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 "jit/jit.h"
35 : #include "lib/bipartite_match.h"
36 : #include "lib/knapsack.h"
37 : #include "miscadmin.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/parse_agg.h"
59 : #include "parser/parsetree.h"
60 : #include "partitioning/partdesc.h"
61 : #include "rewrite/rewriteManip.h"
62 : #include "storage/dsm_impl.h"
63 : #include "utils/lsyscache.h"
64 : #include "utils/rel.h"
65 : #include "utils/selfuncs.h"
66 : #include "utils/syscache.h"
67 :
68 : /* GUC parameters */
69 : double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
70 : int force_parallel_mode = FORCE_PARALLEL_OFF;
71 : bool parallel_leader_participation = true;
72 :
73 : /* Hook for plugins to get control in planner() */
74 : planner_hook_type planner_hook = NULL;
75 :
76 : /* Hook for plugins to get control when grouping_planner() plans upper rels */
77 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
78 :
79 :
80 : /* Expression kind codes for preprocess_expression */
81 : #define EXPRKIND_QUAL 0
82 : #define EXPRKIND_TARGET 1
83 : #define EXPRKIND_RTFUNC 2
84 : #define EXPRKIND_RTFUNC_LATERAL 3
85 : #define EXPRKIND_VALUES 4
86 : #define EXPRKIND_VALUES_LATERAL 5
87 : #define EXPRKIND_LIMIT 6
88 : #define EXPRKIND_APPINFO 7
89 : #define EXPRKIND_PHV 8
90 : #define EXPRKIND_TABLESAMPLE 9
91 : #define EXPRKIND_ARBITER_ELEM 10
92 : #define EXPRKIND_TABLEFUNC 11
93 : #define EXPRKIND_TABLEFUNC_LATERAL 12
94 :
95 : /* Passthrough data for standard_qp_callback */
96 : typedef struct
97 : {
98 : List *activeWindows; /* active windows, if any */
99 : List *groupClause; /* overrides parse->groupClause */
100 : } standard_qp_extra;
101 :
102 : /*
103 : * Data specific to grouping sets
104 : */
105 :
106 : typedef struct
107 : {
108 : List *rollups;
109 : List *hash_sets_idx;
110 : double dNumHashGroups;
111 : bool any_hashable;
112 : Bitmapset *unsortable_refs;
113 : Bitmapset *unhashable_refs;
114 : List *unsortable_sets;
115 : int *tleref_to_colnum_map;
116 : } grouping_sets_data;
117 :
118 : /*
119 : * Temporary structure for use during WindowClause reordering in order to be
120 : * able to sort WindowClauses on partitioning/ordering prefix.
121 : */
122 : typedef struct
123 : {
124 : WindowClause *wc;
125 : List *uniqueOrder; /* A List of unique ordering/partitioning
126 : * clauses per Window */
127 : } WindowClauseSortData;
128 :
129 : /* Local functions */
130 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
131 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
132 : static void inheritance_planner(PlannerInfo *root);
133 : static void grouping_planner(PlannerInfo *root, bool inheritance_update,
134 : double tuple_fraction);
135 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
136 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
137 : int *tleref_to_colnum_map);
138 : static void preprocess_rowmarks(PlannerInfo *root);
139 : static double preprocess_limit(PlannerInfo *root,
140 : double tuple_fraction,
141 : int64 *offset_est, int64 *count_est);
142 : static void remove_useless_groupby_columns(PlannerInfo *root);
143 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
144 : static List *extract_rollup_sets(List *groupingSets);
145 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
146 : static void standard_qp_callback(PlannerInfo *root, void *extra);
147 : static double get_number_of_groups(PlannerInfo *root,
148 : double path_rows,
149 : grouping_sets_data *gd,
150 : List *target_list);
151 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
152 : RelOptInfo *input_rel,
153 : PathTarget *target,
154 : bool target_parallel_safe,
155 : const AggClauseCosts *agg_costs,
156 : grouping_sets_data *gd);
157 : static bool is_degenerate_grouping(PlannerInfo *root);
158 : static void create_degenerate_grouping_paths(PlannerInfo *root,
159 : RelOptInfo *input_rel,
160 : RelOptInfo *grouped_rel);
161 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
162 : PathTarget *target, bool target_parallel_safe,
163 : Node *havingQual);
164 : static void create_ordinary_grouping_paths(PlannerInfo *root,
165 : RelOptInfo *input_rel,
166 : RelOptInfo *grouped_rel,
167 : const AggClauseCosts *agg_costs,
168 : grouping_sets_data *gd,
169 : GroupPathExtraData *extra,
170 : RelOptInfo **partially_grouped_rel_p);
171 : static void consider_groupingsets_paths(PlannerInfo *root,
172 : RelOptInfo *grouped_rel,
173 : Path *path,
174 : bool is_sorted,
175 : bool can_hash,
176 : grouping_sets_data *gd,
177 : const AggClauseCosts *agg_costs,
178 : double dNumGroups);
179 : static RelOptInfo *create_window_paths(PlannerInfo *root,
180 : RelOptInfo *input_rel,
181 : PathTarget *input_target,
182 : PathTarget *output_target,
183 : bool output_target_parallel_safe,
184 : WindowFuncLists *wflists,
185 : List *activeWindows);
186 : static void create_one_window_path(PlannerInfo *root,
187 : RelOptInfo *window_rel,
188 : Path *path,
189 : PathTarget *input_target,
190 : PathTarget *output_target,
191 : WindowFuncLists *wflists,
192 : List *activeWindows);
193 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
194 : RelOptInfo *input_rel);
195 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
196 : RelOptInfo *input_rel,
197 : PathTarget *target,
198 : bool target_parallel_safe,
199 : double limit_tuples);
200 : static PathTarget *make_group_input_target(PlannerInfo *root,
201 : PathTarget *final_target);
202 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
203 : PathTarget *grouping_target,
204 : Node *havingQual);
205 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
206 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
207 : static PathTarget *make_window_input_target(PlannerInfo *root,
208 : PathTarget *final_target,
209 : List *activeWindows);
210 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
211 : List *tlist);
212 : static PathTarget *make_sort_input_target(PlannerInfo *root,
213 : PathTarget *final_target,
214 : bool *have_postponed_srfs);
215 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
216 : List *targets, List *targets_contain_srfs);
217 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
218 : RelOptInfo *grouped_rel,
219 : RelOptInfo *partially_grouped_rel,
220 : const AggClauseCosts *agg_costs,
221 : grouping_sets_data *gd,
222 : double dNumGroups,
223 : GroupPathExtraData *extra);
224 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
225 : RelOptInfo *grouped_rel,
226 : RelOptInfo *input_rel,
227 : grouping_sets_data *gd,
228 : GroupPathExtraData *extra,
229 : bool force_rel_creation);
230 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
231 : static bool can_partial_agg(PlannerInfo *root,
232 : const AggClauseCosts *agg_costs);
233 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
234 : RelOptInfo *rel,
235 : List *scanjoin_targets,
236 : List *scanjoin_targets_contain_srfs,
237 : bool scanjoin_target_parallel_safe,
238 : bool tlist_same_exprs);
239 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
240 : RelOptInfo *input_rel,
241 : RelOptInfo *grouped_rel,
242 : RelOptInfo *partially_grouped_rel,
243 : const AggClauseCosts *agg_costs,
244 : grouping_sets_data *gd,
245 : PartitionwiseAggregateType patype,
246 : GroupPathExtraData *extra);
247 : static bool group_by_has_partkey(RelOptInfo *input_rel,
248 : List *targetList,
249 : List *groupClause);
250 : static int common_prefix_cmp(const void *a, const void *b);
251 :
252 :
253 : /*****************************************************************************
254 : *
255 : * Query optimizer entry point
256 : *
257 : * To support loadable plugins that monitor or modify planner behavior,
258 : * we provide a hook variable that lets a plugin get control before and
259 : * after the standard planning process. The plugin would normally call
260 : * standard_planner().
261 : *
262 : * Note to plugin authors: standard_planner() scribbles on its Query input,
263 : * so you'd better copy that data structure if you want to plan more than once.
264 : *
265 : *****************************************************************************/
266 : PlannedStmt *
267 247288 : planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
268 : {
269 : PlannedStmt *result;
270 :
271 247288 : if (planner_hook)
272 0 : result = (*planner_hook) (parse, cursorOptions, boundParams);
273 : else
274 247288 : result = standard_planner(parse, cursorOptions, boundParams);
275 245762 : return result;
276 : }
277 :
278 : PlannedStmt *
279 247288 : standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
280 : {
281 : PlannedStmt *result;
282 : PlannerGlobal *glob;
283 : double tuple_fraction;
284 : PlannerInfo *root;
285 : RelOptInfo *final_rel;
286 : Path *best_path;
287 : Plan *top_plan;
288 : ListCell *lp,
289 : *lr;
290 :
291 : /*
292 : * Set up global state for this planner invocation. This data is needed
293 : * across all levels of sub-Query that might exist in the given command,
294 : * so we keep it in a separate struct that's linked to by each per-Query
295 : * PlannerInfo.
296 : */
297 247288 : glob = makeNode(PlannerGlobal);
298 :
299 247288 : glob->boundParams = boundParams;
300 247288 : glob->subplans = NIL;
301 247288 : glob->subroots = NIL;
302 247288 : glob->rewindPlanIDs = NULL;
303 247288 : glob->finalrtable = NIL;
304 247288 : glob->finalrowmarks = NIL;
305 247288 : glob->resultRelations = NIL;
306 247288 : glob->rootResultRelations = NIL;
307 247288 : glob->relationOids = NIL;
308 247288 : glob->invalItems = NIL;
309 247288 : glob->paramExecTypes = NIL;
310 247288 : glob->lastPHId = 0;
311 247288 : glob->lastRowMarkId = 0;
312 247288 : glob->lastPlanNodeId = 0;
313 247288 : glob->transientPlan = false;
314 247288 : glob->dependsOnRole = false;
315 :
316 : /*
317 : * Assess whether it's feasible to use parallel mode for this query. We
318 : * can't do this in a standalone backend, or if the command will try to
319 : * modify any data, or if this is a cursor operation, or if GUCs are set
320 : * to values that don't permit parallelism, or if parallel-unsafe
321 : * functions are present in the query tree.
322 : *
323 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
324 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
325 : * the command is writing into a completely new table which workers won't
326 : * be able to see. If the workers could see the table, the fact that
327 : * group locking would cause them to ignore the leader's heavyweight
328 : * relation extension lock and GIN page locks would make this unsafe.
329 : * We'll have to fix that somehow if we want to allow parallel inserts in
330 : * general; updates and deletes have additional problems especially around
331 : * combo CIDs.)
332 : *
333 : * For now, we don't try to use parallel mode if we're running inside a
334 : * parallel worker. We might eventually be able to relax this
335 : * restriction, but for now it seems best not to have parallel workers
336 : * trying to create their own parallel workers.
337 : */
338 247288 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
339 208104 : IsUnderPostmaster &&
340 374890 : parse->commandType == CMD_SELECT &&
341 333488 : !parse->hasModifyingCTE &&
342 333098 : max_parallel_workers_per_gather > 0 &&
343 166396 : !IsParallelWorker())
344 : {
345 : /* all the cheap tests pass, so scan the query tree */
346 166364 : glob->maxParallelHazard = max_parallel_hazard(parse);
347 166364 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
348 : }
349 : else
350 : {
351 : /* skip the query tree scan, just assume it's unsafe */
352 80924 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
353 80924 : glob->parallelModeOK = false;
354 : }
355 :
356 : /*
357 : * glob->parallelModeNeeded is normally set to false here and changed to
358 : * true during plan creation if a Gather or Gather Merge plan is actually
359 : * created (cf. create_gather_plan, create_gather_merge_plan).
360 : *
361 : * However, if force_parallel_mode = on or force_parallel_mode = regress,
362 : * then we impose parallel mode whenever it's safe to do so, even if the
363 : * final plan doesn't use parallelism. It's not safe to do so if the
364 : * query contains anything parallel-unsafe; parallelModeOK will be false
365 : * in that case. Note that parallelModeOK can't change after this point.
366 : * Otherwise, everything in the query is either parallel-safe or
367 : * parallel-restricted, and in either case it should be OK to impose
368 : * parallel-mode restrictions. If that ends up breaking something, then
369 : * either some function the user included in the query is incorrectly
370 : * labelled as parallel-safe or parallel-restricted when in reality it's
371 : * parallel-unsafe, or else the query planner itself has a bug.
372 : */
373 392298 : glob->parallelModeNeeded = glob->parallelModeOK &&
374 145010 : (force_parallel_mode != FORCE_PARALLEL_OFF);
375 :
376 : /* Determine what fraction of the plan is likely to be scanned */
377 247288 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
378 : {
379 : /*
380 : * We have no real idea how many tuples the user will ultimately FETCH
381 : * from a cursor, but it is often the case that he doesn't want 'em
382 : * all, or would prefer a fast-start plan anyway so that he can
383 : * process some of the tuples sooner. Use a GUC parameter to decide
384 : * what fraction to optimize for.
385 : */
386 2064 : tuple_fraction = cursor_tuple_fraction;
387 :
388 : /*
389 : * We document cursor_tuple_fraction as simply being a fraction, which
390 : * means the edge cases 0 and 1 have to be treated specially here. We
391 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
392 : */
393 2064 : if (tuple_fraction >= 1.0)
394 0 : tuple_fraction = 0.0;
395 2064 : else if (tuple_fraction <= 0.0)
396 0 : tuple_fraction = 1e-10;
397 : }
398 : else
399 : {
400 : /* Default assumption is we need all the tuples */
401 245224 : tuple_fraction = 0.0;
402 : }
403 :
404 : /* primary planning entry point (may recurse for subqueries) */
405 247288 : root = subquery_planner(glob, parse, NULL,
406 : false, tuple_fraction);
407 :
408 : /* Select best Path and turn it into a Plan */
409 245882 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
410 245882 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
411 :
412 245882 : top_plan = create_plan(root, best_path);
413 :
414 : /*
415 : * If creating a plan for a scrollable cursor, make sure it can run
416 : * backwards on demand. Add a Material node at the top at need.
417 : */
418 245762 : if (cursorOptions & CURSOR_OPT_SCROLL)
419 : {
420 164 : if (!ExecSupportsBackwardScan(top_plan))
421 16 : top_plan = materialize_finished_plan(top_plan);
422 : }
423 :
424 : /*
425 : * Optionally add a Gather node for testing purposes, provided this is
426 : * actually a safe thing to do.
427 : */
428 245762 : if (force_parallel_mode != FORCE_PARALLEL_OFF && top_plan->parallel_safe)
429 : {
430 18 : Gather *gather = makeNode(Gather);
431 :
432 : /*
433 : * If there are any initPlans attached to the formerly-top plan node,
434 : * move them up to the Gather node; same as we do for Material node in
435 : * materialize_finished_plan.
436 : */
437 18 : gather->plan.initPlan = top_plan->initPlan;
438 18 : top_plan->initPlan = NIL;
439 :
440 18 : gather->plan.targetlist = top_plan->targetlist;
441 18 : gather->plan.qual = NIL;
442 18 : gather->plan.lefttree = top_plan;
443 18 : gather->plan.righttree = NULL;
444 18 : gather->num_workers = 1;
445 18 : gather->single_copy = true;
446 18 : gather->invisible = (force_parallel_mode == FORCE_PARALLEL_REGRESS);
447 :
448 : /*
449 : * Since this Gather has no parallel-aware descendants to signal to,
450 : * we don't need a rescan Param.
451 : */
452 18 : gather->rescan_param = -1;
453 :
454 : /*
455 : * Ideally we'd use cost_gather here, but setting up dummy path data
456 : * to satisfy it doesn't seem much cleaner than knowing what it does.
457 : */
458 18 : gather->plan.startup_cost = top_plan->startup_cost +
459 : parallel_setup_cost;
460 36 : gather->plan.total_cost = top_plan->total_cost +
461 18 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
462 18 : gather->plan.plan_rows = top_plan->plan_rows;
463 18 : gather->plan.plan_width = top_plan->plan_width;
464 18 : gather->plan.parallel_aware = false;
465 18 : gather->plan.parallel_safe = false;
466 :
467 : /* use parallel mode for parallel plans. */
468 18 : root->glob->parallelModeNeeded = true;
469 :
470 18 : top_plan = &gather->plan;
471 : }
472 :
473 : /*
474 : * If any Params were generated, run through the plan tree and compute
475 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
476 : * set_plan_references' tree traversal, but for now it has to be separate
477 : * because we need to visit subplans before not after main plan.
478 : */
479 245762 : if (glob->paramExecTypes != NIL)
480 : {
481 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
482 154246 : forboth(lp, glob->subplans, lr, glob->subroots)
483 : {
484 46230 : Plan *subplan = (Plan *) lfirst(lp);
485 46230 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
486 :
487 46230 : SS_finalize_plan(subroot, subplan);
488 : }
489 108016 : SS_finalize_plan(root, top_plan);
490 : }
491 :
492 : /* final cleanup of the plan */
493 : Assert(glob->finalrtable == NIL);
494 : Assert(glob->finalrowmarks == NIL);
495 : Assert(glob->resultRelations == NIL);
496 : Assert(glob->rootResultRelations == NIL);
497 245762 : top_plan = set_plan_references(root, top_plan);
498 : /* ... and the subplans (both regular subplans and initplans) */
499 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
500 291992 : forboth(lp, glob->subplans, lr, glob->subroots)
501 : {
502 46230 : Plan *subplan = (Plan *) lfirst(lp);
503 46230 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
504 :
505 46230 : lfirst(lp) = set_plan_references(subroot, subplan);
506 : }
507 :
508 : /* build the PlannedStmt result */
509 245762 : result = makeNode(PlannedStmt);
510 :
511 245762 : result->commandType = parse->commandType;
512 245762 : result->queryId = parse->queryId;
513 245762 : result->hasReturning = (parse->returningList != NIL);
514 245762 : result->hasModifyingCTE = parse->hasModifyingCTE;
515 245762 : result->canSetTag = parse->canSetTag;
516 245762 : result->transientPlan = glob->transientPlan;
517 245762 : result->dependsOnRole = glob->dependsOnRole;
518 245762 : result->parallelModeNeeded = glob->parallelModeNeeded;
519 245762 : result->planTree = top_plan;
520 245762 : result->rtable = glob->finalrtable;
521 245762 : result->resultRelations = glob->resultRelations;
522 245762 : result->rootResultRelations = glob->rootResultRelations;
523 245762 : result->subplans = glob->subplans;
524 245762 : result->rewindPlanIDs = glob->rewindPlanIDs;
525 245762 : result->rowMarks = glob->finalrowmarks;
526 245762 : result->relationOids = glob->relationOids;
527 245762 : result->invalItems = glob->invalItems;
528 245762 : result->paramExecTypes = glob->paramExecTypes;
529 : /* utilityStmt should be null, but we might as well copy it */
530 245762 : result->utilityStmt = parse->utilityStmt;
531 245762 : result->stmt_location = parse->stmt_location;
532 245762 : result->stmt_len = parse->stmt_len;
533 :
534 245762 : result->jitFlags = PGJIT_NONE;
535 491524 : if (jit_enabled && jit_above_cost >= 0 &&
536 245762 : top_plan->total_cost > jit_above_cost)
537 : {
538 530 : result->jitFlags |= PGJIT_PERFORM;
539 :
540 : /*
541 : * Decide how much effort should be put into generating better code.
542 : */
543 1060 : if (jit_optimize_above_cost >= 0 &&
544 530 : top_plan->total_cost > jit_optimize_above_cost)
545 442 : result->jitFlags |= PGJIT_OPT3;
546 1060 : if (jit_inline_above_cost >= 0 &&
547 530 : top_plan->total_cost > jit_inline_above_cost)
548 442 : result->jitFlags |= PGJIT_INLINE;
549 :
550 : /*
551 : * Decide which operations should be JITed.
552 : */
553 530 : if (jit_expressions)
554 530 : result->jitFlags |= PGJIT_EXPR;
555 530 : if (jit_tuple_deforming)
556 530 : result->jitFlags |= PGJIT_DEFORM;
557 : }
558 :
559 245762 : if (glob->partition_directory != NULL)
560 5924 : DestroyPartitionDirectory(glob->partition_directory);
561 :
562 245762 : return result;
563 : }
564 :
565 :
566 : /*--------------------
567 : * subquery_planner
568 : * Invokes the planner on a subquery. We recurse to here for each
569 : * sub-SELECT found in the query tree.
570 : *
571 : * glob is the global state for the current planner run.
572 : * parse is the querytree produced by the parser & rewriter.
573 : * parent_root is the immediate parent Query's info (NULL at the top level).
574 : * hasRecursion is true if this is a recursive WITH query.
575 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
576 : * tuple_fraction is interpreted as explained for grouping_planner, below.
577 : *
578 : * Basically, this routine does the stuff that should only be done once
579 : * per Query object. It then calls grouping_planner. At one time,
580 : * grouping_planner could be invoked recursively on the same Query object;
581 : * that's not currently true, but we keep the separation between the two
582 : * routines anyway, in case we need it again someday.
583 : *
584 : * subquery_planner will be called recursively to handle sub-Query nodes
585 : * found within the query's expressions and rangetable.
586 : *
587 : * Returns the PlannerInfo struct ("root") that contains all data generated
588 : * while planning the subquery. In particular, the Path(s) attached to
589 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
590 : * cheapest way(s) to implement the query. The top level will select the
591 : * best Path and pass it through createplan.c to produce a finished Plan.
592 : *--------------------
593 : */
594 : PlannerInfo *
595 301564 : subquery_planner(PlannerGlobal *glob, Query *parse,
596 : PlannerInfo *parent_root,
597 : bool hasRecursion, double tuple_fraction)
598 : {
599 : PlannerInfo *root;
600 : List *newWithCheckOptions;
601 : List *newHaving;
602 : bool hasOuterJoins;
603 : bool hasResultRTEs;
604 : RelOptInfo *final_rel;
605 : ListCell *l;
606 :
607 : /* Create a PlannerInfo data structure for this subquery */
608 301564 : root = makeNode(PlannerInfo);
609 301564 : root->parse = parse;
610 301564 : root->glob = glob;
611 301564 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
612 301564 : root->parent_root = parent_root;
613 301564 : root->plan_params = NIL;
614 301564 : root->outer_params = NULL;
615 301564 : root->planner_cxt = CurrentMemoryContext;
616 301564 : root->init_plans = NIL;
617 301564 : root->cte_plan_ids = NIL;
618 301564 : root->multiexpr_params = NIL;
619 301564 : root->eq_classes = NIL;
620 301564 : root->ec_merging_done = false;
621 301564 : root->append_rel_list = NIL;
622 301564 : root->rowMarks = NIL;
623 301564 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
624 301564 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
625 301564 : root->processed_tlist = NIL;
626 301564 : root->grouping_map = NULL;
627 301564 : root->minmax_aggs = NIL;
628 301564 : root->qual_security_level = 0;
629 301564 : root->inhTargetKind = INHKIND_NONE;
630 301564 : root->hasRecursion = hasRecursion;
631 301564 : if (hasRecursion)
632 320 : root->wt_param_id = assign_special_exec_param(root);
633 : else
634 301244 : root->wt_param_id = -1;
635 301564 : root->non_recursive_path = NULL;
636 301564 : 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 301564 : if (parse->cteList)
643 1170 : 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 301564 : 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 301564 : if (parse->hasSubLinks)
658 29790 : pull_up_sublinks(root);
659 :
660 : /*
661 : * Scan the rangetable for function RTEs, do const-simplification on them,
662 : * and then inline them if possible (producing subqueries that might get
663 : * pulled up next). Recursion issues here are handled in the same way as
664 : * for SubLinks.
665 : */
666 301564 : preprocess_function_rtes(root);
667 :
668 : /*
669 : * Check to see if any subqueries in the jointree can be merged into this
670 : * query.
671 : */
672 301564 : pull_up_subqueries(root);
673 :
674 : /*
675 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
676 : * do this now because it requires applying pull_up_subqueries to the leaf
677 : * queries of the UNION ALL, which weren't touched above because they
678 : * weren't referenced by the jointree (they will be after we do this).
679 : */
680 301560 : if (parse->setOperations)
681 2498 : flatten_simple_union_all(root);
682 :
683 : /*
684 : * Survey the rangetable to see what kinds of entries are present. We can
685 : * skip some later processing if relevant SQL features are not used; for
686 : * example if there are no JOIN RTEs we can avoid the expense of doing
687 : * flatten_join_alias_vars(). This must be done after we have finished
688 : * adding rangetable entries, of course. (Note: actually, processing of
689 : * inherited or partitioned rels can cause RTEs for their child tables to
690 : * get added later; but those must all be RTE_RELATION entries, so they
691 : * don't invalidate the conclusions drawn here.)
692 : */
693 301560 : root->hasJoinRTEs = false;
694 301560 : root->hasLateralRTEs = false;
695 301560 : hasOuterJoins = false;
696 301560 : hasResultRTEs = false;
697 855858 : foreach(l, parse->rtable)
698 : {
699 554298 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
700 :
701 554298 : switch (rte->rtekind)
702 : {
703 : case RTE_RELATION:
704 309892 : if (rte->inh)
705 : {
706 : /*
707 : * Check to see if the relation actually has any children;
708 : * if not, clear the inh flag so we can treat it as a
709 : * plain base relation.
710 : *
711 : * Note: this could give a false-positive result, if the
712 : * rel once had children but no longer does. We used to
713 : * be able to clear rte->inh later on when we discovered
714 : * that, but no more; we have to handle such cases as
715 : * full-fledged inheritance.
716 : */
717 216732 : rte->inh = has_subclass(rte->relid);
718 : }
719 309892 : break;
720 : case RTE_JOIN:
721 48068 : root->hasJoinRTEs = true;
722 48068 : if (IS_OUTER_JOIN(rte->jointype))
723 35396 : hasOuterJoins = true;
724 48068 : break;
725 : case RTE_RESULT:
726 114234 : hasResultRTEs = true;
727 114234 : break;
728 : default:
729 : /* No work here for other RTE types */
730 82104 : break;
731 : }
732 :
733 554298 : if (rte->lateral)
734 1014 : root->hasLateralRTEs = true;
735 :
736 : /*
737 : * We can also determine the maximum security level required for any
738 : * securityQuals now. Addition of inheritance-child RTEs won't affect
739 : * this, because child tables don't have their own securityQuals; see
740 : * expand_single_inheritance_child().
741 : */
742 554298 : if (rte->securityQuals)
743 1152 : root->qual_security_level = Max(root->qual_security_level,
744 : list_length(rte->securityQuals));
745 : }
746 :
747 : /*
748 : * Preprocess RowMark information. We need to do this after subquery
749 : * pullup, so that all base relations are present.
750 : */
751 301560 : preprocess_rowmarks(root);
752 :
753 : /*
754 : * Set hasHavingQual to remember if HAVING clause is present. Needed
755 : * because preprocess_expression will reduce a constant-true condition to
756 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
757 : */
758 301560 : root->hasHavingQual = (parse->havingQual != NULL);
759 :
760 : /* Clear this flag; might get set in distribute_qual_to_rels */
761 301560 : root->hasPseudoConstantQuals = false;
762 :
763 : /*
764 : * Do expression preprocessing on targetlist and quals, as well as other
765 : * random expressions in the querytree. Note that we do not need to
766 : * handle sort/group expressions explicitly, because they are actually
767 : * part of the targetlist.
768 : */
769 300176 : parse->targetList = (List *)
770 301560 : preprocess_expression(root, (Node *) parse->targetList,
771 : EXPRKIND_TARGET);
772 :
773 : /* Constant-folding might have removed all set-returning functions */
774 300176 : if (parse->hasTargetSRFs)
775 3682 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
776 :
777 300176 : newWithCheckOptions = NIL;
778 301330 : foreach(l, parse->withCheckOptions)
779 : {
780 1154 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
781 :
782 1154 : wco->qual = preprocess_expression(root, wco->qual,
783 : EXPRKIND_QUAL);
784 1154 : if (wco->qual != NULL)
785 934 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
786 : }
787 300176 : parse->withCheckOptions = newWithCheckOptions;
788 :
789 300176 : parse->returningList = (List *)
790 300176 : preprocess_expression(root, (Node *) parse->returningList,
791 : EXPRKIND_TARGET);
792 :
793 300176 : preprocess_qual_conditions(root, (Node *) parse->jointree);
794 :
795 300176 : parse->havingQual = preprocess_expression(root, parse->havingQual,
796 : EXPRKIND_QUAL);
797 :
798 301360 : foreach(l, parse->windowClause)
799 : {
800 1184 : WindowClause *wc = lfirst_node(WindowClause, l);
801 :
802 : /* partitionClause/orderClause are sort/group expressions */
803 1184 : wc->startOffset = preprocess_expression(root, wc->startOffset,
804 : EXPRKIND_LIMIT);
805 1184 : wc->endOffset = preprocess_expression(root, wc->endOffset,
806 : EXPRKIND_LIMIT);
807 : }
808 :
809 300176 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
810 : EXPRKIND_LIMIT);
811 300176 : parse->limitCount = preprocess_expression(root, parse->limitCount,
812 : EXPRKIND_LIMIT);
813 :
814 300176 : if (parse->onConflict)
815 : {
816 1960 : parse->onConflict->arbiterElems = (List *)
817 980 : preprocess_expression(root,
818 980 : (Node *) parse->onConflict->arbiterElems,
819 : EXPRKIND_ARBITER_ELEM);
820 1960 : parse->onConflict->arbiterWhere =
821 980 : preprocess_expression(root,
822 980 : parse->onConflict->arbiterWhere,
823 : EXPRKIND_QUAL);
824 1960 : parse->onConflict->onConflictSet = (List *)
825 980 : preprocess_expression(root,
826 980 : (Node *) parse->onConflict->onConflictSet,
827 : EXPRKIND_TARGET);
828 1960 : parse->onConflict->onConflictWhere =
829 980 : preprocess_expression(root,
830 980 : parse->onConflict->onConflictWhere,
831 : EXPRKIND_QUAL);
832 : /* exclRelTlist contains only Vars, so no preprocessing needed */
833 : }
834 :
835 300176 : root->append_rel_list = (List *)
836 300176 : preprocess_expression(root, (Node *) root->append_rel_list,
837 : EXPRKIND_APPINFO);
838 :
839 : /* Also need to preprocess expressions within RTEs */
840 852964 : foreach(l, parse->rtable)
841 : {
842 552788 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
843 : int kind;
844 : ListCell *lcsq;
845 :
846 552788 : if (rte->rtekind == RTE_RELATION)
847 : {
848 309762 : if (rte->tablesample)
849 152 : rte->tablesample = (TableSampleClause *)
850 152 : preprocess_expression(root,
851 152 : (Node *) rte->tablesample,
852 : EXPRKIND_TABLESAMPLE);
853 : }
854 243026 : else if (rte->rtekind == RTE_SUBQUERY)
855 : {
856 : /*
857 : * We don't want to do all preprocessing yet on the subquery's
858 : * expressions, since that will happen when we plan it. But if it
859 : * contains any join aliases of our level, those have to get
860 : * expanded now, because planning of the subquery won't do it.
861 : * That's only possible if the subquery is LATERAL.
862 : */
863 48102 : if (rte->lateral && root->hasJoinRTEs)
864 316 : rte->subquery = (Query *)
865 316 : flatten_join_alias_vars(root->parse,
866 316 : (Node *) rte->subquery);
867 : }
868 194924 : else if (rte->rtekind == RTE_FUNCTION)
869 : {
870 : /* Preprocess the function expression(s) fully */
871 28658 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
872 28658 : rte->functions = (List *)
873 28658 : preprocess_expression(root, (Node *) rte->functions, kind);
874 : }
875 166266 : else if (rte->rtekind == RTE_TABLEFUNC)
876 : {
877 : /* Preprocess the function expression(s) fully */
878 144 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
879 144 : rte->tablefunc = (TableFunc *)
880 144 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
881 : }
882 166122 : else if (rte->rtekind == RTE_VALUES)
883 : {
884 : /* Preprocess the values lists fully */
885 3784 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
886 3784 : rte->values_lists = (List *)
887 3784 : preprocess_expression(root, (Node *) rte->values_lists, kind);
888 : }
889 :
890 : /*
891 : * Process each element of the securityQuals list as if it were a
892 : * separate qual expression (as indeed it is). We need to do it this
893 : * way to get proper canonicalization of AND/OR structure. Note that
894 : * this converts each element into an implicit-AND sublist.
895 : */
896 554096 : foreach(lcsq, rte->securityQuals)
897 : {
898 1308 : lfirst(lcsq) = preprocess_expression(root,
899 1308 : (Node *) lfirst(lcsq),
900 : EXPRKIND_QUAL);
901 : }
902 : }
903 :
904 : /*
905 : * Now that we are done preprocessing expressions, and in particular done
906 : * flattening join alias variables, get rid of the joinaliasvars lists.
907 : * They no longer match what expressions in the rest of the tree look
908 : * like, because we have not preprocessed expressions in those lists (and
909 : * do not want to; for example, expanding a SubLink there would result in
910 : * a useless unreferenced subplan). Leaving them in place simply creates
911 : * a hazard for later scans of the tree. We could try to prevent that by
912 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
913 : * but that doesn't sound very reliable.
914 : */
915 300176 : if (root->hasJoinRTEs)
916 : {
917 208030 : foreach(l, parse->rtable)
918 : {
919 172102 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
920 :
921 172102 : rte->joinaliasvars = NIL;
922 : }
923 : }
924 :
925 : /*
926 : * In some cases we may want to transfer a HAVING clause into WHERE. We
927 : * cannot do so if the HAVING clause contains aggregates (obviously) or
928 : * volatile functions (since a HAVING clause is supposed to be executed
929 : * only once per group). We also can't do this if there are any nonempty
930 : * grouping sets; moving such a clause into WHERE would potentially change
931 : * the results, if any referenced column isn't present in all the grouping
932 : * sets. (If there are only empty grouping sets, then the HAVING clause
933 : * must be degenerate as discussed below.)
934 : *
935 : * Also, it may be that the clause is so expensive to execute that we're
936 : * better off doing it only once per group, despite the loss of
937 : * selectivity. This is hard to estimate short of doing the entire
938 : * planning process twice, so we use a heuristic: clauses containing
939 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
940 : * clause into WHERE, in hopes of eliminating tuples before aggregation
941 : * instead of after.
942 : *
943 : * If the query has explicit grouping then we can simply move such a
944 : * clause into WHERE; any group that fails the clause will not be in the
945 : * output because none of its tuples will reach the grouping or
946 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
947 : * HAVING clause, which we put in WHERE so that query_planner() can use it
948 : * in a gating Result node, but also keep in HAVING to ensure that we
949 : * don't emit a bogus aggregated row. (This could be done better, but it
950 : * seems not worth optimizing.)
951 : *
952 : * Note that both havingQual and parse->jointree->quals are in
953 : * implicitly-ANDed-list form at this point, even though they are declared
954 : * as Node *.
955 : */
956 300176 : newHaving = NIL;
957 300630 : foreach(l, (List *) parse->havingQual)
958 : {
959 454 : Node *havingclause = (Node *) lfirst(l);
960 :
961 880 : if ((parse->groupClause && parse->groupingSets) ||
962 542 : contain_agg_clause(havingclause) ||
963 232 : contain_volatile_functions(havingclause) ||
964 116 : contain_subplans(havingclause))
965 : {
966 : /* keep it in HAVING */
967 338 : newHaving = lappend(newHaving, havingclause);
968 : }
969 116 : else if (parse->groupClause && !parse->groupingSets)
970 : {
971 : /* move it to WHERE */
972 208 : parse->jointree->quals = (Node *)
973 104 : lappend((List *) parse->jointree->quals, havingclause);
974 : }
975 : else
976 : {
977 : /* put a copy in WHERE, keep it in HAVING */
978 24 : parse->jointree->quals = (Node *)
979 12 : lappend((List *) parse->jointree->quals,
980 : copyObject(havingclause));
981 12 : newHaving = lappend(newHaving, havingclause);
982 : }
983 : }
984 300176 : parse->havingQual = (Node *) newHaving;
985 :
986 : /* Remove any redundant GROUP BY columns */
987 300176 : remove_useless_groupby_columns(root);
988 :
989 : /*
990 : * If we have any outer joins, try to reduce them to plain inner joins.
991 : * This step is most easily done after we've done expression
992 : * preprocessing.
993 : */
994 300176 : if (hasOuterJoins)
995 30202 : reduce_outer_joins(root);
996 :
997 : /*
998 : * If we have any RTE_RESULT relations, see if they can be deleted from
999 : * the jointree. This step is most effectively done after we've done
1000 : * expression preprocessing and outer join reduction.
1001 : */
1002 300176 : if (hasResultRTEs)
1003 112658 : remove_useless_result_rtes(root);
1004 :
1005 : /*
1006 : * Do the main planning. If we have an inherited target relation, that
1007 : * needs special processing, else go straight to grouping_planner.
1008 : */
1009 369568 : if (parse->resultRelation &&
1010 69392 : rt_fetch(parse->resultRelation, parse->rtable)->inh)
1011 1224 : inheritance_planner(root);
1012 : else
1013 298952 : grouping_planner(root, false, tuple_fraction);
1014 :
1015 : /*
1016 : * Capture the set of outer-level param IDs we have access to, for use in
1017 : * extParam/allParam calculations later.
1018 : */
1019 300158 : SS_identify_outer_params(root);
1020 :
1021 : /*
1022 : * If any initPlans were created in this query level, adjust the surviving
1023 : * Paths' costs and parallel-safety flags to account for them. The
1024 : * initPlans won't actually get attached to the plan tree till
1025 : * create_plan() runs, but we must include their effects now.
1026 : */
1027 300158 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1028 300158 : SS_charge_for_initplans(root, final_rel);
1029 :
1030 : /*
1031 : * Make sure we've identified the cheapest Path for the final rel. (By
1032 : * doing this here not in grouping_planner, we include initPlan costs in
1033 : * the decision, though it's unlikely that will change anything.)
1034 : */
1035 300158 : set_cheapest(final_rel);
1036 :
1037 300158 : return root;
1038 : }
1039 :
1040 : /*
1041 : * preprocess_expression
1042 : * Do subquery_planner's preprocessing work for an expression,
1043 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1044 : * conditions), a HAVING clause, or a few other things.
1045 : */
1046 : static Node *
1047 2214014 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1048 : {
1049 : /*
1050 : * Fall out quickly if expression is empty. This occurs often enough to
1051 : * be worth checking. Note that null->null is the correct conversion for
1052 : * implicit-AND result format, too.
1053 : */
1054 2214014 : if (expr == NULL)
1055 1689252 : return NULL;
1056 :
1057 : /*
1058 : * If the query has any join RTEs, replace join alias variables with
1059 : * base-relation variables. We must do this first, since any expressions
1060 : * we may extract from the joinaliasvars lists have not been preprocessed.
1061 : * For example, if we did this after sublink processing, sublinks expanded
1062 : * out from join aliases would not get processed. But we can skip this in
1063 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1064 : * they can't contain any Vars of the current query level.
1065 : */
1066 524762 : if (root->hasJoinRTEs &&
1067 234528 : !(kind == EXPRKIND_RTFUNC ||
1068 117174 : kind == EXPRKIND_VALUES ||
1069 : kind == EXPRKIND_TABLESAMPLE ||
1070 : kind == EXPRKIND_TABLEFUNC))
1071 117174 : expr = flatten_join_alias_vars(root->parse, expr);
1072 :
1073 : /*
1074 : * Simplify constant expressions. For function RTEs, this was already
1075 : * done by preprocess_function_rtes ... but we have to do it again if the
1076 : * RTE is LATERAL and might have contained join alias variables.
1077 : *
1078 : * Note: an essential effect of this is to convert named-argument function
1079 : * calls to positional notation and insert the current actual values of
1080 : * any default arguments for functions. To ensure that happens, we *must*
1081 : * process all expressions here. Previous PG versions sometimes skipped
1082 : * const-simplification if it didn't seem worth the trouble, but we can't
1083 : * do that anymore.
1084 : *
1085 : * Note: this also flattens nested AND and OR expressions into N-argument
1086 : * form. All processing of a qual expression after this point must be
1087 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1088 : * with AND directly under AND, nor OR directly under OR.
1089 : */
1090 524762 : if (!(kind == EXPRKIND_RTFUNC ||
1091 346 : (kind == EXPRKIND_RTFUNC_LATERAL && !root->hasJoinRTEs)))
1092 496214 : expr = eval_const_expressions(root, expr);
1093 :
1094 : /*
1095 : * If it's a qual or havingQual, canonicalize it.
1096 : */
1097 523378 : if (kind == EXPRKIND_QUAL)
1098 : {
1099 185890 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1100 :
1101 : #ifdef OPTIMIZER_DEBUG
1102 : printf("After canonicalize_qual()\n");
1103 : pprint(expr);
1104 : #endif
1105 : }
1106 :
1107 : /* Expand SubLinks to SubPlans */
1108 523378 : if (root->parse->hasSubLinks)
1109 104108 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1110 :
1111 : /*
1112 : * XXX do not insert anything here unless you have grokked the comments in
1113 : * SS_replace_correlation_vars ...
1114 : */
1115 :
1116 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1117 523378 : if (root->query_level > 1)
1118 128076 : expr = SS_replace_correlation_vars(root, expr);
1119 :
1120 : /*
1121 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1122 : * don't want to do this before eval_const_expressions, since the latter
1123 : * would be unable to simplify a top-level AND correctly. Also,
1124 : * SS_process_sublinks expects explicit-AND format.)
1125 : */
1126 523378 : if (kind == EXPRKIND_QUAL)
1127 185890 : expr = (Node *) make_ands_implicit((Expr *) expr);
1128 :
1129 523378 : return expr;
1130 : }
1131 :
1132 : /*
1133 : * preprocess_qual_conditions
1134 : * Recursively scan the query's jointree and do subquery_planner's
1135 : * preprocessing work on each qual condition found therein.
1136 : */
1137 : static void
1138 740764 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1139 : {
1140 740764 : if (jtnode == NULL)
1141 0 : return;
1142 740764 : if (IsA(jtnode, RangeTblRef))
1143 : {
1144 : /* nothing to do here */
1145 : }
1146 370062 : else if (IsA(jtnode, FromExpr))
1147 : {
1148 313806 : FromExpr *f = (FromExpr *) jtnode;
1149 : ListCell *l;
1150 :
1151 641882 : foreach(l, f->fromlist)
1152 328076 : preprocess_qual_conditions(root, lfirst(l));
1153 :
1154 313806 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1155 : }
1156 56256 : else if (IsA(jtnode, JoinExpr))
1157 : {
1158 56256 : JoinExpr *j = (JoinExpr *) jtnode;
1159 :
1160 56256 : preprocess_qual_conditions(root, j->larg);
1161 56256 : preprocess_qual_conditions(root, j->rarg);
1162 :
1163 56256 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1164 : }
1165 : else
1166 0 : elog(ERROR, "unrecognized node type: %d",
1167 : (int) nodeTag(jtnode));
1168 : }
1169 :
1170 : /*
1171 : * preprocess_phv_expression
1172 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1173 : *
1174 : * If a LATERAL subquery references an output of another subquery, and that
1175 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1176 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1177 : * and later pull it back up during find_lateral_references, which runs after
1178 : * subquery_planner has preprocessed all the expressions that were in the
1179 : * current query level to start with. So we need to preprocess it then.
1180 : */
1181 : Expr *
1182 24 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1183 : {
1184 24 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1185 : }
1186 :
1187 : /*
1188 : * inheritance_planner
1189 : * Generate Paths in the case where the result relation is an
1190 : * inheritance set.
1191 : *
1192 : * We have to handle this case differently from cases where a source relation
1193 : * is an inheritance set. Source inheritance is expanded at the bottom of the
1194 : * plan tree (see allpaths.c), but target inheritance has to be expanded at
1195 : * the top. The reason is that for UPDATE, each target relation needs a
1196 : * different targetlist matching its own column set. Fortunately,
1197 : * the UPDATE/DELETE target can never be the nullable side of an outer join,
1198 : * so it's OK to generate the plan this way.
1199 : *
1200 : * Returns nothing; the useful output is in the Paths we attach to
1201 : * the (UPPERREL_FINAL, NULL) upperrel stored in *root.
1202 : *
1203 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1204 : * to leave this to the caller.
1205 : */
1206 : static void
1207 1224 : inheritance_planner(PlannerInfo *root)
1208 : {
1209 1224 : Query *parse = root->parse;
1210 1224 : int top_parentRTindex = parse->resultRelation;
1211 : List *select_rtable;
1212 : List *select_appinfos;
1213 : List *child_appinfos;
1214 : List *old_child_rtis;
1215 : List *new_child_rtis;
1216 : Bitmapset *subqueryRTindexes;
1217 : Index next_subquery_rti;
1218 1224 : int nominalRelation = -1;
1219 1224 : Index rootRelation = 0;
1220 1224 : List *final_rtable = NIL;
1221 1224 : List *final_rowmarks = NIL;
1222 1224 : int save_rel_array_size = 0;
1223 1224 : RelOptInfo **save_rel_array = NULL;
1224 1224 : AppendRelInfo **save_append_rel_array = NULL;
1225 1224 : List *subpaths = NIL;
1226 1224 : List *subroots = NIL;
1227 1224 : List *resultRelations = NIL;
1228 1224 : List *withCheckOptionLists = NIL;
1229 1224 : List *returningLists = NIL;
1230 : List *rowMarks;
1231 : RelOptInfo *final_rel;
1232 : ListCell *lc;
1233 : ListCell *lc2;
1234 : Index rti;
1235 : RangeTblEntry *parent_rte;
1236 : Bitmapset *parent_relids;
1237 : Query **parent_parses;
1238 :
1239 : /* Should only get here for UPDATE or DELETE */
1240 : Assert(parse->commandType == CMD_UPDATE ||
1241 : parse->commandType == CMD_DELETE);
1242 :
1243 : /*
1244 : * We generate a modified instance of the original Query for each target
1245 : * relation, plan that, and put all the plans into a list that will be
1246 : * controlled by a single ModifyTable node. All the instances share the
1247 : * same rangetable, but each instance must have its own set of subquery
1248 : * RTEs within the finished rangetable because (1) they are likely to get
1249 : * scribbled on during planning, and (2) it's not inconceivable that
1250 : * subqueries could get planned differently in different cases. We need
1251 : * not create duplicate copies of other RTE kinds, in particular not the
1252 : * target relations, because they don't have either of those issues. Not
1253 : * having to duplicate the target relations is important because doing so
1254 : * (1) would result in a rangetable of length O(N^2) for N targets, with
1255 : * at least O(N^3) work expended here; and (2) would greatly complicate
1256 : * management of the rowMarks list.
1257 : *
1258 : * To begin with, generate a bitmapset of the relids of the subquery RTEs.
1259 : */
1260 1224 : subqueryRTindexes = NULL;
1261 1224 : rti = 1;
1262 2872 : foreach(lc, parse->rtable)
1263 : {
1264 1648 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1265 :
1266 1648 : if (rte->rtekind == RTE_SUBQUERY)
1267 116 : subqueryRTindexes = bms_add_member(subqueryRTindexes, rti);
1268 1648 : rti++;
1269 : }
1270 :
1271 : /*
1272 : * If the parent RTE is a partitioned table, we should use that as the
1273 : * nominal target relation, because the RTEs added for partitioned tables
1274 : * (including the root parent) as child members of the inheritance set do
1275 : * not appear anywhere else in the plan, so the confusion explained below
1276 : * for non-partitioning inheritance cases is not possible.
1277 : */
1278 1224 : parent_rte = rt_fetch(top_parentRTindex, parse->rtable);
1279 : Assert(parent_rte->inh);
1280 1224 : if (parent_rte->relkind == RELKIND_PARTITIONED_TABLE)
1281 : {
1282 850 : nominalRelation = top_parentRTindex;
1283 850 : rootRelation = top_parentRTindex;
1284 : }
1285 :
1286 : /*
1287 : * Before generating the real per-child-relation plans, do a cycle of
1288 : * planning as though the query were a SELECT. The objective here is to
1289 : * find out which child relations need to be processed, using the same
1290 : * expansion and pruning logic as for a SELECT. We'll then pull out the
1291 : * RangeTblEntry-s generated for the child rels, and make use of the
1292 : * AppendRelInfo entries for them to guide the real planning. (This is
1293 : * rather inefficient; we could perhaps stop short of making a full Path
1294 : * tree. But this whole function is inefficient and slated for
1295 : * destruction, so let's not contort query_planner for that.)
1296 : */
1297 : {
1298 : PlannerInfo *subroot;
1299 :
1300 : /*
1301 : * Flat-copy the PlannerInfo to prevent modification of the original.
1302 : */
1303 1224 : subroot = makeNode(PlannerInfo);
1304 1224 : memcpy(subroot, root, sizeof(PlannerInfo));
1305 :
1306 : /*
1307 : * Make a deep copy of the parsetree for this planning cycle to mess
1308 : * around with, and change it to look like a SELECT. (Hack alert: the
1309 : * target RTE still has updatedCols set if this is an UPDATE, so that
1310 : * expand_partitioned_rtentry will correctly update
1311 : * subroot->partColsUpdated.)
1312 : */
1313 1224 : subroot->parse = copyObject(root->parse);
1314 :
1315 1224 : subroot->parse->commandType = CMD_SELECT;
1316 1224 : subroot->parse->resultRelation = 0;
1317 :
1318 : /*
1319 : * Ensure the subroot has its own copy of the original
1320 : * append_rel_list, since it'll be scribbled on. (Note that at this
1321 : * point, the list only contains AppendRelInfos for flattened UNION
1322 : * ALL subqueries.)
1323 : */
1324 1224 : subroot->append_rel_list = copyObject(root->append_rel_list);
1325 :
1326 : /*
1327 : * Better make a private copy of the rowMarks, too.
1328 : */
1329 1224 : subroot->rowMarks = copyObject(root->rowMarks);
1330 :
1331 : /* There shouldn't be any OJ info to translate, as yet */
1332 : Assert(subroot->join_info_list == NIL);
1333 : /* and we haven't created PlaceHolderInfos, either */
1334 : Assert(subroot->placeholder_list == NIL);
1335 :
1336 : /* Generate Path(s) for accessing this result relation */
1337 1224 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1338 :
1339 : /* Extract the info we need. */
1340 1224 : select_rtable = subroot->parse->rtable;
1341 1224 : select_appinfos = subroot->append_rel_list;
1342 :
1343 : /*
1344 : * We need to propagate partColsUpdated back, too. (The later
1345 : * planning cycles will not set this because they won't run
1346 : * expand_partitioned_rtentry for the UPDATE target.)
1347 : */
1348 1224 : root->partColsUpdated = subroot->partColsUpdated;
1349 : }
1350 :
1351 : /*----------
1352 : * Since only one rangetable can exist in the final plan, we need to make
1353 : * sure that it contains all the RTEs needed for any child plan. This is
1354 : * complicated by the need to use separate subquery RTEs for each child.
1355 : * We arrange the final rtable as follows:
1356 : * 1. All original rtable entries (with their original RT indexes).
1357 : * 2. All the relation RTEs generated for children of the target table.
1358 : * 3. Subquery RTEs for children after the first. We need N * (K - 1)
1359 : * RT slots for this, if there are N subqueries and K child tables.
1360 : * 4. Additional RTEs generated during the child planning runs, such as
1361 : * children of inheritable RTEs other than the target table.
1362 : * We assume that each child planning run will create an identical set
1363 : * of type-4 RTEs.
1364 : *
1365 : * So the next thing to do is append the type-2 RTEs (the target table's
1366 : * children) to the original rtable. We look through select_appinfos
1367 : * to find them.
1368 : *
1369 : * To identify which AppendRelInfos are relevant as we thumb through
1370 : * select_appinfos, we need to look for both direct and indirect children
1371 : * of top_parentRTindex, so we use a bitmap of known parent relids.
1372 : * expand_inherited_rtentry() always processes a parent before any of that
1373 : * parent's children, so we should see an intermediate parent before its
1374 : * children.
1375 : *----------
1376 : */
1377 1224 : child_appinfos = NIL;
1378 1224 : old_child_rtis = NIL;
1379 1224 : new_child_rtis = NIL;
1380 1224 : parent_relids = bms_make_singleton(top_parentRTindex);
1381 4154 : foreach(lc, select_appinfos)
1382 : {
1383 2930 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1384 : RangeTblEntry *child_rte;
1385 :
1386 : /* append_rel_list contains all append rels; ignore others */
1387 2930 : if (!bms_is_member(appinfo->parent_relid, parent_relids))
1388 224 : continue;
1389 :
1390 : /* remember relevant AppendRelInfos for use below */
1391 2706 : child_appinfos = lappend(child_appinfos, appinfo);
1392 :
1393 : /* extract RTE for this child rel */
1394 2706 : child_rte = rt_fetch(appinfo->child_relid, select_rtable);
1395 :
1396 : /* and append it to the original rtable */
1397 2706 : parse->rtable = lappend(parse->rtable, child_rte);
1398 :
1399 : /* remember child's index in the SELECT rtable */
1400 2706 : old_child_rtis = lappend_int(old_child_rtis, appinfo->child_relid);
1401 :
1402 : /* and its new index in the final rtable */
1403 2706 : new_child_rtis = lappend_int(new_child_rtis, list_length(parse->rtable));
1404 :
1405 : /* if child is itself partitioned, update parent_relids */
1406 2706 : if (child_rte->inh)
1407 : {
1408 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1409 284 : parent_relids = bms_add_member(parent_relids, appinfo->child_relid);
1410 : }
1411 : }
1412 :
1413 : /*
1414 : * It's possible that the RTIs we just assigned for the child rels in the
1415 : * final rtable are different from what they were in the SELECT query.
1416 : * Adjust the AppendRelInfos so that they will correctly map RT indexes to
1417 : * the final indexes. We can do this left-to-right since no child rel's
1418 : * final RT index could be greater than what it had in the SELECT query.
1419 : */
1420 3930 : forboth(lc, old_child_rtis, lc2, new_child_rtis)
1421 : {
1422 2706 : int old_child_rti = lfirst_int(lc);
1423 2706 : int new_child_rti = lfirst_int(lc2);
1424 :
1425 2706 : if (old_child_rti == new_child_rti)
1426 2690 : continue; /* nothing to do */
1427 :
1428 : Assert(old_child_rti > new_child_rti);
1429 :
1430 16 : ChangeVarNodes((Node *) child_appinfos,
1431 : old_child_rti, new_child_rti, 0);
1432 : }
1433 :
1434 : /*
1435 : * Now set up rangetable entries for subqueries for additional children
1436 : * (the first child will just use the original ones). These all have to
1437 : * look more or less real, or EXPLAIN will get unhappy; so we just make
1438 : * them all clones of the original subqueries.
1439 : */
1440 1224 : next_subquery_rti = list_length(parse->rtable) + 1;
1441 1224 : if (subqueryRTindexes != NULL)
1442 : {
1443 76 : int n_children = list_length(child_appinfos);
1444 :
1445 284 : while (n_children-- > 1)
1446 : {
1447 132 : int oldrti = -1;
1448 :
1449 476 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1450 : {
1451 : RangeTblEntry *subqrte;
1452 :
1453 212 : subqrte = rt_fetch(oldrti, parse->rtable);
1454 212 : parse->rtable = lappend(parse->rtable, copyObject(subqrte));
1455 : }
1456 : }
1457 : }
1458 :
1459 : /*
1460 : * The query for each child is obtained by translating the query for its
1461 : * immediate parent, since the AppendRelInfo data we have shows deltas
1462 : * between parents and children. We use the parent_parses array to
1463 : * remember the appropriate query trees. This is indexed by parent relid.
1464 : * Since the maximum number of parents is limited by the number of RTEs in
1465 : * the SELECT query, we use that number to allocate the array. An extra
1466 : * entry is needed since relids start from 1.
1467 : */
1468 1224 : parent_parses = (Query **) palloc0((list_length(select_rtable) + 1) *
1469 : sizeof(Query *));
1470 1224 : parent_parses[top_parentRTindex] = parse;
1471 :
1472 : /*
1473 : * And now we can get on with generating a plan for each child table.
1474 : */
1475 3930 : foreach(lc, child_appinfos)
1476 : {
1477 2706 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1478 2706 : Index this_subquery_rti = next_subquery_rti;
1479 : Query *parent_parse;
1480 : PlannerInfo *subroot;
1481 : RangeTblEntry *child_rte;
1482 : RelOptInfo *sub_final_rel;
1483 : Path *subpath;
1484 :
1485 : /*
1486 : * expand_inherited_rtentry() always processes a parent before any of
1487 : * that parent's children, so the parent query for this relation
1488 : * should already be available.
1489 : */
1490 2706 : parent_parse = parent_parses[appinfo->parent_relid];
1491 : Assert(parent_parse != NULL);
1492 :
1493 : /*
1494 : * We need a working copy of the PlannerInfo so that we can control
1495 : * propagation of information back to the main copy.
1496 : */
1497 2706 : subroot = makeNode(PlannerInfo);
1498 2706 : memcpy(subroot, root, sizeof(PlannerInfo));
1499 :
1500 : /*
1501 : * Generate modified query with this rel as target. We first apply
1502 : * adjust_appendrel_attrs, which copies the Query and changes
1503 : * references to the parent RTE to refer to the current child RTE,
1504 : * then fool around with subquery RTEs.
1505 : */
1506 2706 : subroot->parse = (Query *)
1507 2706 : adjust_appendrel_attrs(subroot,
1508 : (Node *) parent_parse,
1509 : 1, &appinfo);
1510 :
1511 : /*
1512 : * If there are securityQuals attached to the parent, move them to the
1513 : * child rel (they've already been transformed properly for that).
1514 : */
1515 2706 : parent_rte = rt_fetch(appinfo->parent_relid, subroot->parse->rtable);
1516 2706 : child_rte = rt_fetch(appinfo->child_relid, subroot->parse->rtable);
1517 2706 : child_rte->securityQuals = parent_rte->securityQuals;
1518 2706 : parent_rte->securityQuals = NIL;
1519 :
1520 : /*
1521 : * HACK: setting this to a value other than INHKIND_NONE signals to
1522 : * relation_excluded_by_constraints() to treat the result relation as
1523 : * being an appendrel member.
1524 : */
1525 2706 : subroot->inhTargetKind =
1526 2706 : (rootRelation != 0) ? INHKIND_PARTITIONED : INHKIND_INHERITED;
1527 :
1528 : /*
1529 : * If this child is further partitioned, remember it as a parent.
1530 : * Since a partitioned table does not have any data, we don't need to
1531 : * create a plan for it, and we can stop processing it here. We do,
1532 : * however, need to remember its modified PlannerInfo for use when
1533 : * processing its children, since we'll update their varnos based on
1534 : * the delta from immediate parent to child, not from top to child.
1535 : *
1536 : * Note: a very non-obvious point is that we have not yet added
1537 : * duplicate subquery RTEs to the subroot's rtable. We mustn't,
1538 : * because then its children would have two sets of duplicates,
1539 : * confusing matters.
1540 : */
1541 2706 : if (child_rte->inh)
1542 : {
1543 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1544 284 : parent_parses[appinfo->child_relid] = subroot->parse;
1545 628 : continue;
1546 : }
1547 :
1548 : /*
1549 : * Set the nominal target relation of the ModifyTable node if not
1550 : * already done. If the target is a partitioned table, we already set
1551 : * nominalRelation to refer to the partition root, above. For
1552 : * non-partitioned inheritance cases, we'll use the first child
1553 : * relation (even if it's excluded) as the nominal target relation.
1554 : * Because of the way expand_inherited_rtentry works, that should be
1555 : * the RTE representing the parent table in its role as a simple
1556 : * member of the inheritance set.
1557 : *
1558 : * It would be logically cleaner to *always* use the inheritance
1559 : * parent RTE as the nominal relation; but that RTE is not otherwise
1560 : * referenced in the plan in the non-partitioned inheritance case.
1561 : * Instead the duplicate child RTE created by expand_inherited_rtentry
1562 : * is used elsewhere in the plan, so using the original parent RTE
1563 : * would give rise to confusing use of multiple aliases in EXPLAIN
1564 : * output for what the user will think is the "same" table. OTOH,
1565 : * it's not a problem in the partitioned inheritance case, because
1566 : * there is no duplicate RTE for the parent.
1567 : */
1568 2422 : if (nominalRelation < 0)
1569 374 : nominalRelation = appinfo->child_relid;
1570 :
1571 : /*
1572 : * As above, each child plan run needs its own append_rel_list and
1573 : * rowmarks, which should start out as pristine copies of the
1574 : * originals. There can't be any references to UPDATE/DELETE target
1575 : * rels in them; but there could be subquery references, which we'll
1576 : * fix up in a moment.
1577 : */
1578 2422 : subroot->append_rel_list = copyObject(root->append_rel_list);
1579 2422 : subroot->rowMarks = copyObject(root->rowMarks);
1580 :
1581 : /*
1582 : * If this isn't the first child Query, adjust Vars and jointree
1583 : * entries to reference the appropriate set of subquery RTEs.
1584 : */
1585 2422 : if (final_rtable != NIL && subqueryRTindexes != NULL)
1586 : {
1587 120 : int oldrti = -1;
1588 :
1589 432 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1590 : {
1591 192 : Index newrti = next_subquery_rti++;
1592 :
1593 192 : ChangeVarNodes((Node *) subroot->parse, oldrti, newrti, 0);
1594 192 : ChangeVarNodes((Node *) subroot->append_rel_list,
1595 : oldrti, newrti, 0);
1596 192 : ChangeVarNodes((Node *) subroot->rowMarks, oldrti, newrti, 0);
1597 : }
1598 : }
1599 :
1600 : /* There shouldn't be any OJ info to translate, as yet */
1601 : Assert(subroot->join_info_list == NIL);
1602 : /* and we haven't created PlaceHolderInfos, either */
1603 : Assert(subroot->placeholder_list == NIL);
1604 :
1605 : /* Generate Path(s) for accessing this result relation */
1606 2422 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1607 :
1608 : /*
1609 : * Select cheapest path in case there's more than one. We always run
1610 : * modification queries to conclusion, so we care only for the
1611 : * cheapest-total path.
1612 : */
1613 2422 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
1614 2422 : set_cheapest(sub_final_rel);
1615 2422 : subpath = sub_final_rel->cheapest_total_path;
1616 :
1617 : /*
1618 : * If this child rel was excluded by constraint exclusion, exclude it
1619 : * from the result plan.
1620 : */
1621 2422 : if (IS_DUMMY_REL(sub_final_rel))
1622 60 : continue;
1623 :
1624 : /*
1625 : * If this is the first non-excluded child, its post-planning rtable
1626 : * becomes the initial contents of final_rtable; otherwise, copy its
1627 : * modified subquery RTEs into final_rtable, to ensure we have sane
1628 : * copies of those. Also save the first non-excluded child's version
1629 : * of the rowmarks list; we assume all children will end up with
1630 : * equivalent versions of that.
1631 : */
1632 2362 : if (final_rtable == NIL)
1633 : {
1634 1192 : final_rtable = subroot->parse->rtable;
1635 1192 : final_rowmarks = subroot->rowMarks;
1636 : }
1637 : else
1638 : {
1639 : Assert(list_length(final_rtable) ==
1640 : list_length(subroot->parse->rtable));
1641 1170 : if (subqueryRTindexes != NULL)
1642 : {
1643 120 : int oldrti = -1;
1644 :
1645 432 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1646 : {
1647 192 : Index newrti = this_subquery_rti++;
1648 : RangeTblEntry *subqrte;
1649 : ListCell *newrticell;
1650 :
1651 192 : subqrte = rt_fetch(newrti, subroot->parse->rtable);
1652 192 : newrticell = list_nth_cell(final_rtable, newrti - 1);
1653 192 : lfirst(newrticell) = subqrte;
1654 : }
1655 : }
1656 : }
1657 :
1658 : /*
1659 : * We need to collect all the RelOptInfos from all child plans into
1660 : * the main PlannerInfo, since setrefs.c will need them. We use the
1661 : * last child's simple_rel_array, so we have to propagate forward the
1662 : * RelOptInfos that were already built in previous children.
1663 : */
1664 : Assert(subroot->simple_rel_array_size >= save_rel_array_size);
1665 10216 : for (rti = 1; rti < save_rel_array_size; rti++)
1666 : {
1667 7854 : RelOptInfo *brel = save_rel_array[rti];
1668 :
1669 7854 : if (brel)
1670 2932 : subroot->simple_rel_array[rti] = brel;
1671 : }
1672 2362 : save_rel_array_size = subroot->simple_rel_array_size;
1673 2362 : save_rel_array = subroot->simple_rel_array;
1674 2362 : save_append_rel_array = subroot->append_rel_array;
1675 :
1676 : /*
1677 : * Make sure any initplans from this rel get into the outer list. Note
1678 : * we're effectively assuming all children generate the same
1679 : * init_plans.
1680 : */
1681 2362 : root->init_plans = subroot->init_plans;
1682 :
1683 : /* Build list of sub-paths */
1684 2362 : subpaths = lappend(subpaths, subpath);
1685 :
1686 : /* Build list of modified subroots, too */
1687 2362 : subroots = lappend(subroots, subroot);
1688 :
1689 : /* Build list of target-relation RT indexes */
1690 2362 : resultRelations = lappend_int(resultRelations, appinfo->child_relid);
1691 :
1692 : /* Build lists of per-relation WCO and RETURNING targetlists */
1693 2362 : if (parse->withCheckOptions)
1694 272 : withCheckOptionLists = lappend(withCheckOptionLists,
1695 272 : subroot->parse->withCheckOptions);
1696 2362 : if (parse->returningList)
1697 316 : returningLists = lappend(returningLists,
1698 316 : subroot->parse->returningList);
1699 :
1700 : Assert(!parse->onConflict);
1701 : }
1702 :
1703 : /* Result path must go into outer query's FINAL upperrel */
1704 1224 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1705 :
1706 : /*
1707 : * We don't currently worry about setting final_rel's consider_parallel
1708 : * flag in this case, nor about allowing FDWs or create_upper_paths_hook
1709 : * to get control here.
1710 : */
1711 :
1712 1224 : if (subpaths == NIL)
1713 : {
1714 : /*
1715 : * We managed to exclude every child rel, so generate a dummy path
1716 : * representing the empty set. Although it's clear that no data will
1717 : * be updated or deleted, we will still need to have a ModifyTable
1718 : * node so that any statement triggers are executed. (This could be
1719 : * cleaner if we fixed nodeModifyTable.c to support zero child nodes,
1720 : * but that probably wouldn't be a net win.)
1721 : */
1722 : Path *dummy_path;
1723 :
1724 : /* tlist processing never got done, either */
1725 32 : root->processed_tlist = preprocess_targetlist(root);
1726 32 : final_rel->reltarget = create_pathtarget(root, root->processed_tlist);
1727 :
1728 : /* Make a dummy path, cf set_dummy_rel_pathlist() */
1729 32 : dummy_path = (Path *) create_append_path(NULL, final_rel, NIL, NIL,
1730 : NIL, NULL, 0, false,
1731 : NIL, -1);
1732 :
1733 : /* These lists must be nonempty to make a valid ModifyTable node */
1734 32 : subpaths = list_make1(dummy_path);
1735 32 : subroots = list_make1(root);
1736 32 : resultRelations = list_make1_int(parse->resultRelation);
1737 32 : if (parse->withCheckOptions)
1738 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
1739 32 : if (parse->returningList)
1740 12 : returningLists = list_make1(parse->returningList);
1741 : /* Disable tuple routing, too, just to be safe */
1742 32 : root->partColsUpdated = false;
1743 : }
1744 : else
1745 : {
1746 : /*
1747 : * Put back the final adjusted rtable into the master copy of the
1748 : * Query. (We mustn't do this if we found no non-excluded children,
1749 : * since we never saved an adjusted rtable at all.)
1750 : */
1751 1192 : parse->rtable = final_rtable;
1752 1192 : root->simple_rel_array_size = save_rel_array_size;
1753 1192 : root->simple_rel_array = save_rel_array;
1754 1192 : root->append_rel_array = save_append_rel_array;
1755 :
1756 : /* Must reconstruct master's simple_rte_array, too */
1757 1192 : root->simple_rte_array = (RangeTblEntry **)
1758 1192 : palloc0((list_length(final_rtable) + 1) * sizeof(RangeTblEntry *));
1759 1192 : rti = 1;
1760 5858 : foreach(lc, final_rtable)
1761 : {
1762 4666 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1763 :
1764 4666 : root->simple_rte_array[rti++] = rte;
1765 : }
1766 :
1767 : /* Put back adjusted rowmarks, too */
1768 1192 : root->rowMarks = final_rowmarks;
1769 : }
1770 :
1771 : /*
1772 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node will
1773 : * have dealt with fetching non-locked marked rows, else we need to have
1774 : * ModifyTable do that.
1775 : */
1776 1224 : if (parse->rowMarks)
1777 0 : rowMarks = NIL;
1778 : else
1779 1224 : rowMarks = root->rowMarks;
1780 :
1781 : /* Create Path representing a ModifyTable to do the UPDATE/DELETE work */
1782 1224 : add_path(final_rel, (Path *)
1783 3672 : create_modifytable_path(root, final_rel,
1784 : parse->commandType,
1785 1224 : parse->canSetTag,
1786 : nominalRelation,
1787 : rootRelation,
1788 1224 : root->partColsUpdated,
1789 : resultRelations,
1790 : subpaths,
1791 : subroots,
1792 : withCheckOptionLists,
1793 : returningLists,
1794 : rowMarks,
1795 : NULL,
1796 : assign_special_exec_param(root)));
1797 1224 : }
1798 :
1799 : /*--------------------
1800 : * grouping_planner
1801 : * Perform planning steps related to grouping, aggregation, etc.
1802 : *
1803 : * This function adds all required top-level processing to the scan/join
1804 : * Path(s) produced by query_planner.
1805 : *
1806 : * If inheritance_update is true, we're being called from inheritance_planner
1807 : * and should not include a ModifyTable step in the resulting Path(s).
1808 : * (inheritance_planner will create a single ModifyTable node covering all the
1809 : * target tables.)
1810 : *
1811 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1812 : * tuple_fraction is interpreted as follows:
1813 : * 0: expect all tuples to be retrieved (normal case)
1814 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1815 : * from the plan to be retrieved
1816 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1817 : * expected to be retrieved (ie, a LIMIT specification)
1818 : *
1819 : * Returns nothing; the useful output is in the Paths we attach to the
1820 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1821 : * root->processed_tlist contains the final processed targetlist.
1822 : *
1823 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1824 : * to leave this to the caller.
1825 : *--------------------
1826 : */
1827 : static void
1828 302598 : grouping_planner(PlannerInfo *root, bool inheritance_update,
1829 : double tuple_fraction)
1830 : {
1831 302598 : Query *parse = root->parse;
1832 302598 : int64 offset_est = 0;
1833 302598 : int64 count_est = 0;
1834 302598 : double limit_tuples = -1.0;
1835 302598 : bool have_postponed_srfs = false;
1836 : PathTarget *final_target;
1837 : List *final_targets;
1838 : List *final_targets_contain_srfs;
1839 : bool final_target_parallel_safe;
1840 : RelOptInfo *current_rel;
1841 : RelOptInfo *final_rel;
1842 : FinalPathExtraData extra;
1843 : ListCell *lc;
1844 :
1845 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1846 302598 : if (parse->limitCount || parse->limitOffset)
1847 : {
1848 3054 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1849 : &offset_est, &count_est);
1850 :
1851 : /*
1852 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1853 : * estimate the effects of using a bounded sort.
1854 : */
1855 3054 : if (count_est > 0 && offset_est >= 0)
1856 2820 : limit_tuples = (double) count_est + (double) offset_est;
1857 : }
1858 :
1859 : /* Make tuple_fraction accessible to lower-level routines */
1860 302598 : root->tuple_fraction = tuple_fraction;
1861 :
1862 302598 : if (parse->setOperations)
1863 : {
1864 : /*
1865 : * If there's a top-level ORDER BY, assume we have to fetch all the
1866 : * tuples. This might be too simplistic given all the hackery below
1867 : * to possibly avoid the sort; but the odds of accurate estimates here
1868 : * are pretty low anyway. XXX try to get rid of this in favor of
1869 : * letting plan_set_operations generate both fast-start and
1870 : * cheapest-total paths.
1871 : */
1872 1248 : if (parse->sortClause)
1873 360 : root->tuple_fraction = 0.0;
1874 :
1875 : /*
1876 : * Construct Paths for set operations. The results will not need any
1877 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1878 : * special work for recursive unions is the responsibility of
1879 : * plan_set_operations.
1880 : */
1881 1248 : current_rel = plan_set_operations(root);
1882 :
1883 : /*
1884 : * We should not need to call preprocess_targetlist, since we must be
1885 : * in a SELECT query node. Instead, use the processed_tlist returned
1886 : * by plan_set_operations (since this tells whether it returned any
1887 : * resjunk columns!), and transfer any sort key information from the
1888 : * original tlist.
1889 : */
1890 : Assert(parse->commandType == CMD_SELECT);
1891 :
1892 : /* for safety, copy processed_tlist instead of modifying in-place */
1893 1248 : root->processed_tlist =
1894 1248 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1895 : parse->targetList);
1896 :
1897 : /* Also extract the PathTarget form of the setop result tlist */
1898 1248 : final_target = current_rel->cheapest_total_path->pathtarget;
1899 :
1900 : /* And check whether it's parallel safe */
1901 1248 : final_target_parallel_safe =
1902 1248 : is_parallel_safe(root, (Node *) final_target->exprs);
1903 :
1904 : /* The setop result tlist couldn't contain any SRFs */
1905 : Assert(!parse->hasTargetSRFs);
1906 1248 : final_targets = final_targets_contain_srfs = NIL;
1907 :
1908 : /*
1909 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1910 : * checked already, but let's make sure).
1911 : */
1912 1248 : if (parse->rowMarks)
1913 0 : ereport(ERROR,
1914 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1915 : /*------
1916 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1917 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1918 : LCS_asString(linitial_node(RowMarkClause,
1919 : parse->rowMarks)->strength))));
1920 :
1921 : /*
1922 : * Calculate pathkeys that represent result ordering requirements
1923 : */
1924 : Assert(parse->distinctClause == NIL);
1925 1248 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1926 : parse->sortClause,
1927 : root->processed_tlist);
1928 : }
1929 : else
1930 : {
1931 : /* No set operations, do regular planning */
1932 : PathTarget *sort_input_target;
1933 : List *sort_input_targets;
1934 : List *sort_input_targets_contain_srfs;
1935 : bool sort_input_target_parallel_safe;
1936 : PathTarget *grouping_target;
1937 : List *grouping_targets;
1938 : List *grouping_targets_contain_srfs;
1939 : bool grouping_target_parallel_safe;
1940 : PathTarget *scanjoin_target;
1941 : List *scanjoin_targets;
1942 : List *scanjoin_targets_contain_srfs;
1943 : bool scanjoin_target_parallel_safe;
1944 : bool scanjoin_target_same_exprs;
1945 : bool have_grouping;
1946 : AggClauseCosts agg_costs;
1947 301350 : WindowFuncLists *wflists = NULL;
1948 301350 : List *activeWindows = NIL;
1949 301350 : grouping_sets_data *gset_data = NULL;
1950 : standard_qp_extra qp_extra;
1951 :
1952 : /* A recursive query should always have setOperations */
1953 : Assert(!root->hasRecursion);
1954 :
1955 : /* Preprocess grouping sets and GROUP BY clause, if any */
1956 301350 : if (parse->groupingSets)
1957 : {
1958 436 : gset_data = preprocess_grouping_sets(root);
1959 : }
1960 : else
1961 : {
1962 : /* Preprocess regular GROUP BY clause, if any */
1963 300914 : if (parse->groupClause)
1964 1704 : parse->groupClause = preprocess_groupclause(root, NIL);
1965 : }
1966 :
1967 : /*
1968 : * Preprocess targetlist. Note that much of the remaining planning
1969 : * work will be done with the PathTarget representation of tlists, but
1970 : * we must also maintain the full representation of the final tlist so
1971 : * that we can transfer its decoration (resnames etc) to the topmost
1972 : * tlist of the finished Plan. This is kept in processed_tlist.
1973 : */
1974 301346 : root->processed_tlist = preprocess_targetlist(root);
1975 :
1976 : /*
1977 : * Collect statistics about aggregates for estimating costs, and mark
1978 : * all the aggregates with resolved aggtranstypes. We must do this
1979 : * before slicing and dicing the tlist into various pathtargets, else
1980 : * some copies of the Aggref nodes might escape being marked with the
1981 : * correct transtypes.
1982 : *
1983 : * Note: currently, we do not detect duplicate aggregates here. This
1984 : * may result in somewhat-overestimated cost, which is fine for our
1985 : * purposes since all Paths will get charged the same. But at some
1986 : * point we might wish to do that detection in the planner, rather
1987 : * than during executor startup.
1988 : */
1989 301346 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1990 301346 : if (parse->hasAggs)
1991 : {
1992 23560 : get_agg_clause_costs(root, (Node *) root->processed_tlist,
1993 : AGGSPLIT_SIMPLE, &agg_costs);
1994 23560 : get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
1995 : &agg_costs);
1996 : }
1997 :
1998 : /*
1999 : * Locate any window functions in the tlist. (We don't need to look
2000 : * anywhere else, since expressions used in ORDER BY will be in there
2001 : * too.) Note that they could all have been eliminated by constant
2002 : * folding, in which case we don't need to do any more work.
2003 : */
2004 301346 : if (parse->hasWindowFuncs)
2005 : {
2006 1116 : wflists = find_window_functions((Node *) root->processed_tlist,
2007 1116 : list_length(parse->windowClause));
2008 1116 : if (wflists->numWindowFuncs > 0)
2009 1112 : activeWindows = select_active_windows(root, wflists);
2010 : else
2011 4 : parse->hasWindowFuncs = false;
2012 : }
2013 :
2014 : /*
2015 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
2016 : * adding logic between here and the query_planner() call. Anything
2017 : * that is needed in MIN/MAX-optimizable cases will have to be
2018 : * duplicated in planagg.c.
2019 : */
2020 301346 : if (parse->hasAggs)
2021 23560 : preprocess_minmax_aggregates(root);
2022 :
2023 : /*
2024 : * Figure out whether there's a hard limit on the number of rows that
2025 : * query_planner's result subplan needs to return. Even if we know a
2026 : * hard limit overall, it doesn't apply if the query has any
2027 : * grouping/aggregation operations, or SRFs in the tlist.
2028 : */
2029 600584 : if (parse->groupClause ||
2030 598448 : parse->groupingSets ||
2031 598038 : parse->distinctClause ||
2032 576002 : parse->hasAggs ||
2033 553316 : parse->hasWindowFuncs ||
2034 548714 : parse->hasTargetSRFs ||
2035 272572 : root->hasHavingQual)
2036 28786 : root->limit_tuples = -1.0;
2037 : else
2038 272560 : root->limit_tuples = limit_tuples;
2039 :
2040 : /* Set up data needed by standard_qp_callback */
2041 301346 : qp_extra.activeWindows = activeWindows;
2042 301346 : qp_extra.groupClause = (gset_data
2043 432 : ? (gset_data->rollups ? linitial_node(RollupData, gset_data->rollups)->groupClause : NIL)
2044 301778 : : parse->groupClause);
2045 :
2046 : /*
2047 : * Generate the best unsorted and presorted paths for the scan/join
2048 : * portion of this Query, ie the processing represented by the
2049 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
2050 : * We also generate (in standard_qp_callback) pathkey representations
2051 : * of the query's sort clause, distinct clause, etc.
2052 : */
2053 301346 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
2054 :
2055 : /*
2056 : * Convert the query's result tlist into PathTarget format.
2057 : *
2058 : * Note: this cannot be done before query_planner() has performed
2059 : * appendrel expansion, because that might add resjunk entries to
2060 : * root->processed_tlist. Waiting till afterwards is also helpful
2061 : * because the target width estimates can use per-Var width numbers
2062 : * that were obtained within query_planner().
2063 : */
2064 301336 : final_target = create_pathtarget(root, root->processed_tlist);
2065 301336 : final_target_parallel_safe =
2066 301336 : is_parallel_safe(root, (Node *) final_target->exprs);
2067 :
2068 : /*
2069 : * If ORDER BY was given, consider whether we should use a post-sort
2070 : * projection, and compute the adjusted target for preceding steps if
2071 : * so.
2072 : */
2073 301336 : if (parse->sortClause)
2074 : {
2075 35756 : sort_input_target = make_sort_input_target(root,
2076 : final_target,
2077 : &have_postponed_srfs);
2078 35756 : sort_input_target_parallel_safe =
2079 35756 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
2080 : }
2081 : else
2082 : {
2083 265580 : sort_input_target = final_target;
2084 265580 : sort_input_target_parallel_safe = final_target_parallel_safe;
2085 : }
2086 :
2087 : /*
2088 : * If we have window functions to deal with, the output from any
2089 : * grouping step needs to be what the window functions want;
2090 : * otherwise, it should be sort_input_target.
2091 : */
2092 301336 : if (activeWindows)
2093 : {
2094 1112 : grouping_target = make_window_input_target(root,
2095 : final_target,
2096 : activeWindows);
2097 1112 : grouping_target_parallel_safe =
2098 1112 : is_parallel_safe(root, (Node *) grouping_target->exprs);
2099 : }
2100 : else
2101 : {
2102 300224 : grouping_target = sort_input_target;
2103 300224 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
2104 : }
2105 :
2106 : /*
2107 : * If we have grouping or aggregation to do, the topmost scan/join
2108 : * plan node must emit what the grouping step wants; otherwise, it
2109 : * should emit grouping_target.
2110 : */
2111 1201100 : have_grouping = (parse->groupClause || parse->groupingSets ||
2112 878064 : parse->hasAggs || root->hasHavingQual);
2113 301336 : if (have_grouping)
2114 : {
2115 23820 : scanjoin_target = make_group_input_target(root, final_target);
2116 23820 : scanjoin_target_parallel_safe =
2117 23820 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2118 : }
2119 : else
2120 : {
2121 277516 : scanjoin_target = grouping_target;
2122 277516 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2123 : }
2124 :
2125 : /*
2126 : * If there are any SRFs in the targetlist, we must separate each of
2127 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2128 : * each of the named targets with a SRF-free version, and remember the
2129 : * list of additional projection steps we need to add afterwards.
2130 : */
2131 301336 : if (parse->hasTargetSRFs)
2132 : {
2133 : /* final_target doesn't recompute any SRFs in sort_input_target */
2134 3682 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2135 : &final_targets,
2136 : &final_targets_contain_srfs);
2137 3682 : final_target = linitial_node(PathTarget, final_targets);
2138 : Assert(!linitial_int(final_targets_contain_srfs));
2139 : /* likewise for sort_input_target vs. grouping_target */
2140 3682 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2141 : &sort_input_targets,
2142 : &sort_input_targets_contain_srfs);
2143 3682 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2144 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2145 : /* likewise for grouping_target vs. scanjoin_target */
2146 3682 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
2147 : &grouping_targets,
2148 : &grouping_targets_contain_srfs);
2149 3682 : grouping_target = linitial_node(PathTarget, grouping_targets);
2150 : Assert(!linitial_int(grouping_targets_contain_srfs));
2151 : /* scanjoin_target will not have any SRFs precomputed for it */
2152 3682 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2153 : &scanjoin_targets,
2154 : &scanjoin_targets_contain_srfs);
2155 3682 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2156 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2157 : }
2158 : else
2159 : {
2160 : /* initialize lists; for most of these, dummy values are OK */
2161 297654 : final_targets = final_targets_contain_srfs = NIL;
2162 297654 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2163 297654 : grouping_targets = grouping_targets_contain_srfs = NIL;
2164 297654 : scanjoin_targets = list_make1(scanjoin_target);
2165 297654 : scanjoin_targets_contain_srfs = NIL;
2166 : }
2167 :
2168 : /* Apply scan/join target. */
2169 602672 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2170 301336 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2171 301336 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2172 : scanjoin_targets_contain_srfs,
2173 : scanjoin_target_parallel_safe,
2174 : scanjoin_target_same_exprs);
2175 :
2176 : /*
2177 : * Save the various upper-rel PathTargets we just computed into
2178 : * root->upper_targets[]. The core code doesn't use this, but it
2179 : * provides a convenient place for extensions to get at the info. For
2180 : * consistency, we save all the intermediate targets, even though some
2181 : * of the corresponding upperrels might not be needed for this query.
2182 : */
2183 301336 : root->upper_targets[UPPERREL_FINAL] = final_target;
2184 301336 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2185 301336 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2186 301336 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2187 301336 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2188 :
2189 : /*
2190 : * If we have grouping and/or aggregation, consider ways to implement
2191 : * that. We build a new upperrel representing the output of this
2192 : * phase.
2193 : */
2194 301336 : if (have_grouping)
2195 : {
2196 23820 : current_rel = create_grouping_paths(root,
2197 : current_rel,
2198 : grouping_target,
2199 : grouping_target_parallel_safe,
2200 : &agg_costs,
2201 : gset_data);
2202 : /* Fix things up if grouping_target contains SRFs */
2203 23816 : if (parse->hasTargetSRFs)
2204 84 : adjust_paths_for_srfs(root, current_rel,
2205 : grouping_targets,
2206 : grouping_targets_contain_srfs);
2207 : }
2208 :
2209 : /*
2210 : * If we have window functions, consider ways to implement those. We
2211 : * build a new upperrel representing the output of this phase.
2212 : */
2213 301332 : if (activeWindows)
2214 : {
2215 1112 : current_rel = create_window_paths(root,
2216 : current_rel,
2217 : grouping_target,
2218 : sort_input_target,
2219 : sort_input_target_parallel_safe,
2220 : wflists,
2221 : activeWindows);
2222 : /* Fix things up if sort_input_target contains SRFs */
2223 1112 : if (parse->hasTargetSRFs)
2224 8 : adjust_paths_for_srfs(root, current_rel,
2225 : sort_input_targets,
2226 : sort_input_targets_contain_srfs);
2227 : }
2228 :
2229 : /*
2230 : * If there is a DISTINCT clause, consider ways to implement that. We
2231 : * build a new upperrel representing the output of this phase.
2232 : */
2233 301332 : if (parse->distinctClause)
2234 : {
2235 386 : current_rel = create_distinct_paths(root,
2236 : current_rel);
2237 : }
2238 : } /* end of if (setOperations) */
2239 :
2240 : /*
2241 : * If ORDER BY was given, consider ways to implement that, and generate a
2242 : * new upperrel containing only paths that emit the correct ordering and
2243 : * project the correct final_target. We can apply the original
2244 : * limit_tuples limit in sort costing here, but only if there are no
2245 : * postponed SRFs.
2246 : */
2247 302580 : if (parse->sortClause)
2248 : {
2249 36116 : current_rel = create_ordered_paths(root,
2250 : current_rel,
2251 : final_target,
2252 : final_target_parallel_safe,
2253 : have_postponed_srfs ? -1.0 :
2254 : limit_tuples);
2255 : /* Fix things up if final_target contains SRFs */
2256 36116 : if (parse->hasTargetSRFs)
2257 120 : adjust_paths_for_srfs(root, current_rel,
2258 : final_targets,
2259 : final_targets_contain_srfs);
2260 : }
2261 :
2262 : /*
2263 : * Now we are prepared to build the final-output upperrel.
2264 : */
2265 302580 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2266 :
2267 : /*
2268 : * If the input rel is marked consider_parallel and there's nothing that's
2269 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2270 : * consider_parallel as well. Note that if the query has rowMarks or is
2271 : * not a SELECT, consider_parallel will be false for every relation in the
2272 : * query.
2273 : */
2274 375102 : if (current_rel->consider_parallel &&
2275 145036 : is_parallel_safe(root, parse->limitOffset) &&
2276 72514 : is_parallel_safe(root, parse->limitCount))
2277 72510 : final_rel->consider_parallel = true;
2278 :
2279 : /*
2280 : * If the current_rel belongs to a single FDW, so does the final_rel.
2281 : */
2282 302580 : final_rel->serverid = current_rel->serverid;
2283 302580 : final_rel->userid = current_rel->userid;
2284 302580 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2285 302580 : final_rel->fdwroutine = current_rel->fdwroutine;
2286 :
2287 : /*
2288 : * Generate paths for the final_rel. Insert all surviving paths, with
2289 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2290 : */
2291 610382 : foreach(lc, current_rel->pathlist)
2292 : {
2293 307802 : Path *path = (Path *) lfirst(lc);
2294 :
2295 : /*
2296 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2297 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2298 : * here. If there are only non-locking rowmarks, they should be
2299 : * handled by the ModifyTable node instead. However, root->rowMarks
2300 : * is what goes into the LockRows node.)
2301 : */
2302 307802 : if (parse->rowMarks)
2303 : {
2304 5310 : path = (Path *) create_lockrows_path(root, final_rel, path,
2305 : root->rowMarks,
2306 : assign_special_exec_param(root));
2307 : }
2308 :
2309 : /*
2310 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2311 : */
2312 307802 : if (limit_needed(parse))
2313 : {
2314 3536 : path = (Path *) create_limit_path(root, final_rel, path,
2315 : parse->limitOffset,
2316 : parse->limitCount,
2317 : offset_est, count_est);
2318 : }
2319 :
2320 : /*
2321 : * If this is an INSERT/UPDATE/DELETE, and we're not being called from
2322 : * inheritance_planner, add the ModifyTable node.
2323 : */
2324 307802 : if (parse->commandType != CMD_SELECT && !inheritance_update)
2325 : {
2326 : Index rootRelation;
2327 : List *withCheckOptionLists;
2328 : List *returningLists;
2329 : List *rowMarks;
2330 :
2331 : /*
2332 : * If target is a partition root table, we need to mark the
2333 : * ModifyTable node appropriately for that.
2334 : */
2335 68172 : if (rt_fetch(parse->resultRelation, parse->rtable)->relkind ==
2336 : RELKIND_PARTITIONED_TABLE)
2337 1498 : rootRelation = parse->resultRelation;
2338 : else
2339 66674 : rootRelation = 0;
2340 :
2341 : /*
2342 : * Set up the WITH CHECK OPTION and RETURNING lists-of-lists, if
2343 : * needed.
2344 : */
2345 68172 : if (parse->withCheckOptions)
2346 508 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2347 : else
2348 67664 : withCheckOptionLists = NIL;
2349 :
2350 68172 : if (parse->returningList)
2351 1484 : returningLists = list_make1(parse->returningList);
2352 : else
2353 66688 : returningLists = NIL;
2354 :
2355 : /*
2356 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2357 : * will have dealt with fetching non-locked marked rows, else we
2358 : * need to have ModifyTable do that.
2359 : */
2360 68172 : if (parse->rowMarks)
2361 0 : rowMarks = NIL;
2362 : else
2363 68172 : rowMarks = root->rowMarks;
2364 :
2365 68172 : path = (Path *)
2366 409032 : create_modifytable_path(root, final_rel,
2367 : parse->commandType,
2368 68172 : parse->canSetTag,
2369 68172 : parse->resultRelation,
2370 : rootRelation,
2371 : false,
2372 68172 : list_make1_int(parse->resultRelation),
2373 68172 : list_make1(path),
2374 68172 : list_make1(root),
2375 : withCheckOptionLists,
2376 : returningLists,
2377 : rowMarks,
2378 : parse->onConflict,
2379 : assign_special_exec_param(root));
2380 : }
2381 :
2382 : /* And shove it into final_rel */
2383 307802 : add_path(final_rel, path);
2384 : }
2385 :
2386 : /*
2387 : * Generate partial paths for final_rel, too, if outer query levels might
2388 : * be able to make use of them.
2389 : */
2390 309412 : if (final_rel->consider_parallel && root->query_level > 1 &&
2391 6832 : !limit_needed(parse))
2392 : {
2393 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2394 6798 : foreach(lc, current_rel->partial_pathlist)
2395 : {
2396 44 : Path *partial_path = (Path *) lfirst(lc);
2397 :
2398 44 : add_partial_path(final_rel, partial_path);
2399 : }
2400 : }
2401 :
2402 302580 : extra.limit_needed = limit_needed(parse);
2403 302580 : extra.limit_tuples = limit_tuples;
2404 302580 : extra.count_est = count_est;
2405 302580 : extra.offset_est = offset_est;
2406 :
2407 : /*
2408 : * If there is an FDW that's responsible for all baserels of the query,
2409 : * let it consider adding ForeignPaths.
2410 : */
2411 303572 : if (final_rel->fdwroutine &&
2412 992 : final_rel->fdwroutine->GetForeignUpperPaths)
2413 934 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2414 : current_rel, final_rel,
2415 : &extra);
2416 :
2417 : /* Let extensions possibly add some more paths */
2418 302580 : if (create_upper_paths_hook)
2419 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2420 : current_rel, final_rel, &extra);
2421 :
2422 : /* Note: currently, we leave it to callers to do set_cheapest() */
2423 302580 : }
2424 :
2425 : /*
2426 : * Do preprocessing for groupingSets clause and related data. This handles the
2427 : * preliminary steps of expanding the grouping sets, organizing them into lists
2428 : * of rollups, and preparing annotations which will later be filled in with
2429 : * size estimates.
2430 : */
2431 : static grouping_sets_data *
2432 436 : preprocess_grouping_sets(PlannerInfo *root)
2433 : {
2434 436 : Query *parse = root->parse;
2435 : List *sets;
2436 436 : int maxref = 0;
2437 : ListCell *lc;
2438 : ListCell *lc_set;
2439 436 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2440 :
2441 436 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, -1);
2442 :
2443 436 : gd->any_hashable = false;
2444 436 : gd->unhashable_refs = NULL;
2445 436 : gd->unsortable_refs = NULL;
2446 436 : gd->unsortable_sets = NIL;
2447 :
2448 436 : if (parse->groupClause)
2449 : {
2450 : ListCell *lc;
2451 :
2452 1312 : foreach(lc, parse->groupClause)
2453 : {
2454 904 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2455 904 : Index ref = gc->tleSortGroupRef;
2456 :
2457 904 : if (ref > maxref)
2458 880 : maxref = ref;
2459 :
2460 904 : if (!gc->hashable)
2461 20 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2462 :
2463 904 : if (!OidIsValid(gc->sortop))
2464 28 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2465 : }
2466 : }
2467 :
2468 : /* Allocate workspace array for remapping */
2469 436 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2470 :
2471 : /*
2472 : * If we have any unsortable sets, we must extract them before trying to
2473 : * prepare rollups. Unsortable sets don't go through
2474 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2475 : * here.
2476 : */
2477 436 : if (!bms_is_empty(gd->unsortable_refs))
2478 : {
2479 28 : List *sortable_sets = NIL;
2480 :
2481 84 : foreach(lc, parse->groupingSets)
2482 : {
2483 60 : List *gset = (List *) lfirst(lc);
2484 :
2485 60 : if (bms_overlap_list(gd->unsortable_refs, gset))
2486 : {
2487 32 : GroupingSetData *gs = makeNode(GroupingSetData);
2488 :
2489 32 : gs->set = gset;
2490 32 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2491 :
2492 : /*
2493 : * We must enforce here that an unsortable set is hashable;
2494 : * later code assumes this. Parse analysis only checks that
2495 : * every individual column is either hashable or sortable.
2496 : *
2497 : * Note that passing this test doesn't guarantee we can
2498 : * generate a plan; there might be other showstoppers.
2499 : */
2500 32 : if (bms_overlap_list(gd->unhashable_refs, gset))
2501 4 : ereport(ERROR,
2502 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2503 : errmsg("could not implement GROUP BY"),
2504 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2505 : }
2506 : else
2507 28 : sortable_sets = lappend(sortable_sets, gset);
2508 : }
2509 :
2510 24 : if (sortable_sets)
2511 20 : sets = extract_rollup_sets(sortable_sets);
2512 : else
2513 4 : sets = NIL;
2514 : }
2515 : else
2516 408 : sets = extract_rollup_sets(parse->groupingSets);
2517 :
2518 1124 : foreach(lc_set, sets)
2519 : {
2520 692 : List *current_sets = (List *) lfirst(lc_set);
2521 692 : RollupData *rollup = makeNode(RollupData);
2522 : GroupingSetData *gs;
2523 :
2524 : /*
2525 : * Reorder the current list of grouping sets into correct prefix
2526 : * order. If only one aggregation pass is needed, try to make the
2527 : * list match the ORDER BY clause; if more than one pass is needed, we
2528 : * don't bother with that.
2529 : *
2530 : * Note that this reorders the sets from smallest-member-first to
2531 : * largest-member-first, and applies the GroupingSetData annotations,
2532 : * though the data will be filled in later.
2533 : */
2534 692 : current_sets = reorder_grouping_sets(current_sets,
2535 692 : (list_length(sets) == 1
2536 : ? parse->sortClause
2537 : : NIL));
2538 :
2539 : /*
2540 : * Get the initial (and therefore largest) grouping set.
2541 : */
2542 692 : gs = linitial_node(GroupingSetData, current_sets);
2543 :
2544 : /*
2545 : * Order the groupClause appropriately. If the first grouping set is
2546 : * empty, then the groupClause must also be empty; otherwise we have
2547 : * to force the groupClause to match that grouping set's order.
2548 : *
2549 : * (The first grouping set can be empty even though parse->groupClause
2550 : * is not empty only if all non-empty grouping sets are unsortable.
2551 : * The groupClauses for hashed grouping sets are built later on.)
2552 : */
2553 692 : if (gs->set)
2554 664 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2555 : else
2556 28 : rollup->groupClause = NIL;
2557 :
2558 : /*
2559 : * Is it hashable? We pretend empty sets are hashable even though we
2560 : * actually force them not to be hashed later. But don't bother if
2561 : * there's nothing but empty sets (since in that case we can't hash
2562 : * anything).
2563 : */
2564 1356 : if (gs->set &&
2565 664 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2566 : {
2567 648 : rollup->hashable = true;
2568 648 : gd->any_hashable = true;
2569 : }
2570 :
2571 : /*
2572 : * Now that we've pinned down an order for the groupClause for this
2573 : * list of grouping sets, we need to remap the entries in the grouping
2574 : * sets from sortgrouprefs to plain indices (0-based) into the
2575 : * groupClause for this collection of grouping sets. We keep the
2576 : * original form for later use, though.
2577 : */
2578 692 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2579 : current_sets,
2580 : gd->tleref_to_colnum_map);
2581 692 : rollup->gsets_data = current_sets;
2582 :
2583 692 : gd->rollups = lappend(gd->rollups, rollup);
2584 : }
2585 :
2586 432 : if (gd->unsortable_sets)
2587 : {
2588 : /*
2589 : * We have not yet pinned down a groupclause for this, but we will
2590 : * need index-based lists for estimation purposes. Construct
2591 : * hash_sets_idx based on the entire original groupclause for now.
2592 : */
2593 24 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2594 : gd->unsortable_sets,
2595 : gd->tleref_to_colnum_map);
2596 24 : gd->any_hashable = true;
2597 : }
2598 :
2599 432 : return gd;
2600 : }
2601 :
2602 : /*
2603 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2604 : * (without annotation) mapped to indexes into the given groupclause.
2605 : */
2606 : static List *
2607 1892 : remap_to_groupclause_idx(List *groupClause,
2608 : List *gsets,
2609 : int *tleref_to_colnum_map)
2610 : {
2611 1892 : int ref = 0;
2612 1892 : List *result = NIL;
2613 : ListCell *lc;
2614 :
2615 4556 : foreach(lc, groupClause)
2616 : {
2617 2664 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2618 :
2619 2664 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2620 : }
2621 :
2622 4380 : foreach(lc, gsets)
2623 : {
2624 2488 : List *set = NIL;
2625 : ListCell *lc2;
2626 2488 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2627 :
2628 5448 : foreach(lc2, gs->set)
2629 : {
2630 2960 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2631 : }
2632 :
2633 2488 : result = lappend(result, set);
2634 : }
2635 :
2636 1892 : return result;
2637 : }
2638 :
2639 :
2640 : /*
2641 : * preprocess_rowmarks - set up PlanRowMarks if needed
2642 : */
2643 : static void
2644 301560 : preprocess_rowmarks(PlannerInfo *root)
2645 : {
2646 301560 : Query *parse = root->parse;
2647 : Bitmapset *rels;
2648 : List *prowmarks;
2649 : ListCell *l;
2650 : int i;
2651 :
2652 301560 : if (parse->rowMarks)
2653 : {
2654 : /*
2655 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2656 : * grouping, since grouping renders a reference to individual tuple
2657 : * CTIDs invalid. This is also checked at parse time, but that's
2658 : * insufficient because of rule substitution, query pullup, etc.
2659 : */
2660 4866 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2661 : parse->rowMarks)->strength);
2662 : }
2663 : else
2664 : {
2665 : /*
2666 : * We only need rowmarks for UPDATE, DELETE, or FOR [KEY]
2667 : * UPDATE/SHARE.
2668 : */
2669 583940 : if (parse->commandType != CMD_UPDATE &&
2670 287246 : parse->commandType != CMD_DELETE)
2671 284330 : return;
2672 : }
2673 :
2674 : /*
2675 : * We need to have rowmarks for all base relations except the target. We
2676 : * make a bitmapset of all base rels and then remove the items we don't
2677 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2678 : */
2679 17230 : rels = get_relids_in_jointree((Node *) parse->jointree, false);
2680 17230 : if (parse->resultRelation)
2681 12364 : rels = bms_del_member(rels, parse->resultRelation);
2682 :
2683 : /*
2684 : * Convert RowMarkClauses to PlanRowMark representation.
2685 : */
2686 17230 : prowmarks = NIL;
2687 22318 : foreach(l, parse->rowMarks)
2688 : {
2689 5088 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2690 5088 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2691 : PlanRowMark *newrc;
2692 :
2693 : /*
2694 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2695 : * applied to an update/delete target rel. If that ever becomes
2696 : * possible, we should drop the target from the PlanRowMark list.
2697 : */
2698 : Assert(rc->rti != parse->resultRelation);
2699 :
2700 : /*
2701 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2702 : * can't support true locking. Subqueries that got flattened into the
2703 : * main query should be ignored completely. Any that didn't will get
2704 : * ROW_MARK_COPY items in the next loop.
2705 : */
2706 5088 : if (rte->rtekind != RTE_RELATION)
2707 76 : continue;
2708 :
2709 5012 : rels = bms_del_member(rels, rc->rti);
2710 :
2711 5012 : newrc = makeNode(PlanRowMark);
2712 5012 : newrc->rti = newrc->prti = rc->rti;
2713 5012 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2714 5012 : newrc->markType = select_rowmark_type(rte, rc->strength);
2715 5012 : newrc->allMarkTypes = (1 << newrc->markType);
2716 5012 : newrc->strength = rc->strength;
2717 5012 : newrc->waitPolicy = rc->waitPolicy;
2718 5012 : newrc->isParent = false;
2719 :
2720 5012 : prowmarks = lappend(prowmarks, newrc);
2721 : }
2722 :
2723 : /*
2724 : * Now, add rowmarks for any non-target, non-locked base relations.
2725 : */
2726 17230 : i = 0;
2727 38136 : foreach(l, parse->rtable)
2728 : {
2729 20906 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2730 : PlanRowMark *newrc;
2731 :
2732 20906 : i++;
2733 20906 : if (!bms_is_member(i, rels))
2734 19902 : continue;
2735 :
2736 1004 : newrc = makeNode(PlanRowMark);
2737 1004 : newrc->rti = newrc->prti = i;
2738 1004 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2739 1004 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2740 1004 : newrc->allMarkTypes = (1 << newrc->markType);
2741 1004 : newrc->strength = LCS_NONE;
2742 1004 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2743 1004 : newrc->isParent = false;
2744 :
2745 1004 : prowmarks = lappend(prowmarks, newrc);
2746 : }
2747 :
2748 17230 : root->rowMarks = prowmarks;
2749 : }
2750 :
2751 : /*
2752 : * Select RowMarkType to use for a given table
2753 : */
2754 : RowMarkType
2755 7202 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2756 : {
2757 7202 : if (rte->rtekind != RTE_RELATION)
2758 : {
2759 : /* If it's not a table at all, use ROW_MARK_COPY */
2760 174 : return ROW_MARK_COPY;
2761 : }
2762 7028 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2763 : {
2764 : /* Let the FDW select the rowmark type, if it wants to */
2765 160 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2766 :
2767 160 : if (fdwroutine->GetForeignRowMarkType != NULL)
2768 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2769 : /* Otherwise, use ROW_MARK_COPY by default */
2770 160 : return ROW_MARK_COPY;
2771 : }
2772 : else
2773 : {
2774 : /* Regular table, apply the appropriate lock type */
2775 6868 : switch (strength)
2776 : {
2777 : case LCS_NONE:
2778 :
2779 : /*
2780 : * We don't need a tuple lock, only the ability to re-fetch
2781 : * the row.
2782 : */
2783 1146 : return ROW_MARK_REFERENCE;
2784 : break;
2785 : case LCS_FORKEYSHARE:
2786 4150 : return ROW_MARK_KEYSHARE;
2787 : break;
2788 : case LCS_FORSHARE:
2789 234 : return ROW_MARK_SHARE;
2790 : break;
2791 : case LCS_FORNOKEYUPDATE:
2792 58 : return ROW_MARK_NOKEYEXCLUSIVE;
2793 : break;
2794 : case LCS_FORUPDATE:
2795 1280 : return ROW_MARK_EXCLUSIVE;
2796 : break;
2797 : }
2798 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2799 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2800 : }
2801 : }
2802 :
2803 : /*
2804 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2805 : *
2806 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2807 : * results back in *count_est and *offset_est. These variables are set to
2808 : * 0 if the corresponding clause is not present, and -1 if it's present
2809 : * but we couldn't estimate the value for it. (The "0" convention is OK
2810 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2811 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2812 : * usual practice of never estimating less than one row.) These values will
2813 : * be passed to create_limit_path, which see if you change this code.
2814 : *
2815 : * The return value is the suitably adjusted tuple_fraction to use for
2816 : * planning the query. This adjustment is not overridable, since it reflects
2817 : * plan actions that grouping_planner() will certainly take, not assumptions
2818 : * about context.
2819 : */
2820 : static double
2821 3054 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2822 : int64 *offset_est, int64 *count_est)
2823 : {
2824 3054 : Query *parse = root->parse;
2825 : Node *est;
2826 : double limit_fraction;
2827 :
2828 : /* Should not be called unless LIMIT or OFFSET */
2829 : Assert(parse->limitCount || parse->limitOffset);
2830 :
2831 : /*
2832 : * Try to obtain the clause values. We use estimate_expression_value
2833 : * primarily because it can sometimes do something useful with Params.
2834 : */
2835 3054 : if (parse->limitCount)
2836 : {
2837 2836 : est = estimate_expression_value(root, parse->limitCount);
2838 2836 : if (est && IsA(est, Const))
2839 : {
2840 5664 : if (((Const *) est)->constisnull)
2841 : {
2842 : /* NULL indicates LIMIT ALL, ie, no limit */
2843 0 : *count_est = 0; /* treat as not present */
2844 : }
2845 : else
2846 : {
2847 2832 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2848 2832 : if (*count_est <= 0)
2849 96 : *count_est = 1; /* force to at least 1 */
2850 : }
2851 : }
2852 : else
2853 4 : *count_est = -1; /* can't estimate */
2854 : }
2855 : else
2856 218 : *count_est = 0; /* not present */
2857 :
2858 3054 : if (parse->limitOffset)
2859 : {
2860 536 : est = estimate_expression_value(root, parse->limitOffset);
2861 536 : if (est && IsA(est, Const))
2862 : {
2863 1040 : if (((Const *) est)->constisnull)
2864 : {
2865 : /* Treat NULL as no offset; the executor will too */
2866 0 : *offset_est = 0; /* treat as not present */
2867 : }
2868 : else
2869 : {
2870 520 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2871 520 : if (*offset_est < 0)
2872 0 : *offset_est = 0; /* treat as not present */
2873 : }
2874 : }
2875 : else
2876 16 : *offset_est = -1; /* can't estimate */
2877 : }
2878 : else
2879 2518 : *offset_est = 0; /* not present */
2880 :
2881 3054 : if (*count_est != 0)
2882 : {
2883 : /*
2884 : * A LIMIT clause limits the absolute number of tuples returned.
2885 : * However, if it's not a constant LIMIT then we have to guess; for
2886 : * lack of a better idea, assume 10% of the plan's result is wanted.
2887 : */
2888 2836 : if (*count_est < 0 || *offset_est < 0)
2889 : {
2890 : /* LIMIT or OFFSET is an expression ... punt ... */
2891 16 : limit_fraction = 0.10;
2892 : }
2893 : else
2894 : {
2895 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2896 2820 : limit_fraction = (double) *count_est + (double) *offset_est;
2897 : }
2898 :
2899 : /*
2900 : * If we have absolute limits from both caller and LIMIT, use the
2901 : * smaller value; likewise if they are both fractional. If one is
2902 : * fractional and the other absolute, we can't easily determine which
2903 : * is smaller, but we use the heuristic that the absolute will usually
2904 : * be smaller.
2905 : */
2906 2836 : if (tuple_fraction >= 1.0)
2907 : {
2908 4 : if (limit_fraction >= 1.0)
2909 : {
2910 : /* both absolute */
2911 4 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2912 : }
2913 : else
2914 : {
2915 : /* caller absolute, limit fractional; use caller's value */
2916 : }
2917 : }
2918 2832 : else if (tuple_fraction > 0.0)
2919 : {
2920 94 : if (limit_fraction >= 1.0)
2921 : {
2922 : /* caller fractional, limit absolute; use limit */
2923 94 : tuple_fraction = limit_fraction;
2924 : }
2925 : else
2926 : {
2927 : /* both fractional */
2928 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2929 : }
2930 : }
2931 : else
2932 : {
2933 : /* no info from caller, just use limit */
2934 2738 : tuple_fraction = limit_fraction;
2935 : }
2936 : }
2937 218 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2938 : {
2939 : /*
2940 : * We have an OFFSET but no LIMIT. This acts entirely differently
2941 : * from the LIMIT case: here, we need to increase rather than decrease
2942 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2943 : * tuples to be fetched instead of fewer. This only matters if we got
2944 : * a tuple_fraction > 0, however.
2945 : *
2946 : * As above, use 10% if OFFSET is present but unestimatable.
2947 : */
2948 8 : if (*offset_est < 0)
2949 0 : limit_fraction = 0.10;
2950 : else
2951 8 : limit_fraction = (double) *offset_est;
2952 :
2953 : /*
2954 : * If we have absolute counts from both caller and OFFSET, add them
2955 : * together; likewise if they are both fractional. If one is
2956 : * fractional and the other absolute, we want to take the larger, and
2957 : * we heuristically assume that's the fractional one.
2958 : */
2959 8 : if (tuple_fraction >= 1.0)
2960 : {
2961 0 : if (limit_fraction >= 1.0)
2962 : {
2963 : /* both absolute, so add them together */
2964 0 : tuple_fraction += limit_fraction;
2965 : }
2966 : else
2967 : {
2968 : /* caller absolute, limit fractional; use limit */
2969 0 : tuple_fraction = limit_fraction;
2970 : }
2971 : }
2972 : else
2973 : {
2974 8 : if (limit_fraction >= 1.0)
2975 : {
2976 : /* caller fractional, limit absolute; use caller's value */
2977 : }
2978 : else
2979 : {
2980 : /* both fractional, so add them together */
2981 0 : tuple_fraction += limit_fraction;
2982 0 : if (tuple_fraction >= 1.0)
2983 0 : tuple_fraction = 0.0; /* assume fetch all */
2984 : }
2985 : }
2986 : }
2987 :
2988 3054 : return tuple_fraction;
2989 : }
2990 :
2991 : /*
2992 : * limit_needed - do we actually need a Limit plan node?
2993 : *
2994 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2995 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2996 : * locution for an optimization fence. (Because other places in the planner
2997 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2998 : * an optimization fence --- we're just suppressing unnecessary run-time
2999 : * overhead.)
3000 : *
3001 : * This might look like it could be merged into preprocess_limit, but there's
3002 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3003 : * in preprocess_limit it's good enough to consider estimated values.
3004 : */
3005 : bool
3006 622738 : limit_needed(Query *parse)
3007 : {
3008 : Node *node;
3009 :
3010 622738 : node = parse->limitCount;
3011 622738 : if (node)
3012 : {
3013 6494 : if (IsA(node, Const))
3014 : {
3015 : /* NULL indicates LIMIT ALL, ie, no limit */
3016 6402 : if (!((Const *) node)->constisnull)
3017 6402 : return true; /* LIMIT with a constant value */
3018 : }
3019 : else
3020 92 : return true; /* non-constant LIMIT */
3021 : }
3022 :
3023 616244 : node = parse->limitOffset;
3024 616244 : if (node)
3025 : {
3026 648 : if (IsA(node, Const))
3027 : {
3028 : /* Treat NULL as no offset; the executor would too */
3029 502 : if (!((Const *) node)->constisnull)
3030 : {
3031 502 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3032 :
3033 502 : if (offset != 0)
3034 56 : return true; /* OFFSET with a nonzero value */
3035 : }
3036 : }
3037 : else
3038 146 : return true; /* non-constant OFFSET */
3039 : }
3040 :
3041 616042 : return false; /* don't need a Limit plan node */
3042 : }
3043 :
3044 :
3045 : /*
3046 : * remove_useless_groupby_columns
3047 : * Remove any columns in the GROUP BY clause that are redundant due to
3048 : * being functionally dependent on other GROUP BY columns.
3049 : *
3050 : * Since some other DBMSes do not allow references to ungrouped columns, it's
3051 : * not unusual to find all columns listed in GROUP BY even though listing the
3052 : * primary-key columns would be sufficient. Deleting such excess columns
3053 : * avoids redundant sorting work, so it's worth doing. When we do this, we
3054 : * must mark the plan as dependent on the pkey constraint (compare the
3055 : * parser's check_ungrouped_columns() and check_functional_grouping()).
3056 : *
3057 : * In principle, we could treat any NOT-NULL columns appearing in a UNIQUE
3058 : * index as the determining columns. But as with check_functional_grouping(),
3059 : * there's currently no way to represent dependency on a NOT NULL constraint,
3060 : * so we consider only the pkey for now.
3061 : */
3062 : static void
3063 300176 : remove_useless_groupby_columns(PlannerInfo *root)
3064 : {
3065 300176 : Query *parse = root->parse;
3066 : Bitmapset **groupbyattnos;
3067 : Bitmapset **surplusvars;
3068 : ListCell *lc;
3069 : int relid;
3070 :
3071 : /* No chance to do anything if there are less than two GROUP BY items */
3072 300176 : if (list_length(parse->groupClause) < 2)
3073 299384 : return;
3074 :
3075 : /* Don't fiddle with the GROUP BY clause if the query has grouping sets */
3076 792 : if (parse->groupingSets)
3077 352 : return;
3078 :
3079 : /*
3080 : * Scan the GROUP BY clause to find GROUP BY items that are simple Vars.
3081 : * Fill groupbyattnos[k] with a bitmapset of the column attnos of RTE k
3082 : * that are GROUP BY items.
3083 : */
3084 440 : groupbyattnos = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3085 440 : (list_length(parse->rtable) + 1));
3086 1576 : foreach(lc, parse->groupClause)
3087 : {
3088 1136 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3089 1136 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3090 1136 : Var *var = (Var *) tle->expr;
3091 :
3092 : /*
3093 : * Ignore non-Vars and Vars from other query levels.
3094 : *
3095 : * XXX in principle, stable expressions containing Vars could also be
3096 : * removed, if all the Vars are functionally dependent on other GROUP
3097 : * BY items. But it's not clear that such cases occur often enough to
3098 : * be worth troubling over.
3099 : */
3100 2116 : if (!IsA(var, Var) ||
3101 980 : var->varlevelsup > 0)
3102 156 : continue;
3103 :
3104 : /* OK, remember we have this Var */
3105 980 : relid = var->varno;
3106 : Assert(relid <= list_length(parse->rtable));
3107 1960 : groupbyattnos[relid] = bms_add_member(groupbyattnos[relid],
3108 980 : var->varattno - FirstLowInvalidHeapAttributeNumber);
3109 : }
3110 :
3111 : /*
3112 : * Consider each relation and see if it is possible to remove some of its
3113 : * Vars from GROUP BY. For simplicity and speed, we do the actual removal
3114 : * in a separate pass. Here, we just fill surplusvars[k] with a bitmapset
3115 : * of the column attnos of RTE k that are removable GROUP BY items.
3116 : */
3117 440 : surplusvars = NULL; /* don't allocate array unless required */
3118 440 : relid = 0;
3119 1444 : foreach(lc, parse->rtable)
3120 : {
3121 1004 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
3122 : Bitmapset *relattnos;
3123 : Bitmapset *pkattnos;
3124 : Oid constraintOid;
3125 :
3126 1004 : relid++;
3127 :
3128 : /* Only plain relations could have primary-key constraints */
3129 1004 : if (rte->rtekind != RTE_RELATION)
3130 1262 : continue;
3131 :
3132 : /*
3133 : * We must skip inheritance parent tables as some of the child rels
3134 : * may cause duplicate rows. This cannot happen with partitioned
3135 : * tables, however.
3136 : */
3137 700 : if (rte->inh && rte->relkind != RELKIND_PARTITIONED_TABLE)
3138 4 : continue;
3139 :
3140 : /* Nothing to do unless this rel has multiple Vars in GROUP BY */
3141 696 : relattnos = groupbyattnos[relid];
3142 696 : if (bms_membership(relattnos) != BMS_MULTIPLE)
3143 408 : continue;
3144 :
3145 : /*
3146 : * Can't remove any columns for this rel if there is no suitable
3147 : * (i.e., nondeferrable) primary key constraint.
3148 : */
3149 288 : pkattnos = get_primary_key_attnos(rte->relid, false, &constraintOid);
3150 288 : if (pkattnos == NULL)
3151 242 : continue;
3152 :
3153 : /*
3154 : * If the primary key is a proper subset of relattnos then we have
3155 : * some items in the GROUP BY that can be removed.
3156 : */
3157 46 : if (bms_subset_compare(pkattnos, relattnos) == BMS_SUBSET1)
3158 : {
3159 : /*
3160 : * To easily remember whether we've found anything to do, we don't
3161 : * allocate the surplusvars[] array until we find something.
3162 : */
3163 34 : if (surplusvars == NULL)
3164 30 : surplusvars = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3165 30 : (list_length(parse->rtable) + 1));
3166 :
3167 : /* Remember the attnos of the removable columns */
3168 34 : surplusvars[relid] = bms_difference(relattnos, pkattnos);
3169 :
3170 : /* Also, mark the resulting plan as dependent on this constraint */
3171 34 : parse->constraintDeps = lappend_oid(parse->constraintDeps,
3172 : constraintOid);
3173 : }
3174 : }
3175 :
3176 : /*
3177 : * If we found any surplus Vars, build a new GROUP BY clause without them.
3178 : * (Note: this may leave some TLEs with unreferenced ressortgroupref
3179 : * markings, but that's harmless.)
3180 : */
3181 440 : if (surplusvars != NULL)
3182 : {
3183 30 : List *new_groupby = NIL;
3184 :
3185 150 : foreach(lc, parse->groupClause)
3186 : {
3187 120 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3188 120 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3189 120 : Var *var = (Var *) tle->expr;
3190 :
3191 : /*
3192 : * New list must include non-Vars, outer Vars, and anything not
3193 : * marked as surplus.
3194 : */
3195 240 : if (!IsA(var, Var) ||
3196 240 : var->varlevelsup > 0 ||
3197 120 : !bms_is_member(var->varattno - FirstLowInvalidHeapAttributeNumber,
3198 120 : surplusvars[var->varno]))
3199 66 : new_groupby = lappend(new_groupby, sgc);
3200 : }
3201 :
3202 30 : parse->groupClause = new_groupby;
3203 : }
3204 : }
3205 :
3206 : /*
3207 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3208 : *
3209 : * The idea here is to adjust the ordering of the GROUP BY elements
3210 : * (which in itself is semantically insignificant) to match ORDER BY,
3211 : * thereby allowing a single sort operation to both implement the ORDER BY
3212 : * requirement and set up for a Unique step that implements GROUP BY.
3213 : *
3214 : * In principle it might be interesting to consider other orderings of the
3215 : * GROUP BY elements, which could match the sort ordering of other
3216 : * possible plans (eg an indexscan) and thereby reduce cost. We don't
3217 : * bother with that, though. Hashed grouping will frequently win anyway.
3218 : *
3219 : * Note: we need no comparable processing of the distinctClause because
3220 : * the parser already enforced that that matches ORDER BY.
3221 : *
3222 : * For grouping sets, the order of items is instead forced to agree with that
3223 : * of the grouping set (and items not in the grouping set are skipped). The
3224 : * work of sorting the order of grouping set elements to match the ORDER BY if
3225 : * possible is done elsewhere.
3226 : */
3227 : static List *
3228 3544 : preprocess_groupclause(PlannerInfo *root, List *force)
3229 : {
3230 3544 : Query *parse = root->parse;
3231 3544 : List *new_groupclause = NIL;
3232 : bool partial_match;
3233 : ListCell *sl;
3234 : ListCell *gl;
3235 :
3236 : /* For grouping sets, we need to force the ordering */
3237 3544 : if (force)
3238 : {
3239 4452 : foreach(sl, force)
3240 : {
3241 2612 : Index ref = lfirst_int(sl);
3242 2612 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3243 :
3244 2612 : new_groupclause = lappend(new_groupclause, cl);
3245 : }
3246 :
3247 1840 : return new_groupclause;
3248 : }
3249 :
3250 : /* If no ORDER BY, nothing useful to do here */
3251 1704 : if (parse->sortClause == NIL)
3252 762 : return parse->groupClause;
3253 :
3254 : /*
3255 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3256 : * items, but only as far as we can make a matching prefix.
3257 : *
3258 : * This code assumes that the sortClause contains no duplicate items.
3259 : */
3260 1852 : foreach(sl, parse->sortClause)
3261 : {
3262 1280 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3263 :
3264 1898 : foreach(gl, parse->groupClause)
3265 : {
3266 1528 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3267 :
3268 1528 : if (equal(gc, sc))
3269 : {
3270 910 : new_groupclause = lappend(new_groupclause, gc);
3271 910 : break;
3272 : }
3273 : }
3274 1280 : if (gl == NULL)
3275 370 : break; /* no match, so stop scanning */
3276 : }
3277 :
3278 : /* Did we match all of the ORDER BY list, or just some of it? */
3279 942 : partial_match = (sl != NULL);
3280 :
3281 : /* If no match at all, no point in reordering GROUP BY */
3282 942 : if (new_groupclause == NIL)
3283 144 : return parse->groupClause;
3284 :
3285 : /*
3286 : * Add any remaining GROUP BY items to the new list, but only if we were
3287 : * able to make a complete match. In other words, we only rearrange the
3288 : * GROUP BY list if the result is that one list is a prefix of the other
3289 : * --- otherwise there's no possibility of a common sort. Also, give up
3290 : * if there are any non-sortable GROUP BY items, since then there's no
3291 : * hope anyway.
3292 : */
3293 1740 : foreach(gl, parse->groupClause)
3294 : {
3295 992 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3296 :
3297 992 : if (list_member_ptr(new_groupclause, gc))
3298 894 : continue; /* it matched an ORDER BY item */
3299 98 : if (partial_match)
3300 100 : return parse->groupClause; /* give up, no common sort possible */
3301 48 : if (!OidIsValid(gc->sortop))
3302 0 : return parse->groupClause; /* give up, GROUP BY can't be sorted */
3303 48 : new_groupclause = lappend(new_groupclause, gc);
3304 : }
3305 :
3306 : /* Success --- install the rearranged GROUP BY list */
3307 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3308 748 : return new_groupclause;
3309 : }
3310 :
3311 : /*
3312 : * Extract lists of grouping sets that can be implemented using a single
3313 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3314 : *
3315 : * Input must be sorted with smallest sets first. Result has each sublist
3316 : * sorted with smallest sets first.
3317 : *
3318 : * We want to produce the absolute minimum possible number of lists here to
3319 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3320 : * of finding the minimal partition of a partially-ordered set into chains
3321 : * (which is what we need, taking the list of grouping sets as a poset ordered
3322 : * by set inclusion) can be mapped to the problem of finding the maximum
3323 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3324 : * time with a worst case of no worse than O(n^2.5) and usually much
3325 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3326 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3327 : * half a second on my modest system even with optimization off and assertions
3328 : * on.)
3329 : */
3330 : static List *
3331 428 : extract_rollup_sets(List *groupingSets)
3332 : {
3333 428 : int num_sets_raw = list_length(groupingSets);
3334 428 : int num_empty = 0;
3335 428 : int num_sets = 0; /* distinct sets */
3336 428 : int num_chains = 0;
3337 428 : List *result = NIL;
3338 : List **results;
3339 : List **orig_sets;
3340 : Bitmapset **set_masks;
3341 : int *chains;
3342 : short **adjacency;
3343 : short *adjacency_buf;
3344 : BipartiteMatchState *state;
3345 : int i;
3346 : int j;
3347 : int j_size;
3348 428 : ListCell *lc1 = list_head(groupingSets);
3349 : ListCell *lc;
3350 :
3351 : /*
3352 : * Start by stripping out empty sets. The algorithm doesn't require this,
3353 : * but the planner currently needs all empty sets to be returned in the
3354 : * first list, so we strip them here and add them back after.
3355 : */
3356 1192 : while (lc1 && lfirst(lc1) == NIL)
3357 : {
3358 336 : ++num_empty;
3359 336 : lc1 = lnext(groupingSets, lc1);
3360 : }
3361 :
3362 : /* bail out now if it turns out that all we had were empty sets. */
3363 428 : if (!lc1)
3364 28 : return list_make1(groupingSets);
3365 :
3366 : /*----------
3367 : * We don't strictly need to remove duplicate sets here, but if we don't,
3368 : * they tend to become scattered through the result, which is a bit
3369 : * confusing (and irritating if we ever decide to optimize them out).
3370 : * So we remove them here and add them back after.
3371 : *
3372 : * For each non-duplicate set, we fill in the following:
3373 : *
3374 : * orig_sets[i] = list of the original set lists
3375 : * set_masks[i] = bitmapset for testing inclusion
3376 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3377 : *
3378 : * chains[i] will be the result group this set is assigned to.
3379 : *
3380 : * We index all of these from 1 rather than 0 because it is convenient
3381 : * to leave 0 free for the NIL node in the graph algorithm.
3382 : *----------
3383 : */
3384 400 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3385 400 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3386 400 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3387 400 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3388 :
3389 400 : j_size = 0;
3390 400 : j = 0;
3391 400 : i = 1;
3392 :
3393 1348 : for_each_cell(lc, groupingSets, lc1)
3394 : {
3395 948 : List *candidate = (List *) lfirst(lc);
3396 948 : Bitmapset *candidate_set = NULL;
3397 : ListCell *lc2;
3398 948 : int dup_of = 0;
3399 :
3400 2264 : foreach(lc2, candidate)
3401 : {
3402 1316 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3403 : }
3404 :
3405 : /* we can only be a dup if we're the same length as a previous set */
3406 948 : if (j_size == list_length(candidate))
3407 : {
3408 : int k;
3409 :
3410 824 : for (k = j; k < i; ++k)
3411 : {
3412 528 : if (bms_equal(set_masks[k], candidate_set))
3413 : {
3414 52 : dup_of = k;
3415 52 : break;
3416 : }
3417 : }
3418 : }
3419 600 : else if (j_size < list_length(candidate))
3420 : {
3421 600 : j_size = list_length(candidate);
3422 600 : j = i;
3423 : }
3424 :
3425 948 : if (dup_of > 0)
3426 : {
3427 52 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3428 52 : bms_free(candidate_set);
3429 : }
3430 : else
3431 : {
3432 : int k;
3433 896 : int n_adj = 0;
3434 :
3435 896 : orig_sets[i] = list_make1(candidate);
3436 896 : set_masks[i] = candidate_set;
3437 :
3438 : /* fill in adjacency list; no need to compare equal-size sets */
3439 :
3440 1360 : for (k = j - 1; k > 0; --k)
3441 : {
3442 464 : if (bms_is_subset(set_masks[k], candidate_set))
3443 404 : adjacency_buf[++n_adj] = k;
3444 : }
3445 :
3446 896 : if (n_adj > 0)
3447 : {
3448 232 : adjacency_buf[0] = n_adj;
3449 232 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3450 232 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3451 : }
3452 : else
3453 664 : adjacency[i] = NULL;
3454 :
3455 896 : ++i;
3456 : }
3457 : }
3458 :
3459 400 : num_sets = i - 1;
3460 :
3461 : /*
3462 : * Apply the graph matching algorithm to do the work.
3463 : */
3464 400 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3465 :
3466 : /*
3467 : * Now, the state->pair* fields have the info we need to assign sets to
3468 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3469 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3470 : * it in one pass)
3471 : */
3472 400 : chains = palloc0((num_sets + 1) * sizeof(int));
3473 :
3474 1296 : for (i = 1; i <= num_sets; ++i)
3475 : {
3476 896 : int u = state->pair_vu[i];
3477 896 : int v = state->pair_uv[i];
3478 :
3479 896 : if (u > 0 && u < i)
3480 0 : chains[i] = chains[u];
3481 896 : else if (v > 0 && v < i)
3482 232 : chains[i] = chains[v];
3483 : else
3484 664 : chains[i] = ++num_chains;
3485 : }
3486 :
3487 : /* build result lists. */
3488 400 : results = palloc0((num_chains + 1) * sizeof(List *));
3489 :
3490 1296 : for (i = 1; i <= num_sets; ++i)
3491 : {
3492 896 : int c = chains[i];
3493 :
3494 : Assert(c > 0);
3495 :
3496 896 : results[c] = list_concat(results[c], orig_sets[i]);
3497 : }
3498 :
3499 : /* push any empty sets back on the first list. */
3500 1076 : while (num_empty-- > 0)
3501 276 : results[1] = lcons(NIL, results[1]);
3502 :
3503 : /* make result list */
3504 1064 : for (i = 1; i <= num_chains; ++i)
3505 664 : result = lappend(result, results[i]);
3506 :
3507 : /*
3508 : * Free all the things.
3509 : *
3510 : * (This is over-fussy for small sets but for large sets we could have
3511 : * tied up a nontrivial amount of memory.)
3512 : */
3513 400 : BipartiteMatchFree(state);
3514 400 : pfree(results);
3515 400 : pfree(chains);
3516 1296 : for (i = 1; i <= num_sets; ++i)
3517 896 : if (adjacency[i])
3518 232 : pfree(adjacency[i]);
3519 400 : pfree(adjacency);
3520 400 : pfree(adjacency_buf);
3521 400 : pfree(orig_sets);
3522 1296 : for (i = 1; i <= num_sets; ++i)
3523 896 : bms_free(set_masks[i]);
3524 400 : pfree(set_masks);
3525 :
3526 400 : return result;
3527 : }
3528 :
3529 : /*
3530 : * Reorder the elements of a list of grouping sets such that they have correct
3531 : * prefix relationships. Also inserts the GroupingSetData annotations.
3532 : *
3533 : * The input must be ordered with smallest sets first; the result is returned
3534 : * with largest sets first. Note that the result shares no list substructure
3535 : * with the input, so it's safe for the caller to modify it later.
3536 : *
3537 : * If we're passed in a sortclause, we follow its order of columns to the
3538 : * extent possible, to minimize the chance that we add unnecessary sorts.
3539 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3540 : * gets implemented in one pass.)
3541 : */
3542 : static List *
3543 692 : reorder_grouping_sets(List *groupingsets, List *sortclause)
3544 : {
3545 : ListCell *lc;
3546 692 : List *previous = NIL;
3547 692 : List *result = NIL;
3548 :
3549 1976 : foreach(lc, groupingsets)
3550 : {
3551 1284 : List *candidate = (List *) lfirst(lc);
3552 1284 : List *new_elems = list_difference_int(candidate, previous);
3553 1284 : GroupingSetData *gs = makeNode(GroupingSetData);
3554 :
3555 2860 : while (list_length(sortclause) > list_length(previous) &&
3556 236 : list_length(new_elems) > 0)
3557 : {
3558 128 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3559 128 : int ref = sc->tleSortGroupRef;
3560 :
3561 128 : if (list_member_int(new_elems, ref))
3562 : {
3563 56 : previous = lappend_int(previous, ref);
3564 56 : new_elems = list_delete_int(new_elems, ref);
3565 : }
3566 : else
3567 : {
3568 : /* diverged from the sortclause; give up on it */
3569 72 : sortclause = NIL;
3570 72 : break;
3571 : }
3572 : }
3573 :
3574 1284 : previous = list_concat(previous, new_elems);
3575 :
3576 1284 : gs->set = list_copy(previous);
3577 1284 : result = lcons(gs, result);
3578 : }
3579 :
3580 692 : list_free(previous);
3581 :
3582 692 : return result;
3583 : }
3584 :
3585 : /*
3586 : * Compute query_pathkeys and other pathkeys during plan generation
3587 : */
3588 : static void
3589 301338 : standard_qp_callback(PlannerInfo *root, void *extra)
3590 : {
3591 301338 : Query *parse = root->parse;
3592 301338 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3593 301338 : List *tlist = root->processed_tlist;
3594 301338 : List *activeWindows = qp_extra->activeWindows;
3595 :
3596 : /*
3597 : * Calculate pathkeys that represent grouping/ordering requirements. The
3598 : * sortClause is certainly sort-able, but GROUP BY and DISTINCT might not
3599 : * be, in which case we just leave their pathkeys empty.
3600 : */
3601 303442 : if (qp_extra->groupClause &&
3602 2104 : grouping_is_sortable(qp_extra->groupClause))
3603 2104 : root->group_pathkeys =
3604 2104 : make_pathkeys_for_sortclauses(root,
3605 : qp_extra->groupClause,
3606 : tlist);
3607 : else
3608 299234 : root->group_pathkeys = NIL;
3609 :
3610 : /* We consider only the first (bottom) window in pathkeys logic */
3611 301338 : if (activeWindows != NIL)
3612 : {
3613 1112 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3614 :
3615 1112 : root->window_pathkeys = make_pathkeys_for_window(root,
3616 : wc,
3617 : tlist);
3618 : }
3619 : else
3620 300226 : root->window_pathkeys = NIL;
3621 :
3622 301724 : if (parse->distinctClause &&
3623 386 : grouping_is_sortable(parse->distinctClause))
3624 382 : root->distinct_pathkeys =
3625 382 : make_pathkeys_for_sortclauses(root,
3626 : parse->distinctClause,
3627 : tlist);
3628 : else
3629 300956 : root->distinct_pathkeys = NIL;
3630 :
3631 301338 : root->sort_pathkeys =
3632 301338 : make_pathkeys_for_sortclauses(root,
3633 : parse->sortClause,
3634 : tlist);
3635 :
3636 : /*
3637 : * Figure out whether we want a sorted result from query_planner.
3638 : *
3639 : * If we have a sortable GROUP BY clause, then we want a result sorted
3640 : * properly for grouping. Otherwise, if we have window functions to
3641 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3642 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3643 : * we try to produce output that's sufficiently well sorted for the
3644 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3645 : * by the ORDER BY clause.
3646 : *
3647 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3648 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3649 : * that might just leave us failing to exploit an available sort order at
3650 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3651 : * much easier, since we know that the parser ensured that one is a
3652 : * superset of the other.
3653 : */
3654 301338 : if (root->group_pathkeys)
3655 2030 : root->query_pathkeys = root->group_pathkeys;
3656 299308 : else if (root->window_pathkeys)
3657 968 : root->query_pathkeys = root->window_pathkeys;
3658 596680 : else if (list_length(root->distinct_pathkeys) >
3659 298340 : list_length(root->sort_pathkeys))
3660 236 : root->query_pathkeys = root->distinct_pathkeys;
3661 298104 : else if (root->sort_pathkeys)
3662 34346 : root->query_pathkeys = root->sort_pathkeys;
3663 : else
3664 263758 : root->query_pathkeys = NIL;
3665 301338 : }
3666 :
3667 : /*
3668 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3669 : *
3670 : * path_rows: number of output rows from scan/join step
3671 : * gd: grouping sets data including list of grouping sets and their clauses
3672 : * target_list: target list containing group clause references
3673 : *
3674 : * If doing grouping sets, we also annotate the gsets data with the estimates
3675 : * for each set and each individual rollup list, with a view to later
3676 : * determining whether some combination of them could be hashed instead.
3677 : */
3678 : static double
3679 25780 : get_number_of_groups(PlannerInfo *root,
3680 : double path_rows,
3681 : grouping_sets_data *gd,
3682 : List *target_list)
3683 : {
3684 25780 : Query *parse = root->parse;
3685 : double dNumGroups;
3686 :
3687 25780 : if (parse->groupClause)
3688 : {
3689 : List *groupExprs;
3690 :
3691 3658 : if (parse->groupingSets)
3692 : {
3693 : /* Add up the estimates for each grouping set */
3694 : ListCell *lc;
3695 : ListCell *lc2;
3696 :
3697 : Assert(gd); /* keep Coverity happy */
3698 :
3699 404 : dNumGroups = 0;
3700 :
3701 1068 : foreach(lc, gd->rollups)
3702 : {
3703 664 : RollupData *rollup = lfirst_node(RollupData, lc);
3704 : ListCell *lc;
3705 :
3706 664 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3707 : target_list);
3708 :
3709 664 : rollup->numGroups = 0.0;
3710 :
3711 1888 : forboth(lc, rollup->gsets, lc2, rollup->gsets_data)
3712 : {
3713 1224 : List *gset = (List *) lfirst(lc);
3714 1224 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3715 1224 : double numGroups = estimate_num_groups(root,
3716 : groupExprs,
3717 : path_rows,
3718 : &gset);
3719 :
3720 1224 : gs->numGroups = numGroups;
3721 1224 : rollup->numGroups += numGroups;
3722 : }
3723 :
3724 664 : dNumGroups += rollup->numGroups;
3725 : }
3726 :
3727 404 : if (gd->hash_sets_idx)
3728 : {
3729 : ListCell *lc;
3730 :
3731 24 : gd->dNumHashGroups = 0;
3732 :
3733 24 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3734 : target_list);
3735 :
3736 52 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3737 : {
3738 28 : List *gset = (List *) lfirst(lc);
3739 28 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3740 28 : double numGroups = estimate_num_groups(root,
3741 : groupExprs,
3742 : path_rows,
3743 : &gset);
3744 :
3745 28 : gs->numGroups = numGroups;
3746 28 : gd->dNumHashGroups += numGroups;
3747 : }
3748 :
3749 24 : dNumGroups += gd->dNumHashGroups;
3750 : }
3751 : }
3752 : else
3753 : {
3754 : /* Plain GROUP BY */
3755 3254 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3756 : target_list);
3757 :
3758 3254 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3759 : NULL);
3760 : }
3761 : }
3762 22122 : else if (parse->groupingSets)
3763 : {
3764 : /* Empty grouping sets ... one result row for each one */
3765 28 : dNumGroups = list_length(parse->groupingSets);
3766 : }
3767 22094 : else if (parse->hasAggs || root->hasHavingQual)
3768 : {
3769 : /* Plain aggregation, one result row */
3770 22094 : dNumGroups = 1;
3771 : }
3772 : else
3773 : {
3774 : /* Not grouping */
3775 0 : dNumGroups = 1;
3776 : }
3777 :
3778 25780 : return dNumGroups;
3779 : }
3780 :
3781 : /*
3782 : * create_grouping_paths
3783 : *
3784 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3785 : * Along the way, we also build an upperrel for Paths which are partially
3786 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3787 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3788 : * the only partially grouped paths we build are also partial paths; that
3789 : * is, they need a Gather and then a FinalizeAggregate.
3790 : *
3791 : * input_rel: contains the source-data Paths
3792 : * target: the pathtarget for the result Paths to compute
3793 : * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
3794 : * gd: grouping sets data including list of grouping sets and their clauses
3795 : *
3796 : * Note: all Paths in input_rel are expected to return the target computed
3797 : * by make_group_input_target.
3798 : */
3799 : static RelOptInfo *
3800 23820 : create_grouping_paths(PlannerInfo *root,
3801 : RelOptInfo *input_rel,
3802 : PathTarget *target,
3803 : bool target_parallel_safe,
3804 : const AggClauseCosts *agg_costs,
3805 : grouping_sets_data *gd)
3806 : {
3807 23820 : Query *parse = root->parse;
3808 : RelOptInfo *grouped_rel;
3809 : RelOptInfo *partially_grouped_rel;
3810 :
3811 : /*
3812 : * Create grouping relation to hold fully aggregated grouping and/or
3813 : * aggregation paths.
3814 : */
3815 23820 : grouped_rel = make_grouping_rel(root, input_rel, target,
3816 : target_parallel_safe, parse->havingQual);
3817 :
3818 : /*
3819 : * Create either paths for a degenerate grouping or paths for ordinary
3820 : * grouping, as appropriate.
3821 : */
3822 23820 : if (is_degenerate_grouping(root))
3823 12 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3824 : else
3825 : {
3826 23808 : int flags = 0;
3827 : GroupPathExtraData extra;
3828 :
3829 : /*
3830 : * Determine whether it's possible to perform sort-based
3831 : * implementations of grouping. (Note that if groupClause is empty,
3832 : * grouping_is_sortable() is trivially true, and all the
3833 : * pathkeys_contained_in() tests will succeed too, so that we'll
3834 : * consider every surviving input path.)
3835 : *
3836 : * If we have grouping sets, we might be able to sort some but not all
3837 : * of them; in this case, we need can_sort to be true as long as we
3838 : * must consider any sorted-input plan.
3839 : */
3840 23808 : if ((gd && gd->rollups != NIL)
3841 23380 : || grouping_is_sortable(parse->groupClause))
3842 23804 : flags |= GROUPING_CAN_USE_SORT;
3843 :
3844 : /*
3845 : * Determine whether we should consider hash-based implementations of
3846 : * grouping.
3847 : *
3848 : * Hashed aggregation only applies if we're grouping. If we have
3849 : * grouping sets, some groups might be hashable but others not; in
3850 : * this case we set can_hash true as long as there is nothing globally
3851 : * preventing us from hashing (and we should therefore consider plans
3852 : * with hashes).
3853 : *
3854 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3855 : * BY aggregates. (Doing so would imply storing *all* the input
3856 : * values in the hash table, and/or running many sorts in parallel,
3857 : * either of which seems like a certain loser.) We similarly don't
3858 : * support ordered-set aggregates in hashed aggregation, but that case
3859 : * is also included in the numOrderedAggs count.
3860 : *
3861 : * Note: grouping_is_hashable() is much more expensive to check than
3862 : * the other gating conditions, so we want to do it last.
3863 : */
3864 25916 : if ((parse->groupClause != NIL &&
3865 6028 : agg_costs->numOrderedAggs == 0 &&
3866 1960 : (gd ? gd->any_hashable : grouping_is_hashable(parse->groupClause))))
3867 1956 : flags |= GROUPING_CAN_USE_HASH;
3868 :
3869 : /*
3870 : * Determine whether partial aggregation is possible.
3871 : */
3872 23808 : if (can_partial_agg(root, agg_costs))
3873 8144 : flags |= GROUPING_CAN_PARTIAL_AGG;
3874 :
3875 23808 : extra.flags = flags;
3876 23808 : extra.target_parallel_safe = target_parallel_safe;
3877 23808 : extra.havingQual = parse->havingQual;
3878 23808 : extra.targetList = parse->targetList;
3879 23808 : extra.partial_costs_set = false;
3880 :
3881 : /*
3882 : * Determine whether partitionwise aggregation is in theory possible.
3883 : * It can be disabled by the user, and for now, we don't try to
3884 : * support grouping sets. create_ordinary_grouping_paths() will check
3885 : * additional conditions, such as whether input_rel is partitioned.
3886 : */
3887 23808 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3888 302 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3889 : else
3890 23506 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3891 :
3892 23808 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3893 : agg_costs, gd, &extra,
3894 : &partially_grouped_rel);
3895 : }
3896 :
3897 23816 : set_cheapest(grouped_rel);
3898 23816 : return grouped_rel;
3899 : }
3900 :
3901 : /*
3902 : * make_grouping_rel
3903 : *
3904 : * Create a new grouping rel and set basic properties.
3905 : *
3906 : * input_rel represents the underlying scan/join relation.
3907 : * target is the output expected from the grouping relation.
3908 : */
3909 : static RelOptInfo *
3910 24762 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3911 : PathTarget *target, bool target_parallel_safe,
3912 : Node *havingQual)
3913 : {
3914 : RelOptInfo *grouped_rel;
3915 :
3916 24762 : if (IS_OTHER_REL(input_rel))
3917 : {
3918 942 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3919 : input_rel->relids);
3920 942 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3921 : }
3922 : else
3923 : {
3924 : /*
3925 : * By tradition, the relids set for the main grouping relation is
3926 : * NULL. (This could be changed, but might require adjustments
3927 : * elsewhere.)
3928 : */
3929 23820 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3930 : }
3931 :
3932 : /* Set target. */
3933 24762 : grouped_rel->reltarget = target;
3934 :
3935 : /*
3936 : * If the input relation is not parallel-safe, then the grouped relation
3937 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3938 : * target list and HAVING quals are parallel-safe.
3939 : */
3940 36172 : if (input_rel->consider_parallel && target_parallel_safe &&
3941 11410 : is_parallel_safe(root, (Node *) havingQual))
3942 11398 : grouped_rel->consider_parallel = true;
3943 :
3944 : /*
3945 : * If the input rel belongs to a single FDW, so does the grouped rel.
3946 : */
3947 24762 : grouped_rel->serverid = input_rel->serverid;
3948 24762 : grouped_rel->userid = input_rel->userid;
3949 24762 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3950 24762 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3951 :
3952 24762 : return grouped_rel;
3953 : }
3954 :
3955 : /*
3956 : * is_degenerate_grouping
3957 : *
3958 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3959 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3960 : * grouping sets are all empty).
3961 : */
3962 : static bool
3963 23820 : is_degenerate_grouping(PlannerInfo *root)
3964 : {
3965 23820 : Query *parse = root->parse;
3966 :
3967 48036 : return (root->hasHavingQual || parse->groupingSets) &&
3968 24648 : !parse->hasAggs && parse->groupClause == NIL;
3969 : }
3970 :
3971 : /*
3972 : * create_degenerate_grouping_paths
3973 : *
3974 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3975 : * supposed to emit either zero or one row for each grouping set depending on
3976 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3977 : * either HAVING or the targetlist, so we actually do not need the FROM table
3978 : * at all! We can just throw away the plan-so-far and generate a Result node.
3979 : * This is a sufficiently unusual corner case that it's not worth contorting
3980 : * the structure of this module to avoid having to generate the earlier paths
3981 : * in the first place.
3982 : */
3983 : static void
3984 12 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3985 : RelOptInfo *grouped_rel)
3986 : {
3987 12 : Query *parse = root->parse;
3988 : int nrows;
3989 : Path *path;
3990 :
3991 12 : nrows = list_length(parse->groupingSets);
3992 12 : if (nrows > 1)
3993 : {
3994 : /*
3995 : * Doesn't seem worthwhile writing code to cons up a generate_series
3996 : * or a values scan to emit multiple rows. Instead just make N clones
3997 : * and append them. (With a volatile HAVING clause, this means you
3998 : * might get between 0 and N output rows. Offhand I think that's
3999 : * desired.)
4000 : */
4001 0 : List *paths = NIL;
4002 :
4003 0 : while (--nrows >= 0)
4004 : {
4005 0 : path = (Path *)
4006 : create_group_result_path(root, grouped_rel,
4007 0 : grouped_rel->reltarget,
4008 0 : (List *) parse->havingQual);
4009 0 : paths = lappend(paths, path);
4010 : }
4011 0 : path = (Path *)
4012 : create_append_path(root,
4013 : grouped_rel,
4014 : paths,
4015 : NIL,
4016 : NIL,
4017 : NULL,
4018 : 0,
4019 : false,
4020 : NIL,
4021 : -1);
4022 : }
4023 : else
4024 : {
4025 : /* No grouping sets, or just one, so one output row */
4026 12 : path = (Path *)
4027 : create_group_result_path(root, grouped_rel,
4028 12 : grouped_rel->reltarget,
4029 12 : (List *) parse->havingQual);
4030 : }
4031 :
4032 12 : add_path(grouped_rel, path);
4033 12 : }
4034 :
4035 : /*
4036 : * create_ordinary_grouping_paths
4037 : *
4038 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4039 : *
4040 : * We need to consider sorted and hashed aggregation in the same function,
4041 : * because otherwise (1) it would be harder to throw an appropriate error
4042 : * message if neither way works, and (2) we should not allow hashtable size
4043 : * considerations to dissuade us from using hashing if sorting is not possible.
4044 : *
4045 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4046 : * function creates, or to NULL if it doesn't create one.
4047 : */
4048 : static void
4049 24750 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4050 : RelOptInfo *grouped_rel,
4051 : const AggClauseCosts *agg_costs,
4052 : grouping_sets_data *gd,
4053 : GroupPathExtraData *extra,
4054 : RelOptInfo **partially_grouped_rel_p)
4055 : {
4056 24750 : Path *cheapest_path = input_rel->cheapest_total_path;
4057 24750 : RelOptInfo *partially_grouped_rel = NULL;
4058 : double dNumGroups;
4059 24750 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4060 :
4061 : /*
4062 : * If this is the topmost grouping relation or if the parent relation is
4063 : * doing some form of partitionwise aggregation, then we may be able to do
4064 : * it at this level also. However, if the input relation is not
4065 : * partitioned, partitionwise aggregate is impossible.
4066 : */
4067 25994 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4068 1650 : IS_PARTITIONED_REL(input_rel))
4069 : {
4070 : /*
4071 : * If this is the topmost relation or if the parent relation is doing
4072 : * full partitionwise aggregation, then we can do full partitionwise
4073 : * aggregation provided that the GROUP BY clause contains all of the
4074 : * partitioning columns at this level. Otherwise, we can do at most
4075 : * partial partitionwise aggregation. But if partial aggregation is
4076 : * not supported in general then we can't use it for partitionwise
4077 : * aggregation either.
4078 : */
4079 716 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4080 342 : group_by_has_partkey(input_rel, extra->targetList,
4081 342 : root->parse->groupClause))
4082 192 : patype = PARTITIONWISE_AGGREGATE_FULL;
4083 182 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4084 154 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4085 : else
4086 28 : patype = PARTITIONWISE_AGGREGATE_NONE;
4087 : }
4088 :
4089 : /*
4090 : * Before generating paths for grouped_rel, we first generate any possible
4091 : * partially grouped paths; that way, later code can easily consider both
4092 : * parallel and non-parallel approaches to grouping.
4093 : */
4094 24750 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4095 : {
4096 : bool force_rel_creation;
4097 :
4098 : /*
4099 : * If we're doing partitionwise aggregation at this level, force
4100 : * creation of a partially_grouped_rel so we can add partitionwise
4101 : * paths to it.
4102 : */
4103 9038 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4104 :
4105 9038 : partially_grouped_rel =
4106 9038 : create_partial_grouping_paths(root,
4107 : grouped_rel,
4108 : input_rel,
4109 : gd,
4110 : extra,
4111 : force_rel_creation);
4112 : }
4113 :
4114 : /* Set out parameter. */
4115 24750 : *partially_grouped_rel_p = partially_grouped_rel;
4116 :
4117 : /* Apply partitionwise aggregation technique, if possible. */
4118 24750 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4119 346 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4120 : partially_grouped_rel, agg_costs,
4121 : gd, patype, extra);
4122 :
4123 : /* If we are doing partial aggregation only, return. */
4124 24750 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4125 : {
4126 : Assert(partially_grouped_rel);
4127 :
4128 398 : if (partially_grouped_rel->pathlist)
4129 398 : set_cheapest(partially_grouped_rel);
4130 :
4131 398 : return;
4132 : }
4133 :
4134 : /* Gather any partially grouped partial paths. */
4135 24352 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4136 : {
4137 830 : gather_grouping_paths(root, partially_grouped_rel);
4138 830 : set_cheapest(partially_grouped_rel);
4139 : }
4140 :
4141 : /*
4142 : * Estimate number of groups.
4143 : */
4144 24352 : dNumGroups = get_number_of_groups(root,
4145 : cheapest_path->rows,
4146 : gd,
4147 : extra->targetList);
4148 :
4149 : /* Build final grouping paths */
4150 24352 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4151 : partially_grouped_rel, agg_costs, gd,
4152 : dNumGroups, extra);
4153 :
4154 : /* Give a helpful error if we failed to find any implementation */
4155 24352 : if (grouped_rel->pathlist == NIL)
4156 4 : ereport(ERROR,
4157 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4158 : errmsg("could not implement GROUP BY"),
4159 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4160 :
4161 : /*
4162 : * If there is an FDW that's responsible for all baserels of the query,
4163 : * let it consider adding ForeignPaths.
4164 : */
4165 24622 : if (grouped_rel->fdwroutine &&
4166 274 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4167 274 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4168 : input_rel, grouped_rel,
4169 : extra);
4170 :
4171 : /* Let extensions possibly add some more paths */
4172 24348 : if (create_upper_paths_hook)
4173 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4174 : input_rel, grouped_rel,
4175 : extra);
4176 : }
4177 :
4178 : /*
4179 : * For a given input path, consider the possible ways of doing grouping sets on
4180 : * it, by combinations of hashing and sorting. This can be called multiple
4181 : * times, so it's important that it not scribble on input. No result is
4182 : * returned, but any generated paths are added to grouped_rel.
4183 : */
4184 : static void
4185 844 : consider_groupingsets_paths(PlannerInfo *root,
4186 : RelOptInfo *grouped_rel,
4187 : Path *path,
4188 : bool is_sorted,
4189 : bool can_hash,
4190 : grouping_sets_data *gd,
4191 : const AggClauseCosts *agg_costs,
4192 : double dNumGroups)
4193 : {
4194 844 : Query *parse = root->parse;
4195 :
4196 : /*
4197 : * If we're not being offered sorted input, then only consider plans that
4198 : * can be done entirely by hashing.
4199 : *
4200 : * We can hash everything if it looks like it'll fit in work_mem. But if
4201 : * the input is actually sorted despite not being advertised as such, we
4202 : * prefer to make use of that in order to use less memory.
4203 : *
4204 : * If none of the grouping sets are sortable, then ignore the work_mem
4205 : * limit and generate a path anyway, since otherwise we'll just fail.
4206 : */
4207 844 : if (!is_sorted)
4208 : {
4209 380 : List *new_rollups = NIL;
4210 380 : RollupData *unhashed_rollup = NULL;
4211 : List *sets_data;
4212 380 : List *empty_sets_data = NIL;
4213 380 : List *empty_sets = NIL;
4214 : ListCell *lc;
4215 380 : ListCell *l_start = list_head(gd->rollups);
4216 380 : AggStrategy strat = AGG_HASHED;
4217 : double hashsize;
4218 380 : double exclude_groups = 0.0;
4219 :
4220 : Assert(can_hash);
4221 :
4222 : /*
4223 : * If the input is coincidentally sorted usefully (which can happen
4224 : * even if is_sorted is false, since that only means that our caller
4225 : * has set up the sorting for us), then save some hashtable space by
4226 : * making use of that. But we need to watch out for degenerate cases:
4227 : *
4228 : * 1) If there are any empty grouping sets, then group_pathkeys might
4229 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4230 : * there will be a rollup containing only empty groups, and the
4231 : * pathkeys_contained_in test is vacuously true; this is ok.
4232 : *
4233 : * XXX: the above relies on the fact that group_pathkeys is generated
4234 : * from the first rollup. If we add the ability to consider multiple
4235 : * sort orders for grouping input, this assumption might fail.
4236 : *
4237 : * 2) If there are no empty sets and only unsortable sets, then the
4238 : * rollups list will be empty (and thus l_start == NULL), and
4239 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4240 : * pathkeys_contained_in test doesn't cause us to crash.
4241 : */
4242 756 : if (l_start != NULL &&
4243 376 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4244 : {
4245 8 : unhashed_rollup = lfirst_node(RollupData, l_start);
4246 8 : exclude_groups = unhashed_rollup->numGroups;
4247 8 : l_start = lnext(gd->rollups, l_start);
4248 : }
4249 :
4250 380 : hashsize = estimate_hashagg_tablesize(path,
4251 : agg_costs,
4252 : dNumGroups - exclude_groups);
4253 :
4254 : /*
4255 : * gd->rollups is empty if we have only unsortable columns to work
4256 : * with. Override work_mem in that case; otherwise, we'll rely on the
4257 : * sorted-input case to generate usable mixed paths.
4258 : */
4259 380 : if (hashsize > work_mem * 1024L && gd->rollups)
4260 12 : return; /* nope, won't fit */
4261 :
4262 : /*
4263 : * We need to burst the existing rollups list into individual grouping
4264 : * sets and recompute a groupClause for each set.
4265 : */
4266 368 : sets_data = list_copy(gd->unsortable_sets);
4267 :
4268 904 : for_each_cell(lc, gd->rollups, l_start)
4269 : {
4270 552 : RollupData *rollup = lfirst_node(RollupData, lc);
4271 :
4272 : /*
4273 : * If we find an unhashable rollup that's not been skipped by the
4274 : * "actually sorted" check above, we can't cope; we'd need sorted
4275 : * input (with a different sort order) but we can't get that here.
4276 : * So bail out; we'll get a valid path from the is_sorted case
4277 : * instead.
4278 : *
4279 : * The mere presence of empty grouping sets doesn't make a rollup
4280 : * unhashable (see preprocess_grouping_sets), we handle those
4281 : * specially below.
4282 : */
4283 552 : if (!rollup->hashable)
4284 32 : return;
4285 :
4286 536 : sets_data = list_concat(sets_data, rollup->gsets_data);
4287 : }
4288 1420 : foreach(lc, sets_data)
4289 : {
4290 1068 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4291 1068 : List *gset = gs->set;
4292 : RollupData *rollup;
4293 :
4294 1068 : if (gset == NIL)
4295 : {
4296 : /* Empty grouping sets can't be hashed. */
4297 252 : empty_sets_data = lappend(empty_sets_data, gs);
4298 252 : empty_sets = lappend(empty_sets, NIL);
4299 : }
4300 : else
4301 : {
4302 816 : rollup = makeNode(RollupData);
4303 :
4304 816 : rollup->groupClause = preprocess_groupclause(root, gset);
4305 816 : rollup->gsets_data = list_make1(gs);
4306 816 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4307 : rollup->gsets_data,
4308 : gd->tleref_to_colnum_map);
4309 816 : rollup->numGroups = gs->numGroups;
4310 816 : rollup->hashable = true;
4311 816 : rollup->is_hashed = true;
4312 816 : new_rollups = lappend(new_rollups, rollup);
4313 : }
4314 : }
4315 :
4316 : /*
4317 : * If we didn't find anything nonempty to hash, then bail. We'll
4318 : * generate a path from the is_sorted case.
4319 : */
4320 352 : if (new_rollups == NIL)
4321 0 : return;
4322 :
4323 : /*
4324 : * If there were empty grouping sets they should have been in the
4325 : * first rollup.
4326 : */
4327 : Assert(!unhashed_rollup || !empty_sets);
4328 :
4329 352 : if (unhashed_rollup)
4330 : {
4331 8 : new_rollups = lappend(new_rollups, unhashed_rollup);
4332 8 : strat = AGG_MIXED;
4333 : }
4334 344 : else if (empty_sets)
4335 : {
4336 220 : RollupData *rollup = makeNode(RollupData);
4337 :
4338 220 : rollup->groupClause = NIL;
4339 220 : rollup->gsets_data = empty_sets_data;
4340 220 : rollup->gsets = empty_sets;
4341 220 : rollup->numGroups = list_length(empty_sets);
4342 220 : rollup->hashable = false;
4343 220 : rollup->is_hashed = false;
4344 220 : new_rollups = lappend(new_rollups, rollup);
4345 220 : strat = AGG_MIXED;
4346 : }
4347 :
4348 352 : add_path(grouped_rel, (Path *)
4349 352 : create_groupingsets_path(root,
4350 : grouped_rel,
4351 : path,
4352 352 : (List *) parse->havingQual,
4353 : strat,
4354 : new_rollups,
4355 : agg_costs,
4356 : dNumGroups));
4357 352 : return;
4358 : }
4359 :
4360 : /*
4361 : * If we have sorted input but nothing we can do with it, bail.
4362 : */
4363 464 : if (list_length(gd->rollups) == 0)
4364 0 : return;
4365 :
4366 : /*
4367 : * Given sorted input, we try and make two paths: one sorted and one mixed
4368 : * sort/hash. (We need to try both because hashagg might be disabled, or
4369 : * some columns might not be sortable.)
4370 : *
4371 : * can_hash is passed in as false if some obstacle elsewhere (such as
4372 : * ordered aggs) means that we shouldn't consider hashing at all.
4373 : */
4374 464 : if (can_hash && gd->any_hashable)
4375 : {
4376 412 : List *rollups = NIL;
4377 412 : List *hash_sets = list_copy(gd->unsortable_sets);
4378 412 : double availspace = (work_mem * 1024.0);
4379 : ListCell *lc;
4380 :
4381 : /*
4382 : * Account first for space needed for groups we can't sort at all.
4383 : */
4384 412 : availspace -= estimate_hashagg_tablesize(path,
4385 : agg_costs,
4386 : gd->dNumHashGroups);
4387 :
4388 412 : if (availspace > 0 && list_length(gd->rollups) > 1)
4389 : {
4390 : double scale;
4391 208 : int num_rollups = list_length(gd->rollups);
4392 : int k_capacity;
4393 208 : int *k_weights = palloc(num_rollups * sizeof(int));
4394 208 : Bitmapset *hash_items = NULL;
4395 : int i;
4396 :
4397 : /*
4398 : * We treat this as a knapsack problem: the knapsack capacity
4399 : * represents work_mem, the item weights are the estimated memory
4400 : * usage of the hashtables needed to implement a single rollup,
4401 : * and we really ought to use the cost saving as the item value;
4402 : * however, currently the costs assigned to sort nodes don't
4403 : * reflect the comparison costs well, and so we treat all items as
4404 : * of equal value (each rollup we hash instead saves us one sort).
4405 : *
4406 : * To use the discrete knapsack, we need to scale the values to a
4407 : * reasonably small bounded range. We choose to allow a 5% error
4408 : * margin; we have no more than 4096 rollups in the worst possible
4409 : * case, which with a 5% error margin will require a bit over 42MB
4410 : * of workspace. (Anyone wanting to plan queries that complex had
4411 : * better have the memory for it. In more reasonable cases, with
4412 : * no more than a couple of dozen rollups, the memory usage will
4413 : * be negligible.)
4414 : *
4415 : * k_capacity is naturally bounded, but we clamp the values for
4416 : * scale and weight (below) to avoid overflows or underflows (or
4417 : * uselessly trying to use a scale factor less than 1 byte).
4418 : */
4419 208 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4420 208 : k_capacity = (int) floor(availspace / scale);
4421 :
4422 : /*
4423 : * We leave the first rollup out of consideration since it's the
4424 : * one that matches the input sort order. We assign indexes "i"
4425 : * to only those entries considered for hashing; the second loop,
4426 : * below, must use the same condition.
4427 : */
4428 208 : i = 0;
4429 544 : for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4430 : {
4431 336 : RollupData *rollup = lfirst_node(RollupData, lc);
4432 :
4433 336 : if (rollup->hashable)
4434 : {
4435 336 : double sz = estimate_hashagg_tablesize(path,
4436 : agg_costs,
4437 : rollup->numGroups);
4438 :
4439 : /*
4440 : * If sz is enormous, but work_mem (and hence scale) is
4441 : * small, avoid integer overflow here.
4442 : */
4443 336 : k_weights[i] = (int) Min(floor(sz / scale),
4444 : k_capacity + 1.0);
4445 336 : ++i;
4446 : }
4447 : }
4448 :
4449 : /*
4450 : * Apply knapsack algorithm; compute the set of items which
4451 : * maximizes the value stored (in this case the number of sorts
4452 : * saved) while keeping the total size (approximately) within
4453 : * capacity.
4454 : */
4455 208 : if (i > 0)
4456 208 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4457 :
4458 208 : if (!bms_is_empty(hash_items))
4459 : {
4460 208 : rollups = list_make1(linitial(gd->rollups));
4461 :
4462 208 : i = 0;
4463 544 : for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4464 : {
4465 336 : RollupData *rollup = lfirst_node(RollupData, lc);
4466 :
4467 336 : if (rollup->hashable)
4468 : {
4469 336 : if (bms_is_member(i, hash_items))
4470 312 : hash_sets = list_concat(hash_sets,
4471 312 : rollup->gsets_data);
4472 : else
4473 24 : rollups = lappend(rollups, rollup);
4474 336 : ++i;
4475 : }
4476 : else
4477 0 : rollups = lappend(rollups, rollup);
4478 : }
4479 : }
4480 : }
4481 :
4482 412 : if (!rollups && hash_sets)
4483 16 : rollups = list_copy(gd->rollups);
4484 :
4485 772 : foreach(lc, hash_sets)
4486 : {
4487 360 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4488 360 : RollupData *rollup = makeNode(RollupData);
4489 :
4490 : Assert(gs->set != NIL);
4491 :
4492 360 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4493 360 : rollup->gsets_data = list_make1(gs);
4494 360 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4495 : rollup->gsets_data,
4496 : gd->tleref_to_colnum_map);
4497 360 : rollup->numGroups = gs->numGroups;
4498 360 : rollup->hashable = true;
4499 360 : rollup->is_hashed = true;
4500 360 : rollups = lcons(rollup, rollups);
4501 : }
4502 :
4503 412 : if (rollups)
4504 : {
4505 224 : add_path(grouped_rel, (Path *)
4506 224 : create_groupingsets_path(root,
4507 : grouped_rel,
4508 : path,
4509 224 : (List *) parse->havingQual,
4510 : AGG_MIXED,
4511 : rollups,
4512 : agg_costs,
4513 : dNumGroups));
4514 : }
4515 : }
4516 :
4517 : /*
4518 : * Now try the simple sorted case.
4519 : */
4520 464 : if (!gd->unsortable_sets)
4521 444 : add_path(grouped_rel, (Path *)
4522 888 : create_groupingsets_path(root,
4523 : grouped_rel,
4524 : path,
4525 444 : (List *) parse->havingQual,
4526 : AGG_SORTED,
4527 : gd->rollups,
4528 : agg_costs,
4529 : dNumGroups));
4530 : }
4531 :
4532 : /*
4533 : * create_window_paths
4534 : *
4535 : * Build a new upperrel containing Paths for window-function evaluation.
4536 : *
4537 : * input_rel: contains the source-data Paths
4538 : * input_target: result of make_window_input_target
4539 : * output_target: what the topmost WindowAggPath should return
4540 : * wflists: result of find_window_functions
4541 : * activeWindows: result of select_active_windows
4542 : *
4543 : * Note: all Paths in input_rel are expected to return input_target.
4544 : */
4545 : static RelOptInfo *
4546 1112 : create_window_paths(PlannerInfo *root,
4547 : RelOptInfo *input_rel,
4548 : PathTarget *input_target,
4549 : PathTarget *output_target,
4550 : bool output_target_parallel_safe,
4551 : WindowFuncLists *wflists,
4552 : List *activeWindows)
4553 : {
4554 : RelOptInfo *window_rel;
4555 : ListCell *lc;
4556 :
4557 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4558 1112 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4559 :
4560 : /*
4561 : * If the input relation is not parallel-safe, then the window relation
4562 : * can't be parallel-safe, either. Otherwise, we need to examine the
4563 : * target list and active windows for non-parallel-safe constructs.
4564 : */
4565 1112 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4566 0 : is_parallel_safe(root, (Node *) activeWindows))
4567 0 : window_rel->consider_parallel = true;
4568 :
4569 : /*
4570 : * If the input rel belongs to a single FDW, so does the window rel.
4571 : */
4572 1112 : window_rel->serverid = input_rel->serverid;
4573 1112 : window_rel->userid = input_rel->userid;
4574 1112 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4575 1112 : window_rel->fdwroutine = input_rel->fdwroutine;
4576 :
4577 : /*
4578 : * Consider computing window functions starting from the existing
4579 : * cheapest-total path (which will likely require a sort) as well as any
4580 : * existing paths that satisfy root->window_pathkeys (which won't).
4581 : */
4582 2324 : foreach(lc, input_rel->pathlist)
4583 : {
4584 1212 : Path *path = (Path *) lfirst(lc);
4585 :
4586 1312 : if (path == input_rel->cheapest_total_path ||
4587 100 : pathkeys_contained_in(root->window_pathkeys, path->pathkeys))
4588 1156 : create_one_window_path(root,
4589 : window_rel,
4590 : path,
4591 : input_target,
4592 : output_target,
4593 : wflists,
4594 : activeWindows);
4595 : }
4596 :
4597 : /*
4598 : * If there is an FDW that's responsible for all baserels of the query,
4599 : * let it consider adding ForeignPaths.
4600 : */
4601 1124 : if (window_rel->fdwroutine &&
4602 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4603 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4604 : input_rel, window_rel,
4605 : NULL);
4606 :
4607 : /* Let extensions possibly add some more paths */
4608 1112 : if (create_upper_paths_hook)
4609 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4610 : input_rel, window_rel, NULL);
4611 :
4612 : /* Now choose the best path(s) */
4613 1112 : set_cheapest(window_rel);
4614 :
4615 1112 : return window_rel;
4616 : }
4617 :
4618 : /*
4619 : * Stack window-function implementation steps atop the given Path, and
4620 : * add the result to window_rel.
4621 : *
4622 : * window_rel: upperrel to contain result
4623 : * path: input Path to use (must return input_target)
4624 : * input_target: result of make_window_input_target
4625 : * output_target: what the topmost WindowAggPath should return
4626 : * wflists: result of find_window_functions
4627 : * activeWindows: result of select_active_windows
4628 : */
4629 : static void
4630 1156 : create_one_window_path(PlannerInfo *root,
4631 : RelOptInfo *window_rel,
4632 : Path *path,
4633 : PathTarget *input_target,
4634 : PathTarget *output_target,
4635 : WindowFuncLists *wflists,
4636 : List *activeWindows)
4637 : {
4638 : PathTarget *window_target;
4639 : ListCell *l;
4640 :
4641 : /*
4642 : * Since each window clause could require a different sort order, we stack
4643 : * up a WindowAgg node for each clause, with sort steps between them as
4644 : * needed. (We assume that select_active_windows chose a good order for
4645 : * executing the clauses in.)
4646 : *
4647 : * input_target should contain all Vars and Aggs needed for the result.
4648 : * (In some cases we wouldn't need to propagate all of these all the way
4649 : * to the top, since they might only be needed as inputs to WindowFuncs.
4650 : * It's probably not worth trying to optimize that though.) It must also
4651 : * contain all window partitioning and sorting expressions, to ensure
4652 : * they're computed only once at the bottom of the stack (that's critical
4653 : * for volatile functions). As we climb up the stack, we'll add outputs
4654 : * for the WindowFuncs computed at each level.
4655 : */
4656 1156 : window_target = input_target;
4657 :
4658 2360 : foreach(l, activeWindows)
4659 : {
4660 1204 : WindowClause *wc = lfirst_node(WindowClause, l);
4661 : List *window_pathkeys;
4662 :
4663 1204 : window_pathkeys = make_pathkeys_for_window(root,
4664 : wc,
4665 : root->processed_tlist);
4666 :
4667 : /* Sort if necessary */
4668 1204 : if (!pathkeys_contained_in(window_pathkeys, path->pathkeys))
4669 : {
4670 952 : path = (Path *) create_sort_path(root, window_rel,
4671 : path,
4672 : window_pathkeys,
4673 : -1.0);
4674 : }
4675 :
4676 1204 : if (lnext(activeWindows, l))
4677 : {
4678 : /*
4679 : * Add the current WindowFuncs to the output target for this
4680 : * intermediate WindowAggPath. We must copy window_target to
4681 : * avoid changing the previous path's target.
4682 : *
4683 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4684 : * we do need to account for the increase in tlist width.
4685 : */
4686 : ListCell *lc2;
4687 :
4688 48 : window_target = copy_pathtarget(window_target);
4689 96 : foreach(lc2, wflists->windowFuncs[wc->winref])
4690 : {
4691 48 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4692 :
4693 48 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4694 48 : window_target->width += get_typavgwidth(wfunc->wintype, -1);
4695 : }
4696 : }
4697 : else
4698 : {
4699 : /* Install the goal target in the topmost WindowAgg */
4700 1156 : window_target = output_target;
4701 : }
4702 :
4703 1204 : path = (Path *)
4704 1204 : create_windowagg_path(root, window_rel, path, window_target,
4705 1204 : wflists->windowFuncs[wc->winref],
4706 : wc);
4707 : }
4708 :
4709 1156 : add_path(window_rel, path);
4710 1156 : }
4711 :
4712 : /*
4713 : * create_distinct_paths
4714 : *
4715 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4716 : *
4717 : * input_rel: contains the source-data Paths
4718 : *
4719 : * Note: input paths should already compute the desired pathtarget, since
4720 : * Sort/Unique won't project anything.
4721 : */
4722 : static RelOptInfo *
4723 386 : create_distinct_paths(PlannerInfo *root,
4724 : RelOptInfo *input_rel)
4725 : {
4726 386 : Query *parse = root->parse;
4727 386 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4728 : RelOptInfo *distinct_rel;
4729 : double numDistinctRows;
4730 : bool allow_hash;
4731 : Path *path;
4732 : ListCell *lc;
4733 :
4734 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4735 386 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4736 :
4737 : /*
4738 : * We don't compute anything at this level, so distinct_rel will be
4739 : * parallel-safe if the input rel is parallel-safe. In particular, if
4740 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4741 : * output those expressions, and will not be parallel-safe unless those
4742 : * expressions are parallel-safe.
4743 : */
4744 386 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4745 :
4746 : /*
4747 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4748 : */
4749 386 : distinct_rel->serverid = input_rel->serverid;
4750 386 : distinct_rel->userid = input_rel->userid;
4751 386 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4752 386 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4753 :
4754 : /* Estimate number of distinct rows there will be */
4755 748 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4756 362 : root->hasHavingQual)
4757 : {
4758 : /*
4759 : * If there was grouping or aggregation, use the number of input rows
4760 : * as the estimated number of DISTINCT rows (ie, assume the input is
4761 : * already mostly unique).
4762 : */
4763 24 : numDistinctRows = cheapest_input_path->rows;
4764 : }
4765 : else
4766 : {
4767 : /*
4768 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4769 : */
4770 : List *distinctExprs;
4771 :
4772 362 : distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
4773 : parse->targetList);
4774 362 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4775 : cheapest_input_path->rows,
4776 : NULL);
4777 : }
4778 :
4779 : /*
4780 : * Consider sort-based implementations of DISTINCT, if possible.
4781 : */
4782 386 : if (grouping_is_sortable(parse->distinctClause))
4783 : {
4784 : /*
4785 : * First, if we have any adequately-presorted paths, just stick a
4786 : * Unique node on those. Then consider doing an explicit sort of the
4787 : * cheapest input path and Unique'ing that.
4788 : *
4789 : * When we have DISTINCT ON, we must sort by the more rigorous of
4790 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
4791 : * Also, if we do have to do an explicit sort, we might as well use
4792 : * the more rigorous ordering to avoid a second sort later. (Note
4793 : * that the parser will have ensured that one clause is a prefix of
4794 : * the other.)
4795 : */
4796 : List *needed_pathkeys;
4797 :
4798 418 : if (parse->hasDistinctOn &&
4799 36 : list_length(root->distinct_pathkeys) <
4800 36 : list_length(root->sort_pathkeys))
4801 20 : needed_pathkeys = root->sort_pathkeys;
4802 : else
4803 362 : needed_pathkeys = root->distinct_pathkeys;
4804 :
4805 804 : foreach(lc, input_rel->pathlist)
4806 : {
4807 422 : Path *path = (Path *) lfirst(lc);
4808 :
4809 422 : if (pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4810 : {
4811 30 : add_path(distinct_rel, (Path *)
4812 30 : create_upper_unique_path(root, distinct_rel,
4813 : path,
4814 30 : list_length(root->distinct_pathkeys),
4815 : numDistinctRows));
4816 : }
4817 : }
4818 :
4819 : /* For explicit-sort case, always use the more rigorous clause */
4820 764 : if (list_length(root->distinct_pathkeys) <
4821 382 : list_length(root->sort_pathkeys))
4822 : {
4823 20 : needed_pathkeys = root->sort_pathkeys;
4824 : /* Assert checks that parser didn't mess up... */
4825 : Assert(pathkeys_contained_in(root->distinct_pathkeys,
4826 : needed_pathkeys));
4827 : }
4828 : else
4829 362 : needed_pathkeys = root->distinct_pathkeys;
4830 :
4831 382 : path = cheapest_input_path;
4832 382 : if (!pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4833 372 : path = (Path *) create_sort_path(root, distinct_rel,
4834 : path,
4835 : needed_pathkeys,
4836 : -1.0);
4837 :
4838 382 : add_path(distinct_rel, (Path *)
4839 382 : create_upper_unique_path(root, distinct_rel,
4840 : path,
4841 382 : list_length(root->distinct_pathkeys),
4842 : numDistinctRows));
4843 : }
4844 :
4845 : /*
4846 : * Consider hash-based implementations of DISTINCT, if possible.
4847 : *
4848 : * If we were not able to make any other types of path, we *must* hash or
4849 : * die trying. If we do have other choices, there are several things that
4850 : * should prevent selection of hashing: if the query uses DISTINCT ON
4851 : * (because it won't really have the expected behavior if we hash), or if
4852 : * enable_hashagg is off, or if it looks like the hashtable will exceed
4853 : * work_mem.
4854 : *
4855 : * Note: grouping_is_hashable() is much more expensive to check than the
4856 : * other gating conditions, so we want to do it last.
4857 : */
4858 386 : if (distinct_rel->pathlist == NIL)
4859 4 : allow_hash = true; /* we have no alternatives */
4860 382 : else if (parse->hasDistinctOn || !enable_hashagg)
4861 36 : allow_hash = false; /* policy-based decision not to hash */
4862 : else
4863 : {
4864 : Size hashentrysize;
4865 :
4866 : /* Estimate per-hash-entry space at tuple width... */
4867 346 : hashentrysize = MAXALIGN(cheapest_input_path->pathtarget->width) +
4868 : MAXALIGN(SizeofMinimalTupleHeader);
4869 : /* plus the per-hash-entry overhead */
4870 346 : hashentrysize += hash_agg_entry_size(0);
4871 :
4872 : /* Allow hashing only if hashtable is predicted to fit in work_mem */
4873 346 : allow_hash = (hashentrysize * numDistinctRows <= work_mem * 1024L);
4874 : }
4875 :
4876 386 : if (allow_hash && grouping_is_hashable(parse->distinctClause))
4877 : {
4878 : /* Generate hashed aggregate path --- no sort needed */
4879 350 : add_path(distinct_rel, (Path *)
4880 350 : create_agg_path(root,
4881 : distinct_rel,
4882 : cheapest_input_path,
4883 : cheapest_input_path->pathtarget,
4884 : AGG_HASHED,
4885 : AGGSPLIT_SIMPLE,
4886 : parse->distinctClause,
4887 : NIL,
4888 : NULL,
4889 : numDistinctRows));
4890 : }
4891 :
4892 : /* Give a helpful error if we failed to find any implementation */
4893 386 : if (distinct_rel->pathlist == NIL)
4894 0 : ereport(ERROR,
4895 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4896 : errmsg("could not implement DISTINCT"),
4897 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4898 :
4899 : /*
4900 : * If there is an FDW that's responsible for all baserels of the query,
4901 : * let it consider adding ForeignPaths.
4902 : */
4903 402 : if (distinct_rel->fdwroutine &&
4904 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4905 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_DISTINCT,
4906 : input_rel, distinct_rel,
4907 : NULL);
4908 :
4909 : /* Let extensions possibly add some more paths */
4910 386 : if (create_upper_paths_hook)
4911 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT,
4912 : input_rel, distinct_rel, NULL);
4913 :
4914 : /* Now choose the best path(s) */
4915 386 : set_cheapest(distinct_rel);
4916 :
4917 386 : return distinct_rel;
4918 : }
4919 :
4920 : /*
4921 : * create_ordered_paths
4922 : *
4923 : * Build a new upperrel containing Paths for ORDER BY evaluation.
4924 : *
4925 : * All paths in the result must satisfy the ORDER BY ordering.
4926 : * The only new path we need consider is an explicit sort on the
4927 : * cheapest-total existing path.
4928 : *
4929 : * input_rel: contains the source-data Paths
4930 : * target: the output tlist the result Paths must emit
4931 : * limit_tuples: estimated bound on the number of output tuples,
4932 : * or -1 if no LIMIT or couldn't estimate
4933 : */
4934 : static RelOptInfo *
4935 36116 : create_ordered_paths(PlannerInfo *root,
4936 : RelOptInfo *input_rel,
4937 : PathTarget *target,
4938 : bool target_parallel_safe,
4939 : double limit_tuples)
4940 : {
4941 36116 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4942 : RelOptInfo *ordered_rel;
4943 : ListCell *lc;
4944 :
4945 : /* For now, do all work in the (ORDERED, NULL) upperrel */
4946 36116 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
4947 :
4948 : /*
4949 : * If the input relation is not parallel-safe, then the ordered relation
4950 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4951 : * target list is parallel-safe.
4952 : */
4953 36116 : if (input_rel->consider_parallel && target_parallel_safe)
4954 15430 : ordered_rel->consider_parallel = true;
4955 :
4956 : /*
4957 : * If the input rel belongs to a single FDW, so does the ordered_rel.
4958 : */
4959 36116 : ordered_rel->serverid = input_rel->serverid;
4960 36116 : ordered_rel->userid = input_rel->userid;
4961 36116 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
4962 36116 : ordered_rel->fdwroutine = input_rel->fdwroutine;
4963 :
4964 85988 : foreach(lc, input_rel->pathlist)
4965 : {
4966 49872 : Path *path = (Path *) lfirst(lc);
4967 : bool is_sorted;
4968 :
4969 49872 : is_sorted = pathkeys_contained_in(root->sort_pathkeys,
4970 : path->pathkeys);
4971 49872 : if (path == cheapest_input_path || is_sorted)
4972 : {
4973 48600 : if (!is_sorted)
4974 : {
4975 : /* An explicit sort here can take advantage of LIMIT */
4976 28816 : path = (Path *) create_sort_path(root,
4977 : ordered_rel,
4978 : path,
4979 : root->sort_pathkeys,
4980 : limit_tuples);
4981 : }
4982 :
4983 : /* Add projection step if needed */
4984 48600 : if (path->pathtarget != target)
4985 8810 : path = apply_projection_to_path(root, ordered_rel,
4986 : path, target);
4987 :
4988 48600 : add_path(ordered_rel, path);
4989 : }
4990 : }
4991 :
4992 : /*
4993 : * generate_gather_paths() will have already generated a simple Gather
4994 : * path for the best parallel path, if any, and the loop above will have
4995 : * considered sorting it. Similarly, generate_gather_paths() will also
4996 : * have generated order-preserving Gather Merge plans which can be used
4997 : * without sorting if they happen to match the sort_pathkeys, and the loop
4998 : * above will have handled those as well. However, there's one more
4999 : * possibility: it may make sense to sort the cheapest partial path
5000 : * according to the required output order and then use Gather Merge.
5001 : */
5002 51438 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5003 15322 : input_rel->partial_pathlist != NIL)
5004 : {
5005 : Path *cheapest_partial_path;
5006 :
5007 1822 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5008 :
5009 : /*
5010 : * If cheapest partial path doesn't need a sort, this is redundant
5011 : * with what's already been tried.
5012 : */
5013 1822 : if (!pathkeys_contained_in(root->sort_pathkeys,
5014 : cheapest_partial_path->pathkeys))
5015 : {
5016 : Path *path;
5017 : double total_groups;
5018 :
5019 1814 : path = (Path *) create_sort_path(root,
5020 : ordered_rel,
5021 : cheapest_partial_path,
5022 : root->sort_pathkeys,
5023 : limit_tuples);
5024 :
5025 3628 : total_groups = cheapest_partial_path->rows *
5026 1814 : cheapest_partial_path->parallel_workers;
5027 1814 : path = (Path *)
5028 1814 : create_gather_merge_path(root, ordered_rel,
5029 : path,
5030 : path->pathtarget,
5031 : root->sort_pathkeys, NULL,
5032 : &total_groups);
5033 :
5034 : /* Add projection step if needed */
5035 1814 : if (path->pathtarget != target)
5036 1242 : path = apply_projection_to_path(root, ordered_rel,
5037 : path, target);
5038 :
5039 1814 : add_path(ordered_rel, path);
5040 : }
5041 : }
5042 :
5043 : /*
5044 : * If there is an FDW that's responsible for all baserels of the query,
5045 : * let it consider adding ForeignPaths.
5046 : */
5047 36436 : if (ordered_rel->fdwroutine &&
5048 320 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5049 306 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5050 : input_rel, ordered_rel,
5051 : NULL);
5052 :
5053 : /* Let extensions possibly add some more paths */
5054 36116 : if (create_upper_paths_hook)
5055 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5056 : input_rel, ordered_rel, NULL);
5057 :
5058 : /*
5059 : * No need to bother with set_cheapest here; grouping_planner does not
5060 : * need us to do it.
5061 : */
5062 : Assert(ordered_rel->pathlist != NIL);
5063 :
5064 36116 : return ordered_rel;
5065 : }
5066 :
5067 :
5068 : /*
5069 : * make_group_input_target
5070 : * Generate appropriate PathTarget for initial input to grouping nodes.
5071 : *
5072 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5073 : * the query's final targetlist; for example, it certainly can't emit any
5074 : * aggregate function calls. This routine generates the correct target
5075 : * for the scan/join subplan.
5076 : *
5077 : * The query target list passed from the parser already contains entries
5078 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5079 : * for variables used only in HAVING clauses; so we need to add those
5080 : * variables to the subplan target list. Also, we flatten all expressions
5081 : * except GROUP BY items into their component variables; other expressions
5082 : * will be computed by the upper plan nodes rather than by the subplan.
5083 : * For example, given a query like
5084 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5085 : * we want to pass this targetlist to the subplan:
5086 : * a+b,c,d
5087 : * where the a+b target will be used by the Sort/Group steps, and the
5088 : * other targets will be used for computing the final results.
5089 : *
5090 : * 'final_target' is the query's final target list (in PathTarget form)
5091 : *
5092 : * The result is the PathTarget to be computed by the Paths returned from
5093 : * query_planner().
5094 : */
5095 : static PathTarget *
5096 23820 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5097 : {
5098 23820 : Query *parse = root->parse;
5099 : PathTarget *input_target;
5100 : List *non_group_cols;
5101 : List *non_group_vars;
5102 : int i;
5103 : ListCell *lc;
5104 :
5105 : /*
5106 : * We must build a target containing all grouping columns, plus any other
5107 : * Vars mentioned in the query's targetlist and HAVING qual.
5108 : */
5109 23820 : input_target = create_empty_pathtarget();
5110 23820 : non_group_cols = NIL;
5111 :
5112 23820 : i = 0;
5113 53476 : foreach(lc, final_target->exprs)
5114 : {
5115 29656 : Expr *expr = (Expr *) lfirst(lc);
5116 29656 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5117 :
5118 33770 : if (sgref && parse->groupClause &&
5119 4114 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5120 : {
5121 : /*
5122 : * It's a grouping column, so add it to the input target as-is.
5123 : */
5124 3242 : add_column_to_pathtarget(input_target, expr, sgref);
5125 : }
5126 : else
5127 : {
5128 : /*
5129 : * Non-grouping column, so just remember the expression for later
5130 : * call to pull_var_clause.
5131 : */
5132 26414 : non_group_cols = lappend(non_group_cols, expr);
5133 : }
5134 :
5135 29656 : i++;
5136 : }
5137 :
5138 : /*
5139 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5140 : */
5141 23820 : if (parse->havingQual)
5142 324 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5143 :
5144 : /*
5145 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5146 : * add them to the input target if not already present. (A Var used
5147 : * directly as a GROUP BY item will be present already.) Note this
5148 : * includes Vars used in resjunk items, so we are covering the needs of
5149 : * ORDER BY and window specifications. Vars used within Aggrefs and
5150 : * WindowFuncs will be pulled out here, too.
5151 : */
5152 23820 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5153 : PVC_RECURSE_AGGREGATES |
5154 : PVC_RECURSE_WINDOWFUNCS |
5155 : PVC_INCLUDE_PLACEHOLDERS);
5156 23820 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5157 :
5158 : /* clean up cruft */
5159 23820 : list_free(non_group_vars);
5160 23820 : list_free(non_group_cols);
5161 :
5162 : /* XXX this causes some redundant cost calculation ... */
5163 23820 : return set_pathtarget_cost_width(root, input_target);
5164 : }
5165 :
5166 : /*
5167 : * make_partial_grouping_target
5168 : * Generate appropriate PathTarget for output of partial aggregate
5169 : * (or partial grouping, if there are no aggregates) nodes.
5170 : *
5171 : * A partial aggregation node needs to emit all the same aggregates that
5172 : * a regular aggregation node would, plus any aggregates used in HAVING;
5173 : * except that the Aggref nodes should be marked as partial aggregates.
5174 : *
5175 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5176 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5177 : * these would be Vars that are grouped by or used in grouping expressions.)
5178 : *
5179 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5180 : * havingQual represents the HAVING clause.
5181 : */
5182 : static PathTarget *
5183 1286 : make_partial_grouping_target(PlannerInfo *root,
5184 : PathTarget *grouping_target,
5185 : Node *havingQual)
5186 : {
5187 1286 : Query *parse = root->parse;
5188 : PathTarget *partial_target;
5189 : List *non_group_cols;
5190 : List *non_group_exprs;
5191 : int i;
5192 : ListCell *lc;
5193 :
5194 1286 : partial_target = create_empty_pathtarget();
5195 1286 : non_group_cols = NIL;
5196 :
5197 1286 : i = 0;
5198 4692 : foreach(lc, grouping_target->exprs)
5199 : {
5200 3406 : Expr *expr = (Expr *) lfirst(lc);
5201 3406 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5202 :
5203 5822 : if (sgref && parse->groupClause &&
5204 2416 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5205 : {
5206 : /*
5207 : * It's a grouping column, so add it to the partial_target as-is.
5208 : * (This allows the upper agg step to repeat the grouping calcs.)
5209 : */
5210 1192 : add_column_to_pathtarget(partial_target, expr, sgref);
5211 : }
5212 : else
5213 : {
5214 : /*
5215 : * Non-grouping column, so just remember the expression for later
5216 : * call to pull_var_clause.
5217 : */
5218 2214 : non_group_cols = lappend(non_group_cols, expr);
5219 : }
5220 :
5221 3406 : i++;
5222 : }
5223 :
5224 : /*
5225 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5226 : */
5227 1286 : if (havingQual)
5228 552 : non_group_cols = lappend(non_group_cols, havingQual);
5229 :
5230 : /*
5231 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5232 : * non-group cols (plus HAVING), and add them to the partial_target if not
5233 : * already present. (An expression used directly as a GROUP BY item will
5234 : * be present already.) Note this includes Vars used in resjunk items, so
5235 : * we are covering the needs of ORDER BY and window specifications.
5236 : */
5237 1286 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5238 : PVC_INCLUDE_AGGREGATES |
5239 : PVC_RECURSE_WINDOWFUNCS |
5240 : PVC_INCLUDE_PLACEHOLDERS);
5241 :
5242 1286 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5243 :
5244 : /*
5245 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5246 : * are at the top level of the target list, so we can just scan the list
5247 : * rather than recursing through the expression trees.
5248 : */
5249 5100 : foreach(lc, partial_target->exprs)
5250 : {
5251 3814 : Aggref *aggref = (Aggref *) lfirst(lc);
5252 :
5253 3814 : if (IsA(aggref, Aggref))
5254 : {
5255 : Aggref *newaggref;
5256 :
5257 : /*
5258 : * We shouldn't need to copy the substructure of the Aggref node,
5259 : * but flat-copy the node itself to avoid damaging other trees.
5260 : */
5261 2622 : newaggref = makeNode(Aggref);
5262 2622 : memcpy(newaggref, aggref, sizeof(Aggref));
5263 :
5264 : /* For now, assume serialization is required */
5265 2622 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5266 :
5267 2622 : lfirst(lc) = newaggref;
5268 : }
5269 : }
5270 :
5271 : /* clean up cruft */
5272 1286 : list_free(non_group_exprs);
5273 1286 : list_free(non_group_cols);
5274 :
5275 : /* XXX this causes some redundant cost calculation ... */
5276 1286 : return set_pathtarget_cost_width(root, partial_target);
5277 : }
5278 :
5279 : /*
5280 : * mark_partial_aggref
5281 : * Adjust an Aggref to make it represent a partial-aggregation step.
5282 : *
5283 : * The Aggref node is modified in-place; caller must do any copying required.
5284 : */
5285 : void
5286 4238 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5287 : {
5288 : /* aggtranstype should be computed by this point */
5289 : Assert(OidIsValid(agg->aggtranstype));
5290 : /* ... but aggsplit should still be as the parser left it */
5291 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5292 :
5293 : /* Mark the Aggref with the intended partial-aggregation mode */
5294 4238 : agg->aggsplit = aggsplit;
5295 :
5296 : /*
5297 : * Adjust result type if needed. Normally, a partial aggregate returns
5298 : * the aggregate's transition type; but if that's INTERNAL and we're
5299 : * serializing, it returns BYTEA instead.
5300 : */
5301 4238 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5302 : {
5303 3430 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5304 48 : agg->aggtype = BYTEAOID;
5305 : else
5306 3382 : agg->aggtype = agg->aggtranstype;
5307 : }
5308 4238 : }
5309 :
5310 : /*
5311 : * postprocess_setop_tlist
5312 : * Fix up targetlist returned by plan_set_operations().
5313 : *
5314 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5315 : * NOTE: this would not be good enough if we supported resjunk sort keys
5316 : * for results of set operations --- then, we'd need to project a whole
5317 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5318 : * find any resjunk columns in orig_tlist.
5319 : */
5320 : static List *
5321 1248 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5322 : {
5323 : ListCell *l;
5324 1248 : ListCell *orig_tlist_item = list_head(orig_tlist);
5325 :
5326 3886 : foreach(l, new_tlist)
5327 : {
5328 2638 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5329 : TargetEntry *orig_tle;
5330 :
5331 : /* ignore resjunk columns in setop result */
5332 2638 : if (new_tle->resjunk)
5333 200 : continue;
5334 :
5335 : Assert(orig_tlist_item != NULL);
5336 2438 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5337 2438 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5338 2438 : if (orig_tle->resjunk) /* should not happen */
5339 0 : elog(ERROR, "resjunk output columns are not implemented");
5340 : Assert(new_tle->resno == orig_tle->resno);
5341 2438 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5342 : }
5343 1248 : if (orig_tlist_item != NULL)
5344 0 : elog(ERROR, "resjunk output columns are not implemented");
5345 1248 : return new_tlist;
5346 : }
5347 :
5348 : /*
5349 : * select_active_windows
5350 : * Create a list of the "active" window clauses (ie, those referenced
5351 : * by non-deleted WindowFuncs) in the order they are to be executed.
5352 : */
5353 : static List *
5354 1112 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5355 : {
5356 1112 : List *windowClause = root->parse->windowClause;
5357 1112 : List *result = NIL;
5358 : ListCell *lc;
5359 1112 : int nActive = 0;
5360 1112 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5361 1112 : * list_length(windowClause));
5362 :
5363 : /* First, construct an array of the active windows */
5364 2288 : foreach(lc, windowClause)
5365 : {
5366 1176 : WindowClause *wc = lfirst_node(WindowClause, lc);
5367 :
5368 : /* It's only active if wflists shows some related WindowFuncs */
5369 : Assert(wc->winref <= wflists->maxWinRef);
5370 1176 : if (wflists->windowFuncs[wc->winref] == NIL)
5371 16 : continue;
5372 :
5373 1160 : actives[nActive].wc = wc; /* original clause */
5374 :
5375 : /*
5376 : * For sorting, we want the list of partition keys followed by the
5377 : * list of sort keys. But pathkeys construction will remove duplicates
5378 : * between the two, so we can as well (even though we can't detect all
5379 : * of the duplicates, since some may come from ECs - that might mean
5380 : * we miss optimization chances here). We must, however, ensure that
5381 : * the order of entries is preserved with respect to the ones we do
5382 : * keep.
5383 : *
5384 : * partitionClause and orderClause had their own duplicates removed in
5385 : * parse analysis, so we're only concerned here with removing
5386 : * orderClause entries that also appear in partitionClause.
5387 : */
5388 2320 : actives[nActive].uniqueOrder =
5389 1160 : list_concat_unique(list_copy(wc->partitionClause),
5390 1160 : wc->orderClause);
5391 1160 : nActive++;
5392 : }
5393 :
5394 : /*
5395 : * Sort active windows by their partitioning/ordering clauses, ignoring
5396 : * any framing clauses, so that the windows that need the same sorting are
5397 : * adjacent in the list. When we come to generate paths, this will avoid
5398 : * inserting additional Sort nodes.
5399 : *
5400 : * This is how we implement a specific requirement from the SQL standard,
5401 : * which says that when two or more windows are order-equivalent (i.e.
5402 : * have matching partition and order clauses, even if their names or
5403 : * framing clauses differ), then all peer rows must be presented in the
5404 : * same order in all of them. If we allowed multiple sort nodes for such
5405 : * cases, we'd risk having the peer rows end up in different orders in
5406 : * equivalent windows due to sort instability. (See General Rule 4 of
5407 : * <window clause> in SQL2008 - SQL2016.)
5408 : *
5409 : * Additionally, if the entire list of clauses of one window is a prefix
5410 : * of another, put first the window with stronger sorting requirements.
5411 : * This way we will first sort for stronger window, and won't have to sort
5412 : * again for the weaker one.
5413 : */
5414 1112 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5415 :
5416 : /* build ordered list of the original WindowClause nodes */
5417 2272 : for (int i = 0; i < nActive; i++)
5418 1160 : result = lappend(result, actives[i].wc);
5419 :
5420 1112 : pfree(actives);
5421 :
5422 1112 : return result;
5423 : }
5424 :
5425 : /*
5426 : * common_prefix_cmp
5427 : * QSort comparison function for WindowClauseSortData
5428 : *
5429 : * Sort the windows by the required sorting clauses. First, compare the sort
5430 : * clauses themselves. Second, if one window's clauses are a prefix of another
5431 : * one's clauses, put the window with more sort clauses first.
5432 : */
5433 : static int
5434 48 : common_prefix_cmp(const void *a, const void *b)
5435 : {
5436 48 : const WindowClauseSortData *wcsa = a;
5437 48 : const WindowClauseSortData *wcsb = b;
5438 : ListCell *item_a;
5439 : ListCell *item_b;
5440 :
5441 84 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
5442 : {
5443 56 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
5444 56 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
5445 :
5446 56 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
5447 20 : return -1;
5448 56 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
5449 12 : return 1;
5450 44 : else if (sca->sortop > scb->sortop)
5451 0 : return -1;
5452 44 : else if (sca->sortop < scb->sortop)
5453 8 : return 1;
5454 36 : else if (sca->nulls_first && !scb->nulls_first)
5455 0 : return -1;
5456 36 : else if (!sca->nulls_first && scb->nulls_first)
5457 0 : return 1;
5458 : /* no need to compare eqop, since it is fully determined by sortop */
5459 : }
5460 :
5461 28 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
5462 4 : return -1;
5463 24 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
5464 8 : return 1;
5465 :
5466 16 : return 0;
5467 : }
5468 :
5469 : /*
5470 : * make_window_input_target
5471 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
5472 : *
5473 : * When the query has window functions, this function computes the desired
5474 : * target to be computed by the node just below the first WindowAgg.
5475 : * This tlist must contain all values needed to evaluate the window functions,
5476 : * compute the final target list, and perform any required final sort step.
5477 : * If multiple WindowAggs are needed, each intermediate one adds its window
5478 : * function results onto this base tlist; only the topmost WindowAgg computes
5479 : * the actual desired target list.
5480 : *
5481 : * This function is much like make_group_input_target, though not quite enough
5482 : * like it to share code. As in that function, we flatten most expressions
5483 : * into their component variables. But we do not want to flatten window
5484 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
5485 : * evaluations of them, which would be bad (possibly even resulting in
5486 : * inconsistent answers, if they contain volatile functions).
5487 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
5488 : * make_group_input_target, because we may no longer have access to the
5489 : * individual Vars in them.
5490 : *
5491 : * Another key difference from make_group_input_target is that we don't
5492 : * flatten Aggref expressions, since those are to be computed below the
5493 : * window functions and just referenced like Vars above that.
5494 : *
5495 : * 'final_target' is the query's final target list (in PathTarget form)
5496 : * 'activeWindows' is the list of active windows previously identified by
5497 : * select_active_windows.
5498 : *
5499 : * The result is the PathTarget to be computed by the plan node immediately
5500 : * below the first WindowAgg node.
5501 : */
5502 : static PathTarget *
5503 1112 : make_window_input_target(PlannerInfo *root,
5504 : PathTarget *final_target,
5505 : List *activeWindows)
5506 : {
5507 1112 : Query *parse = root->parse;
5508 : PathTarget *input_target;
5509 : Bitmapset *sgrefs;
5510 : List *flattenable_cols;
5511 : List *flattenable_vars;
5512 : int i;
5513 : ListCell *lc;
5514 :
5515 : Assert(parse->hasWindowFuncs);
5516 :
5517 : /*
5518 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
5519 : * into a bitmapset for convenient reference below.
5520 : */
5521 1112 : sgrefs = NULL;
5522 2272 : foreach(lc, activeWindows)
5523 : {
5524 1160 : WindowClause *wc = lfirst_node(WindowClause, lc);
5525 : ListCell *lc2;
5526 :
5527 1512 : foreach(lc2, wc->partitionClause)
5528 : {
5529 352 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5530 :
5531 352 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5532 : }
5533 2216 : foreach(lc2, wc->orderClause)
5534 : {
5535 1056 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5536 :
5537 1056 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5538 : }
5539 : }
5540 :
5541 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
5542 1224 : foreach(lc, parse->groupClause)
5543 : {
5544 112 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
5545 :
5546 112 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
5547 : }
5548 :
5549 : /*
5550 : * Construct a target containing all the non-flattenable targetlist items,
5551 : * and save aside the others for a moment.
5552 : */
5553 1112 : input_target = create_empty_pathtarget();
5554 1112 : flattenable_cols = NIL;
5555 :
5556 1112 : i = 0;
5557 4776 : foreach(lc, final_target->exprs)
5558 : {
5559 3664 : Expr *expr = (Expr *) lfirst(lc);
5560 3664 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5561 :
5562 : /*
5563 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
5564 : * that such items can't contain window functions, so it's okay to
5565 : * compute them below the WindowAgg nodes.)
5566 : */
5567 3664 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
5568 : {
5569 : /*
5570 : * Don't want to deconstruct this value, so add it to the input
5571 : * target as-is.
5572 : */
5573 1368 : add_column_to_pathtarget(input_target, expr, sgref);
5574 : }
5575 : else
5576 : {
5577 : /*
5578 : * Column is to be flattened, so just remember the expression for
5579 : * later call to pull_var_clause.
5580 : */
5581 2296 : flattenable_cols = lappend(flattenable_cols, expr);
5582 : }
5583 :
5584 3664 : i++;
5585 : }
5586 :
5587 : /*
5588 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
5589 : * add them to the input target if not already present. (Some might be
5590 : * there already because they're used directly as window/group clauses.)
5591 : *
5592 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
5593 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
5594 : * at higher levels. On the other hand, we should recurse into
5595 : * WindowFuncs to make sure their input expressions are available.
5596 : */
5597 1112 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
5598 : PVC_INCLUDE_AGGREGATES |
5599 : PVC_RECURSE_WINDOWFUNCS |
5600 : PVC_INCLUDE_PLACEHOLDERS);
5601 1112 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
5602 :
5603 : /* clean up cruft */
5604 1112 : list_free(flattenable_vars);
5605 1112 : list_free(flattenable_cols);
5606 :
5607 : /* XXX this causes some redundant cost calculation ... */
5608 1112 : return set_pathtarget_cost_width(root, input_target);
5609 : }
5610 :
5611 : /*
5612 : * make_pathkeys_for_window
5613 : * Create a pathkeys list describing the required input ordering
5614 : * for the given WindowClause.
5615 : *
5616 : * The required ordering is first the PARTITION keys, then the ORDER keys.
5617 : * In the future we might try to implement windowing using hashing, in which
5618 : * case the ordering could be relaxed, but for now we always sort.
5619 : */
5620 : static List *
5621 2316 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
5622 : List *tlist)
5623 : {
5624 : List *window_pathkeys;
5625 : List *window_sortclauses;
5626 :
5627 : /* Throw error if can't sort */
5628 2316 : if (!grouping_is_sortable(wc->partitionClause))
5629 0 : ereport(ERROR,
5630 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5631 : errmsg("could not implement window PARTITION BY"),
5632 : errdetail("Window partitioning columns must be of sortable datatypes.")));
5633 2316 : if (!grouping_is_sortable(wc->orderClause))
5634 0 : ereport(ERROR,
5635 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5636 : errmsg("could not implement window ORDER BY"),
5637 : errdetail("Window ordering columns must be of sortable datatypes.")));
5638 :
5639 : /* Okay, make the combined pathkeys */
5640 2316 : window_sortclauses = list_concat_copy(wc->partitionClause, wc->orderClause);
5641 2316 : window_pathkeys = make_pathkeys_for_sortclauses(root,
5642 : window_sortclauses,
5643 : tlist);
5644 2316 : list_free(window_sortclauses);
5645 2316 : return window_pathkeys;
5646 : }
5647 :
5648 : /*
5649 : * make_sort_input_target
5650 : * Generate appropriate PathTarget for initial input to Sort step.
5651 : *
5652 : * If the query has ORDER BY, this function chooses the target to be computed
5653 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
5654 : * project) steps. This might or might not be identical to the query's final
5655 : * output target.
5656 : *
5657 : * The main argument for keeping the sort-input tlist the same as the final
5658 : * is that we avoid a separate projection node (which will be needed if
5659 : * they're different, because Sort can't project). However, there are also
5660 : * advantages to postponing tlist evaluation till after the Sort: it ensures
5661 : * a consistent order of evaluation for any volatile functions in the tlist,
5662 : * and if there's also a LIMIT, we can stop the query without ever computing
5663 : * tlist functions for later rows, which is beneficial for both volatile and
5664 : * expensive functions.
5665 : *
5666 : * Our current policy is to postpone volatile expressions till after the sort
5667 : * unconditionally (assuming that that's possible, ie they are in plain tlist
5668 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
5669 : * postpone set-returning expressions, because running them beforehand would
5670 : * bloat the sort dataset, and because it might cause unexpected output order
5671 : * if the sort isn't stable. However there's a constraint on that: all SRFs
5672 : * in the tlist should be evaluated at the same plan step, so that they can
5673 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
5674 : * mustn't postpone any SRFs. (Note that in principle that policy should
5675 : * probably get applied to the group/window input targetlists too, but we
5676 : * have not done that historically.) Lastly, expensive expressions are
5677 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
5678 : * partial evaluation of the query is possible (if neither is true, we expect
5679 : * to have to evaluate the expressions for every row anyway), or if there are
5680 : * any volatile or set-returning expressions (since once we've put in a
5681 : * projection at all, it won't cost any more to postpone more stuff).
5682 : *
5683 : * Another issue that could potentially be considered here is that
5684 : * evaluating tlist expressions could result in data that's either wider
5685 : * or narrower than the input Vars, thus changing the volume of data that
5686 : * has to go through the Sort. However, we usually have only a very bad
5687 : * idea of the output width of any expression more complex than a Var,
5688 : * so for now it seems too risky to try to optimize on that basis.
5689 : *
5690 : * Note that if we do produce a modified sort-input target, and then the
5691 : * query ends up not using an explicit Sort, no particular harm is done:
5692 : * we'll initially use the modified target for the preceding path nodes,
5693 : * but then change them to the final target with apply_projection_to_path.
5694 : * Moreover, in such a case the guarantees about evaluation order of
5695 : * volatile functions still hold, since the rows are sorted already.
5696 : *
5697 : * This function has some things in common with make_group_input_target and
5698 : * make_window_input_target, though the detailed rules for what to do are
5699 : * different. We never flatten/postpone any grouping or ordering columns;
5700 : * those are needed before the sort. If we do flatten a particular
5701 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
5702 : * computed earlier.
5703 : *
5704 : * 'final_target' is the query's final target list (in PathTarget form)
5705 : * 'have_postponed_srfs' is an output argument, see below
5706 : *
5707 : * The result is the PathTarget to be computed by the plan node immediately
5708 : * below the Sort step (and the Distinct step, if any). This will be
5709 : * exactly final_target if we decide a projection step wouldn't be helpful.
5710 : *
5711 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
5712 : * any set-returning functions to after the Sort.
5713 : */
5714 : static PathTarget *
5715 35756 : make_sort_input_target(PlannerInfo *root,
5716 : PathTarget *final_target,
5717 : bool *have_postponed_srfs)
5718 : {
5719 35756 : Query *parse = root->parse;
5720 : PathTarget *input_target;
5721 : int ncols;
5722 : bool *col_is_srf;
5723 : bool *postpone_col;
5724 : bool have_srf;
5725 : bool have_volatile;
5726 : bool have_expensive;
5727 : bool have_srf_sortcols;
5728 : bool postpone_srfs;
5729 : List *postponable_cols;
5730 : List *postponable_vars;
5731 : int i;
5732 : ListCell *lc;
5733 :
5734 : /* Shouldn't get here unless query has ORDER BY */
5735 : Assert(parse->sortClause);
5736 :
5737 35756 : *have_postponed_srfs = false; /* default result */
5738 :
5739 : /* Inspect tlist and collect per-column information */
5740 35756 : ncols = list_length(final_target->exprs);
5741 35756 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
5742 35756 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
5743 35756 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
5744 :
5745 35756 : i = 0;
5746 278418 : foreach(lc, final_target->exprs)
5747 : {
5748 242662 : Expr *expr = (Expr *) lfirst(lc);
5749 :
5750 : /*
5751 : * If the column has a sortgroupref, assume it has to be evaluated
5752 : * before sorting. Generally such columns would be ORDER BY, GROUP
5753 : * BY, etc targets. One exception is columns that were removed from
5754 : * GROUP BY by remove_useless_groupby_columns() ... but those would
5755 : * only be Vars anyway. There don't seem to be any cases where it
5756 : * would be worth the trouble to double-check.
5757 : */
5758 242662 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
5759 : {
5760 : /*
5761 : * Check for SRF or volatile functions. Check the SRF case first
5762 : * because we must know whether we have any postponed SRFs.
5763 : */
5764 395872 : if (parse->hasTargetSRFs &&
5765 124 : expression_returns_set((Node *) expr))
5766 : {
5767 : /* We'll decide below whether these are postponable */
5768 56 : col_is_srf[i] = true;
5769 56 : have_srf = true;
5770 : }
5771 197818 : else if (contain_volatile_functions((Node *) expr))
5772 : {
5773 : /* Unconditionally postpone */
5774 64 : postpone_col[i] = true;
5775 64 : have_volatile = true;
5776 : }
5777 : else
5778 : {
5779 : /*
5780 : * Else check the cost. XXX it's annoying to have to do this
5781 : * when set_pathtarget_cost_width() just did it. Refactor to
5782 : * allow sharing the work?
5783 : */
5784 : QualCost cost;
5785 :
5786 197754 : cost_qual_eval_node(&cost, (Node *) expr, root);
5787 :
5788 : /*
5789 : * We arbitrarily define "expensive" as "more than 10X
5790 : * cpu_operator_cost". Note this will take in any PL function
5791 : * with default cost.
5792 : */
5793 197754 : if (cost.per_tuple > 10 * cpu_operator_cost)
5794 : {
5795 18208 : postpone_col[i] = true;
5796 18208 : have_expensive = true;
5797 : }
5798 : }
5799 : }
5800 : else
5801 : {
5802 : /* For sortgroupref cols, just check if any contain SRFs */
5803 89576 : if (!have_srf_sortcols &&
5804 44984 : parse->hasTargetSRFs &&
5805 196 : expression_returns_set((Node *) expr))
5806 80 : have_srf_sortcols = true;
5807 : }
5808 :
5809 242662 : i++;
5810 : }
5811 :
5812 : /*
5813 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
5814 : */
5815 35756 : postpone_srfs = (have_srf && !have_srf_sortcols);
5816 :
5817 : /*
5818 : * If we don't need a post-sort projection, just return final_target.
5819 : */
5820 71412 : if (!(postpone_srfs || have_volatile ||
5821 46910 : (have_expensive &&
5822 22500 : (parse->limitCount || root->tuple_fraction > 0))))
5823 35648 : return final_target;
5824 :
5825 : /*
5826 : * Report whether the post-sort projection will contain set-returning
5827 : * functions. This is important because it affects whether the Sort can
5828 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
5829 : * to return.
5830 : */
5831 108 : *have_postponed_srfs = postpone_srfs;
5832 :
5833 : /*
5834 : * Construct the sort-input target, taking all non-postponable columns and
5835 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
5836 : * the postponable ones.
5837 : */
5838 108 : input_target = create_empty_pathtarget();
5839 108 : postponable_cols = NIL;
5840 :
5841 108 : i = 0;
5842 684 : foreach(lc, final_target->exprs)
5843 : {
5844 576 : Expr *expr = (Expr *) lfirst(lc);
5845 :
5846 576 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
5847 112 : postponable_cols = lappend(postponable_cols, expr);
5848 : else
5849 928 : add_column_to_pathtarget(input_target, expr,
5850 928 : get_pathtarget_sortgroupref(final_target, i));
5851 :
5852 576 : i++;
5853 : }
5854 :
5855 : /*
5856 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
5857 : * postponable columns, and add them to the sort-input target if not
5858 : * already present. (Some might be there already.) We mustn't
5859 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
5860 : * would be unable to recompute them.
5861 : */
5862 108 : postponable_vars = pull_var_clause((Node *) postponable_cols,
5863 : PVC_INCLUDE_AGGREGATES |
5864 : PVC_INCLUDE_WINDOWFUNCS |
5865 : PVC_INCLUDE_PLACEHOLDERS);
5866 108 : add_new_columns_to_pathtarget(input_target, postponable_vars);
5867 :
5868 : /* clean up cruft */
5869 108 : list_free(postponable_vars);
5870 108 : list_free(postponable_cols);
5871 :
5872 : /* XXX this represents even more redundant cost calculation ... */
5873 108 : return set_pathtarget_cost_width(root, input_target);
5874 : }
5875 :
5876 : /*
5877 : * get_cheapest_fractional_path
5878 : * Find the cheapest path for retrieving a specified fraction of all
5879 : * the tuples expected to be returned by the given relation.
5880 : *
5881 : * We interpret tuple_fraction the same way as grouping_planner.
5882 : *
5883 : * We assume set_cheapest() has been run on the given rel.
5884 : */
5885 : Path *
5886 292680 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
5887 : {
5888 292680 : Path *best_path = rel->cheapest_total_path;
5889 : ListCell *l;
5890 :
5891 : /* If all tuples will be retrieved, just return the cheapest-total path */
5892 292680 : if (tuple_fraction <= 0.0)
5893 288818 : return best_path;
5894 :
5895 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
5896 3862 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
5897 1600 : tuple_fraction /= best_path->rows;
5898 :
5899 9404 : foreach(l, rel->pathlist)
5900 : {
5901 5542 : Path *path = (Path *) lfirst(l);
5902 :
5903 7222 : if (path == rel->cheapest_total_path ||
5904 1680 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
5905 5072 : continue;
5906 :
5907 470 : best_path = path;
5908 : }
5909 :
5910 3862 : return best_path;
5911 : }
5912 :
5913 : /*
5914 : * adjust_paths_for_srfs
5915 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
5916 : *
5917 : * The executor can only handle set-returning functions that appear at the
5918 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
5919 : * that are not at top level, we need to split up the evaluation into multiple
5920 : * plan levels in which each level satisfies this constraint. This function
5921 : * modifies each Path of an upperrel that (might) compute any SRFs in its
5922 : * output tlist to insert appropriate projection steps.
5923 : *
5924 : * The given targets and targets_contain_srfs lists are from
5925 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
5926 : * target in targets.
5927 : */
5928 : static void
5929 3894 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
5930 : List *targets, List *targets_contain_srfs)
5931 : {
5932 : ListCell *lc;
5933 :
5934 : Assert(list_length(targets) == list_length(targets_contain_srfs));
5935 : Assert(!linitial_int(targets_contain_srfs));
5936 :
5937 : /* If no SRFs appear at this plan level, nothing to do */
5938 3894 : if (list_length(targets) == 1)
5939 212 : return;
5940 :
5941 : /*
5942 : * Stack SRF-evaluation nodes atop each path for the rel.
5943 : *
5944 : * In principle we should re-run set_cheapest() here to identify the
5945 : * cheapest path, but it seems unlikely that adding the same tlist eval
5946 : * costs to all the paths would change that, so we don't bother. Instead,
5947 : * just assume that the cheapest-startup and cheapest-total paths remain
5948 : * so. (There should be no parameterized paths anymore, so we needn't
5949 : * worry about updating cheapest_parameterized_paths.)
5950 : */
5951 7388 : foreach(lc, rel->pathlist)
5952 : {
5953 3706 : Path *subpath = (Path *) lfirst(lc);
5954 3706 : Path *newpath = subpath;
5955 : ListCell *lc1,
5956 : *lc2;
5957 :
5958 : Assert(subpath->param_info == NULL);
5959 11790 : forboth(lc1, targets, lc2, targets_contain_srfs)
5960 : {
5961 8084 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5962 8084 : bool contains_srfs = (bool) lfirst_int(lc2);
5963 :
5964 : /* If this level doesn't contain SRFs, do regular projection */
5965 8084 : if (contains_srfs)
5966 3746 : newpath = (Path *) create_set_projection_path(root,
5967 : rel,
5968 : newpath,
5969 : thistarget);
5970 : else
5971 4338 : newpath = (Path *) apply_projection_to_path(root,
5972 : rel,
5973 : newpath,
5974 : thistarget);
5975 : }
5976 3706 : lfirst(lc) = newpath;
5977 3706 : if (subpath == rel->cheapest_startup_path)
5978 52 : rel->cheapest_startup_path = newpath;
5979 3706 : if (subpath == rel->cheapest_total_path)
5980 52 : rel->cheapest_total_path = newpath;
5981 : }
5982 :
5983 : /* Likewise for partial paths, if any */
5984 3686 : foreach(lc, rel->partial_pathlist)
5985 : {
5986 4 : Path *subpath = (Path *) lfirst(lc);
5987 4 : Path *newpath = subpath;
5988 : ListCell *lc1,
5989 : *lc2;
5990 :
5991 : Assert(subpath->param_info == NULL);
5992 16 : forboth(lc1, targets, lc2, targets_contain_srfs)
5993 : {
5994 12 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5995 12 : bool contains_srfs = (bool) lfirst_int(lc2);
5996 :
5997 : /* If this level doesn't contain SRFs, do regular projection */
5998 12 : if (contains_srfs)
5999 4 : newpath = (Path *) create_set_projection_path(root,
6000 : rel,
6001 : newpath,
6002 : thistarget);
6003 : else
6004 : {
6005 : /* avoid apply_projection_to_path, in case of multiple refs */
6006 8 : newpath = (Path *) create_projection_path(root,
6007 : rel,
6008 : newpath,
6009 : thistarget);
6010 : }
6011 : }
6012 4 : lfirst(lc) = newpath;
6013 : }
6014 : }
6015 :
6016 : /*
6017 : * expression_planner
6018 : * Perform planner's transformations on a standalone expression.
6019 : *
6020 : * Various utility commands need to evaluate expressions that are not part
6021 : * of a plannable query. They can do so using the executor's regular
6022 : * expression-execution machinery, but first the expression has to be fed
6023 : * through here to transform it from parser output to something executable.
6024 : *
6025 : * Currently, we disallow sublinks in standalone expressions, so there's no
6026 : * real "planning" involved here. (That might not always be true though.)
6027 : * What we must do is run eval_const_expressions to ensure that any function
6028 : * calls are converted to positional notation and function default arguments
6029 : * get inserted. The fact that constant subexpressions get simplified is a
6030 : * side-effect that is useful when the expression will get evaluated more than
6031 : * once. Also, we must fix operator function IDs.
6032 : *
6033 : * This does not return any information about dependencies of the expression.
6034 : * Hence callers should use the results only for the duration of the current
6035 : * query. Callers that would like to cache the results for longer should use
6036 : * expression_planner_with_deps, probably via the plancache.
6037 : *
6038 : * Note: this must not make any damaging changes to the passed-in expression
6039 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6040 : * we first do an expression_tree_mutator-based walk, what is returned will
6041 : * be a new node tree.) The result is constructed in the current memory
6042 : * context; beware that this can leak a lot of additional stuff there, too.
6043 : */
6044 : Expr *
6045 84960 : expression_planner(Expr *expr)
6046 : {
6047 : Node *result;
6048 :
6049 : /*
6050 : * Convert named-argument function calls, insert default arguments and
6051 : * simplify constant subexprs
6052 : */
6053 84960 : result = eval_const_expressions(NULL, (Node *) expr);
6054 :
6055 : /* Fill in opfuncid values if missing */
6056 84960 : fix_opfuncids(result);
6057 :
6058 84960 : return (Expr *) result;
6059 : }
6060 :
6061 : /*
6062 : * expression_planner_with_deps
6063 : * Perform planner's transformations on a standalone expression,
6064 : * returning expression dependency information along with the result.
6065 : *
6066 : * This is identical to expression_planner() except that it also returns
6067 : * information about possible dependencies of the expression, ie identities of
6068 : * objects whose definitions affect the result. As in a PlannedStmt, these
6069 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6070 : */
6071 : Expr *
6072 456 : expression_planner_with_deps(Expr *expr,
6073 : List **relationOids,
6074 : List **invalItems)
6075 : {
6076 : Node *result;
6077 : PlannerGlobal glob;
6078 : PlannerInfo root;
6079 :
6080 : /* Make up dummy planner state so we can use setrefs machinery */
6081 456 : MemSet(&glob, 0, sizeof(glob));
6082 456 : glob.type = T_PlannerGlobal;
6083 456 : glob.relationOids = NIL;
6084 456 : glob.invalItems = NIL;
6085 :
6086 456 : MemSet(&root, 0, sizeof(root));
6087 456 : root.type = T_PlannerInfo;
6088 456 : root.glob = &glob;
6089 :
6090 : /*
6091 : * Convert named-argument function calls, insert default arguments and
6092 : * simplify constant subexprs. Collect identities of inlined functions
6093 : * and elided domains, too.
6094 : */
6095 456 : result = eval_const_expressions(&root, (Node *) expr);
6096 :
6097 : /* Fill in opfuncid values if missing */
6098 456 : fix_opfuncids(result);
6099 :
6100 : /*
6101 : * Now walk the finished expression to find anything else we ought to
6102 : * record as an expression dependency.
6103 : */
6104 456 : (void) extract_query_dependencies_walker(result, &root);
6105 :
6106 456 : *relationOids = glob.relationOids;
6107 456 : *invalItems = glob.invalItems;
6108 :
6109 456 : return (Expr *) result;
6110 : }
6111 :
6112 :
6113 : /*
6114 : * plan_cluster_use_sort
6115 : * Use the planner to decide how CLUSTER should implement sorting
6116 : *
6117 : * tableOid is the OID of a table to be clustered on its index indexOid
6118 : * (which is already known to be a btree index). Decide whether it's
6119 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6120 : * Return true to use sorting, false to use an indexscan.
6121 : *
6122 : * Note: caller had better already hold some type of lock on the table.
6123 : */
6124 : bool
6125 74 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6126 : {
6127 : PlannerInfo *root;
6128 : Query *query;
6129 : PlannerGlobal *glob;
6130 : RangeTblEntry *rte;
6131 : RelOptInfo *rel;
6132 : IndexOptInfo *indexInfo;
6133 : QualCost indexExprCost;
6134 : Cost comparisonCost;
6135 : Path *seqScanPath;
6136 : Path seqScanAndSortPath;
6137 : IndexPath *indexScanPath;
6138 : ListCell *lc;
6139 :
6140 : /* We can short-circuit the cost comparison if indexscans are disabled */
6141 74 : if (!enable_indexscan)
6142 20 : return true; /* use sort */
6143 :
6144 : /* Set up mostly-dummy planner state */
6145 54 : query = makeNode(Query);
6146 54 : query->commandType = CMD_SELECT;
6147 :
6148 54 : glob = makeNode(PlannerGlobal);
6149 :
6150 54 : root = makeNode(PlannerInfo);
6151 54 : root->parse = query;
6152 54 : root->glob = glob;
6153 54 : root->query_level = 1;
6154 54 : root->planner_cxt = CurrentMemoryContext;
6155 54 : root->wt_param_id = -1;
6156 :
6157 : /* Build a minimal RTE for the rel */
6158 54 : rte = makeNode(RangeTblEntry);
6159 54 : rte->rtekind = RTE_RELATION;
6160 54 : rte->relid = tableOid;
6161 54 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6162 54 : rte->rellockmode = AccessShareLock;
6163 54 : rte->lateral = false;
6164 54 : rte->inh = false;
6165 54 : rte->inFromCl = true;
6166 54 : query->rtable = list_make1(rte);
6167 :
6168 : /* Set up RTE/RelOptInfo arrays */
6169 54 : setup_simple_rel_arrays(root);
6170 :
6171 : /* Build RelOptInfo */
6172 54 : rel = build_simple_rel(root, 1, NULL);
6173 :
6174 : /* Locate IndexOptInfo for the target index */
6175 54 : indexInfo = NULL;
6176 86 : foreach(lc, rel->indexlist)
6177 : {
6178 86 : indexInfo = lfirst_node(IndexOptInfo, lc);
6179 86 : if (indexInfo->indexoid == indexOid)
6180 54 : break;
6181 : }
6182 :
6183 : /*
6184 : * It's possible that get_relation_info did not generate an IndexOptInfo
6185 : * for the desired index; this could happen if it's not yet reached its
6186 : * indcheckxmin usability horizon, or if it's a system index and we're
6187 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6188 : * trust the index contents but use seqscan-and-sort.
6189 : */
6190 54 : if (lc == NULL) /* not in the list? */
6191 0 : return true; /* use sort */
6192 :
6193 : /*
6194 : * Rather than doing all the pushups that would be needed to use
6195 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6196 : */
6197 54 : rel->rows = rel->tuples;
6198 54 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6199 :
6200 54 : root->total_table_pages = rel->pages;
6201 :
6202 : /*
6203 : * Determine eval cost of the index expressions, if any. We need to
6204 : * charge twice that amount for each tuple comparison that happens during
6205 : * the sort, since tuplesort.c will have to re-evaluate the index
6206 : * expressions each time. (XXX that's pretty inefficient...)
6207 : */
6208 54 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6209 54 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6210 :
6211 : /* Estimate the cost of seq scan + sort */
6212 54 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6213 108 : cost_sort(&seqScanAndSortPath, root, NIL,
6214 54 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6215 : comparisonCost, maintenance_work_mem, -1.0);
6216 :
6217 : /* Estimate the cost of index scan */
6218 54 : indexScanPath = create_index_path(root, indexInfo,
6219 : NIL, NIL, NIL, NIL,
6220 : ForwardScanDirection, false,
6221 : NULL, 1.0, false);
6222 :
6223 54 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6224 : }
6225 :
6226 : /*
6227 : * plan_create_index_workers
6228 : * Use the planner to decide how many parallel worker processes
6229 : * CREATE INDEX should request for use
6230 : *
6231 : * tableOid is the table on which the index is to be built. indexOid is the
6232 : * OID of an index to be created or reindexed (which must be a btree index).
6233 : *
6234 : * Return value is the number of parallel worker processes to request. It
6235 : * may be unsafe to proceed if this is 0. Note that this does not include the
6236 : * leader participating as a worker (value is always a number of parallel
6237 : * worker processes).
6238 : *
6239 : * Note: caller had better already hold some type of lock on the table and
6240 : * index.
6241 : */
6242 : int
6243 22242 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6244 : {
6245 : PlannerInfo *root;
6246 : Query *query;
6247 : PlannerGlobal *glob;
6248 : RangeTblEntry *rte;
6249 : Relation heap;
6250 : Relation index;
6251 : RelOptInfo *rel;
6252 : int parallel_workers;
6253 : BlockNumber heap_blocks;
6254 : double reltuples;
6255 : double allvisfrac;
6256 :
6257 : /* Return immediately when parallelism disabled */
6258 22242 : if (max_parallel_maintenance_workers == 0)
6259 80 : return 0;
6260 :
6261 : /* Set up largely-dummy planner state */
6262 22162 : query = makeNode(Query);
6263 22162 : query->commandType = CMD_SELECT;
6264 :
6265 22162 : glob = makeNode(PlannerGlobal);
6266 :
6267 22162 : root = makeNode(PlannerInfo);
6268 22162 : root->parse = query;
6269 22162 : root->glob = glob;
6270 22162 : root->query_level = 1;
6271 22162 : root->planner_cxt = CurrentMemoryContext;
6272 22162 : root->wt_param_id = -1;
6273 :
6274 : /*
6275 : * Build a minimal RTE.
6276 : *
6277 : * Mark the RTE with inh = true. This is a kludge to prevent
6278 : * get_relation_info() from fetching index info, which is necessary
6279 : * because it does not expect that any IndexOptInfo is currently
6280 : * undergoing REINDEX.
6281 : */
6282 22162 : rte = makeNode(RangeTblEntry);
6283 22162 : rte->rtekind = RTE_RELATION;
6284 22162 : rte->relid = tableOid;
6285 22162 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6286 22162 : rte->rellockmode = AccessShareLock;
6287 22162 : rte->lateral = false;
6288 22162 : rte->inh = true;
6289 22162 : rte->inFromCl = true;
6290 22162 : query->rtable = list_make1(rte);
6291 :
6292 : /* Set up RTE/RelOptInfo arrays */
6293 22162 : setup_simple_rel_arrays(root);
6294 :
6295 : /* Build RelOptInfo */
6296 22162 : rel = build_simple_rel(root, 1, NULL);
6297 :
6298 : /* Rels are assumed already locked by the caller */
6299 22162 : heap = table_open(tableOid, NoLock);
6300 22162 : index = index_open(indexOid, NoLock);
6301 :
6302 : /*
6303 : * Determine if it's safe to proceed.
6304 : *
6305 : * Currently, parallel workers can't access the leader's temporary tables.
6306 : * Furthermore, any index predicate or index expressions must be parallel
6307 : * safe.
6308 : */
6309 42796 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6310 41252 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6311 20618 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6312 : {
6313 1544 : parallel_workers = 0;
6314 1544 : goto done;
6315 : }
6316 :
6317 : /*
6318 : * If parallel_workers storage parameter is set for the table, accept that
6319 : * as the number of parallel worker processes to launch (though still cap
6320 : * at max_parallel_maintenance_workers). Note that we deliberately do not
6321 : * consider any other factor when parallel_workers is set. (e.g., memory
6322 : * use by workers.)
6323 : */
6324 20618 : if (rel->rel_parallel_workers != -1)
6325 : {
6326 14 : parallel_workers = Min(rel->rel_parallel_workers,
6327 : max_parallel_maintenance_workers);
6328 14 : goto done;
6329 : }
6330 :
6331 : /*
6332 : * Estimate heap relation size ourselves, since rel->pages cannot be
6333 : * trusted (heap RTE was marked as inheritance parent)
6334 : */
6335 20604 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6336 :
6337 : /*
6338 : * Determine number of workers to scan the heap relation using generic
6339 : * model
6340 : */
6341 20604 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6342 : max_parallel_maintenance_workers);
6343 :
6344 : /*
6345 : * Cap workers based on available maintenance_work_mem as needed.
6346 : *
6347 : * Note that each tuplesort participant receives an even share of the
6348 : * total maintenance_work_mem budget. Aim to leave participants
6349 : * (including the leader as a participant) with no less than 32MB of
6350 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
6351 : * immediately past the threshold of being capable of launching a single
6352 : * parallel worker to sort.
6353 : */
6354 41440 : while (parallel_workers > 0 &&
6355 158 : maintenance_work_mem / (parallel_workers + 1) < 32768L)
6356 74 : parallel_workers--;
6357 :
6358 : done:
6359 22162 : index_close(index, NoLock);
6360 22162 : table_close(heap, NoLock);
6361 :
6362 22162 : return parallel_workers;
6363 : }
6364 :
6365 : /*
6366 : * add_paths_to_grouping_rel
6367 : *
6368 : * Add non-partial paths to grouping relation.
6369 : */
6370 : static void
6371 24352 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
6372 : RelOptInfo *grouped_rel,
6373 : RelOptInfo *partially_grouped_rel,
6374 : const AggClauseCosts *agg_costs,
6375 : grouping_sets_data *gd, double dNumGroups,
6376 : GroupPathExtraData *extra)
6377 : {
6378 24352 : Query *parse = root->parse;
6379 24352 : Path *cheapest_path = input_rel->cheapest_total_path;
6380 : ListCell *lc;
6381 24352 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6382 24352 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6383 24352 : List *havingQual = (List *) extra->havingQual;
6384 24352 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6385 :
6386 24352 : if (can_sort)
6387 : {
6388 : /*
6389 : * Use any available suitably-sorted path as input, and also consider
6390 : * sorting the cheapest-total path.
6391 : */
6392 50040 : foreach(lc, input_rel->pathlist)
6393 : {
6394 25692 : Path *path = (Path *) lfirst(lc);
6395 : bool is_sorted;
6396 :
6397 25692 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6398 : path->pathkeys);
6399 25692 : if (path == cheapest_path || is_sorted)
6400 : {
6401 : /* Sort the cheapest-total path if it isn't already sorted */
6402 25514 : if (!is_sorted)
6403 2538 : path = (Path *) create_sort_path(root,
6404 : grouped_rel,
6405 : path,
6406 : root->group_pathkeys,
6407 : -1.0);
6408 :
6409 : /* Now decide what to stick atop it */
6410 25514 : if (parse->groupingSets)
6411 : {
6412 464 : consider_groupingsets_paths(root, grouped_rel,
6413 : path, true, can_hash,
6414 : gd, agg_costs, dNumGroups);
6415 : }
6416 25050 : else if (parse->hasAggs)
6417 : {
6418 : /*
6419 : * We have aggregation, possibly with plain GROUP BY. Make
6420 : * an AggPath.
6421 : */
6422 24768 : add_path(grouped_rel, (Path *)
6423 74304 : create_agg_path(root,
6424 : grouped_rel,
6425 : path,
6426 24768 : grouped_rel->reltarget,
6427 24768 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6428 : AGGSPLIT_SIMPLE,
6429 : parse->groupClause,
6430 : havingQual,
6431 : agg_costs,
6432 : dNumGroups));
6433 : }
6434 282 : else if (parse->groupClause)
6435 : {
6436 : /*
6437 : * We have GROUP BY without aggregation or grouping sets.
6438 : * Make a GroupPath.
6439 : */
6440 282 : add_path(grouped_rel, (Path *)
6441 282 : create_group_path(root,
6442 : grouped_rel,
6443 : path,
6444 : parse->groupClause,
6445 : havingQual,
6446 : dNumGroups));
6447 : }
6448 : else
6449 : {
6450 : /* Other cases should have been handled above */
6451 : Assert(false);
6452 : }
6453 : }
6454 : }
6455 :
6456 : /*
6457 : * Instead of operating directly on the input relation, we can
6458 : * consider finalizing a partially aggregated path.
6459 : */
6460 24348 : if (partially_grouped_rel != NULL)
6461 : {
6462 2248 : foreach(lc, partially_grouped_rel->pathlist)
6463 : {
6464 1360 : Path *path = (Path *) lfirst(lc);
6465 :
6466 : /*
6467 : * Insert a Sort node, if required. But there's no point in
6468 : * sorting anything but the cheapest path.
6469 : */
6470 1360 : if (!pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
6471 : {
6472 352 : if (path != partially_grouped_rel->cheapest_total_path)
6473 64 : continue;
6474 288 : path = (Path *) create_sort_path(root,
6475 : grouped_rel,
6476 : path,
6477 : root->group_pathkeys,
6478 : -1.0);
6479 : }
6480 :
6481 1296 : if (parse->hasAggs)
6482 1166 : add_path(grouped_rel, (Path *)
6483 3498 : create_agg_path(root,
6484 : grouped_rel,
6485 : path,
6486 1166 : grouped_rel->reltarget,
6487 1166 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6488 : AGGSPLIT_FINAL_DESERIAL,
6489 : parse->groupClause,
6490 : havingQual,
6491 : agg_final_costs,
6492 : dNumGroups));
6493 : else
6494 130 : add_path(grouped_rel, (Path *)
6495 130 : create_group_path(root,
6496 : grouped_rel,
6497 : path,
6498 : parse->groupClause,
6499 : havingQual,
6500 : dNumGroups));
6501 : }
6502 : }
6503 : }
6504 :
6505 24352 : if (can_hash)
6506 : {
6507 : double hashaggtablesize;
6508 :
6509 2452 : if (parse->groupingSets)
6510 : {
6511 : /*
6512 : * Try for a hash-only groupingsets path over unsorted input.
6513 : */
6514 380 : consider_groupingsets_paths(root, grouped_rel,
6515 : cheapest_path, false, true,
6516 : gd, agg_costs, dNumGroups);
6517 : }
6518 : else
6519 : {
6520 2072 : hashaggtablesize = estimate_hashagg_tablesize(cheapest_path,
6521 : agg_costs,
6522 : dNumGroups);
6523 :
6524 : /*
6525 : * Provided that the estimated size of the hashtable does not
6526 : * exceed work_mem, we'll generate a HashAgg Path, although if we
6527 : * were unable to sort above, then we'd better generate a Path, so
6528 : * that we at least have one.
6529 : */
6530 2072 : if (hashaggtablesize < work_mem * 1024L ||
6531 0 : grouped_rel->pathlist == NIL)
6532 : {
6533 : /*
6534 : * We just need an Agg over the cheapest-total input path,
6535 : * since input order won't matter.
6536 : */
6537 2072 : add_path(grouped_rel, (Path *)
6538 4144 : create_agg_path(root, grouped_rel,
6539 : cheapest_path,
6540 2072 : grouped_rel->reltarget,
6541 : AGG_HASHED,
6542 : AGGSPLIT_SIMPLE,
6543 : parse->groupClause,
6544 : havingQual,
6545 : agg_costs,
6546 : dNumGroups));
6547 : }
6548 : }
6549 :
6550 : /*
6551 : * Generate a Finalize HashAgg Path atop of the cheapest partially
6552 : * grouped path, assuming there is one. Once again, we'll only do this
6553 : * if it looks as though the hash table won't exceed work_mem.
6554 : */
6555 2452 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
6556 : {
6557 466 : Path *path = partially_grouped_rel->cheapest_total_path;
6558 :
6559 466 : hashaggtablesize = estimate_hashagg_tablesize(path,
6560 : agg_final_costs,
6561 : dNumGroups);
6562 :
6563 466 : if (hashaggtablesize < work_mem * 1024L)
6564 466 : add_path(grouped_rel, (Path *)
6565 932 : create_agg_path(root,
6566 : grouped_rel,
6567 : path,
6568 466 : grouped_rel->reltarget,
6569 : AGG_HASHED,
6570 : AGGSPLIT_FINAL_DESERIAL,
6571 : parse->groupClause,
6572 : havingQual,
6573 : agg_final_costs,
6574 : dNumGroups));
6575 : }
6576 : }
6577 :
6578 : /*
6579 : * When partitionwise aggregate is used, we might have fully aggregated
6580 : * paths in the partial pathlist, because add_paths_to_append_rel() will
6581 : * consider a path for grouped_rel consisting of a Parallel Append of
6582 : * non-partial paths from each child.
6583 : */
6584 24352 : if (grouped_rel->partial_pathlist != NIL)
6585 108 : gather_grouping_paths(root, grouped_rel);
6586 24352 : }
6587 :
6588 : /*
6589 : * create_partial_grouping_paths
6590 : *
6591 : * Create a new upper relation representing the result of partial aggregation
6592 : * and populate it with appropriate paths. Note that we don't finalize the
6593 : * lists of paths here, so the caller can add additional partial or non-partial
6594 : * paths and must afterward call gather_grouping_paths and set_cheapest on
6595 : * the returned upper relation.
6596 : *
6597 : * All paths for this new upper relation -- both partial and non-partial --
6598 : * have been partially aggregated but require a subsequent FinalizeAggregate
6599 : * step.
6600 : *
6601 : * NB: This function is allowed to return NULL if it determines that there is
6602 : * no real need to create a new RelOptInfo.
6603 : */
6604 : static RelOptInfo *
6605 9038 : create_partial_grouping_paths(PlannerInfo *root,
6606 : RelOptInfo *grouped_rel,
6607 : RelOptInfo *input_rel,
6608 : grouping_sets_data *gd,
6609 : GroupPathExtraData *extra,
6610 : bool force_rel_creation)
6611 : {
6612 9038 : Query *parse = root->parse;
6613 : RelOptInfo *partially_grouped_rel;
6614 9038 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
6615 9038 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6616 9038 : Path *cheapest_partial_path = NULL;
6617 9038 : Path *cheapest_total_path = NULL;
6618 9038 : double dNumPartialGroups = 0;
6619 9038 : double dNumPartialPartialGroups = 0;
6620 : ListCell *lc;
6621 9038 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6622 9038 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6623 :
6624 : /*
6625 : * Consider whether we should generate partially aggregated non-partial
6626 : * paths. We can only do this if we have a non-partial path, and only if
6627 : * the parent of the input rel is performing partial partitionwise
6628 : * aggregation. (Note that extra->patype is the type of partitionwise
6629 : * aggregation being used at the parent level, not this level.)
6630 : */
6631 18076 : if (input_rel->pathlist != NIL &&
6632 9038 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
6633 398 : cheapest_total_path = input_rel->cheapest_total_path;
6634 :
6635 : /*
6636 : * If parallelism is possible for grouped_rel, then we should consider
6637 : * generating partially-grouped partial paths. However, if the input rel
6638 : * has no partial paths, then we can't.
6639 : */
6640 9038 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
6641 1030 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
6642 :
6643 : /*
6644 : * If we can't partially aggregate partial paths, and we can't partially
6645 : * aggregate non-partial paths, then don't bother creating the new
6646 : * RelOptInfo at all, unless the caller specified force_rel_creation.
6647 : */
6648 9038 : if (cheapest_total_path == NULL &&
6649 7810 : cheapest_partial_path == NULL &&
6650 7810 : !force_rel_creation)
6651 7752 : return NULL;
6652 :
6653 : /*
6654 : * Build a new upper relation to represent the result of partially
6655 : * aggregating the rows from the input relation.
6656 : */
6657 1286 : partially_grouped_rel = fetch_upper_rel(root,
6658 : UPPERREL_PARTIAL_GROUP_AGG,
6659 : grouped_rel->relids);
6660 1286 : partially_grouped_rel->consider_parallel =
6661 1286 : grouped_rel->consider_parallel;
6662 1286 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
6663 1286 : partially_grouped_rel->serverid = grouped_rel->serverid;
6664 1286 : partially_grouped_rel->userid = grouped_rel->userid;
6665 1286 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
6666 1286 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
6667 :
6668 : /*
6669 : * Build target list for partial aggregate paths. These paths cannot just
6670 : * emit the same tlist as regular aggregate paths, because (1) we must
6671 : * include Vars and Aggrefs needed in HAVING, which might not appear in
6672 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
6673 : */
6674 1286 : partially_grouped_rel->reltarget =
6675 1286 : make_partial_grouping_target(root, grouped_rel->reltarget,
6676 : extra->havingQual);
6677 :
6678 1286 : if (!extra->partial_costs_set)
6679 : {
6680 : /*
6681 : * Collect statistics about aggregates for estimating costs of
6682 : * performing aggregation in parallel.
6683 : */
6684 696 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
6685 696 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
6686 696 : if (parse->hasAggs)
6687 : {
6688 : List *partial_target_exprs;
6689 :
6690 : /* partial phase */
6691 632 : partial_target_exprs = partially_grouped_rel->reltarget->exprs;
6692 632 : get_agg_clause_costs(root, (Node *) partial_target_exprs,
6693 : AGGSPLIT_INITIAL_SERIAL,
6694 : agg_partial_costs);
6695 :
6696 : /* final phase */
6697 632 : get_agg_clause_costs(root, (Node *) grouped_rel->reltarget->exprs,
6698 : AGGSPLIT_FINAL_DESERIAL,
6699 : agg_final_costs);
6700 632 : get_agg_clause_costs(root, extra->havingQual,
6701 : AGGSPLIT_FINAL_DESERIAL,
6702 : agg_final_costs);
6703 : }
6704 :
6705 696 : extra->partial_costs_set = true;
6706 : }
6707 :
6708 : /* Estimate number of partial groups. */
6709 1286 : if (cheapest_total_path != NULL)
6710 398 : dNumPartialGroups =
6711 398 : get_number_of_groups(root,
6712 : cheapest_total_path->rows,
6713 : gd,
6714 : extra->targetList);
6715 1286 : if (cheapest_partial_path != NULL)
6716 1030 : dNumPartialPartialGroups =
6717 1030 : get_number_of_groups(root,
6718 : cheapest_partial_path->rows,
6719 : gd,
6720 : extra->targetList);
6721 :
6722 1286 : if (can_sort && cheapest_total_path != NULL)
6723 : {
6724 : /* This should have been checked previously */
6725 : Assert(parse->hasAggs || parse->groupClause);
6726 :
6727 : /*
6728 : * Use any available suitably-sorted path as input, and also consider
6729 : * sorting the cheapest partial path.
6730 : */
6731 802 : foreach(lc, input_rel->pathlist)
6732 : {
6733 404 : Path *path = (Path *) lfirst(lc);
6734 : bool is_sorted;
6735 :
6736 404 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6737 : path->pathkeys);
6738 404 : if (path == cheapest_total_path || is_sorted)
6739 : {
6740 : /* Sort the cheapest partial path, if it isn't already */
6741 404 : if (!is_sorted)
6742 398 : path = (Path *) create_sort_path(root,
6743 : partially_grouped_rel,
6744 : path,
6745 : root->group_pathkeys,
6746 : -1.0);
6747 :
6748 404 : if (parse->hasAggs)
6749 356 : add_path(partially_grouped_rel, (Path *)
6750 1068 : create_agg_path(root,
6751 : partially_grouped_rel,
6752 : path,
6753 356 : partially_grouped_rel->reltarget,
6754 356 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6755 : AGGSPLIT_INITIAL_SERIAL,
6756 : parse->groupClause,
6757 : NIL,
6758 : agg_partial_costs,
6759 : dNumPartialGroups));
6760 : else
6761 48 : add_path(partially_grouped_rel, (Path *)
6762 48 : create_group_path(root,
6763 : partially_grouped_rel,
6764 : path,
6765 : parse->groupClause,
6766 : NIL,
6767 : dNumPartialGroups));
6768 : }
6769 : }
6770 : }
6771 :
6772 1286 : if (can_sort && cheapest_partial_path != NULL)
6773 : {
6774 : /* Similar to above logic, but for partial paths. */
6775 2060 : foreach(lc, input_rel->partial_pathlist)
6776 : {
6777 1030 : Path *path = (Path *) lfirst(lc);
6778 : bool is_sorted;
6779 :
6780 1030 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6781 : path->pathkeys);
6782 1030 : if (path == cheapest_partial_path || is_sorted)
6783 : {
6784 : /* Sort the cheapest partial path, if it isn't already */
6785 1030 : if (!is_sorted)
6786 608 : path = (Path *) create_sort_path(root,
6787 : partially_grouped_rel,
6788 : path,
6789 : root->group_pathkeys,
6790 : -1.0);
6791 :
6792 1030 : if (parse->hasAggs)
6793 974 : add_partial_path(partially_grouped_rel, (Path *)
6794 2922 : create_agg_path(root,
6795 : partially_grouped_rel,
6796 : path,
6797 974 : partially_grouped_rel->reltarget,
6798 974 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6799 : AGGSPLIT_INITIAL_SERIAL,
6800 : parse->groupClause,
6801 : NIL,
6802 : agg_partial_costs,
6803 : dNumPartialPartialGroups));
6804 : else
6805 56 : add_partial_path(partially_grouped_rel, (Path *)
6806 56 : create_group_path(root,
6807 : partially_grouped_rel,
6808 : path,
6809 : parse->groupClause,
6810 : NIL,
6811 : dNumPartialPartialGroups));
6812 : }
6813 : }
6814 : }
6815 :
6816 1286 : if (can_hash && cheapest_total_path != NULL)
6817 : {
6818 : double hashaggtablesize;
6819 :
6820 : /* Checked above */
6821 : Assert(parse->hasAggs || parse->groupClause);
6822 :
6823 398 : hashaggtablesize =
6824 : estimate_hashagg_tablesize(cheapest_total_path,
6825 : agg_partial_costs,
6826 : dNumPartialGroups);
6827 :
6828 : /*
6829 : * Tentatively produce a partial HashAgg Path, depending on if it
6830 : * looks as if the hash table will fit in work_mem.
6831 : */
6832 398 : if (hashaggtablesize < work_mem * 1024L &&
6833 : cheapest_total_path != NULL)
6834 : {
6835 398 : add_path(partially_grouped_rel, (Path *)
6836 796 : create_agg_path(root,
6837 : partially_grouped_rel,
6838 : cheapest_total_path,
6839 398 : partially_grouped_rel->reltarget,
6840 : AGG_HASHED,
6841 : AGGSPLIT_INITIAL_SERIAL,
6842 : parse->groupClause,
6843 : NIL,
6844 : agg_partial_costs,
6845 : dNumPartialGroups));
6846 : }
6847 : }
6848 :
6849 1286 : if (can_hash && cheapest_partial_path != NULL)
6850 : {
6851 : double hashaggtablesize;
6852 :
6853 608 : hashaggtablesize =
6854 : estimate_hashagg_tablesize(cheapest_partial_path,
6855 : agg_partial_costs,
6856 : dNumPartialPartialGroups);
6857 :
6858 : /* Do the same for partial paths. */
6859 608 : if (hashaggtablesize < work_mem * 1024L &&
6860 : cheapest_partial_path != NULL)
6861 : {
6862 608 : add_partial_path(partially_grouped_rel, (Path *)
6863 1216 : create_agg_path(root,
6864 : partially_grouped_rel,
6865 : cheapest_partial_path,
6866 608 : partially_grouped_rel->reltarget,
6867 : AGG_HASHED,
6868 : AGGSPLIT_INITIAL_SERIAL,
6869 : parse->groupClause,
6870 : NIL,
6871 : agg_partial_costs,
6872 : dNumPartialPartialGroups));
6873 : }
6874 : }
6875 :
6876 : /*
6877 : * If there is an FDW that's responsible for all baserels of the query,
6878 : * let it consider adding partially grouped ForeignPaths.
6879 : */
6880 1292 : if (partially_grouped_rel->fdwroutine &&
6881 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
6882 : {
6883 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
6884 :
6885 6 : fdwroutine->GetForeignUpperPaths(root,
6886 : UPPERREL_PARTIAL_GROUP_AGG,
6887 : input_rel, partially_grouped_rel,
6888 : extra);
6889 : }
6890 :
6891 1286 : return partially_grouped_rel;
6892 : }
6893 :
6894 : /*
6895 : * Generate Gather and Gather Merge paths for a grouping relation or partial
6896 : * grouping relation.
6897 : *
6898 : * generate_gather_paths does most of the work, but we also consider a special
6899 : * case: we could try sorting the data by the group_pathkeys and then applying
6900 : * Gather Merge.
6901 : *
6902 : * NB: This function shouldn't be used for anything other than a grouped or
6903 : * partially grouped relation not only because of the fact that it explicitly
6904 : * references group_pathkeys but we pass "true" as the third argument to
6905 : * generate_gather_paths().
6906 : */
6907 : static void
6908 938 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
6909 : {
6910 : Path *cheapest_partial_path;
6911 :
6912 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
6913 938 : generate_gather_paths(root, rel, true);
6914 :
6915 : /* Try cheapest partial path + explicit Sort + Gather Merge. */
6916 938 : cheapest_partial_path = linitial(rel->partial_pathlist);
6917 938 : if (!pathkeys_contained_in(root->group_pathkeys,
6918 : cheapest_partial_path->pathkeys))
6919 : {
6920 : Path *path;
6921 : double total_groups;
6922 :
6923 470 : total_groups =
6924 470 : cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
6925 470 : path = (Path *) create_sort_path(root, rel, cheapest_partial_path,
6926 : root->group_pathkeys,
6927 : -1.0);
6928 470 : path = (Path *)
6929 470 : create_gather_merge_path(root,
6930 : rel,
6931 : path,
6932 470 : rel->reltarget,
6933 : root->group_pathkeys,
6934 : NULL,
6935 : &total_groups);
6936 :
6937 470 : add_path(rel, path);
6938 : }
6939 938 : }
6940 :
6941 : /*
6942 : * can_partial_agg
6943 : *
6944 : * Determines whether or not partial grouping and/or aggregation is possible.
6945 : * Returns true when possible, false otherwise.
6946 : */
6947 : static bool
6948 23808 : can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
6949 : {
6950 23808 : Query *parse = root->parse;
6951 :
6952 23808 : if (!parse->hasAggs && parse->groupClause == NIL)
6953 : {
6954 : /*
6955 : * We don't know how to do parallel aggregation unless we have either
6956 : * some aggregates or a grouping clause.
6957 : */
6958 0 : return false;
6959 : }
6960 23808 : else if (parse->groupingSets)
6961 : {
6962 : /* We don't know how to do grouping sets in parallel. */
6963 432 : return false;
6964 : }
6965 23376 : else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
6966 : {
6967 : /* Insufficient support for partial mode. */
6968 15232 : return false;
6969 : }
6970 :
6971 : /* Everything looks good. */
6972 8144 : return true;
6973 : }
6974 :
6975 : /*
6976 : * apply_scanjoin_target_to_paths
6977 : *
6978 : * Adjust the final scan/join relation, and recursively all of its children,
6979 : * to generate the final scan/join target. It would be more correct to model
6980 : * this as a separate planning step with a new RelOptInfo at the toplevel and
6981 : * for each child relation, but doing it this way is noticeably cheaper.
6982 : * Maybe that problem can be solved at some point, but for now we do this.
6983 : *
6984 : * If tlist_same_exprs is true, then the scan/join target to be applied has
6985 : * the same expressions as the existing reltarget, so we need only insert the
6986 : * appropriate sortgroupref information. By avoiding the creation of
6987 : * projection paths we save effort both immediately and at plan creation time.
6988 : */
6989 : static void
6990 312642 : apply_scanjoin_target_to_paths(PlannerInfo *root,
6991 : RelOptInfo *rel,
6992 : List *scanjoin_targets,
6993 : List *scanjoin_targets_contain_srfs,
6994 : bool scanjoin_target_parallel_safe,
6995 : bool tlist_same_exprs)
6996 : {
6997 312642 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
6998 : PathTarget *scanjoin_target;
6999 : ListCell *lc;
7000 :
7001 : /* This recurses, so be paranoid. */
7002 312642 : check_stack_depth();
7003 :
7004 : /*
7005 : * If the rel is partitioned, we want to drop its existing paths and
7006 : * generate new ones. This function would still be correct if we kept the
7007 : * existing paths: we'd modify them to generate the correct target above
7008 : * the partitioning Append, and then they'd compete on cost with paths
7009 : * generating the target below the Append. However, in our current cost
7010 : * model the latter way is always the same or cheaper cost, so modifying
7011 : * the existing paths would just be useless work. Moreover, when the cost
7012 : * is the same, varying roundoff errors might sometimes allow an existing
7013 : * path to be picked, resulting in undesirable cross-platform plan
7014 : * variations. So we drop old paths and thereby force the work to be done
7015 : * below the Append, except in the case of a non-parallel-safe target.
7016 : *
7017 : * Some care is needed, because we have to allow generate_gather_paths to
7018 : * see the old partial paths in the next stanza. Hence, zap the main
7019 : * pathlist here, then allow generate_gather_paths to add path(s) to the
7020 : * main list, and finally zap the partial pathlist.
7021 : */
7022 312642 : if (rel_is_partitioned)
7023 6682 : rel->pathlist = NIL;
7024 :
7025 : /*
7026 : * If the scan/join target is not parallel-safe, partial paths cannot
7027 : * generate it.
7028 : */
7029 312642 : if (!scanjoin_target_parallel_safe)
7030 : {
7031 : /*
7032 : * Since we can't generate the final scan/join target in parallel
7033 : * workers, this is our last opportunity to use any partial paths that
7034 : * exist; so build Gather path(s) that use them and emit whatever the
7035 : * current reltarget is. We don't do this in the case where the
7036 : * target is parallel-safe, since we will be able to generate superior
7037 : * paths by doing it after the final scan/join target has been
7038 : * applied.
7039 : */
7040 65226 : generate_gather_paths(root, rel, false);
7041 :
7042 : /* Can't use parallel query above this level. */
7043 65226 : rel->partial_pathlist = NIL;
7044 65226 : rel->consider_parallel = false;
7045 : }
7046 :
7047 : /* Finish dropping old paths for a partitioned rel, per comment above */
7048 312642 : if (rel_is_partitioned)
7049 6682 : rel->partial_pathlist = NIL;
7050 :
7051 : /* Extract SRF-free scan/join target. */
7052 312642 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7053 :
7054 : /*
7055 : * Apply the SRF-free scan/join target to each existing path.
7056 : *
7057 : * If the tlist exprs are the same, we can just inject the sortgroupref
7058 : * information into the existing pathtargets. Otherwise, replace each
7059 : * path with a projection path that generates the SRF-free scan/join
7060 : * target. This can't change the ordering of paths within rel->pathlist,
7061 : * so we just modify the list in place.
7062 : */
7063 634376 : foreach(lc, rel->pathlist)
7064 : {
7065 321734 : Path *subpath = (Path *) lfirst(lc);
7066 :
7067 : /* Shouldn't have any parameterized paths anymore */
7068 : Assert(subpath->param_info == NULL);
7069 :
7070 321734 : if (tlist_same_exprs)
7071 204776 : subpath->pathtarget->sortgrouprefs =
7072 102388 : scanjoin_target->sortgrouprefs;
7073 : else
7074 : {
7075 : Path *newpath;
7076 :
7077 219346 : newpath = (Path *) create_projection_path(root, rel, subpath,
7078 : scanjoin_target);
7079 219346 : lfirst(lc) = newpath;
7080 : }
7081 : }
7082 :
7083 : /* Likewise adjust the targets for any partial paths. */
7084 322792 : foreach(lc, rel->partial_pathlist)
7085 : {
7086 10150 : Path *subpath = (Path *) lfirst(lc);
7087 :
7088 : /* Shouldn't have any parameterized paths anymore */
7089 : Assert(subpath->param_info == NULL);
7090 :
7091 10150 : if (tlist_same_exprs)
7092 16480 : subpath->pathtarget->sortgrouprefs =
7093 8240 : scanjoin_target->sortgrouprefs;
7094 : else
7095 : {
7096 : Path *newpath;
7097 :
7098 1910 : newpath = (Path *) create_projection_path(root, rel, subpath,
7099 : scanjoin_target);
7100 1910 : lfirst(lc) = newpath;
7101 : }
7102 : }
7103 :
7104 : /*
7105 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7106 : * atop each existing path. (Note that this function doesn't look at the
7107 : * cheapest-path fields, which is a good thing because they're bogus right
7108 : * now.)
7109 : */
7110 312642 : if (root->parse->hasTargetSRFs)
7111 3682 : adjust_paths_for_srfs(root, rel,
7112 : scanjoin_targets,
7113 : scanjoin_targets_contain_srfs);
7114 :
7115 : /*
7116 : * Update the rel's target to be the final (with SRFs) scan/join target.
7117 : * This now matches the actual output of all the paths, and we might get
7118 : * confused in createplan.c if they don't agree. We must do this now so
7119 : * that any append paths made in the next part will use the correct
7120 : * pathtarget (cf. create_append_path).
7121 : *
7122 : * Note that this is also necessary if GetForeignUpperPaths() gets called
7123 : * on the final scan/join relation or on any of its children, since the
7124 : * FDW might look at the rel's target to create ForeignPaths.
7125 : */
7126 312642 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7127 :
7128 : /*
7129 : * If the relation is partitioned, recursively apply the scan/join target
7130 : * to all partitions, and generate brand-new Append paths in which the
7131 : * scan/join target is computed below the Append rather than above it.
7132 : * Since Append is not projection-capable, that might save a separate
7133 : * Result node, and it also is important for partitionwise aggregate.
7134 : */
7135 312642 : if (rel_is_partitioned)
7136 : {
7137 6682 : List *live_children = NIL;
7138 : int partition_idx;
7139 :
7140 : /* Adjust each partition. */
7141 27230 : for (partition_idx = 0; partition_idx < rel->nparts; partition_idx++)
7142 : {
7143 20548 : RelOptInfo *child_rel = rel->part_rels[partition_idx];
7144 : AppendRelInfo **appinfos;
7145 : int nappinfos;
7146 20548 : List *child_scanjoin_targets = NIL;
7147 : ListCell *lc;
7148 :
7149 : /* Pruned or dummy children can be ignored. */
7150 20548 : if (child_rel == NULL || IS_DUMMY_REL(child_rel))
7151 9242 : continue;
7152 :
7153 : /* Translate scan/join targets for this child. */
7154 11306 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7155 : &nappinfos);
7156 22612 : foreach(lc, scanjoin_targets)
7157 : {
7158 11306 : PathTarget *target = lfirst_node(PathTarget, lc);
7159 :
7160 11306 : target = copy_pathtarget(target);
7161 11306 : target->exprs = (List *)
7162 22612 : adjust_appendrel_attrs(root,
7163 11306 : (Node *) target->exprs,
7164 : nappinfos, appinfos);
7165 11306 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7166 : target);
7167 : }
7168 11306 : pfree(appinfos);
7169 :
7170 : /* Recursion does the real work. */
7171 11306 : apply_scanjoin_target_to_paths(root, child_rel,
7172 : child_scanjoin_targets,
7173 : scanjoin_targets_contain_srfs,
7174 : scanjoin_target_parallel_safe,
7175 : tlist_same_exprs);
7176 :
7177 : /* Save non-dummy children for Append paths. */
7178 11306 : if (!IS_DUMMY_REL(child_rel))
7179 11306 : live_children = lappend(live_children, child_rel);
7180 : }
7181 :
7182 : /* Build new paths for this relation by appending child paths. */
7183 6682 : add_paths_to_append_rel(root, rel, live_children);
7184 : }
7185 :
7186 : /*
7187 : * Consider generating Gather or Gather Merge paths. We must only do this
7188 : * if the relation is parallel safe, and we don't do it for child rels to
7189 : * avoid creating multiple Gather nodes within the same plan. We must do
7190 : * this after all paths have been generated and before set_cheapest, since
7191 : * one of the generated paths may turn out to be the cheapest one.
7192 : */
7193 312642 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7194 73160 : generate_gather_paths(root, rel, false);
7195 :
7196 : /*
7197 : * Reassess which paths are the cheapest, now that we've potentially added
7198 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7199 : * this relation.
7200 : */
7201 312642 : set_cheapest(rel);
7202 312642 : }
7203 :
7204 : /*
7205 : * create_partitionwise_grouping_paths
7206 : *
7207 : * If the partition keys of input relation are part of the GROUP BY clause, all
7208 : * the rows belonging to a given group come from a single partition. This
7209 : * allows aggregation/grouping over a partitioned relation to be broken down
7210 : * into aggregation/grouping on each partition. This should be no worse, and
7211 : * often better, than the normal approach.
7212 : *
7213 : * However, if the GROUP BY clause does not contain all the partition keys,
7214 : * rows from a given group may be spread across multiple partitions. In that
7215 : * case, we perform partial aggregation for each group, append the results,
7216 : * and then finalize aggregation. This is less certain to win than the
7217 : * previous case. It may win if the PartialAggregate stage greatly reduces
7218 : * the number of groups, because fewer rows will pass through the Append node.
7219 : * It may lose if we have lots of small groups.
7220 : */
7221 : static void
7222 346 : create_partitionwise_grouping_paths(PlannerInfo *root,
7223 : RelOptInfo *input_rel,
7224 : RelOptInfo *grouped_rel,
7225 : RelOptInfo *partially_grouped_rel,
7226 : const AggClauseCosts *agg_costs,
7227 : grouping_sets_data *gd,
7228 : PartitionwiseAggregateType patype,
7229 : GroupPathExtraData *extra)
7230 : {
7231 346 : int nparts = input_rel->nparts;
7232 : int cnt_parts;
7233 346 : List *grouped_live_children = NIL;
7234 346 : List *partially_grouped_live_children = NIL;
7235 346 : PathTarget *target = grouped_rel->reltarget;
7236 346 : bool partial_grouping_valid = true;
7237 :
7238 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7239 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7240 : partially_grouped_rel != NULL);
7241 :
7242 : /* Add paths for partitionwise aggregation/grouping. */
7243 1288 : for (cnt_parts = 0; cnt_parts < nparts; cnt_parts++)
7244 : {
7245 942 : RelOptInfo *child_input_rel = input_rel->part_rels[cnt_parts];
7246 942 : PathTarget *child_target = copy_pathtarget(target);
7247 : AppendRelInfo **appinfos;
7248 : int nappinfos;
7249 : GroupPathExtraData child_extra;
7250 : RelOptInfo *child_grouped_rel;
7251 : RelOptInfo *child_partially_grouped_rel;
7252 :
7253 : /* Pruned or dummy children can be ignored. */
7254 942 : if (child_input_rel == NULL || IS_DUMMY_REL(child_input_rel))
7255 0 : continue;
7256 :
7257 : /*
7258 : * Copy the given "extra" structure as is and then override the
7259 : * members specific to this child.
7260 : */
7261 942 : memcpy(&child_extra, extra, sizeof(child_extra));
7262 :
7263 942 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
7264 : &nappinfos);
7265 :
7266 942 : child_target->exprs = (List *)
7267 1884 : adjust_appendrel_attrs(root,
7268 942 : (Node *) target->exprs,
7269 : nappinfos, appinfos);
7270 :
7271 : /* Translate havingQual and targetList. */
7272 942 : child_extra.havingQual = (Node *)
7273 : adjust_appendrel_attrs(root,
7274 : extra->havingQual,
7275 : nappinfos, appinfos);
7276 942 : child_extra.targetList = (List *)
7277 1884 : adjust_appendrel_attrs(root,
7278 942 : (Node *) extra->targetList,
7279 : nappinfos, appinfos);
7280 :
7281 : /*
7282 : * extra->patype was the value computed for our parent rel; patype is
7283 : * the value for this relation. For the child, our value is its
7284 : * parent rel's value.
7285 : */
7286 942 : child_extra.patype = patype;
7287 :
7288 : /*
7289 : * Create grouping relation to hold fully aggregated grouping and/or
7290 : * aggregation paths for the child.
7291 : */
7292 1884 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
7293 : child_target,
7294 942 : extra->target_parallel_safe,
7295 : child_extra.havingQual);
7296 :
7297 : /* Create grouping paths for this child relation. */
7298 942 : create_ordinary_grouping_paths(root, child_input_rel,
7299 : child_grouped_rel,
7300 : agg_costs, gd, &child_extra,
7301 : &child_partially_grouped_rel);
7302 :
7303 942 : if (child_partially_grouped_rel)
7304 : {
7305 590 : partially_grouped_live_children =
7306 590 : lappend(partially_grouped_live_children,
7307 : child_partially_grouped_rel);
7308 : }
7309 : else
7310 352 : partial_grouping_valid = false;
7311 :
7312 942 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7313 : {
7314 544 : set_cheapest(child_grouped_rel);
7315 544 : grouped_live_children = lappend(grouped_live_children,
7316 : child_grouped_rel);
7317 : }
7318 :
7319 942 : pfree(appinfos);
7320 : }
7321 :
7322 : /*
7323 : * Try to create append paths for partially grouped children. For full
7324 : * partitionwise aggregation, we might have paths in the partial_pathlist
7325 : * if parallel aggregation is possible. For partial partitionwise
7326 : * aggregation, we may have paths in both pathlist and partial_pathlist.
7327 : *
7328 : * NB: We must have a partially grouped path for every child in order to
7329 : * generate a partially grouped path for this relation.
7330 : */
7331 346 : if (partially_grouped_rel && partial_grouping_valid)
7332 : {
7333 : Assert(partially_grouped_live_children != NIL);
7334 :
7335 226 : add_paths_to_append_rel(root, partially_grouped_rel,
7336 : partially_grouped_live_children);
7337 :
7338 : /*
7339 : * We need call set_cheapest, since the finalization step will use the
7340 : * cheapest path from the rel.
7341 : */
7342 226 : if (partially_grouped_rel->pathlist)
7343 226 : set_cheapest(partially_grouped_rel);
7344 : }
7345 :
7346 : /* If possible, create append paths for fully grouped children. */
7347 346 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7348 : {
7349 : Assert(grouped_live_children != NIL);
7350 :
7351 192 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
7352 : }
7353 346 : }
7354 :
7355 : /*
7356 : * group_by_has_partkey
7357 : *
7358 : * Returns true, if all the partition keys of the given relation are part of
7359 : * the GROUP BY clauses, false otherwise.
7360 : */
7361 : static bool
7362 342 : group_by_has_partkey(RelOptInfo *input_rel,
7363 : List *targetList,
7364 : List *groupClause)
7365 : {
7366 342 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
7367 342 : int cnt = 0;
7368 : int partnatts;
7369 :
7370 : /* Input relation should be partitioned. */
7371 : Assert(input_rel->part_scheme);
7372 :
7373 : /* Rule out early, if there are no partition keys present. */
7374 342 : if (!input_rel->partexprs)
7375 0 : return false;
7376 :
7377 342 : partnatts = input_rel->part_scheme->partnatts;
7378 :
7379 558 : for (cnt = 0; cnt < partnatts; cnt++)
7380 : {
7381 366 : List *partexprs = input_rel->partexprs[cnt];
7382 : ListCell *lc;
7383 366 : bool found = false;
7384 :
7385 512 : foreach(lc, partexprs)
7386 : {
7387 362 : Expr *partexpr = lfirst(lc);
7388 :
7389 362 : if (list_member(groupexprs, partexpr))
7390 : {
7391 216 : found = true;
7392 216 : break;
7393 : }
7394 : }
7395 :
7396 : /*
7397 : * If none of the partition key expressions match with any of the
7398 : * GROUP BY expression, return false.
7399 : */
7400 366 : if (!found)
7401 150 : return false;
7402 : }
7403 :
7404 192 : return true;
7405 : }
|
