Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * planner.c
4 : * The query optimizer external interface.
5 : *
6 : * Portions Copyright (c) 1996-2020, 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 261292 : planner(Query *parse, const char *query_string, int cursorOptions,
268 : ParamListInfo boundParams)
269 : {
270 : PlannedStmt *result;
271 :
272 261292 : if (planner_hook)
273 212 : result = (*planner_hook) (parse, query_string, cursorOptions, boundParams);
274 : else
275 261080 : result = standard_planner(parse, query_string, cursorOptions, boundParams);
276 259668 : return result;
277 : }
278 :
279 : PlannedStmt *
280 261292 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
281 : ParamListInfo boundParams)
282 : {
283 : PlannedStmt *result;
284 : PlannerGlobal *glob;
285 : double tuple_fraction;
286 : PlannerInfo *root;
287 : RelOptInfo *final_rel;
288 : Path *best_path;
289 : Plan *top_plan;
290 : ListCell *lp,
291 : *lr;
292 :
293 : /*
294 : * Set up global state for this planner invocation. This data is needed
295 : * across all levels of sub-Query that might exist in the given command,
296 : * so we keep it in a separate struct that's linked to by each per-Query
297 : * PlannerInfo.
298 : */
299 261292 : glob = makeNode(PlannerGlobal);
300 :
301 261292 : glob->boundParams = boundParams;
302 261292 : glob->subplans = NIL;
303 261292 : glob->subroots = NIL;
304 261292 : glob->rewindPlanIDs = NULL;
305 261292 : glob->finalrtable = NIL;
306 261292 : glob->finalrowmarks = NIL;
307 261292 : glob->resultRelations = NIL;
308 261292 : glob->rootResultRelations = NIL;
309 261292 : glob->appendRelations = NIL;
310 261292 : glob->relationOids = NIL;
311 261292 : glob->invalItems = NIL;
312 261292 : glob->paramExecTypes = NIL;
313 261292 : glob->lastPHId = 0;
314 261292 : glob->lastRowMarkId = 0;
315 261292 : glob->lastPlanNodeId = 0;
316 261292 : glob->transientPlan = false;
317 261292 : glob->dependsOnRole = false;
318 :
319 : /*
320 : * Assess whether it's feasible to use parallel mode for this query. We
321 : * can't do this in a standalone backend, or if the command will try to
322 : * modify any data, or if this is a cursor operation, or if GUCs are set
323 : * to values that don't permit parallelism, or if parallel-unsafe
324 : * functions are present in the query tree.
325 : *
326 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
327 : * MATERIALIZED VIEW to use parallel plans, but as of now, only the leader
328 : * backend writes into a completely new table. In the future, we can
329 : * extend it to allow workers to write into the table. However, to allow
330 : * parallel updates and deletes, we have to solve other problems,
331 : * especially around 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 261292 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
339 218690 : IsUnderPostmaster &&
340 218690 : parse->commandType == CMD_SELECT &&
341 176398 : !parse->hasModifyingCTE &&
342 176314 : max_parallel_workers_per_gather > 0 &&
343 176012 : !IsParallelWorker())
344 : {
345 : /* all the cheap tests pass, so scan the query tree */
346 175980 : glob->maxParallelHazard = max_parallel_hazard(parse);
347 175980 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
348 : }
349 : else
350 : {
351 : /* skip the query tree scan, just assume it's unsafe */
352 85312 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
353 85312 : 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 414964 : glob->parallelModeNeeded = glob->parallelModeOK &&
374 153672 : (force_parallel_mode != FORCE_PARALLEL_OFF);
375 :
376 : /* Determine what fraction of the plan is likely to be scanned */
377 261292 : 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 2098 : 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 2098 : if (tuple_fraction >= 1.0)
394 0 : tuple_fraction = 0.0;
395 2098 : 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 259194 : tuple_fraction = 0.0;
402 : }
403 :
404 : /* primary planning entry point (may recurse for subqueries) */
405 261292 : root = subquery_planner(glob, parse, NULL,
406 : false, tuple_fraction);
407 :
408 : /* Select best Path and turn it into a Plan */
409 259788 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
410 259788 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
411 :
412 259788 : 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 259668 : 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 259668 : 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 259668 : if (glob->paramExecTypes != NIL)
480 : {
481 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
482 162610 : forboth(lp, glob->subplans, lr, glob->subroots)
483 : {
484 48860 : Plan *subplan = (Plan *) lfirst(lp);
485 48860 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
486 :
487 48860 : SS_finalize_plan(subroot, subplan);
488 : }
489 113750 : 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 : Assert(glob->appendRelations == NIL);
498 259668 : top_plan = set_plan_references(root, top_plan);
499 : /* ... and the subplans (both regular subplans and initplans) */
500 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
501 308528 : forboth(lp, glob->subplans, lr, glob->subroots)
502 : {
503 48860 : Plan *subplan = (Plan *) lfirst(lp);
504 48860 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
505 :
506 48860 : lfirst(lp) = set_plan_references(subroot, subplan);
507 : }
508 :
509 : /* build the PlannedStmt result */
510 259668 : result = makeNode(PlannedStmt);
511 :
512 259668 : result->commandType = parse->commandType;
513 259668 : result->queryId = parse->queryId;
514 259668 : result->hasReturning = (parse->returningList != NIL);
515 259668 : result->hasModifyingCTE = parse->hasModifyingCTE;
516 259668 : result->canSetTag = parse->canSetTag;
517 259668 : result->transientPlan = glob->transientPlan;
518 259668 : result->dependsOnRole = glob->dependsOnRole;
519 259668 : result->parallelModeNeeded = glob->parallelModeNeeded;
520 259668 : result->planTree = top_plan;
521 259668 : result->rtable = glob->finalrtable;
522 259668 : result->resultRelations = glob->resultRelations;
523 259668 : result->rootResultRelations = glob->rootResultRelations;
524 259668 : result->appendRelations = glob->appendRelations;
525 259668 : result->subplans = glob->subplans;
526 259668 : result->rewindPlanIDs = glob->rewindPlanIDs;
527 259668 : result->rowMarks = glob->finalrowmarks;
528 259668 : result->relationOids = glob->relationOids;
529 259668 : result->invalItems = glob->invalItems;
530 259668 : result->paramExecTypes = glob->paramExecTypes;
531 : /* utilityStmt should be null, but we might as well copy it */
532 259668 : result->utilityStmt = parse->utilityStmt;
533 259668 : result->stmt_location = parse->stmt_location;
534 259668 : result->stmt_len = parse->stmt_len;
535 :
536 259668 : result->jitFlags = PGJIT_NONE;
537 259668 : if (jit_enabled && jit_above_cost >= 0 &&
538 259668 : top_plan->total_cost > jit_above_cost)
539 : {
540 608 : result->jitFlags |= PGJIT_PERFORM;
541 :
542 : /*
543 : * Decide how much effort should be put into generating better code.
544 : */
545 608 : if (jit_optimize_above_cost >= 0 &&
546 608 : top_plan->total_cost > jit_optimize_above_cost)
547 462 : result->jitFlags |= PGJIT_OPT3;
548 608 : if (jit_inline_above_cost >= 0 &&
549 608 : top_plan->total_cost > jit_inline_above_cost)
550 462 : result->jitFlags |= PGJIT_INLINE;
551 :
552 : /*
553 : * Decide which operations should be JITed.
554 : */
555 608 : if (jit_expressions)
556 608 : result->jitFlags |= PGJIT_EXPR;
557 608 : if (jit_tuple_deforming)
558 608 : result->jitFlags |= PGJIT_DEFORM;
559 : }
560 :
561 259668 : if (glob->partition_directory != NULL)
562 6608 : DestroyPartitionDirectory(glob->partition_directory);
563 :
564 259668 : return result;
565 : }
566 :
567 :
568 : /*--------------------
569 : * subquery_planner
570 : * Invokes the planner on a subquery. We recurse to here for each
571 : * sub-SELECT found in the query tree.
572 : *
573 : * glob is the global state for the current planner run.
574 : * parse is the querytree produced by the parser & rewriter.
575 : * parent_root is the immediate parent Query's info (NULL at the top level).
576 : * hasRecursion is true if this is a recursive WITH query.
577 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
578 : * tuple_fraction is interpreted as explained for grouping_planner, below.
579 : *
580 : * Basically, this routine does the stuff that should only be done once
581 : * per Query object. It then calls grouping_planner. At one time,
582 : * grouping_planner could be invoked recursively on the same Query object;
583 : * that's not currently true, but we keep the separation between the two
584 : * routines anyway, in case we need it again someday.
585 : *
586 : * subquery_planner will be called recursively to handle sub-Query nodes
587 : * found within the query's expressions and rangetable.
588 : *
589 : * Returns the PlannerInfo struct ("root") that contains all data generated
590 : * while planning the subquery. In particular, the Path(s) attached to
591 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
592 : * cheapest way(s) to implement the query. The top level will select the
593 : * best Path and pass it through createplan.c to produce a finished Plan.
594 : *--------------------
595 : */
596 : PlannerInfo *
597 319108 : subquery_planner(PlannerGlobal *glob, Query *parse,
598 : PlannerInfo *parent_root,
599 : bool hasRecursion, double tuple_fraction)
600 : {
601 : PlannerInfo *root;
602 : List *newWithCheckOptions;
603 : List *newHaving;
604 : bool hasOuterJoins;
605 : bool hasResultRTEs;
606 : RelOptInfo *final_rel;
607 : ListCell *l;
608 :
609 : /* Create a PlannerInfo data structure for this subquery */
610 319108 : root = makeNode(PlannerInfo);
611 319108 : root->parse = parse;
612 319108 : root->glob = glob;
613 319108 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
614 319108 : root->parent_root = parent_root;
615 319108 : root->plan_params = NIL;
616 319108 : root->outer_params = NULL;
617 319108 : root->planner_cxt = CurrentMemoryContext;
618 319108 : root->init_plans = NIL;
619 319108 : root->cte_plan_ids = NIL;
620 319108 : root->multiexpr_params = NIL;
621 319108 : root->eq_classes = NIL;
622 319108 : root->ec_merging_done = false;
623 319108 : root->append_rel_list = NIL;
624 319108 : root->rowMarks = NIL;
625 319108 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
626 319108 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
627 319108 : root->processed_tlist = NIL;
628 319108 : root->grouping_map = NULL;
629 319108 : root->minmax_aggs = NIL;
630 319108 : root->qual_security_level = 0;
631 319108 : root->inhTargetKind = INHKIND_NONE;
632 319108 : root->hasRecursion = hasRecursion;
633 319108 : if (hasRecursion)
634 332 : root->wt_param_id = assign_special_exec_param(root);
635 : else
636 318776 : root->wt_param_id = -1;
637 319108 : root->non_recursive_path = NULL;
638 319108 : root->partColsUpdated = false;
639 :
640 : /*
641 : * If there is a WITH list, process each WITH query and either convert it
642 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
643 : */
644 319108 : if (parse->cteList)
645 1222 : SS_process_ctes(root);
646 :
647 : /*
648 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
649 : * that we don't need so many special cases to deal with that situation.
650 : */
651 319108 : replace_empty_jointree(parse);
652 :
653 : /*
654 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
655 : * to transform them into joins. Note that this step does not descend
656 : * into subqueries; if we pull up any subqueries below, their SubLinks are
657 : * processed just before pulling them up.
658 : */
659 319108 : if (parse->hasSubLinks)
660 31466 : pull_up_sublinks(root);
661 :
662 : /*
663 : * Scan the rangetable for function RTEs, do const-simplification on them,
664 : * and then inline them if possible (producing subqueries that might get
665 : * pulled up next). Recursion issues here are handled in the same way as
666 : * for SubLinks.
667 : */
668 319108 : preprocess_function_rtes(root);
669 :
670 : /*
671 : * Check to see if any subqueries in the jointree can be merged into this
672 : * query.
673 : */
674 319104 : pull_up_subqueries(root);
675 :
676 : /*
677 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
678 : * do this now because it requires applying pull_up_subqueries to the leaf
679 : * queries of the UNION ALL, which weren't touched above because they
680 : * weren't referenced by the jointree (they will be after we do this).
681 : */
682 319100 : if (parse->setOperations)
683 2716 : flatten_simple_union_all(root);
684 :
685 : /*
686 : * Survey the rangetable to see what kinds of entries are present. We can
687 : * skip some later processing if relevant SQL features are not used; for
688 : * example if there are no JOIN RTEs we can avoid the expense of doing
689 : * flatten_join_alias_vars(). This must be done after we have finished
690 : * adding rangetable entries, of course. (Note: actually, processing of
691 : * inherited or partitioned rels can cause RTEs for their child tables to
692 : * get added later; but those must all be RTE_RELATION entries, so they
693 : * don't invalidate the conclusions drawn here.)
694 : */
695 319100 : root->hasJoinRTEs = false;
696 319100 : root->hasLateralRTEs = false;
697 319100 : hasOuterJoins = false;
698 319100 : hasResultRTEs = false;
699 905232 : foreach(l, parse->rtable)
700 : {
701 586132 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
702 :
703 586132 : switch (rte->rtekind)
704 : {
705 328030 : case RTE_RELATION:
706 328030 : if (rte->inh)
707 : {
708 : /*
709 : * Check to see if the relation actually has any children;
710 : * if not, clear the inh flag so we can treat it as a
711 : * plain base relation.
712 : *
713 : * Note: this could give a false-positive result, if the
714 : * rel once had children but no longer does. We used to
715 : * be able to clear rte->inh later on when we discovered
716 : * that, but no more; we have to handle such cases as
717 : * full-fledged inheritance.
718 : */
719 230438 : rte->inh = has_subclass(rte->relid);
720 : }
721 328030 : break;
722 51318 : case RTE_JOIN:
723 51318 : root->hasJoinRTEs = true;
724 51318 : if (IS_OUTER_JOIN(rte->jointype))
725 37662 : hasOuterJoins = true;
726 51318 : break;
727 119892 : case RTE_RESULT:
728 119892 : hasResultRTEs = true;
729 119892 : break;
730 86892 : default:
731 : /* No work here for other RTE types */
732 86892 : break;
733 : }
734 :
735 586132 : if (rte->lateral)
736 1082 : root->hasLateralRTEs = true;
737 :
738 : /*
739 : * We can also determine the maximum security level required for any
740 : * securityQuals now. Addition of inheritance-child RTEs won't affect
741 : * this, because child tables don't have their own securityQuals; see
742 : * expand_single_inheritance_child().
743 : */
744 586132 : if (rte->securityQuals)
745 1152 : root->qual_security_level = Max(root->qual_security_level,
746 : list_length(rte->securityQuals));
747 : }
748 :
749 : /*
750 : * Preprocess RowMark information. We need to do this after subquery
751 : * pullup, so that all base relations are present.
752 : */
753 319100 : preprocess_rowmarks(root);
754 :
755 : /*
756 : * Set hasHavingQual to remember if HAVING clause is present. Needed
757 : * because preprocess_expression will reduce a constant-true condition to
758 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
759 : */
760 319100 : root->hasHavingQual = (parse->havingQual != NULL);
761 :
762 : /* Clear this flag; might get set in distribute_qual_to_rels */
763 319100 : root->hasPseudoConstantQuals = false;
764 :
765 : /*
766 : * Do expression preprocessing on targetlist and quals, as well as other
767 : * random expressions in the querytree. Note that we do not need to
768 : * handle sort/group expressions explicitly, because they are actually
769 : * part of the targetlist.
770 : */
771 317624 : parse->targetList = (List *)
772 319100 : preprocess_expression(root, (Node *) parse->targetList,
773 : EXPRKIND_TARGET);
774 :
775 : /* Constant-folding might have removed all set-returning functions */
776 317624 : if (parse->hasTargetSRFs)
777 3880 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
778 :
779 317624 : newWithCheckOptions = NIL;
780 318778 : foreach(l, parse->withCheckOptions)
781 : {
782 1154 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
783 :
784 1154 : wco->qual = preprocess_expression(root, wco->qual,
785 : EXPRKIND_QUAL);
786 1154 : if (wco->qual != NULL)
787 934 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
788 : }
789 317624 : parse->withCheckOptions = newWithCheckOptions;
790 :
791 317624 : parse->returningList = (List *)
792 317624 : preprocess_expression(root, (Node *) parse->returningList,
793 : EXPRKIND_TARGET);
794 :
795 317624 : preprocess_qual_conditions(root, (Node *) parse->jointree);
796 :
797 317624 : parse->havingQual = preprocess_expression(root, parse->havingQual,
798 : EXPRKIND_QUAL);
799 :
800 318808 : foreach(l, parse->windowClause)
801 : {
802 1184 : WindowClause *wc = lfirst_node(WindowClause, l);
803 :
804 : /* partitionClause/orderClause are sort/group expressions */
805 1184 : wc->startOffset = preprocess_expression(root, wc->startOffset,
806 : EXPRKIND_LIMIT);
807 1184 : wc->endOffset = preprocess_expression(root, wc->endOffset,
808 : EXPRKIND_LIMIT);
809 : }
810 :
811 317624 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
812 : EXPRKIND_LIMIT);
813 317624 : parse->limitCount = preprocess_expression(root, parse->limitCount,
814 : EXPRKIND_LIMIT);
815 :
816 317624 : if (parse->onConflict)
817 : {
818 1968 : parse->onConflict->arbiterElems = (List *)
819 984 : preprocess_expression(root,
820 984 : (Node *) parse->onConflict->arbiterElems,
821 : EXPRKIND_ARBITER_ELEM);
822 1968 : parse->onConflict->arbiterWhere =
823 1968 : preprocess_expression(root,
824 984 : parse->onConflict->arbiterWhere,
825 : EXPRKIND_QUAL);
826 1968 : parse->onConflict->onConflictSet = (List *)
827 984 : preprocess_expression(root,
828 984 : (Node *) parse->onConflict->onConflictSet,
829 : EXPRKIND_TARGET);
830 984 : parse->onConflict->onConflictWhere =
831 1968 : preprocess_expression(root,
832 984 : parse->onConflict->onConflictWhere,
833 : EXPRKIND_QUAL);
834 : /* exclRelTlist contains only Vars, so no preprocessing needed */
835 : }
836 :
837 317624 : root->append_rel_list = (List *)
838 317624 : preprocess_expression(root, (Node *) root->append_rel_list,
839 : EXPRKIND_APPINFO);
840 :
841 : /* Also need to preprocess expressions within RTEs */
842 902154 : foreach(l, parse->rtable)
843 : {
844 584530 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
845 : int kind;
846 : ListCell *lcsq;
847 :
848 584530 : if (rte->rtekind == RTE_RELATION)
849 : {
850 327896 : if (rte->tablesample)
851 152 : rte->tablesample = (TableSampleClause *)
852 152 : preprocess_expression(root,
853 152 : (Node *) rte->tablesample,
854 : EXPRKIND_TABLESAMPLE);
855 : }
856 256634 : else if (rte->rtekind == RTE_SUBQUERY)
857 : {
858 : /*
859 : * We don't want to do all preprocessing yet on the subquery's
860 : * expressions, since that will happen when we plan it. But if it
861 : * contains any join aliases of our level, those have to get
862 : * expanded now, because planning of the subquery won't do it.
863 : * That's only possible if the subquery is LATERAL.
864 : */
865 50854 : if (rte->lateral && root->hasJoinRTEs)
866 352 : rte->subquery = (Query *)
867 352 : flatten_join_alias_vars(root->parse,
868 352 : (Node *) rte->subquery);
869 : }
870 205780 : else if (rte->rtekind == RTE_FUNCTION)
871 : {
872 : /* Preprocess the function expression(s) fully */
873 30518 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
874 30518 : rte->functions = (List *)
875 30518 : preprocess_expression(root, (Node *) rte->functions, kind);
876 : }
877 175262 : else if (rte->rtekind == RTE_TABLEFUNC)
878 : {
879 : /* Preprocess the function expression(s) fully */
880 144 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
881 144 : rte->tablefunc = (TableFunc *)
882 144 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
883 : }
884 175118 : else if (rte->rtekind == RTE_VALUES)
885 : {
886 : /* Preprocess the values lists fully */
887 3928 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
888 3928 : rte->values_lists = (List *)
889 3928 : preprocess_expression(root, (Node *) rte->values_lists, kind);
890 : }
891 :
892 : /*
893 : * Process each element of the securityQuals list as if it were a
894 : * separate qual expression (as indeed it is). We need to do it this
895 : * way to get proper canonicalization of AND/OR structure. Note that
896 : * this converts each element into an implicit-AND sublist.
897 : */
898 585838 : foreach(lcsq, rte->securityQuals)
899 : {
900 1308 : lfirst(lcsq) = preprocess_expression(root,
901 1308 : (Node *) lfirst(lcsq),
902 : EXPRKIND_QUAL);
903 : }
904 : }
905 :
906 : /*
907 : * Now that we are done preprocessing expressions, and in particular done
908 : * flattening join alias variables, get rid of the joinaliasvars lists.
909 : * They no longer match what expressions in the rest of the tree look
910 : * like, because we have not preprocessed expressions in those lists (and
911 : * do not want to; for example, expanding a SubLink there would result in
912 : * a useless unreferenced subplan). Leaving them in place simply creates
913 : * a hazard for later scans of the tree. We could try to prevent that by
914 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
915 : * but that doesn't sound very reliable.
916 : */
917 317624 : if (root->hasJoinRTEs)
918 : {
919 221846 : foreach(l, parse->rtable)
920 : {
921 183430 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
922 :
923 183430 : rte->joinaliasvars = NIL;
924 : }
925 : }
926 :
927 : /*
928 : * In some cases we may want to transfer a HAVING clause into WHERE. We
929 : * cannot do so if the HAVING clause contains aggregates (obviously) or
930 : * volatile functions (since a HAVING clause is supposed to be executed
931 : * only once per group). We also can't do this if there are any nonempty
932 : * grouping sets; moving such a clause into WHERE would potentially change
933 : * the results, if any referenced column isn't present in all the grouping
934 : * sets. (If there are only empty grouping sets, then the HAVING clause
935 : * must be degenerate as discussed below.)
936 : *
937 : * Also, it may be that the clause is so expensive to execute that we're
938 : * better off doing it only once per group, despite the loss of
939 : * selectivity. This is hard to estimate short of doing the entire
940 : * planning process twice, so we use a heuristic: clauses containing
941 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
942 : * clause into WHERE, in hopes of eliminating tuples before aggregation
943 : * instead of after.
944 : *
945 : * If the query has explicit grouping then we can simply move such a
946 : * clause into WHERE; any group that fails the clause will not be in the
947 : * output because none of its tuples will reach the grouping or
948 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
949 : * HAVING clause, which we put in WHERE so that query_planner() can use it
950 : * in a gating Result node, but also keep in HAVING to ensure that we
951 : * don't emit a bogus aggregated row. (This could be done better, but it
952 : * seems not worth optimizing.)
953 : *
954 : * Note that both havingQual and parse->jointree->quals are in
955 : * implicitly-ANDed-list form at this point, even though they are declared
956 : * as Node *.
957 : */
958 317624 : newHaving = NIL;
959 318082 : foreach(l, (List *) parse->havingQual)
960 : {
961 458 : Node *havingclause = (Node *) lfirst(l);
962 :
963 888 : if ((parse->groupClause && parse->groupingSets) ||
964 546 : contain_agg_clause(havingclause) ||
965 232 : contain_volatile_functions(havingclause) ||
966 116 : contain_subplans(havingclause))
967 : {
968 : /* keep it in HAVING */
969 342 : newHaving = lappend(newHaving, havingclause);
970 : }
971 116 : else if (parse->groupClause && !parse->groupingSets)
972 : {
973 : /* move it to WHERE */
974 104 : parse->jointree->quals = (Node *)
975 104 : lappend((List *) parse->jointree->quals, havingclause);
976 : }
977 : else
978 : {
979 : /* put a copy in WHERE, keep it in HAVING */
980 24 : parse->jointree->quals = (Node *)
981 12 : lappend((List *) parse->jointree->quals,
982 : copyObject(havingclause));
983 12 : newHaving = lappend(newHaving, havingclause);
984 : }
985 : }
986 317624 : parse->havingQual = (Node *) newHaving;
987 :
988 : /* Remove any redundant GROUP BY columns */
989 317624 : remove_useless_groupby_columns(root);
990 :
991 : /*
992 : * If we have any outer joins, try to reduce them to plain inner joins.
993 : * This step is most easily done after we've done expression
994 : * preprocessing.
995 : */
996 317624 : if (hasOuterJoins)
997 32068 : reduce_outer_joins(root);
998 :
999 : /*
1000 : * If we have any RTE_RESULT relations, see if they can be deleted from
1001 : * the jointree. This step is most effectively done after we've done
1002 : * expression preprocessing and outer join reduction.
1003 : */
1004 317624 : if (hasResultRTEs)
1005 118216 : remove_useless_result_rtes(root);
1006 :
1007 : /*
1008 : * Do the main planning. If we have an inherited target relation, that
1009 : * needs special processing, else go straight to grouping_planner.
1010 : */
1011 317624 : if (parse->resultRelation &&
1012 73108 : rt_fetch(parse->resultRelation, parse->rtable)->inh)
1013 1340 : inheritance_planner(root);
1014 : else
1015 316284 : grouping_planner(root, false, tuple_fraction);
1016 :
1017 : /*
1018 : * Capture the set of outer-level param IDs we have access to, for use in
1019 : * extParam/allParam calculations later.
1020 : */
1021 317604 : SS_identify_outer_params(root);
1022 :
1023 : /*
1024 : * If any initPlans were created in this query level, adjust the surviving
1025 : * Paths' costs and parallel-safety flags to account for them. The
1026 : * initPlans won't actually get attached to the plan tree till
1027 : * create_plan() runs, but we must include their effects now.
1028 : */
1029 317604 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1030 317604 : SS_charge_for_initplans(root, final_rel);
1031 :
1032 : /*
1033 : * Make sure we've identified the cheapest Path for the final rel. (By
1034 : * doing this here not in grouping_planner, we include initPlan costs in
1035 : * the decision, though it's unlikely that will change anything.)
1036 : */
1037 317604 : set_cheapest(final_rel);
1038 :
1039 317604 : return root;
1040 : }
1041 :
1042 : /*
1043 : * preprocess_expression
1044 : * Do subquery_planner's preprocessing work for an expression,
1045 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1046 : * conditions), a HAVING clause, or a few other things.
1047 : */
1048 : static Node *
1049 2342000 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1050 : {
1051 : /*
1052 : * Fall out quickly if expression is empty. This occurs often enough to
1053 : * be worth checking. Note that null->null is the correct conversion for
1054 : * implicit-AND result format, too.
1055 : */
1056 2342000 : if (expr == NULL)
1057 1785836 : return NULL;
1058 :
1059 : /*
1060 : * If the query has any join RTEs, replace join alias variables with
1061 : * base-relation variables. We must do this first, since any expressions
1062 : * we may extract from the joinaliasvars lists have not been preprocessed.
1063 : * For example, if we did this after sublink processing, sublinks expanded
1064 : * out from join aliases would not get processed. But we can skip this in
1065 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1066 : * they can't contain any Vars of the current query level.
1067 : */
1068 556164 : if (root->hasJoinRTEs &&
1069 250792 : !(kind == EXPRKIND_RTFUNC ||
1070 125306 : kind == EXPRKIND_VALUES ||
1071 : kind == EXPRKIND_TABLESAMPLE ||
1072 : kind == EXPRKIND_TABLEFUNC))
1073 125306 : expr = flatten_join_alias_vars(root->parse, expr);
1074 :
1075 : /*
1076 : * Simplify constant expressions. For function RTEs, this was already
1077 : * done by preprocess_function_rtes ... but we have to do it again if the
1078 : * RTE is LATERAL and might have contained join alias variables.
1079 : *
1080 : * Note: an essential effect of this is to convert named-argument function
1081 : * calls to positional notation and insert the current actual values of
1082 : * any default arguments for functions. To ensure that happens, we *must*
1083 : * process all expressions here. Previous PG versions sometimes skipped
1084 : * const-simplification if it didn't seem worth the trouble, but we can't
1085 : * do that anymore.
1086 : *
1087 : * Note: this also flattens nested AND and OR expressions into N-argument
1088 : * form. All processing of a qual expression after this point must be
1089 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1090 : * with AND directly under AND, nor OR directly under OR.
1091 : */
1092 556164 : if (!(kind == EXPRKIND_RTFUNC ||
1093 370 : (kind == EXPRKIND_RTFUNC_LATERAL && !root->hasJoinRTEs)))
1094 525764 : expr = eval_const_expressions(root, expr);
1095 :
1096 : /*
1097 : * If it's a qual or havingQual, canonicalize it.
1098 : */
1099 554688 : if (kind == EXPRKIND_QUAL)
1100 : {
1101 197474 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1102 :
1103 : #ifdef OPTIMIZER_DEBUG
1104 : printf("After canonicalize_qual()\n");
1105 : pprint(expr);
1106 : #endif
1107 : }
1108 :
1109 : /* Expand SubLinks to SubPlans */
1110 554688 : if (root->parse->hasSubLinks)
1111 109450 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1112 :
1113 : /*
1114 : * XXX do not insert anything here unless you have grokked the comments in
1115 : * SS_replace_correlation_vars ...
1116 : */
1117 :
1118 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1119 554688 : if (root->query_level > 1)
1120 135604 : expr = SS_replace_correlation_vars(root, expr);
1121 :
1122 : /*
1123 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1124 : * don't want to do this before eval_const_expressions, since the latter
1125 : * would be unable to simplify a top-level AND correctly. Also,
1126 : * SS_process_sublinks expects explicit-AND format.)
1127 : */
1128 554688 : if (kind == EXPRKIND_QUAL)
1129 197474 : expr = (Node *) make_ands_implicit((Expr *) expr);
1130 :
1131 554688 : return expr;
1132 : }
1133 :
1134 : /*
1135 : * preprocess_qual_conditions
1136 : * Recursively scan the query's jointree and do subquery_planner's
1137 : * preprocessing work on each qual condition found therein.
1138 : */
1139 : static void
1140 783756 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1141 : {
1142 783756 : if (jtnode == NULL)
1143 0 : return;
1144 783756 : if (IsA(jtnode, RangeTblRef))
1145 : {
1146 : /* nothing to do here */
1147 : }
1148 391240 : else if (IsA(jtnode, FromExpr))
1149 : {
1150 331394 : FromExpr *f = (FromExpr *) jtnode;
1151 : ListCell *l;
1152 :
1153 677834 : foreach(l, f->fromlist)
1154 346440 : preprocess_qual_conditions(root, lfirst(l));
1155 :
1156 331394 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1157 : }
1158 59846 : else if (IsA(jtnode, JoinExpr))
1159 : {
1160 59846 : JoinExpr *j = (JoinExpr *) jtnode;
1161 :
1162 59846 : preprocess_qual_conditions(root, j->larg);
1163 59846 : preprocess_qual_conditions(root, j->rarg);
1164 :
1165 59846 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1166 : }
1167 : else
1168 0 : elog(ERROR, "unrecognized node type: %d",
1169 : (int) nodeTag(jtnode));
1170 : }
1171 :
1172 : /*
1173 : * preprocess_phv_expression
1174 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1175 : *
1176 : * If a LATERAL subquery references an output of another subquery, and that
1177 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1178 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1179 : * and later pull it back up during find_lateral_references, which runs after
1180 : * subquery_planner has preprocessed all the expressions that were in the
1181 : * current query level to start with. So we need to preprocess it then.
1182 : */
1183 : Expr *
1184 32 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1185 : {
1186 32 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1187 : }
1188 :
1189 : /*
1190 : * inheritance_planner
1191 : * Generate Paths in the case where the result relation is an
1192 : * inheritance set.
1193 : *
1194 : * We have to handle this case differently from cases where a source relation
1195 : * is an inheritance set. Source inheritance is expanded at the bottom of the
1196 : * plan tree (see allpaths.c), but target inheritance has to be expanded at
1197 : * the top. The reason is that for UPDATE, each target relation needs a
1198 : * different targetlist matching its own column set. Fortunately,
1199 : * the UPDATE/DELETE target can never be the nullable side of an outer join,
1200 : * so it's OK to generate the plan this way.
1201 : *
1202 : * Returns nothing; the useful output is in the Paths we attach to
1203 : * the (UPPERREL_FINAL, NULL) upperrel stored in *root.
1204 : *
1205 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1206 : * to leave this to the caller.
1207 : */
1208 : static void
1209 1340 : inheritance_planner(PlannerInfo *root)
1210 : {
1211 1340 : Query *parse = root->parse;
1212 1340 : int top_parentRTindex = parse->resultRelation;
1213 : List *select_rtable;
1214 : List *select_appinfos;
1215 : List *child_appinfos;
1216 : List *old_child_rtis;
1217 : List *new_child_rtis;
1218 : Bitmapset *subqueryRTindexes;
1219 : Index next_subquery_rti;
1220 1340 : int nominalRelation = -1;
1221 1340 : Index rootRelation = 0;
1222 1340 : List *final_rtable = NIL;
1223 1340 : List *final_rowmarks = NIL;
1224 1340 : List *final_appendrels = NIL;
1225 1340 : int save_rel_array_size = 0;
1226 1340 : RelOptInfo **save_rel_array = NULL;
1227 1340 : AppendRelInfo **save_append_rel_array = NULL;
1228 1340 : List *subpaths = NIL;
1229 1340 : List *subroots = NIL;
1230 1340 : List *resultRelations = NIL;
1231 1340 : List *withCheckOptionLists = NIL;
1232 1340 : List *returningLists = NIL;
1233 : List *rowMarks;
1234 : RelOptInfo *final_rel;
1235 : ListCell *lc;
1236 : ListCell *lc2;
1237 : Index rti;
1238 : RangeTblEntry *parent_rte;
1239 : Bitmapset *parent_relids;
1240 : Query **parent_parses;
1241 :
1242 : /* Should only get here for UPDATE or DELETE */
1243 : Assert(parse->commandType == CMD_UPDATE ||
1244 : parse->commandType == CMD_DELETE);
1245 :
1246 : /*
1247 : * We generate a modified instance of the original Query for each target
1248 : * relation, plan that, and put all the plans into a list that will be
1249 : * controlled by a single ModifyTable node. All the instances share the
1250 : * same rangetable, but each instance must have its own set of subquery
1251 : * RTEs within the finished rangetable because (1) they are likely to get
1252 : * scribbled on during planning, and (2) it's not inconceivable that
1253 : * subqueries could get planned differently in different cases. We need
1254 : * not create duplicate copies of other RTE kinds, in particular not the
1255 : * target relations, because they don't have either of those issues. Not
1256 : * having to duplicate the target relations is important because doing so
1257 : * (1) would result in a rangetable of length O(N^2) for N targets, with
1258 : * at least O(N^3) work expended here; and (2) would greatly complicate
1259 : * management of the rowMarks list.
1260 : *
1261 : * To begin with, generate a bitmapset of the relids of the subquery RTEs.
1262 : */
1263 1340 : subqueryRTindexes = NULL;
1264 1340 : rti = 1;
1265 3104 : foreach(lc, parse->rtable)
1266 : {
1267 1764 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1268 :
1269 1764 : if (rte->rtekind == RTE_SUBQUERY)
1270 116 : subqueryRTindexes = bms_add_member(subqueryRTindexes, rti);
1271 1764 : rti++;
1272 : }
1273 :
1274 : /*
1275 : * If the parent RTE is a partitioned table, we should use that as the
1276 : * nominal target relation, because the RTEs added for partitioned tables
1277 : * (including the root parent) as child members of the inheritance set do
1278 : * not appear anywhere else in the plan, so the confusion explained below
1279 : * for non-partitioning inheritance cases is not possible.
1280 : */
1281 1340 : parent_rte = rt_fetch(top_parentRTindex, parse->rtable);
1282 : Assert(parent_rte->inh);
1283 1340 : if (parent_rte->relkind == RELKIND_PARTITIONED_TABLE)
1284 : {
1285 954 : nominalRelation = top_parentRTindex;
1286 954 : rootRelation = top_parentRTindex;
1287 : }
1288 :
1289 : /*
1290 : * Before generating the real per-child-relation plans, do a cycle of
1291 : * planning as though the query were a SELECT. The objective here is to
1292 : * find out which child relations need to be processed, using the same
1293 : * expansion and pruning logic as for a SELECT. We'll then pull out the
1294 : * RangeTblEntry-s generated for the child rels, and make use of the
1295 : * AppendRelInfo entries for them to guide the real planning. (This is
1296 : * rather inefficient; we could perhaps stop short of making a full Path
1297 : * tree. But this whole function is inefficient and slated for
1298 : * destruction, so let's not contort query_planner for that.)
1299 : */
1300 : {
1301 : PlannerInfo *subroot;
1302 :
1303 : /*
1304 : * Flat-copy the PlannerInfo to prevent modification of the original.
1305 : */
1306 1340 : subroot = makeNode(PlannerInfo);
1307 1340 : memcpy(subroot, root, sizeof(PlannerInfo));
1308 :
1309 : /*
1310 : * Make a deep copy of the parsetree for this planning cycle to mess
1311 : * around with, and change it to look like a SELECT. (Hack alert: the
1312 : * target RTE still has updatedCols set if this is an UPDATE, so that
1313 : * expand_partitioned_rtentry will correctly update
1314 : * subroot->partColsUpdated.)
1315 : */
1316 1340 : subroot->parse = copyObject(root->parse);
1317 :
1318 1340 : subroot->parse->commandType = CMD_SELECT;
1319 1340 : subroot->parse->resultRelation = 0;
1320 :
1321 : /*
1322 : * Ensure the subroot has its own copy of the original
1323 : * append_rel_list, since it'll be scribbled on. (Note that at this
1324 : * point, the list only contains AppendRelInfos for flattened UNION
1325 : * ALL subqueries.)
1326 : */
1327 1340 : subroot->append_rel_list = copyObject(root->append_rel_list);
1328 :
1329 : /*
1330 : * Better make a private copy of the rowMarks, too.
1331 : */
1332 1340 : subroot->rowMarks = copyObject(root->rowMarks);
1333 :
1334 : /* There shouldn't be any OJ info to translate, as yet */
1335 : Assert(subroot->join_info_list == NIL);
1336 : /* and we haven't created PlaceHolderInfos, either */
1337 : Assert(subroot->placeholder_list == NIL);
1338 :
1339 : /* Generate Path(s) for accessing this result relation */
1340 1340 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1341 :
1342 : /* Extract the info we need. */
1343 1340 : select_rtable = subroot->parse->rtable;
1344 1340 : select_appinfos = subroot->append_rel_list;
1345 :
1346 : /*
1347 : * We need to propagate partColsUpdated back, too. (The later
1348 : * planning cycles will not set this because they won't run
1349 : * expand_partitioned_rtentry for the UPDATE target.)
1350 : */
1351 1340 : root->partColsUpdated = subroot->partColsUpdated;
1352 : }
1353 :
1354 : /*----------
1355 : * Since only one rangetable can exist in the final plan, we need to make
1356 : * sure that it contains all the RTEs needed for any child plan. This is
1357 : * complicated by the need to use separate subquery RTEs for each child.
1358 : * We arrange the final rtable as follows:
1359 : * 1. All original rtable entries (with their original RT indexes).
1360 : * 2. All the relation RTEs generated for children of the target table.
1361 : * 3. Subquery RTEs for children after the first. We need N * (K - 1)
1362 : * RT slots for this, if there are N subqueries and K child tables.
1363 : * 4. Additional RTEs generated during the child planning runs, such as
1364 : * children of inheritable RTEs other than the target table.
1365 : * We assume that each child planning run will create an identical set
1366 : * of type-4 RTEs.
1367 : *
1368 : * So the next thing to do is append the type-2 RTEs (the target table's
1369 : * children) to the original rtable. We look through select_appinfos
1370 : * to find them.
1371 : *
1372 : * To identify which AppendRelInfos are relevant as we thumb through
1373 : * select_appinfos, we need to look for both direct and indirect children
1374 : * of top_parentRTindex, so we use a bitmap of known parent relids.
1375 : * expand_inherited_rtentry() always processes a parent before any of that
1376 : * parent's children, so we should see an intermediate parent before its
1377 : * children.
1378 : *----------
1379 : */
1380 1340 : child_appinfos = NIL;
1381 1340 : old_child_rtis = NIL;
1382 1340 : new_child_rtis = NIL;
1383 1340 : parent_relids = bms_make_singleton(top_parentRTindex);
1384 4430 : foreach(lc, select_appinfos)
1385 : {
1386 3090 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1387 : RangeTblEntry *child_rte;
1388 :
1389 : /* append_rel_list contains all append rels; ignore others */
1390 3090 : if (!bms_is_member(appinfo->parent_relid, parent_relids))
1391 224 : continue;
1392 :
1393 : /* remember relevant AppendRelInfos for use below */
1394 2866 : child_appinfos = lappend(child_appinfos, appinfo);
1395 :
1396 : /* extract RTE for this child rel */
1397 2866 : child_rte = rt_fetch(appinfo->child_relid, select_rtable);
1398 :
1399 : /* and append it to the original rtable */
1400 2866 : parse->rtable = lappend(parse->rtable, child_rte);
1401 :
1402 : /* remember child's index in the SELECT rtable */
1403 2866 : old_child_rtis = lappend_int(old_child_rtis, appinfo->child_relid);
1404 :
1405 : /* and its new index in the final rtable */
1406 2866 : new_child_rtis = lappend_int(new_child_rtis, list_length(parse->rtable));
1407 :
1408 : /* if child is itself partitioned, update parent_relids */
1409 2866 : if (child_rte->inh)
1410 : {
1411 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1412 300 : parent_relids = bms_add_member(parent_relids, appinfo->child_relid);
1413 : }
1414 : }
1415 :
1416 : /*
1417 : * It's possible that the RTIs we just assigned for the child rels in the
1418 : * final rtable are different from what they were in the SELECT query.
1419 : * Adjust the AppendRelInfos so that they will correctly map RT indexes to
1420 : * the final indexes. We can do this left-to-right since no child rel's
1421 : * final RT index could be greater than what it had in the SELECT query.
1422 : */
1423 4206 : forboth(lc, old_child_rtis, lc2, new_child_rtis)
1424 : {
1425 2866 : int old_child_rti = lfirst_int(lc);
1426 2866 : int new_child_rti = lfirst_int(lc2);
1427 :
1428 2866 : if (old_child_rti == new_child_rti)
1429 2850 : continue; /* nothing to do */
1430 :
1431 : Assert(old_child_rti > new_child_rti);
1432 :
1433 16 : ChangeVarNodes((Node *) child_appinfos,
1434 : old_child_rti, new_child_rti, 0);
1435 : }
1436 :
1437 : /*
1438 : * Now set up rangetable entries for subqueries for additional children
1439 : * (the first child will just use the original ones). These all have to
1440 : * look more or less real, or EXPLAIN will get unhappy; so we just make
1441 : * them all clones of the original subqueries.
1442 : */
1443 1340 : next_subquery_rti = list_length(parse->rtable) + 1;
1444 1340 : if (subqueryRTindexes != NULL)
1445 : {
1446 76 : int n_children = list_length(child_appinfos);
1447 :
1448 208 : while (n_children-- > 1)
1449 : {
1450 132 : int oldrti = -1;
1451 :
1452 344 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1453 : {
1454 : RangeTblEntry *subqrte;
1455 :
1456 212 : subqrte = rt_fetch(oldrti, parse->rtable);
1457 212 : parse->rtable = lappend(parse->rtable, copyObject(subqrte));
1458 : }
1459 : }
1460 : }
1461 :
1462 : /*
1463 : * The query for each child is obtained by translating the query for its
1464 : * immediate parent, since the AppendRelInfo data we have shows deltas
1465 : * between parents and children. We use the parent_parses array to
1466 : * remember the appropriate query trees. This is indexed by parent relid.
1467 : * Since the maximum number of parents is limited by the number of RTEs in
1468 : * the SELECT query, we use that number to allocate the array. An extra
1469 : * entry is needed since relids start from 1.
1470 : */
1471 1340 : parent_parses = (Query **) palloc0((list_length(select_rtable) + 1) *
1472 : sizeof(Query *));
1473 1340 : parent_parses[top_parentRTindex] = parse;
1474 :
1475 : /*
1476 : * And now we can get on with generating a plan for each child table.
1477 : */
1478 4206 : foreach(lc, child_appinfos)
1479 : {
1480 2866 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1481 2866 : Index this_subquery_rti = next_subquery_rti;
1482 : Query *parent_parse;
1483 : PlannerInfo *subroot;
1484 : RangeTblEntry *child_rte;
1485 : RelOptInfo *sub_final_rel;
1486 : Path *subpath;
1487 :
1488 : /*
1489 : * expand_inherited_rtentry() always processes a parent before any of
1490 : * that parent's children, so the parent query for this relation
1491 : * should already be available.
1492 : */
1493 2866 : parent_parse = parent_parses[appinfo->parent_relid];
1494 : Assert(parent_parse != NULL);
1495 :
1496 : /*
1497 : * We need a working copy of the PlannerInfo so that we can control
1498 : * propagation of information back to the main copy.
1499 : */
1500 2866 : subroot = makeNode(PlannerInfo);
1501 2866 : memcpy(subroot, root, sizeof(PlannerInfo));
1502 :
1503 : /*
1504 : * Generate modified query with this rel as target. We first apply
1505 : * adjust_appendrel_attrs, which copies the Query and changes
1506 : * references to the parent RTE to refer to the current child RTE,
1507 : * then fool around with subquery RTEs.
1508 : */
1509 2866 : subroot->parse = (Query *)
1510 2866 : adjust_appendrel_attrs(subroot,
1511 : (Node *) parent_parse,
1512 : 1, &appinfo);
1513 :
1514 : /*
1515 : * If there are securityQuals attached to the parent, move them to the
1516 : * child rel (they've already been transformed properly for that).
1517 : */
1518 2866 : parent_rte = rt_fetch(appinfo->parent_relid, subroot->parse->rtable);
1519 2866 : child_rte = rt_fetch(appinfo->child_relid, subroot->parse->rtable);
1520 2866 : child_rte->securityQuals = parent_rte->securityQuals;
1521 2866 : parent_rte->securityQuals = NIL;
1522 :
1523 : /*
1524 : * HACK: setting this to a value other than INHKIND_NONE signals to
1525 : * relation_excluded_by_constraints() to treat the result relation as
1526 : * being an appendrel member.
1527 : */
1528 2866 : subroot->inhTargetKind =
1529 2866 : (rootRelation != 0) ? INHKIND_PARTITIONED : INHKIND_INHERITED;
1530 :
1531 : /*
1532 : * If this child is further partitioned, remember it as a parent.
1533 : * Since a partitioned table does not have any data, we don't need to
1534 : * create a plan for it, and we can stop processing it here. We do,
1535 : * however, need to remember its modified PlannerInfo for use when
1536 : * processing its children, since we'll update their varnos based on
1537 : * the delta from immediate parent to child, not from top to child.
1538 : *
1539 : * Note: a very non-obvious point is that we have not yet added
1540 : * duplicate subquery RTEs to the subroot's rtable. We mustn't,
1541 : * because then its children would have two sets of duplicates,
1542 : * confusing matters.
1543 : */
1544 2866 : if (child_rte->inh)
1545 : {
1546 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1547 300 : parent_parses[appinfo->child_relid] = subroot->parse;
1548 360 : continue;
1549 : }
1550 :
1551 : /*
1552 : * Set the nominal target relation of the ModifyTable node if not
1553 : * already done. If the target is a partitioned table, we already set
1554 : * nominalRelation to refer to the partition root, above. For
1555 : * non-partitioned inheritance cases, we'll use the first child
1556 : * relation (even if it's excluded) as the nominal target relation.
1557 : * Because of the way expand_inherited_rtentry works, that should be
1558 : * the RTE representing the parent table in its role as a simple
1559 : * member of the inheritance set.
1560 : *
1561 : * It would be logically cleaner to *always* use the inheritance
1562 : * parent RTE as the nominal relation; but that RTE is not otherwise
1563 : * referenced in the plan in the non-partitioned inheritance case.
1564 : * Instead the duplicate child RTE created by expand_inherited_rtentry
1565 : * is used elsewhere in the plan, so using the original parent RTE
1566 : * would give rise to confusing use of multiple aliases in EXPLAIN
1567 : * output for what the user will think is the "same" table. OTOH,
1568 : * it's not a problem in the partitioned inheritance case, because
1569 : * there is no duplicate RTE for the parent.
1570 : */
1571 2566 : if (nominalRelation < 0)
1572 386 : nominalRelation = appinfo->child_relid;
1573 :
1574 : /*
1575 : * As above, each child plan run needs its own append_rel_list and
1576 : * rowmarks, which should start out as pristine copies of the
1577 : * originals. There can't be any references to UPDATE/DELETE target
1578 : * rels in them; but there could be subquery references, which we'll
1579 : * fix up in a moment.
1580 : */
1581 2566 : subroot->append_rel_list = copyObject(root->append_rel_list);
1582 2566 : subroot->rowMarks = copyObject(root->rowMarks);
1583 :
1584 : /*
1585 : * If this isn't the first child Query, adjust Vars and jointree
1586 : * entries to reference the appropriate set of subquery RTEs.
1587 : */
1588 2566 : if (final_rtable != NIL && subqueryRTindexes != NULL)
1589 : {
1590 120 : int oldrti = -1;
1591 :
1592 312 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1593 : {
1594 192 : Index newrti = next_subquery_rti++;
1595 :
1596 192 : ChangeVarNodes((Node *) subroot->parse, oldrti, newrti, 0);
1597 192 : ChangeVarNodes((Node *) subroot->append_rel_list,
1598 : oldrti, newrti, 0);
1599 192 : ChangeVarNodes((Node *) subroot->rowMarks, oldrti, newrti, 0);
1600 : }
1601 : }
1602 :
1603 : /* There shouldn't be any OJ info to translate, as yet */
1604 : Assert(subroot->join_info_list == NIL);
1605 : /* and we haven't created PlaceHolderInfos, either */
1606 : Assert(subroot->placeholder_list == NIL);
1607 :
1608 : /* Generate Path(s) for accessing this result relation */
1609 2566 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1610 :
1611 : /*
1612 : * Select cheapest path in case there's more than one. We always run
1613 : * modification queries to conclusion, so we care only for the
1614 : * cheapest-total path.
1615 : */
1616 2566 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
1617 2566 : set_cheapest(sub_final_rel);
1618 2566 : subpath = sub_final_rel->cheapest_total_path;
1619 :
1620 : /*
1621 : * If this child rel was excluded by constraint exclusion, exclude it
1622 : * from the result plan.
1623 : */
1624 2566 : if (IS_DUMMY_REL(sub_final_rel))
1625 60 : continue;
1626 :
1627 : /*
1628 : * If this is the first non-excluded child, its post-planning rtable
1629 : * becomes the initial contents of final_rtable; otherwise, copy its
1630 : * modified subquery RTEs into final_rtable, to ensure we have sane
1631 : * copies of those. Also save the first non-excluded child's version
1632 : * of the rowmarks list; we assume all children will end up with
1633 : * equivalent versions of that. Likewise for append_rel_list.
1634 : */
1635 2506 : if (final_rtable == NIL)
1636 : {
1637 1308 : final_rtable = subroot->parse->rtable;
1638 1308 : final_rowmarks = subroot->rowMarks;
1639 1308 : final_appendrels = subroot->append_rel_list;
1640 : }
1641 : else
1642 : {
1643 : Assert(list_length(final_rtable) ==
1644 : list_length(subroot->parse->rtable));
1645 1198 : if (subqueryRTindexes != NULL)
1646 : {
1647 120 : int oldrti = -1;
1648 :
1649 312 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1650 : {
1651 192 : Index newrti = this_subquery_rti++;
1652 : RangeTblEntry *subqrte;
1653 : ListCell *newrticell;
1654 :
1655 192 : subqrte = rt_fetch(newrti, subroot->parse->rtable);
1656 192 : newrticell = list_nth_cell(final_rtable, newrti - 1);
1657 192 : lfirst(newrticell) = subqrte;
1658 : }
1659 : }
1660 : }
1661 :
1662 : /*
1663 : * We need to collect all the RelOptInfos from all child plans into
1664 : * the main PlannerInfo, since setrefs.c will need them. We use the
1665 : * last child's simple_rel_array, so we have to propagate forward the
1666 : * RelOptInfos that were already built in previous children.
1667 : */
1668 : Assert(subroot->simple_rel_array_size >= save_rel_array_size);
1669 10472 : for (rti = 1; rti < save_rel_array_size; rti++)
1670 : {
1671 7966 : RelOptInfo *brel = save_rel_array[rti];
1672 :
1673 7966 : if (brel)
1674 2974 : subroot->simple_rel_array[rti] = brel;
1675 : }
1676 2506 : save_rel_array_size = subroot->simple_rel_array_size;
1677 2506 : save_rel_array = subroot->simple_rel_array;
1678 2506 : save_append_rel_array = subroot->append_rel_array;
1679 :
1680 : /*
1681 : * Make sure any initplans from this rel get into the outer list. Note
1682 : * we're effectively assuming all children generate the same
1683 : * init_plans.
1684 : */
1685 2506 : root->init_plans = subroot->init_plans;
1686 :
1687 : /* Build list of sub-paths */
1688 2506 : subpaths = lappend(subpaths, subpath);
1689 :
1690 : /* Build list of modified subroots, too */
1691 2506 : subroots = lappend(subroots, subroot);
1692 :
1693 : /* Build list of target-relation RT indexes */
1694 2506 : resultRelations = lappend_int(resultRelations, appinfo->child_relid);
1695 :
1696 : /* Build lists of per-relation WCO and RETURNING targetlists */
1697 2506 : if (parse->withCheckOptions)
1698 272 : withCheckOptionLists = lappend(withCheckOptionLists,
1699 272 : subroot->parse->withCheckOptions);
1700 2506 : if (parse->returningList)
1701 316 : returningLists = lappend(returningLists,
1702 316 : subroot->parse->returningList);
1703 :
1704 : Assert(!parse->onConflict);
1705 : }
1706 :
1707 : /* Result path must go into outer query's FINAL upperrel */
1708 1340 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1709 :
1710 : /*
1711 : * We don't currently worry about setting final_rel's consider_parallel
1712 : * flag in this case, nor about allowing FDWs or create_upper_paths_hook
1713 : * to get control here.
1714 : */
1715 :
1716 1340 : if (subpaths == NIL)
1717 : {
1718 : /*
1719 : * We managed to exclude every child rel, so generate a dummy path
1720 : * representing the empty set. Although it's clear that no data will
1721 : * be updated or deleted, we will still need to have a ModifyTable
1722 : * node so that any statement triggers are executed. (This could be
1723 : * cleaner if we fixed nodeModifyTable.c to support zero child nodes,
1724 : * but that probably wouldn't be a net win.)
1725 : */
1726 : Path *dummy_path;
1727 :
1728 : /* tlist processing never got done, either */
1729 32 : root->processed_tlist = preprocess_targetlist(root);
1730 32 : final_rel->reltarget = create_pathtarget(root, root->processed_tlist);
1731 :
1732 : /* Make a dummy path, cf set_dummy_rel_pathlist() */
1733 32 : dummy_path = (Path *) create_append_path(NULL, final_rel, NIL, NIL,
1734 : NIL, NULL, 0, false,
1735 : NIL, -1);
1736 :
1737 : /* These lists must be nonempty to make a valid ModifyTable node */
1738 32 : subpaths = list_make1(dummy_path);
1739 32 : subroots = list_make1(root);
1740 32 : resultRelations = list_make1_int(parse->resultRelation);
1741 32 : if (parse->withCheckOptions)
1742 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
1743 32 : if (parse->returningList)
1744 12 : returningLists = list_make1(parse->returningList);
1745 : /* Disable tuple routing, too, just to be safe */
1746 32 : root->partColsUpdated = false;
1747 : }
1748 : else
1749 : {
1750 : /*
1751 : * Put back the final adjusted rtable into the master copy of the
1752 : * Query. (We mustn't do this if we found no non-excluded children,
1753 : * since we never saved an adjusted rtable at all.)
1754 : */
1755 1308 : parse->rtable = final_rtable;
1756 1308 : root->simple_rel_array_size = save_rel_array_size;
1757 1308 : root->simple_rel_array = save_rel_array;
1758 1308 : root->append_rel_array = save_append_rel_array;
1759 :
1760 : /* Must reconstruct master's simple_rte_array, too */
1761 1308 : root->simple_rte_array = (RangeTblEntry **)
1762 1308 : palloc0((list_length(final_rtable) + 1) * sizeof(RangeTblEntry *));
1763 1308 : rti = 1;
1764 6250 : foreach(lc, final_rtable)
1765 : {
1766 4942 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1767 :
1768 4942 : root->simple_rte_array[rti++] = rte;
1769 : }
1770 :
1771 : /* Put back adjusted rowmarks and appendrels, too */
1772 1308 : root->rowMarks = final_rowmarks;
1773 1308 : root->append_rel_list = final_appendrels;
1774 : }
1775 :
1776 : /*
1777 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node will
1778 : * have dealt with fetching non-locked marked rows, else we need to have
1779 : * ModifyTable do that.
1780 : */
1781 1340 : if (parse->rowMarks)
1782 0 : rowMarks = NIL;
1783 : else
1784 1340 : rowMarks = root->rowMarks;
1785 :
1786 : /* Create Path representing a ModifyTable to do the UPDATE/DELETE work */
1787 1340 : add_path(final_rel, (Path *)
1788 1340 : create_modifytable_path(root, final_rel,
1789 : parse->commandType,
1790 1340 : parse->canSetTag,
1791 : nominalRelation,
1792 : rootRelation,
1793 1340 : root->partColsUpdated,
1794 : resultRelations,
1795 : subpaths,
1796 : subroots,
1797 : withCheckOptionLists,
1798 : returningLists,
1799 : rowMarks,
1800 : NULL,
1801 : assign_special_exec_param(root)));
1802 1340 : }
1803 :
1804 : /*--------------------
1805 : * grouping_planner
1806 : * Perform planning steps related to grouping, aggregation, etc.
1807 : *
1808 : * This function adds all required top-level processing to the scan/join
1809 : * Path(s) produced by query_planner.
1810 : *
1811 : * If inheritance_update is true, we're being called from inheritance_planner
1812 : * and should not include a ModifyTable step in the resulting Path(s).
1813 : * (inheritance_planner will create a single ModifyTable node covering all the
1814 : * target tables.)
1815 : *
1816 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1817 : * tuple_fraction is interpreted as follows:
1818 : * 0: expect all tuples to be retrieved (normal case)
1819 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1820 : * from the plan to be retrieved
1821 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1822 : * expected to be retrieved (ie, a LIMIT specification)
1823 : *
1824 : * Returns nothing; the useful output is in the Paths we attach to the
1825 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1826 : * root->processed_tlist contains the final processed targetlist.
1827 : *
1828 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1829 : * to leave this to the caller.
1830 : *--------------------
1831 : */
1832 : static void
1833 320190 : grouping_planner(PlannerInfo *root, bool inheritance_update,
1834 : double tuple_fraction)
1835 : {
1836 320190 : Query *parse = root->parse;
1837 320190 : int64 offset_est = 0;
1838 320190 : int64 count_est = 0;
1839 320190 : double limit_tuples = -1.0;
1840 320190 : bool have_postponed_srfs = false;
1841 : PathTarget *final_target;
1842 : List *final_targets;
1843 : List *final_targets_contain_srfs;
1844 : bool final_target_parallel_safe;
1845 : RelOptInfo *current_rel;
1846 : RelOptInfo *final_rel;
1847 : FinalPathExtraData extra;
1848 : ListCell *lc;
1849 :
1850 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1851 320190 : if (parse->limitCount || parse->limitOffset)
1852 : {
1853 3364 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1854 : &offset_est, &count_est);
1855 :
1856 : /*
1857 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1858 : * estimate the effects of using a bounded sort.
1859 : */
1860 3364 : if (count_est > 0 && offset_est >= 0)
1861 3130 : limit_tuples = (double) count_est + (double) offset_est;
1862 : }
1863 :
1864 : /* Make tuple_fraction accessible to lower-level routines */
1865 320190 : root->tuple_fraction = tuple_fraction;
1866 :
1867 320190 : if (parse->setOperations)
1868 : {
1869 : /*
1870 : * If there's a top-level ORDER BY, assume we have to fetch all the
1871 : * tuples. This might be too simplistic given all the hackery below
1872 : * to possibly avoid the sort; but the odds of accurate estimates here
1873 : * are pretty low anyway. XXX try to get rid of this in favor of
1874 : * letting plan_set_operations generate both fast-start and
1875 : * cheapest-total paths.
1876 : */
1877 1354 : if (parse->sortClause)
1878 372 : root->tuple_fraction = 0.0;
1879 :
1880 : /*
1881 : * Construct Paths for set operations. The results will not need any
1882 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1883 : * special work for recursive unions is the responsibility of
1884 : * plan_set_operations.
1885 : */
1886 1354 : current_rel = plan_set_operations(root);
1887 :
1888 : /*
1889 : * We should not need to call preprocess_targetlist, since we must be
1890 : * in a SELECT query node. Instead, use the processed_tlist returned
1891 : * by plan_set_operations (since this tells whether it returned any
1892 : * resjunk columns!), and transfer any sort key information from the
1893 : * original tlist.
1894 : */
1895 : Assert(parse->commandType == CMD_SELECT);
1896 :
1897 : /* for safety, copy processed_tlist instead of modifying in-place */
1898 1354 : root->processed_tlist =
1899 1354 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1900 : parse->targetList);
1901 :
1902 : /* Also extract the PathTarget form of the setop result tlist */
1903 1354 : final_target = current_rel->cheapest_total_path->pathtarget;
1904 :
1905 : /* And check whether it's parallel safe */
1906 : final_target_parallel_safe =
1907 1354 : is_parallel_safe(root, (Node *) final_target->exprs);
1908 :
1909 : /* The setop result tlist couldn't contain any SRFs */
1910 : Assert(!parse->hasTargetSRFs);
1911 1354 : final_targets = final_targets_contain_srfs = NIL;
1912 :
1913 : /*
1914 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1915 : * checked already, but let's make sure).
1916 : */
1917 1354 : if (parse->rowMarks)
1918 0 : ereport(ERROR,
1919 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1920 : /*------
1921 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1922 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1923 : LCS_asString(linitial_node(RowMarkClause,
1924 : parse->rowMarks)->strength))));
1925 :
1926 : /*
1927 : * Calculate pathkeys that represent result ordering requirements
1928 : */
1929 : Assert(parse->distinctClause == NIL);
1930 1354 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1931 : parse->sortClause,
1932 : root->processed_tlist);
1933 : }
1934 : else
1935 : {
1936 : /* No set operations, do regular planning */
1937 : PathTarget *sort_input_target;
1938 : List *sort_input_targets;
1939 : List *sort_input_targets_contain_srfs;
1940 : bool sort_input_target_parallel_safe;
1941 : PathTarget *grouping_target;
1942 : List *grouping_targets;
1943 : List *grouping_targets_contain_srfs;
1944 : bool grouping_target_parallel_safe;
1945 : PathTarget *scanjoin_target;
1946 : List *scanjoin_targets;
1947 : List *scanjoin_targets_contain_srfs;
1948 : bool scanjoin_target_parallel_safe;
1949 : bool scanjoin_target_same_exprs;
1950 : bool have_grouping;
1951 : AggClauseCosts agg_costs;
1952 318836 : WindowFuncLists *wflists = NULL;
1953 318836 : List *activeWindows = NIL;
1954 318836 : grouping_sets_data *gset_data = NULL;
1955 : standard_qp_extra qp_extra;
1956 :
1957 : /* A recursive query should always have setOperations */
1958 : Assert(!root->hasRecursion);
1959 :
1960 : /* Preprocess grouping sets and GROUP BY clause, if any */
1961 318836 : if (parse->groupingSets)
1962 : {
1963 452 : gset_data = preprocess_grouping_sets(root);
1964 : }
1965 : else
1966 : {
1967 : /* Preprocess regular GROUP BY clause, if any */
1968 318384 : if (parse->groupClause)
1969 1774 : parse->groupClause = preprocess_groupclause(root, NIL);
1970 : }
1971 :
1972 : /*
1973 : * Preprocess targetlist. Note that much of the remaining planning
1974 : * work will be done with the PathTarget representation of tlists, but
1975 : * we must also maintain the full representation of the final tlist so
1976 : * that we can transfer its decoration (resnames etc) to the topmost
1977 : * tlist of the finished Plan. This is kept in processed_tlist.
1978 : */
1979 318832 : root->processed_tlist = preprocess_targetlist(root);
1980 :
1981 : /*
1982 : * Collect statistics about aggregates for estimating costs, and mark
1983 : * all the aggregates with resolved aggtranstypes. We must do this
1984 : * before slicing and dicing the tlist into various pathtargets, else
1985 : * some copies of the Aggref nodes might escape being marked with the
1986 : * correct transtypes.
1987 : *
1988 : * Note: currently, we do not detect duplicate aggregates here. This
1989 : * may result in somewhat-overestimated cost, which is fine for our
1990 : * purposes since all Paths will get charged the same. But at some
1991 : * point we might wish to do that detection in the planner, rather
1992 : * than during executor startup.
1993 : */
1994 2550656 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1995 318832 : if (parse->hasAggs)
1996 : {
1997 24856 : get_agg_clause_costs(root, (Node *) root->processed_tlist,
1998 : AGGSPLIT_SIMPLE, &agg_costs);
1999 24856 : get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
2000 : &agg_costs);
2001 : }
2002 :
2003 : /*
2004 : * Locate any window functions in the tlist. (We don't need to look
2005 : * anywhere else, since expressions used in ORDER BY will be in there
2006 : * too.) Note that they could all have been eliminated by constant
2007 : * folding, in which case we don't need to do any more work.
2008 : */
2009 318832 : if (parse->hasWindowFuncs)
2010 : {
2011 1116 : wflists = find_window_functions((Node *) root->processed_tlist,
2012 1116 : list_length(parse->windowClause));
2013 1116 : if (wflists->numWindowFuncs > 0)
2014 1112 : activeWindows = select_active_windows(root, wflists);
2015 : else
2016 4 : parse->hasWindowFuncs = false;
2017 : }
2018 :
2019 : /*
2020 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
2021 : * adding logic between here and the query_planner() call. Anything
2022 : * that is needed in MIN/MAX-optimizable cases will have to be
2023 : * duplicated in planagg.c.
2024 : */
2025 318832 : if (parse->hasAggs)
2026 24856 : preprocess_minmax_aggregates(root);
2027 :
2028 : /*
2029 : * Figure out whether there's a hard limit on the number of rows that
2030 : * query_planner's result subplan needs to return. Even if we know a
2031 : * hard limit overall, it doesn't apply if the query has any
2032 : * grouping/aggregation operations, or SRFs in the tlist.
2033 : */
2034 318832 : if (parse->groupClause ||
2035 316638 : parse->groupingSets ||
2036 316610 : parse->distinctClause ||
2037 316184 : parse->hasAggs ||
2038 293304 : parse->hasWindowFuncs ||
2039 292272 : parse->hasTargetSRFs ||
2040 288504 : root->hasHavingQual)
2041 30340 : root->limit_tuples = -1.0;
2042 : else
2043 288492 : root->limit_tuples = limit_tuples;
2044 :
2045 : /* Set up data needed by standard_qp_callback */
2046 318832 : qp_extra.activeWindows = activeWindows;
2047 318832 : qp_extra.groupClause = (gset_data
2048 448 : ? (gset_data->rollups ? linitial_node(RollupData, gset_data->rollups)->groupClause : NIL)
2049 319280 : : parse->groupClause);
2050 :
2051 : /*
2052 : * Generate the best unsorted and presorted paths for the scan/join
2053 : * portion of this Query, ie the processing represented by the
2054 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
2055 : * We also generate (in standard_qp_callback) pathkey representations
2056 : * of the query's sort clause, distinct clause, etc.
2057 : */
2058 318832 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
2059 :
2060 : /*
2061 : * Convert the query's result tlist into PathTarget format.
2062 : *
2063 : * Note: this cannot be done before query_planner() has performed
2064 : * appendrel expansion, because that might add resjunk entries to
2065 : * root->processed_tlist. Waiting till afterwards is also helpful
2066 : * because the target width estimates can use per-Var width numbers
2067 : * that were obtained within query_planner().
2068 : */
2069 318820 : final_target = create_pathtarget(root, root->processed_tlist);
2070 : final_target_parallel_safe =
2071 318820 : is_parallel_safe(root, (Node *) final_target->exprs);
2072 :
2073 : /*
2074 : * If ORDER BY was given, consider whether we should use a post-sort
2075 : * projection, and compute the adjusted target for preceding steps if
2076 : * so.
2077 : */
2078 318820 : if (parse->sortClause)
2079 : {
2080 38790 : sort_input_target = make_sort_input_target(root,
2081 : final_target,
2082 : &have_postponed_srfs);
2083 : sort_input_target_parallel_safe =
2084 38790 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
2085 : }
2086 : else
2087 : {
2088 280030 : sort_input_target = final_target;
2089 280030 : sort_input_target_parallel_safe = final_target_parallel_safe;
2090 : }
2091 :
2092 : /*
2093 : * If we have window functions to deal with, the output from any
2094 : * grouping step needs to be what the window functions want;
2095 : * otherwise, it should be sort_input_target.
2096 : */
2097 318820 : if (activeWindows)
2098 : {
2099 1112 : grouping_target = make_window_input_target(root,
2100 : final_target,
2101 : activeWindows);
2102 : grouping_target_parallel_safe =
2103 1112 : is_parallel_safe(root, (Node *) grouping_target->exprs);
2104 : }
2105 : else
2106 : {
2107 317708 : grouping_target = sort_input_target;
2108 317708 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
2109 : }
2110 :
2111 : /*
2112 : * If we have grouping or aggregation to do, the topmost scan/join
2113 : * plan node must emit what the grouping step wants; otherwise, it
2114 : * should emit grouping_target.
2115 : */
2116 316626 : have_grouping = (parse->groupClause || parse->groupingSets ||
2117 635446 : parse->hasAggs || root->hasHavingQual);
2118 318820 : if (have_grouping)
2119 : {
2120 25132 : scanjoin_target = make_group_input_target(root, final_target);
2121 : scanjoin_target_parallel_safe =
2122 25132 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2123 : }
2124 : else
2125 : {
2126 293688 : scanjoin_target = grouping_target;
2127 293688 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2128 : }
2129 :
2130 : /*
2131 : * If there are any SRFs in the targetlist, we must separate each of
2132 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2133 : * each of the named targets with a SRF-free version, and remember the
2134 : * list of additional projection steps we need to add afterwards.
2135 : */
2136 318820 : if (parse->hasTargetSRFs)
2137 : {
2138 : /* final_target doesn't recompute any SRFs in sort_input_target */
2139 3880 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2140 : &final_targets,
2141 : &final_targets_contain_srfs);
2142 3880 : final_target = linitial_node(PathTarget, final_targets);
2143 : Assert(!linitial_int(final_targets_contain_srfs));
2144 : /* likewise for sort_input_target vs. grouping_target */
2145 3880 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2146 : &sort_input_targets,
2147 : &sort_input_targets_contain_srfs);
2148 3880 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2149 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2150 : /* likewise for grouping_target vs. scanjoin_target */
2151 3880 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
2152 : &grouping_targets,
2153 : &grouping_targets_contain_srfs);
2154 3880 : grouping_target = linitial_node(PathTarget, grouping_targets);
2155 : Assert(!linitial_int(grouping_targets_contain_srfs));
2156 : /* scanjoin_target will not have any SRFs precomputed for it */
2157 3880 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2158 : &scanjoin_targets,
2159 : &scanjoin_targets_contain_srfs);
2160 3880 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2161 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2162 : }
2163 : else
2164 : {
2165 : /* initialize lists; for most of these, dummy values are OK */
2166 314940 : final_targets = final_targets_contain_srfs = NIL;
2167 314940 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2168 314940 : grouping_targets = grouping_targets_contain_srfs = NIL;
2169 314940 : scanjoin_targets = list_make1(scanjoin_target);
2170 314940 : scanjoin_targets_contain_srfs = NIL;
2171 : }
2172 :
2173 : /* Apply scan/join target. */
2174 318820 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2175 318820 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2176 318820 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2177 : scanjoin_targets_contain_srfs,
2178 : scanjoin_target_parallel_safe,
2179 : scanjoin_target_same_exprs);
2180 :
2181 : /*
2182 : * Save the various upper-rel PathTargets we just computed into
2183 : * root->upper_targets[]. The core code doesn't use this, but it
2184 : * provides a convenient place for extensions to get at the info. For
2185 : * consistency, we save all the intermediate targets, even though some
2186 : * of the corresponding upperrels might not be needed for this query.
2187 : */
2188 318820 : root->upper_targets[UPPERREL_FINAL] = final_target;
2189 318820 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2190 318820 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2191 318820 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2192 318820 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2193 :
2194 : /*
2195 : * If we have grouping and/or aggregation, consider ways to implement
2196 : * that. We build a new upperrel representing the output of this
2197 : * phase.
2198 : */
2199 318820 : if (have_grouping)
2200 : {
2201 25132 : current_rel = create_grouping_paths(root,
2202 : current_rel,
2203 : grouping_target,
2204 : grouping_target_parallel_safe,
2205 : &agg_costs,
2206 : gset_data);
2207 : /* Fix things up if grouping_target contains SRFs */
2208 25128 : if (parse->hasTargetSRFs)
2209 84 : adjust_paths_for_srfs(root, current_rel,
2210 : grouping_targets,
2211 : grouping_targets_contain_srfs);
2212 : }
2213 :
2214 : /*
2215 : * If we have window functions, consider ways to implement those. We
2216 : * build a new upperrel representing the output of this phase.
2217 : */
2218 318816 : if (activeWindows)
2219 : {
2220 1112 : current_rel = create_window_paths(root,
2221 : current_rel,
2222 : grouping_target,
2223 : sort_input_target,
2224 : sort_input_target_parallel_safe,
2225 : wflists,
2226 : activeWindows);
2227 : /* Fix things up if sort_input_target contains SRFs */
2228 1112 : if (parse->hasTargetSRFs)
2229 8 : adjust_paths_for_srfs(root, current_rel,
2230 : sort_input_targets,
2231 : sort_input_targets_contain_srfs);
2232 : }
2233 :
2234 : /*
2235 : * If there is a DISTINCT clause, consider ways to implement that. We
2236 : * build a new upperrel representing the output of this phase.
2237 : */
2238 318816 : if (parse->distinctClause)
2239 : {
2240 430 : current_rel = create_distinct_paths(root,
2241 : current_rel);
2242 : }
2243 : } /* end of if (setOperations) */
2244 :
2245 : /*
2246 : * If ORDER BY was given, consider ways to implement that, and generate a
2247 : * new upperrel containing only paths that emit the correct ordering and
2248 : * project the correct final_target. We can apply the original
2249 : * limit_tuples limit in sort costing here, but only if there are no
2250 : * postponed SRFs.
2251 : */
2252 320170 : if (parse->sortClause)
2253 : {
2254 39162 : current_rel = create_ordered_paths(root,
2255 : current_rel,
2256 : final_target,
2257 : final_target_parallel_safe,
2258 : have_postponed_srfs ? -1.0 :
2259 : limit_tuples);
2260 : /* Fix things up if final_target contains SRFs */
2261 39162 : if (parse->hasTargetSRFs)
2262 124 : adjust_paths_for_srfs(root, current_rel,
2263 : final_targets,
2264 : final_targets_contain_srfs);
2265 : }
2266 :
2267 : /*
2268 : * Now we are prepared to build the final-output upperrel.
2269 : */
2270 320170 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2271 :
2272 : /*
2273 : * If the input rel is marked consider_parallel and there's nothing that's
2274 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2275 : * consider_parallel as well. Note that if the query has rowMarks or is
2276 : * not a SELECT, consider_parallel will be false for every relation in the
2277 : * query.
2278 : */
2279 398148 : if (current_rel->consider_parallel &&
2280 155948 : is_parallel_safe(root, parse->limitOffset) &&
2281 77970 : is_parallel_safe(root, parse->limitCount))
2282 77966 : final_rel->consider_parallel = true;
2283 :
2284 : /*
2285 : * If the current_rel belongs to a single FDW, so does the final_rel.
2286 : */
2287 320170 : final_rel->serverid = current_rel->serverid;
2288 320170 : final_rel->userid = current_rel->userid;
2289 320170 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2290 320170 : final_rel->fdwroutine = current_rel->fdwroutine;
2291 :
2292 : /*
2293 : * Generate paths for the final_rel. Insert all surviving paths, with
2294 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2295 : */
2296 646276 : foreach(lc, current_rel->pathlist)
2297 : {
2298 326106 : Path *path = (Path *) lfirst(lc);
2299 :
2300 : /*
2301 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2302 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2303 : * here. If there are only non-locking rowmarks, they should be
2304 : * handled by the ModifyTable node instead. However, root->rowMarks
2305 : * is what goes into the LockRows node.)
2306 : */
2307 326106 : if (parse->rowMarks)
2308 : {
2309 5380 : path = (Path *) create_lockrows_path(root, final_rel, path,
2310 : root->rowMarks,
2311 : assign_special_exec_param(root));
2312 : }
2313 :
2314 : /*
2315 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2316 : */
2317 326106 : if (limit_needed(parse))
2318 : {
2319 4048 : path = (Path *) create_limit_path(root, final_rel, path,
2320 : parse->limitOffset,
2321 : parse->limitCount,
2322 : parse->limitOption,
2323 : offset_est, count_est);
2324 : }
2325 :
2326 : /*
2327 : * If this is an INSERT/UPDATE/DELETE, and we're not being called from
2328 : * inheritance_planner, add the ModifyTable node.
2329 : */
2330 326106 : if (parse->commandType != CMD_SELECT && !inheritance_update)
2331 : {
2332 : Index rootRelation;
2333 : List *withCheckOptionLists;
2334 : List *returningLists;
2335 : List *rowMarks;
2336 :
2337 : /*
2338 : * If target is a partition root table, we need to mark the
2339 : * ModifyTable node appropriately for that.
2340 : */
2341 71772 : if (rt_fetch(parse->resultRelation, parse->rtable)->relkind ==
2342 : RELKIND_PARTITIONED_TABLE)
2343 1750 : rootRelation = parse->resultRelation;
2344 : else
2345 70022 : rootRelation = 0;
2346 :
2347 : /*
2348 : * Set up the WITH CHECK OPTION and RETURNING lists-of-lists, if
2349 : * needed.
2350 : */
2351 71772 : if (parse->withCheckOptions)
2352 508 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2353 : else
2354 71264 : withCheckOptionLists = NIL;
2355 :
2356 71772 : if (parse->returningList)
2357 1494 : returningLists = list_make1(parse->returningList);
2358 : else
2359 70278 : returningLists = NIL;
2360 :
2361 : /*
2362 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2363 : * will have dealt with fetching non-locked marked rows, else we
2364 : * need to have ModifyTable do that.
2365 : */
2366 71772 : if (parse->rowMarks)
2367 0 : rowMarks = NIL;
2368 : else
2369 71772 : rowMarks = root->rowMarks;
2370 :
2371 : path = (Path *)
2372 71772 : create_modifytable_path(root, final_rel,
2373 : parse->commandType,
2374 71772 : parse->canSetTag,
2375 71772 : parse->resultRelation,
2376 : rootRelation,
2377 : false,
2378 71772 : list_make1_int(parse->resultRelation),
2379 71772 : list_make1(path),
2380 71772 : list_make1(root),
2381 : withCheckOptionLists,
2382 : returningLists,
2383 : rowMarks,
2384 : parse->onConflict,
2385 : assign_special_exec_param(root));
2386 : }
2387 :
2388 : /* And shove it into final_rel */
2389 326106 : add_path(final_rel, path);
2390 : }
2391 :
2392 : /*
2393 : * Generate partial paths for final_rel, too, if outer query levels might
2394 : * be able to make use of them.
2395 : */
2396 320170 : if (final_rel->consider_parallel && root->query_level > 1 &&
2397 7616 : !limit_needed(parse))
2398 : {
2399 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2400 7570 : foreach(lc, current_rel->partial_pathlist)
2401 : {
2402 52 : Path *partial_path = (Path *) lfirst(lc);
2403 :
2404 52 : add_partial_path(final_rel, partial_path);
2405 : }
2406 : }
2407 :
2408 320170 : extra.limit_needed = limit_needed(parse);
2409 320170 : extra.limit_tuples = limit_tuples;
2410 320170 : extra.count_est = count_est;
2411 320170 : extra.offset_est = offset_est;
2412 :
2413 : /*
2414 : * If there is an FDW that's responsible for all baserels of the query,
2415 : * let it consider adding ForeignPaths.
2416 : */
2417 320170 : if (final_rel->fdwroutine &&
2418 1026 : final_rel->fdwroutine->GetForeignUpperPaths)
2419 968 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2420 : current_rel, final_rel,
2421 : &extra);
2422 :
2423 : /* Let extensions possibly add some more paths */
2424 320170 : if (create_upper_paths_hook)
2425 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2426 : current_rel, final_rel, &extra);
2427 :
2428 : /* Note: currently, we leave it to callers to do set_cheapest() */
2429 320170 : }
2430 :
2431 : /*
2432 : * Do preprocessing for groupingSets clause and related data. This handles the
2433 : * preliminary steps of expanding the grouping sets, organizing them into lists
2434 : * of rollups, and preparing annotations which will later be filled in with
2435 : * size estimates.
2436 : */
2437 : static grouping_sets_data *
2438 452 : preprocess_grouping_sets(PlannerInfo *root)
2439 : {
2440 452 : Query *parse = root->parse;
2441 : List *sets;
2442 452 : int maxref = 0;
2443 : ListCell *lc;
2444 : ListCell *lc_set;
2445 452 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2446 :
2447 452 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, -1);
2448 :
2449 452 : gd->any_hashable = false;
2450 452 : gd->unhashable_refs = NULL;
2451 452 : gd->unsortable_refs = NULL;
2452 452 : gd->unsortable_sets = NIL;
2453 :
2454 452 : if (parse->groupClause)
2455 : {
2456 : ListCell *lc;
2457 :
2458 1376 : foreach(lc, parse->groupClause)
2459 : {
2460 952 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2461 952 : Index ref = gc->tleSortGroupRef;
2462 :
2463 952 : if (ref > maxref)
2464 928 : maxref = ref;
2465 :
2466 952 : if (!gc->hashable)
2467 20 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2468 :
2469 952 : if (!OidIsValid(gc->sortop))
2470 28 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2471 : }
2472 : }
2473 :
2474 : /* Allocate workspace array for remapping */
2475 452 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2476 :
2477 : /*
2478 : * If we have any unsortable sets, we must extract them before trying to
2479 : * prepare rollups. Unsortable sets don't go through
2480 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2481 : * here.
2482 : */
2483 452 : if (!bms_is_empty(gd->unsortable_refs))
2484 : {
2485 28 : List *sortable_sets = NIL;
2486 :
2487 84 : foreach(lc, parse->groupingSets)
2488 : {
2489 60 : List *gset = (List *) lfirst(lc);
2490 :
2491 60 : if (bms_overlap_list(gd->unsortable_refs, gset))
2492 : {
2493 32 : GroupingSetData *gs = makeNode(GroupingSetData);
2494 :
2495 32 : gs->set = gset;
2496 32 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2497 :
2498 : /*
2499 : * We must enforce here that an unsortable set is hashable;
2500 : * later code assumes this. Parse analysis only checks that
2501 : * every individual column is either hashable or sortable.
2502 : *
2503 : * Note that passing this test doesn't guarantee we can
2504 : * generate a plan; there might be other showstoppers.
2505 : */
2506 32 : if (bms_overlap_list(gd->unhashable_refs, gset))
2507 4 : ereport(ERROR,
2508 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2509 : errmsg("could not implement GROUP BY"),
2510 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2511 : }
2512 : else
2513 28 : sortable_sets = lappend(sortable_sets, gset);
2514 : }
2515 :
2516 24 : if (sortable_sets)
2517 20 : sets = extract_rollup_sets(sortable_sets);
2518 : else
2519 4 : sets = NIL;
2520 : }
2521 : else
2522 424 : sets = extract_rollup_sets(parse->groupingSets);
2523 :
2524 1188 : foreach(lc_set, sets)
2525 : {
2526 740 : List *current_sets = (List *) lfirst(lc_set);
2527 740 : RollupData *rollup = makeNode(RollupData);
2528 : GroupingSetData *gs;
2529 :
2530 : /*
2531 : * Reorder the current list of grouping sets into correct prefix
2532 : * order. If only one aggregation pass is needed, try to make the
2533 : * list match the ORDER BY clause; if more than one pass is needed, we
2534 : * don't bother with that.
2535 : *
2536 : * Note that this reorders the sets from smallest-member-first to
2537 : * largest-member-first, and applies the GroupingSetData annotations,
2538 : * though the data will be filled in later.
2539 : */
2540 740 : current_sets = reorder_grouping_sets(current_sets,
2541 740 : (list_length(sets) == 1
2542 : ? parse->sortClause
2543 : : NIL));
2544 :
2545 : /*
2546 : * Get the initial (and therefore largest) grouping set.
2547 : */
2548 740 : gs = linitial_node(GroupingSetData, current_sets);
2549 :
2550 : /*
2551 : * Order the groupClause appropriately. If the first grouping set is
2552 : * empty, then the groupClause must also be empty; otherwise we have
2553 : * to force the groupClause to match that grouping set's order.
2554 : *
2555 : * (The first grouping set can be empty even though parse->groupClause
2556 : * is not empty only if all non-empty grouping sets are unsortable.
2557 : * The groupClauses for hashed grouping sets are built later on.)
2558 : */
2559 740 : if (gs->set)
2560 712 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2561 : else
2562 28 : rollup->groupClause = NIL;
2563 :
2564 : /*
2565 : * Is it hashable? We pretend empty sets are hashable even though we
2566 : * actually force them not to be hashed later. But don't bother if
2567 : * there's nothing but empty sets (since in that case we can't hash
2568 : * anything).
2569 : */
2570 740 : if (gs->set &&
2571 712 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2572 : {
2573 696 : rollup->hashable = true;
2574 696 : gd->any_hashable = true;
2575 : }
2576 :
2577 : /*
2578 : * Now that we've pinned down an order for the groupClause for this
2579 : * list of grouping sets, we need to remap the entries in the grouping
2580 : * sets from sortgrouprefs to plain indices (0-based) into the
2581 : * groupClause for this collection of grouping sets. We keep the
2582 : * original form for later use, though.
2583 : */
2584 740 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2585 : current_sets,
2586 : gd->tleref_to_colnum_map);
2587 740 : rollup->gsets_data = current_sets;
2588 :
2589 740 : gd->rollups = lappend(gd->rollups, rollup);
2590 : }
2591 :
2592 448 : if (gd->unsortable_sets)
2593 : {
2594 : /*
2595 : * We have not yet pinned down a groupclause for this, but we will
2596 : * need index-based lists for estimation purposes. Construct
2597 : * hash_sets_idx based on the entire original groupclause for now.
2598 : */
2599 24 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2600 : gd->unsortable_sets,
2601 : gd->tleref_to_colnum_map);
2602 24 : gd->any_hashable = true;
2603 : }
2604 :
2605 448 : return gd;
2606 : }
2607 :
2608 : /*
2609 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2610 : * (without annotation) mapped to indexes into the given groupclause.
2611 : */
2612 : static List *
2613 2060 : remap_to_groupclause_idx(List *groupClause,
2614 : List *gsets,
2615 : int *tleref_to_colnum_map)
2616 : {
2617 2060 : int ref = 0;
2618 2060 : List *result = NIL;
2619 : ListCell *lc;
2620 :
2621 5044 : foreach(lc, groupClause)
2622 : {
2623 2984 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2624 :
2625 2984 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2626 : }
2627 :
2628 4796 : foreach(lc, gsets)
2629 : {
2630 2736 : List *set = NIL;
2631 : ListCell *lc2;
2632 2736 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2633 :
2634 6096 : foreach(lc2, gs->set)
2635 : {
2636 3360 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2637 : }
2638 :
2639 2736 : result = lappend(result, set);
2640 : }
2641 :
2642 2060 : return result;
2643 : }
2644 :
2645 :
2646 : /*
2647 : * preprocess_rowmarks - set up PlanRowMarks if needed
2648 : */
2649 : static void
2650 319100 : preprocess_rowmarks(PlannerInfo *root)
2651 : {
2652 319100 : Query *parse = root->parse;
2653 : Bitmapset *rels;
2654 : List *prowmarks;
2655 : ListCell *l;
2656 : int i;
2657 :
2658 319100 : if (parse->rowMarks)
2659 : {
2660 : /*
2661 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2662 : * grouping, since grouping renders a reference to individual tuple
2663 : * CTIDs invalid. This is also checked at parse time, but that's
2664 : * insufficient because of rule substitution, query pullup, etc.
2665 : */
2666 4932 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2667 : parse->rowMarks)->strength);
2668 : }
2669 : else
2670 : {
2671 : /*
2672 : * We only need rowmarks for UPDATE, DELETE, or FOR [KEY]
2673 : * UPDATE/SHARE.
2674 : */
2675 314168 : if (parse->commandType != CMD_UPDATE &&
2676 304188 : parse->commandType != CMD_DELETE)
2677 301142 : return;
2678 : }
2679 :
2680 : /*
2681 : * We need to have rowmarks for all base relations except the target. We
2682 : * make a bitmapset of all base rels and then remove the items we don't
2683 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2684 : */
2685 17958 : rels = get_relids_in_jointree((Node *) parse->jointree, false);
2686 17958 : if (parse->resultRelation)
2687 13026 : rels = bms_del_member(rels, parse->resultRelation);
2688 :
2689 : /*
2690 : * Convert RowMarkClauses to PlanRowMark representation.
2691 : */
2692 17958 : prowmarks = NIL;
2693 23112 : foreach(l, parse->rowMarks)
2694 : {
2695 5154 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2696 5154 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2697 : PlanRowMark *newrc;
2698 :
2699 : /*
2700 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2701 : * applied to an update/delete target rel. If that ever becomes
2702 : * possible, we should drop the target from the PlanRowMark list.
2703 : */
2704 : Assert(rc->rti != parse->resultRelation);
2705 :
2706 : /*
2707 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2708 : * can't support true locking. Subqueries that got flattened into the
2709 : * main query should be ignored completely. Any that didn't will get
2710 : * ROW_MARK_COPY items in the next loop.
2711 : */
2712 5154 : if (rte->rtekind != RTE_RELATION)
2713 76 : continue;
2714 :
2715 5078 : rels = bms_del_member(rels, rc->rti);
2716 :
2717 5078 : newrc = makeNode(PlanRowMark);
2718 5078 : newrc->rti = newrc->prti = rc->rti;
2719 5078 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2720 5078 : newrc->markType = select_rowmark_type(rte, rc->strength);
2721 5078 : newrc->allMarkTypes = (1 << newrc->markType);
2722 5078 : newrc->strength = rc->strength;
2723 5078 : newrc->waitPolicy = rc->waitPolicy;
2724 5078 : newrc->isParent = false;
2725 :
2726 5078 : prowmarks = lappend(prowmarks, newrc);
2727 : }
2728 :
2729 : /*
2730 : * Now, add rowmarks for any non-target, non-locked base relations.
2731 : */
2732 17958 : i = 0;
2733 39592 : foreach(l, parse->rtable)
2734 : {
2735 21634 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2736 : PlanRowMark *newrc;
2737 :
2738 21634 : i++;
2739 21634 : if (!bms_is_member(i, rels))
2740 20630 : continue;
2741 :
2742 1004 : newrc = makeNode(PlanRowMark);
2743 1004 : newrc->rti = newrc->prti = i;
2744 1004 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2745 1004 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2746 1004 : newrc->allMarkTypes = (1 << newrc->markType);
2747 1004 : newrc->strength = LCS_NONE;
2748 1004 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2749 1004 : newrc->isParent = false;
2750 :
2751 1004 : prowmarks = lappend(prowmarks, newrc);
2752 : }
2753 :
2754 17958 : root->rowMarks = prowmarks;
2755 : }
2756 :
2757 : /*
2758 : * Select RowMarkType to use for a given table
2759 : */
2760 : RowMarkType
2761 7324 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2762 : {
2763 7324 : if (rte->rtekind != RTE_RELATION)
2764 : {
2765 : /* If it's not a table at all, use ROW_MARK_COPY */
2766 174 : return ROW_MARK_COPY;
2767 : }
2768 7150 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2769 : {
2770 : /* Let the FDW select the rowmark type, if it wants to */
2771 160 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2772 :
2773 160 : if (fdwroutine->GetForeignRowMarkType != NULL)
2774 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2775 : /* Otherwise, use ROW_MARK_COPY by default */
2776 160 : return ROW_MARK_COPY;
2777 : }
2778 : else
2779 : {
2780 : /* Regular table, apply the appropriate lock type */
2781 6990 : switch (strength)
2782 : {
2783 1146 : case LCS_NONE:
2784 :
2785 : /*
2786 : * We don't need a tuple lock, only the ability to re-fetch
2787 : * the row.
2788 : */
2789 1146 : return ROW_MARK_REFERENCE;
2790 : break;
2791 4256 : case LCS_FORKEYSHARE:
2792 4256 : return ROW_MARK_KEYSHARE;
2793 : break;
2794 234 : case LCS_FORSHARE:
2795 234 : return ROW_MARK_SHARE;
2796 : break;
2797 58 : case LCS_FORNOKEYUPDATE:
2798 58 : return ROW_MARK_NOKEYEXCLUSIVE;
2799 : break;
2800 1296 : case LCS_FORUPDATE:
2801 1296 : return ROW_MARK_EXCLUSIVE;
2802 : break;
2803 : }
2804 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2805 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2806 : }
2807 : }
2808 :
2809 : /*
2810 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2811 : *
2812 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2813 : * results back in *count_est and *offset_est. These variables are set to
2814 : * 0 if the corresponding clause is not present, and -1 if it's present
2815 : * but we couldn't estimate the value for it. (The "0" convention is OK
2816 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2817 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2818 : * usual practice of never estimating less than one row.) These values will
2819 : * be passed to create_limit_path, which see if you change this code.
2820 : *
2821 : * The return value is the suitably adjusted tuple_fraction to use for
2822 : * planning the query. This adjustment is not overridable, since it reflects
2823 : * plan actions that grouping_planner() will certainly take, not assumptions
2824 : * about context.
2825 : */
2826 : static double
2827 3364 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2828 : int64 *offset_est, int64 *count_est)
2829 : {
2830 3364 : Query *parse = root->parse;
2831 : Node *est;
2832 : double limit_fraction;
2833 :
2834 : /* Should not be called unless LIMIT or OFFSET */
2835 : Assert(parse->limitCount || parse->limitOffset);
2836 :
2837 : /*
2838 : * Try to obtain the clause values. We use estimate_expression_value
2839 : * primarily because it can sometimes do something useful with Params.
2840 : */
2841 3364 : if (parse->limitCount)
2842 : {
2843 3146 : est = estimate_expression_value(root, parse->limitCount);
2844 3146 : if (est && IsA(est, Const))
2845 : {
2846 6284 : if (((Const *) est)->constisnull)
2847 : {
2848 : /* NULL indicates LIMIT ALL, ie, no limit */
2849 0 : *count_est = 0; /* treat as not present */
2850 : }
2851 : else
2852 : {
2853 3142 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2854 3142 : if (*count_est <= 0)
2855 100 : *count_est = 1; /* force to at least 1 */
2856 : }
2857 : }
2858 : else
2859 4 : *count_est = -1; /* can't estimate */
2860 : }
2861 : else
2862 218 : *count_est = 0; /* not present */
2863 :
2864 3364 : if (parse->limitOffset)
2865 : {
2866 536 : est = estimate_expression_value(root, parse->limitOffset);
2867 536 : if (est && IsA(est, Const))
2868 : {
2869 1040 : if (((Const *) est)->constisnull)
2870 : {
2871 : /* Treat NULL as no offset; the executor will too */
2872 0 : *offset_est = 0; /* treat as not present */
2873 : }
2874 : else
2875 : {
2876 520 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2877 520 : if (*offset_est < 0)
2878 0 : *offset_est = 0; /* treat as not present */
2879 : }
2880 : }
2881 : else
2882 16 : *offset_est = -1; /* can't estimate */
2883 : }
2884 : else
2885 2828 : *offset_est = 0; /* not present */
2886 :
2887 3364 : if (*count_est != 0)
2888 : {
2889 : /*
2890 : * A LIMIT clause limits the absolute number of tuples returned.
2891 : * However, if it's not a constant LIMIT then we have to guess; for
2892 : * lack of a better idea, assume 10% of the plan's result is wanted.
2893 : */
2894 3146 : if (*count_est < 0 || *offset_est < 0)
2895 : {
2896 : /* LIMIT or OFFSET is an expression ... punt ... */
2897 16 : limit_fraction = 0.10;
2898 : }
2899 : else
2900 : {
2901 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2902 3130 : limit_fraction = (double) *count_est + (double) *offset_est;
2903 : }
2904 :
2905 : /*
2906 : * If we have absolute limits from both caller and LIMIT, use the
2907 : * smaller value; likewise if they are both fractional. If one is
2908 : * fractional and the other absolute, we can't easily determine which
2909 : * is smaller, but we use the heuristic that the absolute will usually
2910 : * be smaller.
2911 : */
2912 3146 : if (tuple_fraction >= 1.0)
2913 : {
2914 4 : if (limit_fraction >= 1.0)
2915 : {
2916 : /* both absolute */
2917 4 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2918 : }
2919 : else
2920 : {
2921 : /* caller absolute, limit fractional; use caller's value */
2922 : }
2923 : }
2924 3142 : else if (tuple_fraction > 0.0)
2925 : {
2926 106 : if (limit_fraction >= 1.0)
2927 : {
2928 : /* caller fractional, limit absolute; use limit */
2929 106 : tuple_fraction = limit_fraction;
2930 : }
2931 : else
2932 : {
2933 : /* both fractional */
2934 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2935 : }
2936 : }
2937 : else
2938 : {
2939 : /* no info from caller, just use limit */
2940 3036 : tuple_fraction = limit_fraction;
2941 : }
2942 : }
2943 218 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2944 : {
2945 : /*
2946 : * We have an OFFSET but no LIMIT. This acts entirely differently
2947 : * from the LIMIT case: here, we need to increase rather than decrease
2948 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2949 : * tuples to be fetched instead of fewer. This only matters if we got
2950 : * a tuple_fraction > 0, however.
2951 : *
2952 : * As above, use 10% if OFFSET is present but unestimatable.
2953 : */
2954 8 : if (*offset_est < 0)
2955 0 : limit_fraction = 0.10;
2956 : else
2957 8 : limit_fraction = (double) *offset_est;
2958 :
2959 : /*
2960 : * If we have absolute counts from both caller and OFFSET, add them
2961 : * together; likewise if they are both fractional. If one is
2962 : * fractional and the other absolute, we want to take the larger, and
2963 : * we heuristically assume that's the fractional one.
2964 : */
2965 8 : if (tuple_fraction >= 1.0)
2966 : {
2967 0 : if (limit_fraction >= 1.0)
2968 : {
2969 : /* both absolute, so add them together */
2970 0 : tuple_fraction += limit_fraction;
2971 : }
2972 : else
2973 : {
2974 : /* caller absolute, limit fractional; use limit */
2975 0 : tuple_fraction = limit_fraction;
2976 : }
2977 : }
2978 : else
2979 : {
2980 8 : if (limit_fraction >= 1.0)
2981 : {
2982 : /* caller fractional, limit absolute; use caller's value */
2983 : }
2984 : else
2985 : {
2986 : /* both fractional, so add them together */
2987 0 : tuple_fraction += limit_fraction;
2988 0 : if (tuple_fraction >= 1.0)
2989 0 : tuple_fraction = 0.0; /* assume fetch all */
2990 : }
2991 : }
2992 : }
2993 :
2994 3364 : return tuple_fraction;
2995 : }
2996 :
2997 : /*
2998 : * limit_needed - do we actually need a Limit plan node?
2999 : *
3000 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
3001 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
3002 : * locution for an optimization fence. (Because other places in the planner
3003 : * merely check whether parse->limitOffset isn't NULL, it will still work as
3004 : * an optimization fence --- we're just suppressing unnecessary run-time
3005 : * overhead.)
3006 : *
3007 : * This might look like it could be merged into preprocess_limit, but there's
3008 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3009 : * in preprocess_limit it's good enough to consider estimated values.
3010 : */
3011 : bool
3012 660104 : limit_needed(Query *parse)
3013 : {
3014 : Node *node;
3015 :
3016 660104 : node = parse->limitCount;
3017 660104 : if (node)
3018 : {
3019 7364 : if (IsA(node, Const))
3020 : {
3021 : /* NULL indicates LIMIT ALL, ie, no limit */
3022 7244 : if (!((Const *) node)->constisnull)
3023 7244 : return true; /* LIMIT with a constant value */
3024 : }
3025 : else
3026 120 : return true; /* non-constant LIMIT */
3027 : }
3028 :
3029 652740 : node = parse->limitOffset;
3030 652740 : if (node)
3031 : {
3032 648 : if (IsA(node, Const))
3033 : {
3034 : /* Treat NULL as no offset; the executor would too */
3035 502 : if (!((Const *) node)->constisnull)
3036 : {
3037 502 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3038 :
3039 502 : if (offset != 0)
3040 56 : return true; /* OFFSET with a nonzero value */
3041 : }
3042 : }
3043 : else
3044 146 : return true; /* non-constant OFFSET */
3045 : }
3046 :
3047 652538 : return false; /* don't need a Limit plan node */
3048 : }
3049 :
3050 :
3051 : /*
3052 : * remove_useless_groupby_columns
3053 : * Remove any columns in the GROUP BY clause that are redundant due to
3054 : * being functionally dependent on other GROUP BY columns.
3055 : *
3056 : * Since some other DBMSes do not allow references to ungrouped columns, it's
3057 : * not unusual to find all columns listed in GROUP BY even though listing the
3058 : * primary-key columns would be sufficient. Deleting such excess columns
3059 : * avoids redundant sorting work, so it's worth doing.
3060 : *
3061 : * Relcache invalidations will ensure that cached plans become invalidated
3062 : * when the underlying index of the pkey constraint is dropped.
3063 : *
3064 : * Currently, we only make use of pkey constraints for this, however, we may
3065 : * wish to take this further in the future and also use unique constraints
3066 : * which have NOT NULL columns. In that case, plan invalidation will still
3067 : * work since relations will receive a relcache invalidation when a NOT NULL
3068 : * constraint is dropped.
3069 : */
3070 : static void
3071 317624 : remove_useless_groupby_columns(PlannerInfo *root)
3072 : {
3073 317624 : Query *parse = root->parse;
3074 : Bitmapset **groupbyattnos;
3075 : Bitmapset **surplusvars;
3076 : ListCell *lc;
3077 : int relid;
3078 :
3079 : /* No chance to do anything if there are less than two GROUP BY items */
3080 317624 : if (list_length(parse->groupClause) < 2)
3081 316808 : return;
3082 :
3083 : /* Don't fiddle with the GROUP BY clause if the query has grouping sets */
3084 816 : if (parse->groupingSets)
3085 368 : return;
3086 :
3087 : /*
3088 : * Scan the GROUP BY clause to find GROUP BY items that are simple Vars.
3089 : * Fill groupbyattnos[k] with a bitmapset of the column attnos of RTE k
3090 : * that are GROUP BY items.
3091 : */
3092 448 : groupbyattnos = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3093 448 : (list_length(parse->rtable) + 1));
3094 1600 : foreach(lc, parse->groupClause)
3095 : {
3096 1152 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3097 1152 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3098 1152 : Var *var = (Var *) tle->expr;
3099 :
3100 : /*
3101 : * Ignore non-Vars and Vars from other query levels.
3102 : *
3103 : * XXX in principle, stable expressions containing Vars could also be
3104 : * removed, if all the Vars are functionally dependent on other GROUP
3105 : * BY items. But it's not clear that such cases occur often enough to
3106 : * be worth troubling over.
3107 : */
3108 1152 : if (!IsA(var, Var) ||
3109 996 : var->varlevelsup > 0)
3110 156 : continue;
3111 :
3112 : /* OK, remember we have this Var */
3113 996 : relid = var->varno;
3114 : Assert(relid <= list_length(parse->rtable));
3115 996 : groupbyattnos[relid] = bms_add_member(groupbyattnos[relid],
3116 996 : var->varattno - FirstLowInvalidHeapAttributeNumber);
3117 : }
3118 :
3119 : /*
3120 : * Consider each relation and see if it is possible to remove some of its
3121 : * Vars from GROUP BY. For simplicity and speed, we do the actual removal
3122 : * in a separate pass. Here, we just fill surplusvars[k] with a bitmapset
3123 : * of the column attnos of RTE k that are removable GROUP BY items.
3124 : */
3125 448 : surplusvars = NULL; /* don't allocate array unless required */
3126 448 : relid = 0;
3127 1460 : foreach(lc, parse->rtable)
3128 : {
3129 1012 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
3130 : Bitmapset *relattnos;
3131 : Bitmapset *pkattnos;
3132 : Oid constraintOid;
3133 :
3134 1012 : relid++;
3135 :
3136 : /* Only plain relations could have primary-key constraints */
3137 1012 : if (rte->rtekind != RTE_RELATION)
3138 966 : continue;
3139 :
3140 : /*
3141 : * We must skip inheritance parent tables as some of the child rels
3142 : * may cause duplicate rows. This cannot happen with partitioned
3143 : * tables, however.
3144 : */
3145 708 : if (rte->inh && rte->relkind != RELKIND_PARTITIONED_TABLE)
3146 4 : continue;
3147 :
3148 : /* Nothing to do unless this rel has multiple Vars in GROUP BY */
3149 704 : relattnos = groupbyattnos[relid];
3150 704 : if (bms_membership(relattnos) != BMS_MULTIPLE)
3151 408 : continue;
3152 :
3153 : /*
3154 : * Can't remove any columns for this rel if there is no suitable
3155 : * (i.e., nondeferrable) primary key constraint.
3156 : */
3157 296 : pkattnos = get_primary_key_attnos(rte->relid, false, &constraintOid);
3158 296 : if (pkattnos == NULL)
3159 250 : continue;
3160 :
3161 : /*
3162 : * If the primary key is a proper subset of relattnos then we have
3163 : * some items in the GROUP BY that can be removed.
3164 : */
3165 46 : if (bms_subset_compare(pkattnos, relattnos) == BMS_SUBSET1)
3166 : {
3167 : /*
3168 : * To easily remember whether we've found anything to do, we don't
3169 : * allocate the surplusvars[] array until we find something.
3170 : */
3171 34 : if (surplusvars == NULL)
3172 30 : surplusvars = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3173 30 : (list_length(parse->rtable) + 1));
3174 :
3175 : /* Remember the attnos of the removable columns */
3176 34 : surplusvars[relid] = bms_difference(relattnos, pkattnos);
3177 : }
3178 : }
3179 :
3180 : /*
3181 : * If we found any surplus Vars, build a new GROUP BY clause without them.
3182 : * (Note: this may leave some TLEs with unreferenced ressortgroupref
3183 : * markings, but that's harmless.)
3184 : */
3185 448 : if (surplusvars != NULL)
3186 : {
3187 30 : List *new_groupby = NIL;
3188 :
3189 150 : foreach(lc, parse->groupClause)
3190 : {
3191 120 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3192 120 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3193 120 : Var *var = (Var *) tle->expr;
3194 :
3195 : /*
3196 : * New list must include non-Vars, outer Vars, and anything not
3197 : * marked as surplus.
3198 : */
3199 120 : if (!IsA(var, Var) ||
3200 120 : var->varlevelsup > 0 ||
3201 120 : !bms_is_member(var->varattno - FirstLowInvalidHeapAttributeNumber,
3202 120 : surplusvars[var->varno]))
3203 66 : new_groupby = lappend(new_groupby, sgc);
3204 : }
3205 :
3206 30 : parse->groupClause = new_groupby;
3207 : }
3208 : }
3209 :
3210 : /*
3211 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3212 : *
3213 : * The idea here is to adjust the ordering of the GROUP BY elements
3214 : * (which in itself is semantically insignificant) to match ORDER BY,
3215 : * thereby allowing a single sort operation to both implement the ORDER BY
3216 : * requirement and set up for a Unique step that implements GROUP BY.
3217 : *
3218 : * In principle it might be interesting to consider other orderings of the
3219 : * GROUP BY elements, which could match the sort ordering of other
3220 : * possible plans (eg an indexscan) and thereby reduce cost. We don't
3221 : * bother with that, though. Hashed grouping will frequently win anyway.
3222 : *
3223 : * Note: we need no comparable processing of the distinctClause because
3224 : * the parser already enforced that that matches ORDER BY.
3225 : *
3226 : * For grouping sets, the order of items is instead forced to agree with that
3227 : * of the grouping set (and items not in the grouping set are skipped). The
3228 : * work of sorting the order of grouping set elements to match the ORDER BY if
3229 : * possible is done elsewhere.
3230 : */
3231 : static List *
3232 3782 : preprocess_groupclause(PlannerInfo *root, List *force)
3233 : {
3234 3782 : Query *parse = root->parse;
3235 3782 : List *new_groupclause = NIL;
3236 : bool partial_match;
3237 : ListCell *sl;
3238 : ListCell *gl;
3239 :
3240 : /* For grouping sets, we need to force the ordering */
3241 3782 : if (force)
3242 : {
3243 4940 : foreach(sl, force)
3244 : {
3245 2932 : Index ref = lfirst_int(sl);
3246 2932 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3247 :
3248 2932 : new_groupclause = lappend(new_groupclause, cl);
3249 : }
3250 :
3251 2008 : return new_groupclause;
3252 : }
3253 :
3254 : /* If no ORDER BY, nothing useful to do here */
3255 1774 : if (parse->sortClause == NIL)
3256 818 : return parse->groupClause;
3257 :
3258 : /*
3259 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3260 : * items, but only as far as we can make a matching prefix.
3261 : *
3262 : * This code assumes that the sortClause contains no duplicate items.
3263 : */
3264 1884 : foreach(sl, parse->sortClause)
3265 : {
3266 1310 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3267 :
3268 1956 : foreach(gl, parse->groupClause)
3269 : {
3270 1574 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3271 :
3272 1574 : if (equal(gc, sc))
3273 : {
3274 928 : new_groupclause = lappend(new_groupclause, gc);
3275 928 : break;
3276 : }
3277 : }
3278 1310 : if (gl == NULL)
3279 382 : break; /* no match, so stop scanning */
3280 : }
3281 :
3282 : /* Did we match all of the ORDER BY list, or just some of it? */
3283 956 : partial_match = (sl != NULL);
3284 :
3285 : /* If no match at all, no point in reordering GROUP BY */
3286 956 : if (new_groupclause == NIL)
3287 148 : return parse->groupClause;
3288 :
3289 : /*
3290 : * Add any remaining GROUP BY items to the new list, but only if we were
3291 : * able to make a complete match. In other words, we only rearrange the
3292 : * GROUP BY list if the result is that one list is a prefix of the other
3293 : * --- otherwise there's no possibility of a common sort. Also, give up
3294 : * if there are any non-sortable GROUP BY items, since then there's no
3295 : * hope anyway.
3296 : */
3297 1768 : foreach(gl, parse->groupClause)
3298 : {
3299 1010 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3300 :
3301 1010 : if (list_member_ptr(new_groupclause, gc))
3302 912 : continue; /* it matched an ORDER BY item */
3303 98 : if (partial_match)
3304 50 : return parse->groupClause; /* give up, no common sort possible */
3305 48 : if (!OidIsValid(gc->sortop))
3306 0 : return parse->groupClause; /* give up, GROUP BY can't be sorted */
3307 48 : new_groupclause = lappend(new_groupclause, gc);
3308 : }
3309 :
3310 : /* Success --- install the rearranged GROUP BY list */
3311 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3312 758 : return new_groupclause;
3313 : }
3314 :
3315 : /*
3316 : * Extract lists of grouping sets that can be implemented using a single
3317 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3318 : *
3319 : * Input must be sorted with smallest sets first. Result has each sublist
3320 : * sorted with smallest sets first.
3321 : *
3322 : * We want to produce the absolute minimum possible number of lists here to
3323 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3324 : * of finding the minimal partition of a partially-ordered set into chains
3325 : * (which is what we need, taking the list of grouping sets as a poset ordered
3326 : * by set inclusion) can be mapped to the problem of finding the maximum
3327 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3328 : * time with a worst case of no worse than O(n^2.5) and usually much
3329 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3330 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3331 : * half a second on my modest system even with optimization off and assertions
3332 : * on.)
3333 : */
3334 : static List *
3335 444 : extract_rollup_sets(List *groupingSets)
3336 : {
3337 444 : int num_sets_raw = list_length(groupingSets);
3338 444 : int num_empty = 0;
3339 444 : int num_sets = 0; /* distinct sets */
3340 444 : int num_chains = 0;
3341 444 : List *result = NIL;
3342 : List **results;
3343 : List **orig_sets;
3344 : Bitmapset **set_masks;
3345 : int *chains;
3346 : short **adjacency;
3347 : short *adjacency_buf;
3348 : BipartiteMatchState *state;
3349 : int i;
3350 : int j;
3351 : int j_size;
3352 444 : ListCell *lc1 = list_head(groupingSets);
3353 : ListCell *lc;
3354 :
3355 : /*
3356 : * Start by stripping out empty sets. The algorithm doesn't require this,
3357 : * but the planner currently needs all empty sets to be returned in the
3358 : * first list, so we strip them here and add them back after.
3359 : */
3360 796 : while (lc1 && lfirst(lc1) == NIL)
3361 : {
3362 352 : ++num_empty;
3363 352 : lc1 = lnext(groupingSets, lc1);
3364 : }
3365 :
3366 : /* bail out now if it turns out that all we had were empty sets. */
3367 444 : if (!lc1)
3368 28 : return list_make1(groupingSets);
3369 :
3370 : /*----------
3371 : * We don't strictly need to remove duplicate sets here, but if we don't,
3372 : * they tend to become scattered through the result, which is a bit
3373 : * confusing (and irritating if we ever decide to optimize them out).
3374 : * So we remove them here and add them back after.
3375 : *
3376 : * For each non-duplicate set, we fill in the following:
3377 : *
3378 : * orig_sets[i] = list of the original set lists
3379 : * set_masks[i] = bitmapset for testing inclusion
3380 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3381 : *
3382 : * chains[i] will be the result group this set is assigned to.
3383 : *
3384 : * We index all of these from 1 rather than 0 because it is convenient
3385 : * to leave 0 free for the NIL node in the graph algorithm.
3386 : *----------
3387 : */
3388 416 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3389 416 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3390 416 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3391 416 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3392 :
3393 416 : j_size = 0;
3394 416 : j = 0;
3395 416 : i = 1;
3396 :
3397 1476 : for_each_cell(lc, groupingSets, lc1)
3398 : {
3399 1060 : List *candidate = (List *) lfirst(lc);
3400 1060 : Bitmapset *candidate_set = NULL;
3401 : ListCell *lc2;
3402 1060 : int dup_of = 0;
3403 :
3404 2568 : foreach(lc2, candidate)
3405 : {
3406 1508 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3407 : }
3408 :
3409 : /* we can only be a dup if we're the same length as a previous set */
3410 1060 : if (j_size == list_length(candidate))
3411 : {
3412 : int k;
3413 :
3414 984 : for (k = j; k < i; ++k)
3415 : {
3416 624 : if (bms_equal(set_masks[k], candidate_set))
3417 : {
3418 52 : dup_of = k;
3419 52 : break;
3420 : }
3421 : }
3422 : }
3423 648 : else if (j_size < list_length(candidate))
3424 : {
3425 648 : j_size = list_length(candidate);
3426 648 : j = i;
3427 : }
3428 :
3429 1060 : if (dup_of > 0)
3430 : {
3431 52 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3432 52 : bms_free(candidate_set);
3433 : }
3434 : else
3435 : {
3436 : int k;
3437 1008 : int n_adj = 0;
3438 :
3439 1008 : orig_sets[i] = list_make1(candidate);
3440 1008 : set_masks[i] = candidate_set;
3441 :
3442 : /* fill in adjacency list; no need to compare equal-size sets */
3443 :
3444 1712 : for (k = j - 1; k > 0; --k)
3445 : {
3446 704 : if (bms_is_subset(set_masks[k], candidate_set))
3447 596 : adjacency_buf[++n_adj] = k;
3448 : }
3449 :
3450 1008 : if (n_adj > 0)
3451 : {
3452 296 : adjacency_buf[0] = n_adj;
3453 296 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3454 296 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3455 : }
3456 : else
3457 712 : adjacency[i] = NULL;
3458 :
3459 1008 : ++i;
3460 : }
3461 : }
3462 :
3463 416 : num_sets = i - 1;
3464 :
3465 : /*
3466 : * Apply the graph matching algorithm to do the work.
3467 : */
3468 416 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3469 :
3470 : /*
3471 : * Now, the state->pair* fields have the info we need to assign sets to
3472 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3473 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3474 : * it in one pass)
3475 : */
3476 416 : chains = palloc0((num_sets + 1) * sizeof(int));
3477 :
3478 1424 : for (i = 1; i <= num_sets; ++i)
3479 : {
3480 1008 : int u = state->pair_vu[i];
3481 1008 : int v = state->pair_uv[i];
3482 :
3483 1008 : if (u > 0 && u < i)
3484 0 : chains[i] = chains[u];
3485 1008 : else if (v > 0 && v < i)
3486 296 : chains[i] = chains[v];
3487 : else
3488 712 : chains[i] = ++num_chains;
3489 : }
3490 :
3491 : /* build result lists. */
3492 416 : results = palloc0((num_chains + 1) * sizeof(List *));
3493 :
3494 1424 : for (i = 1; i <= num_sets; ++i)
3495 : {
3496 1008 : int c = chains[i];
3497 :
3498 : Assert(c > 0);
3499 :
3500 1008 : results[c] = list_concat(results[c], orig_sets[i]);
3501 : }
3502 :
3503 : /* push any empty sets back on the first list. */
3504 708 : while (num_empty-- > 0)
3505 292 : results[1] = lcons(NIL, results[1]);
3506 :
3507 : /* make result list */
3508 1128 : for (i = 1; i <= num_chains; ++i)
3509 712 : result = lappend(result, results[i]);
3510 :
3511 : /*
3512 : * Free all the things.
3513 : *
3514 : * (This is over-fussy for small sets but for large sets we could have
3515 : * tied up a nontrivial amount of memory.)
3516 : */
3517 416 : BipartiteMatchFree(state);
3518 416 : pfree(results);
3519 416 : pfree(chains);
3520 1424 : for (i = 1; i <= num_sets; ++i)
3521 1008 : if (adjacency[i])
3522 296 : pfree(adjacency[i]);
3523 416 : pfree(adjacency);
3524 416 : pfree(adjacency_buf);
3525 416 : pfree(orig_sets);
3526 1424 : for (i = 1; i <= num_sets; ++i)
3527 1008 : bms_free(set_masks[i]);
3528 416 : pfree(set_masks);
3529 :
3530 416 : return result;
3531 : }
3532 :
3533 : /*
3534 : * Reorder the elements of a list of grouping sets such that they have correct
3535 : * prefix relationships. Also inserts the GroupingSetData annotations.
3536 : *
3537 : * The input must be ordered with smallest sets first; the result is returned
3538 : * with largest sets first. Note that the result shares no list substructure
3539 : * with the input, so it's safe for the caller to modify it later.
3540 : *
3541 : * If we're passed in a sortclause, we follow its order of columns to the
3542 : * extent possible, to minimize the chance that we add unnecessary sorts.
3543 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3544 : * gets implemented in one pass.)
3545 : */
3546 : static List *
3547 740 : reorder_grouping_sets(List *groupingsets, List *sortclause)
3548 : {
3549 : ListCell *lc;
3550 740 : List *previous = NIL;
3551 740 : List *result = NIL;
3552 :
3553 2152 : foreach(lc, groupingsets)
3554 : {
3555 1412 : List *candidate = (List *) lfirst(lc);
3556 1412 : List *new_elems = list_difference_int(candidate, previous);
3557 1412 : GroupingSetData *gs = makeNode(GroupingSetData);
3558 :
3559 1704 : while (list_length(sortclause) > list_length(previous) &&
3560 236 : list_length(new_elems) > 0)
3561 : {
3562 128 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3563 128 : int ref = sc->tleSortGroupRef;
3564 :
3565 128 : if (list_member_int(new_elems, ref))
3566 : {
3567 56 : previous = lappend_int(previous, ref);
3568 56 : new_elems = list_delete_int(new_elems, ref);
3569 : }
3570 : else
3571 : {
3572 : /* diverged from the sortclause; give up on it */
3573 72 : sortclause = NIL;
3574 72 : break;
3575 : }
3576 : }
3577 :
3578 1412 : previous = list_concat(previous, new_elems);
3579 :
3580 1412 : gs->set = list_copy(previous);
3581 1412 : result = lcons(gs, result);
3582 : }
3583 :
3584 740 : list_free(previous);
3585 :
3586 740 : return result;
3587 : }
3588 :
3589 : /*
3590 : * Compute query_pathkeys and other pathkeys during plan generation
3591 : */
3592 : static void
3593 318822 : standard_qp_callback(PlannerInfo *root, void *extra)
3594 : {
3595 318822 : Query *parse = root->parse;
3596 318822 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3597 318822 : List *tlist = root->processed_tlist;
3598 318822 : List *activeWindows = qp_extra->activeWindows;
3599 :
3600 : /*
3601 : * Calculate pathkeys that represent grouping/ordering requirements. The
3602 : * sortClause is certainly sort-able, but GROUP BY and DISTINCT might not
3603 : * be, in which case we just leave their pathkeys empty.
3604 : */
3605 321012 : if (qp_extra->groupClause &&
3606 2190 : grouping_is_sortable(qp_extra->groupClause))
3607 2190 : root->group_pathkeys =
3608 2190 : make_pathkeys_for_sortclauses(root,
3609 : qp_extra->groupClause,
3610 : tlist);
3611 : else
3612 316632 : root->group_pathkeys = NIL;
3613 :
3614 : /* We consider only the first (bottom) window in pathkeys logic */
3615 318822 : if (activeWindows != NIL)
3616 : {
3617 1112 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3618 :
3619 1112 : root->window_pathkeys = make_pathkeys_for_window(root,
3620 : wc,
3621 : tlist);
3622 : }
3623 : else
3624 317710 : root->window_pathkeys = NIL;
3625 :
3626 319252 : if (parse->distinctClause &&
3627 430 : grouping_is_sortable(parse->distinctClause))
3628 426 : root->distinct_pathkeys =
3629 426 : make_pathkeys_for_sortclauses(root,
3630 : parse->distinctClause,
3631 : tlist);
3632 : else
3633 318396 : root->distinct_pathkeys = NIL;
3634 :
3635 318822 : root->sort_pathkeys =
3636 318822 : make_pathkeys_for_sortclauses(root,
3637 : parse->sortClause,
3638 : tlist);
3639 :
3640 : /*
3641 : * Figure out whether we want a sorted result from query_planner.
3642 : *
3643 : * If we have a sortable GROUP BY clause, then we want a result sorted
3644 : * properly for grouping. Otherwise, if we have window functions to
3645 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3646 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3647 : * we try to produce output that's sufficiently well sorted for the
3648 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3649 : * by the ORDER BY clause.
3650 : *
3651 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3652 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3653 : * that might just leave us failing to exploit an available sort order at
3654 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3655 : * much easier, since we know that the parser ensured that one is a
3656 : * superset of the other.
3657 : */
3658 318822 : if (root->group_pathkeys)
3659 2112 : root->query_pathkeys = root->group_pathkeys;
3660 316710 : else if (root->window_pathkeys)
3661 968 : root->query_pathkeys = root->window_pathkeys;
3662 631484 : else if (list_length(root->distinct_pathkeys) >
3663 315742 : list_length(root->sort_pathkeys))
3664 280 : root->query_pathkeys = root->distinct_pathkeys;
3665 315462 : else if (root->sort_pathkeys)
3666 37350 : root->query_pathkeys = root->sort_pathkeys;
3667 : else
3668 278112 : root->query_pathkeys = NIL;
3669 318822 : }
3670 :
3671 : /*
3672 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3673 : *
3674 : * path_rows: number of output rows from scan/join step
3675 : * gd: grouping sets data including list of grouping sets and their clauses
3676 : * target_list: target list containing group clause references
3677 : *
3678 : * If doing grouping sets, we also annotate the gsets data with the estimates
3679 : * for each set and each individual rollup list, with a view to later
3680 : * determining whether some combination of them could be hashed instead.
3681 : */
3682 : static double
3683 27116 : get_number_of_groups(PlannerInfo *root,
3684 : double path_rows,
3685 : grouping_sets_data *gd,
3686 : List *target_list)
3687 : {
3688 27116 : Query *parse = root->parse;
3689 : double dNumGroups;
3690 :
3691 27116 : if (parse->groupClause)
3692 : {
3693 : List *groupExprs;
3694 :
3695 3752 : if (parse->groupingSets)
3696 : {
3697 : /* Add up the estimates for each grouping set */
3698 : ListCell *lc;
3699 : ListCell *lc2;
3700 :
3701 : Assert(gd); /* keep Coverity happy */
3702 :
3703 420 : dNumGroups = 0;
3704 :
3705 1132 : foreach(lc, gd->rollups)
3706 : {
3707 712 : RollupData *rollup = lfirst_node(RollupData, lc);
3708 : ListCell *lc;
3709 :
3710 712 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3711 : target_list);
3712 :
3713 712 : rollup->numGroups = 0.0;
3714 :
3715 2064 : forboth(lc, rollup->gsets, lc2, rollup->gsets_data)
3716 : {
3717 1352 : List *gset = (List *) lfirst(lc);
3718 1352 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3719 1352 : double numGroups = estimate_num_groups(root,
3720 : groupExprs,
3721 : path_rows,
3722 : &gset);
3723 :
3724 1352 : gs->numGroups = numGroups;
3725 1352 : rollup->numGroups += numGroups;
3726 : }
3727 :
3728 712 : dNumGroups += rollup->numGroups;
3729 : }
3730 :
3731 420 : if (gd->hash_sets_idx)
3732 : {
3733 : ListCell *lc;
3734 :
3735 24 : gd->dNumHashGroups = 0;
3736 :
3737 24 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3738 : target_list);
3739 :
3740 52 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3741 : {
3742 28 : List *gset = (List *) lfirst(lc);
3743 28 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3744 28 : double numGroups = estimate_num_groups(root,
3745 : groupExprs,
3746 : path_rows,
3747 : &gset);
3748 :
3749 28 : gs->numGroups = numGroups;
3750 28 : gd->dNumHashGroups += numGroups;
3751 : }
3752 :
3753 24 : dNumGroups += gd->dNumHashGroups;
3754 : }
3755 : }
3756 : else
3757 : {
3758 : /* Plain GROUP BY */
3759 3332 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3760 : target_list);
3761 :
3762 3332 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3763 : NULL);
3764 : }
3765 : }
3766 23364 : else if (parse->groupingSets)
3767 : {
3768 : /* Empty grouping sets ... one result row for each one */
3769 28 : dNumGroups = list_length(parse->groupingSets);
3770 : }
3771 23336 : else if (parse->hasAggs || root->hasHavingQual)
3772 : {
3773 : /* Plain aggregation, one result row */
3774 23336 : dNumGroups = 1;
3775 : }
3776 : else
3777 : {
3778 : /* Not grouping */
3779 0 : dNumGroups = 1;
3780 : }
3781 :
3782 27116 : return dNumGroups;
3783 : }
3784 :
3785 : /*
3786 : * create_grouping_paths
3787 : *
3788 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3789 : * Along the way, we also build an upperrel for Paths which are partially
3790 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3791 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3792 : * the only partially grouped paths we build are also partial paths; that
3793 : * is, they need a Gather and then a FinalizeAggregate.
3794 : *
3795 : * input_rel: contains the source-data Paths
3796 : * target: the pathtarget for the result Paths to compute
3797 : * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
3798 : * gd: grouping sets data including list of grouping sets and their clauses
3799 : *
3800 : * Note: all Paths in input_rel are expected to return the target computed
3801 : * by make_group_input_target.
3802 : */
3803 : static RelOptInfo *
3804 25132 : create_grouping_paths(PlannerInfo *root,
3805 : RelOptInfo *input_rel,
3806 : PathTarget *target,
3807 : bool target_parallel_safe,
3808 : const AggClauseCosts *agg_costs,
3809 : grouping_sets_data *gd)
3810 : {
3811 25132 : Query *parse = root->parse;
3812 : RelOptInfo *grouped_rel;
3813 : RelOptInfo *partially_grouped_rel;
3814 :
3815 : /*
3816 : * Create grouping relation to hold fully aggregated grouping and/or
3817 : * aggregation paths.
3818 : */
3819 25132 : grouped_rel = make_grouping_rel(root, input_rel, target,
3820 : target_parallel_safe, parse->havingQual);
3821 :
3822 : /*
3823 : * Create either paths for a degenerate grouping or paths for ordinary
3824 : * grouping, as appropriate.
3825 : */
3826 25132 : if (is_degenerate_grouping(root))
3827 12 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3828 : else
3829 : {
3830 25120 : int flags = 0;
3831 : GroupPathExtraData extra;
3832 :
3833 : /*
3834 : * Determine whether it's possible to perform sort-based
3835 : * implementations of grouping. (Note that if groupClause is empty,
3836 : * grouping_is_sortable() is trivially true, and all the
3837 : * pathkeys_contained_in() tests will succeed too, so that we'll
3838 : * consider every surviving input path.)
3839 : *
3840 : * If we have grouping sets, we might be able to sort some but not all
3841 : * of them; in this case, we need can_sort to be true as long as we
3842 : * must consider any sorted-input plan.
3843 : */
3844 25120 : if ((gd && gd->rollups != NIL)
3845 24676 : || grouping_is_sortable(parse->groupClause))
3846 25116 : flags |= GROUPING_CAN_USE_SORT;
3847 :
3848 : /*
3849 : * Determine whether we should consider hash-based implementations of
3850 : * grouping.
3851 : *
3852 : * Hashed aggregation only applies if we're grouping. If we have
3853 : * grouping sets, some groups might be hashable but others not; in
3854 : * this case we set can_hash true as long as there is nothing globally
3855 : * preventing us from hashing (and we should therefore consider plans
3856 : * with hashes).
3857 : *
3858 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3859 : * BY aggregates. (Doing so would imply storing *all* the input
3860 : * values in the hash table, and/or running many sorts in parallel,
3861 : * either of which seems like a certain loser.) We similarly don't
3862 : * support ordered-set aggregates in hashed aggregation, but that case
3863 : * is also included in the numOrderedAggs count.
3864 : *
3865 : * Note: grouping_is_hashable() is much more expensive to check than
3866 : * the other gating conditions, so we want to do it last.
3867 : */
3868 25120 : if ((parse->groupClause != NIL &&
3869 4240 : agg_costs->numOrderedAggs == 0 &&
3870 2046 : (gd ? gd->any_hashable : grouping_is_hashable(parse->groupClause))))
3871 2042 : flags |= GROUPING_CAN_USE_HASH;
3872 :
3873 : /*
3874 : * Determine whether partial aggregation is possible.
3875 : */
3876 25120 : if (can_partial_agg(root, agg_costs))
3877 8910 : flags |= GROUPING_CAN_PARTIAL_AGG;
3878 :
3879 25120 : extra.flags = flags;
3880 25120 : extra.target_parallel_safe = target_parallel_safe;
3881 25120 : extra.havingQual = parse->havingQual;
3882 25120 : extra.targetList = parse->targetList;
3883 25120 : extra.partial_costs_set = false;
3884 :
3885 : /*
3886 : * Determine whether partitionwise aggregation is in theory possible.
3887 : * It can be disabled by the user, and for now, we don't try to
3888 : * support grouping sets. create_ordinary_grouping_paths() will check
3889 : * additional conditions, such as whether input_rel is partitioned.
3890 : */
3891 25120 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3892 302 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3893 : else
3894 24818 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3895 :
3896 25120 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3897 : agg_costs, gd, &extra,
3898 : &partially_grouped_rel);
3899 : }
3900 :
3901 25128 : set_cheapest(grouped_rel);
3902 25128 : return grouped_rel;
3903 : }
3904 :
3905 : /*
3906 : * make_grouping_rel
3907 : *
3908 : * Create a new grouping rel and set basic properties.
3909 : *
3910 : * input_rel represents the underlying scan/join relation.
3911 : * target is the output expected from the grouping relation.
3912 : */
3913 : static RelOptInfo *
3914 26074 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3915 : PathTarget *target, bool target_parallel_safe,
3916 : Node *havingQual)
3917 : {
3918 : RelOptInfo *grouped_rel;
3919 :
3920 26074 : if (IS_OTHER_REL(input_rel))
3921 : {
3922 942 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3923 : input_rel->relids);
3924 942 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3925 : }
3926 : else
3927 : {
3928 : /*
3929 : * By tradition, the relids set for the main grouping relation is
3930 : * NULL. (This could be changed, but might require adjustments
3931 : * elsewhere.)
3932 : */
3933 25132 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3934 : }
3935 :
3936 : /* Set target. */
3937 26074 : grouped_rel->reltarget = target;
3938 :
3939 : /*
3940 : * If the input relation is not parallel-safe, then the grouped relation
3941 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3942 : * target list and HAVING quals are parallel-safe.
3943 : */
3944 38210 : if (input_rel->consider_parallel && target_parallel_safe &&
3945 12136 : is_parallel_safe(root, (Node *) havingQual))
3946 12124 : grouped_rel->consider_parallel = true;
3947 :
3948 : /*
3949 : * If the input rel belongs to a single FDW, so does the grouped rel.
3950 : */
3951 26074 : grouped_rel->serverid = input_rel->serverid;
3952 26074 : grouped_rel->userid = input_rel->userid;
3953 26074 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3954 26074 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3955 :
3956 26074 : return grouped_rel;
3957 : }
3958 :
3959 : /*
3960 : * is_degenerate_grouping
3961 : *
3962 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3963 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3964 : * grouping sets are all empty).
3965 : */
3966 : static bool
3967 25132 : is_degenerate_grouping(PlannerInfo *root)
3968 : {
3969 25132 : Query *parse = root->parse;
3970 :
3971 24756 : return (root->hasHavingQual || parse->groupingSets) &&
3972 49888 : !parse->hasAggs && parse->groupClause == NIL;
3973 : }
3974 :
3975 : /*
3976 : * create_degenerate_grouping_paths
3977 : *
3978 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3979 : * supposed to emit either zero or one row for each grouping set depending on
3980 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3981 : * either HAVING or the targetlist, so we actually do not need the FROM table
3982 : * at all! We can just throw away the plan-so-far and generate a Result node.
3983 : * This is a sufficiently unusual corner case that it's not worth contorting
3984 : * the structure of this module to avoid having to generate the earlier paths
3985 : * in the first place.
3986 : */
3987 : static void
3988 12 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3989 : RelOptInfo *grouped_rel)
3990 : {
3991 12 : Query *parse = root->parse;
3992 : int nrows;
3993 : Path *path;
3994 :
3995 12 : nrows = list_length(parse->groupingSets);
3996 12 : if (nrows > 1)
3997 : {
3998 : /*
3999 : * Doesn't seem worthwhile writing code to cons up a generate_series
4000 : * or a values scan to emit multiple rows. Instead just make N clones
4001 : * and append them. (With a volatile HAVING clause, this means you
4002 : * might get between 0 and N output rows. Offhand I think that's
4003 : * desired.)
4004 : */
4005 0 : List *paths = NIL;
4006 :
4007 0 : while (--nrows >= 0)
4008 : {
4009 : path = (Path *)
4010 0 : create_group_result_path(root, grouped_rel,
4011 0 : grouped_rel->reltarget,
4012 0 : (List *) parse->havingQual);
4013 0 : paths = lappend(paths, path);
4014 : }
4015 : path = (Path *)
4016 0 : create_append_path(root,
4017 : grouped_rel,
4018 : paths,
4019 : NIL,
4020 : NIL,
4021 : NULL,
4022 : 0,
4023 : false,
4024 : NIL,
4025 : -1);
4026 : }
4027 : else
4028 : {
4029 : /* No grouping sets, or just one, so one output row */
4030 : path = (Path *)
4031 12 : create_group_result_path(root, grouped_rel,
4032 12 : grouped_rel->reltarget,
4033 12 : (List *) parse->havingQual);
4034 : }
4035 :
4036 12 : add_path(grouped_rel, path);
4037 12 : }
4038 :
4039 : /*
4040 : * create_ordinary_grouping_paths
4041 : *
4042 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4043 : *
4044 : * We need to consider sorted and hashed aggregation in the same function,
4045 : * because otherwise (1) it would be harder to throw an appropriate error
4046 : * message if neither way works, and (2) we should not allow hashtable size
4047 : * considerations to dissuade us from using hashing if sorting is not possible.
4048 : *
4049 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4050 : * function creates, or to NULL if it doesn't create one.
4051 : */
4052 : static void
4053 26062 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4054 : RelOptInfo *grouped_rel,
4055 : const AggClauseCosts *agg_costs,
4056 : grouping_sets_data *gd,
4057 : GroupPathExtraData *extra,
4058 : RelOptInfo **partially_grouped_rel_p)
4059 : {
4060 26062 : Path *cheapest_path = input_rel->cheapest_total_path;
4061 26062 : RelOptInfo *partially_grouped_rel = NULL;
4062 : double dNumGroups;
4063 26062 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4064 :
4065 : /*
4066 : * If this is the topmost grouping relation or if the parent relation is
4067 : * doing some form of partitionwise aggregation, then we may be able to do
4068 : * it at this level also. However, if the input relation is not
4069 : * partitioned, partitionwise aggregate is impossible.
4070 : */
4071 26062 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4072 1244 : IS_PARTITIONED_REL(input_rel))
4073 : {
4074 : /*
4075 : * If this is the topmost relation or if the parent relation is doing
4076 : * full partitionwise aggregation, then we can do full partitionwise
4077 : * aggregation provided that the GROUP BY clause contains all of the
4078 : * partitioning columns at this level. Otherwise, we can do at most
4079 : * partial partitionwise aggregation. But if partial aggregation is
4080 : * not supported in general then we can't use it for partitionwise
4081 : * aggregation either.
4082 : */
4083 716 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4084 342 : group_by_has_partkey(input_rel, extra->targetList,
4085 342 : root->parse->groupClause))
4086 192 : patype = PARTITIONWISE_AGGREGATE_FULL;
4087 182 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4088 154 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4089 : else
4090 28 : patype = PARTITIONWISE_AGGREGATE_NONE;
4091 : }
4092 :
4093 : /*
4094 : * Before generating paths for grouped_rel, we first generate any possible
4095 : * partially grouped paths; that way, later code can easily consider both
4096 : * parallel and non-parallel approaches to grouping.
4097 : */
4098 26062 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4099 : {
4100 : bool force_rel_creation;
4101 :
4102 : /*
4103 : * If we're doing partitionwise aggregation at this level, force
4104 : * creation of a partially_grouped_rel so we can add partitionwise
4105 : * paths to it.
4106 : */
4107 9804 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4108 :
4109 : partially_grouped_rel =
4110 9804 : create_partial_grouping_paths(root,
4111 : grouped_rel,
4112 : input_rel,
4113 : gd,
4114 : extra,
4115 : force_rel_creation);
4116 : }
4117 :
4118 : /* Set out parameter. */
4119 26062 : *partially_grouped_rel_p = partially_grouped_rel;
4120 :
4121 : /* Apply partitionwise aggregation technique, if possible. */
4122 26062 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4123 346 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4124 : partially_grouped_rel, agg_costs,
4125 : gd, patype, extra);
4126 :
4127 : /* If we are doing partial aggregation only, return. */
4128 26062 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4129 : {
4130 : Assert(partially_grouped_rel);
4131 :
4132 398 : if (partially_grouped_rel->pathlist)
4133 398 : set_cheapest(partially_grouped_rel);
4134 :
4135 398 : return;
4136 : }
4137 :
4138 : /* Gather any partially grouped partial paths. */
4139 25664 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4140 : {
4141 854 : gather_grouping_paths(root, partially_grouped_rel);
4142 854 : set_cheapest(partially_grouped_rel);
4143 : }
4144 :
4145 : /*
4146 : * Estimate number of groups.
4147 : */
4148 25664 : dNumGroups = get_number_of_groups(root,
4149 : cheapest_path->rows,
4150 : gd,
4151 : extra->targetList);
4152 :
4153 : /* Build final grouping paths */
4154 25664 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4155 : partially_grouped_rel, agg_costs, gd,
4156 : dNumGroups, extra);
4157 :
4158 : /* Give a helpful error if we failed to find any implementation */
4159 25664 : if (grouped_rel->pathlist == NIL)
4160 4 : ereport(ERROR,
4161 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4162 : errmsg("could not implement GROUP BY"),
4163 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4164 :
4165 : /*
4166 : * If there is an FDW that's responsible for all baserels of the query,
4167 : * let it consider adding ForeignPaths.
4168 : */
4169 25660 : if (grouped_rel->fdwroutine &&
4170 274 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4171 274 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4172 : input_rel, grouped_rel,
4173 : extra);
4174 :
4175 : /* Let extensions possibly add some more paths */
4176 25660 : if (create_upper_paths_hook)
4177 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4178 : input_rel, grouped_rel,
4179 : extra);
4180 : }
4181 :
4182 : /*
4183 : * For a given input path, consider the possible ways of doing grouping sets on
4184 : * it, by combinations of hashing and sorting. This can be called multiple
4185 : * times, so it's important that it not scribble on input. No result is
4186 : * returned, but any generated paths are added to grouped_rel.
4187 : */
4188 : static void
4189 892 : consider_groupingsets_paths(PlannerInfo *root,
4190 : RelOptInfo *grouped_rel,
4191 : Path *path,
4192 : bool is_sorted,
4193 : bool can_hash,
4194 : grouping_sets_data *gd,
4195 : const AggClauseCosts *agg_costs,
4196 : double dNumGroups)
4197 : {
4198 892 : Query *parse = root->parse;
4199 :
4200 : /*
4201 : * If we're not being offered sorted input, then only consider plans that
4202 : * can be done entirely by hashing.
4203 : *
4204 : * We can hash everything if it looks like it'll fit in work_mem. But if
4205 : * the input is actually sorted despite not being advertised as such, we
4206 : * prefer to make use of that in order to use less memory.
4207 : *
4208 : * If none of the grouping sets are sortable, then ignore the work_mem
4209 : * limit and generate a path anyway, since otherwise we'll just fail.
4210 : */
4211 892 : if (!is_sorted)
4212 : {
4213 396 : List *new_rollups = NIL;
4214 396 : RollupData *unhashed_rollup = NULL;
4215 : List *sets_data;
4216 396 : List *empty_sets_data = NIL;
4217 396 : List *empty_sets = NIL;
4218 : ListCell *lc;
4219 396 : ListCell *l_start = list_head(gd->rollups);
4220 396 : AggStrategy strat = AGG_HASHED;
4221 : double hashsize;
4222 396 : double exclude_groups = 0.0;
4223 :
4224 : Assert(can_hash);
4225 :
4226 : /*
4227 : * If the input is coincidentally sorted usefully (which can happen
4228 : * even if is_sorted is false, since that only means that our caller
4229 : * has set up the sorting for us), then save some hashtable space by
4230 : * making use of that. But we need to watch out for degenerate cases:
4231 : *
4232 : * 1) If there are any empty grouping sets, then group_pathkeys might
4233 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4234 : * there will be a rollup containing only empty groups, and the
4235 : * pathkeys_contained_in test is vacuously true; this is ok.
4236 : *
4237 : * XXX: the above relies on the fact that group_pathkeys is generated
4238 : * from the first rollup. If we add the ability to consider multiple
4239 : * sort orders for grouping input, this assumption might fail.
4240 : *
4241 : * 2) If there are no empty sets and only unsortable sets, then the
4242 : * rollups list will be empty (and thus l_start == NULL), and
4243 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4244 : * pathkeys_contained_in test doesn't cause us to crash.
4245 : */
4246 788 : if (l_start != NULL &&
4247 392 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4248 : {
4249 8 : unhashed_rollup = lfirst_node(RollupData, l_start);
4250 8 : exclude_groups = unhashed_rollup->numGroups;
4251 8 : l_start = lnext(gd->rollups, l_start);
4252 : }
4253 :
4254 396 : hashsize = estimate_hashagg_tablesize(path,
4255 : agg_costs,
4256 : dNumGroups - exclude_groups);
4257 :
4258 : /*
4259 : * If we have sortable columns to work with (gd->rollups is non-empty)
4260 : * and enable_groupingsets_hash_disk is disabled, don't generate
4261 : * hash-based paths that will exceed work_mem.
4262 : */
4263 396 : if (!enable_groupingsets_hash_disk &&
4264 380 : hashsize > work_mem * 1024L && gd->rollups)
4265 16 : return; /* nope, won't fit */
4266 :
4267 : /*
4268 : * We need to burst the existing rollups list into individual grouping
4269 : * sets and recompute a groupClause for each set.
4270 : */
4271 380 : sets_data = list_copy(gd->unsortable_sets);
4272 :
4273 960 : for_each_cell(lc, gd->rollups, l_start)
4274 : {
4275 596 : RollupData *rollup = lfirst_node(RollupData, lc);
4276 :
4277 : /*
4278 : * If we find an unhashable rollup that's not been skipped by the
4279 : * "actually sorted" check above, we can't cope; we'd need sorted
4280 : * input (with a different sort order) but we can't get that here.
4281 : * So bail out; we'll get a valid path from the is_sorted case
4282 : * instead.
4283 : *
4284 : * The mere presence of empty grouping sets doesn't make a rollup
4285 : * unhashable (see preprocess_grouping_sets), we handle those
4286 : * specially below.
4287 : */
4288 596 : if (!rollup->hashable)
4289 16 : return;
4290 :
4291 580 : sets_data = list_concat(sets_data, rollup->gsets_data);
4292 : }
4293 1552 : foreach(lc, sets_data)
4294 : {
4295 1188 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4296 1188 : List *gset = gs->set;
4297 : RollupData *rollup;
4298 :
4299 1188 : if (gset == NIL)
4300 : {
4301 : /* Empty grouping sets can't be hashed. */
4302 264 : empty_sets_data = lappend(empty_sets_data, gs);
4303 264 : empty_sets = lappend(empty_sets, NIL);
4304 : }
4305 : else
4306 : {
4307 924 : rollup = makeNode(RollupData);
4308 :
4309 924 : rollup->groupClause = preprocess_groupclause(root, gset);
4310 924 : rollup->gsets_data = list_make1(gs);
4311 924 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4312 : rollup->gsets_data,
4313 : gd->tleref_to_colnum_map);
4314 924 : rollup->numGroups = gs->numGroups;
4315 924 : rollup->hashable = true;
4316 924 : rollup->is_hashed = true;
4317 924 : new_rollups = lappend(new_rollups, rollup);
4318 : }
4319 : }
4320 :
4321 : /*
4322 : * If we didn't find anything nonempty to hash, then bail. We'll
4323 : * generate a path from the is_sorted case.
4324 : */
4325 364 : if (new_rollups == NIL)
4326 0 : return;
4327 :
4328 : /*
4329 : * If there were empty grouping sets they should have been in the
4330 : * first rollup.
4331 : */
4332 : Assert(!unhashed_rollup || !empty_sets);
4333 :
4334 364 : if (unhashed_rollup)
4335 : {
4336 8 : new_rollups = lappend(new_rollups, unhashed_rollup);
4337 8 : strat = AGG_MIXED;
4338 : }
4339 356 : else if (empty_sets)
4340 : {
4341 232 : RollupData *rollup = makeNode(RollupData);
4342 :
4343 232 : rollup->groupClause = NIL;
4344 232 : rollup->gsets_data = empty_sets_data;
4345 232 : rollup->gsets = empty_sets;
4346 232 : rollup->numGroups = list_length(empty_sets);
4347 232 : rollup->hashable = false;
4348 232 : rollup->is_hashed = false;
4349 232 : new_rollups = lappend(new_rollups, rollup);
4350 232 : strat = AGG_MIXED;
4351 : }
4352 :
4353 364 : add_path(grouped_rel, (Path *)
4354 364 : create_groupingsets_path(root,
4355 : grouped_rel,
4356 : path,
4357 364 : (List *) parse->havingQual,
4358 : strat,
4359 : new_rollups,
4360 : agg_costs,
4361 : dNumGroups));
4362 364 : return;
4363 : }
4364 :
4365 : /*
4366 : * If we have sorted input but nothing we can do with it, bail.
4367 : */
4368 496 : if (list_length(gd->rollups) == 0)
4369 0 : return;
4370 :
4371 : /*
4372 : * Given sorted input, we try and make two paths: one sorted and one mixed
4373 : * sort/hash. (We need to try both because hashagg might be disabled, or
4374 : * some columns might not be sortable.)
4375 : *
4376 : * can_hash is passed in as false if some obstacle elsewhere (such as
4377 : * ordered aggs) means that we shouldn't consider hashing at all.
4378 : */
4379 496 : if (can_hash && gd->any_hashable)
4380 : {
4381 444 : List *rollups = NIL;
4382 444 : List *hash_sets = list_copy(gd->unsortable_sets);
4383 444 : double availspace = (work_mem * 1024.0);
4384 : ListCell *lc;
4385 :
4386 : /*
4387 : * Account first for space needed for groups we can't sort at all.
4388 : */
4389 444 : availspace -= estimate_hashagg_tablesize(path,
4390 : agg_costs,
4391 : gd->dNumHashGroups);
4392 :
4393 444 : if (availspace > 0 && list_length(gd->rollups) > 1)
4394 : {
4395 : double scale;
4396 236 : int num_rollups = list_length(gd->rollups);
4397 : int k_capacity;
4398 236 : int *k_weights = palloc(num_rollups * sizeof(int));
4399 236 : Bitmapset *hash_items = NULL;
4400 : int i;
4401 :
4402 : /*
4403 : * We treat this as a knapsack problem: the knapsack capacity
4404 : * represents work_mem, the item weights are the estimated memory
4405 : * usage of the hashtables needed to implement a single rollup,
4406 : * and we really ought to use the cost saving as the item value;
4407 : * however, currently the costs assigned to sort nodes don't
4408 : * reflect the comparison costs well, and so we treat all items as
4409 : * of equal value (each rollup we hash instead saves us one sort).
4410 : *
4411 : * To use the discrete knapsack, we need to scale the values to a
4412 : * reasonably small bounded range. We choose to allow a 5% error
4413 : * margin; we have no more than 4096 rollups in the worst possible
4414 : * case, which with a 5% error margin will require a bit over 42MB
4415 : * of workspace. (Anyone wanting to plan queries that complex had
4416 : * better have the memory for it. In more reasonable cases, with
4417 : * no more than a couple of dozen rollups, the memory usage will
4418 : * be negligible.)
4419 : *
4420 : * k_capacity is naturally bounded, but we clamp the values for
4421 : * scale and weight (below) to avoid overflows or underflows (or
4422 : * uselessly trying to use a scale factor less than 1 byte).
4423 : */
4424 236 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4425 236 : k_capacity = (int) floor(availspace / scale);
4426 :
4427 : /*
4428 : * We leave the first rollup out of consideration since it's the
4429 : * one that matches the input sort order. We assign indexes "i"
4430 : * to only those entries considered for hashing; the second loop,
4431 : * below, must use the same condition.
4432 : */
4433 236 : i = 0;
4434 616 : for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4435 : {
4436 380 : RollupData *rollup = lfirst_node(RollupData, lc);
4437 :
4438 380 : if (rollup->hashable)
4439 : {
4440 380 : double sz = estimate_hashagg_tablesize(path,
4441 : agg_costs,
4442 : rollup->numGroups);
4443 :
4444 : /*
4445 : * If sz is enormous, but work_mem (and hence scale) is
4446 : * small, avoid integer overflow here.
4447 : */
4448 380 : k_weights[i] = (int) Min(floor(sz / scale),
4449 : k_capacity + 1.0);
4450 380 : ++i;
4451 : }
4452 : }
4453 :
4454 : /*
4455 : * Apply knapsack algorithm; compute the set of items which
4456 : * maximizes the value stored (in this case the number of sorts
4457 : * saved) while keeping the total size (approximately) within
4458 : * capacity.
4459 : */
4460 236 : if (i > 0)
4461 236 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4462 :
4463 236 : if (!bms_is_empty(hash_items))
4464 : {
4465 220 : rollups = list_make1(linitial(gd->rollups));
4466 :
4467 220 : i = 0;
4468 568 : for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4469 : {
4470 348 : RollupData *rollup = lfirst_node(RollupData, lc);
4471 :
4472 348 : if (rollup->hashable)
4473 : {
4474 348 : if (bms_is_member(i, hash_items))
4475 324 : hash_sets = list_concat(hash_sets,
4476 324 : rollup->gsets_data);
4477 : else
4478 24 : rollups = lappend(rollups, rollup);
4479 348 : ++i;
4480 : }
4481 : else
4482 0 : rollups = lappend(rollups, rollup);
4483 : }
4484 : }
4485 : }
4486 :
4487 444 : if (!rollups && hash_sets)
4488 16 : rollups = list_copy(gd->rollups);
4489 :
4490 816 : foreach(lc, hash_sets)
4491 : {
4492 372 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4493 372 : RollupData *rollup = makeNode(RollupData);
4494 :
4495 : Assert(gs->set != NIL);
4496 :
4497 372 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4498 372 : rollup->gsets_data = list_make1(gs);
4499 372 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4500 : rollup->gsets_data,
4501 : gd->tleref_to_colnum_map);
4502 372 : rollup->numGroups = gs->numGroups;
4503 372 : rollup->hashable = true;
4504 372 : rollup->is_hashed = true;
4505 372 : rollups = lcons(rollup, rollups);
4506 : }
4507 :
4508 444 : if (rollups)
4509 : {
4510 236 : add_path(grouped_rel, (Path *)
4511 236 : create_groupingsets_path(root,
4512 : grouped_rel,
4513 : path,
4514 236 : (List *) parse->havingQual,
4515 : AGG_MIXED,
4516 : rollups,
4517 : agg_costs,
4518 : dNumGroups));
4519 : }
4520 : }
4521 :
4522 : /*
4523 : * Now try the simple sorted case.
4524 : */
4525 496 : if (!gd->unsortable_sets)
4526 476 : add_path(grouped_rel, (Path *)
4527 952 : create_groupingsets_path(root,
4528 : grouped_rel,
4529 : path,
4530 476 : (List *) parse->havingQual,
4531 : AGG_SORTED,
4532 : gd->rollups,
4533 : agg_costs,
4534 : dNumGroups));
4535 : }
4536 :
4537 : /*
4538 : * create_window_paths
4539 : *
4540 : * Build a new upperrel containing Paths for window-function evaluation.
4541 : *
4542 : * input_rel: contains the source-data Paths
4543 : * input_target: result of make_window_input_target
4544 : * output_target: what the topmost WindowAggPath should return
4545 : * wflists: result of find_window_functions
4546 : * activeWindows: result of select_active_windows
4547 : *
4548 : * Note: all Paths in input_rel are expected to return input_target.
4549 : */
4550 : static RelOptInfo *
4551 1112 : create_window_paths(PlannerInfo *root,
4552 : RelOptInfo *input_rel,
4553 : PathTarget *input_target,
4554 : PathTarget *output_target,
4555 : bool output_target_parallel_safe,
4556 : WindowFuncLists *wflists,
4557 : List *activeWindows)
4558 : {
4559 : RelOptInfo *window_rel;
4560 : ListCell *lc;
4561 :
4562 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4563 1112 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4564 :
4565 : /*
4566 : * If the input relation is not parallel-safe, then the window relation
4567 : * can't be parallel-safe, either. Otherwise, we need to examine the
4568 : * target list and active windows for non-parallel-safe constructs.
4569 : */
4570 1112 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4571 0 : is_parallel_safe(root, (Node *) activeWindows))
4572 0 : window_rel->consider_parallel = true;
4573 :
4574 : /*
4575 : * If the input rel belongs to a single FDW, so does the window rel.
4576 : */
4577 1112 : window_rel->serverid = input_rel->serverid;
4578 1112 : window_rel->userid = input_rel->userid;
4579 1112 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4580 1112 : window_rel->fdwroutine = input_rel->fdwroutine;
4581 :
4582 : /*
4583 : * Consider computing window functions starting from the existing
4584 : * cheapest-total path (which will likely require a sort) as well as any
4585 : * existing paths that satisfy root->window_pathkeys (which won't).
4586 : */
4587 2336 : foreach(lc, input_rel->pathlist)
4588 : {
4589 1224 : Path *path = (Path *) lfirst(lc);
4590 :
4591 1336 : if (path == input_rel->cheapest_total_path ||
4592 112 : pathkeys_contained_in(root->window_pathkeys, path->pathkeys))
4593 1156 : create_one_window_path(root,
4594 : window_rel,
4595 : path,
4596 : input_target,
4597 : output_target,
4598 : wflists,
4599 : activeWindows);
4600 : }
4601 :
4602 : /*
4603 : * If there is an FDW that's responsible for all baserels of the query,
4604 : * let it consider adding ForeignPaths.
4605 : */
4606 1112 : if (window_rel->fdwroutine &&
4607 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4608 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4609 : input_rel, window_rel,
4610 : NULL);
4611 :
4612 : /* Let extensions possibly add some more paths */
4613 1112 : if (create_upper_paths_hook)
4614 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4615 : input_rel, window_rel, NULL);
4616 :
4617 : /* Now choose the best path(s) */
4618 1112 : set_cheapest(window_rel);
4619 :
4620 1112 : return window_rel;
4621 : }
4622 :
4623 : /*
4624 : * Stack window-function implementation steps atop the given Path, and
4625 : * add the result to window_rel.
4626 : *
4627 : * window_rel: upperrel to contain result
4628 : * path: input Path to use (must return input_target)
4629 : * input_target: result of make_window_input_target
4630 : * output_target: what the topmost WindowAggPath should return
4631 : * wflists: result of find_window_functions
4632 : * activeWindows: result of select_active_windows
4633 : */
4634 : static void
4635 1156 : create_one_window_path(PlannerInfo *root,
4636 : RelOptInfo *window_rel,
4637 : Path *path,
4638 : PathTarget *input_target,
4639 : PathTarget *output_target,
4640 : WindowFuncLists *wflists,
4641 : List *activeWindows)
4642 : {
4643 : PathTarget *window_target;
4644 : ListCell *l;
4645 :
4646 : /*
4647 : * Since each window clause could require a different sort order, we stack
4648 : * up a WindowAgg node for each clause, with sort steps between them as
4649 : * needed. (We assume that select_active_windows chose a good order for
4650 : * executing the clauses in.)
4651 : *
4652 : * input_target should contain all Vars and Aggs needed for the result.
4653 : * (In some cases we wouldn't need to propagate all of these all the way
4654 : * to the top, since they might only be needed as inputs to WindowFuncs.
4655 : * It's probably not worth trying to optimize that though.) It must also
4656 : * contain all window partitioning and sorting expressions, to ensure
4657 : * they're computed only once at the bottom of the stack (that's critical
4658 : * for volatile functions). As we climb up the stack, we'll add outputs
4659 : * for the WindowFuncs computed at each level.
4660 : */
4661 1156 : window_target = input_target;
4662 :
4663 2360 : foreach(l, activeWindows)
4664 : {
4665 1204 : WindowClause *wc = lfirst_node(WindowClause, l);
4666 : List *window_pathkeys;
4667 :
4668 1204 : window_pathkeys = make_pathkeys_for_window(root,
4669 : wc,
4670 : root->processed_tlist);
4671 :
4672 : /* Sort if necessary */
4673 1204 : if (!pathkeys_contained_in(window_pathkeys, path->pathkeys))
4674 : {
4675 952 : path = (Path *) create_sort_path(root, window_rel,
4676 : path,
4677 : window_pathkeys,
4678 : -1.0);
4679 : }
4680 :
4681 1204 : if (lnext(activeWindows, l))
4682 : {
4683 : /*
4684 : * Add the current WindowFuncs to the output target for this
4685 : * intermediate WindowAggPath. We must copy window_target to
4686 : * avoid changing the previous path's target.
4687 : *
4688 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4689 : * we do need to account for the increase in tlist width.
4690 : */
4691 : ListCell *lc2;
4692 :
4693 48 : window_target = copy_pathtarget(window_target);
4694 96 : foreach(lc2, wflists->windowFuncs[wc->winref])
4695 : {
4696 48 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4697 :
4698 48 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4699 48 : window_target->width += get_typavgwidth(wfunc->wintype, -1);
4700 : }
4701 : }
4702 : else
4703 : {
4704 : /* Install the goal target in the topmost WindowAgg */
4705 1156 : window_target = output_target;
4706 : }
4707 :
4708 : path = (Path *)
4709 1204 : create_windowagg_path(root, window_rel, path, window_target,
4710 1204 : wflists->windowFuncs[wc->winref],
4711 : wc);
4712 : }
4713 :
4714 1156 : add_path(window_rel, path);
4715 1156 : }
4716 :
4717 : /*
4718 : * create_distinct_paths
4719 : *
4720 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4721 : *
4722 : * input_rel: contains the source-data Paths
4723 : *
4724 : * Note: input paths should already compute the desired pathtarget, since
4725 : * Sort/Unique won't project anything.
4726 : */
4727 : static RelOptInfo *
4728 430 : create_distinct_paths(PlannerInfo *root,
4729 : RelOptInfo *input_rel)
4730 : {
4731 430 : Query *parse = root->parse;
4732 430 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4733 : RelOptInfo *distinct_rel;
4734 : double numDistinctRows;
4735 : bool allow_hash;
4736 : Path *path;
4737 : ListCell *lc;
4738 :
4739 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4740 430 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4741 :
4742 : /*
4743 : * We don't compute anything at this level, so distinct_rel will be
4744 : * parallel-safe if the input rel is parallel-safe. In particular, if
4745 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4746 : * output those expressions, and will not be parallel-safe unless those
4747 : * expressions are parallel-safe.
4748 : */
4749 430 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4750 :
4751 : /*
4752 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4753 : */
4754 430 : distinct_rel->serverid = input_rel->serverid;
4755 430 : distinct_rel->userid = input_rel->userid;
4756 430 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4757 430 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4758 :
4759 : /* Estimate number of distinct rows there will be */
4760 430 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4761 406 : root->hasHavingQual)
4762 : {
4763 : /*
4764 : * If there was grouping or aggregation, use the number of input rows
4765 : * as the estimated number of DISTINCT rows (ie, assume the input is
4766 : * already mostly unique).
4767 : */
4768 24 : numDistinctRows = cheapest_input_path->rows;
4769 : }
4770 : else
4771 : {
4772 : /*
4773 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4774 : */
4775 : List *distinctExprs;
4776 :
4777 406 : distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
4778 : parse->targetList);
4779 406 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4780 : cheapest_input_path->rows,
4781 : NULL);
4782 : }
4783 :
4784 : /*
4785 : * Consider sort-based implementations of DISTINCT, if possible.
4786 : */
4787 430 : if (grouping_is_sortable(parse->distinctClause))
4788 : {
4789 : /*
4790 : * First, if we have any adequately-presorted paths, just stick a
4791 : * Unique node on those. Then consider doing an explicit sort of the
4792 : * cheapest input path and Unique'ing that.
4793 : *
4794 : * When we have DISTINCT ON, we must sort by the more rigorous of
4795 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
4796 : * Also, if we do have to do an explicit sort, we might as well use
4797 : * the more rigorous ordering to avoid a second sort later. (Note
4798 : * that the parser will have ensured that one clause is a prefix of
4799 : * the other.)
4800 : */
4801 : List *needed_pathkeys;
4802 :
4803 462 : if (parse->hasDistinctOn &&
4804 36 : list_length(root->distinct_pathkeys) <
4805 36 : list_length(root->sort_pathkeys))
4806 20 : needed_pathkeys = root->sort_pathkeys;
4807 : else
4808 406 : needed_pathkeys = root->distinct_pathkeys;
4809 :
4810 974 : foreach(lc, input_rel->pathlist)
4811 : {
4812 548 : Path *path = (Path *) lfirst(lc);
4813 :
4814 548 : if (pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4815 : {
4816 30 : add_path(distinct_rel, (Path *)
4817 30 : create_upper_unique_path(root, distinct_rel,
4818 : path,
4819 30 : list_length(root->distinct_pathkeys),
4820 : numDistinctRows));
4821 : }
4822 : }
4823 :
4824 : /* For explicit-sort case, always use the more rigorous clause */
4825 852 : if (list_length(root->distinct_pathkeys) <
4826 426 : list_length(root->sort_pathkeys))
4827 : {
4828 20 : needed_pathkeys = root->sort_pathkeys;
4829 : /* Assert checks that parser didn't mess up... */
4830 : Assert(pathkeys_contained_in(root->distinct_pathkeys,
4831 : needed_pathkeys));
4832 : }
4833 : else
4834 406 : needed_pathkeys = root->distinct_pathkeys;
4835 :
4836 426 : path = cheapest_input_path;
4837 426 : if (!pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4838 416 : path = (Path *) create_sort_path(root, distinct_rel,
4839 : path,
4840 : needed_pathkeys,
4841 : -1.0);
4842 :
4843 426 : add_path(distinct_rel, (Path *)
4844 426 : create_upper_unique_path(root, distinct_rel,
4845 : path,
4846 426 : list_length(root->distinct_pathkeys),
4847 : numDistinctRows));
4848 : }
4849 :
4850 : /*
4851 : * Consider hash-based implementations of DISTINCT, if possible.
4852 : *
4853 : * If we were not able to make any other types of path, we *must* hash or
4854 : * die trying. If we do have other choices, there are several things that
4855 : * should prevent selection of hashing: if the query uses DISTINCT ON
4856 : * (because it won't really have the expected behavior if we hash), or if
4857 : * enable_hashagg is off, or if it looks like the hashtable will exceed
4858 : * work_mem.
4859 : *
4860 : * Note: grouping_is_hashable() is much more expensive to check than the
4861 : * other gating conditions, so we want to do it last.
4862 : */
4863 430 : if (distinct_rel->pathlist == NIL)
4864 4 : allow_hash = true; /* we have no alternatives */
4865 426 : else if (parse->hasDistinctOn || !enable_hashagg)
4866 48 : allow_hash = false; /* policy-based decision not to hash */
4867 : else
4868 : {
4869 378 : Size hashentrysize = hash_agg_entry_size(0, cheapest_input_path->pathtarget->width, 0);
4870 :
4871 378 : allow_hash = enable_hashagg_disk ||
4872 0 : (hashentrysize * numDistinctRows <= work_mem * 1024L);
4873 : }
4874 :
4875 430 : if (allow_hash && grouping_is_hashable(parse->distinctClause))
4876 : {
4877 : /* Generate hashed aggregate path --- no sort needed */
4878 382 : add_path(distinct_rel, (Path *)
4879 382 : create_agg_path(root,
4880 : distinct_rel,
4881 : cheapest_input_path,
4882 : cheapest_input_path->pathtarget,
4883 : AGG_HASHED,
4884 : AGGSPLIT_SIMPLE,
4885 : parse->distinctClause,
4886 : NIL,
4887 : NULL,
4888 : numDistinctRows));
4889 : }
4890 :
4891 : /* Give a helpful error if we failed to find any implementation */
4892 430 : if (distinct_rel->pathlist == NIL)
4893 0 : ereport(ERROR,
4894 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4895 : errmsg("could not implement DISTINCT"),
4896 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4897 :
4898 : /*
4899 : * If there is an FDW that's responsible for all baserels of the query,
4900 : * let it consider adding ForeignPaths.
4901 : */
4902 430 : if (distinct_rel->fdwroutine &&
4903 16 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4904 16 : distinct_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_DISTINCT,
4905 : input_rel, distinct_rel,
4906 : NULL);
4907 :
4908 : /* Let extensions possibly add some more paths */
4909 430 : if (create_upper_paths_hook)
4910 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT,
4911 : input_rel, distinct_rel, NULL);
4912 :
4913 : /* Now choose the best path(s) */
4914 430 : set_cheapest(distinct_rel);
4915 :
4916 430 : return distinct_rel;
4917 : }
4918 :
4919 : /*
4920 : * create_ordered_paths
4921 : *
4922 : * Build a new upperrel containing Paths for ORDER BY evaluation.
4923 : *
4924 : * All paths in the result must satisfy the ORDER BY ordering.
4925 : * The only new paths we need consider are an explicit full sort
4926 : * and incremental sort on the cheapest-total existing path.
4927 : *
4928 : * input_rel: contains the source-data Paths
4929 : * target: the output tlist the result Paths must emit
4930 : * limit_tuples: estimated bound on the number of output tuples,
4931 : * or -1 if no LIMIT or couldn't estimate
4932 : *
4933 : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
4934 : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
4935 : */
4936 : static RelOptInfo *
4937 39162 : create_ordered_paths(PlannerInfo *root,
4938 : RelOptInfo *input_rel,
4939 : PathTarget *target,
4940 : bool target_parallel_safe,
4941 : double limit_tuples)
4942 : {
4943 39162 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4944 : RelOptInfo *ordered_rel;
4945 : ListCell *lc;
4946 :
4947 : /* For now, do all work in the (ORDERED, NULL) upperrel */
4948 39162 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
4949 :
4950 : /*
4951 : * If the input relation is not parallel-safe, then the ordered relation
4952 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4953 : * target list is parallel-safe.
4954 : */
4955 39162 : if (input_rel->consider_parallel && target_parallel_safe)
4956 17154 : ordered_rel->consider_parallel = true;
4957 :
4958 : /*
4959 : * If the input rel belongs to a single FDW, so does the ordered_rel.
4960 : */
4961 39162 : ordered_rel->serverid = input_rel->serverid;
4962 39162 : ordered_rel->userid = input_rel->userid;
4963 39162 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
4964 39162 : ordered_rel->fdwroutine = input_rel->fdwroutine;
4965 :
4966 94370 : foreach(lc, input_rel->pathlist)
4967 : {
4968 55208 : Path *input_path = (Path *) lfirst(lc);
4969 55208 : Path *sorted_path = input_path;
4970 : bool is_sorted;
4971 : int presorted_keys;
4972 :
4973 55208 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
4974 : input_path->pathkeys, &presorted_keys);
4975 :
4976 55208 : if (is_sorted)
4977 : {
4978 : /* Use the input path as is, but add a projection step if needed */
4979 21400 : if (sorted_path->pathtarget != target)
4980 2964 : sorted_path = apply_projection_to_path(root, ordered_rel,
4981 : sorted_path, target);
4982 :
4983 21400 : add_path(ordered_rel, sorted_path);
4984 : }
4985 : else
4986 : {
4987 : /*
4988 : * Try adding an explicit sort, but only to the cheapest total
4989 : * path since a full sort should generally add the same cost to
4990 : * all paths.
4991 : */
4992 33808 : if (input_path == cheapest_input_path)
4993 : {
4994 : /*
4995 : * Sort the cheapest input path. An explicit sort here can
4996 : * take advantage of LIMIT.
4997 : */
4998 31358 : sorted_path = (Path *) create_sort_path(root,
4999 : ordered_rel,
5000 : input_path,
5001 : root->sort_pathkeys,
5002 : limit_tuples);
5003 : /* Add projection step if needed */
5004 31358 : if (sorted_path->pathtarget != target)
5005 6668 : sorted_path = apply_projection_to_path(root, ordered_rel,
5006 : sorted_path, target);
5007 :
5008 31358 : add_path(ordered_rel, sorted_path);
5009 : }
5010 :
5011 : /*
5012 : * If incremental sort is enabled, then try it as well. Unlike
5013 : * with regular sorts, we can't just look at the cheapest path,
5014 : * because the cost of incremental sort depends on how well
5015 : * presorted the path is. Additionally incremental sort may enable
5016 : * a cheaper startup path to win out despite higher total cost.
5017 : */
5018 33808 : if (!enable_incrementalsort)
5019 32544 : continue;
5020 :
5021 : /* Likewise, if the path can't be used for incremental sort. */
5022 33172 : if (!presorted_keys)
5023 31908 : continue;
5024 :
5025 : /* Also consider incremental sort. */
5026 1264 : sorted_path = (Path *) create_incremental_sort_path(root,
5027 : ordered_rel,
5028 : input_path,
5029 : root->sort_pathkeys,
5030 : presorted_keys,
5031 : limit_tuples);
5032 :
5033 : /* Add projection step if needed */
5034 1264 : if (sorted_path->pathtarget != target)
5035 402 : sorted_path = apply_projection_to_path(root, ordered_rel,
5036 : sorted_path, target);
5037 :
5038 1264 : add_path(ordered_rel, sorted_path);
5039 : }
5040 : }
5041 :
5042 : /*
5043 : * generate_gather_paths() will have already generated a simple Gather
5044 : * path for the best parallel path, if any, and the loop above will have
5045 : * considered sorting it. Similarly, generate_gather_paths() will also
5046 : * have generated order-preserving Gather Merge plans which can be used
5047 : * without sorting if they happen to match the sort_pathkeys, and the loop
5048 : * above will have handled those as well. However, there's one more
5049 : * possibility: it may make sense to sort the cheapest partial path
5050 : * according to the required output order and then use Gather Merge.
5051 : */
5052 39162 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5053 17046 : input_rel->partial_pathlist != NIL)
5054 : {
5055 : Path *cheapest_partial_path;
5056 :
5057 2412 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5058 :
5059 : /*
5060 : * If cheapest partial path doesn't need a sort, this is redundant
5061 : * with what's already been tried.
5062 : */
5063 2412 : if (!pathkeys_contained_in(root->sort_pathkeys,
5064 : cheapest_partial_path->pathkeys))
5065 : {
5066 : Path *path;
5067 : double total_groups;
5068 :
5069 2404 : path = (Path *) create_sort_path(root,
5070 : ordered_rel,
5071 : cheapest_partial_path,
5072 : root->sort_pathkeys,
5073 : limit_tuples);
5074 :
5075 4808 : total_groups = cheapest_partial_path->rows *
5076 2404 : cheapest_partial_path->parallel_workers;
5077 : path = (Path *)
5078 2404 : create_gather_merge_path(root, ordered_rel,
5079 : path,
5080 : path->pathtarget,
5081 : root->sort_pathkeys, NULL,
5082 : &total_groups);
5083 :
5084 : /* Add projection step if needed */
5085 2404 : if (path->pathtarget != target)
5086 1434 : path = apply_projection_to_path(root, ordered_rel,
5087 : path, target);
5088 :
5089 2404 : add_path(ordered_rel, path);
5090 : }
5091 :
5092 : /*
5093 : * Consider incremental sort with a gather merge on partial paths.
5094 : *
5095 : * We can also skip the entire loop when we only have a single-item
5096 : * sort_pathkeys because then we can't possibly have a presorted
5097 : * prefix of the list without having the list be fully sorted.
5098 : */
5099 2412 : if (enable_incrementalsort && list_length(root->sort_pathkeys) > 1)
5100 : {
5101 : ListCell *lc;
5102 :
5103 888 : foreach(lc, input_rel->partial_pathlist)
5104 : {
5105 456 : Path *input_path = (Path *) lfirst(lc);
5106 456 : Path *sorted_path = input_path;
5107 : bool is_sorted;
5108 : int presorted_keys;
5109 : double total_groups;
5110 :
5111 : /*
5112 : * We don't care if this is the cheapest partial path - we
5113 : * can't simply skip it, because it may be partially sorted in
5114 : * which case we want to consider adding incremental sort
5115 : * (instead of full sort, which is what happens above).
5116 : */
5117 :
5118 456 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5119 : input_path->pathkeys,
5120 : &presorted_keys);
5121 :
5122 : /* No point in adding incremental sort on fully sorted paths. */
5123 456 : if (is_sorted)
5124 424 : continue;
5125 :
5126 456 : if (presorted_keys == 0)
5127 424 : continue;
5128 :
5129 : /* Since we have presorted keys, consider incremental sort. */
5130 32 : sorted_path = (Path *) create_incremental_sort_path(root,
5131 : ordered_rel,
5132 : input_path,
5133 : root->sort_pathkeys,
5134 : presorted_keys,
5135 : limit_tuples);
5136 64 : total_groups = input_path->rows *
5137 32 : input_path->parallel_workers;
5138 : sorted_path = (Path *)
5139 32 : create_gather_merge_path(root, ordered_rel,
5140 : sorted_path,
5141 : sorted_path->pathtarget,
5142 : root->sort_pathkeys, NULL,
5143 : &total_groups);
5144 :
5145 : /* Add projection step if needed */
5146 32 : if (sorted_path->pathtarget != target)
5147 24 : sorted_path = apply_projection_to_path(root, ordered_rel,
5148 : sorted_path, target);
5149 :
5150 32 : add_path(ordered_rel, sorted_path);
5151 : }
5152 : }
5153 : }
5154 :
5155 : /*
5156 : * If there is an FDW that's responsible for all baserels of the query,
5157 : * let it consider adding ForeignPaths.
5158 : */
5159 39162 : if (ordered_rel->fdwroutine &&
5160 320 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5161 306 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5162 : input_rel, ordered_rel,
5163 : NULL);
5164 :
5165 : /* Let extensions possibly add some more paths */
5166 39162 : if (create_upper_paths_hook)
5167 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5168 : input_rel, ordered_rel, NULL);
5169 :
5170 : /*
5171 : * No need to bother with set_cheapest here; grouping_planner does not
5172 : * need us to do it.
5173 : */
5174 : Assert(ordered_rel->pathlist != NIL);
5175 :
5176 39162 : return ordered_rel;
5177 : }
5178 :
5179 :
5180 : /*
5181 : * make_group_input_target
5182 : * Generate appropriate PathTarget for initial input to grouping nodes.
5183 : *
5184 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5185 : * the query's final targetlist; for example, it certainly can't emit any
5186 : * aggregate function calls. This routine generates the correct target
5187 : * for the scan/join subplan.
5188 : *
5189 : * The query target list passed from the parser already contains entries
5190 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5191 : * for variables used only in HAVING clauses; so we need to add those
5192 : * variables to the subplan target list. Also, we flatten all expressions
5193 : * except GROUP BY items into their component variables; other expressions
5194 : * will be computed by the upper plan nodes rather than by the subplan.
5195 : * For example, given a query like
5196 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5197 : * we want to pass this targetlist to the subplan:
5198 : * a+b,c,d
5199 : * where the a+b target will be used by the Sort/Group steps, and the
5200 : * other targets will be used for computing the final results.
5201 : *
5202 : * 'final_target' is the query's final target list (in PathTarget form)
5203 : *
5204 : * The result is the PathTarget to be computed by the Paths returned from
5205 : * query_planner().
5206 : */
5207 : static PathTarget *
5208 25132 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5209 : {
5210 25132 : Query *parse = root->parse;
5211 : PathTarget *input_target;
5212 : List *non_group_cols;
5213 : List *non_group_vars;
5214 : int i;
5215 : ListCell *lc;
5216 :
5217 : /*
5218 : * We must build a target containing all grouping columns, plus any other
5219 : * Vars mentioned in the query's targetlist and HAVING qual.
5220 : */
5221 25132 : input_target = create_empty_pathtarget();
5222 25132 : non_group_cols = NIL;
5223 :
5224 25132 : i = 0;
5225 56344 : foreach(lc, final_target->exprs)
5226 : {
5227 31212 : Expr *expr = (Expr *) lfirst(lc);
5228 31212 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5229 :
5230 35464 : if (sgref && parse->groupClause &&
5231 4252 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5232 : {
5233 : /*
5234 : * It's a grouping column, so add it to the input target as-is.
5235 : */
5236 3368 : add_column_to_pathtarget(input_target, expr, sgref);
5237 : }
5238 : else
5239 : {
5240 : /*
5241 : * Non-grouping column, so just remember the expression for later
5242 : * call to pull_var_clause.
5243 : */
5244 27844 : non_group_cols = lappend(non_group_cols, expr);
5245 : }
5246 :
5247 31212 : i++;
5248 : }
5249 :
5250 : /*
5251 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5252 : */
5253 25132 : if (parse->havingQual)
5254 328 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5255 :
5256 : /*
5257 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5258 : * add them to the input target if not already present. (A Var used
5259 : * directly as a GROUP BY item will be present already.) Note this
5260 : * includes Vars used in resjunk items, so we are covering the needs of
5261 : * ORDER BY and window specifications. Vars used within Aggrefs and
5262 : * WindowFuncs will be pulled out here, too.
5263 : */
5264 25132 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5265 : PVC_RECURSE_AGGREGATES |
5266 : PVC_RECURSE_WINDOWFUNCS |
5267 : PVC_INCLUDE_PLACEHOLDERS);
5268 25132 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5269 :
5270 : /* clean up cruft */
5271 25132 : list_free(non_group_vars);
5272 25132 : list_free(non_group_cols);
5273 :
5274 : /* XXX this causes some redundant cost calculation ... */
5275 25132 : return set_pathtarget_cost_width(root, input_target);
5276 : }
5277 :
5278 : /*
5279 : * make_partial_grouping_target
5280 : * Generate appropriate PathTarget for output of partial aggregate
5281 : * (or partial grouping, if there are no aggregates) nodes.
5282 : *
5283 : * A partial aggregation node needs to emit all the same aggregates that
5284 : * a regular aggregation node would, plus any aggregates used in HAVING;
5285 : * except that the Aggref nodes should be marked as partial aggregates.
5286 : *
5287 : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5288 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5289 : * these would be Vars that are grouped by or used in grouping expressions.)
5290 : *
5291 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5292 : * havingQual represents the HAVING clause.
5293 : */
5294 : static PathTarget *
5295 1310 : make_partial_grouping_target(PlannerInfo *root,
5296 : PathTarget *grouping_target,
5297 : Node *havingQual)
5298 : {
5299 1310 : Query *parse = root->parse;
5300 : PathTarget *partial_target;
5301 : List *non_group_cols;
5302 : List *non_group_exprs;
5303 : int i;
5304 : ListCell *lc;
5305 :
5306 1310 : partial_target = create_empty_pathtarget();
5307 1310 : non_group_cols = NIL;
5308 :
5309 1310 : i = 0;
5310 4756 : foreach(lc, grouping_target->exprs)
5311 : {
5312 3446 : Expr *expr = (Expr *) lfirst(lc);
5313 3446 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5314 :
5315 5886 : if (sgref && parse->groupClause &&
5316 2440 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5317 : {
5318 : /*
5319 : * It's a grouping column, so add it to the partial_target as-is.
5320 : * (This allows the upper agg step to repeat the grouping calcs.)
5321 : */
5322 1208 : add_column_to_pathtarget(partial_target, expr, sgref);
5323 : }
5324 : else
5325 : {
5326 : /*
5327 : * Non-grouping column, so just remember the expression for later
5328 : * call to pull_var_clause.
5329 : */
5330 2238 : non_group_cols = lappend(non_group_cols, expr);
5331 : }
5332 :
5333 3446 : i++;
5334 : }
5335 :
5336 : /*
5337 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5338 : */
5339 1310 : if (havingQual)
5340 552 : non_group_cols = lappend(non_group_cols, havingQual);
5341 :
5342 : /*
5343 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5344 : * non-group cols (plus HAVING), and add them to the partial_target if not
5345 : * already present. (An expression used directly as a GROUP BY item will
5346 : * be present already.) Note this includes Vars used in resjunk items, so
5347 : * we are covering the needs of ORDER BY and window specifications.
5348 : */
5349 1310 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5350 : PVC_INCLUDE_AGGREGATES |
5351 : PVC_RECURSE_WINDOWFUNCS |
5352 : PVC_INCLUDE_PLACEHOLDERS);
5353 :
5354 1310 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5355 :
5356 : /*
5357 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5358 : * are at the top level of the target list, so we can just scan the list
5359 : * rather than recursing through the expression trees.
5360 : */
5361 5164 : foreach(lc, partial_target->exprs)
5362 : {
5363 3854 : Aggref *aggref = (Aggref *) lfirst(lc);
5364 :
5365 3854 : if (IsA(aggref, Aggref))
5366 : {
5367 : Aggref *newaggref;
5368 :
5369 : /*
5370 : * We shouldn't need to copy the substructure of the Aggref node,
5371 : * but flat-copy the node itself to avoid damaging other trees.
5372 : */
5373 2646 : newaggref = makeNode(Aggref);
5374 2646 : memcpy(newaggref, aggref, sizeof(Aggref));
5375 :
5376 : /* For now, assume serialization is required */
5377 2646 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5378 :
5379 2646 : lfirst(lc) = newaggref;
5380 : }
5381 : }
5382 :
5383 : /* clean up cruft */
5384 1310 : list_free(non_group_exprs);
5385 1310 : list_free(non_group_cols);
5386 :
5387 : /* XXX this causes some redundant cost calculation ... */
5388 1310 : return set_pathtarget_cost_width(root, partial_target);
5389 : }
5390 :
5391 : /*
5392 : * mark_partial_aggref
5393 : * Adjust an Aggref to make it represent a partial-aggregation step.
5394 : *
5395 : * The Aggref node is modified in-place; caller must do any copying required.
5396 : */
5397 : void
5398 4270 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5399 : {
5400 : /* aggtranstype should be computed by this point */
5401 : Assert(OidIsValid(agg->aggtranstype));
5402 : /* ... but aggsplit should still be as the parser left it */
5403 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5404 :
5405 : /* Mark the Aggref with the intended partial-aggregation mode */
5406 4270 : agg->aggsplit = aggsplit;
5407 :
5408 : /*
5409 : * Adjust result type if needed. Normally, a partial aggregate returns
5410 : * the aggregate's transition type; but if that's INTERNAL and we're
5411 : * serializing, it returns BYTEA instead.
5412 : */
5413 4270 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5414 : {
5415 3458 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5416 48 : agg->aggtype = BYTEAOID;
5417 : else
5418 3410 : agg->aggtype = agg->aggtranstype;
5419 : }
5420 4270 : }
5421 :
5422 : /*
5423 : * postprocess_setop_tlist
5424 : * Fix up targetlist returned by plan_set_operations().
5425 : *
5426 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5427 : * NOTE: this would not be good enough if we supported resjunk sort keys
5428 : * for results of set operations --- then, we'd need to project a whole
5429 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5430 : * find any resjunk columns in orig_tlist.
5431 : */
5432 : static List *
5433 1354 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5434 : {
5435 : ListCell *l;
5436 1354 : ListCell *orig_tlist_item = list_head(orig_tlist);
5437 :
5438 4310 : foreach(l, new_tlist)
5439 : {
5440 2956 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5441 : TargetEntry *orig_tle;
5442 :
5443 : /* ignore resjunk columns in setop result */
5444 2956 : if (new_tle->resjunk)
5445 208 : continue;
5446 :
5447 : Assert(orig_tlist_item != NULL);
5448 2748 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5449 2748 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5450 2748 : if (orig_tle->resjunk) /* should not happen */
5451 0 : elog(ERROR, "resjunk output columns are not implemented");
5452 : Assert(new_tle->resno == orig_tle->resno);
5453 2748 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5454 : }
5455 1354 : if (orig_tlist_item != NULL)
5456 0 : elog(ERROR, "resjunk output columns are not implemented");
5457 1354 : return new_tlist;
5458 : }
5459 :
5460 : /*
5461 : * select_active_windows
5462 : * Create a list of the "active" window clauses (ie, those referenced
5463 : * by non-deleted WindowFuncs) in the order they are to be executed.
5464 : */
5465 : static List *
5466 1112 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5467 : {
5468 1112 : List *windowClause = root->parse->windowClause;
5469 1112 : List *result = NIL;
5470 : ListCell *lc;
5471 1112 : int nActive = 0;
5472 1112 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5473 1112 : * list_length(windowClause));
5474 :
5475 : /* First, construct an array of the active windows */
5476 2288 : foreach(lc, windowClause)
5477 : {
5478 1176 : WindowClause *wc = lfirst_node(WindowClause, lc);
5479 :
5480 : /* It's only active if wflists shows some related WindowFuncs */
5481 : Assert(wc->winref <= wflists->maxWinRef);
5482 1176 : if (wflists->windowFuncs[wc->winref] == NIL)
5483 16 : continue;
5484 :
5485 1160 : actives[nActive].wc = wc; /* original clause */
5486 :
5487 : /*
5488 : * For sorting, we want the list of partition keys followed by the
5489 : * list of sort keys. But pathkeys construction will remove duplicates
5490 : * between the two, so we can as well (even though we can't detect all
5491 : * of the duplicates, since some may come from ECs - that might mean
5492 : * we miss optimization chances here). We must, however, ensure that
5493 : * the order of entries is preserved with respect to the ones we do
5494 : * keep.
5495 : *
5496 : * partitionClause and orderClause had their own duplicates removed in
5497 : * parse analysis, so we're only concerned here with removing
5498 : * orderClause entries that also appear in partitionClause.
5499 : */
5500 2320 : actives[nActive].uniqueOrder =
5501 1160 : list_concat_unique(list_copy(wc->partitionClause),
5502 1160 : wc->orderClause);
5503 1160 : nActive++;
5504 : }
5505 :
5506 : /*
5507 : * Sort active windows by their partitioning/ordering clauses, ignoring
5508 : * any framing clauses, so that the windows that need the same sorting are
5509 : * adjacent in the list. When we come to generate paths, this will avoid
5510 : * inserting additional Sort nodes.
5511 : *
5512 : * This is how we implement a specific requirement from the SQL standard,
5513 : * which says that when two or more windows are order-equivalent (i.e.
5514 : * have matching partition and order clauses, even if their names or
5515 : * framing clauses differ), then all peer rows must be presented in the
5516 : * same order in all of them. If we allowed multiple sort nodes for such
5517 : * cases, we'd risk having the peer rows end up in different orders in
5518 : * equivalent windows due to sort instability. (See General Rule 4 of
5519 : * <window clause> in SQL2008 - SQL2016.)
5520 : *
5521 : * Additionally, if the entire list of clauses of one window is a prefix
5522 : * of another, put first the window with stronger sorting requirements.
5523 : * This way we will first sort for stronger window, and won't have to sort
5524 : * again for the weaker one.
5525 : */
5526 1112 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5527 :
5528 : /* build ordered list of the original WindowClause nodes */
5529 2272 : for (int i = 0; i < nActive; i++)
5530 1160 : result = lappend(result, actives[i].wc);
5531 :
5532 1112 : pfree(actives);
5533 :
5534 1112 : return result;
5535 : }
5536 :
5537 : /*
5538 : * common_prefix_cmp
5539 : * QSort comparison function for WindowClauseSortData
5540 : *
5541 : * Sort the windows by the required sorting clauses. First, compare the sort
5542 : * clauses themselves. Second, if one window's clauses are a prefix of another
5543 : * one's clauses, put the window with more sort clauses first.
5544 : */
5545 : static int
5546 48 : common_prefix_cmp(const void *a, const void *b)
5547 : {
5548 48 : const WindowClauseSortData *wcsa = a;
5549 48 : const WindowClauseSortData *wcsb = b;
5550 : ListCell *item_a;
5551 : ListCell *item_b;
5552 :
5553 84 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
5554 : {
5555 56 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
5556 56 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
5557 :
5558 56 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
5559 20 : return -1;
5560 56 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
5561 12 : return 1;
5562 44 : else if (sca->sortop > scb->sortop)
5563 0 : return -1;
5564 44 : else if (sca->sortop < scb->sortop)
5565 8 : return 1;
5566 36 : else if (sca->nulls_first && !scb->nulls_first)
5567 0 : return -1;
5568 36 : else if (!sca->nulls_first && scb->nulls_first)
5569 0 : return 1;
5570 : /* no need to compare eqop, since it is fully determined by sortop */
5571 : }
5572 :
5573 28 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
5574 4 : return -1;
5575 24 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
5576 8 : return 1;
5577 :
5578 16 : return 0;
5579 : }
5580 :
5581 : /*
5582 : * make_window_input_target
5583 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
5584 : *
5585 : * When the query has window functions, this function computes the desired
5586 : * target to be computed by the node just below the first WindowAgg.
5587 : * This tlist must contain all values needed to evaluate the window functions,
5588 : * compute the final target list, and perform any required final sort step.
5589 : * If multiple WindowAggs are needed, each intermediate one adds its window
5590 : * function results onto this base tlist; only the topmost WindowAgg computes
5591 : * the actual desired target list.
5592 : *
5593 : * This function is much like make_group_input_target, though not quite enough
5594 : * like it to share code. As in that function, we flatten most expressions
5595 : * into their component variables. But we do not want to flatten window
5596 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
5597 : * evaluations of them, which would be bad (possibly even resulting in
5598 : * inconsistent answers, if they contain volatile functions).
5599 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
5600 : * make_group_input_target, because we may no longer have access to the
5601 : * individual Vars in them.
5602 : *
5603 : * Another key difference from make_group_input_target is that we don't
5604 : * flatten Aggref expressions, since those are to be computed below the
5605 : * window functions and just referenced like Vars above that.
5606 : *
5607 : * 'final_target' is the query's final target list (in PathTarget form)
5608 : * 'activeWindows' is the list of active windows previously identified by
5609 : * select_active_windows.
5610 : *
5611 : * The result is the PathTarget to be computed by the plan node immediately
5612 : * below the first WindowAgg node.
5613 : */
5614 : static PathTarget *
5615 1112 : make_window_input_target(PlannerInfo *root,
5616 : PathTarget *final_target,
5617 : List *activeWindows)
5618 : {
5619 1112 : Query *parse = root->parse;
5620 : PathTarget *input_target;
5621 : Bitmapset *sgrefs;
5622 : List *flattenable_cols;
5623 : List *flattenable_vars;
5624 : int i;
5625 : ListCell *lc;
5626 :
5627 : Assert(parse->hasWindowFuncs);
5628 :
5629 : /*
5630 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
5631 : * into a bitmapset for convenient reference below.
5632 : */
5633 1112 : sgrefs = NULL;
5634 2272 : foreach(lc, activeWindows)
5635 : {
5636 1160 : WindowClause *wc = lfirst_node(WindowClause, lc);
5637 : ListCell *lc2;
5638 :
5639 1512 : foreach(lc2, wc->partitionClause)
5640 : {
5641 352 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5642 :
5643 352 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5644 : }
5645 2216 : foreach(lc2, wc->orderClause)
5646 : {
5647 1056 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5648 :
5649 1056 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5650 : }
5651 : }
5652 :
5653 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
5654 1224 : foreach(lc, parse->groupClause)
5655 : {
5656 112 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
5657 :
5658 112 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
5659 : }
5660 :
5661 : /*
5662 : * Construct a target containing all the non-flattenable targetlist items,
5663 : * and save aside the others for a moment.
5664 : */
5665 1112 : input_target = create_empty_pathtarget();
5666 1112 : flattenable_cols = NIL;
5667 :
5668 1112 : i = 0;
5669 4752 : foreach(lc, final_target->exprs)
5670 : {
5671 3640 : Expr *expr = (Expr *) lfirst(lc);
5672 3640 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5673 :
5674 : /*
5675 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
5676 : * that such items can't contain window functions, so it's okay to
5677 : * compute them below the WindowAgg nodes.)
5678 : */
5679 3640 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
5680 : {
5681 : /*
5682 : * Don't want to deconstruct this value, so add it to the input
5683 : * target as-is.
5684 : */
5685 1368 : add_column_to_pathtarget(input_target, expr, sgref);
5686 : }
5687 : else
5688 : {
5689 : /*
5690 : * Column is to be flattened, so just remember the expression for
5691 : * later call to pull_var_clause.
5692 : */
5693 2272 : flattenable_cols = lappend(flattenable_cols, expr);
5694 : }
5695 :
5696 3640 : i++;
5697 : }
5698 :
5699 : /*
5700 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
5701 : * add them to the input target if not already present. (Some might be
5702 : * there already because they're used directly as window/group clauses.)
5703 : *
5704 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
5705 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
5706 : * at higher levels. On the other hand, we should recurse into
5707 : * WindowFuncs to make sure their input expressions are available.
5708 : */
5709 1112 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
5710 : PVC_INCLUDE_AGGREGATES |
5711 : PVC_RECURSE_WINDOWFUNCS |
5712 : PVC_INCLUDE_PLACEHOLDERS);
5713 1112 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
5714 :
5715 : /* clean up cruft */
5716 1112 : list_free(flattenable_vars);
5717 1112 : list_free(flattenable_cols);
5718 :
5719 : /* XXX this causes some redundant cost calculation ... */
5720 1112 : return set_pathtarget_cost_width(root, input_target);
5721 : }
5722 :
5723 : /*
5724 : * make_pathkeys_for_window
5725 : * Create a pathkeys list describing the required input ordering
5726 : * for the given WindowClause.
5727 : *
5728 : * The required ordering is first the PARTITION keys, then the ORDER keys.
5729 : * In the future we might try to implement windowing using hashing, in which
5730 : * case the ordering could be relaxed, but for now we always sort.
5731 : */
5732 : static List *
5733 2316 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
5734 : List *tlist)
5735 : {
5736 : List *window_pathkeys;
5737 : List *window_sortclauses;
5738 :
5739 : /* Throw error if can't sort */
5740 2316 : if (!grouping_is_sortable(wc->partitionClause))
5741 0 : ereport(ERROR,
5742 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5743 : errmsg("could not implement window PARTITION BY"),
5744 : errdetail("Window partitioning columns must be of sortable datatypes.")));
5745 2316 : if (!grouping_is_sortable(wc->orderClause))
5746 0 : ereport(ERROR,
5747 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5748 : errmsg("could not implement window ORDER BY"),
5749 : errdetail("Window ordering columns must be of sortable datatypes.")));
5750 :
5751 : /* Okay, make the combined pathkeys */
5752 2316 : window_sortclauses = list_concat_copy(wc->partitionClause, wc->orderClause);
5753 2316 : window_pathkeys = make_pathkeys_for_sortclauses(root,
5754 : window_sortclauses,
5755 : tlist);
5756 2316 : list_free(window_sortclauses);
5757 2316 : return window_pathkeys;
5758 : }
5759 :
5760 : /*
5761 : * make_sort_input_target
5762 : * Generate appropriate PathTarget for initial input to Sort step.
5763 : *
5764 : * If the query has ORDER BY, this function chooses the target to be computed
5765 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
5766 : * project) steps. This might or might not be identical to the query's final
5767 : * output target.
5768 : *
5769 : * The main argument for keeping the sort-input tlist the same as the final
5770 : * is that we avoid a separate projection node (which will be needed if
5771 : * they're different, because Sort can't project). However, there are also
5772 : * advantages to postponing tlist evaluation till after the Sort: it ensures
5773 : * a consistent order of evaluation for any volatile functions in the tlist,
5774 : * and if there's also a LIMIT, we can stop the query without ever computing
5775 : * tlist functions for later rows, which is beneficial for both volatile and
5776 : * expensive functions.
5777 : *
5778 : * Our current policy is to postpone volatile expressions till after the sort
5779 : * unconditionally (assuming that that's possible, ie they are in plain tlist
5780 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
5781 : * postpone set-returning expressions, because running them beforehand would
5782 : * bloat the sort dataset, and because it might cause unexpected output order
5783 : * if the sort isn't stable. However there's a constraint on that: all SRFs
5784 : * in the tlist should be evaluated at the same plan step, so that they can
5785 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
5786 : * mustn't postpone any SRFs. (Note that in principle that policy should
5787 : * probably get applied to the group/window input targetlists too, but we
5788 : * have not done that historically.) Lastly, expensive expressions are
5789 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
5790 : * partial evaluation of the query is possible (if neither is true, we expect
5791 : * to have to evaluate the expressions for every row anyway), or if there are
5792 : * any volatile or set-returning expressions (since once we've put in a
5793 : * projection at all, it won't cost any more to postpone more stuff).
5794 : *
5795 : * Another issue that could potentially be considered here is that
5796 : * evaluating tlist expressions could result in data that's either wider
5797 : * or narrower than the input Vars, thus changing the volume of data that
5798 : * has to go through the Sort. However, we usually have only a very bad
5799 : * idea of the output width of any expression more complex than a Var,
5800 : * so for now it seems too risky to try to optimize on that basis.
5801 : *
5802 : * Note that if we do produce a modified sort-input target, and then the
5803 : * query ends up not using an explicit Sort, no particular harm is done:
5804 : * we'll initially use the modified target for the preceding path nodes,
5805 : * but then change them to the final target with apply_projection_to_path.
5806 : * Moreover, in such a case the guarantees about evaluation order of
5807 : * volatile functions still hold, since the rows are sorted already.
5808 : *
5809 : * This function has some things in common with make_group_input_target and
5810 : * make_window_input_target, though the detailed rules for what to do are
5811 : * different. We never flatten/postpone any grouping or ordering columns;
5812 : * those are needed before the sort. If we do flatten a particular
5813 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
5814 : * computed earlier.
5815 : *
5816 : * 'final_target' is the query's final target list (in PathTarget form)
5817 : * 'have_postponed_srfs' is an output argument, see below
5818 : *
5819 : * The result is the PathTarget to be computed by the plan node immediately
5820 : * below the Sort step (and the Distinct step, if any). This will be
5821 : * exactly final_target if we decide a projection step wouldn't be helpful.
5822 : *
5823 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
5824 : * any set-returning functions to after the Sort.
5825 : */
5826 : static PathTarget *
5827 38790 : make_sort_input_target(PlannerInfo *root,
5828 : PathTarget *final_target,
5829 : bool *have_postponed_srfs)
5830 : {
5831 38790 : Query *parse = root->parse;
5832 : PathTarget *input_target;
5833 : int ncols;
5834 : bool *col_is_srf;
5835 : bool *postpone_col;
5836 : bool have_srf;
5837 : bool have_volatile;
5838 : bool have_expensive;
5839 : bool have_srf_sortcols;
5840 : bool postpone_srfs;
5841 : List *postponable_cols;
5842 : List *postponable_vars;
5843 : int i;
5844 : ListCell *lc;
5845 :
5846 : /* Shouldn't get here unless query has ORDER BY */
5847 : Assert(parse->sortClause);
5848 :
5849 38790 : *have_postponed_srfs = false; /* default result */
5850 :
5851 : /* Inspect tlist and collect per-column information */
5852 38790 : ncols = list_length(final_target->exprs);
5853 38790 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
5854 38790 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
5855 38790 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
5856 :
5857 38790 : i = 0;
5858 299168 : foreach(lc, final_target->exprs)
5859 : {
5860 260378 : Expr *expr = (Expr *) lfirst(lc);
5861 :
5862 : /*
5863 : * If the column has a sortgroupref, assume it has to be evaluated
5864 : * before sorting. Generally such columns would be ORDER BY, GROUP
5865 : * BY, etc targets. One exception is columns that were removed from
5866 : * GROUP BY by remove_useless_groupby_columns() ... but those would
5867 : * only be Vars anyway. There don't seem to be any cases where it
5868 : * would be worth the trouble to double-check.
5869 : */
5870 260378 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
5871 : {
5872 : /*
5873 : * Check for SRF or volatile functions. Check the SRF case first
5874 : * because we must know whether we have any postponed SRFs.
5875 : */
5876 423192 : if (parse->hasTargetSRFs &&
5877 136 : expression_returns_set((Node *) expr))
5878 : {
5879 : /* We'll decide below whether these are postponable */
5880 60 : col_is_srf[i] = true;
5881 60 : have_srf = true;
5882 : }
5883 211468 : else if (contain_volatile_functions((Node *) expr))
5884 : {
5885 : /* Unconditionally postpone */
5886 64 : postpone_col[i] = true;
5887 64 : have_volatile = true;
5888 : }
5889 : else
5890 : {
5891 : /*
5892 : * Else check the cost. XXX it's annoying to have to do this
5893 : * when set_pathtarget_cost_width() just did it. Refactor to
5894 : * allow sharing the work?
5895 : */
5896 : QualCost cost;
5897 :
5898 211404 : cost_qual_eval_node(&cost, (Node *) expr, root);
5899 :
5900 : /*
5901 : * We arbitrarily define "expensive" as "more than 10X
5902 : * cpu_operator_cost". Note this will take in any PL function
5903 : * with default cost.
5904 : */
5905 211404 : if (cost.per_tuple > 10 * cpu_operator_cost)
5906 : {
5907 19582 : postpone_col[i] = true;
5908 19582 : have_expensive = true;
5909 : }
5910 : }
5911 : }
5912 : else
5913 : {
5914 : /* For sortgroupref cols, just check if any contain SRFs */
5915 48850 : if (!have_srf_sortcols &&
5916 49050 : parse->hasTargetSRFs &&
5917 200 : expression_returns_set((Node *) expr))
5918 80 : have_srf_sortcols = true;
5919 : }
5920 :
5921 260378 : i++;
5922 : }
5923 :
5924 : /*
5925 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
5926 : */
5927 38790 : postpone_srfs = (have_srf && !have_srf_sortcols);
5928 :
5929 : /*
5930 : * If we don't need a post-sort projection, just return final_target.
5931 : */
5932 38790 : if (!(postpone_srfs || have_volatile ||
5933 38686 : (have_expensive &&
5934 12110 : (parse->limitCount || root->tuple_fraction > 0))))
5935 38678 : return final_target;
5936 :
5937 : /*
5938 : * Report whether the post-sort projection will contain set-returning
5939 : * functions. This is important because it affects whether the Sort can
5940 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
5941 : * to return.
5942 : */
5943 112 : *have_postponed_srfs = postpone_srfs;
5944 :
5945 : /*
5946 : * Construct the sort-input target, taking all non-postponable columns and
5947 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
5948 : * the postponable ones.
5949 : */
5950 112 : input_target = create_empty_pathtarget();
5951 112 : postponable_cols = NIL;
5952 :
5953 112 : i = 0;
5954 704 : foreach(lc, final_target->exprs)
5955 : {
5956 592 : Expr *expr = (Expr *) lfirst(lc);
5957 :
5958 592 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
5959 116 : postponable_cols = lappend(postponable_cols, expr);
5960 : else
5961 476 : add_column_to_pathtarget(input_target, expr,
5962 476 : get_pathtarget_sortgroupref(final_target, i));
5963 :
5964 592 : i++;
5965 : }
5966 :
5967 : /*
5968 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
5969 : * postponable columns, and add them to the sort-input target if not
5970 : * already present. (Some might be there already.) We mustn't
5971 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
5972 : * would be unable to recompute them.
5973 : */
5974 112 : postponable_vars = pull_var_clause((Node *) postponable_cols,
5975 : PVC_INCLUDE_AGGREGATES |
5976 : PVC_INCLUDE_WINDOWFUNCS |
5977 : PVC_INCLUDE_PLACEHOLDERS);
5978 112 : add_new_columns_to_pathtarget(input_target, postponable_vars);
5979 :
5980 : /* clean up cruft */
5981 112 : list_free(postponable_vars);
5982 112 : list_free(postponable_cols);
5983 :
5984 : /* XXX this represents even more redundant cost calculation ... */
5985 112 : return set_pathtarget_cost_width(root, input_target);
5986 : }
5987 :
5988 : /*
5989 : * get_cheapest_fractional_path
5990 : * Find the cheapest path for retrieving a specified fraction of all
5991 : * the tuples expected to be returned by the given relation.
5992 : *
5993 : * We interpret tuple_fraction the same way as grouping_planner.
5994 : *
5995 : * We assume set_cheapest() has been run on the given rel.
5996 : */
5997 : Path *
5998 309418 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
5999 : {
6000 309418 : Path *best_path = rel->cheapest_total_path;
6001 : ListCell *l;
6002 :
6003 : /* If all tuples will be retrieved, just return the cheapest-total path */
6004 309418 : if (tuple_fraction <= 0.0)
6005 305430 : return best_path;
6006 :
6007 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6008 3988 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6009 1676 : tuple_fraction /= best_path->rows;
6010 :
6011 9806 : foreach(l, rel->pathlist)
6012 : {
6013 5818 : Path *path = (Path *) lfirst(l);
6014 :
6015 7648 : if (path == rel->cheapest_total_path ||
6016 1830 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6017 5338 : continue;
6018 :
6019 480 : best_path = path;
6020 : }
6021 :
6022 3988 : return best_path;
6023 : }
6024 :
6025 : /*
6026 : * adjust_paths_for_srfs
6027 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6028 : *
6029 : * The executor can only handle set-returning functions that appear at the
6030 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6031 : * that are not at top level, we need to split up the evaluation into multiple
6032 : * plan levels in which each level satisfies this constraint. This function
6033 : * modifies each Path of an upperrel that (might) compute any SRFs in its
6034 : * output tlist to insert appropriate projection steps.
6035 : *
6036 : * The given targets and targets_contain_srfs lists are from
6037 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6038 : * target in targets.
6039 : */
6040 : static void
6041 4096 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6042 : List *targets, List *targets_contain_srfs)
6043 : {
6044 : ListCell *lc;
6045 :
6046 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6047 : Assert(!linitial_int(targets_contain_srfs));
6048 :
6049 : /* If no SRFs appear at this plan level, nothing to do */
6050 4096 : if (list_length(targets) == 1)
6051 216 : return;
6052 :
6053 : /*
6054 : * Stack SRF-evaluation nodes atop each path for the rel.
6055 : *
6056 : * In principle we should re-run set_cheapest() here to identify the
6057 : * cheapest path, but it seems unlikely that adding the same tlist eval
6058 : * costs to all the paths would change that, so we don't bother. Instead,
6059 : * just assume that the cheapest-startup and cheapest-total paths remain
6060 : * so. (There should be no parameterized paths anymore, so we needn't
6061 : * worry about updating cheapest_parameterized_paths.)
6062 : */
6063 7784 : foreach(lc, rel->pathlist)
6064 : {
6065 3904 : Path *subpath = (Path *) lfirst(lc);
6066 3904 : Path *newpath = subpath;
6067 : ListCell *lc1,
6068 : *lc2;
6069 :
6070 : Assert(subpath->param_info == NULL);
6071 12388 : forboth(lc1, targets, lc2, targets_contain_srfs)
6072 : {
6073 8484 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6074 8484 : bool contains_srfs = (bool) lfirst_int(lc2);
6075 :
6076 : /* If this level doesn't contain SRFs, do regular projection */
6077 8484 : if (contains_srfs)
6078 3944 : newpath = (Path *) create_set_projection_path(root,
6079 : rel,
6080 : newpath,
6081 : thistarget);
6082 : else
6083 4540 : newpath = (Path *) apply_projection_to_path(root,
6084 : rel,
6085 : newpath,
6086 : thistarget);
6087 : }
6088 3904 : lfirst(lc) = newpath;
6089 3904 : if (subpath == rel->cheapest_startup_path)
6090 52 : rel->cheapest_startup_path = newpath;
6091 3904 : if (subpath == rel->cheapest_total_path)
6092 52 : rel->cheapest_total_path = newpath;
6093 : }
6094 :
6095 : /* Likewise for partial paths, if any */
6096 3884 : foreach(lc, rel->partial_pathlist)
6097 : {
6098 4 : Path *subpath = (Path *) lfirst(lc);
6099 4 : Path *newpath = subpath;
6100 : ListCell *lc1,
6101 : *lc2;
6102 :
6103 : Assert(subpath->param_info == NULL);
6104 16 : forboth(lc1, targets, lc2, targets_contain_srfs)
6105 : {
6106 12 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6107 12 : bool contains_srfs = (bool) lfirst_int(lc2);
6108 :
6109 : /* If this level doesn't contain SRFs, do regular projection */
6110 12 : if (contains_srfs)
6111 4 : newpath = (Path *) create_set_projection_path(root,
6112 : rel,
6113 : newpath,
6114 : thistarget);
6115 : else
6116 : {
6117 : /* avoid apply_projection_to_path, in case of multiple refs */
6118 8 : newpath = (Path *) create_projection_path(root,
6119 : rel,
6120 : newpath,
6121 : thistarget);
6122 : }
6123 : }
6124 4 : lfirst(lc) = newpath;
6125 : }
6126 : }
6127 :
6128 : /*
6129 : * expression_planner
6130 : * Perform planner's transformations on a standalone expression.
6131 : *
6132 : * Various utility commands need to evaluate expressions that are not part
6133 : * of a plannable query. They can do so using the executor's regular
6134 : * expression-execution machinery, but first the expression has to be fed
6135 : * through here to transform it from parser output to something executable.
6136 : *
6137 : * Currently, we disallow sublinks in standalone expressions, so there's no
6138 : * real "planning" involved here. (That might not always be true though.)
6139 : * What we must do is run eval_const_expressions to ensure that any function
6140 : * calls are converted to positional notation and function default arguments
6141 : * get inserted. The fact that constant subexpressions get simplified is a
6142 : * side-effect that is useful when the expression will get evaluated more than
6143 : * once. Also, we must fix operator function IDs.
6144 : *
6145 : * This does not return any information about dependencies of the expression.
6146 : * Hence callers should use the results only for the duration of the current
6147 : * query. Callers that would like to cache the results for longer should use
6148 : * expression_planner_with_deps, probably via the plancache.
6149 : *
6150 : * Note: this must not make any damaging changes to the passed-in expression
6151 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6152 : * we first do an expression_tree_mutator-based walk, what is returned will
6153 : * be a new node tree.) The result is constructed in the current memory
6154 : * context; beware that this can leak a lot of additional stuff there, too.
6155 : */
6156 : Expr *
6157 90804 : expression_planner(Expr *expr)
6158 : {
6159 : Node *result;
6160 :
6161 : /*
6162 : * Convert named-argument function calls, insert default arguments and
6163 : * simplify constant subexprs
6164 : */
6165 90804 : result = eval_const_expressions(NULL, (Node *) expr);
6166 :
6167 : /* Fill in opfuncid values if missing */
6168 90804 : fix_opfuncids(result);
6169 :
6170 90804 : return (Expr *) result;
6171 : }
6172 :
6173 : /*
6174 : * expression_planner_with_deps
6175 : * Perform planner's transformations on a standalone expression,
6176 : * returning expression dependency information along with the result.
6177 : *
6178 : * This is identical to expression_planner() except that it also returns
6179 : * information about possible dependencies of the expression, ie identities of
6180 : * objects whose definitions affect the result. As in a PlannedStmt, these
6181 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6182 : */
6183 : Expr *
6184 474 : expression_planner_with_deps(Expr *expr,
6185 : List **relationOids,
6186 : List **invalItems)
6187 : {
6188 : Node *result;
6189 : PlannerGlobal glob;
6190 : PlannerInfo root;
6191 :
6192 : /* Make up dummy planner state so we can use setrefs machinery */
6193 8532 : MemSet(&glob, 0, sizeof(glob));
6194 474 : glob.type = T_PlannerGlobal;
6195 474 : glob.relationOids = NIL;
6196 474 : glob.invalItems = NIL;
6197 :
6198 32232 : MemSet(&root, 0, sizeof(root));
6199 474 : root.type = T_PlannerInfo;
6200 474 : root.glob = &glob;
6201 :
6202 : /*
6203 : * Convert named-argument function calls, insert default arguments and
6204 : * simplify constant subexprs. Collect identities of inlined functions
6205 : * and elided domains, too.
6206 : */
6207 474 : result = eval_const_expressions(&root, (Node *) expr);
6208 :
6209 : /* Fill in opfuncid values if missing */
6210 474 : fix_opfuncids(result);
6211 :
6212 : /*
6213 : * Now walk the finished expression to find anything else we ought to
6214 : * record as an expression dependency.
6215 : */
6216 474 : (void) extract_query_dependencies_walker(result, &root);
6217 :
6218 474 : *relationOids = glob.relationOids;
6219 474 : *invalItems = glob.invalItems;
6220 :
6221 474 : return (Expr *) result;
6222 : }
6223 :
6224 :
6225 : /*
6226 : * plan_cluster_use_sort
6227 : * Use the planner to decide how CLUSTER should implement sorting
6228 : *
6229 : * tableOid is the OID of a table to be clustered on its index indexOid
6230 : * (which is already known to be a btree index). Decide whether it's
6231 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6232 : * Return true to use sorting, false to use an indexscan.
6233 : *
6234 : * Note: caller had better already hold some type of lock on the table.
6235 : */
6236 : bool
6237 74 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6238 : {
6239 : PlannerInfo *root;
6240 : Query *query;
6241 : PlannerGlobal *glob;
6242 : RangeTblEntry *rte;
6243 : RelOptInfo *rel;
6244 : IndexOptInfo *indexInfo;
6245 : QualCost indexExprCost;
6246 : Cost comparisonCost;
6247 : Path *seqScanPath;
6248 : Path seqScanAndSortPath;
6249 : IndexPath *indexScanPath;
6250 : ListCell *lc;
6251 :
6252 : /* We can short-circuit the cost comparison if indexscans are disabled */
6253 74 : if (!enable_indexscan)
6254 20 : return true; /* use sort */
6255 :
6256 : /* Set up mostly-dummy planner state */
6257 54 : query = makeNode(Query);
6258 54 : query->commandType = CMD_SELECT;
6259 :
6260 54 : glob = makeNode(PlannerGlobal);
6261 :
6262 54 : root = makeNode(PlannerInfo);
6263 54 : root->parse = query;
6264 54 : root->glob = glob;
6265 54 : root->query_level = 1;
6266 54 : root->planner_cxt = CurrentMemoryContext;
6267 54 : root->wt_param_id = -1;
6268 :
6269 : /* Build a minimal RTE for the rel */
6270 54 : rte = makeNode(RangeTblEntry);
6271 54 : rte->rtekind = RTE_RELATION;
6272 54 : rte->relid = tableOid;
6273 54 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6274 54 : rte->rellockmode = AccessShareLock;
6275 54 : rte->lateral = false;
6276 54 : rte->inh = false;
6277 54 : rte->inFromCl = true;
6278 54 : query->rtable = list_make1(rte);
6279 :
6280 : /* Set up RTE/RelOptInfo arrays */
6281 54 : setup_simple_rel_arrays(root);
6282 :
6283 : /* Build RelOptInfo */
6284 54 : rel = build_simple_rel(root, 1, NULL);
6285 :
6286 : /* Locate IndexOptInfo for the target index */
6287 54 : indexInfo = NULL;
6288 86 : foreach(lc, rel->indexlist)
6289 : {
6290 86 : indexInfo = lfirst_node(IndexOptInfo, lc);
6291 86 : if (indexInfo->indexoid == indexOid)
6292 54 : break;
6293 : }
6294 :
6295 : /*
6296 : * It's possible that get_relation_info did not generate an IndexOptInfo
6297 : * for the desired index; this could happen if it's not yet reached its
6298 : * indcheckxmin usability horizon, or if it's a system index and we're
6299 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6300 : * trust the index contents but use seqscan-and-sort.
6301 : */
6302 54 : if (lc == NULL) /* not in the list? */
6303 0 : return true; /* use sort */
6304 :
6305 : /*
6306 : * Rather than doing all the pushups that would be needed to use
6307 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6308 : */
6309 54 : rel->rows = rel->tuples;
6310 54 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6311 :
6312 54 : root->total_table_pages = rel->pages;
6313 :
6314 : /*
6315 : * Determine eval cost of the index expressions, if any. We need to
6316 : * charge twice that amount for each tuple comparison that happens during
6317 : * the sort, since tuplesort.c will have to re-evaluate the index
6318 : * expressions each time. (XXX that's pretty inefficient...)
6319 : */
6320 54 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6321 54 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6322 :
6323 : /* Estimate the cost of seq scan + sort */
6324 54 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6325 108 : cost_sort(&seqScanAndSortPath, root, NIL,
6326 54 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6327 : comparisonCost, maintenance_work_mem, -1.0);
6328 :
6329 : /* Estimate the cost of index scan */
6330 54 : indexScanPath = create_index_path(root, indexInfo,
6331 : NIL, NIL, NIL, NIL,
6332 : ForwardScanDirection, false,
6333 : NULL, 1.0, false);
6334 :
6335 54 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6336 : }
6337 :
6338 : /*
6339 : * plan_create_index_workers
6340 : * Use the planner to decide how many parallel worker processes
6341 : * CREATE INDEX should request for use
6342 : *
6343 : * tableOid is the table on which the index is to be built. indexOid is the
6344 : * OID of an index to be created or reindexed (which must be a btree index).
6345 : *
6346 : * Return value is the number of parallel worker processes to request. It
6347 : * may be unsafe to proceed if this is 0. Note that this does not include the
6348 : * leader participating as a worker (value is always a number of parallel
6349 : * worker processes).
6350 : *
6351 : * Note: caller had better already hold some type of lock on the table and
6352 : * index.
6353 : */
6354 : int
6355 22698 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6356 : {
6357 : PlannerInfo *root;
6358 : Query *query;
6359 : PlannerGlobal *glob;
6360 : RangeTblEntry *rte;
6361 : Relation heap;
6362 : Relation index;
6363 : RelOptInfo *rel;
6364 : int parallel_workers;
6365 : BlockNumber heap_blocks;
6366 : double reltuples;
6367 : double allvisfrac;
6368 :
6369 : /* Return immediately when parallelism disabled */
6370 22698 : if (max_parallel_maintenance_workers == 0)
6371 80 : return 0;
6372 :
6373 : /* Set up largely-dummy planner state */
6374 22618 : query = makeNode(Query);
6375 22618 : query->commandType = CMD_SELECT;
6376 :
6377 22618 : glob = makeNode(PlannerGlobal);
6378 :
6379 22618 : root = makeNode(PlannerInfo);
6380 22618 : root->parse = query;
6381 22618 : root->glob = glob;
6382 22618 : root->query_level = 1;
6383 22618 : root->planner_cxt = CurrentMemoryContext;
6384 22618 : root->wt_param_id = -1;
6385 :
6386 : /*
6387 : * Build a minimal RTE.
6388 : *
6389 : * Mark the RTE with inh = true. This is a kludge to prevent
6390 : * get_relation_info() from fetching index info, which is necessary
6391 : * because it does not expect that any IndexOptInfo is currently
6392 : * undergoing REINDEX.
6393 : */
6394 22618 : rte = makeNode(RangeTblEntry);
6395 22618 : rte->rtekind = RTE_RELATION;
6396 22618 : rte->relid = tableOid;
6397 22618 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6398 22618 : rte->rellockmode = AccessShareLock;
6399 22618 : rte->lateral = false;
6400 22618 : rte->inh = true;
6401 22618 : rte->inFromCl = true;
6402 22618 : query->rtable = list_make1(rte);
6403 :
6404 : /* Set up RTE/RelOptInfo arrays */
6405 22618 : setup_simple_rel_arrays(root);
6406 :
6407 : /* Build RelOptInfo */
6408 22618 : rel = build_simple_rel(root, 1, NULL);
6409 :
6410 : /* Rels are assumed already locked by the caller */
6411 22618 : heap = table_open(tableOid, NoLock);
6412 22618 : index = index_open(indexOid, NoLock);
6413 :
6414 : /*
6415 : * Determine if it's safe to proceed.
6416 : *
6417 : * Currently, parallel workers can't access the leader's temporary tables.
6418 : * Furthermore, any index predicate or index expressions must be parallel
6419 : * safe.
6420 : */
6421 22618 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6422 21612 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6423 21592 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6424 : {
6425 1026 : parallel_workers = 0;
6426 1026 : goto done;
6427 : }
6428 :
6429 : /*
6430 : * If parallel_workers storage parameter is set for the table, accept that
6431 : * as the number of parallel worker processes to launch (though still cap
6432 : * at max_parallel_maintenance_workers). Note that we deliberately do not
6433 : * consider any other factor when parallel_workers is set. (e.g., memory
6434 : * use by workers.)
6435 : */
6436 21592 : if (rel->rel_parallel_workers != -1)
6437 : {
6438 14 : parallel_workers = Min(rel->rel_parallel_workers,
6439 : max_parallel_maintenance_workers);
6440 14 : goto done;
6441 : }
6442 :
6443 : /*
6444 : * Estimate heap relation size ourselves, since rel->pages cannot be
6445 : * trusted (heap RTE was marked as inheritance parent)
6446 : */
6447 21578 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6448 :
6449 : /*
6450 : * Determine number of workers to scan the heap relation using generic
6451 : * model
6452 : */
6453 21578 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6454 : max_parallel_maintenance_workers);
6455 :
6456 : /*
6457 : * Cap workers based on available maintenance_work_mem as needed.
6458 : *
6459 : * Note that each tuplesort participant receives an even share of the
6460 : * total maintenance_work_mem budget. Aim to leave participants
6461 : * (including the leader as a participant) with no less than 32MB of
6462 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
6463 : * immediately past the threshold of being capable of launching a single
6464 : * parallel worker to sort.
6465 : */
6466 21656 : while (parallel_workers > 0 &&
6467 166 : maintenance_work_mem / (parallel_workers + 1) < 32768L)
6468 78 : parallel_workers--;
6469 :
6470 21578 : done:
6471 22618 : index_close(index, NoLock);
6472 22618 : table_close(heap, NoLock);
6473 :
6474 22618 : return parallel_workers;
6475 : }
6476 :
6477 : /*
6478 : * add_paths_to_grouping_rel
6479 : *
6480 : * Add non-partial paths to grouping relation.
6481 : */
6482 : static void
6483 25664 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
6484 : RelOptInfo *grouped_rel,
6485 : RelOptInfo *partially_grouped_rel,
6486 : const AggClauseCosts *agg_costs,
6487 : grouping_sets_data *gd, double dNumGroups,
6488 : GroupPathExtraData *extra)
6489 : {
6490 25664 : Query *parse = root->parse;
6491 25664 : Path *cheapest_path = input_rel->cheapest_total_path;
6492 : ListCell *lc;
6493 25664 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6494 25664 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6495 25664 : List *havingQual = (List *) extra->havingQual;
6496 25664 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6497 :
6498 25664 : if (can_sort)
6499 : {
6500 : /*
6501 : * Use any available suitably-sorted path as input, and also consider
6502 : * sorting the cheapest-total path.
6503 : */
6504 52792 : foreach(lc, input_rel->pathlist)
6505 : {
6506 27132 : Path *path = (Path *) lfirst(lc);
6507 27132 : Path *path_original = path;
6508 : bool is_sorted;
6509 : int presorted_keys;
6510 :
6511 27132 : is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
6512 : path->pathkeys,
6513 : &presorted_keys);
6514 :
6515 27132 : if (path == cheapest_path || is_sorted)
6516 : {
6517 : /* Sort the cheapest-total path if it isn't already sorted */
6518 26850 : if (!is_sorted)
6519 2612 : path = (Path *) create_sort_path(root,
6520 : grouped_rel,
6521 : path,
6522 : root->group_pathkeys,
6523 : -1.0);
6524 :
6525 : /* Now decide what to stick atop it */
6526 26850 : if (parse->groupingSets)
6527 : {
6528 480 : consider_groupingsets_paths(root, grouped_rel,
6529 : path, true, can_hash,
6530 : gd, agg_costs, dNumGroups);
6531 : }
6532 26370 : else if (parse->hasAggs)
6533 : {
6534 : /*
6535 : * We have aggregation, possibly with plain GROUP BY. Make
6536 : * an AggPath.
6537 : */
6538 26072 : add_path(grouped_rel, (Path *)
6539 78216 : create_agg_path(root,
6540 : grouped_rel,
6541 : path,
6542 26072 : grouped_rel->reltarget,
6543 26072 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6544 : AGGSPLIT_SIMPLE,
6545 : parse->groupClause,
6546 : havingQual,
6547 : agg_costs,
6548 : dNumGroups));
6549 : }
6550 298 : else if (parse->groupClause)
6551 : {
6552 : /*
6553 : * We have GROUP BY without aggregation or grouping sets.
6554 : * Make a GroupPath.
6555 : */
6556 298 : add_path(grouped_rel, (Path *)
6557 298 : create_group_path(root,
6558 : grouped_rel,
6559 : path,
6560 : parse->groupClause,
6561 : havingQual,
6562 : dNumGroups));
6563 : }
6564 : else
6565 : {
6566 : /* Other cases should have been handled above */
6567 : Assert(false);
6568 : }
6569 : }
6570 :
6571 : /*
6572 : * Now we may consider incremental sort on this path, but only
6573 : * when the path is not already sorted and when incremental sort
6574 : * is enabled.
6575 : */
6576 27132 : if (is_sorted || !enable_incrementalsort)
6577 27036 : continue;
6578 :
6579 : /* Restore the input path (we might have added Sort on top). */
6580 2074 : path = path_original;
6581 :
6582 : /* no shared prefix, no point in building incremental sort */
6583 2074 : if (presorted_keys == 0)
6584 1978 : continue;
6585 :
6586 : /*
6587 : * We should have already excluded pathkeys of length 1 because
6588 : * then presorted_keys > 0 would imply is_sorted was true.
6589 : */
6590 : Assert(list_length(root->group_pathkeys) != 1);
6591 :
6592 96 : path = (Path *) create_incremental_sort_path(root,
6593 : grouped_rel,
6594 : path,
6595 : root->group_pathkeys,
6596 : presorted_keys,
6597 : -1.0);
6598 :
6599 : /* Now decide what to stick atop it */
6600 96 : if (parse->groupingSets)
6601 : {
6602 16 : consider_groupingsets_paths(root, grouped_rel,
6603 : path, true, can_hash,
6604 : gd, agg_costs, dNumGroups);
6605 : }
6606 80 : else if (parse->hasAggs)
6607 : {
6608 : /*
6609 : * We have aggregation, possibly with plain GROUP BY. Make an
6610 : * AggPath.
6611 : */
6612 64 : add_path(grouped_rel, (Path *)
6613 192 : create_agg_path(root,
6614 : grouped_rel,
6615 : path,
6616 64 : grouped_rel->reltarget,
6617 64 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6618 : AGGSPLIT_SIMPLE,
6619 : parse->groupClause,
6620 : havingQual,
6621 : agg_costs,
6622 : dNumGroups));
6623 : }
6624 16 : else if (parse->groupClause)
6625 : {
6626 : /*
6627 : * We have GROUP BY without aggregation or grouping sets. Make
6628 : * a GroupPath.
6629 : */
6630 16 : add_path(grouped_rel, (Path *)
6631 16 : create_group_path(root,
6632 : grouped_rel,
6633 : path,
6634 : parse->groupClause,
6635 : havingQual,
6636 : dNumGroups));
6637 : }
6638 : else
6639 : {
6640 : /* Other cases should have been handled above */
6641 : Assert(false);
6642 : }
6643 : }
6644 :
6645 : /*
6646 : * Instead of operating directly on the input relation, we can
6647 : * consider finalizing a partially aggregated path.
6648 : */
6649 25660 : if (partially_grouped_rel != NULL)
6650 : {
6651 2304 : foreach(lc, partially_grouped_rel->pathlist)
6652 : {
6653 1392 : Path *path = (Path *) lfirst(lc);
6654 1392 : Path *path_original = path;
6655 : bool is_sorted;
6656 : int presorted_keys;
6657 :
6658 1392 : is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
6659 : path->pathkeys,
6660 : &presorted_keys);
6661 :
6662 : /*
6663 : * Insert a Sort node, if required. But there's no point in
6664 : * sorting anything but the cheapest path.
6665 : */
6666 1392 : if (!is_sorted)
6667 : {
6668 360 : if (path != partially_grouped_rel->cheapest_total_path)
6669 1392 : continue;
6670 296 : path = (Path *) create_sort_path(root,
6671 : grouped_rel,
6672 : path,
6673 : root->group_pathkeys,
6674 : -1.0);
6675 : }
6676 :
6677 1328 : if (parse->hasAggs)
6678 1198 : add_path(grouped_rel, (Path *)
6679 3594 : create_agg_path(root,
6680 : grouped_rel,
6681 : path,
6682 1198 : grouped_rel->reltarget,
6683 1198 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6684 : AGGSPLIT_FINAL_DESERIAL,
6685 : parse->groupClause,
6686 : havingQual,
6687 : agg_final_costs,
6688 : dNumGroups));
6689 : else
6690 130 : add_path(grouped_rel, (Path *)
6691 130 : create_group_path(root,
6692 : grouped_rel,
6693 : path,
6694 : parse->groupClause,
6695 : havingQual,
6696 : dNumGroups));
6697 :
6698 : /*
6699 : * Now we may consider incremental sort on this path, but only
6700 : * when the path is not already sorted and when incremental
6701 : * sort is enabled.
6702 : */
6703 1328 : if (is_sorted || !enable_incrementalsort)
6704 1268 : continue;
6705 :
6706 : /* Restore the input path (we might have added Sort on top). */
6707 60 : path = path_original;
6708 :
6709 : /* no shared prefix, not point in building incremental sort */
6710 60 : if (presorted_keys == 0)
6711 60 : continue;
6712 :
6713 : /*
6714 : * We should have already excluded pathkeys of length 1
6715 : * because then presorted_keys > 0 would imply is_sorted was
6716 : * true.
6717 : */
6718 : Assert(list_length(root->group_pathkeys) != 1);
6719 :
6720 0 : path = (Path *) create_incremental_sort_path(root,
6721 : grouped_rel,
6722 : path,
6723 : root->group_pathkeys,
6724 : presorted_keys,
6725 : -1.0);
6726 :
6727 0 : if (parse->hasAggs)
6728 0 : add_path(grouped_rel, (Path *)
6729 0 : create_agg_path(root,
6730 : grouped_rel,
6731 : path,
6732 0 : grouped_rel->reltarget,
6733 0 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6734 : AGGSPLIT_FINAL_DESERIAL,
6735 : parse->groupClause,
6736 : havingQual,
6737 : agg_final_costs,
6738 : dNumGroups));
6739 : else
6740 0 : add_path(grouped_rel, (Path *)
6741 0 : create_group_path(root,
6742 : grouped_rel,
6743 : path,
6744 : parse->groupClause,
6745 : havingQual,
6746 : dNumGroups));
6747 : }
6748 : }
6749 : }
6750 :
6751 25664 : if (can_hash)
6752 : {
6753 : double hashaggtablesize;
6754 :
6755 2538 : if (parse->groupingSets)
6756 : {
6757 : /*
6758 : * Try for a hash-only groupingsets path over unsorted input.
6759 : */
6760 396 : consider_groupingsets_paths(root, grouped_rel,
6761 : cheapest_path, false, true,
6762 : gd, agg_costs, dNumGroups);
6763 : }
6764 : else
6765 : {
6766 2142 : hashaggtablesize = estimate_hashagg_tablesize(cheapest_path,
6767 : agg_costs,
6768 : dNumGroups);
6769 :
6770 : /*
6771 : * Provided that the estimated size of the hashtable does not
6772 : * exceed work_mem, we'll generate a HashAgg Path, although if we
6773 : * were unable to sort above, then we'd better generate a Path, so
6774 : * that we at least have one.
6775 : */
6776 2142 : if (enable_hashagg_disk ||
6777 0 : hashaggtablesize < work_mem * 1024L ||
6778 0 : grouped_rel->pathlist == NIL)
6779 : {
6780 : /*
6781 : * We just need an Agg over the cheapest-total input path,
6782 : * since input order won't matter.
6783 : */
6784 2142 : add_path(grouped_rel, (Path *)
6785 4284 : create_agg_path(root, grouped_rel,
6786 : cheapest_path,
6787 2142 : grouped_rel->reltarget,
6788 : AGG_HASHED,
6789 : AGGSPLIT_SIMPLE,
6790 : parse->groupClause,
6791 : havingQual,
6792 : agg_costs,
6793 : dNumGroups));
6794 : }
6795 : }
6796 :
6797 : /*
6798 : * Generate a Finalize HashAgg Path atop of the cheapest partially
6799 : * grouped path, assuming there is one. Once again, we'll only do this
6800 : * if it looks as though the hash table won't exceed work_mem.
6801 : */
6802 2538 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
6803 : {
6804 474 : Path *path = partially_grouped_rel->cheapest_total_path;
6805 :
6806 474 : hashaggtablesize = estimate_hashagg_tablesize(path,
6807 : agg_final_costs,
6808 : dNumGroups);
6809 :
6810 474 : if (enable_hashagg_disk ||
6811 0 : hashaggtablesize < work_mem * 1024L)
6812 474 : add_path(grouped_rel, (Path *)
6813 948 : create_agg_path(root,
6814 : grouped_rel,
6815 : path,
6816 474 : grouped_rel->reltarget,
6817 : AGG_HASHED,
6818 : AGGSPLIT_FINAL_DESERIAL,
6819 : parse->groupClause,
6820 : havingQual,
6821 : agg_final_costs,
6822 : dNumGroups));
6823 : }
6824 : }
6825 :
6826 : /*
6827 : * When partitionwise aggregate is used, we might have fully aggregated
6828 : * paths in the partial pathlist, because add_paths_to_append_rel() will
6829 : * consider a path for grouped_rel consisting of a Parallel Append of
6830 : * non-partial paths from each child.
6831 : */
6832 25664 : if (grouped_rel->partial_pathlist != NIL)
6833 108 : gather_grouping_paths(root, grouped_rel);
6834 25664 : }
6835 :
6836 : /*
6837 : * create_partial_grouping_paths
6838 : *
6839 : * Create a new upper relation representing the result of partial aggregation
6840 : * and populate it with appropriate paths. Note that we don't finalize the
6841 : * lists of paths here, so the caller can add additional partial or non-partial
6842 : * paths and must afterward call gather_grouping_paths and set_cheapest on
6843 : * the returned upper relation.
6844 : *
6845 : * All paths for this new upper relation -- both partial and non-partial --
6846 : * have been partially aggregated but require a subsequent FinalizeAggregate
6847 : * step.
6848 : *
6849 : * NB: This function is allowed to return NULL if it determines that there is
6850 : * no real need to create a new RelOptInfo.
6851 : */
6852 : static RelOptInfo *
6853 9804 : create_partial_grouping_paths(PlannerInfo *root,
6854 : RelOptInfo *grouped_rel,
6855 : RelOptInfo *input_rel,
6856 : grouping_sets_data *gd,
6857 : GroupPathExtraData *extra,
6858 : bool force_rel_creation)
6859 : {
6860 9804 : Query *parse = root->parse;
6861 : RelOptInfo *partially_grouped_rel;
6862 9804 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
6863 9804 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6864 9804 : Path *cheapest_partial_path = NULL;
6865 9804 : Path *cheapest_total_path = NULL;
6866 9804 : double dNumPartialGroups = 0;
6867 9804 : double dNumPartialPartialGroups = 0;
6868 : ListCell *lc;
6869 9804 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6870 9804 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6871 :
6872 : /*
6873 : * Consider whether we should generate partially aggregated non-partial
6874 : * paths. We can only do this if we have a non-partial path, and only if
6875 : * the parent of the input rel is performing partial partitionwise
6876 : * aggregation. (Note that extra->patype is the type of partitionwise
6877 : * aggregation being used at the parent level, not this level.)
6878 : */
6879 9804 : if (input_rel->pathlist != NIL &&
6880 9804 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
6881 398 : cheapest_total_path = input_rel->cheapest_total_path;
6882 :
6883 : /*
6884 : * If parallelism is possible for grouped_rel, then we should consider
6885 : * generating partially-grouped partial paths. However, if the input rel
6886 : * has no partial paths, then we can't.
6887 : */
6888 9804 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
6889 1054 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
6890 :
6891 : /*
6892 : * If we can't partially aggregate partial paths, and we can't partially
6893 : * aggregate non-partial paths, then don't bother creating the new
6894 : * RelOptInfo at all, unless the caller specified force_rel_creation.
6895 : */
6896 9804 : if (cheapest_total_path == NULL &&
6897 8552 : cheapest_partial_path == NULL &&
6898 8552 : !force_rel_creation)
6899 8494 : return NULL;
6900 :
6901 : /*
6902 : * Build a new upper relation to represent the result of partially
6903 : * aggregating the rows from the input relation.
6904 : */
6905 1310 : partially_grouped_rel = fetch_upper_rel(root,
6906 : UPPERREL_PARTIAL_GROUP_AGG,
6907 : grouped_rel->relids);
6908 1310 : partially_grouped_rel->consider_parallel =
6909 1310 : grouped_rel->consider_parallel;
6910 1310 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
6911 1310 : partially_grouped_rel->serverid = grouped_rel->serverid;
6912 1310 : partially_grouped_rel->userid = grouped_rel->userid;
6913 1310 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
6914 1310 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
6915 :
6916 : /*
6917 : * Build target list for partial aggregate paths. These paths cannot just
6918 : * emit the same tlist as regular aggregate paths, because (1) we must
6919 : * include Vars and Aggrefs needed in HAVING, which might not appear in
6920 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
6921 : */
6922 1310 : partially_grouped_rel->reltarget =
6923 1310 : make_partial_grouping_target(root, grouped_rel->reltarget,
6924 : extra->havingQual);
6925 :
6926 1310 : if (!extra->partial_costs_set)
6927 : {
6928 : /*
6929 : * Collect statistics about aggregates for estimating costs of
6930 : * performing aggregation in parallel.
6931 : */
6932 5760 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
6933 5760 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
6934 720 : if (parse->hasAggs)
6935 : {
6936 : List *partial_target_exprs;
6937 :
6938 : /* partial phase */
6939 656 : partial_target_exprs = partially_grouped_rel->reltarget->exprs;
6940 656 : get_agg_clause_costs(root, (Node *) partial_target_exprs,
6941 : AGGSPLIT_INITIAL_SERIAL,
6942 : agg_partial_costs);
6943 :
6944 : /* final phase */
6945 656 : get_agg_clause_costs(root, (Node *) grouped_rel->reltarget->exprs,
6946 : AGGSPLIT_FINAL_DESERIAL,
6947 : agg_final_costs);
6948 656 : get_agg_clause_costs(root, extra->havingQual,
6949 : AGGSPLIT_FINAL_DESERIAL,
6950 : agg_final_costs);
6951 : }
6952 :
6953 720 : extra->partial_costs_set = true;
6954 : }
6955 :
6956 : /* Estimate number of partial groups. */
6957 1310 : if (cheapest_total_path != NULL)
6958 : dNumPartialGroups =
6959 398 : get_number_of_groups(root,
6960 : cheapest_total_path->rows,
6961 : gd,
6962 : extra->targetList);
6963 1310 : if (cheapest_partial_path != NULL)
6964 : dNumPartialPartialGroups =
6965 1054 : get_number_of_groups(root,
6966 : cheapest_partial_path->rows,
6967 : gd,
6968 : extra->targetList);
6969 :
6970 1310 : if (can_sort && cheapest_total_path != NULL)
6971 : {
6972 : /* This should have been checked previously */
6973 : Assert(parse->hasAggs || parse->groupClause);
6974 :
6975 : /*
6976 : * Use any available suitably-sorted path as input, and also consider
6977 : * sorting the cheapest partial path.
6978 : */
6979 802 : foreach(lc, input_rel->pathlist)
6980 : {
6981 404 : Path *path = (Path *) lfirst(lc);
6982 : bool is_sorted;
6983 :
6984 404 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6985 : path->pathkeys);
6986 404 : if (path == cheapest_total_path || is_sorted)
6987 : {
6988 : /* Sort the cheapest partial path, if it isn't already */
6989 404 : if (!is_sorted)
6990 398 : path = (Path *) create_sort_path(root,
6991 : partially_grouped_rel,
6992 : path,
6993 : root->group_pathkeys,
6994 : -1.0);
6995 :
6996 404 : if (parse->hasAggs)
6997 356 : add_path(partially_grouped_rel, (Path *)
6998 1068 : create_agg_path(root,
6999 : partially_grouped_rel,
7000 : path,
7001 356 : partially_grouped_rel->reltarget,
7002 356 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7003 : AGGSPLIT_INITIAL_SERIAL,
7004 : parse->groupClause,
7005 : NIL,
7006 : agg_partial_costs,
7007 : dNumPartialGroups));
7008 : else
7009 48 : add_path(partially_grouped_rel, (Path *)
7010 48 : create_group_path(root,
7011 : partially_grouped_rel,
7012 : path,
7013 : parse->groupClause,
7014 : NIL,
7015 : dNumPartialGroups));
7016 : }
7017 : }
7018 :
7019 : /*
7020 : * Consider incremental sort on all partial paths, if enabled.
7021 : *
7022 : * We can also skip the entire loop when we only have a single-item
7023 : * group_pathkeys because then we can't possibly have a presorted
7024 : * prefix of the list without having the list be fully sorted.
7025 : */
7026 398 : if (enable_incrementalsort && list_length(root->group_pathkeys) > 1)
7027 : {
7028 48 : foreach(lc, input_rel->pathlist)
7029 : {
7030 24 : Path *path = (Path *) lfirst(lc);
7031 : bool is_sorted;
7032 : int presorted_keys;
7033 :
7034 24 : is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
7035 : path->pathkeys,
7036 : &presorted_keys);
7037 :
7038 : /* Ignore already sorted paths */
7039 24 : if (is_sorted)
7040 24 : continue;
7041 :
7042 24 : if (presorted_keys == 0)
7043 24 : continue;
7044 :
7045 : /* Since we have presorted keys, consider incremental sort. */
7046 0 : path = (Path *) create_incremental_sort_path(root,
7047 : partially_grouped_rel,
7048 : path,
7049 : root->group_pathkeys,
7050 : presorted_keys,
7051 : -1.0);
7052 :
7053 0 : if (parse->hasAggs)
7054 0 : add_path(partially_grouped_rel, (Path *)
7055 0 : create_agg_path(root,
7056 : partially_grouped_rel,
7057 : path,
7058 0 : partially_grouped_rel->reltarget,
7059 0 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7060 : AGGSPLIT_INITIAL_SERIAL,
7061 : parse->groupClause,
7062 : NIL,
7063 : agg_partial_costs,
7064 : dNumPartialGroups));
7065 : else
7066 0 : add_path(partially_grouped_rel, (Path *)
7067 0 : create_group_path(root,
7068 : partially_grouped_rel,
7069 : path,
7070 : parse->groupClause,
7071 : NIL,
7072 : dNumPartialGroups));
7073 : }
7074 : }
7075 :
7076 : }
7077 :
7078 1310 : if (can_sort && cheapest_partial_path != NULL)
7079 : {
7080 : /* Similar to above logic, but for partial paths. */
7081 2116 : foreach(lc, input_rel->partial_pathlist)
7082 : {
7083 1062 : Path *path = (Path *) lfirst(lc);
7084 1062 : Path *path_original = path;
7085 : bool is_sorted;
7086 : int presorted_keys;
7087 :
7088 1062 : is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
7089 : path->pathkeys,
7090 : &presorted_keys);
7091 :
7092 1062 : if (path == cheapest_partial_path || is_sorted)
7093 : {
7094 : /* Sort the cheapest partial path, if it isn't already */
7095 1054 : if (!is_sorted)
7096 616 : path = (Path *) create_sort_path(root,
7097 : partially_grouped_rel,
7098 : path,
7099 : root->group_pathkeys,
7100 : -1.0);
7101 :
7102 1054 : if (parse->hasAggs)
7103 998 : add_partial_path(partially_grouped_rel, (Path *)
7104 2994 : create_agg_path(root,
7105 : partially_grouped_rel,
7106 : path,
7107 998 : partially_grouped_rel->reltarget,
7108 998 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7109 : AGGSPLIT_INITIAL_SERIAL,
7110 : parse->groupClause,
7111 : NIL,
7112 : agg_partial_costs,
7113 : dNumPartialPartialGroups));
7114 : else
7115 56 : add_partial_path(partially_grouped_rel, (Path *)
7116 56 : create_group_path(root,
7117 : partially_grouped_rel,
7118 : path,
7119 : parse->groupClause,
7120 : NIL,
7121 : dNumPartialPartialGroups));
7122 : }
7123 :
7124 : /*
7125 : * Now we may consider incremental sort on this path, but only
7126 : * when the path is not already sorted and when incremental sort
7127 : * is enabled.
7128 : */
7129 1062 : if (is_sorted || !enable_incrementalsort)
7130 1058 : continue;
7131 :
7132 : /* Restore the input path (we might have added Sort on top). */
7133 136 : path = path_original;
7134 :
7135 : /* no shared prefix, not point in building incremental sort */
7136 136 : if (presorted_keys == 0)
7137 132 : continue;
7138 :
7139 : /*
7140 : * We should have already excluded pathkeys of length 1 because
7141 : * then presorted_keys > 0 would imply is_sorted was true.
7142 : */
7143 : Assert(list_length(root->group_pathkeys) != 1);
7144 :
7145 4 : path = (Path *) create_incremental_sort_path(root,
7146 : partially_grouped_rel,
7147 : path,
7148 : root->group_pathkeys,
7149 : presorted_keys,
7150 : -1.0);
7151 :
7152 4 : if (parse->hasAggs)
7153 4 : add_partial_path(partially_grouped_rel, (Path *)
7154 12 : create_agg_path(root,
7155 : partially_grouped_rel,
7156 : path,
7157 4 : partially_grouped_rel->reltarget,
7158 4 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7159 : AGGSPLIT_INITIAL_SERIAL,
7160 : parse->groupClause,
7161 : NIL,
7162 : agg_partial_costs,
7163 : dNumPartialPartialGroups));
7164 : else
7165 0 : add_partial_path(partially_grouped_rel, (Path *)
7166 0 : create_group_path(root,
7167 : partially_grouped_rel,
7168 : path,
7169 : parse->groupClause,
7170 : NIL,
7171 : dNumPartialPartialGroups));
7172 : }
7173 : }
7174 :
7175 1310 : if (can_hash && cheapest_total_path != NULL)
7176 : {
7177 : double hashaggtablesize;
7178 :
7179 : /* Checked above */
7180 : Assert(parse->hasAggs || parse->groupClause);
7181 :
7182 : hashaggtablesize =
7183 398 : estimate_hashagg_tablesize(cheapest_total_path,
7184 : agg_partial_costs,
7185 : dNumPartialGroups);
7186 :
7187 : /*
7188 : * Tentatively produce a partial HashAgg Path, depending on if it
7189 : * looks as if the hash table will fit in work_mem.
7190 : */
7191 398 : if ((enable_hashagg_disk || hashaggtablesize < work_mem * 1024L) &&
7192 : cheapest_total_path != NULL)
7193 : {
7194 398 : add_path(partially_grouped_rel, (Path *)
7195 796 : create_agg_path(root,
7196 : partially_grouped_rel,
7197 : cheapest_total_path,
7198 398 : partially_grouped_rel->reltarget,
7199 : AGG_HASHED,
7200 : AGGSPLIT_INITIAL_SERIAL,
7201 : parse->groupClause,
7202 : NIL,
7203 : agg_partial_costs,
7204 : dNumPartialGroups));
7205 : }
7206 : }
7207 :
7208 1310 : if (can_hash && cheapest_partial_path != NULL)
7209 : {
7210 : double hashaggtablesize;
7211 :
7212 : hashaggtablesize =
7213 616 : estimate_hashagg_tablesize(cheapest_partial_path,
7214 : agg_partial_costs,
7215 : dNumPartialPartialGroups);
7216 :
7217 : /* Do the same for partial paths. */
7218 616 : if ((enable_hashagg_disk || hashaggtablesize < work_mem * 1024L) &&
7219 : cheapest_partial_path != NULL)
7220 : {
7221 616 : add_partial_path(partially_grouped_rel, (Path *)
7222 1232 : create_agg_path(root,
7223 : partially_grouped_rel,
7224 : cheapest_partial_path,
7225 616 : partially_grouped_rel->reltarget,
7226 : AGG_HASHED,
7227 : AGGSPLIT_INITIAL_SERIAL,
7228 : parse->groupClause,
7229 : NIL,
7230 : agg_partial_costs,
7231 : dNumPartialPartialGroups));
7232 : }
7233 : }
7234 :
7235 : /*
7236 : * If there is an FDW that's responsible for all baserels of the query,
7237 : * let it consider adding partially grouped ForeignPaths.
7238 : */
7239 1310 : if (partially_grouped_rel->fdwroutine &&
7240 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7241 : {
7242 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7243 :
7244 6 : fdwroutine->GetForeignUpperPaths(root,
7245 : UPPERREL_PARTIAL_GROUP_AGG,
7246 : input_rel, partially_grouped_rel,
7247 : extra);
7248 : }
7249 :
7250 1310 : return partially_grouped_rel;
7251 : }
7252 :
7253 : /*
7254 : * Generate Gather and Gather Merge paths for a grouping relation or partial
7255 : * grouping relation.
7256 : *
7257 : * generate_gather_paths does most of the work, but we also consider a special
7258 : * case: we could try sorting the data by the group_pathkeys and then applying
7259 : * Gather Merge.
7260 : *
7261 : * NB: This function shouldn't be used for anything other than a grouped or
7262 : * partially grouped relation not only because of the fact that it explicitly
7263 : * references group_pathkeys but we pass "true" as the third argument to
7264 : * generate_gather_paths().
7265 : */
7266 : static void
7267 962 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7268 : {
7269 : ListCell *lc;
7270 : Path *cheapest_partial_path;
7271 :
7272 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
7273 962 : generate_useful_gather_paths(root, rel, true);
7274 :
7275 : /* Try cheapest partial path + explicit Sort + Gather Merge. */
7276 962 : cheapest_partial_path = linitial(rel->partial_pathlist);
7277 962 : if (!pathkeys_contained_in(root->group_pathkeys,
7278 : cheapest_partial_path->pathkeys))
7279 : {
7280 : Path *path;
7281 : double total_groups;
7282 :
7283 478 : total_groups =
7284 478 : cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
7285 478 : path = (Path *) create_sort_path(root, rel, cheapest_partial_path,
7286 : root->group_pathkeys,
7287 : -1.0);
7288 : path = (Path *)
7289 956 : create_gather_merge_path(root,
7290 : rel,
7291 : path,
7292 478 : rel->reltarget,
7293 : root->group_pathkeys,
7294 : NULL,
7295 : &total_groups);
7296 :
7297 478 : add_path(rel, path);
7298 : }
7299 :
7300 : /*
7301 : * Consider incremental sort on all partial paths, if enabled.
7302 : *
7303 : * We can also skip the entire loop when we only have a single-item
7304 : * group_pathkeys because then we can't possibly have a presorted prefix
7305 : * of the list without having the list be fully sorted.
7306 : */
7307 962 : if (!enable_incrementalsort || list_length(root->group_pathkeys) == 1)
7308 492 : return;
7309 :
7310 : /* also consider incremental sort on partial paths, if enabled */
7311 964 : foreach(lc, rel->partial_pathlist)
7312 : {
7313 494 : Path *path = (Path *) lfirst(lc);
7314 : bool is_sorted;
7315 : int presorted_keys;
7316 : double total_groups;
7317 :
7318 494 : is_sorted = pathkeys_count_contained_in(root->group_pathkeys,
7319 : path->pathkeys,
7320 : &presorted_keys);
7321 :
7322 494 : if (is_sorted)
7323 494 : continue;
7324 :
7325 24 : if (presorted_keys == 0)
7326 24 : continue;
7327 :
7328 0 : path = (Path *) create_incremental_sort_path(root,
7329 : rel,
7330 : path,
7331 : root->group_pathkeys,
7332 : presorted_keys,
7333 : -1.0);
7334 :
7335 : path = (Path *)
7336 0 : create_gather_merge_path(root,
7337 : rel,
7338 : path,
7339 0 : rel->reltarget,
7340 : root->group_pathkeys,
7341 : NULL,
7342 : &total_groups);
7343 :
7344 0 : add_path(rel, path);
7345 : }
7346 : }
7347 :
7348 : /*
7349 : * can_partial_agg
7350 : *
7351 : * Determines whether or not partial grouping and/or aggregation is possible.
7352 : * Returns true when possible, false otherwise.
7353 : */
7354 : static bool
7355 25120 : can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
7356 : {
7357 25120 : Query *parse = root->parse;
7358 :
7359 25120 : if (!parse->hasAggs && parse->groupClause == NIL)
7360 : {
7361 : /*
7362 : * We don't know how to do parallel aggregation unless we have either
7363 : * some aggregates or a grouping clause.
7364 : */
7365 0 : return false;
7366 : }
7367 25120 : else if (parse->groupingSets)
7368 : {
7369 : /* We don't know how to do grouping sets in parallel. */
7370 448 : return false;
7371 : }
7372 24672 : else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
7373 : {
7374 : /* Insufficient support for partial mode. */
7375 15762 : return false;
7376 : }
7377 :
7378 : /* Everything looks good. */
7379 8910 : return true;
7380 : }
7381 :
7382 : /*
7383 : * apply_scanjoin_target_to_paths
7384 : *
7385 : * Adjust the final scan/join relation, and recursively all of its children,
7386 : * to generate the final scan/join target. It would be more correct to model
7387 : * this as a separate planning step with a new RelOptInfo at the toplevel and
7388 : * for each child relation, but doing it this way is noticeably cheaper.
7389 : * Maybe that problem can be solved at some point, but for now we do this.
7390 : *
7391 : * If tlist_same_exprs is true, then the scan/join target to be applied has
7392 : * the same expressions as the existing reltarget, so we need only insert the
7393 : * appropriate sortgroupref information. By avoiding the creation of
7394 : * projection paths we save effort both immediately and at plan creation time.
7395 : */
7396 : static void
7397 331566 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7398 : RelOptInfo *rel,
7399 : List *scanjoin_targets,
7400 : List *scanjoin_targets_contain_srfs,
7401 : bool scanjoin_target_parallel_safe,
7402 : bool tlist_same_exprs)
7403 : {
7404 331566 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
7405 : PathTarget *scanjoin_target;
7406 : ListCell *lc;
7407 :
7408 : /* This recurses, so be paranoid. */
7409 331566 : check_stack_depth();
7410 :
7411 : /*
7412 : * If the rel is partitioned, we want to drop its existing paths and
7413 : * generate new ones. This function would still be correct if we kept the
7414 : * existing paths: we'd modify them to generate the correct target above
7415 : * the partitioning Append, and then they'd compete on cost with paths
7416 : * generating the target below the Append. However, in our current cost
7417 : * model the latter way is always the same or cheaper cost, so modifying
7418 : * the existing paths would just be useless work. Moreover, when the cost
7419 : * is the same, varying roundoff errors might sometimes allow an existing
7420 : * path to be picked, resulting in undesirable cross-platform plan
7421 : * variations. So we drop old paths and thereby force the work to be done
7422 : * below the Append, except in the case of a non-parallel-safe target.
7423 : *
7424 : * Some care is needed, because we have to allow generate_gather_paths to
7425 : * see the old partial paths in the next stanza. Hence, zap the main
7426 : * pathlist here, then allow generate_gather_paths to add path(s) to the
7427 : * main list, and finally zap the partial pathlist.
7428 : */
7429 331566 : if (rel_is_partitioned)
7430 7362 : rel->pathlist = NIL;
7431 :
7432 : /*
7433 : * If the scan/join target is not parallel-safe, partial paths cannot
7434 : * generate it.
7435 : */
7436 331566 : if (!scanjoin_target_parallel_safe)
7437 : {
7438 : /*
7439 : * Since we can't generate the final scan/join target in parallel
7440 : * workers, this is our last opportunity to use any partial paths that
7441 : * exist; so build Gather path(s) that use them and emit whatever the
7442 : * current reltarget is. We don't do this in the case where the
7443 : * target is parallel-safe, since we will be able to generate superior
7444 : * paths by doing it after the final scan/join target has been
7445 : * applied.
7446 : */
7447 69476 : generate_useful_gather_paths(root, rel, false);
7448 :
7449 : /* Can't use parallel query above this level. */
7450 69476 : rel->partial_pathlist = NIL;
7451 69476 : rel->consider_parallel = false;
7452 : }
7453 :
7454 : /* Finish dropping old paths for a partitioned rel, per comment above */
7455 331566 : if (rel_is_partitioned)
7456 7362 : rel->partial_pathlist = NIL;
7457 :
7458 : /* Extract SRF-free scan/join target. */
7459 331566 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7460 :
7461 : /*
7462 : * Apply the SRF-free scan/join target to each existing path.
7463 : *
7464 : * If the tlist exprs are the same, we can just inject the sortgroupref
7465 : * information into the existing pathtargets. Otherwise, replace each
7466 : * path with a projection path that generates the SRF-free scan/join
7467 : * target. This can't change the ordering of paths within rel->pathlist,
7468 : * so we just modify the list in place.
7469 : */
7470 674608 : foreach(lc, rel->pathlist)
7471 : {
7472 343042 : Path *subpath = (Path *) lfirst(lc);
7473 :
7474 : /* Shouldn't have any parameterized paths anymore */
7475 : Assert(subpath->param_info == NULL);
7476 :
7477 343042 : if (tlist_same_exprs)
7478 222476 : subpath->pathtarget->sortgrouprefs =
7479 111238 : scanjoin_target->sortgrouprefs;
7480 : else
7481 : {
7482 : Path *newpath;
7483 :
7484 231804 : newpath = (Path *) create_projection_path(root, rel, subpath,
7485 : scanjoin_target);
7486 231804 : lfirst(lc) = newpath;
7487 : }
7488 : }
7489 :
7490 : /* Likewise adjust the targets for any partial paths. */
7491 344094 : foreach(lc, rel->partial_pathlist)
7492 : {
7493 12528 : Path *subpath = (Path *) lfirst(lc);
7494 :
7495 : /* Shouldn't have any parameterized paths anymore */
7496 : Assert(subpath->param_info == NULL);
7497 :
7498 12528 : if (tlist_same_exprs)
7499 21164 : subpath->pathtarget->sortgrouprefs =
7500 10582 : scanjoin_target->sortgrouprefs;
7501 : else
7502 : {
7503 : Path *newpath;
7504 :
7505 1946 : newpath = (Path *) create_projection_path(root, rel, subpath,
7506 : scanjoin_target);
7507 1946 : lfirst(lc) = newpath;
7508 : }
7509 : }
7510 :
7511 : /*
7512 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7513 : * atop each existing path. (Note that this function doesn't look at the
7514 : * cheapest-path fields, which is a good thing because they're bogus right
7515 : * now.)
7516 : */
7517 331566 : if (root->parse->hasTargetSRFs)
7518 3880 : adjust_paths_for_srfs(root, rel,
7519 : scanjoin_targets,
7520 : scanjoin_targets_contain_srfs);
7521 :
7522 : /*
7523 : * Update the rel's target to be the final (with SRFs) scan/join target.
7524 : * This now matches the actual output of all the paths, and we might get
7525 : * confused in createplan.c if they don't agree. We must do this now so
7526 : * that any append paths made in the next part will use the correct
7527 : * pathtarget (cf. create_append_path).
7528 : *
7529 : * Note that this is also necessary if GetForeignUpperPaths() gets called
7530 : * on the final scan/join relation or on any of its children, since the
7531 : * FDW might look at the rel's target to create ForeignPaths.
7532 : */
7533 331566 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7534 :
7535 : /*
7536 : * If the relation is partitioned, recursively apply the scan/join target
7537 : * to all partitions, and generate brand-new Append paths in which the
7538 : * scan/join target is computed below the Append rather than above it.
7539 : * Since Append is not projection-capable, that might save a separate
7540 : * Result node, and it also is important for partitionwise aggregate.
7541 : */
7542 331566 : if (rel_is_partitioned)
7543 : {
7544 7362 : List *live_children = NIL;
7545 : int partition_idx;
7546 :
7547 : /* Adjust each partition. */
7548 29568 : for (partition_idx = 0; partition_idx < rel->nparts; partition_idx++)
7549 : {
7550 22206 : RelOptInfo *child_rel = rel->part_rels[partition_idx];
7551 : AppendRelInfo **appinfos;
7552 : int nappinfos;
7553 22206 : List *child_scanjoin_targets = NIL;
7554 : ListCell *lc;
7555 :
7556 : /* Pruned or dummy children can be ignored. */
7557 22206 : if (child_rel == NULL || IS_DUMMY_REL(child_rel))
7558 9460 : continue;
7559 :
7560 : /* Translate scan/join targets for this child. */
7561 12746 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7562 : &nappinfos);
7563 25492 : foreach(lc, scanjoin_targets)
7564 : {
7565 12746 : PathTarget *target = lfirst_node(PathTarget, lc);
7566 :
7567 12746 : target = copy_pathtarget(target);
7568 12746 : target->exprs = (List *)
7569 25492 : adjust_appendrel_attrs(root,
7570 12746 : (Node *) target->exprs,
7571 : nappinfos, appinfos);
7572 12746 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7573 : target);
7574 : }
7575 12746 : pfree(appinfos);
7576 :
7577 : /* Recursion does the real work. */
7578 12746 : apply_scanjoin_target_to_paths(root, child_rel,
7579 : child_scanjoin_targets,
7580 : scanjoin_targets_contain_srfs,
7581 : scanjoin_target_parallel_safe,
7582 : tlist_same_exprs);
7583 :
7584 : /* Save non-dummy children for Append paths. */
7585 12746 : if (!IS_DUMMY_REL(child_rel))
7586 12746 : live_children = lappend(live_children, child_rel);
7587 : }
7588 :
7589 : /* Build new paths for this relation by appending child paths. */
7590 7362 : add_paths_to_append_rel(root, rel, live_children);
7591 : }
7592 :
7593 : /*
7594 : * Consider generating Gather or Gather Merge paths. We must only do this
7595 : * if the relation is parallel safe, and we don't do it for child rels to
7596 : * avoid creating multiple Gather nodes within the same plan. We must do
7597 : * this after all paths have been generated and before set_cheapest, since
7598 : * one of the generated paths may turn out to be the cheapest one.
7599 : */
7600 331566 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7601 78608 : generate_useful_gather_paths(root, rel, false);
7602 :
7603 : /*
7604 : * Reassess which paths are the cheapest, now that we've potentially added
7605 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7606 : * this relation.
7607 : */
7608 331566 : set_cheapest(rel);
7609 331566 : }
7610 :
7611 : /*
7612 : * create_partitionwise_grouping_paths
7613 : *
7614 : * If the partition keys of input relation are part of the GROUP BY clause, all
7615 : * the rows belonging to a given group come from a single partition. This
7616 : * allows aggregation/grouping over a partitioned relation to be broken down
7617 : * into aggregation/grouping on each partition. This should be no worse, and
7618 : * often better, than the normal approach.
7619 : *
7620 : * However, if the GROUP BY clause does not contain all the partition keys,
7621 : * rows from a given group may be spread across multiple partitions. In that
7622 : * case, we perform partial aggregation for each group, append the results,
7623 : * and then finalize aggregation. This is less certain to win than the
7624 : * previous case. It may win if the PartialAggregate stage greatly reduces
7625 : * the number of groups, because fewer rows will pass through the Append node.
7626 : * It may lose if we have lots of small groups.
7627 : */
7628 : static void
7629 346 : create_partitionwise_grouping_paths(PlannerInfo *root,
7630 : RelOptInfo *input_rel,
7631 : RelOptInfo *grouped_rel,
7632 : RelOptInfo *partially_grouped_rel,
7633 : const AggClauseCosts *agg_costs,
7634 : grouping_sets_data *gd,
7635 : PartitionwiseAggregateType patype,
7636 : GroupPathExtraData *extra)
7637 : {
7638 346 : int nparts = input_rel->nparts;
7639 : int cnt_parts;
7640 346 : List *grouped_live_children = NIL;
7641 346 : List *partially_grouped_live_children = NIL;
7642 346 : PathTarget *target = grouped_rel->reltarget;
7643 346 : bool partial_grouping_valid = true;
7644 :
7645 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7646 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7647 : partially_grouped_rel != NULL);
7648 :
7649 : /* Add paths for partitionwise aggregation/grouping. */
7650 1288 : for (cnt_parts = 0; cnt_parts < nparts; cnt_parts++)
7651 : {
7652 942 : RelOptInfo *child_input_rel = input_rel->part_rels[cnt_parts];
7653 942 : PathTarget *child_target = copy_pathtarget(target);
7654 : AppendRelInfo **appinfos;
7655 : int nappinfos;
7656 : GroupPathExtraData child_extra;
7657 : RelOptInfo *child_grouped_rel;
7658 : RelOptInfo *child_partially_grouped_rel;
7659 :
7660 : /* Pruned or dummy children can be ignored. */
7661 942 : if (child_input_rel == NULL || IS_DUMMY_REL(child_input_rel))
7662 0 : continue;
7663 :
7664 : /*
7665 : * Copy the given "extra" structure as is and then override the
7666 : * members specific to this child.
7667 : */
7668 942 : memcpy(&child_extra, extra, sizeof(child_extra));
7669 :
7670 942 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
7671 : &nappinfos);
7672 :
7673 942 : child_target->exprs = (List *)
7674 1884 : adjust_appendrel_attrs(root,
7675 942 : (Node *) target->exprs,
7676 : nappinfos, appinfos);
7677 :
7678 : /* Translate havingQual and targetList. */
7679 942 : child_extra.havingQual = (Node *)
7680 : adjust_appendrel_attrs(root,
7681 : extra->havingQual,
7682 : nappinfos, appinfos);
7683 942 : child_extra.targetList = (List *)
7684 1884 : adjust_appendrel_attrs(root,
7685 942 : (Node *) extra->targetList,
7686 : nappinfos, appinfos);
7687 :
7688 : /*
7689 : * extra->patype was the value computed for our parent rel; patype is
7690 : * the value for this relation. For the child, our value is its
7691 : * parent rel's value.
7692 : */
7693 942 : child_extra.patype = patype;
7694 :
7695 : /*
7696 : * Create grouping relation to hold fully aggregated grouping and/or
7697 : * aggregation paths for the child.
7698 : */
7699 1884 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
7700 : child_target,
7701 942 : extra->target_parallel_safe,
7702 : child_extra.havingQual);
7703 :
7704 : /* Create grouping paths for this child relation. */
7705 942 : create_ordinary_grouping_paths(root, child_input_rel,
7706 : child_grouped_rel,
7707 : agg_costs, gd, &child_extra,
7708 : &child_partially_grouped_rel);
7709 :
7710 942 : if (child_partially_grouped_rel)
7711 : {
7712 : partially_grouped_live_children =
7713 590 : lappend(partially_grouped_live_children,
7714 : child_partially_grouped_rel);
7715 : }
7716 : else
7717 352 : partial_grouping_valid = false;
7718 :
7719 942 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7720 : {
7721 544 : set_cheapest(child_grouped_rel);
7722 544 : grouped_live_children = lappend(grouped_live_children,
7723 : child_grouped_rel);
7724 : }
7725 :
7726 942 : pfree(appinfos);
7727 : }
7728 :
7729 : /*
7730 : * Try to create append paths for partially grouped children. For full
7731 : * partitionwise aggregation, we might have paths in the partial_pathlist
7732 : * if parallel aggregation is possible. For partial partitionwise
7733 : * aggregation, we may have paths in both pathlist and partial_pathlist.
7734 : *
7735 : * NB: We must have a partially grouped path for every child in order to
7736 : * generate a partially grouped path for this relation.
7737 : */
7738 346 : if (partially_grouped_rel && partial_grouping_valid)
7739 : {
7740 : Assert(partially_grouped_live_children != NIL);
7741 :
7742 226 : add_paths_to_append_rel(root, partially_grouped_rel,
7743 : partially_grouped_live_children);
7744 :
7745 : /*
7746 : * We need call set_cheapest, since the finalization step will use the
7747 : * cheapest path from the rel.
7748 : */
7749 226 : if (partially_grouped_rel->pathlist)
7750 226 : set_cheapest(partially_grouped_rel);
7751 : }
7752 :
7753 : /* If possible, create append paths for fully grouped children. */
7754 346 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7755 : {
7756 : Assert(grouped_live_children != NIL);
7757 :
7758 192 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
7759 : }
7760 346 : }
7761 :
7762 : /*
7763 : * group_by_has_partkey
7764 : *
7765 : * Returns true, if all the partition keys of the given relation are part of
7766 : * the GROUP BY clauses, false otherwise.
7767 : */
7768 : static bool
7769 342 : group_by_has_partkey(RelOptInfo *input_rel,
7770 : List *targetList,
7771 : List *groupClause)
7772 : {
7773 342 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
7774 342 : int cnt = 0;
7775 : int partnatts;
7776 :
7777 : /* Input relation should be partitioned. */
7778 : Assert(input_rel->part_scheme);
7779 :
7780 : /* Rule out early, if there are no partition keys present. */
7781 342 : if (!input_rel->partexprs)
7782 0 : return false;
7783 :
7784 342 : partnatts = input_rel->part_scheme->partnatts;
7785 :
7786 558 : for (cnt = 0; cnt < partnatts; cnt++)
7787 : {
7788 366 : List *partexprs = input_rel->partexprs[cnt];
7789 : ListCell *lc;
7790 366 : bool found = false;
7791 :
7792 512 : foreach(lc, partexprs)
7793 : {
7794 362 : Expr *partexpr = lfirst(lc);
7795 :
7796 362 : if (list_member(groupexprs, partexpr))
7797 : {
7798 216 : found = true;
7799 216 : break;
7800 : }
7801 : }
7802 :
7803 : /*
7804 : * If none of the partition key expressions match with any of the
7805 : * GROUP BY expression, return false.
7806 : */
7807 366 : if (!found)
7808 150 : return false;
7809 : }
7810 :
7811 192 : return true;
7812 : }
|
