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 461018 : planner(Query *parse, const char *query_string, int cursorOptions,
287 : ParamListInfo boundParams)
288 : {
289 : PlannedStmt *result;
290 :
291 461018 : if (planner_hook)
292 93572 : result = (*planner_hook) (parse, query_string, cursorOptions, boundParams);
293 : else
294 367446 : result = standard_planner(parse, query_string, cursorOptions, boundParams);
295 :
296 456762 : pgstat_report_plan_id(result->planId, false);
297 :
298 456762 : return result;
299 : }
300 :
301 : PlannedStmt *
302 461018 : 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 461018 : glob = makeNode(PlannerGlobal);
322 :
323 461018 : glob->boundParams = boundParams;
324 461018 : glob->subplans = NIL;
325 461018 : glob->subpaths = NIL;
326 461018 : glob->subroots = NIL;
327 461018 : glob->rewindPlanIDs = NULL;
328 461018 : glob->finalrtable = NIL;
329 461018 : glob->finalrteperminfos = NIL;
330 461018 : glob->finalrowmarks = NIL;
331 461018 : glob->resultRelations = NIL;
332 461018 : glob->appendRelations = NIL;
333 461018 : glob->relationOids = NIL;
334 461018 : glob->invalItems = NIL;
335 461018 : glob->paramExecTypes = NIL;
336 461018 : glob->lastPHId = 0;
337 461018 : glob->lastRowMarkId = 0;
338 461018 : glob->lastPlanNodeId = 0;
339 461018 : glob->transientPlan = false;
340 461018 : glob->dependsOnRole = false;
341 :
342 : /*
343 : * Assess whether it's feasible to use parallel mode for this query. We
344 : * can't do this in a standalone backend, or if the command will try to
345 : * modify any data, or if this is a cursor operation, or if GUCs are set
346 : * to values that don't permit parallelism, or if parallel-unsafe
347 : * functions are present in the query tree.
348 : *
349 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
350 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
351 : * the command is writing into a completely new table which workers won't
352 : * be able to see. If the workers could see the table, the fact that
353 : * group locking would cause them to ignore the leader's heavyweight GIN
354 : * page locks would make this unsafe. We'll have to fix that somehow if
355 : * we want to allow parallel inserts in general; updates and deletes have
356 : * additional problems especially around combo CIDs.)
357 : *
358 : * For now, we don't try to use parallel mode if we're running inside a
359 : * parallel worker. We might eventually be able to relax this
360 : * restriction, but for now it seems best not to have parallel workers
361 : * trying to create their own parallel workers.
362 : */
363 461018 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
364 433248 : IsUnderPostmaster &&
365 433248 : parse->commandType == CMD_SELECT &&
366 347336 : !parse->hasModifyingCTE &&
367 347196 : max_parallel_workers_per_gather > 0 &&
368 346598 : !IsParallelWorker())
369 : {
370 : /* all the cheap tests pass, so scan the query tree */
371 346550 : glob->maxParallelHazard = max_parallel_hazard(parse);
372 346550 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
373 : }
374 : else
375 : {
376 : /* skip the query tree scan, just assume it's unsafe */
377 114468 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
378 114468 : glob->parallelModeOK = false;
379 : }
380 :
381 : /*
382 : * glob->parallelModeNeeded is normally set to false here and changed to
383 : * true during plan creation if a Gather or Gather Merge plan is actually
384 : * created (cf. create_gather_plan, create_gather_merge_plan).
385 : *
386 : * However, if debug_parallel_query = on or debug_parallel_query =
387 : * regress, then we impose parallel mode whenever it's safe to do so, even
388 : * if the final plan doesn't use parallelism. It's not safe to do so if
389 : * the query contains anything parallel-unsafe; parallelModeOK will be
390 : * false in that case. Note that parallelModeOK can't change after this
391 : * point. Otherwise, everything in the query is either parallel-safe or
392 : * parallel-restricted, and in either case it should be OK to impose
393 : * parallel-mode restrictions. If that ends up breaking something, then
394 : * either some function the user included in the query is incorrectly
395 : * labeled as parallel-safe or parallel-restricted when in reality it's
396 : * parallel-unsafe, or else the query planner itself has a bug.
397 : */
398 756002 : glob->parallelModeNeeded = glob->parallelModeOK &&
399 294984 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
400 :
401 : /* Determine what fraction of the plan is likely to be scanned */
402 461018 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
403 : {
404 : /*
405 : * We have no real idea how many tuples the user will ultimately FETCH
406 : * from a cursor, but it is often the case that he doesn't want 'em
407 : * all, or would prefer a fast-start plan anyway so that he can
408 : * process some of the tuples sooner. Use a GUC parameter to decide
409 : * what fraction to optimize for.
410 : */
411 4652 : tuple_fraction = cursor_tuple_fraction;
412 :
413 : /*
414 : * We document cursor_tuple_fraction as simply being a fraction, which
415 : * means the edge cases 0 and 1 have to be treated specially here. We
416 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
417 : */
418 4652 : if (tuple_fraction >= 1.0)
419 0 : tuple_fraction = 0.0;
420 4652 : else if (tuple_fraction <= 0.0)
421 0 : tuple_fraction = 1e-10;
422 : }
423 : else
424 : {
425 : /* Default assumption is we need all the tuples */
426 456366 : tuple_fraction = 0.0;
427 : }
428 :
429 : /* primary planning entry point (may recurse for subqueries) */
430 461018 : root = subquery_planner(glob, parse, NULL, false, tuple_fraction, NULL);
431 :
432 : /* Select best Path and turn it into a Plan */
433 457158 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
434 457158 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
435 :
436 457158 : top_plan = create_plan(root, best_path);
437 :
438 : /*
439 : * If creating a plan for a scrollable cursor, make sure it can run
440 : * backwards on demand. Add a Material node at the top at need.
441 : */
442 456762 : if (cursorOptions & CURSOR_OPT_SCROLL)
443 : {
444 266 : if (!ExecSupportsBackwardScan(top_plan))
445 32 : top_plan = materialize_finished_plan(top_plan);
446 : }
447 :
448 : /*
449 : * Optionally add a Gather node for testing purposes, provided this is
450 : * actually a safe thing to do.
451 : *
452 : * We can add Gather even when top_plan has parallel-safe initPlans, but
453 : * then we have to move the initPlans to the Gather node because of
454 : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
455 : * regression tests when debug_parallel_query = regress, because initPlans
456 : * that would normally appear on the top_plan move to the Gather, causing
457 : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
458 : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
459 : */
460 456762 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
461 194 : top_plan->parallel_safe &&
462 128 : (top_plan->initPlan == NIL ||
463 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
464 : {
465 128 : Gather *gather = makeNode(Gather);
466 : Cost initplan_cost;
467 : bool unsafe_initplans;
468 :
469 128 : gather->plan.targetlist = top_plan->targetlist;
470 128 : gather->plan.qual = NIL;
471 128 : gather->plan.lefttree = top_plan;
472 128 : gather->plan.righttree = NULL;
473 128 : gather->num_workers = 1;
474 128 : gather->single_copy = true;
475 128 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
476 :
477 : /* Transfer any initPlans to the new top node */
478 128 : gather->plan.initPlan = top_plan->initPlan;
479 128 : top_plan->initPlan = NIL;
480 :
481 : /*
482 : * Since this Gather has no parallel-aware descendants to signal to,
483 : * we don't need a rescan Param.
484 : */
485 128 : gather->rescan_param = -1;
486 :
487 : /*
488 : * Ideally we'd use cost_gather here, but setting up dummy path data
489 : * to satisfy it doesn't seem much cleaner than knowing what it does.
490 : */
491 128 : gather->plan.startup_cost = top_plan->startup_cost +
492 : parallel_setup_cost;
493 128 : gather->plan.total_cost = top_plan->total_cost +
494 128 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
495 128 : gather->plan.plan_rows = top_plan->plan_rows;
496 128 : gather->plan.plan_width = top_plan->plan_width;
497 128 : gather->plan.parallel_aware = false;
498 128 : gather->plan.parallel_safe = false;
499 :
500 : /*
501 : * Delete the initplans' cost from top_plan. We needn't add it to the
502 : * Gather node, since the above coding already included it there.
503 : */
504 128 : SS_compute_initplan_cost(gather->plan.initPlan,
505 : &initplan_cost, &unsafe_initplans);
506 128 : top_plan->startup_cost -= initplan_cost;
507 128 : top_plan->total_cost -= initplan_cost;
508 :
509 : /* use parallel mode for parallel plans. */
510 128 : root->glob->parallelModeNeeded = true;
511 :
512 128 : top_plan = &gather->plan;
513 : }
514 :
515 : /*
516 : * If any Params were generated, run through the plan tree and compute
517 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
518 : * set_plan_references' tree traversal, but for now it has to be separate
519 : * because we need to visit subplans before not after main plan.
520 : */
521 456762 : if (glob->paramExecTypes != NIL)
522 : {
523 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
524 203760 : forboth(lp, glob->subplans, lr, glob->subroots)
525 : {
526 45710 : Plan *subplan = (Plan *) lfirst(lp);
527 45710 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
528 :
529 45710 : SS_finalize_plan(subroot, subplan);
530 : }
531 158050 : SS_finalize_plan(root, top_plan);
532 : }
533 :
534 : /* final cleanup of the plan */
535 : Assert(glob->finalrtable == NIL);
536 : Assert(glob->finalrteperminfos == NIL);
537 : Assert(glob->finalrowmarks == NIL);
538 : Assert(glob->resultRelations == NIL);
539 : Assert(glob->appendRelations == NIL);
540 456762 : top_plan = set_plan_references(root, top_plan);
541 : /* ... and the subplans (both regular subplans and initplans) */
542 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
543 502472 : forboth(lp, glob->subplans, lr, glob->subroots)
544 : {
545 45710 : Plan *subplan = (Plan *) lfirst(lp);
546 45710 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
547 :
548 45710 : lfirst(lp) = set_plan_references(subroot, subplan);
549 : }
550 :
551 : /* build the PlannedStmt result */
552 456762 : result = makeNode(PlannedStmt);
553 :
554 456762 : result->commandType = parse->commandType;
555 456762 : result->queryId = parse->queryId;
556 456762 : result->hasReturning = (parse->returningList != NIL);
557 456762 : result->hasModifyingCTE = parse->hasModifyingCTE;
558 456762 : result->canSetTag = parse->canSetTag;
559 456762 : result->transientPlan = glob->transientPlan;
560 456762 : result->dependsOnRole = glob->dependsOnRole;
561 456762 : result->parallelModeNeeded = glob->parallelModeNeeded;
562 456762 : result->planTree = top_plan;
563 456762 : result->partPruneInfos = glob->partPruneInfos;
564 456762 : result->rtable = glob->finalrtable;
565 913524 : result->unprunableRelids = bms_difference(glob->allRelids,
566 456762 : glob->prunableRelids);
567 456762 : result->permInfos = glob->finalrteperminfos;
568 456762 : result->resultRelations = glob->resultRelations;
569 456762 : result->firstResultRels = glob->firstResultRels;
570 456762 : result->appendRelations = glob->appendRelations;
571 456762 : result->subplans = glob->subplans;
572 456762 : result->rewindPlanIDs = glob->rewindPlanIDs;
573 456762 : result->rowMarks = glob->finalrowmarks;
574 456762 : result->relationOids = glob->relationOids;
575 456762 : result->invalItems = glob->invalItems;
576 456762 : result->paramExecTypes = glob->paramExecTypes;
577 : /* utilityStmt should be null, but we might as well copy it */
578 456762 : result->utilityStmt = parse->utilityStmt;
579 456762 : result->stmt_location = parse->stmt_location;
580 456762 : result->stmt_len = parse->stmt_len;
581 :
582 456762 : result->jitFlags = PGJIT_NONE;
583 456762 : if (jit_enabled && jit_above_cost >= 0 &&
584 456088 : top_plan->total_cost > jit_above_cost)
585 : {
586 938 : result->jitFlags |= PGJIT_PERFORM;
587 :
588 : /*
589 : * Decide how much effort should be put into generating better code.
590 : */
591 938 : if (jit_optimize_above_cost >= 0 &&
592 938 : top_plan->total_cost > jit_optimize_above_cost)
593 432 : result->jitFlags |= PGJIT_OPT3;
594 938 : if (jit_inline_above_cost >= 0 &&
595 938 : top_plan->total_cost > jit_inline_above_cost)
596 432 : result->jitFlags |= PGJIT_INLINE;
597 :
598 : /*
599 : * Decide which operations should be JITed.
600 : */
601 938 : if (jit_expressions)
602 938 : result->jitFlags |= PGJIT_EXPR;
603 938 : if (jit_tuple_deforming)
604 938 : result->jitFlags |= PGJIT_DEFORM;
605 : }
606 :
607 456762 : if (glob->partition_directory != NULL)
608 11440 : DestroyPartitionDirectory(glob->partition_directory);
609 :
610 456762 : return result;
611 : }
612 :
613 :
614 : /*--------------------
615 : * subquery_planner
616 : * Invokes the planner on a subquery. We recurse to here for each
617 : * sub-SELECT found in the query tree.
618 : *
619 : * glob is the global state for the current planner run.
620 : * parse is the querytree produced by the parser & rewriter.
621 : * parent_root is the immediate parent Query's info (NULL at the top level).
622 : * hasRecursion is true if this is a recursive WITH query.
623 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
624 : * tuple_fraction is interpreted as explained for grouping_planner, below.
625 : * setops is used for set operation subqueries to provide the subquery with
626 : * the context in which it's being used so that Paths correctly sorted for the
627 : * set operation can be generated. NULL when not planning a set operation
628 : * child, or when a child of a set op that isn't interested in sorted input.
629 : *
630 : * Basically, this routine does the stuff that should only be done once
631 : * per Query object. It then calls grouping_planner. At one time,
632 : * grouping_planner could be invoked recursively on the same Query object;
633 : * that's not currently true, but we keep the separation between the two
634 : * routines anyway, in case we need it again someday.
635 : *
636 : * subquery_planner will be called recursively to handle sub-Query nodes
637 : * found within the query's expressions and rangetable.
638 : *
639 : * Returns the PlannerInfo struct ("root") that contains all data generated
640 : * while planning the subquery. In particular, the Path(s) attached to
641 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
642 : * cheapest way(s) to implement the query. The top level will select the
643 : * best Path and pass it through createplan.c to produce a finished Plan.
644 : *--------------------
645 : */
646 : PlannerInfo *
647 534390 : subquery_planner(PlannerGlobal *glob, Query *parse, PlannerInfo *parent_root,
648 : bool hasRecursion, double tuple_fraction,
649 : SetOperationStmt *setops)
650 : {
651 : PlannerInfo *root;
652 : List *newWithCheckOptions;
653 : List *newHaving;
654 : bool hasOuterJoins;
655 : bool hasResultRTEs;
656 : RelOptInfo *final_rel;
657 : ListCell *l;
658 :
659 : /* Create a PlannerInfo data structure for this subquery */
660 534390 : root = makeNode(PlannerInfo);
661 534390 : root->parse = parse;
662 534390 : root->glob = glob;
663 534390 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
664 534390 : root->parent_root = parent_root;
665 534390 : root->plan_params = NIL;
666 534390 : root->outer_params = NULL;
667 534390 : root->planner_cxt = CurrentMemoryContext;
668 534390 : root->init_plans = NIL;
669 534390 : root->cte_plan_ids = NIL;
670 534390 : root->multiexpr_params = NIL;
671 534390 : root->join_domains = NIL;
672 534390 : root->eq_classes = NIL;
673 534390 : root->ec_merging_done = false;
674 534390 : root->last_rinfo_serial = 0;
675 534390 : root->all_result_relids =
676 534390 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
677 534390 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
678 534390 : root->append_rel_list = NIL;
679 534390 : root->row_identity_vars = NIL;
680 534390 : root->rowMarks = NIL;
681 534390 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
682 534390 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
683 534390 : root->processed_groupClause = NIL;
684 534390 : root->processed_distinctClause = NIL;
685 534390 : root->processed_tlist = NIL;
686 534390 : root->update_colnos = NIL;
687 534390 : root->grouping_map = NULL;
688 534390 : root->minmax_aggs = NIL;
689 534390 : root->qual_security_level = 0;
690 534390 : root->hasPseudoConstantQuals = false;
691 534390 : root->hasAlternativeSubPlans = false;
692 534390 : root->placeholdersFrozen = false;
693 534390 : root->hasRecursion = hasRecursion;
694 534390 : if (hasRecursion)
695 1010 : root->wt_param_id = assign_special_exec_param(root);
696 : else
697 533380 : root->wt_param_id = -1;
698 534390 : root->non_recursive_path = NULL;
699 534390 : root->partColsUpdated = false;
700 :
701 : /*
702 : * Create the top-level join domain. This won't have valid contents until
703 : * deconstruct_jointree fills it in, but the node needs to exist before
704 : * that so we can build EquivalenceClasses referencing it.
705 : */
706 534390 : root->join_domains = list_make1(makeNode(JoinDomain));
707 :
708 : /*
709 : * If there is a WITH list, process each WITH query and either convert it
710 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
711 : */
712 534390 : if (parse->cteList)
713 2882 : SS_process_ctes(root);
714 :
715 : /*
716 : * If it's a MERGE command, transform the joinlist as appropriate.
717 : */
718 534384 : transform_MERGE_to_join(parse);
719 :
720 : /*
721 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
722 : * that we don't need so many special cases to deal with that situation.
723 : */
724 534384 : replace_empty_jointree(parse);
725 :
726 : /*
727 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
728 : * to transform them into joins. Note that this step does not descend
729 : * into subqueries; if we pull up any subqueries below, their SubLinks are
730 : * processed just before pulling them up.
731 : */
732 534384 : if (parse->hasSubLinks)
733 37814 : pull_up_sublinks(root);
734 :
735 : /*
736 : * Scan the rangetable for function RTEs, do const-simplification on them,
737 : * and then inline them if possible (producing subqueries that might get
738 : * pulled up next). Recursion issues here are handled in the same way as
739 : * for SubLinks.
740 : */
741 534384 : preprocess_function_rtes(root);
742 :
743 : /*
744 : * Scan the rangetable for relations with virtual generated columns, and
745 : * replace all Var nodes in the query that reference these columns with
746 : * the generation expressions. Recursion issues here are handled in the
747 : * same way as for SubLinks.
748 : */
749 534378 : parse = root->parse = expand_virtual_generated_columns(root);
750 :
751 : /*
752 : * Check to see if any subqueries in the jointree can be merged into this
753 : * query.
754 : */
755 534378 : pull_up_subqueries(root);
756 :
757 : /*
758 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
759 : * do this now because it requires applying pull_up_subqueries to the leaf
760 : * queries of the UNION ALL, which weren't touched above because they
761 : * weren't referenced by the jointree (they will be after we do this).
762 : */
763 534372 : if (parse->setOperations)
764 6602 : flatten_simple_union_all(root);
765 :
766 : /*
767 : * Survey the rangetable to see what kinds of entries are present. We can
768 : * skip some later processing if relevant SQL features are not used; for
769 : * example if there are no JOIN RTEs we can avoid the expense of doing
770 : * flatten_join_alias_vars(). This must be done after we have finished
771 : * adding rangetable entries, of course. (Note: actually, processing of
772 : * inherited or partitioned rels can cause RTEs for their child tables to
773 : * get added later; but those must all be RTE_RELATION entries, so they
774 : * don't invalidate the conclusions drawn here.)
775 : */
776 534372 : root->hasJoinRTEs = false;
777 534372 : root->hasLateralRTEs = false;
778 534372 : root->group_rtindex = 0;
779 534372 : hasOuterJoins = false;
780 534372 : hasResultRTEs = false;
781 1446580 : foreach(l, parse->rtable)
782 : {
783 912208 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
784 :
785 912208 : switch (rte->rtekind)
786 : {
787 474750 : case RTE_RELATION:
788 474750 : if (rte->inh)
789 : {
790 : /*
791 : * Check to see if the relation actually has any children;
792 : * if not, clear the inh flag so we can treat it as a
793 : * plain base relation.
794 : *
795 : * Note: this could give a false-positive result, if the
796 : * rel once had children but no longer does. We used to
797 : * be able to clear rte->inh later on when we discovered
798 : * that, but no more; we have to handle such cases as
799 : * full-fledged inheritance.
800 : */
801 388496 : rte->inh = has_subclass(rte->relid);
802 : }
803 474750 : break;
804 89036 : case RTE_JOIN:
805 89036 : root->hasJoinRTEs = true;
806 89036 : if (IS_OUTER_JOIN(rte->jointype))
807 48974 : hasOuterJoins = true;
808 89036 : break;
809 205826 : case RTE_RESULT:
810 205826 : hasResultRTEs = true;
811 205826 : break;
812 4466 : case RTE_GROUP:
813 : Assert(parse->hasGroupRTE);
814 4466 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
815 4466 : break;
816 138130 : default:
817 : /* No work here for other RTE types */
818 138130 : break;
819 : }
820 :
821 912208 : if (rte->lateral)
822 10548 : root->hasLateralRTEs = true;
823 :
824 : /*
825 : * We can also determine the maximum security level required for any
826 : * securityQuals now. Addition of inheritance-child RTEs won't affect
827 : * this, because child tables don't have their own securityQuals; see
828 : * expand_single_inheritance_child().
829 : */
830 912208 : if (rte->securityQuals)
831 2484 : root->qual_security_level = Max(root->qual_security_level,
832 : list_length(rte->securityQuals));
833 : }
834 :
835 : /*
836 : * If we have now verified that the query target relation is
837 : * non-inheriting, mark it as a leaf target.
838 : */
839 534372 : if (parse->resultRelation)
840 : {
841 92654 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
842 :
843 92654 : if (!rte->inh)
844 89856 : root->leaf_result_relids =
845 89856 : bms_make_singleton(parse->resultRelation);
846 : }
847 :
848 : /*
849 : * Preprocess RowMark information. We need to do this after subquery
850 : * pullup, so that all base relations are present.
851 : */
852 534372 : preprocess_rowmarks(root);
853 :
854 : /*
855 : * Set hasHavingQual to remember if HAVING clause is present. Needed
856 : * because preprocess_expression will reduce a constant-true condition to
857 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
858 : */
859 534372 : root->hasHavingQual = (parse->havingQual != NULL);
860 :
861 : /*
862 : * Do expression preprocessing on targetlist and quals, as well as other
863 : * random expressions in the querytree. Note that we do not need to
864 : * handle sort/group expressions explicitly, because they are actually
865 : * part of the targetlist.
866 : */
867 530596 : parse->targetList = (List *)
868 534372 : preprocess_expression(root, (Node *) parse->targetList,
869 : EXPRKIND_TARGET);
870 :
871 530596 : newWithCheckOptions = NIL;
872 533012 : foreach(l, parse->withCheckOptions)
873 : {
874 2416 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
875 :
876 2416 : wco->qual = preprocess_expression(root, wco->qual,
877 : EXPRKIND_QUAL);
878 2416 : if (wco->qual != NULL)
879 2016 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
880 : }
881 530596 : parse->withCheckOptions = newWithCheckOptions;
882 :
883 530596 : parse->returningList = (List *)
884 530596 : preprocess_expression(root, (Node *) parse->returningList,
885 : EXPRKIND_TARGET);
886 :
887 530596 : preprocess_qual_conditions(root, (Node *) parse->jointree);
888 :
889 530596 : parse->havingQual = preprocess_expression(root, parse->havingQual,
890 : EXPRKIND_QUAL);
891 :
892 533214 : foreach(l, parse->windowClause)
893 : {
894 2618 : WindowClause *wc = lfirst_node(WindowClause, l);
895 :
896 : /* partitionClause/orderClause are sort/group expressions */
897 2618 : wc->startOffset = preprocess_expression(root, wc->startOffset,
898 : EXPRKIND_LIMIT);
899 2618 : wc->endOffset = preprocess_expression(root, wc->endOffset,
900 : EXPRKIND_LIMIT);
901 : }
902 :
903 530596 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
904 : EXPRKIND_LIMIT);
905 530596 : parse->limitCount = preprocess_expression(root, parse->limitCount,
906 : EXPRKIND_LIMIT);
907 :
908 530596 : if (parse->onConflict)
909 : {
910 3628 : parse->onConflict->arbiterElems = (List *)
911 1814 : preprocess_expression(root,
912 1814 : (Node *) parse->onConflict->arbiterElems,
913 : EXPRKIND_ARBITER_ELEM);
914 3628 : parse->onConflict->arbiterWhere =
915 1814 : preprocess_expression(root,
916 1814 : parse->onConflict->arbiterWhere,
917 : EXPRKIND_QUAL);
918 3628 : parse->onConflict->onConflictSet = (List *)
919 1814 : preprocess_expression(root,
920 1814 : (Node *) parse->onConflict->onConflictSet,
921 : EXPRKIND_TARGET);
922 1814 : parse->onConflict->onConflictWhere =
923 1814 : preprocess_expression(root,
924 1814 : parse->onConflict->onConflictWhere,
925 : EXPRKIND_QUAL);
926 : /* exclRelTlist contains only Vars, so no preprocessing needed */
927 : }
928 :
929 533410 : foreach(l, parse->mergeActionList)
930 : {
931 2814 : MergeAction *action = (MergeAction *) lfirst(l);
932 :
933 2814 : action->targetList = (List *)
934 2814 : preprocess_expression(root,
935 2814 : (Node *) action->targetList,
936 : EXPRKIND_TARGET);
937 2814 : action->qual =
938 2814 : preprocess_expression(root,
939 : (Node *) action->qual,
940 : EXPRKIND_QUAL);
941 : }
942 :
943 530596 : parse->mergeJoinCondition =
944 530596 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
945 :
946 530596 : root->append_rel_list = (List *)
947 530596 : preprocess_expression(root, (Node *) root->append_rel_list,
948 : EXPRKIND_APPINFO);
949 :
950 : /* Also need to preprocess expressions within RTEs */
951 1438750 : foreach(l, parse->rtable)
952 : {
953 908154 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
954 : int kind;
955 : ListCell *lcsq;
956 :
957 908154 : if (rte->rtekind == RTE_RELATION)
958 : {
959 474484 : if (rte->tablesample)
960 228 : rte->tablesample = (TableSampleClause *)
961 228 : preprocess_expression(root,
962 228 : (Node *) rte->tablesample,
963 : EXPRKIND_TABLESAMPLE);
964 : }
965 433670 : else if (rte->rtekind == RTE_SUBQUERY)
966 : {
967 : /*
968 : * We don't want to do all preprocessing yet on the subquery's
969 : * expressions, since that will happen when we plan it. But if it
970 : * contains any join aliases of our level, those have to get
971 : * expanded now, because planning of the subquery won't do it.
972 : * That's only possible if the subquery is LATERAL.
973 : */
974 68554 : if (rte->lateral && root->hasJoinRTEs)
975 1228 : rte->subquery = (Query *)
976 1228 : flatten_join_alias_vars(root, root->parse,
977 1228 : (Node *) rte->subquery);
978 : }
979 365116 : else if (rte->rtekind == RTE_FUNCTION)
980 : {
981 : /* Preprocess the function expression(s) fully */
982 55114 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
983 55114 : rte->functions = (List *)
984 55114 : preprocess_expression(root, (Node *) rte->functions, kind);
985 : }
986 310002 : else if (rte->rtekind == RTE_TABLEFUNC)
987 : {
988 : /* Preprocess the function expression(s) fully */
989 626 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
990 626 : rte->tablefunc = (TableFunc *)
991 626 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
992 : }
993 309376 : else if (rte->rtekind == RTE_VALUES)
994 : {
995 : /* Preprocess the values lists fully */
996 8244 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
997 8244 : rte->values_lists = (List *)
998 8244 : preprocess_expression(root, (Node *) rte->values_lists, kind);
999 : }
1000 301132 : else if (rte->rtekind == RTE_GROUP)
1001 : {
1002 : /* Preprocess the groupexprs list fully */
1003 4466 : rte->groupexprs = (List *)
1004 4466 : preprocess_expression(root, (Node *) rte->groupexprs,
1005 : EXPRKIND_GROUPEXPR);
1006 : }
1007 :
1008 : /*
1009 : * Process each element of the securityQuals list as if it were a
1010 : * separate qual expression (as indeed it is). We need to do it this
1011 : * way to get proper canonicalization of AND/OR structure. Note that
1012 : * this converts each element into an implicit-AND sublist.
1013 : */
1014 910996 : foreach(lcsq, rte->securityQuals)
1015 : {
1016 2842 : lfirst(lcsq) = preprocess_expression(root,
1017 2842 : (Node *) lfirst(lcsq),
1018 : EXPRKIND_QUAL);
1019 : }
1020 : }
1021 :
1022 : /*
1023 : * Now that we are done preprocessing expressions, and in particular done
1024 : * flattening join alias variables, get rid of the joinaliasvars lists.
1025 : * They no longer match what expressions in the rest of the tree look
1026 : * like, because we have not preprocessed expressions in those lists (and
1027 : * do not want to; for example, expanding a SubLink there would result in
1028 : * a useless unreferenced subplan). Leaving them in place simply creates
1029 : * a hazard for later scans of the tree. We could try to prevent that by
1030 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1031 : * but that doesn't sound very reliable.
1032 : */
1033 530596 : if (root->hasJoinRTEs)
1034 : {
1035 307460 : foreach(l, parse->rtable)
1036 : {
1037 253554 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1038 :
1039 253554 : rte->joinaliasvars = NIL;
1040 : }
1041 : }
1042 :
1043 : /*
1044 : * Replace any Vars in the subquery's targetlist and havingQual that
1045 : * reference GROUP outputs with the underlying grouping expressions.
1046 : *
1047 : * Note that we need to perform this replacement after we've preprocessed
1048 : * the grouping expressions. This is to ensure that there is only one
1049 : * instance of SubPlan for each SubLink contained within the grouping
1050 : * expressions.
1051 : */
1052 530596 : if (parse->hasGroupRTE)
1053 : {
1054 4466 : parse->targetList = (List *)
1055 4466 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1056 4466 : parse->havingQual =
1057 4466 : flatten_group_exprs(root, root->parse, parse->havingQual);
1058 : }
1059 :
1060 : /* Constant-folding might have removed all set-returning functions */
1061 530596 : if (parse->hasTargetSRFs)
1062 12024 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1063 :
1064 : /*
1065 : * In some cases we may want to transfer a HAVING clause into WHERE. We
1066 : * cannot do so if the HAVING clause contains aggregates (obviously) or
1067 : * volatile functions (since a HAVING clause is supposed to be executed
1068 : * only once per group). We also can't do this if there are any nonempty
1069 : * grouping sets and the clause references any columns that are nullable
1070 : * by the grouping sets; moving such a clause into WHERE would potentially
1071 : * change the results. (If there are only empty grouping sets, then the
1072 : * HAVING clause must be degenerate as discussed below.)
1073 : *
1074 : * Also, it may be that the clause is so expensive to execute that we're
1075 : * better off doing it only once per group, despite the loss of
1076 : * selectivity. This is hard to estimate short of doing the entire
1077 : * planning process twice, so we use a heuristic: clauses containing
1078 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1079 : * clause into WHERE, in hopes of eliminating tuples before aggregation
1080 : * instead of after.
1081 : *
1082 : * If the query has explicit grouping then we can simply move such a
1083 : * clause into WHERE; any group that fails the clause will not be in the
1084 : * output because none of its tuples will reach the grouping or
1085 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
1086 : * HAVING clause, which we put in WHERE so that query_planner() can use it
1087 : * in a gating Result node, but also keep in HAVING to ensure that we
1088 : * don't emit a bogus aggregated row. (This could be done better, but it
1089 : * seems not worth optimizing.)
1090 : *
1091 : * Note that a HAVING clause may contain expressions that are not fully
1092 : * preprocessed. This can happen if these expressions are part of
1093 : * grouping items. In such cases, they are replaced with GROUP Vars in
1094 : * the parser and then replaced back after we've done with expression
1095 : * preprocessing on havingQual. This is not an issue if the clause
1096 : * remains in HAVING, because these expressions will be matched to lower
1097 : * target items in setrefs.c. However, if the clause is moved or copied
1098 : * into WHERE, we need to ensure that these expressions are fully
1099 : * preprocessed.
1100 : *
1101 : * Note that both havingQual and parse->jointree->quals are in
1102 : * implicitly-ANDed-list form at this point, even though they are declared
1103 : * as Node *.
1104 : */
1105 530596 : newHaving = NIL;
1106 531750 : foreach(l, (List *) parse->havingQual)
1107 : {
1108 1154 : Node *havingclause = (Node *) lfirst(l);
1109 :
1110 1488 : if (contain_agg_clause(havingclause) ||
1111 668 : contain_volatile_functions(havingclause) ||
1112 334 : contain_subplans(havingclause) ||
1113 418 : (parse->groupClause && parse->groupingSets &&
1114 84 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1115 : {
1116 : /* keep it in HAVING */
1117 892 : newHaving = lappend(newHaving, havingclause);
1118 : }
1119 262 : else if (parse->groupClause)
1120 : {
1121 : Node *whereclause;
1122 :
1123 : /* Preprocess the HAVING clause fully */
1124 244 : whereclause = preprocess_expression(root, havingclause,
1125 : EXPRKIND_QUAL);
1126 : /* ... and move it to WHERE */
1127 244 : parse->jointree->quals = (Node *)
1128 244 : list_concat((List *) parse->jointree->quals,
1129 : (List *) whereclause);
1130 : }
1131 : else
1132 : {
1133 : Node *whereclause;
1134 :
1135 : /* Preprocess the HAVING clause fully */
1136 18 : whereclause = preprocess_expression(root, copyObject(havingclause),
1137 : EXPRKIND_QUAL);
1138 : /* ... and put a copy in WHERE */
1139 36 : parse->jointree->quals = (Node *)
1140 18 : list_concat((List *) parse->jointree->quals,
1141 : (List *) whereclause);
1142 : /* ... and also keep it in HAVING */
1143 18 : newHaving = lappend(newHaving, havingclause);
1144 : }
1145 : }
1146 530596 : parse->havingQual = (Node *) newHaving;
1147 :
1148 : /*
1149 : * If we have any outer joins, try to reduce them to plain inner joins.
1150 : * This step is most easily done after we've done expression
1151 : * preprocessing.
1152 : */
1153 530596 : if (hasOuterJoins)
1154 34676 : reduce_outer_joins(root);
1155 :
1156 : /*
1157 : * If we have any RTE_RESULT relations, see if they can be deleted from
1158 : * the jointree. We also rely on this processing to flatten single-child
1159 : * FromExprs underneath outer joins. This step is most effectively done
1160 : * after we've done expression preprocessing and outer join reduction.
1161 : */
1162 530596 : if (hasResultRTEs || hasOuterJoins)
1163 235426 : remove_useless_result_rtes(root);
1164 :
1165 : /*
1166 : * Do the main planning.
1167 : */
1168 530596 : grouping_planner(root, tuple_fraction, setops);
1169 :
1170 : /*
1171 : * Capture the set of outer-level param IDs we have access to, for use in
1172 : * extParam/allParam calculations later.
1173 : */
1174 530524 : SS_identify_outer_params(root);
1175 :
1176 : /*
1177 : * If any initPlans were created in this query level, adjust the surviving
1178 : * Paths' costs and parallel-safety flags to account for them. The
1179 : * initPlans won't actually get attached to the plan tree till
1180 : * create_plan() runs, but we must include their effects now.
1181 : */
1182 530524 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1183 530524 : SS_charge_for_initplans(root, final_rel);
1184 :
1185 : /*
1186 : * Make sure we've identified the cheapest Path for the final rel. (By
1187 : * doing this here not in grouping_planner, we include initPlan costs in
1188 : * the decision, though it's unlikely that will change anything.)
1189 : */
1190 530524 : set_cheapest(final_rel);
1191 :
1192 530524 : return root;
1193 : }
1194 :
1195 : /*
1196 : * preprocess_expression
1197 : * Do subquery_planner's preprocessing work for an expression,
1198 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1199 : * conditions), a HAVING clause, or a few other things.
1200 : */
1201 : static Node *
1202 4452414 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1203 : {
1204 : /*
1205 : * Fall out quickly if expression is empty. This occurs often enough to
1206 : * be worth checking. Note that null->null is the correct conversion for
1207 : * implicit-AND result format, too.
1208 : */
1209 4452414 : if (expr == NULL)
1210 3520206 : return NULL;
1211 :
1212 : /*
1213 : * If the query has any join RTEs, replace join alias variables with
1214 : * base-relation variables. We must do this first, since any expressions
1215 : * we may extract from the joinaliasvars lists have not been preprocessed.
1216 : * For example, if we did this after sublink processing, sublinks expanded
1217 : * out from join aliases would not get processed. But we can skip this in
1218 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1219 : * they can't contain any Vars of the current query level.
1220 : */
1221 932208 : if (root->hasJoinRTEs &&
1222 388808 : !(kind == EXPRKIND_RTFUNC ||
1223 194230 : kind == EXPRKIND_VALUES ||
1224 : kind == EXPRKIND_TABLESAMPLE ||
1225 : kind == EXPRKIND_TABLEFUNC))
1226 194212 : expr = flatten_join_alias_vars(root, root->parse, expr);
1227 :
1228 : /*
1229 : * Simplify constant expressions. For function RTEs, this was already
1230 : * done by preprocess_function_rtes. (But note we must do it again for
1231 : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1232 : * un-simplified subexpressions inserted by flattening of subqueries or
1233 : * join alias variables.)
1234 : *
1235 : * Note: an essential effect of this is to convert named-argument function
1236 : * calls to positional notation and insert the current actual values of
1237 : * any default arguments for functions. To ensure that happens, we *must*
1238 : * process all expressions here. Previous PG versions sometimes skipped
1239 : * const-simplification if it didn't seem worth the trouble, but we can't
1240 : * do that anymore.
1241 : *
1242 : * Note: this also flattens nested AND and OR expressions into N-argument
1243 : * form. All processing of a qual expression after this point must be
1244 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1245 : * with AND directly under AND, nor OR directly under OR.
1246 : */
1247 932208 : if (kind != EXPRKIND_RTFUNC)
1248 885502 : expr = eval_const_expressions(root, expr);
1249 :
1250 : /*
1251 : * If it's a qual or havingQual, canonicalize it.
1252 : */
1253 928432 : if (kind == EXPRKIND_QUAL)
1254 : {
1255 328934 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1256 :
1257 : #ifdef OPTIMIZER_DEBUG
1258 : printf("After canonicalize_qual()\n");
1259 : pprint(expr);
1260 : #endif
1261 : }
1262 :
1263 : /*
1264 : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1265 : * hashfuncid of any that might execute more quickly by using hash lookups
1266 : * instead of a linear search.
1267 : */
1268 928432 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1269 : {
1270 846666 : convert_saop_to_hashed_saop(expr);
1271 : }
1272 :
1273 : /* Expand SubLinks to SubPlans */
1274 928432 : if (root->parse->hasSubLinks)
1275 108386 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1276 :
1277 : /*
1278 : * XXX do not insert anything here unless you have grokked the comments in
1279 : * SS_replace_correlation_vars ...
1280 : */
1281 :
1282 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1283 928432 : if (root->query_level > 1)
1284 161578 : expr = SS_replace_correlation_vars(root, expr);
1285 :
1286 : /*
1287 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1288 : * don't want to do this before eval_const_expressions, since the latter
1289 : * would be unable to simplify a top-level AND correctly. Also,
1290 : * SS_process_sublinks expects explicit-AND format.)
1291 : */
1292 928432 : if (kind == EXPRKIND_QUAL)
1293 328934 : expr = (Node *) make_ands_implicit((Expr *) expr);
1294 :
1295 928432 : return expr;
1296 : }
1297 :
1298 : /*
1299 : * preprocess_qual_conditions
1300 : * Recursively scan the query's jointree and do subquery_planner's
1301 : * preprocessing work on each qual condition found therein.
1302 : */
1303 : static void
1304 1313608 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1305 : {
1306 1313608 : if (jtnode == NULL)
1307 0 : return;
1308 1313608 : if (IsA(jtnode, RangeTblRef))
1309 : {
1310 : /* nothing to do here */
1311 : }
1312 642076 : else if (IsA(jtnode, FromExpr))
1313 : {
1314 545272 : FromExpr *f = (FromExpr *) jtnode;
1315 : ListCell *l;
1316 :
1317 1134676 : foreach(l, f->fromlist)
1318 589404 : preprocess_qual_conditions(root, lfirst(l));
1319 :
1320 545272 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1321 : }
1322 96804 : else if (IsA(jtnode, JoinExpr))
1323 : {
1324 96804 : JoinExpr *j = (JoinExpr *) jtnode;
1325 :
1326 96804 : preprocess_qual_conditions(root, j->larg);
1327 96804 : preprocess_qual_conditions(root, j->rarg);
1328 :
1329 96804 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1330 : }
1331 : else
1332 0 : elog(ERROR, "unrecognized node type: %d",
1333 : (int) nodeTag(jtnode));
1334 : }
1335 :
1336 : /*
1337 : * preprocess_phv_expression
1338 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1339 : *
1340 : * If a LATERAL subquery references an output of another subquery, and that
1341 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1342 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1343 : * and later pull it back up during find_lateral_references, which runs after
1344 : * subquery_planner has preprocessed all the expressions that were in the
1345 : * current query level to start with. So we need to preprocess it then.
1346 : */
1347 : Expr *
1348 72 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1349 : {
1350 72 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1351 : }
1352 :
1353 : /*--------------------
1354 : * grouping_planner
1355 : * Perform planning steps related to grouping, aggregation, etc.
1356 : *
1357 : * This function adds all required top-level processing to the scan/join
1358 : * Path(s) produced by query_planner.
1359 : *
1360 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1361 : * tuple_fraction is interpreted as follows:
1362 : * 0: expect all tuples to be retrieved (normal case)
1363 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1364 : * from the plan to be retrieved
1365 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1366 : * expected to be retrieved (ie, a LIMIT specification).
1367 : * setops is used for set operation subqueries to provide the subquery with
1368 : * the context in which it's being used so that Paths correctly sorted for the
1369 : * set operation can be generated. NULL when not planning a set operation
1370 : * child, or when a child of a set op that isn't interested in sorted input.
1371 : *
1372 : * Returns nothing; the useful output is in the Paths we attach to the
1373 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1374 : * root->processed_tlist contains the final processed targetlist.
1375 : *
1376 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1377 : * to leave this to the caller.
1378 : *--------------------
1379 : */
1380 : static void
1381 530596 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1382 : SetOperationStmt *setops)
1383 : {
1384 530596 : Query *parse = root->parse;
1385 530596 : int64 offset_est = 0;
1386 530596 : int64 count_est = 0;
1387 530596 : double limit_tuples = -1.0;
1388 530596 : bool have_postponed_srfs = false;
1389 : PathTarget *final_target;
1390 : List *final_targets;
1391 : List *final_targets_contain_srfs;
1392 : bool final_target_parallel_safe;
1393 : RelOptInfo *current_rel;
1394 : RelOptInfo *final_rel;
1395 : FinalPathExtraData extra;
1396 : ListCell *lc;
1397 :
1398 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1399 530596 : if (parse->limitCount || parse->limitOffset)
1400 : {
1401 4950 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1402 : &offset_est, &count_est);
1403 :
1404 : /*
1405 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1406 : * estimate the effects of using a bounded sort.
1407 : */
1408 4950 : if (count_est > 0 && offset_est >= 0)
1409 4428 : limit_tuples = (double) count_est + (double) offset_est;
1410 : }
1411 :
1412 : /* Make tuple_fraction accessible to lower-level routines */
1413 530596 : root->tuple_fraction = tuple_fraction;
1414 :
1415 530596 : if (parse->setOperations)
1416 : {
1417 : /*
1418 : * Construct Paths for set operations. The results will not need any
1419 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1420 : * special work for recursive unions is the responsibility of
1421 : * plan_set_operations.
1422 : */
1423 6166 : current_rel = plan_set_operations(root);
1424 :
1425 : /*
1426 : * We should not need to call preprocess_targetlist, since we must be
1427 : * in a SELECT query node. Instead, use the processed_tlist returned
1428 : * by plan_set_operations (since this tells whether it returned any
1429 : * resjunk columns!), and transfer any sort key information from the
1430 : * original tlist.
1431 : */
1432 : Assert(parse->commandType == CMD_SELECT);
1433 :
1434 : /* for safety, copy processed_tlist instead of modifying in-place */
1435 6160 : root->processed_tlist =
1436 6160 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1437 : parse->targetList);
1438 :
1439 : /* Also extract the PathTarget form of the setop result tlist */
1440 6160 : final_target = current_rel->cheapest_total_path->pathtarget;
1441 :
1442 : /* And check whether it's parallel safe */
1443 : final_target_parallel_safe =
1444 6160 : is_parallel_safe(root, (Node *) final_target->exprs);
1445 :
1446 : /* The setop result tlist couldn't contain any SRFs */
1447 : Assert(!parse->hasTargetSRFs);
1448 6160 : final_targets = final_targets_contain_srfs = NIL;
1449 :
1450 : /*
1451 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1452 : * checked already, but let's make sure).
1453 : */
1454 6160 : if (parse->rowMarks)
1455 0 : ereport(ERROR,
1456 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1457 : /*------
1458 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1459 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1460 : LCS_asString(linitial_node(RowMarkClause,
1461 : parse->rowMarks)->strength))));
1462 :
1463 : /*
1464 : * Calculate pathkeys that represent result ordering requirements
1465 : */
1466 : Assert(parse->distinctClause == NIL);
1467 6160 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1468 : parse->sortClause,
1469 : root->processed_tlist);
1470 : }
1471 : else
1472 : {
1473 : /* No set operations, do regular planning */
1474 : PathTarget *sort_input_target;
1475 : List *sort_input_targets;
1476 : List *sort_input_targets_contain_srfs;
1477 : bool sort_input_target_parallel_safe;
1478 : PathTarget *grouping_target;
1479 : List *grouping_targets;
1480 : List *grouping_targets_contain_srfs;
1481 : bool grouping_target_parallel_safe;
1482 : PathTarget *scanjoin_target;
1483 : List *scanjoin_targets;
1484 : List *scanjoin_targets_contain_srfs;
1485 : bool scanjoin_target_parallel_safe;
1486 : bool scanjoin_target_same_exprs;
1487 : bool have_grouping;
1488 524430 : WindowFuncLists *wflists = NULL;
1489 524430 : List *activeWindows = NIL;
1490 524430 : grouping_sets_data *gset_data = NULL;
1491 : standard_qp_extra qp_extra;
1492 :
1493 : /* A recursive query should always have setOperations */
1494 : Assert(!root->hasRecursion);
1495 :
1496 : /* Preprocess grouping sets and GROUP BY clause, if any */
1497 524430 : if (parse->groupingSets)
1498 : {
1499 878 : gset_data = preprocess_grouping_sets(root);
1500 : }
1501 523552 : else if (parse->groupClause)
1502 : {
1503 : /* Preprocess regular GROUP BY clause, if any */
1504 3630 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1505 : }
1506 :
1507 : /*
1508 : * Preprocess targetlist. Note that much of the remaining planning
1509 : * work will be done with the PathTarget representation of tlists, but
1510 : * we must also maintain the full representation of the final tlist so
1511 : * that we can transfer its decoration (resnames etc) to the topmost
1512 : * tlist of the finished Plan. This is kept in processed_tlist.
1513 : */
1514 524424 : preprocess_targetlist(root);
1515 :
1516 : /*
1517 : * Mark all the aggregates with resolved aggtranstypes, and detect
1518 : * aggregates that are duplicates or can share transition state. We
1519 : * must do this before slicing and dicing the tlist into various
1520 : * pathtargets, else some copies of the Aggref nodes might escape
1521 : * being marked.
1522 : */
1523 524424 : if (parse->hasAggs)
1524 : {
1525 38350 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1526 38350 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1527 : }
1528 :
1529 : /*
1530 : * Locate any window functions in the tlist. (We don't need to look
1531 : * anywhere else, since expressions used in ORDER BY will be in there
1532 : * too.) Note that they could all have been eliminated by constant
1533 : * folding, in which case we don't need to do any more work.
1534 : */
1535 524424 : if (parse->hasWindowFuncs)
1536 : {
1537 2384 : wflists = find_window_functions((Node *) root->processed_tlist,
1538 2384 : list_length(parse->windowClause));
1539 2384 : if (wflists->numWindowFuncs > 0)
1540 : {
1541 : /*
1542 : * See if any modifications can be made to each WindowClause
1543 : * to allow the executor to execute the WindowFuncs more
1544 : * quickly.
1545 : */
1546 2378 : optimize_window_clauses(root, wflists);
1547 :
1548 : /* Extract the list of windows actually in use. */
1549 2378 : activeWindows = select_active_windows(root, wflists);
1550 :
1551 : /* Make sure they all have names, for EXPLAIN's use. */
1552 2378 : name_active_windows(activeWindows);
1553 : }
1554 : else
1555 6 : parse->hasWindowFuncs = false;
1556 : }
1557 :
1558 : /*
1559 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1560 : * adding logic between here and the query_planner() call. Anything
1561 : * that is needed in MIN/MAX-optimizable cases will have to be
1562 : * duplicated in planagg.c.
1563 : */
1564 524424 : if (parse->hasAggs)
1565 38350 : preprocess_minmax_aggregates(root);
1566 :
1567 : /*
1568 : * Figure out whether there's a hard limit on the number of rows that
1569 : * query_planner's result subplan needs to return. Even if we know a
1570 : * hard limit overall, it doesn't apply if the query has any
1571 : * grouping/aggregation operations, or SRFs in the tlist.
1572 : */
1573 524424 : if (parse->groupClause ||
1574 519964 : parse->groupingSets ||
1575 519922 : parse->distinctClause ||
1576 517282 : parse->hasAggs ||
1577 482822 : parse->hasWindowFuncs ||
1578 480582 : parse->hasTargetSRFs ||
1579 469020 : root->hasHavingQual)
1580 55422 : root->limit_tuples = -1.0;
1581 : else
1582 469002 : root->limit_tuples = limit_tuples;
1583 :
1584 : /* Set up data needed by standard_qp_callback */
1585 524424 : qp_extra.activeWindows = activeWindows;
1586 524424 : qp_extra.gset_data = gset_data;
1587 :
1588 : /*
1589 : * If we're a subquery for a set operation, store the SetOperationStmt
1590 : * in qp_extra.
1591 : */
1592 524424 : qp_extra.setop = setops;
1593 :
1594 : /*
1595 : * Generate the best unsorted and presorted paths for the scan/join
1596 : * portion of this Query, ie the processing represented by the
1597 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1598 : * We also generate (in standard_qp_callback) pathkey representations
1599 : * of the query's sort clause, distinct clause, etc.
1600 : */
1601 524424 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1602 :
1603 : /*
1604 : * Convert the query's result tlist into PathTarget format.
1605 : *
1606 : * Note: this cannot be done before query_planner() has performed
1607 : * appendrel expansion, because that might add resjunk entries to
1608 : * root->processed_tlist. Waiting till afterwards is also helpful
1609 : * because the target width estimates can use per-Var width numbers
1610 : * that were obtained within query_planner().
1611 : */
1612 524370 : final_target = create_pathtarget(root, root->processed_tlist);
1613 : final_target_parallel_safe =
1614 524370 : is_parallel_safe(root, (Node *) final_target->exprs);
1615 :
1616 : /*
1617 : * If ORDER BY was given, consider whether we should use a post-sort
1618 : * projection, and compute the adjusted target for preceding steps if
1619 : * so.
1620 : */
1621 524370 : if (parse->sortClause)
1622 : {
1623 71882 : sort_input_target = make_sort_input_target(root,
1624 : final_target,
1625 : &have_postponed_srfs);
1626 : sort_input_target_parallel_safe =
1627 71882 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1628 : }
1629 : else
1630 : {
1631 452488 : sort_input_target = final_target;
1632 452488 : sort_input_target_parallel_safe = final_target_parallel_safe;
1633 : }
1634 :
1635 : /*
1636 : * If we have window functions to deal with, the output from any
1637 : * grouping step needs to be what the window functions want;
1638 : * otherwise, it should be sort_input_target.
1639 : */
1640 524370 : if (activeWindows)
1641 : {
1642 2378 : grouping_target = make_window_input_target(root,
1643 : final_target,
1644 : activeWindows);
1645 : grouping_target_parallel_safe =
1646 2378 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1647 : }
1648 : else
1649 : {
1650 521992 : grouping_target = sort_input_target;
1651 521992 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1652 : }
1653 :
1654 : /*
1655 : * If we have grouping or aggregation to do, the topmost scan/join
1656 : * plan node must emit what the grouping step wants; otherwise, it
1657 : * should emit grouping_target.
1658 : */
1659 519910 : have_grouping = (parse->groupClause || parse->groupingSets ||
1660 1044280 : parse->hasAggs || root->hasHavingQual);
1661 524370 : if (have_grouping)
1662 : {
1663 39018 : scanjoin_target = make_group_input_target(root, final_target);
1664 : scanjoin_target_parallel_safe =
1665 39018 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1666 : }
1667 : else
1668 : {
1669 485352 : scanjoin_target = grouping_target;
1670 485352 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1671 : }
1672 :
1673 : /*
1674 : * If there are any SRFs in the targetlist, we must separate each of
1675 : * these PathTargets into SRF-computing and SRF-free targets. Replace
1676 : * each of the named targets with a SRF-free version, and remember the
1677 : * list of additional projection steps we need to add afterwards.
1678 : */
1679 524370 : if (parse->hasTargetSRFs)
1680 : {
1681 : /* final_target doesn't recompute any SRFs in sort_input_target */
1682 12024 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1683 : &final_targets,
1684 : &final_targets_contain_srfs);
1685 12024 : final_target = linitial_node(PathTarget, final_targets);
1686 : Assert(!linitial_int(final_targets_contain_srfs));
1687 : /* likewise for sort_input_target vs. grouping_target */
1688 12024 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1689 : &sort_input_targets,
1690 : &sort_input_targets_contain_srfs);
1691 12024 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
1692 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1693 : /* likewise for grouping_target vs. scanjoin_target */
1694 12024 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
1695 : &grouping_targets,
1696 : &grouping_targets_contain_srfs);
1697 12024 : grouping_target = linitial_node(PathTarget, grouping_targets);
1698 : Assert(!linitial_int(grouping_targets_contain_srfs));
1699 : /* scanjoin_target will not have any SRFs precomputed for it */
1700 12024 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1701 : &scanjoin_targets,
1702 : &scanjoin_targets_contain_srfs);
1703 12024 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
1704 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1705 : }
1706 : else
1707 : {
1708 : /* initialize lists; for most of these, dummy values are OK */
1709 512346 : final_targets = final_targets_contain_srfs = NIL;
1710 512346 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1711 512346 : grouping_targets = grouping_targets_contain_srfs = NIL;
1712 512346 : scanjoin_targets = list_make1(scanjoin_target);
1713 512346 : scanjoin_targets_contain_srfs = NIL;
1714 : }
1715 :
1716 : /* Apply scan/join target. */
1717 524370 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1718 524370 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1719 524370 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1720 : scanjoin_targets_contain_srfs,
1721 : scanjoin_target_parallel_safe,
1722 : scanjoin_target_same_exprs);
1723 :
1724 : /*
1725 : * Save the various upper-rel PathTargets we just computed into
1726 : * root->upper_targets[]. The core code doesn't use this, but it
1727 : * provides a convenient place for extensions to get at the info. For
1728 : * consistency, we save all the intermediate targets, even though some
1729 : * of the corresponding upperrels might not be needed for this query.
1730 : */
1731 524370 : root->upper_targets[UPPERREL_FINAL] = final_target;
1732 524370 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1733 524370 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
1734 524370 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
1735 524370 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1736 524370 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1737 :
1738 : /*
1739 : * If we have grouping and/or aggregation, consider ways to implement
1740 : * that. We build a new upperrel representing the output of this
1741 : * phase.
1742 : */
1743 524370 : if (have_grouping)
1744 : {
1745 39018 : current_rel = create_grouping_paths(root,
1746 : current_rel,
1747 : grouping_target,
1748 : grouping_target_parallel_safe,
1749 : gset_data);
1750 : /* Fix things up if grouping_target contains SRFs */
1751 39012 : if (parse->hasTargetSRFs)
1752 420 : adjust_paths_for_srfs(root, current_rel,
1753 : grouping_targets,
1754 : grouping_targets_contain_srfs);
1755 : }
1756 :
1757 : /*
1758 : * If we have window functions, consider ways to implement those. We
1759 : * build a new upperrel representing the output of this phase.
1760 : */
1761 524364 : if (activeWindows)
1762 : {
1763 2378 : current_rel = create_window_paths(root,
1764 : current_rel,
1765 : grouping_target,
1766 : sort_input_target,
1767 : sort_input_target_parallel_safe,
1768 : wflists,
1769 : activeWindows);
1770 : /* Fix things up if sort_input_target contains SRFs */
1771 2378 : if (parse->hasTargetSRFs)
1772 12 : adjust_paths_for_srfs(root, current_rel,
1773 : sort_input_targets,
1774 : sort_input_targets_contain_srfs);
1775 : }
1776 :
1777 : /*
1778 : * If there is a DISTINCT clause, consider ways to implement that. We
1779 : * build a new upperrel representing the output of this phase.
1780 : */
1781 524364 : if (parse->distinctClause)
1782 : {
1783 2674 : current_rel = create_distinct_paths(root,
1784 : current_rel,
1785 : sort_input_target);
1786 : }
1787 : } /* end of if (setOperations) */
1788 :
1789 : /*
1790 : * If ORDER BY was given, consider ways to implement that, and generate a
1791 : * new upperrel containing only paths that emit the correct ordering and
1792 : * project the correct final_target. We can apply the original
1793 : * limit_tuples limit in sort costing here, but only if there are no
1794 : * postponed SRFs.
1795 : */
1796 530524 : if (parse->sortClause)
1797 : {
1798 75838 : current_rel = create_ordered_paths(root,
1799 : current_rel,
1800 : final_target,
1801 : final_target_parallel_safe,
1802 : have_postponed_srfs ? -1.0 :
1803 : limit_tuples);
1804 : /* Fix things up if final_target contains SRFs */
1805 75838 : if (parse->hasTargetSRFs)
1806 196 : adjust_paths_for_srfs(root, current_rel,
1807 : final_targets,
1808 : final_targets_contain_srfs);
1809 : }
1810 :
1811 : /*
1812 : * Now we are prepared to build the final-output upperrel.
1813 : */
1814 530524 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1815 :
1816 : /*
1817 : * If the input rel is marked consider_parallel and there's nothing that's
1818 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1819 : * consider_parallel as well. Note that if the query has rowMarks or is
1820 : * not a SELECT, consider_parallel will be false for every relation in the
1821 : * query.
1822 : */
1823 705272 : if (current_rel->consider_parallel &&
1824 349472 : is_parallel_safe(root, parse->limitOffset) &&
1825 174724 : is_parallel_safe(root, parse->limitCount))
1826 174718 : final_rel->consider_parallel = true;
1827 :
1828 : /*
1829 : * If the current_rel belongs to a single FDW, so does the final_rel.
1830 : */
1831 530524 : final_rel->serverid = current_rel->serverid;
1832 530524 : final_rel->userid = current_rel->userid;
1833 530524 : final_rel->useridiscurrent = current_rel->useridiscurrent;
1834 530524 : final_rel->fdwroutine = current_rel->fdwroutine;
1835 :
1836 : /*
1837 : * Generate paths for the final_rel. Insert all surviving paths, with
1838 : * LockRows, Limit, and/or ModifyTable steps added if needed.
1839 : */
1840 1080696 : foreach(lc, current_rel->pathlist)
1841 : {
1842 550172 : Path *path = (Path *) lfirst(lc);
1843 :
1844 : /*
1845 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1846 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
1847 : * here. If there are only non-locking rowmarks, they should be
1848 : * handled by the ModifyTable node instead. However, root->rowMarks
1849 : * is what goes into the LockRows node.)
1850 : */
1851 550172 : if (parse->rowMarks)
1852 : {
1853 8030 : path = (Path *) create_lockrows_path(root, final_rel, path,
1854 : root->rowMarks,
1855 : assign_special_exec_param(root));
1856 : }
1857 :
1858 : /*
1859 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1860 : */
1861 550172 : if (limit_needed(parse))
1862 : {
1863 5882 : path = (Path *) create_limit_path(root, final_rel, path,
1864 : parse->limitOffset,
1865 : parse->limitCount,
1866 : parse->limitOption,
1867 : offset_est, count_est);
1868 : }
1869 :
1870 : /*
1871 : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
1872 : */
1873 550172 : if (parse->commandType != CMD_SELECT)
1874 : {
1875 : Index rootRelation;
1876 92406 : List *resultRelations = NIL;
1877 92406 : List *updateColnosLists = NIL;
1878 92406 : List *withCheckOptionLists = NIL;
1879 92406 : List *returningLists = NIL;
1880 92406 : List *mergeActionLists = NIL;
1881 92406 : List *mergeJoinConditions = NIL;
1882 : List *rowMarks;
1883 :
1884 92406 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
1885 : {
1886 : /* Inherited UPDATE/DELETE/MERGE */
1887 2768 : RelOptInfo *top_result_rel = find_base_rel(root,
1888 : parse->resultRelation);
1889 2768 : int resultRelation = -1;
1890 :
1891 : /* Pass the root result rel forward to the executor. */
1892 2768 : rootRelation = parse->resultRelation;
1893 :
1894 : /* Add only leaf children to ModifyTable. */
1895 8044 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
1896 : resultRelation)) >= 0)
1897 : {
1898 5276 : RelOptInfo *this_result_rel = find_base_rel(root,
1899 : resultRelation);
1900 :
1901 : /*
1902 : * Also exclude any leaf rels that have turned dummy since
1903 : * being added to the list, for example, by being excluded
1904 : * by constraint exclusion.
1905 : */
1906 5276 : if (IS_DUMMY_REL(this_result_rel))
1907 84 : continue;
1908 :
1909 : /* Build per-target-rel lists needed by ModifyTable */
1910 5192 : resultRelations = lappend_int(resultRelations,
1911 : resultRelation);
1912 5192 : if (parse->commandType == CMD_UPDATE)
1913 : {
1914 3624 : List *update_colnos = root->update_colnos;
1915 :
1916 3624 : if (this_result_rel != top_result_rel)
1917 : update_colnos =
1918 3624 : adjust_inherited_attnums_multilevel(root,
1919 : update_colnos,
1920 : this_result_rel->relid,
1921 : top_result_rel->relid);
1922 3624 : updateColnosLists = lappend(updateColnosLists,
1923 : update_colnos);
1924 : }
1925 5192 : if (parse->withCheckOptions)
1926 : {
1927 486 : List *withCheckOptions = parse->withCheckOptions;
1928 :
1929 486 : if (this_result_rel != top_result_rel)
1930 : withCheckOptions = (List *)
1931 486 : adjust_appendrel_attrs_multilevel(root,
1932 : (Node *) withCheckOptions,
1933 : this_result_rel,
1934 : top_result_rel);
1935 486 : withCheckOptionLists = lappend(withCheckOptionLists,
1936 : withCheckOptions);
1937 : }
1938 5192 : if (parse->returningList)
1939 : {
1940 816 : List *returningList = parse->returningList;
1941 :
1942 816 : if (this_result_rel != top_result_rel)
1943 : returningList = (List *)
1944 816 : adjust_appendrel_attrs_multilevel(root,
1945 : (Node *) returningList,
1946 : this_result_rel,
1947 : top_result_rel);
1948 816 : returningLists = lappend(returningLists,
1949 : returningList);
1950 : }
1951 5192 : if (parse->mergeActionList)
1952 : {
1953 : ListCell *l;
1954 492 : List *mergeActionList = NIL;
1955 :
1956 : /*
1957 : * Copy MergeActions and translate stuff that
1958 : * references attribute numbers.
1959 : */
1960 1554 : foreach(l, parse->mergeActionList)
1961 : {
1962 1062 : MergeAction *action = lfirst(l),
1963 1062 : *leaf_action = copyObject(action);
1964 :
1965 1062 : leaf_action->qual =
1966 1062 : adjust_appendrel_attrs_multilevel(root,
1967 : (Node *) action->qual,
1968 : this_result_rel,
1969 : top_result_rel);
1970 1062 : leaf_action->targetList = (List *)
1971 1062 : adjust_appendrel_attrs_multilevel(root,
1972 1062 : (Node *) action->targetList,
1973 : this_result_rel,
1974 : top_result_rel);
1975 1062 : if (leaf_action->commandType == CMD_UPDATE)
1976 592 : leaf_action->updateColnos =
1977 592 : adjust_inherited_attnums_multilevel(root,
1978 : action->updateColnos,
1979 : this_result_rel->relid,
1980 : top_result_rel->relid);
1981 1062 : mergeActionList = lappend(mergeActionList,
1982 : leaf_action);
1983 : }
1984 :
1985 492 : mergeActionLists = lappend(mergeActionLists,
1986 : mergeActionList);
1987 : }
1988 5192 : if (parse->commandType == CMD_MERGE)
1989 : {
1990 492 : Node *mergeJoinCondition = parse->mergeJoinCondition;
1991 :
1992 492 : if (this_result_rel != top_result_rel)
1993 : mergeJoinCondition =
1994 492 : adjust_appendrel_attrs_multilevel(root,
1995 : mergeJoinCondition,
1996 : this_result_rel,
1997 : top_result_rel);
1998 492 : mergeJoinConditions = lappend(mergeJoinConditions,
1999 : mergeJoinCondition);
2000 : }
2001 : }
2002 :
2003 2768 : if (resultRelations == NIL)
2004 : {
2005 : /*
2006 : * We managed to exclude every child rel, so generate a
2007 : * dummy one-relation plan using info for the top target
2008 : * rel (even though that may not be a leaf target).
2009 : * Although it's clear that no data will be updated or
2010 : * deleted, we still need to have a ModifyTable node so
2011 : * that any statement triggers will be executed. (This
2012 : * could be cleaner if we fixed nodeModifyTable.c to allow
2013 : * zero target relations, but that probably wouldn't be a
2014 : * net win.)
2015 : */
2016 30 : resultRelations = list_make1_int(parse->resultRelation);
2017 30 : if (parse->commandType == CMD_UPDATE)
2018 30 : updateColnosLists = list_make1(root->update_colnos);
2019 30 : if (parse->withCheckOptions)
2020 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2021 30 : if (parse->returningList)
2022 18 : returningLists = list_make1(parse->returningList);
2023 30 : if (parse->mergeActionList)
2024 0 : mergeActionLists = list_make1(parse->mergeActionList);
2025 30 : if (parse->commandType == CMD_MERGE)
2026 0 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2027 : }
2028 : }
2029 : else
2030 : {
2031 : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2032 89638 : rootRelation = 0; /* there's no separate root rel */
2033 89638 : resultRelations = list_make1_int(parse->resultRelation);
2034 89638 : if (parse->commandType == CMD_UPDATE)
2035 11882 : updateColnosLists = list_make1(root->update_colnos);
2036 89638 : if (parse->withCheckOptions)
2037 926 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2038 89638 : if (parse->returningList)
2039 2420 : returningLists = list_make1(parse->returningList);
2040 89638 : if (parse->mergeActionList)
2041 1632 : mergeActionLists = list_make1(parse->mergeActionList);
2042 89638 : if (parse->commandType == CMD_MERGE)
2043 1632 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2044 : }
2045 :
2046 : /*
2047 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2048 : * will have dealt with fetching non-locked marked rows, else we
2049 : * need to have ModifyTable do that.
2050 : */
2051 92406 : if (parse->rowMarks)
2052 0 : rowMarks = NIL;
2053 : else
2054 92406 : rowMarks = root->rowMarks;
2055 :
2056 : path = (Path *)
2057 92406 : create_modifytable_path(root, final_rel,
2058 : path,
2059 : parse->commandType,
2060 92406 : parse->canSetTag,
2061 92406 : parse->resultRelation,
2062 : rootRelation,
2063 92406 : root->partColsUpdated,
2064 : resultRelations,
2065 : updateColnosLists,
2066 : withCheckOptionLists,
2067 : returningLists,
2068 : rowMarks,
2069 : parse->onConflict,
2070 : mergeActionLists,
2071 : mergeJoinConditions,
2072 : assign_special_exec_param(root));
2073 : }
2074 :
2075 : /* And shove it into final_rel */
2076 550172 : add_path(final_rel, path);
2077 : }
2078 :
2079 : /*
2080 : * Generate partial paths for final_rel, too, if outer query levels might
2081 : * be able to make use of them.
2082 : */
2083 530524 : if (final_rel->consider_parallel && root->query_level > 1 &&
2084 25488 : !limit_needed(parse))
2085 : {
2086 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2087 25428 : foreach(lc, current_rel->partial_pathlist)
2088 : {
2089 108 : Path *partial_path = (Path *) lfirst(lc);
2090 :
2091 108 : add_partial_path(final_rel, partial_path);
2092 : }
2093 : }
2094 :
2095 530524 : extra.limit_needed = limit_needed(parse);
2096 530524 : extra.limit_tuples = limit_tuples;
2097 530524 : extra.count_est = count_est;
2098 530524 : extra.offset_est = offset_est;
2099 :
2100 : /*
2101 : * If there is an FDW that's responsible for all baserels of the query,
2102 : * let it consider adding ForeignPaths.
2103 : */
2104 530524 : if (final_rel->fdwroutine &&
2105 1256 : final_rel->fdwroutine->GetForeignUpperPaths)
2106 1188 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2107 : current_rel, final_rel,
2108 : &extra);
2109 :
2110 : /* Let extensions possibly add some more paths */
2111 530524 : if (create_upper_paths_hook)
2112 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2113 : current_rel, final_rel, &extra);
2114 :
2115 : /* Note: currently, we leave it to callers to do set_cheapest() */
2116 530524 : }
2117 :
2118 : /*
2119 : * Do preprocessing for groupingSets clause and related data. This handles the
2120 : * preliminary steps of expanding the grouping sets, organizing them into lists
2121 : * of rollups, and preparing annotations which will later be filled in with
2122 : * size estimates.
2123 : */
2124 : static grouping_sets_data *
2125 878 : preprocess_grouping_sets(PlannerInfo *root)
2126 : {
2127 878 : Query *parse = root->parse;
2128 : List *sets;
2129 878 : int maxref = 0;
2130 : ListCell *lc_set;
2131 878 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2132 :
2133 878 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
2134 :
2135 878 : gd->any_hashable = false;
2136 878 : gd->unhashable_refs = NULL;
2137 878 : gd->unsortable_refs = NULL;
2138 878 : gd->unsortable_sets = NIL;
2139 :
2140 : /*
2141 : * We don't currently make any attempt to optimize the groupClause when
2142 : * there are grouping sets, so just duplicate it in processed_groupClause.
2143 : */
2144 878 : root->processed_groupClause = parse->groupClause;
2145 :
2146 878 : if (parse->groupClause)
2147 : {
2148 : ListCell *lc;
2149 :
2150 2672 : foreach(lc, parse->groupClause)
2151 : {
2152 1836 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2153 1836 : Index ref = gc->tleSortGroupRef;
2154 :
2155 1836 : if (ref > maxref)
2156 1800 : maxref = ref;
2157 :
2158 1836 : if (!gc->hashable)
2159 30 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2160 :
2161 1836 : if (!OidIsValid(gc->sortop))
2162 42 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2163 : }
2164 : }
2165 :
2166 : /* Allocate workspace array for remapping */
2167 878 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2168 :
2169 : /*
2170 : * If we have any unsortable sets, we must extract them before trying to
2171 : * prepare rollups. Unsortable sets don't go through
2172 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2173 : * here.
2174 : */
2175 878 : if (!bms_is_empty(gd->unsortable_refs))
2176 : {
2177 42 : List *sortable_sets = NIL;
2178 : ListCell *lc;
2179 :
2180 126 : foreach(lc, parse->groupingSets)
2181 : {
2182 90 : List *gset = (List *) lfirst(lc);
2183 :
2184 90 : if (bms_overlap_list(gd->unsortable_refs, gset))
2185 : {
2186 48 : GroupingSetData *gs = makeNode(GroupingSetData);
2187 :
2188 48 : gs->set = gset;
2189 48 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2190 :
2191 : /*
2192 : * We must enforce here that an unsortable set is hashable;
2193 : * later code assumes this. Parse analysis only checks that
2194 : * every individual column is either hashable or sortable.
2195 : *
2196 : * Note that passing this test doesn't guarantee we can
2197 : * generate a plan; there might be other showstoppers.
2198 : */
2199 48 : if (bms_overlap_list(gd->unhashable_refs, gset))
2200 6 : ereport(ERROR,
2201 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2202 : errmsg("could not implement GROUP BY"),
2203 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2204 : }
2205 : else
2206 42 : sortable_sets = lappend(sortable_sets, gset);
2207 : }
2208 :
2209 36 : if (sortable_sets)
2210 30 : sets = extract_rollup_sets(sortable_sets);
2211 : else
2212 6 : sets = NIL;
2213 : }
2214 : else
2215 836 : sets = extract_rollup_sets(parse->groupingSets);
2216 :
2217 2298 : foreach(lc_set, sets)
2218 : {
2219 1426 : List *current_sets = (List *) lfirst(lc_set);
2220 1426 : RollupData *rollup = makeNode(RollupData);
2221 : GroupingSetData *gs;
2222 :
2223 : /*
2224 : * Reorder the current list of grouping sets into correct prefix
2225 : * order. If only one aggregation pass is needed, try to make the
2226 : * list match the ORDER BY clause; if more than one pass is needed, we
2227 : * don't bother with that.
2228 : *
2229 : * Note that this reorders the sets from smallest-member-first to
2230 : * largest-member-first, and applies the GroupingSetData annotations,
2231 : * though the data will be filled in later.
2232 : */
2233 1426 : current_sets = reorder_grouping_sets(current_sets,
2234 1426 : (list_length(sets) == 1
2235 : ? parse->sortClause
2236 : : NIL));
2237 :
2238 : /*
2239 : * Get the initial (and therefore largest) grouping set.
2240 : */
2241 1426 : gs = linitial_node(GroupingSetData, current_sets);
2242 :
2243 : /*
2244 : * Order the groupClause appropriately. If the first grouping set is
2245 : * empty, then the groupClause must also be empty; otherwise we have
2246 : * to force the groupClause to match that grouping set's order.
2247 : *
2248 : * (The first grouping set can be empty even though parse->groupClause
2249 : * is not empty only if all non-empty grouping sets are unsortable.
2250 : * The groupClauses for hashed grouping sets are built later on.)
2251 : */
2252 1426 : if (gs->set)
2253 1384 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2254 : else
2255 42 : rollup->groupClause = NIL;
2256 :
2257 : /*
2258 : * Is it hashable? We pretend empty sets are hashable even though we
2259 : * actually force them not to be hashed later. But don't bother if
2260 : * there's nothing but empty sets (since in that case we can't hash
2261 : * anything).
2262 : */
2263 1426 : if (gs->set &&
2264 1384 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2265 : {
2266 1360 : rollup->hashable = true;
2267 1360 : gd->any_hashable = true;
2268 : }
2269 :
2270 : /*
2271 : * Now that we've pinned down an order for the groupClause for this
2272 : * list of grouping sets, we need to remap the entries in the grouping
2273 : * sets from sortgrouprefs to plain indices (0-based) into the
2274 : * groupClause for this collection of grouping sets. We keep the
2275 : * original form for later use, though.
2276 : */
2277 1426 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2278 : current_sets,
2279 : gd->tleref_to_colnum_map);
2280 1426 : rollup->gsets_data = current_sets;
2281 :
2282 1426 : gd->rollups = lappend(gd->rollups, rollup);
2283 : }
2284 :
2285 872 : if (gd->unsortable_sets)
2286 : {
2287 : /*
2288 : * We have not yet pinned down a groupclause for this, but we will
2289 : * need index-based lists for estimation purposes. Construct
2290 : * hash_sets_idx based on the entire original groupclause for now.
2291 : */
2292 36 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2293 : gd->unsortable_sets,
2294 : gd->tleref_to_colnum_map);
2295 36 : gd->any_hashable = true;
2296 : }
2297 :
2298 872 : return gd;
2299 : }
2300 :
2301 : /*
2302 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2303 : * (without annotation) mapped to indexes into the given groupclause.
2304 : */
2305 : static List *
2306 4164 : remap_to_groupclause_idx(List *groupClause,
2307 : List *gsets,
2308 : int *tleref_to_colnum_map)
2309 : {
2310 4164 : int ref = 0;
2311 4164 : List *result = NIL;
2312 : ListCell *lc;
2313 :
2314 10192 : foreach(lc, groupClause)
2315 : {
2316 6028 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2317 :
2318 6028 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2319 : }
2320 :
2321 9630 : foreach(lc, gsets)
2322 : {
2323 5466 : List *set = NIL;
2324 : ListCell *lc2;
2325 5466 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2326 :
2327 12332 : foreach(lc2, gs->set)
2328 : {
2329 6866 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2330 : }
2331 :
2332 5466 : result = lappend(result, set);
2333 : }
2334 :
2335 4164 : return result;
2336 : }
2337 :
2338 :
2339 : /*
2340 : * preprocess_rowmarks - set up PlanRowMarks if needed
2341 : */
2342 : static void
2343 534372 : preprocess_rowmarks(PlannerInfo *root)
2344 : {
2345 534372 : Query *parse = root->parse;
2346 : Bitmapset *rels;
2347 : List *prowmarks;
2348 : ListCell *l;
2349 : int i;
2350 :
2351 534372 : if (parse->rowMarks)
2352 : {
2353 : /*
2354 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2355 : * grouping, since grouping renders a reference to individual tuple
2356 : * CTIDs invalid. This is also checked at parse time, but that's
2357 : * insufficient because of rule substitution, query pullup, etc.
2358 : */
2359 7556 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2360 : parse->rowMarks)->strength);
2361 : }
2362 : else
2363 : {
2364 : /*
2365 : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2366 : * UPDATE/SHARE.
2367 : */
2368 526816 : if (parse->commandType != CMD_UPDATE &&
2369 512950 : parse->commandType != CMD_DELETE &&
2370 508668 : parse->commandType != CMD_MERGE)
2371 506826 : return;
2372 : }
2373 :
2374 : /*
2375 : * We need to have rowmarks for all base relations except the target. We
2376 : * make a bitmapset of all base rels and then remove the items we don't
2377 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2378 : */
2379 27546 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2380 27546 : if (parse->resultRelation)
2381 19990 : rels = bms_del_member(rels, parse->resultRelation);
2382 :
2383 : /*
2384 : * Convert RowMarkClauses to PlanRowMark representation.
2385 : */
2386 27546 : prowmarks = NIL;
2387 35364 : foreach(l, parse->rowMarks)
2388 : {
2389 7818 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2390 7818 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2391 : PlanRowMark *newrc;
2392 :
2393 : /*
2394 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2395 : * applied to an update/delete target rel. If that ever becomes
2396 : * possible, we should drop the target from the PlanRowMark list.
2397 : */
2398 : Assert(rc->rti != parse->resultRelation);
2399 :
2400 : /*
2401 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2402 : * can't support true locking. Subqueries that got flattened into the
2403 : * main query should be ignored completely. Any that didn't will get
2404 : * ROW_MARK_COPY items in the next loop.
2405 : */
2406 7818 : if (rte->rtekind != RTE_RELATION)
2407 108 : continue;
2408 :
2409 7710 : rels = bms_del_member(rels, rc->rti);
2410 :
2411 7710 : newrc = makeNode(PlanRowMark);
2412 7710 : newrc->rti = newrc->prti = rc->rti;
2413 7710 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2414 7710 : newrc->markType = select_rowmark_type(rte, rc->strength);
2415 7710 : newrc->allMarkTypes = (1 << newrc->markType);
2416 7710 : newrc->strength = rc->strength;
2417 7710 : newrc->waitPolicy = rc->waitPolicy;
2418 7710 : newrc->isParent = false;
2419 :
2420 7710 : prowmarks = lappend(prowmarks, newrc);
2421 : }
2422 :
2423 : /*
2424 : * Now, add rowmarks for any non-target, non-locked base relations.
2425 : */
2426 27546 : i = 0;
2427 66392 : foreach(l, parse->rtable)
2428 : {
2429 38846 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2430 : PlanRowMark *newrc;
2431 :
2432 38846 : i++;
2433 38846 : if (!bms_is_member(i, rels))
2434 35114 : continue;
2435 :
2436 3732 : newrc = makeNode(PlanRowMark);
2437 3732 : newrc->rti = newrc->prti = i;
2438 3732 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2439 3732 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2440 3732 : newrc->allMarkTypes = (1 << newrc->markType);
2441 3732 : newrc->strength = LCS_NONE;
2442 3732 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2443 3732 : newrc->isParent = false;
2444 :
2445 3732 : prowmarks = lappend(prowmarks, newrc);
2446 : }
2447 :
2448 27546 : root->rowMarks = prowmarks;
2449 : }
2450 :
2451 : /*
2452 : * Select RowMarkType to use for a given table
2453 : */
2454 : RowMarkType
2455 13654 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2456 : {
2457 13654 : if (rte->rtekind != RTE_RELATION)
2458 : {
2459 : /* If it's not a table at all, use ROW_MARK_COPY */
2460 1482 : return ROW_MARK_COPY;
2461 : }
2462 12172 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2463 : {
2464 : /* Let the FDW select the rowmark type, if it wants to */
2465 200 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2466 :
2467 200 : if (fdwroutine->GetForeignRowMarkType != NULL)
2468 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2469 : /* Otherwise, use ROW_MARK_COPY by default */
2470 200 : return ROW_MARK_COPY;
2471 : }
2472 : else
2473 : {
2474 : /* Regular table, apply the appropriate lock type */
2475 11972 : switch (strength)
2476 : {
2477 2480 : case LCS_NONE:
2478 :
2479 : /*
2480 : * We don't need a tuple lock, only the ability to re-fetch
2481 : * the row.
2482 : */
2483 2480 : return ROW_MARK_REFERENCE;
2484 : break;
2485 7606 : case LCS_FORKEYSHARE:
2486 7606 : return ROW_MARK_KEYSHARE;
2487 : break;
2488 300 : case LCS_FORSHARE:
2489 300 : return ROW_MARK_SHARE;
2490 : break;
2491 72 : case LCS_FORNOKEYUPDATE:
2492 72 : return ROW_MARK_NOKEYEXCLUSIVE;
2493 : break;
2494 1514 : case LCS_FORUPDATE:
2495 1514 : return ROW_MARK_EXCLUSIVE;
2496 : break;
2497 : }
2498 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2499 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2500 : }
2501 : }
2502 :
2503 : /*
2504 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2505 : *
2506 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2507 : * results back in *count_est and *offset_est. These variables are set to
2508 : * 0 if the corresponding clause is not present, and -1 if it's present
2509 : * but we couldn't estimate the value for it. (The "0" convention is OK
2510 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2511 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2512 : * usual practice of never estimating less than one row.) These values will
2513 : * be passed to create_limit_path, which see if you change this code.
2514 : *
2515 : * The return value is the suitably adjusted tuple_fraction to use for
2516 : * planning the query. This adjustment is not overridable, since it reflects
2517 : * plan actions that grouping_planner() will certainly take, not assumptions
2518 : * about context.
2519 : */
2520 : static double
2521 4950 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2522 : int64 *offset_est, int64 *count_est)
2523 : {
2524 4950 : Query *parse = root->parse;
2525 : Node *est;
2526 : double limit_fraction;
2527 :
2528 : /* Should not be called unless LIMIT or OFFSET */
2529 : Assert(parse->limitCount || parse->limitOffset);
2530 :
2531 : /*
2532 : * Try to obtain the clause values. We use estimate_expression_value
2533 : * primarily because it can sometimes do something useful with Params.
2534 : */
2535 4950 : if (parse->limitCount)
2536 : {
2537 4452 : est = estimate_expression_value(root, parse->limitCount);
2538 4452 : if (est && IsA(est, Const))
2539 : {
2540 4446 : if (((Const *) est)->constisnull)
2541 : {
2542 : /* NULL indicates LIMIT ALL, ie, no limit */
2543 0 : *count_est = 0; /* treat as not present */
2544 : }
2545 : else
2546 : {
2547 4446 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2548 4446 : if (*count_est <= 0)
2549 150 : *count_est = 1; /* force to at least 1 */
2550 : }
2551 : }
2552 : else
2553 6 : *count_est = -1; /* can't estimate */
2554 : }
2555 : else
2556 498 : *count_est = 0; /* not present */
2557 :
2558 4950 : if (parse->limitOffset)
2559 : {
2560 870 : est = estimate_expression_value(root, parse->limitOffset);
2561 870 : if (est && IsA(est, Const))
2562 : {
2563 846 : if (((Const *) est)->constisnull)
2564 : {
2565 : /* Treat NULL as no offset; the executor will too */
2566 0 : *offset_est = 0; /* treat as not present */
2567 : }
2568 : else
2569 : {
2570 846 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2571 846 : if (*offset_est < 0)
2572 0 : *offset_est = 0; /* treat as not present */
2573 : }
2574 : }
2575 : else
2576 24 : *offset_est = -1; /* can't estimate */
2577 : }
2578 : else
2579 4080 : *offset_est = 0; /* not present */
2580 :
2581 4950 : if (*count_est != 0)
2582 : {
2583 : /*
2584 : * A LIMIT clause limits the absolute number of tuples returned.
2585 : * However, if it's not a constant LIMIT then we have to guess; for
2586 : * lack of a better idea, assume 10% of the plan's result is wanted.
2587 : */
2588 4452 : if (*count_est < 0 || *offset_est < 0)
2589 : {
2590 : /* LIMIT or OFFSET is an expression ... punt ... */
2591 24 : limit_fraction = 0.10;
2592 : }
2593 : else
2594 : {
2595 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2596 4428 : limit_fraction = (double) *count_est + (double) *offset_est;
2597 : }
2598 :
2599 : /*
2600 : * If we have absolute limits from both caller and LIMIT, use the
2601 : * smaller value; likewise if they are both fractional. If one is
2602 : * fractional and the other absolute, we can't easily determine which
2603 : * is smaller, but we use the heuristic that the absolute will usually
2604 : * be smaller.
2605 : */
2606 4452 : if (tuple_fraction >= 1.0)
2607 : {
2608 6 : if (limit_fraction >= 1.0)
2609 : {
2610 : /* both absolute */
2611 6 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2612 : }
2613 : else
2614 : {
2615 : /* caller absolute, limit fractional; use caller's value */
2616 : }
2617 : }
2618 4446 : else if (tuple_fraction > 0.0)
2619 : {
2620 148 : if (limit_fraction >= 1.0)
2621 : {
2622 : /* caller fractional, limit absolute; use limit */
2623 148 : tuple_fraction = limit_fraction;
2624 : }
2625 : else
2626 : {
2627 : /* both fractional */
2628 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2629 : }
2630 : }
2631 : else
2632 : {
2633 : /* no info from caller, just use limit */
2634 4298 : tuple_fraction = limit_fraction;
2635 : }
2636 : }
2637 498 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2638 : {
2639 : /*
2640 : * We have an OFFSET but no LIMIT. This acts entirely differently
2641 : * from the LIMIT case: here, we need to increase rather than decrease
2642 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2643 : * tuples to be fetched instead of fewer. This only matters if we got
2644 : * a tuple_fraction > 0, however.
2645 : *
2646 : * As above, use 10% if OFFSET is present but unestimatable.
2647 : */
2648 12 : if (*offset_est < 0)
2649 0 : limit_fraction = 0.10;
2650 : else
2651 12 : limit_fraction = (double) *offset_est;
2652 :
2653 : /*
2654 : * If we have absolute counts from both caller and OFFSET, add them
2655 : * together; likewise if they are both fractional. If one is
2656 : * fractional and the other absolute, we want to take the larger, and
2657 : * we heuristically assume that's the fractional one.
2658 : */
2659 12 : if (tuple_fraction >= 1.0)
2660 : {
2661 0 : if (limit_fraction >= 1.0)
2662 : {
2663 : /* both absolute, so add them together */
2664 0 : tuple_fraction += limit_fraction;
2665 : }
2666 : else
2667 : {
2668 : /* caller absolute, limit fractional; use limit */
2669 0 : tuple_fraction = limit_fraction;
2670 : }
2671 : }
2672 : else
2673 : {
2674 12 : if (limit_fraction >= 1.0)
2675 : {
2676 : /* caller fractional, limit absolute; use caller's value */
2677 : }
2678 : else
2679 : {
2680 : /* both fractional, so add them together */
2681 0 : tuple_fraction += limit_fraction;
2682 0 : if (tuple_fraction >= 1.0)
2683 0 : tuple_fraction = 0.0; /* assume fetch all */
2684 : }
2685 : }
2686 : }
2687 :
2688 4950 : return tuple_fraction;
2689 : }
2690 :
2691 : /*
2692 : * limit_needed - do we actually need a Limit plan node?
2693 : *
2694 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2695 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2696 : * locution for an optimization fence. (Because other places in the planner
2697 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2698 : * an optimization fence --- we're just suppressing unnecessary run-time
2699 : * overhead.)
2700 : *
2701 : * This might look like it could be merged into preprocess_limit, but there's
2702 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2703 : * in preprocess_limit it's good enough to consider estimated values.
2704 : */
2705 : bool
2706 1119056 : limit_needed(Query *parse)
2707 : {
2708 : Node *node;
2709 :
2710 1119056 : node = parse->limitCount;
2711 1119056 : if (node)
2712 : {
2713 10614 : if (IsA(node, Const))
2714 : {
2715 : /* NULL indicates LIMIT ALL, ie, no limit */
2716 10414 : if (!((Const *) node)->constisnull)
2717 10414 : return true; /* LIMIT with a constant value */
2718 : }
2719 : else
2720 200 : return true; /* non-constant LIMIT */
2721 : }
2722 :
2723 1108442 : node = parse->limitOffset;
2724 1108442 : if (node)
2725 : {
2726 1456 : if (IsA(node, Const))
2727 : {
2728 : /* Treat NULL as no offset; the executor would too */
2729 1160 : if (!((Const *) node)->constisnull)
2730 : {
2731 1160 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2732 :
2733 1160 : if (offset != 0)
2734 110 : return true; /* OFFSET with a nonzero value */
2735 : }
2736 : }
2737 : else
2738 296 : return true; /* non-constant OFFSET */
2739 : }
2740 :
2741 1108036 : return false; /* don't need a Limit plan node */
2742 : }
2743 :
2744 : /*
2745 : * preprocess_groupclause - do preparatory work on GROUP BY clause
2746 : *
2747 : * The idea here is to adjust the ordering of the GROUP BY elements
2748 : * (which in itself is semantically insignificant) to match ORDER BY,
2749 : * thereby allowing a single sort operation to both implement the ORDER BY
2750 : * requirement and set up for a Unique step that implements GROUP BY.
2751 : * We also consider partial match between GROUP BY and ORDER BY elements,
2752 : * which could allow to implement ORDER BY using the incremental sort.
2753 : *
2754 : * We also consider other orderings of the GROUP BY elements, which could
2755 : * match the sort ordering of other possible plans (eg an indexscan) and
2756 : * thereby reduce cost. This is implemented during the generation of grouping
2757 : * paths. See get_useful_group_keys_orderings() for details.
2758 : *
2759 : * Note: we need no comparable processing of the distinctClause because
2760 : * the parser already enforced that that matches ORDER BY.
2761 : *
2762 : * Note: we return a fresh List, but its elements are the same
2763 : * SortGroupClauses appearing in parse->groupClause. This is important
2764 : * because later processing may modify the processed_groupClause list.
2765 : *
2766 : * For grouping sets, the order of items is instead forced to agree with that
2767 : * of the grouping set (and items not in the grouping set are skipped). The
2768 : * work of sorting the order of grouping set elements to match the ORDER BY if
2769 : * possible is done elsewhere.
2770 : */
2771 : static List *
2772 7716 : preprocess_groupclause(PlannerInfo *root, List *force)
2773 : {
2774 7716 : Query *parse = root->parse;
2775 7716 : List *new_groupclause = NIL;
2776 : ListCell *sl;
2777 : ListCell *gl;
2778 :
2779 : /* For grouping sets, we need to force the ordering */
2780 7716 : if (force)
2781 : {
2782 10036 : foreach(sl, force)
2783 : {
2784 5950 : Index ref = lfirst_int(sl);
2785 5950 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2786 :
2787 5950 : new_groupclause = lappend(new_groupclause, cl);
2788 : }
2789 :
2790 4086 : return new_groupclause;
2791 : }
2792 :
2793 : /* If no ORDER BY, nothing useful to do here */
2794 3630 : if (parse->sortClause == NIL)
2795 2056 : return list_copy(parse->groupClause);
2796 :
2797 : /*
2798 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2799 : * items, but only as far as we can make a matching prefix.
2800 : *
2801 : * This code assumes that the sortClause contains no duplicate items.
2802 : */
2803 3054 : foreach(sl, parse->sortClause)
2804 : {
2805 2126 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2806 :
2807 3222 : foreach(gl, parse->groupClause)
2808 : {
2809 2576 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2810 :
2811 2576 : if (equal(gc, sc))
2812 : {
2813 1480 : new_groupclause = lappend(new_groupclause, gc);
2814 1480 : break;
2815 : }
2816 : }
2817 2126 : if (gl == NULL)
2818 646 : break; /* no match, so stop scanning */
2819 : }
2820 :
2821 :
2822 : /* If no match at all, no point in reordering GROUP BY */
2823 1574 : if (new_groupclause == NIL)
2824 298 : return list_copy(parse->groupClause);
2825 :
2826 : /*
2827 : * Add any remaining GROUP BY items to the new list. We don't require a
2828 : * complete match, because even partial match allows ORDER BY to be
2829 : * implemented using incremental sort. Also, give up if there are any
2830 : * non-sortable GROUP BY items, since then there's no hope anyway.
2831 : */
2832 2922 : foreach(gl, parse->groupClause)
2833 : {
2834 1646 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2835 :
2836 1646 : if (list_member_ptr(new_groupclause, gc))
2837 1480 : continue; /* it matched an ORDER BY item */
2838 166 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
2839 0 : return list_copy(parse->groupClause);
2840 166 : new_groupclause = lappend(new_groupclause, gc);
2841 : }
2842 :
2843 : /* Success --- install the rearranged GROUP BY list */
2844 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
2845 1276 : return new_groupclause;
2846 : }
2847 :
2848 : /*
2849 : * Extract lists of grouping sets that can be implemented using a single
2850 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2851 : *
2852 : * Input must be sorted with smallest sets first. Result has each sublist
2853 : * sorted with smallest sets first.
2854 : *
2855 : * We want to produce the absolute minimum possible number of lists here to
2856 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2857 : * of finding the minimal partition of a partially-ordered set into chains
2858 : * (which is what we need, taking the list of grouping sets as a poset ordered
2859 : * by set inclusion) can be mapped to the problem of finding the maximum
2860 : * cardinality matching on a bipartite graph, which is solvable in polynomial
2861 : * time with a worst case of no worse than O(n^2.5) and usually much
2862 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
2863 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
2864 : * half a second on my modest system even with optimization off and assertions
2865 : * on.)
2866 : */
2867 : static List *
2868 866 : extract_rollup_sets(List *groupingSets)
2869 : {
2870 866 : int num_sets_raw = list_length(groupingSets);
2871 866 : int num_empty = 0;
2872 866 : int num_sets = 0; /* distinct sets */
2873 866 : int num_chains = 0;
2874 866 : List *result = NIL;
2875 : List **results;
2876 : List **orig_sets;
2877 : Bitmapset **set_masks;
2878 : int *chains;
2879 : short **adjacency;
2880 : short *adjacency_buf;
2881 : BipartiteMatchState *state;
2882 : int i;
2883 : int j;
2884 : int j_size;
2885 866 : ListCell *lc1 = list_head(groupingSets);
2886 : ListCell *lc;
2887 :
2888 : /*
2889 : * Start by stripping out empty sets. The algorithm doesn't require this,
2890 : * but the planner currently needs all empty sets to be returned in the
2891 : * first list, so we strip them here and add them back after.
2892 : */
2893 1476 : while (lc1 && lfirst(lc1) == NIL)
2894 : {
2895 610 : ++num_empty;
2896 610 : lc1 = lnext(groupingSets, lc1);
2897 : }
2898 :
2899 : /* bail out now if it turns out that all we had were empty sets. */
2900 866 : if (!lc1)
2901 42 : return list_make1(groupingSets);
2902 :
2903 : /*----------
2904 : * We don't strictly need to remove duplicate sets here, but if we don't,
2905 : * they tend to become scattered through the result, which is a bit
2906 : * confusing (and irritating if we ever decide to optimize them out).
2907 : * So we remove them here and add them back after.
2908 : *
2909 : * For each non-duplicate set, we fill in the following:
2910 : *
2911 : * orig_sets[i] = list of the original set lists
2912 : * set_masks[i] = bitmapset for testing inclusion
2913 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2914 : *
2915 : * chains[i] will be the result group this set is assigned to.
2916 : *
2917 : * We index all of these from 1 rather than 0 because it is convenient
2918 : * to leave 0 free for the NIL node in the graph algorithm.
2919 : *----------
2920 : */
2921 824 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
2922 824 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
2923 824 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
2924 824 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
2925 :
2926 824 : j_size = 0;
2927 824 : j = 0;
2928 824 : i = 1;
2929 :
2930 2936 : for_each_cell(lc, groupingSets, lc1)
2931 : {
2932 2112 : List *candidate = (List *) lfirst(lc);
2933 2112 : Bitmapset *candidate_set = NULL;
2934 : ListCell *lc2;
2935 2112 : int dup_of = 0;
2936 :
2937 5118 : foreach(lc2, candidate)
2938 : {
2939 3006 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
2940 : }
2941 :
2942 : /* we can only be a dup if we're the same length as a previous set */
2943 2112 : if (j_size == list_length(candidate))
2944 : {
2945 : int k;
2946 :
2947 1844 : for (k = j; k < i; ++k)
2948 : {
2949 1188 : if (bms_equal(set_masks[k], candidate_set))
2950 : {
2951 158 : dup_of = k;
2952 158 : break;
2953 : }
2954 : }
2955 : }
2956 1298 : else if (j_size < list_length(candidate))
2957 : {
2958 1298 : j_size = list_length(candidate);
2959 1298 : j = i;
2960 : }
2961 :
2962 2112 : if (dup_of > 0)
2963 : {
2964 158 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
2965 158 : bms_free(candidate_set);
2966 : }
2967 : else
2968 : {
2969 : int k;
2970 1954 : int n_adj = 0;
2971 :
2972 1954 : orig_sets[i] = list_make1(candidate);
2973 1954 : set_masks[i] = candidate_set;
2974 :
2975 : /* fill in adjacency list; no need to compare equal-size sets */
2976 :
2977 3226 : for (k = j - 1; k > 0; --k)
2978 : {
2979 1272 : if (bms_is_subset(set_masks[k], candidate_set))
2980 1110 : adjacency_buf[++n_adj] = k;
2981 : }
2982 :
2983 1954 : if (n_adj > 0)
2984 : {
2985 598 : adjacency_buf[0] = n_adj;
2986 598 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
2987 598 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
2988 : }
2989 : else
2990 1356 : adjacency[i] = NULL;
2991 :
2992 1954 : ++i;
2993 : }
2994 : }
2995 :
2996 824 : num_sets = i - 1;
2997 :
2998 : /*
2999 : * Apply the graph matching algorithm to do the work.
3000 : */
3001 824 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3002 :
3003 : /*
3004 : * Now, the state->pair* fields have the info we need to assign sets to
3005 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3006 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3007 : * it in one pass)
3008 : */
3009 824 : chains = palloc0((num_sets + 1) * sizeof(int));
3010 :
3011 2778 : for (i = 1; i <= num_sets; ++i)
3012 : {
3013 1954 : int u = state->pair_vu[i];
3014 1954 : int v = state->pair_uv[i];
3015 :
3016 1954 : if (u > 0 && u < i)
3017 0 : chains[i] = chains[u];
3018 1954 : else if (v > 0 && v < i)
3019 570 : chains[i] = chains[v];
3020 : else
3021 1384 : chains[i] = ++num_chains;
3022 : }
3023 :
3024 : /* build result lists. */
3025 824 : results = palloc0((num_chains + 1) * sizeof(List *));
3026 :
3027 2778 : for (i = 1; i <= num_sets; ++i)
3028 : {
3029 1954 : int c = chains[i];
3030 :
3031 : Assert(c > 0);
3032 :
3033 1954 : results[c] = list_concat(results[c], orig_sets[i]);
3034 : }
3035 :
3036 : /* push any empty sets back on the first list. */
3037 1344 : while (num_empty-- > 0)
3038 520 : results[1] = lcons(NIL, results[1]);
3039 :
3040 : /* make result list */
3041 2208 : for (i = 1; i <= num_chains; ++i)
3042 1384 : result = lappend(result, results[i]);
3043 :
3044 : /*
3045 : * Free all the things.
3046 : *
3047 : * (This is over-fussy for small sets but for large sets we could have
3048 : * tied up a nontrivial amount of memory.)
3049 : */
3050 824 : BipartiteMatchFree(state);
3051 824 : pfree(results);
3052 824 : pfree(chains);
3053 2778 : for (i = 1; i <= num_sets; ++i)
3054 1954 : if (adjacency[i])
3055 598 : pfree(adjacency[i]);
3056 824 : pfree(adjacency);
3057 824 : pfree(adjacency_buf);
3058 824 : pfree(orig_sets);
3059 2778 : for (i = 1; i <= num_sets; ++i)
3060 1954 : bms_free(set_masks[i]);
3061 824 : pfree(set_masks);
3062 :
3063 824 : return result;
3064 : }
3065 :
3066 : /*
3067 : * Reorder the elements of a list of grouping sets such that they have correct
3068 : * prefix relationships. Also inserts the GroupingSetData annotations.
3069 : *
3070 : * The input must be ordered with smallest sets first; the result is returned
3071 : * with largest sets first. Note that the result shares no list substructure
3072 : * with the input, so it's safe for the caller to modify it later.
3073 : *
3074 : * If we're passed in a sortclause, we follow its order of columns to the
3075 : * extent possible, to minimize the chance that we add unnecessary sorts.
3076 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3077 : * gets implemented in one pass.)
3078 : */
3079 : static List *
3080 1426 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3081 : {
3082 : ListCell *lc;
3083 1426 : List *previous = NIL;
3084 1426 : List *result = NIL;
3085 :
3086 4148 : foreach(lc, groupingSets)
3087 : {
3088 2722 : List *candidate = (List *) lfirst(lc);
3089 2722 : List *new_elems = list_difference_int(candidate, previous);
3090 2722 : GroupingSetData *gs = makeNode(GroupingSetData);
3091 :
3092 2886 : while (list_length(sortclause) > list_length(previous) &&
3093 : new_elems != NIL)
3094 : {
3095 272 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3096 272 : int ref = sc->tleSortGroupRef;
3097 :
3098 272 : if (list_member_int(new_elems, ref))
3099 : {
3100 164 : previous = lappend_int(previous, ref);
3101 164 : new_elems = list_delete_int(new_elems, ref);
3102 : }
3103 : else
3104 : {
3105 : /* diverged from the sortclause; give up on it */
3106 108 : sortclause = NIL;
3107 108 : break;
3108 : }
3109 : }
3110 :
3111 2722 : previous = list_concat(previous, new_elems);
3112 :
3113 2722 : gs->set = list_copy(previous);
3114 2722 : result = lcons(gs, result);
3115 : }
3116 :
3117 1426 : list_free(previous);
3118 :
3119 1426 : return result;
3120 : }
3121 :
3122 : /*
3123 : * has_volatile_pathkey
3124 : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3125 : * containing a volatile function. Otherwise returns false.
3126 : */
3127 : static bool
3128 3092 : has_volatile_pathkey(List *keys)
3129 : {
3130 : ListCell *lc;
3131 :
3132 6328 : foreach(lc, keys)
3133 : {
3134 3254 : PathKey *pathkey = lfirst_node(PathKey, lc);
3135 :
3136 3254 : if (pathkey->pk_eclass->ec_has_volatile)
3137 18 : return true;
3138 : }
3139 :
3140 3074 : return false;
3141 : }
3142 :
3143 : /*
3144 : * adjust_group_pathkeys_for_groupagg
3145 : * Add pathkeys to root->group_pathkeys to reflect the best set of
3146 : * pre-ordered input for ordered aggregates.
3147 : *
3148 : * We define "best" as the pathkeys that suit the largest number of
3149 : * aggregate functions. We find these by looking at the first ORDER BY /
3150 : * DISTINCT aggregate and take the pathkeys for that before searching for
3151 : * other aggregates that require the same or a more strict variation of the
3152 : * same pathkeys. We then repeat that process for any remaining aggregates
3153 : * with different pathkeys and if we find another set of pathkeys that suits a
3154 : * larger number of aggregates then we select those pathkeys instead.
3155 : *
3156 : * When the best pathkeys are found we also mark each Aggref that can use
3157 : * those pathkeys as aggpresorted = true.
3158 : *
3159 : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3160 : * volatile functions, we never make use of these pathkeys. We want to ensure
3161 : * that sorts using volatile functions are done independently in each Aggref
3162 : * rather than once at the query level. If we were to allow this then Aggrefs
3163 : * with compatible sort orders would all transition their rows in the same
3164 : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3165 : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3166 : * better pathkeys to sort on, then the volatile function Aggrefs would be
3167 : * left to perform their sorts individually. To avoid this inconsistent
3168 : * behavior which could make Aggref results depend on what other Aggrefs the
3169 : * query contains, we always force Aggrefs with volatile functions to perform
3170 : * their own sorts.
3171 : */
3172 : static void
3173 2696 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3174 : {
3175 2696 : List *grouppathkeys = root->group_pathkeys;
3176 : List *bestpathkeys;
3177 : Bitmapset *bestaggs;
3178 : Bitmapset *unprocessed_aggs;
3179 : ListCell *lc;
3180 : int i;
3181 :
3182 : /* Shouldn't be here if there are grouping sets */
3183 : Assert(root->parse->groupingSets == NIL);
3184 : /* Shouldn't be here unless there are some ordered aggregates */
3185 : Assert(root->numOrderedAggs > 0);
3186 :
3187 : /* Do nothing if disabled */
3188 2696 : if (!enable_presorted_aggregate)
3189 6 : return;
3190 :
3191 : /*
3192 : * Make a first pass over all AggInfos to collect a Bitmapset containing
3193 : * the indexes of all AggInfos to be processed below.
3194 : */
3195 2690 : unprocessed_aggs = NULL;
3196 6064 : foreach(lc, root->agginfos)
3197 : {
3198 3374 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3199 3374 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3200 :
3201 3374 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3202 264 : continue;
3203 :
3204 : /* Skip unless there's a DISTINCT or ORDER BY clause */
3205 3110 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3206 300 : continue;
3207 :
3208 : /* Additional safety checks are needed if there's a FILTER clause */
3209 2810 : if (aggref->aggfilter != NULL)
3210 : {
3211 : ListCell *lc2;
3212 54 : bool allow_presort = true;
3213 :
3214 : /*
3215 : * When the Aggref has a FILTER clause, it's possible that the
3216 : * filter removes rows that cannot be sorted because the
3217 : * expression to sort by results in an error during its
3218 : * evaluation. This is a problem for presorting as that happens
3219 : * before the FILTER, whereas without presorting, the Aggregate
3220 : * node will apply the FILTER *before* sorting. So that we never
3221 : * try to sort anything that might error, here we aim to skip over
3222 : * any Aggrefs with arguments with expressions which, when
3223 : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3224 : * may be more cases that should be allowed, but more thought
3225 : * needs to be given. Err on the side of caution.
3226 : */
3227 102 : foreach(lc2, aggref->args)
3228 : {
3229 72 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3230 72 : Expr *expr = tle->expr;
3231 :
3232 84 : while (IsA(expr, RelabelType))
3233 12 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3234 :
3235 : /* Common case, Vars and Consts are ok */
3236 72 : if (IsA(expr, Var) || IsA(expr, Const))
3237 48 : continue;
3238 :
3239 : /* Unsupported. Don't try to presort for this Aggref */
3240 24 : allow_presort = false;
3241 24 : break;
3242 : }
3243 :
3244 : /* Skip unsupported Aggrefs */
3245 54 : if (!allow_presort)
3246 24 : continue;
3247 : }
3248 :
3249 2786 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3250 : foreach_current_index(lc));
3251 : }
3252 :
3253 : /*
3254 : * Now process all the unprocessed_aggs to find the best set of pathkeys
3255 : * for the given set of aggregates.
3256 : *
3257 : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3258 : * this during the first loop using the pathkeys for the very first
3259 : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3260 : * a more strict set of compatible pathkeys. Once the outer loop is
3261 : * complete, we mark off all the aggregates with compatible pathkeys then
3262 : * remove those from the unprocessed_aggs and repeat the process to try to
3263 : * find another set of pathkeys that are suitable for a larger number of
3264 : * aggregates. The outer loop will stop when there are not enough
3265 : * unprocessed aggregates for it to be possible to find a set of pathkeys
3266 : * to suit a larger number of aggregates.
3267 : */
3268 2690 : bestpathkeys = NIL;
3269 2690 : bestaggs = NULL;
3270 5314 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3271 : {
3272 2624 : Bitmapset *aggindexes = NULL;
3273 2624 : List *currpathkeys = NIL;
3274 :
3275 2624 : i = -1;
3276 8340 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3277 : {
3278 3092 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3279 3092 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3280 : List *sortlist;
3281 : List *pathkeys;
3282 :
3283 3092 : if (aggref->aggdistinct != NIL)
3284 718 : sortlist = aggref->aggdistinct;
3285 : else
3286 2374 : sortlist = aggref->aggorder;
3287 :
3288 3092 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3289 : aggref->args);
3290 :
3291 : /*
3292 : * Ignore Aggrefs which have volatile functions in their ORDER BY
3293 : * or DISTINCT clause.
3294 : */
3295 3092 : if (has_volatile_pathkey(pathkeys))
3296 : {
3297 18 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3298 18 : continue;
3299 : }
3300 :
3301 : /*
3302 : * When not set yet, take the pathkeys from the first unprocessed
3303 : * aggregate.
3304 : */
3305 3074 : if (currpathkeys == NIL)
3306 : {
3307 2618 : currpathkeys = pathkeys;
3308 :
3309 : /* include the GROUP BY pathkeys, if they exist */
3310 2618 : if (grouppathkeys != NIL)
3311 276 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3312 : currpathkeys);
3313 :
3314 : /* record that we found pathkeys for this aggregate */
3315 2618 : aggindexes = bms_add_member(aggindexes, i);
3316 : }
3317 : else
3318 : {
3319 : /* now look for a stronger set of matching pathkeys */
3320 :
3321 : /* include the GROUP BY pathkeys, if they exist */
3322 456 : if (grouppathkeys != NIL)
3323 288 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3324 : pathkeys);
3325 :
3326 : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3327 456 : switch (compare_pathkeys(currpathkeys, pathkeys))
3328 : {
3329 12 : case PATHKEYS_BETTER2:
3330 : /* 'pathkeys' are stronger, use these ones instead */
3331 12 : currpathkeys = pathkeys;
3332 : /* FALLTHROUGH */
3333 :
3334 66 : case PATHKEYS_BETTER1:
3335 : /* 'pathkeys' are less strict */
3336 : /* FALLTHROUGH */
3337 :
3338 : case PATHKEYS_EQUAL:
3339 : /* mark this aggregate as covered by 'currpathkeys' */
3340 66 : aggindexes = bms_add_member(aggindexes, i);
3341 66 : break;
3342 :
3343 390 : case PATHKEYS_DIFFERENT:
3344 390 : break;
3345 : }
3346 5716 : }
3347 : }
3348 :
3349 : /* remove the aggregates that we've just processed */
3350 2624 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3351 :
3352 : /*
3353 : * If this pass included more aggregates than the previous best then
3354 : * use these ones as the best set.
3355 : */
3356 2624 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3357 : {
3358 2516 : bestaggs = aggindexes;
3359 2516 : bestpathkeys = currpathkeys;
3360 : }
3361 : }
3362 :
3363 : /*
3364 : * If we found any ordered aggregates, update root->group_pathkeys to add
3365 : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3366 : * the original GROUP BY pathkeys already.
3367 : */
3368 2690 : if (bestpathkeys != NIL)
3369 2456 : root->group_pathkeys = bestpathkeys;
3370 :
3371 : /*
3372 : * Now that we've found the best set of aggregates we can set the
3373 : * presorted flag to indicate to the executor that it needn't bother
3374 : * performing a sort for these Aggrefs. We're able to do this now as
3375 : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3376 : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3377 : * of ordered aggregates.
3378 : */
3379 2690 : i = -1;
3380 5242 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3381 : {
3382 2552 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3383 :
3384 5122 : foreach(lc, agginfo->aggrefs)
3385 : {
3386 2570 : Aggref *aggref = lfirst_node(Aggref, lc);
3387 :
3388 2570 : aggref->aggpresorted = true;
3389 : }
3390 : }
3391 : }
3392 :
3393 : /*
3394 : * Compute query_pathkeys and other pathkeys during plan generation
3395 : */
3396 : static void
3397 524406 : standard_qp_callback(PlannerInfo *root, void *extra)
3398 : {
3399 524406 : Query *parse = root->parse;
3400 524406 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3401 524406 : List *tlist = root->processed_tlist;
3402 524406 : List *activeWindows = qp_extra->activeWindows;
3403 :
3404 : /*
3405 : * Calculate pathkeys that represent grouping/ordering and/or ordered
3406 : * aggregate requirements.
3407 : */
3408 524406 : if (qp_extra->gset_data)
3409 : {
3410 : /*
3411 : * With grouping sets, just use the first RollupData's groupClause. We
3412 : * don't make any effort to optimize grouping clauses when there are
3413 : * grouping sets, nor can we combine aggregate ordering keys with
3414 : * grouping.
3415 : */
3416 872 : List *rollups = qp_extra->gset_data->rollups;
3417 872 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3418 :
3419 872 : if (grouping_is_sortable(groupClause))
3420 : {
3421 : bool sortable;
3422 :
3423 : /*
3424 : * The groupClause is logically below the grouping step. So if
3425 : * there is an RTE entry for the grouping step, we need to remove
3426 : * its RT index from the sort expressions before we make PathKeys
3427 : * for them.
3428 : */
3429 872 : root->group_pathkeys =
3430 872 : make_pathkeys_for_sortclauses_extended(root,
3431 : &groupClause,
3432 : tlist,
3433 : false,
3434 872 : parse->hasGroupRTE,
3435 : &sortable,
3436 : false);
3437 : Assert(sortable);
3438 872 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3439 : }
3440 : else
3441 : {
3442 0 : root->group_pathkeys = NIL;
3443 0 : root->num_groupby_pathkeys = 0;
3444 : }
3445 : }
3446 523534 : else if (parse->groupClause || root->numOrderedAggs > 0)
3447 6082 : {
3448 : /*
3449 : * With a plain GROUP BY list, we can remove any grouping items that
3450 : * are proven redundant by EquivalenceClass processing. For example,
3451 : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3452 : * especially common cases, but they're nearly free to detect. Note
3453 : * that we remove redundant items from processed_groupClause but not
3454 : * the original parse->groupClause.
3455 : */
3456 : bool sortable;
3457 :
3458 : /*
3459 : * Convert group clauses into pathkeys. Set the ec_sortref field of
3460 : * EquivalenceClass'es if it's not set yet.
3461 : */
3462 6082 : root->group_pathkeys =
3463 6082 : make_pathkeys_for_sortclauses_extended(root,
3464 : &root->processed_groupClause,
3465 : tlist,
3466 : true,
3467 : false,
3468 : &sortable,
3469 : true);
3470 6082 : if (!sortable)
3471 : {
3472 : /* Can't sort; no point in considering aggregate ordering either */
3473 0 : root->group_pathkeys = NIL;
3474 0 : root->num_groupby_pathkeys = 0;
3475 : }
3476 : else
3477 : {
3478 6082 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3479 : /* If we have ordered aggs, consider adding onto group_pathkeys */
3480 6082 : if (root->numOrderedAggs > 0)
3481 2696 : adjust_group_pathkeys_for_groupagg(root);
3482 : }
3483 : }
3484 : else
3485 : {
3486 517452 : root->group_pathkeys = NIL;
3487 517452 : root->num_groupby_pathkeys = 0;
3488 : }
3489 :
3490 : /* We consider only the first (bottom) window in pathkeys logic */
3491 524406 : if (activeWindows != NIL)
3492 : {
3493 2378 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3494 :
3495 2378 : root->window_pathkeys = make_pathkeys_for_window(root,
3496 : wc,
3497 : tlist);
3498 : }
3499 : else
3500 522028 : root->window_pathkeys = NIL;
3501 :
3502 : /*
3503 : * As with GROUP BY, we can discard any DISTINCT items that are proven
3504 : * redundant by EquivalenceClass processing. The non-redundant list is
3505 : * kept in root->processed_distinctClause, leaving the original
3506 : * parse->distinctClause alone.
3507 : */
3508 524406 : if (parse->distinctClause)
3509 : {
3510 : bool sortable;
3511 :
3512 : /* Make a copy since pathkey processing can modify the list */
3513 2674 : root->processed_distinctClause = list_copy(parse->distinctClause);
3514 2674 : root->distinct_pathkeys =
3515 2674 : make_pathkeys_for_sortclauses_extended(root,
3516 : &root->processed_distinctClause,
3517 : tlist,
3518 : true,
3519 : false,
3520 : &sortable,
3521 : false);
3522 2674 : if (!sortable)
3523 6 : root->distinct_pathkeys = NIL;
3524 : }
3525 : else
3526 521732 : root->distinct_pathkeys = NIL;
3527 :
3528 524406 : root->sort_pathkeys =
3529 524406 : make_pathkeys_for_sortclauses(root,
3530 : parse->sortClause,
3531 : tlist);
3532 :
3533 : /* setting setop_pathkeys might be useful to the union planner */
3534 524406 : if (qp_extra->setop != NULL)
3535 : {
3536 : List *groupClauses;
3537 : bool sortable;
3538 :
3539 12278 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3540 :
3541 12278 : root->setop_pathkeys =
3542 12278 : make_pathkeys_for_sortclauses_extended(root,
3543 : &groupClauses,
3544 : tlist,
3545 : false,
3546 : false,
3547 : &sortable,
3548 : false);
3549 12278 : if (!sortable)
3550 200 : root->setop_pathkeys = NIL;
3551 : }
3552 : else
3553 512128 : root->setop_pathkeys = NIL;
3554 :
3555 : /*
3556 : * Figure out whether we want a sorted result from query_planner.
3557 : *
3558 : * If we have a sortable GROUP BY clause, then we want a result sorted
3559 : * properly for grouping. Otherwise, if we have window functions to
3560 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3561 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3562 : * we try to produce output that's sufficiently well sorted for the
3563 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3564 : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3565 : * for a set operation which can benefit from presorted results and have a
3566 : * sortable targetlist, we want to sort by the target list.
3567 : *
3568 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3569 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3570 : * that might just leave us failing to exploit an available sort order at
3571 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3572 : * much easier, since we know that the parser ensured that one is a
3573 : * superset of the other.
3574 : */
3575 524406 : if (root->group_pathkeys)
3576 6598 : root->query_pathkeys = root->group_pathkeys;
3577 517808 : else if (root->window_pathkeys)
3578 2032 : root->query_pathkeys = root->window_pathkeys;
3579 1031552 : else if (list_length(root->distinct_pathkeys) >
3580 515776 : list_length(root->sort_pathkeys))
3581 2222 : root->query_pathkeys = root->distinct_pathkeys;
3582 513554 : else if (root->sort_pathkeys)
3583 69448 : root->query_pathkeys = root->sort_pathkeys;
3584 444106 : else if (root->setop_pathkeys != NIL)
3585 10878 : root->query_pathkeys = root->setop_pathkeys;
3586 : else
3587 433228 : root->query_pathkeys = NIL;
3588 524406 : }
3589 :
3590 : /*
3591 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3592 : *
3593 : * path_rows: number of output rows from scan/join step
3594 : * gd: grouping sets data including list of grouping sets and their clauses
3595 : * target_list: target list containing group clause references
3596 : *
3597 : * If doing grouping sets, we also annotate the gsets data with the estimates
3598 : * for each set and each individual rollup list, with a view to later
3599 : * determining whether some combination of them could be hashed instead.
3600 : */
3601 : static double
3602 42276 : get_number_of_groups(PlannerInfo *root,
3603 : double path_rows,
3604 : grouping_sets_data *gd,
3605 : List *target_list)
3606 : {
3607 42276 : Query *parse = root->parse;
3608 : double dNumGroups;
3609 :
3610 42276 : if (parse->groupClause)
3611 : {
3612 : List *groupExprs;
3613 :
3614 6940 : if (parse->groupingSets)
3615 : {
3616 : /* Add up the estimates for each grouping set */
3617 : ListCell *lc;
3618 :
3619 : Assert(gd); /* keep Coverity happy */
3620 :
3621 830 : dNumGroups = 0;
3622 :
3623 2214 : foreach(lc, gd->rollups)
3624 : {
3625 1384 : RollupData *rollup = lfirst_node(RollupData, lc);
3626 : ListCell *lc2;
3627 : ListCell *lc3;
3628 :
3629 1384 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3630 : target_list);
3631 :
3632 1384 : rollup->numGroups = 0.0;
3633 :
3634 4016 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
3635 : {
3636 2632 : List *gset = (List *) lfirst(lc2);
3637 2632 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3638 2632 : double numGroups = estimate_num_groups(root,
3639 : groupExprs,
3640 : path_rows,
3641 : &gset,
3642 : NULL);
3643 :
3644 2632 : gs->numGroups = numGroups;
3645 2632 : rollup->numGroups += numGroups;
3646 : }
3647 :
3648 1384 : dNumGroups += rollup->numGroups;
3649 : }
3650 :
3651 830 : if (gd->hash_sets_idx)
3652 : {
3653 : ListCell *lc2;
3654 :
3655 36 : gd->dNumHashGroups = 0;
3656 :
3657 36 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3658 : target_list);
3659 :
3660 78 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3661 : {
3662 42 : List *gset = (List *) lfirst(lc);
3663 42 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3664 42 : double numGroups = estimate_num_groups(root,
3665 : groupExprs,
3666 : path_rows,
3667 : &gset,
3668 : NULL);
3669 :
3670 42 : gs->numGroups = numGroups;
3671 42 : gd->dNumHashGroups += numGroups;
3672 : }
3673 :
3674 36 : dNumGroups += gd->dNumHashGroups;
3675 : }
3676 : }
3677 : else
3678 : {
3679 : /* Plain GROUP BY -- estimate based on optimized groupClause */
3680 6110 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3681 : target_list);
3682 :
3683 6110 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3684 : NULL, NULL);
3685 : }
3686 : }
3687 35336 : else if (parse->groupingSets)
3688 : {
3689 : /* Empty grouping sets ... one result row for each one */
3690 42 : dNumGroups = list_length(parse->groupingSets);
3691 : }
3692 35294 : else if (parse->hasAggs || root->hasHavingQual)
3693 : {
3694 : /* Plain aggregation, one result row */
3695 35294 : dNumGroups = 1;
3696 : }
3697 : else
3698 : {
3699 : /* Not grouping */
3700 0 : dNumGroups = 1;
3701 : }
3702 :
3703 42276 : return dNumGroups;
3704 : }
3705 :
3706 : /*
3707 : * create_grouping_paths
3708 : *
3709 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3710 : * Along the way, we also build an upperrel for Paths which are partially
3711 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3712 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3713 : * the only partially grouped paths we build are also partial paths; that
3714 : * is, they need a Gather and then a FinalizeAggregate.
3715 : *
3716 : * input_rel: contains the source-data Paths
3717 : * target: the pathtarget for the result Paths to compute
3718 : * gd: grouping sets data including list of grouping sets and their clauses
3719 : *
3720 : * Note: all Paths in input_rel are expected to return the target computed
3721 : * by make_group_input_target.
3722 : */
3723 : static RelOptInfo *
3724 39018 : create_grouping_paths(PlannerInfo *root,
3725 : RelOptInfo *input_rel,
3726 : PathTarget *target,
3727 : bool target_parallel_safe,
3728 : grouping_sets_data *gd)
3729 : {
3730 39018 : Query *parse = root->parse;
3731 : RelOptInfo *grouped_rel;
3732 : RelOptInfo *partially_grouped_rel;
3733 : AggClauseCosts agg_costs;
3734 :
3735 234108 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
3736 39018 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3737 :
3738 : /*
3739 : * Create grouping relation to hold fully aggregated grouping and/or
3740 : * aggregation paths.
3741 : */
3742 39018 : grouped_rel = make_grouping_rel(root, input_rel, target,
3743 : target_parallel_safe, parse->havingQual);
3744 :
3745 : /*
3746 : * Create either paths for a degenerate grouping or paths for ordinary
3747 : * grouping, as appropriate.
3748 : */
3749 39018 : if (is_degenerate_grouping(root))
3750 18 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3751 : else
3752 : {
3753 39000 : int flags = 0;
3754 : GroupPathExtraData extra;
3755 :
3756 : /*
3757 : * Determine whether it's possible to perform sort-based
3758 : * implementations of grouping. (Note that if processed_groupClause
3759 : * is empty, grouping_is_sortable() is trivially true, and all the
3760 : * pathkeys_contained_in() tests will succeed too, so that we'll
3761 : * consider every surviving input path.)
3762 : *
3763 : * If we have grouping sets, we might be able to sort some but not all
3764 : * of them; in this case, we need can_sort to be true as long as we
3765 : * must consider any sorted-input plan.
3766 : */
3767 39000 : if ((gd && gd->rollups != NIL)
3768 38134 : || grouping_is_sortable(root->processed_groupClause))
3769 38994 : flags |= GROUPING_CAN_USE_SORT;
3770 :
3771 : /*
3772 : * Determine whether we should consider hash-based implementations of
3773 : * grouping.
3774 : *
3775 : * Hashed aggregation only applies if we're grouping. If we have
3776 : * grouping sets, some groups might be hashable but others not; in
3777 : * this case we set can_hash true as long as there is nothing globally
3778 : * preventing us from hashing (and we should therefore consider plans
3779 : * with hashes).
3780 : *
3781 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3782 : * BY aggregates. (Doing so would imply storing *all* the input
3783 : * values in the hash table, and/or running many sorts in parallel,
3784 : * either of which seems like a certain loser.) We similarly don't
3785 : * support ordered-set aggregates in hashed aggregation, but that case
3786 : * is also included in the numOrderedAggs count.
3787 : *
3788 : * Note: grouping_is_hashable() is much more expensive to check than
3789 : * the other gating conditions, so we want to do it last.
3790 : */
3791 39000 : if ((parse->groupClause != NIL &&
3792 8640 : root->numOrderedAggs == 0 &&
3793 4180 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
3794 4176 : flags |= GROUPING_CAN_USE_HASH;
3795 :
3796 : /*
3797 : * Determine whether partial aggregation is possible.
3798 : */
3799 39000 : if (can_partial_agg(root))
3800 34050 : flags |= GROUPING_CAN_PARTIAL_AGG;
3801 :
3802 39000 : extra.flags = flags;
3803 39000 : extra.target_parallel_safe = target_parallel_safe;
3804 39000 : extra.havingQual = parse->havingQual;
3805 39000 : extra.targetList = parse->targetList;
3806 39000 : extra.partial_costs_set = false;
3807 :
3808 : /*
3809 : * Determine whether partitionwise aggregation is in theory possible.
3810 : * It can be disabled by the user, and for now, we don't try to
3811 : * support grouping sets. create_ordinary_grouping_paths() will check
3812 : * additional conditions, such as whether input_rel is partitioned.
3813 : */
3814 39000 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3815 556 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3816 : else
3817 38444 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3818 :
3819 39000 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3820 : &agg_costs, gd, &extra,
3821 : &partially_grouped_rel);
3822 : }
3823 :
3824 39012 : set_cheapest(grouped_rel);
3825 39012 : return grouped_rel;
3826 : }
3827 :
3828 : /*
3829 : * make_grouping_rel
3830 : *
3831 : * Create a new grouping rel and set basic properties.
3832 : *
3833 : * input_rel represents the underlying scan/join relation.
3834 : * target is the output expected from the grouping relation.
3835 : */
3836 : static RelOptInfo *
3837 40512 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3838 : PathTarget *target, bool target_parallel_safe,
3839 : Node *havingQual)
3840 : {
3841 : RelOptInfo *grouped_rel;
3842 :
3843 40512 : if (IS_OTHER_REL(input_rel))
3844 : {
3845 1494 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3846 : input_rel->relids);
3847 1494 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3848 : }
3849 : else
3850 : {
3851 : /*
3852 : * By tradition, the relids set for the main grouping relation is
3853 : * NULL. (This could be changed, but might require adjustments
3854 : * elsewhere.)
3855 : */
3856 39018 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3857 : }
3858 :
3859 : /* Set target. */
3860 40512 : grouped_rel->reltarget = target;
3861 :
3862 : /*
3863 : * If the input relation is not parallel-safe, then the grouped relation
3864 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3865 : * target list and HAVING quals are parallel-safe.
3866 : */
3867 68592 : if (input_rel->consider_parallel && target_parallel_safe &&
3868 28080 : is_parallel_safe(root, (Node *) havingQual))
3869 28062 : grouped_rel->consider_parallel = true;
3870 :
3871 : /*
3872 : * If the input rel belongs to a single FDW, so does the grouped rel.
3873 : */
3874 40512 : grouped_rel->serverid = input_rel->serverid;
3875 40512 : grouped_rel->userid = input_rel->userid;
3876 40512 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3877 40512 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3878 :
3879 40512 : return grouped_rel;
3880 : }
3881 :
3882 : /*
3883 : * is_degenerate_grouping
3884 : *
3885 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3886 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3887 : * grouping sets are all empty).
3888 : */
3889 : static bool
3890 39018 : is_degenerate_grouping(PlannerInfo *root)
3891 : {
3892 39018 : Query *parse = root->parse;
3893 :
3894 38002 : return (root->hasHavingQual || parse->groupingSets) &&
3895 77020 : !parse->hasAggs && parse->groupClause == NIL;
3896 : }
3897 :
3898 : /*
3899 : * create_degenerate_grouping_paths
3900 : *
3901 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3902 : * supposed to emit either zero or one row for each grouping set depending on
3903 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3904 : * either HAVING or the targetlist, so we actually do not need the FROM table
3905 : * at all! We can just throw away the plan-so-far and generate a Result node.
3906 : * This is a sufficiently unusual corner case that it's not worth contorting
3907 : * the structure of this module to avoid having to generate the earlier paths
3908 : * in the first place.
3909 : */
3910 : static void
3911 18 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3912 : RelOptInfo *grouped_rel)
3913 : {
3914 18 : Query *parse = root->parse;
3915 : int nrows;
3916 : Path *path;
3917 :
3918 18 : nrows = list_length(parse->groupingSets);
3919 18 : if (nrows > 1)
3920 : {
3921 : /*
3922 : * Doesn't seem worthwhile writing code to cons up a generate_series
3923 : * or a values scan to emit multiple rows. Instead just make N clones
3924 : * and append them. (With a volatile HAVING clause, this means you
3925 : * might get between 0 and N output rows. Offhand I think that's
3926 : * desired.)
3927 : */
3928 0 : List *paths = NIL;
3929 :
3930 0 : while (--nrows >= 0)
3931 : {
3932 : path = (Path *)
3933 0 : create_group_result_path(root, grouped_rel,
3934 0 : grouped_rel->reltarget,
3935 0 : (List *) parse->havingQual);
3936 0 : paths = lappend(paths, path);
3937 : }
3938 : path = (Path *)
3939 0 : create_append_path(root,
3940 : grouped_rel,
3941 : paths,
3942 : NIL,
3943 : NIL,
3944 : NULL,
3945 : 0,
3946 : false,
3947 : -1);
3948 : }
3949 : else
3950 : {
3951 : /* No grouping sets, or just one, so one output row */
3952 : path = (Path *)
3953 18 : create_group_result_path(root, grouped_rel,
3954 18 : grouped_rel->reltarget,
3955 18 : (List *) parse->havingQual);
3956 : }
3957 :
3958 18 : add_path(grouped_rel, path);
3959 18 : }
3960 :
3961 : /*
3962 : * create_ordinary_grouping_paths
3963 : *
3964 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
3965 : *
3966 : * We need to consider sorted and hashed aggregation in the same function,
3967 : * because otherwise (1) it would be harder to throw an appropriate error
3968 : * message if neither way works, and (2) we should not allow hashtable size
3969 : * considerations to dissuade us from using hashing if sorting is not possible.
3970 : *
3971 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
3972 : * function creates, or to NULL if it doesn't create one.
3973 : */
3974 : static void
3975 40494 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3976 : RelOptInfo *grouped_rel,
3977 : const AggClauseCosts *agg_costs,
3978 : grouping_sets_data *gd,
3979 : GroupPathExtraData *extra,
3980 : RelOptInfo **partially_grouped_rel_p)
3981 : {
3982 40494 : Path *cheapest_path = input_rel->cheapest_total_path;
3983 40494 : RelOptInfo *partially_grouped_rel = NULL;
3984 : double dNumGroups;
3985 40494 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
3986 :
3987 : /*
3988 : * If this is the topmost grouping relation or if the parent relation is
3989 : * doing some form of partitionwise aggregation, then we may be able to do
3990 : * it at this level also. However, if the input relation is not
3991 : * partitioned, partitionwise aggregate is impossible.
3992 : */
3993 40494 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
3994 2050 : IS_PARTITIONED_REL(input_rel))
3995 : {
3996 : /*
3997 : * If this is the topmost relation or if the parent relation is doing
3998 : * full partitionwise aggregation, then we can do full partitionwise
3999 : * aggregation provided that the GROUP BY clause contains all of the
4000 : * partitioning columns at this level and the collation used by GROUP
4001 : * BY matches the partitioning collation. Otherwise, we can do at
4002 : * most partial partitionwise aggregation. But if partial aggregation
4003 : * is not supported in general then we can't use it for partitionwise
4004 : * aggregation either.
4005 : *
4006 : * Check parse->groupClause not processed_groupClause, because it's
4007 : * okay if some of the partitioning columns were proved redundant.
4008 : */
4009 1160 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4010 556 : group_by_has_partkey(input_rel, extra->targetList,
4011 556 : root->parse->groupClause))
4012 320 : patype = PARTITIONWISE_AGGREGATE_FULL;
4013 284 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4014 242 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4015 : else
4016 42 : patype = PARTITIONWISE_AGGREGATE_NONE;
4017 : }
4018 :
4019 : /*
4020 : * Before generating paths for grouped_rel, we first generate any possible
4021 : * partially grouped paths; that way, later code can easily consider both
4022 : * parallel and non-parallel approaches to grouping.
4023 : */
4024 40494 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4025 : {
4026 : bool force_rel_creation;
4027 :
4028 : /*
4029 : * If we're doing partitionwise aggregation at this level, force
4030 : * creation of a partially_grouped_rel so we can add partitionwise
4031 : * paths to it.
4032 : */
4033 35472 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4034 :
4035 : partially_grouped_rel =
4036 35472 : create_partial_grouping_paths(root,
4037 : grouped_rel,
4038 : input_rel,
4039 : gd,
4040 : extra,
4041 : force_rel_creation);
4042 : }
4043 :
4044 : /* Set out parameter. */
4045 40494 : *partially_grouped_rel_p = partially_grouped_rel;
4046 :
4047 : /* Apply partitionwise aggregation technique, if possible. */
4048 40494 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4049 562 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4050 : partially_grouped_rel, agg_costs,
4051 : gd, patype, extra);
4052 :
4053 : /* If we are doing partial aggregation only, return. */
4054 40494 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4055 : {
4056 : Assert(partially_grouped_rel);
4057 :
4058 618 : if (partially_grouped_rel->pathlist)
4059 618 : set_cheapest(partially_grouped_rel);
4060 :
4061 618 : return;
4062 : }
4063 :
4064 : /* Gather any partially grouped partial paths. */
4065 39876 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4066 : {
4067 1482 : gather_grouping_paths(root, partially_grouped_rel);
4068 1482 : set_cheapest(partially_grouped_rel);
4069 : }
4070 :
4071 : /*
4072 : * Estimate number of groups.
4073 : */
4074 39876 : dNumGroups = get_number_of_groups(root,
4075 : cheapest_path->rows,
4076 : gd,
4077 : extra->targetList);
4078 :
4079 : /* Build final grouping paths */
4080 39876 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4081 : partially_grouped_rel, agg_costs, gd,
4082 : dNumGroups, extra);
4083 :
4084 : /* Give a helpful error if we failed to find any implementation */
4085 39876 : if (grouped_rel->pathlist == NIL)
4086 6 : ereport(ERROR,
4087 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4088 : errmsg("could not implement GROUP BY"),
4089 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4090 :
4091 : /*
4092 : * If there is an FDW that's responsible for all baserels of the query,
4093 : * let it consider adding ForeignPaths.
4094 : */
4095 39870 : if (grouped_rel->fdwroutine &&
4096 336 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4097 336 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4098 : input_rel, grouped_rel,
4099 : extra);
4100 :
4101 : /* Let extensions possibly add some more paths */
4102 39870 : if (create_upper_paths_hook)
4103 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4104 : input_rel, grouped_rel,
4105 : extra);
4106 : }
4107 :
4108 : /*
4109 : * For a given input path, consider the possible ways of doing grouping sets on
4110 : * it, by combinations of hashing and sorting. This can be called multiple
4111 : * times, so it's important that it not scribble on input. No result is
4112 : * returned, but any generated paths are added to grouped_rel.
4113 : */
4114 : static void
4115 1732 : consider_groupingsets_paths(PlannerInfo *root,
4116 : RelOptInfo *grouped_rel,
4117 : Path *path,
4118 : bool is_sorted,
4119 : bool can_hash,
4120 : grouping_sets_data *gd,
4121 : const AggClauseCosts *agg_costs,
4122 : double dNumGroups)
4123 : {
4124 1732 : Query *parse = root->parse;
4125 1732 : Size hash_mem_limit = get_hash_memory_limit();
4126 :
4127 : /*
4128 : * If we're not being offered sorted input, then only consider plans that
4129 : * can be done entirely by hashing.
4130 : *
4131 : * We can hash everything if it looks like it'll fit in hash_mem. But if
4132 : * the input is actually sorted despite not being advertised as such, we
4133 : * prefer to make use of that in order to use less memory.
4134 : *
4135 : * If none of the grouping sets are sortable, then ignore the hash_mem
4136 : * limit and generate a path anyway, since otherwise we'll just fail.
4137 : */
4138 1732 : if (!is_sorted)
4139 : {
4140 794 : List *new_rollups = NIL;
4141 794 : RollupData *unhashed_rollup = NULL;
4142 : List *sets_data;
4143 794 : List *empty_sets_data = NIL;
4144 794 : List *empty_sets = NIL;
4145 : ListCell *lc;
4146 794 : ListCell *l_start = list_head(gd->rollups);
4147 794 : AggStrategy strat = AGG_HASHED;
4148 : double hashsize;
4149 794 : double exclude_groups = 0.0;
4150 :
4151 : Assert(can_hash);
4152 :
4153 : /*
4154 : * If the input is coincidentally sorted usefully (which can happen
4155 : * even if is_sorted is false, since that only means that our caller
4156 : * has set up the sorting for us), then save some hashtable space by
4157 : * making use of that. But we need to watch out for degenerate cases:
4158 : *
4159 : * 1) If there are any empty grouping sets, then group_pathkeys might
4160 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4161 : * there will be a rollup containing only empty groups, and the
4162 : * pathkeys_contained_in test is vacuously true; this is ok.
4163 : *
4164 : * XXX: the above relies on the fact that group_pathkeys is generated
4165 : * from the first rollup. If we add the ability to consider multiple
4166 : * sort orders for grouping input, this assumption might fail.
4167 : *
4168 : * 2) If there are no empty sets and only unsortable sets, then the
4169 : * rollups list will be empty (and thus l_start == NULL), and
4170 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4171 : * pathkeys_contained_in test doesn't cause us to crash.
4172 : */
4173 1582 : if (l_start != NULL &&
4174 788 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4175 : {
4176 12 : unhashed_rollup = lfirst_node(RollupData, l_start);
4177 12 : exclude_groups = unhashed_rollup->numGroups;
4178 12 : l_start = lnext(gd->rollups, l_start);
4179 : }
4180 :
4181 794 : hashsize = estimate_hashagg_tablesize(root,
4182 : path,
4183 : agg_costs,
4184 : dNumGroups - exclude_groups);
4185 :
4186 : /*
4187 : * gd->rollups is empty if we have only unsortable columns to work
4188 : * with. Override hash_mem in that case; otherwise, we'll rely on the
4189 : * sorted-input case to generate usable mixed paths.
4190 : */
4191 794 : if (hashsize > hash_mem_limit && gd->rollups)
4192 18 : return; /* nope, won't fit */
4193 :
4194 : /*
4195 : * We need to burst the existing rollups list into individual grouping
4196 : * sets and recompute a groupClause for each set.
4197 : */
4198 776 : sets_data = list_copy(gd->unsortable_sets);
4199 :
4200 1968 : for_each_cell(lc, gd->rollups, l_start)
4201 : {
4202 1216 : RollupData *rollup = lfirst_node(RollupData, lc);
4203 :
4204 : /*
4205 : * If we find an unhashable rollup that's not been skipped by the
4206 : * "actually sorted" check above, we can't cope; we'd need sorted
4207 : * input (with a different sort order) but we can't get that here.
4208 : * So bail out; we'll get a valid path from the is_sorted case
4209 : * instead.
4210 : *
4211 : * The mere presence of empty grouping sets doesn't make a rollup
4212 : * unhashable (see preprocess_grouping_sets), we handle those
4213 : * specially below.
4214 : */
4215 1216 : if (!rollup->hashable)
4216 24 : return;
4217 :
4218 1192 : sets_data = list_concat(sets_data, rollup->gsets_data);
4219 : }
4220 3150 : foreach(lc, sets_data)
4221 : {
4222 2398 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4223 2398 : List *gset = gs->set;
4224 : RollupData *rollup;
4225 :
4226 2398 : if (gset == NIL)
4227 : {
4228 : /* Empty grouping sets can't be hashed. */
4229 484 : empty_sets_data = lappend(empty_sets_data, gs);
4230 484 : empty_sets = lappend(empty_sets, NIL);
4231 : }
4232 : else
4233 : {
4234 1914 : rollup = makeNode(RollupData);
4235 :
4236 1914 : rollup->groupClause = preprocess_groupclause(root, gset);
4237 1914 : rollup->gsets_data = list_make1(gs);
4238 1914 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4239 : rollup->gsets_data,
4240 : gd->tleref_to_colnum_map);
4241 1914 : rollup->numGroups = gs->numGroups;
4242 1914 : rollup->hashable = true;
4243 1914 : rollup->is_hashed = true;
4244 1914 : new_rollups = lappend(new_rollups, rollup);
4245 : }
4246 : }
4247 :
4248 : /*
4249 : * If we didn't find anything nonempty to hash, then bail. We'll
4250 : * generate a path from the is_sorted case.
4251 : */
4252 752 : if (new_rollups == NIL)
4253 0 : return;
4254 :
4255 : /*
4256 : * If there were empty grouping sets they should have been in the
4257 : * first rollup.
4258 : */
4259 : Assert(!unhashed_rollup || !empty_sets);
4260 :
4261 752 : if (unhashed_rollup)
4262 : {
4263 12 : new_rollups = lappend(new_rollups, unhashed_rollup);
4264 12 : strat = AGG_MIXED;
4265 : }
4266 740 : else if (empty_sets)
4267 : {
4268 436 : RollupData *rollup = makeNode(RollupData);
4269 :
4270 436 : rollup->groupClause = NIL;
4271 436 : rollup->gsets_data = empty_sets_data;
4272 436 : rollup->gsets = empty_sets;
4273 436 : rollup->numGroups = list_length(empty_sets);
4274 436 : rollup->hashable = false;
4275 436 : rollup->is_hashed = false;
4276 436 : new_rollups = lappend(new_rollups, rollup);
4277 436 : strat = AGG_MIXED;
4278 : }
4279 :
4280 752 : add_path(grouped_rel, (Path *)
4281 752 : create_groupingsets_path(root,
4282 : grouped_rel,
4283 : path,
4284 752 : (List *) parse->havingQual,
4285 : strat,
4286 : new_rollups,
4287 : agg_costs));
4288 752 : return;
4289 : }
4290 :
4291 : /*
4292 : * If we have sorted input but nothing we can do with it, bail.
4293 : */
4294 938 : if (gd->rollups == NIL)
4295 0 : return;
4296 :
4297 : /*
4298 : * Given sorted input, we try and make two paths: one sorted and one mixed
4299 : * sort/hash. (We need to try both because hashagg might be disabled, or
4300 : * some columns might not be sortable.)
4301 : *
4302 : * can_hash is passed in as false if some obstacle elsewhere (such as
4303 : * ordered aggs) means that we shouldn't consider hashing at all.
4304 : */
4305 938 : if (can_hash && gd->any_hashable)
4306 : {
4307 860 : List *rollups = NIL;
4308 860 : List *hash_sets = list_copy(gd->unsortable_sets);
4309 860 : double availspace = hash_mem_limit;
4310 : ListCell *lc;
4311 :
4312 : /*
4313 : * Account first for space needed for groups we can't sort at all.
4314 : */
4315 860 : availspace -= estimate_hashagg_tablesize(root,
4316 : path,
4317 : agg_costs,
4318 : gd->dNumHashGroups);
4319 :
4320 860 : if (availspace > 0 && list_length(gd->rollups) > 1)
4321 : {
4322 : double scale;
4323 444 : int num_rollups = list_length(gd->rollups);
4324 : int k_capacity;
4325 444 : int *k_weights = palloc(num_rollups * sizeof(int));
4326 444 : Bitmapset *hash_items = NULL;
4327 : int i;
4328 :
4329 : /*
4330 : * We treat this as a knapsack problem: the knapsack capacity
4331 : * represents hash_mem, the item weights are the estimated memory
4332 : * usage of the hashtables needed to implement a single rollup,
4333 : * and we really ought to use the cost saving as the item value;
4334 : * however, currently the costs assigned to sort nodes don't
4335 : * reflect the comparison costs well, and so we treat all items as
4336 : * of equal value (each rollup we hash instead saves us one sort).
4337 : *
4338 : * To use the discrete knapsack, we need to scale the values to a
4339 : * reasonably small bounded range. We choose to allow a 5% error
4340 : * margin; we have no more than 4096 rollups in the worst possible
4341 : * case, which with a 5% error margin will require a bit over 42MB
4342 : * of workspace. (Anyone wanting to plan queries that complex had
4343 : * better have the memory for it. In more reasonable cases, with
4344 : * no more than a couple of dozen rollups, the memory usage will
4345 : * be negligible.)
4346 : *
4347 : * k_capacity is naturally bounded, but we clamp the values for
4348 : * scale and weight (below) to avoid overflows or underflows (or
4349 : * uselessly trying to use a scale factor less than 1 byte).
4350 : */
4351 444 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4352 444 : k_capacity = (int) floor(availspace / scale);
4353 :
4354 : /*
4355 : * We leave the first rollup out of consideration since it's the
4356 : * one that matches the input sort order. We assign indexes "i"
4357 : * to only those entries considered for hashing; the second loop,
4358 : * below, must use the same condition.
4359 : */
4360 444 : i = 0;
4361 1128 : for_each_from(lc, gd->rollups, 1)
4362 : {
4363 684 : RollupData *rollup = lfirst_node(RollupData, lc);
4364 :
4365 684 : if (rollup->hashable)
4366 : {
4367 684 : double sz = estimate_hashagg_tablesize(root,
4368 : path,
4369 : agg_costs,
4370 : rollup->numGroups);
4371 :
4372 : /*
4373 : * If sz is enormous, but hash_mem (and hence scale) is
4374 : * small, avoid integer overflow here.
4375 : */
4376 684 : k_weights[i] = (int) Min(floor(sz / scale),
4377 : k_capacity + 1.0);
4378 684 : ++i;
4379 : }
4380 : }
4381 :
4382 : /*
4383 : * Apply knapsack algorithm; compute the set of items which
4384 : * maximizes the value stored (in this case the number of sorts
4385 : * saved) while keeping the total size (approximately) within
4386 : * capacity.
4387 : */
4388 444 : if (i > 0)
4389 444 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4390 :
4391 444 : if (!bms_is_empty(hash_items))
4392 : {
4393 444 : rollups = list_make1(linitial(gd->rollups));
4394 :
4395 444 : i = 0;
4396 1128 : for_each_from(lc, gd->rollups, 1)
4397 : {
4398 684 : RollupData *rollup = lfirst_node(RollupData, lc);
4399 :
4400 684 : if (rollup->hashable)
4401 : {
4402 684 : if (bms_is_member(i, hash_items))
4403 648 : hash_sets = list_concat(hash_sets,
4404 648 : rollup->gsets_data);
4405 : else
4406 36 : rollups = lappend(rollups, rollup);
4407 684 : ++i;
4408 : }
4409 : else
4410 0 : rollups = lappend(rollups, rollup);
4411 : }
4412 : }
4413 : }
4414 :
4415 860 : if (!rollups && hash_sets)
4416 24 : rollups = list_copy(gd->rollups);
4417 :
4418 1648 : foreach(lc, hash_sets)
4419 : {
4420 788 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4421 788 : RollupData *rollup = makeNode(RollupData);
4422 :
4423 : Assert(gs->set != NIL);
4424 :
4425 788 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4426 788 : rollup->gsets_data = list_make1(gs);
4427 788 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4428 : rollup->gsets_data,
4429 : gd->tleref_to_colnum_map);
4430 788 : rollup->numGroups = gs->numGroups;
4431 788 : rollup->hashable = true;
4432 788 : rollup->is_hashed = true;
4433 788 : rollups = lcons(rollup, rollups);
4434 : }
4435 :
4436 860 : if (rollups)
4437 : {
4438 468 : add_path(grouped_rel, (Path *)
4439 468 : create_groupingsets_path(root,
4440 : grouped_rel,
4441 : path,
4442 468 : (List *) parse->havingQual,
4443 : AGG_MIXED,
4444 : rollups,
4445 : agg_costs));
4446 : }
4447 : }
4448 :
4449 : /*
4450 : * Now try the simple sorted case.
4451 : */
4452 938 : if (!gd->unsortable_sets)
4453 908 : add_path(grouped_rel, (Path *)
4454 908 : create_groupingsets_path(root,
4455 : grouped_rel,
4456 : path,
4457 908 : (List *) parse->havingQual,
4458 : AGG_SORTED,
4459 : gd->rollups,
4460 : agg_costs));
4461 : }
4462 :
4463 : /*
4464 : * create_window_paths
4465 : *
4466 : * Build a new upperrel containing Paths for window-function evaluation.
4467 : *
4468 : * input_rel: contains the source-data Paths
4469 : * input_target: result of make_window_input_target
4470 : * output_target: what the topmost WindowAggPath should return
4471 : * wflists: result of find_window_functions
4472 : * activeWindows: result of select_active_windows
4473 : *
4474 : * Note: all Paths in input_rel are expected to return input_target.
4475 : */
4476 : static RelOptInfo *
4477 2378 : create_window_paths(PlannerInfo *root,
4478 : RelOptInfo *input_rel,
4479 : PathTarget *input_target,
4480 : PathTarget *output_target,
4481 : bool output_target_parallel_safe,
4482 : WindowFuncLists *wflists,
4483 : List *activeWindows)
4484 : {
4485 : RelOptInfo *window_rel;
4486 : ListCell *lc;
4487 :
4488 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4489 2378 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4490 :
4491 : /*
4492 : * If the input relation is not parallel-safe, then the window relation
4493 : * can't be parallel-safe, either. Otherwise, we need to examine the
4494 : * target list and active windows for non-parallel-safe constructs.
4495 : */
4496 2378 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4497 0 : is_parallel_safe(root, (Node *) activeWindows))
4498 0 : window_rel->consider_parallel = true;
4499 :
4500 : /*
4501 : * If the input rel belongs to a single FDW, so does the window rel.
4502 : */
4503 2378 : window_rel->serverid = input_rel->serverid;
4504 2378 : window_rel->userid = input_rel->userid;
4505 2378 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4506 2378 : window_rel->fdwroutine = input_rel->fdwroutine;
4507 :
4508 : /*
4509 : * Consider computing window functions starting from the existing
4510 : * cheapest-total path (which will likely require a sort) as well as any
4511 : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4512 : */
4513 5078 : foreach(lc, input_rel->pathlist)
4514 : {
4515 2700 : Path *path = (Path *) lfirst(lc);
4516 : int presorted_keys;
4517 :
4518 3022 : if (path == input_rel->cheapest_total_path ||
4519 322 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4520 140 : &presorted_keys) ||
4521 140 : presorted_keys > 0)
4522 2586 : create_one_window_path(root,
4523 : window_rel,
4524 : path,
4525 : input_target,
4526 : output_target,
4527 : wflists,
4528 : activeWindows);
4529 : }
4530 :
4531 : /*
4532 : * If there is an FDW that's responsible for all baserels of the query,
4533 : * let it consider adding ForeignPaths.
4534 : */
4535 2378 : if (window_rel->fdwroutine &&
4536 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4537 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4538 : input_rel, window_rel,
4539 : NULL);
4540 :
4541 : /* Let extensions possibly add some more paths */
4542 2378 : if (create_upper_paths_hook)
4543 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4544 : input_rel, window_rel, NULL);
4545 :
4546 : /* Now choose the best path(s) */
4547 2378 : set_cheapest(window_rel);
4548 :
4549 2378 : return window_rel;
4550 : }
4551 :
4552 : /*
4553 : * Stack window-function implementation steps atop the given Path, and
4554 : * add the result to window_rel.
4555 : *
4556 : * window_rel: upperrel to contain result
4557 : * path: input Path to use (must return input_target)
4558 : * input_target: result of make_window_input_target
4559 : * output_target: what the topmost WindowAggPath should return
4560 : * wflists: result of find_window_functions
4561 : * activeWindows: result of select_active_windows
4562 : */
4563 : static void
4564 2586 : create_one_window_path(PlannerInfo *root,
4565 : RelOptInfo *window_rel,
4566 : Path *path,
4567 : PathTarget *input_target,
4568 : PathTarget *output_target,
4569 : WindowFuncLists *wflists,
4570 : List *activeWindows)
4571 : {
4572 : PathTarget *window_target;
4573 : ListCell *l;
4574 2586 : List *topqual = NIL;
4575 :
4576 : /*
4577 : * Since each window clause could require a different sort order, we stack
4578 : * up a WindowAgg node for each clause, with sort steps between them as
4579 : * needed. (We assume that select_active_windows chose a good order for
4580 : * executing the clauses in.)
4581 : *
4582 : * input_target should contain all Vars and Aggs needed for the result.
4583 : * (In some cases we wouldn't need to propagate all of these all the way
4584 : * to the top, since they might only be needed as inputs to WindowFuncs.
4585 : * It's probably not worth trying to optimize that though.) It must also
4586 : * contain all window partitioning and sorting expressions, to ensure
4587 : * they're computed only once at the bottom of the stack (that's critical
4588 : * for volatile functions). As we climb up the stack, we'll add outputs
4589 : * for the WindowFuncs computed at each level.
4590 : */
4591 2586 : window_target = input_target;
4592 :
4593 5340 : foreach(l, activeWindows)
4594 : {
4595 2754 : WindowClause *wc = lfirst_node(WindowClause, l);
4596 : List *window_pathkeys;
4597 2754 : List *runcondition = NIL;
4598 : int presorted_keys;
4599 : bool is_sorted;
4600 : bool topwindow;
4601 : ListCell *lc2;
4602 :
4603 2754 : window_pathkeys = make_pathkeys_for_window(root,
4604 : wc,
4605 : root->processed_tlist);
4606 :
4607 2754 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4608 : path->pathkeys,
4609 : &presorted_keys);
4610 :
4611 : /* Sort if necessary */
4612 2754 : if (!is_sorted)
4613 : {
4614 : /*
4615 : * No presorted keys or incremental sort disabled, just perform a
4616 : * complete sort.
4617 : */
4618 2106 : if (presorted_keys == 0 || !enable_incremental_sort)
4619 2044 : path = (Path *) create_sort_path(root, window_rel,
4620 : path,
4621 : window_pathkeys,
4622 : -1.0);
4623 : else
4624 : {
4625 : /*
4626 : * Since we have presorted keys and incremental sort is
4627 : * enabled, just use incremental sort.
4628 : */
4629 62 : path = (Path *) create_incremental_sort_path(root,
4630 : window_rel,
4631 : path,
4632 : window_pathkeys,
4633 : presorted_keys,
4634 : -1.0);
4635 : }
4636 : }
4637 :
4638 2754 : if (lnext(activeWindows, l))
4639 : {
4640 : /*
4641 : * Add the current WindowFuncs to the output target for this
4642 : * intermediate WindowAggPath. We must copy window_target to
4643 : * avoid changing the previous path's target.
4644 : *
4645 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4646 : * we do need to account for the increase in tlist width.
4647 : */
4648 168 : int64 tuple_width = window_target->width;
4649 :
4650 168 : window_target = copy_pathtarget(window_target);
4651 384 : foreach(lc2, wflists->windowFuncs[wc->winref])
4652 : {
4653 216 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4654 :
4655 216 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4656 216 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4657 : }
4658 168 : window_target->width = clamp_width_est(tuple_width);
4659 : }
4660 : else
4661 : {
4662 : /* Install the goal target in the topmost WindowAgg */
4663 2586 : window_target = output_target;
4664 : }
4665 :
4666 : /* mark the final item in the list as the top-level window */
4667 2754 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4668 :
4669 : /*
4670 : * Collect the WindowFuncRunConditions from each WindowFunc and
4671 : * convert them into OpExprs
4672 : */
4673 6246 : foreach(lc2, wflists->windowFuncs[wc->winref])
4674 : {
4675 : ListCell *lc3;
4676 3492 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4677 :
4678 3672 : foreach(lc3, wfunc->runCondition)
4679 : {
4680 180 : WindowFuncRunCondition *wfuncrc =
4681 : lfirst_node(WindowFuncRunCondition, lc3);
4682 : Expr *opexpr;
4683 : Expr *leftop;
4684 : Expr *rightop;
4685 :
4686 180 : if (wfuncrc->wfunc_left)
4687 : {
4688 162 : leftop = (Expr *) copyObject(wfunc);
4689 162 : rightop = copyObject(wfuncrc->arg);
4690 : }
4691 : else
4692 : {
4693 18 : leftop = copyObject(wfuncrc->arg);
4694 18 : rightop = (Expr *) copyObject(wfunc);
4695 : }
4696 :
4697 180 : opexpr = make_opclause(wfuncrc->opno,
4698 : BOOLOID,
4699 : false,
4700 : leftop,
4701 : rightop,
4702 : InvalidOid,
4703 : wfuncrc->inputcollid);
4704 :
4705 180 : runcondition = lappend(runcondition, opexpr);
4706 :
4707 180 : if (!topwindow)
4708 24 : topqual = lappend(topqual, opexpr);
4709 : }
4710 : }
4711 :
4712 : path = (Path *)
4713 2754 : create_windowagg_path(root, window_rel, path, window_target,
4714 2754 : wflists->windowFuncs[wc->winref],
4715 : runcondition, wc,
4716 : topwindow ? topqual : NIL, topwindow);
4717 : }
4718 :
4719 2586 : add_path(window_rel, path);
4720 2586 : }
4721 :
4722 : /*
4723 : * create_distinct_paths
4724 : *
4725 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4726 : *
4727 : * input_rel: contains the source-data Paths
4728 : * target: the pathtarget for the result Paths to compute
4729 : *
4730 : * Note: input paths should already compute the desired pathtarget, since
4731 : * Sort/Unique won't project anything.
4732 : */
4733 : static RelOptInfo *
4734 2674 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4735 : PathTarget *target)
4736 : {
4737 : RelOptInfo *distinct_rel;
4738 :
4739 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4740 2674 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4741 :
4742 : /*
4743 : * We don't compute anything at this level, so distinct_rel will be
4744 : * parallel-safe if the input rel is parallel-safe. In particular, if
4745 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4746 : * output those expressions, and will not be parallel-safe unless those
4747 : * expressions are parallel-safe.
4748 : */
4749 2674 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4750 :
4751 : /*
4752 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4753 : */
4754 2674 : distinct_rel->serverid = input_rel->serverid;
4755 2674 : distinct_rel->userid = input_rel->userid;
4756 2674 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4757 2674 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4758 :
4759 : /* build distinct paths based on input_rel's pathlist */
4760 2674 : create_final_distinct_paths(root, input_rel, distinct_rel);
4761 :
4762 : /* now build distinct paths based on input_rel's partial_pathlist */
4763 2674 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4764 :
4765 : /* Give a helpful error if we failed to create any paths */
4766 2674 : if (distinct_rel->pathlist == NIL)
4767 0 : ereport(ERROR,
4768 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4769 : errmsg("could not implement DISTINCT"),
4770 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4771 :
4772 : /*
4773 : * If there is an FDW that's responsible for all baserels of the query,
4774 : * let it consider adding ForeignPaths.
4775 : */
4776 2674 : if (distinct_rel->fdwroutine &&
4777 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4778 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4779 : UPPERREL_DISTINCT,
4780 : input_rel,
4781 : distinct_rel,
4782 : NULL);
4783 :
4784 : /* Let extensions possibly add some more paths */
4785 2674 : if (create_upper_paths_hook)
4786 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4787 : distinct_rel, NULL);
4788 :
4789 : /* Now choose the best path(s) */
4790 2674 : set_cheapest(distinct_rel);
4791 :
4792 2674 : return distinct_rel;
4793 : }
4794 :
4795 : /*
4796 : * create_partial_distinct_paths
4797 : *
4798 : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4799 : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4800 : * paths on top and add a final unique/aggregate path to remove any duplicate
4801 : * produced from combining rows from parallel workers.
4802 : */
4803 : static void
4804 2674 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4805 : RelOptInfo *final_distinct_rel,
4806 : PathTarget *target)
4807 : {
4808 : RelOptInfo *partial_distinct_rel;
4809 : Query *parse;
4810 : List *distinctExprs;
4811 : double numDistinctRows;
4812 : Path *cheapest_partial_path;
4813 : ListCell *lc;
4814 :
4815 : /* nothing to do when there are no partial paths in the input rel */
4816 2674 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4817 2566 : return;
4818 :
4819 108 : parse = root->parse;
4820 :
4821 : /* can't do parallel DISTINCT ON */
4822 108 : if (parse->hasDistinctOn)
4823 0 : return;
4824 :
4825 108 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4826 : NULL);
4827 108 : partial_distinct_rel->reltarget = target;
4828 108 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4829 :
4830 : /*
4831 : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4832 : */
4833 108 : partial_distinct_rel->serverid = input_rel->serverid;
4834 108 : partial_distinct_rel->userid = input_rel->userid;
4835 108 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4836 108 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4837 :
4838 108 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4839 :
4840 108 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4841 : parse->targetList);
4842 :
4843 : /* estimate how many distinct rows we'll get from each worker */
4844 108 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4845 : cheapest_partial_path->rows,
4846 : NULL, NULL);
4847 :
4848 : /*
4849 : * Try sorting the cheapest path and incrementally sorting any paths with
4850 : * presorted keys and put a unique paths atop of those. We'll also
4851 : * attempt to reorder the required pathkeys to match the input path's
4852 : * pathkeys as much as possible, in hopes of avoiding a possible need to
4853 : * re-sort.
4854 : */
4855 108 : if (grouping_is_sortable(root->processed_distinctClause))
4856 : {
4857 234 : foreach(lc, input_rel->partial_pathlist)
4858 : {
4859 126 : Path *input_path = (Path *) lfirst(lc);
4860 : Path *sorted_path;
4861 126 : List *useful_pathkeys_list = NIL;
4862 :
4863 : useful_pathkeys_list =
4864 126 : get_useful_pathkeys_for_distinct(root,
4865 : root->distinct_pathkeys,
4866 : input_path->pathkeys);
4867 : Assert(list_length(useful_pathkeys_list) > 0);
4868 :
4869 390 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
4870 : {
4871 138 : sorted_path = make_ordered_path(root,
4872 : partial_distinct_rel,
4873 : input_path,
4874 : cheapest_partial_path,
4875 : useful_pathkeys,
4876 : -1.0);
4877 :
4878 138 : if (sorted_path == NULL)
4879 12 : continue;
4880 :
4881 : /*
4882 : * An empty distinct_pathkeys means all tuples have the same
4883 : * value for the DISTINCT clause. See
4884 : * create_final_distinct_paths()
4885 : */
4886 126 : if (root->distinct_pathkeys == NIL)
4887 : {
4888 : Node *limitCount;
4889 :
4890 6 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
4891 : sizeof(int64),
4892 : Int64GetDatum(1), false,
4893 : FLOAT8PASSBYVAL);
4894 :
4895 : /*
4896 : * Apply a LimitPath onto the partial path to restrict the
4897 : * tuples from each worker to 1.
4898 : * create_final_distinct_paths will need to apply an
4899 : * additional LimitPath to restrict this to a single row
4900 : * after the Gather node. If the query already has a
4901 : * LIMIT clause, then we could end up with three Limit
4902 : * nodes in the final plan. Consolidating the top two of
4903 : * these could be done, but does not seem worth troubling
4904 : * over.
4905 : */
4906 6 : add_partial_path(partial_distinct_rel, (Path *)
4907 6 : create_limit_path(root, partial_distinct_rel,
4908 : sorted_path,
4909 : NULL,
4910 : limitCount,
4911 : LIMIT_OPTION_COUNT,
4912 : 0, 1));
4913 : }
4914 : else
4915 : {
4916 120 : add_partial_path(partial_distinct_rel, (Path *)
4917 120 : create_upper_unique_path(root, partial_distinct_rel,
4918 : sorted_path,
4919 120 : list_length(root->distinct_pathkeys),
4920 : numDistinctRows));
4921 : }
4922 : }
4923 : }
4924 : }
4925 :
4926 : /*
4927 : * Now try hash aggregate paths, if enabled and hashing is possible. Since
4928 : * we're not on the hook to ensure we do our best to create at least one
4929 : * path here, we treat enable_hashagg as a hard off-switch rather than the
4930 : * slightly softer variant in create_final_distinct_paths.
4931 : */
4932 108 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
4933 : {
4934 78 : add_partial_path(partial_distinct_rel, (Path *)
4935 78 : create_agg_path(root,
4936 : partial_distinct_rel,
4937 : cheapest_partial_path,
4938 : cheapest_partial_path->pathtarget,
4939 : AGG_HASHED,
4940 : AGGSPLIT_SIMPLE,
4941 : root->processed_distinctClause,
4942 : NIL,
4943 : NULL,
4944 : numDistinctRows));
4945 : }
4946 :
4947 : /*
4948 : * If there is an FDW that's responsible for all baserels of the query,
4949 : * let it consider adding ForeignPaths.
4950 : */
4951 108 : if (partial_distinct_rel->fdwroutine &&
4952 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
4953 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4954 : UPPERREL_PARTIAL_DISTINCT,
4955 : input_rel,
4956 : partial_distinct_rel,
4957 : NULL);
4958 :
4959 : /* Let extensions possibly add some more partial paths */
4960 108 : if (create_upper_paths_hook)
4961 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
4962 : input_rel, partial_distinct_rel, NULL);
4963 :
4964 108 : if (partial_distinct_rel->partial_pathlist != NIL)
4965 : {
4966 108 : generate_useful_gather_paths(root, partial_distinct_rel, true);
4967 108 : set_cheapest(partial_distinct_rel);
4968 :
4969 : /*
4970 : * Finally, create paths to distinctify the final result. This step
4971 : * is needed to remove any duplicates due to combining rows from
4972 : * parallel workers.
4973 : */
4974 108 : create_final_distinct_paths(root, partial_distinct_rel,
4975 : final_distinct_rel);
4976 : }
4977 : }
4978 :
4979 : /*
4980 : * create_final_distinct_paths
4981 : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
4982 : *
4983 : * input_rel: contains the source-data paths
4984 : * distinct_rel: destination relation for storing created paths
4985 : */
4986 : static RelOptInfo *
4987 2782 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4988 : RelOptInfo *distinct_rel)
4989 : {
4990 2782 : Query *parse = root->parse;
4991 2782 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4992 : double numDistinctRows;
4993 : bool allow_hash;
4994 :
4995 : /* Estimate number of distinct rows there will be */
4996 2782 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4997 2708 : root->hasHavingQual)
4998 : {
4999 : /*
5000 : * If there was grouping or aggregation, use the number of input rows
5001 : * as the estimated number of DISTINCT rows (ie, assume the input is
5002 : * already mostly unique).
5003 : */
5004 74 : numDistinctRows = cheapest_input_path->rows;
5005 : }
5006 : else
5007 : {
5008 : /*
5009 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5010 : */
5011 : List *distinctExprs;
5012 :
5013 2708 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5014 : parse->targetList);
5015 2708 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5016 : cheapest_input_path->rows,
5017 : NULL, NULL);
5018 : }
5019 :
5020 : /*
5021 : * Consider sort-based implementations of DISTINCT, if possible.
5022 : */
5023 2782 : if (grouping_is_sortable(root->processed_distinctClause))
5024 : {
5025 : /*
5026 : * Firstly, if we have any adequately-presorted paths, just stick a
5027 : * Unique node on those. We also, consider doing an explicit sort of
5028 : * the cheapest input path and Unique'ing that. If any paths have
5029 : * presorted keys then we'll create an incremental sort atop of those
5030 : * before adding a unique node on the top. We'll also attempt to
5031 : * reorder the required pathkeys to match the input path's pathkeys as
5032 : * much as possible, in hopes of avoiding a possible need to re-sort.
5033 : *
5034 : * When we have DISTINCT ON, we must sort by the more rigorous of
5035 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5036 : * Also, if we do have to do an explicit sort, we might as well use
5037 : * the more rigorous ordering to avoid a second sort later. (Note
5038 : * that the parser will have ensured that one clause is a prefix of
5039 : * the other.)
5040 : */
5041 : List *needed_pathkeys;
5042 : ListCell *lc;
5043 2776 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5044 :
5045 3014 : if (parse->hasDistinctOn &&
5046 238 : list_length(root->distinct_pathkeys) <
5047 238 : list_length(root->sort_pathkeys))
5048 54 : needed_pathkeys = root->sort_pathkeys;
5049 : else
5050 2722 : needed_pathkeys = root->distinct_pathkeys;
5051 :
5052 7156 : foreach(lc, input_rel->pathlist)
5053 : {
5054 4380 : Path *input_path = (Path *) lfirst(lc);
5055 : Path *sorted_path;
5056 4380 : List *useful_pathkeys_list = NIL;
5057 :
5058 : useful_pathkeys_list =
5059 4380 : get_useful_pathkeys_for_distinct(root,
5060 : needed_pathkeys,
5061 : input_path->pathkeys);
5062 : Assert(list_length(useful_pathkeys_list) > 0);
5063 :
5064 13626 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
5065 : {
5066 4866 : sorted_path = make_ordered_path(root,
5067 : distinct_rel,
5068 : input_path,
5069 : cheapest_input_path,
5070 : useful_pathkeys,
5071 : limittuples);
5072 :
5073 4866 : if (sorted_path == NULL)
5074 544 : continue;
5075 :
5076 : /*
5077 : * distinct_pathkeys may have become empty if all of the
5078 : * pathkeys were determined to be redundant. If all of the
5079 : * pathkeys are redundant then each DISTINCT target must only
5080 : * allow a single value, therefore all resulting tuples must
5081 : * be identical (or at least indistinguishable by an equality
5082 : * check). We can uniquify these tuples simply by just taking
5083 : * the first tuple. All we do here is add a path to do "LIMIT
5084 : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5085 : * still have a non-NIL sort_pathkeys list, so we must still
5086 : * only do this with paths which are correctly sorted by
5087 : * sort_pathkeys.
5088 : */
5089 4322 : if (root->distinct_pathkeys == NIL)
5090 : {
5091 : Node *limitCount;
5092 :
5093 106 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5094 : sizeof(int64),
5095 : Int64GetDatum(1), false,
5096 : FLOAT8PASSBYVAL);
5097 :
5098 : /*
5099 : * If the query already has a LIMIT clause, then we could
5100 : * end up with a duplicate LimitPath in the final plan.
5101 : * That does not seem worth troubling over too much.
5102 : */
5103 106 : add_path(distinct_rel, (Path *)
5104 106 : create_limit_path(root, distinct_rel, sorted_path,
5105 : NULL, limitCount,
5106 : LIMIT_OPTION_COUNT, 0, 1));
5107 : }
5108 : else
5109 : {
5110 4216 : add_path(distinct_rel, (Path *)
5111 4216 : create_upper_unique_path(root, distinct_rel,
5112 : sorted_path,
5113 4216 : list_length(root->distinct_pathkeys),
5114 : numDistinctRows));
5115 : }
5116 : }
5117 : }
5118 : }
5119 :
5120 : /*
5121 : * Consider hash-based implementations of DISTINCT, if possible.
5122 : *
5123 : * If we were not able to make any other types of path, we *must* hash or
5124 : * die trying. If we do have other choices, there are two things that
5125 : * should prevent selection of hashing: if the query uses DISTINCT ON
5126 : * (because it won't really have the expected behavior if we hash), or if
5127 : * enable_hashagg is off.
5128 : *
5129 : * Note: grouping_is_hashable() is much more expensive to check than the
5130 : * other gating conditions, so we want to do it last.
5131 : */
5132 2782 : if (distinct_rel->pathlist == NIL)
5133 6 : allow_hash = true; /* we have no alternatives */
5134 2776 : else if (parse->hasDistinctOn || !enable_hashagg)
5135 388 : allow_hash = false; /* policy-based decision not to hash */
5136 : else
5137 2388 : allow_hash = true; /* default */
5138 :
5139 2782 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5140 : {
5141 : /* Generate hashed aggregate path --- no sort needed */
5142 2394 : add_path(distinct_rel, (Path *)
5143 2394 : create_agg_path(root,
5144 : distinct_rel,
5145 : cheapest_input_path,
5146 : cheapest_input_path->pathtarget,
5147 : AGG_HASHED,
5148 : AGGSPLIT_SIMPLE,
5149 : root->processed_distinctClause,
5150 : NIL,
5151 : NULL,
5152 : numDistinctRows));
5153 : }
5154 :
5155 2782 : return distinct_rel;
5156 : }
5157 :
5158 : /*
5159 : * get_useful_pathkeys_for_distinct
5160 : * Get useful orderings of pathkeys for distinctClause by reordering
5161 : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5162 : *
5163 : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5164 : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5165 : */
5166 : static List *
5167 4506 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5168 : List *path_pathkeys)
5169 : {
5170 4506 : List *useful_pathkeys_list = NIL;
5171 4506 : List *useful_pathkeys = NIL;
5172 :
5173 : /* always include the given 'needed_pathkeys' */
5174 4506 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5175 : needed_pathkeys);
5176 :
5177 4506 : if (!enable_distinct_reordering)
5178 0 : return useful_pathkeys_list;
5179 :
5180 : /*
5181 : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5182 : * that match 'needed_pathkeys', but only up to the longest matching
5183 : * prefix.
5184 : *
5185 : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5186 : * list matches initial distinctClause pathkeys; otherwise, it won't have
5187 : * the desired behavior.
5188 : */
5189 11020 : foreach_node(PathKey, pathkey, path_pathkeys)
5190 : {
5191 : /*
5192 : * The PathKey nodes are canonical, so they can be checked for
5193 : * equality by simple pointer comparison.
5194 : */
5195 2036 : if (!list_member_ptr(needed_pathkeys, pathkey))
5196 10 : break;
5197 2026 : if (root->parse->hasDistinctOn &&
5198 202 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5199 18 : break;
5200 :
5201 2008 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5202 : }
5203 :
5204 : /* If no match at all, no point in reordering needed_pathkeys */
5205 4506 : if (useful_pathkeys == NIL)
5206 2762 : return useful_pathkeys_list;
5207 :
5208 : /*
5209 : * If not full match, the resulting pathkey list is not useful without
5210 : * incremental sort.
5211 : */
5212 1744 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5213 894 : !enable_incremental_sort)
5214 60 : return useful_pathkeys_list;
5215 :
5216 : /* Append the remaining PathKey nodes in needed_pathkeys */
5217 1684 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5218 : needed_pathkeys);
5219 :
5220 : /*
5221 : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5222 : * just drop it.
5223 : */
5224 1684 : if (compare_pathkeys(needed_pathkeys,
5225 : useful_pathkeys) == PATHKEYS_EQUAL)
5226 1186 : return useful_pathkeys_list;
5227 :
5228 498 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5229 : useful_pathkeys);
5230 :
5231 498 : return useful_pathkeys_list;
5232 : }
5233 :
5234 : /*
5235 : * create_ordered_paths
5236 : *
5237 : * Build a new upperrel containing Paths for ORDER BY evaluation.
5238 : *
5239 : * All paths in the result must satisfy the ORDER BY ordering.
5240 : * The only new paths we need consider are an explicit full sort
5241 : * and incremental sort on the cheapest-total existing path.
5242 : *
5243 : * input_rel: contains the source-data Paths
5244 : * target: the output tlist the result Paths must emit
5245 : * limit_tuples: estimated bound on the number of output tuples,
5246 : * or -1 if no LIMIT or couldn't estimate
5247 : *
5248 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5249 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5250 : */
5251 : static RelOptInfo *
5252 75838 : create_ordered_paths(PlannerInfo *root,
5253 : RelOptInfo *input_rel,
5254 : PathTarget *target,
5255 : bool target_parallel_safe,
5256 : double limit_tuples)
5257 : {
5258 75838 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5259 : RelOptInfo *ordered_rel;
5260 : ListCell *lc;
5261 :
5262 : /* For now, do all work in the (ORDERED, NULL) upperrel */
5263 75838 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5264 :
5265 : /*
5266 : * If the input relation is not parallel-safe, then the ordered relation
5267 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5268 : * target list is parallel-safe.
5269 : */
5270 75838 : if (input_rel->consider_parallel && target_parallel_safe)
5271 52556 : ordered_rel->consider_parallel = true;
5272 :
5273 : /*
5274 : * If the input rel belongs to a single FDW, so does the ordered_rel.
5275 : */
5276 75838 : ordered_rel->serverid = input_rel->serverid;
5277 75838 : ordered_rel->userid = input_rel->userid;
5278 75838 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5279 75838 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5280 :
5281 192406 : foreach(lc, input_rel->pathlist)
5282 : {
5283 116568 : Path *input_path = (Path *) lfirst(lc);
5284 : Path *sorted_path;
5285 : bool is_sorted;
5286 : int presorted_keys;
5287 :
5288 116568 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5289 : input_path->pathkeys, &presorted_keys);
5290 :
5291 116568 : if (is_sorted)
5292 42904 : sorted_path = input_path;
5293 : else
5294 : {
5295 : /*
5296 : * Try at least sorting the cheapest path and also try
5297 : * incrementally sorting any path which is partially sorted
5298 : * already (no need to deal with paths which have presorted keys
5299 : * when incremental sort is disabled unless it's the cheapest
5300 : * input path).
5301 : */
5302 73664 : if (input_path != cheapest_input_path &&
5303 6334 : (presorted_keys == 0 || !enable_incremental_sort))
5304 1878 : continue;
5305 :
5306 : /*
5307 : * We've no need to consider both a sort and incremental sort.
5308 : * We'll just do a sort if there are no presorted keys and an
5309 : * incremental sort when there are presorted keys.
5310 : */
5311 71786 : if (presorted_keys == 0 || !enable_incremental_sort)
5312 66534 : sorted_path = (Path *) create_sort_path(root,
5313 : ordered_rel,
5314 : input_path,
5315 : root->sort_pathkeys,
5316 : limit_tuples);
5317 : else
5318 5252 : sorted_path = (Path *) create_incremental_sort_path(root,
5319 : ordered_rel,
5320 : input_path,
5321 : root->sort_pathkeys,
5322 : presorted_keys,
5323 : limit_tuples);
5324 : }
5325 :
5326 : /*
5327 : * If the pathtarget of the result path has different expressions from
5328 : * the target to be applied, a projection step is needed.
5329 : */
5330 114690 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5331 294 : sorted_path = apply_projection_to_path(root, ordered_rel,
5332 : sorted_path, target);
5333 :
5334 114690 : add_path(ordered_rel, sorted_path);
5335 : }
5336 :
5337 : /*
5338 : * generate_gather_paths() will have already generated a simple Gather
5339 : * path for the best parallel path, if any, and the loop above will have
5340 : * considered sorting it. Similarly, generate_gather_paths() will also
5341 : * have generated order-preserving Gather Merge plans which can be used
5342 : * without sorting if they happen to match the sort_pathkeys, and the loop
5343 : * above will have handled those as well. However, there's one more
5344 : * possibility: it may make sense to sort the cheapest partial path or
5345 : * incrementally sort any partial path that is partially sorted according
5346 : * to the required output order and then use Gather Merge.
5347 : */
5348 75838 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5349 52418 : input_rel->partial_pathlist != NIL)
5350 : {
5351 : Path *cheapest_partial_path;
5352 :
5353 2216 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5354 :
5355 4638 : foreach(lc, input_rel->partial_pathlist)
5356 : {
5357 2422 : Path *input_path = (Path *) lfirst(lc);
5358 : Path *sorted_path;
5359 : bool is_sorted;
5360 : int presorted_keys;
5361 : double total_groups;
5362 :
5363 2422 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5364 : input_path->pathkeys,
5365 : &presorted_keys);
5366 :
5367 2422 : if (is_sorted)
5368 182 : continue;
5369 :
5370 : /*
5371 : * Try at least sorting the cheapest path and also try
5372 : * incrementally sorting any path which is partially sorted
5373 : * already (no need to deal with paths which have presorted keys
5374 : * when incremental sort is disabled unless it's the cheapest
5375 : * partial path).
5376 : */
5377 2240 : if (input_path != cheapest_partial_path &&
5378 42 : (presorted_keys == 0 || !enable_incremental_sort))
5379 0 : continue;
5380 :
5381 : /*
5382 : * We've no need to consider both a sort and incremental sort.
5383 : * We'll just do a sort if there are no presorted keys and an
5384 : * incremental sort when there are presorted keys.
5385 : */
5386 2240 : if (presorted_keys == 0 || !enable_incremental_sort)
5387 2180 : sorted_path = (Path *) create_sort_path(root,
5388 : ordered_rel,
5389 : input_path,
5390 : root->sort_pathkeys,
5391 : limit_tuples);
5392 : else
5393 60 : sorted_path = (Path *) create_incremental_sort_path(root,
5394 : ordered_rel,
5395 : input_path,
5396 : root->sort_pathkeys,
5397 : presorted_keys,
5398 : limit_tuples);
5399 2240 : total_groups = compute_gather_rows(sorted_path);
5400 : sorted_path = (Path *)
5401 2240 : create_gather_merge_path(root, ordered_rel,
5402 : sorted_path,
5403 : sorted_path->pathtarget,
5404 : root->sort_pathkeys, NULL,
5405 : &total_groups);
5406 :
5407 : /*
5408 : * If the pathtarget of the result path has different expressions
5409 : * from the target to be applied, a projection step is needed.
5410 : */
5411 2240 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5412 6 : sorted_path = apply_projection_to_path(root, ordered_rel,
5413 : sorted_path, target);
5414 :
5415 2240 : add_path(ordered_rel, sorted_path);
5416 : }
5417 : }
5418 :
5419 : /*
5420 : * If there is an FDW that's responsible for all baserels of the query,
5421 : * let it consider adding ForeignPaths.
5422 : */
5423 75838 : if (ordered_rel->fdwroutine &&
5424 384 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5425 370 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5426 : input_rel, ordered_rel,
5427 : NULL);
5428 :
5429 : /* Let extensions possibly add some more paths */
5430 75838 : if (create_upper_paths_hook)
5431 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5432 : input_rel, ordered_rel, NULL);
5433 :
5434 : /*
5435 : * No need to bother with set_cheapest here; grouping_planner does not
5436 : * need us to do it.
5437 : */
5438 : Assert(ordered_rel->pathlist != NIL);
5439 :
5440 75838 : return ordered_rel;
5441 : }
5442 :
5443 :
5444 : /*
5445 : * make_group_input_target
5446 : * Generate appropriate PathTarget for initial input to grouping nodes.
5447 : *
5448 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5449 : * the query's final targetlist; for example, it certainly can't emit any
5450 : * aggregate function calls. This routine generates the correct target
5451 : * for the scan/join subplan.
5452 : *
5453 : * The query target list passed from the parser already contains entries
5454 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5455 : * for variables used only in HAVING clauses; so we need to add those
5456 : * variables to the subplan target list. Also, we flatten all expressions
5457 : * except GROUP BY items into their component variables; other expressions
5458 : * will be computed by the upper plan nodes rather than by the subplan.
5459 : * For example, given a query like
5460 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5461 : * we want to pass this targetlist to the subplan:
5462 : * a+b,c,d
5463 : * where the a+b target will be used by the Sort/Group steps, and the
5464 : * other targets will be used for computing the final results.
5465 : *
5466 : * 'final_target' is the query's final target list (in PathTarget form)
5467 : *
5468 : * The result is the PathTarget to be computed by the Paths returned from
5469 : * query_planner().
5470 : */
5471 : static PathTarget *
5472 39018 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5473 : {
5474 39018 : Query *parse = root->parse;
5475 : PathTarget *input_target;
5476 : List *non_group_cols;
5477 : List *non_group_vars;
5478 : int i;
5479 : ListCell *lc;
5480 :
5481 : /*
5482 : * We must build a target containing all grouping columns, plus any other
5483 : * Vars mentioned in the query's targetlist and HAVING qual.
5484 : */
5485 39018 : input_target = create_empty_pathtarget();
5486 39018 : non_group_cols = NIL;
5487 :
5488 39018 : i = 0;
5489 94908 : foreach(lc, final_target->exprs)
5490 : {
5491 55890 : Expr *expr = (Expr *) lfirst(lc);
5492 55890 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5493 :
5494 64636 : if (sgref && root->processed_groupClause &&
5495 8746 : get_sortgroupref_clause_noerr(sgref,
5496 : root->processed_groupClause) != NULL)
5497 : {
5498 : /*
5499 : * It's a grouping column, so add it to the input target as-is.
5500 : *
5501 : * Note that the target is logically below the grouping step. So
5502 : * with grouping sets we need to remove the RT index of the
5503 : * grouping step if there is any from the target expression.
5504 : */
5505 7004 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5506 : {
5507 : Assert(root->group_rtindex > 0);
5508 : expr = (Expr *)
5509 1824 : remove_nulling_relids((Node *) expr,
5510 1824 : bms_make_singleton(root->group_rtindex),
5511 : NULL);
5512 : }
5513 7004 : add_column_to_pathtarget(input_target, expr, sgref);
5514 : }
5515 : else
5516 : {
5517 : /*
5518 : * Non-grouping column, so just remember the expression for later
5519 : * call to pull_var_clause.
5520 : */
5521 48886 : non_group_cols = lappend(non_group_cols, expr);
5522 : }
5523 :
5524 55890 : i++;
5525 : }
5526 :
5527 : /*
5528 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5529 : */
5530 39018 : if (parse->havingQual)
5531 880 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5532 :
5533 : /*
5534 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5535 : * add them to the input target if not already present. (A Var used
5536 : * directly as a GROUP BY item will be present already.) Note this
5537 : * includes Vars used in resjunk items, so we are covering the needs of
5538 : * ORDER BY and window specifications. Vars used within Aggrefs and
5539 : * WindowFuncs will be pulled out here, too.
5540 : *
5541 : * Note that the target is logically below the grouping step. So with
5542 : * grouping sets we need to remove the RT index of the grouping step if
5543 : * there is any from the non-group Vars.
5544 : */
5545 39018 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5546 : PVC_RECURSE_AGGREGATES |
5547 : PVC_RECURSE_WINDOWFUNCS |
5548 : PVC_INCLUDE_PLACEHOLDERS);
5549 39018 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5550 : {
5551 : Assert(root->group_rtindex > 0);
5552 : non_group_vars = (List *)
5553 830 : remove_nulling_relids((Node *) non_group_vars,
5554 830 : bms_make_singleton(root->group_rtindex),
5555 : NULL);
5556 : }
5557 39018 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5558 :
5559 : /* clean up cruft */
5560 39018 : list_free(non_group_vars);
5561 39018 : list_free(non_group_cols);
5562 :
5563 : /* XXX this causes some redundant cost calculation ... */
5564 39018 : return set_pathtarget_cost_width(root, input_target);
5565 : }
5566 :
5567 : /*
5568 : * make_partial_grouping_target
5569 : * Generate appropriate PathTarget for output of partial aggregate
5570 : * (or partial grouping, if there are no aggregates) nodes.
5571 : *
5572 : * A partial aggregation node needs to emit all the same aggregates that
5573 : * a regular aggregation node would, plus any aggregates used in HAVING;
5574 : * except that the Aggref nodes should be marked as partial aggregates.
5575 : *
5576 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5577 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5578 : * these would be Vars that are grouped by or used in grouping expressions.)
5579 : *
5580 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5581 : * havingQual represents the HAVING clause.
5582 : */
5583 : static PathTarget *
5584 2198 : make_partial_grouping_target(PlannerInfo *root,
5585 : PathTarget *grouping_target,
5586 : Node *havingQual)
5587 : {
5588 : PathTarget *partial_target;
5589 : List *non_group_cols;
5590 : List *non_group_exprs;
5591 : int i;
5592 : ListCell *lc;
5593 :
5594 2198 : partial_target = create_empty_pathtarget();
5595 2198 : non_group_cols = NIL;
5596 :
5597 2198 : i = 0;
5598 7814 : foreach(lc, grouping_target->exprs)
5599 : {
5600 5616 : Expr *expr = (Expr *) lfirst(lc);
5601 5616 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5602 :
5603 9436 : if (sgref && root->processed_groupClause &&
5604 3820 : get_sortgroupref_clause_noerr(sgref,
5605 : root->processed_groupClause) != NULL)
5606 : {
5607 : /*
5608 : * It's a grouping column, so add it to the partial_target as-is.
5609 : * (This allows the upper agg step to repeat the grouping calcs.)
5610 : */
5611 1906 : add_column_to_pathtarget(partial_target, expr, sgref);
5612 : }
5613 : else
5614 : {
5615 : /*
5616 : * Non-grouping column, so just remember the expression for later
5617 : * call to pull_var_clause.
5618 : */
5619 3710 : non_group_cols = lappend(non_group_cols, expr);
5620 : }
5621 :
5622 5616 : i++;
5623 : }
5624 :
5625 : /*
5626 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5627 : */
5628 2198 : if (havingQual)
5629 824 : non_group_cols = lappend(non_group_cols, havingQual);
5630 :
5631 : /*
5632 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5633 : * non-group cols (plus HAVING), and add them to the partial_target if not
5634 : * already present. (An expression used directly as a GROUP BY item will
5635 : * be present already.) Note this includes Vars used in resjunk items, so
5636 : * we are covering the needs of ORDER BY and window specifications.
5637 : */
5638 2198 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5639 : PVC_INCLUDE_AGGREGATES |
5640 : PVC_RECURSE_WINDOWFUNCS |
5641 : PVC_INCLUDE_PLACEHOLDERS);
5642 :
5643 2198 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5644 :
5645 : /*
5646 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5647 : * are at the top level of the target list, so we can just scan the list
5648 : * rather than recursing through the expression trees.
5649 : */
5650 8386 : foreach(lc, partial_target->exprs)
5651 : {
5652 6188 : Aggref *aggref = (Aggref *) lfirst(lc);
5653 :
5654 6188 : if (IsA(aggref, Aggref))
5655 : {
5656 : Aggref *newaggref;
5657 :
5658 : /*
5659 : * We shouldn't need to copy the substructure of the Aggref node,
5660 : * but flat-copy the node itself to avoid damaging other trees.
5661 : */
5662 4252 : newaggref = makeNode(Aggref);
5663 4252 : memcpy(newaggref, aggref, sizeof(Aggref));
5664 :
5665 : /* For now, assume serialization is required */
5666 4252 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5667 :
5668 4252 : lfirst(lc) = newaggref;
5669 : }
5670 : }
5671 :
5672 : /* clean up cruft */
5673 2198 : list_free(non_group_exprs);
5674 2198 : list_free(non_group_cols);
5675 :
5676 : /* XXX this causes some redundant cost calculation ... */
5677 2198 : return set_pathtarget_cost_width(root, partial_target);
5678 : }
5679 :
5680 : /*
5681 : * mark_partial_aggref
5682 : * Adjust an Aggref to make it represent a partial-aggregation step.
5683 : *
5684 : * The Aggref node is modified in-place; caller must do any copying required.
5685 : */
5686 : void
5687 7060 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5688 : {
5689 : /* aggtranstype should be computed by this point */
5690 : Assert(OidIsValid(agg->aggtranstype));
5691 : /* ... but aggsplit should still be as the parser left it */
5692 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5693 :
5694 : /* Mark the Aggref with the intended partial-aggregation mode */
5695 7060 : agg->aggsplit = aggsplit;
5696 :
5697 : /*
5698 : * Adjust result type if needed. Normally, a partial aggregate returns
5699 : * the aggregate's transition type; but if that's INTERNAL and we're
5700 : * serializing, it returns BYTEA instead.
5701 : */
5702 7060 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5703 : {
5704 5656 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5705 242 : agg->aggtype = BYTEAOID;
5706 : else
5707 5414 : agg->aggtype = agg->aggtranstype;
5708 : }
5709 7060 : }
5710 :
5711 : /*
5712 : * postprocess_setop_tlist
5713 : * Fix up targetlist returned by plan_set_operations().
5714 : *
5715 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5716 : * NOTE: this would not be good enough if we supported resjunk sort keys
5717 : * for results of set operations --- then, we'd need to project a whole
5718 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5719 : * find any resjunk columns in orig_tlist.
5720 : */
5721 : static List *
5722 6160 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5723 : {
5724 : ListCell *l;
5725 6160 : ListCell *orig_tlist_item = list_head(orig_tlist);
5726 :
5727 24150 : foreach(l, new_tlist)
5728 : {
5729 17990 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5730 : TargetEntry *orig_tle;
5731 :
5732 : /* ignore resjunk columns in setop result */
5733 17990 : if (new_tle->resjunk)
5734 0 : continue;
5735 :
5736 : Assert(orig_tlist_item != NULL);
5737 17990 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5738 17990 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5739 17990 : if (orig_tle->resjunk) /* should not happen */
5740 0 : elog(ERROR, "resjunk output columns are not implemented");
5741 : Assert(new_tle->resno == orig_tle->resno);
5742 17990 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5743 : }
5744 6160 : if (orig_tlist_item != NULL)
5745 0 : elog(ERROR, "resjunk output columns are not implemented");
5746 6160 : return new_tlist;
5747 : }
5748 :
5749 : /*
5750 : * optimize_window_clauses
5751 : * Call each WindowFunc's prosupport function to see if we're able to
5752 : * make any adjustments to any of the WindowClause's so that the executor
5753 : * can execute the window functions in a more optimal way.
5754 : *
5755 : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5756 : * may allow more things to be done here in the future.
5757 : */
5758 : static void
5759 2378 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5760 : {
5761 2378 : List *windowClause = root->parse->windowClause;
5762 : ListCell *lc;
5763 :
5764 4984 : foreach(lc, windowClause)
5765 : {
5766 2606 : WindowClause *wc = lfirst_node(WindowClause, lc);
5767 : ListCell *lc2;
5768 2606 : int optimizedFrameOptions = 0;
5769 :
5770 : Assert(wc->winref <= wflists->maxWinRef);
5771 :
5772 : /* skip any WindowClauses that have no WindowFuncs */
5773 2606 : if (wflists->windowFuncs[wc->winref] == NIL)
5774 24 : continue;
5775 :
5776 3122 : foreach(lc2, wflists->windowFuncs[wc->winref])
5777 : {
5778 : SupportRequestOptimizeWindowClause req;
5779 : SupportRequestOptimizeWindowClause *res;
5780 2624 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5781 : Oid prosupport;
5782 :
5783 2624 : prosupport = get_func_support(wfunc->winfnoid);
5784 :
5785 : /* Check if there's a support function for 'wfunc' */
5786 2624 : if (!OidIsValid(prosupport))
5787 2084 : break; /* can't optimize this WindowClause */
5788 :
5789 760 : req.type = T_SupportRequestOptimizeWindowClause;
5790 760 : req.window_clause = wc;
5791 760 : req.window_func = wfunc;
5792 760 : req.frameOptions = wc->frameOptions;
5793 :
5794 : /* call the support function */
5795 : res = (SupportRequestOptimizeWindowClause *)
5796 760 : DatumGetPointer(OidFunctionCall1(prosupport,
5797 : PointerGetDatum(&req)));
5798 :
5799 : /*
5800 : * Skip to next WindowClause if the support function does not
5801 : * support this request type.
5802 : */
5803 760 : if (res == NULL)
5804 220 : break;
5805 :
5806 : /*
5807 : * Save these frameOptions for the first WindowFunc for this
5808 : * WindowClause.
5809 : */
5810 540 : if (foreach_current_index(lc2) == 0)
5811 516 : optimizedFrameOptions = res->frameOptions;
5812 :
5813 : /*
5814 : * On subsequent WindowFuncs, if the frameOptions are not the same
5815 : * then we're unable to optimize the frameOptions for this
5816 : * WindowClause.
5817 : */
5818 24 : else if (optimizedFrameOptions != res->frameOptions)
5819 0 : break; /* skip to the next WindowClause, if any */
5820 : }
5821 :
5822 : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
5823 2582 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5824 : {
5825 : ListCell *lc3;
5826 :
5827 : /* apply the new frame options */
5828 498 : wc->frameOptions = optimizedFrameOptions;
5829 :
5830 : /*
5831 : * We now check to see if changing the frameOptions has caused
5832 : * this WindowClause to be a duplicate of some other WindowClause.
5833 : * This can only happen if we have multiple WindowClauses, so
5834 : * don't bother if there's only 1.
5835 : */
5836 498 : if (list_length(windowClause) == 1)
5837 408 : continue;
5838 :
5839 : /*
5840 : * Do the duplicate check and reuse the existing WindowClause if
5841 : * we find a duplicate.
5842 : */
5843 228 : foreach(lc3, windowClause)
5844 : {
5845 174 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5846 :
5847 : /* skip over the WindowClause we're currently editing */
5848 174 : if (existing_wc == wc)
5849 54 : continue;
5850 :
5851 : /*
5852 : * Perform the same duplicate check that is done in
5853 : * transformWindowFuncCall.
5854 : */
5855 240 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
5856 120 : equal(wc->orderClause, existing_wc->orderClause) &&
5857 120 : wc->frameOptions == existing_wc->frameOptions &&
5858 72 : equal(wc->startOffset, existing_wc->startOffset) &&
5859 36 : equal(wc->endOffset, existing_wc->endOffset))
5860 : {
5861 : ListCell *lc4;
5862 :
5863 : /*
5864 : * Now move each WindowFunc in 'wc' into 'existing_wc'.
5865 : * This required adjusting each WindowFunc's winref and
5866 : * moving the WindowFuncs in 'wc' to the list of
5867 : * WindowFuncs in 'existing_wc'.
5868 : */
5869 78 : foreach(lc4, wflists->windowFuncs[wc->winref])
5870 : {
5871 42 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
5872 :
5873 42 : wfunc->winref = existing_wc->winref;
5874 : }
5875 :
5876 : /* move list items */
5877 72 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
5878 36 : wflists->windowFuncs[wc->winref]);
5879 36 : wflists->windowFuncs[wc->winref] = NIL;
5880 :
5881 : /*
5882 : * transformWindowFuncCall() should have made sure there
5883 : * are no other duplicates, so we needn't bother looking
5884 : * any further.
5885 : */
5886 36 : break;
5887 : }
5888 : }
5889 : }
5890 : }
5891 2378 : }
5892 :
5893 : /*
5894 : * select_active_windows
5895 : * Create a list of the "active" window clauses (ie, those referenced
5896 : * by non-deleted WindowFuncs) in the order they are to be executed.
5897 : */
5898 : static List *
5899 2378 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5900 : {
5901 2378 : List *windowClause = root->parse->windowClause;
5902 2378 : List *result = NIL;
5903 : ListCell *lc;
5904 2378 : int nActive = 0;
5905 2378 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5906 2378 : * list_length(windowClause));
5907 :
5908 : /* First, construct an array of the active windows */
5909 4984 : foreach(lc, windowClause)
5910 : {
5911 2606 : WindowClause *wc = lfirst_node(WindowClause, lc);
5912 :
5913 : /* It's only active if wflists shows some related WindowFuncs */
5914 : Assert(wc->winref <= wflists->maxWinRef);
5915 2606 : if (wflists->windowFuncs[wc->winref] == NIL)
5916 60 : continue;
5917 :
5918 2546 : actives[nActive].wc = wc; /* original clause */
5919 :
5920 : /*
5921 : * For sorting, we want the list of partition keys followed by the
5922 : * list of sort keys. But pathkeys construction will remove duplicates
5923 : * between the two, so we can as well (even though we can't detect all
5924 : * of the duplicates, since some may come from ECs - that might mean
5925 : * we miss optimization chances here). We must, however, ensure that
5926 : * the order of entries is preserved with respect to the ones we do
5927 : * keep.
5928 : *
5929 : * partitionClause and orderClause had their own duplicates removed in
5930 : * parse analysis, so we're only concerned here with removing
5931 : * orderClause entries that also appear in partitionClause.
5932 : */
5933 5092 : actives[nActive].uniqueOrder =
5934 2546 : list_concat_unique(list_copy(wc->partitionClause),
5935 2546 : wc->orderClause);
5936 2546 : nActive++;
5937 : }
5938 :
5939 : /*
5940 : * Sort active windows by their partitioning/ordering clauses, ignoring
5941 : * any framing clauses, so that the windows that need the same sorting are
5942 : * adjacent in the list. When we come to generate paths, this will avoid
5943 : * inserting additional Sort nodes.
5944 : *
5945 : * This is how we implement a specific requirement from the SQL standard,
5946 : * which says that when two or more windows are order-equivalent (i.e.
5947 : * have matching partition and order clauses, even if their names or
5948 : * framing clauses differ), then all peer rows must be presented in the
5949 : * same order in all of them. If we allowed multiple sort nodes for such
5950 : * cases, we'd risk having the peer rows end up in different orders in
5951 : * equivalent windows due to sort instability. (See General Rule 4 of
5952 : * <window clause> in SQL2008 - SQL2016.)
5953 : *
5954 : * Additionally, if the entire list of clauses of one window is a prefix
5955 : * of another, put first the window with stronger sorting requirements.
5956 : * This way we will first sort for stronger window, and won't have to sort
5957 : * again for the weaker one.
5958 : */
5959 2378 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5960 :
5961 : /* build ordered list of the original WindowClause nodes */
5962 4924 : for (int i = 0; i < nActive; i++)
5963 2546 : result = lappend(result, actives[i].wc);
5964 :
5965 2378 : pfree(actives);
5966 :
5967 2378 : return result;
5968 : }
5969 :
5970 : /*
5971 : * name_active_windows
5972 : * Ensure all active windows have unique names.
5973 : *
5974 : * The parser will have checked that user-assigned window names are unique
5975 : * within the Query. Here we assign made-up names to any unnamed
5976 : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
5977 : * at parse time, because it'd mess up decompilation of views.)
5978 : *
5979 : * activeWindows: result of select_active_windows
5980 : */
5981 : static void
5982 2378 : name_active_windows(List *activeWindows)
5983 : {
5984 2378 : int next_n = 1;
5985 : char newname[16];
5986 : ListCell *lc;
5987 :
5988 4924 : foreach(lc, activeWindows)
5989 : {
5990 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
5991 :
5992 : /* Nothing to do if it has a name already. */
5993 2546 : if (wc->name)
5994 498 : continue;
5995 :
5996 : /* Select a name not currently present in the list. */
5997 : for (;;)
5998 6 : {
5999 : ListCell *lc2;
6000 :
6001 2054 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6002 4456 : foreach(lc2, activeWindows)
6003 : {
6004 2408 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6005 :
6006 2408 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6007 6 : break; /* matched */
6008 : }
6009 2054 : if (lc2 == NULL)
6010 2048 : break; /* reached the end with no match */
6011 : }
6012 2048 : wc->name = pstrdup(newname);
6013 : }
6014 2378 : }
6015 :
6016 : /*
6017 : * common_prefix_cmp
6018 : * QSort comparison function for WindowClauseSortData
6019 : *
6020 : * Sort the windows by the required sorting clauses. First, compare the sort
6021 : * clauses themselves. Second, if one window's clauses are a prefix of another
6022 : * one's clauses, put the window with more sort clauses first.
6023 : *
6024 : * We purposefully sort by the highest tleSortGroupRef first. Since
6025 : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6026 : * and because here we sort the lowest tleSortGroupRefs last, if a
6027 : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6028 : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6029 : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6030 : * reducing the total number of sorts required for the query.
6031 : */
6032 : static int
6033 186 : common_prefix_cmp(const void *a, const void *b)
6034 : {
6035 186 : const WindowClauseSortData *wcsa = a;
6036 186 : const WindowClauseSortData *wcsb = b;
6037 : ListCell *item_a;
6038 : ListCell *item_b;
6039 :
6040 330 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
6041 : {
6042 246 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6043 246 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6044 :
6045 246 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6046 102 : return -1;
6047 234 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6048 66 : return 1;
6049 168 : else if (sca->sortop > scb->sortop)
6050 0 : return -1;
6051 168 : else if (sca->sortop < scb->sortop)
6052 24 : return 1;
6053 144 : else if (sca->nulls_first && !scb->nulls_first)
6054 0 : return -1;
6055 144 : else if (!sca->nulls_first && scb->nulls_first)
6056 0 : return 1;
6057 : /* no need to compare eqop, since it is fully determined by sortop */
6058 : }
6059 :
6060 84 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6061 6 : return -1;
6062 78 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6063 30 : return 1;
6064 :
6065 48 : return 0;
6066 : }
6067 :
6068 : /*
6069 : * make_window_input_target
6070 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6071 : *
6072 : * When the query has window functions, this function computes the desired
6073 : * target to be computed by the node just below the first WindowAgg.
6074 : * This tlist must contain all values needed to evaluate the window functions,
6075 : * compute the final target list, and perform any required final sort step.
6076 : * If multiple WindowAggs are needed, each intermediate one adds its window
6077 : * function results onto this base tlist; only the topmost WindowAgg computes
6078 : * the actual desired target list.
6079 : *
6080 : * This function is much like make_group_input_target, though not quite enough
6081 : * like it to share code. As in that function, we flatten most expressions
6082 : * into their component variables. But we do not want to flatten window
6083 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6084 : * evaluations of them, which would be bad (possibly even resulting in
6085 : * inconsistent answers, if they contain volatile functions).
6086 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6087 : * make_group_input_target, because we may no longer have access to the
6088 : * individual Vars in them.
6089 : *
6090 : * Another key difference from make_group_input_target is that we don't
6091 : * flatten Aggref expressions, since those are to be computed below the
6092 : * window functions and just referenced like Vars above that.
6093 : *
6094 : * 'final_target' is the query's final target list (in PathTarget form)
6095 : * 'activeWindows' is the list of active windows previously identified by
6096 : * select_active_windows.
6097 : *
6098 : * The result is the PathTarget to be computed by the plan node immediately
6099 : * below the first WindowAgg node.
6100 : */
6101 : static PathTarget *
6102 2378 : make_window_input_target(PlannerInfo *root,
6103 : PathTarget *final_target,
6104 : List *activeWindows)
6105 : {
6106 : PathTarget *input_target;
6107 : Bitmapset *sgrefs;
6108 : List *flattenable_cols;
6109 : List *flattenable_vars;
6110 : int i;
6111 : ListCell *lc;
6112 :
6113 : Assert(root->parse->hasWindowFuncs);
6114 :
6115 : /*
6116 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6117 : * into a bitmapset for convenient reference below.
6118 : */
6119 2378 : sgrefs = NULL;
6120 4924 : foreach(lc, activeWindows)
6121 : {
6122 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
6123 : ListCell *lc2;
6124 :
6125 3290 : foreach(lc2, wc->partitionClause)
6126 : {
6127 744 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6128 :
6129 744 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6130 : }
6131 4722 : foreach(lc2, wc->orderClause)
6132 : {
6133 2176 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6134 :
6135 2176 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6136 : }
6137 : }
6138 :
6139 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6140 2564 : foreach(lc, root->processed_groupClause)
6141 : {
6142 186 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6143 :
6144 186 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6145 : }
6146 :
6147 : /*
6148 : * Construct a target containing all the non-flattenable targetlist items,
6149 : * and save aside the others for a moment.
6150 : */
6151 2378 : input_target = create_empty_pathtarget();
6152 2378 : flattenable_cols = NIL;
6153 :
6154 2378 : i = 0;
6155 10238 : foreach(lc, final_target->exprs)
6156 : {
6157 7860 : Expr *expr = (Expr *) lfirst(lc);
6158 7860 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6159 :
6160 : /*
6161 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6162 : * that such items can't contain window functions, so it's okay to
6163 : * compute them below the WindowAgg nodes.)
6164 : */
6165 7860 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6166 : {
6167 : /*
6168 : * Don't want to deconstruct this value, so add it to the input
6169 : * target as-is.
6170 : */
6171 2774 : add_column_to_pathtarget(input_target, expr, sgref);
6172 : }
6173 : else
6174 : {
6175 : /*
6176 : * Column is to be flattened, so just remember the expression for
6177 : * later call to pull_var_clause.
6178 : */
6179 5086 : flattenable_cols = lappend(flattenable_cols, expr);
6180 : }
6181 :
6182 7860 : i++;
6183 : }
6184 :
6185 : /*
6186 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6187 : * add them to the input target if not already present. (Some might be
6188 : * there already because they're used directly as window/group clauses.)
6189 : *
6190 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6191 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6192 : * at higher levels. On the other hand, we should recurse into
6193 : * WindowFuncs to make sure their input expressions are available.
6194 : */
6195 2378 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6196 : PVC_INCLUDE_AGGREGATES |
6197 : PVC_RECURSE_WINDOWFUNCS |
6198 : PVC_INCLUDE_PLACEHOLDERS);
6199 2378 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6200 :
6201 : /* clean up cruft */
6202 2378 : list_free(flattenable_vars);
6203 2378 : list_free(flattenable_cols);
6204 :
6205 : /* XXX this causes some redundant cost calculation ... */
6206 2378 : return set_pathtarget_cost_width(root, input_target);
6207 : }
6208 :
6209 : /*
6210 : * make_pathkeys_for_window
6211 : * Create a pathkeys list describing the required input ordering
6212 : * for the given WindowClause.
6213 : *
6214 : * Modifies wc's partitionClause to remove any clauses which are deemed
6215 : * redundant by the pathkey logic.
6216 : *
6217 : * The required ordering is first the PARTITION keys, then the ORDER keys.
6218 : * In the future we might try to implement windowing using hashing, in which
6219 : * case the ordering could be relaxed, but for now we always sort.
6220 : */
6221 : static List *
6222 5132 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6223 : List *tlist)
6224 : {
6225 5132 : List *window_pathkeys = NIL;
6226 :
6227 : /* Throw error if can't sort */
6228 5132 : if (!grouping_is_sortable(wc->partitionClause))
6229 0 : ereport(ERROR,
6230 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6231 : errmsg("could not implement window PARTITION BY"),
6232 : errdetail("Window partitioning columns must be of sortable datatypes.")));
6233 5132 : if (!grouping_is_sortable(wc->orderClause))
6234 0 : ereport(ERROR,
6235 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6236 : errmsg("could not implement window ORDER BY"),
6237 : errdetail("Window ordering columns must be of sortable datatypes.")));
6238 :
6239 : /*
6240 : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6241 : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6242 : */
6243 5132 : if (wc->partitionClause != NIL)
6244 : {
6245 : bool sortable;
6246 :
6247 1290 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6248 : &wc->partitionClause,
6249 : tlist,
6250 : true,
6251 : false,
6252 : &sortable,
6253 : false);
6254 :
6255 : Assert(sortable);
6256 : }
6257 :
6258 : /*
6259 : * In principle, we could also consider removing redundant ORDER BY items
6260 : * too as doing so does not alter the result of peer row checks done by
6261 : * the executor. However, we must *not* remove the ordering column for
6262 : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6263 : * if it's known to be equal to some partitioning column.
6264 : */
6265 5132 : if (wc->orderClause != NIL)
6266 : {
6267 : List *orderby_pathkeys;
6268 :
6269 4276 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6270 : wc->orderClause,
6271 : tlist);
6272 :
6273 : /* Okay, make the combined pathkeys */
6274 4276 : if (window_pathkeys != NIL)
6275 934 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6276 : else
6277 3342 : window_pathkeys = orderby_pathkeys;
6278 : }
6279 :
6280 5132 : return window_pathkeys;
6281 : }
6282 :
6283 : /*
6284 : * make_sort_input_target
6285 : * Generate appropriate PathTarget for initial input to Sort step.
6286 : *
6287 : * If the query has ORDER BY, this function chooses the target to be computed
6288 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6289 : * project) steps. This might or might not be identical to the query's final
6290 : * output target.
6291 : *
6292 : * The main argument for keeping the sort-input tlist the same as the final
6293 : * is that we avoid a separate projection node (which will be needed if
6294 : * they're different, because Sort can't project). However, there are also
6295 : * advantages to postponing tlist evaluation till after the Sort: it ensures
6296 : * a consistent order of evaluation for any volatile functions in the tlist,
6297 : * and if there's also a LIMIT, we can stop the query without ever computing
6298 : * tlist functions for later rows, which is beneficial for both volatile and
6299 : * expensive functions.
6300 : *
6301 : * Our current policy is to postpone volatile expressions till after the sort
6302 : * unconditionally (assuming that that's possible, ie they are in plain tlist
6303 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6304 : * postpone set-returning expressions, because running them beforehand would
6305 : * bloat the sort dataset, and because it might cause unexpected output order
6306 : * if the sort isn't stable. However there's a constraint on that: all SRFs
6307 : * in the tlist should be evaluated at the same plan step, so that they can
6308 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6309 : * mustn't postpone any SRFs. (Note that in principle that policy should
6310 : * probably get applied to the group/window input targetlists too, but we
6311 : * have not done that historically.) Lastly, expensive expressions are
6312 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6313 : * partial evaluation of the query is possible (if neither is true, we expect
6314 : * to have to evaluate the expressions for every row anyway), or if there are
6315 : * any volatile or set-returning expressions (since once we've put in a
6316 : * projection at all, it won't cost any more to postpone more stuff).
6317 : *
6318 : * Another issue that could potentially be considered here is that
6319 : * evaluating tlist expressions could result in data that's either wider
6320 : * or narrower than the input Vars, thus changing the volume of data that
6321 : * has to go through the Sort. However, we usually have only a very bad
6322 : * idea of the output width of any expression more complex than a Var,
6323 : * so for now it seems too risky to try to optimize on that basis.
6324 : *
6325 : * Note that if we do produce a modified sort-input target, and then the
6326 : * query ends up not using an explicit Sort, no particular harm is done:
6327 : * we'll initially use the modified target for the preceding path nodes,
6328 : * but then change them to the final target with apply_projection_to_path.
6329 : * Moreover, in such a case the guarantees about evaluation order of
6330 : * volatile functions still hold, since the rows are sorted already.
6331 : *
6332 : * This function has some things in common with make_group_input_target and
6333 : * make_window_input_target, though the detailed rules for what to do are
6334 : * different. We never flatten/postpone any grouping or ordering columns;
6335 : * those are needed before the sort. If we do flatten a particular
6336 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6337 : * computed earlier.
6338 : *
6339 : * 'final_target' is the query's final target list (in PathTarget form)
6340 : * 'have_postponed_srfs' is an output argument, see below
6341 : *
6342 : * The result is the PathTarget to be computed by the plan node immediately
6343 : * below the Sort step (and the Distinct step, if any). This will be
6344 : * exactly final_target if we decide a projection step wouldn't be helpful.
6345 : *
6346 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6347 : * any set-returning functions to after the Sort.
6348 : */
6349 : static PathTarget *
6350 71882 : make_sort_input_target(PlannerInfo *root,
6351 : PathTarget *final_target,
6352 : bool *have_postponed_srfs)
6353 : {
6354 71882 : Query *parse = root->parse;
6355 : PathTarget *input_target;
6356 : int ncols;
6357 : bool *col_is_srf;
6358 : bool *postpone_col;
6359 : bool have_srf;
6360 : bool have_volatile;
6361 : bool have_expensive;
6362 : bool have_srf_sortcols;
6363 : bool postpone_srfs;
6364 : List *postponable_cols;
6365 : List *postponable_vars;
6366 : int i;
6367 : ListCell *lc;
6368 :
6369 : /* Shouldn't get here unless query has ORDER BY */
6370 : Assert(parse->sortClause);
6371 :
6372 71882 : *have_postponed_srfs = false; /* default result */
6373 :
6374 : /* Inspect tlist and collect per-column information */
6375 71882 : ncols = list_length(final_target->exprs);
6376 71882 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6377 71882 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6378 71882 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6379 :
6380 71882 : i = 0;
6381 446110 : foreach(lc, final_target->exprs)
6382 : {
6383 374228 : Expr *expr = (Expr *) lfirst(lc);
6384 :
6385 : /*
6386 : * If the column has a sortgroupref, assume it has to be evaluated
6387 : * before sorting. Generally such columns would be ORDER BY, GROUP
6388 : * BY, etc targets. One exception is columns that were removed from
6389 : * GROUP BY by remove_useless_groupby_columns() ... but those would
6390 : * only be Vars anyway. There don't seem to be any cases where it
6391 : * would be worth the trouble to double-check.
6392 : */
6393 374228 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6394 : {
6395 : /*
6396 : * Check for SRF or volatile functions. Check the SRF case first
6397 : * because we must know whether we have any postponed SRFs.
6398 : */
6399 272278 : if (parse->hasTargetSRFs &&
6400 216 : expression_returns_set((Node *) expr))
6401 : {
6402 : /* We'll decide below whether these are postponable */
6403 96 : col_is_srf[i] = true;
6404 96 : have_srf = true;
6405 : }
6406 271966 : else if (contain_volatile_functions((Node *) expr))
6407 : {
6408 : /* Unconditionally postpone */
6409 148 : postpone_col[i] = true;
6410 148 : have_volatile = true;
6411 : }
6412 : else
6413 : {
6414 : /*
6415 : * Else check the cost. XXX it's annoying to have to do this
6416 : * when set_pathtarget_cost_width() just did it. Refactor to
6417 : * allow sharing the work?
6418 : */
6419 : QualCost cost;
6420 :
6421 271818 : cost_qual_eval_node(&cost, (Node *) expr, root);
6422 :
6423 : /*
6424 : * We arbitrarily define "expensive" as "more than 10X
6425 : * cpu_operator_cost". Note this will take in any PL function
6426 : * with default cost.
6427 : */
6428 271818 : if (cost.per_tuple > 10 * cpu_operator_cost)
6429 : {
6430 17216 : postpone_col[i] = true;
6431 17216 : have_expensive = true;
6432 : }
6433 : }
6434 : }
6435 : else
6436 : {
6437 : /* For sortgroupref cols, just check if any contain SRFs */
6438 102166 : if (!have_srf_sortcols &&
6439 102476 : parse->hasTargetSRFs &&
6440 310 : expression_returns_set((Node *) expr))
6441 124 : have_srf_sortcols = true;
6442 : }
6443 :
6444 374228 : i++;
6445 : }
6446 :
6447 : /*
6448 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6449 : */
6450 71882 : postpone_srfs = (have_srf && !have_srf_sortcols);
6451 :
6452 : /*
6453 : * If we don't need a post-sort projection, just return final_target.
6454 : */
6455 71882 : if (!(postpone_srfs || have_volatile ||
6456 71678 : (have_expensive &&
6457 10110 : (parse->limitCount || root->tuple_fraction > 0))))
6458 71642 : return final_target;
6459 :
6460 : /*
6461 : * Report whether the post-sort projection will contain set-returning
6462 : * functions. This is important because it affects whether the Sort can
6463 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6464 : * to return.
6465 : */
6466 240 : *have_postponed_srfs = postpone_srfs;
6467 :
6468 : /*
6469 : * Construct the sort-input target, taking all non-postponable columns and
6470 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6471 : * the postponable ones.
6472 : */
6473 240 : input_target = create_empty_pathtarget();
6474 240 : postponable_cols = NIL;
6475 :
6476 240 : i = 0;
6477 1990 : foreach(lc, final_target->exprs)
6478 : {
6479 1750 : Expr *expr = (Expr *) lfirst(lc);
6480 :
6481 1750 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
6482 298 : postponable_cols = lappend(postponable_cols, expr);
6483 : else
6484 1452 : add_column_to_pathtarget(input_target, expr,
6485 1452 : get_pathtarget_sortgroupref(final_target, i));
6486 :
6487 1750 : i++;
6488 : }
6489 :
6490 : /*
6491 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6492 : * postponable columns, and add them to the sort-input target if not
6493 : * already present. (Some might be there already.) We mustn't
6494 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6495 : * would be unable to recompute them.
6496 : */
6497 240 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6498 : PVC_INCLUDE_AGGREGATES |
6499 : PVC_INCLUDE_WINDOWFUNCS |
6500 : PVC_INCLUDE_PLACEHOLDERS);
6501 240 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6502 :
6503 : /* clean up cruft */
6504 240 : list_free(postponable_vars);
6505 240 : list_free(postponable_cols);
6506 :
6507 : /* XXX this represents even more redundant cost calculation ... */
6508 240 : return set_pathtarget_cost_width(root, input_target);
6509 : }
6510 :
6511 : /*
6512 : * get_cheapest_fractional_path
6513 : * Find the cheapest path for retrieving a specified fraction of all
6514 : * the tuples expected to be returned by the given relation.
6515 : *
6516 : * Do not consider parameterized paths. If the caller needs a path for upper
6517 : * rel, it can't have parameterized paths. If the caller needs an append
6518 : * subpath, it could become limited by the treatment of similar
6519 : * parameterization of all the subpaths.
6520 : *
6521 : * We interpret tuple_fraction the same way as grouping_planner.
6522 : *
6523 : * We assume set_cheapest() has been run on the given rel.
6524 : */
6525 : Path *
6526 499788 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6527 : {
6528 499788 : Path *best_path = rel->cheapest_total_path;
6529 : ListCell *l;
6530 :
6531 : /* If all tuples will be retrieved, just return the cheapest-total path */
6532 499788 : if (tuple_fraction <= 0.0)
6533 490392 : return best_path;
6534 :
6535 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6536 9396 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6537 3888 : tuple_fraction /= best_path->rows;
6538 :
6539 24740 : foreach(l, rel->pathlist)
6540 : {
6541 15344 : Path *path = (Path *) lfirst(l);
6542 :
6543 15344 : if (path->param_info)
6544 188 : continue;
6545 :
6546 20916 : if (path == rel->cheapest_total_path ||
6547 5760 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6548 14666 : continue;
6549 :
6550 490 : best_path = path;
6551 : }
6552 :
6553 9396 : return best_path;
6554 : }
6555 :
6556 : /*
6557 : * adjust_paths_for_srfs
6558 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6559 : *
6560 : * The executor can only handle set-returning functions that appear at the
6561 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6562 : * that are not at top level, we need to split up the evaluation into multiple
6563 : * plan levels in which each level satisfies this constraint. This function
6564 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6565 : * output tlist to insert appropriate projection steps.
6566 : *
6567 : * The given targets and targets_contain_srfs lists are from
6568 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6569 : * target in targets.
6570 : */
6571 : static void
6572 12652 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6573 : List *targets, List *targets_contain_srfs)
6574 : {
6575 : ListCell *lc;
6576 :
6577 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6578 : Assert(!linitial_int(targets_contain_srfs));
6579 :
6580 : /* If no SRFs appear at this plan level, nothing to do */
6581 12652 : if (list_length(targets) == 1)
6582 628 : return;
6583 :
6584 : /*
6585 : * Stack SRF-evaluation nodes atop each path for the rel.
6586 : *
6587 : * In principle we should re-run set_cheapest() here to identify the
6588 : * cheapest path, but it seems unlikely that adding the same tlist eval
6589 : * costs to all the paths would change that, so we don't bother. Instead,
6590 : * just assume that the cheapest-startup and cheapest-total paths remain
6591 : * so. (There should be no parameterized paths anymore, so we needn't
6592 : * worry about updating cheapest_parameterized_paths.)
6593 : */
6594 24072 : foreach(lc, rel->pathlist)
6595 : {
6596 12048 : Path *subpath = (Path *) lfirst(lc);
6597 12048 : Path *newpath = subpath;
6598 : ListCell *lc1,
6599 : *lc2;
6600 :
6601 : Assert(subpath->param_info == NULL);
6602 37318 : forboth(lc1, targets, lc2, targets_contain_srfs)
6603 : {
6604 25270 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6605 25270 : bool contains_srfs = (bool) lfirst_int(lc2);
6606 :
6607 : /* If this level doesn't contain SRFs, do regular projection */
6608 25270 : if (contains_srfs)
6609 12108 : newpath = (Path *) create_set_projection_path(root,
6610 : rel,
6611 : newpath,
6612 : thistarget);
6613 : else
6614 13162 : newpath = (Path *) apply_projection_to_path(root,
6615 : rel,
6616 : newpath,
6617 : thistarget);
6618 : }
6619 12048 : lfirst(lc) = newpath;
6620 12048 : if (subpath == rel->cheapest_startup_path)
6621 372 : rel->cheapest_startup_path = newpath;
6622 12048 : if (subpath == rel->cheapest_total_path)
6623 372 : rel->cheapest_total_path = newpath;
6624 : }
6625 :
6626 : /* Likewise for partial paths, if any */
6627 12030 : foreach(lc, rel->partial_pathlist)
6628 : {
6629 6 : Path *subpath = (Path *) lfirst(lc);
6630 6 : Path *newpath = subpath;
6631 : ListCell *lc1,
6632 : *lc2;
6633 :
6634 : Assert(subpath->param_info == NULL);
6635 24 : forboth(lc1, targets, lc2, targets_contain_srfs)
6636 : {
6637 18 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6638 18 : bool contains_srfs = (bool) lfirst_int(lc2);
6639 :
6640 : /* If this level doesn't contain SRFs, do regular projection */
6641 18 : if (contains_srfs)
6642 6 : newpath = (Path *) create_set_projection_path(root,
6643 : rel,
6644 : newpath,
6645 : thistarget);
6646 : else
6647 : {
6648 : /* avoid apply_projection_to_path, in case of multiple refs */
6649 12 : newpath = (Path *) create_projection_path(root,
6650 : rel,
6651 : newpath,
6652 : thistarget);
6653 : }
6654 : }
6655 6 : lfirst(lc) = newpath;
6656 : }
6657 : }
6658 :
6659 : /*
6660 : * expression_planner
6661 : * Perform planner's transformations on a standalone expression.
6662 : *
6663 : * Various utility commands need to evaluate expressions that are not part
6664 : * of a plannable query. They can do so using the executor's regular
6665 : * expression-execution machinery, but first the expression has to be fed
6666 : * through here to transform it from parser output to something executable.
6667 : *
6668 : * Currently, we disallow sublinks in standalone expressions, so there's no
6669 : * real "planning" involved here. (That might not always be true though.)
6670 : * What we must do is run eval_const_expressions to ensure that any function
6671 : * calls are converted to positional notation and function default arguments
6672 : * get inserted. The fact that constant subexpressions get simplified is a
6673 : * side-effect that is useful when the expression will get evaluated more than
6674 : * once. Also, we must fix operator function IDs.
6675 : *
6676 : * This does not return any information about dependencies of the expression.
6677 : * Hence callers should use the results only for the duration of the current
6678 : * query. Callers that would like to cache the results for longer should use
6679 : * expression_planner_with_deps, probably via the plancache.
6680 : *
6681 : * Note: this must not make any damaging changes to the passed-in expression
6682 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6683 : * we first do an expression_tree_mutator-based walk, what is returned will
6684 : * be a new node tree.) The result is constructed in the current memory
6685 : * context; beware that this can leak a lot of additional stuff there, too.
6686 : */
6687 : Expr *
6688 246452 : expression_planner(Expr *expr)
6689 : {
6690 : Node *result;
6691 :
6692 : /*
6693 : * Convert named-argument function calls, insert default arguments and
6694 : * simplify constant subexprs
6695 : */
6696 246452 : result = eval_const_expressions(NULL, (Node *) expr);
6697 :
6698 : /* Fill in opfuncid values if missing */
6699 246434 : fix_opfuncids(result);
6700 :
6701 246434 : return (Expr *) result;
6702 : }
6703 :
6704 : /*
6705 : * expression_planner_with_deps
6706 : * Perform planner's transformations on a standalone expression,
6707 : * returning expression dependency information along with the result.
6708 : *
6709 : * This is identical to expression_planner() except that it also returns
6710 : * information about possible dependencies of the expression, ie identities of
6711 : * objects whose definitions affect the result. As in a PlannedStmt, these
6712 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6713 : */
6714 : Expr *
6715 350 : expression_planner_with_deps(Expr *expr,
6716 : List **relationOids,
6717 : List **invalItems)
6718 : {
6719 : Node *result;
6720 : PlannerGlobal glob;
6721 : PlannerInfo root;
6722 :
6723 : /* Make up dummy planner state so we can use setrefs machinery */
6724 8050 : MemSet(&glob, 0, sizeof(glob));
6725 350 : glob.type = T_PlannerGlobal;
6726 350 : glob.relationOids = NIL;
6727 350 : glob.invalItems = NIL;
6728 :
6729 31150 : MemSet(&root, 0, sizeof(root));
6730 350 : root.type = T_PlannerInfo;
6731 350 : root.glob = &glob;
6732 :
6733 : /*
6734 : * Convert named-argument function calls, insert default arguments and
6735 : * simplify constant subexprs. Collect identities of inlined functions
6736 : * and elided domains, too.
6737 : */
6738 350 : result = eval_const_expressions(&root, (Node *) expr);
6739 :
6740 : /* Fill in opfuncid values if missing */
6741 350 : fix_opfuncids(result);
6742 :
6743 : /*
6744 : * Now walk the finished expression to find anything else we ought to
6745 : * record as an expression dependency.
6746 : */
6747 350 : (void) extract_query_dependencies_walker(result, &root);
6748 :
6749 350 : *relationOids = glob.relationOids;
6750 350 : *invalItems = glob.invalItems;
6751 :
6752 350 : return (Expr *) result;
6753 : }
6754 :
6755 :
6756 : /*
6757 : * plan_cluster_use_sort
6758 : * Use the planner to decide how CLUSTER should implement sorting
6759 : *
6760 : * tableOid is the OID of a table to be clustered on its index indexOid
6761 : * (which is already known to be a btree index). Decide whether it's
6762 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6763 : * Return true to use sorting, false to use an indexscan.
6764 : *
6765 : * Note: caller had better already hold some type of lock on the table.
6766 : */
6767 : bool
6768 188 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6769 : {
6770 : PlannerInfo *root;
6771 : Query *query;
6772 : PlannerGlobal *glob;
6773 : RangeTblEntry *rte;
6774 : RelOptInfo *rel;
6775 : IndexOptInfo *indexInfo;
6776 : QualCost indexExprCost;
6777 : Cost comparisonCost;
6778 : Path *seqScanPath;
6779 : Path seqScanAndSortPath;
6780 : IndexPath *indexScanPath;
6781 : ListCell *lc;
6782 :
6783 : /* We can short-circuit the cost comparison if indexscans are disabled */
6784 188 : if (!enable_indexscan)
6785 30 : return true; /* use sort */
6786 :
6787 : /* Set up mostly-dummy planner state */
6788 158 : query = makeNode(Query);
6789 158 : query->commandType = CMD_SELECT;
6790 :
6791 158 : glob = makeNode(PlannerGlobal);
6792 :
6793 158 : root = makeNode(PlannerInfo);
6794 158 : root->parse = query;
6795 158 : root->glob = glob;
6796 158 : root->query_level = 1;
6797 158 : root->planner_cxt = CurrentMemoryContext;
6798 158 : root->wt_param_id = -1;
6799 158 : root->join_domains = list_make1(makeNode(JoinDomain));
6800 :
6801 : /* Build a minimal RTE for the rel */
6802 158 : rte = makeNode(RangeTblEntry);
6803 158 : rte->rtekind = RTE_RELATION;
6804 158 : rte->relid = tableOid;
6805 158 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6806 158 : rte->rellockmode = AccessShareLock;
6807 158 : rte->lateral = false;
6808 158 : rte->inh = false;
6809 158 : rte->inFromCl = true;
6810 158 : query->rtable = list_make1(rte);
6811 158 : addRTEPermissionInfo(&query->rteperminfos, rte);
6812 :
6813 : /* Set up RTE/RelOptInfo arrays */
6814 158 : setup_simple_rel_arrays(root);
6815 :
6816 : /* Build RelOptInfo */
6817 158 : rel = build_simple_rel(root, 1, NULL);
6818 :
6819 : /* Locate IndexOptInfo for the target index */
6820 158 : indexInfo = NULL;
6821 196 : foreach(lc, rel->indexlist)
6822 : {
6823 196 : indexInfo = lfirst_node(IndexOptInfo, lc);
6824 196 : if (indexInfo->indexoid == indexOid)
6825 158 : break;
6826 : }
6827 :
6828 : /*
6829 : * It's possible that get_relation_info did not generate an IndexOptInfo
6830 : * for the desired index; this could happen if it's not yet reached its
6831 : * indcheckxmin usability horizon, or if it's a system index and we're
6832 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6833 : * trust the index contents but use seqscan-and-sort.
6834 : */
6835 158 : if (lc == NULL) /* not in the list? */
6836 0 : return true; /* use sort */
6837 :
6838 : /*
6839 : * Rather than doing all the pushups that would be needed to use
6840 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6841 : */
6842 158 : rel->rows = rel->tuples;
6843 158 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6844 :
6845 158 : root->total_table_pages = rel->pages;
6846 :
6847 : /*
6848 : * Determine eval cost of the index expressions, if any. We need to
6849 : * charge twice that amount for each tuple comparison that happens during
6850 : * the sort, since tuplesort.c will have to re-evaluate the index
6851 : * expressions each time. (XXX that's pretty inefficient...)
6852 : */
6853 158 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6854 158 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6855 :
6856 : /* Estimate the cost of seq scan + sort */
6857 158 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6858 158 : cost_sort(&seqScanAndSortPath, root, NIL,
6859 : seqScanPath->disabled_nodes,
6860 158 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6861 : comparisonCost, maintenance_work_mem, -1.0);
6862 :
6863 : /* Estimate the cost of index scan */
6864 158 : indexScanPath = create_index_path(root, indexInfo,
6865 : NIL, NIL, NIL, NIL,
6866 : ForwardScanDirection, false,
6867 : NULL, 1.0, false);
6868 :
6869 158 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6870 : }
6871 :
6872 : /*
6873 : * plan_create_index_workers
6874 : * Use the planner to decide how many parallel worker processes
6875 : * CREATE INDEX should request for use
6876 : *
6877 : * tableOid is the table on which the index is to be built. indexOid is the
6878 : * OID of an index to be created or reindexed (which must be an index with
6879 : * support for parallel builds - currently btree or BRIN).
6880 : *
6881 : * Return value is the number of parallel worker processes to request. It
6882 : * may be unsafe to proceed if this is 0. Note that this does not include the
6883 : * leader participating as a worker (value is always a number of parallel
6884 : * worker processes).
6885 : *
6886 : * Note: caller had better already hold some type of lock on the table and
6887 : * index.
6888 : */
6889 : int
6890 36232 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6891 : {
6892 : PlannerInfo *root;
6893 : Query *query;
6894 : PlannerGlobal *glob;
6895 : RangeTblEntry *rte;
6896 : Relation heap;
6897 : Relation index;
6898 : RelOptInfo *rel;
6899 : int parallel_workers;
6900 : BlockNumber heap_blocks;
6901 : double reltuples;
6902 : double allvisfrac;
6903 :
6904 : /*
6905 : * We don't allow performing parallel operation in standalone backend or
6906 : * when parallelism is disabled.
6907 : */
6908 36232 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
6909 498 : return 0;
6910 :
6911 : /* Set up largely-dummy planner state */
6912 35734 : query = makeNode(Query);
6913 35734 : query->commandType = CMD_SELECT;
6914 :
6915 35734 : glob = makeNode(PlannerGlobal);
6916 :
6917 35734 : root = makeNode(PlannerInfo);
6918 35734 : root->parse = query;
6919 35734 : root->glob = glob;
6920 35734 : root->query_level = 1;
6921 35734 : root->planner_cxt = CurrentMemoryContext;
6922 35734 : root->wt_param_id = -1;
6923 35734 : root->join_domains = list_make1(makeNode(JoinDomain));
6924 :
6925 : /*
6926 : * Build a minimal RTE.
6927 : *
6928 : * Mark the RTE with inh = true. This is a kludge to prevent
6929 : * get_relation_info() from fetching index info, which is necessary
6930 : * because it does not expect that any IndexOptInfo is currently
6931 : * undergoing REINDEX.
6932 : */
6933 35734 : rte = makeNode(RangeTblEntry);
6934 35734 : rte->rtekind = RTE_RELATION;
6935 35734 : rte->relid = tableOid;
6936 35734 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6937 35734 : rte->rellockmode = AccessShareLock;
6938 35734 : rte->lateral = false;
6939 35734 : rte->inh = true;
6940 35734 : rte->inFromCl = true;
6941 35734 : query->rtable = list_make1(rte);
6942 35734 : addRTEPermissionInfo(&query->rteperminfos, rte);
6943 :
6944 : /* Set up RTE/RelOptInfo arrays */
6945 35734 : setup_simple_rel_arrays(root);
6946 :
6947 : /* Build RelOptInfo */
6948 35734 : rel = build_simple_rel(root, 1, NULL);
6949 :
6950 : /* Rels are assumed already locked by the caller */
6951 35734 : heap = table_open(tableOid, NoLock);
6952 35734 : index = index_open(indexOid, NoLock);
6953 :
6954 : /*
6955 : * Determine if it's safe to proceed.
6956 : *
6957 : * Currently, parallel workers can't access the leader's temporary tables.
6958 : * Furthermore, any index predicate or index expressions must be parallel
6959 : * safe.
6960 : */
6961 35734 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6962 33706 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6963 33586 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6964 : {
6965 2148 : parallel_workers = 0;
6966 2148 : goto done;
6967 : }
6968 :
6969 : /*
6970 : * If parallel_workers storage parameter is set for the table, accept that
6971 : * as the number of parallel worker processes to launch (though still cap
6972 : * at max_parallel_maintenance_workers). Note that we deliberately do not
6973 : * consider any other factor when parallel_workers is set. (e.g., memory
6974 : * use by workers.)
6975 : */
6976 33586 : if (rel->rel_parallel_workers != -1)
6977 : {
6978 18 : parallel_workers = Min(rel->rel_parallel_workers,
6979 : max_parallel_maintenance_workers);
6980 18 : goto done;
6981 : }
6982 :
6983 : /*
6984 : * Estimate heap relation size ourselves, since rel->pages cannot be
6985 : * trusted (heap RTE was marked as inheritance parent)
6986 : */
6987 33568 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6988 :
6989 : /*
6990 : * Determine number of workers to scan the heap relation using generic
6991 : * model
6992 : */
6993 33568 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6994 : max_parallel_maintenance_workers);
6995 :
6996 : /*
6997 : * Cap workers based on available maintenance_work_mem as needed.
6998 : *
6999 : * Note that each tuplesort participant receives an even share of the
7000 : * total maintenance_work_mem budget. Aim to leave participants
7001 : * (including the leader as a participant) with no less than 32MB of
7002 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7003 : * immediately past the threshold of being capable of launching a single
7004 : * parallel worker to sort.
7005 : */
7006 33724 : while (parallel_workers > 0 &&
7007 314 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7008 156 : parallel_workers--;
7009 :
7010 33568 : done:
7011 35734 : index_close(index, NoLock);
7012 35734 : table_close(heap, NoLock);
7013 :
7014 35734 : return parallel_workers;
7015 : }
7016 :
7017 : /*
7018 : * add_paths_to_grouping_rel
7019 : *
7020 : * Add non-partial paths to grouping relation.
7021 : */
7022 : static void
7023 39876 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7024 : RelOptInfo *grouped_rel,
7025 : RelOptInfo *partially_grouped_rel,
7026 : const AggClauseCosts *agg_costs,
7027 : grouping_sets_data *gd, double dNumGroups,
7028 : GroupPathExtraData *extra)
7029 : {
7030 39876 : Query *parse = root->parse;
7031 39876 : Path *cheapest_path = input_rel->cheapest_total_path;
7032 : ListCell *lc;
7033 39876 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7034 39876 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7035 39876 : List *havingQual = (List *) extra->havingQual;
7036 39876 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7037 :
7038 39876 : if (can_sort)
7039 : {
7040 : /*
7041 : * Use any available suitably-sorted path as input, and also consider
7042 : * sorting the cheapest-total path and incremental sort on any paths
7043 : * with presorted keys.
7044 : */
7045 82412 : foreach(lc, input_rel->pathlist)
7046 : {
7047 : ListCell *lc2;
7048 42542 : Path *path = (Path *) lfirst(lc);
7049 42542 : Path *path_save = path;
7050 42542 : List *pathkey_orderings = NIL;
7051 :
7052 : /* generate alternative group orderings that might be useful */
7053 42542 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7054 :
7055 : Assert(list_length(pathkey_orderings) > 0);
7056 :
7057 85228 : foreach(lc2, pathkey_orderings)
7058 : {
7059 42686 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7060 :
7061 : /* restore the path (we replace it in the loop) */
7062 42686 : path = path_save;
7063 :
7064 42686 : path = make_ordered_path(root,
7065 : grouped_rel,
7066 : path,
7067 : cheapest_path,
7068 : info->pathkeys,
7069 : -1.0);
7070 42686 : if (path == NULL)
7071 368 : continue;
7072 :
7073 : /* Now decide what to stick atop it */
7074 42318 : if (parse->groupingSets)
7075 : {
7076 938 : consider_groupingsets_paths(root, grouped_rel,
7077 : path, true, can_hash,
7078 : gd, agg_costs, dNumGroups);
7079 : }
7080 41380 : else if (parse->hasAggs)
7081 : {
7082 : /*
7083 : * We have aggregation, possibly with plain GROUP BY. Make
7084 : * an AggPath.
7085 : */
7086 40608 : add_path(grouped_rel, (Path *)
7087 40608 : create_agg_path(root,
7088 : grouped_rel,
7089 : path,
7090 40608 : grouped_rel->reltarget,
7091 40608 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7092 : AGGSPLIT_SIMPLE,
7093 : info->clauses,
7094 : havingQual,
7095 : agg_costs,
7096 : dNumGroups));
7097 : }
7098 772 : else if (parse->groupClause)
7099 : {
7100 : /*
7101 : * We have GROUP BY without aggregation or grouping sets.
7102 : * Make a GroupPath.
7103 : */
7104 772 : add_path(grouped_rel, (Path *)
7105 772 : create_group_path(root,
7106 : grouped_rel,
7107 : path,
7108 : info->clauses,
7109 : havingQual,
7110 : dNumGroups));
7111 : }
7112 : else
7113 : {
7114 : /* Other cases should have been handled above */
7115 : Assert(false);
7116 : }
7117 : }
7118 : }
7119 :
7120 : /*
7121 : * Instead of operating directly on the input relation, we can
7122 : * consider finalizing a partially aggregated path.
7123 : */
7124 39870 : if (partially_grouped_rel != NULL)
7125 : {
7126 3990 : foreach(lc, partially_grouped_rel->pathlist)
7127 : {
7128 : ListCell *lc2;
7129 2410 : Path *path = (Path *) lfirst(lc);
7130 2410 : Path *path_save = path;
7131 2410 : List *pathkey_orderings = NIL;
7132 :
7133 : /* generate alternative group orderings that might be useful */
7134 2410 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7135 :
7136 : Assert(list_length(pathkey_orderings) > 0);
7137 :
7138 : /* process all potentially interesting grouping reorderings */
7139 4820 : foreach(lc2, pathkey_orderings)
7140 : {
7141 2410 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7142 :
7143 : /* restore the path (we replace it in the loop) */
7144 2410 : path = path_save;
7145 :
7146 2410 : path = make_ordered_path(root,
7147 : grouped_rel,
7148 : path,
7149 2410 : partially_grouped_rel->cheapest_total_path,
7150 : info->pathkeys,
7151 : -1.0);
7152 :
7153 2410 : if (path == NULL)
7154 108 : continue;
7155 :
7156 2302 : if (parse->hasAggs)
7157 2054 : add_path(grouped_rel, (Path *)
7158 2054 : create_agg_path(root,
7159 : grouped_rel,
7160 : path,
7161 2054 : grouped_rel->reltarget,
7162 2054 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7163 : AGGSPLIT_FINAL_DESERIAL,
7164 : info->clauses,
7165 : havingQual,
7166 : agg_final_costs,
7167 : dNumGroups));
7168 : else
7169 248 : add_path(grouped_rel, (Path *)
7170 248 : create_group_path(root,
7171 : grouped_rel,
7172 : path,
7173 : info->clauses,
7174 : havingQual,
7175 : dNumGroups));
7176 :
7177 : }
7178 : }
7179 : }
7180 : }
7181 :
7182 39876 : if (can_hash)
7183 : {
7184 4980 : if (parse->groupingSets)
7185 : {
7186 : /*
7187 : * Try for a hash-only groupingsets path over unsorted input.
7188 : */
7189 794 : consider_groupingsets_paths(root, grouped_rel,
7190 : cheapest_path, false, true,
7191 : gd, agg_costs, dNumGroups);
7192 : }
7193 : else
7194 : {
7195 : /*
7196 : * Generate a HashAgg Path. We just need an Agg over the
7197 : * cheapest-total input path, since input order won't matter.
7198 : */
7199 4186 : add_path(grouped_rel, (Path *)
7200 4186 : create_agg_path(root, grouped_rel,
7201 : cheapest_path,
7202 4186 : grouped_rel->reltarget,
7203 : AGG_HASHED,
7204 : AGGSPLIT_SIMPLE,
7205 : root->processed_groupClause,
7206 : havingQual,
7207 : agg_costs,
7208 : dNumGroups));
7209 : }
7210 :
7211 : /*
7212 : * Generate a Finalize HashAgg Path atop of the cheapest partially
7213 : * grouped path, assuming there is one
7214 : */
7215 4980 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7216 : {
7217 784 : Path *path = partially_grouped_rel->cheapest_total_path;
7218 :
7219 784 : add_path(grouped_rel, (Path *)
7220 784 : create_agg_path(root,
7221 : grouped_rel,
7222 : path,
7223 784 : grouped_rel->reltarget,
7224 : AGG_HASHED,
7225 : AGGSPLIT_FINAL_DESERIAL,
7226 : root->processed_groupClause,
7227 : havingQual,
7228 : agg_final_costs,
7229 : dNumGroups));
7230 : }
7231 : }
7232 :
7233 : /*
7234 : * When partitionwise aggregate is used, we might have fully aggregated
7235 : * paths in the partial pathlist, because add_paths_to_append_rel() will
7236 : * consider a path for grouped_rel consisting of a Parallel Append of
7237 : * non-partial paths from each child.
7238 : */
7239 39876 : if (grouped_rel->partial_pathlist != NIL)
7240 162 : gather_grouping_paths(root, grouped_rel);
7241 39876 : }
7242 :
7243 : /*
7244 : * create_partial_grouping_paths
7245 : *
7246 : * Create a new upper relation representing the result of partial aggregation
7247 : * and populate it with appropriate paths. Note that we don't finalize the
7248 : * lists of paths here, so the caller can add additional partial or non-partial
7249 : * paths and must afterward call gather_grouping_paths and set_cheapest on
7250 : * the returned upper relation.
7251 : *
7252 : * All paths for this new upper relation -- both partial and non-partial --
7253 : * have been partially aggregated but require a subsequent FinalizeAggregate
7254 : * step.
7255 : *
7256 : * NB: This function is allowed to return NULL if it determines that there is
7257 : * no real need to create a new RelOptInfo.
7258 : */
7259 : static RelOptInfo *
7260 35472 : create_partial_grouping_paths(PlannerInfo *root,
7261 : RelOptInfo *grouped_rel,
7262 : RelOptInfo *input_rel,
7263 : grouping_sets_data *gd,
7264 : GroupPathExtraData *extra,
7265 : bool force_rel_creation)
7266 : {
7267 35472 : Query *parse = root->parse;
7268 : RelOptInfo *partially_grouped_rel;
7269 35472 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7270 35472 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7271 35472 : Path *cheapest_partial_path = NULL;
7272 35472 : Path *cheapest_total_path = NULL;
7273 35472 : double dNumPartialGroups = 0;
7274 35472 : double dNumPartialPartialGroups = 0;
7275 : ListCell *lc;
7276 35472 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7277 35472 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7278 :
7279 : /*
7280 : * Consider whether we should generate partially aggregated non-partial
7281 : * paths. We can only do this if we have a non-partial path, and only if
7282 : * the parent of the input rel is performing partial partitionwise
7283 : * aggregation. (Note that extra->patype is the type of partitionwise
7284 : * aggregation being used at the parent level, not this level.)
7285 : */
7286 35472 : if (input_rel->pathlist != NIL &&
7287 35472 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7288 618 : cheapest_total_path = input_rel->cheapest_total_path;
7289 :
7290 : /*
7291 : * If parallelism is possible for grouped_rel, then we should consider
7292 : * generating partially-grouped partial paths. However, if the input rel
7293 : * has no partial paths, then we can't.
7294 : */
7295 35472 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7296 1782 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7297 :
7298 : /*
7299 : * If we can't partially aggregate partial paths, and we can't partially
7300 : * aggregate non-partial paths, then don't bother creating the new
7301 : * RelOptInfo at all, unless the caller specified force_rel_creation.
7302 : */
7303 35472 : if (cheapest_total_path == NULL &&
7304 33372 : cheapest_partial_path == NULL &&
7305 33372 : !force_rel_creation)
7306 33274 : return NULL;
7307 :
7308 : /*
7309 : * Build a new upper relation to represent the result of partially
7310 : * aggregating the rows from the input relation.
7311 : */
7312 2198 : partially_grouped_rel = fetch_upper_rel(root,
7313 : UPPERREL_PARTIAL_GROUP_AGG,
7314 : grouped_rel->relids);
7315 2198 : partially_grouped_rel->consider_parallel =
7316 2198 : grouped_rel->consider_parallel;
7317 2198 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7318 2198 : partially_grouped_rel->serverid = grouped_rel->serverid;
7319 2198 : partially_grouped_rel->userid = grouped_rel->userid;
7320 2198 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7321 2198 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7322 :
7323 : /*
7324 : * Build target list for partial aggregate paths. These paths cannot just
7325 : * emit the same tlist as regular aggregate paths, because (1) we must
7326 : * include Vars and Aggrefs needed in HAVING, which might not appear in
7327 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7328 : */
7329 2198 : partially_grouped_rel->reltarget =
7330 2198 : make_partial_grouping_target(root, grouped_rel->reltarget,
7331 : extra->havingQual);
7332 :
7333 2198 : if (!extra->partial_costs_set)
7334 : {
7335 : /*
7336 : * Collect statistics about aggregates for estimating costs of
7337 : * performing aggregation in parallel.
7338 : */
7339 7752 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
7340 7752 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
7341 1292 : if (parse->hasAggs)
7342 : {
7343 : /* partial phase */
7344 1158 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7345 : agg_partial_costs);
7346 :
7347 : /* final phase */
7348 1158 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7349 : agg_final_costs);
7350 : }
7351 :
7352 1292 : extra->partial_costs_set = true;
7353 : }
7354 :
7355 : /* Estimate number of partial groups. */
7356 2198 : if (cheapest_total_path != NULL)
7357 : dNumPartialGroups =
7358 618 : get_number_of_groups(root,
7359 : cheapest_total_path->rows,
7360 : gd,
7361 : extra->targetList);
7362 2198 : if (cheapest_partial_path != NULL)
7363 : dNumPartialPartialGroups =
7364 1782 : get_number_of_groups(root,
7365 : cheapest_partial_path->rows,
7366 : gd,
7367 : extra->targetList);
7368 :
7369 2198 : if (can_sort && cheapest_total_path != NULL)
7370 : {
7371 : /* This should have been checked previously */
7372 : Assert(parse->hasAggs || parse->groupClause);
7373 :
7374 : /*
7375 : * Use any available suitably-sorted path as input, and also consider
7376 : * sorting the cheapest partial path.
7377 : */
7378 1236 : foreach(lc, input_rel->pathlist)
7379 : {
7380 : ListCell *lc2;
7381 618 : Path *path = (Path *) lfirst(lc);
7382 618 : Path *path_save = path;
7383 618 : List *pathkey_orderings = NIL;
7384 :
7385 : /* generate alternative group orderings that might be useful */
7386 618 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7387 :
7388 : Assert(list_length(pathkey_orderings) > 0);
7389 :
7390 : /* process all potentially interesting grouping reorderings */
7391 1236 : foreach(lc2, pathkey_orderings)
7392 : {
7393 618 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7394 :
7395 : /* restore the path (we replace it in the loop) */
7396 618 : path = path_save;
7397 :
7398 618 : path = make_ordered_path(root,
7399 : partially_grouped_rel,
7400 : path,
7401 : cheapest_total_path,
7402 : info->pathkeys,
7403 : -1.0);
7404 :
7405 618 : if (path == NULL)
7406 0 : continue;
7407 :
7408 618 : if (parse->hasAggs)
7409 546 : add_path(partially_grouped_rel, (Path *)
7410 546 : create_agg_path(root,
7411 : partially_grouped_rel,
7412 : path,
7413 546 : partially_grouped_rel->reltarget,
7414 546 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7415 : AGGSPLIT_INITIAL_SERIAL,
7416 : info->clauses,
7417 : NIL,
7418 : agg_partial_costs,
7419 : dNumPartialGroups));
7420 : else
7421 72 : add_path(partially_grouped_rel, (Path *)
7422 72 : create_group_path(root,
7423 : partially_grouped_rel,
7424 : path,
7425 : info->clauses,
7426 : NIL,
7427 : dNumPartialGroups));
7428 : }
7429 : }
7430 : }
7431 :
7432 2198 : if (can_sort && cheapest_partial_path != NULL)
7433 : {
7434 : /* Similar to above logic, but for partial paths. */
7435 3576 : foreach(lc, input_rel->partial_pathlist)
7436 : {
7437 : ListCell *lc2;
7438 1794 : Path *path = (Path *) lfirst(lc);
7439 1794 : Path *path_save = path;
7440 1794 : List *pathkey_orderings = NIL;
7441 :
7442 : /* generate alternative group orderings that might be useful */
7443 1794 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7444 :
7445 : Assert(list_length(pathkey_orderings) > 0);
7446 :
7447 : /* process all potentially interesting grouping reorderings */
7448 3588 : foreach(lc2, pathkey_orderings)
7449 : {
7450 1794 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7451 :
7452 :
7453 : /* restore the path (we replace it in the loop) */
7454 1794 : path = path_save;
7455 :
7456 1794 : path = make_ordered_path(root,
7457 : partially_grouped_rel,
7458 : path,
7459 : cheapest_partial_path,
7460 : info->pathkeys,
7461 : -1.0);
7462 :
7463 1794 : if (path == NULL)
7464 6 : continue;
7465 :
7466 1788 : if (parse->hasAggs)
7467 1666 : add_partial_path(partially_grouped_rel, (Path *)
7468 1666 : create_agg_path(root,
7469 : partially_grouped_rel,
7470 : path,
7471 1666 : partially_grouped_rel->reltarget,
7472 1666 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7473 : AGGSPLIT_INITIAL_SERIAL,
7474 : info->clauses,
7475 : NIL,
7476 : agg_partial_costs,
7477 : dNumPartialPartialGroups));
7478 : else
7479 122 : add_partial_path(partially_grouped_rel, (Path *)
7480 122 : create_group_path(root,
7481 : partially_grouped_rel,
7482 : path,
7483 : info->clauses,
7484 : NIL,
7485 : dNumPartialPartialGroups));
7486 : }
7487 : }
7488 : }
7489 :
7490 : /*
7491 : * Add a partially-grouped HashAgg Path where possible
7492 : */
7493 2198 : if (can_hash && cheapest_total_path != NULL)
7494 : {
7495 : /* Checked above */
7496 : Assert(parse->hasAggs || parse->groupClause);
7497 :
7498 618 : add_path(partially_grouped_rel, (Path *)
7499 618 : create_agg_path(root,
7500 : partially_grouped_rel,
7501 : cheapest_total_path,
7502 618 : partially_grouped_rel->reltarget,
7503 : AGG_HASHED,
7504 : AGGSPLIT_INITIAL_SERIAL,
7505 : root->processed_groupClause,
7506 : NIL,
7507 : agg_partial_costs,
7508 : dNumPartialGroups));
7509 : }
7510 :
7511 : /*
7512 : * Now add a partially-grouped HashAgg partial Path where possible
7513 : */
7514 2198 : if (can_hash && cheapest_partial_path != NULL)
7515 : {
7516 986 : add_partial_path(partially_grouped_rel, (Path *)
7517 986 : create_agg_path(root,
7518 : partially_grouped_rel,
7519 : cheapest_partial_path,
7520 986 : partially_grouped_rel->reltarget,
7521 : AGG_HASHED,
7522 : AGGSPLIT_INITIAL_SERIAL,
7523 : root->processed_groupClause,
7524 : NIL,
7525 : agg_partial_costs,
7526 : dNumPartialPartialGroups));
7527 : }
7528 :
7529 : /*
7530 : * If there is an FDW that's responsible for all baserels of the query,
7531 : * let it consider adding partially grouped ForeignPaths.
7532 : */
7533 2198 : if (partially_grouped_rel->fdwroutine &&
7534 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7535 : {
7536 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7537 :
7538 6 : fdwroutine->GetForeignUpperPaths(root,
7539 : UPPERREL_PARTIAL_GROUP_AGG,
7540 : input_rel, partially_grouped_rel,
7541 : extra);
7542 : }
7543 :
7544 2198 : return partially_grouped_rel;
7545 : }
7546 :
7547 : /*
7548 : * make_ordered_path
7549 : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7550 : * return NULL if it doesn't make sense to generate an ordered path in
7551 : * this case.
7552 : */
7553 : static Path *
7554 52512 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7555 : Path *cheapest_path, List *pathkeys, double limit_tuples)
7556 : {
7557 : bool is_sorted;
7558 : int presorted_keys;
7559 :
7560 52512 : is_sorted = pathkeys_count_contained_in(pathkeys,
7561 : path->pathkeys,
7562 : &presorted_keys);
7563 :
7564 52512 : if (!is_sorted)
7565 : {
7566 : /*
7567 : * Try at least sorting the cheapest path and also try incrementally
7568 : * sorting any path which is partially sorted already (no need to deal
7569 : * with paths which have presorted keys when incremental sort is
7570 : * disabled unless it's the cheapest input path).
7571 : */
7572 12892 : if (path != cheapest_path &&
7573 2034 : (presorted_keys == 0 || !enable_incremental_sort))
7574 1038 : return NULL;
7575 :
7576 : /*
7577 : * We've no need to consider both a sort and incremental sort. We'll
7578 : * just do a sort if there are no presorted keys and an incremental
7579 : * sort when there are presorted keys.
7580 : */
7581 11854 : if (presorted_keys == 0 || !enable_incremental_sort)
7582 10690 : path = (Path *) create_sort_path(root,
7583 : rel,
7584 : path,
7585 : pathkeys,
7586 : limit_tuples);
7587 : else
7588 1164 : path = (Path *) create_incremental_sort_path(root,
7589 : rel,
7590 : path,
7591 : pathkeys,
7592 : presorted_keys,
7593 : limit_tuples);
7594 : }
7595 :
7596 51474 : return path;
7597 : }
7598 :
7599 : /*
7600 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7601 : * grouping relation.
7602 : *
7603 : * generate_useful_gather_paths does most of the work, but we also consider a
7604 : * special case: we could try sorting the data by the group_pathkeys and then
7605 : * applying Gather Merge.
7606 : *
7607 : * NB: This function shouldn't be used for anything other than a grouped or
7608 : * partially grouped relation not only because of the fact that it explicitly
7609 : * references group_pathkeys but we pass "true" as the third argument to
7610 : * generate_useful_gather_paths().
7611 : */
7612 : static void
7613 1644 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7614 : {
7615 : ListCell *lc;
7616 : Path *cheapest_partial_path;
7617 : List *groupby_pathkeys;
7618 :
7619 : /*
7620 : * This occurs after any partial aggregation has taken place, so trim off
7621 : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7622 : */
7623 1644 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7624 18 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7625 : root->num_groupby_pathkeys);
7626 : else
7627 1626 : groupby_pathkeys = root->group_pathkeys;
7628 :
7629 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7630 1644 : generate_useful_gather_paths(root, rel, true);
7631 :
7632 1644 : cheapest_partial_path = linitial(rel->partial_pathlist);
7633 :
7634 : /* XXX Shouldn't this also consider the group-key-reordering? */
7635 3894 : foreach(lc, rel->partial_pathlist)
7636 : {
7637 2250 : Path *path = (Path *) lfirst(lc);
7638 : bool is_sorted;
7639 : int presorted_keys;
7640 : double total_groups;
7641 :
7642 2250 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7643 : path->pathkeys,
7644 : &presorted_keys);
7645 :
7646 2250 : if (is_sorted)
7647 1470 : continue;
7648 :
7649 : /*
7650 : * Try at least sorting the cheapest path and also try incrementally
7651 : * sorting any path which is partially sorted already (no need to deal
7652 : * with paths which have presorted keys when incremental sort is
7653 : * disabled unless it's the cheapest input path).
7654 : */
7655 780 : if (path != cheapest_partial_path &&
7656 0 : (presorted_keys == 0 || !enable_incremental_sort))
7657 0 : continue;
7658 :
7659 : /*
7660 : * We've no need to consider both a sort and incremental sort. We'll
7661 : * just do a sort if there are no presorted keys and an incremental
7662 : * sort when there are presorted keys.
7663 : */
7664 780 : if (presorted_keys == 0 || !enable_incremental_sort)
7665 780 : path = (Path *) create_sort_path(root, rel, path,
7666 : groupby_pathkeys,
7667 : -1.0);
7668 : else
7669 0 : path = (Path *) create_incremental_sort_path(root,
7670 : rel,
7671 : path,
7672 : groupby_pathkeys,
7673 : presorted_keys,
7674 : -1.0);
7675 780 : total_groups = compute_gather_rows(path);
7676 : path = (Path *)
7677 780 : create_gather_merge_path(root,
7678 : rel,
7679 : path,
7680 780 : rel->reltarget,
7681 : groupby_pathkeys,
7682 : NULL,
7683 : &total_groups);
7684 :
7685 780 : add_path(rel, path);
7686 : }
7687 1644 : }
7688 :
7689 : /*
7690 : * can_partial_agg
7691 : *
7692 : * Determines whether or not partial grouping and/or aggregation is possible.
7693 : * Returns true when possible, false otherwise.
7694 : */
7695 : static bool
7696 39000 : can_partial_agg(PlannerInfo *root)
7697 : {
7698 39000 : Query *parse = root->parse;
7699 :
7700 39000 : if (!parse->hasAggs && parse->groupClause == NIL)
7701 : {
7702 : /*
7703 : * We don't know how to do parallel aggregation unless we have either
7704 : * some aggregates or a grouping clause.
7705 : */
7706 0 : return false;
7707 : }
7708 39000 : else if (parse->groupingSets)
7709 : {
7710 : /* We don't know how to do grouping sets in parallel. */
7711 872 : return false;
7712 : }
7713 38128 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7714 : {
7715 : /* Insufficient support for partial mode. */
7716 4078 : return false;
7717 : }
7718 :
7719 : /* Everything looks good. */
7720 34050 : return true;
7721 : }
7722 :
7723 : /*
7724 : * apply_scanjoin_target_to_paths
7725 : *
7726 : * Adjust the final scan/join relation, and recursively all of its children,
7727 : * to generate the final scan/join target. It would be more correct to model
7728 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7729 : * for each child relation, but doing it this way is noticeably cheaper.
7730 : * Maybe that problem can be solved at some point, but for now we do this.
7731 : *
7732 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7733 : * the same expressions as the existing reltarget, so we need only insert the
7734 : * appropriate sortgroupref information. By avoiding the creation of
7735 : * projection paths we save effort both immediately and at plan creation time.
7736 : */
7737 : static void
7738 547078 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7739 : RelOptInfo *rel,
7740 : List *scanjoin_targets,
7741 : List *scanjoin_targets_contain_srfs,
7742 : bool scanjoin_target_parallel_safe,
7743 : bool tlist_same_exprs)
7744 : {
7745 547078 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7746 : PathTarget *scanjoin_target;
7747 : ListCell *lc;
7748 :
7749 : /* This recurses, so be paranoid. */
7750 547078 : check_stack_depth();
7751 :
7752 : /*
7753 : * If the rel is partitioned, we want to drop its existing paths and
7754 : * generate new ones. This function would still be correct if we kept the
7755 : * existing paths: we'd modify them to generate the correct target above
7756 : * the partitioning Append, and then they'd compete on cost with paths
7757 : * generating the target below the Append. However, in our current cost
7758 : * model the latter way is always the same or cheaper cost, so modifying
7759 : * the existing paths would just be useless work. Moreover, when the cost
7760 : * is the same, varying roundoff errors might sometimes allow an existing
7761 : * path to be picked, resulting in undesirable cross-platform plan
7762 : * variations. So we drop old paths and thereby force the work to be done
7763 : * below the Append, except in the case of a non-parallel-safe target.
7764 : *
7765 : * Some care is needed, because we have to allow
7766 : * generate_useful_gather_paths to see the old partial paths in the next
7767 : * stanza. Hence, zap the main pathlist here, then allow
7768 : * generate_useful_gather_paths to add path(s) to the main list, and
7769 : * finally zap the partial pathlist.
7770 : */
7771 547078 : if (rel_is_partitioned)
7772 12460 : rel->pathlist = NIL;
7773 :
7774 : /*
7775 : * If the scan/join target is not parallel-safe, partial paths cannot
7776 : * generate it.
7777 : */
7778 547078 : if (!scanjoin_target_parallel_safe)
7779 : {
7780 : /*
7781 : * Since we can't generate the final scan/join target in parallel
7782 : * workers, this is our last opportunity to use any partial paths that
7783 : * exist; so build Gather path(s) that use them and emit whatever the
7784 : * current reltarget is. We don't do this in the case where the
7785 : * target is parallel-safe, since we will be able to generate superior
7786 : * paths by doing it after the final scan/join target has been
7787 : * applied.
7788 : */
7789 79786 : generate_useful_gather_paths(root, rel, false);
7790 :
7791 : /* Can't use parallel query above this level. */
7792 79786 : rel->partial_pathlist = NIL;
7793 79786 : rel->consider_parallel = false;
7794 : }
7795 :
7796 : /* Finish dropping old paths for a partitioned rel, per comment above */
7797 547078 : if (rel_is_partitioned)
7798 12460 : rel->partial_pathlist = NIL;
7799 :
7800 : /* Extract SRF-free scan/join target. */
7801 547078 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7802 :
7803 : /*
7804 : * Apply the SRF-free scan/join target to each existing path.
7805 : *
7806 : * If the tlist exprs are the same, we can just inject the sortgroupref
7807 : * information into the existing pathtargets. Otherwise, replace each
7808 : * path with a projection path that generates the SRF-free scan/join
7809 : * target. This can't change the ordering of paths within rel->pathlist,
7810 : * so we just modify the list in place.
7811 : */
7812 1136010 : foreach(lc, rel->pathlist)
7813 : {
7814 588932 : Path *subpath = (Path *) lfirst(lc);
7815 :
7816 : /* Shouldn't have any parameterized paths anymore */
7817 : Assert(subpath->param_info == NULL);
7818 :
7819 588932 : if (tlist_same_exprs)
7820 208868 : subpath->pathtarget->sortgrouprefs =
7821 208868 : scanjoin_target->sortgrouprefs;
7822 : else
7823 : {
7824 : Path *newpath;
7825 :
7826 380064 : newpath = (Path *) create_projection_path(root, rel, subpath,
7827 : scanjoin_target);
7828 380064 : lfirst(lc) = newpath;
7829 : }
7830 : }
7831 :
7832 : /* Likewise adjust the targets for any partial paths. */
7833 566668 : foreach(lc, rel->partial_pathlist)
7834 : {
7835 19590 : Path *subpath = (Path *) lfirst(lc);
7836 :
7837 : /* Shouldn't have any parameterized paths anymore */
7838 : Assert(subpath->param_info == NULL);
7839 :
7840 19590 : if (tlist_same_exprs)
7841 15956 : subpath->pathtarget->sortgrouprefs =
7842 15956 : scanjoin_target->sortgrouprefs;
7843 : else
7844 : {
7845 : Path *newpath;
7846 :
7847 3634 : newpath = (Path *) create_projection_path(root, rel, subpath,
7848 : scanjoin_target);
7849 3634 : lfirst(lc) = newpath;
7850 : }
7851 : }
7852 :
7853 : /*
7854 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7855 : * atop each existing path. (Note that this function doesn't look at the
7856 : * cheapest-path fields, which is a good thing because they're bogus right
7857 : * now.)
7858 : */
7859 547078 : if (root->parse->hasTargetSRFs)
7860 12024 : adjust_paths_for_srfs(root, rel,
7861 : scanjoin_targets,
7862 : scanjoin_targets_contain_srfs);
7863 :
7864 : /*
7865 : * Update the rel's target to be the final (with SRFs) scan/join target.
7866 : * This now matches the actual output of all the paths, and we might get
7867 : * confused in createplan.c if they don't agree. We must do this now so
7868 : * that any append paths made in the next part will use the correct
7869 : * pathtarget (cf. create_append_path).
7870 : *
7871 : * Note that this is also necessary if GetForeignUpperPaths() gets called
7872 : * on the final scan/join relation or on any of its children, since the
7873 : * FDW might look at the rel's target to create ForeignPaths.
7874 : */
7875 547078 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7876 :
7877 : /*
7878 : * If the relation is partitioned, recursively apply the scan/join target
7879 : * to all partitions, and generate brand-new Append paths in which the
7880 : * scan/join target is computed below the Append rather than above it.
7881 : * Since Append is not projection-capable, that might save a separate
7882 : * Result node, and it also is important for partitionwise aggregate.
7883 : */
7884 547078 : if (rel_is_partitioned)
7885 : {
7886 12460 : List *live_children = NIL;
7887 : int i;
7888 :
7889 : /* Adjust each partition. */
7890 12460 : i = -1;
7891 35210 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
7892 : {
7893 22750 : RelOptInfo *child_rel = rel->part_rels[i];
7894 : AppendRelInfo **appinfos;
7895 : int nappinfos;
7896 22750 : List *child_scanjoin_targets = NIL;
7897 :
7898 : Assert(child_rel != NULL);
7899 :
7900 : /* Dummy children can be ignored. */
7901 22750 : if (IS_DUMMY_REL(child_rel))
7902 42 : continue;
7903 :
7904 : /* Translate scan/join targets for this child. */
7905 22708 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7906 : &nappinfos);
7907 45416 : foreach(lc, scanjoin_targets)
7908 : {
7909 22708 : PathTarget *target = lfirst_node(PathTarget, lc);
7910 :
7911 22708 : target = copy_pathtarget(target);
7912 22708 : target->exprs = (List *)
7913 22708 : adjust_appendrel_attrs(root,
7914 22708 : (Node *) target->exprs,
7915 : nappinfos, appinfos);
7916 22708 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7917 : target);
7918 : }
7919 22708 : pfree(appinfos);
7920 :
7921 : /* Recursion does the real work. */
7922 22708 : apply_scanjoin_target_to_paths(root, child_rel,
7923 : child_scanjoin_targets,
7924 : scanjoin_targets_contain_srfs,
7925 : scanjoin_target_parallel_safe,
7926 : tlist_same_exprs);
7927 :
7928 : /* Save non-dummy children for Append paths. */
7929 22708 : if (!IS_DUMMY_REL(child_rel))
7930 22708 : live_children = lappend(live_children, child_rel);
7931 : }
7932 :
7933 : /* Build new paths for this relation by appending child paths. */
7934 12460 : add_paths_to_append_rel(root, rel, live_children);
7935 : }
7936 :
7937 : /*
7938 : * Consider generating Gather or Gather Merge paths. We must only do this
7939 : * if the relation is parallel safe, and we don't do it for child rels to
7940 : * avoid creating multiple Gather nodes within the same plan. We must do
7941 : * this after all paths have been generated and before set_cheapest, since
7942 : * one of the generated paths may turn out to be the cheapest one.
7943 : */
7944 547078 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7945 175224 : generate_useful_gather_paths(root, rel, false);
7946 :
7947 : /*
7948 : * Reassess which paths are the cheapest, now that we've potentially added
7949 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7950 : * this relation.
7951 : */
7952 547078 : set_cheapest(rel);
7953 547078 : }
7954 :
7955 : /*
7956 : * create_partitionwise_grouping_paths
7957 : *
7958 : * If the partition keys of input relation are part of the GROUP BY clause, all
7959 : * the rows belonging to a given group come from a single partition. This
7960 : * allows aggregation/grouping over a partitioned relation to be broken down
7961 : * into aggregation/grouping on each partition. This should be no worse, and
7962 : * often better, than the normal approach.
7963 : *
7964 : * However, if the GROUP BY clause does not contain all the partition keys,
7965 : * rows from a given group may be spread across multiple partitions. In that
7966 : * case, we perform partial aggregation for each group, append the results,
7967 : * and then finalize aggregation. This is less certain to win than the
7968 : * previous case. It may win if the PartialAggregate stage greatly reduces
7969 : * the number of groups, because fewer rows will pass through the Append node.
7970 : * It may lose if we have lots of small groups.
7971 : */
7972 : static void
7973 562 : create_partitionwise_grouping_paths(PlannerInfo *root,
7974 : RelOptInfo *input_rel,
7975 : RelOptInfo *grouped_rel,
7976 : RelOptInfo *partially_grouped_rel,
7977 : const AggClauseCosts *agg_costs,
7978 : grouping_sets_data *gd,
7979 : PartitionwiseAggregateType patype,
7980 : GroupPathExtraData *extra)
7981 : {
7982 562 : List *grouped_live_children = NIL;
7983 562 : List *partially_grouped_live_children = NIL;
7984 562 : PathTarget *target = grouped_rel->reltarget;
7985 562 : bool partial_grouping_valid = true;
7986 : int i;
7987 :
7988 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7989 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7990 : partially_grouped_rel != NULL);
7991 :
7992 : /* Add paths for partitionwise aggregation/grouping. */
7993 562 : i = -1;
7994 2056 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
7995 : {
7996 1494 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
7997 : PathTarget *child_target;
7998 : AppendRelInfo **appinfos;
7999 : int nappinfos;
8000 : GroupPathExtraData child_extra;
8001 : RelOptInfo *child_grouped_rel;
8002 : RelOptInfo *child_partially_grouped_rel;
8003 :
8004 : Assert(child_input_rel != NULL);
8005 :
8006 : /* Dummy children can be ignored. */
8007 1494 : if (IS_DUMMY_REL(child_input_rel))
8008 0 : continue;
8009 :
8010 1494 : child_target = copy_pathtarget(target);
8011 :
8012 : /*
8013 : * Copy the given "extra" structure as is and then override the
8014 : * members specific to this child.
8015 : */
8016 1494 : memcpy(&child_extra, extra, sizeof(child_extra));
8017 :
8018 1494 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8019 : &nappinfos);
8020 :
8021 1494 : child_target->exprs = (List *)
8022 1494 : adjust_appendrel_attrs(root,
8023 1494 : (Node *) target->exprs,
8024 : nappinfos, appinfos);
8025 :
8026 : /* Translate havingQual and targetList. */
8027 1494 : child_extra.havingQual = (Node *)
8028 : adjust_appendrel_attrs(root,
8029 : extra->havingQual,
8030 : nappinfos, appinfos);
8031 1494 : child_extra.targetList = (List *)
8032 1494 : adjust_appendrel_attrs(root,
8033 1494 : (Node *) extra->targetList,
8034 : nappinfos, appinfos);
8035 :
8036 : /*
8037 : * extra->patype was the value computed for our parent rel; patype is
8038 : * the value for this relation. For the child, our value is its
8039 : * parent rel's value.
8040 : */
8041 1494 : child_extra.patype = patype;
8042 :
8043 : /*
8044 : * Create grouping relation to hold fully aggregated grouping and/or
8045 : * aggregation paths for the child.
8046 : */
8047 1494 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8048 : child_target,
8049 1494 : extra->target_parallel_safe,
8050 : child_extra.havingQual);
8051 :
8052 : /* Create grouping paths for this child relation. */
8053 1494 : create_ordinary_grouping_paths(root, child_input_rel,
8054 : child_grouped_rel,
8055 : agg_costs, gd, &child_extra,
8056 : &child_partially_grouped_rel);
8057 :
8058 1494 : if (child_partially_grouped_rel)
8059 : {
8060 : partially_grouped_live_children =
8061 906 : lappend(partially_grouped_live_children,
8062 : child_partially_grouped_rel);
8063 : }
8064 : else
8065 588 : partial_grouping_valid = false;
8066 :
8067 1494 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8068 : {
8069 876 : set_cheapest(child_grouped_rel);
8070 876 : grouped_live_children = lappend(grouped_live_children,
8071 : child_grouped_rel);
8072 : }
8073 :
8074 1494 : pfree(appinfos);
8075 : }
8076 :
8077 : /*
8078 : * Try to create append paths for partially grouped children. For full
8079 : * partitionwise aggregation, we might have paths in the partial_pathlist
8080 : * if parallel aggregation is possible. For partial partitionwise
8081 : * aggregation, we may have paths in both pathlist and partial_pathlist.
8082 : *
8083 : * NB: We must have a partially grouped path for every child in order to
8084 : * generate a partially grouped path for this relation.
8085 : */
8086 562 : if (partially_grouped_rel && partial_grouping_valid)
8087 : {
8088 : Assert(partially_grouped_live_children != NIL);
8089 :
8090 350 : add_paths_to_append_rel(root, partially_grouped_rel,
8091 : partially_grouped_live_children);
8092 :
8093 : /*
8094 : * We need call set_cheapest, since the finalization step will use the
8095 : * cheapest path from the rel.
8096 : */
8097 350 : if (partially_grouped_rel->pathlist)
8098 350 : set_cheapest(partially_grouped_rel);
8099 : }
8100 :
8101 : /* If possible, create append paths for fully grouped children. */
8102 562 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8103 : {
8104 : Assert(grouped_live_children != NIL);
8105 :
8106 320 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8107 : }
8108 562 : }
8109 :
8110 : /*
8111 : * group_by_has_partkey
8112 : *
8113 : * Returns true if all the partition keys of the given relation are part of
8114 : * the GROUP BY clauses, including having matching collation, false otherwise.
8115 : */
8116 : static bool
8117 556 : group_by_has_partkey(RelOptInfo *input_rel,
8118 : List *targetList,
8119 : List *groupClause)
8120 : {
8121 556 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8122 556 : int cnt = 0;
8123 : int partnatts;
8124 :
8125 : /* Input relation should be partitioned. */
8126 : Assert(input_rel->part_scheme);
8127 :
8128 : /* Rule out early, if there are no partition keys present. */
8129 556 : if (!input_rel->partexprs)
8130 0 : return false;
8131 :
8132 556 : partnatts = input_rel->part_scheme->partnatts;
8133 :
8134 912 : for (cnt = 0; cnt < partnatts; cnt++)
8135 : {
8136 592 : List *partexprs = input_rel->partexprs[cnt];
8137 : ListCell *lc;
8138 592 : bool found = false;
8139 :
8140 810 : foreach(lc, partexprs)
8141 : {
8142 : ListCell *lg;
8143 586 : Expr *partexpr = lfirst(lc);
8144 586 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8145 :
8146 924 : foreach(lg, groupexprs)
8147 : {
8148 706 : Expr *groupexpr = lfirst(lg);
8149 706 : Oid groupcoll = exprCollation((Node *) groupexpr);
8150 :
8151 : /*
8152 : * Note: we can assume there is at most one RelabelType node;
8153 : * eval_const_expressions() will have simplified if more than
8154 : * one.
8155 : */
8156 706 : if (IsA(groupexpr, RelabelType))
8157 24 : groupexpr = ((RelabelType *) groupexpr)->arg;
8158 :
8159 706 : if (equal(groupexpr, partexpr))
8160 : {
8161 : /*
8162 : * Reject a match if the grouping collation does not match
8163 : * the partitioning collation.
8164 : */
8165 368 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
8166 : partcoll != groupcoll)
8167 12 : return false;
8168 :
8169 356 : found = true;
8170 356 : break;
8171 : }
8172 : }
8173 :
8174 574 : if (found)
8175 356 : break;
8176 : }
8177 :
8178 : /*
8179 : * If none of the partition key expressions match with any of the
8180 : * GROUP BY expression, return false.
8181 : */
8182 580 : if (!found)
8183 224 : return false;
8184 : }
8185 :
8186 320 : return true;
8187 : }
8188 :
8189 : /*
8190 : * generate_setop_child_grouplist
8191 : * Build a SortGroupClause list defining the sort/grouping properties
8192 : * of the child of a set operation.
8193 : *
8194 : * This is similar to generate_setop_grouplist() but differs as the setop
8195 : * child query's targetlist entries may already have a tleSortGroupRef
8196 : * assigned for other purposes, such as GROUP BYs. Here we keep the
8197 : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8198 : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8199 : * any of the columns in the targetlist don't match to the setop's colTypes
8200 : * then we return an empty list. This may leave some TLEs with unreferenced
8201 : * ressortgroupref markings, but that's harmless.
8202 : */
8203 : static List *
8204 12278 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8205 : {
8206 12278 : List *grouplist = copyObject(op->groupClauses);
8207 : ListCell *lg;
8208 : ListCell *lt;
8209 : ListCell *ct;
8210 :
8211 12278 : lg = list_head(grouplist);
8212 12278 : ct = list_head(op->colTypes);
8213 47818 : foreach(lt, targetlist)
8214 : {
8215 35962 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8216 : SortGroupClause *sgc;
8217 : Oid coltype;
8218 :
8219 : /* resjunk columns could have sortgrouprefs. Leave these alone */
8220 35962 : if (tle->resjunk)
8221 0 : continue;
8222 :
8223 : /*
8224 : * We expect every non-resjunk target to have a SortGroupClause and
8225 : * colTypes.
8226 : */
8227 : Assert(lg != NULL);
8228 : Assert(ct != NULL);
8229 35962 : sgc = (SortGroupClause *) lfirst(lg);
8230 35962 : coltype = lfirst_oid(ct);
8231 :
8232 : /* reject if target type isn't the same as the setop target type */
8233 35962 : if (coltype != exprType((Node *) tle->expr))
8234 422 : return NIL;
8235 :
8236 35540 : lg = lnext(grouplist, lg);
8237 35540 : ct = lnext(op->colTypes, ct);
8238 :
8239 : /* assign a tleSortGroupRef, or reuse the existing one */
8240 35540 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8241 : }
8242 :
8243 : Assert(lg == NULL);
8244 : Assert(ct == NULL);
8245 :
8246 11856 : return grouplist;
8247 : }
|
