace
1759 : * each of the named targets with a SRF-free version, and remember the
1760 : * list of additional projection steps we need to add afterwards.
1761 : */
1762 516234 : if (parse->hasTargetSRFs)
1763 : {
1764 : /* final_target doesn't recompute any SRFs in sort_input_target */
1765 11704 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1766 : &final_targets,
1767 : &final_targets_contain_srfs);
1768 11704 : final_target = linitial_node(PathTarget, final_targets);
1769 : Assert(!linitial_int(final_targets_contain_srfs));
1770 : /* likewise for sort_input_target vs. grouping_target */
1771 11704 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1772 : &sort_input_targets,
1773 : &sort_input_targets_contain_srfs);
1774 11704 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
1775 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1776 : /* likewise for grouping_target vs. scanjoin_target */
1777 11704 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
1778 : &grouping_targets,
1779 : &grouping_targets_contain_srfs);
1780 11704 : grouping_target = linitial_node(PathTarget, grouping_targets);
1781 : Assert(!linitial_int(grouping_targets_contain_srfs));
1782 : /* scanjoin_target will not have any SRFs precomputed for it */
1783 11704 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1784 : &scanjoin_targets,
1785 : &scanjoin_targets_contain_srfs);
1786 11704 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1787 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1788 : }
1789 : else
1790 : {
1791 : /* initialize lists; for most of these, dummy values are OK */
1792 504530 : final_targets = final_targets_contain_srfs = NIL;
1793 504530 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1794 504530 : grouping_targets = grouping_targets_contain_srfs = NIL;
1795 504530 : scanjoin_targets = list_make1(scanjoin_target);
1796 504530 : scanjoin_targets_contain_srfs = NIL;
1797 : }
1798 :
1799 : /* Apply scan/join target. */
1800 516234 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1801 516234 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1802 516234 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1803 : scanjoin_targets_contain_srfs,
1804 : scanjoin_target_parallel_safe,
1805 : scanjoin_target_same_exprs);
1806 :
1807 : /*
1808 : * Save the various upper-rel PathTargets we just computed into
1809 : * root->upper_targets[]. The core code doesn't use this, but it
1810 : * provides a convenient place for extensions to get at the info. For
1811 : * consistency, we save all the intermediate targets, even though some
1812 : * of the corresponding upperrels might not be needed for this query.
1813 : */
1814 516234 : root->upper_targets[UPPERREL_FINAL] = final_target;
1815 516234 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1816 516234 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1817 516234 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1818 516234 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1819 516234 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1820 :
1821 : /*
1822 : * If we have grouping and/or aggregation, consider ways to implement
1823 : * that. We build a new upperrel representing the output of this
1824 : * phase.
1825 : */
1826 516234 : if (have_grouping)
1827 : {
1828 39284 : current_rel = create_grouping_paths(root,
1829 : current_rel,
1830 : grouping_target,
1831 : grouping_target_parallel_safe,
1832 : gset_data);
1833 : /* Fix things up if grouping_target contains SRFs */
1834 39278 : if (parse->hasTargetSRFs)
1835 432 : adjust_paths_for_srfs(root, current_rel,
1836 : grouping_targets,
1837 : grouping_targets_contain_srfs);
1838 : }
1839 :
1840 : /*
1841 : * If we have window functions, consider ways to implement those. We
1842 : * build a new upperrel representing the output of this phase.
1843 : */
1844 516228 : if (activeWindows)
1845 : {
1846 2564 : current_rel = create_window_paths(root,
1847 : current_rel,
1848 : grouping_target,
1849 : sort_input_target,
1850 : sort_input_target_parallel_safe,
1851 : wflists,
1852 : activeWindows);
1853 : /* Fix things up if sort_input_target contains SRFs */
1854 2564 : if (parse->hasTargetSRFs)
1855 12 : adjust_paths_for_srfs(root, current_rel,
1856 : sort_input_targets,
1857 : sort_input_targets_contain_srfs);
1858 : }
1859 :
1860 : /*
1861 : * If there is a DISTINCT clause, consider ways to implement that. We
1862 : * build a new upperrel representing the output of this phase.
1863 : */
1864 516228 : if (parse->distinctClause)
1865 : {
1866 2956 : current_rel = create_distinct_paths(root,
1867 : current_rel,
1868 : sort_input_target);
1869 : }
1870 : } /* end of if (setOperations) */
1871 :
1872 : /*
1873 : * If ORDER BY was given, consider ways to implement that, and generate a
1874 : * new upperrel containing only paths that emit the correct ordering and
1875 : * project the correct final_target. We can apply the original
1876 : * limit_tuples limit in sort costing here, but only if there are no
1877 : * postponed SRFs.
1878 : */
1879 522324 : if (parse->sortClause)
1880 : {
1881 74036 : current_rel = create_ordered_paths(root,
1882 : current_rel,
1883 : final_target,
1884 : final_target_parallel_safe,
1885 : have_postponed_srfs ? -1.0 :
1886 : limit_tuples);
1887 : /* Fix things up if final_target contains SRFs */
1888 74036 : if (parse->hasTargetSRFs)
1889 196 : adjust_paths_for_srfs(root, current_rel,
1890 : final_targets,
1891 : final_targets_contain_srfs);
1892 : }
1893 :
1894 : /*
1895 : * Now we are prepared to build the final-output upperrel.
1896 : */
1897 522324 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1898 :
1899 : /*
1900 : * If the input rel is marked consider_parallel and there's nothing that's
1901 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1902 : * consider_parallel as well. Note that if the query has rowMarks or is
1903 : * not a SELECT, consider_parallel will be false for every relation in the
1904 : * query.
1905 : */
1906 697954 : if (current_rel->consider_parallel &&
1907 351236 : is_parallel_safe(root, parse->limitOffset) &&
1908 175606 : is_parallel_safe(root, parse->limitCount))
1909 175600 : final_rel->consider_parallel = true;
1910 :
1911 : /*
1912 : * If the current_rel belongs to a single FDW, so does the final_rel.
1913 : */
1914 522324 : final_rel->serverid = current_rel->serverid;
1915 522324 : final_rel->userid = current_rel->userid;
1916 522324 : final_rel->useridiscurrent = current_rel->useridiscurrent;
1917 522324 : final_rel->fdwroutine = current_rel->fdwroutine;
1918 :
1919 : /*
1920 : * Generate paths for the final_rel. Insert all surviving paths, with
1921 : * LockRows, Limit, and/or ModifyTable steps added if needed.
1922 : */
1923 1065534 : foreach(lc, current_rel->pathlist)
1924 : {
1925 543210 : Path *path = (Path *) lfirst(lc);
1926 :
1927 : /*
1928 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1929 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
1930 : * here. If there are only non-locking rowmarks, they should be
1931 : * handled by the ModifyTable node instead. However, root->rowMarks
1932 : * is what goes into the LockRows node.)
1933 : */
1934 543210 : if (parse->rowMarks)
1935 : {
1936 8350 : path = (Path *) create_lockrows_path(root, final_rel, path,
1937 : root->rowMarks,
1938 : assign_special_exec_param(root));
1939 : }
1940 :
1941 : /*
1942 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1943 : */
1944 543210 : if (limit_needed(parse))
1945 : {
1946 5950 : path = (Path *) create_limit_path(root, final_rel, path,
1947 : parse->limitOffset,
1948 : parse->limitCount,
1949 : parse->limitOption,
1950 : offset_est, count_est);
1951 : }
1952 :
1953 : /*
1954 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
1955 : */
1956 543210 : if (parse->commandType != CMD_SELECT)
1957 : {
1958 : Index rootRelation;
1959 88842 : List *resultRelations = NIL;
1960 88842 : List *updateColnosLists = NIL;
1961 88842 : List *withCheckOptionLists = NIL;
1962 88842 : List *returningLists = NIL;
1963 88842 : List *mergeActionLists = NIL;
1964 88842 : List *mergeJoinConditions = NIL;
1965 : List *rowMarks;
1966 :
1967 88842 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
1968 : {
1969 : /* Inherited UPDATE/DELETE/MERGE */
1970 2868 : RelOptInfo *top_result_rel = find_base_rel(root,
1971 : parse->resultRelation);
1972 2868 : int resultRelation = -1;
1973 :
1974 : /* Pass the root result rel forward to the executor. */
1975 2868 : rootRelation = parse->resultRelation;
1976 :
1977 : /* Add only leaf children to ModifyTable. */
1978 8374 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
1979 8374 : resultRelation)) >= 0)
1980 : {
1981 5506 : RelOptInfo *this_result_rel = find_base_rel(root,
1982 : resultRelation);
1983 :
1984 : /*
1985 : * Also exclude any leaf rels that have turned dummy since
1986 : * being added to the list, for example, by being excluded
1987 : * by constraint exclusion.
1988 : */
1989 5506 : if (IS_DUMMY_REL(this_result_rel))
1990 174 : continue;
1991 :
1992 : /* Build per-target-rel lists needed by ModifyTable */
1993 5332 : resultRelations = lappend_int(resultRelations,
1994 : resultRelation);
1995 5332 : if (parse->commandType == CMD_UPDATE)
1996 : {
1997 3660 : List *update_colnos = root->update_colnos;
1998 :
1999 3660 : if (this_result_rel != top_result_rel)
2000 : update_colnos =
2001 3660 : adjust_inherited_attnums_multilevel(root,
2002 : update_colnos,
2003 : this_result_rel->relid,
2004 : top_result_rel->relid);
2005 3660 : updateColnosLists = lappend(updateColnosLists,
2006 : update_colnos);
2007 : }
2008 5332 : if (parse->withCheckOptions)
2009 : {
2010 504 : List *withCheckOptions = parse->withCheckOptions;
2011 :
2012 504 : if (this_result_rel != top_result_rel)
2013 : withCheckOptions = (List *)
2014 504 : adjust_appendrel_attrs_multilevel(root,
2015 : (Node *) withCheckOptions,
2016 : this_result_rel,
2017 : top_result_rel);
2018 504 : withCheckOptionLists = lappend(withCheckOptionLists,
2019 : withCheckOptions);
2020 : }
2021 5332 : if (parse->returningList)
2022 : {
2023 840 : List *returningList = parse->returningList;
2024 :
2025 840 : if (this_result_rel != top_result_rel)
2026 : returningList = (List *)
2027 840 : adjust_appendrel_attrs_multilevel(root,
2028 : (Node *) returningList,
2029 : this_result_rel,
2030 : top_result_rel);
2031 840 : returningLists = lappend(returningLists,
2032 : returningList);
2033 : }
2034 5332 : if (parse->mergeActionList)
2035 : {
2036 : ListCell *l;
2037 534 : List *mergeActionList = NIL;
2038 :
2039 : /*
2040 : * Copy MergeActions and translate stuff that
2041 : * references attribute numbers.
2042 : */
2043 1662 : foreach(l, parse->mergeActionList)
2044 : {
2045 1128 : MergeAction *action = lfirst(l),
2046 1128 : *leaf_action = copyObject(action);
2047 :
2048 1128 : leaf_action->qual =
2049 1128 : adjust_appendrel_attrs_multilevel(root,
2050 : (Node *) action->qual,
2051 : this_result_rel,
2052 : top_result_rel);
2053 1128 : leaf_action->targetList = (List *)
2054 1128 : adjust_appendrel_attrs_multilevel(root,
2055 1128 : (Node *) action->targetList,
2056 : this_result_rel,
2057 : top_result_rel);
2058 1128 : if (leaf_action->commandType == CMD_UPDATE)
2059 628 : leaf_action->updateColnos =
2060 628 : adjust_inherited_attnums_multilevel(root,
2061 : action->updateColnos,
2062 : this_result_rel->relid,
2063 : top_result_rel->relid);
2064 1128 : mergeActionList = lappend(mergeActionList,
2065 : leaf_action);
2066 : }
2067 :
2068 534 : mergeActionLists = lappend(mergeActionLists,
2069 : mergeActionList);
2070 : }
2071 5332 : if (parse->commandType == CMD_MERGE)
2072 : {
2073 534 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2074 :
2075 534 : if (this_result_rel != top_result_rel)
2076 : mergeJoinCondition =
2077 534 : adjust_appendrel_attrs_multilevel(root,
2078 : mergeJoinCondition,
2079 : this_result_rel,
2080 : top_result_rel);
2081 534 : mergeJoinConditions = lappend(mergeJoinConditions,
2082 : mergeJoinCondition);
2083 : }
2084 : }
2085 :
2086 2868 : if (resultRelations == NIL)
2087 : {
2088 : /*
2089 : * We managed to exclude every child rel, so generate a
2090 : * dummy one-relation plan using info for the top target
2091 : * rel (even though that may not be a leaf target).
2092 : * Although it's clear that no data will be updated or
2093 : * deleted, we still need to have a ModifyTable node so
2094 : * that any statement triggers will be executed. (This
2095 : * could be cleaner if we fixed nodeModifyTable.c to allow
2096 : * zero target relations, but that probably wouldn't be a
2097 : * net win.)
2098 : */
2099 30 : resultRelations = list_make1_int(parse->resultRelation);
2100 30 : if (parse->commandType == CMD_UPDATE)
2101 30 : updateColnosLists = list_make1(root->update_colnos);
2102 30 : if (parse->withCheckOptions)
2103 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2104 30 : if (parse->returningList)
2105 18 : returningLists = list_make1(parse->returningList);
2106 30 : if (parse->mergeActionList)
2107 0 : mergeActionLists = list_make1(parse->mergeActionList);
2108 30 : if (parse->commandType == CMD_MERGE)
2109 0 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2110 : }
2111 : }
2112 : else
2113 : {
2114 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2115 85974 : rootRelation = 0; /* there's no separate root rel */
2116 85974 : resultRelations = list_make1_int(parse->resultRelation);
2117 85974 : if (parse->commandType == CMD_UPDATE)
2118 12092 : updateColnosLists = list_make1(root->update_colnos);
2119 85974 : if (parse->withCheckOptions)
2120 926 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2121 85974 : if (parse->returningList)
2122 2438 : returningLists = list_make1(parse->returningList);
2123 85974 : if (parse->mergeActionList)
2124 1542 : mergeActionLists = list_make1(parse->mergeActionList);
2125 85974 : if (parse->commandType == CMD_MERGE)
2126 1542 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2127 : }
2128 :
2129 : /*
2130 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2131 : * will have dealt with fetching non-locked marked rows, else we
2132 : * need to have ModifyTable do that.
2133 : */
2134 88842 : if (parse->rowMarks)
2135 0 : rowMarks = NIL;
2136 : else
2137 88842 : rowMarks = root->rowMarks;
2138 :
2139 : path = (Path *)
2140 88842 : create_modifytable_path(root, final_rel,
2141 : path,
2142 : parse->commandType,
2143 88842 : parse->canSetTag,
2144 88842 : parse->resultRelation,
2145 : rootRelation,
2146 : resultRelations,
2147 : updateColnosLists,
2148 : withCheckOptionLists,
2149 : returningLists,
2150 : rowMarks,
2151 : parse->onConflict,
2152 : mergeActionLists,
2153 : mergeJoinConditions,
2154 : assign_special_exec_param(root));
2155 : }
2156 :
2157 : /* And shove it into final_rel */
2158 543210 : add_path(final_rel, path);
2159 : }
2160 :
2161 : /*
2162 : * Generate partial paths for final_rel, too, if outer query levels might
2163 : * be able to make use of them.
2164 : */
2165 522324 : if (final_rel->consider_parallel && root->query_level > 1 &&
2166 30108 : !limit_needed(parse))
2167 : {
2168 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2169 30030 : foreach(lc, current_rel->partial_pathlist)
2170 : {
2171 108 : Path *partial_path = (Path *) lfirst(lc);
2172 :
2173 108 : add_partial_path(final_rel, partial_path);
2174 : }
2175 : }
2176 :
2177 522324 : extra.limit_needed = limit_needed(parse);
2178 522324 : extra.limit_tuples = limit_tuples;
2179 522324 : extra.count_est = count_est;
2180 522324 : extra.offset_est = offset_est;
2181 :
2182 : /*
2183 : * If there is an FDW that's responsible for all baserels of the query,
2184 : * let it consider adding ForeignPaths.
2185 : */
2186 522324 : if (final_rel->fdwroutine &&
2187 1260 : final_rel->fdwroutine->GetForeignUpperPaths)
2188 1192 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2189 : current_rel, final_rel,
2190 : &extra);
2191 :
2192 : /* Let extensions possibly add some more paths */
2193 522324 : if (create_upper_paths_hook)
2194 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2195 : current_rel, final_rel, &extra);
2196 :
2197 : /* Note: currently, we leave it to callers to do set_cheapest() */
2198 522324 : }
2199 :
2200 : /*
2201 : * Do preprocessing for groupingSets clause and related data.
2202 : *
2203 : * We expect that parse->groupingSets has already been expanded into a flat
2204 : * list of grouping sets (that is, just integer Lists of ressortgroupref
2205 : * numbers) by expand_grouping_sets(). This function handles the preliminary
2206 : * steps of organizing the grouping sets into lists of rollups, and preparing
2207 : * annotations which will later be filled in with size estimates.
2208 : */
2209 : static grouping_sets_data *
2210 938 : preprocess_grouping_sets(PlannerInfo *root)
2211 : {
2212 938 : Query *parse = root->parse;
2213 : List *sets;
2214 938 : int maxref = 0;
2215 : ListCell *lc_set;
2216 938 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2217 :
2218 : /*
2219 : * We don't currently make any attempt to optimize the groupClause when
2220 : * there are grouping sets, so just duplicate it in processed_groupClause.
2221 : */
2222 938 : root->processed_groupClause = parse->groupClause;
2223 :
2224 : /* Detect unhashable and unsortable grouping expressions */
2225 938 : gd->any_hashable = false;
2226 938 : gd->unhashable_refs = NULL;
2227 938 : gd->unsortable_refs = NULL;
2228 938 : gd->unsortable_sets = NIL;
2229 :
2230 938 : if (parse->groupClause)
2231 : {
2232 : ListCell *lc;
2233 :
2234 2780 : foreach(lc, parse->groupClause)
2235 : {
2236 1908 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2237 1908 : Index ref = gc->tleSortGroupRef;
2238 :
2239 1908 : if (ref > maxref)
2240 1872 : maxref = ref;
2241 :
2242 1908 : if (!gc->hashable)
2243 30 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2244 :
2245 1908 : if (!OidIsValid(gc->sortop))
2246 42 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2247 : }
2248 : }
2249 :
2250 : /* Allocate workspace array for remapping */
2251 938 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2252 :
2253 : /*
2254 : * If we have any unsortable sets, we must extract them before trying to
2255 : * prepare rollups. Unsortable sets don't go through
2256 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2257 : * here.
2258 : */
2259 938 : if (!bms_is_empty(gd->unsortable_refs))
2260 : {
2261 42 : List *sortable_sets = NIL;
2262 : ListCell *lc;
2263 :
2264 126 : foreach(lc, parse->groupingSets)
2265 : {
2266 90 : List *gset = (List *) lfirst(lc);
2267 :
2268 90 : if (bms_overlap_list(gd->unsortable_refs, gset))
2269 : {
2270 48 : GroupingSetData *gs = makeNode(GroupingSetData);
2271 :
2272 48 : gs->set = gset;
2273 48 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2274 :
2275 : /*
2276 : * We must enforce here that an unsortable set is hashable;
2277 : * later code assumes this. Parse analysis only checks that
2278 : * every individual column is either hashable or sortable.
2279 : *
2280 : * Note that passing this test doesn't guarantee we can
2281 : * generate a plan; there might be other showstoppers.
2282 : */
2283 48 : if (bms_overlap_list(gd->unhashable_refs, gset))
2284 6 : ereport(ERROR,
2285 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2286 : errmsg("could not implement GROUP BY"),
2287 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2288 : }
2289 : else
2290 42 : sortable_sets = lappend(sortable_sets, gset);
2291 : }
2292 :
2293 36 : if (sortable_sets)
2294 30 : sets = extract_rollup_sets(sortable_sets);
2295 : else
2296 6 : sets = NIL;
2297 : }
2298 : else
2299 896 : sets = extract_rollup_sets(parse->groupingSets);
2300 :
2301 2442 : foreach(lc_set, sets)
2302 : {
2303 1510 : List *current_sets = (List *) lfirst(lc_set);
2304 1510 : RollupData *rollup = makeNode(RollupData);
2305 : GroupingSetData *gs;
2306 :
2307 : /*
2308 : * Reorder the current list of grouping sets into correct prefix
2309 : * order. If only one aggregation pass is needed, try to make the
2310 : * list match the ORDER BY clause; if more than one pass is needed, we
2311 : * don't bother with that.
2312 : *
2313 : * Note that this reorders the sets from smallest-member-first to
2314 : * largest-member-first, and applies the GroupingSetData annotations,
2315 : * though the data will be filled in later.
2316 : */
2317 1510 : current_sets = reorder_grouping_sets(current_sets,
2318 1510 : (list_length(sets) == 1
2319 : ? parse->sortClause
2320 : : NIL));
2321 :
2322 : /*
2323 : * Get the initial (and therefore largest) grouping set.
2324 : */
2325 1510 : gs = linitial_node(GroupingSetData, current_sets);
2326 :
2327 : /*
2328 : * Order the groupClause appropriately. If the first grouping set is
2329 : * empty, then the groupClause must also be empty; otherwise we have
2330 : * to force the groupClause to match that grouping set's order.
2331 : *
2332 : * (The first grouping set can be empty even though parse->groupClause
2333 : * is not empty only if all non-empty grouping sets are unsortable.
2334 : * The groupClauses for hashed grouping sets are built later on.)
2335 : */
2336 1510 : if (gs->set)
2337 1444 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2338 : else
2339 66 : rollup->groupClause = NIL;
2340 :
2341 : /*
2342 : * Is it hashable? We pretend empty sets are hashable even though we
2343 : * actually force them not to be hashed later. But don't bother if
2344 : * there's nothing but empty sets (since in that case we can't hash
2345 : * anything).
2346 : */
2347 1510 : if (gs->set &&
2348 1444 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2349 : {
2350 1420 : rollup->hashable = true;
2351 1420 : gd->any_hashable = true;
2352 : }
2353 :
2354 : /*
2355 : * Now that we've pinned down an order for the groupClause for this
2356 : * list of grouping sets, we need to remap the entries in the grouping
2357 : * sets from sortgrouprefs to plain indices (0-based) into the
2358 : * groupClause for this collection of grouping sets. We keep the
2359 : * original form for later use, though.
2360 : */
2361 1510 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2362 : current_sets,
2363 : gd->tleref_to_colnum_map);
2364 1510 : rollup->gsets_data = current_sets;
2365 :
2366 1510 : gd->rollups = lappend(gd->rollups, rollup);
2367 : }
2368 :
2369 932 : if (gd->unsortable_sets)
2370 : {
2371 : /*
2372 : * We have not yet pinned down a groupclause for this, but we will
2373 : * need index-based lists for estimation purposes. Construct
2374 : * hash_sets_idx based on the entire original groupclause for now.
2375 : */
2376 36 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2377 : gd->unsortable_sets,
2378 : gd->tleref_to_colnum_map);
2379 36 : gd->any_hashable = true;
2380 : }
2381 :
2382 932 : return gd;
2383 : }
2384 :
2385 : /*
2386 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2387 : * (without annotation) mapped to indexes into the given groupclause.
2388 : */
2389 : static List *
2390 4344 : remap_to_groupclause_idx(List *groupClause,
2391 : List *gsets,
2392 : int *tleref_to_colnum_map)
2393 : {
2394 4344 : int ref = 0;
2395 4344 : List *result = NIL;
2396 : ListCell *lc;
2397 :
2398 10552 : foreach(lc, groupClause)
2399 : {
2400 6208 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2401 :
2402 6208 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2403 : }
2404 :
2405 10014 : foreach(lc, gsets)
2406 : {
2407 5670 : List *set = NIL;
2408 : ListCell *lc2;
2409 5670 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2410 :
2411 12728 : foreach(lc2, gs->set)
2412 : {
2413 7058 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2414 : }
2415 :
2416 5670 : result = lappend(result, set);
2417 : }
2418 :
2419 4344 : return result;
2420 : }
2421 :
2422 :
2423 : /*
2424 : * preprocess_rowmarks - set up PlanRowMarks if needed
2425 : */
2426 : static void
2427 526346 : preprocess_rowmarks(PlannerInfo *root)
2428 : {
2429 526346 : Query *parse = root->parse;
2430 : Bitmapset *rels;
2431 : List *prowmarks;
2432 : ListCell *l;
2433 : int i;
2434 :
2435 526346 : if (parse->rowMarks)
2436 : {
2437 : /*
2438 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2439 : * grouping, since grouping renders a reference to individual tuple
2440 : * CTIDs invalid. This is also checked at parse time, but that's
2441 : * insufficient because of rule substitution, query pullup, etc.
2442 : */
2443 7862 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2444 : parse->rowMarks)->strength);
2445 : }
2446 : else
2447 : {
2448 : /*
2449 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2450 : * UPDATE/SHARE.
2451 : */
2452 518484 : if (parse->commandType != CMD_UPDATE &&
2453 504386 : parse->commandType != CMD_DELETE &&
2454 500038 : parse->commandType != CMD_MERGE)
2455 498250 : return;
2456 : }
2457 :
2458 : /*
2459 : * We need to have rowmarks for all base relations except the target. We
2460 : * make a bitmapset of all base rels and then remove the items we don't
2461 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2462 : */
2463 28096 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2464 28096 : if (parse->resultRelation)
2465 20234 : rels = bms_del_member(rels, parse->resultRelation);
2466 :
2467 : /*
2468 : * Convert RowMarkClauses to PlanRowMark representation.
2469 : */
2470 28096 : prowmarks = NIL;
2471 36188 : foreach(l, parse->rowMarks)
2472 : {
2473 8092 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2474 8092 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2475 : PlanRowMark *newrc;
2476 :
2477 : /*
2478 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2479 : * applied to an update/delete target rel. If that ever becomes
2480 : * possible, we should drop the target from the PlanRowMark list.
2481 : */
2482 : Assert(rc->rti != parse->resultRelation);
2483 :
2484 : /*
2485 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2486 : * can't support true locking. Subqueries that got flattened into the
2487 : * main query should be ignored completely. Any that didn't will get
2488 : * ROW_MARK_COPY items in the next loop.
2489 : */
2490 8092 : if (rte->rtekind != RTE_RELATION)
2491 60 : continue;
2492 :
2493 8032 : rels = bms_del_member(rels, rc->rti);
2494 :
2495 8032 : newrc = makeNode(PlanRowMark);
2496 8032 : newrc->rti = newrc->prti = rc->rti;
2497 8032 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2498 8032 : newrc->markType = select_rowmark_type(rte, rc->strength);
2499 8032 : newrc->allMarkTypes = (1 << newrc->markType);
2500 8032 : newrc->strength = rc->strength;
2501 8032 : newrc->waitPolicy = rc->waitPolicy;
2502 8032 : newrc->isParent = false;
2503 :
2504 8032 : prowmarks = lappend(prowmarks, newrc);
2505 : }
2506 :
2507 : /*
2508 : * Now, add rowmarks for any non-target, non-locked base relations.
2509 : */
2510 28096 : i = 0;
2511 66870 : foreach(l, parse->rtable)
2512 : {
2513 38774 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2514 : PlanRowMark *newrc;
2515 :
2516 38774 : i++;
2517 38774 : if (!bms_is_member(i, rels))
2518 35096 : continue;
2519 :
2520 3678 : newrc = makeNode(PlanRowMark);
2521 3678 : newrc->rti = newrc->prti = i;
2522 3678 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2523 3678 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2524 3678 : newrc->allMarkTypes = (1 << newrc->markType);
2525 3678 : newrc->strength = LCS_NONE;
2526 3678 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2527 3678 : newrc->isParent = false;
2528 :
2529 3678 : prowmarks = lappend(prowmarks, newrc);
2530 : }
2531 :
2532 28096 : root->rowMarks = prowmarks;
2533 : }
2534 :
2535 : /*
2536 : * Select RowMarkType to use for a given table
2537 : */
2538 : RowMarkType
2539 14042 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2540 : {
2541 14042 : if (rte->rtekind != RTE_RELATION)
2542 : {
2543 : /* If it's not a table at all, use ROW_MARK_COPY */
2544 1404 : return ROW_MARK_COPY;
2545 : }
2546 12638 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2547 : {
2548 : /* Let the FDW select the rowmark type, if it wants to */
2549 212 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2550 :
2551 212 : if (fdwroutine->GetForeignRowMarkType != NULL)
2552 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2553 : /* Otherwise, use ROW_MARK_COPY by default */
2554 212 : return ROW_MARK_COPY;
2555 : }
2556 : else
2557 : {
2558 : /* Regular table, apply the appropriate lock type */
2559 12426 : switch (strength)
2560 : {
2561 2504 : case LCS_NONE:
2562 :
2563 : /*
2564 : * We don't need a tuple lock, only the ability to re-fetch
2565 : * the row.
2566 : */
2567 2504 : return ROW_MARK_REFERENCE;
2568 : break;
2569 8038 : case LCS_FORKEYSHARE:
2570 8038 : return ROW_MARK_KEYSHARE;
2571 : break;
2572 300 : case LCS_FORSHARE:
2573 300 : return ROW_MARK_SHARE;
2574 : break;
2575 72 : case LCS_FORNOKEYUPDATE:
2576 72 : return ROW_MARK_NOKEYEXCLUSIVE;
2577 : break;
2578 1512 : case LCS_FORUPDATE:
2579 1512 : return ROW_MARK_EXCLUSIVE;
2580 : break;
2581 : }
2582 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2583 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2584 : }
2585 : }
2586 :
2587 : /*
2588 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2589 : *
2590 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2591 : * results back in *count_est and *offset_est. These variables are set to
2592 : * 0 if the corresponding clause is not present, and -1 if it's present
2593 : * but we couldn't estimate the value for it. (The "0" convention is OK
2594 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2595 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2596 : * usual practice of never estimating less than one row.) These values will
2597 : * be passed to create_limit_path, which see if you change this code.
2598 : *
2599 : * The return value is the suitably adjusted tuple_fraction to use for
2600 : * planning the query. This adjustment is not overridable, since it reflects
2601 : * plan actions that grouping_planner() will certainly take, not assumptions
2602 : * about context.
2603 : */
2604 : static double
2605 5022 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2606 : int64 *offset_est, int64 *count_est)
2607 : {
2608 5022 : Query *parse = root->parse;
2609 : Node *est;
2610 : double limit_fraction;
2611 :
2612 : /* Should not be called unless LIMIT or OFFSET */
2613 : Assert(parse->limitCount || parse->limitOffset);
2614 :
2615 : /*
2616 : * Try to obtain the clause values. We use estimate_expression_value
2617 : * primarily because it can sometimes do something useful with Params.
2618 : */
2619 5022 : if (parse->limitCount)
2620 : {
2621 4496 : est = estimate_expression_value(root, parse->limitCount);
2622 4496 : if (est && IsA(est, Const))
2623 : {
2624 4490 : if (((Const *) est)->constisnull)
2625 : {
2626 : /* NULL indicates LIMIT ALL, ie, no limit */
2627 0 : *count_est = 0; /* treat as not present */
2628 : }
2629 : else
2630 : {
2631 4490 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2632 4490 : if (*count_est <= 0)
2633 150 : *count_est = 1; /* force to at least 1 */
2634 : }
2635 : }
2636 : else
2637 6 : *count_est = -1; /* can't estimate */
2638 : }
2639 : else
2640 526 : *count_est = 0; /* not present */
2641 :
2642 5022 : if (parse->limitOffset)
2643 : {
2644 898 : est = estimate_expression_value(root, parse->limitOffset);
2645 898 : if (est && IsA(est, Const))
2646 : {
2647 874 : if (((Const *) est)->constisnull)
2648 : {
2649 : /* Treat NULL as no offset; the executor will too */
2650 0 : *offset_est = 0; /* treat as not present */
2651 : }
2652 : else
2653 : {
2654 874 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2655 874 : if (*offset_est < 0)
2656 0 : *offset_est = 0; /* treat as not present */
2657 : }
2658 : }
2659 : else
2660 24 : *offset_est = -1; /* can't estimate */
2661 : }
2662 : else
2663 4124 : *offset_est = 0; /* not present */
2664 :
2665 5022 : if (*count_est != 0)
2666 : {
2667 : /*
2668 : * A LIMIT clause limits the absolute number of tuples returned.
2669 : * However, if it's not a constant LIMIT then we have to guess; for
2670 : * lack of a better idea, assume 10% of the plan's result is wanted.
2671 : */
2672 4496 : if (*count_est < 0 || *offset_est < 0)
2673 : {
2674 : /* LIMIT or OFFSET is an expression ... punt ... */
2675 24 : limit_fraction = 0.10;
2676 : }
2677 : else
2678 : {
2679 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2680 4472 : limit_fraction = (double) *count_est + (double) *offset_est;
2681 : }
2682 :
2683 : /*
2684 : * If we have absolute limits from both caller and LIMIT, use the
2685 : * smaller value; likewise if they are both fractional. If one is
2686 : * fractional and the other absolute, we can't easily determine which
2687 : * is smaller, but we use the heuristic that the absolute will usually
2688 : * be smaller.
2689 : */
2690 4496 : if (tuple_fraction >= 1.0)
2691 : {
2692 6 : if (limit_fraction >= 1.0)
2693 : {
2694 : /* both absolute */
2695 6 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2696 : }
2697 : else
2698 : {
2699 : /* caller absolute, limit fractional; use caller's value */
2700 : }
2701 : }
2702 4490 : else if (tuple_fraction > 0.0)
2703 : {
2704 148 : if (limit_fraction >= 1.0)
2705 : {
2706 : /* caller fractional, limit absolute; use limit */
2707 148 : tuple_fraction = limit_fraction;
2708 : }
2709 : else
2710 : {
2711 : /* both fractional */
2712 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2713 : }
2714 : }
2715 : else
2716 : {
2717 : /* no info from caller, just use limit */
2718 4342 : tuple_fraction = limit_fraction;
2719 : }
2720 : }
2721 526 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2722 : {
2723 : /*
2724 : * We have an OFFSET but no LIMIT. This acts entirely differently
2725 : * from the LIMIT case: here, we need to increase rather than decrease
2726 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2727 : * tuples to be fetched instead of fewer. This only matters if we got
2728 : * a tuple_fraction > 0, however.
2729 : *
2730 : * As above, use 10% if OFFSET is present but unestimatable.
2731 : */
2732 14 : if (*offset_est < 0)
2733 0 : limit_fraction = 0.10;
2734 : else
2735 14 : limit_fraction = (double) *offset_est;
2736 :
2737 : /*
2738 : * If we have absolute counts from both caller and OFFSET, add them
2739 : * together; likewise if they are both fractional. If one is
2740 : * fractional and the other absolute, we want to take the larger, and
2741 : * we heuristically assume that's the fractional one.
2742 : */
2743 14 : if (tuple_fraction >= 1.0)
2744 : {
2745 0 : if (limit_fraction >= 1.0)
2746 : {
2747 : /* both absolute, so add them together */
2748 0 : tuple_fraction += limit_fraction;
2749 : }
2750 : else
2751 : {
2752 : /* caller absolute, limit fractional; use limit */
2753 0 : tuple_fraction = limit_fraction;
2754 : }
2755 : }
2756 : else
2757 : {
2758 14 : if (limit_fraction >= 1.0)
2759 : {
2760 : /* caller fractional, limit absolute; use caller's value */
2761 : }
2762 : else
2763 : {
2764 : /* both fractional, so add them together */
2765 0 : tuple_fraction += limit_fraction;
2766 0 : if (tuple_fraction >= 1.0)
2767 0 : tuple_fraction = 0.0; /* assume fetch all */
2768 : }
2769 : }
2770 : }
2771 :
2772 5022 : return tuple_fraction;
2773 : }
2774 :
2775 : /*
2776 : * limit_needed - do we actually need a Limit plan node?
2777 : *
2778 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2779 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2780 : * locution for an optimization fence. (Because other places in the planner
2781 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2782 : * an optimization fence --- we're just suppressing unnecessary run-time
2783 : * overhead.)
2784 : *
2785 : * This might look like it could be merged into preprocess_limit, but there's
2786 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2787 : * in preprocess_limit it's good enough to consider estimated values.
2788 : */
2789 : bool
2790 1114742 : limit_needed(Query *parse)
2791 : {
2792 : Node *node;
2793 :
2794 1114742 : node = parse->limitCount;
2795 1114742 : if (node)
2796 : {
2797 10764 : if (IsA(node, Const))
2798 : {
2799 : /* NULL indicates LIMIT ALL, ie, no limit */
2800 10528 : if (!((Const *) node)->constisnull)
2801 10528 : return true; /* LIMIT with a constant value */
2802 : }
2803 : else
2804 236 : return true; /* non-constant LIMIT */
2805 : }
2806 :
2807 1103978 : node = parse->limitOffset;
2808 1103978 : if (node)
2809 : {
2810 1530 : if (IsA(node, Const))
2811 : {
2812 : /* Treat NULL as no offset; the executor would too */
2813 1222 : if (!((Const *) node)->constisnull)
2814 : {
2815 1222 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2816 :
2817 1222 : if (offset != 0)
2818 142 : return true; /* OFFSET with a nonzero value */
2819 : }
2820 : }
2821 : else
2822 308 : return true; /* non-constant OFFSET */
2823 : }
2824 :
2825 1103528 : return false; /* don't need a Limit plan node */
2826 : }
2827 :
2828 : /*
2829 : * preprocess_groupclause - do preparatory work on GROUP BY clause
2830 : *
2831 : * The idea here is to adjust the ordering of the GROUP BY elements
2832 : * (which in itself is semantically insignificant) to match ORDER BY,
2833 : * thereby allowing a single sort operation to both implement the ORDER BY
2834 : * requirement and set up for a Unique step that implements GROUP BY.
2835 : * We also consider partial match between GROUP BY and ORDER BY elements,
2836 : * which could allow to implement ORDER BY using the incremental sort.
2837 : *
2838 : * We also consider other orderings of the GROUP BY elements, which could
2839 : * match the sort ordering of other possible plans (eg an indexscan) and
2840 : * thereby reduce cost. This is implemented during the generation of grouping
2841 : * paths. See get_useful_group_keys_orderings() for details.
2842 : *
2843 : * Note: we need no comparable processing of the distinctClause because
2844 : * the parser already enforced that that matches ORDER BY.
2845 : *
2846 : * Note: we return a fresh List, but its elements are the same
2847 : * SortGroupClauses appearing in parse->groupClause. This is important
2848 : * because later processing may modify the processed_groupClause list.
2849 : *
2850 : * For grouping sets, the order of items is instead forced to agree with that
2851 : * of the grouping set (and items not in the grouping set are skipped). The
2852 : * work of sorting the order of grouping set elements to match the ORDER BY if
2853 : * possible is done elsewhere.
2854 : */
2855 : static List *
2856 8152 : preprocess_groupclause(PlannerInfo *root, List *force)
2857 : {
2858 8152 : Query *parse = root->parse;
2859 8152 : List *new_groupclause = NIL;
2860 : ListCell *sl;
2861 : ListCell *gl;
2862 :
2863 : /* For grouping sets, we need to force the ordering */
2864 8152 : if (force)
2865 : {
2866 10372 : foreach(sl, force)
2867 : {
2868 6130 : Index ref = lfirst_int(sl);
2869 6130 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2870 :
2871 6130 : new_groupclause = lappend(new_groupclause, cl);
2872 : }
2873 :
2874 4242 : return new_groupclause;
2875 : }
2876 :
2877 : /* If no ORDER BY, nothing useful to do here */
2878 3910 : if (parse->sortClause == NIL)
2879 2106 : return list_copy(parse->groupClause);
2880 :
2881 : /*
2882 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2883 : * items, but only as far as we can make a matching prefix.
2884 : *
2885 : * This code assumes that the sortClause contains no duplicate items.
2886 : */
2887 3514 : foreach(sl, parse->sortClause)
2888 : {
2889 2356 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2890 :
2891 3452 : foreach(gl, parse->groupClause)
2892 : {
2893 2806 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2894 :
2895 2806 : if (equal(gc, sc))
2896 : {
2897 1710 : new_groupclause = lappend(new_groupclause, gc);
2898 1710 : break;
2899 : }
2900 : }
2901 2356 : if (gl == NULL)
2902 646 : break; /* no match, so stop scanning */
2903 : }
2904 :
2905 :
2906 : /* If no match at all, no point in reordering GROUP BY */
2907 1804 : if (new_groupclause == NIL)
2908 298 : return list_copy(parse->groupClause);
2909 :
2910 : /*
2911 : * Add any remaining GROUP BY items to the new list. We don't require a
2912 : * complete match, because even partial match allows ORDER BY to be
2913 : * implemented using incremental sort. Also, give up if there are any
2914 : * non-sortable GROUP BY items, since then there's no hope anyway.
2915 : */
2916 3382 : foreach(gl, parse->groupClause)
2917 : {
2918 1876 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2919 :
2920 1876 : if (list_member_ptr(new_groupclause, gc))
2921 1710 : continue; /* it matched an ORDER BY item */
2922 166 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2923 0 : return list_copy(parse->groupClause);
2924 166 : new_groupclause = lappend(new_groupclause, gc);
2925 : }
2926 :
2927 : /* Success --- install the rearranged GROUP BY list */
2928 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2929 1506 : return new_groupclause;
2930 : }
2931 :
2932 : /*
2933 : * Extract lists of grouping sets that can be implemented using a single
2934 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2935 : *
2936 : * Input must be sorted with smallest sets first. Result has each sublist
2937 : * sorted with smallest sets first.
2938 : *
2939 : * We want to produce the absolute minimum possible number of lists here to
2940 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2941 : * of finding the minimal partition of a partially-ordered set into chains
2942 : * (which is what we need, taking the list of grouping sets as a poset ordered
2943 : * by set inclusion) can be mapped to the problem of finding the maximum
2944 : * cardinality matching on a bipartite graph, which is solvable in polynomial
2945 : * time with a worst case of no worse than O(n^2.5) and usually much
2946 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
2947 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
2948 : * half a second on my modest system even with optimization off and assertions
2949 : * on.)
2950 : */
2951 : static List *
2952 926 : extract_rollup_sets(List *groupingSets)
2953 : {
2954 926 : int num_sets_raw = list_length(groupingSets);
2955 926 : int num_empty = 0;
2956 926 : int num_sets = 0; /* distinct sets */
2957 926 : int num_chains = 0;
2958 926 : List *result = NIL;
2959 : List **results;
2960 : List **orig_sets;
2961 : Bitmapset **set_masks;
2962 : int *chains;
2963 : short **adjacency;
2964 : short *adjacency_buf;
2965 : BipartiteMatchState *state;
2966 : int i;
2967 : int j;
2968 : int j_size;
2969 926 : ListCell *lc1 = list_head(groupingSets);
2970 : ListCell *lc;
2971 :
2972 : /*
2973 : * Start by stripping out empty sets. The algorithm doesn't require this,
2974 : * but the planner currently needs all empty sets to be returned in the
2975 : * first list, so we strip them here and add them back after.
2976 : */
2977 1572 : while (lc1 && lfirst(lc1) == NIL)
2978 : {
2979 646 : ++num_empty;
2980 646 : lc1 = lnext(groupingSets, lc1);
2981 : }
2982 :
2983 : /* bail out now if it turns out that all we had were empty sets. */
2984 926 : if (!lc1)
2985 66 : return list_make1(groupingSets);
2986 :
2987 : /*----------
2988 : * We don't strictly need to remove duplicate sets here, but if we don't,
2989 : * they tend to become scattered through the result, which is a bit
2990 : * confusing (and irritating if we ever decide to optimize them out).
2991 : * So we remove them here and add them back after.
2992 : *
2993 : * For each non-duplicate set, we fill in the following:
2994 : *
2995 : * orig_sets[i] = list of the original set lists
2996 : * set_masks[i] = bitmapset for testing inclusion
2997 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2998 : *
2999 : * chains[i] will be the result group this set is assigned to.
3000 : *
3001 : * We index all of these from 1 rather than 0 because it is convenient
3002 : * to leave 0 free for the NIL node in the graph algorithm.
3003 : *----------
3004 : */
3005 860 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3006 860 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3007 860 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3008 860 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3009 :
3010 860 : j_size = 0;
3011 860 : j = 0;
3012 860 : i = 1;
3013 :
3014 3044 : for_each_cell(lc, groupingSets, lc1)
3015 : {
3016 2184 : List *candidate = (List *) lfirst(lc);
3017 2184 : Bitmapset *candidate_set = NULL;
3018 : ListCell *lc2;
3019 2184 : int dup_of = 0;
3020 :
3021 5274 : foreach(lc2, candidate)
3022 : {
3023 3090 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3024 : }
3025 :
3026 : /* we can only be a dup if we're the same length as a previous set */
3027 2184 : if (j_size == list_length(candidate))
3028 : {
3029 : int k;
3030 :
3031 1928 : for (k = j; k < i; ++k)
3032 : {
3033 1248 : if (bms_equal(set_masks[k], candidate_set))
3034 : {
3035 158 : dup_of = k;
3036 158 : break;
3037 : }
3038 : }
3039 : }
3040 1346 : else if (j_size < list_length(candidate))
3041 : {
3042 1346 : j_size = list_length(candidate);
3043 1346 : j = i;
3044 : }
3045 :
3046 2184 : if (dup_of > 0)
3047 : {
3048 158 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3049 158 : bms_free(candidate_set);
3050 : }
3051 : else
3052 : {
3053 : int k;
3054 2026 : int n_adj = 0;
3055 :
3056 2026 : orig_sets[i] = list_make1(candidate);
3057 2026 : set_masks[i] = candidate_set;
3058 :
3059 : /* fill in adjacency list; no need to compare equal-size sets */
3060 :
3061 3310 : for (k = j - 1; k > 0; --k)
3062 : {
3063 1284 : if (bms_is_subset(set_masks[k], candidate_set))
3064 1122 : adjacency_buf[++n_adj] = k;
3065 : }
3066 :
3067 2026 : if (n_adj > 0)
3068 : {
3069 610 : adjacency_buf[0] = n_adj;
3070 610 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3071 610 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3072 : }
3073 : else
3074 1416 : adjacency[i] = NULL;
3075 :
3076 2026 : ++i;
3077 : }
3078 : }
3079 :
3080 860 : num_sets = i - 1;
3081 :
3082 : /*
3083 : * Apply the graph matching algorithm to do the work.
3084 : */
3085 860 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3086 :
3087 : /*
3088 : * Now, the state->pair* fields have the info we need to assign sets to
3089 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3090 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3091 : * it in one pass)
3092 : */
3093 860 : chains = palloc0((num_sets + 1) * sizeof(int));
3094 :
3095 2886 : for (i = 1; i <= num_sets; ++i)
3096 : {
3097 2026 : int u = state->pair_vu[i];
3098 2026 : int v = state->pair_uv[i];
3099 :
3100 2026 : if (u > 0 && u < i)
3101 0 : chains[i] = chains[u];
3102 2026 : else if (v > 0 && v < i)
3103 582 : chains[i] = chains[v];
3104 : else
3105 1444 : chains[i] = ++num_chains;
3106 : }
3107 :
3108 : /* build result lists. */
3109 860 : results = palloc0((num_chains + 1) * sizeof(List *));
3110 :
3111 2886 : for (i = 1; i <= num_sets; ++i)
3112 : {
3113 2026 : int c = chains[i];
3114 :
3115 : Assert(c > 0);
3116 :
3117 2026 : results[c] = list_concat(results[c], orig_sets[i]);
3118 : }
3119 :
3120 : /* push any empty sets back on the first list. */
3121 1392 : while (num_empty-- > 0)
3122 532 : results[1] = lcons(NIL, results[1]);
3123 :
3124 : /* make result list */
3125 2304 : for (i = 1; i <= num_chains; ++i)
3126 1444 : result = lappend(result, results[i]);
3127 :
3128 : /*
3129 : * Free all the things.
3130 : *
3131 : * (This is over-fussy for small sets but for large sets we could have
3132 : * tied up a nontrivial amount of memory.)
3133 : */
3134 860 : BipartiteMatchFree(state);
3135 860 : pfree(results);
3136 860 : pfree(chains);
3137 2886 : for (i = 1; i <= num_sets; ++i)
3138 2026 : if (adjacency[i])
3139 610 : pfree(adjacency[i]);
3140 860 : pfree(adjacency);
3141 860 : pfree(adjacency_buf);
3142 860 : pfree(orig_sets);
3143 2886 : for (i = 1; i <= num_sets; ++i)
3144 2026 : bms_free(set_masks[i]);
3145 860 : pfree(set_masks);
3146 :
3147 860 : return result;
3148 : }
3149 :
3150 : /*
3151 : * Reorder the elements of a list of grouping sets such that they have correct
3152 : * prefix relationships. Also inserts the GroupingSetData annotations.
3153 : *
3154 : * The input must be ordered with smallest sets first; the result is returned
3155 : * with largest sets first. Note that the result shares no list substructure
3156 : * with the input, so it's safe for the caller to modify it later.
3157 : *
3158 : * If we're passed in a sortclause, we follow its order of columns to the
3159 : * extent possible, to minimize the chance that we add unnecessary sorts.
3160 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3161 : * gets implemented in one pass.)
3162 : */
3163 : static List *
3164 1510 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3165 : {
3166 : ListCell *lc;
3167 1510 : List *previous = NIL;
3168 1510 : List *result = NIL;
3169 :
3170 4340 : foreach(lc, groupingSets)
3171 : {
3172 2830 : List *candidate = (List *) lfirst(lc);
3173 2830 : List *new_elems = list_difference_int(candidate, previous);
3174 2830 : GroupingSetData *gs = makeNode(GroupingSetData);
3175 :
3176 2994 : while (list_length(sortclause) > list_length(previous) &&
3177 : new_elems != NIL)
3178 : {
3179 272 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3180 272 : int ref = sc->tleSortGroupRef;
3181 :
3182 272 : if (list_member_int(new_elems, ref))
3183 : {
3184 164 : previous = lappend_int(previous, ref);
3185 164 : new_elems = list_delete_int(new_elems, ref);
3186 : }
3187 : else
3188 : {
3189 : /* diverged from the sortclause; give up on it */
3190 108 : sortclause = NIL;
3191 108 : break;
3192 : }
3193 : }
3194 :
3195 2830 : previous = list_concat(previous, new_elems);
3196 :
3197 2830 : gs->set = list_copy(previous);
3198 2830 : result = lcons(gs, result);
3199 : }
3200 :
3201 1510 : list_free(previous);
3202 :
3203 1510 : return result;
3204 : }
3205 :
3206 : /*
3207 : * has_volatile_pathkey
3208 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3209 : * containing a volatile function. Otherwise returns false.
3210 : */
3211 : static bool
3212 2814 : has_volatile_pathkey(List *keys)
3213 : {
3214 : ListCell *lc;
3215 :
3216 5772 : foreach(lc, keys)
3217 : {
3218 2976 : PathKey *pathkey = lfirst_node(PathKey, lc);
3219 :
3220 2976 : if (pathkey->pk_eclass->ec_has_volatile)
3221 18 : return true;
3222 : }
3223 :
3224 2796 : return false;
3225 : }
3226 :
3227 : /*
3228 : * adjust_group_pathkeys_for_groupagg
3229 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3230 : * pre-ordered input for ordered aggregates.
3231 : *
3232 : * We define "best" as the pathkeys that suit the largest number of
3233 : * aggregate functions. We find these by looking at the first ORDER BY /
3234 : * DISTINCT aggregate and take the pathkeys for that before searching for
3235 : * other aggregates that require the same or a more strict variation of the
3236 : * same pathkeys. We then repeat that process for any remaining aggregates
3237 : * with different pathkeys and if we find another set of pathkeys that suits a
3238 : * larger number of aggregates then we select those pathkeys instead.
3239 : *
3240 : * When the best pathkeys are found we also mark each Aggref that can use
3241 : * those pathkeys as aggpresorted = true.
3242 : *
3243 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3244 : * volatile functions, we never make use of these pathkeys. We want to ensure
3245 : * that sorts using volatile functions are done independently in each Aggref
3246 : * rather than once at the query level. If we were to allow this then Aggrefs
3247 : * with compatible sort orders would all transition their rows in the same
3248 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3249 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3250 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3251 : * left to perform their sorts individually. To avoid this inconsistent
3252 : * behavior which could make Aggref results depend on what other Aggrefs the
3253 : * query contains, we always force Aggrefs with volatile functions to perform
3254 : * their own sorts.
3255 : */
3256 : static void
3257 2418 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3258 : {
3259 2418 : List *grouppathkeys = root->group_pathkeys;
3260 : List *bestpathkeys;
3261 : Bitmapset *bestaggs;
3262 : Bitmapset *unprocessed_aggs;
3263 : ListCell *lc;
3264 : int i;
3265 :
3266 : /* Shouldn't be here if there are grouping sets */
3267 : Assert(root->parse->groupingSets == NIL);
3268 : /* Shouldn't be here unless there are some ordered aggregates */
3269 : Assert(root->numOrderedAggs > 0);
3270 :
3271 : /* Do nothing if disabled */
3272 2418 : if (!enable_presorted_aggregate)
3273 6 : return;
3274 :
3275 : /*
3276 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3277 : * the indexes of all AggInfos to be processed below.
3278 : */
3279 2412 : unprocessed_aggs = NULL;
3280 5508 : foreach(lc, root->agginfos)
3281 : {
3282 3096 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3283 3096 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3284 :
3285 3096 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3286 264 : continue;
3287 :
3288 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3289 2832 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3290 300 : continue;
3291 :
3292 : /* Additional safety checks are needed if there's a FILTER clause */
3293 2532 : if (aggref->aggfilter != NULL)
3294 : {
3295 : ListCell *lc2;
3296 54 : bool allow_presort = true;
3297 :
3298 : /*
3299 : * When the Aggref has a FILTER clause, it's possible that the
3300 : * filter removes rows that cannot be sorted because the
3301 : * expression to sort by results in an error during its
3302 : * evaluation. This is a problem for presorting as that happens
3303 : * before the FILTER, whereas without presorting, the Aggregate
3304 : * node will apply the FILTER *before* sorting. So that we never
3305 : * try to sort anything that might error, here we aim to skip over
3306 : * any Aggrefs with arguments with expressions which, when
3307 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3308 : * may be more cases that should be allowed, but more thought
3309 : * needs to be given. Err on the side of caution.
3310 : */
3311 102 : foreach(lc2, aggref->args)
3312 : {
3313 72 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3314 72 : Expr *expr = tle->expr;
3315 :
3316 84 : while (IsA(expr, RelabelType))
3317 12 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3318 :
3319 : /* Common case, Vars and Consts are ok */
3320 72 : if (IsA(expr, Var) || IsA(expr, Const))
3321 48 : continue;
3322 :
3323 : /* Unsupported. Don't try to presort for this Aggref */
3324 24 : allow_presort = false;
3325 24 : break;
3326 : }
3327 :
3328 : /* Skip unsupported Aggrefs */
3329 54 : if (!allow_presort)
3330 24 : continue;
3331 : }
3332 :
3333 2508 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3334 : foreach_current_index(lc));
3335 : }
3336 :
3337 : /*
3338 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3339 : * for the given set of aggregates.
3340 : *
3341 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3342 : * this during the first loop using the pathkeys for the very first
3343 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3344 : * a more strict set of compatible pathkeys. Once the outer loop is
3345 : * complete, we mark off all the aggregates with compatible pathkeys then
3346 : * remove those from the unprocessed_aggs and repeat the process to try to
3347 : * find another set of pathkeys that are suitable for a larger number of
3348 : * aggregates. The outer loop will stop when there are not enough
3349 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3350 : * to suit a larger number of aggregates.
3351 : */
3352 2412 : bestpathkeys = NIL;
3353 2412 : bestaggs = NULL;
3354 4758 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3355 : {
3356 2346 : Bitmapset *aggindexes = NULL;
3357 2346 : List *currpathkeys = NIL;
3358 :
3359 2346 : i = -1;
3360 5160 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3361 : {
3362 2814 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3363 2814 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3364 : List *sortlist;
3365 : List *pathkeys;
3366 :
3367 2814 : if (aggref->aggdistinct != NIL)
3368 718 : sortlist = aggref->aggdistinct;
3369 : else
3370 2096 : sortlist = aggref->aggorder;
3371 :
3372 2814 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3373 : aggref->args);
3374 :
3375 : /*
3376 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3377 : * or DISTINCT clause.
3378 : */
3379 2814 : if (has_volatile_pathkey(pathkeys))
3380 : {
3381 18 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3382 18 : continue;
3383 : }
3384 :
3385 : /*
3386 : * When not set yet, take the pathkeys from the first unprocessed
3387 : * aggregate.
3388 : */
3389 2796 : if (currpathkeys == NIL)
3390 : {
3391 2340 : currpathkeys = pathkeys;
3392 :
3393 : /* include the GROUP BY pathkeys, if they exist */
3394 2340 : if (grouppathkeys != NIL)
3395 276 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3396 : currpathkeys);
3397 :
3398 : /* record that we found pathkeys for this aggregate */
3399 2340 : aggindexes = bms_add_member(aggindexes, i);
3400 : }
3401 : else
3402 : {
3403 : /* now look for a stronger set of matching pathkeys */
3404 :
3405 : /* include the GROUP BY pathkeys, if they exist */
3406 456 : if (grouppathkeys != NIL)
3407 288 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3408 : pathkeys);
3409 :
3410 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3411 456 : switch (compare_pathkeys(currpathkeys, pathkeys))
3412 : {
3413 12 : case PATHKEYS_BETTER2:
3414 : /* 'pathkeys' are stronger, use these ones instead */
3415 12 : currpathkeys = pathkeys;
3416 : /* FALLTHROUGH */
3417 :
3418 66 : case PATHKEYS_BETTER1:
3419 : /* 'pathkeys' are less strict */
3420 : /* FALLTHROUGH */
3421 :
3422 : case PATHKEYS_EQUAL:
3423 : /* mark this aggregate as covered by 'currpathkeys' */
3424 66 : aggindexes = bms_add_member(aggindexes, i);
3425 66 : break;
3426 :
3427 390 : case PATHKEYS_DIFFERENT:
3428 390 : break;
3429 : }
3430 : }
3431 : }
3432 :
3433 : /* remove the aggregates that we've just processed */
3434 2346 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3435 :
3436 : /*
3437 : * If this pass included more aggregates than the previous best then
3438 : * use these ones as the best set.
3439 : */
3440 2346 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3441 : {
3442 2238 : bestaggs = aggindexes;
3443 2238 : bestpathkeys = currpathkeys;
3444 : }
3445 : }
3446 :
3447 : /*
3448 : * If we found any ordered aggregates, update root->group_pathkeys to add
3449 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3450 : * the original GROUP BY pathkeys already.
3451 : */
3452 2412 : if (bestpathkeys != NIL)
3453 2178 : root->group_pathkeys = bestpathkeys;
3454 :
3455 : /*
3456 : * Now that we've found the best set of aggregates we can set the
3457 : * presorted flag to indicate to the executor that it needn't bother
3458 : * performing a sort for these Aggrefs. We're able to do this now as
3459 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3460 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3461 : * of ordered aggregates.
3462 : */
3463 2412 : i = -1;
3464 4686 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3465 : {
3466 2274 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3467 :
3468 4566 : foreach(lc, agginfo->aggrefs)
3469 : {
3470 2292 : Aggref *aggref = lfirst_node(Aggref, lc);
3471 :
3472 2292 : aggref->aggpresorted = true;
3473 : }
3474 : }
3475 : }
3476 :
3477 : /*
3478 : * Compute query_pathkeys and other pathkeys during plan generation
3479 : */
3480 : static void
3481 516270 : standard_qp_callback(PlannerInfo *root, void *extra)
3482 : {
3483 516270 : Query *parse = root->parse;
3484 516270 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3485 516270 : List *tlist = root->processed_tlist;
3486 516270 : List *activeWindows = qp_extra->activeWindows;
3487 :
3488 : /*
3489 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3490 : * aggregate requirements.
3491 : */
3492 516270 : if (qp_extra->gset_data)
3493 : {
3494 : /*
3495 : * With grouping sets, just use the first RollupData's groupClause. We
3496 : * don't make any effort to optimize grouping clauses when there are
3497 : * grouping sets, nor can we combine aggregate ordering keys with
3498 : * grouping.
3499 : */
3500 932 : List *rollups = qp_extra->gset_data->rollups;
3501 932 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3502 :
3503 932 : if (grouping_is_sortable(groupClause))
3504 : {
3505 : bool sortable;
3506 :
3507 : /*
3508 : * The groupClause is logically below the grouping step. So if
3509 : * there is an RTE entry for the grouping step, we need to remove
3510 : * its RT index from the sort expressions before we make PathKeys
3511 : * for them.
3512 : */
3513 932 : root->group_pathkeys =
3514 932 : make_pathkeys_for_sortclauses_extended(root,
3515 : &groupClause,
3516 : tlist,
3517 : false,
3518 932 : parse->hasGroupRTE,
3519 : &sortable,
3520 : false);
3521 : Assert(sortable);
3522 932 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3523 : }
3524 : else
3525 : {
3526 0 : root->group_pathkeys = NIL;
3527 0 : root->num_groupby_pathkeys = 0;
3528 : }
3529 : }
3530 515338 : else if (parse->groupClause || root->numOrderedAggs > 0)
3531 6084 : {
3532 : /*
3533 : * With a plain GROUP BY list, we can remove any grouping items that
3534 : * are proven redundant by EquivalenceClass processing. For example,
3535 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3536 : * especially common cases, but they're nearly free to detect. Note
3537 : * that we remove redundant items from processed_groupClause but not
3538 : * the original parse->groupClause.
3539 : */
3540 : bool sortable;
3541 :
3542 : /*
3543 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3544 : * EquivalenceClass'es if it's not set yet.
3545 : */
3546 6084 : root->group_pathkeys =
3547 6084 : make_pathkeys_for_sortclauses_extended(root,
3548 : &root->processed_groupClause,
3549 : tlist,
3550 : true,
3551 : false,
3552 : &sortable,
3553 : true);
3554 6084 : if (!sortable)
3555 : {
3556 : /* Can't sort; no point in considering aggregate ordering either */
3557 0 : root->group_pathkeys = NIL;
3558 0 : root->num_groupby_pathkeys = 0;
3559 : }
3560 : else
3561 : {
3562 6084 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3563 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3564 6084 : if (root->numOrderedAggs > 0)
3565 2418 : adjust_group_pathkeys_for_groupagg(root);
3566 : }
3567 : }
3568 : else
3569 : {
3570 509254 : root->group_pathkeys = NIL;
3571 509254 : root->num_groupby_pathkeys = 0;
3572 : }
3573 :
3574 : /* We consider only the first (bottom) window in pathkeys logic */
3575 516270 : if (activeWindows != NIL)
3576 : {
3577 2564 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3578 :
3579 2564 : root->window_pathkeys = make_pathkeys_for_window(root,
3580 : wc,
3581 : tlist);
3582 : }
3583 : else
3584 513706 : root->window_pathkeys = NIL;
3585 :
3586 : /*
3587 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3588 : * redundant by EquivalenceClass processing. The non-redundant list is
3589 : * kept in root->processed_distinctClause, leaving the original
3590 : * parse->distinctClause alone.
3591 : */
3592 516270 : if (parse->distinctClause)
3593 : {
3594 : bool sortable;
3595 :
3596 : /* Make a copy since pathkey processing can modify the list */
3597 2956 : root->processed_distinctClause = list_copy(parse->distinctClause);
3598 2956 : root->distinct_pathkeys =
3599 2956 : make_pathkeys_for_sortclauses_extended(root,
3600 : &root->processed_distinctClause,
3601 : tlist,
3602 : true,
3603 : false,
3604 : &sortable,
3605 : false);
3606 2956 : if (!sortable)
3607 6 : root->distinct_pathkeys = NIL;
3608 : }
3609 : else
3610 513314 : root->distinct_pathkeys = NIL;
3611 :
3612 516270 : root->sort_pathkeys =
3613 516270 : make_pathkeys_for_sortclauses(root,
3614 : parse->sortClause,
3615 : tlist);
3616 :
3617 : /* setting setop_pathkeys might be useful to the union planner */
3618 516270 : if (qp_extra->setop != NULL)
3619 : {
3620 : List *groupClauses;
3621 : bool sortable;
3622 :
3623 12488 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3624 :
3625 12488 : root->setop_pathkeys =
3626 12488 : make_pathkeys_for_sortclauses_extended(root,
3627 : &groupClauses,
3628 : tlist,
3629 : false,
3630 : false,
3631 : &sortable,
3632 : false);
3633 12488 : if (!sortable)
3634 204 : root->setop_pathkeys = NIL;
3635 : }
3636 : else
3637 503782 : root->setop_pathkeys = NIL;
3638 :
3639 : /*
3640 : * Figure out whether we want a sorted result from query_planner.
3641 : *
3642 : * If we have a sortable GROUP BY clause, then we want a result sorted
3643 : * properly for grouping. Otherwise, if we have window functions to
3644 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3645 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3646 : * we try to produce output that's sufficiently well sorted for the
3647 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3648 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3649 : * for a set operation which can benefit from presorted results and have a
3650 : * sortable targetlist, we want to sort by the target list.
3651 : *
3652 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3653 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3654 : * that might just leave us failing to exploit an available sort order at
3655 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3656 : * much easier, since we know that the parser ensured that one is a
3657 : * superset of the other.
3658 : */
3659 516270 : if (root->group_pathkeys)
3660 6636 : root->query_pathkeys = root->group_pathkeys;
3661 509634 : else if (root->window_pathkeys)
3662 2098 : root->query_pathkeys = root->window_pathkeys;
3663 1015072 : else if (list_length(root->distinct_pathkeys) >
3664 507536 : list_length(root->sort_pathkeys))
3665 2472 : root->query_pathkeys = root->distinct_pathkeys;
3666 505064 : else if (root->sort_pathkeys)
3667 67414 : root->query_pathkeys = root->sort_pathkeys;
3668 437650 : else if (root->setop_pathkeys != NIL)
3669 11092 : root->query_pathkeys = root->setop_pathkeys;
3670 : else
3671 426558 : root->query_pathkeys = NIL;
3672 516270 : }
3673 :
3674 : /*
3675 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3676 : *
3677 : * path_rows: number of output rows from scan/join step
3678 : * gd: grouping sets data including list of grouping sets and their clauses
3679 : * target_list: target list containing group clause references
3680 : *
3681 : * If doing grouping sets, we also annotate the gsets data with the estimates
3682 : * for each set and each individual rollup list, with a view to later
3683 : * determining whether some combination of them could be hashed instead.
3684 : */
3685 : static double
3686 46132 : get_number_of_groups(PlannerInfo *root,
3687 : double path_rows,
3688 : grouping_sets_data *gd,
3689 : List *target_list)
3690 : {
3691 46132 : Query *parse = root->parse;
3692 : double dNumGroups;
3693 :
3694 46132 : if (parse->groupClause)
3695 : {
3696 : List *groupExprs;
3697 :
3698 10066 : if (parse->groupingSets)
3699 : {
3700 : /* Add up the estimates for each grouping set */
3701 : ListCell *lc;
3702 :
3703 : Assert(gd); /* keep Coverity happy */
3704 :
3705 866 : dNumGroups = 0;
3706 :
3707 2310 : foreach(lc, gd->rollups)
3708 : {
3709 1444 : RollupData *rollup = lfirst_node(RollupData, lc);
3710 : ListCell *lc2;
3711 : ListCell *lc3;
3712 :
3713 1444 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3714 : target_list);
3715 :
3716 1444 : rollup->numGroups = 0.0;
3717 :
3718 4160 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3719 : {
3720 2716 : List *gset = (List *) lfirst(lc2);
3721 2716 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3722 2716 : double numGroups = estimate_num_groups(root,
3723 : groupExprs,
3724 : path_rows,
3725 : &gset,
3726 : NULL);
3727 :
3728 2716 : gs->numGroups = numGroups;
3729 2716 : rollup->numGroups += numGroups;
3730 : }
3731 :
3732 1444 : dNumGroups += rollup->numGroups;
3733 : }
3734 :
3735 866 : if (gd->hash_sets_idx)
3736 : {
3737 : ListCell *lc2;
3738 :
3739 36 : gd->dNumHashGroups = 0;
3740 :
3741 36 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3742 : target_list);
3743 :
3744 78 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3745 : {
3746 42 : List *gset = (List *) lfirst(lc);
3747 42 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3748 42 : double numGroups = estimate_num_groups(root,
3749 : groupExprs,
3750 : path_rows,
3751 : &gset,
3752 : NULL);
3753 :
3754 42 : gs->numGroups = numGroups;
3755 42 : gd->dNumHashGroups += numGroups;
3756 : }
3757 :
3758 36 : dNumGroups += gd->dNumHashGroups;
3759 : }
3760 : }
3761 : else
3762 : {
3763 : /* Plain GROUP BY -- estimate based on optimized groupClause */
3764 9200 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3765 : target_list);
3766 :
3767 9200 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3768 : NULL, NULL);
3769 : }
3770 : }
3771 36066 : else if (parse->groupingSets)
3772 : {
3773 : /* Empty grouping sets ... one result row for each one */
3774 60 : dNumGroups = list_length(parse->groupingSets);
3775 : }
3776 36006 : else if (parse->hasAggs || root->hasHavingQual)
3777 : {
3778 : /* Plain aggregation, one result row */
3779 36006 : dNumGroups = 1;
3780 : }
3781 : else
3782 : {
3783 : /* Not grouping */
3784 0 : dNumGroups = 1;
3785 : }
3786 :
3787 46132 : return dNumGroups;
3788 : }
3789 :
3790 : /*
3791 : * create_grouping_paths
3792 : *
3793 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3794 : * Along the way, we also build an upperrel for Paths which are partially
3795 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3796 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3797 : * the only partially grouped paths we build are also partial paths; that
3798 : * is, they need a Gather and then a FinalizeAggregate.
3799 : *
3800 : * input_rel: contains the source-data Paths
3801 : * target: the pathtarget for the result Paths to compute
3802 : * gd: grouping sets data including list of grouping sets and their clauses
3803 : *
3804 : * Note: all Paths in input_rel are expected to return the target computed
3805 : * by make_group_input_target.
3806 : */
3807 : static RelOptInfo *
3808 39284 : create_grouping_paths(PlannerInfo *root,
3809 : RelOptInfo *input_rel,
3810 : PathTarget *target,
3811 : bool target_parallel_safe,
3812 : grouping_sets_data *gd)
3813 : {
3814 39284 : Query *parse = root->parse;
3815 : RelOptInfo *grouped_rel;
3816 : RelOptInfo *partially_grouped_rel;
3817 : AggClauseCosts agg_costs;
3818 :
3819 235704 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3820 39284 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3821 :
3822 : /*
3823 : * Create grouping relation to hold fully aggregated grouping and/or
3824 : * aggregation paths.
3825 : */
3826 39284 : grouped_rel = make_grouping_rel(root, input_rel, target,
3827 : target_parallel_safe, parse->havingQual);
3828 :
3829 : /*
3830 : * Create either paths for a degenerate grouping or paths for ordinary
3831 : * grouping, as appropriate.
3832 : */
3833 39284 : if (is_degenerate_grouping(root))
3834 30 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3835 : else
3836 : {
3837 39254 : int flags = 0;
3838 : GroupPathExtraData extra;
3839 :
3840 : /*
3841 : * Determine whether it's possible to perform sort-based
3842 : * implementations of grouping. (Note that if processed_groupClause
3843 : * is empty, grouping_is_sortable() is trivially true, and all the
3844 : * pathkeys_contained_in() tests will succeed too, so that we'll
3845 : * consider every surviving input path.)
3846 : *
3847 : * If we have grouping sets, we might be able to sort some but not all
3848 : * of them; in this case, we need can_sort to be true as long as we
3849 : * must consider any sorted-input plan.
3850 : */
3851 39254 : if ((gd && gd->rollups != NIL)
3852 38334 : || grouping_is_sortable(root->processed_groupClause))
3853 39248 : flags |= GROUPING_CAN_USE_SORT;
3854 :
3855 : /*
3856 : * Determine whether we should consider hash-based implementations of
3857 : * grouping.
3858 : *
3859 : * Hashed aggregation only applies if we're grouping. If we have
3860 : * grouping sets, some groups might be hashable but others not; in
3861 : * this case we set can_hash true as long as there is nothing globally
3862 : * preventing us from hashing (and we should therefore consider plans
3863 : * with hashes).
3864 : *
3865 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3866 : * BY aggregates. (Doing so would imply storing *all* the input
3867 : * values in the hash table, and/or running many sorts in parallel,
3868 : * either of which seems like a certain loser.) We similarly don't
3869 : * support ordered-set aggregates in hashed aggregation, but that case
3870 : * is also included in the numOrderedAggs count.
3871 : *
3872 : * Note: grouping_is_hashable() is much more expensive to check than
3873 : * the other gating conditions, so we want to do it last.
3874 : */
3875 39254 : if ((parse->groupClause != NIL &&
3876 9272 : root->numOrderedAggs == 0 &&
3877 4496 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3878 4492 : flags |= GROUPING_CAN_USE_HASH;
3879 :
3880 : /*
3881 : * Determine whether partial aggregation is possible.
3882 : */
3883 39254 : if (can_partial_agg(root))
3884 34456 : flags |= GROUPING_CAN_PARTIAL_AGG;
3885 :
3886 39254 : extra.flags = flags;
3887 39254 : extra.target_parallel_safe = target_parallel_safe;
3888 39254 : extra.havingQual = parse->havingQual;
3889 39254 : extra.targetList = parse->targetList;
3890 39254 : extra.partial_costs_set = false;
3891 :
3892 : /*
3893 : * Determine whether partitionwise aggregation is in theory possible.
3894 : * It can be disabled by the user, and for now, we don't try to
3895 : * support grouping sets. create_ordinary_grouping_paths() will check
3896 : * additional conditions, such as whether input_rel is partitioned.
3897 : */
3898 39254 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3899 688 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3900 : else
3901 38566 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3902 :
3903 39254 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3904 : &agg_costs, gd, &extra,
3905 : &partially_grouped_rel);
3906 : }
3907 :
3908 39278 : set_cheapest(grouped_rel);
3909 39278 : return grouped_rel;
3910 : }
3911 :
3912 : /*
3913 : * make_grouping_rel
3914 : *
3915 : * Create a new grouping rel and set basic properties.
3916 : *
3917 : * input_rel represents the underlying scan/join relation.
3918 : * target is the output expected from the grouping relation.
3919 : */
3920 : static RelOptInfo *
3921 41414 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3922 : PathTarget *target, bool target_parallel_safe,
3923 : Node *havingQual)
3924 : {
3925 : RelOptInfo *grouped_rel;
3926 :
3927 41414 : if (IS_OTHER_REL(input_rel))
3928 : {
3929 2130 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3930 : input_rel->relids);
3931 2130 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3932 : }
3933 : else
3934 : {
3935 : /*
3936 : * By tradition, the relids set for the main grouping relation is
3937 : * NULL. (This could be changed, but might require adjustments
3938 : * elsewhere.)
3939 : */
3940 39284 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3941 : }
3942 :
3943 : /* Set target. */
3944 41414 : grouped_rel->reltarget = target;
3945 :
3946 : /*
3947 : * If the input relation is not parallel-safe, then the grouped relation
3948 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3949 : * target list and HAVING quals are parallel-safe.
3950 : */
3951 70442 : if (input_rel->consider_parallel && target_parallel_safe &&
3952 29028 : is_parallel_safe(root, (Node *) havingQual))
3953 29010 : grouped_rel->consider_parallel = true;
3954 :
3955 : /*
3956 : * If the input rel belongs to a single FDW, so does the grouped rel.
3957 : */
3958 41414 : grouped_rel->serverid = input_rel->serverid;
3959 41414 : grouped_rel->userid = input_rel->userid;
3960 41414 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3961 41414 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3962 :
3963 41414 : return grouped_rel;
3964 : }
3965 :
3966 : /*
3967 : * is_degenerate_grouping
3968 : *
3969 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3970 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3971 : * grouping sets are all empty).
3972 : */
3973 : static bool
3974 39284 : is_degenerate_grouping(PlannerInfo *root)
3975 : {
3976 39284 : Query *parse = root->parse;
3977 :
3978 38160 : return (root->hasHavingQual || parse->groupingSets) &&
3979 77444 : !parse->hasAggs && parse->groupClause == NIL;
3980 : }
3981 :
3982 : /*
3983 : * create_degenerate_grouping_paths
3984 : *
3985 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3986 : * supposed to emit either zero or one row for each grouping set depending on
3987 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3988 : * either HAVING or the targetlist, so we actually do not need the FROM table
3989 : * at all! We can just throw away the plan-so-far and generate a Result node.
3990 : * This is a sufficiently unusual corner case that it's not worth contorting
3991 : * the structure of this module to avoid having to generate the earlier paths
3992 : * in the first place.
3993 : */
3994 : static void
3995 30 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3996 : RelOptInfo *grouped_rel)
3997 : {
3998 30 : Query *parse = root->parse;
3999 : int nrows;
4000 : Path *path;
4001 :
4002 30 : nrows = list_length(parse->groupingSets);
4003 30 : if (nrows > 1)
4004 : {
4005 : /*
4006 : * Doesn't seem worthwhile writing code to cons up a generate_series
4007 : * or a values scan to emit multiple rows. Instead just make N clones
4008 : * and append them. (With a volatile HAVING clause, this means you
4009 : * might get between 0 and N output rows. Offhand I think that's
4010 : * desired.)
4011 : */
4012 0 : List *paths = NIL;
4013 :
4014 0 : while (--nrows >= 0)
4015 : {
4016 : path = (Path *)
4017 0 : create_group_result_path(root, grouped_rel,
4018 0 : grouped_rel->reltarget,
4019 0 : (List *) parse->havingQual);
4020 0 : paths = lappend(paths, path);
4021 : }
4022 : path = (Path *)
4023 0 : create_append_path(root,
4024 : grouped_rel,
4025 : paths,
4026 : NIL,
4027 : NIL,
4028 : NULL,
4029 : 0,
4030 : false,
4031 : -1);
4032 : }
4033 : else
4034 : {
4035 : /* No grouping sets, or just one, so one output row */
4036 : path = (Path *)
4037 30 : create_group_result_path(root, grouped_rel,
4038 30 : grouped_rel->reltarget,
4039 30 : (List *) parse->havingQual);
4040 : }
4041 :
4042 30 : add_path(grouped_rel, path);
4043 30 : }
4044 :
4045 : /*
4046 : * create_ordinary_grouping_paths
4047 : *
4048 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4049 : *
4050 : * We need to consider sorted and hashed aggregation in the same function,
4051 : * because otherwise (1) it would be harder to throw an appropriate error
4052 : * message if neither way works, and (2) we should not allow hashtable size
4053 : * considerations to dissuade us from using hashing if sorting is not possible.
4054 : *
4055 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4056 : * function creates, or to NULL if it doesn't create one.
4057 : */
4058 : static void
4059 41384 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4060 : RelOptInfo *grouped_rel,
4061 : const AggClauseCosts *agg_costs,
4062 : grouping_sets_data *gd,
4063 : GroupPathExtraData *extra,
4064 : RelOptInfo **partially_grouped_rel_p)
4065 : {
4066 41384 : RelOptInfo *partially_grouped_rel = NULL;
4067 41384 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4068 :
4069 : /*
4070 : * If this is the topmost grouping relation or if the parent relation is
4071 : * doing some form of partitionwise aggregation, then we may be able to do
4072 : * it at this level also. However, if the input relation is not
4073 : * partitioned, partitionwise aggregate is impossible.
4074 : */
4075 41384 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4076 2818 : IS_PARTITIONED_REL(input_rel))
4077 : {
4078 : /*
4079 : * If this is the topmost relation or if the parent relation is doing
4080 : * full partitionwise aggregation, then we can do full partitionwise
4081 : * aggregation provided that the GROUP BY clause contains all of the
4082 : * partitioning columns at this level and the collation used by GROUP
4083 : * BY matches the partitioning collation. Otherwise, we can do at
4084 : * most partial partitionwise aggregation. But if partial aggregation
4085 : * is not supported in general then we can't use it for partitionwise
4086 : * aggregation either.
4087 : *
4088 : * Check parse->groupClause not processed_groupClause, because it's
4089 : * okay if some of the partitioning columns were proved redundant.
4090 : */
4091 1616 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4092 760 : group_by_has_partkey(input_rel, extra->targetList,
4093 760 : root->parse->groupClause))
4094 476 : patype = PARTITIONWISE_AGGREGATE_FULL;
4095 380 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4096 338 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4097 : else
4098 42 : patype = PARTITIONWISE_AGGREGATE_NONE;
4099 : }
4100 :
4101 : /*
4102 : * Before generating paths for grouped_rel, we first generate any possible
4103 : * partially grouped paths; that way, later code can easily consider both
4104 : * parallel and non-parallel approaches to grouping.
4105 : */
4106 41384 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4107 : {
4108 : bool force_rel_creation;
4109 :
4110 : /*
4111 : * If we're doing partitionwise aggregation at this level, force
4112 : * creation of a partially_grouped_rel so we can add partitionwise
4113 : * paths to it.
4114 : */
4115 36514 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4116 :
4117 : partially_grouped_rel =
4118 36514 : create_partial_grouping_paths(root,
4119 : grouped_rel,
4120 : input_rel,
4121 : gd,
4122 : extra,
4123 : force_rel_creation);
4124 : }
4125 :
4126 : /* Set out parameter. */
4127 41384 : *partially_grouped_rel_p = partially_grouped_rel;
4128 :
4129 : /* Apply partitionwise aggregation technique, if possible. */
4130 41384 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4131 814 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4132 : partially_grouped_rel, agg_costs,
4133 : gd, patype, extra);
4134 :
4135 : /* If we are doing partial aggregation only, return. */
4136 41384 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4137 : {
4138 : Assert(partially_grouped_rel);
4139 :
4140 858 : if (partially_grouped_rel->pathlist)
4141 858 : set_cheapest(partially_grouped_rel);
4142 :
4143 858 : return;
4144 : }
4145 :
4146 : /* Gather any partially grouped partial paths. */
4147 40526 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4148 2008 : gather_grouping_paths(root, partially_grouped_rel);
4149 :
4150 : /* Now choose the best path(s) for partially_grouped_rel. */
4151 40526 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4152 2236 : set_cheapest(partially_grouped_rel);
4153 :
4154 : /* Build final grouping paths */
4155 40526 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4156 : partially_grouped_rel, agg_costs, gd,
4157 : extra);
4158 :
4159 : /* Give a helpful error if we failed to find any implementation */
4160 40526 : if (grouped_rel->pathlist == NIL)
4161 6 : ereport(ERROR,
4162 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4163 : errmsg("could not implement GROUP BY"),
4164 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4165 :
4166 : /*
4167 : * If there is an FDW that's responsible for all baserels of the query,
4168 : * let it consider adding ForeignPaths.
4169 : */
4170 40520 : if (grouped_rel->fdwroutine &&
4171 336 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4172 336 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4173 : input_rel, grouped_rel,
4174 : extra);
4175 :
4176 : /* Let extensions possibly add some more paths */
4177 40520 : if (create_upper_paths_hook)
4178 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4179 : input_rel, grouped_rel,
4180 : extra);
4181 : }
4182 :
4183 : /*
4184 : * For a given input path, consider the possible ways of doing grouping sets on
4185 : * it, by combinations of hashing and sorting. This can be called multiple
4186 : * times, so it's important that it not scribble on input. No result is
4187 : * returned, but any generated paths are added to grouped_rel.
4188 : */
4189 : static void
4190 1822 : consider_groupingsets_paths(PlannerInfo *root,
4191 : RelOptInfo *grouped_rel,
4192 : Path *path,
4193 : bool is_sorted,
4194 : bool can_hash,
4195 : grouping_sets_data *gd,
4196 : const AggClauseCosts *agg_costs,
4197 : double dNumGroups)
4198 : {
4199 1822 : Query *parse = root->parse;
4200 1822 : Size hash_mem_limit = get_hash_memory_limit();
4201 :
4202 : /*
4203 : * If we're not being offered sorted input, then only consider plans that
4204 : * can be done entirely by hashing.
4205 : *
4206 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4207 : * the input is actually sorted despite not being advertised as such, we
4208 : * prefer to make use of that in order to use less memory.
4209 : *
4210 : * If none of the grouping sets are sortable, then ignore the hash_mem
4211 : * limit and generate a path anyway, since otherwise we'll just fail.
4212 : */
4213 1822 : if (!is_sorted)
4214 : {
4215 830 : List *new_rollups = NIL;
4216 830 : RollupData *unhashed_rollup = NULL;
4217 : List *sets_data;
4218 830 : List *empty_sets_data = NIL;
4219 830 : List *empty_sets = NIL;
4220 : ListCell *lc;
4221 830 : ListCell *l_start = list_head(gd->rollups);
4222 830 : AggStrategy strat = AGG_HASHED;
4223 : double hashsize;
4224 830 : double exclude_groups = 0.0;
4225 :
4226 : Assert(can_hash);
4227 :
4228 : /*
4229 : * If the input is coincidentally sorted usefully (which can happen
4230 : * even if is_sorted is false, since that only means that our caller
4231 : * has set up the sorting for us), then save some hashtable space by
4232 : * making use of that. But we need to watch out for degenerate cases:
4233 : *
4234 : * 1) If there are any empty grouping sets, then group_pathkeys might
4235 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4236 : * there will be a rollup containing only empty groups, and the
4237 : * pathkeys_contained_in test is vacuously true; this is ok.
4238 : *
4239 : * XXX: the above relies on the fact that group_pathkeys is generated
4240 : * from the first rollup. If we add the ability to consider multiple
4241 : * sort orders for grouping input, this assumption might fail.
4242 : *
4243 : * 2) If there are no empty sets and only unsortable sets, then the
4244 : * rollups list will be empty (and thus l_start == NULL), and
4245 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4246 : * pathkeys_contained_in test doesn't cause us to crash.
4247 : */
4248 1654 : if (l_start != NULL &&
4249 824 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4250 : {
4251 12 : unhashed_rollup = lfirst_node(RollupData, l_start);
4252 12 : exclude_groups = unhashed_rollup->numGroups;
4253 12 : l_start = lnext(gd->rollups, l_start);
4254 : }
4255 :
4256 830 : hashsize = estimate_hashagg_tablesize(root,
4257 : path,
4258 : agg_costs,
4259 : dNumGroups - exclude_groups);
4260 :
4261 : /*
4262 : * gd->rollups is empty if we have only unsortable columns to work
4263 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4264 : * sorted-input case to generate usable mixed paths.
4265 : */
4266 830 : if (hashsize > hash_mem_limit && gd->rollups)
4267 18 : return; /* nope, won't fit */
4268 :
4269 : /*
4270 : * We need to burst the existing rollups list into individual grouping
4271 : * sets and recompute a groupClause for each set.
4272 : */
4273 812 : sets_data = list_copy(gd->unsortable_sets);
4274 :
4275 2064 : for_each_cell(lc, gd->rollups, l_start)
4276 : {
4277 1276 : RollupData *rollup = lfirst_node(RollupData, lc);
4278 :
4279 : /*
4280 : * If we find an unhashable rollup that's not been skipped by the
4281 : * "actually sorted" check above, we can't cope; we'd need sorted
4282 : * input (with a different sort order) but we can't get that here.
4283 : * So bail out; we'll get a valid path from the is_sorted case
4284 : * instead.
4285 : *
4286 : * The mere presence of empty grouping sets doesn't make a rollup
4287 : * unhashable (see preprocess_grouping_sets), we handle those
4288 : * specially below.
4289 : */
4290 1276 : if (!rollup->hashable)
4291 24 : return;
4292 :
4293 1252 : sets_data = list_concat(sets_data, rollup->gsets_data);
4294 : }
4295 3270 : foreach(lc, sets_data)
4296 : {
4297 2482 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4298 2482 : List *gset = gs->set;
4299 : RollupData *rollup;
4300 :
4301 2482 : if (gset == NIL)
4302 : {
4303 : /* Empty grouping sets can't be hashed. */
4304 496 : empty_sets_data = lappend(empty_sets_data, gs);
4305 496 : empty_sets = lappend(empty_sets, NIL);
4306 : }
4307 : else
4308 : {
4309 1986 : rollup = makeNode(RollupData);
4310 :
4311 1986 : rollup->groupClause = preprocess_groupclause(root, gset);
4312 1986 : rollup->gsets_data = list_make1(gs);
4313 1986 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4314 : rollup->gsets_data,
4315 : gd->tleref_to_colnum_map);
4316 1986 : rollup->numGroups = gs->numGroups;
4317 1986 : rollup->hashable = true;
4318 1986 : rollup->is_hashed = true;
4319 1986 : new_rollups = lappend(new_rollups, rollup);
4320 : }
4321 : }
4322 :
4323 : /*
4324 : * If we didn't find anything nonempty to hash, then bail. We'll
4325 : * generate a path from the is_sorted case.
4326 : */
4327 788 : if (new_rollups == NIL)
4328 0 : return;
4329 :
4330 : /*
4331 : * If there were empty grouping sets they should have been in the
4332 : * first rollup.
4333 : */
4334 : Assert(!unhashed_rollup || !empty_sets);
4335 :
4336 788 : if (unhashed_rollup)
4337 : {
4338 12 : new_rollups = lappend(new_rollups, unhashed_rollup);
4339 12 : strat = AGG_MIXED;
4340 : }
4341 776 : else if (empty_sets)
4342 : {
4343 448 : RollupData *rollup = makeNode(RollupData);
4344 :
4345 448 : rollup->groupClause = NIL;
4346 448 : rollup->gsets_data = empty_sets_data;
4347 448 : rollup->gsets = empty_sets;
4348 448 : rollup->numGroups = list_length(empty_sets);
4349 448 : rollup->hashable = false;
4350 448 : rollup->is_hashed = false;
4351 448 : new_rollups = lappend(new_rollups, rollup);
4352 448 : strat = AGG_MIXED;
4353 : }
4354 :
4355 788 : add_path(grouped_rel, (Path *)
4356 788 : create_groupingsets_path(root,
4357 : grouped_rel,
4358 : path,
4359 788 : (List *) parse->havingQual,
4360 : strat,
4361 : new_rollups,
4362 : agg_costs));
4363 788 : return;
4364 : }
4365 :
4366 : /*
4367 : * If we have sorted input but nothing we can do with it, bail.
4368 : */
4369 992 : if (gd->rollups == NIL)
4370 0 : return;
4371 :
4372 : /*
4373 : * Given sorted input, we try and make two paths: one sorted and one mixed
4374 : * sort/hash. (We need to try both because hashagg might be disabled, or
4375 : * some columns might not be sortable.)
4376 : *
4377 : * can_hash is passed in as false if some obstacle elsewhere (such as
4378 : * ordered aggs) means that we shouldn't consider hashing at all.
4379 : */
4380 992 : if (can_hash && gd->any_hashable)
4381 : {
4382 896 : List *rollups = NIL;
4383 896 : List *hash_sets = list_copy(gd->unsortable_sets);
4384 896 : double availspace = hash_mem_limit;
4385 : ListCell *lc;
4386 :
4387 : /*
4388 : * Account first for space needed for groups we can't sort at all.
4389 : */
4390 896 : availspace -= estimate_hashagg_tablesize(root,
4391 : path,
4392 : agg_costs,
4393 : gd->dNumHashGroups);
4394 :
4395 896 : if (availspace > 0 && list_length(gd->rollups) > 1)
4396 : {
4397 : double scale;
4398 456 : int num_rollups = list_length(gd->rollups);
4399 : int k_capacity;
4400 456 : int *k_weights = palloc(num_rollups * sizeof(int));
4401 456 : Bitmapset *hash_items = NULL;
4402 : int i;
4403 :
4404 : /*
4405 : * We treat this as a knapsack problem: the knapsack capacity
4406 : * represents hash_mem, the item weights are the estimated memory
4407 : * usage of the hashtables needed to implement a single rollup,
4408 : * and we really ought to use the cost saving as the item value;
4409 : * however, currently the costs assigned to sort nodes don't
4410 : * reflect the comparison costs well, and so we treat all items as
4411 : * of equal value (each rollup we hash instead saves us one sort).
4412 : *
4413 : * To use the discrete knapsack, we need to scale the values to a
4414 : * reasonably small bounded range. We choose to allow a 5% error
4415 : * margin; we have no more than 4096 rollups in the worst possible
4416 : * case, which with a 5% error margin will require a bit over 42MB
4417 : * of workspace. (Anyone wanting to plan queries that complex had
4418 : * better have the memory for it. In more reasonable cases, with
4419 : * no more than a couple of dozen rollups, the memory usage will
4420 : * be negligible.)
4421 : *
4422 : * k_capacity is naturally bounded, but we clamp the values for
4423 : * scale and weight (below) to avoid overflows or underflows (or
4424 : * uselessly trying to use a scale factor less than 1 byte).
4425 : */
4426 456 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4427 456 : k_capacity = (int) floor(availspace / scale);
4428 :
4429 : /*
4430 : * We leave the first rollup out of consideration since it's the
4431 : * one that matches the input sort order. We assign indexes "i"
4432 : * to only those entries considered for hashing; the second loop,
4433 : * below, must use the same condition.
4434 : */
4435 456 : i = 0;
4436 1164 : for_each_from(lc, gd->rollups, 1)
4437 : {
4438 708 : RollupData *rollup = lfirst_node(RollupData, lc);
4439 :
4440 708 : if (rollup->hashable)
4441 : {
4442 708 : double sz = estimate_hashagg_tablesize(root,
4443 : path,
4444 : agg_costs,
4445 : rollup->numGroups);
4446 :
4447 : /*
4448 : * If sz is enormous, but hash_mem (and hence scale) is
4449 : * small, avoid integer overflow here.
4450 : */
4451 708 : k_weights[i] = (int) Min(floor(sz / scale),
4452 : k_capacity + 1.0);
4453 708 : ++i;
4454 : }
4455 : }
4456 :
4457 : /*
4458 : * Apply knapsack algorithm; compute the set of items which
4459 : * maximizes the value stored (in this case the number of sorts
4460 : * saved) while keeping the total size (approximately) within
4461 : * capacity.
4462 : */
4463 456 : if (i > 0)
4464 456 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4465 :
4466 456 : if (!bms_is_empty(hash_items))
4467 : {
4468 456 : rollups = list_make1(linitial(gd->rollups));
4469 :
4470 456 : i = 0;
4471 1164 : for_each_from(lc, gd->rollups, 1)
4472 : {
4473 708 : RollupData *rollup = lfirst_node(RollupData, lc);
4474 :
4475 708 : if (rollup->hashable)
4476 : {
4477 708 : if (bms_is_member(i, hash_items))
4478 672 : hash_sets = list_concat(hash_sets,
4479 672 : rollup->gsets_data);
4480 : else
4481 36 : rollups = lappend(rollups, rollup);
4482 708 : ++i;
4483 : }
4484 : else
4485 0 : rollups = lappend(rollups, rollup);
4486 : }
4487 : }
4488 : }
4489 :
4490 896 : if (!rollups && hash_sets)
4491 24 : rollups = list_copy(gd->rollups);
4492 :
4493 1708 : foreach(lc, hash_sets)
4494 : {
4495 812 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4496 812 : RollupData *rollup = makeNode(RollupData);
4497 :
4498 : Assert(gs->set != NIL);
4499 :
4500 812 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4501 812 : rollup->gsets_data = list_make1(gs);
4502 812 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4503 : rollup->gsets_data,
4504 : gd->tleref_to_colnum_map);
4505 812 : rollup->numGroups = gs->numGroups;
4506 812 : rollup->hashable = true;
4507 812 : rollup->is_hashed = true;
4508 812 : rollups = lcons(rollup, rollups);
4509 : }
4510 :
4511 896 : if (rollups)
4512 : {
4513 480 : add_path(grouped_rel, (Path *)
4514 480 : create_groupingsets_path(root,
4515 : grouped_rel,
4516 : path,
4517 480 : (List *) parse->havingQual,
4518 : AGG_MIXED,
4519 : rollups,
4520 : agg_costs));
4521 : }
4522 : }
4523 :
4524 : /*
4525 : * Now try the simple sorted case.
4526 : */
4527 992 : if (!gd->unsortable_sets)
4528 962 : add_path(grouped_rel, (Path *)
4529 962 : create_groupingsets_path(root,
4530 : grouped_rel,
4531 : path,
4532 962 : (List *) parse->havingQual,
4533 : AGG_SORTED,
4534 : gd->rollups,
4535 : agg_costs));
4536 : }
4537 :
4538 : /*
4539 : * create_window_paths
4540 : *
4541 : * Build a new upperrel containing Paths for window-function evaluation.
4542 : *
4543 : * input_rel: contains the source-data Paths
4544 : * input_target: result of make_window_input_target
4545 : * output_target: what the topmost WindowAggPath should return
4546 : * wflists: result of find_window_functions
4547 : * activeWindows: result of select_active_windows
4548 : *
4549 : * Note: all Paths in input_rel are expected to return input_target.
4550 : */
4551 : static RelOptInfo *
4552 2564 : create_window_paths(PlannerInfo *root,
4553 : RelOptInfo *input_rel,
4554 : PathTarget *input_target,
4555 : PathTarget *output_target,
4556 : bool output_target_parallel_safe,
4557 : WindowFuncLists *wflists,
4558 : List *activeWindows)
4559 : {
4560 : RelOptInfo *window_rel;
4561 : ListCell *lc;
4562 :
4563 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4564 2564 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4565 :
4566 : /*
4567 : * If the input relation is not parallel-safe, then the window relation
4568 : * can't be parallel-safe, either. Otherwise, we need to examine the
4569 : * target list and active windows for non-parallel-safe constructs.
4570 : */
4571 2564 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4572 0 : is_parallel_safe(root, (Node *) activeWindows))
4573 0 : window_rel->consider_parallel = true;
4574 :
4575 : /*
4576 : * If the input rel belongs to a single FDW, so does the window rel.
4577 : */
4578 2564 : window_rel->serverid = input_rel->serverid;
4579 2564 : window_rel->userid = input_rel->userid;
4580 2564 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4581 2564 : window_rel->fdwroutine = input_rel->fdwroutine;
4582 :
4583 : /*
4584 : * Consider computing window functions starting from the existing
4585 : * cheapest-total path (which will likely require a sort) as well as any
4586 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4587 : */
4588 5468 : foreach(lc, input_rel->pathlist)
4589 : {
4590 2904 : Path *path = (Path *) lfirst(lc);
4591 : int presorted_keys;
4592 :
4593 3244 : if (path == input_rel->cheapest_total_path ||
4594 340 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4595 152 : &presorted_keys) ||
4596 152 : presorted_keys > 0)
4597 2778 : create_one_window_path(root,
4598 : window_rel,
4599 : path,
4600 : input_target,
4601 : output_target,
4602 : wflists,
4603 : activeWindows);
4604 : }
4605 :
4606 : /*
4607 : * If there is an FDW that's responsible for all baserels of the query,
4608 : * let it consider adding ForeignPaths.
4609 : */
4610 2564 : if (window_rel->fdwroutine &&
4611 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4612 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4613 : input_rel, window_rel,
4614 : NULL);
4615 :
4616 : /* Let extensions possibly add some more paths */
4617 2564 : if (create_upper_paths_hook)
4618 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4619 : input_rel, window_rel, NULL);
4620 :
4621 : /* Now choose the best path(s) */
4622 2564 : set_cheapest(window_rel);
4623 :
4624 2564 : return window_rel;
4625 : }
4626 :
4627 : /*
4628 : * Stack window-function implementation steps atop the given Path, and
4629 : * add the result to window_rel.
4630 : *
4631 : * window_rel: upperrel to contain result
4632 : * path: input Path to use (must return input_target)
4633 : * input_target: result of make_window_input_target
4634 : * output_target: what the topmost WindowAggPath should return
4635 : * wflists: result of find_window_functions
4636 : * activeWindows: result of select_active_windows
4637 : */
4638 : static void
4639 2778 : create_one_window_path(PlannerInfo *root,
4640 : RelOptInfo *window_rel,
4641 : Path *path,
4642 : PathTarget *input_target,
4643 : PathTarget *output_target,
4644 : WindowFuncLists *wflists,
4645 : List *activeWindows)
4646 : {
4647 : PathTarget *window_target;
4648 : ListCell *l;
4649 2778 : List *topqual = NIL;
4650 :
4651 : /*
4652 : * Since each window clause could require a different sort order, we stack
4653 : * up a WindowAgg node for each clause, with sort steps between them as
4654 : * needed. (We assume that select_active_windows chose a good order for
4655 : * executing the clauses in.)
4656 : *
4657 : * input_target should contain all Vars and Aggs needed for the result.
4658 : * (In some cases we wouldn't need to propagate all of these all the way
4659 : * to the top, since they might only be needed as inputs to WindowFuncs.
4660 : * It's probably not worth trying to optimize that though.) It must also
4661 : * contain all window partitioning and sorting expressions, to ensure
4662 : * they're computed only once at the bottom of the stack (that's critical
4663 : * for volatile functions). As we climb up the stack, we'll add outputs
4664 : * for the WindowFuncs computed at each level.
4665 : */
4666 2778 : window_target = input_target;
4667 :
4668 5742 : foreach(l, activeWindows)
4669 : {
4670 2964 : WindowClause *wc = lfirst_node(WindowClause, l);
4671 : List *window_pathkeys;
4672 2964 : List *runcondition = NIL;
4673 : int presorted_keys;
4674 : bool is_sorted;
4675 : bool topwindow;
4676 : ListCell *lc2;
4677 :
4678 2964 : window_pathkeys = make_pathkeys_for_window(root,
4679 : wc,
4680 : root->processed_tlist);
4681 :
4682 2964 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4683 : path->pathkeys,
4684 : &presorted_keys);
4685 :
4686 : /* Sort if necessary */
4687 2964 : if (!is_sorted)
4688 : {
4689 : /*
4690 : * No presorted keys or incremental sort disabled, just perform a
4691 : * complete sort.
4692 : */
4693 2172 : if (presorted_keys == 0 || !enable_incremental_sort)
4694 2110 : path = (Path *) create_sort_path(root, window_rel,
4695 : path,
4696 : window_pathkeys,
4697 : -1.0);
4698 : else
4699 : {
4700 : /*
4701 : * Since we have presorted keys and incremental sort is
4702 : * enabled, just use incremental sort.
4703 : */
4704 62 : path = (Path *) create_incremental_sort_path(root,
4705 : window_rel,
4706 : path,
4707 : window_pathkeys,
4708 : presorted_keys,
4709 : -1.0);
4710 : }
4711 : }
4712 :
4713 2964 : if (lnext(activeWindows, l))
4714 : {
4715 : /*
4716 : * Add the current WindowFuncs to the output target for this
4717 : * intermediate WindowAggPath. We must copy window_target to
4718 : * avoid changing the previous path's target.
4719 : *
4720 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4721 : * we do need to account for the increase in tlist width.
4722 : */
4723 186 : int64 tuple_width = window_target->width;
4724 :
4725 186 : window_target = copy_pathtarget(window_target);
4726 438 : foreach(lc2, wflists->windowFuncs[wc->winref])
4727 : {
4728 252 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4729 :
4730 252 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4731 252 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4732 : }
4733 186 : window_target->width = clamp_width_est(tuple_width);
4734 : }
4735 : else
4736 : {
4737 : /* Install the goal target in the topmost WindowAgg */
4738 2778 : window_target = output_target;
4739 : }
4740 :
4741 : /* mark the final item in the list as the top-level window */
4742 2964 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4743 :
4744 : /*
4745 : * Collect the WindowFuncRunConditions from each WindowFunc and
4746 : * convert them into OpExprs
4747 : */
4748 6798 : foreach(lc2, wflists->windowFuncs[wc->winref])
4749 : {
4750 : ListCell *lc3;
4751 3834 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4752 :
4753 4014 : foreach(lc3, wfunc->runCondition)
4754 : {
4755 180 : WindowFuncRunCondition *wfuncrc =
4756 : lfirst_node(WindowFuncRunCondition, lc3);
4757 : Expr *opexpr;
4758 : Expr *leftop;
4759 : Expr *rightop;
4760 :
4761 180 : if (wfuncrc->wfunc_left)
4762 : {
4763 162 : leftop = (Expr *) copyObject(wfunc);
4764 162 : rightop = copyObject(wfuncrc->arg);
4765 : }
4766 : else
4767 : {
4768 18 : leftop = copyObject(wfuncrc->arg);
4769 18 : rightop = (Expr *) copyObject(wfunc);
4770 : }
4771 :
4772 180 : opexpr = make_opclause(wfuncrc->opno,
4773 : BOOLOID,
4774 : false,
4775 : leftop,
4776 : rightop,
4777 : InvalidOid,
4778 : wfuncrc->inputcollid);
4779 :
4780 180 : runcondition = lappend(runcondition, opexpr);
4781 :
4782 180 : if (!topwindow)
4783 24 : topqual = lappend(topqual, opexpr);
4784 : }
4785 : }
4786 :
4787 : path = (Path *)
4788 2964 : create_windowagg_path(root, window_rel, path, window_target,
4789 2964 : wflists->windowFuncs[wc->winref],
4790 : runcondition, wc,
4791 : topwindow ? topqual : NIL, topwindow);
4792 : }
4793 :
4794 2778 : add_path(window_rel, path);
4795 2778 : }
4796 :
4797 : /*
4798 : * create_distinct_paths
4799 : *
4800 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4801 : *
4802 : * input_rel: contains the source-data Paths
4803 : * target: the pathtarget for the result Paths to compute
4804 : *
4805 : * Note: input paths should already compute the desired pathtarget, since
4806 : * Sort/Unique won't project anything.
4807 : */
4808 : static RelOptInfo *
4809 2956 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4810 : PathTarget *target)
4811 : {
4812 : RelOptInfo *distinct_rel;
4813 :
4814 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4815 2956 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4816 :
4817 : /*
4818 : * We don't compute anything at this level, so distinct_rel will be
4819 : * parallel-safe if the input rel is parallel-safe. In particular, if
4820 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4821 : * output those expressions, and will not be parallel-safe unless those
4822 : * expressions are parallel-safe.
4823 : */
4824 2956 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4825 :
4826 : /*
4827 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4828 : */
4829 2956 : distinct_rel->serverid = input_rel->serverid;
4830 2956 : distinct_rel->userid = input_rel->userid;
4831 2956 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4832 2956 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4833 :
4834 : /* build distinct paths based on input_rel's pathlist */
4835 2956 : create_final_distinct_paths(root, input_rel, distinct_rel);
4836 :
4837 : /* now build distinct paths based on input_rel's partial_pathlist */
4838 2956 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4839 :
4840 : /* Give a helpful error if we failed to create any paths */
4841 2956 : if (distinct_rel->pathlist == NIL)
4842 0 : ereport(ERROR,
4843 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4844 : errmsg("could not implement DISTINCT"),
4845 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4846 :
4847 : /*
4848 : * If there is an FDW that's responsible for all baserels of the query,
4849 : * let it consider adding ForeignPaths.
4850 : */
4851 2956 : if (distinct_rel->fdwroutine &&
4852 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4853 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4854 : UPPERREL_DISTINCT,
4855 : input_rel,
4856 : distinct_rel,
4857 : NULL);
4858 :
4859 : /* Let extensions possibly add some more paths */
4860 2956 : if (create_upper_paths_hook)
4861 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4862 : distinct_rel, NULL);
4863 :
4864 : /* Now choose the best path(s) */
4865 2956 : set_cheapest(distinct_rel);
4866 :
4867 2956 : return distinct_rel;
4868 : }
4869 :
4870 : /*
4871 : * create_partial_distinct_paths
4872 : *
4873 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4874 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4875 : * paths on top and add a final unique/aggregate path to remove any duplicate
4876 : * produced from combining rows from parallel workers.
4877 : */
4878 : static void
4879 2956 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4880 : RelOptInfo *final_distinct_rel,
4881 : PathTarget *target)
4882 : {
4883 : RelOptInfo *partial_distinct_rel;
4884 : Query *parse;
4885 : List *distinctExprs;
4886 : double numDistinctRows;
4887 : Path *cheapest_partial_path;
4888 : ListCell *lc;
4889 :
4890 : /* nothing to do when there are no partial paths in the input rel */
4891 2956 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4892 2848 : return;
4893 :
4894 108 : parse = root->parse;
4895 :
4896 : /* can't do parallel DISTINCT ON */
4897 108 : if (parse->hasDistinctOn)
4898 0 : return;
4899 :
4900 108 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4901 : NULL);
4902 108 : partial_distinct_rel->reltarget = target;
4903 108 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4904 :
4905 : /*
4906 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4907 : */
4908 108 : partial_distinct_rel->serverid = input_rel->serverid;
4909 108 : partial_distinct_rel->userid = input_rel->userid;
4910 108 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4911 108 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4912 :
4913 108 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4914 :
4915 108 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4916 : parse->targetList);
4917 :
4918 : /* estimate how many distinct rows we'll get from each worker */
4919 108 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4920 : cheapest_partial_path->rows,
4921 : NULL, NULL);
4922 :
4923 : /*
4924 : * Try sorting the cheapest path and incrementally sorting any paths with
4925 : * presorted keys and put a unique paths atop of those. We'll also
4926 : * attempt to reorder the required pathkeys to match the input path's
4927 : * pathkeys as much as possible, in hopes of avoiding a possible need to
4928 : * re-sort.
4929 : */
4930 108 : if (grouping_is_sortable(root->processed_distinctClause))
4931 : {
4932 234 : foreach(lc, input_rel->partial_pathlist)
4933 : {
4934 126 : Path *input_path = (Path *) lfirst(lc);
4935 : Path *sorted_path;
4936 126 : List *useful_pathkeys_list = NIL;
4937 :
4938 : useful_pathkeys_list =
4939 126 : get_useful_pathkeys_for_distinct(root,
4940 : root->distinct_pathkeys,
4941 : input_path->pathkeys);
4942 : Assert(list_length(useful_pathkeys_list) > 0);
4943 :
4944 390 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
4945 : {
4946 138 : sorted_path = make_ordered_path(root,
4947 : partial_distinct_rel,
4948 : input_path,
4949 : cheapest_partial_path,
4950 : useful_pathkeys,
4951 : -1.0);
4952 :
4953 138 : if (sorted_path == NULL)
4954 12 : continue;
4955 :
4956 : /*
4957 : * An empty distinct_pathkeys means all tuples have the same
4958 : * value for the DISTINCT clause. See
4959 : * create_final_distinct_paths()
4960 : */
4961 126 : if (root->distinct_pathkeys == NIL)
4962 : {
4963 : Node *limitCount;
4964 :
4965 6 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
4966 : sizeof(int64),
4967 : Int64GetDatum(1), false,
4968 : true);
4969 :
4970 : /*
4971 : * Apply a LimitPath onto the partial path to restrict the
4972 : * tuples from each worker to 1.
4973 : * create_final_distinct_paths will need to apply an
4974 : * additional LimitPath to restrict this to a single row
4975 : * after the Gather node. If the query already has a
4976 : * LIMIT clause, then we could end up with three Limit
4977 : * nodes in the final plan. Consolidating the top two of
4978 : * these could be done, but does not seem worth troubling
4979 : * over.
4980 : */
4981 6 : add_partial_path(partial_distinct_rel, (Path *)
4982 6 : create_limit_path(root, partial_distinct_rel,
4983 : sorted_path,
4984 : NULL,
4985 : limitCount,
4986 : LIMIT_OPTION_COUNT,
4987 : 0, 1));
4988 : }
4989 : else
4990 : {
4991 120 : add_partial_path(partial_distinct_rel, (Path *)
4992 120 : create_unique_path(root, partial_distinct_rel,
4993 : sorted_path,
4994 120 : list_length(root->distinct_pathkeys),
4995 : numDistinctRows));
4996 : }
4997 : }
4998 : }
4999 : }
5000 :
5001 : /*
5002 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5003 : * we're not on the hook to ensure we do our best to create at least one
5004 : * path here, we treat enable_hashagg as a hard off-switch rather than the
5005 : * slightly softer variant in create_final_distinct_paths.
5006 : */
5007 108 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5008 : {
5009 78 : add_partial_path(partial_distinct_rel, (Path *)
5010 78 : create_agg_path(root,
5011 : partial_distinct_rel,
5012 : cheapest_partial_path,
5013 : cheapest_partial_path->pathtarget,
5014 : AGG_HASHED,
5015 : AGGSPLIT_SIMPLE,
5016 : root->processed_distinctClause,
5017 : NIL,
5018 : NULL,
5019 : numDistinctRows));
5020 : }
5021 :
5022 : /*
5023 : * If there is an FDW that's responsible for all baserels of the query,
5024 : * let it consider adding ForeignPaths.
5025 : */
5026 108 : if (partial_distinct_rel->fdwroutine &&
5027 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5028 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5029 : UPPERREL_PARTIAL_DISTINCT,
5030 : input_rel,
5031 : partial_distinct_rel,
5032 : NULL);
5033 :
5034 : /* Let extensions possibly add some more partial paths */
5035 108 : if (create_upper_paths_hook)
5036 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5037 : input_rel, partial_distinct_rel, NULL);
5038 :
5039 108 : if (partial_distinct_rel->partial_pathlist != NIL)
5040 : {
5041 108 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5042 108 : set_cheapest(partial_distinct_rel);
5043 :
5044 : /*
5045 : * Finally, create paths to distinctify the final result. This step
5046 : * is needed to remove any duplicates due to combining rows from
5047 : * parallel workers.
5048 : */
5049 108 : create_final_distinct_paths(root, partial_distinct_rel,
5050 : final_distinct_rel);
5051 : }
5052 : }
5053 :
5054 : /*
5055 : * create_final_distinct_paths
5056 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5057 : *
5058 : * input_rel: contains the source-data paths
5059 : * distinct_rel: destination relation for storing created paths
5060 : */
5061 : static RelOptInfo *
5062 3064 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5063 : RelOptInfo *distinct_rel)
5064 : {
5065 3064 : Query *parse = root->parse;
5066 3064 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5067 : double numDistinctRows;
5068 : bool allow_hash;
5069 :
5070 : /* Estimate number of distinct rows there will be */
5071 3064 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5072 2990 : root->hasHavingQual)
5073 : {
5074 : /*
5075 : * If there was grouping or aggregation, use the number of input rows
5076 : * as the estimated number of DISTINCT rows (ie, assume the input is
5077 : * already mostly unique).
5078 : */
5079 74 : numDistinctRows = cheapest_input_path->rows;
5080 : }
5081 : else
5082 : {
5083 : /*
5084 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5085 : */
5086 : List *distinctExprs;
5087 :
5088 2990 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5089 : parse->targetList);
5090 2990 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5091 : cheapest_input_path->rows,
5092 : NULL, NULL);
5093 : }
5094 :
5095 : /*
5096 : * Consider sort-based implementations of DISTINCT, if possible.
5097 : */
5098 3064 : if (grouping_is_sortable(root->processed_distinctClause))
5099 : {
5100 : /*
5101 : * Firstly, if we have any adequately-presorted paths, just stick a
5102 : * Unique node on those. We also, consider doing an explicit sort of
5103 : * the cheapest input path and Unique'ing that. If any paths have
5104 : * presorted keys then we'll create an incremental sort atop of those
5105 : * before adding a unique node on the top. We'll also attempt to
5106 : * reorder the required pathkeys to match the input path's pathkeys as
5107 : * much as possible, in hopes of avoiding a possible need to re-sort.
5108 : *
5109 : * When we have DISTINCT ON, we must sort by the more rigorous of
5110 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5111 : * Also, if we do have to do an explicit sort, we might as well use
5112 : * the more rigorous ordering to avoid a second sort later. (Note
5113 : * that the parser will have ensured that one clause is a prefix of
5114 : * the other.)
5115 : */
5116 : List *needed_pathkeys;
5117 : ListCell *lc;
5118 3058 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5119 :
5120 3306 : if (parse->hasDistinctOn &&
5121 248 : list_length(root->distinct_pathkeys) <
5122 248 : list_length(root->sort_pathkeys))
5123 54 : needed_pathkeys = root->sort_pathkeys;
5124 : else
5125 3004 : needed_pathkeys = root->distinct_pathkeys;
5126 :
5127 8256 : foreach(lc, input_rel->pathlist)
5128 : {
5129 5198 : Path *input_path = (Path *) lfirst(lc);
5130 : Path *sorted_path;
5131 5198 : List *useful_pathkeys_list = NIL;
5132 :
5133 : useful_pathkeys_list =
5134 5198 : get_useful_pathkeys_for_distinct(root,
5135 : needed_pathkeys,
5136 : input_path->pathkeys);
5137 : Assert(list_length(useful_pathkeys_list) > 0);
5138 :
5139 16398 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5140 : {
5141 6002 : sorted_path = make_ordered_path(root,
5142 : distinct_rel,
5143 : input_path,
5144 : cheapest_input_path,
5145 : useful_pathkeys,
5146 : limittuples);
5147 :
5148 6002 : if (sorted_path == NULL)
5149 862 : continue;
5150 :
5151 : /*
5152 : * distinct_pathkeys may have become empty if all of the
5153 : * pathkeys were determined to be redundant. If all of the
5154 : * pathkeys are redundant then each DISTINCT target must only
5155 : * allow a single value, therefore all resulting tuples must
5156 : * be identical (or at least indistinguishable by an equality
5157 : * check). We can uniquify these tuples simply by just taking
5158 : * the first tuple. All we do here is add a path to do "LIMIT
5159 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5160 : * still have a non-NIL sort_pathkeys list, so we must still
5161 : * only do this with paths which are correctly sorted by
5162 : * sort_pathkeys.
5163 : */
5164 5140 : if (root->distinct_pathkeys == NIL)
5165 : {
5166 : Node *limitCount;
5167 :
5168 138 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5169 : sizeof(int64),
5170 : Int64GetDatum(1), false,
5171 : true);
5172 :
5173 : /*
5174 : * If the query already has a LIMIT clause, then we could
5175 : * end up with a duplicate LimitPath in the final plan.
5176 : * That does not seem worth troubling over too much.
5177 : */
5178 138 : add_path(distinct_rel, (Path *)
5179 138 : create_limit_path(root, distinct_rel, sorted_path,
5180 : NULL, limitCount,
5181 : LIMIT_OPTION_COUNT, 0, 1));
5182 : }
5183 : else
5184 : {
5185 5002 : add_path(distinct_rel, (Path *)
5186 5002 : create_unique_path(root, distinct_rel,
5187 : sorted_path,
5188 5002 : list_length(root->distinct_pathkeys),
5189 : numDistinctRows));
5190 : }
5191 : }
5192 : }
5193 : }
5194 :
5195 : /*
5196 : * Consider hash-based implementations of DISTINCT, if possible.
5197 : *
5198 : * If we were not able to make any other types of path, we *must* hash or
5199 : * die trying. If we do have other choices, there are two things that
5200 : * should prevent selection of hashing: if the query uses DISTINCT ON
5201 : * (because it won't really have the expected behavior if we hash), or if
5202 : * enable_hashagg is off.
5203 : *
5204 : * Note: grouping_is_hashable() is much more expensive to check than the
5205 : * other gating conditions, so we want to do it last.
5206 : */
5207 3064 : if (distinct_rel->pathlist == NIL)
5208 6 : allow_hash = true; /* we have no alternatives */
5209 3058 : else if (parse->hasDistinctOn || !enable_hashagg)
5210 398 : allow_hash = false; /* policy-based decision not to hash */
5211 : else
5212 2660 : allow_hash = true; /* default */
5213 :
5214 3064 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5215 : {
5216 : /* Generate hashed aggregate path --- no sort needed */
5217 2666 : add_path(distinct_rel, (Path *)
5218 2666 : create_agg_path(root,
5219 : distinct_rel,
5220 : cheapest_input_path,
5221 : cheapest_input_path->pathtarget,
5222 : AGG_HASHED,
5223 : AGGSPLIT_SIMPLE,
5224 : root->processed_distinctClause,
5225 : NIL,
5226 : NULL,
5227 : numDistinctRows));
5228 : }
5229 :
5230 3064 : return distinct_rel;
5231 : }
5232 :
5233 : /*
5234 : * get_useful_pathkeys_for_distinct
5235 : * Get useful orderings of pathkeys for distinctClause by reordering
5236 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5237 : *
5238 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5239 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5240 : */
5241 : static List *
5242 5324 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5243 : List *path_pathkeys)
5244 : {
5245 5324 : List *useful_pathkeys_list = NIL;
5246 5324 : List *useful_pathkeys = NIL;
5247 :
5248 : /* always include the given 'needed_pathkeys' */
5249 5324 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5250 : needed_pathkeys);
5251 :
5252 5324 : if (!enable_distinct_reordering)
5253 0 : return useful_pathkeys_list;
5254 :
5255 : /*
5256 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5257 : * that match 'needed_pathkeys', but only up to the longest matching
5258 : * prefix.
5259 : *
5260 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5261 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5262 : * the desired behavior.
5263 : */
5264 13192 : foreach_node(PathKey, pathkey, path_pathkeys)
5265 : {
5266 : /*
5267 : * The PathKey nodes are canonical, so they can be checked for
5268 : * equality by simple pointer comparison.
5269 : */
5270 2572 : if (!list_member_ptr(needed_pathkeys, pathkey))
5271 10 : break;
5272 2562 : if (root->parse->hasDistinctOn &&
5273 200 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5274 18 : break;
5275 :
5276 2544 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5277 : }
5278 :
5279 : /* If no match at all, no point in reordering needed_pathkeys */
5280 5324 : if (useful_pathkeys == NIL)
5281 3044 : return useful_pathkeys_list;
5282 :
5283 : /*
5284 : * If not full match, the resulting pathkey list is not useful without
5285 : * incremental sort.
5286 : */
5287 2280 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5288 1528 : !enable_incremental_sort)
5289 60 : return useful_pathkeys_list;
5290 :
5291 : /* Append the remaining PathKey nodes in needed_pathkeys */
5292 2220 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5293 : needed_pathkeys);
5294 :
5295 : /*
5296 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5297 : * just drop it.
5298 : */
5299 2220 : if (compare_pathkeys(needed_pathkeys,
5300 : useful_pathkeys) == PATHKEYS_EQUAL)
5301 1404 : return useful_pathkeys_list;
5302 :
5303 816 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5304 : useful_pathkeys);
5305 :
5306 816 : return useful_pathkeys_list;
5307 : }
5308 :
5309 : /*
5310 : * create_ordered_paths
5311 : *
5312 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5313 : *
5314 : * All paths in the result must satisfy the ORDER BY ordering.
5315 : * The only new paths we need consider are an explicit full sort
5316 : * and incremental sort on the cheapest-total existing path.
5317 : *
5318 : * input_rel: contains the source-data Paths
5319 : * target: the output tlist the result Paths must emit
5320 : * limit_tuples: estimated bound on the number of output tuples,
5321 : * or -1 if no LIMIT or couldn't estimate
5322 : *
5323 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5324 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5325 : */
5326 : static RelOptInfo *
5327 74036 : create_ordered_paths(PlannerInfo *root,
5328 : RelOptInfo *input_rel,
5329 : PathTarget *target,
5330 : bool target_parallel_safe,
5331 : double limit_tuples)
5332 : {
5333 74036 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5334 : RelOptInfo *ordered_rel;
5335 : ListCell *lc;
5336 :
5337 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5338 74036 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5339 :
5340 : /*
5341 : * If the input relation is not parallel-safe, then the ordered relation
5342 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5343 : * target list is parallel-safe.
5344 : */
5345 74036 : if (input_rel->consider_parallel && target_parallel_safe)
5346 51578 : ordered_rel->consider_parallel = true;
5347 :
5348 : /*
5349 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5350 : */
5351 74036 : ordered_rel->serverid = input_rel->serverid;
5352 74036 : ordered_rel->userid = input_rel->userid;
5353 74036 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5354 74036 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5355 :
5356 187694 : foreach(lc, input_rel->pathlist)
5357 : {
5358 113658 : Path *input_path = (Path *) lfirst(lc);
5359 : Path *sorted_path;
5360 : bool is_sorted;
5361 : int presorted_keys;
5362 :
5363 113658 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5364 : input_path->pathkeys, &presorted_keys);
5365 :
5366 113658 : if (is_sorted)
5367 42528 : sorted_path = input_path;
5368 : else
5369 : {
5370 : /*
5371 : * Try at least sorting the cheapest path and also try
5372 : * incrementally sorting any path which is partially sorted
5373 : * already (no need to deal with paths which have presorted keys
5374 : * when incremental sort is disabled unless it's the cheapest
5375 : * input path).
5376 : */
5377 71130 : if (input_path != cheapest_input_path &&
5378 5976 : (presorted_keys == 0 || !enable_incremental_sort))
5379 1920 : continue;
5380 :
5381 : /*
5382 : * We've no need to consider both a sort and incremental sort.
5383 : * We'll just do a sort if there are no presorted keys and an
5384 : * incremental sort when there are presorted keys.
5385 : */
5386 69210 : if (presorted_keys == 0 || !enable_incremental_sort)
5387 64532 : sorted_path = (Path *) create_sort_path(root,
5388 : ordered_rel,
5389 : input_path,
5390 : root->sort_pathkeys,
5391 : limit_tuples);
5392 : else
5393 4678 : sorted_path = (Path *) create_incremental_sort_path(root,
5394 : ordered_rel,
5395 : input_path,
5396 : root->sort_pathkeys,
5397 : presorted_keys,
5398 : limit_tuples);
5399 : }
5400 :
5401 : /*
5402 : * If the pathtarget of the result path has different expressions from
5403 : * the target to be applied, a projection step is needed.
5404 : */
5405 111738 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5406 294 : sorted_path = apply_projection_to_path(root, ordered_rel,
5407 : sorted_path, target);
5408 :
5409 111738 : add_path(ordered_rel, sorted_path);
5410 : }
5411 :
5412 : /*
5413 : * generate_gather_paths() will have already generated a simple Gather
5414 : * path for the best parallel path, if any, and the loop above will have
5415 : * considered sorting it. Similarly, generate_gather_paths() will also
5416 : * have generated order-preserving Gather Merge plans which can be used
5417 : * without sorting if they happen to match the sort_pathkeys, and the loop
5418 : * above will have handled those as well. However, there's one more
5419 : * possibility: it may make sense to sort the cheapest partial path or
5420 : * incrementally sort any partial path that is partially sorted according
5421 : * to the required output order and then use Gather Merge.
5422 : */
5423 74036 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5424 51374 : input_rel->partial_pathlist != NIL)
5425 : {
5426 : Path *cheapest_partial_path;
5427 :
5428 2686 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5429 :
5430 5578 : foreach(lc, input_rel->partial_pathlist)
5431 : {
5432 2892 : Path *input_path = (Path *) lfirst(lc);
5433 : Path *sorted_path;
5434 : bool is_sorted;
5435 : int presorted_keys;
5436 : double total_groups;
5437 :
5438 2892 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5439 : input_path->pathkeys,
5440 : &presorted_keys);
5441 :
5442 2892 : if (is_sorted)
5443 182 : continue;
5444 :
5445 : /*
5446 : * Try at least sorting the cheapest path and also try
5447 : * incrementally sorting any path which is partially sorted
5448 : * already (no need to deal with paths which have presorted keys
5449 : * when incremental sort is disabled unless it's the cheapest
5450 : * partial path).
5451 : */
5452 2710 : if (input_path != cheapest_partial_path &&
5453 42 : (presorted_keys == 0 || !enable_incremental_sort))
5454 0 : continue;
5455 :
5456 : /*
5457 : * We've no need to consider both a sort and incremental sort.
5458 : * We'll just do a sort if there are no presorted keys and an
5459 : * incremental sort when there are presorted keys.
5460 : */
5461 2710 : if (presorted_keys == 0 || !enable_incremental_sort)
5462 2650 : sorted_path = (Path *) create_sort_path(root,
5463 : ordered_rel,
5464 : input_path,
5465 : root->sort_pathkeys,
5466 : limit_tuples);
5467 : else
5468 60 : sorted_path = (Path *) create_incremental_sort_path(root,
5469 : ordered_rel,
5470 : input_path,
5471 : root->sort_pathkeys,
5472 : presorted_keys,
5473 : limit_tuples);
5474 2710 : total_groups = compute_gather_rows(sorted_path);
5475 : sorted_path = (Path *)
5476 2710 : create_gather_merge_path(root, ordered_rel,
5477 : sorted_path,
5478 : sorted_path->pathtarget,
5479 : root->sort_pathkeys, NULL,
5480 : &total_groups);
5481 :
5482 : /*
5483 : * If the pathtarget of the result path has different expressions
5484 : * from the target to be applied, a projection step is needed.
5485 : */
5486 2710 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5487 6 : sorted_path = apply_projection_to_path(root, ordered_rel,
5488 : sorted_path, target);
5489 :
5490 2710 : add_path(ordered_rel, sorted_path);
5491 : }
5492 : }
5493 :
5494 : /*
5495 : * If there is an FDW that's responsible for all baserels of the query,
5496 : * let it consider adding ForeignPaths.
5497 : */
5498 74036 : if (ordered_rel->fdwroutine &&
5499 384 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5500 370 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5501 : input_rel, ordered_rel,
5502 : NULL);
5503 :
5504 : /* Let extensions possibly add some more paths */
5505 74036 : if (create_upper_paths_hook)
5506 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5507 : input_rel, ordered_rel, NULL);
5508 :
5509 : /*
5510 : * No need to bother with set_cheapest here; grouping_planner does not
5511 : * need us to do it.
5512 : */
5513 : Assert(ordered_rel->pathlist != NIL);
5514 :
5515 74036 : return ordered_rel;
5516 : }
5517 :
5518 :
5519 : /*
5520 : * make_group_input_target
5521 : * Generate appropriate PathTarget for initial input to grouping nodes.
5522 : *
5523 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5524 : * the query's final targetlist; for example, it certainly can't emit any
5525 : * aggregate function calls. This routine generates the correct target
5526 : * for the scan/join subplan.
5527 : *
5528 : * The query target list passed from the parser already contains entries
5529 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5530 : * for variables used only in HAVING clauses; so we need to add those
5531 : * variables to the subplan target list. Also, we flatten all expressions
5532 : * except GROUP BY items into their component variables; other expressions
5533 : * will be computed by the upper plan nodes rather than by the subplan.
5534 : * For example, given a query like
5535 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5536 : * we want to pass this targetlist to the subplan:
5537 : * a+b,c,d
5538 : * where the a+b target will be used by the Sort/Group steps, and the
5539 : * other targets will be used for computing the final results.
5540 : *
5541 : * 'final_target' is the query's final target list (in PathTarget form)
5542 : *
5543 : * The result is the PathTarget to be computed by the Paths returned from
5544 : * query_planner().
5545 : */
5546 : static PathTarget *
5547 39284 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5548 : {
5549 39284 : Query *parse = root->parse;
5550 : PathTarget *input_target;
5551 : List *non_group_cols;
5552 : List *non_group_vars;
5553 : int i;
5554 : ListCell *lc;
5555 :
5556 : /*
5557 : * We must build a target containing all grouping columns, plus any other
5558 : * Vars mentioned in the query's targetlist and HAVING qual.
5559 : */
5560 39284 : input_target = create_empty_pathtarget();
5561 39284 : non_group_cols = NIL;
5562 :
5563 39284 : i = 0;
5564 96390 : foreach(lc, final_target->exprs)
5565 : {
5566 57106 : Expr *expr = (Expr *) lfirst(lc);
5567 57106 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5568 :
5569 66238 : if (sgref && root->processed_groupClause &&
5570 9132 : get_sortgroupref_clause_noerr(sgref,
5571 : root->processed_groupClause) != NULL)
5572 : {
5573 : /*
5574 : * It's a grouping column, so add it to the input target as-is.
5575 : *
5576 : * Note that the target is logically below the grouping step. So
5577 : * with grouping sets we need to remove the RT index of the
5578 : * grouping step if there is any from the target expression.
5579 : */
5580 7376 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5581 : {
5582 : Assert(root->group_rtindex > 0);
5583 : expr = (Expr *)
5584 1896 : remove_nulling_relids((Node *) expr,
5585 1896 : bms_make_singleton(root->group_rtindex),
5586 : NULL);
5587 : }
5588 7376 : add_column_to_pathtarget(input_target, expr, sgref);
5589 : }
5590 : else
5591 : {
5592 : /*
5593 : * Non-grouping column, so just remember the expression for later
5594 : * call to pull_var_clause.
5595 : */
5596 49730 : non_group_cols = lappend(non_group_cols, expr);
5597 : }
5598 :
5599 57106 : i++;
5600 : }
5601 :
5602 : /*
5603 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5604 : */
5605 39284 : if (parse->havingQual)
5606 956 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5607 :
5608 : /*
5609 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5610 : * add them to the input target if not already present. (A Var used
5611 : * directly as a GROUP BY item will be present already.) Note this
5612 : * includes Vars used in resjunk items, so we are covering the needs of
5613 : * ORDER BY and window specifications. Vars used within Aggrefs and
5614 : * WindowFuncs will be pulled out here, too.
5615 : *
5616 : * Note that the target is logically below the grouping step. So with
5617 : * grouping sets we need to remove the RT index of the grouping step if
5618 : * there is any from the non-group Vars.
5619 : */
5620 39284 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5621 : PVC_RECURSE_AGGREGATES |
5622 : PVC_RECURSE_WINDOWFUNCS |
5623 : PVC_INCLUDE_PLACEHOLDERS);
5624 39284 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5625 : {
5626 : Assert(root->group_rtindex > 0);
5627 : non_group_vars = (List *)
5628 866 : remove_nulling_relids((Node *) non_group_vars,
5629 866 : bms_make_singleton(root->group_rtindex),
5630 : NULL);
5631 : }
5632 39284 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5633 :
5634 : /* clean up cruft */
5635 39284 : list_free(non_group_vars);
5636 39284 : list_free(non_group_cols);
5637 :
5638 : /* XXX this causes some redundant cost calculation ... */
5639 39284 : return set_pathtarget_cost_width(root, input_target);
5640 : }
5641 :
5642 : /*
5643 : * make_partial_grouping_target
5644 : * Generate appropriate PathTarget for output of partial aggregate
5645 : * (or partial grouping, if there are no aggregates) nodes.
5646 : *
5647 : * A partial aggregation node needs to emit all the same aggregates that
5648 : * a regular aggregation node would, plus any aggregates used in HAVING;
5649 : * except that the Aggref nodes should be marked as partial aggregates.
5650 : *
5651 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5652 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5653 : * these would be Vars that are grouped by or used in grouping expressions.)
5654 : *
5655 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5656 : * havingQual represents the HAVING clause.
5657 : */
5658 : static PathTarget *
5659 3094 : make_partial_grouping_target(PlannerInfo *root,
5660 : PathTarget *grouping_target,
5661 : Node *havingQual)
5662 : {
5663 : PathTarget *partial_target;
5664 : List *non_group_cols;
5665 : List *non_group_exprs;
5666 : int i;
5667 : ListCell *lc;
5668 :
5669 3094 : partial_target = create_empty_pathtarget();
5670 3094 : non_group_cols = NIL;
5671 :
5672 3094 : i = 0;
5673 11218 : foreach(lc, grouping_target->exprs)
5674 : {
5675 8124 : Expr *expr = (Expr *) lfirst(lc);
5676 8124 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5677 :
5678 12888 : if (sgref && root->processed_groupClause &&
5679 4764 : get_sortgroupref_clause_noerr(sgref,
5680 : root->processed_groupClause) != NULL)
5681 : {
5682 : /*
5683 : * It's a grouping column, so add it to the partial_target as-is.
5684 : * (This allows the upper agg step to repeat the grouping calcs.)
5685 : */
5686 2822 : add_column_to_pathtarget(partial_target, expr, sgref);
5687 : }
5688 : else
5689 : {
5690 : /*
5691 : * Non-grouping column, so just remember the expression for later
5692 : * call to pull_var_clause.
5693 : */
5694 5302 : non_group_cols = lappend(non_group_cols, expr);
5695 : }
5696 :
5697 8124 : i++;
5698 : }
5699 :
5700 : /*
5701 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5702 : */
5703 3094 : if (havingQual)
5704 878 : non_group_cols = lappend(non_group_cols, havingQual);
5705 :
5706 : /*
5707 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5708 : * non-group cols (plus HAVING), and add them to the partial_target if not
5709 : * already present. (An expression used directly as a GROUP BY item will
5710 : * be present already.) Note this includes Vars used in resjunk items, so
5711 : * we are covering the needs of ORDER BY and window specifications.
5712 : */
5713 3094 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5714 : PVC_INCLUDE_AGGREGATES |
5715 : PVC_RECURSE_WINDOWFUNCS |
5716 : PVC_INCLUDE_PLACEHOLDERS);
5717 :
5718 3094 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5719 :
5720 : /*
5721 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5722 : * are at the top level of the target list, so we can just scan the list
5723 : * rather than recursing through the expression trees.
5724 : */
5725 11814 : foreach(lc, partial_target->exprs)
5726 : {
5727 8720 : Aggref *aggref = (Aggref *) lfirst(lc);
5728 :
5729 8720 : if (IsA(aggref, Aggref))
5730 : {
5731 : Aggref *newaggref;
5732 :
5733 : /*
5734 : * We shouldn't need to copy the substructure of the Aggref node,
5735 : * but flat-copy the node itself to avoid damaging other trees.
5736 : */
5737 5868 : newaggref = makeNode(Aggref);
5738 5868 : memcpy(newaggref, aggref, sizeof(Aggref));
5739 :
5740 : /* For now, assume serialization is required */
5741 5868 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5742 :
5743 5868 : lfirst(lc) = newaggref;
5744 : }
5745 : }
5746 :
5747 : /* clean up cruft */
5748 3094 : list_free(non_group_exprs);
5749 3094 : list_free(non_group_cols);
5750 :
5751 : /* XXX this causes some redundant cost calculation ... */
5752 3094 : return set_pathtarget_cost_width(root, partial_target);
5753 : }
5754 :
5755 : /*
5756 : * mark_partial_aggref
5757 : * Adjust an Aggref to make it represent a partial-aggregation step.
5758 : *
5759 : * The Aggref node is modified in-place; caller must do any copying required.
5760 : */
5761 : void
5762 17692 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5763 : {
5764 : /* aggtranstype should be computed by this point */
5765 : Assert(OidIsValid(agg->aggtranstype));
5766 : /* ... but aggsplit should still be as the parser left it */
5767 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5768 :
5769 : /* Mark the Aggref with the intended partial-aggregation mode */
5770 17692 : agg->aggsplit = aggsplit;
5771 :
5772 : /*
5773 : * Adjust result type if needed. Normally, a partial aggregate returns
5774 : * the aggregate's transition type; but if that's INTERNAL and we're
5775 : * serializing, it returns BYTEA instead.
5776 : */
5777 17692 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5778 : {
5779 15446 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5780 314 : agg->aggtype = BYTEAOID;
5781 : else
5782 15132 : agg->aggtype = agg->aggtranstype;
5783 : }
5784 17692 : }
5785 :
5786 : /*
5787 : * postprocess_setop_tlist
5788 : * Fix up targetlist returned by plan_set_operations().
5789 : *
5790 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5791 : * NOTE: this would not be good enough if we supported resjunk sort keys
5792 : * for results of set operations --- then, we'd need to project a whole
5793 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5794 : * find any resjunk columns in orig_tlist.
5795 : */
5796 : static List *
5797 6096 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5798 : {
5799 : ListCell *l;
5800 6096 : ListCell *orig_tlist_item = list_head(orig_tlist);
5801 :
5802 23370 : foreach(l, new_tlist)
5803 : {
5804 17274 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5805 : TargetEntry *orig_tle;
5806 :
5807 : /* ignore resjunk columns in setop result */
5808 17274 : if (new_tle->resjunk)
5809 0 : continue;
5810 :
5811 : Assert(orig_tlist_item != NULL);
5812 17274 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5813 17274 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5814 17274 : if (orig_tle->resjunk) /* should not happen */
5815 0 : elog(ERROR, "resjunk output columns are not implemented");
5816 : Assert(new_tle->resno == orig_tle->resno);
5817 17274 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5818 : }
5819 6096 : if (orig_tlist_item != NULL)
5820 0 : elog(ERROR, "resjunk output columns are not implemented");
5821 6096 : return new_tlist;
5822 : }
5823 :
5824 : /*
5825 : * optimize_window_clauses
5826 : * Call each WindowFunc's prosupport function to see if we're able to
5827 : * make any adjustments to any of the WindowClause's so that the executor
5828 : * can execute the window functions in a more optimal way.
5829 : *
5830 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5831 : * may allow more things to be done here in the future.
5832 : */
5833 : static void
5834 2564 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5835 : {
5836 2564 : List *windowClause = root->parse->windowClause;
5837 : ListCell *lc;
5838 :
5839 5374 : foreach(lc, windowClause)
5840 : {
5841 2810 : WindowClause *wc = lfirst_node(WindowClause, lc);
5842 : ListCell *lc2;
5843 2810 : int optimizedFrameOptions = 0;
5844 :
5845 : Assert(wc->winref <= wflists->maxWinRef);
5846 :
5847 : /* skip any WindowClauses that have no WindowFuncs */
5848 2810 : if (wflists->windowFuncs[wc->winref] == NIL)
5849 24 : continue;
5850 :
5851 3440 : foreach(lc2, wflists->windowFuncs[wc->winref])
5852 : {
5853 : SupportRequestOptimizeWindowClause req;
5854 : SupportRequestOptimizeWindowClause *res;
5855 2828 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5856 : Oid prosupport;
5857 :
5858 2828 : prosupport = get_func_support(wfunc->winfnoid);
5859 :
5860 : /* Check if there's a support function for 'wfunc' */
5861 2828 : if (!OidIsValid(prosupport))
5862 2174 : break; /* can't optimize this WindowClause */
5863 :
5864 880 : req.type = T_SupportRequestOptimizeWindowClause;
5865 880 : req.window_clause = wc;
5866 880 : req.window_func = wfunc;
5867 880 : req.frameOptions = wc->frameOptions;
5868 :
5869 : /* call the support function */
5870 : res = (SupportRequestOptimizeWindowClause *)
5871 880 : DatumGetPointer(OidFunctionCall1(prosupport,
5872 : PointerGetDatum(&req)));
5873 :
5874 : /*
5875 : * Skip to next WindowClause if the support function does not
5876 : * support this request type.
5877 : */
5878 880 : if (res == NULL)
5879 226 : break;
5880 :
5881 : /*
5882 : * Save these frameOptions for the first WindowFunc for this
5883 : * WindowClause.
5884 : */
5885 654 : if (foreach_current_index(lc2) == 0)
5886 630 : optimizedFrameOptions = res->frameOptions;
5887 :
5888 : /*
5889 : * On subsequent WindowFuncs, if the frameOptions are not the same
5890 : * then we're unable to optimize the frameOptions for this
5891 : * WindowClause.
5892 : */
5893 24 : else if (optimizedFrameOptions != res->frameOptions)
5894 0 : break; /* skip to the next WindowClause, if any */
5895 : }
5896 :
5897 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
5898 2786 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5899 : {
5900 : ListCell *lc3;
5901 :
5902 : /* apply the new frame options */
5903 612 : wc->frameOptions = optimizedFrameOptions;
5904 :
5905 : /*
5906 : * We now check to see if changing the frameOptions has caused
5907 : * this WindowClause to be a duplicate of some other WindowClause.
5908 : * This can only happen if we have multiple WindowClauses, so
5909 : * don't bother if there's only 1.
5910 : */
5911 612 : if (list_length(windowClause) == 1)
5912 522 : continue;
5913 :
5914 : /*
5915 : * Do the duplicate check and reuse the existing WindowClause if
5916 : * we find a duplicate.
5917 : */
5918 228 : foreach(lc3, windowClause)
5919 : {
5920 174 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5921 :
5922 : /* skip over the WindowClause we're currently editing */
5923 174 : if (existing_wc == wc)
5924 54 : continue;
5925 :
5926 : /*
5927 : * Perform the same duplicate check that is done in
5928 : * transformWindowFuncCall.
5929 : */
5930 240 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
5931 120 : equal(wc->orderClause, existing_wc->orderClause) &&
5932 120 : wc->frameOptions == existing_wc->frameOptions &&
5933 72 : equal(wc->startOffset, existing_wc->startOffset) &&
5934 36 : equal(wc->endOffset, existing_wc->endOffset))
5935 : {
5936 : ListCell *lc4;
5937 :
5938 : /*
5939 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
5940 : * This required adjusting each WindowFunc's winref and
5941 : * moving the WindowFuncs in 'wc' to the list of
5942 : * WindowFuncs in 'existing_wc'.
5943 : */
5944 78 : foreach(lc4, wflists->windowFuncs[wc->winref])
5945 : {
5946 42 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
5947 :
5948 42 : wfunc->winref = existing_wc->winref;
5949 : }
5950 :
5951 : /* move list items */
5952 72 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
5953 36 : wflists->windowFuncs[wc->winref]);
5954 36 : wflists->windowFuncs[wc->winref] = NIL;
5955 :
5956 : /*
5957 : * transformWindowFuncCall() should have made sure there
5958 : * are no other duplicates, so we needn't bother looking
5959 : * any further.
5960 : */
5961 36 : break;
5962 : }
5963 : }
5964 : }
5965 : }
5966 2564 : }
5967 :
5968 : /*
5969 : * select_active_windows
5970 : * Create a list of the "active" window clauses (ie, those referenced
5971 : * by non-deleted WindowFuncs) in the order they are to be executed.
5972 : */
5973 : static List *
5974 2564 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5975 : {
5976 2564 : List *windowClause = root->parse->windowClause;
5977 2564 : List *result = NIL;
5978 : ListCell *lc;
5979 2564 : int nActive = 0;
5980 2564 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5981 2564 : * list_length(windowClause));
5982 :
5983 : /* First, construct an array of the active windows */
5984 5374 : foreach(lc, windowClause)
5985 : {
5986 2810 : WindowClause *wc = lfirst_node(WindowClause, lc);
5987 :
5988 : /* It's only active if wflists shows some related WindowFuncs */
5989 : Assert(wc->winref <= wflists->maxWinRef);
5990 2810 : if (wflists->windowFuncs[wc->winref] == NIL)
5991 60 : continue;
5992 :
5993 2750 : actives[nActive].wc = wc; /* original clause */
5994 :
5995 : /*
5996 : * For sorting, we want the list of partition keys followed by the
5997 : * list of sort keys. But pathkeys construction will remove duplicates
5998 : * between the two, so we can as well (even though we can't detect all
5999 : * of the duplicates, since some may come from ECs - that might mean
6000 : * we miss optimization chances here). We must, however, ensure that
6001 : * the order of entries is preserved with respect to the ones we do
6002 : * keep.
6003 : *
6004 : * partitionClause and orderClause had their own duplicates removed in
6005 : * parse analysis, so we're only concerned here with removing
6006 : * orderClause entries that also appear in partitionClause.
6007 : */
6008 5500 : actives[nActive].uniqueOrder =
6009 2750 : list_concat_unique(list_copy(wc->partitionClause),
6010 2750 : wc->orderClause);
6011 2750 : nActive++;
6012 : }
6013 :
6014 : /*
6015 : * Sort active windows by their partitioning/ordering clauses, ignoring
6016 : * any framing clauses, so that the windows that need the same sorting are
6017 : * adjacent in the list. When we come to generate paths, this will avoid
6018 : * inserting additional Sort nodes.
6019 : *
6020 : * This is how we implement a specific requirement from the SQL standard,
6021 : * which says that when two or more windows are order-equivalent (i.e.
6022 : * have matching partition and order clauses, even if their names or
6023 : * framing clauses differ), then all peer rows must be presented in the
6024 : * same order in all of them. If we allowed multiple sort nodes for such
6025 : * cases, we'd risk having the peer rows end up in different orders in
6026 : * equivalent windows due to sort instability. (See General Rule 4 of
6027 : * <window clause> in SQL2008 - SQL2016.)
6028 : *
6029 : * Additionally, if the entire list of clauses of one window is a prefix
6030 : * of another, put first the window with stronger sorting requirements.
6031 : * This way we will first sort for stronger window, and won't have to sort
6032 : * again for the weaker one.
6033 : */
6034 2564 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6035 :
6036 : /* build ordered list of the original WindowClause nodes */
6037 5314 : for (int i = 0; i < nActive; i++)
6038 2750 : result = lappend(result, actives[i].wc);
6039 :
6040 2564 : pfree(actives);
6041 :
6042 2564 : return result;
6043 : }
6044 :
6045 : /*
6046 : * name_active_windows
6047 : * Ensure all active windows have unique names.
6048 : *
6049 : * The parser will have checked that user-assigned window names are unique
6050 : * within the Query. Here we assign made-up names to any unnamed
6051 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6052 : * at parse time, because it'd mess up decompilation of views.)
6053 : *
6054 : * activeWindows: result of select_active_windows
6055 : */
6056 : static void
6057 2564 : name_active_windows(List *activeWindows)
6058 : {
6059 2564 : int next_n = 1;
6060 : char newname[16];
6061 : ListCell *lc;
6062 :
6063 5314 : foreach(lc, activeWindows)
6064 : {
6065 2750 : WindowClause *wc = lfirst_node(WindowClause, lc);
6066 :
6067 : /* Nothing to do if it has a name already. */
6068 2750 : if (wc->name)
6069 576 : continue;
6070 :
6071 : /* Select a name not currently present in the list. */
6072 : for (;;)
6073 6 : {
6074 : ListCell *lc2;
6075 :
6076 2180 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6077 4708 : foreach(lc2, activeWindows)
6078 : {
6079 2534 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6080 :
6081 2534 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6082 6 : break; /* matched */
6083 : }
6084 2180 : if (lc2 == NULL)
6085 2174 : break; /* reached the end with no match */
6086 : }
6087 2174 : wc->name = pstrdup(newname);
6088 : }
6089 2564 : }
6090 :
6091 : /*
6092 : * common_prefix_cmp
6093 : * QSort comparison function for WindowClauseSortData
6094 : *
6095 : * Sort the windows by the required sorting clauses. First, compare the sort
6096 : * clauses themselves. Second, if one window's clauses are a prefix of another
6097 : * one's clauses, put the window with more sort clauses first.
6098 : *
6099 : * We purposefully sort by the highest tleSortGroupRef first. Since
6100 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6101 : * and because here we sort the lowest tleSortGroupRefs last, if a
6102 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6103 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6104 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6105 : * reducing the total number of sorts required for the query.
6106 : */
6107 : static int
6108 204 : common_prefix_cmp(const void *a, const void *b)
6109 : {
6110 204 : const WindowClauseSortData *wcsa = a;
6111 204 : const WindowClauseSortData *wcsb = b;
6112 : ListCell *item_a;
6113 : ListCell *item_b;
6114 :
6115 366 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6116 : {
6117 264 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6118 264 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6119 :
6120 264 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6121 102 : return -1;
6122 252 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6123 66 : return 1;
6124 186 : else if (sca->sortop > scb->sortop)
6125 0 : return -1;
6126 186 : else if (sca->sortop < scb->sortop)
6127 24 : return 1;
6128 162 : else if (sca->nulls_first && !scb->nulls_first)
6129 0 : return -1;
6130 162 : else if (!sca->nulls_first && scb->nulls_first)
6131 0 : return 1;
6132 : /* no need to compare eqop, since it is fully determined by sortop */
6133 : }
6134 :
6135 102 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6136 6 : return -1;
6137 96 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6138 30 : return 1;
6139 :
6140 66 : return 0;
6141 : }
6142 :
6143 : /*
6144 : * make_window_input_target
6145 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6146 : *
6147 : * When the query has window functions, this function computes the desired
6148 : * target to be computed by the node just below the first WindowAgg.
6149 : * This tlist must contain all values needed to evaluate the window functions,
6150 : * compute the final target list, and perform any required final sort step.
6151 : * If multiple WindowAggs are needed, each intermediate one adds its window
6152 : * function results onto this base tlist; only the topmost WindowAgg computes
6153 : * the actual desired target list.
6154 : *
6155 : * This function is much like make_group_input_target, though not quite enough
6156 : * like it to share code. As in that function, we flatten most expressions
6157 : * into their component variables. But we do not want to flatten window
6158 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6159 : * evaluations of them, which would be bad (possibly even resulting in
6160 : * inconsistent answers, if they contain volatile functions).
6161 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6162 : * make_group_input_target, because we may no longer have access to the
6163 : * individual Vars in them.
6164 : *
6165 : * Another key difference from make_group_input_target is that we don't
6166 : * flatten Aggref expressions, since those are to be computed below the
6167 : * window functions and just referenced like Vars above that.
6168 : *
6169 : * 'final_target' is the query's final target list (in PathTarget form)
6170 : * 'activeWindows' is the list of active windows previously identified by
6171 : * select_active_windows.
6172 : *
6173 : * The result is the PathTarget to be computed by the plan node immediately
6174 : * below the first WindowAgg node.
6175 : */
6176 : static PathTarget *
6177 2564 : make_window_input_target(PlannerInfo *root,
6178 : PathTarget *final_target,
6179 : List *activeWindows)
6180 : {
6181 : PathTarget *input_target;
6182 : Bitmapset *sgrefs;
6183 : List *flattenable_cols;
6184 : List *flattenable_vars;
6185 : int i;
6186 : ListCell *lc;
6187 :
6188 : Assert(root->parse->hasWindowFuncs);
6189 :
6190 : /*
6191 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6192 : * into a bitmapset for convenient reference below.
6193 : */
6194 2564 : sgrefs = NULL;
6195 5314 : foreach(lc, activeWindows)
6196 : {
6197 2750 : WindowClause *wc = lfirst_node(WindowClause, lc);
6198 : ListCell *lc2;
6199 :
6200 3506 : foreach(lc2, wc->partitionClause)
6201 : {
6202 756 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6203 :
6204 756 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6205 : }
6206 5010 : foreach(lc2, wc->orderClause)
6207 : {
6208 2260 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6209 :
6210 2260 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6211 : }
6212 : }
6213 :
6214 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6215 2756 : foreach(lc, root->processed_groupClause)
6216 : {
6217 192 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6218 :
6219 192 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6220 : }
6221 :
6222 : /*
6223 : * Construct a target containing all the non-flattenable targetlist items,
6224 : * and save aside the others for a moment.
6225 : */
6226 2564 : input_target = create_empty_pathtarget();
6227 2564 : flattenable_cols = NIL;
6228 :
6229 2564 : i = 0;
6230 10892 : foreach(lc, final_target->exprs)
6231 : {
6232 8328 : Expr *expr = (Expr *) lfirst(lc);
6233 8328 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6234 :
6235 : /*
6236 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6237 : * that such items can't contain window functions, so it's okay to
6238 : * compute them below the WindowAgg nodes.)
6239 : */
6240 8328 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6241 : {
6242 : /*
6243 : * Don't want to deconstruct this value, so add it to the input
6244 : * target as-is.
6245 : */
6246 2852 : add_column_to_pathtarget(input_target, expr, sgref);
6247 : }
6248 : else
6249 : {
6250 : /*
6251 : * Column is to be flattened, so just remember the expression for
6252 : * later call to pull_var_clause.
6253 : */
6254 5476 : flattenable_cols = lappend(flattenable_cols, expr);
6255 : }
6256 :
6257 8328 : i++;
6258 : }
6259 :
6260 : /*
6261 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6262 : * add them to the input target if not already present. (Some might be
6263 : * there already because they're used directly as window/group clauses.)
6264 : *
6265 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6266 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6267 : * at higher levels. On the other hand, we should recurse into
6268 : * WindowFuncs to make sure their input expressions are available.
6269 : */
6270 2564 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6271 : PVC_INCLUDE_AGGREGATES |
6272 : PVC_RECURSE_WINDOWFUNCS |
6273 : PVC_INCLUDE_PLACEHOLDERS);
6274 2564 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6275 :
6276 : /* clean up cruft */
6277 2564 : list_free(flattenable_vars);
6278 2564 : list_free(flattenable_cols);
6279 :
6280 : /* XXX this causes some redundant cost calculation ... */
6281 2564 : return set_pathtarget_cost_width(root, input_target);
6282 : }
6283 :
6284 : /*
6285 : * make_pathkeys_for_window
6286 : * Create a pathkeys list describing the required input ordering
6287 : * for the given WindowClause.
6288 : *
6289 : * Modifies wc's partitionClause to remove any clauses which are deemed
6290 : * redundant by the pathkey logic.
6291 : *
6292 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6293 : * In the future we might try to implement windowing using hashing, in which
6294 : * case the ordering could be relaxed, but for now we always sort.
6295 : */
6296 : static List *
6297 5528 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6298 : List *tlist)
6299 : {
6300 5528 : List *window_pathkeys = NIL;
6301 :
6302 : /* Throw error if can't sort */
6303 5528 : if (!grouping_is_sortable(wc->partitionClause))
6304 0 : ereport(ERROR,
6305 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6306 : errmsg("could not implement window PARTITION BY"),
6307 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6308 5528 : if (!grouping_is_sortable(wc->orderClause))
6309 0 : ereport(ERROR,
6310 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6311 : errmsg("could not implement window ORDER BY"),
6312 : errdetail("Window ordering columns must be of sortable datatypes.")));
6313 :
6314 : /*
6315 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6316 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6317 : */
6318 5528 : if (wc->partitionClause != NIL)
6319 : {
6320 : bool sortable;
6321 :
6322 1320 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6323 : &wc->partitionClause,
6324 : tlist,
6325 : true,
6326 : false,
6327 : &sortable,
6328 : false);
6329 :
6330 : Assert(sortable);
6331 : }
6332 :
6333 : /*
6334 : * In principle, we could also consider removing redundant ORDER BY items
6335 : * too as doing so does not alter the result of peer row checks done by
6336 : * the executor. However, we must *not* remove the ordering column for
6337 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6338 : * if it's known to be equal to some partitioning column.
6339 : */
6340 5528 : if (wc->orderClause != NIL)
6341 : {
6342 : List *orderby_pathkeys;
6343 :
6344 4426 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6345 : wc->orderClause,
6346 : tlist);
6347 :
6348 : /* Okay, make the combined pathkeys */
6349 4426 : if (window_pathkeys != NIL)
6350 946 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6351 : else
6352 3480 : window_pathkeys = orderby_pathkeys;
6353 : }
6354 :
6355 5528 : return window_pathkeys;
6356 : }
6357 :
6358 : /*
6359 : * make_sort_input_target
6360 : * Generate appropriate PathTarget for initial input to Sort step.
6361 : *
6362 : * If the query has ORDER BY, this function chooses the target to be computed
6363 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6364 : * project) steps. This might or might not be identical to the query's final
6365 : * output target.
6366 : *
6367 : * The main argument for keeping the sort-input tlist the same as the final
6368 : * is that we avoid a separate projection node (which will be needed if
6369 : * they're different, because Sort can't project). However, there are also
6370 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6371 : * a consistent order of evaluation for any volatile functions in the tlist,
6372 : * and if there's also a LIMIT, we can stop the query without ever computing
6373 : * tlist functions for later rows, which is beneficial for both volatile and
6374 : * expensive functions.
6375 : *
6376 : * Our current policy is to postpone volatile expressions till after the sort
6377 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6378 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6379 : * postpone set-returning expressions, because running them beforehand would
6380 : * bloat the sort dataset, and because it might cause unexpected output order
6381 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6382 : * in the tlist should be evaluated at the same plan step, so that they can
6383 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6384 : * mustn't postpone any SRFs. (Note that in principle that policy should
6385 : * probably get applied to the group/window input targetlists too, but we
6386 : * have not done that historically.) Lastly, expensive expressions are
6387 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6388 : * partial evaluation of the query is possible (if neither is true, we expect
6389 : * to have to evaluate the expressions for every row anyway), or if there are
6390 : * any volatile or set-returning expressions (since once we've put in a
6391 : * projection at all, it won't cost any more to postpone more stuff).
6392 : *
6393 : * Another issue that could potentially be considered here is that
6394 : * evaluating tlist expressions could result in data that's either wider
6395 : * or narrower than the input Vars, thus changing the volume of data that
6396 : * has to go through the Sort. However, we usually have only a very bad
6397 : * idea of the output width of any expression more complex than a Var,
6398 : * so for now it seems too risky to try to optimize on that basis.
6399 : *
6400 : * Note that if we do produce a modified sort-input target, and then the
6401 : * query ends up not using an explicit Sort, no particular harm is done:
6402 : * we'll initially use the modified target for the preceding path nodes,
6403 : * but then change them to the final target with apply_projection_to_path.
6404 : * Moreover, in such a case the guarantees about evaluation order of
6405 : * volatile functions still hold, since the rows are sorted already.
6406 : *
6407 : * This function has some things in common with make_group_input_target and
6408 : * make_window_input_target, though the detailed rules for what to do are
6409 : * different. We never flatten/postpone any grouping or ordering columns;
6410 : * those are needed before the sort. If we do flatten a particular
6411 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6412 : * computed earlier.
6413 : *
6414 : * 'final_target' is the query's final target list (in PathTarget form)
6415 : * 'have_postponed_srfs' is an output argument, see below
6416 : *
6417 : * The result is the PathTarget to be computed by the plan node immediately
6418 : * below the Sort step (and the Distinct step, if any). This will be
6419 : * exactly final_target if we decide a projection step wouldn't be helpful.
6420 : *
6421 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6422 : * any set-returning functions to after the Sort.
6423 : */
6424 : static PathTarget *
6425 70148 : make_sort_input_target(PlannerInfo *root,
6426 : PathTarget *final_target,
6427 : bool *have_postponed_srfs)
6428 : {
6429 70148 : Query *parse = root->parse;
6430 : PathTarget *input_target;
6431 : int ncols;
6432 : bool *col_is_srf;
6433 : bool *postpone_col;
6434 : bool have_srf;
6435 : bool have_volatile;
6436 : bool have_expensive;
6437 : bool have_srf_sortcols;
6438 : bool postpone_srfs;
6439 : List *postponable_cols;
6440 : List *postponable_vars;
6441 : int i;
6442 : ListCell *lc;
6443 :
6444 : /* Shouldn't get here unless query has ORDER BY */
6445 : Assert(parse->sortClause);
6446 :
6447 70148 : *have_postponed_srfs = false; /* default result */
6448 :
6449 : /* Inspect tlist and collect per-column information */
6450 70148 : ncols = list_length(final_target->exprs);
6451 70148 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6452 70148 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6453 70148 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6454 :
6455 70148 : i = 0;
6456 423292 : foreach(lc, final_target->exprs)
6457 : {
6458 353144 : Expr *expr = (Expr *) lfirst(lc);
6459 :
6460 : /*
6461 : * If the column has a sortgroupref, assume it has to be evaluated
6462 : * before sorting. Generally such columns would be ORDER BY, GROUP
6463 : * BY, etc targets. One exception is columns that were removed from
6464 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6465 : * only be Vars anyway. There don't seem to be any cases where it
6466 : * would be worth the trouble to double-check.
6467 : */
6468 353144 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6469 : {
6470 : /*
6471 : * Check for SRF or volatile functions. Check the SRF case first
6472 : * because we must know whether we have any postponed SRFs.
6473 : */
6474 254504 : if (parse->hasTargetSRFs &&
6475 216 : expression_returns_set((Node *) expr))
6476 : {
6477 : /* We'll decide below whether these are postponable */
6478 96 : col_is_srf[i] = true;
6479 96 : have_srf = true;
6480 : }
6481 254192 : else if (contain_volatile_functions((Node *) expr))
6482 : {
6483 : /* Unconditionally postpone */
6484 148 : postpone_col[i] = true;
6485 148 : have_volatile = true;
6486 : }
6487 : else
6488 : {
6489 : /*
6490 : * Else check the cost. XXX it's annoying to have to do this
6491 : * when set_pathtarget_cost_width() just did it. Refactor to
6492 : * allow sharing the work?
6493 : */
6494 : QualCost cost;
6495 :
6496 254044 : cost_qual_eval_node(&cost, (Node *) expr, root);
6497 :
6498 : /*
6499 : * We arbitrarily define "expensive" as "more than 10X
6500 : * cpu_operator_cost". Note this will take in any PL function
6501 : * with default cost.
6502 : */
6503 254044 : if (cost.per_tuple > 10 * cpu_operator_cost)
6504 : {
6505 16512 : postpone_col[i] = true;
6506 16512 : have_expensive = true;
6507 : }
6508 : }
6509 : }
6510 : else
6511 : {
6512 : /* For sortgroupref cols, just check if any contain SRFs */
6513 98856 : if (!have_srf_sortcols &&
6514 99166 : parse->hasTargetSRFs &&
6515 310 : expression_returns_set((Node *) expr))
6516 124 : have_srf_sortcols = true;
6517 : }
6518 :
6519 353144 : i++;
6520 : }
6521 :
6522 : /*
6523 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6524 : */
6525 70148 : postpone_srfs = (have_srf && !have_srf_sortcols);
6526 :
6527 : /*
6528 : * If we don't need a post-sort projection, just return final_target.
6529 : */
6530 70148 : if (!(postpone_srfs || have_volatile ||
6531 69944 : (have_expensive &&
6532 9708 : (parse->limitCount || root->tuple_fraction > 0))))
6533 69908 : return final_target;
6534 :
6535 : /*
6536 : * Report whether the post-sort projection will contain set-returning
6537 : * functions. This is important because it affects whether the Sort can
6538 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6539 : * to return.
6540 : */
6541 240 : *have_postponed_srfs = postpone_srfs;
6542 :
6543 : /*
6544 : * Construct the sort-input target, taking all non-postponable columns and
6545 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6546 : * the postponable ones.
6547 : */
6548 240 : input_target = create_empty_pathtarget();
6549 240 : postponable_cols = NIL;
6550 :
6551 240 : i = 0;
6552 1990 : foreach(lc, final_target->exprs)
6553 : {
6554 1750 : Expr *expr = (Expr *) lfirst(lc);
6555 :
6556 1750 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6557 298 : postponable_cols = lappend(postponable_cols, expr);
6558 : else
6559 1452 : add_column_to_pathtarget(input_target, expr,
6560 1452 : get_pathtarget_sortgroupref(final_target, i));
6561 :
6562 1750 : i++;
6563 : }
6564 :
6565 : /*
6566 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6567 : * postponable columns, and add them to the sort-input target if not
6568 : * already present. (Some might be there already.) We mustn't
6569 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6570 : * would be unable to recompute them.
6571 : */
6572 240 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6573 : PVC_INCLUDE_AGGREGATES |
6574 : PVC_INCLUDE_WINDOWFUNCS |
6575 : PVC_INCLUDE_PLACEHOLDERS);
6576 240 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6577 :
6578 : /* clean up cruft */
6579 240 : list_free(postponable_vars);
6580 240 : list_free(postponable_cols);
6581 :
6582 : /* XXX this represents even more redundant cost calculation ... */
6583 240 : return set_pathtarget_cost_width(root, input_target);
6584 : }
6585 :
6586 : /*
6587 : * get_cheapest_fractional_path
6588 : * Find the cheapest path for retrieving a specified fraction of all
6589 : * the tuples expected to be returned by the given relation.
6590 : *
6591 : * Do not consider parameterized paths. If the caller needs a path for upper
6592 : * rel, it can't have parameterized paths. If the caller needs an append
6593 : * subpath, it could become limited by the treatment of similar
6594 : * parameterization of all the subpaths.
6595 : *
6596 : * We interpret tuple_fraction the same way as grouping_planner.
6597 : *
6598 : * We assume set_cheapest() has been run on the given rel.
6599 : */
6600 : Path *
6601 486364 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6602 : {
6603 486364 : Path *best_path = rel->cheapest_total_path;
6604 : ListCell *l;
6605 :
6606 : /* If all tuples will be retrieved, just return the cheapest-total path */
6607 486364 : if (tuple_fraction <= 0.0)
6608 476972 : return best_path;
6609 :
6610 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6611 9392 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6612 3806 : tuple_fraction /= best_path->rows;
6613 :
6614 24862 : foreach(l, rel->pathlist)
6615 : {
6616 15470 : Path *path = (Path *) lfirst(l);
6617 :
6618 15470 : if (path->param_info)
6619 200 : continue;
6620 :
6621 21148 : if (path == rel->cheapest_total_path ||
6622 5878 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6623 14754 : continue;
6624 :
6625 516 : best_path = path;
6626 : }
6627 :
6628 9392 : return best_path;
6629 : }
6630 :
6631 : /*
6632 : * adjust_paths_for_srfs
6633 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6634 : *
6635 : * The executor can only handle set-returning functions that appear at the
6636 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6637 : * that are not at top level, we need to split up the evaluation into multiple
6638 : * plan levels in which each level satisfies this constraint. This function
6639 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6640 : * output tlist to insert appropriate projection steps.
6641 : *
6642 : * The given targets and targets_contain_srfs lists are from
6643 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6644 : * target in targets.
6645 : */
6646 : static void
6647 12344 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6648 : List *targets, List *targets_contain_srfs)
6649 : {
6650 : ListCell *lc;
6651 :
6652 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6653 : Assert(!linitial_int(targets_contain_srfs));
6654 :
6655 : /* If no SRFs appear at this plan level, nothing to do */
6656 12344 : if (list_length(targets) == 1)
6657 640 : return;
6658 :
6659 : /*
6660 : * Stack SRF-evaluation nodes atop each path for the rel.
6661 : *
6662 : * In principle we should re-run set_cheapest() here to identify the
6663 : * cheapest path, but it seems unlikely that adding the same tlist eval
6664 : * costs to all the paths would change that, so we don't bother. Instead,
6665 : * just assume that the cheapest-startup and cheapest-total paths remain
6666 : * so. (There should be no parameterized paths anymore, so we needn't
6667 : * worry about updating cheapest_parameterized_paths.)
6668 : */
6669 23432 : foreach(lc, rel->pathlist)
6670 : {
6671 11728 : Path *subpath = (Path *) lfirst(lc);
6672 11728 : Path *newpath = subpath;
6673 : ListCell *lc1,
6674 : *lc2;
6675 :
6676 : Assert(subpath->param_info == NULL);
6677 36372 : forboth(lc1, targets, lc2, targets_contain_srfs)
6678 : {
6679 24644 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6680 24644 : bool contains_srfs = (bool) lfirst_int(lc2);
6681 :
6682 : /* If this level doesn't contain SRFs, do regular projection */
6683 24644 : if (contains_srfs)
6684 11788 : newpath = (Path *) create_set_projection_path(root,
6685 : rel,
6686 : newpath,
6687 : thistarget);
6688 : else
6689 12856 : newpath = (Path *) apply_projection_to_path(root,
6690 : rel,
6691 : newpath,
6692 : thistarget);
6693 : }
6694 11728 : lfirst(lc) = newpath;
6695 11728 : if (subpath == rel->cheapest_startup_path)
6696 384 : rel->cheapest_startup_path = newpath;
6697 11728 : if (subpath == rel->cheapest_total_path)
6698 384 : rel->cheapest_total_path = newpath;
6699 : }
6700 :
6701 : /* Likewise for partial paths, if any */
6702 11710 : foreach(lc, rel->partial_pathlist)
6703 : {
6704 6 : Path *subpath = (Path *) lfirst(lc);
6705 6 : Path *newpath = subpath;
6706 : ListCell *lc1,
6707 : *lc2;
6708 :
6709 : Assert(subpath->param_info == NULL);
6710 24 : forboth(lc1, targets, lc2, targets_contain_srfs)
6711 : {
6712 18 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6713 18 : bool contains_srfs = (bool) lfirst_int(lc2);
6714 :
6715 : /* If this level doesn't contain SRFs, do regular projection */
6716 18 : if (contains_srfs)
6717 6 : newpath = (Path *) create_set_projection_path(root,
6718 : rel,
6719 : newpath,
6720 : thistarget);
6721 : else
6722 : {
6723 : /* avoid apply_projection_to_path, in case of multiple refs */
6724 12 : newpath = (Path *) create_projection_path(root,
6725 : rel,
6726 : newpath,
6727 : thistarget);
6728 : }
6729 : }
6730 6 : lfirst(lc) = newpath;
6731 : }
6732 : }
6733 :
6734 : /*
6735 : * expression_planner
6736 : * Perform planner's transformations on a standalone expression.
6737 : *
6738 : * Various utility commands need to evaluate expressions that are not part
6739 : * of a plannable query. They can do so using the executor's regular
6740 : * expression-execution machinery, but first the expression has to be fed
6741 : * through here to transform it from parser output to something executable.
6742 : *
6743 : * Currently, we disallow sublinks in standalone expressions, so there's no
6744 : * real "planning" involved here. (That might not always be true though.)
6745 : * What we must do is run eval_const_expressions to ensure that any function
6746 : * calls are converted to positional notation and function default arguments
6747 : * get inserted. The fact that constant subexpressions get simplified is a
6748 : * side-effect that is useful when the expression will get evaluated more than
6749 : * once. Also, we must fix operator function IDs.
6750 : *
6751 : * This does not return any information about dependencies of the expression.
6752 : * Hence callers should use the results only for the duration of the current
6753 : * query. Callers that would like to cache the results for longer should use
6754 : * expression_planner_with_deps, probably via the plancache.
6755 : *
6756 : * Note: this must not make any damaging changes to the passed-in expression
6757 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6758 : * we first do an expression_tree_mutator-based walk, what is returned will
6759 : * be a new node tree.) The result is constructed in the current memory
6760 : * context; beware that this can leak a lot of additional stuff there, too.
6761 : */
6762 : Expr *
6763 240670 : expression_planner(Expr *expr)
6764 : {
6765 : Node *result;
6766 :
6767 : /*
6768 : * Convert named-argument function calls, insert default arguments and
6769 : * simplify constant subexprs
6770 : */
6771 240670 : result = eval_const_expressions(NULL, (Node *) expr);
6772 :
6773 : /* Fill in opfuncid values if missing */
6774 240652 : fix_opfuncids(result);
6775 :
6776 240652 : return (Expr *) result;
6777 : }
6778 :
6779 : /*
6780 : * expression_planner_with_deps
6781 : * Perform planner's transformations on a standalone expression,
6782 : * returning expression dependency information along with the result.
6783 : *
6784 : * This is identical to expression_planner() except that it also returns
6785 : * information about possible dependencies of the expression, ie identities of
6786 : * objects whose definitions affect the result. As in a PlannedStmt, these
6787 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6788 : */
6789 : Expr *
6790 346 : expression_planner_with_deps(Expr *expr,
6791 : List **relationOids,
6792 : List **invalItems)
6793 : {
6794 : Node *result;
6795 : PlannerGlobal glob;
6796 : PlannerInfo root;
6797 :
6798 : /* Make up dummy planner state so we can use setrefs machinery */
6799 8996 : MemSet(&glob, 0, sizeof(glob));
6800 346 : glob.type = T_PlannerGlobal;
6801 346 : glob.relationOids = NIL;
6802 346 : glob.invalItems = NIL;
6803 :
6804 32178 : MemSet(&root, 0, sizeof(root));
6805 346 : root.type = T_PlannerInfo;
6806 346 : root.glob = &glob;
6807 :
6808 : /*
6809 : * Convert named-argument function calls, insert default arguments and
6810 : * simplify constant subexprs. Collect identities of inlined functions
6811 : * and elided domains, too.
6812 : */
6813 346 : result = eval_const_expressions(&root, (Node *) expr);
6814 :
6815 : /* Fill in opfuncid values if missing */
6816 346 : fix_opfuncids(result);
6817 :
6818 : /*
6819 : * Now walk the finished expression to find anything else we ought to
6820 : * record as an expression dependency.
6821 : */
6822 346 : (void) extract_query_dependencies_walker(result, &root);
6823 :
6824 346 : *relationOids = glob.relationOids;
6825 346 : *invalItems = glob.invalItems;
6826 :
6827 346 : return (Expr *) result;
6828 : }
6829 :
6830 :
6831 : /*
6832 : * plan_cluster_use_sort
6833 : * Use the planner to decide how CLUSTER should implement sorting
6834 : *
6835 : * tableOid is the OID of a table to be clustered on its index indexOid
6836 : * (which is already known to be a btree index). Decide whether it's
6837 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6838 : * Return true to use sorting, false to use an indexscan.
6839 : *
6840 : * Note: caller had better already hold some type of lock on the table.
6841 : */
6842 : bool
6843 188 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6844 : {
6845 : PlannerInfo *root;
6846 : Query *query;
6847 : PlannerGlobal *glob;
6848 : RangeTblEntry *rte;
6849 : RelOptInfo *rel;
6850 : IndexOptInfo *indexInfo;
6851 : QualCost indexExprCost;
6852 : Cost comparisonCost;
6853 : Path *seqScanPath;
6854 : Path seqScanAndSortPath;
6855 : IndexPath *indexScanPath;
6856 : ListCell *lc;
6857 :
6858 : /* We can short-circuit the cost comparison if indexscans are disabled */
6859 188 : if (!enable_indexscan)
6860 30 : return true; /* use sort */
6861 :
6862 : /* Set up mostly-dummy planner state */
6863 158 : query = makeNode(Query);
6864 158 : query->commandType = CMD_SELECT;
6865 :
6866 158 : glob = makeNode(PlannerGlobal);
6867 :
6868 158 : root = makeNode(PlannerInfo);
6869 158 : root->parse = query;
6870 158 : root->glob = glob;
6871 158 : root->query_level = 1;
6872 158 : root->planner_cxt = CurrentMemoryContext;
6873 158 : root->wt_param_id = -1;
6874 158 : root->join_domains = list_make1(makeNode(JoinDomain));
6875 :
6876 : /* Build a minimal RTE for the rel */
6877 158 : rte = makeNode(RangeTblEntry);
6878 158 : rte->rtekind = RTE_RELATION;
6879 158 : rte->relid = tableOid;
6880 158 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6881 158 : rte->rellockmode = AccessShareLock;
6882 158 : rte->lateral = false;
6883 158 : rte->inh = false;
6884 158 : rte->inFromCl = true;
6885 158 : query->rtable = list_make1(rte);
6886 158 : addRTEPermissionInfo(&query->rteperminfos, rte);
6887 :
6888 : /* Set up RTE/RelOptInfo arrays */
6889 158 : setup_simple_rel_arrays(root);
6890 :
6891 : /* Build RelOptInfo */
6892 158 : rel = build_simple_rel(root, 1, NULL);
6893 :
6894 : /* Locate IndexOptInfo for the target index */
6895 158 : indexInfo = NULL;
6896 196 : foreach(lc, rel->indexlist)
6897 : {
6898 196 : indexInfo = lfirst_node(IndexOptInfo, lc);
6899 196 : if (indexInfo->indexoid == indexOid)
6900 158 : break;
6901 : }
6902 :
6903 : /*
6904 : * It's possible that get_relation_info did not generate an IndexOptInfo
6905 : * for the desired index; this could happen if it's not yet reached its
6906 : * indcheckxmin usability horizon, or if it's a system index and we're
6907 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6908 : * trust the index contents but use seqscan-and-sort.
6909 : */
6910 158 : if (lc == NULL) /* not in the list? */
6911 0 : return true; /* use sort */
6912 :
6913 : /*
6914 : * Rather than doing all the pushups that would be needed to use
6915 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6916 : */
6917 158 : rel->rows = rel->tuples;
6918 158 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6919 :
6920 158 : root->total_table_pages = rel->pages;
6921 :
6922 : /*
6923 : * Determine eval cost of the index expressions, if any. We need to
6924 : * charge twice that amount for each tuple comparison that happens during
6925 : * the sort, since tuplesort.c will have to re-evaluate the index
6926 : * expressions each time. (XXX that's pretty inefficient...)
6927 : */
6928 158 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6929 158 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6930 :
6931 : /* Estimate the cost of seq scan + sort */
6932 158 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6933 158 : cost_sort(&seqScanAndSortPath, root, NIL,
6934 : seqScanPath->disabled_nodes,
6935 158 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6936 : comparisonCost, maintenance_work_mem, -1.0);
6937 :
6938 : /* Estimate the cost of index scan */
6939 158 : indexScanPath = create_index_path(root, indexInfo,
6940 : NIL, NIL, NIL, NIL,
6941 : ForwardScanDirection, false,
6942 : NULL, 1.0, false);
6943 :
6944 158 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6945 : }
6946 :
6947 : /*
6948 : * plan_create_index_workers
6949 : * Use the planner to decide how many parallel worker processes
6950 : * CREATE INDEX should request for use
6951 : *
6952 : * tableOid is the table on which the index is to be built. indexOid is the
6953 : * OID of an index to be created or reindexed (which must be an index with
6954 : * support for parallel builds - currently btree, GIN, or BRIN).
6955 : *
6956 : * Return value is the number of parallel worker processes to request. It
6957 : * may be unsafe to proceed if this is 0. Note that this does not include the
6958 : * leader participating as a worker (value is always a number of parallel
6959 : * worker processes).
6960 : *
6961 : * Note: caller had better already hold some type of lock on the table and
6962 : * index.
6963 : */
6964 : int
6965 35256 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6966 : {
6967 : PlannerInfo *root;
6968 : Query *query;
6969 : PlannerGlobal *glob;
6970 : RangeTblEntry *rte;
6971 : Relation heap;
6972 : Relation index;
6973 : RelOptInfo *rel;
6974 : int parallel_workers;
6975 : BlockNumber heap_blocks;
6976 : double reltuples;
6977 : double allvisfrac;
6978 :
6979 : /*
6980 : * We don't allow performing parallel operation in standalone backend or
6981 : * when parallelism is disabled.
6982 : */
6983 35256 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
6984 506 : return 0;
6985 :
6986 : /* Set up largely-dummy planner state */
6987 34750 : query = makeNode(Query);
6988 34750 : query->commandType = CMD_SELECT;
6989 :
6990 34750 : glob = makeNode(PlannerGlobal);
6991 :
6992 34750 : root = makeNode(PlannerInfo);
6993 34750 : root->parse = query;
6994 34750 : root->glob = glob;
6995 34750 : root->query_level = 1;
6996 34750 : root->planner_cxt = CurrentMemoryContext;
6997 34750 : root->wt_param_id = -1;
6998 34750 : root->join_domains = list_make1(makeNode(JoinDomain));
6999 :
7000 : /*
7001 : * Build a minimal RTE.
7002 : *
7003 : * Mark the RTE with inh = true. This is a kludge to prevent
7004 : * get_relation_info() from fetching index info, which is necessary
7005 : * because it does not expect that any IndexOptInfo is currently
7006 : * undergoing REINDEX.
7007 : */
7008 34750 : rte = makeNode(RangeTblEntry);
7009 34750 : rte->rtekind = RTE_RELATION;
7010 34750 : rte->relid = tableOid;
7011 34750 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7012 34750 : rte->rellockmode = AccessShareLock;
7013 34750 : rte->lateral = false;
7014 34750 : rte->inh = true;
7015 34750 : rte->inFromCl = true;
7016 34750 : query->rtable = list_make1(rte);
7017 34750 : addRTEPermissionInfo(&query->rteperminfos, rte);
7018 :
7019 : /* Set up RTE/RelOptInfo arrays */
7020 34750 : setup_simple_rel_arrays(root);
7021 :
7022 : /* Build RelOptInfo */
7023 34750 : rel = build_simple_rel(root, 1, NULL);
7024 :
7025 : /* Rels are assumed already locked by the caller */
7026 34750 : heap = table_open(tableOid, NoLock);
7027 34750 : index = index_open(indexOid, NoLock);
7028 :
7029 : /*
7030 : * Determine if it's safe to proceed.
7031 : *
7032 : * Currently, parallel workers can't access the leader's temporary tables.
7033 : * Furthermore, any index predicate or index expressions must be parallel
7034 : * safe.
7035 : */
7036 34750 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7037 32678 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7038 32558 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7039 : {
7040 2192 : parallel_workers = 0;
7041 2192 : goto done;
7042 : }
7043 :
7044 : /*
7045 : * If parallel_workers storage parameter is set for the table, accept that
7046 : * as the number of parallel worker processes to launch (though still cap
7047 : * at max_parallel_maintenance_workers). Note that we deliberately do not
7048 : * consider any other factor when parallel_workers is set. (e.g., memory
7049 : * use by workers.)
7050 : */
7051 32558 : if (rel->rel_parallel_workers != -1)
7052 : {
7053 20 : parallel_workers = Min(rel->rel_parallel_workers,
7054 : max_parallel_maintenance_workers);
7055 20 : goto done;
7056 : }
7057 :
7058 : /*
7059 : * Estimate heap relation size ourselves, since rel->pages cannot be
7060 : * trusted (heap RTE was marked as inheritance parent)
7061 : */
7062 32538 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7063 :
7064 : /*
7065 : * Determine number of workers to scan the heap relation using generic
7066 : * model
7067 : */
7068 32538 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7069 : max_parallel_maintenance_workers);
7070 :
7071 : /*
7072 : * Cap workers based on available maintenance_work_mem as needed.
7073 : *
7074 : * Note that each tuplesort participant receives an even share of the
7075 : * total maintenance_work_mem budget. Aim to leave participants
7076 : * (including the leader as a participant) with no less than 32MB of
7077 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7078 : * immediately past the threshold of being capable of launching a single
7079 : * parallel worker to sort.
7080 : */
7081 32694 : while (parallel_workers > 0 &&
7082 314 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7083 156 : parallel_workers--;
7084 :
7085 32538 : done:
7086 34750 : index_close(index, NoLock);
7087 34750 : table_close(heap, NoLock);
7088 :
7089 34750 : return parallel_workers;
7090 : }
7091 :
7092 : /*
7093 : * add_paths_to_grouping_rel
7094 : *
7095 : * Add non-partial paths to grouping relation.
7096 : */
7097 : static void
7098 40526 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7099 : RelOptInfo *grouped_rel,
7100 : RelOptInfo *partially_grouped_rel,
7101 : const AggClauseCosts *agg_costs,
7102 : grouping_sets_data *gd,
7103 : GroupPathExtraData *extra)
7104 : {
7105 40526 : Query *parse = root->parse;
7106 40526 : Path *cheapest_path = input_rel->cheapest_total_path;
7107 40526 : Path *cheapest_partially_grouped_path = NULL;
7108 : ListCell *lc;
7109 40526 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7110 40526 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7111 40526 : List *havingQual = (List *) extra->havingQual;
7112 40526 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7113 40526 : double dNumGroups = 0;
7114 40526 : double dNumFinalGroups = 0;
7115 :
7116 : /*
7117 : * Estimate number of groups for non-split aggregation.
7118 : */
7119 40526 : dNumGroups = get_number_of_groups(root,
7120 : cheapest_path->rows,
7121 : gd,
7122 : extra->targetList);
7123 :
7124 40526 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7125 : {
7126 2236 : cheapest_partially_grouped_path =
7127 : partially_grouped_rel->cheapest_total_path;
7128 :
7129 : /*
7130 : * Estimate number of groups for final phase of partial aggregation.
7131 : */
7132 : dNumFinalGroups =
7133 2236 : get_number_of_groups(root,
7134 : cheapest_partially_grouped_path->rows,
7135 : gd,
7136 : extra->targetList);
7137 : }
7138 :
7139 40526 : if (can_sort)
7140 : {
7141 : /*
7142 : * Use any available suitably-sorted path as input, and also consider
7143 : * sorting the cheapest-total path and incremental sort on any paths
7144 : * with presorted keys.
7145 : */
7146 84300 : foreach(lc, input_rel->pathlist)
7147 : {
7148 : ListCell *lc2;
7149 43780 : Path *path = (Path *) lfirst(lc);
7150 43780 : Path *path_save = path;
7151 43780 : List *pathkey_orderings = NIL;
7152 :
7153 : /* generate alternative group orderings that might be useful */
7154 43780 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7155 :
7156 : Assert(list_length(pathkey_orderings) > 0);
7157 :
7158 87704 : foreach(lc2, pathkey_orderings)
7159 : {
7160 43924 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7161 :
7162 : /* restore the path (we replace it in the loop) */
7163 43924 : path = path_save;
7164 :
7165 43924 : path = make_ordered_path(root,
7166 : grouped_rel,
7167 : path,
7168 : cheapest_path,
7169 : info->pathkeys,
7170 : -1.0);
7171 43924 : if (path == NULL)
7172 404 : continue;
7173 :
7174 : /* Now decide what to stick atop it */
7175 43520 : if (parse->groupingSets)
7176 : {
7177 992 : consider_groupingsets_paths(root, grouped_rel,
7178 : path, true, can_hash,
7179 : gd, agg_costs, dNumGroups);
7180 : }
7181 42528 : else if (parse->hasAggs)
7182 : {
7183 : /*
7184 : * We have aggregation, possibly with plain GROUP BY. Make
7185 : * an AggPath.
7186 : */
7187 41744 : add_path(grouped_rel, (Path *)
7188 41744 : create_agg_path(root,
7189 : grouped_rel,
7190 : path,
7191 41744 : grouped_rel->reltarget,
7192 41744 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7193 : AGGSPLIT_SIMPLE,
7194 : info->clauses,
7195 : havingQual,
7196 : agg_costs,
7197 : dNumGroups));
7198 : }
7199 784 : else if (parse->groupClause)
7200 : {
7201 : /*
7202 : * We have GROUP BY without aggregation or grouping sets.
7203 : * Make a GroupPath.
7204 : */
7205 784 : add_path(grouped_rel, (Path *)
7206 784 : create_group_path(root,
7207 : grouped_rel,
7208 : path,
7209 : info->clauses,
7210 : havingQual,
7211 : dNumGroups));
7212 : }
7213 : else
7214 : {
7215 : /* Other cases should have been handled above */
7216 : Assert(false);
7217 : }
7218 : }
7219 : }
7220 :
7221 : /*
7222 : * Instead of operating directly on the input relation, we can
7223 : * consider finalizing a partially aggregated path.
7224 : */
7225 40520 : if (partially_grouped_rel != NULL)
7226 : {
7227 6154 : foreach(lc, partially_grouped_rel->pathlist)
7228 : {
7229 : ListCell *lc2;
7230 3918 : Path *path = (Path *) lfirst(lc);
7231 3918 : Path *path_save = path;
7232 3918 : List *pathkey_orderings = NIL;
7233 :
7234 : /* generate alternative group orderings that might be useful */
7235 3918 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7236 :
7237 : Assert(list_length(pathkey_orderings) > 0);
7238 :
7239 : /* process all potentially interesting grouping reorderings */
7240 7836 : foreach(lc2, pathkey_orderings)
7241 : {
7242 3918 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7243 :
7244 : /* restore the path (we replace it in the loop) */
7245 3918 : path = path_save;
7246 :
7247 3918 : path = make_ordered_path(root,
7248 : grouped_rel,
7249 : path,
7250 : cheapest_partially_grouped_path,
7251 : info->pathkeys,
7252 : -1.0);
7253 :
7254 3918 : if (path == NULL)
7255 204 : continue;
7256 :
7257 3714 : if (parse->hasAggs)
7258 3466 : add_path(grouped_rel, (Path *)
7259 3466 : create_agg_path(root,
7260 : grouped_rel,
7261 : path,
7262 3466 : grouped_rel->reltarget,
7263 3466 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7264 : AGGSPLIT_FINAL_DESERIAL,
7265 : info->clauses,
7266 : havingQual,
7267 : agg_final_costs,
7268 : dNumFinalGroups));
7269 : else
7270 248 : add_path(grouped_rel, (Path *)
7271 248 : create_group_path(root,
7272 : grouped_rel,
7273 : path,
7274 : info->clauses,
7275 : havingQual,
7276 : dNumFinalGroups));
7277 :
7278 : }
7279 : }
7280 : }
7281 : }
7282 :
7283 40526 : if (can_hash)
7284 : {
7285 5692 : if (parse->groupingSets)
7286 : {
7287 : /*
7288 : * Try for a hash-only groupingsets path over unsorted input.
7289 : */
7290 830 : consider_groupingsets_paths(root, grouped_rel,
7291 : cheapest_path, false, true,
7292 : gd, agg_costs, dNumGroups);
7293 : }
7294 : else
7295 : {
7296 : /*
7297 : * Generate a HashAgg Path. We just need an Agg over the
7298 : * cheapest-total input path, since input order won't matter.
7299 : */
7300 4862 : add_path(grouped_rel, (Path *)
7301 4862 : create_agg_path(root, grouped_rel,
7302 : cheapest_path,
7303 4862 : grouped_rel->reltarget,
7304 : AGG_HASHED,
7305 : AGGSPLIT_SIMPLE,
7306 : root->processed_groupClause,
7307 : havingQual,
7308 : agg_costs,
7309 : dNumGroups));
7310 : }
7311 :
7312 : /*
7313 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7314 : * grouped path, assuming there is one
7315 : */
7316 5692 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7317 : {
7318 1442 : add_path(grouped_rel, (Path *)
7319 1442 : create_agg_path(root,
7320 : grouped_rel,
7321 : cheapest_partially_grouped_path,
7322 1442 : grouped_rel->reltarget,
7323 : AGG_HASHED,
7324 : AGGSPLIT_FINAL_DESERIAL,
7325 : root->processed_groupClause,
7326 : havingQual,
7327 : agg_final_costs,
7328 : dNumFinalGroups));
7329 : }
7330 : }
7331 :
7332 : /*
7333 : * When partitionwise aggregate is used, we might have fully aggregated
7334 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7335 : * consider a path for grouped_rel consisting of a Parallel Append of
7336 : * non-partial paths from each child.
7337 : */
7338 40526 : if (grouped_rel->partial_pathlist != NIL)
7339 318 : gather_grouping_paths(root, grouped_rel);
7340 40526 : }
7341 :
7342 : /*
7343 : * create_partial_grouping_paths
7344 : *
7345 : * Create a new upper relation representing the result of partial aggregation
7346 : * and populate it with appropriate paths. Note that we don't finalize the
7347 : * lists of paths here, so the caller can add additional partial or non-partial
7348 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7349 : * the returned upper relation.
7350 : *
7351 : * All paths for this new upper relation -- both partial and non-partial --
7352 : * have been partially aggregated but require a subsequent FinalizeAggregate
7353 : * step.
7354 : *
7355 : * NB: This function is allowed to return NULL if it determines that there is
7356 : * no real need to create a new RelOptInfo.
7357 : */
7358 : static RelOptInfo *
7359 36514 : create_partial_grouping_paths(PlannerInfo *root,
7360 : RelOptInfo *grouped_rel,
7361 : RelOptInfo *input_rel,
7362 : grouping_sets_data *gd,
7363 : GroupPathExtraData *extra,
7364 : bool force_rel_creation)
7365 : {
7366 36514 : Query *parse = root->parse;
7367 : RelOptInfo *partially_grouped_rel;
7368 36514 : RelOptInfo *eager_agg_rel = NULL;
7369 36514 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7370 36514 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7371 36514 : Path *cheapest_partial_path = NULL;
7372 36514 : Path *cheapest_total_path = NULL;
7373 36514 : double dNumPartialGroups = 0;
7374 36514 : double dNumPartialPartialGroups = 0;
7375 : ListCell *lc;
7376 36514 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7377 36514 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7378 :
7379 : /*
7380 : * Check whether any partially aggregated paths have been generated
7381 : * through eager aggregation.
7382 : */
7383 36514 : if (input_rel->grouped_rel &&
7384 958 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7385 958 : input_rel->grouped_rel->pathlist != NIL)
7386 898 : eager_agg_rel = input_rel->grouped_rel;
7387 :
7388 : /*
7389 : * Consider whether we should generate partially aggregated non-partial
7390 : * paths. We can only do this if we have a non-partial path, and only if
7391 : * the parent of the input rel is performing partial partitionwise
7392 : * aggregation. (Note that extra->patype is the type of partitionwise
7393 : * aggregation being used at the parent level, not this level.)
7394 : */
7395 36514 : if (input_rel->pathlist != NIL &&
7396 36514 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7397 858 : cheapest_total_path = input_rel->cheapest_total_path;
7398 :
7399 : /*
7400 : * If parallelism is possible for grouped_rel, then we should consider
7401 : * generating partially-grouped partial paths. However, if the input rel
7402 : * has no partial paths, then we can't.
7403 : */
7404 36514 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7405 2512 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7406 :
7407 : /*
7408 : * If we can't partially aggregate partial paths, and we can't partially
7409 : * aggregate non-partial paths, and no partially aggregated paths were
7410 : * generated by eager aggregation, then don't bother creating the new
7411 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7412 : */
7413 36514 : if (cheapest_total_path == NULL &&
7414 33648 : cheapest_partial_path == NULL &&
7415 33518 : eager_agg_rel == NULL &&
7416 33518 : !force_rel_creation)
7417 33420 : return NULL;
7418 :
7419 : /*
7420 : * Build a new upper relation to represent the result of partially
7421 : * aggregating the rows from the input relation.
7422 : */
7423 3094 : partially_grouped_rel = fetch_upper_rel(root,
7424 : UPPERREL_PARTIAL_GROUP_AGG,
7425 : grouped_rel->relids);
7426 3094 : partially_grouped_rel->consider_parallel =
7427 3094 : grouped_rel->consider_parallel;
7428 3094 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7429 3094 : partially_grouped_rel->serverid = grouped_rel->serverid;
7430 3094 : partially_grouped_rel->userid = grouped_rel->userid;
7431 3094 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7432 3094 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7433 :
7434 : /*
7435 : * Build target list for partial aggregate paths. These paths cannot just
7436 : * emit the same tlist as regular aggregate paths, because (1) we must
7437 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7438 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7439 : */
7440 3094 : partially_grouped_rel->reltarget =
7441 3094 : make_partial_grouping_target(root, grouped_rel->reltarget,
7442 : extra->havingQual);
7443 :
7444 3094 : if (!extra->partial_costs_set)
7445 : {
7446 : /*
7447 : * Collect statistics about aggregates for estimating costs of
7448 : * performing aggregation in parallel.
7449 : */
7450 9312 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7451 9312 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7452 1552 : if (parse->hasAggs)
7453 : {
7454 : /* partial phase */
7455 1418 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7456 : agg_partial_costs);
7457 :
7458 : /* final phase */
7459 1418 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7460 : agg_final_costs);
7461 : }
7462 :
7463 1552 : extra->partial_costs_set = true;
7464 : }
7465 :
7466 : /* Estimate number of partial groups. */
7467 3094 : if (cheapest_total_path != NULL)
7468 : dNumPartialGroups =
7469 858 : get_number_of_groups(root,
7470 : cheapest_total_path->rows,
7471 : gd,
7472 : extra->targetList);
7473 3094 : if (cheapest_partial_path != NULL)
7474 : dNumPartialPartialGroups =
7475 2512 : get_number_of_groups(root,
7476 : cheapest_partial_path->rows,
7477 : gd,
7478 : extra->targetList);
7479 :
7480 3094 : if (can_sort && cheapest_total_path != NULL)
7481 : {
7482 : /* This should have been checked previously */
7483 : Assert(parse->hasAggs || parse->groupClause);
7484 :
7485 : /*
7486 : * Use any available suitably-sorted path as input, and also consider
7487 : * sorting the cheapest partial path.
7488 : */
7489 1716 : foreach(lc, input_rel->pathlist)
7490 : {
7491 : ListCell *lc2;
7492 858 : Path *path = (Path *) lfirst(lc);
7493 858 : Path *path_save = path;
7494 858 : List *pathkey_orderings = NIL;
7495 :
7496 : /* generate alternative group orderings that might be useful */
7497 858 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7498 :
7499 : Assert(list_length(pathkey_orderings) > 0);
7500 :
7501 : /* process all potentially interesting grouping reorderings */
7502 1716 : foreach(lc2, pathkey_orderings)
7503 : {
7504 858 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7505 :
7506 : /* restore the path (we replace it in the loop) */
7507 858 : path = path_save;
7508 :
7509 858 : path = make_ordered_path(root,
7510 : partially_grouped_rel,
7511 : path,
7512 : cheapest_total_path,
7513 : info->pathkeys,
7514 : -1.0);
7515 :
7516 858 : if (path == NULL)
7517 0 : continue;
7518 :
7519 858 : if (parse->hasAggs)
7520 786 : add_path(partially_grouped_rel, (Path *)
7521 786 : create_agg_path(root,
7522 : partially_grouped_rel,
7523 : path,
7524 786 : partially_grouped_rel->reltarget,
7525 786 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7526 : AGGSPLIT_INITIAL_SERIAL,
7527 : info->clauses,
7528 : NIL,
7529 : agg_partial_costs,
7530 : dNumPartialGroups));
7531 : else
7532 72 : add_path(partially_grouped_rel, (Path *)
7533 72 : create_group_path(root,
7534 : partially_grouped_rel,
7535 : path,
7536 : info->clauses,
7537 : NIL,
7538 : dNumPartialGroups));
7539 : }
7540 : }
7541 : }
7542 :
7543 3094 : if (can_sort && cheapest_partial_path != NULL)
7544 : {
7545 : /* Similar to above logic, but for partial paths. */
7546 5360 : foreach(lc, input_rel->partial_pathlist)
7547 : {
7548 : ListCell *lc2;
7549 2848 : Path *path = (Path *) lfirst(lc);
7550 2848 : Path *path_save = path;
7551 2848 : List *pathkey_orderings = NIL;
7552 :
7553 : /* generate alternative group orderings that might be useful */
7554 2848 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7555 :
7556 : Assert(list_length(pathkey_orderings) > 0);
7557 :
7558 : /* process all potentially interesting grouping reorderings */
7559 5696 : foreach(lc2, pathkey_orderings)
7560 : {
7561 2848 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7562 :
7563 :
7564 : /* restore the path (we replace it in the loop) */
7565 2848 : path = path_save;
7566 :
7567 2848 : path = make_ordered_path(root,
7568 : partially_grouped_rel,
7569 : path,
7570 : cheapest_partial_path,
7571 : info->pathkeys,
7572 : -1.0);
7573 :
7574 2848 : if (path == NULL)
7575 6 : continue;
7576 :
7577 2842 : if (parse->hasAggs)
7578 2720 : add_partial_path(partially_grouped_rel, (Path *)
7579 2720 : create_agg_path(root,
7580 : partially_grouped_rel,
7581 : path,
7582 2720 : partially_grouped_rel->reltarget,
7583 2720 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7584 : AGGSPLIT_INITIAL_SERIAL,
7585 : info->clauses,
7586 : NIL,
7587 : agg_partial_costs,
7588 : dNumPartialPartialGroups));
7589 : else
7590 122 : add_partial_path(partially_grouped_rel, (Path *)
7591 122 : create_group_path(root,
7592 : partially_grouped_rel,
7593 : path,
7594 : info->clauses,
7595 : NIL,
7596 : dNumPartialPartialGroups));
7597 : }
7598 : }
7599 : }
7600 :
7601 : /*
7602 : * Add a partially-grouped HashAgg Path where possible
7603 : */
7604 3094 : if (can_hash && cheapest_total_path != NULL)
7605 : {
7606 : /* Checked above */
7607 : Assert(parse->hasAggs || parse->groupClause);
7608 :
7609 858 : add_path(partially_grouped_rel, (Path *)
7610 858 : create_agg_path(root,
7611 : partially_grouped_rel,
7612 : cheapest_total_path,
7613 858 : partially_grouped_rel->reltarget,
7614 : AGG_HASHED,
7615 : AGGSPLIT_INITIAL_SERIAL,
7616 : root->processed_groupClause,
7617 : NIL,
7618 : agg_partial_costs,
7619 : dNumPartialGroups));
7620 : }
7621 :
7622 : /*
7623 : * Now add a partially-grouped HashAgg partial Path where possible
7624 : */
7625 3094 : if (can_hash && cheapest_partial_path != NULL)
7626 : {
7627 1718 : add_partial_path(partially_grouped_rel, (Path *)
7628 1718 : create_agg_path(root,
7629 : partially_grouped_rel,
7630 : cheapest_partial_path,
7631 1718 : partially_grouped_rel->reltarget,
7632 : AGG_HASHED,
7633 : AGGSPLIT_INITIAL_SERIAL,
7634 : root->processed_groupClause,
7635 : NIL,
7636 : agg_partial_costs,
7637 : dNumPartialPartialGroups));
7638 : }
7639 :
7640 : /*
7641 : * Add any partially aggregated paths generated by eager aggregation to
7642 : * the new upper relation after applying projection steps as needed.
7643 : */
7644 3094 : if (eager_agg_rel)
7645 : {
7646 : /* Add the paths */
7647 2348 : foreach(lc, eager_agg_rel->pathlist)
7648 : {
7649 1450 : Path *path = (Path *) lfirst(lc);
7650 :
7651 : /* Shouldn't have any parameterized paths anymore */
7652 : Assert(path->param_info == NULL);
7653 :
7654 1450 : path = (Path *) create_projection_path(root,
7655 : partially_grouped_rel,
7656 : path,
7657 1450 : partially_grouped_rel->reltarget);
7658 :
7659 1450 : add_path(partially_grouped_rel, path);
7660 : }
7661 :
7662 : /*
7663 : * Likewise add the partial paths, but only if parallelism is possible
7664 : * for partially_grouped_rel.
7665 : */
7666 898 : if (partially_grouped_rel->consider_parallel)
7667 : {
7668 2028 : foreach(lc, eager_agg_rel->partial_pathlist)
7669 : {
7670 1212 : Path *path = (Path *) lfirst(lc);
7671 :
7672 : /* Shouldn't have any parameterized paths anymore */
7673 : Assert(path->param_info == NULL);
7674 :
7675 1212 : path = (Path *) create_projection_path(root,
7676 : partially_grouped_rel,
7677 : path,
7678 1212 : partially_grouped_rel->reltarget);
7679 :
7680 1212 : add_partial_path(partially_grouped_rel, path);
7681 : }
7682 : }
7683 : }
7684 :
7685 : /*
7686 : * If there is an FDW that's responsible for all baserels of the query,
7687 : * let it consider adding partially grouped ForeignPaths.
7688 : */
7689 3094 : if (partially_grouped_rel->fdwroutine &&
7690 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7691 : {
7692 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7693 :
7694 6 : fdwroutine->GetForeignUpperPaths(root,
7695 : UPPERREL_PARTIAL_GROUP_AGG,
7696 : input_rel, partially_grouped_rel,
7697 : extra);
7698 : }
7699 :
7700 3094 : return partially_grouped_rel;
7701 : }
7702 :
7703 : /*
7704 : * make_ordered_path
7705 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7706 : * return NULL if it doesn't make sense to generate an ordered path in
7707 : * this case.
7708 : */
7709 : static Path *
7710 57688 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7711 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7712 : {
7713 : bool is_sorted;
7714 : int presorted_keys;
7715 :
7716 57688 : is_sorted = pathkeys_count_contained_in(pathkeys,
7717 : path->pathkeys,
7718 : &presorted_keys);
7719 :
7720 57688 : if (!is_sorted)
7721 : {
7722 : /*
7723 : * Try at least sorting the cheapest path and also try incrementally
7724 : * sorting any path which is partially sorted already (no need to deal
7725 : * with paths which have presorted keys when incremental sort is
7726 : * disabled unless it's the cheapest input path).
7727 : */
7728 16598 : if (path != cheapest_path &&
7729 3132 : (presorted_keys == 0 || !enable_incremental_sort))
7730 1488 : return NULL;
7731 :
7732 : /*
7733 : * We've no need to consider both a sort and incremental sort. We'll
7734 : * just do a sort if there are no presorted keys and an incremental
7735 : * sort when there are presorted keys.
7736 : */
7737 15110 : if (presorted_keys == 0 || !enable_incremental_sort)
7738 13292 : path = (Path *) create_sort_path(root,
7739 : rel,
7740 : path,
7741 : pathkeys,
7742 : limit_tuples);
7743 : else
7744 1818 : path = (Path *) create_incremental_sort_path(root,
7745 : rel,
7746 : path,
7747 : pathkeys,
7748 : presorted_keys,
7749 : limit_tuples);
7750 : }
7751 :
7752 56200 : return path;
7753 : }
7754 :
7755 : /*
7756 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7757 : * grouping relation.
7758 : *
7759 : * generate_useful_gather_paths does most of the work, but we also consider a
7760 : * special case: we could try sorting the data by the group_pathkeys and then
7761 : * applying Gather Merge.
7762 : *
7763 : * NB: This function shouldn't be used for anything other than a grouped or
7764 : * partially grouped relation not only because of the fact that it explicitly
7765 : * references group_pathkeys but we pass "true" as the third argument to
7766 : * generate_useful_gather_paths().
7767 : */
7768 : static void
7769 2326 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7770 : {
7771 : ListCell *lc;
7772 : Path *cheapest_partial_path;
7773 : List *groupby_pathkeys;
7774 :
7775 : /*
7776 : * This occurs after any partial aggregation has taken place, so trim off
7777 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7778 : */
7779 2326 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7780 18 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7781 : root->num_groupby_pathkeys);
7782 : else
7783 2308 : groupby_pathkeys = root->group_pathkeys;
7784 :
7785 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7786 2326 : generate_useful_gather_paths(root, rel, true);
7787 :
7788 2326 : cheapest_partial_path = linitial(rel->partial_pathlist);
7789 :
7790 : /* XXX Shouldn't this also consider the group-key-reordering? */
7791 5786 : foreach(lc, rel->partial_pathlist)
7792 : {
7793 3460 : Path *path = (Path *) lfirst(lc);
7794 : bool is_sorted;
7795 : int presorted_keys;
7796 : double total_groups;
7797 :
7798 3460 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7799 : path->pathkeys,
7800 : &presorted_keys);
7801 :
7802 3460 : if (is_sorted)
7803 1996 : continue;
7804 :
7805 : /*
7806 : * Try at least sorting the cheapest path and also try incrementally
7807 : * sorting any path which is partially sorted already (no need to deal
7808 : * with paths which have presorted keys when incremental sort is
7809 : * disabled unless it's the cheapest input path).
7810 : */
7811 1464 : if (path != cheapest_partial_path &&
7812 0 : (presorted_keys == 0 || !enable_incremental_sort))
7813 0 : continue;
7814 :
7815 : /*
7816 : * We've no need to consider both a sort and incremental sort. We'll
7817 : * just do a sort if there are no presorted keys and an incremental
7818 : * sort when there are presorted keys.
7819 : */
7820 1464 : if (presorted_keys == 0 || !enable_incremental_sort)
7821 1464 : path = (Path *) create_sort_path(root, rel, path,
7822 : groupby_pathkeys,
7823 : -1.0);
7824 : else
7825 0 : path = (Path *) create_incremental_sort_path(root,
7826 : rel,
7827 : path,
7828 : groupby_pathkeys,
7829 : presorted_keys,
7830 : -1.0);
7831 1464 : total_groups = compute_gather_rows(path);
7832 : path = (Path *)
7833 1464 : create_gather_merge_path(root,
7834 : rel,
7835 : path,
7836 1464 : rel->reltarget,
7837 : groupby_pathkeys,
7838 : NULL,
7839 : &total_groups);
7840 :
7841 1464 : add_path(rel, path);
7842 : }
7843 2326 : }
7844 :
7845 : /*
7846 : * can_partial_agg
7847 : *
7848 : * Determines whether or not partial grouping and/or aggregation is possible.
7849 : * Returns true when possible, false otherwise.
7850 : */
7851 : static bool
7852 39254 : can_partial_agg(PlannerInfo *root)
7853 : {
7854 39254 : Query *parse = root->parse;
7855 :
7856 39254 : if (!parse->hasAggs && parse->groupClause == NIL)
7857 : {
7858 : /*
7859 : * We don't know how to do parallel aggregation unless we have either
7860 : * some aggregates or a grouping clause.
7861 : */
7862 0 : return false;
7863 : }
7864 39254 : else if (parse->groupingSets)
7865 : {
7866 : /* We don't know how to do grouping sets in parallel. */
7867 926 : return false;
7868 : }
7869 38328 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7870 : {
7871 : /* Insufficient support for partial mode. */
7872 3872 : return false;
7873 : }
7874 :
7875 : /* Everything looks good. */
7876 34456 : return true;
7877 : }
7878 :
7879 : /*
7880 : * apply_scanjoin_target_to_paths
7881 : *
7882 : * Adjust the final scan/join relation, and recursively all of its children,
7883 : * to generate the final scan/join target. It would be more correct to model
7884 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7885 : * for each child relation, but doing it this way is noticeably cheaper.
7886 : * Maybe that problem can be solved at some point, but for now we do this.
7887 : *
7888 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7889 : * the same expressions as the existing reltarget, so we need only insert the
7890 : * appropriate sortgroupref information. By avoiding the creation of
7891 : * projection paths we save effort both immediately and at plan creation time.
7892 : */
7893 : static void
7894 539874 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7895 : RelOptInfo *rel,
7896 : List *scanjoin_targets,
7897 : List *scanjoin_targets_contain_srfs,
7898 : bool scanjoin_target_parallel_safe,
7899 : bool tlist_same_exprs)
7900 : {
7901 539874 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7902 : PathTarget *scanjoin_target;
7903 : ListCell *lc;
7904 :
7905 : /* This recurses, so be paranoid. */
7906 539874 : check_stack_depth();
7907 :
7908 : /*
7909 : * If the rel is partitioned, we want to drop its existing paths and
7910 : * generate new ones. This function would still be correct if we kept the
7911 : * existing paths: we'd modify them to generate the correct target above
7912 : * the partitioning Append, and then they'd compete on cost with paths
7913 : * generating the target below the Append. However, in our current cost
7914 : * model the latter way is always the same or cheaper cost, so modifying
7915 : * the existing paths would just be useless work. Moreover, when the cost
7916 : * is the same, varying roundoff errors might sometimes allow an existing
7917 : * path to be picked, resulting in undesirable cross-platform plan
7918 : * variations. So we drop old paths and thereby force the work to be done
7919 : * below the Append, except in the case of a non-parallel-safe target.
7920 : *
7921 : * Some care is needed, because we have to allow
7922 : * generate_useful_gather_paths to see the old partial paths in the next
7923 : * stanza. Hence, zap the main pathlist here, then allow
7924 : * generate_useful_gather_paths to add path(s) to the main list, and
7925 : * finally zap the partial pathlist.
7926 : */
7927 539874 : if (rel_is_partitioned)
7928 12906 : rel->pathlist = NIL;
7929 :
7930 : /*
7931 : * If the scan/join target is not parallel-safe, partial paths cannot
7932 : * generate it.
7933 : */
7934 539874 : if (!scanjoin_target_parallel_safe)
7935 : {
7936 : /*
7937 : * Since we can't generate the final scan/join target in parallel
7938 : * workers, this is our last opportunity to use any partial paths that
7939 : * exist; so build Gather path(s) that use them and emit whatever the
7940 : * current reltarget is. We don't do this in the case where the
7941 : * target is parallel-safe, since we will be able to generate superior
7942 : * paths by doing it after the final scan/join target has been
7943 : * applied.
7944 : */
7945 79324 : generate_useful_gather_paths(root, rel, false);
7946 :
7947 : /* Can't use parallel query above this level. */
7948 79324 : rel->partial_pathlist = NIL;
7949 79324 : rel->consider_parallel = false;
7950 : }
7951 :
7952 : /* Finish dropping old paths for a partitioned rel, per comment above */
7953 539874 : if (rel_is_partitioned)
7954 12906 : rel->partial_pathlist = NIL;
7955 :
7956 : /* Extract SRF-free scan/join target. */
7957 539874 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7958 :
7959 : /*
7960 : * Apply the SRF-free scan/join target to each existing path.
7961 : *
7962 : * If the tlist exprs are the same, we can just inject the sortgroupref
7963 : * information into the existing pathtargets. Otherwise, replace each
7964 : * path with a projection path that generates the SRF-free scan/join
7965 : * target. This can't change the ordering of paths within rel->pathlist,
7966 : * so we just modify the list in place.
7967 : */
7968 1120098 : foreach(lc, rel->pathlist)
7969 : {
7970 580224 : Path *subpath = (Path *) lfirst(lc);
7971 :
7972 : /* Shouldn't have any parameterized paths anymore */
7973 : Assert(subpath->param_info == NULL);
7974 :
7975 580224 : if (tlist_same_exprs)
7976 204128 : subpath->pathtarget->sortgrouprefs =
7977 204128 : scanjoin_target->sortgrouprefs;
7978 : else
7979 : {
7980 : Path *newpath;
7981 :
7982 376096 : newpath = (Path *) create_projection_path(root, rel, subpath,
7983 : scanjoin_target);
7984 376096 : lfirst(lc) = newpath;
7985 : }
7986 : }
7987 :
7988 : /* Likewise adjust the targets for any partial paths. */
7989 561066 : foreach(lc, rel->partial_pathlist)
7990 : {
7991 21192 : Path *subpath = (Path *) lfirst(lc);
7992 :
7993 : /* Shouldn't have any parameterized paths anymore */
7994 : Assert(subpath->param_info == NULL);
7995 :
7996 21192 : if (tlist_same_exprs)
7997 17264 : subpath->pathtarget->sortgrouprefs =
7998 17264 : scanjoin_target->sortgrouprefs;
7999 : else
8000 : {
8001 : Path *newpath;
8002 :
8003 3928 : newpath = (Path *) create_projection_path(root, rel, subpath,
8004 : scanjoin_target);
8005 3928 : lfirst(lc) = newpath;
8006 : }
8007 : }
8008 :
8009 : /*
8010 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8011 : * atop each existing path. (Note that this function doesn't look at the
8012 : * cheapest-path fields, which is a good thing because they're bogus right
8013 : * now.)
8014 : */
8015 539874 : if (root->parse->hasTargetSRFs)
8016 11704 : adjust_paths_for_srfs(root, rel,
8017 : scanjoin_targets,
8018 : scanjoin_targets_contain_srfs);
8019 :
8020 : /*
8021 : * Update the rel's target to be the final (with SRFs) scan/join target.
8022 : * This now matches the actual output of all the paths, and we might get
8023 : * confused in createplan.c if they don't agree. We must do this now so
8024 : * that any append paths made in the next part will use the correct
8025 : * pathtarget (cf. create_append_path).
8026 : *
8027 : * Note that this is also necessary if GetForeignUpperPaths() gets called
8028 : * on the final scan/join relation or on any of its children, since the
8029 : * FDW might look at the rel's target to create ForeignPaths.
8030 : */
8031 539874 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8032 :
8033 : /*
8034 : * If the relation is partitioned, recursively apply the scan/join target
8035 : * to all partitions, and generate brand-new Append paths in which the
8036 : * scan/join target is computed below the Append rather than above it.
8037 : * Since Append is not projection-capable, that might save a separate
8038 : * Result node, and it also is important for partitionwise aggregate.
8039 : */
8040 539874 : if (rel_is_partitioned)
8041 : {
8042 12906 : List *live_children = NIL;
8043 : int i;
8044 :
8045 : /* Adjust each partition. */
8046 12906 : i = -1;
8047 36588 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8048 : {
8049 23682 : RelOptInfo *child_rel = rel->part_rels[i];
8050 : AppendRelInfo **appinfos;
8051 : int nappinfos;
8052 23682 : List *child_scanjoin_targets = NIL;
8053 :
8054 : Assert(child_rel != NULL);
8055 :
8056 : /* Dummy children can be ignored. */
8057 23682 : if (IS_DUMMY_REL(child_rel))
8058 42 : continue;
8059 :
8060 : /* Translate scan/join targets for this child. */
8061 23640 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8062 : &nappinfos);
8063 47280 : foreach(lc, scanjoin_targets)
8064 : {
8065 23640 : PathTarget *target = lfirst_node(PathTarget, lc);
8066 :
8067 23640 : target = copy_pathtarget(target);
8068 23640 : target->exprs = (List *)
8069 23640 : adjust_appendrel_attrs(root,
8070 23640 : (Node *) target->exprs,
8071 : nappinfos, appinfos);
8072 23640 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8073 : target);
8074 : }
8075 23640 : pfree(appinfos);
8076 :
8077 : /* Recursion does the real work. */
8078 23640 : apply_scanjoin_target_to_paths(root, child_rel,
8079 : child_scanjoin_targets,
8080 : scanjoin_targets_contain_srfs,
8081 : scanjoin_target_parallel_safe,
8082 : tlist_same_exprs);
8083 :
8084 : /* Save non-dummy children for Append paths. */
8085 23640 : if (!IS_DUMMY_REL(child_rel))
8086 23640 : live_children = lappend(live_children, child_rel);
8087 : }
8088 :
8089 : /* Build new paths for this relation by appending child paths. */
8090 12906 : add_paths_to_append_rel(root, rel, live_children);
8091 : }
8092 :
8093 : /*
8094 : * Consider generating Gather or Gather Merge paths. We must only do this
8095 : * if the relation is parallel safe, and we don't do it for child rels to
8096 : * avoid creating multiple Gather nodes within the same plan. We must do
8097 : * this after all paths have been generated and before set_cheapest, since
8098 : * one of the generated paths may turn out to be the cheapest one.
8099 : */
8100 539874 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
8101 176164 : generate_useful_gather_paths(root, rel, false);
8102 :
8103 : /*
8104 : * Reassess which paths are the cheapest, now that we've potentially added
8105 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8106 : * this relation.
8107 : */
8108 539874 : set_cheapest(rel);
8109 539874 : }
8110 :
8111 : /*
8112 : * create_partitionwise_grouping_paths
8113 : *
8114 : * If the partition keys of input relation are part of the GROUP BY clause, all
8115 : * the rows belonging to a given group come from a single partition. This
8116 : * allows aggregation/grouping over a partitioned relation to be broken down
8117 : * into aggregation/grouping on each partition. This should be no worse, and
8118 : * often better, than the normal approach.
8119 : *
8120 : * However, if the GROUP BY clause does not contain all the partition keys,
8121 : * rows from a given group may be spread across multiple partitions. In that
8122 : * case, we perform partial aggregation for each group, append the results,
8123 : * and then finalize aggregation. This is less certain to win than the
8124 : * previous case. It may win if the PartialAggregate stage greatly reduces
8125 : * the number of groups, because fewer rows will pass through the Append node.
8126 : * It may lose if we have lots of small groups.
8127 : */
8128 : static void
8129 814 : create_partitionwise_grouping_paths(PlannerInfo *root,
8130 : RelOptInfo *input_rel,
8131 : RelOptInfo *grouped_rel,
8132 : RelOptInfo *partially_grouped_rel,
8133 : const AggClauseCosts *agg_costs,
8134 : grouping_sets_data *gd,
8135 : PartitionwiseAggregateType patype,
8136 : GroupPathExtraData *extra)
8137 : {
8138 814 : List *grouped_live_children = NIL;
8139 814 : List *partially_grouped_live_children = NIL;
8140 814 : PathTarget *target = grouped_rel->reltarget;
8141 814 : bool partial_grouping_valid = true;
8142 : int i;
8143 :
8144 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8145 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8146 : partially_grouped_rel != NULL);
8147 :
8148 : /* Add paths for partitionwise aggregation/grouping. */
8149 814 : i = -1;
8150 2944 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8151 : {
8152 2130 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8153 : PathTarget *child_target;
8154 : AppendRelInfo **appinfos;
8155 : int nappinfos;
8156 : GroupPathExtraData child_extra;
8157 : RelOptInfo *child_grouped_rel;
8158 : RelOptInfo *child_partially_grouped_rel;
8159 :
8160 : Assert(child_input_rel != NULL);
8161 :
8162 : /* Dummy children can be ignored. */
8163 2130 : if (IS_DUMMY_REL(child_input_rel))
8164 0 : continue;
8165 :
8166 2130 : child_target = copy_pathtarget(target);
8167 :
8168 : /*
8169 : * Copy the given "extra" structure as is and then override the
8170 : * members specific to this child.
8171 : */
8172 2130 : memcpy(&child_extra, extra, sizeof(child_extra));
8173 :
8174 2130 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8175 : &nappinfos);
8176 :
8177 2130 : child_target->exprs = (List *)
8178 2130 : adjust_appendrel_attrs(root,
8179 2130 : (Node *) target->exprs,
8180 : nappinfos, appinfos);
8181 :
8182 : /* Translate havingQual and targetList. */
8183 2130 : child_extra.havingQual = (Node *)
8184 : adjust_appendrel_attrs(root,
8185 : extra->havingQual,
8186 : nappinfos, appinfos);
8187 2130 : child_extra.targetList = (List *)
8188 2130 : adjust_appendrel_attrs(root,
8189 2130 : (Node *) extra->targetList,
8190 : nappinfos, appinfos);
8191 :
8192 : /*
8193 : * extra->patype was the value computed for our parent rel; patype is
8194 : * the value for this relation. For the child, our value is its
8195 : * parent rel's value.
8196 : */
8197 2130 : child_extra.patype = patype;
8198 :
8199 : /*
8200 : * Create grouping relation to hold fully aggregated grouping and/or
8201 : * aggregation paths for the child.
8202 : */
8203 2130 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8204 : child_target,
8205 2130 : extra->target_parallel_safe,
8206 : child_extra.havingQual);
8207 :
8208 : /* Create grouping paths for this child relation. */
8209 2130 : create_ordinary_grouping_paths(root, child_input_rel,
8210 : child_grouped_rel,
8211 : agg_costs, gd, &child_extra,
8212 : &child_partially_grouped_rel);
8213 :
8214 2130 : if (child_partially_grouped_rel)
8215 : {
8216 : partially_grouped_live_children =
8217 1542 : lappend(partially_grouped_live_children,
8218 : child_partially_grouped_rel);
8219 : }
8220 : else
8221 588 : partial_grouping_valid = false;
8222 :
8223 2130 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8224 : {
8225 1272 : set_cheapest(child_grouped_rel);
8226 1272 : grouped_live_children = lappend(grouped_live_children,
8227 : child_grouped_rel);
8228 : }
8229 :
8230 2130 : pfree(appinfos);
8231 : }
8232 :
8233 : /*
8234 : * Try to create append paths for partially grouped children. For full
8235 : * partitionwise aggregation, we might have paths in the partial_pathlist
8236 : * if parallel aggregation is possible. For partial partitionwise
8237 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8238 : *
8239 : * NB: We must have a partially grouped path for every child in order to
8240 : * generate a partially grouped path for this relation.
8241 : */
8242 814 : if (partially_grouped_rel && partial_grouping_valid)
8243 : {
8244 : Assert(partially_grouped_live_children != NIL);
8245 :
8246 602 : add_paths_to_append_rel(root, partially_grouped_rel,
8247 : partially_grouped_live_children);
8248 : }
8249 :
8250 : /* If possible, create append paths for fully grouped children. */
8251 814 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8252 : {
8253 : Assert(grouped_live_children != NIL);
8254 :
8255 476 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8256 : }
8257 814 : }
8258 :
8259 : /*
8260 : * group_by_has_partkey
8261 : *
8262 : * Returns true if all the partition keys of the given relation are part of
8263 : * the GROUP BY clauses, including having matching collation, false otherwise.
8264 : */
8265 : static bool
8266 760 : group_by_has_partkey(RelOptInfo *input_rel,
8267 : List *targetList,
8268 : List *groupClause)
8269 : {
8270 760 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8271 760 : int cnt = 0;
8272 : int partnatts;
8273 :
8274 : /* Input relation should be partitioned. */
8275 : Assert(input_rel->part_scheme);
8276 :
8277 : /* Rule out early, if there are no partition keys present. */
8278 760 : if (!input_rel->partexprs)
8279 0 : return false;
8280 :
8281 760 : partnatts = input_rel->part_scheme->partnatts;
8282 :
8283 1272 : for (cnt = 0; cnt < partnatts; cnt++)
8284 : {
8285 796 : List *partexprs = input_rel->partexprs[cnt];
8286 : ListCell *lc;
8287 796 : bool found = false;
8288 :
8289 1194 : foreach(lc, partexprs)
8290 : {
8291 : ListCell *lg;
8292 922 : Expr *partexpr = lfirst(lc);
8293 922 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8294 :
8295 1440 : foreach(lg, groupexprs)
8296 : {
8297 1042 : Expr *groupexpr = lfirst(lg);
8298 1042 : Oid groupcoll = exprCollation((Node *) groupexpr);
8299 :
8300 : /*
8301 : * Note: we can assume there is at most one RelabelType node;
8302 : * eval_const_expressions() will have simplified if more than
8303 : * one.
8304 : */
8305 1042 : if (IsA(groupexpr, RelabelType))
8306 24 : groupexpr = ((RelabelType *) groupexpr)->arg;
8307 :
8308 1042 : if (equal(groupexpr, partexpr))
8309 : {
8310 : /*
8311 : * Reject a match if the grouping collation does not match
8312 : * the partitioning collation.
8313 : */
8314 524 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8315 : partcoll != groupcoll)
8316 12 : return false;
8317 :
8318 512 : found = true;
8319 512 : break;
8320 : }
8321 : }
8322 :
8323 910 : if (found)
8324 512 : break;
8325 : }
8326 :
8327 : /*
8328 : * If none of the partition key expressions match with any of the
8329 : * GROUP BY expression, return false.
8330 : */
8331 784 : if (!found)
8332 272 : return false;
8333 : }
8334 :
8335 476 : return true;
8336 : }
8337 :
8338 : /*
8339 : * generate_setop_child_grouplist
8340 : * Build a SortGroupClause list defining the sort/grouping properties
8341 : * of the child of a set operation.
8342 : *
8343 : * This is similar to generate_setop_grouplist() but differs as the setop
8344 : * child query's targetlist entries may already have a tleSortGroupRef
8345 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8346 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8347 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8348 : * any of the columns in the targetlist don't match to the setop's colTypes
8349 : * then we return an empty list. This may leave some TLEs with unreferenced
8350 : * ressortgroupref markings, but that's harmless.
8351 : */
8352 : static List *
8353 12488 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8354 : {
8355 12488 : List *grouplist = copyObject(op->groupClauses);
8356 : ListCell *lg;
8357 : ListCell *lt;
8358 : ListCell *ct;
8359 :
8360 12488 : lg = list_head(grouplist);
8361 12488 : ct = list_head(op->colTypes);
8362 48090 : foreach(lt, targetlist)
8363 : {
8364 36016 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8365 : SortGroupClause *sgc;
8366 : Oid coltype;
8367 :
8368 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8369 36016 : if (tle->resjunk)
8370 0 : continue;
8371 :
8372 : /*
8373 : * We expect every non-resjunk target to have a SortGroupClause and
8374 : * colTypes.
8375 : */
8376 : Assert(lg != NULL);
8377 : Assert(ct != NULL);
8378 36016 : sgc = (SortGroupClause *) lfirst(lg);
8379 36016 : coltype = lfirst_oid(ct);
8380 :
8381 : /* reject if target type isn't the same as the setop target type */
8382 36016 : if (coltype != exprType((Node *) tle->expr))
8383 414 : return NIL;
8384 :
8385 35602 : lg = lnext(grouplist, lg);
8386 35602 : ct = lnext(op->colTypes, ct);
8387 :
8388 : /* assign a tleSortGroupRef, or reuse the existing one */
8389 35602 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8390 : }
8391 :
8392 : Assert(lg == NULL);
8393 : Assert(ct == NULL);
8394 :
8395 12074 : return grouplist;
8396 : }
8397 :
8398 : /*
8399 : * create_unique_paths
8400 : * Build a new RelOptInfo containing Paths that represent elimination of
8401 : * distinct rows from the input data. Distinct-ness is defined according to
8402 : * the needs of the semijoin represented by sjinfo. If it is not possible
8403 : * to identify how to make the data unique, NULL is returned.
8404 : *
8405 : * If used at all, this is likely to be called repeatedly on the same rel,
8406 : * so we cache the result.
8407 : */
8408 : RelOptInfo *
8409 8776 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8410 : {
8411 : RelOptInfo *unique_rel;
8412 8776 : List *sortPathkeys = NIL;
8413 8776 : List *groupClause = NIL;
8414 : MemoryContext oldcontext;
8415 :
8416 : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8417 : Assert(sjinfo->jointype == JOIN_SEMI);
8418 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8419 :
8420 : /* If result already cached, return it */
8421 8776 : if (rel->unique_rel)
8422 1820 : return rel->unique_rel;
8423 :
8424 : /* If it's not possible to unique-ify, return NULL */
8425 6956 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8426 132 : return NULL;
8427 :
8428 : /*
8429 : * Punt if this is a child relation and we failed to build a unique-ified
8430 : * relation for its parent. This can happen if all the RHS columns were
8431 : * found to be equated to constants when unique-ifying the parent table,
8432 : * leaving no columns to unique-ify.
8433 : */
8434 6824 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
8435 12 : return NULL;
8436 :
8437 : /*
8438 : * When called during GEQO join planning, we are in a short-lived memory
8439 : * context. We must make sure that the unique rel and any subsidiary data
8440 : * structures created for a baserel survive the GEQO cycle, else the
8441 : * baserel is trashed for future GEQO cycles. On the other hand, when we
8442 : * are creating those for a joinrel during GEQO, we don't want them to
8443 : * clutter the main planning context. Upshot is that the best solution is
8444 : * to explicitly allocate memory in the same context the given RelOptInfo
8445 : * is in.
8446 : */
8447 6812 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8448 :
8449 6812 : unique_rel = makeNode(RelOptInfo);
8450 6812 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8451 :
8452 : /*
8453 : * clear path info
8454 : */
8455 6812 : unique_rel->pathlist = NIL;
8456 6812 : unique_rel->ppilist = NIL;
8457 6812 : unique_rel->partial_pathlist = NIL;
8458 6812 : unique_rel->cheapest_startup_path = NULL;
8459 6812 : unique_rel->cheapest_total_path = NULL;
8460 6812 : unique_rel->cheapest_parameterized_paths = NIL;
8461 :
8462 : /*
8463 : * Build the target list for the unique rel. We also build the pathkeys
8464 : * that represent the ordering requirements for the sort-based
8465 : * implementation, and the list of SortGroupClause nodes that represent
8466 : * the columns to be grouped on for the hash-based implementation.
8467 : *
8468 : * For a child rel, we can construct these fields from those of its
8469 : * parent.
8470 : */
8471 6812 : if (IS_OTHER_REL(rel))
8472 432 : {
8473 : PathTarget *child_unique_target;
8474 : PathTarget *parent_unique_target;
8475 :
8476 432 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8477 :
8478 432 : child_unique_target = copy_pathtarget(parent_unique_target);
8479 :
8480 : /* Translate the target expressions */
8481 432 : child_unique_target->exprs = (List *)
8482 432 : adjust_appendrel_attrs_multilevel(root,
8483 432 : (Node *) parent_unique_target->exprs,
8484 : rel,
8485 432 : rel->top_parent);
8486 :
8487 432 : unique_rel->reltarget = child_unique_target;
8488 :
8489 432 : sortPathkeys = rel->top_parent->unique_pathkeys;
8490 432 : groupClause = rel->top_parent->unique_groupclause;
8491 : }
8492 : else
8493 : {
8494 : List *newtlist;
8495 : int nextresno;
8496 6380 : List *sortList = NIL;
8497 : ListCell *lc1;
8498 : ListCell *lc2;
8499 :
8500 : /*
8501 : * The values we are supposed to unique-ify may be expressions in the
8502 : * variables of the input rel's targetlist. We have to add any such
8503 : * expressions to the unique rel's targetlist.
8504 : *
8505 : * To complicate matters, some of the values to be unique-ified may be
8506 : * known redundant by the EquivalenceClass machinery (e.g., because
8507 : * they have been equated to constants). There is no need to compare
8508 : * such values during unique-ification, and indeed we had better not
8509 : * try because the Vars involved may not have propagated as high as
8510 : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8511 : * detect such cases, which is a tad inefficient but it doesn't seem
8512 : * worth building specialized infrastructure for this.
8513 : */
8514 6380 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8515 6380 : nextresno = list_length(newtlist) + 1;
8516 :
8517 12994 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
8518 : {
8519 6614 : Expr *uniqexpr = lfirst(lc1);
8520 6614 : Oid in_oper = lfirst_oid(lc2);
8521 : Oid sortop;
8522 : TargetEntry *tle;
8523 6614 : bool made_tle = false;
8524 :
8525 6614 : tle = tlist_member(uniqexpr, newtlist);
8526 6614 : if (!tle)
8527 : {
8528 3184 : tle = makeTargetEntry((Expr *) uniqexpr,
8529 : nextresno,
8530 : NULL,
8531 : false);
8532 3184 : newtlist = lappend(newtlist, tle);
8533 3184 : nextresno++;
8534 3184 : made_tle = true;
8535 : }
8536 :
8537 : /*
8538 : * Try to build an ORDER BY list to sort the input compatibly. We
8539 : * do this for each sortable clause even when the clauses are not
8540 : * all sortable, so that we can detect clauses that are redundant
8541 : * according to the pathkey machinery.
8542 : */
8543 6614 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8544 6614 : if (OidIsValid(sortop))
8545 : {
8546 : Oid eqop;
8547 : SortGroupClause *sortcl;
8548 :
8549 : /*
8550 : * The Unique node will need equality operators. Normally
8551 : * these are the same as the IN clause operators, but if those
8552 : * are cross-type operators then the equality operators are
8553 : * the ones for the IN clause operators' RHS datatype.
8554 : */
8555 6614 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8556 6614 : if (!OidIsValid(eqop)) /* shouldn't happen */
8557 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8558 : sortop);
8559 :
8560 6614 : sortcl = makeNode(SortGroupClause);
8561 6614 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8562 6614 : sortcl->eqop = eqop;
8563 6614 : sortcl->sortop = sortop;
8564 6614 : sortcl->reverse_sort = false;
8565 6614 : sortcl->nulls_first = false;
8566 6614 : sortcl->hashable = false; /* no need to make this accurate */
8567 6614 : sortList = lappend(sortList, sortcl);
8568 :
8569 : /*
8570 : * At each step, convert the SortGroupClause list to pathkey
8571 : * form. If the just-added SortGroupClause is redundant, the
8572 : * result will be shorter than the SortGroupClause list.
8573 : */
8574 6614 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8575 : newtlist);
8576 6614 : if (list_length(sortPathkeys) != list_length(sortList))
8577 : {
8578 : /* Drop the redundant SortGroupClause */
8579 2052 : sortList = list_delete_last(sortList);
8580 : Assert(list_length(sortPathkeys) == list_length(sortList));
8581 : /* Undo tlist addition, if we made one */
8582 2052 : if (made_tle)
8583 : {
8584 12 : newtlist = list_delete_last(newtlist);
8585 12 : nextresno--;
8586 : }
8587 : /* We need not consider this clause for hashing, either */
8588 2052 : continue;
8589 : }
8590 : }
8591 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8592 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8593 : in_oper);
8594 :
8595 4562 : if (sjinfo->semi_can_hash)
8596 : {
8597 : /* Create a GROUP BY list for the Agg node to use */
8598 : Oid eq_oper;
8599 : SortGroupClause *groupcl;
8600 :
8601 : /*
8602 : * Get the hashable equality operators for the Agg node to
8603 : * use. Normally these are the same as the IN clause
8604 : * operators, but if those are cross-type operators then the
8605 : * equality operators are the ones for the IN clause
8606 : * operators' RHS datatype.
8607 : */
8608 4562 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8609 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8610 : in_oper);
8611 :
8612 4562 : groupcl = makeNode(SortGroupClause);
8613 4562 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8614 4562 : groupcl->eqop = eq_oper;
8615 4562 : groupcl->sortop = sortop;
8616 4562 : groupcl->reverse_sort = false;
8617 4562 : groupcl->nulls_first = false;
8618 4562 : groupcl->hashable = true;
8619 4562 : groupClause = lappend(groupClause, groupcl);
8620 : }
8621 : }
8622 :
8623 : /*
8624 : * Done building the sortPathkeys and groupClause. But the
8625 : * sortPathkeys are bogus if not all the clauses were sortable.
8626 : */
8627 6380 : if (!sjinfo->semi_can_btree)
8628 0 : sortPathkeys = NIL;
8629 :
8630 : /*
8631 : * It can happen that all the RHS columns are equated to constants.
8632 : * We'd have to do something special to unique-ify in that case, and
8633 : * it's such an unlikely-in-the-real-world case that it's not worth
8634 : * the effort. So just punt if we found no columns to unique-ify.
8635 : */
8636 6380 : if (sortPathkeys == NIL && groupClause == NIL)
8637 : {
8638 1950 : MemoryContextSwitchTo(oldcontext);
8639 1950 : return NULL;
8640 : }
8641 :
8642 : /* Convert the required targetlist back to PathTarget form */
8643 4430 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8644 : }
8645 :
8646 : /* build unique paths based on input rel's pathlist */
8647 4862 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8648 : sjinfo, unique_rel);
8649 :
8650 : /* build unique paths based on input rel's partial_pathlist */
8651 4862 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8652 : sjinfo, unique_rel);
8653 :
8654 : /* Now choose the best path(s) */
8655 4862 : set_cheapest(unique_rel);
8656 :
8657 : /*
8658 : * There shouldn't be any partial paths for the unique relation;
8659 : * otherwise, we won't be able to properly guarantee uniqueness.
8660 : */
8661 : Assert(unique_rel->partial_pathlist == NIL);
8662 :
8663 : /* Cache the result */
8664 4862 : rel->unique_rel = unique_rel;
8665 4862 : rel->unique_pathkeys = sortPathkeys;
8666 4862 : rel->unique_groupclause = groupClause;
8667 :
8668 4862 : MemoryContextSwitchTo(oldcontext);
8669 :
8670 4862 : return unique_rel;
8671 : }
8672 :
8673 : /*
8674 : * create_final_unique_paths
8675 : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8676 : */
8677 : static void
8678 8496 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8679 : List *sortPathkeys, List *groupClause,
8680 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8681 : {
8682 8496 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8683 :
8684 : /* Estimate number of output rows */
8685 8496 : unique_rel->rows = estimate_num_groups(root,
8686 : sjinfo->semi_rhs_exprs,
8687 : cheapest_input_path->rows,
8688 : NULL,
8689 : NULL);
8690 :
8691 : /* Consider sort-based implementations, if possible. */
8692 8496 : if (sjinfo->semi_can_btree)
8693 : {
8694 : ListCell *lc;
8695 :
8696 : /*
8697 : * Use any available suitably-sorted path as input, and also consider
8698 : * sorting the cheapest-total path and incremental sort on any paths
8699 : * with presorted keys.
8700 : *
8701 : * To save planning time, we ignore parameterized input paths unless
8702 : * they are the cheapest-total path.
8703 : */
8704 18548 : foreach(lc, input_rel->pathlist)
8705 : {
8706 10052 : Path *input_path = (Path *) lfirst(lc);
8707 : Path *path;
8708 : bool is_sorted;
8709 : int presorted_keys;
8710 :
8711 : /*
8712 : * Ignore parameterized paths that are not the cheapest-total
8713 : * path.
8714 : */
8715 10052 : if (input_path->param_info &&
8716 : input_path != cheapest_input_path)
8717 916 : continue;
8718 :
8719 9180 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8720 : input_path->pathkeys,
8721 : &presorted_keys);
8722 :
8723 : /*
8724 : * Ignore paths that are not suitably or partially sorted, unless
8725 : * they are the cheapest total path (no need to deal with paths
8726 : * which have presorted keys when incremental sort is disabled).
8727 : */
8728 9180 : if (!is_sorted && input_path != cheapest_input_path &&
8729 92 : (presorted_keys == 0 || !enable_incremental_sort))
8730 44 : continue;
8731 :
8732 : /*
8733 : * Make a separate ProjectionPath in case we need a Result node.
8734 : */
8735 9136 : path = (Path *) create_projection_path(root,
8736 : unique_rel,
8737 : input_path,
8738 9136 : unique_rel->reltarget);
8739 :
8740 9136 : if (!is_sorted)
8741 : {
8742 : /*
8743 : * We've no need to consider both a sort and incremental sort.
8744 : * We'll just do a sort if there are no presorted keys and an
8745 : * incremental sort when there are presorted keys.
8746 : */
8747 4842 : if (presorted_keys == 0 || !enable_incremental_sort)
8748 4794 : path = (Path *) create_sort_path(root,
8749 : unique_rel,
8750 : path,
8751 : sortPathkeys,
8752 : -1.0);
8753 : else
8754 48 : path = (Path *) create_incremental_sort_path(root,
8755 : unique_rel,
8756 : path,
8757 : sortPathkeys,
8758 : presorted_keys,
8759 : -1.0);
8760 : }
8761 :
8762 9136 : path = (Path *) create_unique_path(root, unique_rel, path,
8763 : list_length(sortPathkeys),
8764 : unique_rel->rows);
8765 :
8766 9136 : add_path(unique_rel, path);
8767 : }
8768 : }
8769 :
8770 : /* Consider hash-based implementation, if possible. */
8771 8496 : if (sjinfo->semi_can_hash)
8772 : {
8773 : Path *path;
8774 :
8775 : /*
8776 : * Make a separate ProjectionPath in case we need a Result node.
8777 : */
8778 8496 : path = (Path *) create_projection_path(root,
8779 : unique_rel,
8780 : cheapest_input_path,
8781 8496 : unique_rel->reltarget);
8782 :
8783 8496 : path = (Path *) create_agg_path(root,
8784 : unique_rel,
8785 : path,
8786 : cheapest_input_path->pathtarget,
8787 : AGG_HASHED,
8788 : AGGSPLIT_SIMPLE,
8789 : groupClause,
8790 : NIL,
8791 : NULL,
8792 : unique_rel->rows);
8793 :
8794 8496 : add_path(unique_rel, path);
8795 : }
8796 8496 : }
8797 :
8798 : /*
8799 : * create_partial_unique_paths
8800 : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
8801 : */
8802 : static void
8803 4862 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8804 : List *sortPathkeys, List *groupClause,
8805 : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8806 : {
8807 : RelOptInfo *partial_unique_rel;
8808 : Path *cheapest_partial_path;
8809 :
8810 : /* nothing to do when there are no partial paths in the input rel */
8811 4862 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
8812 1228 : return;
8813 :
8814 : /*
8815 : * nothing to do if there's anything in the targetlist that's
8816 : * parallel-restricted.
8817 : */
8818 3634 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
8819 0 : return;
8820 :
8821 3634 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
8822 :
8823 3634 : partial_unique_rel = makeNode(RelOptInfo);
8824 3634 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
8825 :
8826 : /*
8827 : * clear path info
8828 : */
8829 3634 : partial_unique_rel->pathlist = NIL;
8830 3634 : partial_unique_rel->ppilist = NIL;
8831 3634 : partial_unique_rel->partial_pathlist = NIL;
8832 3634 : partial_unique_rel->cheapest_startup_path = NULL;
8833 3634 : partial_unique_rel->cheapest_total_path = NULL;
8834 3634 : partial_unique_rel->cheapest_parameterized_paths = NIL;
8835 :
8836 : /* Estimate number of output rows */
8837 3634 : partial_unique_rel->rows = estimate_num_groups(root,
8838 : sjinfo->semi_rhs_exprs,
8839 : cheapest_partial_path->rows,
8840 : NULL,
8841 : NULL);
8842 3634 : partial_unique_rel->reltarget = unique_rel->reltarget;
8843 :
8844 : /* Consider sort-based implementations, if possible. */
8845 3634 : if (sjinfo->semi_can_btree)
8846 : {
8847 : ListCell *lc;
8848 :
8849 : /*
8850 : * Use any available suitably-sorted path as input, and also consider
8851 : * sorting the cheapest partial path and incremental sort on any paths
8852 : * with presorted keys.
8853 : */
8854 7580 : foreach(lc, input_rel->partial_pathlist)
8855 : {
8856 3946 : Path *input_path = (Path *) lfirst(lc);
8857 : Path *path;
8858 : bool is_sorted;
8859 : int presorted_keys;
8860 :
8861 3946 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8862 : input_path->pathkeys,
8863 : &presorted_keys);
8864 :
8865 : /*
8866 : * Ignore paths that are not suitably or partially sorted, unless
8867 : * they are the cheapest partial path (no need to deal with paths
8868 : * which have presorted keys when incremental sort is disabled).
8869 : */
8870 3946 : if (!is_sorted && input_path != cheapest_partial_path &&
8871 0 : (presorted_keys == 0 || !enable_incremental_sort))
8872 0 : continue;
8873 :
8874 : /*
8875 : * Make a separate ProjectionPath in case we need a Result node.
8876 : */
8877 3946 : path = (Path *) create_projection_path(root,
8878 : partial_unique_rel,
8879 : input_path,
8880 3946 : partial_unique_rel->reltarget);
8881 :
8882 3946 : if (!is_sorted)
8883 : {
8884 : /*
8885 : * We've no need to consider both a sort and incremental sort.
8886 : * We'll just do a sort if there are no presorted keys and an
8887 : * incremental sort when there are presorted keys.
8888 : */
8889 3586 : if (presorted_keys == 0 || !enable_incremental_sort)
8890 3586 : path = (Path *) create_sort_path(root,
8891 : partial_unique_rel,
8892 : path,
8893 : sortPathkeys,
8894 : -1.0);
8895 : else
8896 0 : path = (Path *) create_incremental_sort_path(root,
8897 : partial_unique_rel,
8898 : path,
8899 : sortPathkeys,
8900 : presorted_keys,
8901 : -1.0);
8902 : }
8903 :
8904 3946 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
8905 : list_length(sortPathkeys),
8906 : partial_unique_rel->rows);
8907 :
8908 3946 : add_partial_path(partial_unique_rel, path);
8909 : }
8910 : }
8911 :
8912 : /* Consider hash-based implementation, if possible. */
8913 3634 : if (sjinfo->semi_can_hash)
8914 : {
8915 : Path *path;
8916 :
8917 : /*
8918 : * Make a separate ProjectionPath in case we need a Result node.
8919 : */
8920 3634 : path = (Path *) create_projection_path(root,
8921 : partial_unique_rel,
8922 : cheapest_partial_path,
8923 3634 : partial_unique_rel->reltarget);
8924 :
8925 3634 : path = (Path *) create_agg_path(root,
8926 : partial_unique_rel,
8927 : path,
8928 : cheapest_partial_path->pathtarget,
8929 : AGG_HASHED,
8930 : AGGSPLIT_SIMPLE,
8931 : groupClause,
8932 : NIL,
8933 : NULL,
8934 : partial_unique_rel->rows);
8935 :
8936 3634 : add_partial_path(partial_unique_rel, path);
8937 : }
8938 :
8939 3634 : if (partial_unique_rel->partial_pathlist != NIL)
8940 : {
8941 3634 : generate_useful_gather_paths(root, partial_unique_rel, true);
8942 3634 : set_cheapest(partial_unique_rel);
8943 :
8944 : /*
8945 : * Finally, create paths to unique-ify the final result. This step is
8946 : * needed to remove any duplicates due to combining rows from parallel
8947 : * workers.
8948 : */
8949 3634 : create_final_unique_paths(root, partial_unique_rel,
8950 : sortPathkeys, groupClause,
8951 : sjinfo, unique_rel);
8952 : }
8953 : }
8954 :
8955 : /*
8956 : * Choose a unique name for some subroot.
8957 : *
8958 : * Modifies glob->subplanNames to track names already used.
8959 : */
8960 : char *
8961 77594 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
8962 : {
8963 : unsigned n;
8964 :
8965 : /*
8966 : * If a numeric suffix is not required, then search the list of
8967 : * previously-assigned names for a match. If none is found, then we can
8968 : * use the provided name without modification.
8969 : */
8970 77594 : if (!always_number)
8971 : {
8972 19842 : bool found = false;
8973 :
8974 49532 : foreach_ptr(char, subplan_name, glob->subplanNames)
8975 : {
8976 15468 : if (strcmp(subplan_name, name) == 0)
8977 : {
8978 5620 : found = true;
8979 5620 : break;
8980 : }
8981 : }
8982 :
8983 19842 : if (!found)
8984 : {
8985 : /* pstrdup here is just to avoid cast-away-const */
8986 14222 : char *chosen_name = pstrdup(name);
8987 :
8988 14222 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
8989 14222 : return chosen_name;
8990 : }
8991 : }
8992 :
8993 : /*
8994 : * If a numeric suffix is required or if the un-suffixed name is already
8995 : * in use, then loop until we find a positive integer that produces a
8996 : * novel name.
8997 : */
8998 63372 : for (n = 1; true; ++n)
8999 54626 : {
9000 117998 : char *proposed_name = psprintf("%s_%u", name, n);
9001 117998 : bool found = false;
9002 :
9003 450326 : foreach_ptr(char, subplan_name, glob->subplanNames)
9004 : {
9005 268956 : if (strcmp(subplan_name, proposed_name) == 0)
9006 : {
9007 54626 : found = true;
9008 54626 : break;
9009 : }
9010 : }
9011 :
9012 117998 : if (!found)
9013 : {
9014 63372 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9015 63372 : return proposed_name;
9016 : }
9017 :
9018 54626 : pfree(proposed_name);
9019 : }
9020 : }
|
