Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * planner.c
4 : * The query optimizer external interface.
5 : *
6 : * Portions Copyright (c) 1996-2019, 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/htup_details.h"
23 : #include "access/parallel.h"
24 : #include "access/sysattr.h"
25 : #include "access/table.h"
26 : #include "access/xact.h"
27 : #include "catalog/pg_constraint.h"
28 : #include "catalog/pg_inherits.h"
29 : #include "catalog/pg_proc.h"
30 : #include "catalog/pg_type.h"
31 : #include "executor/executor.h"
32 : #include "executor/nodeAgg.h"
33 : #include "foreign/fdwapi.h"
34 : #include "miscadmin.h"
35 : #include "jit/jit.h"
36 : #include "lib/bipartite_match.h"
37 : #include "lib/knapsack.h"
38 : #include "nodes/makefuncs.h"
39 : #include "nodes/nodeFuncs.h"
40 : #ifdef OPTIMIZER_DEBUG
41 : #include "nodes/print.h"
42 : #endif
43 : #include "optimizer/appendinfo.h"
44 : #include "optimizer/clauses.h"
45 : #include "optimizer/cost.h"
46 : #include "optimizer/inherit.h"
47 : #include "optimizer/optimizer.h"
48 : #include "optimizer/paramassign.h"
49 : #include "optimizer/pathnode.h"
50 : #include "optimizer/paths.h"
51 : #include "optimizer/plancat.h"
52 : #include "optimizer/planmain.h"
53 : #include "optimizer/planner.h"
54 : #include "optimizer/prep.h"
55 : #include "optimizer/subselect.h"
56 : #include "optimizer/tlist.h"
57 : #include "parser/analyze.h"
58 : #include "parser/parsetree.h"
59 : #include "parser/parse_agg.h"
60 : #include "partitioning/partdesc.h"
61 : #include "rewrite/rewriteManip.h"
62 : #include "storage/dsm_impl.h"
63 : #include "utils/rel.h"
64 : #include "utils/selfuncs.h"
65 : #include "utils/lsyscache.h"
66 : #include "utils/syscache.h"
67 :
68 :
69 : /* GUC parameters */
70 : double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
71 : int force_parallel_mode = FORCE_PARALLEL_OFF;
72 : bool parallel_leader_participation = true;
73 :
74 : /* Hook for plugins to get control in planner() */
75 : planner_hook_type planner_hook = NULL;
76 :
77 : /* Hook for plugins to get control when grouping_planner() plans upper rels */
78 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
79 :
80 :
81 : /* Expression kind codes for preprocess_expression */
82 : #define EXPRKIND_QUAL 0
83 : #define EXPRKIND_TARGET 1
84 : #define EXPRKIND_RTFUNC 2
85 : #define EXPRKIND_RTFUNC_LATERAL 3
86 : #define EXPRKIND_VALUES 4
87 : #define EXPRKIND_VALUES_LATERAL 5
88 : #define EXPRKIND_LIMIT 6
89 : #define EXPRKIND_APPINFO 7
90 : #define EXPRKIND_PHV 8
91 : #define EXPRKIND_TABLESAMPLE 9
92 : #define EXPRKIND_ARBITER_ELEM 10
93 : #define EXPRKIND_TABLEFUNC 11
94 : #define EXPRKIND_TABLEFUNC_LATERAL 12
95 :
96 : /* Passthrough data for standard_qp_callback */
97 : typedef struct
98 : {
99 : List *activeWindows; /* active windows, if any */
100 : List *groupClause; /* overrides parse->groupClause */
101 : } standard_qp_extra;
102 :
103 : /*
104 : * Data specific to grouping sets
105 : */
106 :
107 : typedef struct
108 : {
109 : List *rollups;
110 : List *hash_sets_idx;
111 : double dNumHashGroups;
112 : bool any_hashable;
113 : Bitmapset *unsortable_refs;
114 : Bitmapset *unhashable_refs;
115 : List *unsortable_sets;
116 : int *tleref_to_colnum_map;
117 : } grouping_sets_data;
118 :
119 : /*
120 : * Temporary structure for use during WindowClause reordering in order to be
121 : * able to sort WindowClauses on partitioning/ordering prefix.
122 : */
123 : typedef struct
124 : {
125 : WindowClause *wc;
126 : List *uniqueOrder; /* A List of unique ordering/partitioning
127 : * clauses per Window */
128 : } WindowClauseSortData;
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 inheritance_planner(PlannerInfo *root);
134 : static void grouping_planner(PlannerInfo *root, bool inheritance_update,
135 : double tuple_fraction);
136 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
137 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
138 : int *tleref_to_colnum_map);
139 : static void preprocess_rowmarks(PlannerInfo *root);
140 : static double preprocess_limit(PlannerInfo *root,
141 : double tuple_fraction,
142 : int64 *offset_est, int64 *count_est);
143 : static void remove_useless_groupby_columns(PlannerInfo *root);
144 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
145 : static List *extract_rollup_sets(List *groupingSets);
146 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
147 : static void standard_qp_callback(PlannerInfo *root, void *extra);
148 : static double get_number_of_groups(PlannerInfo *root,
149 : double path_rows,
150 : grouping_sets_data *gd,
151 : List *target_list);
152 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
153 : RelOptInfo *input_rel,
154 : PathTarget *target,
155 : bool target_parallel_safe,
156 : const AggClauseCosts *agg_costs,
157 : grouping_sets_data *gd);
158 : static bool is_degenerate_grouping(PlannerInfo *root);
159 : static void create_degenerate_grouping_paths(PlannerInfo *root,
160 : RelOptInfo *input_rel,
161 : RelOptInfo *grouped_rel);
162 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
163 : PathTarget *target, bool target_parallel_safe,
164 : Node *havingQual);
165 : static void create_ordinary_grouping_paths(PlannerInfo *root,
166 : RelOptInfo *input_rel,
167 : RelOptInfo *grouped_rel,
168 : const AggClauseCosts *agg_costs,
169 : grouping_sets_data *gd,
170 : GroupPathExtraData *extra,
171 : RelOptInfo **partially_grouped_rel_p);
172 : static void consider_groupingsets_paths(PlannerInfo *root,
173 : RelOptInfo *grouped_rel,
174 : Path *path,
175 : bool is_sorted,
176 : bool can_hash,
177 : grouping_sets_data *gd,
178 : const AggClauseCosts *agg_costs,
179 : double dNumGroups);
180 : static RelOptInfo *create_window_paths(PlannerInfo *root,
181 : RelOptInfo *input_rel,
182 : PathTarget *input_target,
183 : PathTarget *output_target,
184 : bool output_target_parallel_safe,
185 : WindowFuncLists *wflists,
186 : List *activeWindows);
187 : static void create_one_window_path(PlannerInfo *root,
188 : RelOptInfo *window_rel,
189 : Path *path,
190 : PathTarget *input_target,
191 : PathTarget *output_target,
192 : WindowFuncLists *wflists,
193 : List *activeWindows);
194 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
195 : RelOptInfo *input_rel);
196 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
197 : RelOptInfo *input_rel,
198 : PathTarget *target,
199 : bool target_parallel_safe,
200 : double limit_tuples);
201 : static PathTarget *make_group_input_target(PlannerInfo *root,
202 : PathTarget *final_target);
203 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
204 : PathTarget *grouping_target,
205 : Node *havingQual);
206 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
207 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
208 : static PathTarget *make_window_input_target(PlannerInfo *root,
209 : PathTarget *final_target,
210 : List *activeWindows);
211 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
212 : List *tlist);
213 : static PathTarget *make_sort_input_target(PlannerInfo *root,
214 : PathTarget *final_target,
215 : bool *have_postponed_srfs);
216 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
217 : List *targets, List *targets_contain_srfs);
218 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
219 : RelOptInfo *grouped_rel,
220 : RelOptInfo *partially_grouped_rel,
221 : const AggClauseCosts *agg_costs,
222 : grouping_sets_data *gd,
223 : double dNumGroups,
224 : GroupPathExtraData *extra);
225 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
226 : RelOptInfo *grouped_rel,
227 : RelOptInfo *input_rel,
228 : grouping_sets_data *gd,
229 : GroupPathExtraData *extra,
230 : bool force_rel_creation);
231 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
232 : static bool can_partial_agg(PlannerInfo *root,
233 : const AggClauseCosts *agg_costs);
234 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
235 : RelOptInfo *rel,
236 : List *scanjoin_targets,
237 : List *scanjoin_targets_contain_srfs,
238 : bool scanjoin_target_parallel_safe,
239 : bool tlist_same_exprs);
240 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
241 : RelOptInfo *input_rel,
242 : RelOptInfo *grouped_rel,
243 : RelOptInfo *partially_grouped_rel,
244 : const AggClauseCosts *agg_costs,
245 : grouping_sets_data *gd,
246 : PartitionwiseAggregateType patype,
247 : GroupPathExtraData *extra);
248 : static bool group_by_has_partkey(RelOptInfo *input_rel,
249 : List *targetList,
250 : List *groupClause);
251 : static int common_prefix_cmp(const void *a, const void *b);
252 :
253 :
254 : /*****************************************************************************
255 : *
256 : * Query optimizer entry point
257 : *
258 : * To support loadable plugins that monitor or modify planner behavior,
259 : * we provide a hook variable that lets a plugin get control before and
260 : * after the standard planning process. The plugin would normally call
261 : * standard_planner().
262 : *
263 : * Note to plugin authors: standard_planner() scribbles on its Query input,
264 : * so you'd better copy that data structure if you want to plan more than once.
265 : *
266 : *****************************************************************************/
267 : PlannedStmt *
268 236424 : planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
269 : {
270 : PlannedStmt *result;
271 :
272 236424 : if (planner_hook)
273 0 : result = (*planner_hook) (parse, cursorOptions, boundParams);
274 : else
275 236424 : result = standard_planner(parse, cursorOptions, boundParams);
276 234924 : return result;
277 : }
278 :
279 : PlannedStmt *
280 236424 : standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
281 : {
282 : PlannedStmt *result;
283 : PlannerGlobal *glob;
284 : double tuple_fraction;
285 : PlannerInfo *root;
286 : RelOptInfo *final_rel;
287 : Path *best_path;
288 : Plan *top_plan;
289 : ListCell *lp,
290 : *lr;
291 :
292 : /*
293 : * Set up global state for this planner invocation. This data is needed
294 : * across all levels of sub-Query that might exist in the given command,
295 : * so we keep it in a separate struct that's linked to by each per-Query
296 : * PlannerInfo.
297 : */
298 236424 : glob = makeNode(PlannerGlobal);
299 :
300 236424 : glob->boundParams = boundParams;
301 236424 : glob->subplans = NIL;
302 236424 : glob->subroots = NIL;
303 236424 : glob->rewindPlanIDs = NULL;
304 236424 : glob->finalrtable = NIL;
305 236424 : glob->finalrowmarks = NIL;
306 236424 : glob->resultRelations = NIL;
307 236424 : glob->rootResultRelations = NIL;
308 236424 : glob->relationOids = NIL;
309 236424 : glob->invalItems = NIL;
310 236424 : glob->paramExecTypes = NIL;
311 236424 : glob->lastPHId = 0;
312 236424 : glob->lastRowMarkId = 0;
313 236424 : glob->lastPlanNodeId = 0;
314 236424 : glob->transientPlan = false;
315 236424 : glob->dependsOnRole = false;
316 :
317 : /*
318 : * Assess whether it's feasible to use parallel mode for this query. We
319 : * can't do this in a standalone backend, or if the command will try to
320 : * modify any data, or if this is a cursor operation, or if GUCs are set
321 : * to values that don't permit parallelism, or if parallel-unsafe
322 : * functions are present in the query tree.
323 : *
324 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
325 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
326 : * the command is writing into a completely new table which workers won't
327 : * be able to see. If the workers could see the table, the fact that
328 : * group locking would cause them to ignore the leader's heavyweight
329 : * relation extension lock and GIN page locks would make this unsafe.
330 : * We'll have to fix that somehow if we want to allow parallel inserts in
331 : * general; updates and deletes have additional problems especially around
332 : * combo CIDs.)
333 : *
334 : * For now, we don't try to use parallel mode if we're running inside a
335 : * parallel worker. We might eventually be able to relax this
336 : * restriction, but for now it seems best not to have parallel workers
337 : * trying to create their own parallel workers.
338 : */
339 236424 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
340 194556 : IsUnderPostmaster &&
341 351468 : parse->commandType == CMD_SELECT &&
342 313740 : !parse->hasModifyingCTE &&
343 313350 : max_parallel_workers_per_gather > 0 &&
344 156522 : !IsParallelWorker())
345 : {
346 : /* all the cheap tests pass, so scan the query tree */
347 156518 : glob->maxParallelHazard = max_parallel_hazard(parse);
348 156518 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
349 : }
350 : else
351 : {
352 : /* skip the query tree scan, just assume it's unsafe */
353 79906 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
354 79906 : glob->parallelModeOK = false;
355 : }
356 :
357 : /*
358 : * glob->parallelModeNeeded is normally set to false here and changed to
359 : * true during plan creation if a Gather or Gather Merge plan is actually
360 : * created (cf. create_gather_plan, create_gather_merge_plan).
361 : *
362 : * However, if force_parallel_mode = on or force_parallel_mode = regress,
363 : * then we impose parallel mode whenever it's safe to do so, even if the
364 : * final plan doesn't use parallelism. It's not safe to do so if the
365 : * query contains anything parallel-unsafe; parallelModeOK will be false
366 : * in that case. Note that parallelModeOK can't change after this point.
367 : * Otherwise, everything in the query is either parallel-safe or
368 : * parallel-restricted, and in either case it should be OK to impose
369 : * parallel-mode restrictions. If that ends up breaking something, then
370 : * either some function the user included in the query is incorrectly
371 : * labelled as parallel-safe or parallel-restricted when in reality it's
372 : * parallel-unsafe, or else the query planner itself has a bug.
373 : */
374 371086 : glob->parallelModeNeeded = glob->parallelModeOK &&
375 134662 : (force_parallel_mode != FORCE_PARALLEL_OFF);
376 :
377 : /* Determine what fraction of the plan is likely to be scanned */
378 236424 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
379 : {
380 : /*
381 : * We have no real idea how many tuples the user will ultimately FETCH
382 : * from a cursor, but it is often the case that he doesn't want 'em
383 : * all, or would prefer a fast-start plan anyway so that he can
384 : * process some of the tuples sooner. Use a GUC parameter to decide
385 : * what fraction to optimize for.
386 : */
387 1974 : tuple_fraction = cursor_tuple_fraction;
388 :
389 : /*
390 : * We document cursor_tuple_fraction as simply being a fraction, which
391 : * means the edge cases 0 and 1 have to be treated specially here. We
392 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
393 : */
394 1974 : if (tuple_fraction >= 1.0)
395 0 : tuple_fraction = 0.0;
396 1974 : else if (tuple_fraction <= 0.0)
397 0 : tuple_fraction = 1e-10;
398 : }
399 : else
400 : {
401 : /* Default assumption is we need all the tuples */
402 234450 : tuple_fraction = 0.0;
403 : }
404 :
405 : /* primary planning entry point (may recurse for subqueries) */
406 236424 : root = subquery_planner(glob, parse, NULL,
407 : false, tuple_fraction);
408 :
409 : /* Select best Path and turn it into a Plan */
410 235044 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
411 235044 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
412 :
413 235044 : top_plan = create_plan(root, best_path);
414 :
415 : /*
416 : * If creating a plan for a scrollable cursor, make sure it can run
417 : * backwards on demand. Add a Material node at the top at need.
418 : */
419 234924 : if (cursorOptions & CURSOR_OPT_SCROLL)
420 : {
421 148 : if (!ExecSupportsBackwardScan(top_plan))
422 16 : top_plan = materialize_finished_plan(top_plan);
423 : }
424 :
425 : /*
426 : * Optionally add a Gather node for testing purposes, provided this is
427 : * actually a safe thing to do.
428 : */
429 234924 : if (force_parallel_mode != FORCE_PARALLEL_OFF && top_plan->parallel_safe)
430 : {
431 14 : Gather *gather = makeNode(Gather);
432 :
433 : /*
434 : * If there are any initPlans attached to the formerly-top plan node,
435 : * move them up to the Gather node; same as we do for Material node in
436 : * materialize_finished_plan.
437 : */
438 14 : gather->plan.initPlan = top_plan->initPlan;
439 14 : top_plan->initPlan = NIL;
440 :
441 14 : gather->plan.targetlist = top_plan->targetlist;
442 14 : gather->plan.qual = NIL;
443 14 : gather->plan.lefttree = top_plan;
444 14 : gather->plan.righttree = NULL;
445 14 : gather->num_workers = 1;
446 14 : gather->single_copy = true;
447 14 : gather->invisible = (force_parallel_mode == FORCE_PARALLEL_REGRESS);
448 :
449 : /*
450 : * Since this Gather has no parallel-aware descendants to signal to,
451 : * we don't need a rescan Param.
452 : */
453 14 : gather->rescan_param = -1;
454 :
455 : /*
456 : * Ideally we'd use cost_gather here, but setting up dummy path data
457 : * to satisfy it doesn't seem much cleaner than knowing what it does.
458 : */
459 14 : gather->plan.startup_cost = top_plan->startup_cost +
460 : parallel_setup_cost;
461 28 : gather->plan.total_cost = top_plan->total_cost +
462 14 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
463 14 : gather->plan.plan_rows = top_plan->plan_rows;
464 14 : gather->plan.plan_width = top_plan->plan_width;
465 14 : gather->plan.parallel_aware = false;
466 14 : gather->plan.parallel_safe = false;
467 :
468 : /* use parallel mode for parallel plans. */
469 14 : root->glob->parallelModeNeeded = true;
470 :
471 14 : top_plan = &gather->plan;
472 : }
473 :
474 : /*
475 : * If any Params were generated, run through the plan tree and compute
476 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
477 : * set_plan_references' tree traversal, but for now it has to be separate
478 : * because we need to visit subplans before not after main plan.
479 : */
480 234924 : if (glob->paramExecTypes != NIL)
481 : {
482 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
483 151410 : forboth(lp, glob->subplans, lr, glob->subroots)
484 : {
485 44812 : Plan *subplan = (Plan *) lfirst(lp);
486 44812 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
487 :
488 44812 : SS_finalize_plan(subroot, subplan);
489 : }
490 106598 : SS_finalize_plan(root, top_plan);
491 : }
492 :
493 : /* final cleanup of the plan */
494 : Assert(glob->finalrtable == NIL);
495 : Assert(glob->finalrowmarks == NIL);
496 : Assert(glob->resultRelations == NIL);
497 : Assert(glob->rootResultRelations == NIL);
498 234924 : top_plan = set_plan_references(root, top_plan);
499 : /* ... and the subplans (both regular subplans and initplans) */
500 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
501 279736 : forboth(lp, glob->subplans, lr, glob->subroots)
502 : {
503 44812 : Plan *subplan = (Plan *) lfirst(lp);
504 44812 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
505 :
506 44812 : lfirst(lp) = set_plan_references(subroot, subplan);
507 : }
508 :
509 : /* build the PlannedStmt result */
510 234924 : result = makeNode(PlannedStmt);
511 :
512 234924 : result->commandType = parse->commandType;
513 234924 : result->queryId = parse->queryId;
514 234924 : result->hasReturning = (parse->returningList != NIL);
515 234924 : result->hasModifyingCTE = parse->hasModifyingCTE;
516 234924 : result->canSetTag = parse->canSetTag;
517 234924 : result->transientPlan = glob->transientPlan;
518 234924 : result->dependsOnRole = glob->dependsOnRole;
519 234924 : result->parallelModeNeeded = glob->parallelModeNeeded;
520 234924 : result->planTree = top_plan;
521 234924 : result->rtable = glob->finalrtable;
522 234924 : result->resultRelations = glob->resultRelations;
523 234924 : result->rootResultRelations = glob->rootResultRelations;
524 234924 : result->subplans = glob->subplans;
525 234924 : result->rewindPlanIDs = glob->rewindPlanIDs;
526 234924 : result->rowMarks = glob->finalrowmarks;
527 234924 : result->relationOids = glob->relationOids;
528 234924 : result->invalItems = glob->invalItems;
529 234924 : result->paramExecTypes = glob->paramExecTypes;
530 : /* utilityStmt should be null, but we might as well copy it */
531 234924 : result->utilityStmt = parse->utilityStmt;
532 234924 : result->stmt_location = parse->stmt_location;
533 234924 : result->stmt_len = parse->stmt_len;
534 :
535 234924 : result->jitFlags = PGJIT_NONE;
536 469848 : if (jit_enabled && jit_above_cost >= 0 &&
537 234924 : top_plan->total_cost > jit_above_cost)
538 : {
539 536 : result->jitFlags |= PGJIT_PERFORM;
540 :
541 : /*
542 : * Decide how much effort should be put into generating better code.
543 : */
544 1072 : if (jit_optimize_above_cost >= 0 &&
545 536 : top_plan->total_cost > jit_optimize_above_cost)
546 466 : result->jitFlags |= PGJIT_OPT3;
547 1072 : if (jit_inline_above_cost >= 0 &&
548 536 : top_plan->total_cost > jit_inline_above_cost)
549 466 : result->jitFlags |= PGJIT_INLINE;
550 :
551 : /*
552 : * Decide which operations should be JITed.
553 : */
554 536 : if (jit_expressions)
555 536 : result->jitFlags |= PGJIT_EXPR;
556 536 : if (jit_tuple_deforming)
557 536 : result->jitFlags |= PGJIT_DEFORM;
558 : }
559 :
560 234924 : if (glob->partition_directory != NULL)
561 3128 : DestroyPartitionDirectory(glob->partition_directory);
562 :
563 234924 : return result;
564 : }
565 :
566 :
567 : /*--------------------
568 : * subquery_planner
569 : * Invokes the planner on a subquery. We recurse to here for each
570 : * sub-SELECT found in the query tree.
571 : *
572 : * glob is the global state for the current planner run.
573 : * parse is the querytree produced by the parser & rewriter.
574 : * parent_root is the immediate parent Query's info (NULL at the top level).
575 : * hasRecursion is true if this is a recursive WITH query.
576 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
577 : * tuple_fraction is interpreted as explained for grouping_planner, below.
578 : *
579 : * Basically, this routine does the stuff that should only be done once
580 : * per Query object. It then calls grouping_planner. At one time,
581 : * grouping_planner could be invoked recursively on the same Query object;
582 : * that's not currently true, but we keep the separation between the two
583 : * routines anyway, in case we need it again someday.
584 : *
585 : * subquery_planner will be called recursively to handle sub-Query nodes
586 : * found within the query's expressions and rangetable.
587 : *
588 : * Returns the PlannerInfo struct ("root") that contains all data generated
589 : * while planning the subquery. In particular, the Path(s) attached to
590 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
591 : * cheapest way(s) to implement the query. The top level will select the
592 : * best Path and pass it through createplan.c to produce a finished Plan.
593 : *--------------------
594 : */
595 : PlannerInfo *
596 289088 : subquery_planner(PlannerGlobal *glob, Query *parse,
597 : PlannerInfo *parent_root,
598 : bool hasRecursion, double tuple_fraction)
599 : {
600 : PlannerInfo *root;
601 : List *newWithCheckOptions;
602 : List *newHaving;
603 : bool hasOuterJoins;
604 : bool hasResultRTEs;
605 : RelOptInfo *final_rel;
606 : ListCell *l;
607 :
608 : /* Create a PlannerInfo data structure for this subquery */
609 289088 : root = makeNode(PlannerInfo);
610 289088 : root->parse = parse;
611 289088 : root->glob = glob;
612 289088 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
613 289088 : root->parent_root = parent_root;
614 289088 : root->plan_params = NIL;
615 289088 : root->outer_params = NULL;
616 289088 : root->planner_cxt = CurrentMemoryContext;
617 289088 : root->init_plans = NIL;
618 289088 : root->cte_plan_ids = NIL;
619 289088 : root->multiexpr_params = NIL;
620 289088 : root->eq_classes = NIL;
621 289088 : root->append_rel_list = NIL;
622 289088 : root->rowMarks = NIL;
623 289088 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
624 289088 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
625 289088 : root->processed_tlist = NIL;
626 289088 : root->grouping_map = NULL;
627 289088 : root->minmax_aggs = NIL;
628 289088 : root->qual_security_level = 0;
629 289088 : root->inhTargetKind = INHKIND_NONE;
630 289088 : root->hasRecursion = hasRecursion;
631 289088 : if (hasRecursion)
632 326 : root->wt_param_id = assign_special_exec_param(root);
633 : else
634 288762 : root->wt_param_id = -1;
635 289088 : root->non_recursive_path = NULL;
636 289088 : root->partColsUpdated = false;
637 :
638 : /*
639 : * If there is a WITH list, process each WITH query and either convert it
640 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
641 : */
642 289088 : if (parse->cteList)
643 1158 : SS_process_ctes(root);
644 :
645 : /*
646 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
647 : * that we don't need so many special cases to deal with that situation.
648 : */
649 289088 : replace_empty_jointree(parse);
650 :
651 : /*
652 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
653 : * to transform them into joins. Note that this step does not descend
654 : * into subqueries; if we pull up any subqueries below, their SubLinks are
655 : * processed just before pulling them up.
656 : */
657 289088 : if (parse->hasSubLinks)
658 28834 : pull_up_sublinks(root);
659 :
660 : /*
661 : * Scan the rangetable for set-returning functions, and inline them if
662 : * possible (producing subqueries that might get pulled up next).
663 : * Recursion issues here are handled in the same way as for SubLinks.
664 : */
665 289088 : inline_set_returning_functions(root);
666 :
667 : /*
668 : * Check to see if any subqueries in the jointree can be merged into this
669 : * query.
670 : */
671 289088 : pull_up_subqueries(root);
672 :
673 : /*
674 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
675 : * do this now because it requires applying pull_up_subqueries to the leaf
676 : * queries of the UNION ALL, which weren't touched above because they
677 : * weren't referenced by the jointree (they will be after we do this).
678 : */
679 289084 : if (parse->setOperations)
680 2444 : flatten_simple_union_all(root);
681 :
682 : /*
683 : * Survey the rangetable to see what kinds of entries are present. We can
684 : * skip some later processing if relevant SQL features are not used; for
685 : * example if there are no JOIN RTEs we can avoid the expense of doing
686 : * flatten_join_alias_vars(). This must be done after we have finished
687 : * adding rangetable entries, of course. (Note: actually, processing of
688 : * inherited or partitioned rels can cause RTEs for their child tables to
689 : * get added later; but those must all be RTE_RELATION entries, so they
690 : * don't invalidate the conclusions drawn here.)
691 : */
692 289084 : root->hasJoinRTEs = false;
693 289084 : root->hasLateralRTEs = false;
694 289084 : hasOuterJoins = false;
695 289084 : hasResultRTEs = false;
696 828466 : foreach(l, parse->rtable)
697 : {
698 539382 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
699 :
700 539382 : switch (rte->rtekind)
701 : {
702 : case RTE_RELATION:
703 300806 : if (rte->inh)
704 : {
705 : /*
706 : * Check to see if the relation actually has any children;
707 : * if not, clear the inh flag so we can treat it as a
708 : * plain base relation.
709 : *
710 : * Note: this could give a false-positive result, if the
711 : * rel once had children but no longer does. We used to
712 : * be able to clear rte->inh later on when we discovered
713 : * that, but no more; we have to handle such cases as
714 : * full-fledged inheritance.
715 : */
716 206936 : rte->inh = has_subclass(rte->relid);
717 : }
718 300806 : break;
719 : case RTE_JOIN:
720 48108 : root->hasJoinRTEs = true;
721 48108 : if (IS_OUTER_JOIN(rte->jointype))
722 34442 : hasOuterJoins = true;
723 48108 : break;
724 : case RTE_RESULT:
725 109754 : hasResultRTEs = true;
726 109754 : break;
727 : default:
728 : /* No work here for other RTE types */
729 80714 : break;
730 : }
731 :
732 539382 : if (rte->lateral)
733 952 : root->hasLateralRTEs = true;
734 :
735 : /*
736 : * We can also determine the maximum security level required for any
737 : * securityQuals now. Addition of inheritance-child RTEs won't affect
738 : * this, because child tables don't have their own securityQuals; see
739 : * expand_single_inheritance_child().
740 : */
741 539382 : if (rte->securityQuals)
742 1124 : root->qual_security_level = Max(root->qual_security_level,
743 : list_length(rte->securityQuals));
744 : }
745 :
746 : /*
747 : * Preprocess RowMark information. We need to do this after subquery
748 : * pullup, so that all base relations are present.
749 : */
750 289084 : preprocess_rowmarks(root);
751 :
752 : /*
753 : * Set hasHavingQual to remember if HAVING clause is present. Needed
754 : * because preprocess_expression will reduce a constant-true condition to
755 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
756 : */
757 289084 : root->hasHavingQual = (parse->havingQual != NULL);
758 :
759 : /* Clear this flag; might get set in distribute_qual_to_rels */
760 289084 : root->hasPseudoConstantQuals = false;
761 :
762 : /*
763 : * Do expression preprocessing on targetlist and quals, as well as other
764 : * random expressions in the querytree. Note that we do not need to
765 : * handle sort/group expressions explicitly, because they are actually
766 : * part of the targetlist.
767 : */
768 287726 : parse->targetList = (List *)
769 289084 : preprocess_expression(root, (Node *) parse->targetList,
770 : EXPRKIND_TARGET);
771 :
772 : /* Constant-folding might have removed all set-returning functions */
773 287726 : if (parse->hasTargetSRFs)
774 3578 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
775 :
776 287726 : newWithCheckOptions = NIL;
777 288868 : foreach(l, parse->withCheckOptions)
778 : {
779 1142 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
780 :
781 1142 : wco->qual = preprocess_expression(root, wco->qual,
782 : EXPRKIND_QUAL);
783 1142 : if (wco->qual != NULL)
784 922 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
785 : }
786 287726 : parse->withCheckOptions = newWithCheckOptions;
787 :
788 287726 : parse->returningList = (List *)
789 287726 : preprocess_expression(root, (Node *) parse->returningList,
790 : EXPRKIND_TARGET);
791 :
792 287726 : preprocess_qual_conditions(root, (Node *) parse->jointree);
793 :
794 287726 : parse->havingQual = preprocess_expression(root, parse->havingQual,
795 : EXPRKIND_QUAL);
796 :
797 288874 : foreach(l, parse->windowClause)
798 : {
799 1148 : WindowClause *wc = lfirst_node(WindowClause, l);
800 :
801 : /* partitionClause/orderClause are sort/group expressions */
802 1148 : wc->startOffset = preprocess_expression(root, wc->startOffset,
803 : EXPRKIND_LIMIT);
804 1148 : wc->endOffset = preprocess_expression(root, wc->endOffset,
805 : EXPRKIND_LIMIT);
806 : }
807 :
808 287726 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
809 : EXPRKIND_LIMIT);
810 287726 : parse->limitCount = preprocess_expression(root, parse->limitCount,
811 : EXPRKIND_LIMIT);
812 :
813 287726 : if (parse->onConflict)
814 : {
815 1896 : parse->onConflict->arbiterElems = (List *)
816 948 : preprocess_expression(root,
817 948 : (Node *) parse->onConflict->arbiterElems,
818 : EXPRKIND_ARBITER_ELEM);
819 1896 : parse->onConflict->arbiterWhere =
820 948 : preprocess_expression(root,
821 948 : parse->onConflict->arbiterWhere,
822 : EXPRKIND_QUAL);
823 1896 : parse->onConflict->onConflictSet = (List *)
824 948 : preprocess_expression(root,
825 948 : (Node *) parse->onConflict->onConflictSet,
826 : EXPRKIND_TARGET);
827 1896 : parse->onConflict->onConflictWhere =
828 948 : preprocess_expression(root,
829 948 : parse->onConflict->onConflictWhere,
830 : EXPRKIND_QUAL);
831 : /* exclRelTlist contains only Vars, so no preprocessing needed */
832 : }
833 :
834 287726 : root->append_rel_list = (List *)
835 287726 : preprocess_expression(root, (Node *) root->append_rel_list,
836 : EXPRKIND_APPINFO);
837 :
838 : /* Also need to preprocess expressions within RTEs */
839 825640 : foreach(l, parse->rtable)
840 : {
841 537914 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
842 : int kind;
843 : ListCell *lcsq;
844 :
845 537914 : if (rte->rtekind == RTE_RELATION)
846 : {
847 300692 : if (rte->tablesample)
848 152 : rte->tablesample = (TableSampleClause *)
849 152 : preprocess_expression(root,
850 152 : (Node *) rte->tablesample,
851 : EXPRKIND_TABLESAMPLE);
852 : }
853 237222 : else if (rte->rtekind == RTE_SUBQUERY)
854 : {
855 : /*
856 : * We don't want to do all preprocessing yet on the subquery's
857 : * expressions, since that will happen when we plan it. But if it
858 : * contains any join aliases of our level, those have to get
859 : * expanded now, because planning of the subquery won't do it.
860 : * That's only possible if the subquery is LATERAL.
861 : */
862 47214 : if (rte->lateral && root->hasJoinRTEs)
863 304 : rte->subquery = (Query *)
864 304 : flatten_join_alias_vars(root->parse,
865 304 : (Node *) rte->subquery);
866 : }
867 190008 : else if (rte->rtekind == RTE_FUNCTION)
868 : {
869 : /* Preprocess the function expression(s) fully */
870 28204 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
871 28204 : rte->functions = (List *)
872 28204 : preprocess_expression(root, (Node *) rte->functions, kind);
873 : }
874 161804 : else if (rte->rtekind == RTE_TABLEFUNC)
875 : {
876 : /* Preprocess the function expression(s) fully */
877 140 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
878 140 : rte->tablefunc = (TableFunc *)
879 140 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
880 : }
881 161664 : else if (rte->rtekind == RTE_VALUES)
882 : {
883 : /* Preprocess the values lists fully */
884 3738 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
885 3738 : rte->values_lists = (List *)
886 3738 : preprocess_expression(root, (Node *) rte->values_lists, kind);
887 : }
888 :
889 : /*
890 : * Process each element of the securityQuals list as if it were a
891 : * separate qual expression (as indeed it is). We need to do it this
892 : * way to get proper canonicalization of AND/OR structure. Note that
893 : * this converts each element into an implicit-AND sublist.
894 : */
895 539190 : foreach(lcsq, rte->securityQuals)
896 : {
897 1276 : lfirst(lcsq) = preprocess_expression(root,
898 1276 : (Node *) lfirst(lcsq),
899 : EXPRKIND_QUAL);
900 : }
901 : }
902 :
903 : /*
904 : * Now that we are done preprocessing expressions, and in particular done
905 : * flattening join alias variables, get rid of the joinaliasvars lists.
906 : * They no longer match what expressions in the rest of the tree look
907 : * like, because we have not preprocessed expressions in those lists (and
908 : * do not want to; for example, expanding a SubLink there would result in
909 : * a useless unreferenced subplan). Leaving them in place simply creates
910 : * a hazard for later scans of the tree. We could try to prevent that by
911 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
912 : * but that doesn't sound very reliable.
913 : */
914 287726 : if (root->hasJoinRTEs)
915 : {
916 208674 : foreach(l, parse->rtable)
917 : {
918 172544 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
919 :
920 172544 : rte->joinaliasvars = NIL;
921 : }
922 : }
923 :
924 : /*
925 : * In some cases we may want to transfer a HAVING clause into WHERE. We
926 : * cannot do so if the HAVING clause contains aggregates (obviously) or
927 : * volatile functions (since a HAVING clause is supposed to be executed
928 : * only once per group). We also can't do this if there are any nonempty
929 : * grouping sets; moving such a clause into WHERE would potentially change
930 : * the results, if any referenced column isn't present in all the grouping
931 : * sets. (If there are only empty grouping sets, then the HAVING clause
932 : * must be degenerate as discussed below.)
933 : *
934 : * Also, it may be that the clause is so expensive to execute that we're
935 : * better off doing it only once per group, despite the loss of
936 : * selectivity. This is hard to estimate short of doing the entire
937 : * planning process twice, so we use a heuristic: clauses containing
938 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
939 : * clause into WHERE, in hopes of eliminating tuples before aggregation
940 : * instead of after.
941 : *
942 : * If the query has explicit grouping then we can simply move such a
943 : * clause into WHERE; any group that fails the clause will not be in the
944 : * output because none of its tuples will reach the grouping or
945 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
946 : * HAVING clause, which we put in WHERE so that query_planner() can use it
947 : * in a gating Result node, but also keep in HAVING to ensure that we
948 : * don't emit a bogus aggregated row. (This could be done better, but it
949 : * seems not worth optimizing.)
950 : *
951 : * Note that both havingQual and parse->jointree->quals are in
952 : * implicitly-ANDed-list form at this point, even though they are declared
953 : * as Node *.
954 : */
955 287726 : newHaving = NIL;
956 288164 : foreach(l, (List *) parse->havingQual)
957 : {
958 438 : Node *havingclause = (Node *) lfirst(l);
959 :
960 848 : if ((parse->groupClause && parse->groupingSets) ||
961 526 : contain_agg_clause(havingclause) ||
962 232 : contain_volatile_functions(havingclause) ||
963 116 : contain_subplans(havingclause))
964 : {
965 : /* keep it in HAVING */
966 322 : newHaving = lappend(newHaving, havingclause);
967 : }
968 116 : else if (parse->groupClause && !parse->groupingSets)
969 : {
970 : /* move it to WHERE */
971 208 : parse->jointree->quals = (Node *)
972 104 : lappend((List *) parse->jointree->quals, havingclause);
973 : }
974 : else
975 : {
976 : /* put a copy in WHERE, keep it in HAVING */
977 24 : parse->jointree->quals = (Node *)
978 12 : lappend((List *) parse->jointree->quals,
979 : copyObject(havingclause));
980 12 : newHaving = lappend(newHaving, havingclause);
981 : }
982 : }
983 287726 : parse->havingQual = (Node *) newHaving;
984 :
985 : /* Remove any redundant GROUP BY columns */
986 287726 : remove_useless_groupby_columns(root);
987 :
988 : /*
989 : * If we have any outer joins, try to reduce them to plain inner joins.
990 : * This step is most easily done after we've done expression
991 : * preprocessing.
992 : */
993 287726 : if (hasOuterJoins)
994 29514 : reduce_outer_joins(root);
995 :
996 : /*
997 : * If we have any RTE_RESULT relations, see if they can be deleted from
998 : * the jointree. This step is most effectively done after we've done
999 : * expression preprocessing and outer join reduction.
1000 : */
1001 287726 : if (hasResultRTEs)
1002 108212 : remove_useless_result_rtes(root);
1003 :
1004 : /*
1005 : * Do the main planning. If we have an inherited target relation, that
1006 : * needs special processing, else go straight to grouping_planner.
1007 : */
1008 357162 : if (parse->resultRelation &&
1009 69436 : rt_fetch(parse->resultRelation, parse->rtable)->inh)
1010 974 : inheritance_planner(root);
1011 : else
1012 286752 : grouping_planner(root, false, tuple_fraction);
1013 :
1014 : /*
1015 : * Capture the set of outer-level param IDs we have access to, for use in
1016 : * extParam/allParam calculations later.
1017 : */
1018 287708 : SS_identify_outer_params(root);
1019 :
1020 : /*
1021 : * If any initPlans were created in this query level, adjust the surviving
1022 : * Paths' costs and parallel-safety flags to account for them. The
1023 : * initPlans won't actually get attached to the plan tree till
1024 : * create_plan() runs, but we must include their effects now.
1025 : */
1026 287708 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1027 287708 : SS_charge_for_initplans(root, final_rel);
1028 :
1029 : /*
1030 : * Make sure we've identified the cheapest Path for the final rel. (By
1031 : * doing this here not in grouping_planner, we include initPlan costs in
1032 : * the decision, though it's unlikely that will change anything.)
1033 : */
1034 287708 : set_cheapest(final_rel);
1035 :
1036 287708 : return root;
1037 : }
1038 :
1039 : /*
1040 : * preprocess_expression
1041 : * Do subquery_planner's preprocessing work for an expression,
1042 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1043 : * conditions), a HAVING clause, or a few other things.
1044 : */
1045 : static Node *
1046 2125474 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1047 : {
1048 : /*
1049 : * Fall out quickly if expression is empty. This occurs often enough to
1050 : * be worth checking. Note that null->null is the correct conversion for
1051 : * implicit-AND result format, too.
1052 : */
1053 2125474 : if (expr == NULL)
1054 1621124 : return NULL;
1055 :
1056 : /*
1057 : * If the query has any join RTEs, replace join alias variables with
1058 : * base-relation variables. We must do this first, since any expressions
1059 : * we may extract from the joinaliasvars lists have not been preprocessed.
1060 : * For example, if we did this after sublink processing, sublinks expanded
1061 : * out from join aliases would not get processed. But we can skip this in
1062 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1063 : * they can't contain any Vars of the current query level.
1064 : */
1065 504350 : if (root->hasJoinRTEs &&
1066 235324 : !(kind == EXPRKIND_RTFUNC ||
1067 117550 : kind == EXPRKIND_VALUES ||
1068 : kind == EXPRKIND_TABLESAMPLE ||
1069 : kind == EXPRKIND_TABLEFUNC))
1070 117550 : expr = flatten_join_alias_vars(root->parse, expr);
1071 :
1072 : /*
1073 : * Simplify constant expressions.
1074 : *
1075 : * Note: an essential effect of this is to convert named-argument function
1076 : * calls to positional notation and insert the current actual values of
1077 : * any default arguments for functions. To ensure that happens, we *must*
1078 : * process all expressions here. Previous PG versions sometimes skipped
1079 : * const-simplification if it didn't seem worth the trouble, but we can't
1080 : * do that anymore.
1081 : *
1082 : * Note: this also flattens nested AND and OR expressions into N-argument
1083 : * form. All processing of a qual expression after this point must be
1084 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1085 : * with AND directly under AND, nor OR directly under OR.
1086 : */
1087 504350 : expr = eval_const_expressions(root, expr);
1088 :
1089 : /*
1090 : * If it's a qual or havingQual, canonicalize it.
1091 : */
1092 502992 : if (kind == EXPRKIND_QUAL)
1093 : {
1094 178746 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1095 :
1096 : #ifdef OPTIMIZER_DEBUG
1097 : printf("After canonicalize_qual()\n");
1098 : pprint(expr);
1099 : #endif
1100 : }
1101 :
1102 : /* Expand SubLinks to SubPlans */
1103 502992 : if (root->parse->hasSubLinks)
1104 101016 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1105 :
1106 : /*
1107 : * XXX do not insert anything here unless you have grokked the comments in
1108 : * SS_replace_correlation_vars ...
1109 : */
1110 :
1111 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1112 502992 : if (root->query_level > 1)
1113 123800 : expr = SS_replace_correlation_vars(root, expr);
1114 :
1115 : /*
1116 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1117 : * don't want to do this before eval_const_expressions, since the latter
1118 : * would be unable to simplify a top-level AND correctly. Also,
1119 : * SS_process_sublinks expects explicit-AND format.)
1120 : */
1121 502992 : if (kind == EXPRKIND_QUAL)
1122 178746 : expr = (Node *) make_ands_implicit((Expr *) expr);
1123 :
1124 502992 : return expr;
1125 : }
1126 :
1127 : /*
1128 : * preprocess_qual_conditions
1129 : * Recursively scan the query's jointree and do subquery_planner's
1130 : * preprocessing work on each qual condition found therein.
1131 : */
1132 : static void
1133 713176 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1134 : {
1135 713176 : if (jtnode == NULL)
1136 0 : return;
1137 713176 : if (IsA(jtnode, RangeTblRef))
1138 : {
1139 : /* nothing to do here */
1140 : }
1141 356996 : else if (IsA(jtnode, FromExpr))
1142 : {
1143 300970 : FromExpr *f = (FromExpr *) jtnode;
1144 : ListCell *l;
1145 :
1146 614368 : foreach(l, f->fromlist)
1147 313398 : preprocess_qual_conditions(root, lfirst(l));
1148 :
1149 300970 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1150 : }
1151 56026 : else if (IsA(jtnode, JoinExpr))
1152 : {
1153 56026 : JoinExpr *j = (JoinExpr *) jtnode;
1154 :
1155 56026 : preprocess_qual_conditions(root, j->larg);
1156 56026 : preprocess_qual_conditions(root, j->rarg);
1157 :
1158 56026 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1159 : }
1160 : else
1161 0 : elog(ERROR, "unrecognized node type: %d",
1162 : (int) nodeTag(jtnode));
1163 : }
1164 :
1165 : /*
1166 : * preprocess_phv_expression
1167 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1168 : *
1169 : * If a LATERAL subquery references an output of another subquery, and that
1170 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1171 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1172 : * and later pull it back up during find_lateral_references, which runs after
1173 : * subquery_planner has preprocessed all the expressions that were in the
1174 : * current query level to start with. So we need to preprocess it then.
1175 : */
1176 : Expr *
1177 24 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1178 : {
1179 24 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1180 : }
1181 :
1182 : /*
1183 : * inheritance_planner
1184 : * Generate Paths in the case where the result relation is an
1185 : * inheritance set.
1186 : *
1187 : * We have to handle this case differently from cases where a source relation
1188 : * is an inheritance set. Source inheritance is expanded at the bottom of the
1189 : * plan tree (see allpaths.c), but target inheritance has to be expanded at
1190 : * the top. The reason is that for UPDATE, each target relation needs a
1191 : * different targetlist matching its own column set. Fortunately,
1192 : * the UPDATE/DELETE target can never be the nullable side of an outer join,
1193 : * so it's OK to generate the plan this way.
1194 : *
1195 : * Returns nothing; the useful output is in the Paths we attach to
1196 : * the (UPPERREL_FINAL, NULL) upperrel stored in *root.
1197 : *
1198 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1199 : * to leave this to the caller.
1200 : */
1201 : static void
1202 974 : inheritance_planner(PlannerInfo *root)
1203 : {
1204 974 : Query *parse = root->parse;
1205 974 : int top_parentRTindex = parse->resultRelation;
1206 : List *select_rtable;
1207 : List *select_appinfos;
1208 : List *child_appinfos;
1209 : List *old_child_rtis;
1210 : List *new_child_rtis;
1211 : Bitmapset *subqueryRTindexes;
1212 : Index next_subquery_rti;
1213 974 : int nominalRelation = -1;
1214 974 : Index rootRelation = 0;
1215 974 : List *final_rtable = NIL;
1216 974 : List *final_rowmarks = NIL;
1217 974 : int save_rel_array_size = 0;
1218 974 : RelOptInfo **save_rel_array = NULL;
1219 974 : AppendRelInfo **save_append_rel_array = NULL;
1220 974 : List *subpaths = NIL;
1221 974 : List *subroots = NIL;
1222 974 : List *resultRelations = NIL;
1223 974 : List *withCheckOptionLists = NIL;
1224 974 : List *returningLists = NIL;
1225 : List *rowMarks;
1226 : RelOptInfo *final_rel;
1227 : ListCell *lc;
1228 : ListCell *lc2;
1229 : Index rti;
1230 : RangeTblEntry *parent_rte;
1231 : Bitmapset *parent_relids;
1232 : Query **parent_parses;
1233 :
1234 : /* Should only get here for UPDATE or DELETE */
1235 : Assert(parse->commandType == CMD_UPDATE ||
1236 : parse->commandType == CMD_DELETE);
1237 :
1238 : /*
1239 : * We generate a modified instance of the original Query for each target
1240 : * relation, plan that, and put all the plans into a list that will be
1241 : * controlled by a single ModifyTable node. All the instances share the
1242 : * same rangetable, but each instance must have its own set of subquery
1243 : * RTEs within the finished rangetable because (1) they are likely to get
1244 : * scribbled on during planning, and (2) it's not inconceivable that
1245 : * subqueries could get planned differently in different cases. We need
1246 : * not create duplicate copies of other RTE kinds, in particular not the
1247 : * target relations, because they don't have either of those issues. Not
1248 : * having to duplicate the target relations is important because doing so
1249 : * (1) would result in a rangetable of length O(N^2) for N targets, with
1250 : * at least O(N^3) work expended here; and (2) would greatly complicate
1251 : * management of the rowMarks list.
1252 : *
1253 : * To begin with, generate a bitmapset of the relids of the subquery RTEs.
1254 : */
1255 974 : subqueryRTindexes = NULL;
1256 974 : rti = 1;
1257 2356 : foreach(lc, parse->rtable)
1258 : {
1259 1382 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1260 :
1261 1382 : if (rte->rtekind == RTE_SUBQUERY)
1262 108 : subqueryRTindexes = bms_add_member(subqueryRTindexes, rti);
1263 1382 : rti++;
1264 : }
1265 :
1266 : /*
1267 : * If the parent RTE is a partitioned table, we should use that as the
1268 : * nominal target relation, because the RTEs added for partitioned tables
1269 : * (including the root parent) as child members of the inheritance set do
1270 : * not appear anywhere else in the plan, so the confusion explained below
1271 : * for non-partitioning inheritance cases is not possible.
1272 : */
1273 974 : parent_rte = rt_fetch(top_parentRTindex, parse->rtable);
1274 : Assert(parent_rte->inh);
1275 974 : if (parent_rte->relkind == RELKIND_PARTITIONED_TABLE)
1276 : {
1277 600 : nominalRelation = top_parentRTindex;
1278 600 : rootRelation = top_parentRTindex;
1279 : }
1280 :
1281 : /*
1282 : * Before generating the real per-child-relation plans, do a cycle of
1283 : * planning as though the query were a SELECT. The objective here is to
1284 : * find out which child relations need to be processed, using the same
1285 : * expansion and pruning logic as for a SELECT. We'll then pull out the
1286 : * RangeTblEntry-s generated for the child rels, and make use of the
1287 : * AppendRelInfo entries for them to guide the real planning. (This is
1288 : * rather inefficient; we could perhaps stop short of making a full Path
1289 : * tree. But this whole function is inefficient and slated for
1290 : * destruction, so let's not contort query_planner for that.)
1291 : */
1292 : {
1293 : PlannerInfo *subroot;
1294 :
1295 : /*
1296 : * Flat-copy the PlannerInfo to prevent modification of the original.
1297 : */
1298 974 : subroot = makeNode(PlannerInfo);
1299 974 : memcpy(subroot, root, sizeof(PlannerInfo));
1300 :
1301 : /*
1302 : * Make a deep copy of the parsetree for this planning cycle to mess
1303 : * around with, and change it to look like a SELECT. (Hack alert: the
1304 : * target RTE still has updatedCols set if this is an UPDATE, so that
1305 : * expand_partitioned_rtentry will correctly update
1306 : * subroot->partColsUpdated.)
1307 : */
1308 974 : subroot->parse = copyObject(root->parse);
1309 :
1310 974 : subroot->parse->commandType = CMD_SELECT;
1311 974 : subroot->parse->resultRelation = 0;
1312 :
1313 : /*
1314 : * Ensure the subroot has its own copy of the original
1315 : * append_rel_list, since it'll be scribbled on. (Note that at this
1316 : * point, the list only contains AppendRelInfos for flattened UNION
1317 : * ALL subqueries.)
1318 : */
1319 974 : subroot->append_rel_list = copyObject(root->append_rel_list);
1320 :
1321 : /*
1322 : * Better make a private copy of the rowMarks, too.
1323 : */
1324 974 : subroot->rowMarks = copyObject(root->rowMarks);
1325 :
1326 : /* There shouldn't be any OJ info to translate, as yet */
1327 : Assert(subroot->join_info_list == NIL);
1328 : /* and we haven't created PlaceHolderInfos, either */
1329 : Assert(subroot->placeholder_list == NIL);
1330 :
1331 : /* Generate Path(s) for accessing this result relation */
1332 974 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1333 :
1334 : /* Extract the info we need. */
1335 974 : select_rtable = subroot->parse->rtable;
1336 974 : select_appinfos = subroot->append_rel_list;
1337 :
1338 : /*
1339 : * We need to propagate partColsUpdated back, too. (The later
1340 : * planning cycles will not set this because they won't run
1341 : * expand_partitioned_rtentry for the UPDATE target.)
1342 : */
1343 974 : root->partColsUpdated = subroot->partColsUpdated;
1344 : }
1345 :
1346 : /*----------
1347 : * Since only one rangetable can exist in the final plan, we need to make
1348 : * sure that it contains all the RTEs needed for any child plan. This is
1349 : * complicated by the need to use separate subquery RTEs for each child.
1350 : * We arrange the final rtable as follows:
1351 : * 1. All original rtable entries (with their original RT indexes).
1352 : * 2. All the relation RTEs generated for children of the target table.
1353 : * 3. Subquery RTEs for children after the first. We need N * (K - 1)
1354 : * RT slots for this, if there are N subqueries and K child tables.
1355 : * 4. Additional RTEs generated during the child planning runs, such as
1356 : * children of inheritable RTEs other than the target table.
1357 : * We assume that each child planning run will create an identical set
1358 : * of type-4 RTEs.
1359 : *
1360 : * So the next thing to do is append the type-2 RTEs (the target table's
1361 : * children) to the original rtable. We look through select_appinfos
1362 : * to find them.
1363 : *
1364 : * To identify which AppendRelInfos are relevant as we thumb through
1365 : * select_appinfos, we need to look for both direct and indirect children
1366 : * of top_parentRTindex, so we use a bitmap of known parent relids.
1367 : * expand_inherited_rtentry() always processes a parent before any of that
1368 : * parent's children, so we should see an intermediate parent before its
1369 : * children.
1370 : *----------
1371 : */
1372 974 : child_appinfos = NIL;
1373 974 : old_child_rtis = NIL;
1374 974 : new_child_rtis = NIL;
1375 974 : parent_relids = bms_make_singleton(top_parentRTindex);
1376 3616 : foreach(lc, select_appinfos)
1377 : {
1378 2642 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1379 : RangeTblEntry *child_rte;
1380 :
1381 : /* append_rel_list contains all append rels; ignore others */
1382 2642 : if (!bms_is_member(appinfo->parent_relid, parent_relids))
1383 224 : continue;
1384 :
1385 : /* remember relevant AppendRelInfos for use below */
1386 2418 : child_appinfos = lappend(child_appinfos, appinfo);
1387 :
1388 : /* extract RTE for this child rel */
1389 2418 : child_rte = rt_fetch(appinfo->child_relid, select_rtable);
1390 :
1391 : /* and append it to the original rtable */
1392 2418 : parse->rtable = lappend(parse->rtable, child_rte);
1393 :
1394 : /* remember child's index in the SELECT rtable */
1395 2418 : old_child_rtis = lappend_int(old_child_rtis, appinfo->child_relid);
1396 :
1397 : /* and its new index in the final rtable */
1398 2418 : new_child_rtis = lappend_int(new_child_rtis, list_length(parse->rtable));
1399 :
1400 : /* if child is itself partitioned, update parent_relids */
1401 2418 : if (child_rte->inh)
1402 : {
1403 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1404 276 : parent_relids = bms_add_member(parent_relids, appinfo->child_relid);
1405 : }
1406 : }
1407 :
1408 : /*
1409 : * It's possible that the RTIs we just assigned for the child rels in the
1410 : * final rtable are different from what they were in the SELECT query.
1411 : * Adjust the AppendRelInfos so that they will correctly map RT indexes to
1412 : * the final indexes. We can do this left-to-right since no child rel's
1413 : * final RT index could be greater than what it had in the SELECT query.
1414 : */
1415 3392 : forboth(lc, old_child_rtis, lc2, new_child_rtis)
1416 : {
1417 2418 : int old_child_rti = lfirst_int(lc);
1418 2418 : int new_child_rti = lfirst_int(lc2);
1419 :
1420 2418 : if (old_child_rti == new_child_rti)
1421 2402 : continue; /* nothing to do */
1422 :
1423 : Assert(old_child_rti > new_child_rti);
1424 :
1425 16 : ChangeVarNodes((Node *) child_appinfos,
1426 : old_child_rti, new_child_rti, 0);
1427 : }
1428 :
1429 : /*
1430 : * Now set up rangetable entries for subqueries for additional children
1431 : * (the first child will just use the original ones). These all have to
1432 : * look more or less real, or EXPLAIN will get unhappy; so we just make
1433 : * them all clones of the original subqueries.
1434 : */
1435 974 : next_subquery_rti = list_length(parse->rtable) + 1;
1436 974 : if (subqueryRTindexes != NULL)
1437 : {
1438 68 : int n_children = list_length(child_appinfos);
1439 :
1440 256 : while (n_children-- > 1)
1441 : {
1442 120 : int oldrti = -1;
1443 :
1444 432 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1445 : {
1446 : RangeTblEntry *subqrte;
1447 :
1448 192 : subqrte = rt_fetch(oldrti, parse->rtable);
1449 192 : parse->rtable = lappend(parse->rtable, copyObject(subqrte));
1450 : }
1451 : }
1452 : }
1453 :
1454 : /*
1455 : * The query for each child is obtained by translating the query for its
1456 : * immediate parent, since the AppendRelInfo data we have shows deltas
1457 : * between parents and children. We use the parent_parses array to
1458 : * remember the appropriate query trees. This is indexed by parent relid.
1459 : * Since the maximum number of parents is limited by the number of RTEs in
1460 : * the SELECT query, we use that number to allocate the array. An extra
1461 : * entry is needed since relids start from 1.
1462 : */
1463 974 : parent_parses = (Query **) palloc0((list_length(select_rtable) + 1) *
1464 : sizeof(Query *));
1465 974 : parent_parses[top_parentRTindex] = parse;
1466 :
1467 : /*
1468 : * And now we can get on with generating a plan for each child table.
1469 : */
1470 3392 : foreach(lc, child_appinfos)
1471 : {
1472 2418 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1473 2418 : Index this_subquery_rti = next_subquery_rti;
1474 : Query *parent_parse;
1475 : PlannerInfo *subroot;
1476 : RangeTblEntry *child_rte;
1477 : RelOptInfo *sub_final_rel;
1478 : Path *subpath;
1479 :
1480 : /*
1481 : * expand_inherited_rtentry() always processes a parent before any of
1482 : * that parent's children, so the parent query for this relation
1483 : * should already be available.
1484 : */
1485 2418 : parent_parse = parent_parses[appinfo->parent_relid];
1486 : Assert(parent_parse != NULL);
1487 :
1488 : /*
1489 : * We need a working copy of the PlannerInfo so that we can control
1490 : * propagation of information back to the main copy.
1491 : */
1492 2418 : subroot = makeNode(PlannerInfo);
1493 2418 : memcpy(subroot, root, sizeof(PlannerInfo));
1494 :
1495 : /*
1496 : * Generate modified query with this rel as target. We first apply
1497 : * adjust_appendrel_attrs, which copies the Query and changes
1498 : * references to the parent RTE to refer to the current child RTE,
1499 : * then fool around with subquery RTEs.
1500 : */
1501 2418 : subroot->parse = (Query *)
1502 2418 : adjust_appendrel_attrs(subroot,
1503 : (Node *) parent_parse,
1504 : 1, &appinfo);
1505 :
1506 : /*
1507 : * If there are securityQuals attached to the parent, move them to the
1508 : * child rel (they've already been transformed properly for that).
1509 : */
1510 2418 : parent_rte = rt_fetch(appinfo->parent_relid, subroot->parse->rtable);
1511 2418 : child_rte = rt_fetch(appinfo->child_relid, subroot->parse->rtable);
1512 2418 : child_rte->securityQuals = parent_rte->securityQuals;
1513 2418 : parent_rte->securityQuals = NIL;
1514 :
1515 : /*
1516 : * Mark whether we're planning a query to a partitioned table or an
1517 : * inheritance parent.
1518 : */
1519 2418 : subroot->inhTargetKind =
1520 2418 : (rootRelation != 0) ? INHKIND_PARTITIONED : INHKIND_INHERITED;
1521 :
1522 : /*
1523 : * If this child is further partitioned, remember it as a parent.
1524 : * Since a partitioned table does not have any data, we don't need to
1525 : * create a plan for it, and we can stop processing it here. We do,
1526 : * however, need to remember its modified PlannerInfo for use when
1527 : * processing its children, since we'll update their varnos based on
1528 : * the delta from immediate parent to child, not from top to child.
1529 : *
1530 : * Note: a very non-obvious point is that we have not yet added
1531 : * duplicate subquery RTEs to the subroot's rtable. We mustn't,
1532 : * because then its children would have two sets of duplicates,
1533 : * confusing matters.
1534 : */
1535 2418 : if (child_rte->inh)
1536 : {
1537 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1538 276 : parent_parses[appinfo->child_relid] = subroot->parse;
1539 712 : continue;
1540 : }
1541 :
1542 : /*
1543 : * Set the nominal target relation of the ModifyTable node if not
1544 : * already done. If the target is a partitioned table, we already set
1545 : * nominalRelation to refer to the partition root, above. For
1546 : * non-partitioned inheritance cases, we'll use the first child
1547 : * relation (even if it's excluded) as the nominal target relation.
1548 : * Because of the way expand_inherited_rtentry works, that should be
1549 : * the RTE representing the parent table in its role as a simple
1550 : * member of the inheritance set.
1551 : *
1552 : * It would be logically cleaner to *always* use the inheritance
1553 : * parent RTE as the nominal relation; but that RTE is not otherwise
1554 : * referenced in the plan in the non-partitioned inheritance case.
1555 : * Instead the duplicate child RTE created by expand_inherited_rtentry
1556 : * is used elsewhere in the plan, so using the original parent RTE
1557 : * would give rise to confusing use of multiple aliases in EXPLAIN
1558 : * output for what the user will think is the "same" table. OTOH,
1559 : * it's not a problem in the partitioned inheritance case, because
1560 : * there is no duplicate RTE for the parent.
1561 : */
1562 2142 : if (nominalRelation < 0)
1563 374 : nominalRelation = appinfo->child_relid;
1564 :
1565 : /*
1566 : * As above, each child plan run needs its own append_rel_list and
1567 : * rowmarks, which should start out as pristine copies of the
1568 : * originals. There can't be any references to UPDATE/DELETE target
1569 : * rels in them; but there could be subquery references, which we'll
1570 : * fix up in a moment.
1571 : */
1572 2142 : subroot->append_rel_list = copyObject(root->append_rel_list);
1573 2142 : subroot->rowMarks = copyObject(root->rowMarks);
1574 :
1575 : /*
1576 : * If this isn't the first child Query, adjust Vars and jointree
1577 : * entries to reference the appropriate set of subquery RTEs.
1578 : */
1579 2142 : if (final_rtable != NIL && subqueryRTindexes != NULL)
1580 : {
1581 108 : int oldrti = -1;
1582 :
1583 388 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1584 : {
1585 172 : Index newrti = next_subquery_rti++;
1586 :
1587 172 : ChangeVarNodes((Node *) subroot->parse, oldrti, newrti, 0);
1588 172 : ChangeVarNodes((Node *) subroot->append_rel_list,
1589 : oldrti, newrti, 0);
1590 172 : ChangeVarNodes((Node *) subroot->rowMarks, oldrti, newrti, 0);
1591 : }
1592 : }
1593 :
1594 : /* There shouldn't be any OJ info to translate, as yet */
1595 : Assert(subroot->join_info_list == NIL);
1596 : /* and we haven't created PlaceHolderInfos, either */
1597 : Assert(subroot->placeholder_list == NIL);
1598 :
1599 : /* Generate Path(s) for accessing this result relation */
1600 2142 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1601 :
1602 : /*
1603 : * Select cheapest path in case there's more than one. We always run
1604 : * modification queries to conclusion, so we care only for the
1605 : * cheapest-total path.
1606 : */
1607 2142 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
1608 2142 : set_cheapest(sub_final_rel);
1609 2142 : subpath = sub_final_rel->cheapest_total_path;
1610 :
1611 : /*
1612 : * If this child rel was excluded by constraint exclusion, exclude it
1613 : * from the result plan.
1614 : */
1615 2142 : if (IS_DUMMY_REL(sub_final_rel))
1616 160 : continue;
1617 :
1618 : /*
1619 : * If this is the first non-excluded child, its post-planning rtable
1620 : * becomes the initial contents of final_rtable; otherwise, copy its
1621 : * modified subquery RTEs into final_rtable, to ensure we have sane
1622 : * copies of those. Also save the first non-excluded child's version
1623 : * of the rowmarks list; we assume all children will end up with
1624 : * equivalent versions of that.
1625 : */
1626 1982 : if (final_rtable == NIL)
1627 : {
1628 942 : final_rtable = subroot->parse->rtable;
1629 942 : final_rowmarks = subroot->rowMarks;
1630 : }
1631 : else
1632 : {
1633 : Assert(list_length(final_rtable) ==
1634 : list_length(subroot->parse->rtable));
1635 1040 : if (subqueryRTindexes != NULL)
1636 : {
1637 108 : int oldrti = -1;
1638 :
1639 388 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1640 : {
1641 172 : Index newrti = this_subquery_rti++;
1642 : RangeTblEntry *subqrte;
1643 : ListCell *newrticell;
1644 :
1645 172 : subqrte = rt_fetch(newrti, subroot->parse->rtable);
1646 172 : newrticell = list_nth_cell(final_rtable, newrti - 1);
1647 172 : lfirst(newrticell) = subqrte;
1648 : }
1649 : }
1650 : }
1651 :
1652 : /*
1653 : * We need to collect all the RelOptInfos from all child plans into
1654 : * the main PlannerInfo, since setrefs.c will need them. We use the
1655 : * last child's simple_rel_array, so we have to propagate forward the
1656 : * RelOptInfos that were already built in previous children.
1657 : */
1658 : Assert(subroot->simple_rel_array_size >= save_rel_array_size);
1659 8510 : for (rti = 1; rti < save_rel_array_size; rti++)
1660 : {
1661 6528 : RelOptInfo *brel = save_rel_array[rti];
1662 :
1663 6528 : if (brel)
1664 2450 : subroot->simple_rel_array[rti] = brel;
1665 : }
1666 1982 : save_rel_array_size = subroot->simple_rel_array_size;
1667 1982 : save_rel_array = subroot->simple_rel_array;
1668 1982 : save_append_rel_array = subroot->append_rel_array;
1669 :
1670 : /*
1671 : * Make sure any initplans from this rel get into the outer list. Note
1672 : * we're effectively assuming all children generate the same
1673 : * init_plans.
1674 : */
1675 1982 : root->init_plans = subroot->init_plans;
1676 :
1677 : /* Build list of sub-paths */
1678 1982 : subpaths = lappend(subpaths, subpath);
1679 :
1680 : /* Build list of modified subroots, too */
1681 1982 : subroots = lappend(subroots, subroot);
1682 :
1683 : /* Build list of target-relation RT indexes */
1684 1982 : resultRelations = lappend_int(resultRelations, appinfo->child_relid);
1685 :
1686 : /* Build lists of per-relation WCO and RETURNING targetlists */
1687 1982 : if (parse->withCheckOptions)
1688 192 : withCheckOptionLists = lappend(withCheckOptionLists,
1689 192 : subroot->parse->withCheckOptions);
1690 1982 : if (parse->returningList)
1691 284 : returningLists = lappend(returningLists,
1692 284 : subroot->parse->returningList);
1693 :
1694 : Assert(!parse->onConflict);
1695 : }
1696 :
1697 : /* Result path must go into outer query's FINAL upperrel */
1698 974 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1699 :
1700 : /*
1701 : * We don't currently worry about setting final_rel's consider_parallel
1702 : * flag in this case, nor about allowing FDWs or create_upper_paths_hook
1703 : * to get control here.
1704 : */
1705 :
1706 974 : if (subpaths == NIL)
1707 : {
1708 : /*
1709 : * We managed to exclude every child rel, so generate a dummy path
1710 : * representing the empty set. Although it's clear that no data will
1711 : * be updated or deleted, we will still need to have a ModifyTable
1712 : * node so that any statement triggers are executed. (This could be
1713 : * cleaner if we fixed nodeModifyTable.c to support zero child nodes,
1714 : * but that probably wouldn't be a net win.)
1715 : */
1716 : Path *dummy_path;
1717 :
1718 : /* tlist processing never got done, either */
1719 32 : root->processed_tlist = preprocess_targetlist(root);
1720 32 : final_rel->reltarget = create_pathtarget(root, root->processed_tlist);
1721 :
1722 : /* Make a dummy path, cf set_dummy_rel_pathlist() */
1723 32 : dummy_path = (Path *) create_append_path(NULL, final_rel, NIL, NIL,
1724 : NIL, NULL, 0, false,
1725 : NIL, -1);
1726 :
1727 : /* These lists must be nonempty to make a valid ModifyTable node */
1728 32 : subpaths = list_make1(dummy_path);
1729 32 : subroots = list_make1(root);
1730 32 : resultRelations = list_make1_int(parse->resultRelation);
1731 32 : if (parse->withCheckOptions)
1732 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
1733 32 : if (parse->returningList)
1734 12 : returningLists = list_make1(parse->returningList);
1735 : /* Disable tuple routing, too, just to be safe */
1736 32 : root->partColsUpdated = false;
1737 : }
1738 : else
1739 : {
1740 : /*
1741 : * Put back the final adjusted rtable into the master copy of the
1742 : * Query. (We mustn't do this if we found no non-excluded children,
1743 : * since we never saved an adjusted rtable at all.)
1744 : */
1745 942 : parse->rtable = final_rtable;
1746 942 : root->simple_rel_array_size = save_rel_array_size;
1747 942 : root->simple_rel_array = save_rel_array;
1748 942 : root->append_rel_array = save_append_rel_array;
1749 :
1750 : /* Must reconstruct master's simple_rte_array, too */
1751 942 : root->simple_rte_array = (RangeTblEntry **)
1752 942 : palloc0((list_length(final_rtable) + 1) * sizeof(RangeTblEntry *));
1753 942 : rti = 1;
1754 5034 : foreach(lc, final_rtable)
1755 : {
1756 4092 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1757 :
1758 4092 : root->simple_rte_array[rti++] = rte;
1759 : }
1760 :
1761 : /* Put back adjusted rowmarks, too */
1762 942 : root->rowMarks = final_rowmarks;
1763 : }
1764 :
1765 : /*
1766 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node will
1767 : * have dealt with fetching non-locked marked rows, else we need to have
1768 : * ModifyTable do that.
1769 : */
1770 974 : if (parse->rowMarks)
1771 0 : rowMarks = NIL;
1772 : else
1773 974 : rowMarks = root->rowMarks;
1774 :
1775 : /* Create Path representing a ModifyTable to do the UPDATE/DELETE work */
1776 974 : add_path(final_rel, (Path *)
1777 2922 : create_modifytable_path(root, final_rel,
1778 : parse->commandType,
1779 974 : parse->canSetTag,
1780 : nominalRelation,
1781 : rootRelation,
1782 974 : root->partColsUpdated,
1783 : resultRelations,
1784 : subpaths,
1785 : subroots,
1786 : withCheckOptionLists,
1787 : returningLists,
1788 : rowMarks,
1789 : NULL,
1790 : assign_special_exec_param(root)));
1791 974 : }
1792 :
1793 : /*--------------------
1794 : * grouping_planner
1795 : * Perform planning steps related to grouping, aggregation, etc.
1796 : *
1797 : * This function adds all required top-level processing to the scan/join
1798 : * Path(s) produced by query_planner.
1799 : *
1800 : * If inheritance_update is true, we're being called from inheritance_planner
1801 : * and should not include a ModifyTable step in the resulting Path(s).
1802 : * (inheritance_planner will create a single ModifyTable node covering all the
1803 : * target tables.)
1804 : *
1805 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1806 : * tuple_fraction is interpreted as follows:
1807 : * 0: expect all tuples to be retrieved (normal case)
1808 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1809 : * from the plan to be retrieved
1810 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1811 : * expected to be retrieved (ie, a LIMIT specification)
1812 : *
1813 : * Returns nothing; the useful output is in the Paths we attach to the
1814 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1815 : * root->processed_tlist contains the final processed targetlist.
1816 : *
1817 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1818 : * to leave this to the caller.
1819 : *--------------------
1820 : */
1821 : static void
1822 289868 : grouping_planner(PlannerInfo *root, bool inheritance_update,
1823 : double tuple_fraction)
1824 : {
1825 289868 : Query *parse = root->parse;
1826 289868 : int64 offset_est = 0;
1827 289868 : int64 count_est = 0;
1828 289868 : double limit_tuples = -1.0;
1829 289868 : bool have_postponed_srfs = false;
1830 : PathTarget *final_target;
1831 : List *final_targets;
1832 : List *final_targets_contain_srfs;
1833 : bool final_target_parallel_safe;
1834 : RelOptInfo *current_rel;
1835 : RelOptInfo *final_rel;
1836 : FinalPathExtraData extra;
1837 : ListCell *lc;
1838 :
1839 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1840 289868 : if (parse->limitCount || parse->limitOffset)
1841 : {
1842 2890 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1843 : &offset_est, &count_est);
1844 :
1845 : /*
1846 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1847 : * estimate the effects of using a bounded sort.
1848 : */
1849 2890 : if (count_est > 0 && offset_est >= 0)
1850 2672 : limit_tuples = (double) count_est + (double) offset_est;
1851 : }
1852 :
1853 : /* Make tuple_fraction accessible to lower-level routines */
1854 289868 : root->tuple_fraction = tuple_fraction;
1855 :
1856 289868 : if (parse->setOperations)
1857 : {
1858 : /*
1859 : * If there's a top-level ORDER BY, assume we have to fetch all the
1860 : * tuples. This might be too simplistic given all the hackery below
1861 : * to possibly avoid the sort; but the odds of accurate estimates here
1862 : * are pretty low anyway. XXX try to get rid of this in favor of
1863 : * letting plan_set_operations generate both fast-start and
1864 : * cheapest-total paths.
1865 : */
1866 1278 : if (parse->sortClause)
1867 204 : root->tuple_fraction = 0.0;
1868 :
1869 : /*
1870 : * Construct Paths for set operations. The results will not need any
1871 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1872 : * special work for recursive unions is the responsibility of
1873 : * plan_set_operations.
1874 : */
1875 1278 : current_rel = plan_set_operations(root);
1876 :
1877 : /*
1878 : * We should not need to call preprocess_targetlist, since we must be
1879 : * in a SELECT query node. Instead, use the processed_tlist returned
1880 : * by plan_set_operations (since this tells whether it returned any
1881 : * resjunk columns!), and transfer any sort key information from the
1882 : * original tlist.
1883 : */
1884 : Assert(parse->commandType == CMD_SELECT);
1885 :
1886 : /* for safety, copy processed_tlist instead of modifying in-place */
1887 1278 : root->processed_tlist =
1888 1278 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1889 : parse->targetList);
1890 :
1891 : /* Also extract the PathTarget form of the setop result tlist */
1892 1278 : final_target = current_rel->cheapest_total_path->pathtarget;
1893 :
1894 : /* And check whether it's parallel safe */
1895 1278 : final_target_parallel_safe =
1896 1278 : is_parallel_safe(root, (Node *) final_target->exprs);
1897 :
1898 : /* The setop result tlist couldn't contain any SRFs */
1899 : Assert(!parse->hasTargetSRFs);
1900 1278 : final_targets = final_targets_contain_srfs = NIL;
1901 :
1902 : /*
1903 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1904 : * checked already, but let's make sure).
1905 : */
1906 1278 : if (parse->rowMarks)
1907 0 : ereport(ERROR,
1908 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1909 : /*------
1910 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1911 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1912 : LCS_asString(linitial_node(RowMarkClause,
1913 : parse->rowMarks)->strength))));
1914 :
1915 : /*
1916 : * Calculate pathkeys that represent result ordering requirements
1917 : */
1918 : Assert(parse->distinctClause == NIL);
1919 1278 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1920 : parse->sortClause,
1921 : root->processed_tlist);
1922 : }
1923 : else
1924 : {
1925 : /* No set operations, do regular planning */
1926 : PathTarget *sort_input_target;
1927 : List *sort_input_targets;
1928 : List *sort_input_targets_contain_srfs;
1929 : bool sort_input_target_parallel_safe;
1930 : PathTarget *grouping_target;
1931 : List *grouping_targets;
1932 : List *grouping_targets_contain_srfs;
1933 : bool grouping_target_parallel_safe;
1934 : PathTarget *scanjoin_target;
1935 : List *scanjoin_targets;
1936 : List *scanjoin_targets_contain_srfs;
1937 : bool scanjoin_target_parallel_safe;
1938 : bool scanjoin_target_same_exprs;
1939 : bool have_grouping;
1940 : AggClauseCosts agg_costs;
1941 288590 : WindowFuncLists *wflists = NULL;
1942 288590 : List *activeWindows = NIL;
1943 288590 : grouping_sets_data *gset_data = NULL;
1944 : standard_qp_extra qp_extra;
1945 :
1946 : /* A recursive query should always have setOperations */
1947 : Assert(!root->hasRecursion);
1948 :
1949 : /* Preprocess grouping sets and GROUP BY clause, if any */
1950 288590 : if (parse->groupingSets)
1951 : {
1952 416 : gset_data = preprocess_grouping_sets(root);
1953 : }
1954 : else
1955 : {
1956 : /* Preprocess regular GROUP BY clause, if any */
1957 288174 : if (parse->groupClause)
1958 1650 : parse->groupClause = preprocess_groupclause(root, NIL);
1959 : }
1960 :
1961 : /*
1962 : * Preprocess targetlist. Note that much of the remaining planning
1963 : * work will be done with the PathTarget representation of tlists, but
1964 : * we must also maintain the full representation of the final tlist so
1965 : * that we can transfer its decoration (resnames etc) to the topmost
1966 : * tlist of the finished Plan. This is kept in processed_tlist.
1967 : */
1968 288586 : root->processed_tlist = preprocess_targetlist(root);
1969 :
1970 : /*
1971 : * Collect statistics about aggregates for estimating costs, and mark
1972 : * all the aggregates with resolved aggtranstypes. We must do this
1973 : * before slicing and dicing the tlist into various pathtargets, else
1974 : * some copies of the Aggref nodes might escape being marked with the
1975 : * correct transtypes.
1976 : *
1977 : * Note: currently, we do not detect duplicate aggregates here. This
1978 : * may result in somewhat-overestimated cost, which is fine for our
1979 : * purposes since all Paths will get charged the same. But at some
1980 : * point we might wish to do that detection in the planner, rather
1981 : * than during executor startup.
1982 : */
1983 288586 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1984 288586 : if (parse->hasAggs)
1985 : {
1986 23364 : get_agg_clause_costs(root, (Node *) root->processed_tlist,
1987 : AGGSPLIT_SIMPLE, &agg_costs);
1988 23364 : get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
1989 : &agg_costs);
1990 : }
1991 :
1992 : /*
1993 : * Locate any window functions in the tlist. (We don't need to look
1994 : * anywhere else, since expressions used in ORDER BY will be in there
1995 : * too.) Note that they could all have been eliminated by constant
1996 : * folding, in which case we don't need to do any more work.
1997 : */
1998 288586 : if (parse->hasWindowFuncs)
1999 : {
2000 1080 : wflists = find_window_functions((Node *) root->processed_tlist,
2001 1080 : list_length(parse->windowClause));
2002 1080 : if (wflists->numWindowFuncs > 0)
2003 1076 : activeWindows = select_active_windows(root, wflists);
2004 : else
2005 4 : parse->hasWindowFuncs = false;
2006 : }
2007 :
2008 : /*
2009 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
2010 : * adding logic between here and the query_planner() call. Anything
2011 : * that is needed in MIN/MAX-optimizable cases will have to be
2012 : * duplicated in planagg.c.
2013 : */
2014 288586 : if (parse->hasAggs)
2015 23364 : preprocess_minmax_aggregates(root);
2016 :
2017 : /*
2018 : * Figure out whether there's a hard limit on the number of rows that
2019 : * query_planner's result subplan needs to return. Even if we know a
2020 : * hard limit overall, it doesn't apply if the query has any
2021 : * grouping/aggregation operations, or SRFs in the tlist.
2022 : */
2023 575138 : if (parse->groupClause ||
2024 573076 : parse->groupingSets ||
2025 572716 : parse->distinctClause ||
2026 550876 : parse->hasAggs ||
2027 528372 : parse->hasWindowFuncs ||
2028 523910 : parse->hasTargetSRFs ||
2029 260222 : root->hasHavingQual)
2030 28376 : root->limit_tuples = -1.0;
2031 : else
2032 260210 : root->limit_tuples = limit_tuples;
2033 :
2034 : /* Set up data needed by standard_qp_callback */
2035 288586 : qp_extra.activeWindows = activeWindows;
2036 288586 : qp_extra.groupClause = (gset_data
2037 412 : ? (gset_data->rollups ? linitial_node(RollupData, gset_data->rollups)->groupClause : NIL)
2038 288998 : : parse->groupClause);
2039 :
2040 : /*
2041 : * Generate the best unsorted and presorted paths for the scan/join
2042 : * portion of this Query, ie the processing represented by the
2043 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
2044 : * We also generate (in standard_qp_callback) pathkey representations
2045 : * of the query's sort clause, distinct clause, etc.
2046 : */
2047 288586 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
2048 :
2049 : /*
2050 : * Convert the query's result tlist into PathTarget format.
2051 : *
2052 : * Note: this cannot be done before query_planner() has performed
2053 : * appendrel expansion, because that might add resjunk entries to
2054 : * root->processed_tlist. Waiting till afterwards is also helpful
2055 : * because the target width estimates can use per-Var width numbers
2056 : * that were obtained within query_planner().
2057 : */
2058 288576 : final_target = create_pathtarget(root, root->processed_tlist);
2059 288576 : final_target_parallel_safe =
2060 288576 : is_parallel_safe(root, (Node *) final_target->exprs);
2061 :
2062 : /*
2063 : * If ORDER BY was given, consider whether we should use a post-sort
2064 : * projection, and compute the adjusted target for preceding steps if
2065 : * so.
2066 : */
2067 288576 : if (parse->sortClause)
2068 : {
2069 33874 : sort_input_target = make_sort_input_target(root,
2070 : final_target,
2071 : &have_postponed_srfs);
2072 33874 : sort_input_target_parallel_safe =
2073 33874 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
2074 : }
2075 : else
2076 : {
2077 254702 : sort_input_target = final_target;
2078 254702 : sort_input_target_parallel_safe = final_target_parallel_safe;
2079 : }
2080 :
2081 : /*
2082 : * If we have window functions to deal with, the output from any
2083 : * grouping step needs to be what the window functions want;
2084 : * otherwise, it should be sort_input_target.
2085 : */
2086 288576 : if (activeWindows)
2087 : {
2088 1076 : grouping_target = make_window_input_target(root,
2089 : final_target,
2090 : activeWindows);
2091 1076 : grouping_target_parallel_safe =
2092 1076 : is_parallel_safe(root, (Node *) grouping_target->exprs);
2093 : }
2094 : else
2095 : {
2096 287500 : grouping_target = sort_input_target;
2097 287500 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
2098 : }
2099 :
2100 : /*
2101 : * If we have grouping or aggregation to do, the topmost scan/join
2102 : * plan node must emit what the grouping step wants; otherwise, it
2103 : * should emit grouping_target.
2104 : */
2105 1150208 : have_grouping = (parse->groupClause || parse->groupingSets ||
2106 840078 : parse->hasAggs || root->hasHavingQual);
2107 288576 : if (have_grouping)
2108 : {
2109 23600 : scanjoin_target = make_group_input_target(root, final_target);
2110 23600 : scanjoin_target_parallel_safe =
2111 23600 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2112 : }
2113 : else
2114 : {
2115 264976 : scanjoin_target = grouping_target;
2116 264976 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2117 : }
2118 :
2119 : /*
2120 : * If there are any SRFs in the targetlist, we must separate each of
2121 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2122 : * each of the named targets with a SRF-free version, and remember the
2123 : * list of additional projection steps we need to add afterwards.
2124 : */
2125 288576 : if (parse->hasTargetSRFs)
2126 : {
2127 : /* final_target doesn't recompute any SRFs in sort_input_target */
2128 3578 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2129 : &final_targets,
2130 : &final_targets_contain_srfs);
2131 3578 : final_target = linitial_node(PathTarget, final_targets);
2132 : Assert(!linitial_int(final_targets_contain_srfs));
2133 : /* likewise for sort_input_target vs. grouping_target */
2134 3578 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2135 : &sort_input_targets,
2136 : &sort_input_targets_contain_srfs);
2137 3578 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2138 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2139 : /* likewise for grouping_target vs. scanjoin_target */
2140 3578 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
2141 : &grouping_targets,
2142 : &grouping_targets_contain_srfs);
2143 3578 : grouping_target = linitial_node(PathTarget, grouping_targets);
2144 : Assert(!linitial_int(grouping_targets_contain_srfs));
2145 : /* scanjoin_target will not have any SRFs precomputed for it */
2146 3578 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2147 : &scanjoin_targets,
2148 : &scanjoin_targets_contain_srfs);
2149 3578 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2150 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2151 : }
2152 : else
2153 : {
2154 : /* initialize lists; for most of these, dummy values are OK */
2155 284998 : final_targets = final_targets_contain_srfs = NIL;
2156 284998 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2157 284998 : grouping_targets = grouping_targets_contain_srfs = NIL;
2158 284998 : scanjoin_targets = list_make1(scanjoin_target);
2159 284998 : scanjoin_targets_contain_srfs = NIL;
2160 : }
2161 :
2162 : /* Apply scan/join target. */
2163 577152 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2164 288576 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2165 288576 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2166 : scanjoin_targets_contain_srfs,
2167 : scanjoin_target_parallel_safe,
2168 : scanjoin_target_same_exprs);
2169 :
2170 : /*
2171 : * Save the various upper-rel PathTargets we just computed into
2172 : * root->upper_targets[]. The core code doesn't use this, but it
2173 : * provides a convenient place for extensions to get at the info. For
2174 : * consistency, we save all the intermediate targets, even though some
2175 : * of the corresponding upperrels might not be needed for this query.
2176 : */
2177 288576 : root->upper_targets[UPPERREL_FINAL] = final_target;
2178 288576 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2179 288576 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2180 288576 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2181 288576 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2182 :
2183 : /*
2184 : * If we have grouping and/or aggregation, consider ways to implement
2185 : * that. We build a new upperrel representing the output of this
2186 : * phase.
2187 : */
2188 288576 : if (have_grouping)
2189 : {
2190 23600 : current_rel = create_grouping_paths(root,
2191 : current_rel,
2192 : grouping_target,
2193 : grouping_target_parallel_safe,
2194 : &agg_costs,
2195 : gset_data);
2196 : /* Fix things up if grouping_target contains SRFs */
2197 23596 : if (parse->hasTargetSRFs)
2198 84 : adjust_paths_for_srfs(root, current_rel,
2199 : grouping_targets,
2200 : grouping_targets_contain_srfs);
2201 : }
2202 :
2203 : /*
2204 : * If we have window functions, consider ways to implement those. We
2205 : * build a new upperrel representing the output of this phase.
2206 : */
2207 288572 : if (activeWindows)
2208 : {
2209 1076 : current_rel = create_window_paths(root,
2210 : current_rel,
2211 : grouping_target,
2212 : sort_input_target,
2213 : sort_input_target_parallel_safe,
2214 : wflists,
2215 : activeWindows);
2216 : /* Fix things up if sort_input_target contains SRFs */
2217 1076 : if (parse->hasTargetSRFs)
2218 8 : adjust_paths_for_srfs(root, current_rel,
2219 : sort_input_targets,
2220 : sort_input_targets_contain_srfs);
2221 : }
2222 :
2223 : /*
2224 : * If there is a DISTINCT clause, consider ways to implement that. We
2225 : * build a new upperrel representing the output of this phase.
2226 : */
2227 288572 : if (parse->distinctClause)
2228 : {
2229 336 : current_rel = create_distinct_paths(root,
2230 : current_rel);
2231 : }
2232 : } /* end of if (setOperations) */
2233 :
2234 : /*
2235 : * If ORDER BY was given, consider ways to implement that, and generate a
2236 : * new upperrel containing only paths that emit the correct ordering and
2237 : * project the correct final_target. We can apply the original
2238 : * limit_tuples limit in sort costing here, but only if there are no
2239 : * postponed SRFs.
2240 : */
2241 289850 : if (parse->sortClause)
2242 : {
2243 34078 : current_rel = create_ordered_paths(root,
2244 : current_rel,
2245 : final_target,
2246 : final_target_parallel_safe,
2247 : have_postponed_srfs ? -1.0 :
2248 : limit_tuples);
2249 : /* Fix things up if final_target contains SRFs */
2250 34078 : if (parse->hasTargetSRFs)
2251 116 : adjust_paths_for_srfs(root, current_rel,
2252 : final_targets,
2253 : final_targets_contain_srfs);
2254 : }
2255 :
2256 : /*
2257 : * Now we are prepared to build the final-output upperrel.
2258 : */
2259 289850 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2260 :
2261 : /*
2262 : * If the input rel is marked consider_parallel and there's nothing that's
2263 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2264 : * consider_parallel as well. Note that if the query has rowMarks or is
2265 : * not a SELECT, consider_parallel will be false for every relation in the
2266 : * query.
2267 : */
2268 356924 : if (current_rel->consider_parallel &&
2269 134140 : is_parallel_safe(root, parse->limitOffset) &&
2270 67066 : is_parallel_safe(root, parse->limitCount))
2271 67062 : final_rel->consider_parallel = true;
2272 :
2273 : /*
2274 : * If the current_rel belongs to a single FDW, so does the final_rel.
2275 : */
2276 289850 : final_rel->serverid = current_rel->serverid;
2277 289850 : final_rel->userid = current_rel->userid;
2278 289850 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2279 289850 : final_rel->fdwroutine = current_rel->fdwroutine;
2280 :
2281 : /*
2282 : * Generate paths for the final_rel. Insert all surviving paths, with
2283 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2284 : */
2285 584800 : foreach(lc, current_rel->pathlist)
2286 : {
2287 294950 : Path *path = (Path *) lfirst(lc);
2288 :
2289 : /*
2290 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2291 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2292 : * here. If there are only non-locking rowmarks, they should be
2293 : * handled by the ModifyTable node instead. However, root->rowMarks
2294 : * is what goes into the LockRows node.)
2295 : */
2296 294950 : if (parse->rowMarks)
2297 : {
2298 5186 : path = (Path *) create_lockrows_path(root, final_rel, path,
2299 : root->rowMarks,
2300 : assign_special_exec_param(root));
2301 : }
2302 :
2303 : /*
2304 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2305 : */
2306 294950 : if (limit_needed(parse))
2307 : {
2308 3338 : path = (Path *) create_limit_path(root, final_rel, path,
2309 : parse->limitOffset,
2310 : parse->limitCount,
2311 : offset_est, count_est);
2312 : }
2313 :
2314 : /*
2315 : * If this is an INSERT/UPDATE/DELETE, and we're not being called from
2316 : * inheritance_planner, add the ModifyTable node.
2317 : */
2318 294950 : if (parse->commandType != CMD_SELECT && !inheritance_update)
2319 : {
2320 : Index rootRelation;
2321 : List *withCheckOptionLists;
2322 : List *returningLists;
2323 : List *rowMarks;
2324 :
2325 : /*
2326 : * If target is a partition root table, we need to mark the
2327 : * ModifyTable node appropriately for that.
2328 : */
2329 68466 : if (rt_fetch(parse->resultRelation, parse->rtable)->relkind ==
2330 : RELKIND_PARTITIONED_TABLE)
2331 1408 : rootRelation = parse->resultRelation;
2332 : else
2333 67058 : rootRelation = 0;
2334 :
2335 : /*
2336 : * Set up the WITH CHECK OPTION and RETURNING lists-of-lists, if
2337 : * needed.
2338 : */
2339 68466 : if (parse->withCheckOptions)
2340 496 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2341 : else
2342 67970 : withCheckOptionLists = NIL;
2343 :
2344 68466 : if (parse->returningList)
2345 1152 : returningLists = list_make1(parse->returningList);
2346 : else
2347 67314 : returningLists = NIL;
2348 :
2349 : /*
2350 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2351 : * will have dealt with fetching non-locked marked rows, else we
2352 : * need to have ModifyTable do that.
2353 : */
2354 68466 : if (parse->rowMarks)
2355 0 : rowMarks = NIL;
2356 : else
2357 68466 : rowMarks = root->rowMarks;
2358 :
2359 68466 : path = (Path *)
2360 205398 : create_modifytable_path(root, final_rel,
2361 : parse->commandType,
2362 68466 : parse->canSetTag,
2363 68466 : parse->resultRelation,
2364 : rootRelation,
2365 : false,
2366 : list_make1_int(parse->resultRelation),
2367 : list_make1(path),
2368 : list_make1(root),
2369 : withCheckOptionLists,
2370 : returningLists,
2371 : rowMarks,
2372 : parse->onConflict,
2373 : assign_special_exec_param(root));
2374 : }
2375 :
2376 : /* And shove it into final_rel */
2377 294950 : add_path(final_rel, path);
2378 : }
2379 :
2380 : /*
2381 : * Generate partial paths for final_rel, too, if outer query levels might
2382 : * be able to make use of them.
2383 : */
2384 295754 : if (final_rel->consider_parallel && root->query_level > 1 &&
2385 5904 : !limit_needed(parse))
2386 : {
2387 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2388 5864 : foreach(lc, current_rel->partial_pathlist)
2389 : {
2390 36 : Path *partial_path = (Path *) lfirst(lc);
2391 :
2392 36 : add_partial_path(final_rel, partial_path);
2393 : }
2394 : }
2395 :
2396 289850 : extra.limit_needed = limit_needed(parse);
2397 289850 : extra.limit_tuples = limit_tuples;
2398 289850 : extra.count_est = count_est;
2399 289850 : extra.offset_est = offset_est;
2400 :
2401 : /*
2402 : * If there is an FDW that's responsible for all baserels of the query,
2403 : * let it consider adding ForeignPaths.
2404 : */
2405 290814 : if (final_rel->fdwroutine &&
2406 964 : final_rel->fdwroutine->GetForeignUpperPaths)
2407 906 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2408 : current_rel, final_rel,
2409 : &extra);
2410 :
2411 : /* Let extensions possibly add some more paths */
2412 289850 : if (create_upper_paths_hook)
2413 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2414 : current_rel, final_rel, &extra);
2415 :
2416 : /* Note: currently, we leave it to callers to do set_cheapest() */
2417 289850 : }
2418 :
2419 : /*
2420 : * Do preprocessing for groupingSets clause and related data. This handles the
2421 : * preliminary steps of expanding the grouping sets, organizing them into lists
2422 : * of rollups, and preparing annotations which will later be filled in with
2423 : * size estimates.
2424 : */
2425 : static grouping_sets_data *
2426 416 : preprocess_grouping_sets(PlannerInfo *root)
2427 : {
2428 416 : Query *parse = root->parse;
2429 : List *sets;
2430 416 : int maxref = 0;
2431 : ListCell *lc;
2432 : ListCell *lc_set;
2433 416 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2434 :
2435 416 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, -1);
2436 :
2437 416 : gd->any_hashable = false;
2438 416 : gd->unhashable_refs = NULL;
2439 416 : gd->unsortable_refs = NULL;
2440 416 : gd->unsortable_sets = NIL;
2441 :
2442 416 : if (parse->groupClause)
2443 : {
2444 : ListCell *lc;
2445 :
2446 1248 : foreach(lc, parse->groupClause)
2447 : {
2448 860 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2449 860 : Index ref = gc->tleSortGroupRef;
2450 :
2451 860 : if (ref > maxref)
2452 844 : maxref = ref;
2453 :
2454 860 : if (!gc->hashable)
2455 20 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2456 :
2457 860 : if (!OidIsValid(gc->sortop))
2458 28 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2459 : }
2460 : }
2461 :
2462 : /* Allocate workspace array for remapping */
2463 416 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2464 :
2465 : /*
2466 : * If we have any unsortable sets, we must extract them before trying to
2467 : * prepare rollups. Unsortable sets don't go through
2468 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2469 : * here.
2470 : */
2471 416 : if (!bms_is_empty(gd->unsortable_refs))
2472 : {
2473 28 : List *sortable_sets = NIL;
2474 :
2475 84 : foreach(lc, parse->groupingSets)
2476 : {
2477 60 : List *gset = (List *) lfirst(lc);
2478 :
2479 60 : if (bms_overlap_list(gd->unsortable_refs, gset))
2480 : {
2481 32 : GroupingSetData *gs = makeNode(GroupingSetData);
2482 :
2483 32 : gs->set = gset;
2484 32 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2485 :
2486 : /*
2487 : * We must enforce here that an unsortable set is hashable;
2488 : * later code assumes this. Parse analysis only checks that
2489 : * every individual column is either hashable or sortable.
2490 : *
2491 : * Note that passing this test doesn't guarantee we can
2492 : * generate a plan; there might be other showstoppers.
2493 : */
2494 32 : if (bms_overlap_list(gd->unhashable_refs, gset))
2495 4 : ereport(ERROR,
2496 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2497 : errmsg("could not implement GROUP BY"),
2498 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2499 : }
2500 : else
2501 28 : sortable_sets = lappend(sortable_sets, gset);
2502 : }
2503 :
2504 24 : if (sortable_sets)
2505 20 : sets = extract_rollup_sets(sortable_sets);
2506 : else
2507 4 : sets = NIL;
2508 : }
2509 : else
2510 388 : sets = extract_rollup_sets(parse->groupingSets);
2511 :
2512 1068 : foreach(lc_set, sets)
2513 : {
2514 656 : List *current_sets = (List *) lfirst(lc_set);
2515 656 : RollupData *rollup = makeNode(RollupData);
2516 : GroupingSetData *gs;
2517 :
2518 : /*
2519 : * Reorder the current list of grouping sets into correct prefix
2520 : * order. If only one aggregation pass is needed, try to make the
2521 : * list match the ORDER BY clause; if more than one pass is needed, we
2522 : * don't bother with that.
2523 : *
2524 : * Note that this reorders the sets from smallest-member-first to
2525 : * largest-member-first, and applies the GroupingSetData annotations,
2526 : * though the data will be filled in later.
2527 : */
2528 656 : current_sets = reorder_grouping_sets(current_sets,
2529 656 : (list_length(sets) == 1
2530 : ? parse->sortClause
2531 : : NIL));
2532 :
2533 : /*
2534 : * Get the initial (and therefore largest) grouping set.
2535 : */
2536 656 : gs = linitial_node(GroupingSetData, current_sets);
2537 :
2538 : /*
2539 : * Order the groupClause appropriately. If the first grouping set is
2540 : * empty, then the groupClause must also be empty; otherwise we have
2541 : * to force the groupClause to match that grouping set's order.
2542 : *
2543 : * (The first grouping set can be empty even though parse->groupClause
2544 : * is not empty only if all non-empty grouping sets are unsortable.
2545 : * The groupClauses for hashed grouping sets are built later on.)
2546 : */
2547 656 : if (gs->set)
2548 628 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2549 : else
2550 28 : rollup->groupClause = NIL;
2551 :
2552 : /*
2553 : * Is it hashable? We pretend empty sets are hashable even though we
2554 : * actually force them not to be hashed later. But don't bother if
2555 : * there's nothing but empty sets (since in that case we can't hash
2556 : * anything).
2557 : */
2558 1284 : if (gs->set &&
2559 628 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2560 : {
2561 612 : rollup->hashable = true;
2562 612 : gd->any_hashable = true;
2563 : }
2564 :
2565 : /*
2566 : * Now that we've pinned down an order for the groupClause for this
2567 : * list of grouping sets, we need to remap the entries in the grouping
2568 : * sets from sortgrouprefs to plain indices (0-based) into the
2569 : * groupClause for this collection of grouping sets. We keep the
2570 : * original form for later use, though.
2571 : */
2572 656 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2573 : current_sets,
2574 : gd->tleref_to_colnum_map);
2575 656 : rollup->gsets_data = current_sets;
2576 :
2577 656 : gd->rollups = lappend(gd->rollups, rollup);
2578 : }
2579 :
2580 412 : if (gd->unsortable_sets)
2581 : {
2582 : /*
2583 : * We have not yet pinned down a groupclause for this, but we will
2584 : * need index-based lists for estimation purposes. Construct
2585 : * hash_sets_idx based on the entire original groupclause for now.
2586 : */
2587 24 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2588 : gd->unsortable_sets,
2589 : gd->tleref_to_colnum_map);
2590 24 : gd->any_hashable = true;
2591 : }
2592 :
2593 412 : return gd;
2594 : }
2595 :
2596 : /*
2597 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2598 : * (without annotation) mapped to indexes into the given groupclause.
2599 : */
2600 : static List *
2601 1800 : remap_to_groupclause_idx(List *groupClause,
2602 : List *gsets,
2603 : int *tleref_to_colnum_map)
2604 : {
2605 1800 : int ref = 0;
2606 1800 : List *result = NIL;
2607 : ListCell *lc;
2608 :
2609 4356 : foreach(lc, groupClause)
2610 : {
2611 2556 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2612 :
2613 2556 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2614 : }
2615 :
2616 4176 : foreach(lc, gsets)
2617 : {
2618 2376 : List *set = NIL;
2619 : ListCell *lc2;
2620 2376 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2621 :
2622 5224 : foreach(lc2, gs->set)
2623 : {
2624 2848 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2625 : }
2626 :
2627 2376 : result = lappend(result, set);
2628 : }
2629 :
2630 1800 : return result;
2631 : }
2632 :
2633 :
2634 : /*
2635 : * preprocess_rowmarks - set up PlanRowMarks if needed
2636 : */
2637 : static void
2638 289084 : preprocess_rowmarks(PlannerInfo *root)
2639 : {
2640 289084 : Query *parse = root->parse;
2641 : Bitmapset *rels;
2642 : List *prowmarks;
2643 : ListCell *l;
2644 : int i;
2645 :
2646 289084 : if (parse->rowMarks)
2647 : {
2648 : /*
2649 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2650 : * grouping, since grouping renders a reference to individual tuple
2651 : * CTIDs invalid. This is also checked at parse time, but that's
2652 : * insufficient because of rule substitution, query pullup, etc.
2653 : */
2654 4742 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2655 : parse->rowMarks)->strength);
2656 : }
2657 : else
2658 : {
2659 : /*
2660 : * We only need rowmarks for UPDATE, DELETE, or FOR [KEY]
2661 : * UPDATE/SHARE.
2662 : */
2663 559482 : if (parse->commandType != CMD_UPDATE &&
2664 275140 : parse->commandType != CMD_DELETE)
2665 272368 : return;
2666 : }
2667 :
2668 : /*
2669 : * We need to have rowmarks for all base relations except the target. We
2670 : * make a bitmapset of all base rels and then remove the items we don't
2671 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2672 : */
2673 16716 : rels = get_relids_in_jointree((Node *) parse->jointree, false);
2674 16716 : if (parse->resultRelation)
2675 11974 : rels = bms_del_member(rels, parse->resultRelation);
2676 :
2677 : /*
2678 : * Convert RowMarkClauses to PlanRowMark representation.
2679 : */
2680 16716 : prowmarks = NIL;
2681 21664 : foreach(l, parse->rowMarks)
2682 : {
2683 4948 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2684 4948 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2685 : PlanRowMark *newrc;
2686 :
2687 : /*
2688 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2689 : * applied to an update/delete target rel. If that ever becomes
2690 : * possible, we should drop the target from the PlanRowMark list.
2691 : */
2692 : Assert(rc->rti != parse->resultRelation);
2693 :
2694 : /*
2695 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2696 : * can't support true locking. Subqueries that got flattened into the
2697 : * main query should be ignored completely. Any that didn't will get
2698 : * ROW_MARK_COPY items in the next loop.
2699 : */
2700 4948 : if (rte->rtekind != RTE_RELATION)
2701 76 : continue;
2702 :
2703 4872 : rels = bms_del_member(rels, rc->rti);
2704 :
2705 4872 : newrc = makeNode(PlanRowMark);
2706 4872 : newrc->rti = newrc->prti = rc->rti;
2707 4872 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2708 4872 : newrc->markType = select_rowmark_type(rte, rc->strength);
2709 4872 : newrc->allMarkTypes = (1 << newrc->markType);
2710 4872 : newrc->strength = rc->strength;
2711 4872 : newrc->waitPolicy = rc->waitPolicy;
2712 4872 : newrc->isParent = false;
2713 :
2714 4872 : prowmarks = lappend(prowmarks, newrc);
2715 : }
2716 :
2717 : /*
2718 : * Now, add rowmarks for any non-target, non-locked base relations.
2719 : */
2720 16716 : i = 0;
2721 37042 : foreach(l, parse->rtable)
2722 : {
2723 20326 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2724 : PlanRowMark *newrc;
2725 :
2726 20326 : i++;
2727 20326 : if (!bms_is_member(i, rels))
2728 19340 : continue;
2729 :
2730 986 : newrc = makeNode(PlanRowMark);
2731 986 : newrc->rti = newrc->prti = i;
2732 986 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2733 986 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2734 986 : newrc->allMarkTypes = (1 << newrc->markType);
2735 986 : newrc->strength = LCS_NONE;
2736 986 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2737 986 : newrc->isParent = false;
2738 :
2739 986 : prowmarks = lappend(prowmarks, newrc);
2740 : }
2741 :
2742 16716 : root->rowMarks = prowmarks;
2743 : }
2744 :
2745 : /*
2746 : * Select RowMarkType to use for a given table
2747 : */
2748 : RowMarkType
2749 6792 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2750 : {
2751 6792 : if (rte->rtekind != RTE_RELATION)
2752 : {
2753 : /* If it's not a table at all, use ROW_MARK_COPY */
2754 160 : return ROW_MARK_COPY;
2755 : }
2756 6632 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2757 : {
2758 : /* Let the FDW select the rowmark type, if it wants to */
2759 160 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2760 :
2761 160 : if (fdwroutine->GetForeignRowMarkType != NULL)
2762 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2763 : /* Otherwise, use ROW_MARK_COPY by default */
2764 160 : return ROW_MARK_COPY;
2765 : }
2766 : else
2767 : {
2768 : /* Regular table, apply the appropriate lock type */
2769 6472 : switch (strength)
2770 : {
2771 : case LCS_NONE:
2772 :
2773 : /*
2774 : * We don't need a tuple lock, only the ability to re-fetch
2775 : * the row.
2776 : */
2777 1142 : return ROW_MARK_REFERENCE;
2778 : break;
2779 : case LCS_FORKEYSHARE:
2780 3842 : return ROW_MARK_KEYSHARE;
2781 : break;
2782 : case LCS_FORSHARE:
2783 218 : return ROW_MARK_SHARE;
2784 : break;
2785 : case LCS_FORNOKEYUPDATE:
2786 50 : return ROW_MARK_NOKEYEXCLUSIVE;
2787 : break;
2788 : case LCS_FORUPDATE:
2789 1220 : return ROW_MARK_EXCLUSIVE;
2790 : break;
2791 : }
2792 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2793 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2794 : }
2795 : }
2796 :
2797 : /*
2798 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2799 : *
2800 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2801 : * results back in *count_est and *offset_est. These variables are set to
2802 : * 0 if the corresponding clause is not present, and -1 if it's present
2803 : * but we couldn't estimate the value for it. (The "0" convention is OK
2804 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2805 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2806 : * usual practice of never estimating less than one row.) These values will
2807 : * be passed to create_limit_path, which see if you change this code.
2808 : *
2809 : * The return value is the suitably adjusted tuple_fraction to use for
2810 : * planning the query. This adjustment is not overridable, since it reflects
2811 : * plan actions that grouping_planner() will certainly take, not assumptions
2812 : * about context.
2813 : */
2814 : static double
2815 2890 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2816 : int64 *offset_est, int64 *count_est)
2817 : {
2818 2890 : Query *parse = root->parse;
2819 : Node *est;
2820 : double limit_fraction;
2821 :
2822 : /* Should not be called unless LIMIT or OFFSET */
2823 : Assert(parse->limitCount || parse->limitOffset);
2824 :
2825 : /*
2826 : * Try to obtain the clause values. We use estimate_expression_value
2827 : * primarily because it can sometimes do something useful with Params.
2828 : */
2829 2890 : if (parse->limitCount)
2830 : {
2831 2688 : est = estimate_expression_value(root, parse->limitCount);
2832 2688 : if (est && IsA(est, Const))
2833 : {
2834 5368 : if (((Const *) est)->constisnull)
2835 : {
2836 : /* NULL indicates LIMIT ALL, ie, no limit */
2837 0 : *count_est = 0; /* treat as not present */
2838 : }
2839 : else
2840 : {
2841 2684 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2842 2684 : if (*count_est <= 0)
2843 96 : *count_est = 1; /* force to at least 1 */
2844 : }
2845 : }
2846 : else
2847 4 : *count_est = -1; /* can't estimate */
2848 : }
2849 : else
2850 202 : *count_est = 0; /* not present */
2851 :
2852 2890 : if (parse->limitOffset)
2853 : {
2854 520 : est = estimate_expression_value(root, parse->limitOffset);
2855 520 : if (est && IsA(est, Const))
2856 : {
2857 1008 : if (((Const *) est)->constisnull)
2858 : {
2859 : /* Treat NULL as no offset; the executor will too */
2860 0 : *offset_est = 0; /* treat as not present */
2861 : }
2862 : else
2863 : {
2864 504 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2865 504 : if (*offset_est < 0)
2866 0 : *offset_est = 0; /* treat as not present */
2867 : }
2868 : }
2869 : else
2870 16 : *offset_est = -1; /* can't estimate */
2871 : }
2872 : else
2873 2370 : *offset_est = 0; /* not present */
2874 :
2875 2890 : if (*count_est != 0)
2876 : {
2877 : /*
2878 : * A LIMIT clause limits the absolute number of tuples returned.
2879 : * However, if it's not a constant LIMIT then we have to guess; for
2880 : * lack of a better idea, assume 10% of the plan's result is wanted.
2881 : */
2882 2688 : if (*count_est < 0 || *offset_est < 0)
2883 : {
2884 : /* LIMIT or OFFSET is an expression ... punt ... */
2885 16 : limit_fraction = 0.10;
2886 : }
2887 : else
2888 : {
2889 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2890 2672 : limit_fraction = (double) *count_est + (double) *offset_est;
2891 : }
2892 :
2893 : /*
2894 : * If we have absolute limits from both caller and LIMIT, use the
2895 : * smaller value; likewise if they are both fractional. If one is
2896 : * fractional and the other absolute, we can't easily determine which
2897 : * is smaller, but we use the heuristic that the absolute will usually
2898 : * be smaller.
2899 : */
2900 2688 : if (tuple_fraction >= 1.0)
2901 : {
2902 4 : if (limit_fraction >= 1.0)
2903 : {
2904 : /* both absolute */
2905 4 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2906 : }
2907 : else
2908 : {
2909 : /* caller absolute, limit fractional; use caller's value */
2910 : }
2911 : }
2912 2684 : else if (tuple_fraction > 0.0)
2913 : {
2914 94 : if (limit_fraction >= 1.0)
2915 : {
2916 : /* caller fractional, limit absolute; use limit */
2917 94 : tuple_fraction = limit_fraction;
2918 : }
2919 : else
2920 : {
2921 : /* both fractional */
2922 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2923 : }
2924 : }
2925 : else
2926 : {
2927 : /* no info from caller, just use limit */
2928 2590 : tuple_fraction = limit_fraction;
2929 : }
2930 : }
2931 202 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2932 : {
2933 : /*
2934 : * We have an OFFSET but no LIMIT. This acts entirely differently
2935 : * from the LIMIT case: here, we need to increase rather than decrease
2936 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2937 : * tuples to be fetched instead of fewer. This only matters if we got
2938 : * a tuple_fraction > 0, however.
2939 : *
2940 : * As above, use 10% if OFFSET is present but unestimatable.
2941 : */
2942 8 : if (*offset_est < 0)
2943 0 : limit_fraction = 0.10;
2944 : else
2945 8 : limit_fraction = (double) *offset_est;
2946 :
2947 : /*
2948 : * If we have absolute counts from both caller and OFFSET, add them
2949 : * together; likewise if they are both fractional. If one is
2950 : * fractional and the other absolute, we want to take the larger, and
2951 : * we heuristically assume that's the fractional one.
2952 : */
2953 8 : if (tuple_fraction >= 1.0)
2954 : {
2955 0 : if (limit_fraction >= 1.0)
2956 : {
2957 : /* both absolute, so add them together */
2958 0 : tuple_fraction += limit_fraction;
2959 : }
2960 : else
2961 : {
2962 : /* caller absolute, limit fractional; use limit */
2963 0 : tuple_fraction = limit_fraction;
2964 : }
2965 : }
2966 : else
2967 : {
2968 8 : if (limit_fraction >= 1.0)
2969 : {
2970 : /* caller fractional, limit absolute; use caller's value */
2971 : }
2972 : else
2973 : {
2974 : /* both fractional, so add them together */
2975 0 : tuple_fraction += limit_fraction;
2976 0 : if (tuple_fraction >= 1.0)
2977 0 : tuple_fraction = 0.0; /* assume fetch all */
2978 : }
2979 : }
2980 : }
2981 :
2982 2890 : return tuple_fraction;
2983 : }
2984 :
2985 : /*
2986 : * limit_needed - do we actually need a Limit plan node?
2987 : *
2988 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2989 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2990 : * locution for an optimization fence. (Because other places in the planner
2991 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2992 : * an optimization fence --- we're just suppressing unnecessary run-time
2993 : * overhead.)
2994 : *
2995 : * This might look like it could be merged into preprocess_limit, but there's
2996 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2997 : * in preprocess_limit it's good enough to consider estimated values.
2998 : */
2999 : bool
3000 596076 : limit_needed(Query *parse)
3001 : {
3002 : Node *node;
3003 :
3004 596076 : node = parse->limitCount;
3005 596076 : if (node)
3006 : {
3007 6162 : if (IsA(node, Const))
3008 : {
3009 : /* NULL indicates LIMIT ALL, ie, no limit */
3010 6070 : if (!((Const *) node)->constisnull)
3011 6070 : return true; /* LIMIT with a constant value */
3012 : }
3013 : else
3014 92 : return true; /* non-constant LIMIT */
3015 : }
3016 :
3017 589914 : node = parse->limitOffset;
3018 589914 : if (node)
3019 : {
3020 616 : if (IsA(node, Const))
3021 : {
3022 : /* Treat NULL as no offset; the executor would too */
3023 470 : if (!((Const *) node)->constisnull)
3024 : {
3025 470 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3026 :
3027 470 : if (offset != 0)
3028 24 : return true; /* OFFSET with a nonzero value */
3029 : }
3030 : }
3031 : else
3032 146 : return true; /* non-constant OFFSET */
3033 : }
3034 :
3035 589744 : return false; /* don't need a Limit plan node */
3036 : }
3037 :
3038 :
3039 : /*
3040 : * remove_useless_groupby_columns
3041 : * Remove any columns in the GROUP BY clause that are redundant due to
3042 : * being functionally dependent on other GROUP BY columns.
3043 : *
3044 : * Since some other DBMSes do not allow references to ungrouped columns, it's
3045 : * not unusual to find all columns listed in GROUP BY even though listing the
3046 : * primary-key columns would be sufficient. Deleting such excess columns
3047 : * avoids redundant sorting work, so it's worth doing. When we do this, we
3048 : * must mark the plan as dependent on the pkey constraint (compare the
3049 : * parser's check_ungrouped_columns() and check_functional_grouping()).
3050 : *
3051 : * In principle, we could treat any NOT-NULL columns appearing in a UNIQUE
3052 : * index as the determining columns. But as with check_functional_grouping(),
3053 : * there's currently no way to represent dependency on a NOT NULL constraint,
3054 : * so we consider only the pkey for now.
3055 : */
3056 : static void
3057 287726 : remove_useless_groupby_columns(PlannerInfo *root)
3058 : {
3059 287726 : Query *parse = root->parse;
3060 : Bitmapset **groupbyattnos;
3061 : Bitmapset **surplusvars;
3062 : ListCell *lc;
3063 : int relid;
3064 :
3065 : /* No chance to do anything if there are less than two GROUP BY items */
3066 287726 : if (list_length(parse->groupClause) < 2)
3067 286990 : return;
3068 :
3069 : /* Don't fiddle with the GROUP BY clause if the query has grouping sets */
3070 736 : if (parse->groupingSets)
3071 332 : return;
3072 :
3073 : /*
3074 : * Scan the GROUP BY clause to find GROUP BY items that are simple Vars.
3075 : * Fill groupbyattnos[k] with a bitmapset of the column attnos of RTE k
3076 : * that are GROUP BY items.
3077 : */
3078 404 : groupbyattnos = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3079 404 : (list_length(parse->rtable) + 1));
3080 1440 : foreach(lc, parse->groupClause)
3081 : {
3082 1036 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3083 1036 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3084 1036 : Var *var = (Var *) tle->expr;
3085 :
3086 : /*
3087 : * Ignore non-Vars and Vars from other query levels.
3088 : *
3089 : * XXX in principle, stable expressions containing Vars could also be
3090 : * removed, if all the Vars are functionally dependent on other GROUP
3091 : * BY items. But it's not clear that such cases occur often enough to
3092 : * be worth troubling over.
3093 : */
3094 1944 : if (!IsA(var, Var) ||
3095 908 : var->varlevelsup > 0)
3096 128 : continue;
3097 :
3098 : /* OK, remember we have this Var */
3099 908 : relid = var->varno;
3100 : Assert(relid <= list_length(parse->rtable));
3101 1816 : groupbyattnos[relid] = bms_add_member(groupbyattnos[relid],
3102 908 : var->varattno - FirstLowInvalidHeapAttributeNumber);
3103 : }
3104 :
3105 : /*
3106 : * Consider each relation and see if it is possible to remove some of its
3107 : * Vars from GROUP BY. For simplicity and speed, we do the actual removal
3108 : * in a separate pass. Here, we just fill surplusvars[k] with a bitmapset
3109 : * of the column attnos of RTE k that are removable GROUP BY items.
3110 : */
3111 404 : surplusvars = NULL; /* don't allocate array unless required */
3112 404 : relid = 0;
3113 1248 : foreach(lc, parse->rtable)
3114 : {
3115 844 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
3116 : Bitmapset *relattnos;
3117 : Bitmapset *pkattnos;
3118 : Oid constraintOid;
3119 :
3120 844 : relid++;
3121 :
3122 : /* Only plain relations could have primary-key constraints */
3123 844 : if (rte->rtekind != RTE_RELATION)
3124 1030 : continue;
3125 :
3126 : /* Nothing to do unless this rel has multiple Vars in GROUP BY */
3127 620 : relattnos = groupbyattnos[relid];
3128 620 : if (bms_membership(relattnos) != BMS_MULTIPLE)
3129 344 : continue;
3130 :
3131 : /*
3132 : * Can't remove any columns for this rel if there is no suitable
3133 : * (i.e., nondeferrable) primary key constraint.
3134 : */
3135 276 : pkattnos = get_primary_key_attnos(rte->relid, false, &constraintOid);
3136 276 : if (pkattnos == NULL)
3137 238 : continue;
3138 :
3139 : /*
3140 : * If the primary key is a proper subset of relattnos then we have
3141 : * some items in the GROUP BY that can be removed.
3142 : */
3143 38 : if (bms_subset_compare(pkattnos, relattnos) == BMS_SUBSET1)
3144 : {
3145 : /*
3146 : * To easily remember whether we've found anything to do, we don't
3147 : * allocate the surplusvars[] array until we find something.
3148 : */
3149 26 : if (surplusvars == NULL)
3150 22 : surplusvars = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3151 22 : (list_length(parse->rtable) + 1));
3152 :
3153 : /* Remember the attnos of the removable columns */
3154 26 : surplusvars[relid] = bms_difference(relattnos, pkattnos);
3155 :
3156 : /* Also, mark the resulting plan as dependent on this constraint */
3157 26 : parse->constraintDeps = lappend_oid(parse->constraintDeps,
3158 : constraintOid);
3159 : }
3160 : }
3161 :
3162 : /*
3163 : * If we found any surplus Vars, build a new GROUP BY clause without them.
3164 : * (Note: this may leave some TLEs with unreferenced ressortgroupref
3165 : * markings, but that's harmless.)
3166 : */
3167 404 : if (surplusvars != NULL)
3168 : {
3169 22 : List *new_groupby = NIL;
3170 :
3171 110 : foreach(lc, parse->groupClause)
3172 : {
3173 88 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3174 88 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3175 88 : Var *var = (Var *) tle->expr;
3176 :
3177 : /*
3178 : * New list must include non-Vars, outer Vars, and anything not
3179 : * marked as surplus.
3180 : */
3181 176 : if (!IsA(var, Var) ||
3182 176 : var->varlevelsup > 0 ||
3183 88 : !bms_is_member(var->varattno - FirstLowInvalidHeapAttributeNumber,
3184 88 : surplusvars[var->varno]))
3185 50 : new_groupby = lappend(new_groupby, sgc);
3186 : }
3187 :
3188 22 : parse->groupClause = new_groupby;
3189 : }
3190 : }
3191 :
3192 : /*
3193 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3194 : *
3195 : * The idea here is to adjust the ordering of the GROUP BY elements
3196 : * (which in itself is semantically insignificant) to match ORDER BY,
3197 : * thereby allowing a single sort operation to both implement the ORDER BY
3198 : * requirement and set up for a Unique step that implements GROUP BY.
3199 : *
3200 : * In principle it might be interesting to consider other orderings of the
3201 : * GROUP BY elements, which could match the sort ordering of other
3202 : * possible plans (eg an indexscan) and thereby reduce cost. We don't
3203 : * bother with that, though. Hashed grouping will frequently win anyway.
3204 : *
3205 : * Note: we need no comparable processing of the distinctClause because
3206 : * the parser already enforced that that matches ORDER BY.
3207 : *
3208 : * For grouping sets, the order of items is instead forced to agree with that
3209 : * of the grouping set (and items not in the grouping set are skipped). The
3210 : * work of sorting the order of grouping set elements to match the ORDER BY if
3211 : * possible is done elsewhere.
3212 : */
3213 : static List *
3214 3398 : preprocess_groupclause(PlannerInfo *root, List *force)
3215 : {
3216 3398 : Query *parse = root->parse;
3217 3398 : List *new_groupclause = NIL;
3218 : bool partial_match;
3219 : ListCell *sl;
3220 : ListCell *gl;
3221 :
3222 : /* For grouping sets, we need to force the ordering */
3223 3398 : if (force)
3224 : {
3225 4252 : foreach(sl, force)
3226 : {
3227 2504 : Index ref = lfirst_int(sl);
3228 2504 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3229 :
3230 2504 : new_groupclause = lappend(new_groupclause, cl);
3231 : }
3232 :
3233 1748 : return new_groupclause;
3234 : }
3235 :
3236 : /* If no ORDER BY, nothing useful to do here */
3237 1650 : if (parse->sortClause == NIL)
3238 742 : return parse->groupClause;
3239 :
3240 : /*
3241 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3242 : * items, but only as far as we can make a matching prefix.
3243 : *
3244 : * This code assumes that the sortClause contains no duplicate items.
3245 : */
3246 1792 : foreach(sl, parse->sortClause)
3247 : {
3248 1238 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3249 :
3250 1832 : foreach(gl, parse->groupClause)
3251 : {
3252 1478 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3253 :
3254 1478 : if (equal(gc, sc))
3255 : {
3256 884 : new_groupclause = lappend(new_groupclause, gc);
3257 884 : break;
3258 : }
3259 : }
3260 1238 : if (gl == NULL)
3261 354 : break; /* no match, so stop scanning */
3262 : }
3263 :
3264 : /* Did we match all of the ORDER BY list, or just some of it? */
3265 908 : partial_match = (sl != NULL);
3266 :
3267 : /* If no match at all, no point in reordering GROUP BY */
3268 908 : if (new_groupclause == NIL)
3269 128 : return parse->groupClause;
3270 :
3271 : /*
3272 : * Add any remaining GROUP BY items to the new list, but only if we were
3273 : * able to make a complete match. In other words, we only rearrange the
3274 : * GROUP BY list if the result is that one list is a prefix of the other
3275 : * --- otherwise there's no possibility of a common sort. Also, give up
3276 : * if there are any non-sortable GROUP BY items, since then there's no
3277 : * hope anyway.
3278 : */
3279 1684 : foreach(gl, parse->groupClause)
3280 : {
3281 954 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3282 :
3283 954 : if (list_member_ptr(new_groupclause, gc))
3284 868 : continue; /* it matched an ORDER BY item */
3285 86 : if (partial_match)
3286 50 : return parse->groupClause; /* give up, no common sort possible */
3287 36 : if (!OidIsValid(gc->sortop))
3288 0 : return parse->groupClause; /* give up, GROUP BY can't be sorted */
3289 36 : new_groupclause = lappend(new_groupclause, gc);
3290 : }
3291 :
3292 : /* Success --- install the rearranged GROUP BY list */
3293 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3294 730 : return new_groupclause;
3295 : }
3296 :
3297 : /*
3298 : * Extract lists of grouping sets that can be implemented using a single
3299 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3300 : *
3301 : * Input must be sorted with smallest sets first. Result has each sublist
3302 : * sorted with smallest sets first.
3303 : *
3304 : * We want to produce the absolute minimum possible number of lists here to
3305 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3306 : * of finding the minimal partition of a partially-ordered set into chains
3307 : * (which is what we need, taking the list of grouping sets as a poset ordered
3308 : * by set inclusion) can be mapped to the problem of finding the maximum
3309 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3310 : * time with a worst case of no worse than O(n^2.5) and usually much
3311 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3312 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3313 : * half a second on my modest system even with optimization off and assertions
3314 : * on.)
3315 : */
3316 : static List *
3317 408 : extract_rollup_sets(List *groupingSets)
3318 : {
3319 408 : int num_sets_raw = list_length(groupingSets);
3320 408 : int num_empty = 0;
3321 408 : int num_sets = 0; /* distinct sets */
3322 408 : int num_chains = 0;
3323 408 : List *result = NIL;
3324 : List **results;
3325 : List **orig_sets;
3326 : Bitmapset **set_masks;
3327 : int *chains;
3328 : short **adjacency;
3329 : short *adjacency_buf;
3330 : BipartiteMatchState *state;
3331 : int i;
3332 : int j;
3333 : int j_size;
3334 408 : ListCell *lc1 = list_head(groupingSets);
3335 : ListCell *lc;
3336 :
3337 : /*
3338 : * Start by stripping out empty sets. The algorithm doesn't require this,
3339 : * but the planner currently needs all empty sets to be returned in the
3340 : * first list, so we strip them here and add them back after.
3341 : */
3342 1136 : while (lc1 && lfirst(lc1) == NIL)
3343 : {
3344 320 : ++num_empty;
3345 320 : lc1 = lnext(lc1);
3346 : }
3347 :
3348 : /* bail out now if it turns out that all we had were empty sets. */
3349 408 : if (!lc1)
3350 28 : return list_make1(groupingSets);
3351 :
3352 : /*----------
3353 : * We don't strictly need to remove duplicate sets here, but if we don't,
3354 : * they tend to become scattered through the result, which is a bit
3355 : * confusing (and irritating if we ever decide to optimize them out).
3356 : * So we remove them here and add them back after.
3357 : *
3358 : * For each non-duplicate set, we fill in the following:
3359 : *
3360 : * orig_sets[i] = list of the original set lists
3361 : * set_masks[i] = bitmapset for testing inclusion
3362 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3363 : *
3364 : * chains[i] will be the result group this set is assigned to.
3365 : *
3366 : * We index all of these from 1 rather than 0 because it is convenient
3367 : * to leave 0 free for the NIL node in the graph algorithm.
3368 : *----------
3369 : */
3370 380 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3371 380 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3372 380 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3373 380 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3374 :
3375 380 : j_size = 0;
3376 380 : j = 0;
3377 380 : i = 1;
3378 :
3379 1288 : for_each_cell(lc, lc1)
3380 : {
3381 908 : List *candidate = (List *) lfirst(lc);
3382 908 : Bitmapset *candidate_set = NULL;
3383 : ListCell *lc2;
3384 908 : int dup_of = 0;
3385 :
3386 2176 : foreach(lc2, candidate)
3387 : {
3388 1268 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3389 : }
3390 :
3391 : /* we can only be a dup if we're the same length as a previous set */
3392 908 : if (j_size == list_length(candidate))
3393 : {
3394 : int k;
3395 :
3396 792 : for (k = j; k < i; ++k)
3397 : {
3398 512 : if (bms_equal(set_masks[k], candidate_set))
3399 : {
3400 52 : dup_of = k;
3401 52 : break;
3402 : }
3403 : }
3404 : }
3405 576 : else if (j_size < list_length(candidate))
3406 : {
3407 576 : j_size = list_length(candidate);
3408 576 : j = i;
3409 : }
3410 :
3411 908 : if (dup_of > 0)
3412 : {
3413 52 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3414 52 : bms_free(candidate_set);
3415 : }
3416 : else
3417 : {
3418 : int k;
3419 856 : int n_adj = 0;
3420 :
3421 856 : orig_sets[i] = list_make1(candidate);
3422 856 : set_masks[i] = candidate_set;
3423 :
3424 : /* fill in adjacency list; no need to compare equal-size sets */
3425 :
3426 1316 : for (k = j - 1; k > 0; --k)
3427 : {
3428 460 : if (bms_is_subset(set_masks[k], candidate_set))
3429 400 : adjacency_buf[++n_adj] = k;
3430 : }
3431 :
3432 856 : if (n_adj > 0)
3433 : {
3434 228 : adjacency_buf[0] = n_adj;
3435 228 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3436 228 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3437 : }
3438 : else
3439 628 : adjacency[i] = NULL;
3440 :
3441 856 : ++i;
3442 : }
3443 : }
3444 :
3445 380 : num_sets = i - 1;
3446 :
3447 : /*
3448 : * Apply the graph matching algorithm to do the work.
3449 : */
3450 380 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3451 :
3452 : /*
3453 : * Now, the state->pair* fields have the info we need to assign sets to
3454 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3455 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3456 : * it in one pass)
3457 : */
3458 380 : chains = palloc0((num_sets + 1) * sizeof(int));
3459 :
3460 1236 : for (i = 1; i <= num_sets; ++i)
3461 : {
3462 856 : int u = state->pair_vu[i];
3463 856 : int v = state->pair_uv[i];
3464 :
3465 856 : if (u > 0 && u < i)
3466 0 : chains[i] = chains[u];
3467 856 : else if (v > 0 && v < i)
3468 228 : chains[i] = chains[v];
3469 : else
3470 628 : chains[i] = ++num_chains;
3471 : }
3472 :
3473 : /* build result lists. */
3474 380 : results = palloc0((num_chains + 1) * sizeof(List *));
3475 :
3476 1236 : for (i = 1; i <= num_sets; ++i)
3477 : {
3478 856 : int c = chains[i];
3479 :
3480 : Assert(c > 0);
3481 :
3482 856 : results[c] = list_concat(results[c], orig_sets[i]);
3483 : }
3484 :
3485 : /* push any empty sets back on the first list. */
3486 1020 : while (num_empty-- > 0)
3487 260 : results[1] = lcons(NIL, results[1]);
3488 :
3489 : /* make result list */
3490 1008 : for (i = 1; i <= num_chains; ++i)
3491 628 : result = lappend(result, results[i]);
3492 :
3493 : /*
3494 : * Free all the things.
3495 : *
3496 : * (This is over-fussy for small sets but for large sets we could have
3497 : * tied up a nontrivial amount of memory.)
3498 : */
3499 380 : BipartiteMatchFree(state);
3500 380 : pfree(results);
3501 380 : pfree(chains);
3502 1236 : for (i = 1; i <= num_sets; ++i)
3503 856 : if (adjacency[i])
3504 228 : pfree(adjacency[i]);
3505 380 : pfree(adjacency);
3506 380 : pfree(adjacency_buf);
3507 380 : pfree(orig_sets);
3508 1236 : for (i = 1; i <= num_sets; ++i)
3509 856 : bms_free(set_masks[i]);
3510 380 : pfree(set_masks);
3511 :
3512 380 : return result;
3513 : }
3514 :
3515 : /*
3516 : * Reorder the elements of a list of grouping sets such that they have correct
3517 : * prefix relationships. Also inserts the GroupingSetData annotations.
3518 : *
3519 : * The input must be ordered with smallest sets first; the result is returned
3520 : * with largest sets first. Note that the result shares no list substructure
3521 : * with the input, so it's safe for the caller to modify it later.
3522 : *
3523 : * If we're passed in a sortclause, we follow its order of columns to the
3524 : * extent possible, to minimize the chance that we add unnecessary sorts.
3525 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3526 : * gets implemented in one pass.)
3527 : */
3528 : static List *
3529 656 : reorder_grouping_sets(List *groupingsets, List *sortclause)
3530 : {
3531 : ListCell *lc;
3532 : ListCell *lc2;
3533 656 : List *previous = NIL;
3534 656 : List *result = NIL;
3535 :
3536 1884 : foreach(lc, groupingsets)
3537 : {
3538 1228 : List *candidate = (List *) lfirst(lc);
3539 1228 : List *new_elems = list_difference_int(candidate, previous);
3540 1228 : GroupingSetData *gs = makeNode(GroupingSetData);
3541 :
3542 1228 : if (list_length(new_elems) > 0)
3543 : {
3544 1748 : while (list_length(sortclause) > list_length(previous))
3545 : {
3546 120 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3547 120 : int ref = sc->tleSortGroupRef;
3548 :
3549 120 : if (list_member_int(new_elems, ref))
3550 : {
3551 36 : previous = lappend_int(previous, ref);
3552 36 : new_elems = list_delete_int(new_elems, ref);
3553 : }
3554 : else
3555 : {
3556 : /* diverged from the sortclause; give up on it */
3557 84 : sortclause = NIL;
3558 84 : break;
3559 : }
3560 : }
3561 :
3562 1784 : foreach(lc2, new_elems)
3563 : {
3564 928 : previous = lappend_int(previous, lfirst_int(lc2));
3565 : }
3566 : }
3567 :
3568 1228 : gs->set = list_copy(previous);
3569 1228 : result = lcons(gs, result);
3570 1228 : list_free(new_elems);
3571 : }
3572 :
3573 656 : list_free(previous);
3574 :
3575 656 : return result;
3576 : }
3577 :
3578 : /*
3579 : * Compute query_pathkeys and other pathkeys during plan generation
3580 : */
3581 : static void
3582 288578 : standard_qp_callback(PlannerInfo *root, void *extra)
3583 : {
3584 288578 : Query *parse = root->parse;
3585 288578 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3586 288578 : List *tlist = root->processed_tlist;
3587 288578 : List *activeWindows = qp_extra->activeWindows;
3588 :
3589 : /*
3590 : * Calculate pathkeys that represent grouping/ordering requirements. The
3591 : * sortClause is certainly sort-able, but GROUP BY and DISTINCT might not
3592 : * be, in which case we just leave their pathkeys empty.
3593 : */
3594 290608 : if (qp_extra->groupClause &&
3595 2030 : grouping_is_sortable(qp_extra->groupClause))
3596 2030 : root->group_pathkeys =
3597 2030 : make_pathkeys_for_sortclauses(root,
3598 : qp_extra->groupClause,
3599 : tlist);
3600 : else
3601 286548 : root->group_pathkeys = NIL;
3602 :
3603 : /* We consider only the first (bottom) window in pathkeys logic */
3604 288578 : if (activeWindows != NIL)
3605 : {
3606 1076 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3607 :
3608 1076 : root->window_pathkeys = make_pathkeys_for_window(root,
3609 : wc,
3610 : tlist);
3611 : }
3612 : else
3613 287502 : root->window_pathkeys = NIL;
3614 :
3615 288914 : if (parse->distinctClause &&
3616 336 : grouping_is_sortable(parse->distinctClause))
3617 332 : root->distinct_pathkeys =
3618 332 : make_pathkeys_for_sortclauses(root,
3619 : parse->distinctClause,
3620 : tlist);
3621 : else
3622 288246 : root->distinct_pathkeys = NIL;
3623 :
3624 288578 : root->sort_pathkeys =
3625 288578 : make_pathkeys_for_sortclauses(root,
3626 : parse->sortClause,
3627 : tlist);
3628 :
3629 : /*
3630 : * Figure out whether we want a sorted result from query_planner.
3631 : *
3632 : * If we have a sortable GROUP BY clause, then we want a result sorted
3633 : * properly for grouping. Otherwise, if we have window functions to
3634 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3635 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3636 : * we try to produce output that's sufficiently well sorted for the
3637 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3638 : * by the ORDER BY clause.
3639 : *
3640 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3641 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3642 : * that might just leave us failing to exploit an available sort order at
3643 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3644 : * much easier, since we know that the parser ensured that one is a
3645 : * superset of the other.
3646 : */
3647 288578 : if (root->group_pathkeys)
3648 1960 : root->query_pathkeys = root->group_pathkeys;
3649 286618 : else if (root->window_pathkeys)
3650 932 : root->query_pathkeys = root->window_pathkeys;
3651 571372 : else if (list_length(root->distinct_pathkeys) >
3652 285686 : list_length(root->sort_pathkeys))
3653 186 : root->query_pathkeys = root->distinct_pathkeys;
3654 285500 : else if (root->sort_pathkeys)
3655 32518 : root->query_pathkeys = root->sort_pathkeys;
3656 : else
3657 252982 : root->query_pathkeys = NIL;
3658 288578 : }
3659 :
3660 : /*
3661 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3662 : *
3663 : * path_rows: number of output rows from scan/join step
3664 : * gd: grouping sets data including list of grouping sets and their clauses
3665 : * target_list: target list containing group clause references
3666 : *
3667 : * If doing grouping sets, we also annotate the gsets data with the estimates
3668 : * for each set and each individual rollup list, with a view to later
3669 : * determining whether some combination of them could be hashed instead.
3670 : */
3671 : static double
3672 25564 : get_number_of_groups(PlannerInfo *root,
3673 : double path_rows,
3674 : grouping_sets_data *gd,
3675 : List *target_list)
3676 : {
3677 25564 : Query *parse = root->parse;
3678 : double dNumGroups;
3679 :
3680 25564 : if (parse->groupClause)
3681 : {
3682 : List *groupExprs;
3683 :
3684 3552 : if (parse->groupingSets)
3685 : {
3686 : /* Add up the estimates for each grouping set */
3687 : ListCell *lc;
3688 : ListCell *lc2;
3689 :
3690 : Assert(gd); /* keep Coverity happy */
3691 :
3692 384 : dNumGroups = 0;
3693 :
3694 1012 : foreach(lc, gd->rollups)
3695 : {
3696 628 : RollupData *rollup = lfirst_node(RollupData, lc);
3697 : ListCell *lc;
3698 :
3699 628 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3700 : target_list);
3701 :
3702 628 : rollup->numGroups = 0.0;
3703 :
3704 1796 : forboth(lc, rollup->gsets, lc2, rollup->gsets_data)
3705 : {
3706 1168 : List *gset = (List *) lfirst(lc);
3707 1168 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3708 1168 : double numGroups = estimate_num_groups(root,
3709 : groupExprs,
3710 : path_rows,
3711 : &gset);
3712 :
3713 1168 : gs->numGroups = numGroups;
3714 1168 : rollup->numGroups += numGroups;
3715 : }
3716 :
3717 628 : dNumGroups += rollup->numGroups;
3718 : }
3719 :
3720 384 : if (gd->hash_sets_idx)
3721 : {
3722 : ListCell *lc;
3723 :
3724 24 : gd->dNumHashGroups = 0;
3725 :
3726 24 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3727 : target_list);
3728 :
3729 52 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3730 : {
3731 28 : List *gset = (List *) lfirst(lc);
3732 28 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3733 28 : double numGroups = estimate_num_groups(root,
3734 : groupExprs,
3735 : path_rows,
3736 : &gset);
3737 :
3738 28 : gs->numGroups = numGroups;
3739 28 : gd->dNumHashGroups += numGroups;
3740 : }
3741 :
3742 24 : dNumGroups += gd->dNumHashGroups;
3743 : }
3744 : }
3745 : else
3746 : {
3747 : /* Plain GROUP BY */
3748 3168 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3749 : target_list);
3750 :
3751 3168 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3752 : NULL);
3753 : }
3754 : }
3755 22012 : else if (parse->groupingSets)
3756 : {
3757 : /* Empty grouping sets ... one result row for each one */
3758 28 : dNumGroups = list_length(parse->groupingSets);
3759 : }
3760 21984 : else if (parse->hasAggs || root->hasHavingQual)
3761 : {
3762 : /* Plain aggregation, one result row */
3763 21984 : dNumGroups = 1;
3764 : }
3765 : else
3766 : {
3767 : /* Not grouping */
3768 0 : dNumGroups = 1;
3769 : }
3770 :
3771 25564 : return dNumGroups;
3772 : }
3773 :
3774 : /*
3775 : * create_grouping_paths
3776 : *
3777 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3778 : * Along the way, we also build an upperrel for Paths which are partially
3779 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3780 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3781 : * the only partially grouped paths we build are also partial paths; that
3782 : * is, they need a Gather and then a FinalizeAggregate.
3783 : *
3784 : * input_rel: contains the source-data Paths
3785 : * target: the pathtarget for the result Paths to compute
3786 : * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
3787 : * gd: grouping sets data including list of grouping sets and their clauses
3788 : *
3789 : * Note: all Paths in input_rel are expected to return the target computed
3790 : * by make_group_input_target.
3791 : */
3792 : static RelOptInfo *
3793 23600 : create_grouping_paths(PlannerInfo *root,
3794 : RelOptInfo *input_rel,
3795 : PathTarget *target,
3796 : bool target_parallel_safe,
3797 : const AggClauseCosts *agg_costs,
3798 : grouping_sets_data *gd)
3799 : {
3800 23600 : Query *parse = root->parse;
3801 : RelOptInfo *grouped_rel;
3802 : RelOptInfo *partially_grouped_rel;
3803 :
3804 : /*
3805 : * Create grouping relation to hold fully aggregated grouping and/or
3806 : * aggregation paths.
3807 : */
3808 23600 : grouped_rel = make_grouping_rel(root, input_rel, target,
3809 : target_parallel_safe, parse->havingQual);
3810 :
3811 : /*
3812 : * Create either paths for a degenerate grouping or paths for ordinary
3813 : * grouping, as appropriate.
3814 : */
3815 23600 : if (is_degenerate_grouping(root))
3816 12 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3817 : else
3818 : {
3819 23588 : int flags = 0;
3820 : GroupPathExtraData extra;
3821 :
3822 : /*
3823 : * Determine whether it's possible to perform sort-based
3824 : * implementations of grouping. (Note that if groupClause is empty,
3825 : * grouping_is_sortable() is trivially true, and all the
3826 : * pathkeys_contained_in() tests will succeed too, so that we'll
3827 : * consider every surviving input path.)
3828 : *
3829 : * If we have grouping sets, we might be able to sort some but not all
3830 : * of them; in this case, we need can_sort to be true as long as we
3831 : * must consider any sorted-input plan.
3832 : */
3833 23588 : if ((gd && gd->rollups != NIL)
3834 23180 : || grouping_is_sortable(parse->groupClause))
3835 23584 : flags |= GROUPING_CAN_USE_SORT;
3836 :
3837 : /*
3838 : * Determine whether we should consider hash-based implementations of
3839 : * grouping.
3840 : *
3841 : * Hashed aggregation only applies if we're grouping. If we have
3842 : * grouping sets, some groups might be hashable but others not; in
3843 : * this case we set can_hash true as long as there is nothing globally
3844 : * preventing us from hashing (and we should therefore consider plans
3845 : * with hashes).
3846 : *
3847 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3848 : * BY aggregates. (Doing so would imply storing *all* the input
3849 : * values in the hash table, and/or running many sorts in parallel,
3850 : * either of which seems like a certain loser.) We similarly don't
3851 : * support ordered-set aggregates in hashed aggregation, but that case
3852 : * is also included in the numOrderedAggs count.
3853 : *
3854 : * Note: grouping_is_hashable() is much more expensive to check than
3855 : * the other gating conditions, so we want to do it last.
3856 : */
3857 25622 : if ((parse->groupClause != NIL &&
3858 5838 : agg_costs->numOrderedAggs == 0 &&
3859 1902 : (gd ? gd->any_hashable : grouping_is_hashable(parse->groupClause))))
3860 1898 : flags |= GROUPING_CAN_USE_HASH;
3861 :
3862 : /*
3863 : * Determine whether partial aggregation is possible.
3864 : */
3865 23588 : if (can_partial_agg(root, agg_costs))
3866 8038 : flags |= GROUPING_CAN_PARTIAL_AGG;
3867 :
3868 23588 : extra.flags = flags;
3869 23588 : extra.target_parallel_safe = target_parallel_safe;
3870 23588 : extra.havingQual = parse->havingQual;
3871 23588 : extra.targetList = parse->targetList;
3872 23588 : extra.partial_costs_set = false;
3873 :
3874 : /*
3875 : * Determine whether partitionwise aggregation is in theory possible.
3876 : * It can be disabled by the user, and for now, we don't try to
3877 : * support grouping sets. create_ordinary_grouping_paths() will check
3878 : * additional conditions, such as whether input_rel is partitioned.
3879 : */
3880 23588 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3881 294 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3882 : else
3883 23294 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3884 :
3885 23588 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3886 : agg_costs, gd, &extra,
3887 : &partially_grouped_rel);
3888 : }
3889 :
3890 23596 : set_cheapest(grouped_rel);
3891 23596 : return grouped_rel;
3892 : }
3893 :
3894 : /*
3895 : * make_grouping_rel
3896 : *
3897 : * Create a new grouping rel and set basic properties.
3898 : *
3899 : * input_rel represents the underlying scan/join relation.
3900 : * target is the output expected from the grouping relation.
3901 : */
3902 : static RelOptInfo *
3903 24518 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3904 : PathTarget *target, bool target_parallel_safe,
3905 : Node *havingQual)
3906 : {
3907 : RelOptInfo *grouped_rel;
3908 :
3909 24518 : if (IS_OTHER_REL(input_rel))
3910 : {
3911 918 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3912 : input_rel->relids);
3913 918 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3914 : }
3915 : else
3916 : {
3917 : /*
3918 : * By tradition, the relids set for the main grouping relation is
3919 : * NULL. (This could be changed, but might require adjustments
3920 : * elsewhere.)
3921 : */
3922 23600 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3923 : }
3924 :
3925 : /* Set target. */
3926 24518 : grouped_rel->reltarget = target;
3927 :
3928 : /*
3929 : * If the input relation is not parallel-safe, then the grouped relation
3930 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3931 : * target list and HAVING quals are parallel-safe.
3932 : */
3933 36010 : if (input_rel->consider_parallel && target_parallel_safe &&
3934 11492 : is_parallel_safe(root, (Node *) havingQual))
3935 11480 : grouped_rel->consider_parallel = true;
3936 :
3937 : /*
3938 : * If the input rel belongs to a single FDW, so does the grouped rel.
3939 : */
3940 24518 : grouped_rel->serverid = input_rel->serverid;
3941 24518 : grouped_rel->userid = input_rel->userid;
3942 24518 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3943 24518 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3944 :
3945 24518 : return grouped_rel;
3946 : }
3947 :
3948 : /*
3949 : * is_degenerate_grouping
3950 : *
3951 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3952 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3953 : * grouping sets are all empty).
3954 : */
3955 : static bool
3956 23600 : is_degenerate_grouping(PlannerInfo *root)
3957 : {
3958 23600 : Query *parse = root->parse;
3959 :
3960 47576 : return (root->hasHavingQual || parse->groupingSets) &&
3961 24388 : !parse->hasAggs && parse->groupClause == NIL;
3962 : }
3963 :
3964 : /*
3965 : * create_degenerate_grouping_paths
3966 : *
3967 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3968 : * supposed to emit either zero or one row for each grouping set depending on
3969 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3970 : * either HAVING or the targetlist, so we actually do not need the FROM table
3971 : * at all! We can just throw away the plan-so-far and generate a Result node.
3972 : * This is a sufficiently unusual corner case that it's not worth contorting
3973 : * the structure of this module to avoid having to generate the earlier paths
3974 : * in the first place.
3975 : */
3976 : static void
3977 12 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3978 : RelOptInfo *grouped_rel)
3979 : {
3980 12 : Query *parse = root->parse;
3981 : int nrows;
3982 : Path *path;
3983 :
3984 12 : nrows = list_length(parse->groupingSets);
3985 12 : if (nrows > 1)
3986 : {
3987 : /*
3988 : * Doesn't seem worthwhile writing code to cons up a generate_series
3989 : * or a values scan to emit multiple rows. Instead just make N clones
3990 : * and append them. (With a volatile HAVING clause, this means you
3991 : * might get between 0 and N output rows. Offhand I think that's
3992 : * desired.)
3993 : */
3994 0 : List *paths = NIL;
3995 :
3996 0 : while (--nrows >= 0)
3997 : {
3998 0 : path = (Path *)
3999 : create_group_result_path(root, grouped_rel,
4000 0 : grouped_rel->reltarget,
4001 0 : (List *) parse->havingQual);
4002 0 : paths = lappend(paths, path);
4003 : }
4004 0 : path = (Path *)
4005 : create_append_path(root,
4006 : grouped_rel,
4007 : paths,
4008 : NIL,
4009 : NIL,
4010 : NULL,
4011 : 0,
4012 : false,
4013 : NIL,
4014 : -1);
4015 : }
4016 : else
4017 : {
4018 : /* No grouping sets, or just one, so one output row */
4019 12 : path = (Path *)
4020 : create_group_result_path(root, grouped_rel,
4021 12 : grouped_rel->reltarget,
4022 12 : (List *) parse->havingQual);
4023 : }
4024 :
4025 12 : add_path(grouped_rel, path);
4026 12 : }
4027 :
4028 : /*
4029 : * create_ordinary_grouping_paths
4030 : *
4031 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4032 : *
4033 : * We need to consider sorted and hashed aggregation in the same function,
4034 : * because otherwise (1) it would be harder to throw an appropriate error
4035 : * message if neither way works, and (2) we should not allow hashtable size
4036 : * considerations to dissuade us from using hashing if sorting is not possible.
4037 : *
4038 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4039 : * function creates, or to NULL if it doesn't create one.
4040 : */
4041 : static void
4042 24506 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4043 : RelOptInfo *grouped_rel,
4044 : const AggClauseCosts *agg_costs,
4045 : grouping_sets_data *gd,
4046 : GroupPathExtraData *extra,
4047 : RelOptInfo **partially_grouped_rel_p)
4048 : {
4049 24506 : Path *cheapest_path = input_rel->cheapest_total_path;
4050 24506 : RelOptInfo *partially_grouped_rel = NULL;
4051 : double dNumGroups;
4052 24506 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4053 :
4054 : /*
4055 : * If this is the topmost grouping relation or if the parent relation is
4056 : * doing some form of partitionwise aggregation, then we may be able to do
4057 : * it at this level also. However, if the input relation is not
4058 : * partitioned, partitionwise aggregate is impossible.
4059 : */
4060 25718 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4061 1610 : IS_PARTITIONED_REL(input_rel))
4062 : {
4063 : /*
4064 : * If this is the topmost relation or if the parent relation is doing
4065 : * full partitionwise aggregation, then we can do full partitionwise
4066 : * aggregation provided that the GROUP BY clause contains all of the
4067 : * partitioning columns at this level. Otherwise, we can do at most
4068 : * partial partitionwise aggregation. But if partial aggregation is
4069 : * not supported in general then we can't use it for partitionwise
4070 : * aggregation either.
4071 : */
4072 700 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4073 334 : group_by_has_partkey(input_rel, extra->targetList,
4074 334 : root->parse->groupClause))
4075 192 : patype = PARTITIONWISE_AGGREGATE_FULL;
4076 174 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4077 146 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4078 : else
4079 28 : patype = PARTITIONWISE_AGGREGATE_NONE;
4080 : }
4081 :
4082 : /*
4083 : * Before generating paths for grouped_rel, we first generate any possible
4084 : * partially grouped paths; that way, later code can easily consider both
4085 : * parallel and non-parallel approaches to grouping.
4086 : */
4087 24506 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4088 : {
4089 : bool force_rel_creation;
4090 :
4091 : /*
4092 : * If we're doing partitionwise aggregation at this level, force
4093 : * creation of a partially_grouped_rel so we can add partitionwise
4094 : * paths to it.
4095 : */
4096 8908 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4097 :
4098 8908 : partially_grouped_rel =
4099 8908 : create_partial_grouping_paths(root,
4100 : grouped_rel,
4101 : input_rel,
4102 : gd,
4103 : extra,
4104 : force_rel_creation);
4105 : }
4106 :
4107 : /* Set out parameter. */
4108 24506 : *partially_grouped_rel_p = partially_grouped_rel;
4109 :
4110 : /* Apply partitionwise aggregation technique, if possible. */
4111 24506 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4112 338 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4113 : partially_grouped_rel, agg_costs,
4114 : gd, patype, extra);
4115 :
4116 : /* If we are doing partial aggregation only, return. */
4117 24506 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4118 : {
4119 : Assert(partially_grouped_rel);
4120 :
4121 374 : if (partially_grouped_rel->pathlist)
4122 374 : set_cheapest(partially_grouped_rel);
4123 :
4124 374 : return;
4125 : }
4126 :
4127 : /* Gather any partially grouped partial paths. */
4128 24132 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4129 : {
4130 858 : gather_grouping_paths(root, partially_grouped_rel);
4131 858 : set_cheapest(partially_grouped_rel);
4132 : }
4133 :
4134 : /*
4135 : * Estimate number of groups.
4136 : */
4137 24132 : dNumGroups = get_number_of_groups(root,
4138 : cheapest_path->rows,
4139 : gd,
4140 : extra->targetList);
4141 :
4142 : /* Build final grouping paths */
4143 24132 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4144 : partially_grouped_rel, agg_costs, gd,
4145 : dNumGroups, extra);
4146 :
4147 : /* Give a helpful error if we failed to find any implementation */
4148 24132 : if (grouped_rel->pathlist == NIL)
4149 4 : ereport(ERROR,
4150 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4151 : errmsg("could not implement GROUP BY"),
4152 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4153 :
4154 : /*
4155 : * If there is an FDW that's responsible for all baserels of the query,
4156 : * let it consider adding ForeignPaths.
4157 : */
4158 24392 : if (grouped_rel->fdwroutine &&
4159 264 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4160 264 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4161 : input_rel, grouped_rel,
4162 : extra);
4163 :
4164 : /* Let extensions possibly add some more paths */
4165 24128 : if (create_upper_paths_hook)
4166 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4167 : input_rel, grouped_rel,
4168 : extra);
4169 : }
4170 :
4171 : /*
4172 : * For a given input path, consider the possible ways of doing grouping sets on
4173 : * it, by combinations of hashing and sorting. This can be called multiple
4174 : * times, so it's important that it not scribble on input. No result is
4175 : * returned, but any generated paths are added to grouped_rel.
4176 : */
4177 : static void
4178 796 : consider_groupingsets_paths(PlannerInfo *root,
4179 : RelOptInfo *grouped_rel,
4180 : Path *path,
4181 : bool is_sorted,
4182 : bool can_hash,
4183 : grouping_sets_data *gd,
4184 : const AggClauseCosts *agg_costs,
4185 : double dNumGroups)
4186 : {
4187 796 : Query *parse = root->parse;
4188 :
4189 : /*
4190 : * If we're not being offered sorted input, then only consider plans that
4191 : * can be done entirely by hashing.
4192 : *
4193 : * We can hash everything if it looks like it'll fit in work_mem. But if
4194 : * the input is actually sorted despite not being advertised as such, we
4195 : * prefer to make use of that in order to use less memory.
4196 : *
4197 : * If none of the grouping sets are sortable, then ignore the work_mem
4198 : * limit and generate a path anyway, since otherwise we'll just fail.
4199 : */
4200 796 : if (!is_sorted)
4201 : {
4202 360 : List *new_rollups = NIL;
4203 360 : RollupData *unhashed_rollup = NULL;
4204 : List *sets_data;
4205 360 : List *empty_sets_data = NIL;
4206 360 : List *empty_sets = NIL;
4207 : ListCell *lc;
4208 360 : ListCell *l_start = list_head(gd->rollups);
4209 360 : AggStrategy strat = AGG_HASHED;
4210 : double hashsize;
4211 360 : double exclude_groups = 0.0;
4212 :
4213 : Assert(can_hash);
4214 :
4215 : /*
4216 : * If the input is coincidentally sorted usefully (which can happen
4217 : * even if is_sorted is false, since that only means that our caller
4218 : * has set up the sorting for us), then save some hashtable space by
4219 : * making use of that. But we need to watch out for degenerate cases:
4220 : *
4221 : * 1) If there are any empty grouping sets, then group_pathkeys might
4222 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4223 : * there will be a rollup containing only empty groups, and the
4224 : * pathkeys_contained_in test is vacuously true; this is ok.
4225 : *
4226 : * XXX: the above relies on the fact that group_pathkeys is generated
4227 : * from the first rollup. If we add the ability to consider multiple
4228 : * sort orders for grouping input, this assumption might fail.
4229 : *
4230 : * 2) If there are no empty sets and only unsortable sets, then the
4231 : * rollups list will be empty (and thus l_start == NULL), and
4232 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4233 : * pathkeys_contain_in test doesn't cause us to crash.
4234 : */
4235 716 : if (l_start != NULL &&
4236 356 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4237 : {
4238 0 : unhashed_rollup = lfirst_node(RollupData, l_start);
4239 0 : exclude_groups = unhashed_rollup->numGroups;
4240 0 : l_start = lnext(l_start);
4241 : }
4242 :
4243 360 : hashsize = estimate_hashagg_tablesize(path,
4244 : agg_costs,
4245 : dNumGroups - exclude_groups);
4246 :
4247 : /*
4248 : * gd->rollups is empty if we have only unsortable columns to work
4249 : * with. Override work_mem in that case; otherwise, we'll rely on the
4250 : * sorted-input case to generate usable mixed paths.
4251 : */
4252 360 : if (hashsize > work_mem * 1024L && gd->rollups)
4253 12 : return; /* nope, won't fit */
4254 :
4255 : /*
4256 : * We need to burst the existing rollups list into individual grouping
4257 : * sets and recompute a groupClause for each set.
4258 : */
4259 348 : sets_data = list_copy(gd->unsortable_sets);
4260 :
4261 856 : for_each_cell(lc, l_start)
4262 : {
4263 524 : RollupData *rollup = lfirst_node(RollupData, lc);
4264 :
4265 : /*
4266 : * If we find an unhashable rollup that's not been skipped by the
4267 : * "actually sorted" check above, we can't cope; we'd need sorted
4268 : * input (with a different sort order) but we can't get that here.
4269 : * So bail out; we'll get a valid path from the is_sorted case
4270 : * instead.
4271 : *
4272 : * The mere presence of empty grouping sets doesn't make a rollup
4273 : * unhashable (see preprocess_grouping_sets), we handle those
4274 : * specially below.
4275 : */
4276 524 : if (!rollup->hashable)
4277 16 : return;
4278 : else
4279 508 : sets_data = list_concat(sets_data, list_copy(rollup->gsets_data));
4280 : }
4281 1360 : foreach(lc, sets_data)
4282 : {
4283 1028 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4284 1028 : List *gset = gs->set;
4285 : RollupData *rollup;
4286 :
4287 1028 : if (gset == NIL)
4288 : {
4289 : /* Empty grouping sets can't be hashed. */
4290 244 : empty_sets_data = lappend(empty_sets_data, gs);
4291 244 : empty_sets = lappend(empty_sets, NIL);
4292 : }
4293 : else
4294 : {
4295 784 : rollup = makeNode(RollupData);
4296 :
4297 784 : rollup->groupClause = preprocess_groupclause(root, gset);
4298 784 : rollup->gsets_data = list_make1(gs);
4299 784 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4300 : rollup->gsets_data,
4301 : gd->tleref_to_colnum_map);
4302 784 : rollup->numGroups = gs->numGroups;
4303 784 : rollup->hashable = true;
4304 784 : rollup->is_hashed = true;
4305 784 : new_rollups = lappend(new_rollups, rollup);
4306 : }
4307 : }
4308 :
4309 : /*
4310 : * If we didn't find anything nonempty to hash, then bail. We'll
4311 : * generate a path from the is_sorted case.
4312 : */
4313 332 : if (new_rollups == NIL)
4314 0 : return;
4315 :
4316 : /*
4317 : * If there were empty grouping sets they should have been in the
4318 : * first rollup.
4319 : */
4320 : Assert(!unhashed_rollup || !empty_sets);
4321 :
4322 332 : if (unhashed_rollup)
4323 : {
4324 0 : new_rollups = lappend(new_rollups, unhashed_rollup);
4325 0 : strat = AGG_MIXED;
4326 : }
4327 332 : else if (empty_sets)
4328 : {
4329 212 : RollupData *rollup = makeNode(RollupData);
4330 :
4331 212 : rollup->groupClause = NIL;
4332 212 : rollup->gsets_data = empty_sets_data;
4333 212 : rollup->gsets = empty_sets;
4334 212 : rollup->numGroups = list_length(empty_sets);
4335 212 : rollup->hashable = false;
4336 212 : rollup->is_hashed = false;
4337 212 : new_rollups = lappend(new_rollups, rollup);
4338 212 : strat = AGG_MIXED;
4339 : }
4340 :
4341 332 : add_path(grouped_rel, (Path *)
4342 332 : create_groupingsets_path(root,
4343 : grouped_rel,
4344 : path,
4345 332 : (List *) parse->havingQual,
4346 : strat,
4347 : new_rollups,
4348 : agg_costs,
4349 : dNumGroups));
4350 332 : return;
4351 : }
4352 :
4353 : /*
4354 : * If we have sorted input but nothing we can do with it, bail.
4355 : */
4356 436 : if (list_length(gd->rollups) == 0)
4357 0 : return;
4358 :
4359 : /*
4360 : * Given sorted input, we try and make two paths: one sorted and one mixed
4361 : * sort/hash. (We need to try both because hashagg might be disabled, or
4362 : * some columns might not be sortable.)
4363 : *
4364 : * can_hash is passed in as false if some obstacle elsewhere (such as
4365 : * ordered aggs) means that we shouldn't consider hashing at all.
4366 : */
4367 436 : if (can_hash && gd->any_hashable)
4368 : {
4369 384 : List *rollups = NIL;
4370 384 : List *hash_sets = list_copy(gd->unsortable_sets);
4371 384 : double availspace = (work_mem * 1024.0);
4372 : ListCell *lc;
4373 :
4374 : /*
4375 : * Account first for space needed for groups we can't sort at all.
4376 : */
4377 384 : availspace -= estimate_hashagg_tablesize(path,
4378 : agg_costs,
4379 : gd->dNumHashGroups);
4380 :
4381 384 : if (availspace > 0 && list_length(gd->rollups) > 1)
4382 : {
4383 : double scale;
4384 184 : int num_rollups = list_length(gd->rollups);
4385 : int k_capacity;
4386 184 : int *k_weights = palloc(num_rollups * sizeof(int));
4387 184 : Bitmapset *hash_items = NULL;
4388 : int i;
4389 :
4390 : /*
4391 : * We treat this as a knapsack problem: the knapsack capacity
4392 : * represents work_mem, the item weights are the estimated memory
4393 : * usage of the hashtables needed to implement a single rollup,
4394 : * and we really ought to use the cost saving as the item value;
4395 : * however, currently the costs assigned to sort nodes don't
4396 : * reflect the comparison costs well, and so we treat all items as
4397 : * of equal value (each rollup we hash instead saves us one sort).
4398 : *
4399 : * To use the discrete knapsack, we need to scale the values to a
4400 : * reasonably small bounded range. We choose to allow a 5% error
4401 : * margin; we have no more than 4096 rollups in the worst possible
4402 : * case, which with a 5% error margin will require a bit over 42MB
4403 : * of workspace. (Anyone wanting to plan queries that complex had
4404 : * better have the memory for it. In more reasonable cases, with
4405 : * no more than a couple of dozen rollups, the memory usage will
4406 : * be negligible.)
4407 : *
4408 : * k_capacity is naturally bounded, but we clamp the values for
4409 : * scale and weight (below) to avoid overflows or underflows (or
4410 : * uselessly trying to use a scale factor less than 1 byte).
4411 : */
4412 184 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4413 184 : k_capacity = (int) floor(availspace / scale);
4414 :
4415 : /*
4416 : * We leave the first rollup out of consideration since it's the
4417 : * one that matches the input sort order. We assign indexes "i"
4418 : * to only those entries considered for hashing; the second loop,
4419 : * below, must use the same condition.
4420 : */
4421 184 : i = 0;
4422 496 : for_each_cell(lc, lnext(list_head(gd->rollups)))
4423 : {
4424 312 : RollupData *rollup = lfirst_node(RollupData, lc);
4425 :
4426 312 : if (rollup->hashable)
4427 : {
4428 312 : double sz = estimate_hashagg_tablesize(path,
4429 : agg_costs,
4430 : rollup->numGroups);
4431 :
4432 : /*
4433 : * If sz is enormous, but work_mem (and hence scale) is
4434 : * small, avoid integer overflow here.
4435 : */
4436 312 : k_weights[i] = (int) Min(floor(sz / scale),
4437 : k_capacity + 1.0);
4438 312 : ++i;
4439 : }
4440 : }
4441 :
4442 : /*
4443 : * Apply knapsack algorithm; compute the set of items which
4444 : * maximizes the value stored (in this case the number of sorts
4445 : * saved) while keeping the total size (approximately) within
4446 : * capacity.
4447 : */
4448 184 : if (i > 0)
4449 184 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4450 :
4451 184 : if (!bms_is_empty(hash_items))
4452 : {
4453 184 : rollups = list_make1(linitial(gd->rollups));
4454 :
4455 184 : i = 0;
4456 496 : for_each_cell(lc, lnext(list_head(gd->rollups)))
4457 : {
4458 312 : RollupData *rollup = lfirst_node(RollupData, lc);
4459 :
4460 312 : if (rollup->hashable)
4461 : {
4462 312 : if (bms_is_member(i, hash_items))
4463 288 : hash_sets = list_concat(hash_sets,
4464 288 : list_copy(rollup->gsets_data));
4465 : else
4466 24 : rollups = lappend(rollups, rollup);
4467 312 : ++i;
4468 : }
4469 : else
4470 0 : rollups = lappend(rollups, rollup);
4471 : }
4472 : }
4473 : }
4474 :
4475 384 : if (!rollups && hash_sets)
4476 16 : rollups = list_copy(gd->rollups);
4477 :
4478 720 : foreach(lc, hash_sets)
4479 : {
4480 336 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4481 336 : RollupData *rollup = makeNode(RollupData);
4482 :
4483 : Assert(gs->set != NIL);
4484 :
4485 336 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4486 336 : rollup->gsets_data = list_make1(gs);
4487 336 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4488 : rollup->gsets_data,
4489 : gd->tleref_to_colnum_map);
4490 336 : rollup->numGroups = gs->numGroups;
4491 336 : rollup->hashable = true;
4492 336 : rollup->is_hashed = true;
4493 336 : rollups = lcons(rollup, rollups);
4494 : }
4495 :
4496 384 : if (rollups)
4497 : {
4498 200 : add_path(grouped_rel, (Path *)
4499 200 : create_groupingsets_path(root,
4500 : grouped_rel,
4501 : path,
4502 200 : (List *) parse->havingQual,
4503 : AGG_MIXED,
4504 : rollups,
4505 : agg_costs,
4506 : dNumGroups));
4507 : }
4508 : }
4509 :
4510 : /*
4511 : * Now try the simple sorted case.
4512 : */
4513 436 : if (!gd->unsortable_sets)
4514 416 : add_path(grouped_rel, (Path *)
4515 832 : create_groupingsets_path(root,
4516 : grouped_rel,
4517 : path,
4518 416 : (List *) parse->havingQual,
4519 : AGG_SORTED,
4520 : gd->rollups,
4521 : agg_costs,
4522 : dNumGroups));
4523 : }
4524 :
4525 : /*
4526 : * create_window_paths
4527 : *
4528 : * Build a new upperrel containing Paths for window-function evaluation.
4529 : *
4530 : * input_rel: contains the source-data Paths
4531 : * input_target: result of make_window_input_target
4532 : * output_target: what the topmost WindowAggPath should return
4533 : * wflists: result of find_window_functions
4534 : * activeWindows: result of select_active_windows
4535 : *
4536 : * Note: all Paths in input_rel are expected to return input_target.
4537 : */
4538 : static RelOptInfo *
4539 1076 : create_window_paths(PlannerInfo *root,
4540 : RelOptInfo *input_rel,
4541 : PathTarget *input_target,
4542 : PathTarget *output_target,
4543 : bool output_target_parallel_safe,
4544 : WindowFuncLists *wflists,
4545 : List *activeWindows)
4546 : {
4547 : RelOptInfo *window_rel;
4548 : ListCell *lc;
4549 :
4550 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4551 1076 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4552 :
4553 : /*
4554 : * If the input relation is not parallel-safe, then the window relation
4555 : * can't be parallel-safe, either. Otherwise, we need to examine the
4556 : * target list and active windows for non-parallel-safe constructs.
4557 : */
4558 1076 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4559 0 : is_parallel_safe(root, (Node *) activeWindows))
4560 0 : window_rel->consider_parallel = true;
4561 :
4562 : /*
4563 : * If the input rel belongs to a single FDW, so does the window rel.
4564 : */
4565 1076 : window_rel->serverid = input_rel->serverid;
4566 1076 : window_rel->userid = input_rel->userid;
4567 1076 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4568 1076 : window_rel->fdwroutine = input_rel->fdwroutine;
4569 :
4570 : /*
4571 : * Consider computing window functions starting from the existing
4572 : * cheapest-total path (which will likely require a sort) as well as any
4573 : * existing paths that satisfy root->window_pathkeys (which won't).
4574 : */
4575 2244 : foreach(lc, input_rel->pathlist)
4576 : {
4577 1168 : Path *path = (Path *) lfirst(lc);
4578 :
4579 1260 : if (path == input_rel->cheapest_total_path ||
4580 92 : pathkeys_contained_in(root->window_pathkeys, path->pathkeys))
4581 1112 : create_one_window_path(root,
4582 : window_rel,
4583 : path,
4584 : input_target,
4585 : output_target,
4586 : wflists,
4587 : activeWindows);
4588 : }
4589 :
4590 : /*
4591 : * If there is an FDW that's responsible for all baserels of the query,
4592 : * let it consider adding ForeignPaths.
4593 : */
4594 1088 : if (window_rel->fdwroutine &&
4595 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4596 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4597 : input_rel, window_rel,
4598 : NULL);
4599 :
4600 : /* Let extensions possibly add some more paths */
4601 1076 : if (create_upper_paths_hook)
4602 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4603 : input_rel, window_rel, NULL);
4604 :
4605 : /* Now choose the best path(s) */
4606 1076 : set_cheapest(window_rel);
4607 :
4608 1076 : return window_rel;
4609 : }
4610 :
4611 : /*
4612 : * Stack window-function implementation steps atop the given Path, and
4613 : * add the result to window_rel.
4614 : *
4615 : * window_rel: upperrel to contain result
4616 : * path: input Path to use (must return input_target)
4617 : * input_target: result of make_window_input_target
4618 : * output_target: what the topmost WindowAggPath should return
4619 : * wflists: result of find_window_functions
4620 : * activeWindows: result of select_active_windows
4621 : */
4622 : static void
4623 1112 : create_one_window_path(PlannerInfo *root,
4624 : RelOptInfo *window_rel,
4625 : Path *path,
4626 : PathTarget *input_target,
4627 : PathTarget *output_target,
4628 : WindowFuncLists *wflists,
4629 : List *activeWindows)
4630 : {
4631 : PathTarget *window_target;
4632 : ListCell *l;
4633 :
4634 : /*
4635 : * Since each window clause could require a different sort order, we stack
4636 : * up a WindowAgg node for each clause, with sort steps between them as
4637 : * needed. (We assume that select_active_windows chose a good order for
4638 : * executing the clauses in.)
4639 : *
4640 : * input_target should contain all Vars and Aggs needed for the result.
4641 : * (In some cases we wouldn't need to propagate all of these all the way
4642 : * to the top, since they might only be needed as inputs to WindowFuncs.
4643 : * It's probably not worth trying to optimize that though.) It must also
4644 : * contain all window partitioning and sorting expressions, to ensure
4645 : * they're computed only once at the bottom of the stack (that's critical
4646 : * for volatile functions). As we climb up the stack, we'll add outputs
4647 : * for the WindowFuncs computed at each level.
4648 : */
4649 1112 : window_target = input_target;
4650 :
4651 2272 : foreach(l, activeWindows)
4652 : {
4653 1160 : WindowClause *wc = lfirst_node(WindowClause, l);
4654 : List *window_pathkeys;
4655 :
4656 1160 : window_pathkeys = make_pathkeys_for_window(root,
4657 : wc,
4658 : root->processed_tlist);
4659 :
4660 : /* Sort if necessary */
4661 1160 : if (!pathkeys_contained_in(window_pathkeys, path->pathkeys))
4662 : {
4663 932 : path = (Path *) create_sort_path(root, window_rel,
4664 : path,
4665 : window_pathkeys,
4666 : -1.0);
4667 : }
4668 :
4669 1160 : if (lnext(l))
4670 : {
4671 : /*
4672 : * Add the current WindowFuncs to the output target for this
4673 : * intermediate WindowAggPath. We must copy window_target to
4674 : * avoid changing the previous path's target.
4675 : *
4676 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4677 : * we do need to account for the increase in tlist width.
4678 : */
4679 : ListCell *lc2;
4680 :
4681 48 : window_target = copy_pathtarget(window_target);
4682 96 : foreach(lc2, wflists->windowFuncs[wc->winref])
4683 : {
4684 48 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4685 :
4686 48 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4687 48 : window_target->width += get_typavgwidth(wfunc->wintype, -1);
4688 : }
4689 : }
4690 : else
4691 : {
4692 : /* Install the goal target in the topmost WindowAgg */
4693 1112 : window_target = output_target;
4694 : }
4695 :
4696 1160 : path = (Path *)
4697 1160 : create_windowagg_path(root, window_rel, path, window_target,
4698 1160 : wflists->windowFuncs[wc->winref],
4699 : wc);
4700 : }
4701 :
4702 1112 : add_path(window_rel, path);
4703 1112 : }
4704 :
4705 : /*
4706 : * create_distinct_paths
4707 : *
4708 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4709 : *
4710 : * input_rel: contains the source-data Paths
4711 : *
4712 : * Note: input paths should already compute the desired pathtarget, since
4713 : * Sort/Unique won't project anything.
4714 : */
4715 : static RelOptInfo *
4716 336 : create_distinct_paths(PlannerInfo *root,
4717 : RelOptInfo *input_rel)
4718 : {
4719 336 : Query *parse = root->parse;
4720 336 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4721 : RelOptInfo *distinct_rel;
4722 : double numDistinctRows;
4723 : bool allow_hash;
4724 : Path *path;
4725 : ListCell *lc;
4726 :
4727 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4728 336 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4729 :
4730 : /*
4731 : * We don't compute anything at this level, so distinct_rel will be
4732 : * parallel-safe if the input rel is parallel-safe. In particular, if
4733 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4734 : * output those expressions, and will not be parallel-safe unless those
4735 : * expressions are parallel-safe.
4736 : */
4737 336 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4738 :
4739 : /*
4740 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4741 : */
4742 336 : distinct_rel->serverid = input_rel->serverid;
4743 336 : distinct_rel->userid = input_rel->userid;
4744 336 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4745 336 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4746 :
4747 : /* Estimate number of distinct rows there will be */
4748 648 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4749 312 : root->hasHavingQual)
4750 : {
4751 : /*
4752 : * If there was grouping or aggregation, use the number of input rows
4753 : * as the estimated number of DISTINCT rows (ie, assume the input is
4754 : * already mostly unique).
4755 : */
4756 24 : numDistinctRows = cheapest_input_path->rows;
4757 : }
4758 : else
4759 : {
4760 : /*
4761 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4762 : */
4763 : List *distinctExprs;
4764 :
4765 312 : distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
4766 : parse->targetList);
4767 312 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4768 : cheapest_input_path->rows,
4769 : NULL);
4770 : }
4771 :
4772 : /*
4773 : * Consider sort-based implementations of DISTINCT, if possible.
4774 : */
4775 336 : if (grouping_is_sortable(parse->distinctClause))
4776 : {
4777 : /*
4778 : * First, if we have any adequately-presorted paths, just stick a
4779 : * Unique node on those. Then consider doing an explicit sort of the
4780 : * cheapest input path and Unique'ing that.
4781 : *
4782 : * When we have DISTINCT ON, we must sort by the more rigorous of
4783 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
4784 : * Also, if we do have to do an explicit sort, we might as well use
4785 : * the more rigorous ordering to avoid a second sort later. (Note
4786 : * that the parser will have ensured that one clause is a prefix of
4787 : * the other.)
4788 : */
4789 : List *needed_pathkeys;
4790 :
4791 368 : if (parse->hasDistinctOn &&
4792 36 : list_length(root->distinct_pathkeys) <
4793 36 : list_length(root->sort_pathkeys))
4794 20 : needed_pathkeys = root->sort_pathkeys;
4795 : else
4796 312 : needed_pathkeys = root->distinct_pathkeys;
4797 :
4798 688 : foreach(lc, input_rel->pathlist)
4799 : {
4800 356 : Path *path = (Path *) lfirst(lc);
4801 :
4802 356 : if (pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4803 : {
4804 14 : add_path(distinct_rel, (Path *)
4805 14 : create_upper_unique_path(root, distinct_rel,
4806 : path,
4807 14 : list_length(root->distinct_pathkeys),
4808 : numDistinctRows));
4809 : }
4810 : }
4811 :
4812 : /* For explicit-sort case, always use the more rigorous clause */
4813 664 : if (list_length(root->distinct_pathkeys) <
4814 332 : list_length(root->sort_pathkeys))
4815 : {
4816 20 : needed_pathkeys = root->sort_pathkeys;
4817 : /* Assert checks that parser didn't mess up... */
4818 : Assert(pathkeys_contained_in(root->distinct_pathkeys,
4819 : needed_pathkeys));
4820 : }
4821 : else
4822 312 : needed_pathkeys = root->distinct_pathkeys;
4823 :
4824 332 : path = cheapest_input_path;
4825 332 : if (!pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4826 322 : path = (Path *) create_sort_path(root, distinct_rel,
4827 : path,
4828 : needed_pathkeys,
4829 : -1.0);
4830 :
4831 332 : add_path(distinct_rel, (Path *)
4832 332 : create_upper_unique_path(root, distinct_rel,
4833 : path,
4834 332 : list_length(root->distinct_pathkeys),
4835 : numDistinctRows));
4836 : }
4837 :
4838 : /*
4839 : * Consider hash-based implementations of DISTINCT, if possible.
4840 : *
4841 : * If we were not able to make any other types of path, we *must* hash or
4842 : * die trying. If we do have other choices, there are several things that
4843 : * should prevent selection of hashing: if the query uses DISTINCT ON
4844 : * (because it won't really have the expected behavior if we hash), or if
4845 : * enable_hashagg is off, or if it looks like the hashtable will exceed
4846 : * work_mem.
4847 : *
4848 : * Note: grouping_is_hashable() is much more expensive to check than the
4849 : * other gating conditions, so we want to do it last.
4850 : */
4851 336 : if (distinct_rel->pathlist == NIL)
4852 4 : allow_hash = true; /* we have no alternatives */
4853 332 : else if (parse->hasDistinctOn || !enable_hashagg)
4854 36 : allow_hash = false; /* policy-based decision not to hash */
4855 : else
4856 : {
4857 : Size hashentrysize;
4858 :
4859 : /* Estimate per-hash-entry space at tuple width... */
4860 296 : hashentrysize = MAXALIGN(cheapest_input_path->pathtarget->width) +
4861 : MAXALIGN(SizeofMinimalTupleHeader);
4862 : /* plus the per-hash-entry overhead */
4863 296 : hashentrysize += hash_agg_entry_size(0);
4864 :
4865 : /* Allow hashing only if hashtable is predicted to fit in work_mem */
4866 296 : allow_hash = (hashentrysize * numDistinctRows <= work_mem * 1024L);
4867 : }
4868 :
4869 336 : if (allow_hash && grouping_is_hashable(parse->distinctClause))
4870 : {
4871 : /* Generate hashed aggregate path --- no sort needed */
4872 300 : add_path(distinct_rel, (Path *)
4873 300 : create_agg_path(root,
4874 : distinct_rel,
4875 : cheapest_input_path,
4876 : cheapest_input_path->pathtarget,
4877 : AGG_HASHED,
4878 : AGGSPLIT_SIMPLE,
4879 : parse->distinctClause,
4880 : NIL,
4881 : NULL,
4882 : numDistinctRows));
4883 : }
4884 :
4885 : /* Give a helpful error if we failed to find any implementation */
4886 336 : if (distinct_rel->pathlist == NIL)
4887 0 : ereport(ERROR,
4888 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4889 : errmsg("could not implement DISTINCT"),
4890 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4891 :
4892 : /*
4893 : * If there is an FDW that's responsible for all baserels of the query,
4894 : * let it consider adding ForeignPaths.
4895 : */
4896 352 : if (distinct_rel->fdwroutine &&
4897 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4898 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_DISTINCT,
4899 : input_rel, distinct_rel,
4900 : NULL);
4901 :
4902 : /* Let extensions possibly add some more paths */
4903 336 : if (create_upper_paths_hook)
4904 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT,
4905 : input_rel, distinct_rel, NULL);
4906 :
4907 : /* Now choose the best path(s) */
4908 336 : set_cheapest(distinct_rel);
4909 :
4910 336 : return distinct_rel;
4911 : }
4912 :
4913 : /*
4914 : * create_ordered_paths
4915 : *
4916 : * Build a new upperrel containing Paths for ORDER BY evaluation.
4917 : *
4918 : * All paths in the result must satisfy the ORDER BY ordering.
4919 : * The only new path we need consider is an explicit sort on the
4920 : * cheapest-total existing path.
4921 : *
4922 : * input_rel: contains the source-data Paths
4923 : * target: the output tlist the result Paths must emit
4924 : * limit_tuples: estimated bound on the number of output tuples,
4925 : * or -1 if no LIMIT or couldn't estimate
4926 : */
4927 : static RelOptInfo *
4928 34078 : create_ordered_paths(PlannerInfo *root,
4929 : RelOptInfo *input_rel,
4930 : PathTarget *target,
4931 : bool target_parallel_safe,
4932 : double limit_tuples)
4933 : {
4934 34078 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4935 : RelOptInfo *ordered_rel;
4936 : ListCell *lc;
4937 :
4938 : /* For now, do all work in the (ORDERED, NULL) upperrel */
4939 34078 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
4940 :
4941 : /*
4942 : * If the input relation is not parallel-safe, then the ordered relation
4943 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4944 : * target list is parallel-safe.
4945 : */
4946 34078 : if (input_rel->consider_parallel && target_parallel_safe)
4947 13994 : ordered_rel->consider_parallel = true;
4948 :
4949 : /*
4950 : * If the input rel belongs to a single FDW, so does the ordered_rel.
4951 : */
4952 34078 : ordered_rel->serverid = input_rel->serverid;
4953 34078 : ordered_rel->userid = input_rel->userid;
4954 34078 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
4955 34078 : ordered_rel->fdwroutine = input_rel->fdwroutine;
4956 :
4957 81468 : foreach(lc, input_rel->pathlist)
4958 : {
4959 47390 : Path *path = (Path *) lfirst(lc);
4960 : bool is_sorted;
4961 :
4962 47390 : is_sorted = pathkeys_contained_in(root->sort_pathkeys,
4963 : path->pathkeys);
4964 47390 : if (path == cheapest_input_path || is_sorted)
4965 : {
4966 46126 : if (!is_sorted)
4967 : {
4968 : /* An explicit sort here can take advantage of LIMIT */
4969 27142 : path = (Path *) create_sort_path(root,
4970 : ordered_rel,
4971 : path,
4972 : root->sort_pathkeys,
4973 : limit_tuples);
4974 : }
4975 :
4976 : /* Add projection step if needed */
4977 46126 : if (path->pathtarget != target)
4978 8796 : path = apply_projection_to_path(root, ordered_rel,
4979 : path, target);
4980 :
4981 46126 : add_path(ordered_rel, path);
4982 : }
4983 : }
4984 :
4985 : /*
4986 : * generate_gather_paths() will have already generated a simple Gather
4987 : * path for the best parallel path, if any, and the loop above will have
4988 : * considered sorting it. Similarly, generate_gather_paths() will also
4989 : * have generated order-preserving Gather Merge plans which can be used
4990 : * without sorting if they happen to match the sort_pathkeys, and the loop
4991 : * above will have handled those as well. However, there's one more
4992 : * possibility: it may make sense to sort the cheapest partial path
4993 : * according to the required output order and then use Gather Merge.
4994 : */
4995 47964 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
4996 13886 : input_rel->partial_pathlist != NIL)
4997 : {
4998 : Path *cheapest_partial_path;
4999 :
5000 1766 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5001 :
5002 : /*
5003 : * If cheapest partial path doesn't need a sort, this is redundant
5004 : * with what's already been tried.
5005 : */
5006 1766 : if (!pathkeys_contained_in(root->sort_pathkeys,
5007 : cheapest_partial_path->pathkeys))
5008 : {
5009 : Path *path;
5010 : double total_groups;
5011 :
5012 1758 : path = (Path *) create_sort_path(root,
5013 : ordered_rel,
5014 : cheapest_partial_path,
5015 : root->sort_pathkeys,
5016 : limit_tuples);
5017 :
5018 3516 : total_groups = cheapest_partial_path->rows *
5019 1758 : cheapest_partial_path->parallel_workers;
5020 1758 : path = (Path *)
5021 1758 : create_gather_merge_path(root, ordered_rel,
5022 : path,
5023 : path->pathtarget,
5024 : root->sort_pathkeys, NULL,
5025 : &total_groups);
5026 :
5027 : /* Add projection step if needed */
5028 1758 : if (path->pathtarget != target)
5029 1166 : path = apply_projection_to_path(root, ordered_rel,
5030 : path, target);
5031 :
5032 1758 : add_path(ordered_rel, path);
5033 : }
5034 : }
5035 :
5036 : /*
5037 : * If there is an FDW that's responsible for all baserels of the query,
5038 : * let it consider adding ForeignPaths.
5039 : */
5040 34398 : if (ordered_rel->fdwroutine &&
5041 320 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5042 306 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5043 : input_rel, ordered_rel,
5044 : NULL);
5045 :
5046 : /* Let extensions possibly add some more paths */
5047 34078 : if (create_upper_paths_hook)
5048 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5049 : input_rel, ordered_rel, NULL);
5050 :
5051 : /*
5052 : * No need to bother with set_cheapest here; grouping_planner does not
5053 : * need us to do it.
5054 : */
5055 : Assert(ordered_rel->pathlist != NIL);
5056 :
5057 34078 : return ordered_rel;
5058 : }
5059 :
5060 :
5061 : /*
5062 : * make_group_input_target
5063 : * Generate appropriate PathTarget for initial input to grouping nodes.
5064 : *
5065 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5066 : * the query's final targetlist; for example, it certainly can't emit any
5067 : * aggregate function calls. This routine generates the correct target
5068 : * for the scan/join subplan.
5069 : *
5070 : * The query target list passed from the parser already contains entries
5071 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5072 : * for variables used only in HAVING clauses; so we need to add those
5073 : * variables to the subplan target list. Also, we flatten all expressions
5074 : * except GROUP BY items into their component variables; other expressions
5075 : * will be computed by the upper plan nodes rather than by the subplan.
5076 : * For example, given a query like
5077 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5078 : * we want to pass this targetlist to the subplan:
5079 : * a+b,c,d
5080 : * where the a+b target will be used by the Sort/Group steps, and the
5081 : * other targets will be used for computing the final results.
5082 : *
5083 : * 'final_target' is the query's final target list (in PathTarget form)
5084 : *
5085 : * The result is the PathTarget to be computed by the Paths returned from
5086 : * query_planner().
5087 : */
5088 : static PathTarget *
5089 23600 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5090 : {
5091 23600 : Query *parse = root->parse;
5092 : PathTarget *input_target;
5093 : List *non_group_cols;
5094 : List *non_group_vars;
5095 : int i;
5096 : ListCell *lc;
5097 :
5098 : /*
5099 : * We must build a target containing all grouping columns, plus any other
5100 : * Vars mentioned in the query's targetlist and HAVING qual.
5101 : */
5102 23600 : input_target = create_empty_pathtarget();
5103 23600 : non_group_cols = NIL;
5104 :
5105 23600 : i = 0;
5106 52728 : foreach(lc, final_target->exprs)
5107 : {
5108 29128 : Expr *expr = (Expr *) lfirst(lc);
5109 29128 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5110 :
5111 33000 : if (sgref && parse->groupClause &&
5112 3872 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5113 : {
5114 : /*
5115 : * It's a grouping column, so add it to the input target as-is.
5116 : */
5117 3096 : add_column_to_pathtarget(input_target, expr, sgref);
5118 : }
5119 : else
5120 : {
5121 : /*
5122 : * Non-grouping column, so just remember the expression for later
5123 : * call to pull_var_clause.
5124 : */
5125 26032 : non_group_cols = lappend(non_group_cols, expr);
5126 : }
5127 :
5128 29128 : i++;
5129 : }
5130 :
5131 : /*
5132 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5133 : */
5134 23600 : if (parse->havingQual)
5135 308 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5136 :
5137 : /*
5138 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5139 : * add them to the input target if not already present. (A Var used
5140 : * directly as a GROUP BY item will be present already.) Note this
5141 : * includes Vars used in resjunk items, so we are covering the needs of
5142 : * ORDER BY and window specifications. Vars used within Aggrefs and
5143 : * WindowFuncs will be pulled out here, too.
5144 : */
5145 23600 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5146 : PVC_RECURSE_AGGREGATES |
5147 : PVC_RECURSE_WINDOWFUNCS |
5148 : PVC_INCLUDE_PLACEHOLDERS);
5149 23600 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5150 :
5151 : /* clean up cruft */
5152 23600 : list_free(non_group_vars);
5153 23600 : list_free(non_group_cols);
5154 :
5155 : /* XXX this causes some redundant cost calculation ... */
5156 23600 : return set_pathtarget_cost_width(root, input_target);
5157 : }
5158 :
5159 : /*
5160 : * make_partial_grouping_target
5161 : * Generate appropriate PathTarget for output of partial aggregate
5162 : * (or partial grouping, if there are no aggregates) nodes.
5163 : *
5164 : * A partial aggregation node needs to emit all the same aggregates that
5165 : * a regular aggregation node would, plus any aggregates used in HAVING;
5166 : * except that the Aggref nodes should be marked as partial aggregates.
5167 : *
5168 : * In addition, we'd better emit any Vars and PlaceholderVars that are
5169 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5170 : * these would be Vars that are grouped by or used in grouping expressions.)
5171 : *
5172 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5173 : * havingQual represents the HAVING clause.
5174 : */
5175 : static PathTarget *
5176 1290 : make_partial_grouping_target(PlannerInfo *root,
5177 : PathTarget *grouping_target,
5178 : Node *havingQual)
5179 : {
5180 1290 : Query *parse = root->parse;
5181 : PathTarget *partial_target;
5182 : List *non_group_cols;
5183 : List *non_group_exprs;
5184 : int i;
5185 : ListCell *lc;
5186 :
5187 1290 : partial_target = create_empty_pathtarget();
5188 1290 : non_group_cols = NIL;
5189 :
5190 1290 : i = 0;
5191 4660 : foreach(lc, grouping_target->exprs)
5192 : {
5193 3370 : Expr *expr = (Expr *) lfirst(lc);
5194 3370 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5195 :
5196 5722 : if (sgref && parse->groupClause &&
5197 2352 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5198 : {
5199 : /*
5200 : * It's a grouping column, so add it to the partial_target as-is.
5201 : * (This allows the upper agg step to repeat the grouping calcs.)
5202 : */
5203 1128 : add_column_to_pathtarget(partial_target, expr, sgref);
5204 : }
5205 : else
5206 : {
5207 : /*
5208 : * Non-grouping column, so just remember the expression for later
5209 : * call to pull_var_clause.
5210 : */
5211 2242 : non_group_cols = lappend(non_group_cols, expr);
5212 : }
5213 :
5214 3370 : i++;
5215 : }
5216 :
5217 : /*
5218 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5219 : */
5220 1290 : if (havingQual)
5221 552 : non_group_cols = lappend(non_group_cols, havingQual);
5222 :
5223 : /*
5224 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5225 : * non-group cols (plus HAVING), and add them to the partial_target if not
5226 : * already present. (An expression used directly as a GROUP BY item will
5227 : * be present already.) Note this includes Vars used in resjunk items, so
5228 : * we are covering the needs of ORDER BY and window specifications.
5229 : */
5230 1290 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5231 : PVC_INCLUDE_AGGREGATES |
5232 : PVC_RECURSE_WINDOWFUNCS |
5233 : PVC_INCLUDE_PLACEHOLDERS);
5234 :
5235 1290 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5236 :
5237 : /*
5238 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5239 : * are at the top level of the target list, so we can just scan the list
5240 : * rather than recursing through the expression trees.
5241 : */
5242 5068 : foreach(lc, partial_target->exprs)
5243 : {
5244 3778 : Aggref *aggref = (Aggref *) lfirst(lc);
5245 :
5246 3778 : if (IsA(aggref, Aggref))
5247 : {
5248 : Aggref *newaggref;
5249 :
5250 : /*
5251 : * We shouldn't need to copy the substructure of the Aggref node,
5252 : * but flat-copy the node itself to avoid damaging other trees.
5253 : */
5254 2650 : newaggref = makeNode(Aggref);
5255 2650 : memcpy(newaggref, aggref, sizeof(Aggref));
5256 :
5257 : /* For now, assume serialization is required */
5258 2650 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5259 :
5260 2650 : lfirst(lc) = newaggref;
5261 : }
5262 : }
5263 :
5264 : /* clean up cruft */
5265 1290 : list_free(non_group_exprs);
5266 1290 : list_free(non_group_cols);
5267 :
5268 : /* XXX this causes some redundant cost calculation ... */
5269 1290 : return set_pathtarget_cost_width(root, partial_target);
5270 : }
5271 :
5272 : /*
5273 : * mark_partial_aggref
5274 : * Adjust an Aggref to make it represent a partial-aggregation step.
5275 : *
5276 : * The Aggref node is modified in-place; caller must do any copying required.
5277 : */
5278 : void
5279 4322 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5280 : {
5281 : /* aggtranstype should be computed by this point */
5282 : Assert(OidIsValid(agg->aggtranstype));
5283 : /* ... but aggsplit should still be as the parser left it */
5284 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5285 :
5286 : /* Mark the Aggref with the intended partial-aggregation mode */
5287 4322 : agg->aggsplit = aggsplit;
5288 :
5289 : /*
5290 : * Adjust result type if needed. Normally, a partial aggregate returns
5291 : * the aggregate's transition type; but if that's INTERNAL and we're
5292 : * serializing, it returns BYTEA instead.
5293 : */
5294 4322 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5295 : {
5296 3486 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5297 48 : agg->aggtype = BYTEAOID;
5298 : else
5299 3438 : agg->aggtype = agg->aggtranstype;
5300 : }
5301 4322 : }
5302 :
5303 : /*
5304 : * postprocess_setop_tlist
5305 : * Fix up targetlist returned by plan_set_operations().
5306 : *
5307 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5308 : * NOTE: this would not be good enough if we supported resjunk sort keys
5309 : * for results of set operations --- then, we'd need to project a whole
5310 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5311 : * find any resjunk columns in orig_tlist.
5312 : */
5313 : static List *
5314 1278 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5315 : {
5316 : ListCell *l;
5317 1278 : ListCell *orig_tlist_item = list_head(orig_tlist);
5318 :
5319 3522 : foreach(l, new_tlist)
5320 : {
5321 2244 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5322 : TargetEntry *orig_tle;
5323 :
5324 : /* ignore resjunk columns in setop result */
5325 2244 : if (new_tle->resjunk)
5326 388 : continue;
5327 :
5328 : Assert(orig_tlist_item != NULL);
5329 1856 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5330 1856 : orig_tlist_item = lnext(orig_tlist_item);
5331 1856 : if (orig_tle->resjunk) /* should not happen */
5332 0 : elog(ERROR, "resjunk output columns are not implemented");
5333 : Assert(new_tle->resno == orig_tle->resno);
5334 1856 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5335 : }
5336 1278 : if (orig_tlist_item != NULL)
5337 0 : elog(ERROR, "resjunk output columns are not implemented");
5338 1278 : return new_tlist;
5339 : }
5340 :
5341 : /*
5342 : * select_active_windows
5343 : * Create a list of the "active" window clauses (ie, those referenced
5344 : * by non-deleted WindowFuncs) in the order they are to be executed.
5345 : */
5346 : static List *
5347 1076 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5348 : {
5349 1076 : List *windowClause = root->parse->windowClause;
5350 1076 : List *result = NIL;
5351 : ListCell *lc;
5352 1076 : int nActive = 0;
5353 1076 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5354 1076 : * list_length(windowClause));
5355 :
5356 : /* First, construct an array of the active windows */
5357 2216 : foreach(lc, windowClause)
5358 : {
5359 1140 : WindowClause *wc = lfirst_node(WindowClause, lc);
5360 :
5361 : /* It's only active if wflists shows some related WindowFuncs */
5362 : Assert(wc->winref <= wflists->maxWinRef);
5363 1140 : if (wflists->windowFuncs[wc->winref] == NIL)
5364 16 : continue;
5365 :
5366 1124 : actives[nActive].wc = wc; /* original clause */
5367 :
5368 : /*
5369 : * For sorting, we want the list of partition keys followed by the
5370 : * list of sort keys. But pathkeys construction will remove duplicates
5371 : * between the two, so we can as well (even though we can't detect all
5372 : * of the duplicates, since some may come from ECs - that might mean
5373 : * we miss optimization chances here). We must, however, ensure that
5374 : * the order of entries is preserved with respect to the ones we do
5375 : * keep.
5376 : *
5377 : * partitionClause and orderClause had their own duplicates removed in
5378 : * parse analysis, so we're only concerned here with removing
5379 : * orderClause entries that also appear in partitionClause.
5380 : */
5381 2248 : actives[nActive].uniqueOrder =
5382 1124 : list_concat_unique(list_copy(wc->partitionClause),
5383 : wc->orderClause);
5384 1124 : nActive++;
5385 : }
5386 :
5387 : /*
5388 : * Sort active windows by their partitioning/ordering clauses, ignoring
5389 : * any framing clauses, so that the windows that need the same sorting are
5390 : * adjacent in the list. When we come to generate paths, this will avoid
5391 : * inserting additional Sort nodes.
5392 : *
5393 : * This is how we implement a specific requirement from the SQL standard,
5394 : * which says that when two or more windows are order-equivalent (i.e.
5395 : * have matching partition and order clauses, even if their names or
5396 : * framing clauses differ), then all peer rows must be presented in the
5397 : * same order in all of them. If we allowed multiple sort nodes for such
5398 : * cases, we'd risk having the peer rows end up in different orders in
5399 : * equivalent windows due to sort instability. (See General Rule 4 of
5400 : * <window clause> in SQL2008 - SQL2016.)
5401 : *
5402 : * Additionally, if the entire list of clauses of one window is a prefix
5403 : * of another, put first the window with stronger sorting requirements.
5404 : * This way we will first sort for stronger window, and won't have to sort
5405 : * again for the weaker one.
5406 : */
5407 1076 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5408 :
5409 : /* build ordered list of the original WindowClause nodes */
5410 2200 : for (int i = 0; i < nActive; i++)
5411 1124 : result = lappend(result, actives[i].wc);
5412 :
5413 1076 : pfree(actives);
5414 :
5415 1076 : return result;
5416 : }
5417 :
5418 : /*
5419 : * common_prefix_cmp
5420 : * QSort comparison function for WindowClauseSortData
5421 : *
5422 : * Sort the windows by the required sorting clauses. First, compare the sort
5423 : * clauses themselves. Second, if one window's clauses are a prefix of another
5424 : * one's clauses, put the window with more sort clauses first.
5425 : */
5426 : static int
5427 48 : common_prefix_cmp(const void *a, const void *b)
5428 : {
5429 48 : const WindowClauseSortData *wcsa = a;
5430 48 : const WindowClauseSortData *wcsb = b;
5431 : ListCell *item_a;
5432 : ListCell *item_b;
5433 :
5434 84 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
5435 : {
5436 56 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
5437 56 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
5438 :
5439 56 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
5440 0 : return -1;
5441 56 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
5442 12 : return 1;
5443 44 : else if (sca->sortop > scb->sortop)
5444 0 : return -1;
5445 44 : else if (sca->sortop < scb->sortop)
5446 8 : return 1;
5447 36 : else if (sca->nulls_first && !scb->nulls_first)
5448 0 : return -1;
5449 36 : else if (!sca->nulls_first && scb->nulls_first)
5450 0 : return 1;
5451 : /* no need to compare eqop, since it is fully determined by sortop */
5452 : }
5453 :
5454 28 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
5455 4 : return -1;
5456 24 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
5457 8 : return 1;
5458 :
5459 16 : return 0;
5460 : }
5461 :
5462 : /*
5463 : * make_window_input_target
5464 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
5465 : *
5466 : * When the query has window functions, this function computes the desired
5467 : * target to be computed by the node just below the first WindowAgg.
5468 : * This tlist must contain all values needed to evaluate the window functions,
5469 : * compute the final target list, and perform any required final sort step.
5470 : * If multiple WindowAggs are needed, each intermediate one adds its window
5471 : * function results onto this base tlist; only the topmost WindowAgg computes
5472 : * the actual desired target list.
5473 : *
5474 : * This function is much like make_group_input_target, though not quite enough
5475 : * like it to share code. As in that function, we flatten most expressions
5476 : * into their component variables. But we do not want to flatten window
5477 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
5478 : * evaluations of them, which would be bad (possibly even resulting in
5479 : * inconsistent answers, if they contain volatile functions).
5480 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
5481 : * make_group_input_target, because we may no longer have access to the
5482 : * individual Vars in them.
5483 : *
5484 : * Another key difference from make_group_input_target is that we don't
5485 : * flatten Aggref expressions, since those are to be computed below the
5486 : * window functions and just referenced like Vars above that.
5487 : *
5488 : * 'final_target' is the query's final target list (in PathTarget form)
5489 : * 'activeWindows' is the list of active windows previously identified by
5490 : * select_active_windows.
5491 : *
5492 : * The result is the PathTarget to be computed by the plan node immediately
5493 : * below the first WindowAgg node.
5494 : */
5495 : static PathTarget *
5496 1076 : make_window_input_target(PlannerInfo *root,
5497 : PathTarget *final_target,
5498 : List *activeWindows)
5499 : {
5500 1076 : Query *parse = root->parse;
5501 : PathTarget *input_target;
5502 : Bitmapset *sgrefs;
5503 : List *flattenable_cols;
5504 : List *flattenable_vars;
5505 : int i;
5506 : ListCell *lc;
5507 :
5508 : Assert(parse->hasWindowFuncs);
5509 :
5510 : /*
5511 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
5512 : * into a bitmapset for convenient reference below.
5513 : */
5514 1076 : sgrefs = NULL;
5515 2200 : foreach(lc, activeWindows)
5516 : {
5517 1124 : WindowClause *wc = lfirst_node(WindowClause, lc);
5518 : ListCell *lc2;
5519 :
5520 1476 : foreach(lc2, wc->partitionClause)
5521 : {
5522 352 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5523 :
5524 352 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5525 : }
5526 2144 : foreach(lc2, wc->orderClause)
5527 : {
5528 1020 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5529 :
5530 1020 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5531 : }
5532 : }
5533 :
5534 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
5535 1188 : foreach(lc, parse->groupClause)
5536 : {
5537 112 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
5538 :
5539 112 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
5540 : }
5541 :
5542 : /*
5543 : * Construct a target containing all the non-flattenable targetlist items,
5544 : * and save aside the others for a moment.
5545 : */
5546 1076 : input_target = create_empty_pathtarget();
5547 1076 : flattenable_cols = NIL;
5548 :
5549 1076 : i = 0;
5550 4644 : foreach(lc, final_target->exprs)
5551 : {
5552 3568 : Expr *expr = (Expr *) lfirst(lc);
5553 3568 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5554 :
5555 : /*
5556 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
5557 : * that such items can't contain window functions, so it's okay to
5558 : * compute them below the WindowAgg nodes.)
5559 : */
5560 3568 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
5561 : {
5562 : /*
5563 : * Don't want to deconstruct this value, so add it to the input
5564 : * target as-is.
5565 : */
5566 1332 : add_column_to_pathtarget(input_target, expr, sgref);
5567 : }
5568 : else
5569 : {
5570 : /*
5571 : * Column is to be flattened, so just remember the expression for
5572 : * later call to pull_var_clause.
5573 : */
5574 2236 : flattenable_cols = lappend(flattenable_cols, expr);
5575 : }
5576 :
5577 3568 : i++;
5578 : }
5579 :
5580 : /*
5581 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
5582 : * add them to the input target if not already present. (Some might be
5583 : * there already because they're used directly as window/group clauses.)
5584 : *
5585 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
5586 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
5587 : * at higher levels. On the other hand, we should recurse into
5588 : * WindowFuncs to make sure their input expressions are available.
5589 : */
5590 1076 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
5591 : PVC_INCLUDE_AGGREGATES |
5592 : PVC_RECURSE_WINDOWFUNCS |
5593 : PVC_INCLUDE_PLACEHOLDERS);
5594 1076 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
5595 :
5596 : /* clean up cruft */
5597 1076 : list_free(flattenable_vars);
5598 1076 : list_free(flattenable_cols);
5599 :
5600 : /* XXX this causes some redundant cost calculation ... */
5601 1076 : return set_pathtarget_cost_width(root, input_target);
5602 : }
5603 :
5604 : /*
5605 : * make_pathkeys_for_window
5606 : * Create a pathkeys list describing the required input ordering
5607 : * for the given WindowClause.
5608 : *
5609 : * The required ordering is first the PARTITION keys, then the ORDER keys.
5610 : * In the future we might try to implement windowing using hashing, in which
5611 : * case the ordering could be relaxed, but for now we always sort.
5612 : */
5613 : static List *
5614 2236 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
5615 : List *tlist)
5616 : {
5617 : List *window_pathkeys;
5618 : List *window_sortclauses;
5619 :
5620 : /* Throw error if can't sort */
5621 2236 : if (!grouping_is_sortable(wc->partitionClause))
5622 0 : ereport(ERROR,
5623 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5624 : errmsg("could not implement window PARTITION BY"),
5625 : errdetail("Window partitioning columns must be of sortable datatypes.")));
5626 2236 : if (!grouping_is_sortable(wc->orderClause))
5627 0 : ereport(ERROR,
5628 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5629 : errmsg("could not implement window ORDER BY"),
5630 : errdetail("Window ordering columns must be of sortable datatypes.")));
5631 :
5632 : /* Okay, make the combined pathkeys */
5633 2236 : window_sortclauses = list_concat(list_copy(wc->partitionClause),
5634 2236 : list_copy(wc->orderClause));
5635 2236 : window_pathkeys = make_pathkeys_for_sortclauses(root,
5636 : window_sortclauses,
5637 : tlist);
5638 2236 : list_free(window_sortclauses);
5639 2236 : return window_pathkeys;
5640 : }
5641 :
5642 : /*
5643 : * make_sort_input_target
5644 : * Generate appropriate PathTarget for initial input to Sort step.
5645 : *
5646 : * If the query has ORDER BY, this function chooses the target to be computed
5647 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
5648 : * project) steps. This might or might not be identical to the query's final
5649 : * output target.
5650 : *
5651 : * The main argument for keeping the sort-input tlist the same as the final
5652 : * is that we avoid a separate projection node (which will be needed if
5653 : * they're different, because Sort can't project). However, there are also
5654 : * advantages to postponing tlist evaluation till after the Sort: it ensures
5655 : * a consistent order of evaluation for any volatile functions in the tlist,
5656 : * and if there's also a LIMIT, we can stop the query without ever computing
5657 : * tlist functions for later rows, which is beneficial for both volatile and
5658 : * expensive functions.
5659 : *
5660 : * Our current policy is to postpone volatile expressions till after the sort
5661 : * unconditionally (assuming that that's possible, ie they are in plain tlist
5662 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
5663 : * postpone set-returning expressions, because running them beforehand would
5664 : * bloat the sort dataset, and because it might cause unexpected output order
5665 : * if the sort isn't stable. However there's a constraint on that: all SRFs
5666 : * in the tlist should be evaluated at the same plan step, so that they can
5667 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
5668 : * mustn't postpone any SRFs. (Note that in principle that policy should
5669 : * probably get applied to the group/window input targetlists too, but we
5670 : * have not done that historically.) Lastly, expensive expressions are
5671 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
5672 : * partial evaluation of the query is possible (if neither is true, we expect
5673 : * to have to evaluate the expressions for every row anyway), or if there are
5674 : * any volatile or set-returning expressions (since once we've put in a
5675 : * projection at all, it won't cost any more to postpone more stuff).
5676 : *
5677 : * Another issue that could potentially be considered here is that
5678 : * evaluating tlist expressions could result in data that's either wider
5679 : * or narrower than the input Vars, thus changing the volume of data that
5680 : * has to go through the Sort. However, we usually have only a very bad
5681 : * idea of the output width of any expression more complex than a Var,
5682 : * so for now it seems too risky to try to optimize on that basis.
5683 : *
5684 : * Note that if we do produce a modified sort-input target, and then the
5685 : * query ends up not using an explicit Sort, no particular harm is done:
5686 : * we'll initially use the modified target for the preceding path nodes,
5687 : * but then change them to the final target with apply_projection_to_path.
5688 : * Moreover, in such a case the guarantees about evaluation order of
5689 : * volatile functions still hold, since the rows are sorted already.
5690 : *
5691 : * This function has some things in common with make_group_input_target and
5692 : * make_window_input_target, though the detailed rules for what to do are
5693 : * different. We never flatten/postpone any grouping or ordering columns;
5694 : * those are needed before the sort. If we do flatten a particular
5695 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
5696 : * computed earlier.
5697 : *
5698 : * 'final_target' is the query's final target list (in PathTarget form)
5699 : * 'have_postponed_srfs' is an output argument, see below
5700 : *
5701 : * The result is the PathTarget to be computed by the plan node immediately
5702 : * below the Sort step (and the Distinct step, if any). This will be
5703 : * exactly final_target if we decide a projection step wouldn't be helpful.
5704 : *
5705 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
5706 : * any set-returning functions to after the Sort.
5707 : */
5708 : static PathTarget *
5709 33874 : make_sort_input_target(PlannerInfo *root,
5710 : PathTarget *final_target,
5711 : bool *have_postponed_srfs)
5712 : {
5713 33874 : Query *parse = root->parse;
5714 : PathTarget *input_target;
5715 : int ncols;
5716 : bool *col_is_srf;
5717 : bool *postpone_col;
5718 : bool have_srf;
5719 : bool have_volatile;
5720 : bool have_expensive;
5721 : bool have_srf_sortcols;
5722 : bool postpone_srfs;
5723 : List *postponable_cols;
5724 : List *postponable_vars;
5725 : int i;
5726 : ListCell *lc;
5727 :
5728 : /* Shouldn't get here unless query has ORDER BY */
5729 : Assert(parse->sortClause);
5730 :
5731 33874 : *have_postponed_srfs = false; /* default result */
5732 :
5733 : /* Inspect tlist and collect per-column information */
5734 33874 : ncols = list_length(final_target->exprs);
5735 33874 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
5736 33874 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
5737 33874 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
5738 :
5739 33874 : i = 0;
5740 267818 : foreach(lc, final_target->exprs)
5741 : {
5742 233944 : Expr *expr = (Expr *) lfirst(lc);
5743 :
5744 : /*
5745 : * If the column has a sortgroupref, assume it has to be evaluated
5746 : * before sorting. Generally such columns would be ORDER BY, GROUP
5747 : * BY, etc targets. One exception is columns that were removed from
5748 : * GROUP BY by remove_useless_groupby_columns() ... but those would
5749 : * only be Vars anyway. There don't seem to be any cases where it
5750 : * would be worth the trouble to double-check.
5751 : */
5752 233944 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
5753 : {
5754 : /*
5755 : * Check for SRF or volatile functions. Check the SRF case first
5756 : * because we must know whether we have any postponed SRFs.
5757 : */
5758 382752 : if (parse->hasTargetSRFs &&
5759 124 : expression_returns_set((Node *) expr))
5760 : {
5761 : /* We'll decide below whether these are postponable */
5762 56 : col_is_srf[i] = true;
5763 56 : have_srf = true;
5764 : }
5765 191258 : else if (contain_volatile_functions((Node *) expr))
5766 : {
5767 : /* Unconditionally postpone */
5768 64 : postpone_col[i] = true;
5769 64 : have_volatile = true;
5770 : }
5771 : else
5772 : {
5773 : /*
5774 : * Else check the cost. XXX it's annoying to have to do this
5775 : * when set_pathtarget_cost_width() just did it. Refactor to
5776 : * allow sharing the work?
5777 : */
5778 : QualCost cost;
5779 :
5780 191194 : cost_qual_eval_node(&cost, (Node *) expr, root);
5781 :
5782 : /*
5783 : * We arbitrarily define "expensive" as "more than 10X
5784 : * cpu_operator_cost". Note this will take in any PL function
5785 : * with default cost.
5786 : */
5787 191194 : if (cost.per_tuple > 10 * cpu_operator_cost)
5788 : {
5789 17802 : postpone_col[i] = true;
5790 17802 : have_expensive = true;
5791 : }
5792 : }
5793 : }
5794 : else
5795 : {
5796 : /* For sortgroupref cols, just check if any contain SRFs */
5797 85260 : if (!have_srf_sortcols &&
5798 42822 : parse->hasTargetSRFs &&
5799 192 : expression_returns_set((Node *) expr))
5800 76 : have_srf_sortcols = true;
5801 : }
5802 :
5803 233944 : i++;
5804 : }
5805 :
5806 : /*
5807 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
5808 : */
5809 33874 : postpone_srfs = (have_srf && !have_srf_sortcols);
5810 :
5811 : /*
5812 : * If we don't need a post-sort projection, just return final_target.
5813 : */
5814 67648 : if (!(postpone_srfs || have_volatile ||
5815 44670 : (have_expensive &&
5816 21784 : (parse->limitCount || root->tuple_fraction > 0))))
5817 33766 : return final_target;
5818 :
5819 : /*
5820 : * Report whether the post-sort projection will contain set-returning
5821 : * functions. This is important because it affects whether the Sort can
5822 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
5823 : * to return.
5824 : */
5825 108 : *have_postponed_srfs = postpone_srfs;
5826 :
5827 : /*
5828 : * Construct the sort-input target, taking all non-postponable columns and
5829 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
5830 : * the postponable ones.
5831 : */
5832 108 : input_target = create_empty_pathtarget();
5833 108 : postponable_cols = NIL;
5834 :
5835 108 : i = 0;
5836 684 : foreach(lc, final_target->exprs)
5837 : {
5838 576 : Expr *expr = (Expr *) lfirst(lc);
5839 :
5840 576 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
5841 112 : postponable_cols = lappend(postponable_cols, expr);
5842 : else
5843 928 : add_column_to_pathtarget(input_target, expr,
5844 928 : get_pathtarget_sortgroupref(final_target, i));
5845 :
5846 576 : i++;
5847 : }
5848 :
5849 : /*
5850 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
5851 : * postponable columns, and add them to the sort-input target if not
5852 : * already present. (Some might be there already.) We mustn't
5853 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
5854 : * would be unable to recompute them.
5855 : */
5856 108 : postponable_vars = pull_var_clause((Node *) postponable_cols,
5857 : PVC_INCLUDE_AGGREGATES |
5858 : PVC_INCLUDE_WINDOWFUNCS |
5859 : PVC_INCLUDE_PLACEHOLDERS);
5860 108 : add_new_columns_to_pathtarget(input_target, postponable_vars);
5861 :
5862 : /* clean up cruft */
5863 108 : list_free(postponable_vars);
5864 108 : list_free(postponable_cols);
5865 :
5866 : /* XXX this represents even more redundant cost calculation ... */
5867 108 : return set_pathtarget_cost_width(root, input_target);
5868 : }
5869 :
5870 : /*
5871 : * get_cheapest_fractional_path
5872 : * Find the cheapest path for retrieving a specified fraction of all
5873 : * the tuples expected to be returned by the given relation.
5874 : *
5875 : * We interpret tuple_fraction the same way as grouping_planner.
5876 : *
5877 : * We assume set_cheapest() has been run on the given rel.
5878 : */
5879 : Path *
5880 280624 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
5881 : {
5882 280624 : Path *best_path = rel->cheapest_total_path;
5883 : ListCell *l;
5884 :
5885 : /* If all tuples will be retrieved, just return the cheapest-total path */
5886 280624 : if (tuple_fraction <= 0.0)
5887 277122 : return best_path;
5888 :
5889 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
5890 3502 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
5891 1298 : tuple_fraction /= best_path->rows;
5892 :
5893 8710 : foreach(l, rel->pathlist)
5894 : {
5895 5208 : Path *path = (Path *) lfirst(l);
5896 :
5897 6914 : if (path == rel->cheapest_total_path ||
5898 1706 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
5899 4728 : continue;
5900 :
5901 480 : best_path = path;
5902 : }
5903 :
5904 3502 : return best_path;
5905 : }
5906 :
5907 : /*
5908 : * adjust_paths_for_srfs
5909 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
5910 : *
5911 : * The executor can only handle set-returning functions that appear at the
5912 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
5913 : * that are not at top level, we need to split up the evaluation into multiple
5914 : * plan levels in which each level satisfies this constraint. This function
5915 : * modifies each Path of an upperrel that (might) compute any SRFs in its
5916 : * output tlist to insert appropriate projection steps.
5917 : *
5918 : * The given targets and targets_contain_srfs lists are from
5919 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
5920 : * target in targets.
5921 : */
5922 : static void
5923 3786 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
5924 : List *targets, List *targets_contain_srfs)
5925 : {
5926 : ListCell *lc;
5927 :
5928 : Assert(list_length(targets) == list_length(targets_contain_srfs));
5929 : Assert(!linitial_int(targets_contain_srfs));
5930 :
5931 : /* If no SRFs appear at this plan level, nothing to do */
5932 3786 : if (list_length(targets) == 1)
5933 208 : return;
5934 :
5935 : /*
5936 : * Stack SRF-evaluation nodes atop each path for the rel.
5937 : *
5938 : * In principle we should re-run set_cheapest() here to identify the
5939 : * cheapest path, but it seems unlikely that adding the same tlist eval
5940 : * costs to all the paths would change that, so we don't bother. Instead,
5941 : * just assume that the cheapest-startup and cheapest-total paths remain
5942 : * so. (There should be no parameterized paths anymore, so we needn't
5943 : * worry about updating cheapest_parameterized_paths.)
5944 : */
5945 7180 : foreach(lc, rel->pathlist)
5946 : {
5947 3602 : Path *subpath = (Path *) lfirst(lc);
5948 3602 : Path *newpath = subpath;
5949 : ListCell *lc1,
5950 : *lc2;
5951 :
5952 : Assert(subpath->param_info == NULL);
5953 11478 : forboth(lc1, targets, lc2, targets_contain_srfs)
5954 : {
5955 7876 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5956 7876 : bool contains_srfs = (bool) lfirst_int(lc2);
5957 :
5958 : /* If this level doesn't contain SRFs, do regular projection */
5959 7876 : if (contains_srfs)
5960 3642 : newpath = (Path *) create_set_projection_path(root,
5961 : rel,
5962 : newpath,
5963 : thistarget);
5964 : else
5965 4234 : newpath = (Path *) apply_projection_to_path(root,
5966 : rel,
5967 : newpath,
5968 : thistarget);
5969 : }
5970 3602 : lfirst(lc) = newpath;
5971 3602 : if (subpath == rel->cheapest_startup_path)
5972 52 : rel->cheapest_startup_path = newpath;
5973 3602 : if (subpath == rel->cheapest_total_path)
5974 52 : rel->cheapest_total_path = newpath;
5975 : }
5976 :
5977 : /* Likewise for partial paths, if any */
5978 3582 : foreach(lc, rel->partial_pathlist)
5979 : {
5980 4 : Path *subpath = (Path *) lfirst(lc);
5981 4 : Path *newpath = subpath;
5982 : ListCell *lc1,
5983 : *lc2;
5984 :
5985 : Assert(subpath->param_info == NULL);
5986 16 : forboth(lc1, targets, lc2, targets_contain_srfs)
5987 : {
5988 12 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5989 12 : bool contains_srfs = (bool) lfirst_int(lc2);
5990 :
5991 : /* If this level doesn't contain SRFs, do regular projection */
5992 12 : if (contains_srfs)
5993 4 : newpath = (Path *) create_set_projection_path(root,
5994 : rel,
5995 : newpath,
5996 : thistarget);
5997 : else
5998 : {
5999 : /* avoid apply_projection_to_path, in case of multiple refs */
6000 8 : newpath = (Path *) create_projection_path(root,
6001 : rel,
6002 : newpath,
6003 : thistarget);
6004 : }
6005 : }
6006 4 : lfirst(lc) = newpath;
6007 : }
6008 : }
6009 :
6010 : /*
6011 : * expression_planner
6012 : * Perform planner's transformations on a standalone expression.
6013 : *
6014 : * Various utility commands need to evaluate expressions that are not part
6015 : * of a plannable query. They can do so using the executor's regular
6016 : * expression-execution machinery, but first the expression has to be fed
6017 : * through here to transform it from parser output to something executable.
6018 : *
6019 : * Currently, we disallow sublinks in standalone expressions, so there's no
6020 : * real "planning" involved here. (That might not always be true though.)
6021 : * What we must do is run eval_const_expressions to ensure that any function
6022 : * calls are converted to positional notation and function default arguments
6023 : * get inserted. The fact that constant subexpressions get simplified is a
6024 : * side-effect that is useful when the expression will get evaluated more than
6025 : * once. Also, we must fix operator function IDs.
6026 : *
6027 : * This does not return any information about dependencies of the expression.
6028 : * Hence callers should use the results only for the duration of the current
6029 : * query. Callers that would like to cache the results for longer should use
6030 : * expression_planner_with_deps, probably via the plancache.
6031 : *
6032 : * Note: this must not make any damaging changes to the passed-in expression
6033 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6034 : * we first do an expression_tree_mutator-based walk, what is returned will
6035 : * be a new node tree.) The result is constructed in the current memory
6036 : * context; beware that this can leak a lot of additional stuff there, too.
6037 : */
6038 : Expr *
6039 81926 : expression_planner(Expr *expr)
6040 : {
6041 : Node *result;
6042 :
6043 : /*
6044 : * Convert named-argument function calls, insert default arguments and
6045 : * simplify constant subexprs
6046 : */
6047 81926 : result = eval_const_expressions(NULL, (Node *) expr);
6048 :
6049 : /* Fill in opfuncid values if missing */
6050 81926 : fix_opfuncids(result);
6051 :
6052 81926 : return (Expr *) result;
6053 : }
6054 :
6055 : /*
6056 : * expression_planner_with_deps
6057 : * Perform planner's transformations on a standalone expression,
6058 : * returning expression dependency information along with the result.
6059 : *
6060 : * This is identical to expression_planner() except that it also returns
6061 : * information about possible dependencies of the expression, ie identities of
6062 : * objects whose definitions affect the result. As in a PlannedStmt, these
6063 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6064 : */
6065 : Expr *
6066 338 : expression_planner_with_deps(Expr *expr,
6067 : List **relationOids,
6068 : List **invalItems)
6069 : {
6070 : Node *result;
6071 : PlannerGlobal glob;
6072 : PlannerInfo root;
6073 :
6074 : /* Make up dummy planner state so we can use setrefs machinery */
6075 338 : MemSet(&glob, 0, sizeof(glob));
6076 338 : glob.type = T_PlannerGlobal;
6077 338 : glob.relationOids = NIL;
6078 338 : glob.invalItems = NIL;
6079 :
6080 338 : MemSet(&root, 0, sizeof(root));
6081 338 : root.type = T_PlannerInfo;
6082 338 : root.glob = &glob;
6083 :
6084 : /*
6085 : * Convert named-argument function calls, insert default arguments and
6086 : * simplify constant subexprs. Collect identities of inlined functions
6087 : * and elided domains, too.
6088 : */
6089 338 : result = eval_const_expressions(&root, (Node *) expr);
6090 :
6091 : /* Fill in opfuncid values if missing */
6092 338 : fix_opfuncids(result);
6093 :
6094 : /*
6095 : * Now walk the finished expression to find anything else we ought to
6096 : * record as an expression dependency.
6097 : */
6098 338 : (void) extract_query_dependencies_walker(result, &root);
6099 :
6100 338 : *relationOids = glob.relationOids;
6101 338 : *invalItems = glob.invalItems;
6102 :
6103 338 : return (Expr *) result;
6104 : }
6105 :
6106 :
6107 : /*
6108 : * plan_cluster_use_sort
6109 : * Use the planner to decide how CLUSTER should implement sorting
6110 : *
6111 : * tableOid is the OID of a table to be clustered on its index indexOid
6112 : * (which is already known to be a btree index). Decide whether it's
6113 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6114 : * Return true to use sorting, false to use an indexscan.
6115 : *
6116 : * Note: caller had better already hold some type of lock on the table.
6117 : */
6118 : bool
6119 50 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6120 : {
6121 : PlannerInfo *root;
6122 : Query *query;
6123 : PlannerGlobal *glob;
6124 : RangeTblEntry *rte;
6125 : RelOptInfo *rel;
6126 : IndexOptInfo *indexInfo;
6127 : QualCost indexExprCost;
6128 : Cost comparisonCost;
6129 : Path *seqScanPath;
6130 : Path seqScanAndSortPath;
6131 : IndexPath *indexScanPath;
6132 : ListCell *lc;
6133 :
6134 : /* We can short-circuit the cost comparison if indexscans are disabled */
6135 50 : if (!enable_indexscan)
6136 4 : return true; /* use sort */
6137 :
6138 : /* Set up mostly-dummy planner state */
6139 46 : query = makeNode(Query);
6140 46 : query->commandType = CMD_SELECT;
6141 :
6142 46 : glob = makeNode(PlannerGlobal);
6143 :
6144 46 : root = makeNode(PlannerInfo);
6145 46 : root->parse = query;
6146 46 : root->glob = glob;
6147 46 : root->query_level = 1;
6148 46 : root->planner_cxt = CurrentMemoryContext;
6149 46 : root->wt_param_id = -1;
6150 :
6151 : /* Build a minimal RTE for the rel */
6152 46 : rte = makeNode(RangeTblEntry);
6153 46 : rte->rtekind = RTE_RELATION;
6154 46 : rte->relid = tableOid;
6155 46 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6156 46 : rte->rellockmode = AccessShareLock;
6157 46 : rte->lateral = false;
6158 46 : rte->inh = false;
6159 46 : rte->inFromCl = true;
6160 46 : query->rtable = list_make1(rte);
6161 :
6162 : /* Set up RTE/RelOptInfo arrays */
6163 46 : setup_simple_rel_arrays(root);
6164 :
6165 : /* Build RelOptInfo */
6166 46 : rel = build_simple_rel(root, 1, NULL);
6167 :
6168 : /* Locate IndexOptInfo for the target index */
6169 46 : indexInfo = NULL;
6170 74 : foreach(lc, rel->indexlist)
6171 : {
6172 74 : indexInfo = lfirst_node(IndexOptInfo, lc);
6173 74 : if (indexInfo->indexoid == indexOid)
6174 46 : break;
6175 : }
6176 :
6177 : /*
6178 : * It's possible that get_relation_info did not generate an IndexOptInfo
6179 : * for the desired index; this could happen if it's not yet reached its
6180 : * indcheckxmin usability horizon, or if it's a system index and we're
6181 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6182 : * trust the index contents but use seqscan-and-sort.
6183 : */
6184 46 : if (lc == NULL) /* not in the list? */
6185 0 : return true; /* use sort */
6186 :
6187 : /*
6188 : * Rather than doing all the pushups that would be needed to use
6189 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6190 : */
6191 46 : rel->rows = rel->tuples;
6192 46 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6193 :
6194 46 : root->total_table_pages = rel->pages;
6195 :
6196 : /*
6197 : * Determine eval cost of the index expressions, if any. We need to
6198 : * charge twice that amount for each tuple comparison that happens during
6199 : * the sort, since tuplesort.c will have to re-evaluate the index
6200 : * expressions each time. (XXX that's pretty inefficient...)
6201 : */
6202 46 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6203 46 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6204 :
6205 : /* Estimate the cost of seq scan + sort */
6206 46 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6207 92 : cost_sort(&seqScanAndSortPath, root, NIL,
6208 46 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6209 : comparisonCost, maintenance_work_mem, -1.0);
6210 :
6211 : /* Estimate the cost of index scan */
6212 46 : indexScanPath = create_index_path(root, indexInfo,
6213 : NIL, NIL, NIL, NIL,
6214 : ForwardScanDirection, false,
6215 : NULL, 1.0, false);
6216 :
6217 46 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6218 : }
6219 :
6220 : /*
6221 : * plan_create_index_workers
6222 : * Use the planner to decide how many parallel worker processes
6223 : * CREATE INDEX should request for use
6224 : *
6225 : * tableOid is the table on which the index is to be built. indexOid is the
6226 : * OID of an index to be created or reindexed (which must be a btree index).
6227 : *
6228 : * Return value is the number of parallel worker processes to request. It
6229 : * may be unsafe to proceed if this is 0. Note that this does not include the
6230 : * leader participating as a worker (value is always a number of parallel
6231 : * worker processes).
6232 : *
6233 : * Note: caller had better already hold some type of lock on the table and
6234 : * index.
6235 : */
6236 : int
6237 21824 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6238 : {
6239 : PlannerInfo *root;
6240 : Query *query;
6241 : PlannerGlobal *glob;
6242 : RangeTblEntry *rte;
6243 : Relation heap;
6244 : Relation index;
6245 : RelOptInfo *rel;
6246 : int parallel_workers;
6247 : BlockNumber heap_blocks;
6248 : double reltuples;
6249 : double allvisfrac;
6250 :
6251 : /* Return immediately when parallelism disabled */
6252 21824 : if (max_parallel_maintenance_workers == 0)
6253 0 : return 0;
6254 :
6255 : /* Set up largely-dummy planner state */
6256 21824 : query = makeNode(Query);
6257 21824 : query->commandType = CMD_SELECT;
6258 :
6259 21824 : glob = makeNode(PlannerGlobal);
6260 :
6261 21824 : root = makeNode(PlannerInfo);
6262 21824 : root->parse = query;
6263 21824 : root->glob = glob;
6264 21824 : root->query_level = 1;
6265 21824 : root->planner_cxt = CurrentMemoryContext;
6266 21824 : root->wt_param_id = -1;
6267 :
6268 : /*
6269 : * Build a minimal RTE.
6270 : *
6271 : * Mark the RTE with inh = true. This is a kludge to prevent
6272 : * get_relation_info() from fetching index info, which is necessary
6273 : * because it does not expect that any IndexOptInfo is currently
6274 : * undergoing REINDEX.
6275 : */
6276 21824 : rte = makeNode(RangeTblEntry);
6277 21824 : rte->rtekind = RTE_RELATION;
6278 21824 : rte->relid = tableOid;
6279 21824 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6280 21824 : rte->rellockmode = AccessShareLock;
6281 21824 : rte->lateral = false;
6282 21824 : rte->inh = true;
6283 21824 : rte->inFromCl = true;
6284 21824 : query->rtable = list_make1(rte);
6285 :
6286 : /* Set up RTE/RelOptInfo arrays */
6287 21824 : setup_simple_rel_arrays(root);
6288 :
6289 : /* Build RelOptInfo */
6290 21824 : rel = build_simple_rel(root, 1, NULL);
6291 :
6292 : /* Rels are assumed already locked by the caller */
6293 21824 : heap = table_open(tableOid, NoLock);
6294 21824 : index = index_open(indexOid, NoLock);
6295 :
6296 : /*
6297 : * Determine if it's safe to proceed.
6298 : *
6299 : * Currently, parallel workers can't access the leader's temporary tables.
6300 : * Furthermore, any index predicate or index expressions must be parallel
6301 : * safe.
6302 : */
6303 42152 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6304 40646 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6305 20318 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6306 : {
6307 1506 : parallel_workers = 0;
6308 1506 : goto done;
6309 : }
6310 :
6311 : /*
6312 : * If parallel_workers storage parameter is set for the table, accept that
6313 : * as the number of parallel worker processes to launch (though still cap
6314 : * at max_parallel_maintenance_workers). Note that we deliberately do not
6315 : * consider any other factor when parallel_workers is set. (e.g., memory
6316 : * use by workers.)
6317 : */
6318 20318 : if (rel->rel_parallel_workers != -1)
6319 : {
6320 14 : parallel_workers = Min(rel->rel_parallel_workers,
6321 : max_parallel_maintenance_workers);
6322 14 : goto done;
6323 : }
6324 :
6325 : /*
6326 : * Estimate heap relation size ourselves, since rel->pages cannot be
6327 : * trusted (heap RTE was marked as inheritance parent)
6328 : */
6329 20304 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6330 :
6331 : /*
6332 : * Determine number of workers to scan the heap relation using generic
6333 : * model
6334 : */
6335 20304 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6336 : max_parallel_maintenance_workers);
6337 :
6338 : /*
6339 : * Cap workers based on available maintenance_work_mem as needed.
6340 : *
6341 : * Note that each tuplesort participant receives an even share of the
6342 : * total maintenance_work_mem budget. Aim to leave participants
6343 : * (including the leader as a participant) with no less than 32MB of
6344 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
6345 : * immediately past the threshold of being capable of launching a single
6346 : * parallel worker to sort.
6347 : */
6348 40840 : while (parallel_workers > 0 &&
6349 158 : maintenance_work_mem / (parallel_workers + 1) < 32768L)
6350 74 : parallel_workers--;
6351 :
6352 : done:
6353 21824 : index_close(index, NoLock);
6354 21824 : table_close(heap, NoLock);
6355 :
6356 21824 : return parallel_workers;
6357 : }
6358 :
6359 : /*
6360 : * add_paths_to_grouping_rel
6361 : *
6362 : * Add non-partial paths to grouping relation.
6363 : */
6364 : static void
6365 24132 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
6366 : RelOptInfo *grouped_rel,
6367 : RelOptInfo *partially_grouped_rel,
6368 : const AggClauseCosts *agg_costs,
6369 : grouping_sets_data *gd, double dNumGroups,
6370 : GroupPathExtraData *extra)
6371 : {
6372 24132 : Query *parse = root->parse;
6373 24132 : Path *cheapest_path = input_rel->cheapest_total_path;
6374 : ListCell *lc;
6375 24132 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6376 24132 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6377 24132 : List *havingQual = (List *) extra->havingQual;
6378 24132 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6379 :
6380 24132 : if (can_sort)
6381 : {
6382 : /*
6383 : * Use any available suitably-sorted path as input, and also consider
6384 : * sorting the cheapest-total path.
6385 : */
6386 49552 : foreach(lc, input_rel->pathlist)
6387 : {
6388 25424 : Path *path = (Path *) lfirst(lc);
6389 : bool is_sorted;
6390 :
6391 25424 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6392 : path->pathkeys);
6393 25424 : if (path == cheapest_path || is_sorted)
6394 : {
6395 : /* Sort the cheapest-total path if it isn't already sorted */
6396 25314 : if (!is_sorted)
6397 2492 : path = (Path *) create_sort_path(root,
6398 : grouped_rel,
6399 : path,
6400 : root->group_pathkeys,
6401 : -1.0);
6402 :
6403 : /* Now decide what to stick atop it */
6404 25314 : if (parse->groupingSets)
6405 : {
6406 436 : consider_groupingsets_paths(root, grouped_rel,
6407 : path, true, can_hash,
6408 : gd, agg_costs, dNumGroups);
6409 : }
6410 24878 : else if (parse->hasAggs)
6411 : {
6412 : /*
6413 : * We have aggregation, possibly with plain GROUP BY. Make
6414 : * an AggPath.
6415 : */
6416 24624 : add_path(grouped_rel, (Path *)
6417 73872 : create_agg_path(root,
6418 : grouped_rel,
6419 : path,
6420 24624 : grouped_rel->reltarget,
6421 24624 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6422 : AGGSPLIT_SIMPLE,
6423 : parse->groupClause,
6424 : havingQual,
6425 : agg_costs,
6426 : dNumGroups));
6427 : }
6428 254 : else if (parse->groupClause)
6429 : {
6430 : /*
6431 : * We have GROUP BY without aggregation or grouping sets.
6432 : * Make a GroupPath.
6433 : */
6434 254 : add_path(grouped_rel, (Path *)
6435 254 : create_group_path(root,
6436 : grouped_rel,
6437 : path,
6438 : parse->groupClause,
6439 : havingQual,
6440 : dNumGroups));
6441 : }
6442 : else
6443 : {
6444 : /* Other cases should have been handled above */
6445 : Assert(false);
6446 : }
6447 : }
6448 : }
6449 :
6450 : /*
6451 : * Instead of operating directly on the input relation, we can
6452 : * consider finalizing a partially aggregated path.
6453 : */
6454 24128 : if (partially_grouped_rel != NULL)
6455 : {
6456 2288 : foreach(lc, partially_grouped_rel->pathlist)
6457 : {
6458 1372 : Path *path = (Path *) lfirst(lc);
6459 :
6460 : /*
6461 : * Insert a Sort node, if required. But there's no point in
6462 : * sorting anything but the cheapest path.
6463 : */
6464 1372 : if (!pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
6465 : {
6466 344 : if (path != partially_grouped_rel->cheapest_total_path)
6467 64 : continue;
6468 280 : path = (Path *) create_sort_path(root,
6469 : grouped_rel,
6470 : path,
6471 : root->group_pathkeys,
6472 : -1.0);
6473 : }
6474 :
6475 1308 : if (parse->hasAggs)
6476 1202 : add_path(grouped_rel, (Path *)
6477 3606 : create_agg_path(root,
6478 : grouped_rel,
6479 : path,
6480 1202 : grouped_rel->reltarget,
6481 1202 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6482 : AGGSPLIT_FINAL_DESERIAL,
6483 : parse->groupClause,
6484 : havingQual,
6485 : agg_final_costs,
6486 : dNumGroups));
6487 : else
6488 106 : add_path(grouped_rel, (Path *)
6489 106 : create_group_path(root,
6490 : grouped_rel,
6491 : path,
6492 : parse->groupClause,
6493 : havingQual,
6494 : dNumGroups));
6495 : }
6496 : }
6497 : }
6498 :
6499 24132 : if (can_hash)
6500 : {
6501 : double hashaggtablesize;
6502 :
6503 2394 : if (parse->groupingSets)
6504 : {
6505 : /*
6506 : * Try for a hash-only groupingsets path over unsorted input.
6507 : */
6508 360 : consider_groupingsets_paths(root, grouped_rel,
6509 : cheapest_path, false, true,
6510 : gd, agg_costs, dNumGroups);
6511 : }
6512 : else
6513 : {
6514 2034 : hashaggtablesize = estimate_hashagg_tablesize(cheapest_path,
6515 : agg_costs,
6516 : dNumGroups);
6517 :
6518 : /*
6519 : * Provided that the estimated size of the hashtable does not
6520 : * exceed work_mem, we'll generate a HashAgg Path, although if we
6521 : * were unable to sort above, then we'd better generate a Path, so
6522 : * that we at least have one.
6523 : */
6524 2034 : if (hashaggtablesize < work_mem * 1024L ||
6525 0 : grouped_rel->pathlist == NIL)
6526 : {
6527 : /*
6528 : * We just need an Agg over the cheapest-total input path,
6529 : * since input order won't matter.
6530 : */
6531 2034 : add_path(grouped_rel, (Path *)
6532 4068 : create_agg_path(root, grouped_rel,
6533 : cheapest_path,
6534 2034 : grouped_rel->reltarget,
6535 : AGG_HASHED,
6536 : AGGSPLIT_SIMPLE,
6537 : parse->groupClause,
6538 : havingQual,
6539 : agg_costs,
6540 : dNumGroups));
6541 : }
6542 : }
6543 :
6544 : /*
6545 : * Generate a Finalize HashAgg Path atop of the cheapest partially
6546 : * grouped path, assuming there is one. Once again, we'll only do this
6547 : * if it looks as though the hash table won't exceed work_mem.
6548 : */
6549 2394 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
6550 : {
6551 458 : Path *path = partially_grouped_rel->cheapest_total_path;
6552 :
6553 458 : hashaggtablesize = estimate_hashagg_tablesize(path,
6554 : agg_final_costs,
6555 : dNumGroups);
6556 :
6557 458 : if (hashaggtablesize < work_mem * 1024L)
6558 458 : add_path(grouped_rel, (Path *)
6559 916 : create_agg_path(root,
6560 : grouped_rel,
6561 : path,
6562 458 : grouped_rel->reltarget,
6563 : AGG_HASHED,
6564 : AGGSPLIT_FINAL_DESERIAL,
6565 : parse->groupClause,
6566 : havingQual,
6567 : agg_final_costs,
6568 : dNumGroups));
6569 : }
6570 : }
6571 :
6572 : /*
6573 : * When partitionwise aggregate is used, we might have fully aggregated
6574 : * paths in the partial pathlist, because add_paths_to_append_rel() will
6575 : * consider a path for grouped_rel consisting of a Parallel Append of
6576 : * non-partial paths from each child.
6577 : */
6578 24132 : if (grouped_rel->partial_pathlist != NIL)
6579 108 : gather_grouping_paths(root, grouped_rel);
6580 24132 : }
6581 :
6582 : /*
6583 : * create_partial_grouping_paths
6584 : *
6585 : * Create a new upper relation representing the result of partial aggregation
6586 : * and populate it with appropriate paths. Note that we don't finalize the
6587 : * lists of paths here, so the caller can add additional partial or non-partial
6588 : * paths and must afterward call gather_grouping_paths and set_cheapest on
6589 : * the returned upper relation.
6590 : *
6591 : * All paths for this new upper relation -- both partial and non-partial --
6592 : * have been partially aggregated but require a subsequent FinalizeAggregate
6593 : * step.
6594 : *
6595 : * NB: This function is allowed to return NULL if it determines that there is
6596 : * no real need to create a new RelOptInfo.
6597 : */
6598 : static RelOptInfo *
6599 8908 : create_partial_grouping_paths(PlannerInfo *root,
6600 : RelOptInfo *grouped_rel,
6601 : RelOptInfo *input_rel,
6602 : grouping_sets_data *gd,
6603 : GroupPathExtraData *extra,
6604 : bool force_rel_creation)
6605 : {
6606 8908 : Query *parse = root->parse;
6607 : RelOptInfo *partially_grouped_rel;
6608 8908 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
6609 8908 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6610 8908 : Path *cheapest_partial_path = NULL;
6611 8908 : Path *cheapest_total_path = NULL;
6612 8908 : double dNumPartialGroups = 0;
6613 8908 : double dNumPartialPartialGroups = 0;
6614 : ListCell *lc;
6615 8908 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6616 8908 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6617 :
6618 : /*
6619 : * Consider whether we should generate partially aggregated non-partial
6620 : * paths. We can only do this if we have a non-partial path, and only if
6621 : * the parent of the input rel is performing partial partitionwise
6622 : * aggregation. (Note that extra->patype is the type of partitionwise
6623 : * aggregation being used at the parent level, not this level.)
6624 : */
6625 17816 : if (input_rel->pathlist != NIL &&
6626 8908 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
6627 374 : cheapest_total_path = input_rel->cheapest_total_path;
6628 :
6629 : /*
6630 : * If parallelism is possible for grouped_rel, then we should consider
6631 : * generating partially-grouped partial paths. However, if the input rel
6632 : * has no partial paths, then we can't.
6633 : */
6634 8908 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
6635 1058 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
6636 :
6637 : /*
6638 : * If we can't partially aggregate partial paths, and we can't partially
6639 : * aggregate non-partial paths, then don't bother creating the new
6640 : * RelOptInfo at all, unless the caller specified force_rel_creation.
6641 : */
6642 8908 : if (cheapest_total_path == NULL &&
6643 7676 : cheapest_partial_path == NULL &&
6644 7676 : !force_rel_creation)
6645 7618 : return NULL;
6646 :
6647 : /*
6648 : * Build a new upper relation to represent the result of partially
6649 : * aggregating the rows from the input relation.
6650 : */
6651 1290 : partially_grouped_rel = fetch_upper_rel(root,
6652 : UPPERREL_PARTIAL_GROUP_AGG,
6653 : grouped_rel->relids);
6654 1290 : partially_grouped_rel->consider_parallel =
6655 1290 : grouped_rel->consider_parallel;
6656 1290 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
6657 1290 : partially_grouped_rel->serverid = grouped_rel->serverid;
6658 1290 : partially_grouped_rel->userid = grouped_rel->userid;
6659 1290 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
6660 1290 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
6661 :
6662 : /*
6663 : * Build target list for partial aggregate paths. These paths cannot just
6664 : * emit the same tlist as regular aggregate paths, because (1) we must
6665 : * include Vars and Aggrefs needed in HAVING, which might not appear in
6666 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
6667 : */
6668 1290 : partially_grouped_rel->reltarget =
6669 1290 : make_partial_grouping_target(root, grouped_rel->reltarget,
6670 : extra->havingQual);
6671 :
6672 1290 : if (!extra->partial_costs_set)
6673 : {
6674 : /*
6675 : * Collect statistics about aggregates for estimating costs of
6676 : * performing aggregation in parallel.
6677 : */
6678 724 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
6679 724 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
6680 724 : if (parse->hasAggs)
6681 : {
6682 : List *partial_target_exprs;
6683 :
6684 : /* partial phase */
6685 668 : partial_target_exprs = partially_grouped_rel->reltarget->exprs;
6686 668 : get_agg_clause_costs(root, (Node *) partial_target_exprs,
6687 : AGGSPLIT_INITIAL_SERIAL,
6688 : agg_partial_costs);
6689 :
6690 : /* final phase */
6691 668 : get_agg_clause_costs(root, (Node *) grouped_rel->reltarget->exprs,
6692 : AGGSPLIT_FINAL_DESERIAL,
6693 : agg_final_costs);
6694 668 : get_agg_clause_costs(root, extra->havingQual,
6695 : AGGSPLIT_FINAL_DESERIAL,
6696 : agg_final_costs);
6697 : }
6698 :
6699 724 : extra->partial_costs_set = true;
6700 : }
6701 :
6702 : /* Estimate number of partial groups. */
6703 1290 : if (cheapest_total_path != NULL)
6704 374 : dNumPartialGroups =
6705 374 : get_number_of_groups(root,
6706 : cheapest_total_path->rows,
6707 : gd,
6708 : extra->targetList);
6709 1290 : if (cheapest_partial_path != NULL)
6710 1058 : dNumPartialPartialGroups =
6711 1058 : get_number_of_groups(root,
6712 : cheapest_partial_path->rows,
6713 : gd,
6714 : extra->targetList);
6715 :
6716 1290 : if (can_sort && cheapest_total_path != NULL)
6717 : {
6718 : /* This should have been checked previously */
6719 : Assert(parse->hasAggs || parse->groupClause);
6720 :
6721 : /*
6722 : * Use any available suitably-sorted path as input, and also consider
6723 : * sorting the cheapest partial path.
6724 : */
6725 754 : foreach(lc, input_rel->pathlist)
6726 : {
6727 380 : Path *path = (Path *) lfirst(lc);
6728 : bool is_sorted;
6729 :
6730 380 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6731 : path->pathkeys);
6732 380 : if (path == cheapest_total_path || is_sorted)
6733 : {
6734 : /* Sort the cheapest partial path, if it isn't already */
6735 380 : if (!is_sorted)
6736 374 : path = (Path *) create_sort_path(root,
6737 : partially_grouped_rel,
6738 : path,
6739 : root->group_pathkeys,
6740 : -1.0);
6741 :
6742 380 : if (parse->hasAggs)
6743 356 : add_path(partially_grouped_rel, (Path *)
6744 1068 : create_agg_path(root,
6745 : partially_grouped_rel,
6746 : path,
6747 356 : partially_grouped_rel->reltarget,
6748 356 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6749 : AGGSPLIT_INITIAL_SERIAL,
6750 : parse->groupClause,
6751 : NIL,
6752 : agg_partial_costs,
6753 : dNumPartialGroups));
6754 : else
6755 24 : add_path(partially_grouped_rel, (Path *)
6756 24 : create_group_path(root,
6757 : partially_grouped_rel,
6758 : path,
6759 : parse->groupClause,
6760 : NIL,
6761 : dNumPartialGroups));
6762 : }
6763 : }
6764 : }
6765 :
6766 1290 : if (can_sort && cheapest_partial_path != NULL)
6767 : {
6768 : /* Similar to above logic, but for partial paths. */
6769 2116 : foreach(lc, input_rel->partial_pathlist)
6770 : {
6771 1058 : Path *path = (Path *) lfirst(lc);
6772 : bool is_sorted;
6773 :
6774 1058 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6775 : path->pathkeys);
6776 1058 : if (path == cheapest_partial_path || is_sorted)
6777 : {
6778 : /* Sort the cheapest partial path, if it isn't already */
6779 1058 : if (!is_sorted)
6780 600 : path = (Path *) create_sort_path(root,
6781 : partially_grouped_rel,
6782 : path,
6783 : root->group_pathkeys,
6784 : -1.0);
6785 :
6786 1058 : if (parse->hasAggs)
6787 1010 : add_partial_path(partially_grouped_rel, (Path *)
6788 3030 : create_agg_path(root,
6789 : partially_grouped_rel,
6790 : path,
6791 1010 : partially_grouped_rel->reltarget,
6792 1010 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6793 : AGGSPLIT_INITIAL_SERIAL,
6794 : parse->groupClause,
6795 : NIL,
6796 : agg_partial_costs,
6797 : dNumPartialPartialGroups));
6798 : else
6799 48 : add_partial_path(partially_grouped_rel, (Path *)
6800 48 : create_group_path(root,
6801 : partially_grouped_rel,
6802 : path,
6803 : parse->groupClause,
6804 : NIL,
6805 : dNumPartialPartialGroups));
6806 : }
6807 : }
6808 : }
6809 :
6810 1290 : if (can_hash && cheapest_total_path != NULL)
6811 : {
6812 : double hashaggtablesize;
6813 :
6814 : /* Checked above */
6815 : Assert(parse->hasAggs || parse->groupClause);
6816 :
6817 374 : hashaggtablesize =
6818 : estimate_hashagg_tablesize(cheapest_total_path,
6819 : agg_partial_costs,
6820 : dNumPartialGroups);
6821 :
6822 : /*
6823 : * Tentatively produce a partial HashAgg Path, depending on if it
6824 : * looks as if the hash table will fit in work_mem.
6825 : */
6826 374 : if (hashaggtablesize < work_mem * 1024L &&
6827 : cheapest_total_path != NULL)
6828 : {
6829 374 : add_path(partially_grouped_rel, (Path *)
6830 748 : create_agg_path(root,
6831 : partially_grouped_rel,
6832 : cheapest_total_path,
6833 374 : partially_grouped_rel->reltarget,
6834 : AGG_HASHED,
6835 : AGGSPLIT_INITIAL_SERIAL,
6836 : parse->groupClause,
6837 : NIL,
6838 : agg_partial_costs,
6839 : dNumPartialGroups));
6840 : }
6841 : }
6842 :
6843 1290 : if (can_hash && cheapest_partial_path != NULL)
6844 : {
6845 : double hashaggtablesize;
6846 :
6847 600 : hashaggtablesize =
6848 : estimate_hashagg_tablesize(cheapest_partial_path,
6849 : agg_partial_costs,
6850 : dNumPartialPartialGroups);
6851 :
6852 : /* Do the same for partial paths. */
6853 600 : if (hashaggtablesize < work_mem * 1024L &&
6854 : cheapest_partial_path != NULL)
6855 : {
6856 600 : add_partial_path(partially_grouped_rel, (Path *)
6857 1200 : create_agg_path(root,
6858 : partially_grouped_rel,
6859 : cheapest_partial_path,
6860 600 : partially_grouped_rel->reltarget,
6861 : AGG_HASHED,
6862 : AGGSPLIT_INITIAL_SERIAL,
6863 : parse->groupClause,
6864 : NIL,
6865 : agg_partial_costs,
6866 : dNumPartialPartialGroups));
6867 : }
6868 : }
6869 :
6870 : /*
6871 : * If there is an FDW that's responsible for all baserels of the query,
6872 : * let it consider adding partially grouped ForeignPaths.
6873 : */
6874 1296 : if (partially_grouped_rel->fdwroutine &&
6875 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
6876 : {
6877 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
6878 :
6879 6 : fdwroutine->GetForeignUpperPaths(root,
6880 : UPPERREL_PARTIAL_GROUP_AGG,
6881 : input_rel, partially_grouped_rel,
6882 : extra);
6883 : }
6884 :
6885 1290 : return partially_grouped_rel;
6886 : }
6887 :
6888 : /*
6889 : * Generate Gather and Gather Merge paths for a grouping relation or partial
6890 : * grouping relation.
6891 : *
6892 : * generate_gather_paths does most of the work, but we also consider a special
6893 : * case: we could try sorting the data by the group_pathkeys and then applying
6894 : * Gather Merge.
6895 : *
6896 : * NB: This function shouldn't be used for anything other than a grouped or
6897 : * partially grouped relation not only because of the fact that it explicitly
6898 : * references group_pathkeys but we pass "true" as the third argument to
6899 : * generate_gather_paths().
6900 : */
6901 : static void
6902 966 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
6903 : {
6904 : Path *cheapest_partial_path;
6905 :
6906 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
6907 966 : generate_gather_paths(root, rel, true);
6908 :
6909 : /* Try cheapest partial path + explicit Sort + Gather Merge. */
6910 966 : cheapest_partial_path = linitial(rel->partial_pathlist);
6911 966 : if (!pathkeys_contained_in(root->group_pathkeys,
6912 : cheapest_partial_path->pathkeys))
6913 : {
6914 : Path *path;
6915 : double total_groups;
6916 :
6917 470 : total_groups =
6918 470 : cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
6919 470 : path = (Path *) create_sort_path(root, rel, cheapest_partial_path,
6920 : root->group_pathkeys,
6921 : -1.0);
6922 470 : path = (Path *)
6923 470 : create_gather_merge_path(root,
6924 : rel,
6925 : path,
6926 470 : rel->reltarget,
6927 : root->group_pathkeys,
6928 : NULL,
6929 : &total_groups);
6930 :
6931 470 : add_path(rel, path);
6932 : }
6933 966 : }
6934 :
6935 : /*
6936 : * can_partial_agg
6937 : *
6938 : * Determines whether or not partial grouping and/or aggregation is possible.
6939 : * Returns true when possible, false otherwise.
6940 : */
6941 : static bool
6942 23588 : can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
6943 : {
6944 23588 : Query *parse = root->parse;
6945 :
6946 23588 : if (!parse->hasAggs && parse->groupClause == NIL)
6947 : {
6948 : /*
6949 : * We don't know how to do parallel aggregation unless we have either
6950 : * some aggregates or a grouping clause.
6951 : */
6952 0 : return false;
6953 : }
6954 23588 : else if (parse->groupingSets)
6955 : {
6956 : /* We don't know how to do grouping sets in parallel. */
6957 412 : return false;
6958 : }
6959 23176 : else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
6960 : {
6961 : /* Insufficient support for partial mode. */
6962 15138 : return false;
6963 : }
6964 :
6965 : /* Everything looks good. */
6966 8038 : return true;
6967 : }
6968 :
6969 : /*
6970 : * apply_scanjoin_target_to_paths
6971 : *
6972 : * Adjust the final scan/join relation, and recursively all of its children,
6973 : * to generate the final scan/join target. It would be more correct to model
6974 : * this as a separate planning step with a new RelOptInfo at the toplevel and
6975 : * for each child relation, but doing it this way is noticeably cheaper.
6976 : * Maybe that problem can be solved at some point, but for now we do this.
6977 : *
6978 : * If tlist_same_exprs is true, then the scan/join target to be applied has
6979 : * the same expressions as the existing reltarget, so we need only insert the
6980 : * appropriate sortgroupref information. By avoiding the creation of
6981 : * projection paths we save effort both immediately and at plan creation time.
6982 : */
6983 : static void
6984 298408 : apply_scanjoin_target_to_paths(PlannerInfo *root,
6985 : RelOptInfo *rel,
6986 : List *scanjoin_targets,
6987 : List *scanjoin_targets_contain_srfs,
6988 : bool scanjoin_target_parallel_safe,
6989 : bool tlist_same_exprs)
6990 : {
6991 298408 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
6992 : PathTarget *scanjoin_target;
6993 : ListCell *lc;
6994 :
6995 : /* This recurses, so be paranoid. */
6996 298408 : check_stack_depth();
6997 :
6998 : /*
6999 : * If the rel is partitioned, we want to drop its existing paths and
7000 : * generate new ones. This function would still be correct if we kept the
7001 : * existing paths: we'd modify them to generate the correct target above
7002 : * the partitioning Append, and then they'd compete on cost with paths
7003 : * generating the target below the Append. However, in our current cost
7004 : * model the latter way is always the same or cheaper cost, so modifying
7005 : * the existing paths would just be useless work. Moreover, when the cost
7006 : * is the same, varying roundoff errors might sometimes allow an existing
7007 : * path to be picked, resulting in undesirable cross-platform plan
7008 : * variations. So we drop old paths and thereby force the work to be done
7009 : * below the Append, except in the case of a non-parallel-safe target.
7010 : *
7011 : * Some care is needed, because we have to allow generate_gather_paths to
7012 : * see the old partial paths in the next stanza. Hence, zap the main
7013 : * pathlist here, then allow generate_gather_paths to add path(s) to the
7014 : * main list, and finally zap the partial pathlist.
7015 : */
7016 298408 : if (rel_is_partitioned)
7017 4370 : rel->pathlist = NIL;
7018 :
7019 : /*
7020 : * If the scan/join target is not parallel-safe, partial paths cannot
7021 : * generate it.
7022 : */
7023 298408 : if (!scanjoin_target_parallel_safe)
7024 : {
7025 : /*
7026 : * Since we can't generate the final scan/join target in parallel
7027 : * workers, this is our last opportunity to use any partial paths that
7028 : * exist; so build Gather path(s) that use them and emit whatever the
7029 : * current reltarget is. We don't do this in the case where the
7030 : * target is parallel-safe, since we will be able to generate superior
7031 : * paths by doing it after the final scan/join target has been
7032 : * applied.
7033 : */
7034 68174 : generate_gather_paths(root, rel, false);
7035 :
7036 : /* Can't use parallel query above this level. */
7037 68174 : rel->partial_pathlist = NIL;
7038 68174 : rel->consider_parallel = false;
7039 : }
7040 :
7041 : /* Finish dropping old paths for a partitioned rel, per comment above */
7042 298408 : if (rel_is_partitioned)
7043 4370 : rel->partial_pathlist = NIL;
7044 :
7045 : /* Extract SRF-free scan/join target. */
7046 298408 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7047 :
7048 : /*
7049 : * Apply the SRF-free scan/join target to each existing path.
7050 : *
7051 : * If the tlist exprs are the same, we can just inject the sortgroupref
7052 : * information into the existing pathtargets. Otherwise, replace each
7053 : * path with a projection path that generates the SRF-free scan/join
7054 : * target. This can't change the ordering of paths within rel->pathlist,
7055 : * so we just modify the list in place.
7056 : */
7057 607684 : foreach(lc, rel->pathlist)
7058 : {
7059 309276 : Path *subpath = (Path *) lfirst(lc);
7060 :
7061 : /* Shouldn't have any parameterized paths anymore */
7062 : Assert(subpath->param_info == NULL);
7063 :
7064 309276 : if (tlist_same_exprs)
7065 195240 : subpath->pathtarget->sortgrouprefs =
7066 97620 : scanjoin_target->sortgrouprefs;
7067 : else
7068 : {
7069 : Path *newpath;
7070 :
7071 211656 : newpath = (Path *) create_projection_path(root, rel, subpath,
7072 : scanjoin_target);
7073 211656 : lfirst(lc) = newpath;
7074 : }
7075 : }
7076 :
7077 : /* Likewise adjust the targets for any partial paths. */
7078 308046 : foreach(lc, rel->partial_pathlist)
7079 : {
7080 9638 : Path *subpath = (Path *) lfirst(lc);
7081 :
7082 : /* Shouldn't have any parameterized paths anymore */
7083 : Assert(subpath->param_info == NULL);
7084 :
7085 9638 : if (tlist_same_exprs)
7086 15460 : subpath->pathtarget->sortgrouprefs =
7087 7730 : scanjoin_target->sortgrouprefs;
7088 : else
7089 : {
7090 : Path *newpath;
7091 :
7092 1908 : newpath = (Path *) create_projection_path(root, rel, subpath,
7093 : scanjoin_target);
7094 1908 : lfirst(lc) = newpath;
7095 : }
7096 : }
7097 :
7098 : /*
7099 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7100 : * atop each existing path. (Note that this function doesn't look at the
7101 : * cheapest-path fields, which is a good thing because they're bogus right
7102 : * now.)
7103 : */
7104 298408 : if (root->parse->hasTargetSRFs)
7105 3578 : adjust_paths_for_srfs(root, rel,
7106 : scanjoin_targets,
7107 : scanjoin_targets_contain_srfs);
7108 :
7109 : /*
7110 : * Update the rel's target to be the final (with SRFs) scan/join target.
7111 : * This now matches the actual output of all the paths, and we might get
7112 : * confused in createplan.c if they don't agree. We must do this now so
7113 : * that any append paths made in the next part will use the correct
7114 : * pathtarget (cf. create_append_path).
7115 : */
7116 298408 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7117 :
7118 : /*
7119 : * If the relation is partitioned, recursively apply the scan/join target
7120 : * to all partitions, and generate brand-new Append paths in which the
7121 : * scan/join target is computed below the Append rather than above it.
7122 : * Since Append is not projection-capable, that might save a separate
7123 : * Result node, and it also is important for partitionwise aggregate.
7124 : */
7125 298408 : if (rel_is_partitioned)
7126 : {
7127 4370 : List *live_children = NIL;
7128 : int partition_idx;
7129 :
7130 : /* Adjust each partition. */
7131 17748 : for (partition_idx = 0; partition_idx < rel->nparts; partition_idx++)
7132 : {
7133 13378 : RelOptInfo *child_rel = rel->part_rels[partition_idx];
7134 : AppendRelInfo **appinfos;
7135 : int nappinfos;
7136 13378 : List *child_scanjoin_targets = NIL;
7137 : ListCell *lc;
7138 :
7139 : /* Pruned or dummy children can be ignored. */
7140 13378 : if (child_rel == NULL || IS_DUMMY_REL(child_rel))
7141 3546 : continue;
7142 :
7143 : /* Translate scan/join targets for this child. */
7144 9832 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7145 : &nappinfos);
7146 19664 : foreach(lc, scanjoin_targets)
7147 : {
7148 9832 : PathTarget *target = lfirst_node(PathTarget, lc);
7149 :
7150 9832 : target = copy_pathtarget(target);
7151 9832 : target->exprs = (List *)
7152 19664 : adjust_appendrel_attrs(root,
7153 9832 : (Node *) target->exprs,
7154 : nappinfos, appinfos);
7155 9832 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7156 : target);
7157 : }
7158 9832 : pfree(appinfos);
7159 :
7160 : /* Recursion does the real work. */
7161 9832 : apply_scanjoin_target_to_paths(root, child_rel,
7162 : child_scanjoin_targets,
7163 : scanjoin_targets_contain_srfs,
7164 : scanjoin_target_parallel_safe,
7165 : tlist_same_exprs);
7166 :
7167 : /* Save non-dummy children for Append paths. */
7168 9832 : if (!IS_DUMMY_REL(child_rel))
7169 9832 : live_children = lappend(live_children, child_rel);
7170 : }
7171 :
7172 : /* Build new paths for this relation by appending child paths. */
7173 4370 : add_paths_to_append_rel(root, rel, live_children);
7174 : }
7175 :
7176 : /*
7177 : * Consider generating Gather or Gather Merge paths. We must only do this
7178 : * if the relation is parallel safe, and we don't do it for child rels to
7179 : * avoid creating multiple Gather nodes within the same plan. We must do
7180 : * this after all paths have been generated and before set_cheapest, since
7181 : * one of the generated paths may turn out to be the cheapest one.
7182 : */
7183 298408 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7184 67840 : generate_gather_paths(root, rel, false);
7185 :
7186 : /*
7187 : * Reassess which paths are the cheapest, now that we've potentially added
7188 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7189 : * this relation.
7190 : */
7191 298408 : set_cheapest(rel);
7192 298408 : }
7193 :
7194 : /*
7195 : * create_partitionwise_grouping_paths
7196 : *
7197 : * If the partition keys of input relation are part of the GROUP BY clause, all
7198 : * the rows belonging to a given group come from a single partition. This
7199 : * allows aggregation/grouping over a partitioned relation to be broken down
7200 : * into aggregation/grouping on each partition. This should be no worse, and
7201 : * often better, than the normal approach.
7202 : *
7203 : * However, if the GROUP BY clause does not contain all the partition keys,
7204 : * rows from a given group may be spread across multiple partitions. In that
7205 : * case, we perform partial aggregation for each group, append the results,
7206 : * and then finalize aggregation. This is less certain to win than the
7207 : * previous case. It may win if the PartialAggregate stage greatly reduces
7208 : * the number of groups, because fewer rows will pass through the Append node.
7209 : * It may lose if we have lots of small groups.
7210 : */
7211 : static void
7212 338 : create_partitionwise_grouping_paths(PlannerInfo *root,
7213 : RelOptInfo *input_rel,
7214 : RelOptInfo *grouped_rel,
7215 : RelOptInfo *partially_grouped_rel,
7216 : const AggClauseCosts *agg_costs,
7217 : grouping_sets_data *gd,
7218 : PartitionwiseAggregateType patype,
7219 : GroupPathExtraData *extra)
7220 : {
7221 338 : int nparts = input_rel->nparts;
7222 : int cnt_parts;
7223 338 : List *grouped_live_children = NIL;
7224 338 : List *partially_grouped_live_children = NIL;
7225 338 : PathTarget *target = grouped_rel->reltarget;
7226 338 : bool partial_grouping_valid = true;
7227 :
7228 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7229 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7230 : partially_grouped_rel != NULL);
7231 :
7232 : /* Add paths for partitionwise aggregation/grouping. */
7233 1256 : for (cnt_parts = 0; cnt_parts < nparts; cnt_parts++)
7234 : {
7235 918 : RelOptInfo *child_input_rel = input_rel->part_rels[cnt_parts];
7236 918 : PathTarget *child_target = copy_pathtarget(target);
7237 : AppendRelInfo **appinfos;
7238 : int nappinfos;
7239 : GroupPathExtraData child_extra;
7240 : RelOptInfo *child_grouped_rel;
7241 : RelOptInfo *child_partially_grouped_rel;
7242 :
7243 : /* Pruned or dummy children can be ignored. */
7244 918 : if (child_input_rel == NULL || IS_DUMMY_REL(child_input_rel))
7245 0 : continue;
7246 :
7247 : /*
7248 : * Copy the given "extra" structure as is and then override the
7249 : * members specific to this child.
7250 : */
7251 918 : memcpy(&child_extra, extra, sizeof(child_extra));
7252 :
7253 918 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
7254 : &nappinfos);
7255 :
7256 918 : child_target->exprs = (List *)
7257 1836 : adjust_appendrel_attrs(root,
7258 918 : (Node *) target->exprs,
7259 : nappinfos, appinfos);
7260 :
7261 : /* Translate havingQual and targetList. */
7262 918 : child_extra.havingQual = (Node *)
7263 : adjust_appendrel_attrs(root,
7264 : extra->havingQual,
7265 : nappinfos, appinfos);
7266 918 : child_extra.targetList = (List *)
7267 1836 : adjust_appendrel_attrs(root,
7268 918 : (Node *) extra->targetList,
7269 : nappinfos, appinfos);
7270 :
7271 : /*
7272 : * extra->patype was the value computed for our parent rel; patype is
7273 : * the value for this relation. For the child, our value is its
7274 : * parent rel's value.
7275 : */
7276 918 : child_extra.patype = patype;
7277 :
7278 : /*
7279 : * Create grouping relation to hold fully aggregated grouping and/or
7280 : * aggregation paths for the child.
7281 : */
7282 1836 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
7283 : child_target,
7284 918 : extra->target_parallel_safe,
7285 : child_extra.havingQual);
7286 :
7287 : /* Create grouping paths for this child relation. */
7288 918 : create_ordinary_grouping_paths(root, child_input_rel,
7289 : child_grouped_rel,
7290 : agg_costs, gd, &child_extra,
7291 : &child_partially_grouped_rel);
7292 :
7293 918 : if (child_partially_grouped_rel)
7294 : {
7295 566 : partially_grouped_live_children =
7296 566 : lappend(partially_grouped_live_children,
7297 : child_partially_grouped_rel);
7298 : }
7299 : else
7300 352 : partial_grouping_valid = false;
7301 :
7302 918 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7303 : {
7304 544 : set_cheapest(child_grouped_rel);
7305 544 : grouped_live_children = lappend(grouped_live_children,
7306 : child_grouped_rel);
7307 : }
7308 :
7309 918 : pfree(appinfos);
7310 : }
7311 :
7312 : /*
7313 : * Try to create append paths for partially grouped children. For full
7314 : * partitionwise aggregation, we might have paths in the partial_pathlist
7315 : * if parallel aggregation is possible. For partial partitionwise
7316 : * aggregation, we may have paths in both pathlist and partial_pathlist.
7317 : *
7318 : * NB: We must have a partially grouped path for every child in order to
7319 : * generate a partially grouped path for this relation.
7320 : */
7321 338 : if (partially_grouped_rel && partial_grouping_valid)
7322 : {
7323 : Assert(partially_grouped_live_children != NIL);
7324 :
7325 218 : add_paths_to_append_rel(root, partially_grouped_rel,
7326 : partially_grouped_live_children);
7327 :
7328 : /*
7329 : * We need call set_cheapest, since the finalization step will use the
7330 : * cheapest path from the rel.
7331 : */
7332 218 : if (partially_grouped_rel->pathlist)
7333 218 : set_cheapest(partially_grouped_rel);
7334 : }
7335 :
7336 : /* If possible, create append paths for fully grouped children. */
7337 338 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7338 : {
7339 : Assert(grouped_live_children != NIL);
7340 :
7341 192 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
7342 : }
7343 338 : }
7344 :
7345 : /*
7346 : * group_by_has_partkey
7347 : *
7348 : * Returns true, if all the partition keys of the given relation are part of
7349 : * the GROUP BY clauses, false otherwise.
7350 : */
7351 : static bool
7352 334 : group_by_has_partkey(RelOptInfo *input_rel,
7353 : List *targetList,
7354 : List *groupClause)
7355 : {
7356 334 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
7357 334 : int cnt = 0;
7358 : int partnatts;
7359 :
7360 : /* Input relation should be partitioned. */
7361 : Assert(input_rel->part_scheme);
7362 :
7363 : /* Rule out early, if there are no partition keys present. */
7364 334 : if (!input_rel->partexprs)
7365 0 : return false;
7366 :
7367 334 : partnatts = input_rel->part_scheme->partnatts;
7368 :
7369 550 : for (cnt = 0; cnt < partnatts; cnt++)
7370 : {
7371 358 : List *partexprs = input_rel->partexprs[cnt];
7372 : ListCell *lc;
7373 358 : bool found = false;
7374 :
7375 504 : foreach(lc, partexprs)
7376 : {
7377 362 : Expr *partexpr = lfirst(lc);
7378 :
7379 362 : if (list_member(groupexprs, partexpr))
7380 : {
7381 216 : found = true;
7382 216 : break;
7383 : }
7384 : }
7385 :
7386 : /*
7387 : * If none of the partition key expressions match with any of the
7388 : * GROUP BY expression, return false.
7389 : */
7390 358 : if (!found)
7391 142 : return false;
7392 : }
7393 :
7394 192 : return true;
7395 : }
|
