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/backend_status.h"
62 : #include "utils/lsyscache.h"
63 : #include "utils/rel.h"
64 : #include "utils/selfuncs.h"
65 :
66 : /* GUC parameters */
67 : double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
68 : int debug_parallel_query = DEBUG_PARALLEL_OFF;
69 : bool parallel_leader_participation = true;
70 : bool enable_distinct_reordering = true;
71 :
72 : /* Hook for plugins to get control in planner() */
73 : planner_hook_type planner_hook = NULL;
74 :
75 : /* Hook for plugins to get control when grouping_planner() plans upper rels */
76 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
77 :
78 :
79 : /* Expression kind codes for preprocess_expression */
80 : #define EXPRKIND_QUAL 0
81 : #define EXPRKIND_TARGET 1
82 : #define EXPRKIND_RTFUNC 2
83 : #define EXPRKIND_RTFUNC_LATERAL 3
84 : #define EXPRKIND_VALUES 4
85 : #define EXPRKIND_VALUES_LATERAL 5
86 : #define EXPRKIND_LIMIT 6
87 : #define EXPRKIND_APPINFO 7
88 : #define EXPRKIND_PHV 8
89 : #define EXPRKIND_TABLESAMPLE 9
90 : #define EXPRKIND_ARBITER_ELEM 10
91 : #define EXPRKIND_TABLEFUNC 11
92 : #define EXPRKIND_TABLEFUNC_LATERAL 12
93 : #define EXPRKIND_GROUPEXPR 13
94 :
95 : /*
96 : * Data specific to grouping sets
97 : */
98 : typedef struct
99 : {
100 : List *rollups;
101 : List *hash_sets_idx;
102 : double dNumHashGroups;
103 : bool any_hashable;
104 : Bitmapset *unsortable_refs;
105 : Bitmapset *unhashable_refs;
106 : List *unsortable_sets;
107 : int *tleref_to_colnum_map;
108 : } grouping_sets_data;
109 :
110 : /*
111 : * Temporary structure for use during WindowClause reordering in order to be
112 : * able to sort WindowClauses on partitioning/ordering prefix.
113 : */
114 : typedef struct
115 : {
116 : WindowClause *wc;
117 : List *uniqueOrder; /* A List of unique ordering/partitioning
118 : * clauses per Window */
119 : } WindowClauseSortData;
120 :
121 : /* Passthrough data for standard_qp_callback */
122 : typedef struct
123 : {
124 : List *activeWindows; /* active windows, if any */
125 : grouping_sets_data *gset_data; /* grouping sets data, if any */
126 : SetOperationStmt *setop; /* parent set operation or NULL if not a
127 : * subquery belonging to a set operation */
128 : } standard_qp_extra;
129 :
130 : /* Local functions */
131 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
132 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
133 : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
134 : SetOperationStmt *setops);
135 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
136 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
137 : int *tleref_to_colnum_map);
138 : static void preprocess_rowmarks(PlannerInfo *root);
139 : static double preprocess_limit(PlannerInfo *root,
140 : double tuple_fraction,
141 : int64 *offset_est, int64 *count_est);
142 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
143 : static List *extract_rollup_sets(List *groupingSets);
144 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
145 : static void standard_qp_callback(PlannerInfo *root, void *extra);
146 : static double get_number_of_groups(PlannerInfo *root,
147 : double path_rows,
148 : grouping_sets_data *gd,
149 : List *target_list);
150 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
151 : RelOptInfo *input_rel,
152 : PathTarget *target,
153 : bool target_parallel_safe,
154 : grouping_sets_data *gd);
155 : static bool is_degenerate_grouping(PlannerInfo *root);
156 : static void create_degenerate_grouping_paths(PlannerInfo *root,
157 : RelOptInfo *input_rel,
158 : RelOptInfo *grouped_rel);
159 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
160 : PathTarget *target, bool target_parallel_safe,
161 : Node *havingQual);
162 : static void create_ordinary_grouping_paths(PlannerInfo *root,
163 : RelOptInfo *input_rel,
164 : RelOptInfo *grouped_rel,
165 : const AggClauseCosts *agg_costs,
166 : grouping_sets_data *gd,
167 : GroupPathExtraData *extra,
168 : RelOptInfo **partially_grouped_rel_p);
169 : static void consider_groupingsets_paths(PlannerInfo *root,
170 : RelOptInfo *grouped_rel,
171 : Path *path,
172 : bool is_sorted,
173 : bool can_hash,
174 : grouping_sets_data *gd,
175 : const AggClauseCosts *agg_costs,
176 : double dNumGroups);
177 : static RelOptInfo *create_window_paths(PlannerInfo *root,
178 : RelOptInfo *input_rel,
179 : PathTarget *input_target,
180 : PathTarget *output_target,
181 : bool output_target_parallel_safe,
182 : WindowFuncLists *wflists,
183 : List *activeWindows);
184 : static void create_one_window_path(PlannerInfo *root,
185 : RelOptInfo *window_rel,
186 : Path *path,
187 : PathTarget *input_target,
188 : PathTarget *output_target,
189 : WindowFuncLists *wflists,
190 : List *activeWindows);
191 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
192 : RelOptInfo *input_rel,
193 : PathTarget *target);
194 : static void create_partial_distinct_paths(PlannerInfo *root,
195 : RelOptInfo *input_rel,
196 : RelOptInfo *final_distinct_rel,
197 : PathTarget *target);
198 : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
199 : RelOptInfo *input_rel,
200 : RelOptInfo *distinct_rel);
201 : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
202 : List *needed_pathkeys,
203 : List *path_pathkeys);
204 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
205 : RelOptInfo *input_rel,
206 : PathTarget *target,
207 : bool target_parallel_safe,
208 : double limit_tuples);
209 : static PathTarget *make_group_input_target(PlannerInfo *root,
210 : PathTarget *final_target);
211 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
212 : PathTarget *grouping_target,
213 : Node *havingQual);
214 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
215 : static void optimize_window_clauses(PlannerInfo *root,
216 : WindowFuncLists *wflists);
217 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
218 : static void name_active_windows(List *activeWindows);
219 : static PathTarget *make_window_input_target(PlannerInfo *root,
220 : PathTarget *final_target,
221 : List *activeWindows);
222 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
223 : List *tlist);
224 : static PathTarget *make_sort_input_target(PlannerInfo *root,
225 : PathTarget *final_target,
226 : bool *have_postponed_srfs);
227 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
228 : List *targets, List *targets_contain_srfs);
229 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
230 : RelOptInfo *grouped_rel,
231 : RelOptInfo *partially_grouped_rel,
232 : const AggClauseCosts *agg_costs,
233 : grouping_sets_data *gd,
234 : double dNumGroups,
235 : GroupPathExtraData *extra);
236 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
237 : RelOptInfo *grouped_rel,
238 : RelOptInfo *input_rel,
239 : grouping_sets_data *gd,
240 : GroupPathExtraData *extra,
241 : bool force_rel_creation);
242 : static Path *make_ordered_path(PlannerInfo *root,
243 : RelOptInfo *rel,
244 : Path *path,
245 : Path *cheapest_path,
246 : List *pathkeys,
247 : double limit_tuples);
248 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
249 : static bool can_partial_agg(PlannerInfo *root);
250 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
251 : RelOptInfo *rel,
252 : List *scanjoin_targets,
253 : List *scanjoin_targets_contain_srfs,
254 : bool scanjoin_target_parallel_safe,
255 : bool tlist_same_exprs);
256 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
257 : RelOptInfo *input_rel,
258 : RelOptInfo *grouped_rel,
259 : RelOptInfo *partially_grouped_rel,
260 : const AggClauseCosts *agg_costs,
261 : grouping_sets_data *gd,
262 : PartitionwiseAggregateType patype,
263 : GroupPathExtraData *extra);
264 : static bool group_by_has_partkey(RelOptInfo *input_rel,
265 : List *targetList,
266 : List *groupClause);
267 : static int common_prefix_cmp(const void *a, const void *b);
268 : static List *generate_setop_child_grouplist(SetOperationStmt *op,
269 : List *targetlist);
270 :
271 :
272 : /*****************************************************************************
273 : *
274 : * Query optimizer entry point
275 : *
276 : * To support loadable plugins that monitor or modify planner behavior,
277 : * we provide a hook variable that lets a plugin get control before and
278 : * after the standard planning process. The plugin would normally call
279 : * standard_planner().
280 : *
281 : * Note to plugin authors: standard_planner() scribbles on its Query input,
282 : * so you'd better copy that data structure if you want to plan more than once.
283 : *
284 : *****************************************************************************/
285 : PlannedStmt *
286 463060 : planner(Query *parse, const char *query_string, int cursorOptions,
287 : ParamListInfo boundParams)
288 : {
289 : PlannedStmt *result;
290 :
291 463060 : if (planner_hook)
292 93620 : result = (*planner_hook) (parse, query_string, cursorOptions, boundParams);
293 : else
294 369440 : result = standard_planner(parse, query_string, cursorOptions, boundParams);
295 :
296 458804 : pgstat_report_plan_id(result->planId, false);
297 :
298 458804 : return result;
299 : }
300 :
301 : PlannedStmt *
302 463060 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
303 : ParamListInfo boundParams)
304 : {
305 : PlannedStmt *result;
306 : PlannerGlobal *glob;
307 : double tuple_fraction;
308 : PlannerInfo *root;
309 : RelOptInfo *final_rel;
310 : Path *best_path;
311 : Plan *top_plan;
312 : ListCell *lp,
313 : *lr;
314 :
315 : /*
316 : * Set up global state for this planner invocation. This data is needed
317 : * across all levels of sub-Query that might exist in the given command,
318 : * so we keep it in a separate struct that's linked to by each per-Query
319 : * PlannerInfo.
320 : */
321 463060 : glob = makeNode(PlannerGlobal);
322 :
323 463060 : glob->boundParams = boundParams;
324 463060 : glob->subplans = NIL;
325 463060 : glob->subpaths = NIL;
326 463060 : glob->subroots = NIL;
327 463060 : glob->rewindPlanIDs = NULL;
328 463060 : glob->finalrtable = NIL;
329 463060 : glob->allRelids = NULL;
330 463060 : glob->prunableRelids = NULL;
331 463060 : glob->finalrteperminfos = NIL;
332 463060 : glob->finalrowmarks = NIL;
333 463060 : glob->resultRelations = NIL;
334 463060 : glob->firstResultRels = NIL;
335 463060 : glob->appendRelations = NIL;
336 463060 : glob->partPruneInfos = NIL;
337 463060 : glob->relationOids = NIL;
338 463060 : glob->invalItems = NIL;
339 463060 : glob->paramExecTypes = NIL;
340 463060 : glob->lastPHId = 0;
341 463060 : glob->lastRowMarkId = 0;
342 463060 : glob->lastPlanNodeId = 0;
343 463060 : glob->transientPlan = false;
344 463060 : glob->dependsOnRole = false;
345 463060 : glob->partition_directory = NULL;
346 :
347 : /*
348 : * Assess whether it's feasible to use parallel mode for this query. We
349 : * can't do this in a standalone backend, or if the command will try to
350 : * modify any data, or if this is a cursor operation, or if GUCs are set
351 : * to values that don't permit parallelism, or if parallel-unsafe
352 : * functions are present in the query tree.
353 : *
354 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
355 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
356 : * the command is writing into a completely new table which workers won't
357 : * be able to see. If the workers could see the table, the fact that
358 : * group locking would cause them to ignore the leader's heavyweight GIN
359 : * page locks would make this unsafe. We'll have to fix that somehow if
360 : * we want to allow parallel inserts in general; updates and deletes have
361 : * additional problems especially around combo CIDs.)
362 : *
363 : * For now, we don't try to use parallel mode if we're running inside a
364 : * parallel worker. We might eventually be able to relax this
365 : * restriction, but for now it seems best not to have parallel workers
366 : * trying to create their own parallel workers.
367 : */
368 463060 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
369 435254 : IsUnderPostmaster &&
370 435254 : parse->commandType == CMD_SELECT &&
371 349254 : !parse->hasModifyingCTE &&
372 349114 : max_parallel_workers_per_gather > 0 &&
373 348516 : !IsParallelWorker())
374 : {
375 : /* all the cheap tests pass, so scan the query tree */
376 348468 : glob->maxParallelHazard = max_parallel_hazard(parse);
377 348468 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
378 : }
379 : else
380 : {
381 : /* skip the query tree scan, just assume it's unsafe */
382 114592 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
383 114592 : glob->parallelModeOK = false;
384 : }
385 :
386 : /*
387 : * glob->parallelModeNeeded is normally set to false here and changed to
388 : * true during plan creation if a Gather or Gather Merge plan is actually
389 : * created (cf. create_gather_plan, create_gather_merge_plan).
390 : *
391 : * However, if debug_parallel_query = on or debug_parallel_query =
392 : * regress, then we impose parallel mode whenever it's safe to do so, even
393 : * if the final plan doesn't use parallelism. It's not safe to do so if
394 : * the query contains anything parallel-unsafe; parallelModeOK will be
395 : * false in that case. Note that parallelModeOK can't change after this
396 : * point. Otherwise, everything in the query is either parallel-safe or
397 : * parallel-restricted, and in either case it should be OK to impose
398 : * parallel-mode restrictions. If that ends up breaking something, then
399 : * either some function the user included in the query is incorrectly
400 : * labeled as parallel-safe or parallel-restricted when in reality it's
401 : * parallel-unsafe, or else the query planner itself has a bug.
402 : */
403 759946 : glob->parallelModeNeeded = glob->parallelModeOK &&
404 296886 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
405 :
406 : /* Determine what fraction of the plan is likely to be scanned */
407 463060 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
408 : {
409 : /*
410 : * We have no real idea how many tuples the user will ultimately FETCH
411 : * from a cursor, but it is often the case that he doesn't want 'em
412 : * all, or would prefer a fast-start plan anyway so that he can
413 : * process some of the tuples sooner. Use a GUC parameter to decide
414 : * what fraction to optimize for.
415 : */
416 4670 : tuple_fraction = cursor_tuple_fraction;
417 :
418 : /*
419 : * We document cursor_tuple_fraction as simply being a fraction, which
420 : * means the edge cases 0 and 1 have to be treated specially here. We
421 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
422 : */
423 4670 : if (tuple_fraction >= 1.0)
424 0 : tuple_fraction = 0.0;
425 4670 : else if (tuple_fraction <= 0.0)
426 0 : tuple_fraction = 1e-10;
427 : }
428 : else
429 : {
430 : /* Default assumption is we need all the tuples */
431 458390 : tuple_fraction = 0.0;
432 : }
433 :
434 : /* primary planning entry point (may recurse for subqueries) */
435 463060 : root = subquery_planner(glob, parse, NULL, false, tuple_fraction, NULL);
436 :
437 : /* Select best Path and turn it into a Plan */
438 459200 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
439 459200 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
440 :
441 459200 : top_plan = create_plan(root, best_path);
442 :
443 : /*
444 : * If creating a plan for a scrollable cursor, make sure it can run
445 : * backwards on demand. Add a Material node at the top at need.
446 : */
447 458804 : if (cursorOptions & CURSOR_OPT_SCROLL)
448 : {
449 266 : if (!ExecSupportsBackwardScan(top_plan))
450 32 : top_plan = materialize_finished_plan(top_plan);
451 : }
452 :
453 : /*
454 : * Optionally add a Gather node for testing purposes, provided this is
455 : * actually a safe thing to do.
456 : *
457 : * We can add Gather even when top_plan has parallel-safe initPlans, but
458 : * then we have to move the initPlans to the Gather node because of
459 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
460 : * regression tests when debug_parallel_query = regress, because initPlans
461 : * that would normally appear on the top_plan move to the Gather, causing
462 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
463 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
464 : */
465 458804 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
466 194 : top_plan->parallel_safe &&
467 128 : (top_plan->initPlan == NIL ||
468 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
469 : {
470 128 : Gather *gather = makeNode(Gather);
471 : Cost initplan_cost;
472 : bool unsafe_initplans;
473 :
474 128 : gather->plan.targetlist = top_plan->targetlist;
475 128 : gather->plan.qual = NIL;
476 128 : gather->plan.lefttree = top_plan;
477 128 : gather->plan.righttree = NULL;
478 128 : gather->num_workers = 1;
479 128 : gather->single_copy = true;
480 128 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
481 :
482 : /* Transfer any initPlans to the new top node */
483 128 : gather->plan.initPlan = top_plan->initPlan;
484 128 : top_plan->initPlan = NIL;
485 :
486 : /*
487 : * Since this Gather has no parallel-aware descendants to signal to,
488 : * we don't need a rescan Param.
489 : */
490 128 : gather->rescan_param = -1;
491 :
492 : /*
493 : * Ideally we'd use cost_gather here, but setting up dummy path data
494 : * to satisfy it doesn't seem much cleaner than knowing what it does.
495 : */
496 128 : gather->plan.startup_cost = top_plan->startup_cost +
497 : parallel_setup_cost;
498 128 : gather->plan.total_cost = top_plan->total_cost +
499 128 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
500 128 : gather->plan.plan_rows = top_plan->plan_rows;
501 128 : gather->plan.plan_width = top_plan->plan_width;
502 128 : gather->plan.parallel_aware = false;
503 128 : gather->plan.parallel_safe = false;
504 :
505 : /*
506 : * Delete the initplans' cost from top_plan. We needn't add it to the
507 : * Gather node, since the above coding already included it there.
508 : */
509 128 : SS_compute_initplan_cost(gather->plan.initPlan,
510 : &initplan_cost, &unsafe_initplans);
511 128 : top_plan->startup_cost -= initplan_cost;
512 128 : top_plan->total_cost -= initplan_cost;
513 :
514 : /* use parallel mode for parallel plans. */
515 128 : root->glob->parallelModeNeeded = true;
516 :
517 128 : top_plan = &gather->plan;
518 : }
519 :
520 : /*
521 : * If any Params were generated, run through the plan tree and compute
522 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
523 : * set_plan_references' tree traversal, but for now it has to be separate
524 : * because we need to visit subplans before not after main plan.
525 : */
526 458804 : if (glob->paramExecTypes != NIL)
527 : {
528 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
529 203902 : forboth(lp, glob->subplans, lr, glob->subroots)
530 : {
531 45712 : Plan *subplan = (Plan *) lfirst(lp);
532 45712 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
533 :
534 45712 : SS_finalize_plan(subroot, subplan);
535 : }
536 158190 : SS_finalize_plan(root, top_plan);
537 : }
538 :
539 : /* final cleanup of the plan */
540 : Assert(glob->finalrtable == NIL);
541 : Assert(glob->finalrteperminfos == NIL);
542 : Assert(glob->finalrowmarks == NIL);
543 : Assert(glob->resultRelations == NIL);
544 : Assert(glob->appendRelations == NIL);
545 458804 : top_plan = set_plan_references(root, top_plan);
546 : /* ... and the subplans (both regular subplans and initplans) */
547 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
548 504516 : forboth(lp, glob->subplans, lr, glob->subroots)
549 : {
550 45712 : Plan *subplan = (Plan *) lfirst(lp);
551 45712 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
552 :
553 45712 : lfirst(lp) = set_plan_references(subroot, subplan);
554 : }
555 :
556 : /* build the PlannedStmt result */
557 458804 : result = makeNode(PlannedStmt);
558 :
559 458804 : result->commandType = parse->commandType;
560 458804 : result->queryId = parse->queryId;
561 458804 : result->hasReturning = (parse->returningList != NIL);
562 458804 : result->hasModifyingCTE = parse->hasModifyingCTE;
563 458804 : result->canSetTag = parse->canSetTag;
564 458804 : result->transientPlan = glob->transientPlan;
565 458804 : result->dependsOnRole = glob->dependsOnRole;
566 458804 : result->parallelModeNeeded = glob->parallelModeNeeded;
567 458804 : result->planTree = top_plan;
568 458804 : result->partPruneInfos = glob->partPruneInfos;
569 458804 : result->rtable = glob->finalrtable;
570 917608 : result->unprunableRelids = bms_difference(glob->allRelids,
571 458804 : glob->prunableRelids);
572 458804 : result->permInfos = glob->finalrteperminfos;
573 458804 : result->resultRelations = glob->resultRelations;
574 458804 : result->firstResultRels = glob->firstResultRels;
575 458804 : result->appendRelations = glob->appendRelations;
576 458804 : result->subplans = glob->subplans;
577 458804 : result->rewindPlanIDs = glob->rewindPlanIDs;
578 458804 : result->rowMarks = glob->finalrowmarks;
579 458804 : result->relationOids = glob->relationOids;
580 458804 : result->invalItems = glob->invalItems;
581 458804 : result->paramExecTypes = glob->paramExecTypes;
582 : /* utilityStmt should be null, but we might as well copy it */
583 458804 : result->utilityStmt = parse->utilityStmt;
584 458804 : result->stmt_location = parse->stmt_location;
585 458804 : result->stmt_len = parse->stmt_len;
586 :
587 458804 : result->jitFlags = PGJIT_NONE;
588 458804 : if (jit_enabled && jit_above_cost >= 0 &&
589 458102 : top_plan->total_cost > jit_above_cost)
590 : {
591 944 : result->jitFlags |= PGJIT_PERFORM;
592 :
593 : /*
594 : * Decide how much effort should be put into generating better code.
595 : */
596 944 : if (jit_optimize_above_cost >= 0 &&
597 944 : top_plan->total_cost > jit_optimize_above_cost)
598 432 : result->jitFlags |= PGJIT_OPT3;
599 944 : if (jit_inline_above_cost >= 0 &&
600 944 : top_plan->total_cost > jit_inline_above_cost)
601 432 : result->jitFlags |= PGJIT_INLINE;
602 :
603 : /*
604 : * Decide which operations should be JITed.
605 : */
606 944 : if (jit_expressions)
607 944 : result->jitFlags |= PGJIT_EXPR;
608 944 : if (jit_tuple_deforming)
609 944 : result->jitFlags |= PGJIT_DEFORM;
610 : }
611 :
612 458804 : if (glob->partition_directory != NULL)
613 11530 : DestroyPartitionDirectory(glob->partition_directory);
614 :
615 458804 : return result;
616 : }
617 :
618 :
619 : /*--------------------
620 : * subquery_planner
621 : * Invokes the planner on a subquery. We recurse to here for each
622 : * sub-SELECT found in the query tree.
623 : *
624 : * glob is the global state for the current planner run.
625 : * parse is the querytree produced by the parser & rewriter.
626 : * parent_root is the immediate parent Query's info (NULL at the top level).
627 : * hasRecursion is true if this is a recursive WITH query.
628 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
629 : * tuple_fraction is interpreted as explained for grouping_planner, below.
630 : * setops is used for set operation subqueries to provide the subquery with
631 : * the context in which it's being used so that Paths correctly sorted for the
632 : * set operation can be generated. NULL when not planning a set operation
633 : * child, or when a child of a set op that isn't interested in sorted input.
634 : *
635 : * Basically, this routine does the stuff that should only be done once
636 : * per Query object. It then calls grouping_planner. At one time,
637 : * grouping_planner could be invoked recursively on the same Query object;
638 : * that's not currently true, but we keep the separation between the two
639 : * routines anyway, in case we need it again someday.
640 : *
641 : * subquery_planner will be called recursively to handle sub-Query nodes
642 : * found within the query's expressions and rangetable.
643 : *
644 : * Returns the PlannerInfo struct ("root") that contains all data generated
645 : * while planning the subquery. In particular, the Path(s) attached to
646 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
647 : * cheapest way(s) to implement the query. The top level will select the
648 : * best Path and pass it through createplan.c to produce a finished Plan.
649 : *--------------------
650 : */
651 : PlannerInfo *
652 536500 : subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root,
653 : bool hasRecursion, double tuple_fraction,
654 : SetOperationStmt *setops)
655 : {
656 : PlannerInfo *root;
657 : List *newWithCheckOptions;
658 : List *newHaving;
659 : bool hasOuterJoins;
660 : bool hasResultRTEs;
661 : RelOptInfo *final_rel;
662 : ListCell *l;
663 :
664 : /* Create a PlannerInfo data structure for this subquery */
665 536500 : root = makeNode(PlannerInfo);
666 536500 : root->parse = parse;
667 536500 : root->glob = glob;
668 536500 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
669 536500 : root->parent_root = parent_root;
670 536500 : root->plan_params = NIL;
671 536500 : root->outer_params = NULL;
672 536500 : root->planner_cxt = CurrentMemoryContext;
673 536500 : root->init_plans = NIL;
674 536500 : root->cte_plan_ids = NIL;
675 536500 : root->multiexpr_params = NIL;
676 536500 : root->join_domains = NIL;
677 536500 : root->eq_classes = NIL;
678 536500 : root->ec_merging_done = false;
679 536500 : root->last_rinfo_serial = 0;
680 536500 : root->all_result_relids =
681 536500 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
682 536500 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
683 536500 : root->append_rel_list = NIL;
684 536500 : root->row_identity_vars = NIL;
685 536500 : root->rowMarks = NIL;
686 536500 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
687 536500 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
688 536500 : root->processed_groupClause = NIL;
689 536500 : root->processed_distinctClause = NIL;
690 536500 : root->processed_tlist = NIL;
691 536500 : root->update_colnos = NIL;
692 536500 : root->grouping_map = NULL;
693 536500 : root->minmax_aggs = NIL;
694 536500 : root->qual_security_level = 0;
695 536500 : root->hasPseudoConstantQuals = false;
696 536500 : root->hasAlternativeSubPlans = false;
697 536500 : root->placeholdersFrozen = false;
698 536500 : root->hasRecursion = hasRecursion;
699 536500 : if (hasRecursion)
700 1010 : root->wt_param_id = assign_special_exec_param(root);
701 : else
702 535490 : root->wt_param_id = -1;
703 536500 : root->non_recursive_path = NULL;
704 536500 : root->partColsUpdated = false;
705 :
706 : /*
707 : * Create the top-level join domain. This won't have valid contents until
708 : * deconstruct_jointree fills it in, but the node needs to exist before
709 : * that so we can build EquivalenceClasses referencing it.
710 : */
711 536500 : root->join_domains = list_make1(makeNode(JoinDomain));
712 :
713 : /*
714 : * If there is a WITH list, process each WITH query and either convert it
715 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
716 : */
717 536500 : if (parse->cteList)
718 2882 : SS_process_ctes(root);
719 :
720 : /*
721 : * If it's a MERGE command, transform the joinlist as appropriate.
722 : */
723 536494 : transform_MERGE_to_join(parse);
724 :
725 : /*
726 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
727 : * that we don't need so many special cases to deal with that situation.
728 : */
729 536494 : replace_empty_jointree(parse);
730 :
731 : /*
732 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
733 : * to transform them into joins. Note that this step does not descend
734 : * into subqueries; if we pull up any subqueries below, their SubLinks are
735 : * processed just before pulling them up.
736 : */
737 536494 : if (parse->hasSubLinks)
738 37816 : pull_up_sublinks(root);
739 :
740 : /*
741 : * Scan the rangetable for function RTEs, do const-simplification on them,
742 : * and then inline them if possible (producing subqueries that might get
743 : * pulled up next). Recursion issues here are handled in the same way as
744 : * for SubLinks.
745 : */
746 536494 : preprocess_function_rtes(root);
747 :
748 : /*
749 : * Scan the rangetable for relations with virtual generated columns, and
750 : * replace all Var nodes in the query that reference these columns with
751 : * the generation expressions. Recursion issues here are handled in the
752 : * same way as for SubLinks.
753 : */
754 536488 : parse = root->parse = expand_virtual_generated_columns(root);
755 :
756 : /*
757 : * Check to see if any subqueries in the jointree can be merged into this
758 : * query.
759 : */
760 536488 : pull_up_subqueries(root);
761 :
762 : /*
763 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
764 : * do this now because it requires applying pull_up_subqueries to the leaf
765 : * queries of the UNION ALL, which weren't touched above because they
766 : * weren't referenced by the jointree (they will be after we do this).
767 : */
768 536482 : if (parse->setOperations)
769 6602 : flatten_simple_union_all(root);
770 :
771 : /*
772 : * Survey the rangetable to see what kinds of entries are present. We can
773 : * skip some later processing if relevant SQL features are not used; for
774 : * example if there are no JOIN RTEs we can avoid the expense of doing
775 : * flatten_join_alias_vars(). This must be done after we have finished
776 : * adding rangetable entries, of course. (Note: actually, processing of
777 : * inherited or partitioned rels can cause RTEs for their child tables to
778 : * get added later; but those must all be RTE_RELATION entries, so they
779 : * don't invalidate the conclusions drawn here.)
780 : */
781 536482 : root->hasJoinRTEs = false;
782 536482 : root->hasLateralRTEs = false;
783 536482 : root->group_rtindex = 0;
784 536482 : hasOuterJoins = false;
785 536482 : hasResultRTEs = false;
786 1450016 : foreach(l, parse->rtable)
787 : {
788 913534 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
789 :
790 913534 : switch (rte->rtekind)
791 : {
792 477574 : case RTE_RELATION:
793 477574 : if (rte->inh)
794 : {
795 : /*
796 : * Check to see if the relation actually has any children;
797 : * if not, clear the inh flag so we can treat it as a
798 : * plain base relation.
799 : *
800 : * Note: this could give a false-positive result, if the
801 : * rel once had children but no longer does. We used to
802 : * be able to clear rte->inh later on when we discovered
803 : * that, but no more; we have to handle such cases as
804 : * full-fledged inheritance.
805 : */
806 391174 : rte->inh = has_subclass(rte->relid);
807 : }
808 477574 : break;
809 88780 : case RTE_JOIN:
810 88780 : root->hasJoinRTEs = true;
811 88780 : if (IS_OUTER_JOIN(rte->jointype))
812 48780 : hasOuterJoins = true;
813 88780 : break;
814 205284 : case RTE_RESULT:
815 205284 : hasResultRTEs = true;
816 205284 : break;
817 4468 : case RTE_GROUP:
818 : Assert(parse->hasGroupRTE);
819 4468 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
820 4468 : break;
821 137428 : default:
822 : /* No work here for other RTE types */
823 137428 : break;
824 : }
825 :
826 913534 : if (rte->lateral)
827 10502 : root->hasLateralRTEs = true;
828 :
829 : /*
830 : * We can also determine the maximum security level required for any
831 : * securityQuals now. Addition of inheritance-child RTEs won't affect
832 : * this, because child tables don't have their own securityQuals; see
833 : * expand_single_inheritance_child().
834 : */
835 913534 : if (rte->securityQuals)
836 2484 : root->qual_security_level = Max(root->qual_security_level,
837 : list_length(rte->securityQuals));
838 : }
839 :
840 : /*
841 : * If we have now verified that the query target relation is
842 : * non-inheriting, mark it as a leaf target.
843 : */
844 536482 : if (parse->resultRelation)
845 : {
846 92742 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
847 :
848 92742 : if (!rte->inh)
849 89926 : root->leaf_result_relids =
850 89926 : bms_make_singleton(parse->resultRelation);
851 : }
852 :
853 : /*
854 : * Preprocess RowMark information. We need to do this after subquery
855 : * pullup, so that all base relations are present.
856 : */
857 536482 : preprocess_rowmarks(root);
858 :
859 : /*
860 : * Set hasHavingQual to remember if HAVING clause is present. Needed
861 : * because preprocess_expression will reduce a constant-true condition to
862 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
863 : */
864 536482 : root->hasHavingQual = (parse->havingQual != NULL);
865 :
866 : /*
867 : * Do expression preprocessing on targetlist and quals, as well as other
868 : * random expressions in the querytree. Note that we do not need to
869 : * handle sort/group expressions explicitly, because they are actually
870 : * part of the targetlist.
871 : */
872 532706 : parse->targetList = (List *)
873 536482 : preprocess_expression(root, (Node *) parse->targetList,
874 : EXPRKIND_TARGET);
875 :
876 532706 : newWithCheckOptions = NIL;
877 535122 : foreach(l, parse->withCheckOptions)
878 : {
879 2416 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
880 :
881 2416 : wco->qual = preprocess_expression(root, wco->qual,
882 : EXPRKIND_QUAL);
883 2416 : if (wco->qual != NULL)
884 2016 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
885 : }
886 532706 : parse->withCheckOptions = newWithCheckOptions;
887 :
888 532706 : parse->returningList = (List *)
889 532706 : preprocess_expression(root, (Node *) parse->returningList,
890 : EXPRKIND_TARGET);
891 :
892 532706 : preprocess_qual_conditions(root, (Node *) parse->jointree);
893 :
894 532706 : parse->havingQual = preprocess_expression(root, parse->havingQual,
895 : EXPRKIND_QUAL);
896 :
897 535324 : foreach(l, parse->windowClause)
898 : {
899 2618 : WindowClause *wc = lfirst_node(WindowClause, l);
900 :
901 : /* partitionClause/orderClause are sort/group expressions */
902 2618 : wc->startOffset = preprocess_expression(root, wc->startOffset,
903 : EXPRKIND_LIMIT);
904 2618 : wc->endOffset = preprocess_expression(root, wc->endOffset,
905 : EXPRKIND_LIMIT);
906 : }
907 :
908 532706 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
909 : EXPRKIND_LIMIT);
910 532706 : parse->limitCount = preprocess_expression(root, parse->limitCount,
911 : EXPRKIND_LIMIT);
912 :
913 532706 : if (parse->onConflict)
914 : {
915 3628 : parse->onConflict->arbiterElems = (List *)
916 1814 : preprocess_expression(root,
917 1814 : (Node *) parse->onConflict->arbiterElems,
918 : EXPRKIND_ARBITER_ELEM);
919 3628 : parse->onConflict->arbiterWhere =
920 1814 : preprocess_expression(root,
921 1814 : parse->onConflict->arbiterWhere,
922 : EXPRKIND_QUAL);
923 3628 : parse->onConflict->onConflictSet = (List *)
924 1814 : preprocess_expression(root,
925 1814 : (Node *) parse->onConflict->onConflictSet,
926 : EXPRKIND_TARGET);
927 1814 : parse->onConflict->onConflictWhere =
928 1814 : preprocess_expression(root,
929 1814 : parse->onConflict->onConflictWhere,
930 : EXPRKIND_QUAL);
931 : /* exclRelTlist contains only Vars, so no preprocessing needed */
932 : }
933 :
934 535520 : foreach(l, parse->mergeActionList)
935 : {
936 2814 : MergeAction *action = (MergeAction *) lfirst(l);
937 :
938 2814 : action->targetList = (List *)
939 2814 : preprocess_expression(root,
940 2814 : (Node *) action->targetList,
941 : EXPRKIND_TARGET);
942 2814 : action->qual =
943 2814 : preprocess_expression(root,
944 : (Node *) action->qual,
945 : EXPRKIND_QUAL);
946 : }
947 :
948 532706 : parse->mergeJoinCondition =
949 532706 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
950 :
951 532706 : root->append_rel_list = (List *)
952 532706 : preprocess_expression(root, (Node *) root->append_rel_list,
953 : EXPRKIND_APPINFO);
954 :
955 : /* Also need to preprocess expressions within RTEs */
956 1442186 : foreach(l, parse->rtable)
957 : {
958 909480 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
959 : int kind;
960 : ListCell *lcsq;
961 :
962 909480 : if (rte->rtekind == RTE_RELATION)
963 : {
964 477308 : if (rte->tablesample)
965 228 : rte->tablesample = (TableSampleClause *)
966 228 : preprocess_expression(root,
967 228 : (Node *) rte->tablesample,
968 : EXPRKIND_TABLESAMPLE);
969 : }
970 432172 : else if (rte->rtekind == RTE_SUBQUERY)
971 : {
972 : /*
973 : * We don't want to do all preprocessing yet on the subquery's
974 : * expressions, since that will happen when we plan it. But if it
975 : * contains any join aliases of our level, those have to get
976 : * expanded now, because planning of the subquery won't do it.
977 : * That's only possible if the subquery is LATERAL.
978 : */
979 68348 : if (rte->lateral && root->hasJoinRTEs)
980 1228 : rte->subquery = (Query *)
981 1228 : flatten_join_alias_vars(root, root->parse,
982 1228 : (Node *) rte->subquery);
983 : }
984 363824 : else if (rte->rtekind == RTE_FUNCTION)
985 : {
986 : /* Preprocess the function expression(s) fully */
987 54596 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
988 54596 : rte->functions = (List *)
989 54596 : preprocess_expression(root, (Node *) rte->functions, kind);
990 : }
991 309228 : else if (rte->rtekind == RTE_TABLEFUNC)
992 : {
993 : /* Preprocess the function expression(s) fully */
994 626 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
995 626 : rte->tablefunc = (TableFunc *)
996 626 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
997 : }
998 308602 : else if (rte->rtekind == RTE_VALUES)
999 : {
1000 : /* Preprocess the values lists fully */
1001 8266 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1002 8266 : rte->values_lists = (List *)
1003 8266 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1004 : }
1005 300336 : else if (rte->rtekind == RTE_GROUP)
1006 : {
1007 : /* Preprocess the groupexprs list fully */
1008 4468 : rte->groupexprs = (List *)
1009 4468 : preprocess_expression(root, (Node *) rte->groupexprs,
1010 : EXPRKIND_GROUPEXPR);
1011 : }
1012 :
1013 : /*
1014 : * Process each element of the securityQuals list as if it were a
1015 : * separate qual expression (as indeed it is). We need to do it this
1016 : * way to get proper canonicalization of AND/OR structure. Note that
1017 : * this converts each element into an implicit-AND sublist.
1018 : */
1019 912322 : foreach(lcsq, rte->securityQuals)
1020 : {
1021 2842 : lfirst(lcsq) = preprocess_expression(root,
1022 2842 : (Node *) lfirst(lcsq),
1023 : EXPRKIND_QUAL);
1024 : }
1025 : }
1026 :
1027 : /*
1028 : * Now that we are done preprocessing expressions, and in particular done
1029 : * flattening join alias variables, get rid of the joinaliasvars lists.
1030 : * They no longer match what expressions in the rest of the tree look
1031 : * like, because we have not preprocessed expressions in those lists (and
1032 : * do not want to; for example, expanding a SubLink there would result in
1033 : * a useless unreferenced subplan). Leaving them in place simply creates
1034 : * a hazard for later scans of the tree. We could try to prevent that by
1035 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1036 : * but that doesn't sound very reliable.
1037 : */
1038 532706 : if (root->hasJoinRTEs)
1039 : {
1040 306462 : foreach(l, parse->rtable)
1041 : {
1042 252666 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1043 :
1044 252666 : rte->joinaliasvars = NIL;
1045 : }
1046 : }
1047 :
1048 : /*
1049 : * Replace any Vars in the subquery's targetlist and havingQual that
1050 : * reference GROUP outputs with the underlying grouping expressions.
1051 : *
1052 : * Note that we need to perform this replacement after we've preprocessed
1053 : * the grouping expressions. This is to ensure that there is only one
1054 : * instance of SubPlan for each SubLink contained within the grouping
1055 : * expressions.
1056 : */
1057 532706 : if (parse->hasGroupRTE)
1058 : {
1059 4468 : parse->targetList = (List *)
1060 4468 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1061 4468 : parse->havingQual =
1062 4468 : flatten_group_exprs(root, root->parse, parse->havingQual);
1063 : }
1064 :
1065 : /* Constant-folding might have removed all set-returning functions */
1066 532706 : if (parse->hasTargetSRFs)
1067 12020 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1068 :
1069 : /*
1070 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1071 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1072 : * volatile functions (since a HAVING clause is supposed to be executed
1073 : * only once per group). We also can't do this if there are any nonempty
1074 : * grouping sets and the clause references any columns that are nullable
1075 : * by the grouping sets; moving such a clause into WHERE would potentially
1076 : * change the results. (If there are only empty grouping sets, then the
1077 : * HAVING clause must be degenerate as discussed below.)
1078 : *
1079 : * Also, it may be that the clause is so expensive to execute that we're
1080 : * better off doing it only once per group, despite the loss of
1081 : * selectivity. This is hard to estimate short of doing the entire
1082 : * planning process twice, so we use a heuristic: clauses containing
1083 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1084 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1085 : * instead of after.
1086 : *
1087 : * If the query has explicit grouping then we can simply move such a
1088 : * clause into WHERE; any group that fails the clause will not be in the
1089 : * output because none of its tuples will reach the grouping or
1090 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
1091 : * HAVING clause, which we put in WHERE so that query_planner() can use it
1092 : * in a gating Result node, but also keep in HAVING to ensure that we
1093 : * don't emit a bogus aggregated row. (This could be done better, but it
1094 : * seems not worth optimizing.)
1095 : *
1096 : * Note that a HAVING clause may contain expressions that are not fully
1097 : * preprocessed. This can happen if these expressions are part of
1098 : * grouping items. In such cases, they are replaced with GROUP Vars in
1099 : * the parser and then replaced back after we've done with expression
1100 : * preprocessing on havingQual. This is not an issue if the clause
1101 : * remains in HAVING, because these expressions will be matched to lower
1102 : * target items in setrefs.c. However, if the clause is moved or copied
1103 : * into WHERE, we need to ensure that these expressions are fully
1104 : * preprocessed.
1105 : *
1106 : * Note that both havingQual and parse->jointree->quals are in
1107 : * implicitly-ANDed-list form at this point, even though they are declared
1108 : * as Node *.
1109 : */
1110 532706 : newHaving = NIL;
1111 533862 : foreach(l, (List *) parse->havingQual)
1112 : {
1113 1156 : Node *havingclause = (Node *) lfirst(l);
1114 :
1115 1490 : if (contain_agg_clause(havingclause) ||
1116 668 : contain_volatile_functions(havingclause) ||
1117 334 : contain_subplans(havingclause) ||
1118 418 : (parse->groupClause && parse->groupingSets &&
1119 84 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1120 : {
1121 : /* keep it in HAVING */
1122 894 : newHaving = lappend(newHaving, havingclause);
1123 : }
1124 262 : else if (parse->groupClause)
1125 : {
1126 : Node *whereclause;
1127 :
1128 : /* Preprocess the HAVING clause fully */
1129 244 : whereclause = preprocess_expression(root, havingclause,
1130 : EXPRKIND_QUAL);
1131 : /* ... and move it to WHERE */
1132 244 : parse->jointree->quals = (Node *)
1133 244 : list_concat((List *) parse->jointree->quals,
1134 : (List *) whereclause);
1135 : }
1136 : else
1137 : {
1138 : Node *whereclause;
1139 :
1140 : /* Preprocess the HAVING clause fully */
1141 18 : whereclause = preprocess_expression(root, copyObject(havingclause),
1142 : EXPRKIND_QUAL);
1143 : /* ... and put a copy in WHERE */
1144 36 : parse->jointree->quals = (Node *)
1145 18 : list_concat((List *) parse->jointree->quals,
1146 : (List *) whereclause);
1147 : /* ... and also keep it in HAVING */
1148 18 : newHaving = lappend(newHaving, havingclause);
1149 : }
1150 : }
1151 532706 : parse->havingQual = (Node *) newHaving;
1152 :
1153 : /*
1154 : * If we have any outer joins, try to reduce them to plain inner joins.
1155 : * This step is most easily done after we've done expression
1156 : * preprocessing.
1157 : */
1158 532706 : if (hasOuterJoins)
1159 34476 : reduce_outer_joins(root);
1160 :
1161 : /*
1162 : * If we have any RTE_RESULT relations, see if they can be deleted from
1163 : * the jointree. We also rely on this processing to flatten single-child
1164 : * FromExprs underneath outer joins. This step is most effectively done
1165 : * after we've done expression preprocessing and outer join reduction.
1166 : */
1167 532706 : if (hasResultRTEs || hasOuterJoins)
1168 234684 : remove_useless_result_rtes(root);
1169 :
1170 : /*
1171 : * Do the main planning.
1172 : */
1173 532706 : grouping_planner(root, tuple_fraction, setops);
1174 :
1175 : /*
1176 : * Capture the set of outer-level param IDs we have access to, for use in
1177 : * extParam/allParam calculations later.
1178 : */
1179 532634 : SS_identify_outer_params(root);
1180 :
1181 : /*
1182 : * If any initPlans were created in this query level, adjust the surviving
1183 : * Paths' costs and parallel-safety flags to account for them. The
1184 : * initPlans won't actually get attached to the plan tree till
1185 : * create_plan() runs, but we must include their effects now.
1186 : */
1187 532634 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1188 532634 : SS_charge_for_initplans(root, final_rel);
1189 :
1190 : /*
1191 : * Make sure we've identified the cheapest Path for the final rel. (By
1192 : * doing this here not in grouping_planner, we include initPlan costs in
1193 : * the decision, though it's unlikely that will change anything.)
1194 : */
1195 532634 : set_cheapest(final_rel);
1196 :
1197 532634 : return root;
1198 : }
1199 :
1200 : /*
1201 : * preprocess_expression
1202 : * Do subquery_planner's preprocessing work for an expression,
1203 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1204 : * conditions), a HAVING clause, or a few other things.
1205 : */
1206 : static Node *
1207 4468538 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1208 : {
1209 : /*
1210 : * Fall out quickly if expression is empty. This occurs often enough to
1211 : * be worth checking. Note that null->null is the correct conversion for
1212 : * implicit-AND result format, too.
1213 : */
1214 4468538 : if (expr == NULL)
1215 3532384 : return NULL;
1216 :
1217 : /*
1218 : * If the query has any join RTEs, replace join alias variables with
1219 : * base-relation variables. We must do this first, since any expressions
1220 : * we may extract from the joinaliasvars lists have not been preprocessed.
1221 : * For example, if we did this after sublink processing, sublinks expanded
1222 : * out from join aliases would not get processed. But we can skip this in
1223 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1224 : * they can't contain any Vars of the current query level.
1225 : */
1226 936154 : if (root->hasJoinRTEs &&
1227 387864 : !(kind == EXPRKIND_RTFUNC ||
1228 193758 : kind == EXPRKIND_VALUES ||
1229 : kind == EXPRKIND_TABLESAMPLE ||
1230 : kind == EXPRKIND_TABLEFUNC))
1231 193740 : expr = flatten_join_alias_vars(root, root->parse, expr);
1232 :
1233 : /*
1234 : * Simplify constant expressions. For function RTEs, this was already
1235 : * done by preprocess_function_rtes. (But note we must do it again for
1236 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1237 : * un-simplified subexpressions inserted by flattening of subqueries or
1238 : * join alias variables.)
1239 : *
1240 : * Note: an essential effect of this is to convert named-argument function
1241 : * calls to positional notation and insert the current actual values of
1242 : * any default arguments for functions. To ensure that happens, we *must*
1243 : * process all expressions here. Previous PG versions sometimes skipped
1244 : * const-simplification if it didn't seem worth the trouble, but we can't
1245 : * do that anymore.
1246 : *
1247 : * Note: this also flattens nested AND and OR expressions into N-argument
1248 : * form. All processing of a qual expression after this point must be
1249 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1250 : * with AND directly under AND, nor OR directly under OR.
1251 : */
1252 936154 : if (kind != EXPRKIND_RTFUNC)
1253 889920 : expr = eval_const_expressions(root, expr);
1254 :
1255 : /*
1256 : * If it's a qual or havingQual, canonicalize it.
1257 : */
1258 932378 : if (kind == EXPRKIND_QUAL)
1259 : {
1260 331268 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1261 :
1262 : #ifdef OPTIMIZER_DEBUG
1263 : printf("After canonicalize_qual()\n");
1264 : pprint(expr);
1265 : #endif
1266 : }
1267 :
1268 : /*
1269 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1270 : * hashfuncid of any that might execute more quickly by using hash lookups
1271 : * instead of a linear search.
1272 : */
1273 932378 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1274 : {
1275 851106 : convert_saop_to_hashed_saop(expr);
1276 : }
1277 :
1278 : /* Expand SubLinks to SubPlans */
1279 932378 : if (root->parse->hasSubLinks)
1280 108400 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1281 :
1282 : /*
1283 : * XXX do not insert anything here unless you have grokked the comments in
1284 : * SS_replace_correlation_vars ...
1285 : */
1286 :
1287 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1288 932378 : if (root->query_level > 1)
1289 161904 : expr = SS_replace_correlation_vars(root, expr);
1290 :
1291 : /*
1292 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1293 : * don't want to do this before eval_const_expressions, since the latter
1294 : * would be unable to simplify a top-level AND correctly. Also,
1295 : * SS_process_sublinks expects explicit-AND format.)
1296 : */
1297 932378 : if (kind == EXPRKIND_QUAL)
1298 331268 : expr = (Node *) make_ands_implicit((Expr *) expr);
1299 :
1300 932378 : return expr;
1301 : }
1302 :
1303 : /*
1304 : * preprocess_qual_conditions
1305 : * Recursively scan the query's jointree and do subquery_planner's
1306 : * preprocessing work on each qual condition found therein.
1307 : */
1308 : static void
1309 1317236 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1310 : {
1311 1317236 : if (jtnode == NULL)
1312 0 : return;
1313 1317236 : if (IsA(jtnode, RangeTblRef))
1314 : {
1315 : /* nothing to do here */
1316 : }
1317 643924 : else if (IsA(jtnode, FromExpr))
1318 : {
1319 547376 : FromExpr *f = (FromExpr *) jtnode;
1320 : ListCell *l;
1321 :
1322 1138810 : foreach(l, f->fromlist)
1323 591434 : preprocess_qual_conditions(root, lfirst(l));
1324 :
1325 547376 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1326 : }
1327 96548 : else if (IsA(jtnode, JoinExpr))
1328 : {
1329 96548 : JoinExpr *j = (JoinExpr *) jtnode;
1330 :
1331 96548 : preprocess_qual_conditions(root, j->larg);
1332 96548 : preprocess_qual_conditions(root, j->rarg);
1333 :
1334 96548 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1335 : }
1336 : else
1337 0 : elog(ERROR, "unrecognized node type: %d",
1338 : (int) nodeTag(jtnode));
1339 : }
1340 :
1341 : /*
1342 : * preprocess_phv_expression
1343 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1344 : *
1345 : * If a LATERAL subquery references an output of another subquery, and that
1346 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1347 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1348 : * and later pull it back up during find_lateral_references, which runs after
1349 : * subquery_planner has preprocessed all the expressions that were in the
1350 : * current query level to start with. So we need to preprocess it then.
1351 : */
1352 : Expr *
1353 72 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1354 : {
1355 72 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1356 : }
1357 :
1358 : /*--------------------
1359 : * grouping_planner
1360 : * Perform planning steps related to grouping, aggregation, etc.
1361 : *
1362 : * This function adds all required top-level processing to the scan/join
1363 : * Path(s) produced by query_planner.
1364 : *
1365 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1366 : * tuple_fraction is interpreted as follows:
1367 : * 0: expect all tuples to be retrieved (normal case)
1368 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1369 : * from the plan to be retrieved
1370 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1371 : * expected to be retrieved (ie, a LIMIT specification).
1372 : * setops is used for set operation subqueries to provide the subquery with
1373 : * the context in which it's being used so that Paths correctly sorted for the
1374 : * set operation can be generated. NULL when not planning a set operation
1375 : * child, or when a child of a set op that isn't interested in sorted input.
1376 : *
1377 : * Returns nothing; the useful output is in the Paths we attach to the
1378 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1379 : * root->processed_tlist contains the final processed targetlist.
1380 : *
1381 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1382 : * to leave this to the caller.
1383 : *--------------------
1384 : */
1385 : static void
1386 532706 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1387 : SetOperationStmt *setops)
1388 : {
1389 532706 : Query *parse = root->parse;
1390 532706 : int64 offset_est = 0;
1391 532706 : int64 count_est = 0;
1392 532706 : double limit_tuples = -1.0;
1393 532706 : bool have_postponed_srfs = false;
1394 : PathTarget *final_target;
1395 : List *final_targets;
1396 : List *final_targets_contain_srfs;
1397 : bool final_target_parallel_safe;
1398 : RelOptInfo *current_rel;
1399 : RelOptInfo *final_rel;
1400 : FinalPathExtraData extra;
1401 : ListCell *lc;
1402 :
1403 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1404 532706 : if (parse->limitCount || parse->limitOffset)
1405 : {
1406 4950 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1407 : &offset_est, &count_est);
1408 :
1409 : /*
1410 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1411 : * estimate the effects of using a bounded sort.
1412 : */
1413 4950 : if (count_est > 0 && offset_est >= 0)
1414 4428 : limit_tuples = (double) count_est + (double) offset_est;
1415 : }
1416 :
1417 : /* Make tuple_fraction accessible to lower-level routines */
1418 532706 : root->tuple_fraction = tuple_fraction;
1419 :
1420 532706 : if (parse->setOperations)
1421 : {
1422 : /*
1423 : * Construct Paths for set operations. The results will not need any
1424 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1425 : * special work for recursive unions is the responsibility of
1426 : * plan_set_operations.
1427 : */
1428 6166 : current_rel = plan_set_operations(root);
1429 :
1430 : /*
1431 : * We should not need to call preprocess_targetlist, since we must be
1432 : * in a SELECT query node. Instead, use the processed_tlist returned
1433 : * by plan_set_operations (since this tells whether it returned any
1434 : * resjunk columns!), and transfer any sort key information from the
1435 : * original tlist.
1436 : */
1437 : Assert(parse->commandType == CMD_SELECT);
1438 :
1439 : /* for safety, copy processed_tlist instead of modifying in-place */
1440 6160 : root->processed_tlist =
1441 6160 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1442 : parse->targetList);
1443 :
1444 : /* Also extract the PathTarget form of the setop result tlist */
1445 6160 : final_target = current_rel->cheapest_total_path->pathtarget;
1446 :
1447 : /* And check whether it's parallel safe */
1448 : final_target_parallel_safe =
1449 6160 : is_parallel_safe(root, (Node *) final_target->exprs);
1450 :
1451 : /* The setop result tlist couldn't contain any SRFs */
1452 : Assert(!parse->hasTargetSRFs);
1453 6160 : final_targets = final_targets_contain_srfs = NIL;
1454 :
1455 : /*
1456 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1457 : * checked already, but let's make sure).
1458 : */
1459 6160 : if (parse->rowMarks)
1460 0 : ereport(ERROR,
1461 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1462 : /*------
1463 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1464 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1465 : LCS_asString(linitial_node(RowMarkClause,
1466 : parse->rowMarks)->strength))));
1467 :
1468 : /*
1469 : * Calculate pathkeys that represent result ordering requirements
1470 : */
1471 : Assert(parse->distinctClause == NIL);
1472 6160 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1473 : parse->sortClause,
1474 : root->processed_tlist);
1475 : }
1476 : else
1477 : {
1478 : /* No set operations, do regular planning */
1479 : PathTarget *sort_input_target;
1480 : List *sort_input_targets;
1481 : List *sort_input_targets_contain_srfs;
1482 : bool sort_input_target_parallel_safe;
1483 : PathTarget *grouping_target;
1484 : List *grouping_targets;
1485 : List *grouping_targets_contain_srfs;
1486 : bool grouping_target_parallel_safe;
1487 : PathTarget *scanjoin_target;
1488 : List *scanjoin_targets;
1489 : List *scanjoin_targets_contain_srfs;
1490 : bool scanjoin_target_parallel_safe;
1491 : bool scanjoin_target_same_exprs;
1492 : bool have_grouping;
1493 526540 : WindowFuncLists *wflists = NULL;
1494 526540 : List *activeWindows = NIL;
1495 526540 : grouping_sets_data *gset_data = NULL;
1496 : standard_qp_extra qp_extra;
1497 :
1498 : /* A recursive query should always have setOperations */
1499 : Assert(!root->hasRecursion);
1500 :
1501 : /* Preprocess grouping sets and GROUP BY clause, if any */
1502 526540 : if (parse->groupingSets)
1503 : {
1504 878 : gset_data = preprocess_grouping_sets(root);
1505 : }
1506 525662 : else if (parse->groupClause)
1507 : {
1508 : /* Preprocess regular GROUP BY clause, if any */
1509 3632 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1510 : }
1511 :
1512 : /*
1513 : * Preprocess targetlist. Note that much of the remaining planning
1514 : * work will be done with the PathTarget representation of tlists, but
1515 : * we must also maintain the full representation of the final tlist so
1516 : * that we can transfer its decoration (resnames etc) to the topmost
1517 : * tlist of the finished Plan. This is kept in processed_tlist.
1518 : */
1519 526534 : preprocess_targetlist(root);
1520 :
1521 : /*
1522 : * Mark all the aggregates with resolved aggtranstypes, and detect
1523 : * aggregates that are duplicates or can share transition state. We
1524 : * must do this before slicing and dicing the tlist into various
1525 : * pathtargets, else some copies of the Aggref nodes might escape
1526 : * being marked.
1527 : */
1528 526534 : if (parse->hasAggs)
1529 : {
1530 41048 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1531 41048 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1532 : }
1533 :
1534 : /*
1535 : * Locate any window functions in the tlist. (We don't need to look
1536 : * anywhere else, since expressions used in ORDER BY will be in there
1537 : * too.) Note that they could all have been eliminated by constant
1538 : * folding, in which case we don't need to do any more work.
1539 : */
1540 526534 : if (parse->hasWindowFuncs)
1541 : {
1542 2384 : wflists = find_window_functions((Node *) root->processed_tlist,
1543 2384 : list_length(parse->windowClause));
1544 2384 : if (wflists->numWindowFuncs > 0)
1545 : {
1546 : /*
1547 : * See if any modifications can be made to each WindowClause
1548 : * to allow the executor to execute the WindowFuncs more
1549 : * quickly.
1550 : */
1551 2378 : optimize_window_clauses(root, wflists);
1552 :
1553 : /* Extract the list of windows actually in use. */
1554 2378 : activeWindows = select_active_windows(root, wflists);
1555 :
1556 : /* Make sure they all have names, for EXPLAIN's use. */
1557 2378 : name_active_windows(activeWindows);
1558 : }
1559 : else
1560 6 : parse->hasWindowFuncs = false;
1561 : }
1562 :
1563 : /*
1564 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1565 : * adding logic between here and the query_planner() call. Anything
1566 : * that is needed in MIN/MAX-optimizable cases will have to be
1567 : * duplicated in planagg.c.
1568 : */
1569 526534 : if (parse->hasAggs)
1570 41048 : preprocess_minmax_aggregates(root);
1571 :
1572 : /*
1573 : * Figure out whether there's a hard limit on the number of rows that
1574 : * query_planner's result subplan needs to return. Even if we know a
1575 : * hard limit overall, it doesn't apply if the query has any
1576 : * grouping/aggregation operations, or SRFs in the tlist.
1577 : */
1578 526534 : if (parse->groupClause ||
1579 522072 : parse->groupingSets ||
1580 522030 : parse->distinctClause ||
1581 519388 : parse->hasAggs ||
1582 482232 : parse->hasWindowFuncs ||
1583 479992 : parse->hasTargetSRFs ||
1584 468434 : root->hasHavingQual)
1585 58118 : root->limit_tuples = -1.0;
1586 : else
1587 468416 : root->limit_tuples = limit_tuples;
1588 :
1589 : /* Set up data needed by standard_qp_callback */
1590 526534 : qp_extra.activeWindows = activeWindows;
1591 526534 : qp_extra.gset_data = gset_data;
1592 :
1593 : /*
1594 : * If we're a subquery for a set operation, store the SetOperationStmt
1595 : * in qp_extra.
1596 : */
1597 526534 : qp_extra.setop = setops;
1598 :
1599 : /*
1600 : * Generate the best unsorted and presorted paths for the scan/join
1601 : * portion of this Query, ie the processing represented by the
1602 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1603 : * We also generate (in standard_qp_callback) pathkey representations
1604 : * of the query's sort clause, distinct clause, etc.
1605 : */
1606 526534 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1607 :
1608 : /*
1609 : * Convert the query's result tlist into PathTarget format.
1610 : *
1611 : * Note: this cannot be done before query_planner() has performed
1612 : * appendrel expansion, because that might add resjunk entries to
1613 : * root->processed_tlist. Waiting till afterwards is also helpful
1614 : * because the target width estimates can use per-Var width numbers
1615 : * that were obtained within query_planner().
1616 : */
1617 526480 : final_target = create_pathtarget(root, root->processed_tlist);
1618 : final_target_parallel_safe =
1619 526480 : is_parallel_safe(root, (Node *) final_target->exprs);
1620 :
1621 : /*
1622 : * If ORDER BY was given, consider whether we should use a post-sort
1623 : * projection, and compute the adjusted target for preceding steps if
1624 : * so.
1625 : */
1626 526480 : if (parse->sortClause)
1627 : {
1628 71966 : sort_input_target = make_sort_input_target(root,
1629 : final_target,
1630 : &have_postponed_srfs);
1631 : sort_input_target_parallel_safe =
1632 71966 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1633 : }
1634 : else
1635 : {
1636 454514 : sort_input_target = final_target;
1637 454514 : sort_input_target_parallel_safe = final_target_parallel_safe;
1638 : }
1639 :
1640 : /*
1641 : * If we have window functions to deal with, the output from any
1642 : * grouping step needs to be what the window functions want;
1643 : * otherwise, it should be sort_input_target.
1644 : */
1645 526480 : if (activeWindows)
1646 : {
1647 2378 : grouping_target = make_window_input_target(root,
1648 : final_target,
1649 : activeWindows);
1650 : grouping_target_parallel_safe =
1651 2378 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1652 : }
1653 : else
1654 : {
1655 524102 : grouping_target = sort_input_target;
1656 524102 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1657 : }
1658 :
1659 : /*
1660 : * If we have grouping or aggregation to do, the topmost scan/join
1661 : * plan node must emit what the grouping step wants; otherwise, it
1662 : * should emit grouping_target.
1663 : */
1664 522018 : have_grouping = (parse->groupClause || parse->groupingSets ||
1665 1048498 : parse->hasAggs || root->hasHavingQual);
1666 526480 : if (have_grouping)
1667 : {
1668 41716 : scanjoin_target = make_group_input_target(root, final_target);
1669 : scanjoin_target_parallel_safe =
1670 41716 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1671 : }
1672 : else
1673 : {
1674 484764 : scanjoin_target = grouping_target;
1675 484764 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1676 : }
1677 :
1678 : /*
1679 : * If there are any SRFs in the targetlist, we must separate each of
1680 : * these PathTargets into SRF-computing and SRF-free targets. Replace
1681 : * each of the named targets with a SRF-free version, and remember the
1682 : * list of additional projection steps we need to add afterwards.
1683 : */
1684 526480 : if (parse->hasTargetSRFs)
1685 : {
1686 : /* final_target doesn't recompute any SRFs in sort_input_target */
1687 12020 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1688 : &final_targets,
1689 : &final_targets_contain_srfs);
1690 12020 : final_target = linitial_node(PathTarget, final_targets);
1691 : Assert(!linitial_int(final_targets_contain_srfs));
1692 : /* likewise for sort_input_target vs. grouping_target */
1693 12020 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1694 : &sort_input_targets,
1695 : &sort_input_targets_contain_srfs);
1696 12020 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
1697 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1698 : /* likewise for grouping_target vs. scanjoin_target */
1699 12020 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
1700 : &grouping_targets,
1701 : &grouping_targets_contain_srfs);
1702 12020 : grouping_target = linitial_node(PathTarget, grouping_targets);
1703 : Assert(!linitial_int(grouping_targets_contain_srfs));
1704 : /* scanjoin_target will not have any SRFs precomputed for it */
1705 12020 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1706 : &scanjoin_targets,
1707 : &scanjoin_targets_contain_srfs);
1708 12020 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1709 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1710 : }
1711 : else
1712 : {
1713 : /* initialize lists; for most of these, dummy values are OK */
1714 514460 : final_targets = final_targets_contain_srfs = NIL;
1715 514460 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1716 514460 : grouping_targets = grouping_targets_contain_srfs = NIL;
1717 514460 : scanjoin_targets = list_make1(scanjoin_target);
1718 514460 : scanjoin_targets_contain_srfs = NIL;
1719 : }
1720 :
1721 : /* Apply scan/join target. */
1722 526480 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1723 526480 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1724 526480 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1725 : scanjoin_targets_contain_srfs,
1726 : scanjoin_target_parallel_safe,
1727 : scanjoin_target_same_exprs);
1728 :
1729 : /*
1730 : * Save the various upper-rel PathTargets we just computed into
1731 : * root->upper_targets[]. The core code doesn't use this, but it
1732 : * provides a convenient place for extensions to get at the info. For
1733 : * consistency, we save all the intermediate targets, even though some
1734 : * of the corresponding upperrels might not be needed for this query.
1735 : */
1736 526480 : root->upper_targets[UPPERREL_FINAL] = final_target;
1737 526480 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1738 526480 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1739 526480 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1740 526480 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1741 526480 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1742 :
1743 : /*
1744 : * If we have grouping and/or aggregation, consider ways to implement
1745 : * that. We build a new upperrel representing the output of this
1746 : * phase.
1747 : */
1748 526480 : if (have_grouping)
1749 : {
1750 41716 : current_rel = create_grouping_paths(root,
1751 : current_rel,
1752 : grouping_target,
1753 : grouping_target_parallel_safe,
1754 : gset_data);
1755 : /* Fix things up if grouping_target contains SRFs */
1756 41710 : if (parse->hasTargetSRFs)
1757 420 : adjust_paths_for_srfs(root, current_rel,
1758 : grouping_targets,
1759 : grouping_targets_contain_srfs);
1760 : }
1761 :
1762 : /*
1763 : * If we have window functions, consider ways to implement those. We
1764 : * build a new upperrel representing the output of this phase.
1765 : */
1766 526474 : if (activeWindows)
1767 : {
1768 2378 : current_rel = create_window_paths(root,
1769 : current_rel,
1770 : grouping_target,
1771 : sort_input_target,
1772 : sort_input_target_parallel_safe,
1773 : wflists,
1774 : activeWindows);
1775 : /* Fix things up if sort_input_target contains SRFs */
1776 2378 : if (parse->hasTargetSRFs)
1777 12 : adjust_paths_for_srfs(root, current_rel,
1778 : sort_input_targets,
1779 : sort_input_targets_contain_srfs);
1780 : }
1781 :
1782 : /*
1783 : * If there is a DISTINCT clause, consider ways to implement that. We
1784 : * build a new upperrel representing the output of this phase.
1785 : */
1786 526474 : if (parse->distinctClause)
1787 : {
1788 2676 : current_rel = create_distinct_paths(root,
1789 : current_rel,
1790 : sort_input_target);
1791 : }
1792 : } /* end of if (setOperations) */
1793 :
1794 : /*
1795 : * If ORDER BY was given, consider ways to implement that, and generate a
1796 : * new upperrel containing only paths that emit the correct ordering and
1797 : * project the correct final_target. We can apply the original
1798 : * limit_tuples limit in sort costing here, but only if there are no
1799 : * postponed SRFs.
1800 : */
1801 532634 : if (parse->sortClause)
1802 : {
1803 75922 : current_rel = create_ordered_paths(root,
1804 : current_rel,
1805 : final_target,
1806 : final_target_parallel_safe,
1807 : have_postponed_srfs ? -1.0 :
1808 : limit_tuples);
1809 : /* Fix things up if final_target contains SRFs */
1810 75922 : if (parse->hasTargetSRFs)
1811 196 : adjust_paths_for_srfs(root, current_rel,
1812 : final_targets,
1813 : final_targets_contain_srfs);
1814 : }
1815 :
1816 : /*
1817 : * Now we are prepared to build the final-output upperrel.
1818 : */
1819 532634 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1820 :
1821 : /*
1822 : * If the input rel is marked consider_parallel and there's nothing that's
1823 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1824 : * consider_parallel as well. Note that if the query has rowMarks or is
1825 : * not a SELECT, consider_parallel will be false for every relation in the
1826 : * query.
1827 : */
1828 710192 : if (current_rel->consider_parallel &&
1829 355092 : is_parallel_safe(root, parse->limitOffset) &&
1830 177534 : is_parallel_safe(root, parse->limitCount))
1831 177528 : final_rel->consider_parallel = true;
1832 :
1833 : /*
1834 : * If the current_rel belongs to a single FDW, so does the final_rel.
1835 : */
1836 532634 : final_rel->serverid = current_rel->serverid;
1837 532634 : final_rel->userid = current_rel->userid;
1838 532634 : final_rel->useridiscurrent = current_rel->useridiscurrent;
1839 532634 : final_rel->fdwroutine = current_rel->fdwroutine;
1840 :
1841 : /*
1842 : * Generate paths for the final_rel. Insert all surviving paths, with
1843 : * LockRows, Limit, and/or ModifyTable steps added if needed.
1844 : */
1845 1084650 : foreach(lc, current_rel->pathlist)
1846 : {
1847 552016 : Path *path = (Path *) lfirst(lc);
1848 :
1849 : /*
1850 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1851 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
1852 : * here. If there are only non-locking rowmarks, they should be
1853 : * handled by the ModifyTable node instead. However, root->rowMarks
1854 : * is what goes into the LockRows node.)
1855 : */
1856 552016 : if (parse->rowMarks)
1857 : {
1858 8162 : path = (Path *) create_lockrows_path(root, final_rel, path,
1859 : root->rowMarks,
1860 : assign_special_exec_param(root));
1861 : }
1862 :
1863 : /*
1864 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1865 : */
1866 552016 : if (limit_needed(parse))
1867 : {
1868 5856 : path = (Path *) create_limit_path(root, final_rel, path,
1869 : parse->limitOffset,
1870 : parse->limitCount,
1871 : parse->limitOption,
1872 : offset_est, count_est);
1873 : }
1874 :
1875 : /*
1876 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
1877 : */
1878 552016 : if (parse->commandType != CMD_SELECT)
1879 : {
1880 : Index rootRelation;
1881 92494 : List *resultRelations = NIL;
1882 92494 : List *updateColnosLists = NIL;
1883 92494 : List *withCheckOptionLists = NIL;
1884 92494 : List *returningLists = NIL;
1885 92494 : List *mergeActionLists = NIL;
1886 92494 : List *mergeJoinConditions = NIL;
1887 : List *rowMarks;
1888 :
1889 92494 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
1890 : {
1891 : /* Inherited UPDATE/DELETE/MERGE */
1892 2786 : RelOptInfo *top_result_rel = find_base_rel(root,
1893 : parse->resultRelation);
1894 2786 : int resultRelation = -1;
1895 :
1896 : /* Pass the root result rel forward to the executor. */
1897 2786 : rootRelation = parse->resultRelation;
1898 :
1899 : /* Add only leaf children to ModifyTable. */
1900 8080 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
1901 : resultRelation)) >= 0)
1902 : {
1903 5294 : RelOptInfo *this_result_rel = find_base_rel(root,
1904 : resultRelation);
1905 :
1906 : /*
1907 : * Also exclude any leaf rels that have turned dummy since
1908 : * being added to the list, for example, by being excluded
1909 : * by constraint exclusion.
1910 : */
1911 5294 : if (IS_DUMMY_REL(this_result_rel))
1912 84 : continue;
1913 :
1914 : /* Build per-target-rel lists needed by ModifyTable */
1915 5210 : resultRelations = lappend_int(resultRelations,
1916 : resultRelation);
1917 5210 : if (parse->commandType == CMD_UPDATE)
1918 : {
1919 3624 : List *update_colnos = root->update_colnos;
1920 :
1921 3624 : if (this_result_rel != top_result_rel)
1922 : update_colnos =
1923 3624 : adjust_inherited_attnums_multilevel(root,
1924 : update_colnos,
1925 : this_result_rel->relid,
1926 : top_result_rel->relid);
1927 3624 : updateColnosLists = lappend(updateColnosLists,
1928 : update_colnos);
1929 : }
1930 5210 : if (parse->withCheckOptions)
1931 : {
1932 486 : List *withCheckOptions = parse->withCheckOptions;
1933 :
1934 486 : if (this_result_rel != top_result_rel)
1935 : withCheckOptions = (List *)
1936 486 : adjust_appendrel_attrs_multilevel(root,
1937 : (Node *) withCheckOptions,
1938 : this_result_rel,
1939 : top_result_rel);
1940 486 : withCheckOptionLists = lappend(withCheckOptionLists,
1941 : withCheckOptions);
1942 : }
1943 5210 : if (parse->returningList)
1944 : {
1945 834 : List *returningList = parse->returningList;
1946 :
1947 834 : if (this_result_rel != top_result_rel)
1948 : returningList = (List *)
1949 834 : adjust_appendrel_attrs_multilevel(root,
1950 : (Node *) returningList,
1951 : this_result_rel,
1952 : top_result_rel);
1953 834 : returningLists = lappend(returningLists,
1954 : returningList);
1955 : }
1956 5210 : if (parse->mergeActionList)
1957 : {
1958 : ListCell *l;
1959 492 : List *mergeActionList = NIL;
1960 :
1961 : /*
1962 : * Copy MergeActions and translate stuff that
1963 : * references attribute numbers.
1964 : */
1965 1554 : foreach(l, parse->mergeActionList)
1966 : {
1967 1062 : MergeAction *action = lfirst(l),
1968 1062 : *leaf_action = copyObject(action);
1969 :
1970 1062 : leaf_action->qual =
1971 1062 : adjust_appendrel_attrs_multilevel(root,
1972 : (Node *) action->qual,
1973 : this_result_rel,
1974 : top_result_rel);
1975 1062 : leaf_action->targetList = (List *)
1976 1062 : adjust_appendrel_attrs_multilevel(root,
1977 1062 : (Node *) action->targetList,
1978 : this_result_rel,
1979 : top_result_rel);
1980 1062 : if (leaf_action->commandType == CMD_UPDATE)
1981 592 : leaf_action->updateColnos =
1982 592 : adjust_inherited_attnums_multilevel(root,
1983 : action->updateColnos,
1984 : this_result_rel->relid,
1985 : top_result_rel->relid);
1986 1062 : mergeActionList = lappend(mergeActionList,
1987 : leaf_action);
1988 : }
1989 :
1990 492 : mergeActionLists = lappend(mergeActionLists,
1991 : mergeActionList);
1992 : }
1993 5210 : if (parse->commandType == CMD_MERGE)
1994 : {
1995 492 : Node *mergeJoinCondition = parse->mergeJoinCondition;
1996 :
1997 492 : if (this_result_rel != top_result_rel)
1998 : mergeJoinCondition =
1999 492 : adjust_appendrel_attrs_multilevel(root,
2000 : mergeJoinCondition,
2001 : this_result_rel,
2002 : top_result_rel);
2003 492 : mergeJoinConditions = lappend(mergeJoinConditions,
2004 : mergeJoinCondition);
2005 : }
2006 : }
2007 :
2008 2786 : if (resultRelations == NIL)
2009 : {
2010 : /*
2011 : * We managed to exclude every child rel, so generate a
2012 : * dummy one-relation plan using info for the top target
2013 : * rel (even though that may not be a leaf target).
2014 : * Although it's clear that no data will be updated or
2015 : * deleted, we still need to have a ModifyTable node so
2016 : * that any statement triggers will be executed. (This
2017 : * could be cleaner if we fixed nodeModifyTable.c to allow
2018 : * zero target relations, but that probably wouldn't be a
2019 : * net win.)
2020 : */
2021 30 : resultRelations = list_make1_int(parse->resultRelation);
2022 30 : if (parse->commandType == CMD_UPDATE)
2023 30 : updateColnosLists = list_make1(root->update_colnos);
2024 30 : if (parse->withCheckOptions)
2025 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2026 30 : if (parse->returningList)
2027 18 : returningLists = list_make1(parse->returningList);
2028 30 : if (parse->mergeActionList)
2029 0 : mergeActionLists = list_make1(parse->mergeActionList);
2030 30 : if (parse->commandType == CMD_MERGE)
2031 0 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2032 : }
2033 : }
2034 : else
2035 : {
2036 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2037 89708 : rootRelation = 0; /* there's no separate root rel */
2038 89708 : resultRelations = list_make1_int(parse->resultRelation);
2039 89708 : if (parse->commandType == CMD_UPDATE)
2040 11880 : updateColnosLists = list_make1(root->update_colnos);
2041 89708 : if (parse->withCheckOptions)
2042 926 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2043 89708 : if (parse->returningList)
2044 2418 : returningLists = list_make1(parse->returningList);
2045 89708 : if (parse->mergeActionList)
2046 1632 : mergeActionLists = list_make1(parse->mergeActionList);
2047 89708 : if (parse->commandType == CMD_MERGE)
2048 1632 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2049 : }
2050 :
2051 : /*
2052 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2053 : * will have dealt with fetching non-locked marked rows, else we
2054 : * need to have ModifyTable do that.
2055 : */
2056 92494 : if (parse->rowMarks)
2057 0 : rowMarks = NIL;
2058 : else
2059 92494 : rowMarks = root->rowMarks;
2060 :
2061 : path = (Path *)
2062 92494 : create_modifytable_path(root, final_rel,
2063 : path,
2064 : parse->commandType,
2065 92494 : parse->canSetTag,
2066 92494 : parse->resultRelation,
2067 : rootRelation,
2068 92494 : root->partColsUpdated,
2069 : resultRelations,
2070 : updateColnosLists,
2071 : withCheckOptionLists,
2072 : returningLists,
2073 : rowMarks,
2074 : parse->onConflict,
2075 : mergeActionLists,
2076 : mergeJoinConditions,
2077 : assign_special_exec_param(root));
2078 : }
2079 :
2080 : /* And shove it into final_rel */
2081 552016 : add_path(final_rel, path);
2082 : }
2083 :
2084 : /*
2085 : * Generate partial paths for final_rel, too, if outer query levels might
2086 : * be able to make use of them.
2087 : */
2088 532634 : if (final_rel->consider_parallel && root->query_level > 1 &&
2089 25554 : !limit_needed(parse))
2090 : {
2091 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2092 25494 : foreach(lc, current_rel->partial_pathlist)
2093 : {
2094 108 : Path *partial_path = (Path *) lfirst(lc);
2095 :
2096 108 : add_partial_path(final_rel, partial_path);
2097 : }
2098 : }
2099 :
2100 532634 : extra.limit_needed = limit_needed(parse);
2101 532634 : extra.limit_tuples = limit_tuples;
2102 532634 : extra.count_est = count_est;
2103 532634 : extra.offset_est = offset_est;
2104 :
2105 : /*
2106 : * If there is an FDW that's responsible for all baserels of the query,
2107 : * let it consider adding ForeignPaths.
2108 : */
2109 532634 : if (final_rel->fdwroutine &&
2110 1254 : final_rel->fdwroutine->GetForeignUpperPaths)
2111 1186 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2112 : current_rel, final_rel,
2113 : &extra);
2114 :
2115 : /* Let extensions possibly add some more paths */
2116 532634 : if (create_upper_paths_hook)
2117 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2118 : current_rel, final_rel, &extra);
2119 :
2120 : /* Note: currently, we leave it to callers to do set_cheapest() */
2121 532634 : }
2122 :
2123 : /*
2124 : * Do preprocessing for groupingSets clause and related data. This handles the
2125 : * preliminary steps of expanding the grouping sets, organizing them into lists
2126 : * of rollups, and preparing annotations which will later be filled in with
2127 : * size estimates.
2128 : */
2129 : static grouping_sets_data *
2130 878 : preprocess_grouping_sets(PlannerInfo *root)
2131 : {
2132 878 : Query *parse = root->parse;
2133 : List *sets;
2134 878 : int maxref = 0;
2135 : ListCell *lc_set;
2136 878 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2137 :
2138 878 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
2139 :
2140 878 : gd->any_hashable = false;
2141 878 : gd->unhashable_refs = NULL;
2142 878 : gd->unsortable_refs = NULL;
2143 878 : gd->unsortable_sets = NIL;
2144 :
2145 : /*
2146 : * We don't currently make any attempt to optimize the groupClause when
2147 : * there are grouping sets, so just duplicate it in processed_groupClause.
2148 : */
2149 878 : root->processed_groupClause = parse->groupClause;
2150 :
2151 878 : if (parse->groupClause)
2152 : {
2153 : ListCell *lc;
2154 :
2155 2672 : foreach(lc, parse->groupClause)
2156 : {
2157 1836 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2158 1836 : Index ref = gc->tleSortGroupRef;
2159 :
2160 1836 : if (ref > maxref)
2161 1800 : maxref = ref;
2162 :
2163 1836 : if (!gc->hashable)
2164 30 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2165 :
2166 1836 : if (!OidIsValid(gc->sortop))
2167 42 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2168 : }
2169 : }
2170 :
2171 : /* Allocate workspace array for remapping */
2172 878 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2173 :
2174 : /*
2175 : * If we have any unsortable sets, we must extract them before trying to
2176 : * prepare rollups. Unsortable sets don't go through
2177 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2178 : * here.
2179 : */
2180 878 : if (!bms_is_empty(gd->unsortable_refs))
2181 : {
2182 42 : List *sortable_sets = NIL;
2183 : ListCell *lc;
2184 :
2185 126 : foreach(lc, parse->groupingSets)
2186 : {
2187 90 : List *gset = (List *) lfirst(lc);
2188 :
2189 90 : if (bms_overlap_list(gd->unsortable_refs, gset))
2190 : {
2191 48 : GroupingSetData *gs = makeNode(GroupingSetData);
2192 :
2193 48 : gs->set = gset;
2194 48 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2195 :
2196 : /*
2197 : * We must enforce here that an unsortable set is hashable;
2198 : * later code assumes this. Parse analysis only checks that
2199 : * every individual column is either hashable or sortable.
2200 : *
2201 : * Note that passing this test doesn't guarantee we can
2202 : * generate a plan; there might be other showstoppers.
2203 : */
2204 48 : if (bms_overlap_list(gd->unhashable_refs, gset))
2205 6 : ereport(ERROR,
2206 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2207 : errmsg("could not implement GROUP BY"),
2208 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2209 : }
2210 : else
2211 42 : sortable_sets = lappend(sortable_sets, gset);
2212 : }
2213 :
2214 36 : if (sortable_sets)
2215 30 : sets = extract_rollup_sets(sortable_sets);
2216 : else
2217 6 : sets = NIL;
2218 : }
2219 : else
2220 836 : sets = extract_rollup_sets(parse->groupingSets);
2221 :
2222 2298 : foreach(lc_set, sets)
2223 : {
2224 1426 : List *current_sets = (List *) lfirst(lc_set);
2225 1426 : RollupData *rollup = makeNode(RollupData);
2226 : GroupingSetData *gs;
2227 :
2228 : /*
2229 : * Reorder the current list of grouping sets into correct prefix
2230 : * order. If only one aggregation pass is needed, try to make the
2231 : * list match the ORDER BY clause; if more than one pass is needed, we
2232 : * don't bother with that.
2233 : *
2234 : * Note that this reorders the sets from smallest-member-first to
2235 : * largest-member-first, and applies the GroupingSetData annotations,
2236 : * though the data will be filled in later.
2237 : */
2238 1426 : current_sets = reorder_grouping_sets(current_sets,
2239 1426 : (list_length(sets) == 1
2240 : ? parse->sortClause
2241 : : NIL));
2242 :
2243 : /*
2244 : * Get the initial (and therefore largest) grouping set.
2245 : */
2246 1426 : gs = linitial_node(GroupingSetData, current_sets);
2247 :
2248 : /*
2249 : * Order the groupClause appropriately. If the first grouping set is
2250 : * empty, then the groupClause must also be empty; otherwise we have
2251 : * to force the groupClause to match that grouping set's order.
2252 : *
2253 : * (The first grouping set can be empty even though parse->groupClause
2254 : * is not empty only if all non-empty grouping sets are unsortable.
2255 : * The groupClauses for hashed grouping sets are built later on.)
2256 : */
2257 1426 : if (gs->set)
2258 1384 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2259 : else
2260 42 : rollup->groupClause = NIL;
2261 :
2262 : /*
2263 : * Is it hashable? We pretend empty sets are hashable even though we
2264 : * actually force them not to be hashed later. But don't bother if
2265 : * there's nothing but empty sets (since in that case we can't hash
2266 : * anything).
2267 : */
2268 1426 : if (gs->set &&
2269 1384 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2270 : {
2271 1360 : rollup->hashable = true;
2272 1360 : gd->any_hashable = true;
2273 : }
2274 :
2275 : /*
2276 : * Now that we've pinned down an order for the groupClause for this
2277 : * list of grouping sets, we need to remap the entries in the grouping
2278 : * sets from sortgrouprefs to plain indices (0-based) into the
2279 : * groupClause for this collection of grouping sets. We keep the
2280 : * original form for later use, though.
2281 : */
2282 1426 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2283 : current_sets,
2284 : gd->tleref_to_colnum_map);
2285 1426 : rollup->gsets_data = current_sets;
2286 :
2287 1426 : gd->rollups = lappend(gd->rollups, rollup);
2288 : }
2289 :
2290 872 : if (gd->unsortable_sets)
2291 : {
2292 : /*
2293 : * We have not yet pinned down a groupclause for this, but we will
2294 : * need index-based lists for estimation purposes. Construct
2295 : * hash_sets_idx based on the entire original groupclause for now.
2296 : */
2297 36 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2298 : gd->unsortable_sets,
2299 : gd->tleref_to_colnum_map);
2300 36 : gd->any_hashable = true;
2301 : }
2302 :
2303 872 : return gd;
2304 : }
2305 :
2306 : /*
2307 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2308 : * (without annotation) mapped to indexes into the given groupclause.
2309 : */
2310 : static List *
2311 4164 : remap_to_groupclause_idx(List *groupClause,
2312 : List *gsets,
2313 : int *tleref_to_colnum_map)
2314 : {
2315 4164 : int ref = 0;
2316 4164 : List *result = NIL;
2317 : ListCell *lc;
2318 :
2319 10192 : foreach(lc, groupClause)
2320 : {
2321 6028 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2322 :
2323 6028 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2324 : }
2325 :
2326 9630 : foreach(lc, gsets)
2327 : {
2328 5466 : List *set = NIL;
2329 : ListCell *lc2;
2330 5466 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2331 :
2332 12332 : foreach(lc2, gs->set)
2333 : {
2334 6866 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2335 : }
2336 :
2337 5466 : result = lappend(result, set);
2338 : }
2339 :
2340 4164 : return result;
2341 : }
2342 :
2343 :
2344 : /*
2345 : * preprocess_rowmarks - set up PlanRowMarks if needed
2346 : */
2347 : static void
2348 536482 : preprocess_rowmarks(PlannerInfo *root)
2349 : {
2350 536482 : Query *parse = root->parse;
2351 : Bitmapset *rels;
2352 : List *prowmarks;
2353 : ListCell *l;
2354 : int i;
2355 :
2356 536482 : if (parse->rowMarks)
2357 : {
2358 : /*
2359 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2360 : * grouping, since grouping renders a reference to individual tuple
2361 : * CTIDs invalid. This is also checked at parse time, but that's
2362 : * insufficient because of rule substitution, query pullup, etc.
2363 : */
2364 7674 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2365 : parse->rowMarks)->strength);
2366 : }
2367 : else
2368 : {
2369 : /*
2370 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2371 : * UPDATE/SHARE.
2372 : */
2373 528808 : if (parse->commandType != CMD_UPDATE &&
2374 514944 : parse->commandType != CMD_DELETE &&
2375 510644 : parse->commandType != CMD_MERGE)
2376 508802 : return;
2377 : }
2378 :
2379 : /*
2380 : * We need to have rowmarks for all base relations except the target. We
2381 : * make a bitmapset of all base rels and then remove the items we don't
2382 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2383 : */
2384 27680 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2385 27680 : if (parse->resultRelation)
2386 20006 : rels = bms_del_member(rels, parse->resultRelation);
2387 :
2388 : /*
2389 : * Convert RowMarkClauses to PlanRowMark representation.
2390 : */
2391 27680 : prowmarks = NIL;
2392 35632 : foreach(l, parse->rowMarks)
2393 : {
2394 7952 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2395 7952 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2396 : PlanRowMark *newrc;
2397 :
2398 : /*
2399 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2400 : * applied to an update/delete target rel. If that ever becomes
2401 : * possible, we should drop the target from the PlanRowMark list.
2402 : */
2403 : Assert(rc->rti != parse->resultRelation);
2404 :
2405 : /*
2406 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2407 : * can't support true locking. Subqueries that got flattened into the
2408 : * main query should be ignored completely. Any that didn't will get
2409 : * ROW_MARK_COPY items in the next loop.
2410 : */
2411 7952 : if (rte->rtekind != RTE_RELATION)
2412 108 : continue;
2413 :
2414 7844 : rels = bms_del_member(rels, rc->rti);
2415 :
2416 7844 : newrc = makeNode(PlanRowMark);
2417 7844 : newrc->rti = newrc->prti = rc->rti;
2418 7844 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2419 7844 : newrc->markType = select_rowmark_type(rte, rc->strength);
2420 7844 : newrc->allMarkTypes = (1 << newrc->markType);
2421 7844 : newrc->strength = rc->strength;
2422 7844 : newrc->waitPolicy = rc->waitPolicy;
2423 7844 : newrc->isParent = false;
2424 :
2425 7844 : prowmarks = lappend(prowmarks, newrc);
2426 : }
2427 :
2428 : /*
2429 : * Now, add rowmarks for any non-target, non-locked base relations.
2430 : */
2431 27680 : i = 0;
2432 66680 : foreach(l, parse->rtable)
2433 : {
2434 39000 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2435 : PlanRowMark *newrc;
2436 :
2437 39000 : i++;
2438 39000 : if (!bms_is_member(i, rels))
2439 35268 : continue;
2440 :
2441 3732 : newrc = makeNode(PlanRowMark);
2442 3732 : newrc->rti = newrc->prti = i;
2443 3732 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2444 3732 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2445 3732 : newrc->allMarkTypes = (1 << newrc->markType);
2446 3732 : newrc->strength = LCS_NONE;
2447 3732 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2448 3732 : newrc->isParent = false;
2449 :
2450 3732 : prowmarks = lappend(prowmarks, newrc);
2451 : }
2452 :
2453 27680 : root->rowMarks = prowmarks;
2454 : }
2455 :
2456 : /*
2457 : * Select RowMarkType to use for a given table
2458 : */
2459 : RowMarkType
2460 13908 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2461 : {
2462 13908 : if (rte->rtekind != RTE_RELATION)
2463 : {
2464 : /* If it's not a table at all, use ROW_MARK_COPY */
2465 1482 : return ROW_MARK_COPY;
2466 : }
2467 12426 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2468 : {
2469 : /* Let the FDW select the rowmark type, if it wants to */
2470 212 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2471 :
2472 212 : if (fdwroutine->GetForeignRowMarkType != NULL)
2473 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2474 : /* Otherwise, use ROW_MARK_COPY by default */
2475 212 : return ROW_MARK_COPY;
2476 : }
2477 : else
2478 : {
2479 : /* Regular table, apply the appropriate lock type */
2480 12214 : switch (strength)
2481 : {
2482 2480 : case LCS_NONE:
2483 :
2484 : /*
2485 : * We don't need a tuple lock, only the ability to re-fetch
2486 : * the row.
2487 : */
2488 2480 : return ROW_MARK_REFERENCE;
2489 : break;
2490 7838 : case LCS_FORKEYSHARE:
2491 7838 : return ROW_MARK_KEYSHARE;
2492 : break;
2493 300 : case LCS_FORSHARE:
2494 300 : return ROW_MARK_SHARE;
2495 : break;
2496 72 : case LCS_FORNOKEYUPDATE:
2497 72 : return ROW_MARK_NOKEYEXCLUSIVE;
2498 : break;
2499 1524 : case LCS_FORUPDATE:
2500 1524 : return ROW_MARK_EXCLUSIVE;
2501 : break;
2502 : }
2503 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2504 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2505 : }
2506 : }
2507 :
2508 : /*
2509 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2510 : *
2511 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2512 : * results back in *count_est and *offset_est. These variables are set to
2513 : * 0 if the corresponding clause is not present, and -1 if it's present
2514 : * but we couldn't estimate the value for it. (The "0" convention is OK
2515 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2516 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2517 : * usual practice of never estimating less than one row.) These values will
2518 : * be passed to create_limit_path, which see if you change this code.
2519 : *
2520 : * The return value is the suitably adjusted tuple_fraction to use for
2521 : * planning the query. This adjustment is not overridable, since it reflects
2522 : * plan actions that grouping_planner() will certainly take, not assumptions
2523 : * about context.
2524 : */
2525 : static double
2526 4950 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2527 : int64 *offset_est, int64 *count_est)
2528 : {
2529 4950 : Query *parse = root->parse;
2530 : Node *est;
2531 : double limit_fraction;
2532 :
2533 : /* Should not be called unless LIMIT or OFFSET */
2534 : Assert(parse->limitCount || parse->limitOffset);
2535 :
2536 : /*
2537 : * Try to obtain the clause values. We use estimate_expression_value
2538 : * primarily because it can sometimes do something useful with Params.
2539 : */
2540 4950 : if (parse->limitCount)
2541 : {
2542 4452 : est = estimate_expression_value(root, parse->limitCount);
2543 4452 : if (est && IsA(est, Const))
2544 : {
2545 4446 : if (((Const *) est)->constisnull)
2546 : {
2547 : /* NULL indicates LIMIT ALL, ie, no limit */
2548 0 : *count_est = 0; /* treat as not present */
2549 : }
2550 : else
2551 : {
2552 4446 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2553 4446 : if (*count_est <= 0)
2554 150 : *count_est = 1; /* force to at least 1 */
2555 : }
2556 : }
2557 : else
2558 6 : *count_est = -1; /* can't estimate */
2559 : }
2560 : else
2561 498 : *count_est = 0; /* not present */
2562 :
2563 4950 : if (parse->limitOffset)
2564 : {
2565 870 : est = estimate_expression_value(root, parse->limitOffset);
2566 870 : if (est && IsA(est, Const))
2567 : {
2568 846 : if (((Const *) est)->constisnull)
2569 : {
2570 : /* Treat NULL as no offset; the executor will too */
2571 0 : *offset_est = 0; /* treat as not present */
2572 : }
2573 : else
2574 : {
2575 846 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2576 846 : if (*offset_est < 0)
2577 0 : *offset_est = 0; /* treat as not present */
2578 : }
2579 : }
2580 : else
2581 24 : *offset_est = -1; /* can't estimate */
2582 : }
2583 : else
2584 4080 : *offset_est = 0; /* not present */
2585 :
2586 4950 : if (*count_est != 0)
2587 : {
2588 : /*
2589 : * A LIMIT clause limits the absolute number of tuples returned.
2590 : * However, if it's not a constant LIMIT then we have to guess; for
2591 : * lack of a better idea, assume 10% of the plan's result is wanted.
2592 : */
2593 4452 : if (*count_est < 0 || *offset_est < 0)
2594 : {
2595 : /* LIMIT or OFFSET is an expression ... punt ... */
2596 24 : limit_fraction = 0.10;
2597 : }
2598 : else
2599 : {
2600 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2601 4428 : limit_fraction = (double) *count_est + (double) *offset_est;
2602 : }
2603 :
2604 : /*
2605 : * If we have absolute limits from both caller and LIMIT, use the
2606 : * smaller value; likewise if they are both fractional. If one is
2607 : * fractional and the other absolute, we can't easily determine which
2608 : * is smaller, but we use the heuristic that the absolute will usually
2609 : * be smaller.
2610 : */
2611 4452 : if (tuple_fraction >= 1.0)
2612 : {
2613 6 : if (limit_fraction >= 1.0)
2614 : {
2615 : /* both absolute */
2616 6 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2617 : }
2618 : else
2619 : {
2620 : /* caller absolute, limit fractional; use caller's value */
2621 : }
2622 : }
2623 4446 : else if (tuple_fraction > 0.0)
2624 : {
2625 148 : if (limit_fraction >= 1.0)
2626 : {
2627 : /* caller fractional, limit absolute; use limit */
2628 148 : tuple_fraction = limit_fraction;
2629 : }
2630 : else
2631 : {
2632 : /* both fractional */
2633 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2634 : }
2635 : }
2636 : else
2637 : {
2638 : /* no info from caller, just use limit */
2639 4298 : tuple_fraction = limit_fraction;
2640 : }
2641 : }
2642 498 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2643 : {
2644 : /*
2645 : * We have an OFFSET but no LIMIT. This acts entirely differently
2646 : * from the LIMIT case: here, we need to increase rather than decrease
2647 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2648 : * tuples to be fetched instead of fewer. This only matters if we got
2649 : * a tuple_fraction > 0, however.
2650 : *
2651 : * As above, use 10% if OFFSET is present but unestimatable.
2652 : */
2653 12 : if (*offset_est < 0)
2654 0 : limit_fraction = 0.10;
2655 : else
2656 12 : limit_fraction = (double) *offset_est;
2657 :
2658 : /*
2659 : * If we have absolute counts from both caller and OFFSET, add them
2660 : * together; likewise if they are both fractional. If one is
2661 : * fractional and the other absolute, we want to take the larger, and
2662 : * we heuristically assume that's the fractional one.
2663 : */
2664 12 : if (tuple_fraction >= 1.0)
2665 : {
2666 0 : if (limit_fraction >= 1.0)
2667 : {
2668 : /* both absolute, so add them together */
2669 0 : tuple_fraction += limit_fraction;
2670 : }
2671 : else
2672 : {
2673 : /* caller absolute, limit fractional; use limit */
2674 0 : tuple_fraction = limit_fraction;
2675 : }
2676 : }
2677 : else
2678 : {
2679 12 : if (limit_fraction >= 1.0)
2680 : {
2681 : /* caller fractional, limit absolute; use caller's value */
2682 : }
2683 : else
2684 : {
2685 : /* both fractional, so add them together */
2686 0 : tuple_fraction += limit_fraction;
2687 0 : if (tuple_fraction >= 1.0)
2688 0 : tuple_fraction = 0.0; /* assume fetch all */
2689 : }
2690 : }
2691 : }
2692 :
2693 4950 : return tuple_fraction;
2694 : }
2695 :
2696 : /*
2697 : * limit_needed - do we actually need a Limit plan node?
2698 : *
2699 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2700 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2701 : * locution for an optimization fence. (Because other places in the planner
2702 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2703 : * an optimization fence --- we're just suppressing unnecessary run-time
2704 : * overhead.)
2705 : *
2706 : * This might look like it could be merged into preprocess_limit, but there's
2707 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2708 : * in preprocess_limit it's good enough to consider estimated values.
2709 : */
2710 : bool
2711 1123140 : limit_needed(Query *parse)
2712 : {
2713 : Node *node;
2714 :
2715 1123140 : node = parse->limitCount;
2716 1123140 : if (node)
2717 : {
2718 10588 : if (IsA(node, Const))
2719 : {
2720 : /* NULL indicates LIMIT ALL, ie, no limit */
2721 10388 : if (!((Const *) node)->constisnull)
2722 10388 : return true; /* LIMIT with a constant value */
2723 : }
2724 : else
2725 200 : return true; /* non-constant LIMIT */
2726 : }
2727 :
2728 1112552 : node = parse->limitOffset;
2729 1112552 : if (node)
2730 : {
2731 1456 : if (IsA(node, Const))
2732 : {
2733 : /* Treat NULL as no offset; the executor would too */
2734 1160 : if (!((Const *) node)->constisnull)
2735 : {
2736 1160 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2737 :
2738 1160 : if (offset != 0)
2739 110 : return true; /* OFFSET with a nonzero value */
2740 : }
2741 : }
2742 : else
2743 296 : return true; /* non-constant OFFSET */
2744 : }
2745 :
2746 1112146 : return false; /* don't need a Limit plan node */
2747 : }
2748 :
2749 : /*
2750 : * preprocess_groupclause - do preparatory work on GROUP BY clause
2751 : *
2752 : * The idea here is to adjust the ordering of the GROUP BY elements
2753 : * (which in itself is semantically insignificant) to match ORDER BY,
2754 : * thereby allowing a single sort operation to both implement the ORDER BY
2755 : * requirement and set up for a Unique step that implements GROUP BY.
2756 : * We also consider partial match between GROUP BY and ORDER BY elements,
2757 : * which could allow to implement ORDER BY using the incremental sort.
2758 : *
2759 : * We also consider other orderings of the GROUP BY elements, which could
2760 : * match the sort ordering of other possible plans (eg an indexscan) and
2761 : * thereby reduce cost. This is implemented during the generation of grouping
2762 : * paths. See get_useful_group_keys_orderings() for details.
2763 : *
2764 : * Note: we need no comparable processing of the distinctClause because
2765 : * the parser already enforced that that matches ORDER BY.
2766 : *
2767 : * Note: we return a fresh List, but its elements are the same
2768 : * SortGroupClauses appearing in parse->groupClause. This is important
2769 : * because later processing may modify the processed_groupClause list.
2770 : *
2771 : * For grouping sets, the order of items is instead forced to agree with that
2772 : * of the grouping set (and items not in the grouping set are skipped). The
2773 : * work of sorting the order of grouping set elements to match the ORDER BY if
2774 : * possible is done elsewhere.
2775 : */
2776 : static List *
2777 7718 : preprocess_groupclause(PlannerInfo *root, List *force)
2778 : {
2779 7718 : Query *parse = root->parse;
2780 7718 : List *new_groupclause = NIL;
2781 : ListCell *sl;
2782 : ListCell *gl;
2783 :
2784 : /* For grouping sets, we need to force the ordering */
2785 7718 : if (force)
2786 : {
2787 10036 : foreach(sl, force)
2788 : {
2789 5950 : Index ref = lfirst_int(sl);
2790 5950 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2791 :
2792 5950 : new_groupclause = lappend(new_groupclause, cl);
2793 : }
2794 :
2795 4086 : return new_groupclause;
2796 : }
2797 :
2798 : /* If no ORDER BY, nothing useful to do here */
2799 3632 : if (parse->sortClause == NIL)
2800 2056 : return list_copy(parse->groupClause);
2801 :
2802 : /*
2803 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2804 : * items, but only as far as we can make a matching prefix.
2805 : *
2806 : * This code assumes that the sortClause contains no duplicate items.
2807 : */
2808 3058 : foreach(sl, parse->sortClause)
2809 : {
2810 2128 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2811 :
2812 3224 : foreach(gl, parse->groupClause)
2813 : {
2814 2578 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2815 :
2816 2578 : if (equal(gc, sc))
2817 : {
2818 1482 : new_groupclause = lappend(new_groupclause, gc);
2819 1482 : break;
2820 : }
2821 : }
2822 2128 : if (gl == NULL)
2823 646 : break; /* no match, so stop scanning */
2824 : }
2825 :
2826 :
2827 : /* If no match at all, no point in reordering GROUP BY */
2828 1576 : if (new_groupclause == NIL)
2829 298 : return list_copy(parse->groupClause);
2830 :
2831 : /*
2832 : * Add any remaining GROUP BY items to the new list. We don't require a
2833 : * complete match, because even partial match allows ORDER BY to be
2834 : * implemented using incremental sort. Also, give up if there are any
2835 : * non-sortable GROUP BY items, since then there's no hope anyway.
2836 : */
2837 2926 : foreach(gl, parse->groupClause)
2838 : {
2839 1648 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2840 :
2841 1648 : if (list_member_ptr(new_groupclause, gc))
2842 1482 : continue; /* it matched an ORDER BY item */
2843 166 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2844 0 : return list_copy(parse->groupClause);
2845 166 : new_groupclause = lappend(new_groupclause, gc);
2846 : }
2847 :
2848 : /* Success --- install the rearranged GROUP BY list */
2849 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2850 1278 : return new_groupclause;
2851 : }
2852 :
2853 : /*
2854 : * Extract lists of grouping sets that can be implemented using a single
2855 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2856 : *
2857 : * Input must be sorted with smallest sets first. Result has each sublist
2858 : * sorted with smallest sets first.
2859 : *
2860 : * We want to produce the absolute minimum possible number of lists here to
2861 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2862 : * of finding the minimal partition of a partially-ordered set into chains
2863 : * (which is what we need, taking the list of grouping sets as a poset ordered
2864 : * by set inclusion) can be mapped to the problem of finding the maximum
2865 : * cardinality matching on a bipartite graph, which is solvable in polynomial
2866 : * time with a worst case of no worse than O(n^2.5) and usually much
2867 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
2868 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
2869 : * half a second on my modest system even with optimization off and assertions
2870 : * on.)
2871 : */
2872 : static List *
2873 866 : extract_rollup_sets(List *groupingSets)
2874 : {
2875 866 : int num_sets_raw = list_length(groupingSets);
2876 866 : int num_empty = 0;
2877 866 : int num_sets = 0; /* distinct sets */
2878 866 : int num_chains = 0;
2879 866 : List *result = NIL;
2880 : List **results;
2881 : List **orig_sets;
2882 : Bitmapset **set_masks;
2883 : int *chains;
2884 : short **adjacency;
2885 : short *adjacency_buf;
2886 : BipartiteMatchState *state;
2887 : int i;
2888 : int j;
2889 : int j_size;
2890 866 : ListCell *lc1 = list_head(groupingSets);
2891 : ListCell *lc;
2892 :
2893 : /*
2894 : * Start by stripping out empty sets. The algorithm doesn't require this,
2895 : * but the planner currently needs all empty sets to be returned in the
2896 : * first list, so we strip them here and add them back after.
2897 : */
2898 1476 : while (lc1 && lfirst(lc1) == NIL)
2899 : {
2900 610 : ++num_empty;
2901 610 : lc1 = lnext(groupingSets, lc1);
2902 : }
2903 :
2904 : /* bail out now if it turns out that all we had were empty sets. */
2905 866 : if (!lc1)
2906 42 : return list_make1(groupingSets);
2907 :
2908 : /*----------
2909 : * We don't strictly need to remove duplicate sets here, but if we don't,
2910 : * they tend to become scattered through the result, which is a bit
2911 : * confusing (and irritating if we ever decide to optimize them out).
2912 : * So we remove them here and add them back after.
2913 : *
2914 : * For each non-duplicate set, we fill in the following:
2915 : *
2916 : * orig_sets[i] = list of the original set lists
2917 : * set_masks[i] = bitmapset for testing inclusion
2918 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2919 : *
2920 : * chains[i] will be the result group this set is assigned to.
2921 : *
2922 : * We index all of these from 1 rather than 0 because it is convenient
2923 : * to leave 0 free for the NIL node in the graph algorithm.
2924 : *----------
2925 : */
2926 824 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
2927 824 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
2928 824 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
2929 824 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
2930 :
2931 824 : j_size = 0;
2932 824 : j = 0;
2933 824 : i = 1;
2934 :
2935 2936 : for_each_cell(lc, groupingSets, lc1)
2936 : {
2937 2112 : List *candidate = (List *) lfirst(lc);
2938 2112 : Bitmapset *candidate_set = NULL;
2939 : ListCell *lc2;
2940 2112 : int dup_of = 0;
2941 :
2942 5118 : foreach(lc2, candidate)
2943 : {
2944 3006 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
2945 : }
2946 :
2947 : /* we can only be a dup if we're the same length as a previous set */
2948 2112 : if (j_size == list_length(candidate))
2949 : {
2950 : int k;
2951 :
2952 1844 : for (k = j; k < i; ++k)
2953 : {
2954 1188 : if (bms_equal(set_masks[k], candidate_set))
2955 : {
2956 158 : dup_of = k;
2957 158 : break;
2958 : }
2959 : }
2960 : }
2961 1298 : else if (j_size < list_length(candidate))
2962 : {
2963 1298 : j_size = list_length(candidate);
2964 1298 : j = i;
2965 : }
2966 :
2967 2112 : if (dup_of > 0)
2968 : {
2969 158 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
2970 158 : bms_free(candidate_set);
2971 : }
2972 : else
2973 : {
2974 : int k;
2975 1954 : int n_adj = 0;
2976 :
2977 1954 : orig_sets[i] = list_make1(candidate);
2978 1954 : set_masks[i] = candidate_set;
2979 :
2980 : /* fill in adjacency list; no need to compare equal-size sets */
2981 :
2982 3226 : for (k = j - 1; k > 0; --k)
2983 : {
2984 1272 : if (bms_is_subset(set_masks[k], candidate_set))
2985 1110 : adjacency_buf[++n_adj] = k;
2986 : }
2987 :
2988 1954 : if (n_adj > 0)
2989 : {
2990 598 : adjacency_buf[0] = n_adj;
2991 598 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
2992 598 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
2993 : }
2994 : else
2995 1356 : adjacency[i] = NULL;
2996 :
2997 1954 : ++i;
2998 : }
2999 : }
3000 :
3001 824 : num_sets = i - 1;
3002 :
3003 : /*
3004 : * Apply the graph matching algorithm to do the work.
3005 : */
3006 824 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3007 :
3008 : /*
3009 : * Now, the state->pair* fields have the info we need to assign sets to
3010 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3011 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3012 : * it in one pass)
3013 : */
3014 824 : chains = palloc0((num_sets + 1) * sizeof(int));
3015 :
3016 2778 : for (i = 1; i <= num_sets; ++i)
3017 : {
3018 1954 : int u = state->pair_vu[i];
3019 1954 : int v = state->pair_uv[i];
3020 :
3021 1954 : if (u > 0 && u < i)
3022 0 : chains[i] = chains[u];
3023 1954 : else if (v > 0 && v < i)
3024 570 : chains[i] = chains[v];
3025 : else
3026 1384 : chains[i] = ++num_chains;
3027 : }
3028 :
3029 : /* build result lists. */
3030 824 : results = palloc0((num_chains + 1) * sizeof(List *));
3031 :
3032 2778 : for (i = 1; i <= num_sets; ++i)
3033 : {
3034 1954 : int c = chains[i];
3035 :
3036 : Assert(c > 0);
3037 :
3038 1954 : results[c] = list_concat(results[c], orig_sets[i]);
3039 : }
3040 :
3041 : /* push any empty sets back on the first list. */
3042 1344 : while (num_empty-- > 0)
3043 520 : results[1] = lcons(NIL, results[1]);
3044 :
3045 : /* make result list */
3046 2208 : for (i = 1; i <= num_chains; ++i)
3047 1384 : result = lappend(result, results[i]);
3048 :
3049 : /*
3050 : * Free all the things.
3051 : *
3052 : * (This is over-fussy for small sets but for large sets we could have
3053 : * tied up a nontrivial amount of memory.)
3054 : */
3055 824 : BipartiteMatchFree(state);
3056 824 : pfree(results);
3057 824 : pfree(chains);
3058 2778 : for (i = 1; i <= num_sets; ++i)
3059 1954 : if (adjacency[i])
3060 598 : pfree(adjacency[i]);
3061 824 : pfree(adjacency);
3062 824 : pfree(adjacency_buf);
3063 824 : pfree(orig_sets);
3064 2778 : for (i = 1; i <= num_sets; ++i)
3065 1954 : bms_free(set_masks[i]);
3066 824 : pfree(set_masks);
3067 :
3068 824 : return result;
3069 : }
3070 :
3071 : /*
3072 : * Reorder the elements of a list of grouping sets such that they have correct
3073 : * prefix relationships. Also inserts the GroupingSetData annotations.
3074 : *
3075 : * The input must be ordered with smallest sets first; the result is returned
3076 : * with largest sets first. Note that the result shares no list substructure
3077 : * with the input, so it's safe for the caller to modify it later.
3078 : *
3079 : * If we're passed in a sortclause, we follow its order of columns to the
3080 : * extent possible, to minimize the chance that we add unnecessary sorts.
3081 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3082 : * gets implemented in one pass.)
3083 : */
3084 : static List *
3085 1426 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3086 : {
3087 : ListCell *lc;
3088 1426 : List *previous = NIL;
3089 1426 : List *result = NIL;
3090 :
3091 4148 : foreach(lc, groupingSets)
3092 : {
3093 2722 : List *candidate = (List *) lfirst(lc);
3094 2722 : List *new_elems = list_difference_int(candidate, previous);
3095 2722 : GroupingSetData *gs = makeNode(GroupingSetData);
3096 :
3097 2886 : while (list_length(sortclause) > list_length(previous) &&
3098 : new_elems != NIL)
3099 : {
3100 272 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3101 272 : int ref = sc->tleSortGroupRef;
3102 :
3103 272 : if (list_member_int(new_elems, ref))
3104 : {
3105 164 : previous = lappend_int(previous, ref);
3106 164 : new_elems = list_delete_int(new_elems, ref);
3107 : }
3108 : else
3109 : {
3110 : /* diverged from the sortclause; give up on it */
3111 108 : sortclause = NIL;
3112 108 : break;
3113 : }
3114 : }
3115 :
3116 2722 : previous = list_concat(previous, new_elems);
3117 :
3118 2722 : gs->set = list_copy(previous);
3119 2722 : result = lcons(gs, result);
3120 : }
3121 :
3122 1426 : list_free(previous);
3123 :
3124 1426 : return result;
3125 : }
3126 :
3127 : /*
3128 : * has_volatile_pathkey
3129 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3130 : * containing a volatile function. Otherwise returns false.
3131 : */
3132 : static bool
3133 3092 : has_volatile_pathkey(List *keys)
3134 : {
3135 : ListCell *lc;
3136 :
3137 6328 : foreach(lc, keys)
3138 : {
3139 3254 : PathKey *pathkey = lfirst_node(PathKey, lc);
3140 :
3141 3254 : if (pathkey->pk_eclass->ec_has_volatile)
3142 18 : return true;
3143 : }
3144 :
3145 3074 : return false;
3146 : }
3147 :
3148 : /*
3149 : * adjust_group_pathkeys_for_groupagg
3150 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3151 : * pre-ordered input for ordered aggregates.
3152 : *
3153 : * We define "best" as the pathkeys that suit the largest number of
3154 : * aggregate functions. We find these by looking at the first ORDER BY /
3155 : * DISTINCT aggregate and take the pathkeys for that before searching for
3156 : * other aggregates that require the same or a more strict variation of the
3157 : * same pathkeys. We then repeat that process for any remaining aggregates
3158 : * with different pathkeys and if we find another set of pathkeys that suits a
3159 : * larger number of aggregates then we select those pathkeys instead.
3160 : *
3161 : * When the best pathkeys are found we also mark each Aggref that can use
3162 : * those pathkeys as aggpresorted = true.
3163 : *
3164 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3165 : * volatile functions, we never make use of these pathkeys. We want to ensure
3166 : * that sorts using volatile functions are done independently in each Aggref
3167 : * rather than once at the query level. If we were to allow this then Aggrefs
3168 : * with compatible sort orders would all transition their rows in the same
3169 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3170 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3171 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3172 : * left to perform their sorts individually. To avoid this inconsistent
3173 : * behavior which could make Aggref results depend on what other Aggrefs the
3174 : * query contains, we always force Aggrefs with volatile functions to perform
3175 : * their own sorts.
3176 : */
3177 : static void
3178 2696 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3179 : {
3180 2696 : List *grouppathkeys = root->group_pathkeys;
3181 : List *bestpathkeys;
3182 : Bitmapset *bestaggs;
3183 : Bitmapset *unprocessed_aggs;
3184 : ListCell *lc;
3185 : int i;
3186 :
3187 : /* Shouldn't be here if there are grouping sets */
3188 : Assert(root->parse->groupingSets == NIL);
3189 : /* Shouldn't be here unless there are some ordered aggregates */
3190 : Assert(root->numOrderedAggs > 0);
3191 :
3192 : /* Do nothing if disabled */
3193 2696 : if (!enable_presorted_aggregate)
3194 6 : return;
3195 :
3196 : /*
3197 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3198 : * the indexes of all AggInfos to be processed below.
3199 : */
3200 2690 : unprocessed_aggs = NULL;
3201 6064 : foreach(lc, root->agginfos)
3202 : {
3203 3374 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3204 3374 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3205 :
3206 3374 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3207 264 : continue;
3208 :
3209 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3210 3110 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3211 300 : continue;
3212 :
3213 : /* Additional safety checks are needed if there's a FILTER clause */
3214 2810 : if (aggref->aggfilter != NULL)
3215 : {
3216 : ListCell *lc2;
3217 54 : bool allow_presort = true;
3218 :
3219 : /*
3220 : * When the Aggref has a FILTER clause, it's possible that the
3221 : * filter removes rows that cannot be sorted because the
3222 : * expression to sort by results in an error during its
3223 : * evaluation. This is a problem for presorting as that happens
3224 : * before the FILTER, whereas without presorting, the Aggregate
3225 : * node will apply the FILTER *before* sorting. So that we never
3226 : * try to sort anything that might error, here we aim to skip over
3227 : * any Aggrefs with arguments with expressions which, when
3228 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3229 : * may be more cases that should be allowed, but more thought
3230 : * needs to be given. Err on the side of caution.
3231 : */
3232 102 : foreach(lc2, aggref->args)
3233 : {
3234 72 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3235 72 : Expr *expr = tle->expr;
3236 :
3237 84 : while (IsA(expr, RelabelType))
3238 12 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3239 :
3240 : /* Common case, Vars and Consts are ok */
3241 72 : if (IsA(expr, Var) || IsA(expr, Const))
3242 48 : continue;
3243 :
3244 : /* Unsupported. Don't try to presort for this Aggref */
3245 24 : allow_presort = false;
3246 24 : break;
3247 : }
3248 :
3249 : /* Skip unsupported Aggrefs */
3250 54 : if (!allow_presort)
3251 24 : continue;
3252 : }
3253 :
3254 2786 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3255 : foreach_current_index(lc));
3256 : }
3257 :
3258 : /*
3259 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3260 : * for the given set of aggregates.
3261 : *
3262 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3263 : * this during the first loop using the pathkeys for the very first
3264 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3265 : * a more strict set of compatible pathkeys. Once the outer loop is
3266 : * complete, we mark off all the aggregates with compatible pathkeys then
3267 : * remove those from the unprocessed_aggs and repeat the process to try to
3268 : * find another set of pathkeys that are suitable for a larger number of
3269 : * aggregates. The outer loop will stop when there are not enough
3270 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3271 : * to suit a larger number of aggregates.
3272 : */
3273 2690 : bestpathkeys = NIL;
3274 2690 : bestaggs = NULL;
3275 5314 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3276 : {
3277 2624 : Bitmapset *aggindexes = NULL;
3278 2624 : List *currpathkeys = NIL;
3279 :
3280 2624 : i = -1;
3281 8340 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3282 : {
3283 3092 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3284 3092 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3285 : List *sortlist;
3286 : List *pathkeys;
3287 :
3288 3092 : if (aggref->aggdistinct != NIL)
3289 718 : sortlist = aggref->aggdistinct;
3290 : else
3291 2374 : sortlist = aggref->aggorder;
3292 :
3293 3092 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3294 : aggref->args);
3295 :
3296 : /*
3297 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3298 : * or DISTINCT clause.
3299 : */
3300 3092 : if (has_volatile_pathkey(pathkeys))
3301 : {
3302 18 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3303 18 : continue;
3304 : }
3305 :
3306 : /*
3307 : * When not set yet, take the pathkeys from the first unprocessed
3308 : * aggregate.
3309 : */
3310 3074 : if (currpathkeys == NIL)
3311 : {
3312 2618 : currpathkeys = pathkeys;
3313 :
3314 : /* include the GROUP BY pathkeys, if they exist */
3315 2618 : if (grouppathkeys != NIL)
3316 276 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3317 : currpathkeys);
3318 :
3319 : /* record that we found pathkeys for this aggregate */
3320 2618 : aggindexes = bms_add_member(aggindexes, i);
3321 : }
3322 : else
3323 : {
3324 : /* now look for a stronger set of matching pathkeys */
3325 :
3326 : /* include the GROUP BY pathkeys, if they exist */
3327 456 : if (grouppathkeys != NIL)
3328 288 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3329 : pathkeys);
3330 :
3331 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3332 456 : switch (compare_pathkeys(currpathkeys, pathkeys))
3333 : {
3334 12 : case PATHKEYS_BETTER2:
3335 : /* 'pathkeys' are stronger, use these ones instead */
3336 12 : currpathkeys = pathkeys;
3337 : /* FALLTHROUGH */
3338 :
3339 66 : case PATHKEYS_BETTER1:
3340 : /* 'pathkeys' are less strict */
3341 : /* FALLTHROUGH */
3342 :
3343 : case PATHKEYS_EQUAL:
3344 : /* mark this aggregate as covered by 'currpathkeys' */
3345 66 : aggindexes = bms_add_member(aggindexes, i);
3346 66 : break;
3347 :
3348 390 : case PATHKEYS_DIFFERENT:
3349 390 : break;
3350 : }
3351 5716 : }
3352 : }
3353 :
3354 : /* remove the aggregates that we've just processed */
3355 2624 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3356 :
3357 : /*
3358 : * If this pass included more aggregates than the previous best then
3359 : * use these ones as the best set.
3360 : */
3361 2624 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3362 : {
3363 2516 : bestaggs = aggindexes;
3364 2516 : bestpathkeys = currpathkeys;
3365 : }
3366 : }
3367 :
3368 : /*
3369 : * If we found any ordered aggregates, update root->group_pathkeys to add
3370 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3371 : * the original GROUP BY pathkeys already.
3372 : */
3373 2690 : if (bestpathkeys != NIL)
3374 2456 : root->group_pathkeys = bestpathkeys;
3375 :
3376 : /*
3377 : * Now that we've found the best set of aggregates we can set the
3378 : * presorted flag to indicate to the executor that it needn't bother
3379 : * performing a sort for these Aggrefs. We're able to do this now as
3380 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3381 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3382 : * of ordered aggregates.
3383 : */
3384 2690 : i = -1;
3385 5242 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3386 : {
3387 2552 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3388 :
3389 5122 : foreach(lc, agginfo->aggrefs)
3390 : {
3391 2570 : Aggref *aggref = lfirst_node(Aggref, lc);
3392 :
3393 2570 : aggref->aggpresorted = true;
3394 : }
3395 : }
3396 : }
3397 :
3398 : /*
3399 : * Compute query_pathkeys and other pathkeys during plan generation
3400 : */
3401 : static void
3402 526516 : standard_qp_callback(PlannerInfo *root, void *extra)
3403 : {
3404 526516 : Query *parse = root->parse;
3405 526516 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3406 526516 : List *tlist = root->processed_tlist;
3407 526516 : List *activeWindows = qp_extra->activeWindows;
3408 :
3409 : /*
3410 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3411 : * aggregate requirements.
3412 : */
3413 526516 : if (qp_extra->gset_data)
3414 : {
3415 : /*
3416 : * With grouping sets, just use the first RollupData's groupClause. We
3417 : * don't make any effort to optimize grouping clauses when there are
3418 : * grouping sets, nor can we combine aggregate ordering keys with
3419 : * grouping.
3420 : */
3421 872 : List *rollups = qp_extra->gset_data->rollups;
3422 872 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3423 :
3424 872 : if (grouping_is_sortable(groupClause))
3425 : {
3426 : bool sortable;
3427 :
3428 : /*
3429 : * The groupClause is logically below the grouping step. So if
3430 : * there is an RTE entry for the grouping step, we need to remove
3431 : * its RT index from the sort expressions before we make PathKeys
3432 : * for them.
3433 : */
3434 872 : root->group_pathkeys =
3435 872 : make_pathkeys_for_sortclauses_extended(root,
3436 : &groupClause,
3437 : tlist,
3438 : false,
3439 872 : parse->hasGroupRTE,
3440 : &sortable,
3441 : false);
3442 : Assert(sortable);
3443 872 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3444 : }
3445 : else
3446 : {
3447 0 : root->group_pathkeys = NIL;
3448 0 : root->num_groupby_pathkeys = 0;
3449 : }
3450 : }
3451 525644 : else if (parse->groupClause || root->numOrderedAggs > 0)
3452 6084 : {
3453 : /*
3454 : * With a plain GROUP BY list, we can remove any grouping items that
3455 : * are proven redundant by EquivalenceClass processing. For example,
3456 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3457 : * especially common cases, but they're nearly free to detect. Note
3458 : * that we remove redundant items from processed_groupClause but not
3459 : * the original parse->groupClause.
3460 : */
3461 : bool sortable;
3462 :
3463 : /*
3464 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3465 : * EquivalenceClass'es if it's not set yet.
3466 : */
3467 6084 : root->group_pathkeys =
3468 6084 : make_pathkeys_for_sortclauses_extended(root,
3469 : &root->processed_groupClause,
3470 : tlist,
3471 : true,
3472 : false,
3473 : &sortable,
3474 : true);
3475 6084 : if (!sortable)
3476 : {
3477 : /* Can't sort; no point in considering aggregate ordering either */
3478 0 : root->group_pathkeys = NIL;
3479 0 : root->num_groupby_pathkeys = 0;
3480 : }
3481 : else
3482 : {
3483 6084 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3484 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3485 6084 : if (root->numOrderedAggs > 0)
3486 2696 : adjust_group_pathkeys_for_groupagg(root);
3487 : }
3488 : }
3489 : else
3490 : {
3491 519560 : root->group_pathkeys = NIL;
3492 519560 : root->num_groupby_pathkeys = 0;
3493 : }
3494 :
3495 : /* We consider only the first (bottom) window in pathkeys logic */
3496 526516 : if (activeWindows != NIL)
3497 : {
3498 2378 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3499 :
3500 2378 : root->window_pathkeys = make_pathkeys_for_window(root,
3501 : wc,
3502 : tlist);
3503 : }
3504 : else
3505 524138 : root->window_pathkeys = NIL;
3506 :
3507 : /*
3508 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3509 : * redundant by EquivalenceClass processing. The non-redundant list is
3510 : * kept in root->processed_distinctClause, leaving the original
3511 : * parse->distinctClause alone.
3512 : */
3513 526516 : if (parse->distinctClause)
3514 : {
3515 : bool sortable;
3516 :
3517 : /* Make a copy since pathkey processing can modify the list */
3518 2676 : root->processed_distinctClause = list_copy(parse->distinctClause);
3519 2676 : root->distinct_pathkeys =
3520 2676 : make_pathkeys_for_sortclauses_extended(root,
3521 : &root->processed_distinctClause,
3522 : tlist,
3523 : true,
3524 : false,
3525 : &sortable,
3526 : false);
3527 2676 : if (!sortable)
3528 6 : root->distinct_pathkeys = NIL;
3529 : }
3530 : else
3531 523840 : root->distinct_pathkeys = NIL;
3532 :
3533 526516 : root->sort_pathkeys =
3534 526516 : make_pathkeys_for_sortclauses(root,
3535 : parse->sortClause,
3536 : tlist);
3537 :
3538 : /* setting setop_pathkeys might be useful to the union planner */
3539 526516 : if (qp_extra->setop != NULL)
3540 : {
3541 : List *groupClauses;
3542 : bool sortable;
3543 :
3544 12278 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3545 :
3546 12278 : root->setop_pathkeys =
3547 12278 : make_pathkeys_for_sortclauses_extended(root,
3548 : &groupClauses,
3549 : tlist,
3550 : false,
3551 : false,
3552 : &sortable,
3553 : false);
3554 12278 : if (!sortable)
3555 200 : root->setop_pathkeys = NIL;
3556 : }
3557 : else
3558 514238 : root->setop_pathkeys = NIL;
3559 :
3560 : /*
3561 : * Figure out whether we want a sorted result from query_planner.
3562 : *
3563 : * If we have a sortable GROUP BY clause, then we want a result sorted
3564 : * properly for grouping. Otherwise, if we have window functions to
3565 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3566 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3567 : * we try to produce output that's sufficiently well sorted for the
3568 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3569 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3570 : * for a set operation which can benefit from presorted results and have a
3571 : * sortable targetlist, we want to sort by the target list.
3572 : *
3573 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3574 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3575 : * that might just leave us failing to exploit an available sort order at
3576 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3577 : * much easier, since we know that the parser ensured that one is a
3578 : * superset of the other.
3579 : */
3580 526516 : if (root->group_pathkeys)
3581 6600 : root->query_pathkeys = root->group_pathkeys;
3582 519916 : else if (root->window_pathkeys)
3583 2032 : root->query_pathkeys = root->window_pathkeys;
3584 1035768 : else if (list_length(root->distinct_pathkeys) >
3585 517884 : list_length(root->sort_pathkeys))
3586 2224 : root->query_pathkeys = root->distinct_pathkeys;
3587 515660 : else if (root->sort_pathkeys)
3588 69530 : root->query_pathkeys = root->sort_pathkeys;
3589 446130 : else if (root->setop_pathkeys != NIL)
3590 10878 : root->query_pathkeys = root->setop_pathkeys;
3591 : else
3592 435252 : root->query_pathkeys = NIL;
3593 526516 : }
3594 :
3595 : /*
3596 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3597 : *
3598 : * path_rows: number of output rows from scan/join step
3599 : * gd: grouping sets data including list of grouping sets and their clauses
3600 : * target_list: target list containing group clause references
3601 : *
3602 : * If doing grouping sets, we also annotate the gsets data with the estimates
3603 : * for each set and each individual rollup list, with a view to later
3604 : * determining whether some combination of them could be hashed instead.
3605 : */
3606 : static double
3607 44974 : get_number_of_groups(PlannerInfo *root,
3608 : double path_rows,
3609 : grouping_sets_data *gd,
3610 : List *target_list)
3611 : {
3612 44974 : Query *parse = root->parse;
3613 : double dNumGroups;
3614 :
3615 44974 : if (parse->groupClause)
3616 : {
3617 : List *groupExprs;
3618 :
3619 6942 : if (parse->groupingSets)
3620 : {
3621 : /* Add up the estimates for each grouping set */
3622 : ListCell *lc;
3623 :
3624 : Assert(gd); /* keep Coverity happy */
3625 :
3626 830 : dNumGroups = 0;
3627 :
3628 2214 : foreach(lc, gd->rollups)
3629 : {
3630 1384 : RollupData *rollup = lfirst_node(RollupData, lc);
3631 : ListCell *lc2;
3632 : ListCell *lc3;
3633 :
3634 1384 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3635 : target_list);
3636 :
3637 1384 : rollup->numGroups = 0.0;
3638 :
3639 4016 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3640 : {
3641 2632 : List *gset = (List *) lfirst(lc2);
3642 2632 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3643 2632 : double numGroups = estimate_num_groups(root,
3644 : groupExprs,
3645 : path_rows,
3646 : &gset,
3647 : NULL);
3648 :
3649 2632 : gs->numGroups = numGroups;
3650 2632 : rollup->numGroups += numGroups;
3651 : }
3652 :
3653 1384 : dNumGroups += rollup->numGroups;
3654 : }
3655 :
3656 830 : if (gd->hash_sets_idx)
3657 : {
3658 : ListCell *lc2;
3659 :
3660 36 : gd->dNumHashGroups = 0;
3661 :
3662 36 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3663 : target_list);
3664 :
3665 78 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3666 : {
3667 42 : List *gset = (List *) lfirst(lc);
3668 42 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3669 42 : double numGroups = estimate_num_groups(root,
3670 : groupExprs,
3671 : path_rows,
3672 : &gset,
3673 : NULL);
3674 :
3675 42 : gs->numGroups = numGroups;
3676 42 : gd->dNumHashGroups += numGroups;
3677 : }
3678 :
3679 36 : dNumGroups += gd->dNumHashGroups;
3680 : }
3681 : }
3682 : else
3683 : {
3684 : /* Plain GROUP BY -- estimate based on optimized groupClause */
3685 6112 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3686 : target_list);
3687 :
3688 6112 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3689 : NULL, NULL);
3690 : }
3691 : }
3692 38032 : else if (parse->groupingSets)
3693 : {
3694 : /* Empty grouping sets ... one result row for each one */
3695 42 : dNumGroups = list_length(parse->groupingSets);
3696 : }
3697 37990 : else if (parse->hasAggs || root->hasHavingQual)
3698 : {
3699 : /* Plain aggregation, one result row */
3700 37990 : dNumGroups = 1;
3701 : }
3702 : else
3703 : {
3704 : /* Not grouping */
3705 0 : dNumGroups = 1;
3706 : }
3707 :
3708 44974 : return dNumGroups;
3709 : }
3710 :
3711 : /*
3712 : * create_grouping_paths
3713 : *
3714 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3715 : * Along the way, we also build an upperrel for Paths which are partially
3716 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3717 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3718 : * the only partially grouped paths we build are also partial paths; that
3719 : * is, they need a Gather and then a FinalizeAggregate.
3720 : *
3721 : * input_rel: contains the source-data Paths
3722 : * target: the pathtarget for the result Paths to compute
3723 : * gd: grouping sets data including list of grouping sets and their clauses
3724 : *
3725 : * Note: all Paths in input_rel are expected to return the target computed
3726 : * by make_group_input_target.
3727 : */
3728 : static RelOptInfo *
3729 41716 : create_grouping_paths(PlannerInfo *root,
3730 : RelOptInfo *input_rel,
3731 : PathTarget *target,
3732 : bool target_parallel_safe,
3733 : grouping_sets_data *gd)
3734 : {
3735 41716 : Query *parse = root->parse;
3736 : RelOptInfo *grouped_rel;
3737 : RelOptInfo *partially_grouped_rel;
3738 : AggClauseCosts agg_costs;
3739 :
3740 250296 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3741 41716 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3742 :
3743 : /*
3744 : * Create grouping relation to hold fully aggregated grouping and/or
3745 : * aggregation paths.
3746 : */
3747 41716 : grouped_rel = make_grouping_rel(root, input_rel, target,
3748 : target_parallel_safe, parse->havingQual);
3749 :
3750 : /*
3751 : * Create either paths for a degenerate grouping or paths for ordinary
3752 : * grouping, as appropriate.
3753 : */
3754 41716 : if (is_degenerate_grouping(root))
3755 18 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3756 : else
3757 : {
3758 41698 : int flags = 0;
3759 : GroupPathExtraData extra;
3760 :
3761 : /*
3762 : * Determine whether it's possible to perform sort-based
3763 : * implementations of grouping. (Note that if processed_groupClause
3764 : * is empty, grouping_is_sortable() is trivially true, and all the
3765 : * pathkeys_contained_in() tests will succeed too, so that we'll
3766 : * consider every surviving input path.)
3767 : *
3768 : * If we have grouping sets, we might be able to sort some but not all
3769 : * of them; in this case, we need can_sort to be true as long as we
3770 : * must consider any sorted-input plan.
3771 : */
3772 41698 : if ((gd && gd->rollups != NIL)
3773 40832 : || grouping_is_sortable(root->processed_groupClause))
3774 41692 : flags |= GROUPING_CAN_USE_SORT;
3775 :
3776 : /*
3777 : * Determine whether we should consider hash-based implementations of
3778 : * grouping.
3779 : *
3780 : * Hashed aggregation only applies if we're grouping. If we have
3781 : * grouping sets, some groups might be hashable but others not; in
3782 : * this case we set can_hash true as long as there is nothing globally
3783 : * preventing us from hashing (and we should therefore consider plans
3784 : * with hashes).
3785 : *
3786 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3787 : * BY aggregates. (Doing so would imply storing *all* the input
3788 : * values in the hash table, and/or running many sorts in parallel,
3789 : * either of which seems like a certain loser.) We similarly don't
3790 : * support ordered-set aggregates in hashed aggregation, but that case
3791 : * is also included in the numOrderedAggs count.
3792 : *
3793 : * Note: grouping_is_hashable() is much more expensive to check than
3794 : * the other gating conditions, so we want to do it last.
3795 : */
3796 41698 : if ((parse->groupClause != NIL &&
3797 8644 : root->numOrderedAggs == 0 &&
3798 4182 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3799 4178 : flags |= GROUPING_CAN_USE_HASH;
3800 :
3801 : /*
3802 : * Determine whether partial aggregation is possible.
3803 : */
3804 41698 : if (can_partial_agg(root))
3805 36746 : flags |= GROUPING_CAN_PARTIAL_AGG;
3806 :
3807 41698 : extra.flags = flags;
3808 41698 : extra.target_parallel_safe = target_parallel_safe;
3809 41698 : extra.havingQual = parse->havingQual;
3810 41698 : extra.targetList = parse->targetList;
3811 41698 : extra.partial_costs_set = false;
3812 :
3813 : /*
3814 : * Determine whether partitionwise aggregation is in theory possible.
3815 : * It can be disabled by the user, and for now, we don't try to
3816 : * support grouping sets. create_ordinary_grouping_paths() will check
3817 : * additional conditions, such as whether input_rel is partitioned.
3818 : */
3819 41698 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3820 556 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3821 : else
3822 41142 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3823 :
3824 41698 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3825 : &agg_costs, gd, &extra,
3826 : &partially_grouped_rel);
3827 : }
3828 :
3829 41710 : set_cheapest(grouped_rel);
3830 41710 : return grouped_rel;
3831 : }
3832 :
3833 : /*
3834 : * make_grouping_rel
3835 : *
3836 : * Create a new grouping rel and set basic properties.
3837 : *
3838 : * input_rel represents the underlying scan/join relation.
3839 : * target is the output expected from the grouping relation.
3840 : */
3841 : static RelOptInfo *
3842 43210 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3843 : PathTarget *target, bool target_parallel_safe,
3844 : Node *havingQual)
3845 : {
3846 : RelOptInfo *grouped_rel;
3847 :
3848 43210 : if (IS_OTHER_REL(input_rel))
3849 : {
3850 1494 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3851 : input_rel->relids);
3852 1494 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3853 : }
3854 : else
3855 : {
3856 : /*
3857 : * By tradition, the relids set for the main grouping relation is
3858 : * NULL. (This could be changed, but might require adjustments
3859 : * elsewhere.)
3860 : */
3861 41716 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3862 : }
3863 :
3864 : /* Set target. */
3865 43210 : grouped_rel->reltarget = target;
3866 :
3867 : /*
3868 : * If the input relation is not parallel-safe, then the grouped relation
3869 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3870 : * target list and HAVING quals are parallel-safe.
3871 : */
3872 73984 : if (input_rel->consider_parallel && target_parallel_safe &&
3873 30774 : is_parallel_safe(root, (Node *) havingQual))
3874 30756 : grouped_rel->consider_parallel = true;
3875 :
3876 : /*
3877 : * If the input rel belongs to a single FDW, so does the grouped rel.
3878 : */
3879 43210 : grouped_rel->serverid = input_rel->serverid;
3880 43210 : grouped_rel->userid = input_rel->userid;
3881 43210 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3882 43210 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3883 :
3884 43210 : return grouped_rel;
3885 : }
3886 :
3887 : /*
3888 : * is_degenerate_grouping
3889 : *
3890 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3891 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3892 : * grouping sets are all empty).
3893 : */
3894 : static bool
3895 41716 : is_degenerate_grouping(PlannerInfo *root)
3896 : {
3897 41716 : Query *parse = root->parse;
3898 :
3899 40698 : return (root->hasHavingQual || parse->groupingSets) &&
3900 82414 : !parse->hasAggs && parse->groupClause == NIL;
3901 : }
3902 :
3903 : /*
3904 : * create_degenerate_grouping_paths
3905 : *
3906 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3907 : * supposed to emit either zero or one row for each grouping set depending on
3908 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3909 : * either HAVING or the targetlist, so we actually do not need the FROM table
3910 : * at all! We can just throw away the plan-so-far and generate a Result node.
3911 : * This is a sufficiently unusual corner case that it's not worth contorting
3912 : * the structure of this module to avoid having to generate the earlier paths
3913 : * in the first place.
3914 : */
3915 : static void
3916 18 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3917 : RelOptInfo *grouped_rel)
3918 : {
3919 18 : Query *parse = root->parse;
3920 : int nrows;
3921 : Path *path;
3922 :
3923 18 : nrows = list_length(parse->groupingSets);
3924 18 : if (nrows > 1)
3925 : {
3926 : /*
3927 : * Doesn't seem worthwhile writing code to cons up a generate_series
3928 : * or a values scan to emit multiple rows. Instead just make N clones
3929 : * and append them. (With a volatile HAVING clause, this means you
3930 : * might get between 0 and N output rows. Offhand I think that's
3931 : * desired.)
3932 : */
3933 0 : List *paths = NIL;
3934 :
3935 0 : while (--nrows >= 0)
3936 : {
3937 : path = (Path *)
3938 0 : create_group_result_path(root, grouped_rel,
3939 0 : grouped_rel->reltarget,
3940 0 : (List *) parse->havingQual);
3941 0 : paths = lappend(paths, path);
3942 : }
3943 : path = (Path *)
3944 0 : create_append_path(root,
3945 : grouped_rel,
3946 : paths,
3947 : NIL,
3948 : NIL,
3949 : NULL,
3950 : 0,
3951 : false,
3952 : -1);
3953 : }
3954 : else
3955 : {
3956 : /* No grouping sets, or just one, so one output row */
3957 : path = (Path *)
3958 18 : create_group_result_path(root, grouped_rel,
3959 18 : grouped_rel->reltarget,
3960 18 : (List *) parse->havingQual);
3961 : }
3962 :
3963 18 : add_path(grouped_rel, path);
3964 18 : }
3965 :
3966 : /*
3967 : * create_ordinary_grouping_paths
3968 : *
3969 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
3970 : *
3971 : * We need to consider sorted and hashed aggregation in the same function,
3972 : * because otherwise (1) it would be harder to throw an appropriate error
3973 : * message if neither way works, and (2) we should not allow hashtable size
3974 : * considerations to dissuade us from using hashing if sorting is not possible.
3975 : *
3976 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
3977 : * function creates, or to NULL if it doesn't create one.
3978 : */
3979 : static void
3980 43192 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3981 : RelOptInfo *grouped_rel,
3982 : const AggClauseCosts *agg_costs,
3983 : grouping_sets_data *gd,
3984 : GroupPathExtraData *extra,
3985 : RelOptInfo **partially_grouped_rel_p)
3986 : {
3987 43192 : Path *cheapest_path = input_rel->cheapest_total_path;
3988 43192 : RelOptInfo *partially_grouped_rel = NULL;
3989 : double dNumGroups;
3990 43192 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
3991 :
3992 : /*
3993 : * If this is the topmost grouping relation or if the parent relation is
3994 : * doing some form of partitionwise aggregation, then we may be able to do
3995 : * it at this level also. However, if the input relation is not
3996 : * partitioned, partitionwise aggregate is impossible.
3997 : */
3998 43192 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
3999 2050 : IS_PARTITIONED_REL(input_rel))
4000 : {
4001 : /*
4002 : * If this is the topmost relation or if the parent relation is doing
4003 : * full partitionwise aggregation, then we can do full partitionwise
4004 : * aggregation provided that the GROUP BY clause contains all of the
4005 : * partitioning columns at this level and the collation used by GROUP
4006 : * BY matches the partitioning collation. Otherwise, we can do at
4007 : * most partial partitionwise aggregation. But if partial aggregation
4008 : * is not supported in general then we can't use it for partitionwise
4009 : * aggregation either.
4010 : *
4011 : * Check parse->groupClause not processed_groupClause, because it's
4012 : * okay if some of the partitioning columns were proved redundant.
4013 : */
4014 1160 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4015 556 : group_by_has_partkey(input_rel, extra->targetList,
4016 556 : root->parse->groupClause))
4017 320 : patype = PARTITIONWISE_AGGREGATE_FULL;
4018 284 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4019 242 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4020 : else
4021 42 : patype = PARTITIONWISE_AGGREGATE_NONE;
4022 : }
4023 :
4024 : /*
4025 : * Before generating paths for grouped_rel, we first generate any possible
4026 : * partially grouped paths; that way, later code can easily consider both
4027 : * parallel and non-parallel approaches to grouping.
4028 : */
4029 43192 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4030 : {
4031 : bool force_rel_creation;
4032 :
4033 : /*
4034 : * If we're doing partitionwise aggregation at this level, force
4035 : * creation of a partially_grouped_rel so we can add partitionwise
4036 : * paths to it.
4037 : */
4038 38168 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4039 :
4040 : partially_grouped_rel =
4041 38168 : create_partial_grouping_paths(root,
4042 : grouped_rel,
4043 : input_rel,
4044 : gd,
4045 : extra,
4046 : force_rel_creation);
4047 : }
4048 :
4049 : /* Set out parameter. */
4050 43192 : *partially_grouped_rel_p = partially_grouped_rel;
4051 :
4052 : /* Apply partitionwise aggregation technique, if possible. */
4053 43192 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4054 562 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4055 : partially_grouped_rel, agg_costs,
4056 : gd, patype, extra);
4057 :
4058 : /* If we are doing partial aggregation only, return. */
4059 43192 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4060 : {
4061 : Assert(partially_grouped_rel);
4062 :
4063 618 : if (partially_grouped_rel->pathlist)
4064 618 : set_cheapest(partially_grouped_rel);
4065 :
4066 618 : return;
4067 : }
4068 :
4069 : /* Gather any partially grouped partial paths. */
4070 42574 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4071 : {
4072 1482 : gather_grouping_paths(root, partially_grouped_rel);
4073 1482 : set_cheapest(partially_grouped_rel);
4074 : }
4075 :
4076 : /*
4077 : * Estimate number of groups.
4078 : */
4079 42574 : dNumGroups = get_number_of_groups(root,
4080 : cheapest_path->rows,
4081 : gd,
4082 : extra->targetList);
4083 :
4084 : /* Build final grouping paths */
4085 42574 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4086 : partially_grouped_rel, agg_costs, gd,
4087 : dNumGroups, extra);
4088 :
4089 : /* Give a helpful error if we failed to find any implementation */
4090 42574 : if (grouped_rel->pathlist == NIL)
4091 6 : ereport(ERROR,
4092 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4093 : errmsg("could not implement GROUP BY"),
4094 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4095 :
4096 : /*
4097 : * If there is an FDW that's responsible for all baserels of the query,
4098 : * let it consider adding ForeignPaths.
4099 : */
4100 42568 : if (grouped_rel->fdwroutine &&
4101 334 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4102 334 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4103 : input_rel, grouped_rel,
4104 : extra);
4105 :
4106 : /* Let extensions possibly add some more paths */
4107 42568 : if (create_upper_paths_hook)
4108 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4109 : input_rel, grouped_rel,
4110 : extra);
4111 : }
4112 :
4113 : /*
4114 : * For a given input path, consider the possible ways of doing grouping sets on
4115 : * it, by combinations of hashing and sorting. This can be called multiple
4116 : * times, so it's important that it not scribble on input. No result is
4117 : * returned, but any generated paths are added to grouped_rel.
4118 : */
4119 : static void
4120 1732 : consider_groupingsets_paths(PlannerInfo *root,
4121 : RelOptInfo *grouped_rel,
4122 : Path *path,
4123 : bool is_sorted,
4124 : bool can_hash,
4125 : grouping_sets_data *gd,
4126 : const AggClauseCosts *agg_costs,
4127 : double dNumGroups)
4128 : {
4129 1732 : Query *parse = root->parse;
4130 1732 : Size hash_mem_limit = get_hash_memory_limit();
4131 :
4132 : /*
4133 : * If we're not being offered sorted input, then only consider plans that
4134 : * can be done entirely by hashing.
4135 : *
4136 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4137 : * the input is actually sorted despite not being advertised as such, we
4138 : * prefer to make use of that in order to use less memory.
4139 : *
4140 : * If none of the grouping sets are sortable, then ignore the hash_mem
4141 : * limit and generate a path anyway, since otherwise we'll just fail.
4142 : */
4143 1732 : if (!is_sorted)
4144 : {
4145 794 : List *new_rollups = NIL;
4146 794 : RollupData *unhashed_rollup = NULL;
4147 : List *sets_data;
4148 794 : List *empty_sets_data = NIL;
4149 794 : List *empty_sets = NIL;
4150 : ListCell *lc;
4151 794 : ListCell *l_start = list_head(gd->rollups);
4152 794 : AggStrategy strat = AGG_HASHED;
4153 : double hashsize;
4154 794 : double exclude_groups = 0.0;
4155 :
4156 : Assert(can_hash);
4157 :
4158 : /*
4159 : * If the input is coincidentally sorted usefully (which can happen
4160 : * even if is_sorted is false, since that only means that our caller
4161 : * has set up the sorting for us), then save some hashtable space by
4162 : * making use of that. But we need to watch out for degenerate cases:
4163 : *
4164 : * 1) If there are any empty grouping sets, then group_pathkeys might
4165 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4166 : * there will be a rollup containing only empty groups, and the
4167 : * pathkeys_contained_in test is vacuously true; this is ok.
4168 : *
4169 : * XXX: the above relies on the fact that group_pathkeys is generated
4170 : * from the first rollup. If we add the ability to consider multiple
4171 : * sort orders for grouping input, this assumption might fail.
4172 : *
4173 : * 2) If there are no empty sets and only unsortable sets, then the
4174 : * rollups list will be empty (and thus l_start == NULL), and
4175 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4176 : * pathkeys_contained_in test doesn't cause us to crash.
4177 : */
4178 1582 : if (l_start != NULL &&
4179 788 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4180 : {
4181 12 : unhashed_rollup = lfirst_node(RollupData, l_start);
4182 12 : exclude_groups = unhashed_rollup->numGroups;
4183 12 : l_start = lnext(gd->rollups, l_start);
4184 : }
4185 :
4186 794 : hashsize = estimate_hashagg_tablesize(root,
4187 : path,
4188 : agg_costs,
4189 : dNumGroups - exclude_groups);
4190 :
4191 : /*
4192 : * gd->rollups is empty if we have only unsortable columns to work
4193 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4194 : * sorted-input case to generate usable mixed paths.
4195 : */
4196 794 : if (hashsize > hash_mem_limit && gd->rollups)
4197 18 : return; /* nope, won't fit */
4198 :
4199 : /*
4200 : * We need to burst the existing rollups list into individual grouping
4201 : * sets and recompute a groupClause for each set.
4202 : */
4203 776 : sets_data = list_copy(gd->unsortable_sets);
4204 :
4205 1968 : for_each_cell(lc, gd->rollups, l_start)
4206 : {
4207 1216 : RollupData *rollup = lfirst_node(RollupData, lc);
4208 :
4209 : /*
4210 : * If we find an unhashable rollup that's not been skipped by the
4211 : * "actually sorted" check above, we can't cope; we'd need sorted
4212 : * input (with a different sort order) but we can't get that here.
4213 : * So bail out; we'll get a valid path from the is_sorted case
4214 : * instead.
4215 : *
4216 : * The mere presence of empty grouping sets doesn't make a rollup
4217 : * unhashable (see preprocess_grouping_sets), we handle those
4218 : * specially below.
4219 : */
4220 1216 : if (!rollup->hashable)
4221 24 : return;
4222 :
4223 1192 : sets_data = list_concat(sets_data, rollup->gsets_data);
4224 : }
4225 3150 : foreach(lc, sets_data)
4226 : {
4227 2398 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4228 2398 : List *gset = gs->set;
4229 : RollupData *rollup;
4230 :
4231 2398 : if (gset == NIL)
4232 : {
4233 : /* Empty grouping sets can't be hashed. */
4234 484 : empty_sets_data = lappend(empty_sets_data, gs);
4235 484 : empty_sets = lappend(empty_sets, NIL);
4236 : }
4237 : else
4238 : {
4239 1914 : rollup = makeNode(RollupData);
4240 :
4241 1914 : rollup->groupClause = preprocess_groupclause(root, gset);
4242 1914 : rollup->gsets_data = list_make1(gs);
4243 1914 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4244 : rollup->gsets_data,
4245 : gd->tleref_to_colnum_map);
4246 1914 : rollup->numGroups = gs->numGroups;
4247 1914 : rollup->hashable = true;
4248 1914 : rollup->is_hashed = true;
4249 1914 : new_rollups = lappend(new_rollups, rollup);
4250 : }
4251 : }
4252 :
4253 : /*
4254 : * If we didn't find anything nonempty to hash, then bail. We'll
4255 : * generate a path from the is_sorted case.
4256 : */
4257 752 : if (new_rollups == NIL)
4258 0 : return;
4259 :
4260 : /*
4261 : * If there were empty grouping sets they should have been in the
4262 : * first rollup.
4263 : */
4264 : Assert(!unhashed_rollup || !empty_sets);
4265 :
4266 752 : if (unhashed_rollup)
4267 : {
4268 12 : new_rollups = lappend(new_rollups, unhashed_rollup);
4269 12 : strat = AGG_MIXED;
4270 : }
4271 740 : else if (empty_sets)
4272 : {
4273 436 : RollupData *rollup = makeNode(RollupData);
4274 :
4275 436 : rollup->groupClause = NIL;
4276 436 : rollup->gsets_data = empty_sets_data;
4277 436 : rollup->gsets = empty_sets;
4278 436 : rollup->numGroups = list_length(empty_sets);
4279 436 : rollup->hashable = false;
4280 436 : rollup->is_hashed = false;
4281 436 : new_rollups = lappend(new_rollups, rollup);
4282 436 : strat = AGG_MIXED;
4283 : }
4284 :
4285 752 : add_path(grouped_rel, (Path *)
4286 752 : create_groupingsets_path(root,
4287 : grouped_rel,
4288 : path,
4289 752 : (List *) parse->havingQual,
4290 : strat,
4291 : new_rollups,
4292 : agg_costs));
4293 752 : return;
4294 : }
4295 :
4296 : /*
4297 : * If we have sorted input but nothing we can do with it, bail.
4298 : */
4299 938 : if (gd->rollups == NIL)
4300 0 : return;
4301 :
4302 : /*
4303 : * Given sorted input, we try and make two paths: one sorted and one mixed
4304 : * sort/hash. (We need to try both because hashagg might be disabled, or
4305 : * some columns might not be sortable.)
4306 : *
4307 : * can_hash is passed in as false if some obstacle elsewhere (such as
4308 : * ordered aggs) means that we shouldn't consider hashing at all.
4309 : */
4310 938 : if (can_hash && gd->any_hashable)
4311 : {
4312 860 : List *rollups = NIL;
4313 860 : List *hash_sets = list_copy(gd->unsortable_sets);
4314 860 : double availspace = hash_mem_limit;
4315 : ListCell *lc;
4316 :
4317 : /*
4318 : * Account first for space needed for groups we can't sort at all.
4319 : */
4320 860 : availspace -= estimate_hashagg_tablesize(root,
4321 : path,
4322 : agg_costs,
4323 : gd->dNumHashGroups);
4324 :
4325 860 : if (availspace > 0 && list_length(gd->rollups) > 1)
4326 : {
4327 : double scale;
4328 444 : int num_rollups = list_length(gd->rollups);
4329 : int k_capacity;
4330 444 : int *k_weights = palloc(num_rollups * sizeof(int));
4331 444 : Bitmapset *hash_items = NULL;
4332 : int i;
4333 :
4334 : /*
4335 : * We treat this as a knapsack problem: the knapsack capacity
4336 : * represents hash_mem, the item weights are the estimated memory
4337 : * usage of the hashtables needed to implement a single rollup,
4338 : * and we really ought to use the cost saving as the item value;
4339 : * however, currently the costs assigned to sort nodes don't
4340 : * reflect the comparison costs well, and so we treat all items as
4341 : * of equal value (each rollup we hash instead saves us one sort).
4342 : *
4343 : * To use the discrete knapsack, we need to scale the values to a
4344 : * reasonably small bounded range. We choose to allow a 5% error
4345 : * margin; we have no more than 4096 rollups in the worst possible
4346 : * case, which with a 5% error margin will require a bit over 42MB
4347 : * of workspace. (Anyone wanting to plan queries that complex had
4348 : * better have the memory for it. In more reasonable cases, with
4349 : * no more than a couple of dozen rollups, the memory usage will
4350 : * be negligible.)
4351 : *
4352 : * k_capacity is naturally bounded, but we clamp the values for
4353 : * scale and weight (below) to avoid overflows or underflows (or
4354 : * uselessly trying to use a scale factor less than 1 byte).
4355 : */
4356 444 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4357 444 : k_capacity = (int) floor(availspace / scale);
4358 :
4359 : /*
4360 : * We leave the first rollup out of consideration since it's the
4361 : * one that matches the input sort order. We assign indexes "i"
4362 : * to only those entries considered for hashing; the second loop,
4363 : * below, must use the same condition.
4364 : */
4365 444 : i = 0;
4366 1128 : for_each_from(lc, gd->rollups, 1)
4367 : {
4368 684 : RollupData *rollup = lfirst_node(RollupData, lc);
4369 :
4370 684 : if (rollup->hashable)
4371 : {
4372 684 : double sz = estimate_hashagg_tablesize(root,
4373 : path,
4374 : agg_costs,
4375 : rollup->numGroups);
4376 :
4377 : /*
4378 : * If sz is enormous, but hash_mem (and hence scale) is
4379 : * small, avoid integer overflow here.
4380 : */
4381 684 : k_weights[i] = (int) Min(floor(sz / scale),
4382 : k_capacity + 1.0);
4383 684 : ++i;
4384 : }
4385 : }
4386 :
4387 : /*
4388 : * Apply knapsack algorithm; compute the set of items which
4389 : * maximizes the value stored (in this case the number of sorts
4390 : * saved) while keeping the total size (approximately) within
4391 : * capacity.
4392 : */
4393 444 : if (i > 0)
4394 444 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4395 :
4396 444 : if (!bms_is_empty(hash_items))
4397 : {
4398 444 : rollups = list_make1(linitial(gd->rollups));
4399 :
4400 444 : i = 0;
4401 1128 : for_each_from(lc, gd->rollups, 1)
4402 : {
4403 684 : RollupData *rollup = lfirst_node(RollupData, lc);
4404 :
4405 684 : if (rollup->hashable)
4406 : {
4407 684 : if (bms_is_member(i, hash_items))
4408 648 : hash_sets = list_concat(hash_sets,
4409 648 : rollup->gsets_data);
4410 : else
4411 36 : rollups = lappend(rollups, rollup);
4412 684 : ++i;
4413 : }
4414 : else
4415 0 : rollups = lappend(rollups, rollup);
4416 : }
4417 : }
4418 : }
4419 :
4420 860 : if (!rollups && hash_sets)
4421 24 : rollups = list_copy(gd->rollups);
4422 :
4423 1648 : foreach(lc, hash_sets)
4424 : {
4425 788 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4426 788 : RollupData *rollup = makeNode(RollupData);
4427 :
4428 : Assert(gs->set != NIL);
4429 :
4430 788 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4431 788 : rollup->gsets_data = list_make1(gs);
4432 788 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4433 : rollup->gsets_data,
4434 : gd->tleref_to_colnum_map);
4435 788 : rollup->numGroups = gs->numGroups;
4436 788 : rollup->hashable = true;
4437 788 : rollup->is_hashed = true;
4438 788 : rollups = lcons(rollup, rollups);
4439 : }
4440 :
4441 860 : if (rollups)
4442 : {
4443 468 : add_path(grouped_rel, (Path *)
4444 468 : create_groupingsets_path(root,
4445 : grouped_rel,
4446 : path,
4447 468 : (List *) parse->havingQual,
4448 : AGG_MIXED,
4449 : rollups,
4450 : agg_costs));
4451 : }
4452 : }
4453 :
4454 : /*
4455 : * Now try the simple sorted case.
4456 : */
4457 938 : if (!gd->unsortable_sets)
4458 908 : add_path(grouped_rel, (Path *)
4459 908 : create_groupingsets_path(root,
4460 : grouped_rel,
4461 : path,
4462 908 : (List *) parse->havingQual,
4463 : AGG_SORTED,
4464 : gd->rollups,
4465 : agg_costs));
4466 : }
4467 :
4468 : /*
4469 : * create_window_paths
4470 : *
4471 : * Build a new upperrel containing Paths for window-function evaluation.
4472 : *
4473 : * input_rel: contains the source-data Paths
4474 : * input_target: result of make_window_input_target
4475 : * output_target: what the topmost WindowAggPath should return
4476 : * wflists: result of find_window_functions
4477 : * activeWindows: result of select_active_windows
4478 : *
4479 : * Note: all Paths in input_rel are expected to return input_target.
4480 : */
4481 : static RelOptInfo *
4482 2378 : create_window_paths(PlannerInfo *root,
4483 : RelOptInfo *input_rel,
4484 : PathTarget *input_target,
4485 : PathTarget *output_target,
4486 : bool output_target_parallel_safe,
4487 : WindowFuncLists *wflists,
4488 : List *activeWindows)
4489 : {
4490 : RelOptInfo *window_rel;
4491 : ListCell *lc;
4492 :
4493 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4494 2378 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4495 :
4496 : /*
4497 : * If the input relation is not parallel-safe, then the window relation
4498 : * can't be parallel-safe, either. Otherwise, we need to examine the
4499 : * target list and active windows for non-parallel-safe constructs.
4500 : */
4501 2378 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4502 0 : is_parallel_safe(root, (Node *) activeWindows))
4503 0 : window_rel->consider_parallel = true;
4504 :
4505 : /*
4506 : * If the input rel belongs to a single FDW, so does the window rel.
4507 : */
4508 2378 : window_rel->serverid = input_rel->serverid;
4509 2378 : window_rel->userid = input_rel->userid;
4510 2378 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4511 2378 : window_rel->fdwroutine = input_rel->fdwroutine;
4512 :
4513 : /*
4514 : * Consider computing window functions starting from the existing
4515 : * cheapest-total path (which will likely require a sort) as well as any
4516 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4517 : */
4518 5078 : foreach(lc, input_rel->pathlist)
4519 : {
4520 2700 : Path *path = (Path *) lfirst(lc);
4521 : int presorted_keys;
4522 :
4523 3022 : if (path == input_rel->cheapest_total_path ||
4524 322 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4525 140 : &presorted_keys) ||
4526 140 : presorted_keys > 0)
4527 2586 : create_one_window_path(root,
4528 : window_rel,
4529 : path,
4530 : input_target,
4531 : output_target,
4532 : wflists,
4533 : activeWindows);
4534 : }
4535 :
4536 : /*
4537 : * If there is an FDW that's responsible for all baserels of the query,
4538 : * let it consider adding ForeignPaths.
4539 : */
4540 2378 : if (window_rel->fdwroutine &&
4541 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4542 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4543 : input_rel, window_rel,
4544 : NULL);
4545 :
4546 : /* Let extensions possibly add some more paths */
4547 2378 : if (create_upper_paths_hook)
4548 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4549 : input_rel, window_rel, NULL);
4550 :
4551 : /* Now choose the best path(s) */
4552 2378 : set_cheapest(window_rel);
4553 :
4554 2378 : return window_rel;
4555 : }
4556 :
4557 : /*
4558 : * Stack window-function implementation steps atop the given Path, and
4559 : * add the result to window_rel.
4560 : *
4561 : * window_rel: upperrel to contain result
4562 : * path: input Path to use (must return input_target)
4563 : * input_target: result of make_window_input_target
4564 : * output_target: what the topmost WindowAggPath should return
4565 : * wflists: result of find_window_functions
4566 : * activeWindows: result of select_active_windows
4567 : */
4568 : static void
4569 2586 : create_one_window_path(PlannerInfo *root,
4570 : RelOptInfo *window_rel,
4571 : Path *path,
4572 : PathTarget *input_target,
4573 : PathTarget *output_target,
4574 : WindowFuncLists *wflists,
4575 : List *activeWindows)
4576 : {
4577 : PathTarget *window_target;
4578 : ListCell *l;
4579 2586 : List *topqual = NIL;
4580 :
4581 : /*
4582 : * Since each window clause could require a different sort order, we stack
4583 : * up a WindowAgg node for each clause, with sort steps between them as
4584 : * needed. (We assume that select_active_windows chose a good order for
4585 : * executing the clauses in.)
4586 : *
4587 : * input_target should contain all Vars and Aggs needed for the result.
4588 : * (In some cases we wouldn't need to propagate all of these all the way
4589 : * to the top, since they might only be needed as inputs to WindowFuncs.
4590 : * It's probably not worth trying to optimize that though.) It must also
4591 : * contain all window partitioning and sorting expressions, to ensure
4592 : * they're computed only once at the bottom of the stack (that's critical
4593 : * for volatile functions). As we climb up the stack, we'll add outputs
4594 : * for the WindowFuncs computed at each level.
4595 : */
4596 2586 : window_target = input_target;
4597 :
4598 5340 : foreach(l, activeWindows)
4599 : {
4600 2754 : WindowClause *wc = lfirst_node(WindowClause, l);
4601 : List *window_pathkeys;
4602 2754 : List *runcondition = NIL;
4603 : int presorted_keys;
4604 : bool is_sorted;
4605 : bool topwindow;
4606 : ListCell *lc2;
4607 :
4608 2754 : window_pathkeys = make_pathkeys_for_window(root,
4609 : wc,
4610 : root->processed_tlist);
4611 :
4612 2754 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4613 : path->pathkeys,
4614 : &presorted_keys);
4615 :
4616 : /* Sort if necessary */
4617 2754 : if (!is_sorted)
4618 : {
4619 : /*
4620 : * No presorted keys or incremental sort disabled, just perform a
4621 : * complete sort.
4622 : */
4623 2106 : if (presorted_keys == 0 || !enable_incremental_sort)
4624 2044 : path = (Path *) create_sort_path(root, window_rel,
4625 : path,
4626 : window_pathkeys,
4627 : -1.0);
4628 : else
4629 : {
4630 : /*
4631 : * Since we have presorted keys and incremental sort is
4632 : * enabled, just use incremental sort.
4633 : */
4634 62 : path = (Path *) create_incremental_sort_path(root,
4635 : window_rel,
4636 : path,
4637 : window_pathkeys,
4638 : presorted_keys,
4639 : -1.0);
4640 : }
4641 : }
4642 :
4643 2754 : if (lnext(activeWindows, l))
4644 : {
4645 : /*
4646 : * Add the current WindowFuncs to the output target for this
4647 : * intermediate WindowAggPath. We must copy window_target to
4648 : * avoid changing the previous path's target.
4649 : *
4650 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4651 : * we do need to account for the increase in tlist width.
4652 : */
4653 168 : int64 tuple_width = window_target->width;
4654 :
4655 168 : window_target = copy_pathtarget(window_target);
4656 384 : foreach(lc2, wflists->windowFuncs[wc->winref])
4657 : {
4658 216 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4659 :
4660 216 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4661 216 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4662 : }
4663 168 : window_target->width = clamp_width_est(tuple_width);
4664 : }
4665 : else
4666 : {
4667 : /* Install the goal target in the topmost WindowAgg */
4668 2586 : window_target = output_target;
4669 : }
4670 :
4671 : /* mark the final item in the list as the top-level window */
4672 2754 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4673 :
4674 : /*
4675 : * Collect the WindowFuncRunConditions from each WindowFunc and
4676 : * convert them into OpExprs
4677 : */
4678 6246 : foreach(lc2, wflists->windowFuncs[wc->winref])
4679 : {
4680 : ListCell *lc3;
4681 3492 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4682 :
4683 3672 : foreach(lc3, wfunc->runCondition)
4684 : {
4685 180 : WindowFuncRunCondition *wfuncrc =
4686 : lfirst_node(WindowFuncRunCondition, lc3);
4687 : Expr *opexpr;
4688 : Expr *leftop;
4689 : Expr *rightop;
4690 :
4691 180 : if (wfuncrc->wfunc_left)
4692 : {
4693 162 : leftop = (Expr *) copyObject(wfunc);
4694 162 : rightop = copyObject(wfuncrc->arg);
4695 : }
4696 : else
4697 : {
4698 18 : leftop = copyObject(wfuncrc->arg);
4699 18 : rightop = (Expr *) copyObject(wfunc);
4700 : }
4701 :
4702 180 : opexpr = make_opclause(wfuncrc->opno,
4703 : BOOLOID,
4704 : false,
4705 : leftop,
4706 : rightop,
4707 : InvalidOid,
4708 : wfuncrc->inputcollid);
4709 :
4710 180 : runcondition = lappend(runcondition, opexpr);
4711 :
4712 180 : if (!topwindow)
4713 24 : topqual = lappend(topqual, opexpr);
4714 : }
4715 : }
4716 :
4717 : path = (Path *)
4718 2754 : create_windowagg_path(root, window_rel, path, window_target,
4719 2754 : wflists->windowFuncs[wc->winref],
4720 : runcondition, wc,
4721 : topwindow ? topqual : NIL, topwindow);
4722 : }
4723 :
4724 2586 : add_path(window_rel, path);
4725 2586 : }
4726 :
4727 : /*
4728 : * create_distinct_paths
4729 : *
4730 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4731 : *
4732 : * input_rel: contains the source-data Paths
4733 : * target: the pathtarget for the result Paths to compute
4734 : *
4735 : * Note: input paths should already compute the desired pathtarget, since
4736 : * Sort/Unique won't project anything.
4737 : */
4738 : static RelOptInfo *
4739 2676 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4740 : PathTarget *target)
4741 : {
4742 : RelOptInfo *distinct_rel;
4743 :
4744 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4745 2676 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4746 :
4747 : /*
4748 : * We don't compute anything at this level, so distinct_rel will be
4749 : * parallel-safe if the input rel is parallel-safe. In particular, if
4750 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4751 : * output those expressions, and will not be parallel-safe unless those
4752 : * expressions are parallel-safe.
4753 : */
4754 2676 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4755 :
4756 : /*
4757 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4758 : */
4759 2676 : distinct_rel->serverid = input_rel->serverid;
4760 2676 : distinct_rel->userid = input_rel->userid;
4761 2676 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4762 2676 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4763 :
4764 : /* build distinct paths based on input_rel's pathlist */
4765 2676 : create_final_distinct_paths(root, input_rel, distinct_rel);
4766 :
4767 : /* now build distinct paths based on input_rel's partial_pathlist */
4768 2676 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4769 :
4770 : /* Give a helpful error if we failed to create any paths */
4771 2676 : if (distinct_rel->pathlist == NIL)
4772 0 : ereport(ERROR,
4773 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4774 : errmsg("could not implement DISTINCT"),
4775 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4776 :
4777 : /*
4778 : * If there is an FDW that's responsible for all baserels of the query,
4779 : * let it consider adding ForeignPaths.
4780 : */
4781 2676 : if (distinct_rel->fdwroutine &&
4782 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4783 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4784 : UPPERREL_DISTINCT,
4785 : input_rel,
4786 : distinct_rel,
4787 : NULL);
4788 :
4789 : /* Let extensions possibly add some more paths */
4790 2676 : if (create_upper_paths_hook)
4791 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4792 : distinct_rel, NULL);
4793 :
4794 : /* Now choose the best path(s) */
4795 2676 : set_cheapest(distinct_rel);
4796 :
4797 2676 : return distinct_rel;
4798 : }
4799 :
4800 : /*
4801 : * create_partial_distinct_paths
4802 : *
4803 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4804 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4805 : * paths on top and add a final unique/aggregate path to remove any duplicate
4806 : * produced from combining rows from parallel workers.
4807 : */
4808 : static void
4809 2676 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4810 : RelOptInfo *final_distinct_rel,
4811 : PathTarget *target)
4812 : {
4813 : RelOptInfo *partial_distinct_rel;
4814 : Query *parse;
4815 : List *distinctExprs;
4816 : double numDistinctRows;
4817 : Path *cheapest_partial_path;
4818 : ListCell *lc;
4819 :
4820 : /* nothing to do when there are no partial paths in the input rel */
4821 2676 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4822 2568 : return;
4823 :
4824 108 : parse = root->parse;
4825 :
4826 : /* can't do parallel DISTINCT ON */
4827 108 : if (parse->hasDistinctOn)
4828 0 : return;
4829 :
4830 108 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4831 : NULL);
4832 108 : partial_distinct_rel->reltarget = target;
4833 108 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4834 :
4835 : /*
4836 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4837 : */
4838 108 : partial_distinct_rel->serverid = input_rel->serverid;
4839 108 : partial_distinct_rel->userid = input_rel->userid;
4840 108 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4841 108 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4842 :
4843 108 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4844 :
4845 108 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4846 : parse->targetList);
4847 :
4848 : /* estimate how many distinct rows we'll get from each worker */
4849 108 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4850 : cheapest_partial_path->rows,
4851 : NULL, NULL);
4852 :
4853 : /*
4854 : * Try sorting the cheapest path and incrementally sorting any paths with
4855 : * presorted keys and put a unique paths atop of those. We'll also
4856 : * attempt to reorder the required pathkeys to match the input path's
4857 : * pathkeys as much as possible, in hopes of avoiding a possible need to
4858 : * re-sort.
4859 : */
4860 108 : if (grouping_is_sortable(root->processed_distinctClause))
4861 : {
4862 234 : foreach(lc, input_rel->partial_pathlist)
4863 : {
4864 126 : Path *input_path = (Path *) lfirst(lc);
4865 : Path *sorted_path;
4866 126 : List *useful_pathkeys_list = NIL;
4867 :
4868 : useful_pathkeys_list =
4869 126 : get_useful_pathkeys_for_distinct(root,
4870 : root->distinct_pathkeys,
4871 : input_path->pathkeys);
4872 : Assert(list_length(useful_pathkeys_list) > 0);
4873 :
4874 390 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
4875 : {
4876 138 : sorted_path = make_ordered_path(root,
4877 : partial_distinct_rel,
4878 : input_path,
4879 : cheapest_partial_path,
4880 : useful_pathkeys,
4881 : -1.0);
4882 :
4883 138 : if (sorted_path == NULL)
4884 12 : continue;
4885 :
4886 : /*
4887 : * An empty distinct_pathkeys means all tuples have the same
4888 : * value for the DISTINCT clause. See
4889 : * create_final_distinct_paths()
4890 : */
4891 126 : if (root->distinct_pathkeys == NIL)
4892 : {
4893 : Node *limitCount;
4894 :
4895 6 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
4896 : sizeof(int64),
4897 : Int64GetDatum(1), false,
4898 : FLOAT8PASSBYVAL);
4899 :
4900 : /*
4901 : * Apply a LimitPath onto the partial path to restrict the
4902 : * tuples from each worker to 1.
4903 : * create_final_distinct_paths will need to apply an
4904 : * additional LimitPath to restrict this to a single row
4905 : * after the Gather node. If the query already has a
4906 : * LIMIT clause, then we could end up with three Limit
4907 : * nodes in the final plan. Consolidating the top two of
4908 : * these could be done, but does not seem worth troubling
4909 : * over.
4910 : */
4911 6 : add_partial_path(partial_distinct_rel, (Path *)
4912 6 : create_limit_path(root, partial_distinct_rel,
4913 : sorted_path,
4914 : NULL,
4915 : limitCount,
4916 : LIMIT_OPTION_COUNT,
4917 : 0, 1));
4918 : }
4919 : else
4920 : {
4921 120 : add_partial_path(partial_distinct_rel, (Path *)
4922 120 : create_upper_unique_path(root, partial_distinct_rel,
4923 : sorted_path,
4924 120 : list_length(root->distinct_pathkeys),
4925 : numDistinctRows));
4926 : }
4927 : }
4928 : }
4929 : }
4930 :
4931 : /*
4932 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
4933 : * we're not on the hook to ensure we do our best to create at least one
4934 : * path here, we treat enable_hashagg as a hard off-switch rather than the
4935 : * slightly softer variant in create_final_distinct_paths.
4936 : */
4937 108 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
4938 : {
4939 78 : add_partial_path(partial_distinct_rel, (Path *)
4940 78 : create_agg_path(root,
4941 : partial_distinct_rel,
4942 : cheapest_partial_path,
4943 : cheapest_partial_path->pathtarget,
4944 : AGG_HASHED,
4945 : AGGSPLIT_SIMPLE,
4946 : root->processed_distinctClause,
4947 : NIL,
4948 : NULL,
4949 : numDistinctRows));
4950 : }
4951 :
4952 : /*
4953 : * If there is an FDW that's responsible for all baserels of the query,
4954 : * let it consider adding ForeignPaths.
4955 : */
4956 108 : if (partial_distinct_rel->fdwroutine &&
4957 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
4958 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4959 : UPPERREL_PARTIAL_DISTINCT,
4960 : input_rel,
4961 : partial_distinct_rel,
4962 : NULL);
4963 :
4964 : /* Let extensions possibly add some more partial paths */
4965 108 : if (create_upper_paths_hook)
4966 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
4967 : input_rel, partial_distinct_rel, NULL);
4968 :
4969 108 : if (partial_distinct_rel->partial_pathlist != NIL)
4970 : {
4971 108 : generate_useful_gather_paths(root, partial_distinct_rel, true);
4972 108 : set_cheapest(partial_distinct_rel);
4973 :
4974 : /*
4975 : * Finally, create paths to distinctify the final result. This step
4976 : * is needed to remove any duplicates due to combining rows from
4977 : * parallel workers.
4978 : */
4979 108 : create_final_distinct_paths(root, partial_distinct_rel,
4980 : final_distinct_rel);
4981 : }
4982 : }
4983 :
4984 : /*
4985 : * create_final_distinct_paths
4986 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
4987 : *
4988 : * input_rel: contains the source-data paths
4989 : * distinct_rel: destination relation for storing created paths
4990 : */
4991 : static RelOptInfo *
4992 2784 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4993 : RelOptInfo *distinct_rel)
4994 : {
4995 2784 : Query *parse = root->parse;
4996 2784 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4997 : double numDistinctRows;
4998 : bool allow_hash;
4999 :
5000 : /* Estimate number of distinct rows there will be */
5001 2784 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
5002 2710 : root->hasHavingQual)
5003 : {
5004 : /*
5005 : * If there was grouping or aggregation, use the number of input rows
5006 : * as the estimated number of DISTINCT rows (ie, assume the input is
5007 : * already mostly unique).
5008 : */
5009 74 : numDistinctRows = cheapest_input_path->rows;
5010 : }
5011 : else
5012 : {
5013 : /*
5014 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5015 : */
5016 : List *distinctExprs;
5017 :
5018 2710 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5019 : parse->targetList);
5020 2710 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5021 : cheapest_input_path->rows,
5022 : NULL, NULL);
5023 : }
5024 :
5025 : /*
5026 : * Consider sort-based implementations of DISTINCT, if possible.
5027 : */
5028 2784 : if (grouping_is_sortable(root->processed_distinctClause))
5029 : {
5030 : /*
5031 : * Firstly, if we have any adequately-presorted paths, just stick a
5032 : * Unique node on those. We also, consider doing an explicit sort of
5033 : * the cheapest input path and Unique'ing that. If any paths have
5034 : * presorted keys then we'll create an incremental sort atop of those
5035 : * before adding a unique node on the top. We'll also attempt to
5036 : * reorder the required pathkeys to match the input path's pathkeys as
5037 : * much as possible, in hopes of avoiding a possible need to re-sort.
5038 : *
5039 : * When we have DISTINCT ON, we must sort by the more rigorous of
5040 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5041 : * Also, if we do have to do an explicit sort, we might as well use
5042 : * the more rigorous ordering to avoid a second sort later. (Note
5043 : * that the parser will have ensured that one clause is a prefix of
5044 : * the other.)
5045 : */
5046 : List *needed_pathkeys;
5047 : ListCell *lc;
5048 2778 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5049 :
5050 3016 : if (parse->hasDistinctOn &&
5051 238 : list_length(root->distinct_pathkeys) <
5052 238 : list_length(root->sort_pathkeys))
5053 54 : needed_pathkeys = root->sort_pathkeys;
5054 : else
5055 2724 : needed_pathkeys = root->distinct_pathkeys;
5056 :
5057 7160 : foreach(lc, input_rel->pathlist)
5058 : {
5059 4382 : Path *input_path = (Path *) lfirst(lc);
5060 : Path *sorted_path;
5061 4382 : List *useful_pathkeys_list = NIL;
5062 :
5063 : useful_pathkeys_list =
5064 4382 : get_useful_pathkeys_for_distinct(root,
5065 : needed_pathkeys,
5066 : input_path->pathkeys);
5067 : Assert(list_length(useful_pathkeys_list) > 0);
5068 :
5069 13632 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5070 : {
5071 4868 : sorted_path = make_ordered_path(root,
5072 : distinct_rel,
5073 : input_path,
5074 : cheapest_input_path,
5075 : useful_pathkeys,
5076 : limittuples);
5077 :
5078 4868 : if (sorted_path == NULL)
5079 544 : continue;
5080 :
5081 : /*
5082 : * distinct_pathkeys may have become empty if all of the
5083 : * pathkeys were determined to be redundant. If all of the
5084 : * pathkeys are redundant then each DISTINCT target must only
5085 : * allow a single value, therefore all resulting tuples must
5086 : * be identical (or at least indistinguishable by an equality
5087 : * check). We can uniquify these tuples simply by just taking
5088 : * the first tuple. All we do here is add a path to do "LIMIT
5089 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5090 : * still have a non-NIL sort_pathkeys list, so we must still
5091 : * only do this with paths which are correctly sorted by
5092 : * sort_pathkeys.
5093 : */
5094 4324 : if (root->distinct_pathkeys == NIL)
5095 : {
5096 : Node *limitCount;
5097 :
5098 106 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5099 : sizeof(int64),
5100 : Int64GetDatum(1), false,
5101 : FLOAT8PASSBYVAL);
5102 :
5103 : /*
5104 : * If the query already has a LIMIT clause, then we could
5105 : * end up with a duplicate LimitPath in the final plan.
5106 : * That does not seem worth troubling over too much.
5107 : */
5108 106 : add_path(distinct_rel, (Path *)
5109 106 : create_limit_path(root, distinct_rel, sorted_path,
5110 : NULL, limitCount,
5111 : LIMIT_OPTION_COUNT, 0, 1));
5112 : }
5113 : else
5114 : {
5115 4218 : add_path(distinct_rel, (Path *)
5116 4218 : create_upper_unique_path(root, distinct_rel,
5117 : sorted_path,
5118 4218 : list_length(root->distinct_pathkeys),
5119 : numDistinctRows));
5120 : }
5121 : }
5122 : }
5123 : }
5124 :
5125 : /*
5126 : * Consider hash-based implementations of DISTINCT, if possible.
5127 : *
5128 : * If we were not able to make any other types of path, we *must* hash or
5129 : * die trying. If we do have other choices, there are two things that
5130 : * should prevent selection of hashing: if the query uses DISTINCT ON
5131 : * (because it won't really have the expected behavior if we hash), or if
5132 : * enable_hashagg is off.
5133 : *
5134 : * Note: grouping_is_hashable() is much more expensive to check than the
5135 : * other gating conditions, so we want to do it last.
5136 : */
5137 2784 : if (distinct_rel->pathlist == NIL)
5138 6 : allow_hash = true; /* we have no alternatives */
5139 2778 : else if (parse->hasDistinctOn || !enable_hashagg)
5140 388 : allow_hash = false; /* policy-based decision not to hash */
5141 : else
5142 2390 : allow_hash = true; /* default */
5143 :
5144 2784 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5145 : {
5146 : /* Generate hashed aggregate path --- no sort needed */
5147 2396 : add_path(distinct_rel, (Path *)
5148 2396 : create_agg_path(root,
5149 : distinct_rel,
5150 : cheapest_input_path,
5151 : cheapest_input_path->pathtarget,
5152 : AGG_HASHED,
5153 : AGGSPLIT_SIMPLE,
5154 : root->processed_distinctClause,
5155 : NIL,
5156 : NULL,
5157 : numDistinctRows));
5158 : }
5159 :
5160 2784 : return distinct_rel;
5161 : }
5162 :
5163 : /*
5164 : * get_useful_pathkeys_for_distinct
5165 : * Get useful orderings of pathkeys for distinctClause by reordering
5166 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5167 : *
5168 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5169 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5170 : */
5171 : static List *
5172 4508 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5173 : List *path_pathkeys)
5174 : {
5175 4508 : List *useful_pathkeys_list = NIL;
5176 4508 : List *useful_pathkeys = NIL;
5177 :
5178 : /* always include the given 'needed_pathkeys' */
5179 4508 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5180 : needed_pathkeys);
5181 :
5182 4508 : if (!enable_distinct_reordering)
5183 0 : return useful_pathkeys_list;
5184 :
5185 : /*
5186 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5187 : * that match 'needed_pathkeys', but only up to the longest matching
5188 : * prefix.
5189 : *
5190 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5191 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5192 : * the desired behavior.
5193 : */
5194 11024 : foreach_node(PathKey, pathkey, path_pathkeys)
5195 : {
5196 : /*
5197 : * The PathKey nodes are canonical, so they can be checked for
5198 : * equality by simple pointer comparison.
5199 : */
5200 2036 : if (!list_member_ptr(needed_pathkeys, pathkey))
5201 10 : break;
5202 2026 : if (root->parse->hasDistinctOn &&
5203 202 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5204 18 : break;
5205 :
5206 2008 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5207 : }
5208 :
5209 : /* If no match at all, no point in reordering needed_pathkeys */
5210 4508 : if (useful_pathkeys == NIL)
5211 2764 : return useful_pathkeys_list;
5212 :
5213 : /*
5214 : * If not full match, the resulting pathkey list is not useful without
5215 : * incremental sort.
5216 : */
5217 1744 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5218 894 : !enable_incremental_sort)
5219 60 : return useful_pathkeys_list;
5220 :
5221 : /* Append the remaining PathKey nodes in needed_pathkeys */
5222 1684 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5223 : needed_pathkeys);
5224 :
5225 : /*
5226 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5227 : * just drop it.
5228 : */
5229 1684 : if (compare_pathkeys(needed_pathkeys,
5230 : useful_pathkeys) == PATHKEYS_EQUAL)
5231 1186 : return useful_pathkeys_list;
5232 :
5233 498 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5234 : useful_pathkeys);
5235 :
5236 498 : return useful_pathkeys_list;
5237 : }
5238 :
5239 : /*
5240 : * create_ordered_paths
5241 : *
5242 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5243 : *
5244 : * All paths in the result must satisfy the ORDER BY ordering.
5245 : * The only new paths we need consider are an explicit full sort
5246 : * and incremental sort on the cheapest-total existing path.
5247 : *
5248 : * input_rel: contains the source-data Paths
5249 : * target: the output tlist the result Paths must emit
5250 : * limit_tuples: estimated bound on the number of output tuples,
5251 : * or -1 if no LIMIT or couldn't estimate
5252 : *
5253 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5254 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5255 : */
5256 : static RelOptInfo *
5257 75922 : create_ordered_paths(PlannerInfo *root,
5258 : RelOptInfo *input_rel,
5259 : PathTarget *target,
5260 : bool target_parallel_safe,
5261 : double limit_tuples)
5262 : {
5263 75922 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5264 : RelOptInfo *ordered_rel;
5265 : ListCell *lc;
5266 :
5267 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5268 75922 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5269 :
5270 : /*
5271 : * If the input relation is not parallel-safe, then the ordered relation
5272 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5273 : * target list is parallel-safe.
5274 : */
5275 75922 : if (input_rel->consider_parallel && target_parallel_safe)
5276 52640 : ordered_rel->consider_parallel = true;
5277 :
5278 : /*
5279 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5280 : */
5281 75922 : ordered_rel->serverid = input_rel->serverid;
5282 75922 : ordered_rel->userid = input_rel->userid;
5283 75922 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5284 75922 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5285 :
5286 191890 : foreach(lc, input_rel->pathlist)
5287 : {
5288 115968 : Path *input_path = (Path *) lfirst(lc);
5289 : Path *sorted_path;
5290 : bool is_sorted;
5291 : int presorted_keys;
5292 :
5293 115968 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5294 : input_path->pathkeys, &presorted_keys);
5295 :
5296 115968 : if (is_sorted)
5297 42908 : sorted_path = input_path;
5298 : else
5299 : {
5300 : /*
5301 : * Try at least sorting the cheapest path and also try
5302 : * incrementally sorting any path which is partially sorted
5303 : * already (no need to deal with paths which have presorted keys
5304 : * when incremental sort is disabled unless it's the cheapest
5305 : * input path).
5306 : */
5307 73060 : if (input_path != cheapest_input_path &&
5308 6298 : (presorted_keys == 0 || !enable_incremental_sort))
5309 1836 : continue;
5310 :
5311 : /*
5312 : * We've no need to consider both a sort and incremental sort.
5313 : * We'll just do a sort if there are no presorted keys and an
5314 : * incremental sort when there are presorted keys.
5315 : */
5316 71224 : if (presorted_keys == 0 || !enable_incremental_sort)
5317 65896 : sorted_path = (Path *) create_sort_path(root,
5318 : ordered_rel,
5319 : input_path,
5320 : root->sort_pathkeys,
5321 : limit_tuples);
5322 : else
5323 5328 : sorted_path = (Path *) create_incremental_sort_path(root,
5324 : ordered_rel,
5325 : input_path,
5326 : root->sort_pathkeys,
5327 : presorted_keys,
5328 : limit_tuples);
5329 : }
5330 :
5331 : /*
5332 : * If the pathtarget of the result path has different expressions from
5333 : * the target to be applied, a projection step is needed.
5334 : */
5335 114132 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5336 294 : sorted_path = apply_projection_to_path(root, ordered_rel,
5337 : sorted_path, target);
5338 :
5339 114132 : add_path(ordered_rel, sorted_path);
5340 : }
5341 :
5342 : /*
5343 : * generate_gather_paths() will have already generated a simple Gather
5344 : * path for the best parallel path, if any, and the loop above will have
5345 : * considered sorting it. Similarly, generate_gather_paths() will also
5346 : * have generated order-preserving Gather Merge plans which can be used
5347 : * without sorting if they happen to match the sort_pathkeys, and the loop
5348 : * above will have handled those as well. However, there's one more
5349 : * possibility: it may make sense to sort the cheapest partial path or
5350 : * incrementally sort any partial path that is partially sorted according
5351 : * to the required output order and then use Gather Merge.
5352 : */
5353 75922 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5354 52502 : input_rel->partial_pathlist != NIL)
5355 : {
5356 : Path *cheapest_partial_path;
5357 :
5358 2216 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5359 :
5360 4638 : foreach(lc, input_rel->partial_pathlist)
5361 : {
5362 2422 : Path *input_path = (Path *) lfirst(lc);
5363 : Path *sorted_path;
5364 : bool is_sorted;
5365 : int presorted_keys;
5366 : double total_groups;
5367 :
5368 2422 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5369 : input_path->pathkeys,
5370 : &presorted_keys);
5371 :
5372 2422 : if (is_sorted)
5373 182 : continue;
5374 :
5375 : /*
5376 : * Try at least sorting the cheapest path and also try
5377 : * incrementally sorting any path which is partially sorted
5378 : * already (no need to deal with paths which have presorted keys
5379 : * when incremental sort is disabled unless it's the cheapest
5380 : * partial path).
5381 : */
5382 2240 : if (input_path != cheapest_partial_path &&
5383 42 : (presorted_keys == 0 || !enable_incremental_sort))
5384 0 : continue;
5385 :
5386 : /*
5387 : * We've no need to consider both a sort and incremental sort.
5388 : * We'll just do a sort if there are no presorted keys and an
5389 : * incremental sort when there are presorted keys.
5390 : */
5391 2240 : if (presorted_keys == 0 || !enable_incremental_sort)
5392 2180 : sorted_path = (Path *) create_sort_path(root,
5393 : ordered_rel,
5394 : input_path,
5395 : root->sort_pathkeys,
5396 : limit_tuples);
5397 : else
5398 60 : sorted_path = (Path *) create_incremental_sort_path(root,
5399 : ordered_rel,
5400 : input_path,
5401 : root->sort_pathkeys,
5402 : presorted_keys,
5403 : limit_tuples);
5404 2240 : total_groups = compute_gather_rows(sorted_path);
5405 : sorted_path = (Path *)
5406 2240 : create_gather_merge_path(root, ordered_rel,
5407 : sorted_path,
5408 : sorted_path->pathtarget,
5409 : root->sort_pathkeys, NULL,
5410 : &total_groups);
5411 :
5412 : /*
5413 : * If the pathtarget of the result path has different expressions
5414 : * from the target to be applied, a projection step is needed.
5415 : */
5416 2240 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5417 6 : sorted_path = apply_projection_to_path(root, ordered_rel,
5418 : sorted_path, target);
5419 :
5420 2240 : add_path(ordered_rel, sorted_path);
5421 : }
5422 : }
5423 :
5424 : /*
5425 : * If there is an FDW that's responsible for all baserels of the query,
5426 : * let it consider adding ForeignPaths.
5427 : */
5428 75922 : if (ordered_rel->fdwroutine &&
5429 384 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5430 370 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5431 : input_rel, ordered_rel,
5432 : NULL);
5433 :
5434 : /* Let extensions possibly add some more paths */
5435 75922 : if (create_upper_paths_hook)
5436 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5437 : input_rel, ordered_rel, NULL);
5438 :
5439 : /*
5440 : * No need to bother with set_cheapest here; grouping_planner does not
5441 : * need us to do it.
5442 : */
5443 : Assert(ordered_rel->pathlist != NIL);
5444 :
5445 75922 : return ordered_rel;
5446 : }
5447 :
5448 :
5449 : /*
5450 : * make_group_input_target
5451 : * Generate appropriate PathTarget for initial input to grouping nodes.
5452 : *
5453 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5454 : * the query's final targetlist; for example, it certainly can't emit any
5455 : * aggregate function calls. This routine generates the correct target
5456 : * for the scan/join subplan.
5457 : *
5458 : * The query target list passed from the parser already contains entries
5459 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5460 : * for variables used only in HAVING clauses; so we need to add those
5461 : * variables to the subplan target list. Also, we flatten all expressions
5462 : * except GROUP BY items into their component variables; other expressions
5463 : * will be computed by the upper plan nodes rather than by the subplan.
5464 : * For example, given a query like
5465 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5466 : * we want to pass this targetlist to the subplan:
5467 : * a+b,c,d
5468 : * where the a+b target will be used by the Sort/Group steps, and the
5469 : * other targets will be used for computing the final results.
5470 : *
5471 : * 'final_target' is the query's final target list (in PathTarget form)
5472 : *
5473 : * The result is the PathTarget to be computed by the Paths returned from
5474 : * query_planner().
5475 : */
5476 : static PathTarget *
5477 41716 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5478 : {
5479 41716 : Query *parse = root->parse;
5480 : PathTarget *input_target;
5481 : List *non_group_cols;
5482 : List *non_group_vars;
5483 : int i;
5484 : ListCell *lc;
5485 :
5486 : /*
5487 : * We must build a target containing all grouping columns, plus any other
5488 : * Vars mentioned in the query's targetlist and HAVING qual.
5489 : */
5490 41716 : input_target = create_empty_pathtarget();
5491 41716 : non_group_cols = NIL;
5492 :
5493 41716 : i = 0;
5494 100308 : foreach(lc, final_target->exprs)
5495 : {
5496 58592 : Expr *expr = (Expr *) lfirst(lc);
5497 58592 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5498 :
5499 67340 : if (sgref && root->processed_groupClause &&
5500 8748 : get_sortgroupref_clause_noerr(sgref,
5501 : root->processed_groupClause) != NULL)
5502 : {
5503 : /*
5504 : * It's a grouping column, so add it to the input target as-is.
5505 : *
5506 : * Note that the target is logically below the grouping step. So
5507 : * with grouping sets we need to remove the RT index of the
5508 : * grouping step if there is any from the target expression.
5509 : */
5510 7006 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5511 : {
5512 : Assert(root->group_rtindex > 0);
5513 : expr = (Expr *)
5514 1824 : remove_nulling_relids((Node *) expr,
5515 1824 : bms_make_singleton(root->group_rtindex),
5516 : NULL);
5517 : }
5518 7006 : add_column_to_pathtarget(input_target, expr, sgref);
5519 : }
5520 : else
5521 : {
5522 : /*
5523 : * Non-grouping column, so just remember the expression for later
5524 : * call to pull_var_clause.
5525 : */
5526 51586 : non_group_cols = lappend(non_group_cols, expr);
5527 : }
5528 :
5529 58592 : i++;
5530 : }
5531 :
5532 : /*
5533 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5534 : */
5535 41716 : if (parse->havingQual)
5536 882 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5537 :
5538 : /*
5539 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5540 : * add them to the input target if not already present. (A Var used
5541 : * directly as a GROUP BY item will be present already.) Note this
5542 : * includes Vars used in resjunk items, so we are covering the needs of
5543 : * ORDER BY and window specifications. Vars used within Aggrefs and
5544 : * WindowFuncs will be pulled out here, too.
5545 : *
5546 : * Note that the target is logically below the grouping step. So with
5547 : * grouping sets we need to remove the RT index of the grouping step if
5548 : * there is any from the non-group Vars.
5549 : */
5550 41716 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5551 : PVC_RECURSE_AGGREGATES |
5552 : PVC_RECURSE_WINDOWFUNCS |
5553 : PVC_INCLUDE_PLACEHOLDERS);
5554 41716 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5555 : {
5556 : Assert(root->group_rtindex > 0);
5557 : non_group_vars = (List *)
5558 830 : remove_nulling_relids((Node *) non_group_vars,
5559 830 : bms_make_singleton(root->group_rtindex),
5560 : NULL);
5561 : }
5562 41716 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5563 :
5564 : /* clean up cruft */
5565 41716 : list_free(non_group_vars);
5566 41716 : list_free(non_group_cols);
5567 :
5568 : /* XXX this causes some redundant cost calculation ... */
5569 41716 : return set_pathtarget_cost_width(root, input_target);
5570 : }
5571 :
5572 : /*
5573 : * make_partial_grouping_target
5574 : * Generate appropriate PathTarget for output of partial aggregate
5575 : * (or partial grouping, if there are no aggregates) nodes.
5576 : *
5577 : * A partial aggregation node needs to emit all the same aggregates that
5578 : * a regular aggregation node would, plus any aggregates used in HAVING;
5579 : * except that the Aggref nodes should be marked as partial aggregates.
5580 : *
5581 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5582 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5583 : * these would be Vars that are grouped by or used in grouping expressions.)
5584 : *
5585 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5586 : * havingQual represents the HAVING clause.
5587 : */
5588 : static PathTarget *
5589 2198 : make_partial_grouping_target(PlannerInfo *root,
5590 : PathTarget *grouping_target,
5591 : Node *havingQual)
5592 : {
5593 : PathTarget *partial_target;
5594 : List *non_group_cols;
5595 : List *non_group_exprs;
5596 : int i;
5597 : ListCell *lc;
5598 :
5599 2198 : partial_target = create_empty_pathtarget();
5600 2198 : non_group_cols = NIL;
5601 :
5602 2198 : i = 0;
5603 7814 : foreach(lc, grouping_target->exprs)
5604 : {
5605 5616 : Expr *expr = (Expr *) lfirst(lc);
5606 5616 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5607 :
5608 9436 : if (sgref && root->processed_groupClause &&
5609 3820 : get_sortgroupref_clause_noerr(sgref,
5610 : root->processed_groupClause) != NULL)
5611 : {
5612 : /*
5613 : * It's a grouping column, so add it to the partial_target as-is.
5614 : * (This allows the upper agg step to repeat the grouping calcs.)
5615 : */
5616 1906 : add_column_to_pathtarget(partial_target, expr, sgref);
5617 : }
5618 : else
5619 : {
5620 : /*
5621 : * Non-grouping column, so just remember the expression for later
5622 : * call to pull_var_clause.
5623 : */
5624 3710 : non_group_cols = lappend(non_group_cols, expr);
5625 : }
5626 :
5627 5616 : i++;
5628 : }
5629 :
5630 : /*
5631 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5632 : */
5633 2198 : if (havingQual)
5634 824 : non_group_cols = lappend(non_group_cols, havingQual);
5635 :
5636 : /*
5637 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5638 : * non-group cols (plus HAVING), and add them to the partial_target if not
5639 : * already present. (An expression used directly as a GROUP BY item will
5640 : * be present already.) Note this includes Vars used in resjunk items, so
5641 : * we are covering the needs of ORDER BY and window specifications.
5642 : */
5643 2198 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5644 : PVC_INCLUDE_AGGREGATES |
5645 : PVC_RECURSE_WINDOWFUNCS |
5646 : PVC_INCLUDE_PLACEHOLDERS);
5647 :
5648 2198 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5649 :
5650 : /*
5651 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5652 : * are at the top level of the target list, so we can just scan the list
5653 : * rather than recursing through the expression trees.
5654 : */
5655 8386 : foreach(lc, partial_target->exprs)
5656 : {
5657 6188 : Aggref *aggref = (Aggref *) lfirst(lc);
5658 :
5659 6188 : if (IsA(aggref, Aggref))
5660 : {
5661 : Aggref *newaggref;
5662 :
5663 : /*
5664 : * We shouldn't need to copy the substructure of the Aggref node,
5665 : * but flat-copy the node itself to avoid damaging other trees.
5666 : */
5667 4252 : newaggref = makeNode(Aggref);
5668 4252 : memcpy(newaggref, aggref, sizeof(Aggref));
5669 :
5670 : /* For now, assume serialization is required */
5671 4252 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5672 :
5673 4252 : lfirst(lc) = newaggref;
5674 : }
5675 : }
5676 :
5677 : /* clean up cruft */
5678 2198 : list_free(non_group_exprs);
5679 2198 : list_free(non_group_cols);
5680 :
5681 : /* XXX this causes some redundant cost calculation ... */
5682 2198 : return set_pathtarget_cost_width(root, partial_target);
5683 : }
5684 :
5685 : /*
5686 : * mark_partial_aggref
5687 : * Adjust an Aggref to make it represent a partial-aggregation step.
5688 : *
5689 : * The Aggref node is modified in-place; caller must do any copying required.
5690 : */
5691 : void
5692 7060 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5693 : {
5694 : /* aggtranstype should be computed by this point */
5695 : Assert(OidIsValid(agg->aggtranstype));
5696 : /* ... but aggsplit should still be as the parser left it */
5697 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5698 :
5699 : /* Mark the Aggref with the intended partial-aggregation mode */
5700 7060 : agg->aggsplit = aggsplit;
5701 :
5702 : /*
5703 : * Adjust result type if needed. Normally, a partial aggregate returns
5704 : * the aggregate's transition type; but if that's INTERNAL and we're
5705 : * serializing, it returns BYTEA instead.
5706 : */
5707 7060 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5708 : {
5709 5656 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5710 242 : agg->aggtype = BYTEAOID;
5711 : else
5712 5414 : agg->aggtype = agg->aggtranstype;
5713 : }
5714 7060 : }
5715 :
5716 : /*
5717 : * postprocess_setop_tlist
5718 : * Fix up targetlist returned by plan_set_operations().
5719 : *
5720 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5721 : * NOTE: this would not be good enough if we supported resjunk sort keys
5722 : * for results of set operations --- then, we'd need to project a whole
5723 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5724 : * find any resjunk columns in orig_tlist.
5725 : */
5726 : static List *
5727 6160 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5728 : {
5729 : ListCell *l;
5730 6160 : ListCell *orig_tlist_item = list_head(orig_tlist);
5731 :
5732 24150 : foreach(l, new_tlist)
5733 : {
5734 17990 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5735 : TargetEntry *orig_tle;
5736 :
5737 : /* ignore resjunk columns in setop result */
5738 17990 : if (new_tle->resjunk)
5739 0 : continue;
5740 :
5741 : Assert(orig_tlist_item != NULL);
5742 17990 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5743 17990 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5744 17990 : if (orig_tle->resjunk) /* should not happen */
5745 0 : elog(ERROR, "resjunk output columns are not implemented");
5746 : Assert(new_tle->resno == orig_tle->resno);
5747 17990 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5748 : }
5749 6160 : if (orig_tlist_item != NULL)
5750 0 : elog(ERROR, "resjunk output columns are not implemented");
5751 6160 : return new_tlist;
5752 : }
5753 :
5754 : /*
5755 : * optimize_window_clauses
5756 : * Call each WindowFunc's prosupport function to see if we're able to
5757 : * make any adjustments to any of the WindowClause's so that the executor
5758 : * can execute the window functions in a more optimal way.
5759 : *
5760 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5761 : * may allow more things to be done here in the future.
5762 : */
5763 : static void
5764 2378 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5765 : {
5766 2378 : List *windowClause = root->parse->windowClause;
5767 : ListCell *lc;
5768 :
5769 4984 : foreach(lc, windowClause)
5770 : {
5771 2606 : WindowClause *wc = lfirst_node(WindowClause, lc);
5772 : ListCell *lc2;
5773 2606 : int optimizedFrameOptions = 0;
5774 :
5775 : Assert(wc->winref <= wflists->maxWinRef);
5776 :
5777 : /* skip any WindowClauses that have no WindowFuncs */
5778 2606 : if (wflists->windowFuncs[wc->winref] == NIL)
5779 24 : continue;
5780 :
5781 3122 : foreach(lc2, wflists->windowFuncs[wc->winref])
5782 : {
5783 : SupportRequestOptimizeWindowClause req;
5784 : SupportRequestOptimizeWindowClause *res;
5785 2624 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5786 : Oid prosupport;
5787 :
5788 2624 : prosupport = get_func_support(wfunc->winfnoid);
5789 :
5790 : /* Check if there's a support function for 'wfunc' */
5791 2624 : if (!OidIsValid(prosupport))
5792 2084 : break; /* can't optimize this WindowClause */
5793 :
5794 760 : req.type = T_SupportRequestOptimizeWindowClause;
5795 760 : req.window_clause = wc;
5796 760 : req.window_func = wfunc;
5797 760 : req.frameOptions = wc->frameOptions;
5798 :
5799 : /* call the support function */
5800 : res = (SupportRequestOptimizeWindowClause *)
5801 760 : DatumGetPointer(OidFunctionCall1(prosupport,
5802 : PointerGetDatum(&req)));
5803 :
5804 : /*
5805 : * Skip to next WindowClause if the support function does not
5806 : * support this request type.
5807 : */
5808 760 : if (res == NULL)
5809 220 : break;
5810 :
5811 : /*
5812 : * Save these frameOptions for the first WindowFunc for this
5813 : * WindowClause.
5814 : */
5815 540 : if (foreach_current_index(lc2) == 0)
5816 516 : optimizedFrameOptions = res->frameOptions;
5817 :
5818 : /*
5819 : * On subsequent WindowFuncs, if the frameOptions are not the same
5820 : * then we're unable to optimize the frameOptions for this
5821 : * WindowClause.
5822 : */
5823 24 : else if (optimizedFrameOptions != res->frameOptions)
5824 0 : break; /* skip to the next WindowClause, if any */
5825 : }
5826 :
5827 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
5828 2582 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5829 : {
5830 : ListCell *lc3;
5831 :
5832 : /* apply the new frame options */
5833 498 : wc->frameOptions = optimizedFrameOptions;
5834 :
5835 : /*
5836 : * We now check to see if changing the frameOptions has caused
5837 : * this WindowClause to be a duplicate of some other WindowClause.
5838 : * This can only happen if we have multiple WindowClauses, so
5839 : * don't bother if there's only 1.
5840 : */
5841 498 : if (list_length(windowClause) == 1)
5842 408 : continue;
5843 :
5844 : /*
5845 : * Do the duplicate check and reuse the existing WindowClause if
5846 : * we find a duplicate.
5847 : */
5848 228 : foreach(lc3, windowClause)
5849 : {
5850 174 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5851 :
5852 : /* skip over the WindowClause we're currently editing */
5853 174 : if (existing_wc == wc)
5854 54 : continue;
5855 :
5856 : /*
5857 : * Perform the same duplicate check that is done in
5858 : * transformWindowFuncCall.
5859 : */
5860 240 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
5861 120 : equal(wc->orderClause, existing_wc->orderClause) &&
5862 120 : wc->frameOptions == existing_wc->frameOptions &&
5863 72 : equal(wc->startOffset, existing_wc->startOffset) &&
5864 36 : equal(wc->endOffset, existing_wc->endOffset))
5865 : {
5866 : ListCell *lc4;
5867 :
5868 : /*
5869 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
5870 : * This required adjusting each WindowFunc's winref and
5871 : * moving the WindowFuncs in 'wc' to the list of
5872 : * WindowFuncs in 'existing_wc'.
5873 : */
5874 78 : foreach(lc4, wflists->windowFuncs[wc->winref])
5875 : {
5876 42 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
5877 :
5878 42 : wfunc->winref = existing_wc->winref;
5879 : }
5880 :
5881 : /* move list items */
5882 72 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
5883 36 : wflists->windowFuncs[wc->winref]);
5884 36 : wflists->windowFuncs[wc->winref] = NIL;
5885 :
5886 : /*
5887 : * transformWindowFuncCall() should have made sure there
5888 : * are no other duplicates, so we needn't bother looking
5889 : * any further.
5890 : */
5891 36 : break;
5892 : }
5893 : }
5894 : }
5895 : }
5896 2378 : }
5897 :
5898 : /*
5899 : * select_active_windows
5900 : * Create a list of the "active" window clauses (ie, those referenced
5901 : * by non-deleted WindowFuncs) in the order they are to be executed.
5902 : */
5903 : static List *
5904 2378 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5905 : {
5906 2378 : List *windowClause = root->parse->windowClause;
5907 2378 : List *result = NIL;
5908 : ListCell *lc;
5909 2378 : int nActive = 0;
5910 2378 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5911 2378 : * list_length(windowClause));
5912 :
5913 : /* First, construct an array of the active windows */
5914 4984 : foreach(lc, windowClause)
5915 : {
5916 2606 : WindowClause *wc = lfirst_node(WindowClause, lc);
5917 :
5918 : /* It's only active if wflists shows some related WindowFuncs */
5919 : Assert(wc->winref <= wflists->maxWinRef);
5920 2606 : if (wflists->windowFuncs[wc->winref] == NIL)
5921 60 : continue;
5922 :
5923 2546 : actives[nActive].wc = wc; /* original clause */
5924 :
5925 : /*
5926 : * For sorting, we want the list of partition keys followed by the
5927 : * list of sort keys. But pathkeys construction will remove duplicates
5928 : * between the two, so we can as well (even though we can't detect all
5929 : * of the duplicates, since some may come from ECs - that might mean
5930 : * we miss optimization chances here). We must, however, ensure that
5931 : * the order of entries is preserved with respect to the ones we do
5932 : * keep.
5933 : *
5934 : * partitionClause and orderClause had their own duplicates removed in
5935 : * parse analysis, so we're only concerned here with removing
5936 : * orderClause entries that also appear in partitionClause.
5937 : */
5938 5092 : actives[nActive].uniqueOrder =
5939 2546 : list_concat_unique(list_copy(wc->partitionClause),
5940 2546 : wc->orderClause);
5941 2546 : nActive++;
5942 : }
5943 :
5944 : /*
5945 : * Sort active windows by their partitioning/ordering clauses, ignoring
5946 : * any framing clauses, so that the windows that need the same sorting are
5947 : * adjacent in the list. When we come to generate paths, this will avoid
5948 : * inserting additional Sort nodes.
5949 : *
5950 : * This is how we implement a specific requirement from the SQL standard,
5951 : * which says that when two or more windows are order-equivalent (i.e.
5952 : * have matching partition and order clauses, even if their names or
5953 : * framing clauses differ), then all peer rows must be presented in the
5954 : * same order in all of them. If we allowed multiple sort nodes for such
5955 : * cases, we'd risk having the peer rows end up in different orders in
5956 : * equivalent windows due to sort instability. (See General Rule 4 of
5957 : * <window clause> in SQL2008 - SQL2016.)
5958 : *
5959 : * Additionally, if the entire list of clauses of one window is a prefix
5960 : * of another, put first the window with stronger sorting requirements.
5961 : * This way we will first sort for stronger window, and won't have to sort
5962 : * again for the weaker one.
5963 : */
5964 2378 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5965 :
5966 : /* build ordered list of the original WindowClause nodes */
5967 4924 : for (int i = 0; i < nActive; i++)
5968 2546 : result = lappend(result, actives[i].wc);
5969 :
5970 2378 : pfree(actives);
5971 :
5972 2378 : return result;
5973 : }
5974 :
5975 : /*
5976 : * name_active_windows
5977 : * Ensure all active windows have unique names.
5978 : *
5979 : * The parser will have checked that user-assigned window names are unique
5980 : * within the Query. Here we assign made-up names to any unnamed
5981 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
5982 : * at parse time, because it'd mess up decompilation of views.)
5983 : *
5984 : * activeWindows: result of select_active_windows
5985 : */
5986 : static void
5987 2378 : name_active_windows(List *activeWindows)
5988 : {
5989 2378 : int next_n = 1;
5990 : char newname[16];
5991 : ListCell *lc;
5992 :
5993 4924 : foreach(lc, activeWindows)
5994 : {
5995 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
5996 :
5997 : /* Nothing to do if it has a name already. */
5998 2546 : if (wc->name)
5999 498 : continue;
6000 :
6001 : /* Select a name not currently present in the list. */
6002 : for (;;)
6003 6 : {
6004 : ListCell *lc2;
6005 :
6006 2054 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6007 4456 : foreach(lc2, activeWindows)
6008 : {
6009 2408 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6010 :
6011 2408 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6012 6 : break; /* matched */
6013 : }
6014 2054 : if (lc2 == NULL)
6015 2048 : break; /* reached the end with no match */
6016 : }
6017 2048 : wc->name = pstrdup(newname);
6018 : }
6019 2378 : }
6020 :
6021 : /*
6022 : * common_prefix_cmp
6023 : * QSort comparison function for WindowClauseSortData
6024 : *
6025 : * Sort the windows by the required sorting clauses. First, compare the sort
6026 : * clauses themselves. Second, if one window's clauses are a prefix of another
6027 : * one's clauses, put the window with more sort clauses first.
6028 : *
6029 : * We purposefully sort by the highest tleSortGroupRef first. Since
6030 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6031 : * and because here we sort the lowest tleSortGroupRefs last, if a
6032 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6033 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6034 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6035 : * reducing the total number of sorts required for the query.
6036 : */
6037 : static int
6038 186 : common_prefix_cmp(const void *a, const void *b)
6039 : {
6040 186 : const WindowClauseSortData *wcsa = a;
6041 186 : const WindowClauseSortData *wcsb = b;
6042 : ListCell *item_a;
6043 : ListCell *item_b;
6044 :
6045 330 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6046 : {
6047 246 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6048 246 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6049 :
6050 246 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6051 102 : return -1;
6052 234 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6053 66 : return 1;
6054 168 : else if (sca->sortop > scb->sortop)
6055 0 : return -1;
6056 168 : else if (sca->sortop < scb->sortop)
6057 24 : return 1;
6058 144 : else if (sca->nulls_first && !scb->nulls_first)
6059 0 : return -1;
6060 144 : else if (!sca->nulls_first && scb->nulls_first)
6061 0 : return 1;
6062 : /* no need to compare eqop, since it is fully determined by sortop */
6063 : }
6064 :
6065 84 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6066 6 : return -1;
6067 78 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6068 30 : return 1;
6069 :
6070 48 : return 0;
6071 : }
6072 :
6073 : /*
6074 : * make_window_input_target
6075 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6076 : *
6077 : * When the query has window functions, this function computes the desired
6078 : * target to be computed by the node just below the first WindowAgg.
6079 : * This tlist must contain all values needed to evaluate the window functions,
6080 : * compute the final target list, and perform any required final sort step.
6081 : * If multiple WindowAggs are needed, each intermediate one adds its window
6082 : * function results onto this base tlist; only the topmost WindowAgg computes
6083 : * the actual desired target list.
6084 : *
6085 : * This function is much like make_group_input_target, though not quite enough
6086 : * like it to share code. As in that function, we flatten most expressions
6087 : * into their component variables. But we do not want to flatten window
6088 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6089 : * evaluations of them, which would be bad (possibly even resulting in
6090 : * inconsistent answers, if they contain volatile functions).
6091 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6092 : * make_group_input_target, because we may no longer have access to the
6093 : * individual Vars in them.
6094 : *
6095 : * Another key difference from make_group_input_target is that we don't
6096 : * flatten Aggref expressions, since those are to be computed below the
6097 : * window functions and just referenced like Vars above that.
6098 : *
6099 : * 'final_target' is the query's final target list (in PathTarget form)
6100 : * 'activeWindows' is the list of active windows previously identified by
6101 : * select_active_windows.
6102 : *
6103 : * The result is the PathTarget to be computed by the plan node immediately
6104 : * below the first WindowAgg node.
6105 : */
6106 : static PathTarget *
6107 2378 : make_window_input_target(PlannerInfo *root,
6108 : PathTarget *final_target,
6109 : List *activeWindows)
6110 : {
6111 : PathTarget *input_target;
6112 : Bitmapset *sgrefs;
6113 : List *flattenable_cols;
6114 : List *flattenable_vars;
6115 : int i;
6116 : ListCell *lc;
6117 :
6118 : Assert(root->parse->hasWindowFuncs);
6119 :
6120 : /*
6121 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6122 : * into a bitmapset for convenient reference below.
6123 : */
6124 2378 : sgrefs = NULL;
6125 4924 : foreach(lc, activeWindows)
6126 : {
6127 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
6128 : ListCell *lc2;
6129 :
6130 3290 : foreach(lc2, wc->partitionClause)
6131 : {
6132 744 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6133 :
6134 744 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6135 : }
6136 4722 : foreach(lc2, wc->orderClause)
6137 : {
6138 2176 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6139 :
6140 2176 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6141 : }
6142 : }
6143 :
6144 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6145 2564 : foreach(lc, root->processed_groupClause)
6146 : {
6147 186 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6148 :
6149 186 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6150 : }
6151 :
6152 : /*
6153 : * Construct a target containing all the non-flattenable targetlist items,
6154 : * and save aside the others for a moment.
6155 : */
6156 2378 : input_target = create_empty_pathtarget();
6157 2378 : flattenable_cols = NIL;
6158 :
6159 2378 : i = 0;
6160 10238 : foreach(lc, final_target->exprs)
6161 : {
6162 7860 : Expr *expr = (Expr *) lfirst(lc);
6163 7860 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6164 :
6165 : /*
6166 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6167 : * that such items can't contain window functions, so it's okay to
6168 : * compute them below the WindowAgg nodes.)
6169 : */
6170 7860 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6171 : {
6172 : /*
6173 : * Don't want to deconstruct this value, so add it to the input
6174 : * target as-is.
6175 : */
6176 2774 : add_column_to_pathtarget(input_target, expr, sgref);
6177 : }
6178 : else
6179 : {
6180 : /*
6181 : * Column is to be flattened, so just remember the expression for
6182 : * later call to pull_var_clause.
6183 : */
6184 5086 : flattenable_cols = lappend(flattenable_cols, expr);
6185 : }
6186 :
6187 7860 : i++;
6188 : }
6189 :
6190 : /*
6191 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6192 : * add them to the input target if not already present. (Some might be
6193 : * there already because they're used directly as window/group clauses.)
6194 : *
6195 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6196 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6197 : * at higher levels. On the other hand, we should recurse into
6198 : * WindowFuncs to make sure their input expressions are available.
6199 : */
6200 2378 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6201 : PVC_INCLUDE_AGGREGATES |
6202 : PVC_RECURSE_WINDOWFUNCS |
6203 : PVC_INCLUDE_PLACEHOLDERS);
6204 2378 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6205 :
6206 : /* clean up cruft */
6207 2378 : list_free(flattenable_vars);
6208 2378 : list_free(flattenable_cols);
6209 :
6210 : /* XXX this causes some redundant cost calculation ... */
6211 2378 : return set_pathtarget_cost_width(root, input_target);
6212 : }
6213 :
6214 : /*
6215 : * make_pathkeys_for_window
6216 : * Create a pathkeys list describing the required input ordering
6217 : * for the given WindowClause.
6218 : *
6219 : * Modifies wc's partitionClause to remove any clauses which are deemed
6220 : * redundant by the pathkey logic.
6221 : *
6222 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6223 : * In the future we might try to implement windowing using hashing, in which
6224 : * case the ordering could be relaxed, but for now we always sort.
6225 : */
6226 : static List *
6227 5132 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6228 : List *tlist)
6229 : {
6230 5132 : List *window_pathkeys = NIL;
6231 :
6232 : /* Throw error if can't sort */
6233 5132 : if (!grouping_is_sortable(wc->partitionClause))
6234 0 : ereport(ERROR,
6235 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6236 : errmsg("could not implement window PARTITION BY"),
6237 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6238 5132 : if (!grouping_is_sortable(wc->orderClause))
6239 0 : ereport(ERROR,
6240 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6241 : errmsg("could not implement window ORDER BY"),
6242 : errdetail("Window ordering columns must be of sortable datatypes.")));
6243 :
6244 : /*
6245 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6246 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6247 : */
6248 5132 : if (wc->partitionClause != NIL)
6249 : {
6250 : bool sortable;
6251 :
6252 1290 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6253 : &wc->partitionClause,
6254 : tlist,
6255 : true,
6256 : false,
6257 : &sortable,
6258 : false);
6259 :
6260 : Assert(sortable);
6261 : }
6262 :
6263 : /*
6264 : * In principle, we could also consider removing redundant ORDER BY items
6265 : * too as doing so does not alter the result of peer row checks done by
6266 : * the executor. However, we must *not* remove the ordering column for
6267 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6268 : * if it's known to be equal to some partitioning column.
6269 : */
6270 5132 : if (wc->orderClause != NIL)
6271 : {
6272 : List *orderby_pathkeys;
6273 :
6274 4276 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6275 : wc->orderClause,
6276 : tlist);
6277 :
6278 : /* Okay, make the combined pathkeys */
6279 4276 : if (window_pathkeys != NIL)
6280 934 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6281 : else
6282 3342 : window_pathkeys = orderby_pathkeys;
6283 : }
6284 :
6285 5132 : return window_pathkeys;
6286 : }
6287 :
6288 : /*
6289 : * make_sort_input_target
6290 : * Generate appropriate PathTarget for initial input to Sort step.
6291 : *
6292 : * If the query has ORDER BY, this function chooses the target to be computed
6293 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6294 : * project) steps. This might or might not be identical to the query's final
6295 : * output target.
6296 : *
6297 : * The main argument for keeping the sort-input tlist the same as the final
6298 : * is that we avoid a separate projection node (which will be needed if
6299 : * they're different, because Sort can't project). However, there are also
6300 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6301 : * a consistent order of evaluation for any volatile functions in the tlist,
6302 : * and if there's also a LIMIT, we can stop the query without ever computing
6303 : * tlist functions for later rows, which is beneficial for both volatile and
6304 : * expensive functions.
6305 : *
6306 : * Our current policy is to postpone volatile expressions till after the sort
6307 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6308 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6309 : * postpone set-returning expressions, because running them beforehand would
6310 : * bloat the sort dataset, and because it might cause unexpected output order
6311 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6312 : * in the tlist should be evaluated at the same plan step, so that they can
6313 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6314 : * mustn't postpone any SRFs. (Note that in principle that policy should
6315 : * probably get applied to the group/window input targetlists too, but we
6316 : * have not done that historically.) Lastly, expensive expressions are
6317 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6318 : * partial evaluation of the query is possible (if neither is true, we expect
6319 : * to have to evaluate the expressions for every row anyway), or if there are
6320 : * any volatile or set-returning expressions (since once we've put in a
6321 : * projection at all, it won't cost any more to postpone more stuff).
6322 : *
6323 : * Another issue that could potentially be considered here is that
6324 : * evaluating tlist expressions could result in data that's either wider
6325 : * or narrower than the input Vars, thus changing the volume of data that
6326 : * has to go through the Sort. However, we usually have only a very bad
6327 : * idea of the output width of any expression more complex than a Var,
6328 : * so for now it seems too risky to try to optimize on that basis.
6329 : *
6330 : * Note that if we do produce a modified sort-input target, and then the
6331 : * query ends up not using an explicit Sort, no particular harm is done:
6332 : * we'll initially use the modified target for the preceding path nodes,
6333 : * but then change them to the final target with apply_projection_to_path.
6334 : * Moreover, in such a case the guarantees about evaluation order of
6335 : * volatile functions still hold, since the rows are sorted already.
6336 : *
6337 : * This function has some things in common with make_group_input_target and
6338 : * make_window_input_target, though the detailed rules for what to do are
6339 : * different. We never flatten/postpone any grouping or ordering columns;
6340 : * those are needed before the sort. If we do flatten a particular
6341 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6342 : * computed earlier.
6343 : *
6344 : * 'final_target' is the query's final target list (in PathTarget form)
6345 : * 'have_postponed_srfs' is an output argument, see below
6346 : *
6347 : * The result is the PathTarget to be computed by the plan node immediately
6348 : * below the Sort step (and the Distinct step, if any). This will be
6349 : * exactly final_target if we decide a projection step wouldn't be helpful.
6350 : *
6351 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6352 : * any set-returning functions to after the Sort.
6353 : */
6354 : static PathTarget *
6355 71966 : make_sort_input_target(PlannerInfo *root,
6356 : PathTarget *final_target,
6357 : bool *have_postponed_srfs)
6358 : {
6359 71966 : Query *parse = root->parse;
6360 : PathTarget *input_target;
6361 : int ncols;
6362 : bool *col_is_srf;
6363 : bool *postpone_col;
6364 : bool have_srf;
6365 : bool have_volatile;
6366 : bool have_expensive;
6367 : bool have_srf_sortcols;
6368 : bool postpone_srfs;
6369 : List *postponable_cols;
6370 : List *postponable_vars;
6371 : int i;
6372 : ListCell *lc;
6373 :
6374 : /* Shouldn't get here unless query has ORDER BY */
6375 : Assert(parse->sortClause);
6376 :
6377 71966 : *have_postponed_srfs = false; /* default result */
6378 :
6379 : /* Inspect tlist and collect per-column information */
6380 71966 : ncols = list_length(final_target->exprs);
6381 71966 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6382 71966 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6383 71966 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6384 :
6385 71966 : i = 0;
6386 447454 : foreach(lc, final_target->exprs)
6387 : {
6388 375488 : Expr *expr = (Expr *) lfirst(lc);
6389 :
6390 : /*
6391 : * If the column has a sortgroupref, assume it has to be evaluated
6392 : * before sorting. Generally such columns would be ORDER BY, GROUP
6393 : * BY, etc targets. One exception is columns that were removed from
6394 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6395 : * only be Vars anyway. There don't seem to be any cases where it
6396 : * would be worth the trouble to double-check.
6397 : */
6398 375488 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6399 : {
6400 : /*
6401 : * Check for SRF or volatile functions. Check the SRF case first
6402 : * because we must know whether we have any postponed SRFs.
6403 : */
6404 273320 : if (parse->hasTargetSRFs &&
6405 216 : expression_returns_set((Node *) expr))
6406 : {
6407 : /* We'll decide below whether these are postponable */
6408 96 : col_is_srf[i] = true;
6409 96 : have_srf = true;
6410 : }
6411 273008 : else if (contain_volatile_functions((Node *) expr))
6412 : {
6413 : /* Unconditionally postpone */
6414 148 : postpone_col[i] = true;
6415 148 : have_volatile = true;
6416 : }
6417 : else
6418 : {
6419 : /*
6420 : * Else check the cost. XXX it's annoying to have to do this
6421 : * when set_pathtarget_cost_width() just did it. Refactor to
6422 : * allow sharing the work?
6423 : */
6424 : QualCost cost;
6425 :
6426 272860 : cost_qual_eval_node(&cost, (Node *) expr, root);
6427 :
6428 : /*
6429 : * We arbitrarily define "expensive" as "more than 10X
6430 : * cpu_operator_cost". Note this will take in any PL function
6431 : * with default cost.
6432 : */
6433 272860 : if (cost.per_tuple > 10 * cpu_operator_cost)
6434 : {
6435 17216 : postpone_col[i] = true;
6436 17216 : have_expensive = true;
6437 : }
6438 : }
6439 : }
6440 : else
6441 : {
6442 : /* For sortgroupref cols, just check if any contain SRFs */
6443 102384 : if (!have_srf_sortcols &&
6444 102694 : parse->hasTargetSRFs &&
6445 310 : expression_returns_set((Node *) expr))
6446 124 : have_srf_sortcols = true;
6447 : }
6448 :
6449 375488 : i++;
6450 : }
6451 :
6452 : /*
6453 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6454 : */
6455 71966 : postpone_srfs = (have_srf && !have_srf_sortcols);
6456 :
6457 : /*
6458 : * If we don't need a post-sort projection, just return final_target.
6459 : */
6460 71966 : if (!(postpone_srfs || have_volatile ||
6461 71762 : (have_expensive &&
6462 10110 : (parse->limitCount || root->tuple_fraction > 0))))
6463 71726 : return final_target;
6464 :
6465 : /*
6466 : * Report whether the post-sort projection will contain set-returning
6467 : * functions. This is important because it affects whether the Sort can
6468 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6469 : * to return.
6470 : */
6471 240 : *have_postponed_srfs = postpone_srfs;
6472 :
6473 : /*
6474 : * Construct the sort-input target, taking all non-postponable columns and
6475 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6476 : * the postponable ones.
6477 : */
6478 240 : input_target = create_empty_pathtarget();
6479 240 : postponable_cols = NIL;
6480 :
6481 240 : i = 0;
6482 1990 : foreach(lc, final_target->exprs)
6483 : {
6484 1750 : Expr *expr = (Expr *) lfirst(lc);
6485 :
6486 1750 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6487 298 : postponable_cols = lappend(postponable_cols, expr);
6488 : else
6489 1452 : add_column_to_pathtarget(input_target, expr,
6490 1452 : get_pathtarget_sortgroupref(final_target, i));
6491 :
6492 1750 : i++;
6493 : }
6494 :
6495 : /*
6496 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6497 : * postponable columns, and add them to the sort-input target if not
6498 : * already present. (Some might be there already.) We mustn't
6499 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6500 : * would be unable to recompute them.
6501 : */
6502 240 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6503 : PVC_INCLUDE_AGGREGATES |
6504 : PVC_INCLUDE_WINDOWFUNCS |
6505 : PVC_INCLUDE_PLACEHOLDERS);
6506 240 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6507 :
6508 : /* clean up cruft */
6509 240 : list_free(postponable_vars);
6510 240 : list_free(postponable_cols);
6511 :
6512 : /* XXX this represents even more redundant cost calculation ... */
6513 240 : return set_pathtarget_cost_width(root, input_target);
6514 : }
6515 :
6516 : /*
6517 : * get_cheapest_fractional_path
6518 : * Find the cheapest path for retrieving a specified fraction of all
6519 : * the tuples expected to be returned by the given relation.
6520 : *
6521 : * Do not consider parameterized paths. If the caller needs a path for upper
6522 : * rel, it can't have parameterized paths. If the caller needs an append
6523 : * subpath, it could become limited by the treatment of similar
6524 : * parameterization of all the subpaths.
6525 : *
6526 : * We interpret tuple_fraction the same way as grouping_planner.
6527 : *
6528 : * We assume set_cheapest() has been run on the given rel.
6529 : */
6530 : Path *
6531 501832 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6532 : {
6533 501832 : Path *best_path = rel->cheapest_total_path;
6534 : ListCell *l;
6535 :
6536 : /* If all tuples will be retrieved, just return the cheapest-total path */
6537 501832 : if (tuple_fraction <= 0.0)
6538 492416 : return best_path;
6539 :
6540 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6541 9416 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6542 3890 : tuple_fraction /= best_path->rows;
6543 :
6544 24538 : foreach(l, rel->pathlist)
6545 : {
6546 15122 : Path *path = (Path *) lfirst(l);
6547 :
6548 15122 : if (path->param_info)
6549 188 : continue;
6550 :
6551 20452 : if (path == rel->cheapest_total_path ||
6552 5518 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6553 14442 : continue;
6554 :
6555 492 : best_path = path;
6556 : }
6557 :
6558 9416 : return best_path;
6559 : }
6560 :
6561 : /*
6562 : * adjust_paths_for_srfs
6563 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6564 : *
6565 : * The executor can only handle set-returning functions that appear at the
6566 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6567 : * that are not at top level, we need to split up the evaluation into multiple
6568 : * plan levels in which each level satisfies this constraint. This function
6569 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6570 : * output tlist to insert appropriate projection steps.
6571 : *
6572 : * The given targets and targets_contain_srfs lists are from
6573 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6574 : * target in targets.
6575 : */
6576 : static void
6577 12648 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6578 : List *targets, List *targets_contain_srfs)
6579 : {
6580 : ListCell *lc;
6581 :
6582 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6583 : Assert(!linitial_int(targets_contain_srfs));
6584 :
6585 : /* If no SRFs appear at this plan level, nothing to do */
6586 12648 : if (list_length(targets) == 1)
6587 628 : return;
6588 :
6589 : /*
6590 : * Stack SRF-evaluation nodes atop each path for the rel.
6591 : *
6592 : * In principle we should re-run set_cheapest() here to identify the
6593 : * cheapest path, but it seems unlikely that adding the same tlist eval
6594 : * costs to all the paths would change that, so we don't bother. Instead,
6595 : * just assume that the cheapest-startup and cheapest-total paths remain
6596 : * so. (There should be no parameterized paths anymore, so we needn't
6597 : * worry about updating cheapest_parameterized_paths.)
6598 : */
6599 24066 : foreach(lc, rel->pathlist)
6600 : {
6601 12046 : Path *subpath = (Path *) lfirst(lc);
6602 12046 : Path *newpath = subpath;
6603 : ListCell *lc1,
6604 : *lc2;
6605 :
6606 : Assert(subpath->param_info == NULL);
6607 37314 : forboth(lc1, targets, lc2, targets_contain_srfs)
6608 : {
6609 25268 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6610 25268 : bool contains_srfs = (bool) lfirst_int(lc2);
6611 :
6612 : /* If this level doesn't contain SRFs, do regular projection */
6613 25268 : if (contains_srfs)
6614 12106 : newpath = (Path *) create_set_projection_path(root,
6615 : rel,
6616 : newpath,
6617 : thistarget);
6618 : else
6619 13162 : newpath = (Path *) apply_projection_to_path(root,
6620 : rel,
6621 : newpath,
6622 : thistarget);
6623 : }
6624 12046 : lfirst(lc) = newpath;
6625 12046 : if (subpath == rel->cheapest_startup_path)
6626 372 : rel->cheapest_startup_path = newpath;
6627 12046 : if (subpath == rel->cheapest_total_path)
6628 372 : rel->cheapest_total_path = newpath;
6629 : }
6630 :
6631 : /* Likewise for partial paths, if any */
6632 12026 : foreach(lc, rel->partial_pathlist)
6633 : {
6634 6 : Path *subpath = (Path *) lfirst(lc);
6635 6 : Path *newpath = subpath;
6636 : ListCell *lc1,
6637 : *lc2;
6638 :
6639 : Assert(subpath->param_info == NULL);
6640 24 : forboth(lc1, targets, lc2, targets_contain_srfs)
6641 : {
6642 18 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6643 18 : bool contains_srfs = (bool) lfirst_int(lc2);
6644 :
6645 : /* If this level doesn't contain SRFs, do regular projection */
6646 18 : if (contains_srfs)
6647 6 : newpath = (Path *) create_set_projection_path(root,
6648 : rel,
6649 : newpath,
6650 : thistarget);
6651 : else
6652 : {
6653 : /* avoid apply_projection_to_path, in case of multiple refs */
6654 12 : newpath = (Path *) create_projection_path(root,
6655 : rel,
6656 : newpath,
6657 : thistarget);
6658 : }
6659 : }
6660 6 : lfirst(lc) = newpath;
6661 : }
6662 : }
6663 :
6664 : /*
6665 : * expression_planner
6666 : * Perform planner's transformations on a standalone expression.
6667 : *
6668 : * Various utility commands need to evaluate expressions that are not part
6669 : * of a plannable query. They can do so using the executor's regular
6670 : * expression-execution machinery, but first the expression has to be fed
6671 : * through here to transform it from parser output to something executable.
6672 : *
6673 : * Currently, we disallow sublinks in standalone expressions, so there's no
6674 : * real "planning" involved here. (That might not always be true though.)
6675 : * What we must do is run eval_const_expressions to ensure that any function
6676 : * calls are converted to positional notation and function default arguments
6677 : * get inserted. The fact that constant subexpressions get simplified is a
6678 : * side-effect that is useful when the expression will get evaluated more than
6679 : * once. Also, we must fix operator function IDs.
6680 : *
6681 : * This does not return any information about dependencies of the expression.
6682 : * Hence callers should use the results only for the duration of the current
6683 : * query. Callers that would like to cache the results for longer should use
6684 : * expression_planner_with_deps, probably via the plancache.
6685 : *
6686 : * Note: this must not make any damaging changes to the passed-in expression
6687 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6688 : * we first do an expression_tree_mutator-based walk, what is returned will
6689 : * be a new node tree.) The result is constructed in the current memory
6690 : * context; beware that this can leak a lot of additional stuff there, too.
6691 : */
6692 : Expr *
6693 247038 : expression_planner(Expr *expr)
6694 : {
6695 : Node *result;
6696 :
6697 : /*
6698 : * Convert named-argument function calls, insert default arguments and
6699 : * simplify constant subexprs
6700 : */
6701 247038 : result = eval_const_expressions(NULL, (Node *) expr);
6702 :
6703 : /* Fill in opfuncid values if missing */
6704 247020 : fix_opfuncids(result);
6705 :
6706 247020 : return (Expr *) result;
6707 : }
6708 :
6709 : /*
6710 : * expression_planner_with_deps
6711 : * Perform planner's transformations on a standalone expression,
6712 : * returning expression dependency information along with the result.
6713 : *
6714 : * This is identical to expression_planner() except that it also returns
6715 : * information about possible dependencies of the expression, ie identities of
6716 : * objects whose definitions affect the result. As in a PlannedStmt, these
6717 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6718 : */
6719 : Expr *
6720 344 : expression_planner_with_deps(Expr *expr,
6721 : List **relationOids,
6722 : List **invalItems)
6723 : {
6724 : Node *result;
6725 : PlannerGlobal glob;
6726 : PlannerInfo root;
6727 :
6728 : /* Make up dummy planner state so we can use setrefs machinery */
6729 7912 : MemSet(&glob, 0, sizeof(glob));
6730 344 : glob.type = T_PlannerGlobal;
6731 344 : glob.relationOids = NIL;
6732 344 : glob.invalItems = NIL;
6733 :
6734 30616 : MemSet(&root, 0, sizeof(root));
6735 344 : root.type = T_PlannerInfo;
6736 344 : root.glob = &glob;
6737 :
6738 : /*
6739 : * Convert named-argument function calls, insert default arguments and
6740 : * simplify constant subexprs. Collect identities of inlined functions
6741 : * and elided domains, too.
6742 : */
6743 344 : result = eval_const_expressions(&root, (Node *) expr);
6744 :
6745 : /* Fill in opfuncid values if missing */
6746 344 : fix_opfuncids(result);
6747 :
6748 : /*
6749 : * Now walk the finished expression to find anything else we ought to
6750 : * record as an expression dependency.
6751 : */
6752 344 : (void) extract_query_dependencies_walker(result, &root);
6753 :
6754 344 : *relationOids = glob.relationOids;
6755 344 : *invalItems = glob.invalItems;
6756 :
6757 344 : return (Expr *) result;
6758 : }
6759 :
6760 :
6761 : /*
6762 : * plan_cluster_use_sort
6763 : * Use the planner to decide how CLUSTER should implement sorting
6764 : *
6765 : * tableOid is the OID of a table to be clustered on its index indexOid
6766 : * (which is already known to be a btree index). Decide whether it's
6767 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6768 : * Return true to use sorting, false to use an indexscan.
6769 : *
6770 : * Note: caller had better already hold some type of lock on the table.
6771 : */
6772 : bool
6773 190 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6774 : {
6775 : PlannerInfo *root;
6776 : Query *query;
6777 : PlannerGlobal *glob;
6778 : RangeTblEntry *rte;
6779 : RelOptInfo *rel;
6780 : IndexOptInfo *indexInfo;
6781 : QualCost indexExprCost;
6782 : Cost comparisonCost;
6783 : Path *seqScanPath;
6784 : Path seqScanAndSortPath;
6785 : IndexPath *indexScanPath;
6786 : ListCell *lc;
6787 :
6788 : /* We can short-circuit the cost comparison if indexscans are disabled */
6789 190 : if (!enable_indexscan)
6790 30 : return true; /* use sort */
6791 :
6792 : /* Set up mostly-dummy planner state */
6793 160 : query = makeNode(Query);
6794 160 : query->commandType = CMD_SELECT;
6795 :
6796 160 : glob = makeNode(PlannerGlobal);
6797 :
6798 160 : root = makeNode(PlannerInfo);
6799 160 : root->parse = query;
6800 160 : root->glob = glob;
6801 160 : root->query_level = 1;
6802 160 : root->planner_cxt = CurrentMemoryContext;
6803 160 : root->wt_param_id = -1;
6804 160 : root->join_domains = list_make1(makeNode(JoinDomain));
6805 :
6806 : /* Build a minimal RTE for the rel */
6807 160 : rte = makeNode(RangeTblEntry);
6808 160 : rte->rtekind = RTE_RELATION;
6809 160 : rte->relid = tableOid;
6810 160 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6811 160 : rte->rellockmode = AccessShareLock;
6812 160 : rte->lateral = false;
6813 160 : rte->inh = false;
6814 160 : rte->inFromCl = true;
6815 160 : query->rtable = list_make1(rte);
6816 160 : addRTEPermissionInfo(&query->rteperminfos, rte);
6817 :
6818 : /* Set up RTE/RelOptInfo arrays */
6819 160 : setup_simple_rel_arrays(root);
6820 :
6821 : /* Build RelOptInfo */
6822 160 : rel = build_simple_rel(root, 1, NULL);
6823 :
6824 : /* Locate IndexOptInfo for the target index */
6825 160 : indexInfo = NULL;
6826 198 : foreach(lc, rel->indexlist)
6827 : {
6828 198 : indexInfo = lfirst_node(IndexOptInfo, lc);
6829 198 : if (indexInfo->indexoid == indexOid)
6830 160 : break;
6831 : }
6832 :
6833 : /*
6834 : * It's possible that get_relation_info did not generate an IndexOptInfo
6835 : * for the desired index; this could happen if it's not yet reached its
6836 : * indcheckxmin usability horizon, or if it's a system index and we're
6837 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6838 : * trust the index contents but use seqscan-and-sort.
6839 : */
6840 160 : if (lc == NULL) /* not in the list? */
6841 0 : return true; /* use sort */
6842 :
6843 : /*
6844 : * Rather than doing all the pushups that would be needed to use
6845 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6846 : */
6847 160 : rel->rows = rel->tuples;
6848 160 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6849 :
6850 160 : root->total_table_pages = rel->pages;
6851 :
6852 : /*
6853 : * Determine eval cost of the index expressions, if any. We need to
6854 : * charge twice that amount for each tuple comparison that happens during
6855 : * the sort, since tuplesort.c will have to re-evaluate the index
6856 : * expressions each time. (XXX that's pretty inefficient...)
6857 : */
6858 160 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6859 160 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6860 :
6861 : /* Estimate the cost of seq scan + sort */
6862 160 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6863 160 : cost_sort(&seqScanAndSortPath, root, NIL,
6864 : seqScanPath->disabled_nodes,
6865 160 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6866 : comparisonCost, maintenance_work_mem, -1.0);
6867 :
6868 : /* Estimate the cost of index scan */
6869 160 : indexScanPath = create_index_path(root, indexInfo,
6870 : NIL, NIL, NIL, NIL,
6871 : ForwardScanDirection, false,
6872 : NULL, 1.0, false);
6873 :
6874 160 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6875 : }
6876 :
6877 : /*
6878 : * plan_create_index_workers
6879 : * Use the planner to decide how many parallel worker processes
6880 : * CREATE INDEX should request for use
6881 : *
6882 : * tableOid is the table on which the index is to be built. indexOid is the
6883 : * OID of an index to be created or reindexed (which must be an index with
6884 : * support for parallel builds - currently btree or BRIN).
6885 : *
6886 : * Return value is the number of parallel worker processes to request. It
6887 : * may be unsafe to proceed if this is 0. Note that this does not include the
6888 : * leader participating as a worker (value is always a number of parallel
6889 : * worker processes).
6890 : *
6891 : * Note: caller had better already hold some type of lock on the table and
6892 : * index.
6893 : */
6894 : int
6895 36578 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6896 : {
6897 : PlannerInfo *root;
6898 : Query *query;
6899 : PlannerGlobal *glob;
6900 : RangeTblEntry *rte;
6901 : Relation heap;
6902 : Relation index;
6903 : RelOptInfo *rel;
6904 : int parallel_workers;
6905 : BlockNumber heap_blocks;
6906 : double reltuples;
6907 : double allvisfrac;
6908 :
6909 : /*
6910 : * We don't allow performing parallel operation in standalone backend or
6911 : * when parallelism is disabled.
6912 : */
6913 36578 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
6914 498 : return 0;
6915 :
6916 : /* Set up largely-dummy planner state */
6917 36080 : query = makeNode(Query);
6918 36080 : query->commandType = CMD_SELECT;
6919 :
6920 36080 : glob = makeNode(PlannerGlobal);
6921 :
6922 36080 : root = makeNode(PlannerInfo);
6923 36080 : root->parse = query;
6924 36080 : root->glob = glob;
6925 36080 : root->query_level = 1;
6926 36080 : root->planner_cxt = CurrentMemoryContext;
6927 36080 : root->wt_param_id = -1;
6928 36080 : root->join_domains = list_make1(makeNode(JoinDomain));
6929 :
6930 : /*
6931 : * Build a minimal RTE.
6932 : *
6933 : * Mark the RTE with inh = true. This is a kludge to prevent
6934 : * get_relation_info() from fetching index info, which is necessary
6935 : * because it does not expect that any IndexOptInfo is currently
6936 : * undergoing REINDEX.
6937 : */
6938 36080 : rte = makeNode(RangeTblEntry);
6939 36080 : rte->rtekind = RTE_RELATION;
6940 36080 : rte->relid = tableOid;
6941 36080 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6942 36080 : rte->rellockmode = AccessShareLock;
6943 36080 : rte->lateral = false;
6944 36080 : rte->inh = true;
6945 36080 : rte->inFromCl = true;
6946 36080 : query->rtable = list_make1(rte);
6947 36080 : addRTEPermissionInfo(&query->rteperminfos, rte);
6948 :
6949 : /* Set up RTE/RelOptInfo arrays */
6950 36080 : setup_simple_rel_arrays(root);
6951 :
6952 : /* Build RelOptInfo */
6953 36080 : rel = build_simple_rel(root, 1, NULL);
6954 :
6955 : /* Rels are assumed already locked by the caller */
6956 36080 : heap = table_open(tableOid, NoLock);
6957 36080 : index = index_open(indexOid, NoLock);
6958 :
6959 : /*
6960 : * Determine if it's safe to proceed.
6961 : *
6962 : * Currently, parallel workers can't access the leader's temporary tables.
6963 : * Furthermore, any index predicate or index expressions must be parallel
6964 : * safe.
6965 : */
6966 36080 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6967 34052 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6968 33932 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6969 : {
6970 2148 : parallel_workers = 0;
6971 2148 : goto done;
6972 : }
6973 :
6974 : /*
6975 : * If parallel_workers storage parameter is set for the table, accept that
6976 : * as the number of parallel worker processes to launch (though still cap
6977 : * at max_parallel_maintenance_workers). Note that we deliberately do not
6978 : * consider any other factor when parallel_workers is set. (e.g., memory
6979 : * use by workers.)
6980 : */
6981 33932 : if (rel->rel_parallel_workers != -1)
6982 : {
6983 18 : parallel_workers = Min(rel->rel_parallel_workers,
6984 : max_parallel_maintenance_workers);
6985 18 : goto done;
6986 : }
6987 :
6988 : /*
6989 : * Estimate heap relation size ourselves, since rel->pages cannot be
6990 : * trusted (heap RTE was marked as inheritance parent)
6991 : */
6992 33914 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6993 :
6994 : /*
6995 : * Determine number of workers to scan the heap relation using generic
6996 : * model
6997 : */
6998 33914 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6999 : max_parallel_maintenance_workers);
7000 :
7001 : /*
7002 : * Cap workers based on available maintenance_work_mem as needed.
7003 : *
7004 : * Note that each tuplesort participant receives an even share of the
7005 : * total maintenance_work_mem budget. Aim to leave participants
7006 : * (including the leader as a participant) with no less than 32MB of
7007 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7008 : * immediately past the threshold of being capable of launching a single
7009 : * parallel worker to sort.
7010 : */
7011 34070 : while (parallel_workers > 0 &&
7012 314 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7013 156 : parallel_workers--;
7014 :
7015 33914 : done:
7016 36080 : index_close(index, NoLock);
7017 36080 : table_close(heap, NoLock);
7018 :
7019 36080 : return parallel_workers;
7020 : }
7021 :
7022 : /*
7023 : * add_paths_to_grouping_rel
7024 : *
7025 : * Add non-partial paths to grouping relation.
7026 : */
7027 : static void
7028 42574 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7029 : RelOptInfo *grouped_rel,
7030 : RelOptInfo *partially_grouped_rel,
7031 : const AggClauseCosts *agg_costs,
7032 : grouping_sets_data *gd, double dNumGroups,
7033 : GroupPathExtraData *extra)
7034 : {
7035 42574 : Query *parse = root->parse;
7036 42574 : Path *cheapest_path = input_rel->cheapest_total_path;
7037 : ListCell *lc;
7038 42574 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7039 42574 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7040 42574 : List *havingQual = (List *) extra->havingQual;
7041 42574 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7042 :
7043 42574 : if (can_sort)
7044 : {
7045 : /*
7046 : * Use any available suitably-sorted path as input, and also consider
7047 : * sorting the cheapest-total path and incremental sort on any paths
7048 : * with presorted keys.
7049 : */
7050 87808 : foreach(lc, input_rel->pathlist)
7051 : {
7052 : ListCell *lc2;
7053 45240 : Path *path = (Path *) lfirst(lc);
7054 45240 : Path *path_save = path;
7055 45240 : List *pathkey_orderings = NIL;
7056 :
7057 : /* generate alternative group orderings that might be useful */
7058 45240 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7059 :
7060 : Assert(list_length(pathkey_orderings) > 0);
7061 :
7062 90624 : foreach(lc2, pathkey_orderings)
7063 : {
7064 45384 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7065 :
7066 : /* restore the path (we replace it in the loop) */
7067 45384 : path = path_save;
7068 :
7069 45384 : path = make_ordered_path(root,
7070 : grouped_rel,
7071 : path,
7072 : cheapest_path,
7073 : info->pathkeys,
7074 : -1.0);
7075 45384 : if (path == NULL)
7076 368 : continue;
7077 :
7078 : /* Now decide what to stick atop it */
7079 45016 : if (parse->groupingSets)
7080 : {
7081 938 : consider_groupingsets_paths(root, grouped_rel,
7082 : path, true, can_hash,
7083 : gd, agg_costs, dNumGroups);
7084 : }
7085 44078 : else if (parse->hasAggs)
7086 : {
7087 : /*
7088 : * We have aggregation, possibly with plain GROUP BY. Make
7089 : * an AggPath.
7090 : */
7091 43306 : add_path(grouped_rel, (Path *)
7092 43306 : create_agg_path(root,
7093 : grouped_rel,
7094 : path,
7095 43306 : grouped_rel->reltarget,
7096 43306 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7097 : AGGSPLIT_SIMPLE,
7098 : info->clauses,
7099 : havingQual,
7100 : agg_costs,
7101 : dNumGroups));
7102 : }
7103 772 : else if (parse->groupClause)
7104 : {
7105 : /*
7106 : * We have GROUP BY without aggregation or grouping sets.
7107 : * Make a GroupPath.
7108 : */
7109 772 : add_path(grouped_rel, (Path *)
7110 772 : create_group_path(root,
7111 : grouped_rel,
7112 : path,
7113 : info->clauses,
7114 : havingQual,
7115 : dNumGroups));
7116 : }
7117 : else
7118 : {
7119 : /* Other cases should have been handled above */
7120 : Assert(false);
7121 : }
7122 : }
7123 : }
7124 :
7125 : /*
7126 : * Instead of operating directly on the input relation, we can
7127 : * consider finalizing a partially aggregated path.
7128 : */
7129 42568 : if (partially_grouped_rel != NULL)
7130 : {
7131 3990 : foreach(lc, partially_grouped_rel->pathlist)
7132 : {
7133 : ListCell *lc2;
7134 2410 : Path *path = (Path *) lfirst(lc);
7135 2410 : Path *path_save = path;
7136 2410 : List *pathkey_orderings = NIL;
7137 :
7138 : /* generate alternative group orderings that might be useful */
7139 2410 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7140 :
7141 : Assert(list_length(pathkey_orderings) > 0);
7142 :
7143 : /* process all potentially interesting grouping reorderings */
7144 4820 : foreach(lc2, pathkey_orderings)
7145 : {
7146 2410 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7147 :
7148 : /* restore the path (we replace it in the loop) */
7149 2410 : path = path_save;
7150 :
7151 2410 : path = make_ordered_path(root,
7152 : grouped_rel,
7153 : path,
7154 2410 : partially_grouped_rel->cheapest_total_path,
7155 : info->pathkeys,
7156 : -1.0);
7157 :
7158 2410 : if (path == NULL)
7159 108 : continue;
7160 :
7161 2302 : if (parse->hasAggs)
7162 2054 : add_path(grouped_rel, (Path *)
7163 2054 : create_agg_path(root,
7164 : grouped_rel,
7165 : path,
7166 2054 : grouped_rel->reltarget,
7167 2054 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7168 : AGGSPLIT_FINAL_DESERIAL,
7169 : info->clauses,
7170 : havingQual,
7171 : agg_final_costs,
7172 : dNumGroups));
7173 : else
7174 248 : add_path(grouped_rel, (Path *)
7175 248 : create_group_path(root,
7176 : grouped_rel,
7177 : path,
7178 : info->clauses,
7179 : havingQual,
7180 : dNumGroups));
7181 :
7182 : }
7183 : }
7184 : }
7185 : }
7186 :
7187 42574 : if (can_hash)
7188 : {
7189 4982 : if (parse->groupingSets)
7190 : {
7191 : /*
7192 : * Try for a hash-only groupingsets path over unsorted input.
7193 : */
7194 794 : consider_groupingsets_paths(root, grouped_rel,
7195 : cheapest_path, false, true,
7196 : gd, agg_costs, dNumGroups);
7197 : }
7198 : else
7199 : {
7200 : /*
7201 : * Generate a HashAgg Path. We just need an Agg over the
7202 : * cheapest-total input path, since input order won't matter.
7203 : */
7204 4188 : add_path(grouped_rel, (Path *)
7205 4188 : create_agg_path(root, grouped_rel,
7206 : cheapest_path,
7207 4188 : grouped_rel->reltarget,
7208 : AGG_HASHED,
7209 : AGGSPLIT_SIMPLE,
7210 : root->processed_groupClause,
7211 : havingQual,
7212 : agg_costs,
7213 : dNumGroups));
7214 : }
7215 :
7216 : /*
7217 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7218 : * grouped path, assuming there is one
7219 : */
7220 4982 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7221 : {
7222 784 : Path *path = partially_grouped_rel->cheapest_total_path;
7223 :
7224 784 : add_path(grouped_rel, (Path *)
7225 784 : create_agg_path(root,
7226 : grouped_rel,
7227 : path,
7228 784 : grouped_rel->reltarget,
7229 : AGG_HASHED,
7230 : AGGSPLIT_FINAL_DESERIAL,
7231 : root->processed_groupClause,
7232 : havingQual,
7233 : agg_final_costs,
7234 : dNumGroups));
7235 : }
7236 : }
7237 :
7238 : /*
7239 : * When partitionwise aggregate is used, we might have fully aggregated
7240 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7241 : * consider a path for grouped_rel consisting of a Parallel Append of
7242 : * non-partial paths from each child.
7243 : */
7244 42574 : if (grouped_rel->partial_pathlist != NIL)
7245 162 : gather_grouping_paths(root, grouped_rel);
7246 42574 : }
7247 :
7248 : /*
7249 : * create_partial_grouping_paths
7250 : *
7251 : * Create a new upper relation representing the result of partial aggregation
7252 : * and populate it with appropriate paths. Note that we don't finalize the
7253 : * lists of paths here, so the caller can add additional partial or non-partial
7254 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7255 : * the returned upper relation.
7256 : *
7257 : * All paths for this new upper relation -- both partial and non-partial --
7258 : * have been partially aggregated but require a subsequent FinalizeAggregate
7259 : * step.
7260 : *
7261 : * NB: This function is allowed to return NULL if it determines that there is
7262 : * no real need to create a new RelOptInfo.
7263 : */
7264 : static RelOptInfo *
7265 38168 : create_partial_grouping_paths(PlannerInfo *root,
7266 : RelOptInfo *grouped_rel,
7267 : RelOptInfo *input_rel,
7268 : grouping_sets_data *gd,
7269 : GroupPathExtraData *extra,
7270 : bool force_rel_creation)
7271 : {
7272 38168 : Query *parse = root->parse;
7273 : RelOptInfo *partially_grouped_rel;
7274 38168 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7275 38168 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7276 38168 : Path *cheapest_partial_path = NULL;
7277 38168 : Path *cheapest_total_path = NULL;
7278 38168 : double dNumPartialGroups = 0;
7279 38168 : double dNumPartialPartialGroups = 0;
7280 : ListCell *lc;
7281 38168 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7282 38168 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7283 :
7284 : /*
7285 : * Consider whether we should generate partially aggregated non-partial
7286 : * paths. We can only do this if we have a non-partial path, and only if
7287 : * the parent of the input rel is performing partial partitionwise
7288 : * aggregation. (Note that extra->patype is the type of partitionwise
7289 : * aggregation being used at the parent level, not this level.)
7290 : */
7291 38168 : if (input_rel->pathlist != NIL &&
7292 38168 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7293 618 : cheapest_total_path = input_rel->cheapest_total_path;
7294 :
7295 : /*
7296 : * If parallelism is possible for grouped_rel, then we should consider
7297 : * generating partially-grouped partial paths. However, if the input rel
7298 : * has no partial paths, then we can't.
7299 : */
7300 38168 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7301 1782 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7302 :
7303 : /*
7304 : * If we can't partially aggregate partial paths, and we can't partially
7305 : * aggregate non-partial paths, then don't bother creating the new
7306 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7307 : */
7308 38168 : if (cheapest_total_path == NULL &&
7309 36068 : cheapest_partial_path == NULL &&
7310 36068 : !force_rel_creation)
7311 35970 : return NULL;
7312 :
7313 : /*
7314 : * Build a new upper relation to represent the result of partially
7315 : * aggregating the rows from the input relation.
7316 : */
7317 2198 : partially_grouped_rel = fetch_upper_rel(root,
7318 : UPPERREL_PARTIAL_GROUP_AGG,
7319 : grouped_rel->relids);
7320 2198 : partially_grouped_rel->consider_parallel =
7321 2198 : grouped_rel->consider_parallel;
7322 2198 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7323 2198 : partially_grouped_rel->serverid = grouped_rel->serverid;
7324 2198 : partially_grouped_rel->userid = grouped_rel->userid;
7325 2198 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7326 2198 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7327 :
7328 : /*
7329 : * Build target list for partial aggregate paths. These paths cannot just
7330 : * emit the same tlist as regular aggregate paths, because (1) we must
7331 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7332 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7333 : */
7334 2198 : partially_grouped_rel->reltarget =
7335 2198 : make_partial_grouping_target(root, grouped_rel->reltarget,
7336 : extra->havingQual);
7337 :
7338 2198 : if (!extra->partial_costs_set)
7339 : {
7340 : /*
7341 : * Collect statistics about aggregates for estimating costs of
7342 : * performing aggregation in parallel.
7343 : */
7344 7752 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7345 7752 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7346 1292 : if (parse->hasAggs)
7347 : {
7348 : /* partial phase */
7349 1158 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7350 : agg_partial_costs);
7351 :
7352 : /* final phase */
7353 1158 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7354 : agg_final_costs);
7355 : }
7356 :
7357 1292 : extra->partial_costs_set = true;
7358 : }
7359 :
7360 : /* Estimate number of partial groups. */
7361 2198 : if (cheapest_total_path != NULL)
7362 : dNumPartialGroups =
7363 618 : get_number_of_groups(root,
7364 : cheapest_total_path->rows,
7365 : gd,
7366 : extra->targetList);
7367 2198 : if (cheapest_partial_path != NULL)
7368 : dNumPartialPartialGroups =
7369 1782 : get_number_of_groups(root,
7370 : cheapest_partial_path->rows,
7371 : gd,
7372 : extra->targetList);
7373 :
7374 2198 : if (can_sort && cheapest_total_path != NULL)
7375 : {
7376 : /* This should have been checked previously */
7377 : Assert(parse->hasAggs || parse->groupClause);
7378 :
7379 : /*
7380 : * Use any available suitably-sorted path as input, and also consider
7381 : * sorting the cheapest partial path.
7382 : */
7383 1236 : foreach(lc, input_rel->pathlist)
7384 : {
7385 : ListCell *lc2;
7386 618 : Path *path = (Path *) lfirst(lc);
7387 618 : Path *path_save = path;
7388 618 : List *pathkey_orderings = NIL;
7389 :
7390 : /* generate alternative group orderings that might be useful */
7391 618 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7392 :
7393 : Assert(list_length(pathkey_orderings) > 0);
7394 :
7395 : /* process all potentially interesting grouping reorderings */
7396 1236 : foreach(lc2, pathkey_orderings)
7397 : {
7398 618 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7399 :
7400 : /* restore the path (we replace it in the loop) */
7401 618 : path = path_save;
7402 :
7403 618 : path = make_ordered_path(root,
7404 : partially_grouped_rel,
7405 : path,
7406 : cheapest_total_path,
7407 : info->pathkeys,
7408 : -1.0);
7409 :
7410 618 : if (path == NULL)
7411 0 : continue;
7412 :
7413 618 : if (parse->hasAggs)
7414 546 : add_path(partially_grouped_rel, (Path *)
7415 546 : create_agg_path(root,
7416 : partially_grouped_rel,
7417 : path,
7418 546 : partially_grouped_rel->reltarget,
7419 546 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7420 : AGGSPLIT_INITIAL_SERIAL,
7421 : info->clauses,
7422 : NIL,
7423 : agg_partial_costs,
7424 : dNumPartialGroups));
7425 : else
7426 72 : add_path(partially_grouped_rel, (Path *)
7427 72 : create_group_path(root,
7428 : partially_grouped_rel,
7429 : path,
7430 : info->clauses,
7431 : NIL,
7432 : dNumPartialGroups));
7433 : }
7434 : }
7435 : }
7436 :
7437 2198 : if (can_sort && cheapest_partial_path != NULL)
7438 : {
7439 : /* Similar to above logic, but for partial paths. */
7440 3576 : foreach(lc, input_rel->partial_pathlist)
7441 : {
7442 : ListCell *lc2;
7443 1794 : Path *path = (Path *) lfirst(lc);
7444 1794 : Path *path_save = path;
7445 1794 : List *pathkey_orderings = NIL;
7446 :
7447 : /* generate alternative group orderings that might be useful */
7448 1794 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7449 :
7450 : Assert(list_length(pathkey_orderings) > 0);
7451 :
7452 : /* process all potentially interesting grouping reorderings */
7453 3588 : foreach(lc2, pathkey_orderings)
7454 : {
7455 1794 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7456 :
7457 :
7458 : /* restore the path (we replace it in the loop) */
7459 1794 : path = path_save;
7460 :
7461 1794 : path = make_ordered_path(root,
7462 : partially_grouped_rel,
7463 : path,
7464 : cheapest_partial_path,
7465 : info->pathkeys,
7466 : -1.0);
7467 :
7468 1794 : if (path == NULL)
7469 6 : continue;
7470 :
7471 1788 : if (parse->hasAggs)
7472 1666 : add_partial_path(partially_grouped_rel, (Path *)
7473 1666 : create_agg_path(root,
7474 : partially_grouped_rel,
7475 : path,
7476 1666 : partially_grouped_rel->reltarget,
7477 1666 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7478 : AGGSPLIT_INITIAL_SERIAL,
7479 : info->clauses,
7480 : NIL,
7481 : agg_partial_costs,
7482 : dNumPartialPartialGroups));
7483 : else
7484 122 : add_partial_path(partially_grouped_rel, (Path *)
7485 122 : create_group_path(root,
7486 : partially_grouped_rel,
7487 : path,
7488 : info->clauses,
7489 : NIL,
7490 : dNumPartialPartialGroups));
7491 : }
7492 : }
7493 : }
7494 :
7495 : /*
7496 : * Add a partially-grouped HashAgg Path where possible
7497 : */
7498 2198 : if (can_hash && cheapest_total_path != NULL)
7499 : {
7500 : /* Checked above */
7501 : Assert(parse->hasAggs || parse->groupClause);
7502 :
7503 618 : add_path(partially_grouped_rel, (Path *)
7504 618 : create_agg_path(root,
7505 : partially_grouped_rel,
7506 : cheapest_total_path,
7507 618 : partially_grouped_rel->reltarget,
7508 : AGG_HASHED,
7509 : AGGSPLIT_INITIAL_SERIAL,
7510 : root->processed_groupClause,
7511 : NIL,
7512 : agg_partial_costs,
7513 : dNumPartialGroups));
7514 : }
7515 :
7516 : /*
7517 : * Now add a partially-grouped HashAgg partial Path where possible
7518 : */
7519 2198 : if (can_hash && cheapest_partial_path != NULL)
7520 : {
7521 986 : add_partial_path(partially_grouped_rel, (Path *)
7522 986 : create_agg_path(root,
7523 : partially_grouped_rel,
7524 : cheapest_partial_path,
7525 986 : partially_grouped_rel->reltarget,
7526 : AGG_HASHED,
7527 : AGGSPLIT_INITIAL_SERIAL,
7528 : root->processed_groupClause,
7529 : NIL,
7530 : agg_partial_costs,
7531 : dNumPartialPartialGroups));
7532 : }
7533 :
7534 : /*
7535 : * If there is an FDW that's responsible for all baserels of the query,
7536 : * let it consider adding partially grouped ForeignPaths.
7537 : */
7538 2198 : if (partially_grouped_rel->fdwroutine &&
7539 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7540 : {
7541 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7542 :
7543 6 : fdwroutine->GetForeignUpperPaths(root,
7544 : UPPERREL_PARTIAL_GROUP_AGG,
7545 : input_rel, partially_grouped_rel,
7546 : extra);
7547 : }
7548 :
7549 2198 : return partially_grouped_rel;
7550 : }
7551 :
7552 : /*
7553 : * make_ordered_path
7554 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7555 : * return NULL if it doesn't make sense to generate an ordered path in
7556 : * this case.
7557 : */
7558 : static Path *
7559 55212 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7560 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7561 : {
7562 : bool is_sorted;
7563 : int presorted_keys;
7564 :
7565 55212 : is_sorted = pathkeys_count_contained_in(pathkeys,
7566 : path->pathkeys,
7567 : &presorted_keys);
7568 :
7569 55212 : if (!is_sorted)
7570 : {
7571 : /*
7572 : * Try at least sorting the cheapest path and also try incrementally
7573 : * sorting any path which is partially sorted already (no need to deal
7574 : * with paths which have presorted keys when incremental sort is
7575 : * disabled unless it's the cheapest input path).
7576 : */
7577 12896 : if (path != cheapest_path &&
7578 2034 : (presorted_keys == 0 || !enable_incremental_sort))
7579 1038 : return NULL;
7580 :
7581 : /*
7582 : * We've no need to consider both a sort and incremental sort. We'll
7583 : * just do a sort if there are no presorted keys and an incremental
7584 : * sort when there are presorted keys.
7585 : */
7586 11858 : if (presorted_keys == 0 || !enable_incremental_sort)
7587 10694 : path = (Path *) create_sort_path(root,
7588 : rel,
7589 : path,
7590 : pathkeys,
7591 : limit_tuples);
7592 : else
7593 1164 : path = (Path *) create_incremental_sort_path(root,
7594 : rel,
7595 : path,
7596 : pathkeys,
7597 : presorted_keys,
7598 : limit_tuples);
7599 : }
7600 :
7601 54174 : return path;
7602 : }
7603 :
7604 : /*
7605 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7606 : * grouping relation.
7607 : *
7608 : * generate_useful_gather_paths does most of the work, but we also consider a
7609 : * special case: we could try sorting the data by the group_pathkeys and then
7610 : * applying Gather Merge.
7611 : *
7612 : * NB: This function shouldn't be used for anything other than a grouped or
7613 : * partially grouped relation not only because of the fact that it explicitly
7614 : * references group_pathkeys but we pass "true" as the third argument to
7615 : * generate_useful_gather_paths().
7616 : */
7617 : static void
7618 1644 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7619 : {
7620 : ListCell *lc;
7621 : Path *cheapest_partial_path;
7622 : List *groupby_pathkeys;
7623 :
7624 : /*
7625 : * This occurs after any partial aggregation has taken place, so trim off
7626 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7627 : */
7628 1644 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7629 18 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7630 : root->num_groupby_pathkeys);
7631 : else
7632 1626 : groupby_pathkeys = root->group_pathkeys;
7633 :
7634 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7635 1644 : generate_useful_gather_paths(root, rel, true);
7636 :
7637 1644 : cheapest_partial_path = linitial(rel->partial_pathlist);
7638 :
7639 : /* XXX Shouldn't this also consider the group-key-reordering? */
7640 3894 : foreach(lc, rel->partial_pathlist)
7641 : {
7642 2250 : Path *path = (Path *) lfirst(lc);
7643 : bool is_sorted;
7644 : int presorted_keys;
7645 : double total_groups;
7646 :
7647 2250 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7648 : path->pathkeys,
7649 : &presorted_keys);
7650 :
7651 2250 : if (is_sorted)
7652 1470 : continue;
7653 :
7654 : /*
7655 : * Try at least sorting the cheapest path and also try incrementally
7656 : * sorting any path which is partially sorted already (no need to deal
7657 : * with paths which have presorted keys when incremental sort is
7658 : * disabled unless it's the cheapest input path).
7659 : */
7660 780 : if (path != cheapest_partial_path &&
7661 0 : (presorted_keys == 0 || !enable_incremental_sort))
7662 0 : continue;
7663 :
7664 : /*
7665 : * We've no need to consider both a sort and incremental sort. We'll
7666 : * just do a sort if there are no presorted keys and an incremental
7667 : * sort when there are presorted keys.
7668 : */
7669 780 : if (presorted_keys == 0 || !enable_incremental_sort)
7670 780 : path = (Path *) create_sort_path(root, rel, path,
7671 : groupby_pathkeys,
7672 : -1.0);
7673 : else
7674 0 : path = (Path *) create_incremental_sort_path(root,
7675 : rel,
7676 : path,
7677 : groupby_pathkeys,
7678 : presorted_keys,
7679 : -1.0);
7680 780 : total_groups = compute_gather_rows(path);
7681 : path = (Path *)
7682 780 : create_gather_merge_path(root,
7683 : rel,
7684 : path,
7685 780 : rel->reltarget,
7686 : groupby_pathkeys,
7687 : NULL,
7688 : &total_groups);
7689 :
7690 780 : add_path(rel, path);
7691 : }
7692 1644 : }
7693 :
7694 : /*
7695 : * can_partial_agg
7696 : *
7697 : * Determines whether or not partial grouping and/or aggregation is possible.
7698 : * Returns true when possible, false otherwise.
7699 : */
7700 : static bool
7701 41698 : can_partial_agg(PlannerInfo *root)
7702 : {
7703 41698 : Query *parse = root->parse;
7704 :
7705 41698 : if (!parse->hasAggs && parse->groupClause == NIL)
7706 : {
7707 : /*
7708 : * We don't know how to do parallel aggregation unless we have either
7709 : * some aggregates or a grouping clause.
7710 : */
7711 0 : return false;
7712 : }
7713 41698 : else if (parse->groupingSets)
7714 : {
7715 : /* We don't know how to do grouping sets in parallel. */
7716 872 : return false;
7717 : }
7718 40826 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7719 : {
7720 : /* Insufficient support for partial mode. */
7721 4080 : return false;
7722 : }
7723 :
7724 : /* Everything looks good. */
7725 36746 : return true;
7726 : }
7727 :
7728 : /*
7729 : * apply_scanjoin_target_to_paths
7730 : *
7731 : * Adjust the final scan/join relation, and recursively all of its children,
7732 : * to generate the final scan/join target. It would be more correct to model
7733 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7734 : * for each child relation, but doing it this way is noticeably cheaper.
7735 : * Maybe that problem can be solved at some point, but for now we do this.
7736 : *
7737 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7738 : * the same expressions as the existing reltarget, so we need only insert the
7739 : * appropriate sortgroupref information. By avoiding the creation of
7740 : * projection paths we save effort both immediately and at plan creation time.
7741 : */
7742 : static void
7743 549320 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7744 : RelOptInfo *rel,
7745 : List *scanjoin_targets,
7746 : List *scanjoin_targets_contain_srfs,
7747 : bool scanjoin_target_parallel_safe,
7748 : bool tlist_same_exprs)
7749 : {
7750 549320 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7751 : PathTarget *scanjoin_target;
7752 : ListCell *lc;
7753 :
7754 : /* This recurses, so be paranoid. */
7755 549320 : check_stack_depth();
7756 :
7757 : /*
7758 : * If the rel is partitioned, we want to drop its existing paths and
7759 : * generate new ones. This function would still be correct if we kept the
7760 : * existing paths: we'd modify them to generate the correct target above
7761 : * the partitioning Append, and then they'd compete on cost with paths
7762 : * generating the target below the Append. However, in our current cost
7763 : * model the latter way is always the same or cheaper cost, so modifying
7764 : * the existing paths would just be useless work. Moreover, when the cost
7765 : * is the same, varying roundoff errors might sometimes allow an existing
7766 : * path to be picked, resulting in undesirable cross-platform plan
7767 : * variations. So we drop old paths and thereby force the work to be done
7768 : * below the Append, except in the case of a non-parallel-safe target.
7769 : *
7770 : * Some care is needed, because we have to allow
7771 : * generate_useful_gather_paths to see the old partial paths in the next
7772 : * stanza. Hence, zap the main pathlist here, then allow
7773 : * generate_useful_gather_paths to add path(s) to the main list, and
7774 : * finally zap the partial pathlist.
7775 : */
7776 549320 : if (rel_is_partitioned)
7777 12544 : rel->pathlist = NIL;
7778 :
7779 : /*
7780 : * If the scan/join target is not parallel-safe, partial paths cannot
7781 : * generate it.
7782 : */
7783 549320 : if (!scanjoin_target_parallel_safe)
7784 : {
7785 : /*
7786 : * Since we can't generate the final scan/join target in parallel
7787 : * workers, this is our last opportunity to use any partial paths that
7788 : * exist; so build Gather path(s) that use them and emit whatever the
7789 : * current reltarget is. We don't do this in the case where the
7790 : * target is parallel-safe, since we will be able to generate superior
7791 : * paths by doing it after the final scan/join target has been
7792 : * applied.
7793 : */
7794 79822 : generate_useful_gather_paths(root, rel, false);
7795 :
7796 : /* Can't use parallel query above this level. */
7797 79822 : rel->partial_pathlist = NIL;
7798 79822 : rel->consider_parallel = false;
7799 : }
7800 :
7801 : /* Finish dropping old paths for a partitioned rel, per comment above */
7802 549320 : if (rel_is_partitioned)
7803 12544 : rel->partial_pathlist = NIL;
7804 :
7805 : /* Extract SRF-free scan/join target. */
7806 549320 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7807 :
7808 : /*
7809 : * Apply the SRF-free scan/join target to each existing path.
7810 : *
7811 : * If the tlist exprs are the same, we can just inject the sortgroupref
7812 : * information into the existing pathtargets. Otherwise, replace each
7813 : * path with a projection path that generates the SRF-free scan/join
7814 : * target. This can't change the ordering of paths within rel->pathlist,
7815 : * so we just modify the list in place.
7816 : */
7817 1139672 : foreach(lc, rel->pathlist)
7818 : {
7819 590352 : Path *subpath = (Path *) lfirst(lc);
7820 :
7821 : /* Shouldn't have any parameterized paths anymore */
7822 : Assert(subpath->param_info == NULL);
7823 :
7824 590352 : if (tlist_same_exprs)
7825 211352 : subpath->pathtarget->sortgrouprefs =
7826 211352 : scanjoin_target->sortgrouprefs;
7827 : else
7828 : {
7829 : Path *newpath;
7830 :
7831 379000 : newpath = (Path *) create_projection_path(root, rel, subpath,
7832 : scanjoin_target);
7833 379000 : lfirst(lc) = newpath;
7834 : }
7835 : }
7836 :
7837 : /* Likewise adjust the targets for any partial paths. */
7838 568904 : foreach(lc, rel->partial_pathlist)
7839 : {
7840 19584 : Path *subpath = (Path *) lfirst(lc);
7841 :
7842 : /* Shouldn't have any parameterized paths anymore */
7843 : Assert(subpath->param_info == NULL);
7844 :
7845 19584 : if (tlist_same_exprs)
7846 15950 : subpath->pathtarget->sortgrouprefs =
7847 15950 : scanjoin_target->sortgrouprefs;
7848 : else
7849 : {
7850 : Path *newpath;
7851 :
7852 3634 : newpath = (Path *) create_projection_path(root, rel, subpath,
7853 : scanjoin_target);
7854 3634 : lfirst(lc) = newpath;
7855 : }
7856 : }
7857 :
7858 : /*
7859 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7860 : * atop each existing path. (Note that this function doesn't look at the
7861 : * cheapest-path fields, which is a good thing because they're bogus right
7862 : * now.)
7863 : */
7864 549320 : if (root->parse->hasTargetSRFs)
7865 12020 : adjust_paths_for_srfs(root, rel,
7866 : scanjoin_targets,
7867 : scanjoin_targets_contain_srfs);
7868 :
7869 : /*
7870 : * Update the rel's target to be the final (with SRFs) scan/join target.
7871 : * This now matches the actual output of all the paths, and we might get
7872 : * confused in createplan.c if they don't agree. We must do this now so
7873 : * that any append paths made in the next part will use the correct
7874 : * pathtarget (cf. create_append_path).
7875 : *
7876 : * Note that this is also necessary if GetForeignUpperPaths() gets called
7877 : * on the final scan/join relation or on any of its children, since the
7878 : * FDW might look at the rel's target to create ForeignPaths.
7879 : */
7880 549320 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7881 :
7882 : /*
7883 : * If the relation is partitioned, recursively apply the scan/join target
7884 : * to all partitions, and generate brand-new Append paths in which the
7885 : * scan/join target is computed below the Append rather than above it.
7886 : * Since Append is not projection-capable, that might save a separate
7887 : * Result node, and it also is important for partitionwise aggregate.
7888 : */
7889 549320 : if (rel_is_partitioned)
7890 : {
7891 12544 : List *live_children = NIL;
7892 : int i;
7893 :
7894 : /* Adjust each partition. */
7895 12544 : i = -1;
7896 35426 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
7897 : {
7898 22882 : RelOptInfo *child_rel = rel->part_rels[i];
7899 : AppendRelInfo **appinfos;
7900 : int nappinfos;
7901 22882 : List *child_scanjoin_targets = NIL;
7902 :
7903 : Assert(child_rel != NULL);
7904 :
7905 : /* Dummy children can be ignored. */
7906 22882 : if (IS_DUMMY_REL(child_rel))
7907 42 : continue;
7908 :
7909 : /* Translate scan/join targets for this child. */
7910 22840 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7911 : &nappinfos);
7912 45680 : foreach(lc, scanjoin_targets)
7913 : {
7914 22840 : PathTarget *target = lfirst_node(PathTarget, lc);
7915 :
7916 22840 : target = copy_pathtarget(target);
7917 22840 : target->exprs = (List *)
7918 22840 : adjust_appendrel_attrs(root,
7919 22840 : (Node *) target->exprs,
7920 : nappinfos, appinfos);
7921 22840 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7922 : target);
7923 : }
7924 22840 : pfree(appinfos);
7925 :
7926 : /* Recursion does the real work. */
7927 22840 : apply_scanjoin_target_to_paths(root, child_rel,
7928 : child_scanjoin_targets,
7929 : scanjoin_targets_contain_srfs,
7930 : scanjoin_target_parallel_safe,
7931 : tlist_same_exprs);
7932 :
7933 : /* Save non-dummy children for Append paths. */
7934 22840 : if (!IS_DUMMY_REL(child_rel))
7935 22840 : live_children = lappend(live_children, child_rel);
7936 : }
7937 :
7938 : /* Build new paths for this relation by appending child paths. */
7939 12544 : add_paths_to_append_rel(root, rel, live_children);
7940 : }
7941 :
7942 : /*
7943 : * Consider generating Gather or Gather Merge paths. We must only do this
7944 : * if the relation is parallel safe, and we don't do it for child rels to
7945 : * avoid creating multiple Gather nodes within the same plan. We must do
7946 : * this after all paths have been generated and before set_cheapest, since
7947 : * one of the generated paths may turn out to be the cheapest one.
7948 : */
7949 549320 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7950 178034 : generate_useful_gather_paths(root, rel, false);
7951 :
7952 : /*
7953 : * Reassess which paths are the cheapest, now that we've potentially added
7954 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7955 : * this relation.
7956 : */
7957 549320 : set_cheapest(rel);
7958 549320 : }
7959 :
7960 : /*
7961 : * create_partitionwise_grouping_paths
7962 : *
7963 : * If the partition keys of input relation are part of the GROUP BY clause, all
7964 : * the rows belonging to a given group come from a single partition. This
7965 : * allows aggregation/grouping over a partitioned relation to be broken down
7966 : * into aggregation/grouping on each partition. This should be no worse, and
7967 : * often better, than the normal approach.
7968 : *
7969 : * However, if the GROUP BY clause does not contain all the partition keys,
7970 : * rows from a given group may be spread across multiple partitions. In that
7971 : * case, we perform partial aggregation for each group, append the results,
7972 : * and then finalize aggregation. This is less certain to win than the
7973 : * previous case. It may win if the PartialAggregate stage greatly reduces
7974 : * the number of groups, because fewer rows will pass through the Append node.
7975 : * It may lose if we have lots of small groups.
7976 : */
7977 : static void
7978 562 : create_partitionwise_grouping_paths(PlannerInfo *root,
7979 : RelOptInfo *input_rel,
7980 : RelOptInfo *grouped_rel,
7981 : RelOptInfo *partially_grouped_rel,
7982 : const AggClauseCosts *agg_costs,
7983 : grouping_sets_data *gd,
7984 : PartitionwiseAggregateType patype,
7985 : GroupPathExtraData *extra)
7986 : {
7987 562 : List *grouped_live_children = NIL;
7988 562 : List *partially_grouped_live_children = NIL;
7989 562 : PathTarget *target = grouped_rel->reltarget;
7990 562 : bool partial_grouping_valid = true;
7991 : int i;
7992 :
7993 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7994 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7995 : partially_grouped_rel != NULL);
7996 :
7997 : /* Add paths for partitionwise aggregation/grouping. */
7998 562 : i = -1;
7999 2056 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8000 : {
8001 1494 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8002 : PathTarget *child_target;
8003 : AppendRelInfo **appinfos;
8004 : int nappinfos;
8005 : GroupPathExtraData child_extra;
8006 : RelOptInfo *child_grouped_rel;
8007 : RelOptInfo *child_partially_grouped_rel;
8008 :
8009 : Assert(child_input_rel != NULL);
8010 :
8011 : /* Dummy children can be ignored. */
8012 1494 : if (IS_DUMMY_REL(child_input_rel))
8013 0 : continue;
8014 :
8015 1494 : child_target = copy_pathtarget(target);
8016 :
8017 : /*
8018 : * Copy the given "extra" structure as is and then override the
8019 : * members specific to this child.
8020 : */
8021 1494 : memcpy(&child_extra, extra, sizeof(child_extra));
8022 :
8023 1494 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8024 : &nappinfos);
8025 :
8026 1494 : child_target->exprs = (List *)
8027 1494 : adjust_appendrel_attrs(root,
8028 1494 : (Node *) target->exprs,
8029 : nappinfos, appinfos);
8030 :
8031 : /* Translate havingQual and targetList. */
8032 1494 : child_extra.havingQual = (Node *)
8033 : adjust_appendrel_attrs(root,
8034 : extra->havingQual,
8035 : nappinfos, appinfos);
8036 1494 : child_extra.targetList = (List *)
8037 1494 : adjust_appendrel_attrs(root,
8038 1494 : (Node *) extra->targetList,
8039 : nappinfos, appinfos);
8040 :
8041 : /*
8042 : * extra->patype was the value computed for our parent rel; patype is
8043 : * the value for this relation. For the child, our value is its
8044 : * parent rel's value.
8045 : */
8046 1494 : child_extra.patype = patype;
8047 :
8048 : /*
8049 : * Create grouping relation to hold fully aggregated grouping and/or
8050 : * aggregation paths for the child.
8051 : */
8052 1494 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8053 : child_target,
8054 1494 : extra->target_parallel_safe,
8055 : child_extra.havingQual);
8056 :
8057 : /* Create grouping paths for this child relation. */
8058 1494 : create_ordinary_grouping_paths(root, child_input_rel,
8059 : child_grouped_rel,
8060 : agg_costs, gd, &child_extra,
8061 : &child_partially_grouped_rel);
8062 :
8063 1494 : if (child_partially_grouped_rel)
8064 : {
8065 : partially_grouped_live_children =
8066 906 : lappend(partially_grouped_live_children,
8067 : child_partially_grouped_rel);
8068 : }
8069 : else
8070 588 : partial_grouping_valid = false;
8071 :
8072 1494 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8073 : {
8074 876 : set_cheapest(child_grouped_rel);
8075 876 : grouped_live_children = lappend(grouped_live_children,
8076 : child_grouped_rel);
8077 : }
8078 :
8079 1494 : pfree(appinfos);
8080 : }
8081 :
8082 : /*
8083 : * Try to create append paths for partially grouped children. For full
8084 : * partitionwise aggregation, we might have paths in the partial_pathlist
8085 : * if parallel aggregation is possible. For partial partitionwise
8086 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8087 : *
8088 : * NB: We must have a partially grouped path for every child in order to
8089 : * generate a partially grouped path for this relation.
8090 : */
8091 562 : if (partially_grouped_rel && partial_grouping_valid)
8092 : {
8093 : Assert(partially_grouped_live_children != NIL);
8094 :
8095 350 : add_paths_to_append_rel(root, partially_grouped_rel,
8096 : partially_grouped_live_children);
8097 :
8098 : /*
8099 : * We need call set_cheapest, since the finalization step will use the
8100 : * cheapest path from the rel.
8101 : */
8102 350 : if (partially_grouped_rel->pathlist)
8103 350 : set_cheapest(partially_grouped_rel);
8104 : }
8105 :
8106 : /* If possible, create append paths for fully grouped children. */
8107 562 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8108 : {
8109 : Assert(grouped_live_children != NIL);
8110 :
8111 320 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8112 : }
8113 562 : }
8114 :
8115 : /*
8116 : * group_by_has_partkey
8117 : *
8118 : * Returns true if all the partition keys of the given relation are part of
8119 : * the GROUP BY clauses, including having matching collation, false otherwise.
8120 : */
8121 : static bool
8122 556 : group_by_has_partkey(RelOptInfo *input_rel,
8123 : List *targetList,
8124 : List *groupClause)
8125 : {
8126 556 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8127 556 : int cnt = 0;
8128 : int partnatts;
8129 :
8130 : /* Input relation should be partitioned. */
8131 : Assert(input_rel->part_scheme);
8132 :
8133 : /* Rule out early, if there are no partition keys present. */
8134 556 : if (!input_rel->partexprs)
8135 0 : return false;
8136 :
8137 556 : partnatts = input_rel->part_scheme->partnatts;
8138 :
8139 912 : for (cnt = 0; cnt < partnatts; cnt++)
8140 : {
8141 592 : List *partexprs = input_rel->partexprs[cnt];
8142 : ListCell *lc;
8143 592 : bool found = false;
8144 :
8145 810 : foreach(lc, partexprs)
8146 : {
8147 : ListCell *lg;
8148 586 : Expr *partexpr = lfirst(lc);
8149 586 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8150 :
8151 924 : foreach(lg, groupexprs)
8152 : {
8153 706 : Expr *groupexpr = lfirst(lg);
8154 706 : Oid groupcoll = exprCollation((Node *) groupexpr);
8155 :
8156 : /*
8157 : * Note: we can assume there is at most one RelabelType node;
8158 : * eval_const_expressions() will have simplified if more than
8159 : * one.
8160 : */
8161 706 : if (IsA(groupexpr, RelabelType))
8162 24 : groupexpr = ((RelabelType *) groupexpr)->arg;
8163 :
8164 706 : if (equal(groupexpr, partexpr))
8165 : {
8166 : /*
8167 : * Reject a match if the grouping collation does not match
8168 : * the partitioning collation.
8169 : */
8170 368 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8171 : partcoll != groupcoll)
8172 12 : return false;
8173 :
8174 356 : found = true;
8175 356 : break;
8176 : }
8177 : }
8178 :
8179 574 : if (found)
8180 356 : break;
8181 : }
8182 :
8183 : /*
8184 : * If none of the partition key expressions match with any of the
8185 : * GROUP BY expression, return false.
8186 : */
8187 580 : if (!found)
8188 224 : return false;
8189 : }
8190 :
8191 320 : return true;
8192 : }
8193 :
8194 : /*
8195 : * generate_setop_child_grouplist
8196 : * Build a SortGroupClause list defining the sort/grouping properties
8197 : * of the child of a set operation.
8198 : *
8199 : * This is similar to generate_setop_grouplist() but differs as the setop
8200 : * child query's targetlist entries may already have a tleSortGroupRef
8201 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8202 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8203 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8204 : * any of the columns in the targetlist don't match to the setop's colTypes
8205 : * then we return an empty list. This may leave some TLEs with unreferenced
8206 : * ressortgroupref markings, but that's harmless.
8207 : */
8208 : static List *
8209 12278 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8210 : {
8211 12278 : List *grouplist = copyObject(op->groupClauses);
8212 : ListCell *lg;
8213 : ListCell *lt;
8214 : ListCell *ct;
8215 :
8216 12278 : lg = list_head(grouplist);
8217 12278 : ct = list_head(op->colTypes);
8218 47818 : foreach(lt, targetlist)
8219 : {
8220 35962 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8221 : SortGroupClause *sgc;
8222 : Oid coltype;
8223 :
8224 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8225 35962 : if (tle->resjunk)
8226 0 : continue;
8227 :
8228 : /*
8229 : * We expect every non-resjunk target to have a SortGroupClause and
8230 : * colTypes.
8231 : */
8232 : Assert(lg != NULL);
8233 : Assert(ct != NULL);
8234 35962 : sgc = (SortGroupClause *) lfirst(lg);
8235 35962 : coltype = lfirst_oid(ct);
8236 :
8237 : /* reject if target type isn't the same as the setop target type */
8238 35962 : if (coltype != exprType((Node *) tle->expr))
8239 422 : return NIL;
8240 :
8241 35540 : lg = lnext(grouplist, lg);
8242 35540 : ct = lnext(op->colTypes, ct);
8243 :
8244 : /* assign a tleSortGroupRef, or reuse the existing one */
8245 35540 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8246 : }
8247 :
8248 : Assert(lg == NULL);
8249 : Assert(ct == NULL);
8250 :
8251 11856 : return grouplist;
8252 : }
|
