Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * planner.c
4 : * The query optimizer external interface.
5 : *
6 : * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/optimizer/plan/planner.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 :
16 : #include "postgres.h"
17 :
18 : #include <limits.h>
19 : #include <math.h>
20 :
21 : #include "access/genam.h"
22 : #include "access/htup_details.h"
23 : #include "access/parallel.h"
24 : #include "access/sysattr.h"
25 : #include "access/table.h"
26 : #include "access/xact.h"
27 : #include "catalog/pg_constraint.h"
28 : #include "catalog/pg_inherits.h"
29 : #include "catalog/pg_proc.h"
30 : #include "catalog/pg_type.h"
31 : #include "executor/executor.h"
32 : #include "executor/nodeAgg.h"
33 : #include "foreign/fdwapi.h"
34 : #include "miscadmin.h"
35 : #include "jit/jit.h"
36 : #include "lib/bipartite_match.h"
37 : #include "lib/knapsack.h"
38 : #include "nodes/makefuncs.h"
39 : #include "nodes/nodeFuncs.h"
40 : #ifdef OPTIMIZER_DEBUG
41 : #include "nodes/print.h"
42 : #endif
43 : #include "optimizer/appendinfo.h"
44 : #include "optimizer/clauses.h"
45 : #include "optimizer/cost.h"
46 : #include "optimizer/inherit.h"
47 : #include "optimizer/optimizer.h"
48 : #include "optimizer/paramassign.h"
49 : #include "optimizer/pathnode.h"
50 : #include "optimizer/paths.h"
51 : #include "optimizer/plancat.h"
52 : #include "optimizer/planmain.h"
53 : #include "optimizer/planner.h"
54 : #include "optimizer/prep.h"
55 : #include "optimizer/subselect.h"
56 : #include "optimizer/tlist.h"
57 : #include "parser/analyze.h"
58 : #include "parser/parsetree.h"
59 : #include "parser/parse_agg.h"
60 : #include "partitioning/partdesc.h"
61 : #include "rewrite/rewriteManip.h"
62 : #include "storage/dsm_impl.h"
63 : #include "utils/rel.h"
64 : #include "utils/selfuncs.h"
65 : #include "utils/lsyscache.h"
66 : #include "utils/syscache.h"
67 :
68 :
69 : /* GUC parameters */
70 : double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
71 : int force_parallel_mode = FORCE_PARALLEL_OFF;
72 : bool parallel_leader_participation = true;
73 :
74 : /* Hook for plugins to get control in planner() */
75 : planner_hook_type planner_hook = NULL;
76 :
77 : /* Hook for plugins to get control when grouping_planner() plans upper rels */
78 : create_upper_paths_hook_type create_upper_paths_hook = NULL;
79 :
80 :
81 : /* Expression kind codes for preprocess_expression */
82 : #define EXPRKIND_QUAL 0
83 : #define EXPRKIND_TARGET 1
84 : #define EXPRKIND_RTFUNC 2
85 : #define EXPRKIND_RTFUNC_LATERAL 3
86 : #define EXPRKIND_VALUES 4
87 : #define EXPRKIND_VALUES_LATERAL 5
88 : #define EXPRKIND_LIMIT 6
89 : #define EXPRKIND_APPINFO 7
90 : #define EXPRKIND_PHV 8
91 : #define EXPRKIND_TABLESAMPLE 9
92 : #define EXPRKIND_ARBITER_ELEM 10
93 : #define EXPRKIND_TABLEFUNC 11
94 : #define EXPRKIND_TABLEFUNC_LATERAL 12
95 :
96 : /* Passthrough data for standard_qp_callback */
97 : typedef struct
98 : {
99 : List *activeWindows; /* active windows, if any */
100 : List *groupClause; /* overrides parse->groupClause */
101 : } standard_qp_extra;
102 :
103 : /*
104 : * Data specific to grouping sets
105 : */
106 :
107 : typedef struct
108 : {
109 : List *rollups;
110 : List *hash_sets_idx;
111 : double dNumHashGroups;
112 : bool any_hashable;
113 : Bitmapset *unsortable_refs;
114 : Bitmapset *unhashable_refs;
115 : List *unsortable_sets;
116 : int *tleref_to_colnum_map;
117 : } grouping_sets_data;
118 :
119 : /*
120 : * Temporary structure for use during WindowClause reordering in order to be
121 : * able to sort WindowClauses on partitioning/ordering prefix.
122 : */
123 : typedef struct
124 : {
125 : WindowClause *wc;
126 : List *uniqueOrder; /* A List of unique ordering/partitioning
127 : * clauses per Window */
128 : } WindowClauseSortData;
129 :
130 : /* Local functions */
131 : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
132 : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
133 : static void inheritance_planner(PlannerInfo *root);
134 : static void grouping_planner(PlannerInfo *root, bool inheritance_update,
135 : double tuple_fraction);
136 : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
137 : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
138 : int *tleref_to_colnum_map);
139 : static void preprocess_rowmarks(PlannerInfo *root);
140 : static double preprocess_limit(PlannerInfo *root,
141 : double tuple_fraction,
142 : int64 *offset_est, int64 *count_est);
143 : static void remove_useless_groupby_columns(PlannerInfo *root);
144 : static List *preprocess_groupclause(PlannerInfo *root, List *force);
145 : static List *extract_rollup_sets(List *groupingSets);
146 : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
147 : static void standard_qp_callback(PlannerInfo *root, void *extra);
148 : static double get_number_of_groups(PlannerInfo *root,
149 : double path_rows,
150 : grouping_sets_data *gd,
151 : List *target_list);
152 : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
153 : RelOptInfo *input_rel,
154 : PathTarget *target,
155 : bool target_parallel_safe,
156 : const AggClauseCosts *agg_costs,
157 : grouping_sets_data *gd);
158 : static bool is_degenerate_grouping(PlannerInfo *root);
159 : static void create_degenerate_grouping_paths(PlannerInfo *root,
160 : RelOptInfo *input_rel,
161 : RelOptInfo *grouped_rel);
162 : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
163 : PathTarget *target, bool target_parallel_safe,
164 : Node *havingQual);
165 : static void create_ordinary_grouping_paths(PlannerInfo *root,
166 : RelOptInfo *input_rel,
167 : RelOptInfo *grouped_rel,
168 : const AggClauseCosts *agg_costs,
169 : grouping_sets_data *gd,
170 : GroupPathExtraData *extra,
171 : RelOptInfo **partially_grouped_rel_p);
172 : static void consider_groupingsets_paths(PlannerInfo *root,
173 : RelOptInfo *grouped_rel,
174 : Path *path,
175 : bool is_sorted,
176 : bool can_hash,
177 : grouping_sets_data *gd,
178 : const AggClauseCosts *agg_costs,
179 : double dNumGroups);
180 : static RelOptInfo *create_window_paths(PlannerInfo *root,
181 : RelOptInfo *input_rel,
182 : PathTarget *input_target,
183 : PathTarget *output_target,
184 : bool output_target_parallel_safe,
185 : WindowFuncLists *wflists,
186 : List *activeWindows);
187 : static void create_one_window_path(PlannerInfo *root,
188 : RelOptInfo *window_rel,
189 : Path *path,
190 : PathTarget *input_target,
191 : PathTarget *output_target,
192 : WindowFuncLists *wflists,
193 : List *activeWindows);
194 : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
195 : RelOptInfo *input_rel);
196 : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
197 : RelOptInfo *input_rel,
198 : PathTarget *target,
199 : bool target_parallel_safe,
200 : double limit_tuples);
201 : static PathTarget *make_group_input_target(PlannerInfo *root,
202 : PathTarget *final_target);
203 : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
204 : PathTarget *grouping_target,
205 : Node *havingQual);
206 : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
207 : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
208 : static PathTarget *make_window_input_target(PlannerInfo *root,
209 : PathTarget *final_target,
210 : List *activeWindows);
211 : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
212 : List *tlist);
213 : static PathTarget *make_sort_input_target(PlannerInfo *root,
214 : PathTarget *final_target,
215 : bool *have_postponed_srfs);
216 : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
217 : List *targets, List *targets_contain_srfs);
218 : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
219 : RelOptInfo *grouped_rel,
220 : RelOptInfo *partially_grouped_rel,
221 : const AggClauseCosts *agg_costs,
222 : grouping_sets_data *gd,
223 : double dNumGroups,
224 : GroupPathExtraData *extra);
225 : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
226 : RelOptInfo *grouped_rel,
227 : RelOptInfo *input_rel,
228 : grouping_sets_data *gd,
229 : GroupPathExtraData *extra,
230 : bool force_rel_creation);
231 : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
232 : static bool can_partial_agg(PlannerInfo *root,
233 : const AggClauseCosts *agg_costs);
234 : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
235 : RelOptInfo *rel,
236 : List *scanjoin_targets,
237 : List *scanjoin_targets_contain_srfs,
238 : bool scanjoin_target_parallel_safe,
239 : bool tlist_same_exprs);
240 : static void create_partitionwise_grouping_paths(PlannerInfo *root,
241 : RelOptInfo *input_rel,
242 : RelOptInfo *grouped_rel,
243 : RelOptInfo *partially_grouped_rel,
244 : const AggClauseCosts *agg_costs,
245 : grouping_sets_data *gd,
246 : PartitionwiseAggregateType patype,
247 : GroupPathExtraData *extra);
248 : static bool group_by_has_partkey(RelOptInfo *input_rel,
249 : List *targetList,
250 : List *groupClause);
251 : static int common_prefix_cmp(const void *a, const void *b);
252 :
253 :
254 : /*****************************************************************************
255 : *
256 : * Query optimizer entry point
257 : *
258 : * To support loadable plugins that monitor or modify planner behavior,
259 : * we provide a hook variable that lets a plugin get control before and
260 : * after the standard planning process. The plugin would normally call
261 : * standard_planner().
262 : *
263 : * Note to plugin authors: standard_planner() scribbles on its Query input,
264 : * so you'd better copy that data structure if you want to plan more than once.
265 : *
266 : *****************************************************************************/
267 : PlannedStmt *
268 243250 : planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
269 : {
270 : PlannedStmt *result;
271 :
272 243250 : if (planner_hook)
273 0 : result = (*planner_hook) (parse, cursorOptions, boundParams);
274 : else
275 243250 : result = standard_planner(parse, cursorOptions, boundParams);
276 241748 : return result;
277 : }
278 :
279 : PlannedStmt *
280 243250 : standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
281 : {
282 : PlannedStmt *result;
283 : PlannerGlobal *glob;
284 : double tuple_fraction;
285 : PlannerInfo *root;
286 : RelOptInfo *final_rel;
287 : Path *best_path;
288 : Plan *top_plan;
289 : ListCell *lp,
290 : *lr;
291 :
292 : /*
293 : * Set up global state for this planner invocation. This data is needed
294 : * across all levels of sub-Query that might exist in the given command,
295 : * so we keep it in a separate struct that's linked to by each per-Query
296 : * PlannerInfo.
297 : */
298 243250 : glob = makeNode(PlannerGlobal);
299 :
300 243250 : glob->boundParams = boundParams;
301 243250 : glob->subplans = NIL;
302 243250 : glob->subroots = NIL;
303 243250 : glob->rewindPlanIDs = NULL;
304 243250 : glob->finalrtable = NIL;
305 243250 : glob->finalrowmarks = NIL;
306 243250 : glob->resultRelations = NIL;
307 243250 : glob->rootResultRelations = NIL;
308 243250 : glob->relationOids = NIL;
309 243250 : glob->invalItems = NIL;
310 243250 : glob->paramExecTypes = NIL;
311 243250 : glob->lastPHId = 0;
312 243250 : glob->lastRowMarkId = 0;
313 243250 : glob->lastPlanNodeId = 0;
314 243250 : glob->transientPlan = false;
315 243250 : glob->dependsOnRole = false;
316 :
317 : /*
318 : * Assess whether it's feasible to use parallel mode for this query. We
319 : * can't do this in a standalone backend, or if the command will try to
320 : * modify any data, or if this is a cursor operation, or if GUCs are set
321 : * to values that don't permit parallelism, or if parallel-unsafe
322 : * functions are present in the query tree.
323 : *
324 : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
325 : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
326 : * the command is writing into a completely new table which workers won't
327 : * be able to see. If the workers could see the table, the fact that
328 : * group locking would cause them to ignore the leader's heavyweight
329 : * relation extension lock and GIN page locks would make this unsafe.
330 : * We'll have to fix that somehow if we want to allow parallel inserts in
331 : * general; updates and deletes have additional problems especially around
332 : * combo CIDs.)
333 : *
334 : * For now, we don't try to use parallel mode if we're running inside a
335 : * parallel worker. We might eventually be able to relax this
336 : * restriction, but for now it seems best not to have parallel workers
337 : * trying to create their own parallel workers.
338 : */
339 243250 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
340 200598 : IsUnderPostmaster &&
341 360606 : parse->commandType == CMD_SELECT &&
342 319932 : !parse->hasModifyingCTE &&
343 319542 : max_parallel_workers_per_gather > 0 &&
344 159618 : !IsParallelWorker())
345 : {
346 : /* all the cheap tests pass, so scan the query tree */
347 159614 : glob->maxParallelHazard = max_parallel_hazard(parse);
348 159614 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
349 : }
350 : else
351 : {
352 : /* skip the query tree scan, just assume it's unsafe */
353 83636 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
354 83636 : glob->parallelModeOK = false;
355 : }
356 :
357 : /*
358 : * glob->parallelModeNeeded is normally set to false here and changed to
359 : * true during plan creation if a Gather or Gather Merge plan is actually
360 : * created (cf. create_gather_plan, create_gather_merge_plan).
361 : *
362 : * However, if force_parallel_mode = on or force_parallel_mode = regress,
363 : * then we impose parallel mode whenever it's safe to do so, even if the
364 : * final plan doesn't use parallelism. It's not safe to do so if the
365 : * query contains anything parallel-unsafe; parallelModeOK will be false
366 : * in that case. Note that parallelModeOK can't change after this point.
367 : * Otherwise, everything in the query is either parallel-safe or
368 : * parallel-restricted, and in either case it should be OK to impose
369 : * parallel-mode restrictions. If that ends up breaking something, then
370 : * either some function the user included in the query is incorrectly
371 : * labelled as parallel-safe or parallel-restricted when in reality it's
372 : * parallel-unsafe, or else the query planner itself has a bug.
373 : */
374 381068 : glob->parallelModeNeeded = glob->parallelModeOK &&
375 137818 : (force_parallel_mode != FORCE_PARALLEL_OFF);
376 :
377 : /* Determine what fraction of the plan is likely to be scanned */
378 243250 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
379 : {
380 : /*
381 : * We have no real idea how many tuples the user will ultimately FETCH
382 : * from a cursor, but it is often the case that he doesn't want 'em
383 : * all, or would prefer a fast-start plan anyway so that he can
384 : * process some of the tuples sooner. Use a GUC parameter to decide
385 : * what fraction to optimize for.
386 : */
387 2062 : tuple_fraction = cursor_tuple_fraction;
388 :
389 : /*
390 : * We document cursor_tuple_fraction as simply being a fraction, which
391 : * means the edge cases 0 and 1 have to be treated specially here. We
392 : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
393 : */
394 2062 : if (tuple_fraction >= 1.0)
395 0 : tuple_fraction = 0.0;
396 2062 : else if (tuple_fraction <= 0.0)
397 0 : tuple_fraction = 1e-10;
398 : }
399 : else
400 : {
401 : /* Default assumption is we need all the tuples */
402 241188 : tuple_fraction = 0.0;
403 : }
404 :
405 : /* primary planning entry point (may recurse for subqueries) */
406 243250 : root = subquery_planner(glob, parse, NULL,
407 : false, tuple_fraction);
408 :
409 : /* Select best Path and turn it into a Plan */
410 241868 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
411 241868 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
412 :
413 241868 : top_plan = create_plan(root, best_path);
414 :
415 : /*
416 : * If creating a plan for a scrollable cursor, make sure it can run
417 : * backwards on demand. Add a Material node at the top at need.
418 : */
419 241748 : if (cursorOptions & CURSOR_OPT_SCROLL)
420 : {
421 148 : if (!ExecSupportsBackwardScan(top_plan))
422 16 : top_plan = materialize_finished_plan(top_plan);
423 : }
424 :
425 : /*
426 : * Optionally add a Gather node for testing purposes, provided this is
427 : * actually a safe thing to do.
428 : */
429 241748 : if (force_parallel_mode != FORCE_PARALLEL_OFF && top_plan->parallel_safe)
430 : {
431 14 : Gather *gather = makeNode(Gather);
432 :
433 : /*
434 : * If there are any initPlans attached to the formerly-top plan node,
435 : * move them up to the Gather node; same as we do for Material node in
436 : * materialize_finished_plan.
437 : */
438 14 : gather->plan.initPlan = top_plan->initPlan;
439 14 : top_plan->initPlan = NIL;
440 :
441 14 : gather->plan.targetlist = top_plan->targetlist;
442 14 : gather->plan.qual = NIL;
443 14 : gather->plan.lefttree = top_plan;
444 14 : gather->plan.righttree = NULL;
445 14 : gather->num_workers = 1;
446 14 : gather->single_copy = true;
447 14 : gather->invisible = (force_parallel_mode == FORCE_PARALLEL_REGRESS);
448 :
449 : /*
450 : * Since this Gather has no parallel-aware descendants to signal to,
451 : * we don't need a rescan Param.
452 : */
453 14 : gather->rescan_param = -1;
454 :
455 : /*
456 : * Ideally we'd use cost_gather here, but setting up dummy path data
457 : * to satisfy it doesn't seem much cleaner than knowing what it does.
458 : */
459 14 : gather->plan.startup_cost = top_plan->startup_cost +
460 : parallel_setup_cost;
461 28 : gather->plan.total_cost = top_plan->total_cost +
462 14 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
463 14 : gather->plan.plan_rows = top_plan->plan_rows;
464 14 : gather->plan.plan_width = top_plan->plan_width;
465 14 : gather->plan.parallel_aware = false;
466 14 : gather->plan.parallel_safe = false;
467 :
468 : /* use parallel mode for parallel plans. */
469 14 : root->glob->parallelModeNeeded = true;
470 :
471 14 : top_plan = &gather->plan;
472 : }
473 :
474 : /*
475 : * If any Params were generated, run through the plan tree and compute
476 : * each plan node's extParam/allParam sets. Ideally we'd merge this into
477 : * set_plan_references' tree traversal, but for now it has to be separate
478 : * because we need to visit subplans before not after main plan.
479 : */
480 241748 : if (glob->paramExecTypes != NIL)
481 : {
482 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
483 156236 : forboth(lp, glob->subplans, lr, glob->subroots)
484 : {
485 45474 : Plan *subplan = (Plan *) lfirst(lp);
486 45474 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
487 :
488 45474 : SS_finalize_plan(subroot, subplan);
489 : }
490 110762 : SS_finalize_plan(root, top_plan);
491 : }
492 :
493 : /* final cleanup of the plan */
494 : Assert(glob->finalrtable == NIL);
495 : Assert(glob->finalrowmarks == NIL);
496 : Assert(glob->resultRelations == NIL);
497 : Assert(glob->rootResultRelations == NIL);
498 241748 : 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 287222 : forboth(lp, glob->subplans, lr, glob->subroots)
502 : {
503 45474 : Plan *subplan = (Plan *) lfirst(lp);
504 45474 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
505 :
506 45474 : lfirst(lp) = set_plan_references(subroot, subplan);
507 : }
508 :
509 : /* build the PlannedStmt result */
510 241748 : result = makeNode(PlannedStmt);
511 :
512 241748 : result->commandType = parse->commandType;
513 241748 : result->queryId = parse->queryId;
514 241748 : result->hasReturning = (parse->returningList != NIL);
515 241748 : result->hasModifyingCTE = parse->hasModifyingCTE;
516 241748 : result->canSetTag = parse->canSetTag;
517 241748 : result->transientPlan = glob->transientPlan;
518 241748 : result->dependsOnRole = glob->dependsOnRole;
519 241748 : result->parallelModeNeeded = glob->parallelModeNeeded;
520 241748 : result->planTree = top_plan;
521 241748 : result->rtable = glob->finalrtable;
522 241748 : result->resultRelations = glob->resultRelations;
523 241748 : result->rootResultRelations = glob->rootResultRelations;
524 241748 : result->subplans = glob->subplans;
525 241748 : result->rewindPlanIDs = glob->rewindPlanIDs;
526 241748 : result->rowMarks = glob->finalrowmarks;
527 241748 : result->relationOids = glob->relationOids;
528 241748 : result->invalItems = glob->invalItems;
529 241748 : result->paramExecTypes = glob->paramExecTypes;
530 : /* utilityStmt should be null, but we might as well copy it */
531 241748 : result->utilityStmt = parse->utilityStmt;
532 241748 : result->stmt_location = parse->stmt_location;
533 241748 : result->stmt_len = parse->stmt_len;
534 :
535 241748 : result->jitFlags = PGJIT_NONE;
536 483496 : if (jit_enabled && jit_above_cost >= 0 &&
537 241748 : top_plan->total_cost > jit_above_cost)
538 : {
539 526 : result->jitFlags |= PGJIT_PERFORM;
540 :
541 : /*
542 : * Decide how much effort should be put into generating better code.
543 : */
544 1052 : if (jit_optimize_above_cost >= 0 &&
545 526 : top_plan->total_cost > jit_optimize_above_cost)
546 438 : result->jitFlags |= PGJIT_OPT3;
547 1052 : if (jit_inline_above_cost >= 0 &&
548 526 : top_plan->total_cost > jit_inline_above_cost)
549 438 : result->jitFlags |= PGJIT_INLINE;
550 :
551 : /*
552 : * Decide which operations should be JITed.
553 : */
554 526 : if (jit_expressions)
555 526 : result->jitFlags |= PGJIT_EXPR;
556 526 : if (jit_tuple_deforming)
557 526 : result->jitFlags |= PGJIT_DEFORM;
558 : }
559 :
560 241748 : if (glob->partition_directory != NULL)
561 3214 : DestroyPartitionDirectory(glob->partition_directory);
562 :
563 241748 : return result;
564 : }
565 :
566 :
567 : /*--------------------
568 : * subquery_planner
569 : * Invokes the planner on a subquery. We recurse to here for each
570 : * sub-SELECT found in the query tree.
571 : *
572 : * glob is the global state for the current planner run.
573 : * parse is the querytree produced by the parser & rewriter.
574 : * parent_root is the immediate parent Query's info (NULL at the top level).
575 : * hasRecursion is true if this is a recursive WITH query.
576 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
577 : * tuple_fraction is interpreted as explained for grouping_planner, below.
578 : *
579 : * Basically, this routine does the stuff that should only be done once
580 : * per Query object. It then calls grouping_planner. At one time,
581 : * grouping_planner could be invoked recursively on the same Query object;
582 : * that's not currently true, but we keep the separation between the two
583 : * routines anyway, in case we need it again someday.
584 : *
585 : * subquery_planner will be called recursively to handle sub-Query nodes
586 : * found within the query's expressions and rangetable.
587 : *
588 : * Returns the PlannerInfo struct ("root") that contains all data generated
589 : * while planning the subquery. In particular, the Path(s) attached to
590 : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
591 : * cheapest way(s) to implement the query. The top level will select the
592 : * best Path and pass it through createplan.c to produce a finished Plan.
593 : *--------------------
594 : */
595 : PlannerInfo *
596 296118 : subquery_planner(PlannerGlobal *glob, Query *parse,
597 : PlannerInfo *parent_root,
598 : bool hasRecursion, double tuple_fraction)
599 : {
600 : PlannerInfo *root;
601 : List *newWithCheckOptions;
602 : List *newHaving;
603 : bool hasOuterJoins;
604 : bool hasResultRTEs;
605 : RelOptInfo *final_rel;
606 : ListCell *l;
607 :
608 : /* Create a PlannerInfo data structure for this subquery */
609 296118 : root = makeNode(PlannerInfo);
610 296118 : root->parse = parse;
611 296118 : root->glob = glob;
612 296118 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
613 296118 : root->parent_root = parent_root;
614 296118 : root->plan_params = NIL;
615 296118 : root->outer_params = NULL;
616 296118 : root->planner_cxt = CurrentMemoryContext;
617 296118 : root->init_plans = NIL;
618 296118 : root->cte_plan_ids = NIL;
619 296118 : root->multiexpr_params = NIL;
620 296118 : root->eq_classes = NIL;
621 296118 : root->append_rel_list = NIL;
622 296118 : root->rowMarks = NIL;
623 296118 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
624 296118 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
625 296118 : root->processed_tlist = NIL;
626 296118 : root->grouping_map = NULL;
627 296118 : root->minmax_aggs = NIL;
628 296118 : root->qual_security_level = 0;
629 296118 : root->inhTargetKind = INHKIND_NONE;
630 296118 : root->hasRecursion = hasRecursion;
631 296118 : if (hasRecursion)
632 318 : root->wt_param_id = assign_special_exec_param(root);
633 : else
634 295800 : root->wt_param_id = -1;
635 296118 : root->non_recursive_path = NULL;
636 296118 : root->partColsUpdated = false;
637 :
638 : /*
639 : * If there is a WITH list, process each WITH query and either convert it
640 : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
641 : */
642 296118 : if (parse->cteList)
643 1162 : SS_process_ctes(root);
644 :
645 : /*
646 : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
647 : * that we don't need so many special cases to deal with that situation.
648 : */
649 296118 : replace_empty_jointree(parse);
650 :
651 : /*
652 : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
653 : * to transform them into joins. Note that this step does not descend
654 : * into subqueries; if we pull up any subqueries below, their SubLinks are
655 : * processed just before pulling them up.
656 : */
657 296118 : if (parse->hasSubLinks)
658 29268 : pull_up_sublinks(root);
659 :
660 : /*
661 : * Scan the rangetable for set-returning functions, and inline them if
662 : * possible (producing subqueries that might get pulled up next).
663 : * Recursion issues here are handled in the same way as for SubLinks.
664 : */
665 296118 : inline_set_returning_functions(root);
666 :
667 : /*
668 : * Check to see if any subqueries in the jointree can be merged into this
669 : * query.
670 : */
671 296118 : pull_up_subqueries(root);
672 :
673 : /*
674 : * If this is a simple UNION ALL query, flatten it into an appendrel. We
675 : * do this now because it requires applying pull_up_subqueries to the leaf
676 : * queries of the UNION ALL, which weren't touched above because they
677 : * weren't referenced by the jointree (they will be after we do this).
678 : */
679 296114 : if (parse->setOperations)
680 2262 : flatten_simple_union_all(root);
681 :
682 : /*
683 : * Survey the rangetable to see what kinds of entries are present. We can
684 : * skip some later processing if relevant SQL features are not used; for
685 : * example if there are no JOIN RTEs we can avoid the expense of doing
686 : * flatten_join_alias_vars(). This must be done after we have finished
687 : * adding rangetable entries, of course. (Note: actually, processing of
688 : * inherited or partitioned rels can cause RTEs for their child tables to
689 : * get added later; but those must all be RTE_RELATION entries, so they
690 : * don't invalidate the conclusions drawn here.)
691 : */
692 296114 : root->hasJoinRTEs = false;
693 296114 : root->hasLateralRTEs = false;
694 296114 : hasOuterJoins = false;
695 296114 : hasResultRTEs = false;
696 846502 : foreach(l, parse->rtable)
697 : {
698 550388 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
699 :
700 550388 : switch (rte->rtekind)
701 : {
702 : case RTE_RELATION:
703 309146 : if (rte->inh)
704 : {
705 : /*
706 : * Check to see if the relation actually has any children;
707 : * if not, clear the inh flag so we can treat it as a
708 : * plain base relation.
709 : *
710 : * Note: this could give a false-positive result, if the
711 : * rel once had children but no longer does. We used to
712 : * be able to clear rte->inh later on when we discovered
713 : * that, but no more; we have to handle such cases as
714 : * full-fledged inheritance.
715 : */
716 212296 : rte->inh = has_subclass(rte->relid);
717 : }
718 309146 : break;
719 : case RTE_JOIN:
720 48180 : root->hasJoinRTEs = true;
721 48180 : if (IS_OUTER_JOIN(rte->jointype))
722 34594 : hasOuterJoins = true;
723 48180 : break;
724 : case RTE_RESULT:
725 112606 : hasResultRTEs = true;
726 112606 : break;
727 : default:
728 : /* No work here for other RTE types */
729 80456 : break;
730 : }
731 :
732 550388 : if (rte->lateral)
733 992 : root->hasLateralRTEs = true;
734 :
735 : /*
736 : * We can also determine the maximum security level required for any
737 : * securityQuals now. Addition of inheritance-child RTEs won't affect
738 : * this, because child tables don't have their own securityQuals; see
739 : * expand_single_inheritance_child().
740 : */
741 550388 : if (rte->securityQuals)
742 1140 : root->qual_security_level = Max(root->qual_security_level,
743 : list_length(rte->securityQuals));
744 : }
745 :
746 : /*
747 : * Preprocess RowMark information. We need to do this after subquery
748 : * pullup, so that all base relations are present.
749 : */
750 296114 : preprocess_rowmarks(root);
751 :
752 : /*
753 : * Set hasHavingQual to remember if HAVING clause is present. Needed
754 : * because preprocess_expression will reduce a constant-true condition to
755 : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
756 : */
757 296114 : root->hasHavingQual = (parse->havingQual != NULL);
758 :
759 : /* Clear this flag; might get set in distribute_qual_to_rels */
760 296114 : root->hasPseudoConstantQuals = false;
761 :
762 : /*
763 : * Do expression preprocessing on targetlist and quals, as well as other
764 : * random expressions in the querytree. Note that we do not need to
765 : * handle sort/group expressions explicitly, because they are actually
766 : * part of the targetlist.
767 : */
768 294754 : parse->targetList = (List *)
769 296114 : preprocess_expression(root, (Node *) parse->targetList,
770 : EXPRKIND_TARGET);
771 :
772 : /* Constant-folding might have removed all set-returning functions */
773 294754 : if (parse->hasTargetSRFs)
774 3276 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
775 :
776 294754 : newWithCheckOptions = NIL;
777 295896 : foreach(l, parse->withCheckOptions)
778 : {
779 1142 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
780 :
781 1142 : wco->qual = preprocess_expression(root, wco->qual,
782 : EXPRKIND_QUAL);
783 1142 : if (wco->qual != NULL)
784 922 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
785 : }
786 294754 : parse->withCheckOptions = newWithCheckOptions;
787 :
788 294754 : parse->returningList = (List *)
789 294754 : preprocess_expression(root, (Node *) parse->returningList,
790 : EXPRKIND_TARGET);
791 :
792 294754 : preprocess_qual_conditions(root, (Node *) parse->jointree);
793 :
794 294754 : parse->havingQual = preprocess_expression(root, parse->havingQual,
795 : EXPRKIND_QUAL);
796 :
797 295926 : foreach(l, parse->windowClause)
798 : {
799 1172 : WindowClause *wc = lfirst_node(WindowClause, l);
800 :
801 : /* partitionClause/orderClause are sort/group expressions */
802 1172 : wc->startOffset = preprocess_expression(root, wc->startOffset,
803 : EXPRKIND_LIMIT);
804 1172 : wc->endOffset = preprocess_expression(root, wc->endOffset,
805 : EXPRKIND_LIMIT);
806 : }
807 :
808 294754 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
809 : EXPRKIND_LIMIT);
810 294754 : parse->limitCount = preprocess_expression(root, parse->limitCount,
811 : EXPRKIND_LIMIT);
812 :
813 294754 : if (parse->onConflict)
814 : {
815 1920 : parse->onConflict->arbiterElems = (List *)
816 960 : preprocess_expression(root,
817 960 : (Node *) parse->onConflict->arbiterElems,
818 : EXPRKIND_ARBITER_ELEM);
819 1920 : parse->onConflict->arbiterWhere =
820 960 : preprocess_expression(root,
821 960 : parse->onConflict->arbiterWhere,
822 : EXPRKIND_QUAL);
823 1920 : parse->onConflict->onConflictSet = (List *)
824 960 : preprocess_expression(root,
825 960 : (Node *) parse->onConflict->onConflictSet,
826 : EXPRKIND_TARGET);
827 1920 : parse->onConflict->onConflictWhere =
828 960 : preprocess_expression(root,
829 960 : parse->onConflict->onConflictWhere,
830 : EXPRKIND_QUAL);
831 : /* exclRelTlist contains only Vars, so no preprocessing needed */
832 : }
833 :
834 294754 : root->append_rel_list = (List *)
835 294754 : preprocess_expression(root, (Node *) root->append_rel_list,
836 : EXPRKIND_APPINFO);
837 :
838 : /* Also need to preprocess expressions within RTEs */
839 843672 : foreach(l, parse->rtable)
840 : {
841 548918 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
842 : int kind;
843 : ListCell *lcsq;
844 :
845 548918 : if (rte->rtekind == RTE_RELATION)
846 : {
847 309032 : if (rte->tablesample)
848 152 : rte->tablesample = (TableSampleClause *)
849 152 : preprocess_expression(root,
850 152 : (Node *) rte->tablesample,
851 : EXPRKIND_TABLESAMPLE);
852 : }
853 239886 : else if (rte->rtekind == RTE_SUBQUERY)
854 : {
855 : /*
856 : * We don't want to do all preprocessing yet on the subquery's
857 : * expressions, since that will happen when we plan it. But if it
858 : * contains any join aliases of our level, those have to get
859 : * expanded now, because planning of the subquery won't do it.
860 : * That's only possible if the subquery is LATERAL.
861 : */
862 46888 : if (rte->lateral && root->hasJoinRTEs)
863 304 : rte->subquery = (Query *)
864 304 : flatten_join_alias_vars(root->parse,
865 304 : (Node *) rte->subquery);
866 : }
867 192998 : else if (rte->rtekind == RTE_FUNCTION)
868 : {
869 : /* Preprocess the function expression(s) fully */
870 28290 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
871 28290 : rte->functions = (List *)
872 28290 : preprocess_expression(root, (Node *) rte->functions, kind);
873 : }
874 164708 : else if (rte->rtekind == RTE_TABLEFUNC)
875 : {
876 : /* Preprocess the function expression(s) fully */
877 140 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
878 140 : rte->tablefunc = (TableFunc *)
879 140 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
880 : }
881 164568 : else if (rte->rtekind == RTE_VALUES)
882 : {
883 : /* Preprocess the values lists fully */
884 3734 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
885 3734 : rte->values_lists = (List *)
886 3734 : preprocess_expression(root, (Node *) rte->values_lists, kind);
887 : }
888 :
889 : /*
890 : * Process each element of the securityQuals list as if it were a
891 : * separate qual expression (as indeed it is). We need to do it this
892 : * way to get proper canonicalization of AND/OR structure. Note that
893 : * this converts each element into an implicit-AND sublist.
894 : */
895 550214 : foreach(lcsq, rte->securityQuals)
896 : {
897 1296 : lfirst(lcsq) = preprocess_expression(root,
898 1296 : (Node *) lfirst(lcsq),
899 : EXPRKIND_QUAL);
900 : }
901 : }
902 :
903 : /*
904 : * Now that we are done preprocessing expressions, and in particular done
905 : * flattening join alias variables, get rid of the joinaliasvars lists.
906 : * They no longer match what expressions in the rest of the tree look
907 : * like, because we have not preprocessed expressions in those lists (and
908 : * do not want to; for example, expanding a SubLink there would result in
909 : * a useless unreferenced subplan). Leaving them in place simply creates
910 : * a hazard for later scans of the tree. We could try to prevent that by
911 : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
912 : * but that doesn't sound very reliable.
913 : */
914 294754 : if (root->hasJoinRTEs)
915 : {
916 208638 : foreach(l, parse->rtable)
917 : {
918 172220 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
919 :
920 172220 : rte->joinaliasvars = NIL;
921 : }
922 : }
923 :
924 : /*
925 : * In some cases we may want to transfer a HAVING clause into WHERE. We
926 : * cannot do so if the HAVING clause contains aggregates (obviously) or
927 : * volatile functions (since a HAVING clause is supposed to be executed
928 : * only once per group). We also can't do this if there are any nonempty
929 : * grouping sets; moving such a clause into WHERE would potentially change
930 : * the results, if any referenced column isn't present in all the grouping
931 : * sets. (If there are only empty grouping sets, then the HAVING clause
932 : * must be degenerate as discussed below.)
933 : *
934 : * Also, it may be that the clause is so expensive to execute that we're
935 : * better off doing it only once per group, despite the loss of
936 : * selectivity. This is hard to estimate short of doing the entire
937 : * planning process twice, so we use a heuristic: clauses containing
938 : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
939 : * clause into WHERE, in hopes of eliminating tuples before aggregation
940 : * instead of after.
941 : *
942 : * If the query has explicit grouping then we can simply move such a
943 : * clause into WHERE; any group that fails the clause will not be in the
944 : * output because none of its tuples will reach the grouping or
945 : * aggregation stage. Otherwise we must have a degenerate (variable-free)
946 : * HAVING clause, which we put in WHERE so that query_planner() can use it
947 : * in a gating Result node, but also keep in HAVING to ensure that we
948 : * don't emit a bogus aggregated row. (This could be done better, but it
949 : * seems not worth optimizing.)
950 : *
951 : * Note that both havingQual and parse->jointree->quals are in
952 : * implicitly-ANDed-list form at this point, even though they are declared
953 : * as Node *.
954 : */
955 294754 : newHaving = NIL;
956 295200 : foreach(l, (List *) parse->havingQual)
957 : {
958 446 : Node *havingclause = (Node *) lfirst(l);
959 :
960 864 : if ((parse->groupClause && parse->groupingSets) ||
961 542 : contain_agg_clause(havingclause) ||
962 248 : contain_volatile_functions(havingclause) ||
963 124 : contain_subplans(havingclause))
964 : {
965 : /* keep it in HAVING */
966 322 : newHaving = lappend(newHaving, havingclause);
967 : }
968 124 : else if (parse->groupClause && !parse->groupingSets)
969 : {
970 : /* move it to WHERE */
971 224 : parse->jointree->quals = (Node *)
972 112 : lappend((List *) parse->jointree->quals, havingclause);
973 : }
974 : else
975 : {
976 : /* put a copy in WHERE, keep it in HAVING */
977 24 : parse->jointree->quals = (Node *)
978 12 : lappend((List *) parse->jointree->quals,
979 : copyObject(havingclause));
980 12 : newHaving = lappend(newHaving, havingclause);
981 : }
982 : }
983 294754 : parse->havingQual = (Node *) newHaving;
984 :
985 : /* Remove any redundant GROUP BY columns */
986 294754 : remove_useless_groupby_columns(root);
987 :
988 : /*
989 : * If we have any outer joins, try to reduce them to plain inner joins.
990 : * This step is most easily done after we've done expression
991 : * preprocessing.
992 : */
993 294754 : if (hasOuterJoins)
994 29626 : reduce_outer_joins(root);
995 :
996 : /*
997 : * If we have any RTE_RESULT relations, see if they can be deleted from
998 : * the jointree. This step is most effectively done after we've done
999 : * expression preprocessing and outer join reduction.
1000 : */
1001 294754 : if (hasResultRTEs)
1002 111062 : remove_useless_result_rtes(root);
1003 :
1004 : /*
1005 : * Do the main planning. If we have an inherited target relation, that
1006 : * needs special processing, else go straight to grouping_planner.
1007 : */
1008 367742 : if (parse->resultRelation &&
1009 72988 : rt_fetch(parse->resultRelation, parse->rtable)->inh)
1010 996 : inheritance_planner(root);
1011 : else
1012 293758 : grouping_planner(root, false, tuple_fraction);
1013 :
1014 : /*
1015 : * Capture the set of outer-level param IDs we have access to, for use in
1016 : * extParam/allParam calculations later.
1017 : */
1018 294736 : SS_identify_outer_params(root);
1019 :
1020 : /*
1021 : * If any initPlans were created in this query level, adjust the surviving
1022 : * Paths' costs and parallel-safety flags to account for them. The
1023 : * initPlans won't actually get attached to the plan tree till
1024 : * create_plan() runs, but we must include their effects now.
1025 : */
1026 294736 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1027 294736 : SS_charge_for_initplans(root, final_rel);
1028 :
1029 : /*
1030 : * Make sure we've identified the cheapest Path for the final rel. (By
1031 : * doing this here not in grouping_planner, we include initPlan costs in
1032 : * the decision, though it's unlikely that will change anything.)
1033 : */
1034 294736 : set_cheapest(final_rel);
1035 :
1036 294736 : return root;
1037 : }
1038 :
1039 : /*
1040 : * preprocess_expression
1041 : * Do subquery_planner's preprocessing work for an expression,
1042 : * which can be a targetlist, a WHERE clause (including JOIN/ON
1043 : * conditions), a HAVING clause, or a few other things.
1044 : */
1045 : static Node *
1046 2175406 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1047 : {
1048 : /*
1049 : * Fall out quickly if expression is empty. This occurs often enough to
1050 : * be worth checking. Note that null->null is the correct conversion for
1051 : * implicit-AND result format, too.
1052 : */
1053 2175406 : if (expr == NULL)
1054 1659914 : return NULL;
1055 :
1056 : /*
1057 : * If the query has any join RTEs, replace join alias variables with
1058 : * base-relation variables. We must do this first, since any expressions
1059 : * we may extract from the joinaliasvars lists have not been preprocessed.
1060 : * For example, if we did this after sublink processing, sublinks expanded
1061 : * out from join aliases would not get processed. But we can skip this in
1062 : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1063 : * they can't contain any Vars of the current query level.
1064 : */
1065 515492 : if (root->hasJoinRTEs &&
1066 236872 : !(kind == EXPRKIND_RTFUNC ||
1067 118346 : kind == EXPRKIND_VALUES ||
1068 : kind == EXPRKIND_TABLESAMPLE ||
1069 : kind == EXPRKIND_TABLEFUNC))
1070 118346 : expr = flatten_join_alias_vars(root->parse, expr);
1071 :
1072 : /*
1073 : * Simplify constant expressions.
1074 : *
1075 : * Note: an essential effect of this is to convert named-argument function
1076 : * calls to positional notation and insert the current actual values of
1077 : * any default arguments for functions. To ensure that happens, we *must*
1078 : * process all expressions here. Previous PG versions sometimes skipped
1079 : * const-simplification if it didn't seem worth the trouble, but we can't
1080 : * do that anymore.
1081 : *
1082 : * Note: this also flattens nested AND and OR expressions into N-argument
1083 : * form. All processing of a qual expression after this point must be
1084 : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1085 : * with AND directly under AND, nor OR directly under OR.
1086 : */
1087 515492 : expr = eval_const_expressions(root, expr);
1088 :
1089 : /*
1090 : * If it's a qual or havingQual, canonicalize it.
1091 : */
1092 514132 : if (kind == EXPRKIND_QUAL)
1093 : {
1094 183014 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1095 :
1096 : #ifdef OPTIMIZER_DEBUG
1097 : printf("After canonicalize_qual()\n");
1098 : pprint(expr);
1099 : #endif
1100 : }
1101 :
1102 : /* Expand SubLinks to SubPlans */
1103 514132 : if (root->parse->hasSubLinks)
1104 102686 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1105 :
1106 : /*
1107 : * XXX do not insert anything here unless you have grokked the comments in
1108 : * SS_replace_correlation_vars ...
1109 : */
1110 :
1111 : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1112 514132 : if (root->query_level > 1)
1113 124984 : expr = SS_replace_correlation_vars(root, expr);
1114 :
1115 : /*
1116 : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1117 : * don't want to do this before eval_const_expressions, since the latter
1118 : * would be unable to simplify a top-level AND correctly. Also,
1119 : * SS_process_sublinks expects explicit-AND format.)
1120 : */
1121 514132 : if (kind == EXPRKIND_QUAL)
1122 183014 : expr = (Node *) make_ands_implicit((Expr *) expr);
1123 :
1124 514132 : return expr;
1125 : }
1126 :
1127 : /*
1128 : * preprocess_qual_conditions
1129 : * Recursively scan the query's jointree and do subquery_planner's
1130 : * preprocessing work on each qual condition found therein.
1131 : */
1132 : static void
1133 728910 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1134 : {
1135 728910 : if (jtnode == NULL)
1136 0 : return;
1137 728910 : if (IsA(jtnode, RangeTblRef))
1138 : {
1139 : /* nothing to do here */
1140 : }
1141 364560 : else if (IsA(jtnode, FromExpr))
1142 : {
1143 308270 : FromExpr *f = (FromExpr *) jtnode;
1144 : ListCell *l;
1145 :
1146 629846 : foreach(l, f->fromlist)
1147 321576 : preprocess_qual_conditions(root, lfirst(l));
1148 :
1149 308270 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1150 : }
1151 56290 : else if (IsA(jtnode, JoinExpr))
1152 : {
1153 56290 : JoinExpr *j = (JoinExpr *) jtnode;
1154 :
1155 56290 : preprocess_qual_conditions(root, j->larg);
1156 56290 : preprocess_qual_conditions(root, j->rarg);
1157 :
1158 56290 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1159 : }
1160 : else
1161 0 : elog(ERROR, "unrecognized node type: %d",
1162 : (int) nodeTag(jtnode));
1163 : }
1164 :
1165 : /*
1166 : * preprocess_phv_expression
1167 : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1168 : *
1169 : * If a LATERAL subquery references an output of another subquery, and that
1170 : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1171 : * join, then we'll push the PlaceHolderVar expression down into the subquery
1172 : * and later pull it back up during find_lateral_references, which runs after
1173 : * subquery_planner has preprocessed all the expressions that were in the
1174 : * current query level to start with. So we need to preprocess it then.
1175 : */
1176 : Expr *
1177 24 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1178 : {
1179 24 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1180 : }
1181 :
1182 : /*
1183 : * inheritance_planner
1184 : * Generate Paths in the case where the result relation is an
1185 : * inheritance set.
1186 : *
1187 : * We have to handle this case differently from cases where a source relation
1188 : * is an inheritance set. Source inheritance is expanded at the bottom of the
1189 : * plan tree (see allpaths.c), but target inheritance has to be expanded at
1190 : * the top. The reason is that for UPDATE, each target relation needs a
1191 : * different targetlist matching its own column set. Fortunately,
1192 : * the UPDATE/DELETE target can never be the nullable side of an outer join,
1193 : * so it's OK to generate the plan this way.
1194 : *
1195 : * Returns nothing; the useful output is in the Paths we attach to
1196 : * the (UPPERREL_FINAL, NULL) upperrel stored in *root.
1197 : *
1198 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1199 : * to leave this to the caller.
1200 : */
1201 : static void
1202 996 : inheritance_planner(PlannerInfo *root)
1203 : {
1204 996 : Query *parse = root->parse;
1205 996 : int top_parentRTindex = parse->resultRelation;
1206 : List *select_rtable;
1207 : List *select_appinfos;
1208 : List *child_appinfos;
1209 : List *old_child_rtis;
1210 : List *new_child_rtis;
1211 : Bitmapset *subqueryRTindexes;
1212 : Index next_subquery_rti;
1213 996 : int nominalRelation = -1;
1214 996 : Index rootRelation = 0;
1215 996 : List *final_rtable = NIL;
1216 996 : List *final_rowmarks = NIL;
1217 996 : int save_rel_array_size = 0;
1218 996 : RelOptInfo **save_rel_array = NULL;
1219 996 : AppendRelInfo **save_append_rel_array = NULL;
1220 996 : List *subpaths = NIL;
1221 996 : List *subroots = NIL;
1222 996 : List *resultRelations = NIL;
1223 996 : List *withCheckOptionLists = NIL;
1224 996 : List *returningLists = NIL;
1225 : List *rowMarks;
1226 : RelOptInfo *final_rel;
1227 : ListCell *lc;
1228 : ListCell *lc2;
1229 : Index rti;
1230 : RangeTblEntry *parent_rte;
1231 : Bitmapset *parent_relids;
1232 : Query **parent_parses;
1233 :
1234 : /* Should only get here for UPDATE or DELETE */
1235 : Assert(parse->commandType == CMD_UPDATE ||
1236 : parse->commandType == CMD_DELETE);
1237 :
1238 : /*
1239 : * We generate a modified instance of the original Query for each target
1240 : * relation, plan that, and put all the plans into a list that will be
1241 : * controlled by a single ModifyTable node. All the instances share the
1242 : * same rangetable, but each instance must have its own set of subquery
1243 : * RTEs within the finished rangetable because (1) they are likely to get
1244 : * scribbled on during planning, and (2) it's not inconceivable that
1245 : * subqueries could get planned differently in different cases. We need
1246 : * not create duplicate copies of other RTE kinds, in particular not the
1247 : * target relations, because they don't have either of those issues. Not
1248 : * having to duplicate the target relations is important because doing so
1249 : * (1) would result in a rangetable of length O(N^2) for N targets, with
1250 : * at least O(N^3) work expended here; and (2) would greatly complicate
1251 : * management of the rowMarks list.
1252 : *
1253 : * To begin with, generate a bitmapset of the relids of the subquery RTEs.
1254 : */
1255 996 : subqueryRTindexes = NULL;
1256 996 : rti = 1;
1257 2416 : foreach(lc, parse->rtable)
1258 : {
1259 1420 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1260 :
1261 1420 : if (rte->rtekind == RTE_SUBQUERY)
1262 116 : subqueryRTindexes = bms_add_member(subqueryRTindexes, rti);
1263 1420 : rti++;
1264 : }
1265 :
1266 : /*
1267 : * If the parent RTE is a partitioned table, we should use that as the
1268 : * nominal target relation, because the RTEs added for partitioned tables
1269 : * (including the root parent) as child members of the inheritance set do
1270 : * not appear anywhere else in the plan, so the confusion explained below
1271 : * for non-partitioning inheritance cases is not possible.
1272 : */
1273 996 : parent_rte = rt_fetch(top_parentRTindex, parse->rtable);
1274 : Assert(parent_rte->inh);
1275 996 : if (parent_rte->relkind == RELKIND_PARTITIONED_TABLE)
1276 : {
1277 622 : nominalRelation = top_parentRTindex;
1278 622 : rootRelation = top_parentRTindex;
1279 : }
1280 :
1281 : /*
1282 : * Before generating the real per-child-relation plans, do a cycle of
1283 : * planning as though the query were a SELECT. The objective here is to
1284 : * find out which child relations need to be processed, using the same
1285 : * expansion and pruning logic as for a SELECT. We'll then pull out the
1286 : * RangeTblEntry-s generated for the child rels, and make use of the
1287 : * AppendRelInfo entries for them to guide the real planning. (This is
1288 : * rather inefficient; we could perhaps stop short of making a full Path
1289 : * tree. But this whole function is inefficient and slated for
1290 : * destruction, so let's not contort query_planner for that.)
1291 : */
1292 : {
1293 : PlannerInfo *subroot;
1294 :
1295 : /*
1296 : * Flat-copy the PlannerInfo to prevent modification of the original.
1297 : */
1298 996 : subroot = makeNode(PlannerInfo);
1299 996 : memcpy(subroot, root, sizeof(PlannerInfo));
1300 :
1301 : /*
1302 : * Make a deep copy of the parsetree for this planning cycle to mess
1303 : * around with, and change it to look like a SELECT. (Hack alert: the
1304 : * target RTE still has updatedCols set if this is an UPDATE, so that
1305 : * expand_partitioned_rtentry will correctly update
1306 : * subroot->partColsUpdated.)
1307 : */
1308 996 : subroot->parse = copyObject(root->parse);
1309 :
1310 996 : subroot->parse->commandType = CMD_SELECT;
1311 996 : subroot->parse->resultRelation = 0;
1312 :
1313 : /*
1314 : * Ensure the subroot has its own copy of the original
1315 : * append_rel_list, since it'll be scribbled on. (Note that at this
1316 : * point, the list only contains AppendRelInfos for flattened UNION
1317 : * ALL subqueries.)
1318 : */
1319 996 : subroot->append_rel_list = copyObject(root->append_rel_list);
1320 :
1321 : /*
1322 : * Better make a private copy of the rowMarks, too.
1323 : */
1324 996 : subroot->rowMarks = copyObject(root->rowMarks);
1325 :
1326 : /* There shouldn't be any OJ info to translate, as yet */
1327 : Assert(subroot->join_info_list == NIL);
1328 : /* and we haven't created PlaceHolderInfos, either */
1329 : Assert(subroot->placeholder_list == NIL);
1330 :
1331 : /* Generate Path(s) for accessing this result relation */
1332 996 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1333 :
1334 : /* Extract the info we need. */
1335 996 : select_rtable = subroot->parse->rtable;
1336 996 : select_appinfos = subroot->append_rel_list;
1337 :
1338 : /*
1339 : * We need to propagate partColsUpdated back, too. (The later
1340 : * planning cycles will not set this because they won't run
1341 : * expand_partitioned_rtentry for the UPDATE target.)
1342 : */
1343 996 : root->partColsUpdated = subroot->partColsUpdated;
1344 : }
1345 :
1346 : /*----------
1347 : * Since only one rangetable can exist in the final plan, we need to make
1348 : * sure that it contains all the RTEs needed for any child plan. This is
1349 : * complicated by the need to use separate subquery RTEs for each child.
1350 : * We arrange the final rtable as follows:
1351 : * 1. All original rtable entries (with their original RT indexes).
1352 : * 2. All the relation RTEs generated for children of the target table.
1353 : * 3. Subquery RTEs for children after the first. We need N * (K - 1)
1354 : * RT slots for this, if there are N subqueries and K child tables.
1355 : * 4. Additional RTEs generated during the child planning runs, such as
1356 : * children of inheritable RTEs other than the target table.
1357 : * We assume that each child planning run will create an identical set
1358 : * of type-4 RTEs.
1359 : *
1360 : * So the next thing to do is append the type-2 RTEs (the target table's
1361 : * children) to the original rtable. We look through select_appinfos
1362 : * to find them.
1363 : *
1364 : * To identify which AppendRelInfos are relevant as we thumb through
1365 : * select_appinfos, we need to look for both direct and indirect children
1366 : * of top_parentRTindex, so we use a bitmap of known parent relids.
1367 : * expand_inherited_rtentry() always processes a parent before any of that
1368 : * parent's children, so we should see an intermediate parent before its
1369 : * children.
1370 : *----------
1371 : */
1372 996 : child_appinfos = NIL;
1373 996 : old_child_rtis = NIL;
1374 996 : new_child_rtis = NIL;
1375 996 : parent_relids = bms_make_singleton(top_parentRTindex);
1376 3682 : foreach(lc, select_appinfos)
1377 : {
1378 2686 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1379 : RangeTblEntry *child_rte;
1380 :
1381 : /* append_rel_list contains all append rels; ignore others */
1382 2686 : if (!bms_is_member(appinfo->parent_relid, parent_relids))
1383 224 : continue;
1384 :
1385 : /* remember relevant AppendRelInfos for use below */
1386 2462 : child_appinfos = lappend(child_appinfos, appinfo);
1387 :
1388 : /* extract RTE for this child rel */
1389 2462 : child_rte = rt_fetch(appinfo->child_relid, select_rtable);
1390 :
1391 : /* and append it to the original rtable */
1392 2462 : parse->rtable = lappend(parse->rtable, child_rte);
1393 :
1394 : /* remember child's index in the SELECT rtable */
1395 2462 : old_child_rtis = lappend_int(old_child_rtis, appinfo->child_relid);
1396 :
1397 : /* and its new index in the final rtable */
1398 2462 : new_child_rtis = lappend_int(new_child_rtis, list_length(parse->rtable));
1399 :
1400 : /* if child is itself partitioned, update parent_relids */
1401 2462 : if (child_rte->inh)
1402 : {
1403 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1404 276 : parent_relids = bms_add_member(parent_relids, appinfo->child_relid);
1405 : }
1406 : }
1407 :
1408 : /*
1409 : * It's possible that the RTIs we just assigned for the child rels in the
1410 : * final rtable are different from what they were in the SELECT query.
1411 : * Adjust the AppendRelInfos so that they will correctly map RT indexes to
1412 : * the final indexes. We can do this left-to-right since no child rel's
1413 : * final RT index could be greater than what it had in the SELECT query.
1414 : */
1415 3458 : forboth(lc, old_child_rtis, lc2, new_child_rtis)
1416 : {
1417 2462 : int old_child_rti = lfirst_int(lc);
1418 2462 : int new_child_rti = lfirst_int(lc2);
1419 :
1420 2462 : if (old_child_rti == new_child_rti)
1421 2446 : continue; /* nothing to do */
1422 :
1423 : Assert(old_child_rti > new_child_rti);
1424 :
1425 16 : ChangeVarNodes((Node *) child_appinfos,
1426 : old_child_rti, new_child_rti, 0);
1427 : }
1428 :
1429 : /*
1430 : * Now set up rangetable entries for subqueries for additional children
1431 : * (the first child will just use the original ones). These all have to
1432 : * look more or less real, or EXPLAIN will get unhappy; so we just make
1433 : * them all clones of the original subqueries.
1434 : */
1435 996 : next_subquery_rti = list_length(parse->rtable) + 1;
1436 996 : if (subqueryRTindexes != NULL)
1437 : {
1438 76 : int n_children = list_length(child_appinfos);
1439 :
1440 280 : while (n_children-- > 1)
1441 : {
1442 128 : int oldrti = -1;
1443 :
1444 456 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1445 : {
1446 : RangeTblEntry *subqrte;
1447 :
1448 200 : subqrte = rt_fetch(oldrti, parse->rtable);
1449 200 : parse->rtable = lappend(parse->rtable, copyObject(subqrte));
1450 : }
1451 : }
1452 : }
1453 :
1454 : /*
1455 : * The query for each child is obtained by translating the query for its
1456 : * immediate parent, since the AppendRelInfo data we have shows deltas
1457 : * between parents and children. We use the parent_parses array to
1458 : * remember the appropriate query trees. This is indexed by parent relid.
1459 : * Since the maximum number of parents is limited by the number of RTEs in
1460 : * the SELECT query, we use that number to allocate the array. An extra
1461 : * entry is needed since relids start from 1.
1462 : */
1463 996 : parent_parses = (Query **) palloc0((list_length(select_rtable) + 1) *
1464 : sizeof(Query *));
1465 996 : parent_parses[top_parentRTindex] = parse;
1466 :
1467 : /*
1468 : * And now we can get on with generating a plan for each child table.
1469 : */
1470 3458 : foreach(lc, child_appinfos)
1471 : {
1472 2462 : AppendRelInfo *appinfo = lfirst_node(AppendRelInfo, lc);
1473 2462 : Index this_subquery_rti = next_subquery_rti;
1474 : Query *parent_parse;
1475 : PlannerInfo *subroot;
1476 : RangeTblEntry *child_rte;
1477 : RelOptInfo *sub_final_rel;
1478 : Path *subpath;
1479 :
1480 : /*
1481 : * expand_inherited_rtentry() always processes a parent before any of
1482 : * that parent's children, so the parent query for this relation
1483 : * should already be available.
1484 : */
1485 2462 : parent_parse = parent_parses[appinfo->parent_relid];
1486 : Assert(parent_parse != NULL);
1487 :
1488 : /*
1489 : * We need a working copy of the PlannerInfo so that we can control
1490 : * propagation of information back to the main copy.
1491 : */
1492 2462 : subroot = makeNode(PlannerInfo);
1493 2462 : memcpy(subroot, root, sizeof(PlannerInfo));
1494 :
1495 : /*
1496 : * Generate modified query with this rel as target. We first apply
1497 : * adjust_appendrel_attrs, which copies the Query and changes
1498 : * references to the parent RTE to refer to the current child RTE,
1499 : * then fool around with subquery RTEs.
1500 : */
1501 2462 : subroot->parse = (Query *)
1502 2462 : adjust_appendrel_attrs(subroot,
1503 : (Node *) parent_parse,
1504 : 1, &appinfo);
1505 :
1506 : /*
1507 : * If there are securityQuals attached to the parent, move them to the
1508 : * child rel (they've already been transformed properly for that).
1509 : */
1510 2462 : parent_rte = rt_fetch(appinfo->parent_relid, subroot->parse->rtable);
1511 2462 : child_rte = rt_fetch(appinfo->child_relid, subroot->parse->rtable);
1512 2462 : child_rte->securityQuals = parent_rte->securityQuals;
1513 2462 : parent_rte->securityQuals = NIL;
1514 :
1515 : /*
1516 : * HACK: setting this to a value other than INHKIND_NONE signals to
1517 : * relation_excluded_by_constraints() to treat the result relation as
1518 : * being an appendrel member.
1519 : */
1520 2462 : subroot->inhTargetKind =
1521 2462 : (rootRelation != 0) ? INHKIND_PARTITIONED : INHKIND_INHERITED;
1522 :
1523 : /*
1524 : * If this child is further partitioned, remember it as a parent.
1525 : * Since a partitioned table does not have any data, we don't need to
1526 : * create a plan for it, and we can stop processing it here. We do,
1527 : * however, need to remember its modified PlannerInfo for use when
1528 : * processing its children, since we'll update their varnos based on
1529 : * the delta from immediate parent to child, not from top to child.
1530 : *
1531 : * Note: a very non-obvious point is that we have not yet added
1532 : * duplicate subquery RTEs to the subroot's rtable. We mustn't,
1533 : * because then its children would have two sets of duplicates,
1534 : * confusing matters.
1535 : */
1536 2462 : if (child_rte->inh)
1537 : {
1538 : Assert(child_rte->relkind == RELKIND_PARTITIONED_TABLE);
1539 276 : parent_parses[appinfo->child_relid] = subroot->parse;
1540 612 : continue;
1541 : }
1542 :
1543 : /*
1544 : * Set the nominal target relation of the ModifyTable node if not
1545 : * already done. If the target is a partitioned table, we already set
1546 : * nominalRelation to refer to the partition root, above. For
1547 : * non-partitioned inheritance cases, we'll use the first child
1548 : * relation (even if it's excluded) as the nominal target relation.
1549 : * Because of the way expand_inherited_rtentry works, that should be
1550 : * the RTE representing the parent table in its role as a simple
1551 : * member of the inheritance set.
1552 : *
1553 : * It would be logically cleaner to *always* use the inheritance
1554 : * parent RTE as the nominal relation; but that RTE is not otherwise
1555 : * referenced in the plan in the non-partitioned inheritance case.
1556 : * Instead the duplicate child RTE created by expand_inherited_rtentry
1557 : * is used elsewhere in the plan, so using the original parent RTE
1558 : * would give rise to confusing use of multiple aliases in EXPLAIN
1559 : * output for what the user will think is the "same" table. OTOH,
1560 : * it's not a problem in the partitioned inheritance case, because
1561 : * there is no duplicate RTE for the parent.
1562 : */
1563 2186 : if (nominalRelation < 0)
1564 374 : nominalRelation = appinfo->child_relid;
1565 :
1566 : /*
1567 : * As above, each child plan run needs its own append_rel_list and
1568 : * rowmarks, which should start out as pristine copies of the
1569 : * originals. There can't be any references to UPDATE/DELETE target
1570 : * rels in them; but there could be subquery references, which we'll
1571 : * fix up in a moment.
1572 : */
1573 2186 : subroot->append_rel_list = copyObject(root->append_rel_list);
1574 2186 : subroot->rowMarks = copyObject(root->rowMarks);
1575 :
1576 : /*
1577 : * If this isn't the first child Query, adjust Vars and jointree
1578 : * entries to reference the appropriate set of subquery RTEs.
1579 : */
1580 2186 : if (final_rtable != NIL && subqueryRTindexes != NULL)
1581 : {
1582 116 : int oldrti = -1;
1583 :
1584 412 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1585 : {
1586 180 : Index newrti = next_subquery_rti++;
1587 :
1588 180 : ChangeVarNodes((Node *) subroot->parse, oldrti, newrti, 0);
1589 180 : ChangeVarNodes((Node *) subroot->append_rel_list,
1590 : oldrti, newrti, 0);
1591 180 : ChangeVarNodes((Node *) subroot->rowMarks, oldrti, newrti, 0);
1592 : }
1593 : }
1594 :
1595 : /* There shouldn't be any OJ info to translate, as yet */
1596 : Assert(subroot->join_info_list == NIL);
1597 : /* and we haven't created PlaceHolderInfos, either */
1598 : Assert(subroot->placeholder_list == NIL);
1599 :
1600 : /* Generate Path(s) for accessing this result relation */
1601 2186 : grouping_planner(subroot, true, 0.0 /* retrieve all tuples */ );
1602 :
1603 : /*
1604 : * Select cheapest path in case there's more than one. We always run
1605 : * modification queries to conclusion, so we care only for the
1606 : * cheapest-total path.
1607 : */
1608 2186 : sub_final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
1609 2186 : set_cheapest(sub_final_rel);
1610 2186 : subpath = sub_final_rel->cheapest_total_path;
1611 :
1612 : /*
1613 : * If this child rel was excluded by constraint exclusion, exclude it
1614 : * from the result plan.
1615 : */
1616 2186 : if (IS_DUMMY_REL(sub_final_rel))
1617 60 : continue;
1618 :
1619 : /*
1620 : * If this is the first non-excluded child, its post-planning rtable
1621 : * becomes the initial contents of final_rtable; otherwise, copy its
1622 : * modified subquery RTEs into final_rtable, to ensure we have sane
1623 : * copies of those. Also save the first non-excluded child's version
1624 : * of the rowmarks list; we assume all children will end up with
1625 : * equivalent versions of that.
1626 : */
1627 2126 : if (final_rtable == NIL)
1628 : {
1629 964 : final_rtable = subroot->parse->rtable;
1630 964 : final_rowmarks = subroot->rowMarks;
1631 : }
1632 : else
1633 : {
1634 : Assert(list_length(final_rtable) ==
1635 : list_length(subroot->parse->rtable));
1636 1162 : if (subqueryRTindexes != NULL)
1637 : {
1638 116 : int oldrti = -1;
1639 :
1640 412 : while ((oldrti = bms_next_member(subqueryRTindexes, oldrti)) >= 0)
1641 : {
1642 180 : Index newrti = this_subquery_rti++;
1643 : RangeTblEntry *subqrte;
1644 : ListCell *newrticell;
1645 :
1646 180 : subqrte = rt_fetch(newrti, subroot->parse->rtable);
1647 180 : newrticell = list_nth_cell(final_rtable, newrti - 1);
1648 180 : lfirst(newrticell) = subqrte;
1649 : }
1650 : }
1651 : }
1652 :
1653 : /*
1654 : * We need to collect all the RelOptInfos from all child plans into
1655 : * the main PlannerInfo, since setrefs.c will need them. We use the
1656 : * last child's simple_rel_array, so we have to propagate forward the
1657 : * RelOptInfos that were already built in previous children.
1658 : */
1659 : Assert(subroot->simple_rel_array_size >= save_rel_array_size);
1660 9824 : for (rti = 1; rti < save_rel_array_size; rti++)
1661 : {
1662 7698 : RelOptInfo *brel = save_rel_array[rti];
1663 :
1664 7698 : if (brel)
1665 2880 : subroot->simple_rel_array[rti] = brel;
1666 : }
1667 2126 : save_rel_array_size = subroot->simple_rel_array_size;
1668 2126 : save_rel_array = subroot->simple_rel_array;
1669 2126 : save_append_rel_array = subroot->append_rel_array;
1670 :
1671 : /*
1672 : * Make sure any initplans from this rel get into the outer list. Note
1673 : * we're effectively assuming all children generate the same
1674 : * init_plans.
1675 : */
1676 2126 : root->init_plans = subroot->init_plans;
1677 :
1678 : /* Build list of sub-paths */
1679 2126 : subpaths = lappend(subpaths, subpath);
1680 :
1681 : /* Build list of modified subroots, too */
1682 2126 : subroots = lappend(subroots, subroot);
1683 :
1684 : /* Build list of target-relation RT indexes */
1685 2126 : resultRelations = lappend_int(resultRelations, appinfo->child_relid);
1686 :
1687 : /* Build lists of per-relation WCO and RETURNING targetlists */
1688 2126 : if (parse->withCheckOptions)
1689 272 : withCheckOptionLists = lappend(withCheckOptionLists,
1690 272 : subroot->parse->withCheckOptions);
1691 2126 : if (parse->returningList)
1692 316 : returningLists = lappend(returningLists,
1693 316 : subroot->parse->returningList);
1694 :
1695 : Assert(!parse->onConflict);
1696 : }
1697 :
1698 : /* Result path must go into outer query's FINAL upperrel */
1699 996 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1700 :
1701 : /*
1702 : * We don't currently worry about setting final_rel's consider_parallel
1703 : * flag in this case, nor about allowing FDWs or create_upper_paths_hook
1704 : * to get control here.
1705 : */
1706 :
1707 996 : if (subpaths == NIL)
1708 : {
1709 : /*
1710 : * We managed to exclude every child rel, so generate a dummy path
1711 : * representing the empty set. Although it's clear that no data will
1712 : * be updated or deleted, we will still need to have a ModifyTable
1713 : * node so that any statement triggers are executed. (This could be
1714 : * cleaner if we fixed nodeModifyTable.c to support zero child nodes,
1715 : * but that probably wouldn't be a net win.)
1716 : */
1717 : Path *dummy_path;
1718 :
1719 : /* tlist processing never got done, either */
1720 32 : root->processed_tlist = preprocess_targetlist(root);
1721 32 : final_rel->reltarget = create_pathtarget(root, root->processed_tlist);
1722 :
1723 : /* Make a dummy path, cf set_dummy_rel_pathlist() */
1724 32 : dummy_path = (Path *) create_append_path(NULL, final_rel, NIL, NIL,
1725 : NIL, NULL, 0, false,
1726 : NIL, -1);
1727 :
1728 : /* These lists must be nonempty to make a valid ModifyTable node */
1729 32 : subpaths = list_make1(dummy_path);
1730 32 : subroots = list_make1(root);
1731 32 : resultRelations = list_make1_int(parse->resultRelation);
1732 32 : if (parse->withCheckOptions)
1733 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
1734 32 : if (parse->returningList)
1735 12 : returningLists = list_make1(parse->returningList);
1736 : /* Disable tuple routing, too, just to be safe */
1737 32 : root->partColsUpdated = false;
1738 : }
1739 : else
1740 : {
1741 : /*
1742 : * Put back the final adjusted rtable into the master copy of the
1743 : * Query. (We mustn't do this if we found no non-excluded children,
1744 : * since we never saved an adjusted rtable at all.)
1745 : */
1746 964 : parse->rtable = final_rtable;
1747 964 : root->simple_rel_array_size = save_rel_array_size;
1748 964 : root->simple_rel_array = save_rel_array;
1749 964 : root->append_rel_array = save_append_rel_array;
1750 :
1751 : /* Must reconstruct master's simple_rte_array, too */
1752 964 : root->simple_rte_array = (RangeTblEntry **)
1753 964 : palloc0((list_length(final_rtable) + 1) * sizeof(RangeTblEntry *));
1754 964 : rti = 1;
1755 5146 : foreach(lc, final_rtable)
1756 : {
1757 4182 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1758 :
1759 4182 : root->simple_rte_array[rti++] = rte;
1760 : }
1761 :
1762 : /* Put back adjusted rowmarks, too */
1763 964 : root->rowMarks = final_rowmarks;
1764 : }
1765 :
1766 : /*
1767 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node will
1768 : * have dealt with fetching non-locked marked rows, else we need to have
1769 : * ModifyTable do that.
1770 : */
1771 996 : if (parse->rowMarks)
1772 0 : rowMarks = NIL;
1773 : else
1774 996 : rowMarks = root->rowMarks;
1775 :
1776 : /* Create Path representing a ModifyTable to do the UPDATE/DELETE work */
1777 996 : add_path(final_rel, (Path *)
1778 2988 : create_modifytable_path(root, final_rel,
1779 : parse->commandType,
1780 996 : parse->canSetTag,
1781 : nominalRelation,
1782 : rootRelation,
1783 996 : root->partColsUpdated,
1784 : resultRelations,
1785 : subpaths,
1786 : subroots,
1787 : withCheckOptionLists,
1788 : returningLists,
1789 : rowMarks,
1790 : NULL,
1791 : assign_special_exec_param(root)));
1792 996 : }
1793 :
1794 : /*--------------------
1795 : * grouping_planner
1796 : * Perform planning steps related to grouping, aggregation, etc.
1797 : *
1798 : * This function adds all required top-level processing to the scan/join
1799 : * Path(s) produced by query_planner.
1800 : *
1801 : * If inheritance_update is true, we're being called from inheritance_planner
1802 : * and should not include a ModifyTable step in the resulting Path(s).
1803 : * (inheritance_planner will create a single ModifyTable node covering all the
1804 : * target tables.)
1805 : *
1806 : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1807 : * tuple_fraction is interpreted as follows:
1808 : * 0: expect all tuples to be retrieved (normal case)
1809 : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1810 : * from the plan to be retrieved
1811 : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1812 : * expected to be retrieved (ie, a LIMIT specification)
1813 : *
1814 : * Returns nothing; the useful output is in the Paths we attach to the
1815 : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1816 : * root->processed_tlist contains the final processed targetlist.
1817 : *
1818 : * Note that we have not done set_cheapest() on the final rel; it's convenient
1819 : * to leave this to the caller.
1820 : *--------------------
1821 : */
1822 : static void
1823 296940 : grouping_planner(PlannerInfo *root, bool inheritance_update,
1824 : double tuple_fraction)
1825 : {
1826 296940 : Query *parse = root->parse;
1827 296940 : int64 offset_est = 0;
1828 296940 : int64 count_est = 0;
1829 296940 : double limit_tuples = -1.0;
1830 296940 : bool have_postponed_srfs = false;
1831 : PathTarget *final_target;
1832 : List *final_targets;
1833 : List *final_targets_contain_srfs;
1834 : bool final_target_parallel_safe;
1835 : RelOptInfo *current_rel;
1836 : RelOptInfo *final_rel;
1837 : FinalPathExtraData extra;
1838 : ListCell *lc;
1839 :
1840 : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1841 296940 : if (parse->limitCount || parse->limitOffset)
1842 : {
1843 2922 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1844 : &offset_est, &count_est);
1845 :
1846 : /*
1847 : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1848 : * estimate the effects of using a bounded sort.
1849 : */
1850 2922 : if (count_est > 0 && offset_est >= 0)
1851 2704 : limit_tuples = (double) count_est + (double) offset_est;
1852 : }
1853 :
1854 : /* Make tuple_fraction accessible to lower-level routines */
1855 296940 : root->tuple_fraction = tuple_fraction;
1856 :
1857 296940 : if (parse->setOperations)
1858 : {
1859 : /*
1860 : * If there's a top-level ORDER BY, assume we have to fetch all the
1861 : * tuples. This might be too simplistic given all the hackery below
1862 : * to possibly avoid the sort; but the odds of accurate estimates here
1863 : * are pretty low anyway. XXX try to get rid of this in favor of
1864 : * letting plan_set_operations generate both fast-start and
1865 : * cheapest-total paths.
1866 : */
1867 1088 : if (parse->sortClause)
1868 204 : root->tuple_fraction = 0.0;
1869 :
1870 : /*
1871 : * Construct Paths for set operations. The results will not need any
1872 : * work except perhaps a top-level sort and/or LIMIT. Note that any
1873 : * special work for recursive unions is the responsibility of
1874 : * plan_set_operations.
1875 : */
1876 1088 : current_rel = plan_set_operations(root);
1877 :
1878 : /*
1879 : * We should not need to call preprocess_targetlist, since we must be
1880 : * in a SELECT query node. Instead, use the processed_tlist returned
1881 : * by plan_set_operations (since this tells whether it returned any
1882 : * resjunk columns!), and transfer any sort key information from the
1883 : * original tlist.
1884 : */
1885 : Assert(parse->commandType == CMD_SELECT);
1886 :
1887 : /* for safety, copy processed_tlist instead of modifying in-place */
1888 1088 : root->processed_tlist =
1889 1088 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1890 : parse->targetList);
1891 :
1892 : /* Also extract the PathTarget form of the setop result tlist */
1893 1088 : final_target = current_rel->cheapest_total_path->pathtarget;
1894 :
1895 : /* And check whether it's parallel safe */
1896 1088 : final_target_parallel_safe =
1897 1088 : is_parallel_safe(root, (Node *) final_target->exprs);
1898 :
1899 : /* The setop result tlist couldn't contain any SRFs */
1900 : Assert(!parse->hasTargetSRFs);
1901 1088 : final_targets = final_targets_contain_srfs = NIL;
1902 :
1903 : /*
1904 : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1905 : * checked already, but let's make sure).
1906 : */
1907 1088 : if (parse->rowMarks)
1908 0 : ereport(ERROR,
1909 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1910 : /*------
1911 : translator: %s is a SQL row locking clause such as FOR UPDATE */
1912 : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1913 : LCS_asString(linitial_node(RowMarkClause,
1914 : parse->rowMarks)->strength))));
1915 :
1916 : /*
1917 : * Calculate pathkeys that represent result ordering requirements
1918 : */
1919 : Assert(parse->distinctClause == NIL);
1920 1088 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1921 : parse->sortClause,
1922 : root->processed_tlist);
1923 : }
1924 : else
1925 : {
1926 : /* No set operations, do regular planning */
1927 : PathTarget *sort_input_target;
1928 : List *sort_input_targets;
1929 : List *sort_input_targets_contain_srfs;
1930 : bool sort_input_target_parallel_safe;
1931 : PathTarget *grouping_target;
1932 : List *grouping_targets;
1933 : List *grouping_targets_contain_srfs;
1934 : bool grouping_target_parallel_safe;
1935 : PathTarget *scanjoin_target;
1936 : List *scanjoin_targets;
1937 : List *scanjoin_targets_contain_srfs;
1938 : bool scanjoin_target_parallel_safe;
1939 : bool scanjoin_target_same_exprs;
1940 : bool have_grouping;
1941 : AggClauseCosts agg_costs;
1942 295852 : WindowFuncLists *wflists = NULL;
1943 295852 : List *activeWindows = NIL;
1944 295852 : grouping_sets_data *gset_data = NULL;
1945 : standard_qp_extra qp_extra;
1946 :
1947 : /* A recursive query should always have setOperations */
1948 : Assert(!root->hasRecursion);
1949 :
1950 : /* Preprocess grouping sets and GROUP BY clause, if any */
1951 295852 : if (parse->groupingSets)
1952 : {
1953 420 : gset_data = preprocess_grouping_sets(root);
1954 : }
1955 : else
1956 : {
1957 : /* Preprocess regular GROUP BY clause, if any */
1958 295432 : if (parse->groupClause)
1959 1672 : parse->groupClause = preprocess_groupclause(root, NIL);
1960 : }
1961 :
1962 : /*
1963 : * Preprocess targetlist. Note that much of the remaining planning
1964 : * work will be done with the PathTarget representation of tlists, but
1965 : * we must also maintain the full representation of the final tlist so
1966 : * that we can transfer its decoration (resnames etc) to the topmost
1967 : * tlist of the finished Plan. This is kept in processed_tlist.
1968 : */
1969 295848 : root->processed_tlist = preprocess_targetlist(root);
1970 :
1971 : /*
1972 : * Collect statistics about aggregates for estimating costs, and mark
1973 : * all the aggregates with resolved aggtranstypes. We must do this
1974 : * before slicing and dicing the tlist into various pathtargets, else
1975 : * some copies of the Aggref nodes might escape being marked with the
1976 : * correct transtypes.
1977 : *
1978 : * Note: currently, we do not detect duplicate aggregates here. This
1979 : * may result in somewhat-overestimated cost, which is fine for our
1980 : * purposes since all Paths will get charged the same. But at some
1981 : * point we might wish to do that detection in the planner, rather
1982 : * than during executor startup.
1983 : */
1984 295848 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
1985 295848 : if (parse->hasAggs)
1986 : {
1987 23372 : get_agg_clause_costs(root, (Node *) root->processed_tlist,
1988 : AGGSPLIT_SIMPLE, &agg_costs);
1989 23372 : get_agg_clause_costs(root, parse->havingQual, AGGSPLIT_SIMPLE,
1990 : &agg_costs);
1991 : }
1992 :
1993 : /*
1994 : * Locate any window functions in the tlist. (We don't need to look
1995 : * anywhere else, since expressions used in ORDER BY will be in there
1996 : * too.) Note that they could all have been eliminated by constant
1997 : * folding, in which case we don't need to do any more work.
1998 : */
1999 295848 : if (parse->hasWindowFuncs)
2000 : {
2001 1104 : wflists = find_window_functions((Node *) root->processed_tlist,
2002 1104 : list_length(parse->windowClause));
2003 1104 : if (wflists->numWindowFuncs > 0)
2004 1100 : activeWindows = select_active_windows(root, wflists);
2005 : else
2006 4 : parse->hasWindowFuncs = false;
2007 : }
2008 :
2009 : /*
2010 : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
2011 : * adding logic between here and the query_planner() call. Anything
2012 : * that is needed in MIN/MAX-optimizable cases will have to be
2013 : * duplicated in planagg.c.
2014 : */
2015 295848 : if (parse->hasAggs)
2016 23372 : preprocess_minmax_aggregates(root);
2017 :
2018 : /*
2019 : * Figure out whether there's a hard limit on the number of rows that
2020 : * query_planner's result subplan needs to return. Even if we know a
2021 : * hard limit overall, it doesn't apply if the query has any
2022 : * grouping/aggregation operations, or SRFs in the tlist.
2023 : */
2024 589636 : if (parse->groupClause ||
2025 587548 : parse->groupingSets ||
2026 587140 : parse->distinctClause ||
2027 565254 : parse->hasAggs ||
2028 542728 : parse->hasWindowFuncs ||
2029 538544 : parse->hasTargetSRFs ||
2030 267690 : root->hasHavingQual)
2031 28170 : root->limit_tuples = -1.0;
2032 : else
2033 267678 : root->limit_tuples = limit_tuples;
2034 :
2035 : /* Set up data needed by standard_qp_callback */
2036 295848 : qp_extra.activeWindows = activeWindows;
2037 295848 : qp_extra.groupClause = (gset_data
2038 416 : ? (gset_data->rollups ? linitial_node(RollupData, gset_data->rollups)->groupClause : NIL)
2039 296264 : : parse->groupClause);
2040 :
2041 : /*
2042 : * Generate the best unsorted and presorted paths for the scan/join
2043 : * portion of this Query, ie the processing represented by the
2044 : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
2045 : * We also generate (in standard_qp_callback) pathkey representations
2046 : * of the query's sort clause, distinct clause, etc.
2047 : */
2048 295848 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
2049 :
2050 : /*
2051 : * Convert the query's result tlist into PathTarget format.
2052 : *
2053 : * Note: this cannot be done before query_planner() has performed
2054 : * appendrel expansion, because that might add resjunk entries to
2055 : * root->processed_tlist. Waiting till afterwards is also helpful
2056 : * because the target width estimates can use per-Var width numbers
2057 : * that were obtained within query_planner().
2058 : */
2059 295838 : final_target = create_pathtarget(root, root->processed_tlist);
2060 295838 : final_target_parallel_safe =
2061 295838 : is_parallel_safe(root, (Node *) final_target->exprs);
2062 :
2063 : /*
2064 : * If ORDER BY was given, consider whether we should use a post-sort
2065 : * projection, and compute the adjusted target for preceding steps if
2066 : * so.
2067 : */
2068 295838 : if (parse->sortClause)
2069 : {
2070 34554 : sort_input_target = make_sort_input_target(root,
2071 : final_target,
2072 : &have_postponed_srfs);
2073 34554 : sort_input_target_parallel_safe =
2074 34554 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
2075 : }
2076 : else
2077 : {
2078 261284 : sort_input_target = final_target;
2079 261284 : sort_input_target_parallel_safe = final_target_parallel_safe;
2080 : }
2081 :
2082 : /*
2083 : * If we have window functions to deal with, the output from any
2084 : * grouping step needs to be what the window functions want;
2085 : * otherwise, it should be sort_input_target.
2086 : */
2087 295838 : if (activeWindows)
2088 : {
2089 1100 : grouping_target = make_window_input_target(root,
2090 : final_target,
2091 : activeWindows);
2092 1100 : grouping_target_parallel_safe =
2093 1100 : is_parallel_safe(root, (Node *) grouping_target->exprs);
2094 : }
2095 : else
2096 : {
2097 294738 : grouping_target = sort_input_target;
2098 294738 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
2099 : }
2100 :
2101 : /*
2102 : * If we have grouping or aggregation to do, the topmost scan/join
2103 : * plan node must emit what the grouping step wants; otherwise, it
2104 : * should emit grouping_target.
2105 : */
2106 1179204 : have_grouping = (parse->groupClause || parse->groupingSets ||
2107 861814 : parse->hasAggs || root->hasHavingQual);
2108 295838 : if (have_grouping)
2109 : {
2110 23624 : scanjoin_target = make_group_input_target(root, final_target);
2111 23624 : scanjoin_target_parallel_safe =
2112 23624 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
2113 : }
2114 : else
2115 : {
2116 272214 : scanjoin_target = grouping_target;
2117 272214 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
2118 : }
2119 :
2120 : /*
2121 : * If there are any SRFs in the targetlist, we must separate each of
2122 : * these PathTargets into SRF-computing and SRF-free targets. Replace
2123 : * each of the named targets with a SRF-free version, and remember the
2124 : * list of additional projection steps we need to add afterwards.
2125 : */
2126 295838 : if (parse->hasTargetSRFs)
2127 : {
2128 : /* final_target doesn't recompute any SRFs in sort_input_target */
2129 3276 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2130 : &final_targets,
2131 : &final_targets_contain_srfs);
2132 3276 : final_target = linitial_node(PathTarget, final_targets);
2133 : Assert(!linitial_int(final_targets_contain_srfs));
2134 : /* likewise for sort_input_target vs. grouping_target */
2135 3276 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2136 : &sort_input_targets,
2137 : &sort_input_targets_contain_srfs);
2138 3276 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2139 : Assert(!linitial_int(sort_input_targets_contain_srfs));
2140 : /* likewise for grouping_target vs. scanjoin_target */
2141 3276 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
2142 : &grouping_targets,
2143 : &grouping_targets_contain_srfs);
2144 3276 : grouping_target = linitial_node(PathTarget, grouping_targets);
2145 : Assert(!linitial_int(grouping_targets_contain_srfs));
2146 : /* scanjoin_target will not have any SRFs precomputed for it */
2147 3276 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2148 : &scanjoin_targets,
2149 : &scanjoin_targets_contain_srfs);
2150 3276 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2151 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2152 : }
2153 : else
2154 : {
2155 : /* initialize lists; for most of these, dummy values are OK */
2156 292562 : final_targets = final_targets_contain_srfs = NIL;
2157 292562 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2158 292562 : grouping_targets = grouping_targets_contain_srfs = NIL;
2159 292562 : scanjoin_targets = list_make1(scanjoin_target);
2160 292562 : scanjoin_targets_contain_srfs = NIL;
2161 : }
2162 :
2163 : /* Apply scan/join target. */
2164 591676 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2165 295838 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2166 295838 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2167 : scanjoin_targets_contain_srfs,
2168 : scanjoin_target_parallel_safe,
2169 : scanjoin_target_same_exprs);
2170 :
2171 : /*
2172 : * Save the various upper-rel PathTargets we just computed into
2173 : * root->upper_targets[]. The core code doesn't use this, but it
2174 : * provides a convenient place for extensions to get at the info. For
2175 : * consistency, we save all the intermediate targets, even though some
2176 : * of the corresponding upperrels might not be needed for this query.
2177 : */
2178 295838 : root->upper_targets[UPPERREL_FINAL] = final_target;
2179 295838 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2180 295838 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2181 295838 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2182 295838 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2183 :
2184 : /*
2185 : * If we have grouping and/or aggregation, consider ways to implement
2186 : * that. We build a new upperrel representing the output of this
2187 : * phase.
2188 : */
2189 295838 : if (have_grouping)
2190 : {
2191 23624 : current_rel = create_grouping_paths(root,
2192 : current_rel,
2193 : grouping_target,
2194 : grouping_target_parallel_safe,
2195 : &agg_costs,
2196 : gset_data);
2197 : /* Fix things up if grouping_target contains SRFs */
2198 23620 : if (parse->hasTargetSRFs)
2199 84 : adjust_paths_for_srfs(root, current_rel,
2200 : grouping_targets,
2201 : grouping_targets_contain_srfs);
2202 : }
2203 :
2204 : /*
2205 : * If we have window functions, consider ways to implement those. We
2206 : * build a new upperrel representing the output of this phase.
2207 : */
2208 295834 : if (activeWindows)
2209 : {
2210 1100 : current_rel = create_window_paths(root,
2211 : current_rel,
2212 : grouping_target,
2213 : sort_input_target,
2214 : sort_input_target_parallel_safe,
2215 : wflists,
2216 : activeWindows);
2217 : /* Fix things up if sort_input_target contains SRFs */
2218 1100 : if (parse->hasTargetSRFs)
2219 8 : adjust_paths_for_srfs(root, current_rel,
2220 : sort_input_targets,
2221 : sort_input_targets_contain_srfs);
2222 : }
2223 :
2224 : /*
2225 : * If there is a DISTINCT clause, consider ways to implement that. We
2226 : * build a new upperrel representing the output of this phase.
2227 : */
2228 295834 : if (parse->distinctClause)
2229 : {
2230 384 : current_rel = create_distinct_paths(root,
2231 : current_rel);
2232 : }
2233 : } /* end of if (setOperations) */
2234 :
2235 : /*
2236 : * If ORDER BY was given, consider ways to implement that, and generate a
2237 : * new upperrel containing only paths that emit the correct ordering and
2238 : * project the correct final_target. We can apply the original
2239 : * limit_tuples limit in sort costing here, but only if there are no
2240 : * postponed SRFs.
2241 : */
2242 296922 : if (parse->sortClause)
2243 : {
2244 34758 : current_rel = create_ordered_paths(root,
2245 : current_rel,
2246 : final_target,
2247 : final_target_parallel_safe,
2248 : have_postponed_srfs ? -1.0 :
2249 : limit_tuples);
2250 : /* Fix things up if final_target contains SRFs */
2251 34758 : if (parse->hasTargetSRFs)
2252 116 : adjust_paths_for_srfs(root, current_rel,
2253 : final_targets,
2254 : final_targets_contain_srfs);
2255 : }
2256 :
2257 : /*
2258 : * Now we are prepared to build the final-output upperrel.
2259 : */
2260 296922 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2261 :
2262 : /*
2263 : * If the input rel is marked consider_parallel and there's nothing that's
2264 : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2265 : * consider_parallel as well. Note that if the query has rowMarks or is
2266 : * not a SELECT, consider_parallel will be false for every relation in the
2267 : * query.
2268 : */
2269 367246 : if (current_rel->consider_parallel &&
2270 140640 : is_parallel_safe(root, parse->limitOffset) &&
2271 70316 : is_parallel_safe(root, parse->limitCount))
2272 70312 : final_rel->consider_parallel = true;
2273 :
2274 : /*
2275 : * If the current_rel belongs to a single FDW, so does the final_rel.
2276 : */
2277 296922 : final_rel->serverid = current_rel->serverid;
2278 296922 : final_rel->userid = current_rel->userid;
2279 296922 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2280 296922 : final_rel->fdwroutine = current_rel->fdwroutine;
2281 :
2282 : /*
2283 : * Generate paths for the final_rel. Insert all surviving paths, with
2284 : * LockRows, Limit, and/or ModifyTable steps added if needed.
2285 : */
2286 598996 : foreach(lc, current_rel->pathlist)
2287 : {
2288 302074 : Path *path = (Path *) lfirst(lc);
2289 :
2290 : /*
2291 : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2292 : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2293 : * here. If there are only non-locking rowmarks, they should be
2294 : * handled by the ModifyTable node instead. However, root->rowMarks
2295 : * is what goes into the LockRows node.)
2296 : */
2297 302074 : if (parse->rowMarks)
2298 : {
2299 5258 : path = (Path *) create_lockrows_path(root, final_rel, path,
2300 : root->rowMarks,
2301 : assign_special_exec_param(root));
2302 : }
2303 :
2304 : /*
2305 : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2306 : */
2307 302074 : if (limit_needed(parse))
2308 : {
2309 3372 : path = (Path *) create_limit_path(root, final_rel, path,
2310 : parse->limitOffset,
2311 : parse->limitCount,
2312 : offset_est, count_est);
2313 : }
2314 :
2315 : /*
2316 : * If this is an INSERT/UPDATE/DELETE, and we're not being called from
2317 : * inheritance_planner, add the ModifyTable node.
2318 : */
2319 302074 : if (parse->commandType != CMD_SELECT && !inheritance_update)
2320 : {
2321 : Index rootRelation;
2322 : List *withCheckOptionLists;
2323 : List *returningLists;
2324 : List *rowMarks;
2325 :
2326 : /*
2327 : * If target is a partition root table, we need to mark the
2328 : * ModifyTable node appropriately for that.
2329 : */
2330 71996 : if (rt_fetch(parse->resultRelation, parse->rtable)->relkind ==
2331 : RELKIND_PARTITIONED_TABLE)
2332 1460 : rootRelation = parse->resultRelation;
2333 : else
2334 70536 : rootRelation = 0;
2335 :
2336 : /*
2337 : * Set up the WITH CHECK OPTION and RETURNING lists-of-lists, if
2338 : * needed.
2339 : */
2340 71996 : if (parse->withCheckOptions)
2341 496 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2342 : else
2343 71500 : withCheckOptionLists = NIL;
2344 :
2345 71996 : if (parse->returningList)
2346 1152 : returningLists = list_make1(parse->returningList);
2347 : else
2348 70844 : returningLists = NIL;
2349 :
2350 : /*
2351 : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2352 : * will have dealt with fetching non-locked marked rows, else we
2353 : * need to have ModifyTable do that.
2354 : */
2355 71996 : if (parse->rowMarks)
2356 0 : rowMarks = NIL;
2357 : else
2358 71996 : rowMarks = root->rowMarks;
2359 :
2360 71996 : path = (Path *)
2361 431976 : create_modifytable_path(root, final_rel,
2362 : parse->commandType,
2363 71996 : parse->canSetTag,
2364 71996 : parse->resultRelation,
2365 : rootRelation,
2366 : false,
2367 71996 : list_make1_int(parse->resultRelation),
2368 71996 : list_make1(path),
2369 71996 : list_make1(root),
2370 : withCheckOptionLists,
2371 : returningLists,
2372 : rowMarks,
2373 : parse->onConflict,
2374 : assign_special_exec_param(root));
2375 : }
2376 :
2377 : /* And shove it into final_rel */
2378 302074 : add_path(final_rel, path);
2379 : }
2380 :
2381 : /*
2382 : * Generate partial paths for final_rel, too, if outer query levels might
2383 : * be able to make use of them.
2384 : */
2385 303076 : if (final_rel->consider_parallel && root->query_level > 1 &&
2386 6154 : !limit_needed(parse))
2387 : {
2388 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2389 6112 : foreach(lc, current_rel->partial_pathlist)
2390 : {
2391 36 : Path *partial_path = (Path *) lfirst(lc);
2392 :
2393 36 : add_partial_path(final_rel, partial_path);
2394 : }
2395 : }
2396 :
2397 296922 : extra.limit_needed = limit_needed(parse);
2398 296922 : extra.limit_tuples = limit_tuples;
2399 296922 : extra.count_est = count_est;
2400 296922 : extra.offset_est = offset_est;
2401 :
2402 : /*
2403 : * If there is an FDW that's responsible for all baserels of the query,
2404 : * let it consider adding ForeignPaths.
2405 : */
2406 297912 : if (final_rel->fdwroutine &&
2407 990 : final_rel->fdwroutine->GetForeignUpperPaths)
2408 932 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2409 : current_rel, final_rel,
2410 : &extra);
2411 :
2412 : /* Let extensions possibly add some more paths */
2413 296922 : if (create_upper_paths_hook)
2414 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2415 : current_rel, final_rel, &extra);
2416 :
2417 : /* Note: currently, we leave it to callers to do set_cheapest() */
2418 296922 : }
2419 :
2420 : /*
2421 : * Do preprocessing for groupingSets clause and related data. This handles the
2422 : * preliminary steps of expanding the grouping sets, organizing them into lists
2423 : * of rollups, and preparing annotations which will later be filled in with
2424 : * size estimates.
2425 : */
2426 : static grouping_sets_data *
2427 420 : preprocess_grouping_sets(PlannerInfo *root)
2428 : {
2429 420 : Query *parse = root->parse;
2430 : List *sets;
2431 420 : int maxref = 0;
2432 : ListCell *lc;
2433 : ListCell *lc_set;
2434 420 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2435 :
2436 420 : parse->groupingSets = expand_grouping_sets(parse->groupingSets, -1);
2437 :
2438 420 : gd->any_hashable = false;
2439 420 : gd->unhashable_refs = NULL;
2440 420 : gd->unsortable_refs = NULL;
2441 420 : gd->unsortable_sets = NIL;
2442 :
2443 420 : if (parse->groupClause)
2444 : {
2445 : ListCell *lc;
2446 :
2447 1264 : foreach(lc, parse->groupClause)
2448 : {
2449 872 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2450 872 : Index ref = gc->tleSortGroupRef;
2451 :
2452 872 : if (ref > maxref)
2453 848 : maxref = ref;
2454 :
2455 872 : if (!gc->hashable)
2456 20 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2457 :
2458 872 : if (!OidIsValid(gc->sortop))
2459 28 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2460 : }
2461 : }
2462 :
2463 : /* Allocate workspace array for remapping */
2464 420 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2465 :
2466 : /*
2467 : * If we have any unsortable sets, we must extract them before trying to
2468 : * prepare rollups. Unsortable sets don't go through
2469 : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2470 : * here.
2471 : */
2472 420 : if (!bms_is_empty(gd->unsortable_refs))
2473 : {
2474 28 : List *sortable_sets = NIL;
2475 :
2476 84 : foreach(lc, parse->groupingSets)
2477 : {
2478 60 : List *gset = (List *) lfirst(lc);
2479 :
2480 60 : if (bms_overlap_list(gd->unsortable_refs, gset))
2481 : {
2482 32 : GroupingSetData *gs = makeNode(GroupingSetData);
2483 :
2484 32 : gs->set = gset;
2485 32 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2486 :
2487 : /*
2488 : * We must enforce here that an unsortable set is hashable;
2489 : * later code assumes this. Parse analysis only checks that
2490 : * every individual column is either hashable or sortable.
2491 : *
2492 : * Note that passing this test doesn't guarantee we can
2493 : * generate a plan; there might be other showstoppers.
2494 : */
2495 32 : if (bms_overlap_list(gd->unhashable_refs, gset))
2496 4 : ereport(ERROR,
2497 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2498 : errmsg("could not implement GROUP BY"),
2499 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2500 : }
2501 : else
2502 28 : sortable_sets = lappend(sortable_sets, gset);
2503 : }
2504 :
2505 24 : if (sortable_sets)
2506 20 : sets = extract_rollup_sets(sortable_sets);
2507 : else
2508 4 : sets = NIL;
2509 : }
2510 : else
2511 392 : sets = extract_rollup_sets(parse->groupingSets);
2512 :
2513 1076 : foreach(lc_set, sets)
2514 : {
2515 660 : List *current_sets = (List *) lfirst(lc_set);
2516 660 : RollupData *rollup = makeNode(RollupData);
2517 : GroupingSetData *gs;
2518 :
2519 : /*
2520 : * Reorder the current list of grouping sets into correct prefix
2521 : * order. If only one aggregation pass is needed, try to make the
2522 : * list match the ORDER BY clause; if more than one pass is needed, we
2523 : * don't bother with that.
2524 : *
2525 : * Note that this reorders the sets from smallest-member-first to
2526 : * largest-member-first, and applies the GroupingSetData annotations,
2527 : * though the data will be filled in later.
2528 : */
2529 660 : current_sets = reorder_grouping_sets(current_sets,
2530 660 : (list_length(sets) == 1
2531 : ? parse->sortClause
2532 : : NIL));
2533 :
2534 : /*
2535 : * Get the initial (and therefore largest) grouping set.
2536 : */
2537 660 : gs = linitial_node(GroupingSetData, current_sets);
2538 :
2539 : /*
2540 : * Order the groupClause appropriately. If the first grouping set is
2541 : * empty, then the groupClause must also be empty; otherwise we have
2542 : * to force the groupClause to match that grouping set's order.
2543 : *
2544 : * (The first grouping set can be empty even though parse->groupClause
2545 : * is not empty only if all non-empty grouping sets are unsortable.
2546 : * The groupClauses for hashed grouping sets are built later on.)
2547 : */
2548 660 : if (gs->set)
2549 632 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2550 : else
2551 28 : rollup->groupClause = NIL;
2552 :
2553 : /*
2554 : * Is it hashable? We pretend empty sets are hashable even though we
2555 : * actually force them not to be hashed later. But don't bother if
2556 : * there's nothing but empty sets (since in that case we can't hash
2557 : * anything).
2558 : */
2559 1292 : if (gs->set &&
2560 632 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2561 : {
2562 616 : rollup->hashable = true;
2563 616 : gd->any_hashable = true;
2564 : }
2565 :
2566 : /*
2567 : * Now that we've pinned down an order for the groupClause for this
2568 : * list of grouping sets, we need to remap the entries in the grouping
2569 : * sets from sortgrouprefs to plain indices (0-based) into the
2570 : * groupClause for this collection of grouping sets. We keep the
2571 : * original form for later use, though.
2572 : */
2573 660 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2574 : current_sets,
2575 : gd->tleref_to_colnum_map);
2576 660 : rollup->gsets_data = current_sets;
2577 :
2578 660 : gd->rollups = lappend(gd->rollups, rollup);
2579 : }
2580 :
2581 416 : if (gd->unsortable_sets)
2582 : {
2583 : /*
2584 : * We have not yet pinned down a groupclause for this, but we will
2585 : * need index-based lists for estimation purposes. Construct
2586 : * hash_sets_idx based on the entire original groupclause for now.
2587 : */
2588 24 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2589 : gd->unsortable_sets,
2590 : gd->tleref_to_colnum_map);
2591 24 : gd->any_hashable = true;
2592 : }
2593 :
2594 416 : return gd;
2595 : }
2596 :
2597 : /*
2598 : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2599 : * (without annotation) mapped to indexes into the given groupclause.
2600 : */
2601 : static List *
2602 1812 : remap_to_groupclause_idx(List *groupClause,
2603 : List *gsets,
2604 : int *tleref_to_colnum_map)
2605 : {
2606 1812 : int ref = 0;
2607 1812 : List *result = NIL;
2608 : ListCell *lc;
2609 :
2610 4396 : foreach(lc, groupClause)
2611 : {
2612 2584 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2613 :
2614 2584 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2615 : }
2616 :
2617 4204 : foreach(lc, gsets)
2618 : {
2619 2392 : List *set = NIL;
2620 : ListCell *lc2;
2621 2392 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2622 :
2623 5272 : foreach(lc2, gs->set)
2624 : {
2625 2880 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2626 : }
2627 :
2628 2392 : result = lappend(result, set);
2629 : }
2630 :
2631 1812 : return result;
2632 : }
2633 :
2634 :
2635 : /*
2636 : * preprocess_rowmarks - set up PlanRowMarks if needed
2637 : */
2638 : static void
2639 296114 : preprocess_rowmarks(PlannerInfo *root)
2640 : {
2641 296114 : Query *parse = root->parse;
2642 : Bitmapset *rels;
2643 : List *prowmarks;
2644 : ListCell *l;
2645 : int i;
2646 :
2647 296114 : if (parse->rowMarks)
2648 : {
2649 : /*
2650 : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2651 : * grouping, since grouping renders a reference to individual tuple
2652 : * CTIDs invalid. This is also checked at parse time, but that's
2653 : * insufficient because of rule substitution, query pullup, etc.
2654 : */
2655 4814 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2656 : parse->rowMarks)->strength);
2657 : }
2658 : else
2659 : {
2660 : /*
2661 : * We only need rowmarks for UPDATE, DELETE, or FOR [KEY]
2662 : * UPDATE/SHARE.
2663 : */
2664 573322 : if (parse->commandType != CMD_UPDATE &&
2665 282022 : parse->commandType != CMD_DELETE)
2666 279180 : return;
2667 : }
2668 :
2669 : /*
2670 : * We need to have rowmarks for all base relations except the target. We
2671 : * make a bitmapset of all base rels and then remove the items we don't
2672 : * need or have FOR [KEY] UPDATE/SHARE marks for.
2673 : */
2674 16934 : rels = get_relids_in_jointree((Node *) parse->jointree, false);
2675 16934 : if (parse->resultRelation)
2676 12120 : rels = bms_del_member(rels, parse->resultRelation);
2677 :
2678 : /*
2679 : * Convert RowMarkClauses to PlanRowMark representation.
2680 : */
2681 16934 : prowmarks = NIL;
2682 21970 : foreach(l, parse->rowMarks)
2683 : {
2684 5036 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2685 5036 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2686 : PlanRowMark *newrc;
2687 :
2688 : /*
2689 : * Currently, it is syntactically impossible to have FOR UPDATE et al
2690 : * applied to an update/delete target rel. If that ever becomes
2691 : * possible, we should drop the target from the PlanRowMark list.
2692 : */
2693 : Assert(rc->rti != parse->resultRelation);
2694 :
2695 : /*
2696 : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2697 : * can't support true locking. Subqueries that got flattened into the
2698 : * main query should be ignored completely. Any that didn't will get
2699 : * ROW_MARK_COPY items in the next loop.
2700 : */
2701 5036 : if (rte->rtekind != RTE_RELATION)
2702 76 : continue;
2703 :
2704 4960 : rels = bms_del_member(rels, rc->rti);
2705 :
2706 4960 : newrc = makeNode(PlanRowMark);
2707 4960 : newrc->rti = newrc->prti = rc->rti;
2708 4960 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2709 4960 : newrc->markType = select_rowmark_type(rte, rc->strength);
2710 4960 : newrc->allMarkTypes = (1 << newrc->markType);
2711 4960 : newrc->strength = rc->strength;
2712 4960 : newrc->waitPolicy = rc->waitPolicy;
2713 4960 : newrc->isParent = false;
2714 :
2715 4960 : prowmarks = lappend(prowmarks, newrc);
2716 : }
2717 :
2718 : /*
2719 : * Now, add rowmarks for any non-target, non-locked base relations.
2720 : */
2721 16934 : i = 0;
2722 37544 : foreach(l, parse->rtable)
2723 : {
2724 20610 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2725 : PlanRowMark *newrc;
2726 :
2727 20610 : i++;
2728 20610 : if (!bms_is_member(i, rels))
2729 19606 : continue;
2730 :
2731 1004 : newrc = makeNode(PlanRowMark);
2732 1004 : newrc->rti = newrc->prti = i;
2733 1004 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2734 1004 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2735 1004 : newrc->allMarkTypes = (1 << newrc->markType);
2736 1004 : newrc->strength = LCS_NONE;
2737 1004 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2738 1004 : newrc->isParent = false;
2739 :
2740 1004 : prowmarks = lappend(prowmarks, newrc);
2741 : }
2742 :
2743 16934 : root->rowMarks = prowmarks;
2744 : }
2745 :
2746 : /*
2747 : * Select RowMarkType to use for a given table
2748 : */
2749 : RowMarkType
2750 6898 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2751 : {
2752 6898 : if (rte->rtekind != RTE_RELATION)
2753 : {
2754 : /* If it's not a table at all, use ROW_MARK_COPY */
2755 174 : return ROW_MARK_COPY;
2756 : }
2757 6724 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2758 : {
2759 : /* Let the FDW select the rowmark type, if it wants to */
2760 160 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2761 :
2762 160 : if (fdwroutine->GetForeignRowMarkType != NULL)
2763 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2764 : /* Otherwise, use ROW_MARK_COPY by default */
2765 160 : return ROW_MARK_COPY;
2766 : }
2767 : else
2768 : {
2769 : /* Regular table, apply the appropriate lock type */
2770 6564 : switch (strength)
2771 : {
2772 : case LCS_NONE:
2773 :
2774 : /*
2775 : * We don't need a tuple lock, only the ability to re-fetch
2776 : * the row.
2777 : */
2778 1146 : return ROW_MARK_REFERENCE;
2779 : break;
2780 : case LCS_FORKEYSHARE:
2781 3860 : return ROW_MARK_KEYSHARE;
2782 : break;
2783 : case LCS_FORSHARE:
2784 234 : return ROW_MARK_SHARE;
2785 : break;
2786 : case LCS_FORNOKEYUPDATE:
2787 58 : return ROW_MARK_NOKEYEXCLUSIVE;
2788 : break;
2789 : case LCS_FORUPDATE:
2790 1266 : return ROW_MARK_EXCLUSIVE;
2791 : break;
2792 : }
2793 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2794 : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2795 : }
2796 : }
2797 :
2798 : /*
2799 : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2800 : *
2801 : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2802 : * results back in *count_est and *offset_est. These variables are set to
2803 : * 0 if the corresponding clause is not present, and -1 if it's present
2804 : * but we couldn't estimate the value for it. (The "0" convention is OK
2805 : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2806 : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2807 : * usual practice of never estimating less than one row.) These values will
2808 : * be passed to create_limit_path, which see if you change this code.
2809 : *
2810 : * The return value is the suitably adjusted tuple_fraction to use for
2811 : * planning the query. This adjustment is not overridable, since it reflects
2812 : * plan actions that grouping_planner() will certainly take, not assumptions
2813 : * about context.
2814 : */
2815 : static double
2816 2922 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2817 : int64 *offset_est, int64 *count_est)
2818 : {
2819 2922 : Query *parse = root->parse;
2820 : Node *est;
2821 : double limit_fraction;
2822 :
2823 : /* Should not be called unless LIMIT or OFFSET */
2824 : Assert(parse->limitCount || parse->limitOffset);
2825 :
2826 : /*
2827 : * Try to obtain the clause values. We use estimate_expression_value
2828 : * primarily because it can sometimes do something useful with Params.
2829 : */
2830 2922 : if (parse->limitCount)
2831 : {
2832 2720 : est = estimate_expression_value(root, parse->limitCount);
2833 2720 : if (est && IsA(est, Const))
2834 : {
2835 5432 : if (((Const *) est)->constisnull)
2836 : {
2837 : /* NULL indicates LIMIT ALL, ie, no limit */
2838 0 : *count_est = 0; /* treat as not present */
2839 : }
2840 : else
2841 : {
2842 2716 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2843 2716 : if (*count_est <= 0)
2844 96 : *count_est = 1; /* force to at least 1 */
2845 : }
2846 : }
2847 : else
2848 4 : *count_est = -1; /* can't estimate */
2849 : }
2850 : else
2851 202 : *count_est = 0; /* not present */
2852 :
2853 2922 : if (parse->limitOffset)
2854 : {
2855 520 : est = estimate_expression_value(root, parse->limitOffset);
2856 520 : if (est && IsA(est, Const))
2857 : {
2858 1008 : if (((Const *) est)->constisnull)
2859 : {
2860 : /* Treat NULL as no offset; the executor will too */
2861 0 : *offset_est = 0; /* treat as not present */
2862 : }
2863 : else
2864 : {
2865 504 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2866 504 : if (*offset_est < 0)
2867 0 : *offset_est = 0; /* treat as not present */
2868 : }
2869 : }
2870 : else
2871 16 : *offset_est = -1; /* can't estimate */
2872 : }
2873 : else
2874 2402 : *offset_est = 0; /* not present */
2875 :
2876 2922 : if (*count_est != 0)
2877 : {
2878 : /*
2879 : * A LIMIT clause limits the absolute number of tuples returned.
2880 : * However, if it's not a constant LIMIT then we have to guess; for
2881 : * lack of a better idea, assume 10% of the plan's result is wanted.
2882 : */
2883 2720 : if (*count_est < 0 || *offset_est < 0)
2884 : {
2885 : /* LIMIT or OFFSET is an expression ... punt ... */
2886 16 : limit_fraction = 0.10;
2887 : }
2888 : else
2889 : {
2890 : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2891 2704 : limit_fraction = (double) *count_est + (double) *offset_est;
2892 : }
2893 :
2894 : /*
2895 : * If we have absolute limits from both caller and LIMIT, use the
2896 : * smaller value; likewise if they are both fractional. If one is
2897 : * fractional and the other absolute, we can't easily determine which
2898 : * is smaller, but we use the heuristic that the absolute will usually
2899 : * be smaller.
2900 : */
2901 2720 : if (tuple_fraction >= 1.0)
2902 : {
2903 4 : if (limit_fraction >= 1.0)
2904 : {
2905 : /* both absolute */
2906 4 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2907 : }
2908 : else
2909 : {
2910 : /* caller absolute, limit fractional; use caller's value */
2911 : }
2912 : }
2913 2716 : else if (tuple_fraction > 0.0)
2914 : {
2915 94 : if (limit_fraction >= 1.0)
2916 : {
2917 : /* caller fractional, limit absolute; use limit */
2918 94 : tuple_fraction = limit_fraction;
2919 : }
2920 : else
2921 : {
2922 : /* both fractional */
2923 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2924 : }
2925 : }
2926 : else
2927 : {
2928 : /* no info from caller, just use limit */
2929 2622 : tuple_fraction = limit_fraction;
2930 : }
2931 : }
2932 202 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2933 : {
2934 : /*
2935 : * We have an OFFSET but no LIMIT. This acts entirely differently
2936 : * from the LIMIT case: here, we need to increase rather than decrease
2937 : * the caller's tuple_fraction, because the OFFSET acts to cause more
2938 : * tuples to be fetched instead of fewer. This only matters if we got
2939 : * a tuple_fraction > 0, however.
2940 : *
2941 : * As above, use 10% if OFFSET is present but unestimatable.
2942 : */
2943 8 : if (*offset_est < 0)
2944 0 : limit_fraction = 0.10;
2945 : else
2946 8 : limit_fraction = (double) *offset_est;
2947 :
2948 : /*
2949 : * If we have absolute counts from both caller and OFFSET, add them
2950 : * together; likewise if they are both fractional. If one is
2951 : * fractional and the other absolute, we want to take the larger, and
2952 : * we heuristically assume that's the fractional one.
2953 : */
2954 8 : if (tuple_fraction >= 1.0)
2955 : {
2956 0 : if (limit_fraction >= 1.0)
2957 : {
2958 : /* both absolute, so add them together */
2959 0 : tuple_fraction += limit_fraction;
2960 : }
2961 : else
2962 : {
2963 : /* caller absolute, limit fractional; use limit */
2964 0 : tuple_fraction = limit_fraction;
2965 : }
2966 : }
2967 : else
2968 : {
2969 8 : if (limit_fraction >= 1.0)
2970 : {
2971 : /* caller fractional, limit absolute; use caller's value */
2972 : }
2973 : else
2974 : {
2975 : /* both fractional, so add them together */
2976 0 : tuple_fraction += limit_fraction;
2977 0 : if (tuple_fraction >= 1.0)
2978 0 : tuple_fraction = 0.0; /* assume fetch all */
2979 : }
2980 : }
2981 : }
2982 :
2983 2922 : return tuple_fraction;
2984 : }
2985 :
2986 : /*
2987 : * limit_needed - do we actually need a Limit plan node?
2988 : *
2989 : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2990 : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2991 : * locution for an optimization fence. (Because other places in the planner
2992 : * merely check whether parse->limitOffset isn't NULL, it will still work as
2993 : * an optimization fence --- we're just suppressing unnecessary run-time
2994 : * overhead.)
2995 : *
2996 : * This might look like it could be merged into preprocess_limit, but there's
2997 : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2998 : * in preprocess_limit it's good enough to consider estimated values.
2999 : */
3000 : bool
3001 610414 : limit_needed(Query *parse)
3002 : {
3003 : Node *node;
3004 :
3005 610414 : node = parse->limitCount;
3006 610414 : if (node)
3007 : {
3008 6230 : if (IsA(node, Const))
3009 : {
3010 : /* NULL indicates LIMIT ALL, ie, no limit */
3011 6138 : if (!((Const *) node)->constisnull)
3012 6138 : return true; /* LIMIT with a constant value */
3013 : }
3014 : else
3015 92 : return true; /* non-constant LIMIT */
3016 : }
3017 :
3018 604184 : node = parse->limitOffset;
3019 604184 : if (node)
3020 : {
3021 616 : if (IsA(node, Const))
3022 : {
3023 : /* Treat NULL as no offset; the executor would too */
3024 470 : if (!((Const *) node)->constisnull)
3025 : {
3026 470 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3027 :
3028 470 : if (offset != 0)
3029 24 : return true; /* OFFSET with a nonzero value */
3030 : }
3031 : }
3032 : else
3033 146 : return true; /* non-constant OFFSET */
3034 : }
3035 :
3036 604014 : return false; /* don't need a Limit plan node */
3037 : }
3038 :
3039 :
3040 : /*
3041 : * remove_useless_groupby_columns
3042 : * Remove any columns in the GROUP BY clause that are redundant due to
3043 : * being functionally dependent on other GROUP BY columns.
3044 : *
3045 : * Since some other DBMSes do not allow references to ungrouped columns, it's
3046 : * not unusual to find all columns listed in GROUP BY even though listing the
3047 : * primary-key columns would be sufficient. Deleting such excess columns
3048 : * avoids redundant sorting work, so it's worth doing. When we do this, we
3049 : * must mark the plan as dependent on the pkey constraint (compare the
3050 : * parser's check_ungrouped_columns() and check_functional_grouping()).
3051 : *
3052 : * In principle, we could treat any NOT-NULL columns appearing in a UNIQUE
3053 : * index as the determining columns. But as with check_functional_grouping(),
3054 : * there's currently no way to represent dependency on a NOT NULL constraint,
3055 : * so we consider only the pkey for now.
3056 : */
3057 : static void
3058 294754 : remove_useless_groupby_columns(PlannerInfo *root)
3059 : {
3060 294754 : Query *parse = root->parse;
3061 : Bitmapset **groupbyattnos;
3062 : Bitmapset **surplusvars;
3063 : ListCell *lc;
3064 : int relid;
3065 :
3066 : /* No chance to do anything if there are less than two GROUP BY items */
3067 294754 : if (list_length(parse->groupClause) < 2)
3068 293994 : return;
3069 :
3070 : /* Don't fiddle with the GROUP BY clause if the query has grouping sets */
3071 760 : if (parse->groupingSets)
3072 336 : return;
3073 :
3074 : /*
3075 : * Scan the GROUP BY clause to find GROUP BY items that are simple Vars.
3076 : * Fill groupbyattnos[k] with a bitmapset of the column attnos of RTE k
3077 : * that are GROUP BY items.
3078 : */
3079 424 : groupbyattnos = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3080 424 : (list_length(parse->rtable) + 1));
3081 1532 : foreach(lc, parse->groupClause)
3082 : {
3083 1108 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3084 1108 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3085 1108 : Var *var = (Var *) tle->expr;
3086 :
3087 : /*
3088 : * Ignore non-Vars and Vars from other query levels.
3089 : *
3090 : * XXX in principle, stable expressions containing Vars could also be
3091 : * removed, if all the Vars are functionally dependent on other GROUP
3092 : * BY items. But it's not clear that such cases occur often enough to
3093 : * be worth troubling over.
3094 : */
3095 2088 : if (!IsA(var, Var) ||
3096 980 : var->varlevelsup > 0)
3097 128 : continue;
3098 :
3099 : /* OK, remember we have this Var */
3100 980 : relid = var->varno;
3101 : Assert(relid <= list_length(parse->rtable));
3102 1960 : groupbyattnos[relid] = bms_add_member(groupbyattnos[relid],
3103 980 : var->varattno - FirstLowInvalidHeapAttributeNumber);
3104 : }
3105 :
3106 : /*
3107 : * Consider each relation and see if it is possible to remove some of its
3108 : * Vars from GROUP BY. For simplicity and speed, we do the actual removal
3109 : * in a separate pass. Here, we just fill surplusvars[k] with a bitmapset
3110 : * of the column attnos of RTE k that are removable GROUP BY items.
3111 : */
3112 424 : surplusvars = NULL; /* don't allocate array unless required */
3113 424 : relid = 0;
3114 1336 : foreach(lc, parse->rtable)
3115 : {
3116 912 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
3117 : Bitmapset *relattnos;
3118 : Bitmapset *pkattnos;
3119 : Oid constraintOid;
3120 :
3121 912 : relid++;
3122 :
3123 : /* Only plain relations could have primary-key constraints */
3124 912 : if (rte->rtekind != RTE_RELATION)
3125 1106 : continue;
3126 :
3127 : /*
3128 : * We must skip inheritance parent tables as some of the child rels
3129 : * may cause duplicate rows. This cannot happen with partitioned
3130 : * tables, however.
3131 : */
3132 672 : if (rte->inh && rte->relkind != RELKIND_PARTITIONED_TABLE)
3133 4 : continue;
3134 :
3135 : /* Nothing to do unless this rel has multiple Vars in GROUP BY */
3136 668 : relattnos = groupbyattnos[relid];
3137 668 : if (bms_membership(relattnos) != BMS_MULTIPLE)
3138 376 : continue;
3139 :
3140 : /*
3141 : * Can't remove any columns for this rel if there is no suitable
3142 : * (i.e., nondeferrable) primary key constraint.
3143 : */
3144 292 : pkattnos = get_primary_key_attnos(rte->relid, false, &constraintOid);
3145 292 : if (pkattnos == NULL)
3146 246 : continue;
3147 :
3148 : /*
3149 : * If the primary key is a proper subset of relattnos then we have
3150 : * some items in the GROUP BY that can be removed.
3151 : */
3152 46 : if (bms_subset_compare(pkattnos, relattnos) == BMS_SUBSET1)
3153 : {
3154 : /*
3155 : * To easily remember whether we've found anything to do, we don't
3156 : * allocate the surplusvars[] array until we find something.
3157 : */
3158 34 : if (surplusvars == NULL)
3159 30 : surplusvars = (Bitmapset **) palloc0(sizeof(Bitmapset *) *
3160 30 : (list_length(parse->rtable) + 1));
3161 :
3162 : /* Remember the attnos of the removable columns */
3163 34 : surplusvars[relid] = bms_difference(relattnos, pkattnos);
3164 :
3165 : /* Also, mark the resulting plan as dependent on this constraint */
3166 34 : parse->constraintDeps = lappend_oid(parse->constraintDeps,
3167 : constraintOid);
3168 : }
3169 : }
3170 :
3171 : /*
3172 : * If we found any surplus Vars, build a new GROUP BY clause without them.
3173 : * (Note: this may leave some TLEs with unreferenced ressortgroupref
3174 : * markings, but that's harmless.)
3175 : */
3176 424 : if (surplusvars != NULL)
3177 : {
3178 30 : List *new_groupby = NIL;
3179 :
3180 150 : foreach(lc, parse->groupClause)
3181 : {
3182 120 : SortGroupClause *sgc = lfirst_node(SortGroupClause, lc);
3183 120 : TargetEntry *tle = get_sortgroupclause_tle(sgc, parse->targetList);
3184 120 : Var *var = (Var *) tle->expr;
3185 :
3186 : /*
3187 : * New list must include non-Vars, outer Vars, and anything not
3188 : * marked as surplus.
3189 : */
3190 240 : if (!IsA(var, Var) ||
3191 240 : var->varlevelsup > 0 ||
3192 120 : !bms_is_member(var->varattno - FirstLowInvalidHeapAttributeNumber,
3193 120 : surplusvars[var->varno]))
3194 66 : new_groupby = lappend(new_groupby, sgc);
3195 : }
3196 :
3197 30 : parse->groupClause = new_groupby;
3198 : }
3199 : }
3200 :
3201 : /*
3202 : * preprocess_groupclause - do preparatory work on GROUP BY clause
3203 : *
3204 : * The idea here is to adjust the ordering of the GROUP BY elements
3205 : * (which in itself is semantically insignificant) to match ORDER BY,
3206 : * thereby allowing a single sort operation to both implement the ORDER BY
3207 : * requirement and set up for a Unique step that implements GROUP BY.
3208 : *
3209 : * In principle it might be interesting to consider other orderings of the
3210 : * GROUP BY elements, which could match the sort ordering of other
3211 : * possible plans (eg an indexscan) and thereby reduce cost. We don't
3212 : * bother with that, though. Hashed grouping will frequently win anyway.
3213 : *
3214 : * Note: we need no comparable processing of the distinctClause because
3215 : * the parser already enforced that that matches ORDER BY.
3216 : *
3217 : * For grouping sets, the order of items is instead forced to agree with that
3218 : * of the grouping set (and items not in the grouping set are skipped). The
3219 : * work of sorting the order of grouping set elements to match the ORDER BY if
3220 : * possible is done elsewhere.
3221 : */
3222 : static List *
3223 3432 : preprocess_groupclause(PlannerInfo *root, List *force)
3224 : {
3225 3432 : Query *parse = root->parse;
3226 3432 : List *new_groupclause = NIL;
3227 : bool partial_match;
3228 : ListCell *sl;
3229 : ListCell *gl;
3230 :
3231 : /* For grouping sets, we need to force the ordering */
3232 3432 : if (force)
3233 : {
3234 4292 : foreach(sl, force)
3235 : {
3236 2532 : Index ref = lfirst_int(sl);
3237 2532 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3238 :
3239 2532 : new_groupclause = lappend(new_groupclause, cl);
3240 : }
3241 :
3242 1760 : return new_groupclause;
3243 : }
3244 :
3245 : /* If no ORDER BY, nothing useful to do here */
3246 1672 : if (parse->sortClause == NIL)
3247 766 : return parse->groupClause;
3248 :
3249 : /*
3250 : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3251 : * items, but only as far as we can make a matching prefix.
3252 : *
3253 : * This code assumes that the sortClause contains no duplicate items.
3254 : */
3255 1788 : foreach(sl, parse->sortClause)
3256 : {
3257 1236 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3258 :
3259 1830 : foreach(gl, parse->groupClause)
3260 : {
3261 1476 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3262 :
3263 1476 : if (equal(gc, sc))
3264 : {
3265 882 : new_groupclause = lappend(new_groupclause, gc);
3266 882 : break;
3267 : }
3268 : }
3269 1236 : if (gl == NULL)
3270 354 : break; /* no match, so stop scanning */
3271 : }
3272 :
3273 : /* Did we match all of the ORDER BY list, or just some of it? */
3274 906 : partial_match = (sl != NULL);
3275 :
3276 : /* If no match at all, no point in reordering GROUP BY */
3277 906 : if (new_groupclause == NIL)
3278 128 : return parse->groupClause;
3279 :
3280 : /*
3281 : * Add any remaining GROUP BY items to the new list, but only if we were
3282 : * able to make a complete match. In other words, we only rearrange the
3283 : * GROUP BY list if the result is that one list is a prefix of the other
3284 : * --- otherwise there's no possibility of a common sort. Also, give up
3285 : * if there are any non-sortable GROUP BY items, since then there's no
3286 : * hope anyway.
3287 : */
3288 1680 : foreach(gl, parse->groupClause)
3289 : {
3290 952 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3291 :
3292 952 : if (list_member_ptr(new_groupclause, gc))
3293 866 : continue; /* it matched an ORDER BY item */
3294 86 : if (partial_match)
3295 100 : return parse->groupClause; /* give up, no common sort possible */
3296 36 : if (!OidIsValid(gc->sortop))
3297 0 : return parse->groupClause; /* give up, GROUP BY can't be sorted */
3298 36 : new_groupclause = lappend(new_groupclause, gc);
3299 : }
3300 :
3301 : /* Success --- install the rearranged GROUP BY list */
3302 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3303 728 : return new_groupclause;
3304 : }
3305 :
3306 : /*
3307 : * Extract lists of grouping sets that can be implemented using a single
3308 : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3309 : *
3310 : * Input must be sorted with smallest sets first. Result has each sublist
3311 : * sorted with smallest sets first.
3312 : *
3313 : * We want to produce the absolute minimum possible number of lists here to
3314 : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3315 : * of finding the minimal partition of a partially-ordered set into chains
3316 : * (which is what we need, taking the list of grouping sets as a poset ordered
3317 : * by set inclusion) can be mapped to the problem of finding the maximum
3318 : * cardinality matching on a bipartite graph, which is solvable in polynomial
3319 : * time with a worst case of no worse than O(n^2.5) and usually much
3320 : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3321 : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3322 : * half a second on my modest system even with optimization off and assertions
3323 : * on.)
3324 : */
3325 : static List *
3326 412 : extract_rollup_sets(List *groupingSets)
3327 : {
3328 412 : int num_sets_raw = list_length(groupingSets);
3329 412 : int num_empty = 0;
3330 412 : int num_sets = 0; /* distinct sets */
3331 412 : int num_chains = 0;
3332 412 : List *result = NIL;
3333 : List **results;
3334 : List **orig_sets;
3335 : Bitmapset **set_masks;
3336 : int *chains;
3337 : short **adjacency;
3338 : short *adjacency_buf;
3339 : BipartiteMatchState *state;
3340 : int i;
3341 : int j;
3342 : int j_size;
3343 412 : ListCell *lc1 = list_head(groupingSets);
3344 : ListCell *lc;
3345 :
3346 : /*
3347 : * Start by stripping out empty sets. The algorithm doesn't require this,
3348 : * but the planner currently needs all empty sets to be returned in the
3349 : * first list, so we strip them here and add them back after.
3350 : */
3351 1144 : while (lc1 && lfirst(lc1) == NIL)
3352 : {
3353 320 : ++num_empty;
3354 320 : lc1 = lnext(groupingSets, lc1);
3355 : }
3356 :
3357 : /* bail out now if it turns out that all we had were empty sets. */
3358 412 : if (!lc1)
3359 28 : return list_make1(groupingSets);
3360 :
3361 : /*----------
3362 : * We don't strictly need to remove duplicate sets here, but if we don't,
3363 : * they tend to become scattered through the result, which is a bit
3364 : * confusing (and irritating if we ever decide to optimize them out).
3365 : * So we remove them here and add them back after.
3366 : *
3367 : * For each non-duplicate set, we fill in the following:
3368 : *
3369 : * orig_sets[i] = list of the original set lists
3370 : * set_masks[i] = bitmapset for testing inclusion
3371 : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3372 : *
3373 : * chains[i] will be the result group this set is assigned to.
3374 : *
3375 : * We index all of these from 1 rather than 0 because it is convenient
3376 : * to leave 0 free for the NIL node in the graph algorithm.
3377 : *----------
3378 : */
3379 384 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3380 384 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3381 384 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3382 384 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3383 :
3384 384 : j_size = 0;
3385 384 : j = 0;
3386 384 : i = 1;
3387 :
3388 1300 : for_each_cell(lc, groupingSets, lc1)
3389 : {
3390 916 : List *candidate = (List *) lfirst(lc);
3391 916 : Bitmapset *candidate_set = NULL;
3392 : ListCell *lc2;
3393 916 : int dup_of = 0;
3394 :
3395 2200 : foreach(lc2, candidate)
3396 : {
3397 1284 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3398 : }
3399 :
3400 : /* we can only be a dup if we're the same length as a previous set */
3401 916 : if (j_size == list_length(candidate))
3402 : {
3403 : int k;
3404 :
3405 792 : for (k = j; k < i; ++k)
3406 : {
3407 512 : if (bms_equal(set_masks[k], candidate_set))
3408 : {
3409 52 : dup_of = k;
3410 52 : break;
3411 : }
3412 : }
3413 : }
3414 584 : else if (j_size < list_length(candidate))
3415 : {
3416 584 : j_size = list_length(candidate);
3417 584 : j = i;
3418 : }
3419 :
3420 916 : if (dup_of > 0)
3421 : {
3422 52 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3423 52 : bms_free(candidate_set);
3424 : }
3425 : else
3426 : {
3427 : int k;
3428 864 : int n_adj = 0;
3429 :
3430 864 : orig_sets[i] = list_make1(candidate);
3431 864 : set_masks[i] = candidate_set;
3432 :
3433 : /* fill in adjacency list; no need to compare equal-size sets */
3434 :
3435 1328 : for (k = j - 1; k > 0; --k)
3436 : {
3437 464 : if (bms_is_subset(set_masks[k], candidate_set))
3438 404 : adjacency_buf[++n_adj] = k;
3439 : }
3440 :
3441 864 : if (n_adj > 0)
3442 : {
3443 232 : adjacency_buf[0] = n_adj;
3444 232 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3445 232 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3446 : }
3447 : else
3448 632 : adjacency[i] = NULL;
3449 :
3450 864 : ++i;
3451 : }
3452 : }
3453 :
3454 384 : num_sets = i - 1;
3455 :
3456 : /*
3457 : * Apply the graph matching algorithm to do the work.
3458 : */
3459 384 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3460 :
3461 : /*
3462 : * Now, the state->pair* fields have the info we need to assign sets to
3463 : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3464 : * pair_vu[v] = u (both will be true, but we check both so that we can do
3465 : * it in one pass)
3466 : */
3467 384 : chains = palloc0((num_sets + 1) * sizeof(int));
3468 :
3469 1248 : for (i = 1; i <= num_sets; ++i)
3470 : {
3471 864 : int u = state->pair_vu[i];
3472 864 : int v = state->pair_uv[i];
3473 :
3474 864 : if (u > 0 && u < i)
3475 0 : chains[i] = chains[u];
3476 864 : else if (v > 0 && v < i)
3477 232 : chains[i] = chains[v];
3478 : else
3479 632 : chains[i] = ++num_chains;
3480 : }
3481 :
3482 : /* build result lists. */
3483 384 : results = palloc0((num_chains + 1) * sizeof(List *));
3484 :
3485 1248 : for (i = 1; i <= num_sets; ++i)
3486 : {
3487 864 : int c = chains[i];
3488 :
3489 : Assert(c > 0);
3490 :
3491 864 : results[c] = list_concat(results[c], orig_sets[i]);
3492 : }
3493 :
3494 : /* push any empty sets back on the first list. */
3495 1028 : while (num_empty-- > 0)
3496 260 : results[1] = lcons(NIL, results[1]);
3497 :
3498 : /* make result list */
3499 1016 : for (i = 1; i <= num_chains; ++i)
3500 632 : result = lappend(result, results[i]);
3501 :
3502 : /*
3503 : * Free all the things.
3504 : *
3505 : * (This is over-fussy for small sets but for large sets we could have
3506 : * tied up a nontrivial amount of memory.)
3507 : */
3508 384 : BipartiteMatchFree(state);
3509 384 : pfree(results);
3510 384 : pfree(chains);
3511 1248 : for (i = 1; i <= num_sets; ++i)
3512 864 : if (adjacency[i])
3513 232 : pfree(adjacency[i]);
3514 384 : pfree(adjacency);
3515 384 : pfree(adjacency_buf);
3516 384 : pfree(orig_sets);
3517 1248 : for (i = 1; i <= num_sets; ++i)
3518 864 : bms_free(set_masks[i]);
3519 384 : pfree(set_masks);
3520 :
3521 384 : return result;
3522 : }
3523 :
3524 : /*
3525 : * Reorder the elements of a list of grouping sets such that they have correct
3526 : * prefix relationships. Also inserts the GroupingSetData annotations.
3527 : *
3528 : * The input must be ordered with smallest sets first; the result is returned
3529 : * with largest sets first. Note that the result shares no list substructure
3530 : * with the input, so it's safe for the caller to modify it later.
3531 : *
3532 : * If we're passed in a sortclause, we follow its order of columns to the
3533 : * extent possible, to minimize the chance that we add unnecessary sorts.
3534 : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3535 : * gets implemented in one pass.)
3536 : */
3537 : static List *
3538 660 : reorder_grouping_sets(List *groupingsets, List *sortclause)
3539 : {
3540 : ListCell *lc;
3541 660 : List *previous = NIL;
3542 660 : List *result = NIL;
3543 :
3544 1896 : foreach(lc, groupingsets)
3545 : {
3546 1236 : List *candidate = (List *) lfirst(lc);
3547 1236 : List *new_elems = list_difference_int(candidate, previous);
3548 1236 : GroupingSetData *gs = makeNode(GroupingSetData);
3549 :
3550 2764 : while (list_length(sortclause) > list_length(previous) &&
3551 236 : list_length(new_elems) > 0)
3552 : {
3553 128 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3554 128 : int ref = sc->tleSortGroupRef;
3555 :
3556 128 : if (list_member_int(new_elems, ref))
3557 : {
3558 56 : previous = lappend_int(previous, ref);
3559 56 : new_elems = list_delete_int(new_elems, ref);
3560 : }
3561 : else
3562 : {
3563 : /* diverged from the sortclause; give up on it */
3564 72 : sortclause = NIL;
3565 72 : break;
3566 : }
3567 : }
3568 :
3569 : /*
3570 : * Safe to use list_concat (which shares cells of the second arg)
3571 : * because we know that new_elems does not share cells with anything.
3572 : */
3573 1236 : previous = list_concat(previous, new_elems);
3574 :
3575 1236 : gs->set = list_copy(previous);
3576 1236 : result = lcons(gs, result);
3577 : }
3578 :
3579 660 : list_free(previous);
3580 :
3581 660 : return result;
3582 : }
3583 :
3584 : /*
3585 : * Compute query_pathkeys and other pathkeys during plan generation
3586 : */
3587 : static void
3588 295840 : standard_qp_callback(PlannerInfo *root, void *extra)
3589 : {
3590 295840 : Query *parse = root->parse;
3591 295840 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3592 295840 : List *tlist = root->processed_tlist;
3593 295840 : List *activeWindows = qp_extra->activeWindows;
3594 :
3595 : /*
3596 : * Calculate pathkeys that represent grouping/ordering requirements. The
3597 : * sortClause is certainly sort-able, but GROUP BY and DISTINCT might not
3598 : * be, in which case we just leave their pathkeys empty.
3599 : */
3600 297896 : if (qp_extra->groupClause &&
3601 2056 : grouping_is_sortable(qp_extra->groupClause))
3602 2056 : root->group_pathkeys =
3603 2056 : make_pathkeys_for_sortclauses(root,
3604 : qp_extra->groupClause,
3605 : tlist);
3606 : else
3607 293784 : root->group_pathkeys = NIL;
3608 :
3609 : /* We consider only the first (bottom) window in pathkeys logic */
3610 295840 : if (activeWindows != NIL)
3611 : {
3612 1100 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3613 :
3614 1100 : root->window_pathkeys = make_pathkeys_for_window(root,
3615 : wc,
3616 : tlist);
3617 : }
3618 : else
3619 294740 : root->window_pathkeys = NIL;
3620 :
3621 296224 : if (parse->distinctClause &&
3622 384 : grouping_is_sortable(parse->distinctClause))
3623 380 : root->distinct_pathkeys =
3624 380 : make_pathkeys_for_sortclauses(root,
3625 : parse->distinctClause,
3626 : tlist);
3627 : else
3628 295460 : root->distinct_pathkeys = NIL;
3629 :
3630 295840 : root->sort_pathkeys =
3631 295840 : make_pathkeys_for_sortclauses(root,
3632 : parse->sortClause,
3633 : tlist);
3634 :
3635 : /*
3636 : * Figure out whether we want a sorted result from query_planner.
3637 : *
3638 : * If we have a sortable GROUP BY clause, then we want a result sorted
3639 : * properly for grouping. Otherwise, if we have window functions to
3640 : * evaluate, we try to sort for the first window. Otherwise, if there's a
3641 : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3642 : * we try to produce output that's sufficiently well sorted for the
3643 : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3644 : * by the ORDER BY clause.
3645 : *
3646 : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3647 : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3648 : * that might just leave us failing to exploit an available sort order at
3649 : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3650 : * much easier, since we know that the parser ensured that one is a
3651 : * superset of the other.
3652 : */
3653 295840 : if (root->group_pathkeys)
3654 1982 : root->query_pathkeys = root->group_pathkeys;
3655 293858 : else if (root->window_pathkeys)
3656 956 : root->query_pathkeys = root->window_pathkeys;
3657 585804 : else if (list_length(root->distinct_pathkeys) >
3658 292902 : list_length(root->sort_pathkeys))
3659 234 : root->query_pathkeys = root->distinct_pathkeys;
3660 292668 : else if (root->sort_pathkeys)
3661 33196 : root->query_pathkeys = root->sort_pathkeys;
3662 : else
3663 259472 : root->query_pathkeys = NIL;
3664 295840 : }
3665 :
3666 : /*
3667 : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3668 : *
3669 : * path_rows: number of output rows from scan/join step
3670 : * gd: grouping sets data including list of grouping sets and their clauses
3671 : * target_list: target list containing group clause references
3672 : *
3673 : * If doing grouping sets, we also annotate the gsets data with the estimates
3674 : * for each set and each individual rollup list, with a view to later
3675 : * determining whether some combination of them could be hashed instead.
3676 : */
3677 : static double
3678 25588 : get_number_of_groups(PlannerInfo *root,
3679 : double path_rows,
3680 : grouping_sets_data *gd,
3681 : List *target_list)
3682 : {
3683 25588 : Query *parse = root->parse;
3684 : double dNumGroups;
3685 :
3686 25588 : if (parse->groupClause)
3687 : {
3688 : List *groupExprs;
3689 :
3690 3578 : if (parse->groupingSets)
3691 : {
3692 : /* Add up the estimates for each grouping set */
3693 : ListCell *lc;
3694 : ListCell *lc2;
3695 :
3696 : Assert(gd); /* keep Coverity happy */
3697 :
3698 388 : dNumGroups = 0;
3699 :
3700 1020 : foreach(lc, gd->rollups)
3701 : {
3702 632 : RollupData *rollup = lfirst_node(RollupData, lc);
3703 : ListCell *lc;
3704 :
3705 632 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3706 : target_list);
3707 :
3708 632 : rollup->numGroups = 0.0;
3709 :
3710 1808 : forboth(lc, rollup->gsets, lc2, rollup->gsets_data)
3711 : {
3712 1176 : List *gset = (List *) lfirst(lc);
3713 1176 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3714 1176 : double numGroups = estimate_num_groups(root,
3715 : groupExprs,
3716 : path_rows,
3717 : &gset);
3718 :
3719 1176 : gs->numGroups = numGroups;
3720 1176 : rollup->numGroups += numGroups;
3721 : }
3722 :
3723 632 : dNumGroups += rollup->numGroups;
3724 : }
3725 :
3726 388 : if (gd->hash_sets_idx)
3727 : {
3728 : ListCell *lc;
3729 :
3730 24 : gd->dNumHashGroups = 0;
3731 :
3732 24 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3733 : target_list);
3734 :
3735 52 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
3736 : {
3737 28 : List *gset = (List *) lfirst(lc);
3738 28 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3739 28 : double numGroups = estimate_num_groups(root,
3740 : groupExprs,
3741 : path_rows,
3742 : &gset);
3743 :
3744 28 : gs->numGroups = numGroups;
3745 28 : gd->dNumHashGroups += numGroups;
3746 : }
3747 :
3748 24 : dNumGroups += gd->dNumHashGroups;
3749 : }
3750 : }
3751 : else
3752 : {
3753 : /* Plain GROUP BY */
3754 3190 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3755 : target_list);
3756 :
3757 3190 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3758 : NULL);
3759 : }
3760 : }
3761 22010 : else if (parse->groupingSets)
3762 : {
3763 : /* Empty grouping sets ... one result row for each one */
3764 28 : dNumGroups = list_length(parse->groupingSets);
3765 : }
3766 21982 : else if (parse->hasAggs || root->hasHavingQual)
3767 : {
3768 : /* Plain aggregation, one result row */
3769 21982 : dNumGroups = 1;
3770 : }
3771 : else
3772 : {
3773 : /* Not grouping */
3774 0 : dNumGroups = 1;
3775 : }
3776 :
3777 25588 : return dNumGroups;
3778 : }
3779 :
3780 : /*
3781 : * create_grouping_paths
3782 : *
3783 : * Build a new upperrel containing Paths for grouping and/or aggregation.
3784 : * Along the way, we also build an upperrel for Paths which are partially
3785 : * grouped and/or aggregated. A partially grouped and/or aggregated path
3786 : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3787 : * the only partially grouped paths we build are also partial paths; that
3788 : * is, they need a Gather and then a FinalizeAggregate.
3789 : *
3790 : * input_rel: contains the source-data Paths
3791 : * target: the pathtarget for the result Paths to compute
3792 : * agg_costs: cost info about all aggregates in query (in AGGSPLIT_SIMPLE mode)
3793 : * gd: grouping sets data including list of grouping sets and their clauses
3794 : *
3795 : * Note: all Paths in input_rel are expected to return the target computed
3796 : * by make_group_input_target.
3797 : */
3798 : static RelOptInfo *
3799 23624 : create_grouping_paths(PlannerInfo *root,
3800 : RelOptInfo *input_rel,
3801 : PathTarget *target,
3802 : bool target_parallel_safe,
3803 : const AggClauseCosts *agg_costs,
3804 : grouping_sets_data *gd)
3805 : {
3806 23624 : Query *parse = root->parse;
3807 : RelOptInfo *grouped_rel;
3808 : RelOptInfo *partially_grouped_rel;
3809 :
3810 : /*
3811 : * Create grouping relation to hold fully aggregated grouping and/or
3812 : * aggregation paths.
3813 : */
3814 23624 : grouped_rel = make_grouping_rel(root, input_rel, target,
3815 : target_parallel_safe, parse->havingQual);
3816 :
3817 : /*
3818 : * Create either paths for a degenerate grouping or paths for ordinary
3819 : * grouping, as appropriate.
3820 : */
3821 23624 : if (is_degenerate_grouping(root))
3822 12 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3823 : else
3824 : {
3825 23612 : int flags = 0;
3826 : GroupPathExtraData extra;
3827 :
3828 : /*
3829 : * Determine whether it's possible to perform sort-based
3830 : * implementations of grouping. (Note that if groupClause is empty,
3831 : * grouping_is_sortable() is trivially true, and all the
3832 : * pathkeys_contained_in() tests will succeed too, so that we'll
3833 : * consider every surviving input path.)
3834 : *
3835 : * If we have grouping sets, we might be able to sort some but not all
3836 : * of them; in this case, we need can_sort to be true as long as we
3837 : * must consider any sorted-input plan.
3838 : */
3839 23612 : if ((gd && gd->rollups != NIL)
3840 23200 : || grouping_is_sortable(parse->groupClause))
3841 23608 : flags |= GROUPING_CAN_USE_SORT;
3842 :
3843 : /*
3844 : * Determine whether we should consider hash-based implementations of
3845 : * grouping.
3846 : *
3847 : * Hashed aggregation only applies if we're grouping. If we have
3848 : * grouping sets, some groups might be hashable but others not; in
3849 : * this case we set can_hash true as long as there is nothing globally
3850 : * preventing us from hashing (and we should therefore consider plans
3851 : * with hashes).
3852 : *
3853 : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3854 : * BY aggregates. (Doing so would imply storing *all* the input
3855 : * values in the hash table, and/or running many sorts in parallel,
3856 : * either of which seems like a certain loser.) We similarly don't
3857 : * support ordered-set aggregates in hashed aggregation, but that case
3858 : * is also included in the numOrderedAggs count.
3859 : *
3860 : * Note: grouping_is_hashable() is much more expensive to check than
3861 : * the other gating conditions, so we want to do it last.
3862 : */
3863 25672 : if ((parse->groupClause != NIL &&
3864 5916 : agg_costs->numOrderedAggs == 0 &&
3865 1928 : (gd ? gd->any_hashable : grouping_is_hashable(parse->groupClause))))
3866 1924 : flags |= GROUPING_CAN_USE_HASH;
3867 :
3868 : /*
3869 : * Determine whether partial aggregation is possible.
3870 : */
3871 23612 : if (can_partial_agg(root, agg_costs))
3872 8106 : flags |= GROUPING_CAN_PARTIAL_AGG;
3873 :
3874 23612 : extra.flags = flags;
3875 23612 : extra.target_parallel_safe = target_parallel_safe;
3876 23612 : extra.havingQual = parse->havingQual;
3877 23612 : extra.targetList = parse->targetList;
3878 23612 : extra.partial_costs_set = false;
3879 :
3880 : /*
3881 : * Determine whether partitionwise aggregation is in theory possible.
3882 : * It can be disabled by the user, and for now, we don't try to
3883 : * support grouping sets. create_ordinary_grouping_paths() will check
3884 : * additional conditions, such as whether input_rel is partitioned.
3885 : */
3886 23612 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3887 294 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3888 : else
3889 23318 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3890 :
3891 23612 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3892 : agg_costs, gd, &extra,
3893 : &partially_grouped_rel);
3894 : }
3895 :
3896 23620 : set_cheapest(grouped_rel);
3897 23620 : return grouped_rel;
3898 : }
3899 :
3900 : /*
3901 : * make_grouping_rel
3902 : *
3903 : * Create a new grouping rel and set basic properties.
3904 : *
3905 : * input_rel represents the underlying scan/join relation.
3906 : * target is the output expected from the grouping relation.
3907 : */
3908 : static RelOptInfo *
3909 24542 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3910 : PathTarget *target, bool target_parallel_safe,
3911 : Node *havingQual)
3912 : {
3913 : RelOptInfo *grouped_rel;
3914 :
3915 24542 : if (IS_OTHER_REL(input_rel))
3916 : {
3917 918 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3918 : input_rel->relids);
3919 918 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3920 : }
3921 : else
3922 : {
3923 : /*
3924 : * By tradition, the relids set for the main grouping relation is
3925 : * NULL. (This could be changed, but might require adjustments
3926 : * elsewhere.)
3927 : */
3928 23624 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3929 : }
3930 :
3931 : /* Set target. */
3932 24542 : grouped_rel->reltarget = target;
3933 :
3934 : /*
3935 : * If the input relation is not parallel-safe, then the grouped relation
3936 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3937 : * target list and HAVING quals are parallel-safe.
3938 : */
3939 35886 : if (input_rel->consider_parallel && target_parallel_safe &&
3940 11344 : is_parallel_safe(root, (Node *) havingQual))
3941 11332 : grouped_rel->consider_parallel = true;
3942 :
3943 : /*
3944 : * If the input rel belongs to a single FDW, so does the grouped rel.
3945 : */
3946 24542 : grouped_rel->serverid = input_rel->serverid;
3947 24542 : grouped_rel->userid = input_rel->userid;
3948 24542 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3949 24542 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3950 :
3951 24542 : return grouped_rel;
3952 : }
3953 :
3954 : /*
3955 : * is_degenerate_grouping
3956 : *
3957 : * A degenerate grouping is one in which the query has a HAVING qual and/or
3958 : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3959 : * grouping sets are all empty).
3960 : */
3961 : static bool
3962 23624 : is_degenerate_grouping(PlannerInfo *root)
3963 : {
3964 23624 : Query *parse = root->parse;
3965 :
3966 47628 : return (root->hasHavingQual || parse->groupingSets) &&
3967 24428 : !parse->hasAggs && parse->groupClause == NIL;
3968 : }
3969 :
3970 : /*
3971 : * create_degenerate_grouping_paths
3972 : *
3973 : * When the grouping is degenerate (see is_degenerate_grouping), we are
3974 : * supposed to emit either zero or one row for each grouping set depending on
3975 : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3976 : * either HAVING or the targetlist, so we actually do not need the FROM table
3977 : * at all! We can just throw away the plan-so-far and generate a Result node.
3978 : * This is a sufficiently unusual corner case that it's not worth contorting
3979 : * the structure of this module to avoid having to generate the earlier paths
3980 : * in the first place.
3981 : */
3982 : static void
3983 12 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3984 : RelOptInfo *grouped_rel)
3985 : {
3986 12 : Query *parse = root->parse;
3987 : int nrows;
3988 : Path *path;
3989 :
3990 12 : nrows = list_length(parse->groupingSets);
3991 12 : if (nrows > 1)
3992 : {
3993 : /*
3994 : * Doesn't seem worthwhile writing code to cons up a generate_series
3995 : * or a values scan to emit multiple rows. Instead just make N clones
3996 : * and append them. (With a volatile HAVING clause, this means you
3997 : * might get between 0 and N output rows. Offhand I think that's
3998 : * desired.)
3999 : */
4000 0 : List *paths = NIL;
4001 :
4002 0 : while (--nrows >= 0)
4003 : {
4004 0 : path = (Path *)
4005 : create_group_result_path(root, grouped_rel,
4006 0 : grouped_rel->reltarget,
4007 0 : (List *) parse->havingQual);
4008 0 : paths = lappend(paths, path);
4009 : }
4010 0 : path = (Path *)
4011 : create_append_path(root,
4012 : grouped_rel,
4013 : paths,
4014 : NIL,
4015 : NIL,
4016 : NULL,
4017 : 0,
4018 : false,
4019 : NIL,
4020 : -1);
4021 : }
4022 : else
4023 : {
4024 : /* No grouping sets, or just one, so one output row */
4025 12 : path = (Path *)
4026 : create_group_result_path(root, grouped_rel,
4027 12 : grouped_rel->reltarget,
4028 12 : (List *) parse->havingQual);
4029 : }
4030 :
4031 12 : add_path(grouped_rel, path);
4032 12 : }
4033 :
4034 : /*
4035 : * create_ordinary_grouping_paths
4036 : *
4037 : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4038 : *
4039 : * We need to consider sorted and hashed aggregation in the same function,
4040 : * because otherwise (1) it would be harder to throw an appropriate error
4041 : * message if neither way works, and (2) we should not allow hashtable size
4042 : * considerations to dissuade us from using hashing if sorting is not possible.
4043 : *
4044 : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4045 : * function creates, or to NULL if it doesn't create one.
4046 : */
4047 : static void
4048 24530 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4049 : RelOptInfo *grouped_rel,
4050 : const AggClauseCosts *agg_costs,
4051 : grouping_sets_data *gd,
4052 : GroupPathExtraData *extra,
4053 : RelOptInfo **partially_grouped_rel_p)
4054 : {
4055 24530 : Path *cheapest_path = input_rel->cheapest_total_path;
4056 24530 : RelOptInfo *partially_grouped_rel = NULL;
4057 : double dNumGroups;
4058 24530 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4059 :
4060 : /*
4061 : * If this is the topmost grouping relation or if the parent relation is
4062 : * doing some form of partitionwise aggregation, then we may be able to do
4063 : * it at this level also. However, if the input relation is not
4064 : * partitioned, partitionwise aggregate is impossible.
4065 : */
4066 25742 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4067 1610 : IS_PARTITIONED_REL(input_rel))
4068 : {
4069 : /*
4070 : * If this is the topmost relation or if the parent relation is doing
4071 : * full partitionwise aggregation, then we can do full partitionwise
4072 : * aggregation provided that the GROUP BY clause contains all of the
4073 : * partitioning columns at this level. Otherwise, we can do at most
4074 : * partial partitionwise aggregation. But if partial aggregation is
4075 : * not supported in general then we can't use it for partitionwise
4076 : * aggregation either.
4077 : */
4078 700 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4079 334 : group_by_has_partkey(input_rel, extra->targetList,
4080 334 : root->parse->groupClause))
4081 192 : patype = PARTITIONWISE_AGGREGATE_FULL;
4082 174 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4083 146 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4084 : else
4085 28 : patype = PARTITIONWISE_AGGREGATE_NONE;
4086 : }
4087 :
4088 : /*
4089 : * Before generating paths for grouped_rel, we first generate any possible
4090 : * partially grouped paths; that way, later code can easily consider both
4091 : * parallel and non-parallel approaches to grouping.
4092 : */
4093 24530 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4094 : {
4095 : bool force_rel_creation;
4096 :
4097 : /*
4098 : * If we're doing partitionwise aggregation at this level, force
4099 : * creation of a partially_grouped_rel so we can add partitionwise
4100 : * paths to it.
4101 : */
4102 8976 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4103 :
4104 8976 : partially_grouped_rel =
4105 8976 : create_partial_grouping_paths(root,
4106 : grouped_rel,
4107 : input_rel,
4108 : gd,
4109 : extra,
4110 : force_rel_creation);
4111 : }
4112 :
4113 : /* Set out parameter. */
4114 24530 : *partially_grouped_rel_p = partially_grouped_rel;
4115 :
4116 : /* Apply partitionwise aggregation technique, if possible. */
4117 24530 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4118 338 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4119 : partially_grouped_rel, agg_costs,
4120 : gd, patype, extra);
4121 :
4122 : /* If we are doing partial aggregation only, return. */
4123 24530 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4124 : {
4125 : Assert(partially_grouped_rel);
4126 :
4127 374 : if (partially_grouped_rel->pathlist)
4128 374 : set_cheapest(partially_grouped_rel);
4129 :
4130 374 : return;
4131 : }
4132 :
4133 : /* Gather any partially grouped partial paths. */
4134 24156 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4135 : {
4136 858 : gather_grouping_paths(root, partially_grouped_rel);
4137 858 : set_cheapest(partially_grouped_rel);
4138 : }
4139 :
4140 : /*
4141 : * Estimate number of groups.
4142 : */
4143 24156 : dNumGroups = get_number_of_groups(root,
4144 : cheapest_path->rows,
4145 : gd,
4146 : extra->targetList);
4147 :
4148 : /* Build final grouping paths */
4149 24156 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4150 : partially_grouped_rel, agg_costs, gd,
4151 : dNumGroups, extra);
4152 :
4153 : /* Give a helpful error if we failed to find any implementation */
4154 24156 : if (grouped_rel->pathlist == NIL)
4155 4 : ereport(ERROR,
4156 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4157 : errmsg("could not implement GROUP BY"),
4158 : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4159 :
4160 : /*
4161 : * If there is an FDW that's responsible for all baserels of the query,
4162 : * let it consider adding ForeignPaths.
4163 : */
4164 24424 : if (grouped_rel->fdwroutine &&
4165 272 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4166 272 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4167 : input_rel, grouped_rel,
4168 : extra);
4169 :
4170 : /* Let extensions possibly add some more paths */
4171 24152 : if (create_upper_paths_hook)
4172 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4173 : input_rel, grouped_rel,
4174 : extra);
4175 : }
4176 :
4177 : /*
4178 : * For a given input path, consider the possible ways of doing grouping sets on
4179 : * it, by combinations of hashing and sorting. This can be called multiple
4180 : * times, so it's important that it not scribble on input. No result is
4181 : * returned, but any generated paths are added to grouped_rel.
4182 : */
4183 : static void
4184 804 : consider_groupingsets_paths(PlannerInfo *root,
4185 : RelOptInfo *grouped_rel,
4186 : Path *path,
4187 : bool is_sorted,
4188 : bool can_hash,
4189 : grouping_sets_data *gd,
4190 : const AggClauseCosts *agg_costs,
4191 : double dNumGroups)
4192 : {
4193 804 : Query *parse = root->parse;
4194 :
4195 : /*
4196 : * If we're not being offered sorted input, then only consider plans that
4197 : * can be done entirely by hashing.
4198 : *
4199 : * We can hash everything if it looks like it'll fit in work_mem. But if
4200 : * the input is actually sorted despite not being advertised as such, we
4201 : * prefer to make use of that in order to use less memory.
4202 : *
4203 : * If none of the grouping sets are sortable, then ignore the work_mem
4204 : * limit and generate a path anyway, since otherwise we'll just fail.
4205 : */
4206 804 : if (!is_sorted)
4207 : {
4208 364 : List *new_rollups = NIL;
4209 364 : RollupData *unhashed_rollup = NULL;
4210 : List *sets_data;
4211 364 : List *empty_sets_data = NIL;
4212 364 : List *empty_sets = NIL;
4213 : ListCell *lc;
4214 364 : ListCell *l_start = list_head(gd->rollups);
4215 364 : AggStrategy strat = AGG_HASHED;
4216 : double hashsize;
4217 364 : double exclude_groups = 0.0;
4218 :
4219 : Assert(can_hash);
4220 :
4221 : /*
4222 : * If the input is coincidentally sorted usefully (which can happen
4223 : * even if is_sorted is false, since that only means that our caller
4224 : * has set up the sorting for us), then save some hashtable space by
4225 : * making use of that. But we need to watch out for degenerate cases:
4226 : *
4227 : * 1) If there are any empty grouping sets, then group_pathkeys might
4228 : * be NIL if all non-empty grouping sets are unsortable. In this case,
4229 : * there will be a rollup containing only empty groups, and the
4230 : * pathkeys_contained_in test is vacuously true; this is ok.
4231 : *
4232 : * XXX: the above relies on the fact that group_pathkeys is generated
4233 : * from the first rollup. If we add the ability to consider multiple
4234 : * sort orders for grouping input, this assumption might fail.
4235 : *
4236 : * 2) If there are no empty sets and only unsortable sets, then the
4237 : * rollups list will be empty (and thus l_start == NULL), and
4238 : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4239 : * pathkeys_contain_in test doesn't cause us to crash.
4240 : */
4241 724 : if (l_start != NULL &&
4242 360 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4243 : {
4244 0 : unhashed_rollup = lfirst_node(RollupData, l_start);
4245 0 : exclude_groups = unhashed_rollup->numGroups;
4246 0 : l_start = lnext(gd->rollups, l_start);
4247 : }
4248 :
4249 364 : hashsize = estimate_hashagg_tablesize(path,
4250 : agg_costs,
4251 : dNumGroups - exclude_groups);
4252 :
4253 : /*
4254 : * gd->rollups is empty if we have only unsortable columns to work
4255 : * with. Override work_mem in that case; otherwise, we'll rely on the
4256 : * sorted-input case to generate usable mixed paths.
4257 : */
4258 364 : if (hashsize > work_mem * 1024L && gd->rollups)
4259 12 : return; /* nope, won't fit */
4260 :
4261 : /*
4262 : * We need to burst the existing rollups list into individual grouping
4263 : * sets and recompute a groupClause for each set.
4264 : */
4265 352 : sets_data = list_copy(gd->unsortable_sets);
4266 :
4267 864 : for_each_cell(lc, gd->rollups, l_start)
4268 : {
4269 528 : RollupData *rollup = lfirst_node(RollupData, lc);
4270 :
4271 : /*
4272 : * If we find an unhashable rollup that's not been skipped by the
4273 : * "actually sorted" check above, we can't cope; we'd need sorted
4274 : * input (with a different sort order) but we can't get that here.
4275 : * So bail out; we'll get a valid path from the is_sorted case
4276 : * instead.
4277 : *
4278 : * The mere presence of empty grouping sets doesn't make a rollup
4279 : * unhashable (see preprocess_grouping_sets), we handle those
4280 : * specially below.
4281 : */
4282 528 : if (!rollup->hashable)
4283 32 : return;
4284 : else
4285 512 : sets_data = list_concat(sets_data, list_copy(rollup->gsets_data));
4286 : }
4287 1372 : foreach(lc, sets_data)
4288 : {
4289 1036 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4290 1036 : List *gset = gs->set;
4291 : RollupData *rollup;
4292 :
4293 1036 : if (gset == NIL)
4294 : {
4295 : /* Empty grouping sets can't be hashed. */
4296 244 : empty_sets_data = lappend(empty_sets_data, gs);
4297 244 : empty_sets = lappend(empty_sets, NIL);
4298 : }
4299 : else
4300 : {
4301 792 : rollup = makeNode(RollupData);
4302 :
4303 792 : rollup->groupClause = preprocess_groupclause(root, gset);
4304 792 : rollup->gsets_data = list_make1(gs);
4305 792 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4306 : rollup->gsets_data,
4307 : gd->tleref_to_colnum_map);
4308 792 : rollup->numGroups = gs->numGroups;
4309 792 : rollup->hashable = true;
4310 792 : rollup->is_hashed = true;
4311 792 : new_rollups = lappend(new_rollups, rollup);
4312 : }
4313 : }
4314 :
4315 : /*
4316 : * If we didn't find anything nonempty to hash, then bail. We'll
4317 : * generate a path from the is_sorted case.
4318 : */
4319 336 : if (new_rollups == NIL)
4320 0 : return;
4321 :
4322 : /*
4323 : * If there were empty grouping sets they should have been in the
4324 : * first rollup.
4325 : */
4326 : Assert(!unhashed_rollup || !empty_sets);
4327 :
4328 336 : if (unhashed_rollup)
4329 : {
4330 0 : new_rollups = lappend(new_rollups, unhashed_rollup);
4331 0 : strat = AGG_MIXED;
4332 : }
4333 336 : else if (empty_sets)
4334 : {
4335 212 : RollupData *rollup = makeNode(RollupData);
4336 :
4337 212 : rollup->groupClause = NIL;
4338 212 : rollup->gsets_data = empty_sets_data;
4339 212 : rollup->gsets = empty_sets;
4340 212 : rollup->numGroups = list_length(empty_sets);
4341 212 : rollup->hashable = false;
4342 212 : rollup->is_hashed = false;
4343 212 : new_rollups = lappend(new_rollups, rollup);
4344 212 : strat = AGG_MIXED;
4345 : }
4346 :
4347 336 : add_path(grouped_rel, (Path *)
4348 336 : create_groupingsets_path(root,
4349 : grouped_rel,
4350 : path,
4351 336 : (List *) parse->havingQual,
4352 : strat,
4353 : new_rollups,
4354 : agg_costs,
4355 : dNumGroups));
4356 336 : return;
4357 : }
4358 :
4359 : /*
4360 : * If we have sorted input but nothing we can do with it, bail.
4361 : */
4362 440 : if (list_length(gd->rollups) == 0)
4363 0 : return;
4364 :
4365 : /*
4366 : * Given sorted input, we try and make two paths: one sorted and one mixed
4367 : * sort/hash. (We need to try both because hashagg might be disabled, or
4368 : * some columns might not be sortable.)
4369 : *
4370 : * can_hash is passed in as false if some obstacle elsewhere (such as
4371 : * ordered aggs) means that we shouldn't consider hashing at all.
4372 : */
4373 440 : if (can_hash && gd->any_hashable)
4374 : {
4375 388 : List *rollups = NIL;
4376 388 : List *hash_sets = list_copy(gd->unsortable_sets);
4377 388 : double availspace = (work_mem * 1024.0);
4378 : ListCell *lc;
4379 :
4380 : /*
4381 : * Account first for space needed for groups we can't sort at all.
4382 : */
4383 388 : availspace -= estimate_hashagg_tablesize(path,
4384 : agg_costs,
4385 : gd->dNumHashGroups);
4386 :
4387 388 : if (availspace > 0 && list_length(gd->rollups) > 1)
4388 : {
4389 : double scale;
4390 184 : int num_rollups = list_length(gd->rollups);
4391 : int k_capacity;
4392 184 : int *k_weights = palloc(num_rollups * sizeof(int));
4393 184 : Bitmapset *hash_items = NULL;
4394 : int i;
4395 :
4396 : /*
4397 : * We treat this as a knapsack problem: the knapsack capacity
4398 : * represents work_mem, the item weights are the estimated memory
4399 : * usage of the hashtables needed to implement a single rollup,
4400 : * and we really ought to use the cost saving as the item value;
4401 : * however, currently the costs assigned to sort nodes don't
4402 : * reflect the comparison costs well, and so we treat all items as
4403 : * of equal value (each rollup we hash instead saves us one sort).
4404 : *
4405 : * To use the discrete knapsack, we need to scale the values to a
4406 : * reasonably small bounded range. We choose to allow a 5% error
4407 : * margin; we have no more than 4096 rollups in the worst possible
4408 : * case, which with a 5% error margin will require a bit over 42MB
4409 : * of workspace. (Anyone wanting to plan queries that complex had
4410 : * better have the memory for it. In more reasonable cases, with
4411 : * no more than a couple of dozen rollups, the memory usage will
4412 : * be negligible.)
4413 : *
4414 : * k_capacity is naturally bounded, but we clamp the values for
4415 : * scale and weight (below) to avoid overflows or underflows (or
4416 : * uselessly trying to use a scale factor less than 1 byte).
4417 : */
4418 184 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4419 184 : k_capacity = (int) floor(availspace / scale);
4420 :
4421 : /*
4422 : * We leave the first rollup out of consideration since it's the
4423 : * one that matches the input sort order. We assign indexes "i"
4424 : * to only those entries considered for hashing; the second loop,
4425 : * below, must use the same condition.
4426 : */
4427 184 : i = 0;
4428 496 : for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4429 : {
4430 312 : RollupData *rollup = lfirst_node(RollupData, lc);
4431 :
4432 312 : if (rollup->hashable)
4433 : {
4434 312 : double sz = estimate_hashagg_tablesize(path,
4435 : agg_costs,
4436 : rollup->numGroups);
4437 :
4438 : /*
4439 : * If sz is enormous, but work_mem (and hence scale) is
4440 : * small, avoid integer overflow here.
4441 : */
4442 312 : k_weights[i] = (int) Min(floor(sz / scale),
4443 : k_capacity + 1.0);
4444 312 : ++i;
4445 : }
4446 : }
4447 :
4448 : /*
4449 : * Apply knapsack algorithm; compute the set of items which
4450 : * maximizes the value stored (in this case the number of sorts
4451 : * saved) while keeping the total size (approximately) within
4452 : * capacity.
4453 : */
4454 184 : if (i > 0)
4455 184 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4456 :
4457 184 : if (!bms_is_empty(hash_items))
4458 : {
4459 184 : rollups = list_make1(linitial(gd->rollups));
4460 :
4461 184 : i = 0;
4462 496 : for_each_cell(lc, gd->rollups, list_second_cell(gd->rollups))
4463 : {
4464 312 : RollupData *rollup = lfirst_node(RollupData, lc);
4465 :
4466 312 : if (rollup->hashable)
4467 : {
4468 312 : if (bms_is_member(i, hash_items))
4469 288 : hash_sets = list_concat(hash_sets,
4470 288 : list_copy(rollup->gsets_data));
4471 : else
4472 24 : rollups = lappend(rollups, rollup);
4473 312 : ++i;
4474 : }
4475 : else
4476 0 : rollups = lappend(rollups, rollup);
4477 : }
4478 : }
4479 : }
4480 :
4481 388 : if (!rollups && hash_sets)
4482 16 : rollups = list_copy(gd->rollups);
4483 :
4484 724 : foreach(lc, hash_sets)
4485 : {
4486 336 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4487 336 : RollupData *rollup = makeNode(RollupData);
4488 :
4489 : Assert(gs->set != NIL);
4490 :
4491 336 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4492 336 : rollup->gsets_data = list_make1(gs);
4493 336 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4494 : rollup->gsets_data,
4495 : gd->tleref_to_colnum_map);
4496 336 : rollup->numGroups = gs->numGroups;
4497 336 : rollup->hashable = true;
4498 336 : rollup->is_hashed = true;
4499 336 : rollups = lcons(rollup, rollups);
4500 : }
4501 :
4502 388 : if (rollups)
4503 : {
4504 200 : add_path(grouped_rel, (Path *)
4505 200 : create_groupingsets_path(root,
4506 : grouped_rel,
4507 : path,
4508 200 : (List *) parse->havingQual,
4509 : AGG_MIXED,
4510 : rollups,
4511 : agg_costs,
4512 : dNumGroups));
4513 : }
4514 : }
4515 :
4516 : /*
4517 : * Now try the simple sorted case.
4518 : */
4519 440 : if (!gd->unsortable_sets)
4520 420 : add_path(grouped_rel, (Path *)
4521 840 : create_groupingsets_path(root,
4522 : grouped_rel,
4523 : path,
4524 420 : (List *) parse->havingQual,
4525 : AGG_SORTED,
4526 : gd->rollups,
4527 : agg_costs,
4528 : dNumGroups));
4529 : }
4530 :
4531 : /*
4532 : * create_window_paths
4533 : *
4534 : * Build a new upperrel containing Paths for window-function evaluation.
4535 : *
4536 : * input_rel: contains the source-data Paths
4537 : * input_target: result of make_window_input_target
4538 : * output_target: what the topmost WindowAggPath should return
4539 : * wflists: result of find_window_functions
4540 : * activeWindows: result of select_active_windows
4541 : *
4542 : * Note: all Paths in input_rel are expected to return input_target.
4543 : */
4544 : static RelOptInfo *
4545 1100 : create_window_paths(PlannerInfo *root,
4546 : RelOptInfo *input_rel,
4547 : PathTarget *input_target,
4548 : PathTarget *output_target,
4549 : bool output_target_parallel_safe,
4550 : WindowFuncLists *wflists,
4551 : List *activeWindows)
4552 : {
4553 : RelOptInfo *window_rel;
4554 : ListCell *lc;
4555 :
4556 : /* For now, do all work in the (WINDOW, NULL) upperrel */
4557 1100 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4558 :
4559 : /*
4560 : * If the input relation is not parallel-safe, then the window relation
4561 : * can't be parallel-safe, either. Otherwise, we need to examine the
4562 : * target list and active windows for non-parallel-safe constructs.
4563 : */
4564 1100 : if (input_rel->consider_parallel && output_target_parallel_safe &&
4565 0 : is_parallel_safe(root, (Node *) activeWindows))
4566 0 : window_rel->consider_parallel = true;
4567 :
4568 : /*
4569 : * If the input rel belongs to a single FDW, so does the window rel.
4570 : */
4571 1100 : window_rel->serverid = input_rel->serverid;
4572 1100 : window_rel->userid = input_rel->userid;
4573 1100 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4574 1100 : window_rel->fdwroutine = input_rel->fdwroutine;
4575 :
4576 : /*
4577 : * Consider computing window functions starting from the existing
4578 : * cheapest-total path (which will likely require a sort) as well as any
4579 : * existing paths that satisfy root->window_pathkeys (which won't).
4580 : */
4581 2300 : foreach(lc, input_rel->pathlist)
4582 : {
4583 1200 : Path *path = (Path *) lfirst(lc);
4584 :
4585 1300 : if (path == input_rel->cheapest_total_path ||
4586 100 : pathkeys_contained_in(root->window_pathkeys, path->pathkeys))
4587 1144 : create_one_window_path(root,
4588 : window_rel,
4589 : path,
4590 : input_target,
4591 : output_target,
4592 : wflists,
4593 : activeWindows);
4594 : }
4595 :
4596 : /*
4597 : * If there is an FDW that's responsible for all baserels of the query,
4598 : * let it consider adding ForeignPaths.
4599 : */
4600 1112 : if (window_rel->fdwroutine &&
4601 12 : window_rel->fdwroutine->GetForeignUpperPaths)
4602 12 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4603 : input_rel, window_rel,
4604 : NULL);
4605 :
4606 : /* Let extensions possibly add some more paths */
4607 1100 : if (create_upper_paths_hook)
4608 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4609 : input_rel, window_rel, NULL);
4610 :
4611 : /* Now choose the best path(s) */
4612 1100 : set_cheapest(window_rel);
4613 :
4614 1100 : return window_rel;
4615 : }
4616 :
4617 : /*
4618 : * Stack window-function implementation steps atop the given Path, and
4619 : * add the result to window_rel.
4620 : *
4621 : * window_rel: upperrel to contain result
4622 : * path: input Path to use (must return input_target)
4623 : * input_target: result of make_window_input_target
4624 : * output_target: what the topmost WindowAggPath should return
4625 : * wflists: result of find_window_functions
4626 : * activeWindows: result of select_active_windows
4627 : */
4628 : static void
4629 1144 : create_one_window_path(PlannerInfo *root,
4630 : RelOptInfo *window_rel,
4631 : Path *path,
4632 : PathTarget *input_target,
4633 : PathTarget *output_target,
4634 : WindowFuncLists *wflists,
4635 : List *activeWindows)
4636 : {
4637 : PathTarget *window_target;
4638 : ListCell *l;
4639 :
4640 : /*
4641 : * Since each window clause could require a different sort order, we stack
4642 : * up a WindowAgg node for each clause, with sort steps between them as
4643 : * needed. (We assume that select_active_windows chose a good order for
4644 : * executing the clauses in.)
4645 : *
4646 : * input_target should contain all Vars and Aggs needed for the result.
4647 : * (In some cases we wouldn't need to propagate all of these all the way
4648 : * to the top, since they might only be needed as inputs to WindowFuncs.
4649 : * It's probably not worth trying to optimize that though.) It must also
4650 : * contain all window partitioning and sorting expressions, to ensure
4651 : * they're computed only once at the bottom of the stack (that's critical
4652 : * for volatile functions). As we climb up the stack, we'll add outputs
4653 : * for the WindowFuncs computed at each level.
4654 : */
4655 1144 : window_target = input_target;
4656 :
4657 2336 : foreach(l, activeWindows)
4658 : {
4659 1192 : WindowClause *wc = lfirst_node(WindowClause, l);
4660 : List *window_pathkeys;
4661 :
4662 1192 : window_pathkeys = make_pathkeys_for_window(root,
4663 : wc,
4664 : root->processed_tlist);
4665 :
4666 : /* Sort if necessary */
4667 1192 : if (!pathkeys_contained_in(window_pathkeys, path->pathkeys))
4668 : {
4669 940 : path = (Path *) create_sort_path(root, window_rel,
4670 : path,
4671 : window_pathkeys,
4672 : -1.0);
4673 : }
4674 :
4675 1192 : if (lnext(activeWindows, l))
4676 : {
4677 : /*
4678 : * Add the current WindowFuncs to the output target for this
4679 : * intermediate WindowAggPath. We must copy window_target to
4680 : * avoid changing the previous path's target.
4681 : *
4682 : * Note: a WindowFunc adds nothing to the target's eval costs; but
4683 : * we do need to account for the increase in tlist width.
4684 : */
4685 : ListCell *lc2;
4686 :
4687 48 : window_target = copy_pathtarget(window_target);
4688 96 : foreach(lc2, wflists->windowFuncs[wc->winref])
4689 : {
4690 48 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4691 :
4692 48 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4693 48 : window_target->width += get_typavgwidth(wfunc->wintype, -1);
4694 : }
4695 : }
4696 : else
4697 : {
4698 : /* Install the goal target in the topmost WindowAgg */
4699 1144 : window_target = output_target;
4700 : }
4701 :
4702 1192 : path = (Path *)
4703 1192 : create_windowagg_path(root, window_rel, path, window_target,
4704 1192 : wflists->windowFuncs[wc->winref],
4705 : wc);
4706 : }
4707 :
4708 1144 : add_path(window_rel, path);
4709 1144 : }
4710 :
4711 : /*
4712 : * create_distinct_paths
4713 : *
4714 : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4715 : *
4716 : * input_rel: contains the source-data Paths
4717 : *
4718 : * Note: input paths should already compute the desired pathtarget, since
4719 : * Sort/Unique won't project anything.
4720 : */
4721 : static RelOptInfo *
4722 384 : create_distinct_paths(PlannerInfo *root,
4723 : RelOptInfo *input_rel)
4724 : {
4725 384 : Query *parse = root->parse;
4726 384 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4727 : RelOptInfo *distinct_rel;
4728 : double numDistinctRows;
4729 : bool allow_hash;
4730 : Path *path;
4731 : ListCell *lc;
4732 :
4733 : /* For now, do all work in the (DISTINCT, NULL) upperrel */
4734 384 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4735 :
4736 : /*
4737 : * We don't compute anything at this level, so distinct_rel will be
4738 : * parallel-safe if the input rel is parallel-safe. In particular, if
4739 : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4740 : * output those expressions, and will not be parallel-safe unless those
4741 : * expressions are parallel-safe.
4742 : */
4743 384 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4744 :
4745 : /*
4746 : * If the input rel belongs to a single FDW, so does the distinct_rel.
4747 : */
4748 384 : distinct_rel->serverid = input_rel->serverid;
4749 384 : distinct_rel->userid = input_rel->userid;
4750 384 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4751 384 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4752 :
4753 : /* Estimate number of distinct rows there will be */
4754 744 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
4755 360 : root->hasHavingQual)
4756 : {
4757 : /*
4758 : * If there was grouping or aggregation, use the number of input rows
4759 : * as the estimated number of DISTINCT rows (ie, assume the input is
4760 : * already mostly unique).
4761 : */
4762 24 : numDistinctRows = cheapest_input_path->rows;
4763 : }
4764 : else
4765 : {
4766 : /*
4767 : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
4768 : */
4769 : List *distinctExprs;
4770 :
4771 360 : distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
4772 : parse->targetList);
4773 360 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4774 : cheapest_input_path->rows,
4775 : NULL);
4776 : }
4777 :
4778 : /*
4779 : * Consider sort-based implementations of DISTINCT, if possible.
4780 : */
4781 384 : if (grouping_is_sortable(parse->distinctClause))
4782 : {
4783 : /*
4784 : * First, if we have any adequately-presorted paths, just stick a
4785 : * Unique node on those. Then consider doing an explicit sort of the
4786 : * cheapest input path and Unique'ing that.
4787 : *
4788 : * When we have DISTINCT ON, we must sort by the more rigorous of
4789 : * DISTINCT and ORDER BY, else it won't have the desired behavior.
4790 : * Also, if we do have to do an explicit sort, we might as well use
4791 : * the more rigorous ordering to avoid a second sort later. (Note
4792 : * that the parser will have ensured that one clause is a prefix of
4793 : * the other.)
4794 : */
4795 : List *needed_pathkeys;
4796 :
4797 416 : if (parse->hasDistinctOn &&
4798 36 : list_length(root->distinct_pathkeys) <
4799 36 : list_length(root->sort_pathkeys))
4800 20 : needed_pathkeys = root->sort_pathkeys;
4801 : else
4802 360 : needed_pathkeys = root->distinct_pathkeys;
4803 :
4804 800 : foreach(lc, input_rel->pathlist)
4805 : {
4806 420 : Path *path = (Path *) lfirst(lc);
4807 :
4808 420 : if (pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4809 : {
4810 30 : add_path(distinct_rel, (Path *)
4811 30 : create_upper_unique_path(root, distinct_rel,
4812 : path,
4813 30 : list_length(root->distinct_pathkeys),
4814 : numDistinctRows));
4815 : }
4816 : }
4817 :
4818 : /* For explicit-sort case, always use the more rigorous clause */
4819 760 : if (list_length(root->distinct_pathkeys) <
4820 380 : list_length(root->sort_pathkeys))
4821 : {
4822 20 : needed_pathkeys = root->sort_pathkeys;
4823 : /* Assert checks that parser didn't mess up... */
4824 : Assert(pathkeys_contained_in(root->distinct_pathkeys,
4825 : needed_pathkeys));
4826 : }
4827 : else
4828 360 : needed_pathkeys = root->distinct_pathkeys;
4829 :
4830 380 : path = cheapest_input_path;
4831 380 : if (!pathkeys_contained_in(needed_pathkeys, path->pathkeys))
4832 370 : path = (Path *) create_sort_path(root, distinct_rel,
4833 : path,
4834 : needed_pathkeys,
4835 : -1.0);
4836 :
4837 380 : add_path(distinct_rel, (Path *)
4838 380 : create_upper_unique_path(root, distinct_rel,
4839 : path,
4840 380 : list_length(root->distinct_pathkeys),
4841 : numDistinctRows));
4842 : }
4843 :
4844 : /*
4845 : * Consider hash-based implementations of DISTINCT, if possible.
4846 : *
4847 : * If we were not able to make any other types of path, we *must* hash or
4848 : * die trying. If we do have other choices, there are several things that
4849 : * should prevent selection of hashing: if the query uses DISTINCT ON
4850 : * (because it won't really have the expected behavior if we hash), or if
4851 : * enable_hashagg is off, or if it looks like the hashtable will exceed
4852 : * work_mem.
4853 : *
4854 : * Note: grouping_is_hashable() is much more expensive to check than the
4855 : * other gating conditions, so we want to do it last.
4856 : */
4857 384 : if (distinct_rel->pathlist == NIL)
4858 4 : allow_hash = true; /* we have no alternatives */
4859 380 : else if (parse->hasDistinctOn || !enable_hashagg)
4860 36 : allow_hash = false; /* policy-based decision not to hash */
4861 : else
4862 : {
4863 : Size hashentrysize;
4864 :
4865 : /* Estimate per-hash-entry space at tuple width... */
4866 344 : hashentrysize = MAXALIGN(cheapest_input_path->pathtarget->width) +
4867 : MAXALIGN(SizeofMinimalTupleHeader);
4868 : /* plus the per-hash-entry overhead */
4869 344 : hashentrysize += hash_agg_entry_size(0);
4870 :
4871 : /* Allow hashing only if hashtable is predicted to fit in work_mem */
4872 344 : allow_hash = (hashentrysize * numDistinctRows <= work_mem * 1024L);
4873 : }
4874 :
4875 384 : if (allow_hash && grouping_is_hashable(parse->distinctClause))
4876 : {
4877 : /* Generate hashed aggregate path --- no sort needed */
4878 348 : add_path(distinct_rel, (Path *)
4879 348 : 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 384 : 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 400 : 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 384 : 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 384 : set_cheapest(distinct_rel);
4915 :
4916 384 : 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 path we need consider is an explicit sort on the
4926 : * 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 : static RelOptInfo *
4934 34758 : create_ordered_paths(PlannerInfo *root,
4935 : RelOptInfo *input_rel,
4936 : PathTarget *target,
4937 : bool target_parallel_safe,
4938 : double limit_tuples)
4939 : {
4940 34758 : Path *cheapest_input_path = input_rel->cheapest_total_path;
4941 : RelOptInfo *ordered_rel;
4942 : ListCell *lc;
4943 :
4944 : /* For now, do all work in the (ORDERED, NULL) upperrel */
4945 34758 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
4946 :
4947 : /*
4948 : * If the input relation is not parallel-safe, then the ordered relation
4949 : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4950 : * target list is parallel-safe.
4951 : */
4952 34758 : if (input_rel->consider_parallel && target_parallel_safe)
4953 14434 : ordered_rel->consider_parallel = true;
4954 :
4955 : /*
4956 : * If the input rel belongs to a single FDW, so does the ordered_rel.
4957 : */
4958 34758 : ordered_rel->serverid = input_rel->serverid;
4959 34758 : ordered_rel->userid = input_rel->userid;
4960 34758 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
4961 34758 : ordered_rel->fdwroutine = input_rel->fdwroutine;
4962 :
4963 83010 : foreach(lc, input_rel->pathlist)
4964 : {
4965 48252 : Path *path = (Path *) lfirst(lc);
4966 : bool is_sorted;
4967 :
4968 48252 : is_sorted = pathkeys_contained_in(root->sort_pathkeys,
4969 : path->pathkeys);
4970 48252 : if (path == cheapest_input_path || is_sorted)
4971 : {
4972 46978 : if (!is_sorted)
4973 : {
4974 : /* An explicit sort here can take advantage of LIMIT */
4975 27616 : path = (Path *) create_sort_path(root,
4976 : ordered_rel,
4977 : path,
4978 : root->sort_pathkeys,
4979 : limit_tuples);
4980 : }
4981 :
4982 : /* Add projection step if needed */
4983 46978 : if (path->pathtarget != target)
4984 8766 : path = apply_projection_to_path(root, ordered_rel,
4985 : path, target);
4986 :
4987 46978 : add_path(ordered_rel, path);
4988 : }
4989 : }
4990 :
4991 : /*
4992 : * generate_gather_paths() will have already generated a simple Gather
4993 : * path for the best parallel path, if any, and the loop above will have
4994 : * considered sorting it. Similarly, generate_gather_paths() will also
4995 : * have generated order-preserving Gather Merge plans which can be used
4996 : * without sorting if they happen to match the sort_pathkeys, and the loop
4997 : * above will have handled those as well. However, there's one more
4998 : * possibility: it may make sense to sort the cheapest partial path
4999 : * according to the required output order and then use Gather Merge.
5000 : */
5001 49084 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5002 14326 : input_rel->partial_pathlist != NIL)
5003 : {
5004 : Path *cheapest_partial_path;
5005 :
5006 1778 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5007 :
5008 : /*
5009 : * If cheapest partial path doesn't need a sort, this is redundant
5010 : * with what's already been tried.
5011 : */
5012 1778 : if (!pathkeys_contained_in(root->sort_pathkeys,
5013 : cheapest_partial_path->pathkeys))
5014 : {
5015 : Path *path;
5016 : double total_groups;
5017 :
5018 1770 : path = (Path *) create_sort_path(root,
5019 : ordered_rel,
5020 : cheapest_partial_path,
5021 : root->sort_pathkeys,
5022 : limit_tuples);
5023 :
5024 3540 : total_groups = cheapest_partial_path->rows *
5025 1770 : cheapest_partial_path->parallel_workers;
5026 1770 : path = (Path *)
5027 1770 : create_gather_merge_path(root, ordered_rel,
5028 : path,
5029 : path->pathtarget,
5030 : root->sort_pathkeys, NULL,
5031 : &total_groups);
5032 :
5033 : /* Add projection step if needed */
5034 1770 : if (path->pathtarget != target)
5035 1178 : path = apply_projection_to_path(root, ordered_rel,
5036 : path, target);
5037 :
5038 1770 : add_path(ordered_rel, path);
5039 : }
5040 : }
5041 :
5042 : /*
5043 : * If there is an FDW that's responsible for all baserels of the query,
5044 : * let it consider adding ForeignPaths.
5045 : */
5046 35078 : if (ordered_rel->fdwroutine &&
5047 320 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5048 306 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5049 : input_rel, ordered_rel,
5050 : NULL);
5051 :
5052 : /* Let extensions possibly add some more paths */
5053 34758 : if (create_upper_paths_hook)
5054 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5055 : input_rel, ordered_rel, NULL);
5056 :
5057 : /*
5058 : * No need to bother with set_cheapest here; grouping_planner does not
5059 : * need us to do it.
5060 : */
5061 : Assert(ordered_rel->pathlist != NIL);
5062 :
5063 34758 : return ordered_rel;
5064 : }
5065 :
5066 :
5067 : /*
5068 : * make_group_input_target
5069 : * Generate appropriate PathTarget for initial input to grouping nodes.
5070 : *
5071 : * If there is grouping or aggregation, the scan/join subplan cannot emit
5072 : * the query's final targetlist; for example, it certainly can't emit any
5073 : * aggregate function calls. This routine generates the correct target
5074 : * for the scan/join subplan.
5075 : *
5076 : * The query target list passed from the parser already contains entries
5077 : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5078 : * for variables used only in HAVING clauses; so we need to add those
5079 : * variables to the subplan target list. Also, we flatten all expressions
5080 : * except GROUP BY items into their component variables; other expressions
5081 : * will be computed by the upper plan nodes rather than by the subplan.
5082 : * For example, given a query like
5083 : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5084 : * we want to pass this targetlist to the subplan:
5085 : * a+b,c,d
5086 : * where the a+b target will be used by the Sort/Group steps, and the
5087 : * other targets will be used for computing the final results.
5088 : *
5089 : * 'final_target' is the query's final target list (in PathTarget form)
5090 : *
5091 : * The result is the PathTarget to be computed by the Paths returned from
5092 : * query_planner().
5093 : */
5094 : static PathTarget *
5095 23624 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5096 : {
5097 23624 : Query *parse = root->parse;
5098 : PathTarget *input_target;
5099 : List *non_group_cols;
5100 : List *non_group_vars;
5101 : int i;
5102 : ListCell *lc;
5103 :
5104 : /*
5105 : * We must build a target containing all grouping columns, plus any other
5106 : * Vars mentioned in the query's targetlist and HAVING qual.
5107 : */
5108 23624 : input_target = create_empty_pathtarget();
5109 23624 : non_group_cols = NIL;
5110 :
5111 23624 : i = 0;
5112 53004 : foreach(lc, final_target->exprs)
5113 : {
5114 29380 : Expr *expr = (Expr *) lfirst(lc);
5115 29380 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5116 :
5117 33338 : if (sgref && parse->groupClause &&
5118 3958 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5119 : {
5120 : /*
5121 : * It's a grouping column, so add it to the input target as-is.
5122 : */
5123 3166 : add_column_to_pathtarget(input_target, expr, sgref);
5124 : }
5125 : else
5126 : {
5127 : /*
5128 : * Non-grouping column, so just remember the expression for later
5129 : * call to pull_var_clause.
5130 : */
5131 26214 : non_group_cols = lappend(non_group_cols, expr);
5132 : }
5133 :
5134 29380 : i++;
5135 : }
5136 :
5137 : /*
5138 : * If there's a HAVING clause, we'll need the Vars it uses, too.
5139 : */
5140 23624 : if (parse->havingQual)
5141 308 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5142 :
5143 : /*
5144 : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5145 : * add them to the input target if not already present. (A Var used
5146 : * directly as a GROUP BY item will be present already.) Note this
5147 : * includes Vars used in resjunk items, so we are covering the needs of
5148 : * ORDER BY and window specifications. Vars used within Aggrefs and
5149 : * WindowFuncs will be pulled out here, too.
5150 : */
5151 23624 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5152 : PVC_RECURSE_AGGREGATES |
5153 : PVC_RECURSE_WINDOWFUNCS |
5154 : PVC_INCLUDE_PLACEHOLDERS);
5155 23624 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5156 :
5157 : /* clean up cruft */
5158 23624 : list_free(non_group_vars);
5159 23624 : list_free(non_group_cols);
5160 :
5161 : /* XXX this causes some redundant cost calculation ... */
5162 23624 : return set_pathtarget_cost_width(root, input_target);
5163 : }
5164 :
5165 : /*
5166 : * make_partial_grouping_target
5167 : * Generate appropriate PathTarget for output of partial aggregate
5168 : * (or partial grouping, if there are no aggregates) nodes.
5169 : *
5170 : * A partial aggregation node needs to emit all the same aggregates that
5171 : * a regular aggregation node would, plus any aggregates used in HAVING;
5172 : * except that the Aggref nodes should be marked as partial aggregates.
5173 : *
5174 : * In addition, we'd better emit any Vars and PlaceholderVars that are
5175 : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5176 : * these would be Vars that are grouped by or used in grouping expressions.)
5177 : *
5178 : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5179 : * havingQual represents the HAVING clause.
5180 : */
5181 : static PathTarget *
5182 1290 : make_partial_grouping_target(PlannerInfo *root,
5183 : PathTarget *grouping_target,
5184 : Node *havingQual)
5185 : {
5186 1290 : Query *parse = root->parse;
5187 : PathTarget *partial_target;
5188 : List *non_group_cols;
5189 : List *non_group_exprs;
5190 : int i;
5191 : ListCell *lc;
5192 :
5193 1290 : partial_target = create_empty_pathtarget();
5194 1290 : non_group_cols = NIL;
5195 :
5196 1290 : i = 0;
5197 4668 : foreach(lc, grouping_target->exprs)
5198 : {
5199 3378 : Expr *expr = (Expr *) lfirst(lc);
5200 3378 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5201 :
5202 5730 : if (sgref && parse->groupClause &&
5203 2352 : get_sortgroupref_clause_noerr(sgref, parse->groupClause) != NULL)
5204 : {
5205 : /*
5206 : * It's a grouping column, so add it to the partial_target as-is.
5207 : * (This allows the upper agg step to repeat the grouping calcs.)
5208 : */
5209 1128 : add_column_to_pathtarget(partial_target, expr, sgref);
5210 : }
5211 : else
5212 : {
5213 : /*
5214 : * Non-grouping column, so just remember the expression for later
5215 : * call to pull_var_clause.
5216 : */
5217 2250 : non_group_cols = lappend(non_group_cols, expr);
5218 : }
5219 :
5220 3378 : i++;
5221 : }
5222 :
5223 : /*
5224 : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5225 : */
5226 1290 : if (havingQual)
5227 552 : non_group_cols = lappend(non_group_cols, havingQual);
5228 :
5229 : /*
5230 : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5231 : * non-group cols (plus HAVING), and add them to the partial_target if not
5232 : * already present. (An expression used directly as a GROUP BY item will
5233 : * be present already.) Note this includes Vars used in resjunk items, so
5234 : * we are covering the needs of ORDER BY and window specifications.
5235 : */
5236 1290 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5237 : PVC_INCLUDE_AGGREGATES |
5238 : PVC_RECURSE_WINDOWFUNCS |
5239 : PVC_INCLUDE_PLACEHOLDERS);
5240 :
5241 1290 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5242 :
5243 : /*
5244 : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5245 : * are at the top level of the target list, so we can just scan the list
5246 : * rather than recursing through the expression trees.
5247 : */
5248 5076 : foreach(lc, partial_target->exprs)
5249 : {
5250 3786 : Aggref *aggref = (Aggref *) lfirst(lc);
5251 :
5252 3786 : if (IsA(aggref, Aggref))
5253 : {
5254 : Aggref *newaggref;
5255 :
5256 : /*
5257 : * We shouldn't need to copy the substructure of the Aggref node,
5258 : * but flat-copy the node itself to avoid damaging other trees.
5259 : */
5260 2658 : newaggref = makeNode(Aggref);
5261 2658 : memcpy(newaggref, aggref, sizeof(Aggref));
5262 :
5263 : /* For now, assume serialization is required */
5264 2658 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5265 :
5266 2658 : lfirst(lc) = newaggref;
5267 : }
5268 : }
5269 :
5270 : /* clean up cruft */
5271 1290 : list_free(non_group_exprs);
5272 1290 : list_free(non_group_cols);
5273 :
5274 : /* XXX this causes some redundant cost calculation ... */
5275 1290 : return set_pathtarget_cost_width(root, partial_target);
5276 : }
5277 :
5278 : /*
5279 : * mark_partial_aggref
5280 : * Adjust an Aggref to make it represent a partial-aggregation step.
5281 : *
5282 : * The Aggref node is modified in-place; caller must do any copying required.
5283 : */
5284 : void
5285 4346 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5286 : {
5287 : /* aggtranstype should be computed by this point */
5288 : Assert(OidIsValid(agg->aggtranstype));
5289 : /* ... but aggsplit should still be as the parser left it */
5290 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5291 :
5292 : /* Mark the Aggref with the intended partial-aggregation mode */
5293 4346 : agg->aggsplit = aggsplit;
5294 :
5295 : /*
5296 : * Adjust result type if needed. Normally, a partial aggregate returns
5297 : * the aggregate's transition type; but if that's INTERNAL and we're
5298 : * serializing, it returns BYTEA instead.
5299 : */
5300 4346 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5301 : {
5302 3502 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5303 48 : agg->aggtype = BYTEAOID;
5304 : else
5305 3454 : agg->aggtype = agg->aggtranstype;
5306 : }
5307 4346 : }
5308 :
5309 : /*
5310 : * postprocess_setop_tlist
5311 : * Fix up targetlist returned by plan_set_operations().
5312 : *
5313 : * We need to transpose sort key info from the orig_tlist into new_tlist.
5314 : * NOTE: this would not be good enough if we supported resjunk sort keys
5315 : * for results of set operations --- then, we'd need to project a whole
5316 : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5317 : * find any resjunk columns in orig_tlist.
5318 : */
5319 : static List *
5320 1088 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5321 : {
5322 : ListCell *l;
5323 1088 : ListCell *orig_tlist_item = list_head(orig_tlist);
5324 :
5325 2938 : foreach(l, new_tlist)
5326 : {
5327 1850 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5328 : TargetEntry *orig_tle;
5329 :
5330 : /* ignore resjunk columns in setop result */
5331 1850 : if (new_tle->resjunk)
5332 200 : continue;
5333 :
5334 : Assert(orig_tlist_item != NULL);
5335 1650 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
5336 1650 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
5337 1650 : if (orig_tle->resjunk) /* should not happen */
5338 0 : elog(ERROR, "resjunk output columns are not implemented");
5339 : Assert(new_tle->resno == orig_tle->resno);
5340 1650 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5341 : }
5342 1088 : if (orig_tlist_item != NULL)
5343 0 : elog(ERROR, "resjunk output columns are not implemented");
5344 1088 : return new_tlist;
5345 : }
5346 :
5347 : /*
5348 : * select_active_windows
5349 : * Create a list of the "active" window clauses (ie, those referenced
5350 : * by non-deleted WindowFuncs) in the order they are to be executed.
5351 : */
5352 : static List *
5353 1100 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5354 : {
5355 1100 : List *windowClause = root->parse->windowClause;
5356 1100 : List *result = NIL;
5357 : ListCell *lc;
5358 1100 : int nActive = 0;
5359 1100 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5360 1100 : * list_length(windowClause));
5361 :
5362 : /* First, construct an array of the active windows */
5363 2264 : foreach(lc, windowClause)
5364 : {
5365 1164 : WindowClause *wc = lfirst_node(WindowClause, lc);
5366 :
5367 : /* It's only active if wflists shows some related WindowFuncs */
5368 : Assert(wc->winref <= wflists->maxWinRef);
5369 1164 : if (wflists->windowFuncs[wc->winref] == NIL)
5370 16 : continue;
5371 :
5372 1148 : actives[nActive].wc = wc; /* original clause */
5373 :
5374 : /*
5375 : * For sorting, we want the list of partition keys followed by the
5376 : * list of sort keys. But pathkeys construction will remove duplicates
5377 : * between the two, so we can as well (even though we can't detect all
5378 : * of the duplicates, since some may come from ECs - that might mean
5379 : * we miss optimization chances here). We must, however, ensure that
5380 : * the order of entries is preserved with respect to the ones we do
5381 : * keep.
5382 : *
5383 : * partitionClause and orderClause had their own duplicates removed in
5384 : * parse analysis, so we're only concerned here with removing
5385 : * orderClause entries that also appear in partitionClause.
5386 : */
5387 2296 : actives[nActive].uniqueOrder =
5388 1148 : list_concat_unique(list_copy(wc->partitionClause),
5389 1148 : wc->orderClause);
5390 1148 : nActive++;
5391 : }
5392 :
5393 : /*
5394 : * Sort active windows by their partitioning/ordering clauses, ignoring
5395 : * any framing clauses, so that the windows that need the same sorting are
5396 : * adjacent in the list. When we come to generate paths, this will avoid
5397 : * inserting additional Sort nodes.
5398 : *
5399 : * This is how we implement a specific requirement from the SQL standard,
5400 : * which says that when two or more windows are order-equivalent (i.e.
5401 : * have matching partition and order clauses, even if their names or
5402 : * framing clauses differ), then all peer rows must be presented in the
5403 : * same order in all of them. If we allowed multiple sort nodes for such
5404 : * cases, we'd risk having the peer rows end up in different orders in
5405 : * equivalent windows due to sort instability. (See General Rule 4 of
5406 : * <window clause> in SQL2008 - SQL2016.)
5407 : *
5408 : * Additionally, if the entire list of clauses of one window is a prefix
5409 : * of another, put first the window with stronger sorting requirements.
5410 : * This way we will first sort for stronger window, and won't have to sort
5411 : * again for the weaker one.
5412 : */
5413 1100 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
5414 :
5415 : /* build ordered list of the original WindowClause nodes */
5416 2248 : for (int i = 0; i < nActive; i++)
5417 1148 : result = lappend(result, actives[i].wc);
5418 :
5419 1100 : pfree(actives);
5420 :
5421 1100 : return result;
5422 : }
5423 :
5424 : /*
5425 : * common_prefix_cmp
5426 : * QSort comparison function for WindowClauseSortData
5427 : *
5428 : * Sort the windows by the required sorting clauses. First, compare the sort
5429 : * clauses themselves. Second, if one window's clauses are a prefix of another
5430 : * one's clauses, put the window with more sort clauses first.
5431 : */
5432 : static int
5433 48 : common_prefix_cmp(const void *a, const void *b)
5434 : {
5435 48 : const WindowClauseSortData *wcsa = a;
5436 48 : const WindowClauseSortData *wcsb = b;
5437 : ListCell *item_a;
5438 : ListCell *item_b;
5439 :
5440 84 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
5441 : {
5442 56 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
5443 56 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
5444 :
5445 56 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
5446 20 : return -1;
5447 56 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
5448 12 : return 1;
5449 44 : else if (sca->sortop > scb->sortop)
5450 0 : return -1;
5451 44 : else if (sca->sortop < scb->sortop)
5452 8 : return 1;
5453 36 : else if (sca->nulls_first && !scb->nulls_first)
5454 0 : return -1;
5455 36 : else if (!sca->nulls_first && scb->nulls_first)
5456 0 : return 1;
5457 : /* no need to compare eqop, since it is fully determined by sortop */
5458 : }
5459 :
5460 28 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
5461 4 : return -1;
5462 24 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
5463 8 : return 1;
5464 :
5465 16 : return 0;
5466 : }
5467 :
5468 : /*
5469 : * make_window_input_target
5470 : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
5471 : *
5472 : * When the query has window functions, this function computes the desired
5473 : * target to be computed by the node just below the first WindowAgg.
5474 : * This tlist must contain all values needed to evaluate the window functions,
5475 : * compute the final target list, and perform any required final sort step.
5476 : * If multiple WindowAggs are needed, each intermediate one adds its window
5477 : * function results onto this base tlist; only the topmost WindowAgg computes
5478 : * the actual desired target list.
5479 : *
5480 : * This function is much like make_group_input_target, though not quite enough
5481 : * like it to share code. As in that function, we flatten most expressions
5482 : * into their component variables. But we do not want to flatten window
5483 : * PARTITION BY/ORDER BY clauses, since that might result in multiple
5484 : * evaluations of them, which would be bad (possibly even resulting in
5485 : * inconsistent answers, if they contain volatile functions).
5486 : * Also, we must not flatten GROUP BY clauses that were left unflattened by
5487 : * make_group_input_target, because we may no longer have access to the
5488 : * individual Vars in them.
5489 : *
5490 : * Another key difference from make_group_input_target is that we don't
5491 : * flatten Aggref expressions, since those are to be computed below the
5492 : * window functions and just referenced like Vars above that.
5493 : *
5494 : * 'final_target' is the query's final target list (in PathTarget form)
5495 : * 'activeWindows' is the list of active windows previously identified by
5496 : * select_active_windows.
5497 : *
5498 : * The result is the PathTarget to be computed by the plan node immediately
5499 : * below the first WindowAgg node.
5500 : */
5501 : static PathTarget *
5502 1100 : make_window_input_target(PlannerInfo *root,
5503 : PathTarget *final_target,
5504 : List *activeWindows)
5505 : {
5506 1100 : Query *parse = root->parse;
5507 : PathTarget *input_target;
5508 : Bitmapset *sgrefs;
5509 : List *flattenable_cols;
5510 : List *flattenable_vars;
5511 : int i;
5512 : ListCell *lc;
5513 :
5514 : Assert(parse->hasWindowFuncs);
5515 :
5516 : /*
5517 : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
5518 : * into a bitmapset for convenient reference below.
5519 : */
5520 1100 : sgrefs = NULL;
5521 2248 : foreach(lc, activeWindows)
5522 : {
5523 1148 : WindowClause *wc = lfirst_node(WindowClause, lc);
5524 : ListCell *lc2;
5525 :
5526 1500 : foreach(lc2, wc->partitionClause)
5527 : {
5528 352 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5529 :
5530 352 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5531 : }
5532 2192 : foreach(lc2, wc->orderClause)
5533 : {
5534 1044 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
5535 :
5536 1044 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
5537 : }
5538 : }
5539 :
5540 : /* Add in sortgroupref numbers of GROUP BY clauses, too */
5541 1212 : foreach(lc, parse->groupClause)
5542 : {
5543 112 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
5544 :
5545 112 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
5546 : }
5547 :
5548 : /*
5549 : * Construct a target containing all the non-flattenable targetlist items,
5550 : * and save aside the others for a moment.
5551 : */
5552 1100 : input_target = create_empty_pathtarget();
5553 1100 : flattenable_cols = NIL;
5554 :
5555 1100 : i = 0;
5556 4740 : foreach(lc, final_target->exprs)
5557 : {
5558 3640 : Expr *expr = (Expr *) lfirst(lc);
5559 3640 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5560 :
5561 : /*
5562 : * Don't want to deconstruct window clauses or GROUP BY items. (Note
5563 : * that such items can't contain window functions, so it's okay to
5564 : * compute them below the WindowAgg nodes.)
5565 : */
5566 3640 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
5567 : {
5568 : /*
5569 : * Don't want to deconstruct this value, so add it to the input
5570 : * target as-is.
5571 : */
5572 1356 : add_column_to_pathtarget(input_target, expr, sgref);
5573 : }
5574 : else
5575 : {
5576 : /*
5577 : * Column is to be flattened, so just remember the expression for
5578 : * later call to pull_var_clause.
5579 : */
5580 2284 : flattenable_cols = lappend(flattenable_cols, expr);
5581 : }
5582 :
5583 3640 : i++;
5584 : }
5585 :
5586 : /*
5587 : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
5588 : * add them to the input target if not already present. (Some might be
5589 : * there already because they're used directly as window/group clauses.)
5590 : *
5591 : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
5592 : * Aggrefs are placed in the Agg node's tlist and not left to be computed
5593 : * at higher levels. On the other hand, we should recurse into
5594 : * WindowFuncs to make sure their input expressions are available.
5595 : */
5596 1100 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
5597 : PVC_INCLUDE_AGGREGATES |
5598 : PVC_RECURSE_WINDOWFUNCS |
5599 : PVC_INCLUDE_PLACEHOLDERS);
5600 1100 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
5601 :
5602 : /* clean up cruft */
5603 1100 : list_free(flattenable_vars);
5604 1100 : list_free(flattenable_cols);
5605 :
5606 : /* XXX this causes some redundant cost calculation ... */
5607 1100 : return set_pathtarget_cost_width(root, input_target);
5608 : }
5609 :
5610 : /*
5611 : * make_pathkeys_for_window
5612 : * Create a pathkeys list describing the required input ordering
5613 : * for the given WindowClause.
5614 : *
5615 : * The required ordering is first the PARTITION keys, then the ORDER keys.
5616 : * In the future we might try to implement windowing using hashing, in which
5617 : * case the ordering could be relaxed, but for now we always sort.
5618 : */
5619 : static List *
5620 2292 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
5621 : List *tlist)
5622 : {
5623 : List *window_pathkeys;
5624 : List *window_sortclauses;
5625 :
5626 : /* Throw error if can't sort */
5627 2292 : if (!grouping_is_sortable(wc->partitionClause))
5628 0 : ereport(ERROR,
5629 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5630 : errmsg("could not implement window PARTITION BY"),
5631 : errdetail("Window partitioning columns must be of sortable datatypes.")));
5632 2292 : if (!grouping_is_sortable(wc->orderClause))
5633 0 : ereport(ERROR,
5634 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5635 : errmsg("could not implement window ORDER BY"),
5636 : errdetail("Window ordering columns must be of sortable datatypes.")));
5637 :
5638 : /* Okay, make the combined pathkeys */
5639 2292 : window_sortclauses = list_concat(list_copy(wc->partitionClause),
5640 2292 : list_copy(wc->orderClause));
5641 2292 : window_pathkeys = make_pathkeys_for_sortclauses(root,
5642 : window_sortclauses,
5643 : tlist);
5644 2292 : list_free(window_sortclauses);
5645 2292 : return window_pathkeys;
5646 : }
5647 :
5648 : /*
5649 : * make_sort_input_target
5650 : * Generate appropriate PathTarget for initial input to Sort step.
5651 : *
5652 : * If the query has ORDER BY, this function chooses the target to be computed
5653 : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
5654 : * project) steps. This might or might not be identical to the query's final
5655 : * output target.
5656 : *
5657 : * The main argument for keeping the sort-input tlist the same as the final
5658 : * is that we avoid a separate projection node (which will be needed if
5659 : * they're different, because Sort can't project). However, there are also
5660 : * advantages to postponing tlist evaluation till after the Sort: it ensures
5661 : * a consistent order of evaluation for any volatile functions in the tlist,
5662 : * and if there's also a LIMIT, we can stop the query without ever computing
5663 : * tlist functions for later rows, which is beneficial for both volatile and
5664 : * expensive functions.
5665 : *
5666 : * Our current policy is to postpone volatile expressions till after the sort
5667 : * unconditionally (assuming that that's possible, ie they are in plain tlist
5668 : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
5669 : * postpone set-returning expressions, because running them beforehand would
5670 : * bloat the sort dataset, and because it might cause unexpected output order
5671 : * if the sort isn't stable. However there's a constraint on that: all SRFs
5672 : * in the tlist should be evaluated at the same plan step, so that they can
5673 : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
5674 : * mustn't postpone any SRFs. (Note that in principle that policy should
5675 : * probably get applied to the group/window input targetlists too, but we
5676 : * have not done that historically.) Lastly, expensive expressions are
5677 : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
5678 : * partial evaluation of the query is possible (if neither is true, we expect
5679 : * to have to evaluate the expressions for every row anyway), or if there are
5680 : * any volatile or set-returning expressions (since once we've put in a
5681 : * projection at all, it won't cost any more to postpone more stuff).
5682 : *
5683 : * Another issue that could potentially be considered here is that
5684 : * evaluating tlist expressions could result in data that's either wider
5685 : * or narrower than the input Vars, thus changing the volume of data that
5686 : * has to go through the Sort. However, we usually have only a very bad
5687 : * idea of the output width of any expression more complex than a Var,
5688 : * so for now it seems too risky to try to optimize on that basis.
5689 : *
5690 : * Note that if we do produce a modified sort-input target, and then the
5691 : * query ends up not using an explicit Sort, no particular harm is done:
5692 : * we'll initially use the modified target for the preceding path nodes,
5693 : * but then change them to the final target with apply_projection_to_path.
5694 : * Moreover, in such a case the guarantees about evaluation order of
5695 : * volatile functions still hold, since the rows are sorted already.
5696 : *
5697 : * This function has some things in common with make_group_input_target and
5698 : * make_window_input_target, though the detailed rules for what to do are
5699 : * different. We never flatten/postpone any grouping or ordering columns;
5700 : * those are needed before the sort. If we do flatten a particular
5701 : * expression, we leave Aggref and WindowFunc nodes alone, since those were
5702 : * computed earlier.
5703 : *
5704 : * 'final_target' is the query's final target list (in PathTarget form)
5705 : * 'have_postponed_srfs' is an output argument, see below
5706 : *
5707 : * The result is the PathTarget to be computed by the plan node immediately
5708 : * below the Sort step (and the Distinct step, if any). This will be
5709 : * exactly final_target if we decide a projection step wouldn't be helpful.
5710 : *
5711 : * In addition, *have_postponed_srfs is set to true if we choose to postpone
5712 : * any set-returning functions to after the Sort.
5713 : */
5714 : static PathTarget *
5715 34554 : make_sort_input_target(PlannerInfo *root,
5716 : PathTarget *final_target,
5717 : bool *have_postponed_srfs)
5718 : {
5719 34554 : Query *parse = root->parse;
5720 : PathTarget *input_target;
5721 : int ncols;
5722 : bool *col_is_srf;
5723 : bool *postpone_col;
5724 : bool have_srf;
5725 : bool have_volatile;
5726 : bool have_expensive;
5727 : bool have_srf_sortcols;
5728 : bool postpone_srfs;
5729 : List *postponable_cols;
5730 : List *postponable_vars;
5731 : int i;
5732 : ListCell *lc;
5733 :
5734 : /* Shouldn't get here unless query has ORDER BY */
5735 : Assert(parse->sortClause);
5736 :
5737 34554 : *have_postponed_srfs = false; /* default result */
5738 :
5739 : /* Inspect tlist and collect per-column information */
5740 34554 : ncols = list_length(final_target->exprs);
5741 34554 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
5742 34554 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
5743 34554 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
5744 :
5745 34554 : i = 0;
5746 272126 : foreach(lc, final_target->exprs)
5747 : {
5748 237572 : Expr *expr = (Expr *) lfirst(lc);
5749 :
5750 : /*
5751 : * If the column has a sortgroupref, assume it has to be evaluated
5752 : * before sorting. Generally such columns would be ORDER BY, GROUP
5753 : * BY, etc targets. One exception is columns that were removed from
5754 : * GROUP BY by remove_useless_groupby_columns() ... but those would
5755 : * only be Vars anyway. There don't seem to be any cases where it
5756 : * would be worth the trouble to double-check.
5757 : */
5758 237572 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
5759 : {
5760 : /*
5761 : * Check for SRF or volatile functions. Check the SRF case first
5762 : * because we must know whether we have any postponed SRFs.
5763 : */
5764 388900 : if (parse->hasTargetSRFs &&
5765 124 : expression_returns_set((Node *) expr))
5766 : {
5767 : /* We'll decide below whether these are postponable */
5768 56 : col_is_srf[i] = true;
5769 56 : have_srf = true;
5770 : }
5771 194332 : else if (contain_volatile_functions((Node *) expr))
5772 : {
5773 : /* Unconditionally postpone */
5774 64 : postpone_col[i] = true;
5775 64 : have_volatile = true;
5776 : }
5777 : else
5778 : {
5779 : /*
5780 : * Else check the cost. XXX it's annoying to have to do this
5781 : * when set_pathtarget_cost_width() just did it. Refactor to
5782 : * allow sharing the work?
5783 : */
5784 : QualCost cost;
5785 :
5786 194268 : cost_qual_eval_node(&cost, (Node *) expr, root);
5787 :
5788 : /*
5789 : * We arbitrarily define "expensive" as "more than 10X
5790 : * cpu_operator_cost". Note this will take in any PL function
5791 : * with default cost.
5792 : */
5793 194268 : if (cost.per_tuple > 10 * cpu_operator_cost)
5794 : {
5795 17896 : postpone_col[i] = true;
5796 17896 : have_expensive = true;
5797 : }
5798 : }
5799 : }
5800 : else
5801 : {
5802 : /* For sortgroupref cols, just check if any contain SRFs */
5803 86368 : if (!have_srf_sortcols &&
5804 43376 : parse->hasTargetSRFs &&
5805 192 : expression_returns_set((Node *) expr))
5806 76 : have_srf_sortcols = true;
5807 : }
5808 :
5809 237572 : i++;
5810 : }
5811 :
5812 : /*
5813 : * We can postpone SRFs if we have some but none are in sortgroupref cols.
5814 : */
5815 34554 : postpone_srfs = (have_srf && !have_srf_sortcols);
5816 :
5817 : /*
5818 : * If we don't need a post-sort projection, just return final_target.
5819 : */
5820 69008 : if (!(postpone_srfs || have_volatile ||
5821 45492 : (have_expensive &&
5822 22068 : (parse->limitCount || root->tuple_fraction > 0))))
5823 34446 : return final_target;
5824 :
5825 : /*
5826 : * Report whether the post-sort projection will contain set-returning
5827 : * functions. This is important because it affects whether the Sort can
5828 : * rely on the query's LIMIT (if any) to bound the number of rows it needs
5829 : * to return.
5830 : */
5831 108 : *have_postponed_srfs = postpone_srfs;
5832 :
5833 : /*
5834 : * Construct the sort-input target, taking all non-postponable columns and
5835 : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
5836 : * the postponable ones.
5837 : */
5838 108 : input_target = create_empty_pathtarget();
5839 108 : postponable_cols = NIL;
5840 :
5841 108 : i = 0;
5842 684 : foreach(lc, final_target->exprs)
5843 : {
5844 576 : Expr *expr = (Expr *) lfirst(lc);
5845 :
5846 576 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
5847 112 : postponable_cols = lappend(postponable_cols, expr);
5848 : else
5849 928 : add_column_to_pathtarget(input_target, expr,
5850 928 : get_pathtarget_sortgroupref(final_target, i));
5851 :
5852 576 : i++;
5853 : }
5854 :
5855 : /*
5856 : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
5857 : * postponable columns, and add them to the sort-input target if not
5858 : * already present. (Some might be there already.) We mustn't
5859 : * deconstruct Aggrefs or WindowFuncs here, since the projection node
5860 : * would be unable to recompute them.
5861 : */
5862 108 : postponable_vars = pull_var_clause((Node *) postponable_cols,
5863 : PVC_INCLUDE_AGGREGATES |
5864 : PVC_INCLUDE_WINDOWFUNCS |
5865 : PVC_INCLUDE_PLACEHOLDERS);
5866 108 : add_new_columns_to_pathtarget(input_target, postponable_vars);
5867 :
5868 : /* clean up cruft */
5869 108 : list_free(postponable_vars);
5870 108 : list_free(postponable_cols);
5871 :
5872 : /* XXX this represents even more redundant cost calculation ... */
5873 108 : return set_pathtarget_cost_width(root, input_target);
5874 : }
5875 :
5876 : /*
5877 : * get_cheapest_fractional_path
5878 : * Find the cheapest path for retrieving a specified fraction of all
5879 : * the tuples expected to be returned by the given relation.
5880 : *
5881 : * We interpret tuple_fraction the same way as grouping_planner.
5882 : *
5883 : * We assume set_cheapest() has been run on the given rel.
5884 : */
5885 : Path *
5886 287460 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
5887 : {
5888 287460 : Path *best_path = rel->cheapest_total_path;
5889 : ListCell *l;
5890 :
5891 : /* If all tuples will be retrieved, just return the cheapest-total path */
5892 287460 : if (tuple_fraction <= 0.0)
5893 283646 : return best_path;
5894 :
5895 : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
5896 3814 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
5897 1562 : tuple_fraction /= best_path->rows;
5898 :
5899 9304 : foreach(l, rel->pathlist)
5900 : {
5901 5490 : Path *path = (Path *) lfirst(l);
5902 :
5903 7166 : if (path == rel->cheapest_total_path ||
5904 1676 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
5905 5024 : continue;
5906 :
5907 466 : best_path = path;
5908 : }
5909 :
5910 3814 : return best_path;
5911 : }
5912 :
5913 : /*
5914 : * adjust_paths_for_srfs
5915 : * Fix up the Paths of the given upperrel to handle tSRFs properly.
5916 : *
5917 : * The executor can only handle set-returning functions that appear at the
5918 : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
5919 : * that are not at top level, we need to split up the evaluation into multiple
5920 : * plan levels in which each level satisfies this constraint. This function
5921 : * modifies each Path of an upperrel that (might) compute any SRFs in its
5922 : * output tlist to insert appropriate projection steps.
5923 : *
5924 : * The given targets and targets_contain_srfs lists are from
5925 : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
5926 : * target in targets.
5927 : */
5928 : static void
5929 3484 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
5930 : List *targets, List *targets_contain_srfs)
5931 : {
5932 : ListCell *lc;
5933 :
5934 : Assert(list_length(targets) == list_length(targets_contain_srfs));
5935 : Assert(!linitial_int(targets_contain_srfs));
5936 :
5937 : /* If no SRFs appear at this plan level, nothing to do */
5938 3484 : if (list_length(targets) == 1)
5939 208 : return;
5940 :
5941 : /*
5942 : * Stack SRF-evaluation nodes atop each path for the rel.
5943 : *
5944 : * In principle we should re-run set_cheapest() here to identify the
5945 : * cheapest path, but it seems unlikely that adding the same tlist eval
5946 : * costs to all the paths would change that, so we don't bother. Instead,
5947 : * just assume that the cheapest-startup and cheapest-total paths remain
5948 : * so. (There should be no parameterized paths anymore, so we needn't
5949 : * worry about updating cheapest_parameterized_paths.)
5950 : */
5951 6576 : foreach(lc, rel->pathlist)
5952 : {
5953 3300 : Path *subpath = (Path *) lfirst(lc);
5954 3300 : Path *newpath = subpath;
5955 : ListCell *lc1,
5956 : *lc2;
5957 :
5958 : Assert(subpath->param_info == NULL);
5959 10572 : forboth(lc1, targets, lc2, targets_contain_srfs)
5960 : {
5961 7272 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5962 7272 : bool contains_srfs = (bool) lfirst_int(lc2);
5963 :
5964 : /* If this level doesn't contain SRFs, do regular projection */
5965 7272 : if (contains_srfs)
5966 3340 : newpath = (Path *) create_set_projection_path(root,
5967 : rel,
5968 : newpath,
5969 : thistarget);
5970 : else
5971 3932 : newpath = (Path *) apply_projection_to_path(root,
5972 : rel,
5973 : newpath,
5974 : thistarget);
5975 : }
5976 3300 : lfirst(lc) = newpath;
5977 3300 : if (subpath == rel->cheapest_startup_path)
5978 52 : rel->cheapest_startup_path = newpath;
5979 3300 : if (subpath == rel->cheapest_total_path)
5980 52 : rel->cheapest_total_path = newpath;
5981 : }
5982 :
5983 : /* Likewise for partial paths, if any */
5984 3280 : foreach(lc, rel->partial_pathlist)
5985 : {
5986 4 : Path *subpath = (Path *) lfirst(lc);
5987 4 : Path *newpath = subpath;
5988 : ListCell *lc1,
5989 : *lc2;
5990 :
5991 : Assert(subpath->param_info == NULL);
5992 16 : forboth(lc1, targets, lc2, targets_contain_srfs)
5993 : {
5994 12 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
5995 12 : bool contains_srfs = (bool) lfirst_int(lc2);
5996 :
5997 : /* If this level doesn't contain SRFs, do regular projection */
5998 12 : if (contains_srfs)
5999 4 : newpath = (Path *) create_set_projection_path(root,
6000 : rel,
6001 : newpath,
6002 : thistarget);
6003 : else
6004 : {
6005 : /* avoid apply_projection_to_path, in case of multiple refs */
6006 8 : newpath = (Path *) create_projection_path(root,
6007 : rel,
6008 : newpath,
6009 : thistarget);
6010 : }
6011 : }
6012 4 : lfirst(lc) = newpath;
6013 : }
6014 : }
6015 :
6016 : /*
6017 : * expression_planner
6018 : * Perform planner's transformations on a standalone expression.
6019 : *
6020 : * Various utility commands need to evaluate expressions that are not part
6021 : * of a plannable query. They can do so using the executor's regular
6022 : * expression-execution machinery, but first the expression has to be fed
6023 : * through here to transform it from parser output to something executable.
6024 : *
6025 : * Currently, we disallow sublinks in standalone expressions, so there's no
6026 : * real "planning" involved here. (That might not always be true though.)
6027 : * What we must do is run eval_const_expressions to ensure that any function
6028 : * calls are converted to positional notation and function default arguments
6029 : * get inserted. The fact that constant subexpressions get simplified is a
6030 : * side-effect that is useful when the expression will get evaluated more than
6031 : * once. Also, we must fix operator function IDs.
6032 : *
6033 : * This does not return any information about dependencies of the expression.
6034 : * Hence callers should use the results only for the duration of the current
6035 : * query. Callers that would like to cache the results for longer should use
6036 : * expression_planner_with_deps, probably via the plancache.
6037 : *
6038 : * Note: this must not make any damaging changes to the passed-in expression
6039 : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6040 : * we first do an expression_tree_mutator-based walk, what is returned will
6041 : * be a new node tree.) The result is constructed in the current memory
6042 : * context; beware that this can leak a lot of additional stuff there, too.
6043 : */
6044 : Expr *
6045 83904 : expression_planner(Expr *expr)
6046 : {
6047 : Node *result;
6048 :
6049 : /*
6050 : * Convert named-argument function calls, insert default arguments and
6051 : * simplify constant subexprs
6052 : */
6053 83904 : result = eval_const_expressions(NULL, (Node *) expr);
6054 :
6055 : /* Fill in opfuncid values if missing */
6056 83904 : fix_opfuncids(result);
6057 :
6058 83904 : return (Expr *) result;
6059 : }
6060 :
6061 : /*
6062 : * expression_planner_with_deps
6063 : * Perform planner's transformations on a standalone expression,
6064 : * returning expression dependency information along with the result.
6065 : *
6066 : * This is identical to expression_planner() except that it also returns
6067 : * information about possible dependencies of the expression, ie identities of
6068 : * objects whose definitions affect the result. As in a PlannedStmt, these
6069 : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6070 : */
6071 : Expr *
6072 324 : expression_planner_with_deps(Expr *expr,
6073 : List **relationOids,
6074 : List **invalItems)
6075 : {
6076 : Node *result;
6077 : PlannerGlobal glob;
6078 : PlannerInfo root;
6079 :
6080 : /* Make up dummy planner state so we can use setrefs machinery */
6081 324 : MemSet(&glob, 0, sizeof(glob));
6082 324 : glob.type = T_PlannerGlobal;
6083 324 : glob.relationOids = NIL;
6084 324 : glob.invalItems = NIL;
6085 :
6086 324 : MemSet(&root, 0, sizeof(root));
6087 324 : root.type = T_PlannerInfo;
6088 324 : root.glob = &glob;
6089 :
6090 : /*
6091 : * Convert named-argument function calls, insert default arguments and
6092 : * simplify constant subexprs. Collect identities of inlined functions
6093 : * and elided domains, too.
6094 : */
6095 324 : result = eval_const_expressions(&root, (Node *) expr);
6096 :
6097 : /* Fill in opfuncid values if missing */
6098 324 : fix_opfuncids(result);
6099 :
6100 : /*
6101 : * Now walk the finished expression to find anything else we ought to
6102 : * record as an expression dependency.
6103 : */
6104 324 : (void) extract_query_dependencies_walker(result, &root);
6105 :
6106 324 : *relationOids = glob.relationOids;
6107 324 : *invalItems = glob.invalItems;
6108 :
6109 324 : return (Expr *) result;
6110 : }
6111 :
6112 :
6113 : /*
6114 : * plan_cluster_use_sort
6115 : * Use the planner to decide how CLUSTER should implement sorting
6116 : *
6117 : * tableOid is the OID of a table to be clustered on its index indexOid
6118 : * (which is already known to be a btree index). Decide whether it's
6119 : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6120 : * Return true to use sorting, false to use an indexscan.
6121 : *
6122 : * Note: caller had better already hold some type of lock on the table.
6123 : */
6124 : bool
6125 50 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6126 : {
6127 : PlannerInfo *root;
6128 : Query *query;
6129 : PlannerGlobal *glob;
6130 : RangeTblEntry *rte;
6131 : RelOptInfo *rel;
6132 : IndexOptInfo *indexInfo;
6133 : QualCost indexExprCost;
6134 : Cost comparisonCost;
6135 : Path *seqScanPath;
6136 : Path seqScanAndSortPath;
6137 : IndexPath *indexScanPath;
6138 : ListCell *lc;
6139 :
6140 : /* We can short-circuit the cost comparison if indexscans are disabled */
6141 50 : if (!enable_indexscan)
6142 4 : return true; /* use sort */
6143 :
6144 : /* Set up mostly-dummy planner state */
6145 46 : query = makeNode(Query);
6146 46 : query->commandType = CMD_SELECT;
6147 :
6148 46 : glob = makeNode(PlannerGlobal);
6149 :
6150 46 : root = makeNode(PlannerInfo);
6151 46 : root->parse = query;
6152 46 : root->glob = glob;
6153 46 : root->query_level = 1;
6154 46 : root->planner_cxt = CurrentMemoryContext;
6155 46 : root->wt_param_id = -1;
6156 :
6157 : /* Build a minimal RTE for the rel */
6158 46 : rte = makeNode(RangeTblEntry);
6159 46 : rte->rtekind = RTE_RELATION;
6160 46 : rte->relid = tableOid;
6161 46 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6162 46 : rte->rellockmode = AccessShareLock;
6163 46 : rte->lateral = false;
6164 46 : rte->inh = false;
6165 46 : rte->inFromCl = true;
6166 46 : query->rtable = list_make1(rte);
6167 :
6168 : /* Set up RTE/RelOptInfo arrays */
6169 46 : setup_simple_rel_arrays(root);
6170 :
6171 : /* Build RelOptInfo */
6172 46 : rel = build_simple_rel(root, 1, NULL);
6173 :
6174 : /* Locate IndexOptInfo for the target index */
6175 46 : indexInfo = NULL;
6176 74 : foreach(lc, rel->indexlist)
6177 : {
6178 74 : indexInfo = lfirst_node(IndexOptInfo, lc);
6179 74 : if (indexInfo->indexoid == indexOid)
6180 46 : break;
6181 : }
6182 :
6183 : /*
6184 : * It's possible that get_relation_info did not generate an IndexOptInfo
6185 : * for the desired index; this could happen if it's not yet reached its
6186 : * indcheckxmin usability horizon, or if it's a system index and we're
6187 : * ignoring system indexes. In such cases we should tell CLUSTER to not
6188 : * trust the index contents but use seqscan-and-sort.
6189 : */
6190 46 : if (lc == NULL) /* not in the list? */
6191 0 : return true; /* use sort */
6192 :
6193 : /*
6194 : * Rather than doing all the pushups that would be needed to use
6195 : * set_baserel_size_estimates, just do a quick hack for rows and width.
6196 : */
6197 46 : rel->rows = rel->tuples;
6198 46 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6199 :
6200 46 : root->total_table_pages = rel->pages;
6201 :
6202 : /*
6203 : * Determine eval cost of the index expressions, if any. We need to
6204 : * charge twice that amount for each tuple comparison that happens during
6205 : * the sort, since tuplesort.c will have to re-evaluate the index
6206 : * expressions each time. (XXX that's pretty inefficient...)
6207 : */
6208 46 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6209 46 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6210 :
6211 : /* Estimate the cost of seq scan + sort */
6212 46 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6213 92 : cost_sort(&seqScanAndSortPath, root, NIL,
6214 46 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6215 : comparisonCost, maintenance_work_mem, -1.0);
6216 :
6217 : /* Estimate the cost of index scan */
6218 46 : indexScanPath = create_index_path(root, indexInfo,
6219 : NIL, NIL, NIL, NIL,
6220 : ForwardScanDirection, false,
6221 : NULL, 1.0, false);
6222 :
6223 46 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6224 : }
6225 :
6226 : /*
6227 : * plan_create_index_workers
6228 : * Use the planner to decide how many parallel worker processes
6229 : * CREATE INDEX should request for use
6230 : *
6231 : * tableOid is the table on which the index is to be built. indexOid is the
6232 : * OID of an index to be created or reindexed (which must be a btree index).
6233 : *
6234 : * Return value is the number of parallel worker processes to request. It
6235 : * may be unsafe to proceed if this is 0. Note that this does not include the
6236 : * leader participating as a worker (value is always a number of parallel
6237 : * worker processes).
6238 : *
6239 : * Note: caller had better already hold some type of lock on the table and
6240 : * index.
6241 : */
6242 : int
6243 22404 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6244 : {
6245 : PlannerInfo *root;
6246 : Query *query;
6247 : PlannerGlobal *glob;
6248 : RangeTblEntry *rte;
6249 : Relation heap;
6250 : Relation index;
6251 : RelOptInfo *rel;
6252 : int parallel_workers;
6253 : BlockNumber heap_blocks;
6254 : double reltuples;
6255 : double allvisfrac;
6256 :
6257 : /* Return immediately when parallelism disabled */
6258 22404 : if (max_parallel_maintenance_workers == 0)
6259 0 : return 0;
6260 :
6261 : /* Set up largely-dummy planner state */
6262 22404 : query = makeNode(Query);
6263 22404 : query->commandType = CMD_SELECT;
6264 :
6265 22404 : glob = makeNode(PlannerGlobal);
6266 :
6267 22404 : root = makeNode(PlannerInfo);
6268 22404 : root->parse = query;
6269 22404 : root->glob = glob;
6270 22404 : root->query_level = 1;
6271 22404 : root->planner_cxt = CurrentMemoryContext;
6272 22404 : root->wt_param_id = -1;
6273 :
6274 : /*
6275 : * Build a minimal RTE.
6276 : *
6277 : * Mark the RTE with inh = true. This is a kludge to prevent
6278 : * get_relation_info() from fetching index info, which is necessary
6279 : * because it does not expect that any IndexOptInfo is currently
6280 : * undergoing REINDEX.
6281 : */
6282 22404 : rte = makeNode(RangeTblEntry);
6283 22404 : rte->rtekind = RTE_RELATION;
6284 22404 : rte->relid = tableOid;
6285 22404 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
6286 22404 : rte->rellockmode = AccessShareLock;
6287 22404 : rte->lateral = false;
6288 22404 : rte->inh = true;
6289 22404 : rte->inFromCl = true;
6290 22404 : query->rtable = list_make1(rte);
6291 :
6292 : /* Set up RTE/RelOptInfo arrays */
6293 22404 : setup_simple_rel_arrays(root);
6294 :
6295 : /* Build RelOptInfo */
6296 22404 : rel = build_simple_rel(root, 1, NULL);
6297 :
6298 : /* Rels are assumed already locked by the caller */
6299 22404 : heap = table_open(tableOid, NoLock);
6300 22404 : index = index_open(indexOid, NoLock);
6301 :
6302 : /*
6303 : * Determine if it's safe to proceed.
6304 : *
6305 : * Currently, parallel workers can't access the leader's temporary tables.
6306 : * Furthermore, any index predicate or index expressions must be parallel
6307 : * safe.
6308 : */
6309 43292 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
6310 41760 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
6311 20872 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
6312 : {
6313 1532 : parallel_workers = 0;
6314 1532 : goto done;
6315 : }
6316 :
6317 : /*
6318 : * If parallel_workers storage parameter is set for the table, accept that
6319 : * as the number of parallel worker processes to launch (though still cap
6320 : * at max_parallel_maintenance_workers). Note that we deliberately do not
6321 : * consider any other factor when parallel_workers is set. (e.g., memory
6322 : * use by workers.)
6323 : */
6324 20872 : if (rel->rel_parallel_workers != -1)
6325 : {
6326 14 : parallel_workers = Min(rel->rel_parallel_workers,
6327 : max_parallel_maintenance_workers);
6328 14 : goto done;
6329 : }
6330 :
6331 : /*
6332 : * Estimate heap relation size ourselves, since rel->pages cannot be
6333 : * trusted (heap RTE was marked as inheritance parent)
6334 : */
6335 20858 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
6336 :
6337 : /*
6338 : * Determine number of workers to scan the heap relation using generic
6339 : * model
6340 : */
6341 20858 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
6342 : max_parallel_maintenance_workers);
6343 :
6344 : /*
6345 : * Cap workers based on available maintenance_work_mem as needed.
6346 : *
6347 : * Note that each tuplesort participant receives an even share of the
6348 : * total maintenance_work_mem budget. Aim to leave participants
6349 : * (including the leader as a participant) with no less than 32MB of
6350 : * memory. This leaves cases where maintenance_work_mem is set to 64MB
6351 : * immediately past the threshold of being capable of launching a single
6352 : * parallel worker to sort.
6353 : */
6354 41948 : while (parallel_workers > 0 &&
6355 158 : maintenance_work_mem / (parallel_workers + 1) < 32768L)
6356 74 : parallel_workers--;
6357 :
6358 : done:
6359 22404 : index_close(index, NoLock);
6360 22404 : table_close(heap, NoLock);
6361 :
6362 22404 : return parallel_workers;
6363 : }
6364 :
6365 : /*
6366 : * add_paths_to_grouping_rel
6367 : *
6368 : * Add non-partial paths to grouping relation.
6369 : */
6370 : static void
6371 24156 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
6372 : RelOptInfo *grouped_rel,
6373 : RelOptInfo *partially_grouped_rel,
6374 : const AggClauseCosts *agg_costs,
6375 : grouping_sets_data *gd, double dNumGroups,
6376 : GroupPathExtraData *extra)
6377 : {
6378 24156 : Query *parse = root->parse;
6379 24156 : Path *cheapest_path = input_rel->cheapest_total_path;
6380 : ListCell *lc;
6381 24156 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6382 24156 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6383 24156 : List *havingQual = (List *) extra->havingQual;
6384 24156 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6385 :
6386 24156 : if (can_sort)
6387 : {
6388 : /*
6389 : * Use any available suitably-sorted path as input, and also consider
6390 : * sorting the cheapest-total path.
6391 : */
6392 49608 : foreach(lc, input_rel->pathlist)
6393 : {
6394 25456 : Path *path = (Path *) lfirst(lc);
6395 : bool is_sorted;
6396 :
6397 25456 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6398 : path->pathkeys);
6399 25456 : if (path == cheapest_path || is_sorted)
6400 : {
6401 : /* Sort the cheapest-total path if it isn't already sorted */
6402 25346 : if (!is_sorted)
6403 2514 : path = (Path *) create_sort_path(root,
6404 : grouped_rel,
6405 : path,
6406 : root->group_pathkeys,
6407 : -1.0);
6408 :
6409 : /* Now decide what to stick atop it */
6410 25346 : if (parse->groupingSets)
6411 : {
6412 440 : consider_groupingsets_paths(root, grouped_rel,
6413 : path, true, can_hash,
6414 : gd, agg_costs, dNumGroups);
6415 : }
6416 24906 : else if (parse->hasAggs)
6417 : {
6418 : /*
6419 : * We have aggregation, possibly with plain GROUP BY. Make
6420 : * an AggPath.
6421 : */
6422 24632 : add_path(grouped_rel, (Path *)
6423 73896 : create_agg_path(root,
6424 : grouped_rel,
6425 : path,
6426 24632 : grouped_rel->reltarget,
6427 24632 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6428 : AGGSPLIT_SIMPLE,
6429 : parse->groupClause,
6430 : havingQual,
6431 : agg_costs,
6432 : dNumGroups));
6433 : }
6434 274 : else if (parse->groupClause)
6435 : {
6436 : /*
6437 : * We have GROUP BY without aggregation or grouping sets.
6438 : * Make a GroupPath.
6439 : */
6440 274 : add_path(grouped_rel, (Path *)
6441 274 : create_group_path(root,
6442 : grouped_rel,
6443 : path,
6444 : parse->groupClause,
6445 : havingQual,
6446 : dNumGroups));
6447 : }
6448 : else
6449 : {
6450 : /* Other cases should have been handled above */
6451 : Assert(false);
6452 : }
6453 : }
6454 : }
6455 :
6456 : /*
6457 : * Instead of operating directly on the input relation, we can
6458 : * consider finalizing a partially aggregated path.
6459 : */
6460 24152 : if (partially_grouped_rel != NULL)
6461 : {
6462 2288 : foreach(lc, partially_grouped_rel->pathlist)
6463 : {
6464 1372 : Path *path = (Path *) lfirst(lc);
6465 :
6466 : /*
6467 : * Insert a Sort node, if required. But there's no point in
6468 : * sorting anything but the cheapest path.
6469 : */
6470 1372 : if (!pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
6471 : {
6472 344 : if (path != partially_grouped_rel->cheapest_total_path)
6473 64 : continue;
6474 280 : path = (Path *) create_sort_path(root,
6475 : grouped_rel,
6476 : path,
6477 : root->group_pathkeys,
6478 : -1.0);
6479 : }
6480 :
6481 1308 : if (parse->hasAggs)
6482 1202 : add_path(grouped_rel, (Path *)
6483 3606 : create_agg_path(root,
6484 : grouped_rel,
6485 : path,
6486 1202 : grouped_rel->reltarget,
6487 1202 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6488 : AGGSPLIT_FINAL_DESERIAL,
6489 : parse->groupClause,
6490 : havingQual,
6491 : agg_final_costs,
6492 : dNumGroups));
6493 : else
6494 106 : add_path(grouped_rel, (Path *)
6495 106 : create_group_path(root,
6496 : grouped_rel,
6497 : path,
6498 : parse->groupClause,
6499 : havingQual,
6500 : dNumGroups));
6501 : }
6502 : }
6503 : }
6504 :
6505 24156 : if (can_hash)
6506 : {
6507 : double hashaggtablesize;
6508 :
6509 2420 : if (parse->groupingSets)
6510 : {
6511 : /*
6512 : * Try for a hash-only groupingsets path over unsorted input.
6513 : */
6514 364 : consider_groupingsets_paths(root, grouped_rel,
6515 : cheapest_path, false, true,
6516 : gd, agg_costs, dNumGroups);
6517 : }
6518 : else
6519 : {
6520 2056 : hashaggtablesize = estimate_hashagg_tablesize(cheapest_path,
6521 : agg_costs,
6522 : dNumGroups);
6523 :
6524 : /*
6525 : * Provided that the estimated size of the hashtable does not
6526 : * exceed work_mem, we'll generate a HashAgg Path, although if we
6527 : * were unable to sort above, then we'd better generate a Path, so
6528 : * that we at least have one.
6529 : */
6530 2056 : if (hashaggtablesize < work_mem * 1024L ||
6531 0 : grouped_rel->pathlist == NIL)
6532 : {
6533 : /*
6534 : * We just need an Agg over the cheapest-total input path,
6535 : * since input order won't matter.
6536 : */
6537 2056 : add_path(grouped_rel, (Path *)
6538 4112 : create_agg_path(root, grouped_rel,
6539 : cheapest_path,
6540 2056 : grouped_rel->reltarget,
6541 : AGG_HASHED,
6542 : AGGSPLIT_SIMPLE,
6543 : parse->groupClause,
6544 : havingQual,
6545 : agg_costs,
6546 : dNumGroups));
6547 : }
6548 : }
6549 :
6550 : /*
6551 : * Generate a Finalize HashAgg Path atop of the cheapest partially
6552 : * grouped path, assuming there is one. Once again, we'll only do this
6553 : * if it looks as though the hash table won't exceed work_mem.
6554 : */
6555 2420 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
6556 : {
6557 458 : Path *path = partially_grouped_rel->cheapest_total_path;
6558 :
6559 458 : hashaggtablesize = estimate_hashagg_tablesize(path,
6560 : agg_final_costs,
6561 : dNumGroups);
6562 :
6563 458 : if (hashaggtablesize < work_mem * 1024L)
6564 458 : add_path(grouped_rel, (Path *)
6565 916 : create_agg_path(root,
6566 : grouped_rel,
6567 : path,
6568 458 : grouped_rel->reltarget,
6569 : AGG_HASHED,
6570 : AGGSPLIT_FINAL_DESERIAL,
6571 : parse->groupClause,
6572 : havingQual,
6573 : agg_final_costs,
6574 : dNumGroups));
6575 : }
6576 : }
6577 :
6578 : /*
6579 : * When partitionwise aggregate is used, we might have fully aggregated
6580 : * paths in the partial pathlist, because add_paths_to_append_rel() will
6581 : * consider a path for grouped_rel consisting of a Parallel Append of
6582 : * non-partial paths from each child.
6583 : */
6584 24156 : if (grouped_rel->partial_pathlist != NIL)
6585 108 : gather_grouping_paths(root, grouped_rel);
6586 24156 : }
6587 :
6588 : /*
6589 : * create_partial_grouping_paths
6590 : *
6591 : * Create a new upper relation representing the result of partial aggregation
6592 : * and populate it with appropriate paths. Note that we don't finalize the
6593 : * lists of paths here, so the caller can add additional partial or non-partial
6594 : * paths and must afterward call gather_grouping_paths and set_cheapest on
6595 : * the returned upper relation.
6596 : *
6597 : * All paths for this new upper relation -- both partial and non-partial --
6598 : * have been partially aggregated but require a subsequent FinalizeAggregate
6599 : * step.
6600 : *
6601 : * NB: This function is allowed to return NULL if it determines that there is
6602 : * no real need to create a new RelOptInfo.
6603 : */
6604 : static RelOptInfo *
6605 8976 : create_partial_grouping_paths(PlannerInfo *root,
6606 : RelOptInfo *grouped_rel,
6607 : RelOptInfo *input_rel,
6608 : grouping_sets_data *gd,
6609 : GroupPathExtraData *extra,
6610 : bool force_rel_creation)
6611 : {
6612 8976 : Query *parse = root->parse;
6613 : RelOptInfo *partially_grouped_rel;
6614 8976 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
6615 8976 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
6616 8976 : Path *cheapest_partial_path = NULL;
6617 8976 : Path *cheapest_total_path = NULL;
6618 8976 : double dNumPartialGroups = 0;
6619 8976 : double dNumPartialPartialGroups = 0;
6620 : ListCell *lc;
6621 8976 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
6622 8976 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
6623 :
6624 : /*
6625 : * Consider whether we should generate partially aggregated non-partial
6626 : * paths. We can only do this if we have a non-partial path, and only if
6627 : * the parent of the input rel is performing partial partitionwise
6628 : * aggregation. (Note that extra->patype is the type of partitionwise
6629 : * aggregation being used at the parent level, not this level.)
6630 : */
6631 17952 : if (input_rel->pathlist != NIL &&
6632 8976 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
6633 374 : cheapest_total_path = input_rel->cheapest_total_path;
6634 :
6635 : /*
6636 : * If parallelism is possible for grouped_rel, then we should consider
6637 : * generating partially-grouped partial paths. However, if the input rel
6638 : * has no partial paths, then we can't.
6639 : */
6640 8976 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
6641 1058 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
6642 :
6643 : /*
6644 : * If we can't partially aggregate partial paths, and we can't partially
6645 : * aggregate non-partial paths, then don't bother creating the new
6646 : * RelOptInfo at all, unless the caller specified force_rel_creation.
6647 : */
6648 8976 : if (cheapest_total_path == NULL &&
6649 7744 : cheapest_partial_path == NULL &&
6650 7744 : !force_rel_creation)
6651 7686 : return NULL;
6652 :
6653 : /*
6654 : * Build a new upper relation to represent the result of partially
6655 : * aggregating the rows from the input relation.
6656 : */
6657 1290 : partially_grouped_rel = fetch_upper_rel(root,
6658 : UPPERREL_PARTIAL_GROUP_AGG,
6659 : grouped_rel->relids);
6660 1290 : partially_grouped_rel->consider_parallel =
6661 1290 : grouped_rel->consider_parallel;
6662 1290 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
6663 1290 : partially_grouped_rel->serverid = grouped_rel->serverid;
6664 1290 : partially_grouped_rel->userid = grouped_rel->userid;
6665 1290 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
6666 1290 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
6667 :
6668 : /*
6669 : * Build target list for partial aggregate paths. These paths cannot just
6670 : * emit the same tlist as regular aggregate paths, because (1) we must
6671 : * include Vars and Aggrefs needed in HAVING, which might not appear in
6672 : * the result tlist, and (2) the Aggrefs must be set in partial mode.
6673 : */
6674 1290 : partially_grouped_rel->reltarget =
6675 1290 : make_partial_grouping_target(root, grouped_rel->reltarget,
6676 : extra->havingQual);
6677 :
6678 1290 : if (!extra->partial_costs_set)
6679 : {
6680 : /*
6681 : * Collect statistics about aggregates for estimating costs of
6682 : * performing aggregation in parallel.
6683 : */
6684 724 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
6685 724 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
6686 724 : if (parse->hasAggs)
6687 : {
6688 : List *partial_target_exprs;
6689 :
6690 : /* partial phase */
6691 668 : partial_target_exprs = partially_grouped_rel->reltarget->exprs;
6692 668 : get_agg_clause_costs(root, (Node *) partial_target_exprs,
6693 : AGGSPLIT_INITIAL_SERIAL,
6694 : agg_partial_costs);
6695 :
6696 : /* final phase */
6697 668 : get_agg_clause_costs(root, (Node *) grouped_rel->reltarget->exprs,
6698 : AGGSPLIT_FINAL_DESERIAL,
6699 : agg_final_costs);
6700 668 : get_agg_clause_costs(root, extra->havingQual,
6701 : AGGSPLIT_FINAL_DESERIAL,
6702 : agg_final_costs);
6703 : }
6704 :
6705 724 : extra->partial_costs_set = true;
6706 : }
6707 :
6708 : /* Estimate number of partial groups. */
6709 1290 : if (cheapest_total_path != NULL)
6710 374 : dNumPartialGroups =
6711 374 : get_number_of_groups(root,
6712 : cheapest_total_path->rows,
6713 : gd,
6714 : extra->targetList);
6715 1290 : if (cheapest_partial_path != NULL)
6716 1058 : dNumPartialPartialGroups =
6717 1058 : get_number_of_groups(root,
6718 : cheapest_partial_path->rows,
6719 : gd,
6720 : extra->targetList);
6721 :
6722 1290 : if (can_sort && cheapest_total_path != NULL)
6723 : {
6724 : /* This should have been checked previously */
6725 : Assert(parse->hasAggs || parse->groupClause);
6726 :
6727 : /*
6728 : * Use any available suitably-sorted path as input, and also consider
6729 : * sorting the cheapest partial path.
6730 : */
6731 754 : foreach(lc, input_rel->pathlist)
6732 : {
6733 380 : Path *path = (Path *) lfirst(lc);
6734 : bool is_sorted;
6735 :
6736 380 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6737 : path->pathkeys);
6738 380 : if (path == cheapest_total_path || is_sorted)
6739 : {
6740 : /* Sort the cheapest partial path, if it isn't already */
6741 380 : if (!is_sorted)
6742 374 : path = (Path *) create_sort_path(root,
6743 : partially_grouped_rel,
6744 : path,
6745 : root->group_pathkeys,
6746 : -1.0);
6747 :
6748 380 : if (parse->hasAggs)
6749 356 : add_path(partially_grouped_rel, (Path *)
6750 1068 : create_agg_path(root,
6751 : partially_grouped_rel,
6752 : path,
6753 356 : partially_grouped_rel->reltarget,
6754 356 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6755 : AGGSPLIT_INITIAL_SERIAL,
6756 : parse->groupClause,
6757 : NIL,
6758 : agg_partial_costs,
6759 : dNumPartialGroups));
6760 : else
6761 24 : add_path(partially_grouped_rel, (Path *)
6762 24 : create_group_path(root,
6763 : partially_grouped_rel,
6764 : path,
6765 : parse->groupClause,
6766 : NIL,
6767 : dNumPartialGroups));
6768 : }
6769 : }
6770 : }
6771 :
6772 1290 : if (can_sort && cheapest_partial_path != NULL)
6773 : {
6774 : /* Similar to above logic, but for partial paths. */
6775 2116 : foreach(lc, input_rel->partial_pathlist)
6776 : {
6777 1058 : Path *path = (Path *) lfirst(lc);
6778 : bool is_sorted;
6779 :
6780 1058 : is_sorted = pathkeys_contained_in(root->group_pathkeys,
6781 : path->pathkeys);
6782 1058 : if (path == cheapest_partial_path || is_sorted)
6783 : {
6784 : /* Sort the cheapest partial path, if it isn't already */
6785 1058 : if (!is_sorted)
6786 600 : path = (Path *) create_sort_path(root,
6787 : partially_grouped_rel,
6788 : path,
6789 : root->group_pathkeys,
6790 : -1.0);
6791 :
6792 1058 : if (parse->hasAggs)
6793 1010 : add_partial_path(partially_grouped_rel, (Path *)
6794 3030 : create_agg_path(root,
6795 : partially_grouped_rel,
6796 : path,
6797 1010 : partially_grouped_rel->reltarget,
6798 1010 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
6799 : AGGSPLIT_INITIAL_SERIAL,
6800 : parse->groupClause,
6801 : NIL,
6802 : agg_partial_costs,
6803 : dNumPartialPartialGroups));
6804 : else
6805 48 : add_partial_path(partially_grouped_rel, (Path *)
6806 48 : create_group_path(root,
6807 : partially_grouped_rel,
6808 : path,
6809 : parse->groupClause,
6810 : NIL,
6811 : dNumPartialPartialGroups));
6812 : }
6813 : }
6814 : }
6815 :
6816 1290 : if (can_hash && cheapest_total_path != NULL)
6817 : {
6818 : double hashaggtablesize;
6819 :
6820 : /* Checked above */
6821 : Assert(parse->hasAggs || parse->groupClause);
6822 :
6823 374 : hashaggtablesize =
6824 : estimate_hashagg_tablesize(cheapest_total_path,
6825 : agg_partial_costs,
6826 : dNumPartialGroups);
6827 :
6828 : /*
6829 : * Tentatively produce a partial HashAgg Path, depending on if it
6830 : * looks as if the hash table will fit in work_mem.
6831 : */
6832 374 : if (hashaggtablesize < work_mem * 1024L &&
6833 : cheapest_total_path != NULL)
6834 : {
6835 374 : add_path(partially_grouped_rel, (Path *)
6836 748 : create_agg_path(root,
6837 : partially_grouped_rel,
6838 : cheapest_total_path,
6839 374 : partially_grouped_rel->reltarget,
6840 : AGG_HASHED,
6841 : AGGSPLIT_INITIAL_SERIAL,
6842 : parse->groupClause,
6843 : NIL,
6844 : agg_partial_costs,
6845 : dNumPartialGroups));
6846 : }
6847 : }
6848 :
6849 1290 : if (can_hash && cheapest_partial_path != NULL)
6850 : {
6851 : double hashaggtablesize;
6852 :
6853 600 : hashaggtablesize =
6854 : estimate_hashagg_tablesize(cheapest_partial_path,
6855 : agg_partial_costs,
6856 : dNumPartialPartialGroups);
6857 :
6858 : /* Do the same for partial paths. */
6859 600 : if (hashaggtablesize < work_mem * 1024L &&
6860 : cheapest_partial_path != NULL)
6861 : {
6862 600 : add_partial_path(partially_grouped_rel, (Path *)
6863 1200 : create_agg_path(root,
6864 : partially_grouped_rel,
6865 : cheapest_partial_path,
6866 600 : partially_grouped_rel->reltarget,
6867 : AGG_HASHED,
6868 : AGGSPLIT_INITIAL_SERIAL,
6869 : parse->groupClause,
6870 : NIL,
6871 : agg_partial_costs,
6872 : dNumPartialPartialGroups));
6873 : }
6874 : }
6875 :
6876 : /*
6877 : * If there is an FDW that's responsible for all baserels of the query,
6878 : * let it consider adding partially grouped ForeignPaths.
6879 : */
6880 1296 : if (partially_grouped_rel->fdwroutine &&
6881 6 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
6882 : {
6883 6 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
6884 :
6885 6 : fdwroutine->GetForeignUpperPaths(root,
6886 : UPPERREL_PARTIAL_GROUP_AGG,
6887 : input_rel, partially_grouped_rel,
6888 : extra);
6889 : }
6890 :
6891 1290 : return partially_grouped_rel;
6892 : }
6893 :
6894 : /*
6895 : * Generate Gather and Gather Merge paths for a grouping relation or partial
6896 : * grouping relation.
6897 : *
6898 : * generate_gather_paths does most of the work, but we also consider a special
6899 : * case: we could try sorting the data by the group_pathkeys and then applying
6900 : * Gather Merge.
6901 : *
6902 : * NB: This function shouldn't be used for anything other than a grouped or
6903 : * partially grouped relation not only because of the fact that it explicitly
6904 : * references group_pathkeys but we pass "true" as the third argument to
6905 : * generate_gather_paths().
6906 : */
6907 : static void
6908 966 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
6909 : {
6910 : Path *cheapest_partial_path;
6911 :
6912 : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
6913 966 : generate_gather_paths(root, rel, true);
6914 :
6915 : /* Try cheapest partial path + explicit Sort + Gather Merge. */
6916 966 : cheapest_partial_path = linitial(rel->partial_pathlist);
6917 966 : if (!pathkeys_contained_in(root->group_pathkeys,
6918 : cheapest_partial_path->pathkeys))
6919 : {
6920 : Path *path;
6921 : double total_groups;
6922 :
6923 470 : total_groups =
6924 470 : cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
6925 470 : path = (Path *) create_sort_path(root, rel, cheapest_partial_path,
6926 : root->group_pathkeys,
6927 : -1.0);
6928 470 : path = (Path *)
6929 470 : create_gather_merge_path(root,
6930 : rel,
6931 : path,
6932 470 : rel->reltarget,
6933 : root->group_pathkeys,
6934 : NULL,
6935 : &total_groups);
6936 :
6937 470 : add_path(rel, path);
6938 : }
6939 966 : }
6940 :
6941 : /*
6942 : * can_partial_agg
6943 : *
6944 : * Determines whether or not partial grouping and/or aggregation is possible.
6945 : * Returns true when possible, false otherwise.
6946 : */
6947 : static bool
6948 23612 : can_partial_agg(PlannerInfo *root, const AggClauseCosts *agg_costs)
6949 : {
6950 23612 : Query *parse = root->parse;
6951 :
6952 23612 : if (!parse->hasAggs && parse->groupClause == NIL)
6953 : {
6954 : /*
6955 : * We don't know how to do parallel aggregation unless we have either
6956 : * some aggregates or a grouping clause.
6957 : */
6958 0 : return false;
6959 : }
6960 23612 : else if (parse->groupingSets)
6961 : {
6962 : /* We don't know how to do grouping sets in parallel. */
6963 416 : return false;
6964 : }
6965 23196 : else if (agg_costs->hasNonPartial || agg_costs->hasNonSerial)
6966 : {
6967 : /* Insufficient support for partial mode. */
6968 15090 : return false;
6969 : }
6970 :
6971 : /* Everything looks good. */
6972 8106 : return true;
6973 : }
6974 :
6975 : /*
6976 : * apply_scanjoin_target_to_paths
6977 : *
6978 : * Adjust the final scan/join relation, and recursively all of its children,
6979 : * to generate the final scan/join target. It would be more correct to model
6980 : * this as a separate planning step with a new RelOptInfo at the toplevel and
6981 : * for each child relation, but doing it this way is noticeably cheaper.
6982 : * Maybe that problem can be solved at some point, but for now we do this.
6983 : *
6984 : * If tlist_same_exprs is true, then the scan/join target to be applied has
6985 : * the same expressions as the existing reltarget, so we need only insert the
6986 : * appropriate sortgroupref information. By avoiding the creation of
6987 : * projection paths we save effort both immediately and at plan creation time.
6988 : */
6989 : static void
6990 305826 : apply_scanjoin_target_to_paths(PlannerInfo *root,
6991 : RelOptInfo *rel,
6992 : List *scanjoin_targets,
6993 : List *scanjoin_targets_contain_srfs,
6994 : bool scanjoin_target_parallel_safe,
6995 : bool tlist_same_exprs)
6996 : {
6997 305826 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
6998 : PathTarget *scanjoin_target;
6999 : ListCell *lc;
7000 :
7001 : /* This recurses, so be paranoid. */
7002 305826 : check_stack_depth();
7003 :
7004 : /*
7005 : * If the rel is partitioned, we want to drop its existing paths and
7006 : * generate new ones. This function would still be correct if we kept the
7007 : * existing paths: we'd modify them to generate the correct target above
7008 : * the partitioning Append, and then they'd compete on cost with paths
7009 : * generating the target below the Append. However, in our current cost
7010 : * model the latter way is always the same or cheaper cost, so modifying
7011 : * the existing paths would just be useless work. Moreover, when the cost
7012 : * is the same, varying roundoff errors might sometimes allow an existing
7013 : * path to be picked, resulting in undesirable cross-platform plan
7014 : * variations. So we drop old paths and thereby force the work to be done
7015 : * below the Append, except in the case of a non-parallel-safe target.
7016 : *
7017 : * Some care is needed, because we have to allow generate_gather_paths to
7018 : * see the old partial paths in the next stanza. Hence, zap the main
7019 : * pathlist here, then allow generate_gather_paths to add path(s) to the
7020 : * main list, and finally zap the partial pathlist.
7021 : */
7022 305826 : if (rel_is_partitioned)
7023 4444 : rel->pathlist = NIL;
7024 :
7025 : /*
7026 : * If the scan/join target is not parallel-safe, partial paths cannot
7027 : * generate it.
7028 : */
7029 305826 : if (!scanjoin_target_parallel_safe)
7030 : {
7031 : /*
7032 : * Since we can't generate the final scan/join target in parallel
7033 : * workers, this is our last opportunity to use any partial paths that
7034 : * exist; so build Gather path(s) that use them and emit whatever the
7035 : * current reltarget is. We don't do this in the case where the
7036 : * target is parallel-safe, since we will be able to generate superior
7037 : * paths by doing it after the final scan/join target has been
7038 : * applied.
7039 : */
7040 68300 : generate_gather_paths(root, rel, false);
7041 :
7042 : /* Can't use parallel query above this level. */
7043 68300 : rel->partial_pathlist = NIL;
7044 68300 : rel->consider_parallel = false;
7045 : }
7046 :
7047 : /* Finish dropping old paths for a partitioned rel, per comment above */
7048 305826 : if (rel_is_partitioned)
7049 4444 : rel->partial_pathlist = NIL;
7050 :
7051 : /* Extract SRF-free scan/join target. */
7052 305826 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7053 :
7054 : /*
7055 : * Apply the SRF-free scan/join target to each existing path.
7056 : *
7057 : * If the tlist exprs are the same, we can just inject the sortgroupref
7058 : * information into the existing pathtargets. Otherwise, replace each
7059 : * path with a projection path that generates the SRF-free scan/join
7060 : * target. This can't change the ordering of paths within rel->pathlist,
7061 : * so we just modify the list in place.
7062 : */
7063 622632 : foreach(lc, rel->pathlist)
7064 : {
7065 316806 : Path *subpath = (Path *) lfirst(lc);
7066 :
7067 : /* Shouldn't have any parameterized paths anymore */
7068 : Assert(subpath->param_info == NULL);
7069 :
7070 316806 : if (tlist_same_exprs)
7071 201428 : subpath->pathtarget->sortgrouprefs =
7072 100714 : scanjoin_target->sortgrouprefs;
7073 : else
7074 : {
7075 : Path *newpath;
7076 :
7077 216092 : newpath = (Path *) create_projection_path(root, rel, subpath,
7078 : scanjoin_target);
7079 216092 : lfirst(lc) = newpath;
7080 : }
7081 : }
7082 :
7083 : /* Likewise adjust the targets for any partial paths. */
7084 315616 : foreach(lc, rel->partial_pathlist)
7085 : {
7086 9790 : Path *subpath = (Path *) lfirst(lc);
7087 :
7088 : /* Shouldn't have any parameterized paths anymore */
7089 : Assert(subpath->param_info == NULL);
7090 :
7091 9790 : if (tlist_same_exprs)
7092 15756 : subpath->pathtarget->sortgrouprefs =
7093 7878 : scanjoin_target->sortgrouprefs;
7094 : else
7095 : {
7096 : Path *newpath;
7097 :
7098 1912 : newpath = (Path *) create_projection_path(root, rel, subpath,
7099 : scanjoin_target);
7100 1912 : lfirst(lc) = newpath;
7101 : }
7102 : }
7103 :
7104 : /*
7105 : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
7106 : * atop each existing path. (Note that this function doesn't look at the
7107 : * cheapest-path fields, which is a good thing because they're bogus right
7108 : * now.)
7109 : */
7110 305826 : if (root->parse->hasTargetSRFs)
7111 3276 : adjust_paths_for_srfs(root, rel,
7112 : scanjoin_targets,
7113 : scanjoin_targets_contain_srfs);
7114 :
7115 : /*
7116 : * Update the rel's target to be the final (with SRFs) scan/join target.
7117 : * This now matches the actual output of all the paths, and we might get
7118 : * confused in createplan.c if they don't agree. We must do this now so
7119 : * that any append paths made in the next part will use the correct
7120 : * pathtarget (cf. create_append_path).
7121 : *
7122 : * Note that this is also necessary if GetForeignUpperPaths() gets called
7123 : * on the final scan/join relation or on any of its children, since the
7124 : * FDW might look at the rel's target to create ForeignPaths.
7125 : */
7126 305826 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
7127 :
7128 : /*
7129 : * If the relation is partitioned, recursively apply the scan/join target
7130 : * to all partitions, and generate brand-new Append paths in which the
7131 : * scan/join target is computed below the Append rather than above it.
7132 : * Since Append is not projection-capable, that might save a separate
7133 : * Result node, and it also is important for partitionwise aggregate.
7134 : */
7135 305826 : if (rel_is_partitioned)
7136 : {
7137 4444 : List *live_children = NIL;
7138 : int partition_idx;
7139 :
7140 : /* Adjust each partition. */
7141 18014 : for (partition_idx = 0; partition_idx < rel->nparts; partition_idx++)
7142 : {
7143 13570 : RelOptInfo *child_rel = rel->part_rels[partition_idx];
7144 : AppendRelInfo **appinfos;
7145 : int nappinfos;
7146 13570 : List *child_scanjoin_targets = NIL;
7147 : ListCell *lc;
7148 :
7149 : /* Pruned or dummy children can be ignored. */
7150 13570 : if (child_rel == NULL || IS_DUMMY_REL(child_rel))
7151 3582 : continue;
7152 :
7153 : /* Translate scan/join targets for this child. */
7154 9988 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
7155 : &nappinfos);
7156 19976 : foreach(lc, scanjoin_targets)
7157 : {
7158 9988 : PathTarget *target = lfirst_node(PathTarget, lc);
7159 :
7160 9988 : target = copy_pathtarget(target);
7161 9988 : target->exprs = (List *)
7162 19976 : adjust_appendrel_attrs(root,
7163 9988 : (Node *) target->exprs,
7164 : nappinfos, appinfos);
7165 9988 : child_scanjoin_targets = lappend(child_scanjoin_targets,
7166 : target);
7167 : }
7168 9988 : pfree(appinfos);
7169 :
7170 : /* Recursion does the real work. */
7171 9988 : apply_scanjoin_target_to_paths(root, child_rel,
7172 : child_scanjoin_targets,
7173 : scanjoin_targets_contain_srfs,
7174 : scanjoin_target_parallel_safe,
7175 : tlist_same_exprs);
7176 :
7177 : /* Save non-dummy children for Append paths. */
7178 9988 : if (!IS_DUMMY_REL(child_rel))
7179 9988 : live_children = lappend(live_children, child_rel);
7180 : }
7181 :
7182 : /* Build new paths for this relation by appending child paths. */
7183 4444 : add_paths_to_append_rel(root, rel, live_children);
7184 : }
7185 :
7186 : /*
7187 : * Consider generating Gather or Gather Merge paths. We must only do this
7188 : * if the relation is parallel safe, and we don't do it for child rels to
7189 : * avoid creating multiple Gather nodes within the same plan. We must do
7190 : * this after all paths have been generated and before set_cheapest, since
7191 : * one of the generated paths may turn out to be the cheapest one.
7192 : */
7193 305826 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
7194 71122 : generate_gather_paths(root, rel, false);
7195 :
7196 : /*
7197 : * Reassess which paths are the cheapest, now that we've potentially added
7198 : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
7199 : * this relation.
7200 : */
7201 305826 : set_cheapest(rel);
7202 305826 : }
7203 :
7204 : /*
7205 : * create_partitionwise_grouping_paths
7206 : *
7207 : * If the partition keys of input relation are part of the GROUP BY clause, all
7208 : * the rows belonging to a given group come from a single partition. This
7209 : * allows aggregation/grouping over a partitioned relation to be broken down
7210 : * into aggregation/grouping on each partition. This should be no worse, and
7211 : * often better, than the normal approach.
7212 : *
7213 : * However, if the GROUP BY clause does not contain all the partition keys,
7214 : * rows from a given group may be spread across multiple partitions. In that
7215 : * case, we perform partial aggregation for each group, append the results,
7216 : * and then finalize aggregation. This is less certain to win than the
7217 : * previous case. It may win if the PartialAggregate stage greatly reduces
7218 : * the number of groups, because fewer rows will pass through the Append node.
7219 : * It may lose if we have lots of small groups.
7220 : */
7221 : static void
7222 338 : create_partitionwise_grouping_paths(PlannerInfo *root,
7223 : RelOptInfo *input_rel,
7224 : RelOptInfo *grouped_rel,
7225 : RelOptInfo *partially_grouped_rel,
7226 : const AggClauseCosts *agg_costs,
7227 : grouping_sets_data *gd,
7228 : PartitionwiseAggregateType patype,
7229 : GroupPathExtraData *extra)
7230 : {
7231 338 : int nparts = input_rel->nparts;
7232 : int cnt_parts;
7233 338 : List *grouped_live_children = NIL;
7234 338 : List *partially_grouped_live_children = NIL;
7235 338 : PathTarget *target = grouped_rel->reltarget;
7236 338 : bool partial_grouping_valid = true;
7237 :
7238 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
7239 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
7240 : partially_grouped_rel != NULL);
7241 :
7242 : /* Add paths for partitionwise aggregation/grouping. */
7243 1256 : for (cnt_parts = 0; cnt_parts < nparts; cnt_parts++)
7244 : {
7245 918 : RelOptInfo *child_input_rel = input_rel->part_rels[cnt_parts];
7246 918 : PathTarget *child_target = copy_pathtarget(target);
7247 : AppendRelInfo **appinfos;
7248 : int nappinfos;
7249 : GroupPathExtraData child_extra;
7250 : RelOptInfo *child_grouped_rel;
7251 : RelOptInfo *child_partially_grouped_rel;
7252 :
7253 : /* Pruned or dummy children can be ignored. */
7254 918 : if (child_input_rel == NULL || IS_DUMMY_REL(child_input_rel))
7255 0 : continue;
7256 :
7257 : /*
7258 : * Copy the given "extra" structure as is and then override the
7259 : * members specific to this child.
7260 : */
7261 918 : memcpy(&child_extra, extra, sizeof(child_extra));
7262 :
7263 918 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
7264 : &nappinfos);
7265 :
7266 918 : child_target->exprs = (List *)
7267 1836 : adjust_appendrel_attrs(root,
7268 918 : (Node *) target->exprs,
7269 : nappinfos, appinfos);
7270 :
7271 : /* Translate havingQual and targetList. */
7272 918 : child_extra.havingQual = (Node *)
7273 : adjust_appendrel_attrs(root,
7274 : extra->havingQual,
7275 : nappinfos, appinfos);
7276 918 : child_extra.targetList = (List *)
7277 1836 : adjust_appendrel_attrs(root,
7278 918 : (Node *) extra->targetList,
7279 : nappinfos, appinfos);
7280 :
7281 : /*
7282 : * extra->patype was the value computed for our parent rel; patype is
7283 : * the value for this relation. For the child, our value is its
7284 : * parent rel's value.
7285 : */
7286 918 : child_extra.patype = patype;
7287 :
7288 : /*
7289 : * Create grouping relation to hold fully aggregated grouping and/or
7290 : * aggregation paths for the child.
7291 : */
7292 1836 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
7293 : child_target,
7294 918 : extra->target_parallel_safe,
7295 : child_extra.havingQual);
7296 :
7297 : /* Create grouping paths for this child relation. */
7298 918 : create_ordinary_grouping_paths(root, child_input_rel,
7299 : child_grouped_rel,
7300 : agg_costs, gd, &child_extra,
7301 : &child_partially_grouped_rel);
7302 :
7303 918 : if (child_partially_grouped_rel)
7304 : {
7305 566 : partially_grouped_live_children =
7306 566 : lappend(partially_grouped_live_children,
7307 : child_partially_grouped_rel);
7308 : }
7309 : else
7310 352 : partial_grouping_valid = false;
7311 :
7312 918 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7313 : {
7314 544 : set_cheapest(child_grouped_rel);
7315 544 : grouped_live_children = lappend(grouped_live_children,
7316 : child_grouped_rel);
7317 : }
7318 :
7319 918 : pfree(appinfos);
7320 : }
7321 :
7322 : /*
7323 : * Try to create append paths for partially grouped children. For full
7324 : * partitionwise aggregation, we might have paths in the partial_pathlist
7325 : * if parallel aggregation is possible. For partial partitionwise
7326 : * aggregation, we may have paths in both pathlist and partial_pathlist.
7327 : *
7328 : * NB: We must have a partially grouped path for every child in order to
7329 : * generate a partially grouped path for this relation.
7330 : */
7331 338 : if (partially_grouped_rel && partial_grouping_valid)
7332 : {
7333 : Assert(partially_grouped_live_children != NIL);
7334 :
7335 218 : add_paths_to_append_rel(root, partially_grouped_rel,
7336 : partially_grouped_live_children);
7337 :
7338 : /*
7339 : * We need call set_cheapest, since the finalization step will use the
7340 : * cheapest path from the rel.
7341 : */
7342 218 : if (partially_grouped_rel->pathlist)
7343 218 : set_cheapest(partially_grouped_rel);
7344 : }
7345 :
7346 : /* If possible, create append paths for fully grouped children. */
7347 338 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
7348 : {
7349 : Assert(grouped_live_children != NIL);
7350 :
7351 192 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
7352 : }
7353 338 : }
7354 :
7355 : /*
7356 : * group_by_has_partkey
7357 : *
7358 : * Returns true, if all the partition keys of the given relation are part of
7359 : * the GROUP BY clauses, false otherwise.
7360 : */
7361 : static bool
7362 334 : group_by_has_partkey(RelOptInfo *input_rel,
7363 : List *targetList,
7364 : List *groupClause)
7365 : {
7366 334 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
7367 334 : int cnt = 0;
7368 : int partnatts;
7369 :
7370 : /* Input relation should be partitioned. */
7371 : Assert(input_rel->part_scheme);
7372 :
7373 : /* Rule out early, if there are no partition keys present. */
7374 334 : if (!input_rel->partexprs)
7375 0 : return false;
7376 :
7377 334 : partnatts = input_rel->part_scheme->partnatts;
7378 :
7379 550 : for (cnt = 0; cnt < partnatts; cnt++)
7380 : {
7381 358 : List *partexprs = input_rel->partexprs[cnt];
7382 : ListCell *lc;
7383 358 : bool found = false;
7384 :
7385 504 : foreach(lc, partexprs)
7386 : {
7387 362 : Expr *partexpr = lfirst(lc);
7388 :
7389 362 : if (list_member(groupexprs, partexpr))
7390 : {
7391 216 : found = true;
7392 216 : break;
7393 : }
7394 : }
7395 :
7396 : /*
7397 : * If none of the partition key expressions match with any of the
7398 : * GROUP BY expression, return false.
7399 : */
7400 358 : if (!found)
7401 142 : return false;
7402 : }
7403 :
7404 192 : return true;
7405 : }
|
