Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * planner.c
4 : : * The query optimizer external interface.
5 : : *
6 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/optimizer/plan/planner.c
12 : : *
13 : : *-------------------------------------------------------------------------
14 : : */
15 : :
16 : : #include "postgres.h"
17 : :
18 : : #include <limits.h>
19 : : #include <math.h>
20 : :
21 : : #include "access/genam.h"
22 : : #include "access/parallel.h"
23 : : #include "access/sysattr.h"
24 : : #include "access/table.h"
25 : : #include "catalog/pg_aggregate.h"
26 : : #include "catalog/pg_inherits.h"
27 : : #include "catalog/pg_proc.h"
28 : : #include "catalog/pg_type.h"
29 : : #include "executor/executor.h"
30 : : #include "foreign/fdwapi.h"
31 : : #include "jit/jit.h"
32 : : #include "lib/bipartite_match.h"
33 : : #include "lib/knapsack.h"
34 : : #include "miscadmin.h"
35 : : #include "nodes/makefuncs.h"
36 : : #include "nodes/nodeFuncs.h"
37 : : #ifdef OPTIMIZER_DEBUG
38 : : #include "nodes/print.h"
39 : : #endif
40 : : #include "nodes/supportnodes.h"
41 : : #include "optimizer/appendinfo.h"
42 : : #include "optimizer/clauses.h"
43 : : #include "optimizer/cost.h"
44 : : #include "optimizer/optimizer.h"
45 : : #include "optimizer/paramassign.h"
46 : : #include "optimizer/pathnode.h"
47 : : #include "optimizer/paths.h"
48 : : #include "optimizer/plancat.h"
49 : : #include "optimizer/planmain.h"
50 : : #include "optimizer/planner.h"
51 : : #include "optimizer/prep.h"
52 : : #include "optimizer/subselect.h"
53 : : #include "optimizer/tlist.h"
54 : : #include "parser/analyze.h"
55 : : #include "parser/parse_agg.h"
56 : : #include "parser/parse_clause.h"
57 : : #include "parser/parse_relation.h"
58 : : #include "parser/parsetree.h"
59 : : #include "partitioning/partdesc.h"
60 : : #include "rewrite/rewriteManip.h"
61 : : #include "utils/acl.h"
62 : : #include "utils/backend_status.h"
63 : : #include "utils/lsyscache.h"
64 : : #include "utils/rel.h"
65 : : #include "utils/selfuncs.h"
66 : :
67 : : /* GUC parameters */
68 : : double cursor_tuple_fraction = DEFAULT_CURSOR_TUPLE_FRACTION;
69 : : int debug_parallel_query = DEBUG_PARALLEL_OFF;
70 : : bool parallel_leader_participation = true;
71 : : bool enable_distinct_reordering = true;
72 : :
73 : : /* Hook for plugins to get control in planner() */
74 : : planner_hook_type planner_hook = NULL;
75 : :
76 : : /* Hook for plugins to get control after PlannerGlobal is initialized */
77 : : planner_setup_hook_type planner_setup_hook = NULL;
78 : :
79 : : /* Hook for plugins to get control before PlannerGlobal is discarded */
80 : : planner_shutdown_hook_type planner_shutdown_hook = NULL;
81 : :
82 : : /* Hook for plugins to get control when grouping_planner() plans upper rels */
83 : : create_upper_paths_hook_type create_upper_paths_hook = NULL;
84 : :
85 : :
86 : : /* Expression kind codes for preprocess_expression */
87 : : #define EXPRKIND_QUAL 0
88 : : #define EXPRKIND_TARGET 1
89 : : #define EXPRKIND_RTFUNC 2
90 : : #define EXPRKIND_RTFUNC_LATERAL 3
91 : : #define EXPRKIND_VALUES 4
92 : : #define EXPRKIND_VALUES_LATERAL 5
93 : : #define EXPRKIND_LIMIT 6
94 : : #define EXPRKIND_APPINFO 7
95 : : #define EXPRKIND_PHV 8
96 : : #define EXPRKIND_TABLESAMPLE 9
97 : : #define EXPRKIND_ARBITER_ELEM 10
98 : : #define EXPRKIND_TABLEFUNC 11
99 : : #define EXPRKIND_TABLEFUNC_LATERAL 12
100 : : #define EXPRKIND_GROUPEXPR 13
101 : :
102 : : /*
103 : : * Data specific to grouping sets
104 : : */
105 : : typedef struct
106 : : {
107 : : List *rollups;
108 : : List *hash_sets_idx;
109 : : double dNumHashGroups;
110 : : bool any_hashable;
111 : : Bitmapset *unsortable_refs;
112 : : Bitmapset *unhashable_refs;
113 : : List *unsortable_sets;
114 : : int *tleref_to_colnum_map;
115 : : } grouping_sets_data;
116 : :
117 : : /*
118 : : * Temporary structure for use during WindowClause reordering in order to be
119 : : * able to sort WindowClauses on partitioning/ordering prefix.
120 : : */
121 : : typedef struct
122 : : {
123 : : WindowClause *wc;
124 : : List *uniqueOrder; /* A List of unique ordering/partitioning
125 : : * clauses per Window */
126 : : } WindowClauseSortData;
127 : :
128 : : /* Passthrough data for standard_qp_callback */
129 : : typedef struct
130 : : {
131 : : List *activeWindows; /* active windows, if any */
132 : : grouping_sets_data *gset_data; /* grouping sets data, if any */
133 : : SetOperationStmt *setop; /* parent set operation or NULL if not a
134 : : * subquery belonging to a set operation */
135 : : } standard_qp_extra;
136 : :
137 : : /*
138 : : * Context for find_having_conflicts. This is the callback context passed to
139 : : * expression_has_grouping_conflict in clauses.c.
140 : : */
141 : : typedef struct
142 : : {
143 : : Query *parse;
144 : : Index group_rtindex;
145 : : } having_grouping_ctx;
146 : :
147 : : /* Local functions */
148 : : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
149 : : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
150 : : static Bitmapset *find_having_conflicts(Query *parse, Index group_rtindex);
151 : : static Oid having_var_grouping_eqop(Var *var, void *context);
152 : : static Oid group_var_eqop(Query *parse, Var *var);
153 : : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
154 : : SetOperationStmt *setops);
155 : : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
156 : : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
157 : : int *tleref_to_colnum_map);
158 : : static void preprocess_rowmarks(PlannerInfo *root);
159 : : static double preprocess_limit(PlannerInfo *root,
160 : : double tuple_fraction,
161 : : int64 *offset_est, int64 *count_est);
162 : : static List *preprocess_groupclause(PlannerInfo *root, List *force);
163 : : static List *extract_rollup_sets(List *groupingSets);
164 : : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
165 : : static void standard_qp_callback(PlannerInfo *root, void *extra);
166 : : static double get_number_of_groups(PlannerInfo *root,
167 : : double path_rows,
168 : : grouping_sets_data *gd,
169 : : List *target_list);
170 : : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
171 : : RelOptInfo *input_rel,
172 : : PathTarget *target,
173 : : bool target_parallel_safe,
174 : : grouping_sets_data *gd);
175 : : static bool is_degenerate_grouping(PlannerInfo *root);
176 : : static void create_degenerate_grouping_paths(PlannerInfo *root,
177 : : RelOptInfo *input_rel,
178 : : RelOptInfo *grouped_rel);
179 : : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
180 : : PathTarget *target, bool target_parallel_safe,
181 : : Node *havingQual);
182 : : static void create_ordinary_grouping_paths(PlannerInfo *root,
183 : : RelOptInfo *input_rel,
184 : : RelOptInfo *grouped_rel,
185 : : const AggClauseCosts *agg_costs,
186 : : grouping_sets_data *gd,
187 : : GroupPathExtraData *extra,
188 : : RelOptInfo **partially_grouped_rel_p);
189 : : static void consider_groupingsets_paths(PlannerInfo *root,
190 : : RelOptInfo *grouped_rel,
191 : : Path *path,
192 : : bool is_sorted,
193 : : bool can_hash,
194 : : grouping_sets_data *gd,
195 : : const AggClauseCosts *agg_costs,
196 : : double dNumGroups);
197 : : static RelOptInfo *create_window_paths(PlannerInfo *root,
198 : : RelOptInfo *input_rel,
199 : : PathTarget *input_target,
200 : : PathTarget *output_target,
201 : : bool output_target_parallel_safe,
202 : : WindowFuncLists *wflists,
203 : : List *activeWindows);
204 : : static void create_one_window_path(PlannerInfo *root,
205 : : RelOptInfo *window_rel,
206 : : Path *path,
207 : : PathTarget *input_target,
208 : : PathTarget *output_target,
209 : : WindowFuncLists *wflists,
210 : : List *activeWindows);
211 : : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
212 : : RelOptInfo *input_rel,
213 : : PathTarget *target);
214 : : static void create_partial_distinct_paths(PlannerInfo *root,
215 : : RelOptInfo *input_rel,
216 : : RelOptInfo *final_distinct_rel,
217 : : PathTarget *target);
218 : : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
219 : : RelOptInfo *input_rel,
220 : : RelOptInfo *distinct_rel);
221 : : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
222 : : List *needed_pathkeys,
223 : : List *path_pathkeys);
224 : : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
225 : : RelOptInfo *input_rel,
226 : : PathTarget *target,
227 : : bool target_parallel_safe,
228 : : double limit_tuples);
229 : : static PathTarget *make_group_input_target(PlannerInfo *root,
230 : : PathTarget *final_target);
231 : : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
232 : : PathTarget *grouping_target,
233 : : Node *havingQual);
234 : : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
235 : : static void optimize_window_clauses(PlannerInfo *root,
236 : : WindowFuncLists *wflists);
237 : : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
238 : : static void name_active_windows(List *activeWindows);
239 : : static PathTarget *make_window_input_target(PlannerInfo *root,
240 : : PathTarget *final_target,
241 : : List *activeWindows);
242 : : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
243 : : List *tlist);
244 : : static PathTarget *make_sort_input_target(PlannerInfo *root,
245 : : PathTarget *final_target,
246 : : bool *have_postponed_srfs);
247 : : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
248 : : List *targets, List *targets_contain_srfs);
249 : : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
250 : : RelOptInfo *grouped_rel,
251 : : RelOptInfo *partially_grouped_rel,
252 : : const AggClauseCosts *agg_costs,
253 : : grouping_sets_data *gd,
254 : : GroupPathExtraData *extra);
255 : : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
256 : : RelOptInfo *grouped_rel,
257 : : RelOptInfo *input_rel,
258 : : grouping_sets_data *gd,
259 : : GroupPathExtraData *extra,
260 : : bool force_rel_creation);
261 : : static Path *make_ordered_path(PlannerInfo *root,
262 : : RelOptInfo *rel,
263 : : Path *path,
264 : : Path *cheapest_path,
265 : : List *pathkeys,
266 : : double limit_tuples);
267 : : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
268 : : static bool can_partial_agg(PlannerInfo *root);
269 : : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
270 : : RelOptInfo *rel,
271 : : List *scanjoin_targets,
272 : : List *scanjoin_targets_contain_srfs,
273 : : bool scanjoin_target_parallel_safe,
274 : : bool tlist_same_exprs);
275 : : static void create_partitionwise_grouping_paths(PlannerInfo *root,
276 : : RelOptInfo *input_rel,
277 : : RelOptInfo *grouped_rel,
278 : : RelOptInfo *partially_grouped_rel,
279 : : const AggClauseCosts *agg_costs,
280 : : grouping_sets_data *gd,
281 : : PartitionwiseAggregateType patype,
282 : : GroupPathExtraData *extra);
283 : : static bool group_by_has_partkey(RelOptInfo *input_rel,
284 : : List *targetList,
285 : : List *groupClause);
286 : : static int common_prefix_cmp(const void *a, const void *b);
287 : : static List *generate_setop_child_grouplist(SetOperationStmt *op,
288 : : List *targetlist);
289 : : static void create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
290 : : List *sortPathkeys, List *groupClause,
291 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
292 : : static void create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
293 : : List *sortPathkeys, List *groupClause,
294 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
295 : :
296 : :
297 : : /*****************************************************************************
298 : : *
299 : : * Query optimizer entry point
300 : : *
301 : : * Inputs:
302 : : * parse: an analyzed-and-rewritten query tree for an optimizable statement
303 : : * query_string: source text for the query tree (used for error reports)
304 : : * cursorOptions: bitmask of CURSOR_OPT_XXX flags, see parsenodes.h
305 : : * boundParams: passed-in parameter values, or NULL if none
306 : : * es: ExplainState if being called from EXPLAIN, else NULL
307 : : *
308 : : * The result is a PlannedStmt tree.
309 : : *
310 : : * PARAM_EXTERN Param nodes within the parse tree can be replaced by Consts
311 : : * using values from boundParams, if those values are marked PARAM_FLAG_CONST.
312 : : * Parameter values not so marked are still relied on for estimation purposes.
313 : : *
314 : : * The ExplainState pointer is not currently used by the core planner, but it
315 : : * is passed through to some planner hooks so that they can report information
316 : : * back to EXPLAIN extension hooks.
317 : : *
318 : : * To support loadable plugins that monitor or modify planner behavior,
319 : : * we provide a hook variable that lets a plugin get control before and
320 : : * after the standard planning process. The plugin would normally call
321 : : * standard_planner().
322 : : *
323 : : * Note to plugin authors: standard_planner() scribbles on its Query input,
324 : : * so you'd better copy that data structure if you want to plan more than once.
325 : : *
326 : : *****************************************************************************/
327 : : PlannedStmt *
328 : 333217 : planner(Query *parse, const char *query_string, int cursorOptions,
329 : : ParamListInfo boundParams, ExplainState *es)
330 : : {
331 : : PlannedStmt *result;
332 : :
333 [ + + ]: 333217 : if (planner_hook)
334 : 50944 : result = (*planner_hook) (parse, query_string, cursorOptions,
335 : : boundParams, es);
336 : : else
337 : 282273 : result = standard_planner(parse, query_string, cursorOptions,
338 : : boundParams, es);
339 : :
340 : 329141 : pgstat_report_plan_id(result->planId, false);
341 : :
342 : 329141 : return result;
343 : : }
344 : :
345 : : PlannedStmt *
346 : 333217 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
347 : : ParamListInfo boundParams, ExplainState *es)
348 : : {
349 : : PlannedStmt *result;
350 : : PlannerGlobal *glob;
351 : : double tuple_fraction;
352 : : PlannerInfo *root;
353 : : RelOptInfo *final_rel;
354 : : Path *best_path;
355 : : Plan *top_plan;
356 : : ListCell *lp,
357 : : *lr,
358 : : *lc;
359 : :
360 : : /*
361 : : * Set up global state for this planner invocation. This data is needed
362 : : * across all levels of sub-Query that might exist in the given command,
363 : : * so we keep it in a separate struct that's linked to by each per-Query
364 : : * PlannerInfo.
365 : : */
366 : 333217 : glob = makeNode(PlannerGlobal);
367 : :
368 : 333217 : glob->boundParams = boundParams;
369 : 333217 : glob->subplans = NIL;
370 : 333217 : glob->subpaths = NIL;
371 : 333217 : glob->subroots = NIL;
372 : 333217 : glob->rewindPlanIDs = NULL;
373 : 333217 : glob->finalrtable = NIL;
374 : 333217 : glob->allRelids = NULL;
375 : 333217 : glob->prunableRelids = NULL;
376 : 333217 : glob->finalrteperminfos = NIL;
377 : 333217 : glob->finalrowmarks = NIL;
378 : 333217 : glob->resultRelations = NIL;
379 : 333217 : glob->appendRelations = NIL;
380 : 333217 : glob->partPruneInfos = NIL;
381 : 333217 : glob->relationOids = NIL;
382 : 333217 : glob->invalItems = NIL;
383 : 333217 : glob->paramExecTypes = NIL;
384 : 333217 : glob->lastPHId = 0;
385 : 333217 : glob->lastRowMarkId = 0;
386 : 333217 : glob->lastPlanNodeId = 0;
387 : 333217 : glob->transientPlan = false;
388 : 333217 : glob->dependsOnRole = false;
389 : 333217 : glob->partition_directory = NULL;
390 : 333217 : glob->rel_notnullatts_hash = NULL;
391 : :
392 : : /*
393 : : * Assess whether it's feasible to use parallel mode for this query. We
394 : : * can't do this in a standalone backend, or if the command will try to
395 : : * modify any data, or if this is a cursor operation, or if GUCs are set
396 : : * to values that don't permit parallelism, or if parallel-unsafe
397 : : * functions are present in the query tree.
398 : : *
399 : : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
400 : : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
401 : : * the command is writing into a completely new table which workers won't
402 : : * be able to see. If the workers could see the table, the fact that
403 : : * group locking would cause them to ignore the leader's heavyweight GIN
404 : : * page locks would make this unsafe. We'll have to fix that somehow if
405 : : * we want to allow parallel inserts in general; updates and deletes have
406 : : * additional problems especially around combo CIDs.)
407 : : *
408 : : * For now, we don't try to use parallel mode if we're running inside a
409 : : * parallel worker. We might eventually be able to relax this
410 : : * restriction, but for now it seems best not to have parallel workers
411 : : * trying to create their own parallel workers.
412 : : */
413 [ + + + + ]: 333217 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
414 : 314957 : IsUnderPostmaster &&
415 [ + + ]: 314957 : parse->commandType == CMD_SELECT &&
416 [ + + ]: 256034 : !parse->hasModifyingCTE &&
417 [ + + ]: 255910 : max_parallel_workers_per_gather > 0 &&
418 [ + + ]: 255215 : !IsParallelWorker())
419 : : {
420 : : /* all the cheap tests pass, so scan the query tree */
421 : 255175 : glob->maxParallelHazard = max_parallel_hazard(parse);
422 : 255175 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
423 : : }
424 : : else
425 : : {
426 : : /* skip the query tree scan, just assume it's unsafe */
427 : 78042 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
428 : 78042 : glob->parallelModeOK = false;
429 : : }
430 : :
431 : : /*
432 : : * glob->parallelModeNeeded is normally set to false here and changed to
433 : : * true during plan creation if a Gather or Gather Merge plan is actually
434 : : * created (cf. create_gather_plan, create_gather_merge_plan).
435 : : *
436 : : * However, if debug_parallel_query = on or debug_parallel_query =
437 : : * regress, then we impose parallel mode whenever it's safe to do so, even
438 : : * if the final plan doesn't use parallelism. It's not safe to do so if
439 : : * the query contains anything parallel-unsafe; parallelModeOK will be
440 : : * false in that case. Note that parallelModeOK can't change after this
441 : : * point. Otherwise, everything in the query is either parallel-safe or
442 : : * parallel-restricted, and in either case it should be OK to impose
443 : : * parallel-mode restrictions. If that ends up breaking something, then
444 : : * either some function the user included in the query is incorrectly
445 : : * labeled as parallel-safe or parallel-restricted when in reality it's
446 : : * parallel-unsafe, or else the query planner itself has a bug.
447 : : */
448 [ + + ]: 553675 : glob->parallelModeNeeded = glob->parallelModeOK &&
449 [ + + ]: 220458 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
450 : :
451 : : /* Determine what fraction of the plan is likely to be scanned */
452 [ + + ]: 333217 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
453 : : {
454 : : /*
455 : : * We have no real idea how many tuples the user will ultimately FETCH
456 : : * from a cursor, but it is often the case that he doesn't want 'em
457 : : * all, or would prefer a fast-start plan anyway so that he can
458 : : * process some of the tuples sooner. Use a GUC parameter to decide
459 : : * what fraction to optimize for.
460 : : */
461 : 3248 : tuple_fraction = cursor_tuple_fraction;
462 : :
463 : : /*
464 : : * We document cursor_tuple_fraction as simply being a fraction, which
465 : : * means the edge cases 0 and 1 have to be treated specially here. We
466 : : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
467 : : */
468 [ - + ]: 3248 : if (tuple_fraction >= 1.0)
469 : 0 : tuple_fraction = 0.0;
470 [ - + ]: 3248 : else if (tuple_fraction <= 0.0)
471 : 0 : tuple_fraction = 1e-10;
472 : : }
473 : : else
474 : : {
475 : : /* Default assumption is we need all the tuples */
476 : 329969 : tuple_fraction = 0.0;
477 : : }
478 : :
479 : : /*
480 : : * Compute the initial path generation strategy mask.
481 : : *
482 : : * Some strategies, such as PGS_FOREIGNJOIN, have no corresponding enable_*
483 : : * GUC, and so the corresponding bits are always set in the default
484 : : * strategy mask.
485 : : *
486 : : * It may seem surprising that enable_indexscan sets both PGS_INDEXSCAN
487 : : * and PGS_INDEXONLYSCAN. However, the historical behavior of this GUC
488 : : * corresponds to this exactly: enable_indexscan=off disables both
489 : : * index-scan and index-only scan paths, whereas enable_indexonlyscan=off
490 : : * converts the index-only scan paths that we would have considered into
491 : : * index scan paths.
492 : : */
493 : 333217 : glob->default_pgs_mask = PGS_APPEND | PGS_MERGE_APPEND | PGS_FOREIGNJOIN |
494 : : PGS_GATHER | PGS_CONSIDER_NONPARTIAL;
495 [ + - ]: 333217 : if (enable_tidscan)
496 : 333217 : glob->default_pgs_mask |= PGS_TIDSCAN;
497 [ + + ]: 333217 : if (enable_seqscan)
498 : 316071 : glob->default_pgs_mask |= PGS_SEQSCAN;
499 [ + + ]: 333217 : if (enable_indexscan)
500 : 330980 : glob->default_pgs_mask |= PGS_INDEXSCAN | PGS_INDEXONLYSCAN;
501 [ + + ]: 333217 : if (enable_indexonlyscan)
502 : 331790 : glob->default_pgs_mask |= PGS_CONSIDER_INDEXONLY;
503 [ + + ]: 333217 : if (enable_bitmapscan)
504 : 323752 : glob->default_pgs_mask |= PGS_BITMAPSCAN;
505 [ + + ]: 333217 : if (enable_mergejoin)
506 : : {
507 : 331295 : glob->default_pgs_mask |= PGS_MERGEJOIN_PLAIN;
508 [ + + ]: 331295 : if (enable_material)
509 : 331228 : glob->default_pgs_mask |= PGS_MERGEJOIN_MATERIALIZE;
510 : : }
511 [ + + ]: 333217 : if (enable_nestloop)
512 : : {
513 : 332912 : glob->default_pgs_mask |= PGS_NESTLOOP_PLAIN;
514 [ + + ]: 332912 : if (enable_material)
515 : 332619 : glob->default_pgs_mask |= PGS_NESTLOOP_MATERIALIZE;
516 [ + + ]: 332912 : if (enable_memoize)
517 : 332793 : glob->default_pgs_mask |= PGS_NESTLOOP_MEMOIZE;
518 : : }
519 [ + + ]: 333217 : if (enable_hashjoin)
520 : 330957 : glob->default_pgs_mask |= PGS_HASHJOIN;
521 [ + - ]: 333217 : if (enable_gathermerge)
522 : 333217 : glob->default_pgs_mask |= PGS_GATHER_MERGE;
523 [ + + ]: 333217 : if (enable_partitionwise_join)
524 : 2074 : glob->default_pgs_mask |= PGS_CONSIDER_PARTITIONWISE;
525 : :
526 : : /* Allow plugins to take control after we've initialized "glob" */
527 [ + + ]: 333217 : if (planner_setup_hook)
528 : 90046 : (*planner_setup_hook) (glob, parse, query_string, cursorOptions,
529 : : &tuple_fraction, es);
530 : :
531 : : /* primary planning entry point (may recurse for subqueries) */
532 : 332433 : root = subquery_planner(glob, parse, NULL, NULL, NULL, false,
533 : : tuple_fraction, NULL);
534 : :
535 : : /* Select best Path and turn it into a Plan */
536 : 329411 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
537 : 329411 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
538 : :
539 : 329411 : top_plan = create_plan(root, best_path);
540 : :
541 : : /*
542 : : * If creating a plan for a scrollable cursor, make sure it can run
543 : : * backwards on demand. Add a Material node at the top at need.
544 : : */
545 [ + + ]: 329141 : if (cursorOptions & CURSOR_OPT_SCROLL)
546 : : {
547 [ + + ]: 221 : if (!ExecSupportsBackwardScan(top_plan))
548 : 26 : top_plan = materialize_finished_plan(top_plan);
549 : : }
550 : :
551 : : /*
552 : : * Optionally add a Gather node for testing purposes, provided this is
553 : : * actually a safe thing to do.
554 : : *
555 : : * We can add Gather even when top_plan has parallel-safe initPlans, but
556 : : * then we have to move the initPlans to the Gather node because of
557 : : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
558 : : * regression tests when debug_parallel_query = regress, because initPlans
559 : : * that would normally appear on the top_plan move to the Gather, causing
560 : : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
561 : : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
562 : : */
563 [ + + ]: 329141 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
564 [ + + ]: 157 : top_plan->parallel_safe &&
565 [ - + ]: 104 : (top_plan->initPlan == NIL ||
566 [ # # ]: 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
567 : : {
568 : 104 : Gather *gather = makeNode(Gather);
569 : : Cost initplan_cost;
570 : : bool unsafe_initplans;
571 : :
572 : 104 : gather->plan.targetlist = top_plan->targetlist;
573 : 104 : gather->plan.qual = NIL;
574 : 104 : gather->plan.lefttree = top_plan;
575 : 104 : gather->plan.righttree = NULL;
576 : 104 : gather->num_workers = 1;
577 : 104 : gather->single_copy = true;
578 : 104 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
579 : :
580 : : /* Transfer any initPlans to the new top node */
581 : 104 : gather->plan.initPlan = top_plan->initPlan;
582 : 104 : top_plan->initPlan = NIL;
583 : :
584 : : /*
585 : : * Since this Gather has no parallel-aware descendants to signal to,
586 : : * we don't need a rescan Param.
587 : : */
588 : 104 : gather->rescan_param = -1;
589 : :
590 : : /*
591 : : * Ideally we'd use cost_gather here, but setting up dummy path data
592 : : * to satisfy it doesn't seem much cleaner than knowing what it does.
593 : : */
594 : 104 : gather->plan.startup_cost = top_plan->startup_cost +
595 : : parallel_setup_cost;
596 : 104 : gather->plan.total_cost = top_plan->total_cost +
597 : 104 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
598 : 104 : gather->plan.plan_rows = top_plan->plan_rows;
599 : 104 : gather->plan.plan_width = top_plan->plan_width;
600 : 104 : gather->plan.parallel_aware = false;
601 : 104 : gather->plan.parallel_safe = false;
602 : :
603 : : /*
604 : : * Delete the initplans' cost from top_plan. We needn't add it to the
605 : : * Gather node, since the above coding already included it there.
606 : : */
607 : 104 : SS_compute_initplan_cost(gather->plan.initPlan,
608 : : &initplan_cost, &unsafe_initplans);
609 : 104 : top_plan->startup_cost -= initplan_cost;
610 : 104 : top_plan->total_cost -= initplan_cost;
611 : :
612 : : /* use parallel mode for parallel plans. */
613 : 104 : root->glob->parallelModeNeeded = true;
614 : :
615 : 104 : top_plan = &gather->plan;
616 : : }
617 : :
618 : : /*
619 : : * If any Params were generated, run through the plan tree and compute
620 : : * each plan node's extParam/allParam sets. Ideally we'd merge this into
621 : : * set_plan_references' tree traversal, but for now it has to be separate
622 : : * because we need to visit subplans before not after main plan.
623 : : */
624 [ + + ]: 329141 : if (glob->paramExecTypes != NIL)
625 : : {
626 : : Assert(list_length(glob->subplans) == list_length(glob->subroots));
627 [ + + + + : 142872 : forboth(lp, glob->subplans, lr, glob->subroots)
+ + + + +
+ + - +
+ ]
628 : : {
629 : 32304 : Plan *subplan = (Plan *) lfirst(lp);
630 : 32304 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
631 : :
632 : 32304 : SS_finalize_plan(subroot, subplan);
633 : : }
634 : 110568 : SS_finalize_plan(root, top_plan);
635 : : }
636 : :
637 : : /* final cleanup of the plan */
638 : : Assert(glob->finalrtable == NIL);
639 : : Assert(glob->finalrteperminfos == NIL);
640 : : Assert(glob->finalrowmarks == NIL);
641 : : Assert(glob->resultRelations == NIL);
642 : : Assert(glob->appendRelations == NIL);
643 : 329141 : top_plan = set_plan_references(root, top_plan);
644 : : /* ... and the subplans (both regular subplans and initplans) */
645 : : Assert(list_length(glob->subplans) == list_length(glob->subroots));
646 [ + + + + : 361445 : forboth(lp, glob->subplans, lr, glob->subroots)
+ + + + +
+ + - +
+ ]
647 : : {
648 : 32304 : Plan *subplan = (Plan *) lfirst(lp);
649 : 32304 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
650 : :
651 : 32304 : lfirst(lp) = set_plan_references(subroot, subplan);
652 : : }
653 : :
654 : : /* build the PlannedStmt result */
655 : 329141 : result = makeNode(PlannedStmt);
656 : :
657 : 329141 : result->commandType = parse->commandType;
658 : 329141 : result->queryId = parse->queryId;
659 : 329141 : result->planOrigin = PLAN_STMT_STANDARD;
660 : 329141 : result->hasReturning = (parse->returningList != NIL);
661 : 329141 : result->hasModifyingCTE = parse->hasModifyingCTE;
662 : 329141 : result->canSetTag = parse->canSetTag;
663 : 329141 : result->transientPlan = glob->transientPlan;
664 : 329141 : result->dependsOnRole = glob->dependsOnRole;
665 : 329141 : result->parallelModeNeeded = glob->parallelModeNeeded;
666 : 329141 : result->planTree = top_plan;
667 : 329141 : result->partPruneInfos = glob->partPruneInfos;
668 : 329141 : result->rtable = glob->finalrtable;
669 : 658282 : result->unprunableRelids = bms_difference(glob->allRelids,
670 : 329141 : glob->prunableRelids);
671 : 329141 : result->permInfos = glob->finalrteperminfos;
672 : 329141 : result->subrtinfos = glob->subrtinfos;
673 : 329141 : result->appendRelations = glob->appendRelations;
674 : 329141 : result->subplans = glob->subplans;
675 : 329141 : result->rewindPlanIDs = glob->rewindPlanIDs;
676 : 329141 : result->rowMarks = glob->finalrowmarks;
677 : :
678 : : /*
679 : : * Compute resultRelationRelids and rowMarkRelids from resultRelations and
680 : : * rowMarks. These can be used for cheap membership checks.
681 : : */
682 [ + + + + : 395803 : foreach(lc, glob->resultRelations)
+ + ]
683 : 66662 : result->resultRelationRelids = bms_add_member(result->resultRelationRelids,
684 : : lfirst_int(lc));
685 [ + + + + : 339948 : foreach(lc, glob->finalrowmarks)
+ + ]
686 : 10807 : result->rowMarkRelids = bms_add_member(result->rowMarkRelids,
687 : 10807 : ((PlanRowMark *) lfirst(lc))->rti);
688 : :
689 : 329141 : result->relationOids = glob->relationOids;
690 : 329141 : result->invalItems = glob->invalItems;
691 : 329141 : result->paramExecTypes = glob->paramExecTypes;
692 : : /* utilityStmt should be null, but we might as well copy it */
693 : 329141 : result->utilityStmt = parse->utilityStmt;
694 : 329141 : result->elidedNodes = glob->elidedNodes;
695 : 329141 : result->stmt_location = parse->stmt_location;
696 : 329141 : result->stmt_len = parse->stmt_len;
697 : :
698 : 329141 : result->jitFlags = PGJIT_NONE;
699 [ - + - - ]: 329141 : if (jit_enabled && jit_above_cost >= 0 &&
700 [ # # ]: 0 : top_plan->total_cost > jit_above_cost)
701 : : {
702 : 0 : result->jitFlags |= PGJIT_PERFORM;
703 : :
704 : : /*
705 : : * Decide how much effort should be put into generating better code.
706 : : */
707 [ # # ]: 0 : if (jit_optimize_above_cost >= 0 &&
708 [ # # ]: 0 : top_plan->total_cost > jit_optimize_above_cost)
709 : 0 : result->jitFlags |= PGJIT_OPT3;
710 [ # # ]: 0 : if (jit_inline_above_cost >= 0 &&
711 [ # # ]: 0 : top_plan->total_cost > jit_inline_above_cost)
712 : 0 : result->jitFlags |= PGJIT_INLINE;
713 : :
714 : : /*
715 : : * Decide which operations should be JITed.
716 : : */
717 [ # # ]: 0 : if (jit_expressions)
718 : 0 : result->jitFlags |= PGJIT_EXPR;
719 [ # # ]: 0 : if (jit_tuple_deforming)
720 : 0 : result->jitFlags |= PGJIT_DEFORM;
721 : : }
722 : :
723 : : /* Allow plugins to take control before we discard "glob" */
724 [ + + ]: 329141 : if (planner_shutdown_hook)
725 : 88475 : (*planner_shutdown_hook) (glob, parse, query_string, result);
726 : :
727 [ + + ]: 329141 : if (glob->partition_directory != NULL)
728 : 9205 : DestroyPartitionDirectory(glob->partition_directory);
729 : :
730 : 329141 : return result;
731 : : }
732 : :
733 : :
734 : : /*--------------------
735 : : * subquery_planner
736 : : * Invokes the planner on a subquery. We recurse to here for each
737 : : * sub-SELECT found in the query tree.
738 : : *
739 : : * glob is the global state for the current planner run.
740 : : * parse is the querytree produced by the parser & rewriter.
741 : : * plan_name is the name to assign to this subplan (NULL at the top level).
742 : : * parent_root is the immediate parent Query's info (NULL at the top level).
743 : : * alternative_root is a previously created PlannerInfo for which this query
744 : : * level is an alternative implementation, or else NULL.
745 : : * hasRecursion is true if this is a recursive WITH query.
746 : : * tuple_fraction is the fraction of tuples we expect will be retrieved.
747 : : * tuple_fraction is interpreted as explained for grouping_planner, below.
748 : : * setops is used for set operation subqueries to provide the subquery with
749 : : * the context in which it's being used so that Paths correctly sorted for the
750 : : * set operation can be generated. NULL when not planning a set operation
751 : : * child, or when a child of a set op that isn't interested in sorted input.
752 : : *
753 : : * Basically, this routine does the stuff that should only be done once
754 : : * per Query object. It then calls grouping_planner. At one time,
755 : : * grouping_planner could be invoked recursively on the same Query object;
756 : : * that's not currently true, but we keep the separation between the two
757 : : * routines anyway, in case we need it again someday.
758 : : *
759 : : * subquery_planner will be called recursively to handle sub-Query nodes
760 : : * found within the query's expressions and rangetable.
761 : : *
762 : : * Returns the PlannerInfo struct ("root") that contains all data generated
763 : : * while planning the subquery. In particular, the Path(s) attached to
764 : : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
765 : : * cheapest way(s) to implement the query. The top level will select the
766 : : * best Path and pass it through createplan.c to produce a finished Plan.
767 : : *--------------------
768 : : */
769 : : PlannerInfo *
770 : 394546 : subquery_planner(PlannerGlobal *glob, Query *parse, char *plan_name,
771 : : PlannerInfo *parent_root, PlannerInfo *alternative_root,
772 : : bool hasRecursion, double tuple_fraction,
773 : : SetOperationStmt *setops)
774 : : {
775 : : PlannerInfo *root;
776 : : List *newWithCheckOptions;
777 : : List *newHaving;
778 : : Bitmapset *havingPushdownConflicts;
779 : : int havingIdx;
780 : : bool hasOuterJoins;
781 : : bool hasResultRTEs;
782 : : RelOptInfo *final_rel;
783 : : ListCell *l;
784 : :
785 : : /* Create a PlannerInfo data structure for this subquery */
786 : 394546 : root = makeNode(PlannerInfo);
787 : 394546 : root->parse = parse;
788 : 394546 : root->glob = glob;
789 [ + + ]: 394546 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
790 : 394546 : root->plan_name = plan_name;
791 [ + + ]: 394546 : if (alternative_root != NULL)
792 : 1263 : root->alternative_plan_name = alternative_root->plan_name;
793 : : else
794 : 393283 : root->alternative_plan_name = plan_name;
795 : 394546 : root->parent_root = parent_root;
796 : 394546 : root->plan_params = NIL;
797 : 394546 : root->outer_params = NULL;
798 : 394546 : root->planner_cxt = CurrentMemoryContext;
799 : 394546 : root->init_plans = NIL;
800 : 394546 : root->cte_plan_ids = NIL;
801 : 394546 : root->multiexpr_params = NIL;
802 : 394546 : root->join_domains = NIL;
803 : 394546 : root->eq_classes = NIL;
804 : 394546 : root->ec_merging_done = false;
805 : 394546 : root->last_rinfo_serial = 0;
806 : 394546 : root->all_result_relids =
807 [ + + ]: 394546 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
808 : 394546 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
809 : 394546 : root->append_rel_list = NIL;
810 : 394546 : root->row_identity_vars = NIL;
811 : 394546 : root->rowMarks = NIL;
812 : 394546 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
813 : 394546 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
814 : 394546 : root->processed_groupClause = NIL;
815 : 394546 : root->processed_distinctClause = NIL;
816 : 394546 : root->processed_tlist = NIL;
817 : 394546 : root->update_colnos = NIL;
818 : 394546 : root->grouping_map = NULL;
819 : 394546 : root->minmax_aggs = NIL;
820 : 394546 : root->qual_security_level = 0;
821 : 394546 : root->hasPseudoConstantQuals = false;
822 : 394546 : root->hasAlternativeSubPlans = false;
823 : 394546 : root->placeholdersFrozen = false;
824 : 394546 : root->hasRecursion = hasRecursion;
825 : 394546 : root->assumeReplanning = false;
826 [ + + ]: 394546 : if (hasRecursion)
827 : 638 : root->wt_param_id = assign_special_exec_param(root);
828 : : else
829 : 393908 : root->wt_param_id = -1;
830 : 394546 : root->non_recursive_path = NULL;
831 : :
832 : : /*
833 : : * Create the top-level join domain. This won't have valid contents until
834 : : * deconstruct_jointree fills it in, but the node needs to exist before
835 : : * that so we can build EquivalenceClasses referencing it.
836 : : */
837 : 394546 : root->join_domains = list_make1(makeNode(JoinDomain));
838 : :
839 : : /*
840 : : * If there is a WITH list, process each WITH query and either convert it
841 : : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
842 : : */
843 [ + + ]: 394546 : if (parse->cteList)
844 : 2180 : SS_process_ctes(root);
845 : :
846 : : /*
847 : : * If it's a MERGE command, transform the joinlist as appropriate.
848 : : */
849 : 394542 : transform_MERGE_to_join(parse);
850 : :
851 : : /*
852 : : * Reject FOR PORTION OF on a generated column. We can't write to a
853 : : * virtual generated column, and a stored generated column should be
854 : : * written by its own expression.
855 : : *
856 : : * We do this in the planner rather than parse analysis so that updatable
857 : : * views have been rewritten; otherwise they would mask which columns are
858 : : * generated. We need to check before preprocess_relation_rtes(), so that
859 : : * for virtual generated columns we still have the rangeVar. After that
860 : : * it is replaced by the column's expression.
861 : : *
862 : : * XXX: We plan to implement PERIODs as stored generated columns, so later
863 : : * we will loosen this restriction if the column belongs to a PERIOD.
864 : : */
865 [ + + ]: 394542 : if (parse->forPortionOf)
866 : : {
867 : 1071 : ForPortionOfExpr *forPortionOf = parse->forPortionOf;
868 : 1071 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
869 : :
870 [ + + ]: 1071 : if (get_attgenerated(rte->relid, forPortionOf->rangeVar->varattno))
871 [ + - ]: 16 : ereport(ERROR,
872 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
873 : : errmsg("cannot use generated column \"%s\" in FOR PORTION OF",
874 : : forPortionOf->range_name)));
875 : : }
876 : :
877 : : /*
878 : : * Scan the rangetable for relation RTEs and retrieve the necessary
879 : : * catalog information for each relation. Using this information, clear
880 : : * the inh flag for any relation that has no children, collect not-null
881 : : * attribute numbers for any relation that has column not-null
882 : : * constraints, and expand virtual generated columns for any relation that
883 : : * contains them. Note that this step does not descend into sublinks and
884 : : * subqueries; if we pull up any sublinks or subqueries below, their
885 : : * relation RTEs are processed just before pulling them up.
886 : : */
887 : 394526 : parse = root->parse = preprocess_relation_rtes(root);
888 : :
889 : : /*
890 : : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
891 : : * that we don't need so many special cases to deal with that situation.
892 : : */
893 : 394526 : replace_empty_jointree(parse);
894 : :
895 : : /*
896 : : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
897 : : * to transform them into joins. Note that this step does not descend
898 : : * into subqueries; if we pull up any subqueries below, their SubLinks are
899 : : * processed just before pulling them up.
900 : : */
901 [ + + ]: 394526 : if (parse->hasSubLinks)
902 : 30203 : pull_up_sublinks(root);
903 : :
904 : : /*
905 : : * Scan the rangetable for function RTEs, do const-simplification on them,
906 : : * and then inline them if possible (producing subqueries that might get
907 : : * pulled up next). Recursion issues here are handled in the same way as
908 : : * for SubLinks.
909 : : */
910 : 394526 : preprocess_function_rtes(root);
911 : :
912 : : /*
913 : : * Check to see if any subqueries in the jointree can be merged into this
914 : : * query.
915 : : */
916 : 394522 : pull_up_subqueries(root);
917 : :
918 : : /*
919 : : * If this is a simple UNION ALL query, flatten it into an appendrel. We
920 : : * do this now because it requires applying pull_up_subqueries to the leaf
921 : : * queries of the UNION ALL, which weren't touched above because they
922 : : * weren't referenced by the jointree (they will be after we do this).
923 : : */
924 [ + + ]: 394522 : if (parse->setOperations)
925 : 5476 : flatten_simple_union_all(root);
926 : :
927 : : /*
928 : : * Survey the rangetable to see what kinds of entries are present. We can
929 : : * skip some later processing if relevant SQL features are not used; for
930 : : * example if there are no JOIN RTEs we can avoid the expense of doing
931 : : * flatten_join_alias_vars(). This must be done after we have finished
932 : : * adding rangetable entries, of course. (Note: actually, processing of
933 : : * inherited or partitioned rels can cause RTEs for their child tables to
934 : : * get added later; but those must all be RTE_RELATION entries, so they
935 : : * don't invalidate the conclusions drawn here.)
936 : : */
937 : 394522 : root->hasJoinRTEs = false;
938 : 394522 : root->hasLateralRTEs = false;
939 : 394522 : root->group_rtindex = 0;
940 : 394522 : hasOuterJoins = false;
941 : 394522 : hasResultRTEs = false;
942 [ + - + + : 1086580 : foreach(l, parse->rtable)
+ + ]
943 : : {
944 : 692058 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
945 : :
946 [ + + + + ]: 692058 : switch (rte->rtekind)
947 : : {
948 : 71339 : case RTE_JOIN:
949 : 71339 : root->hasJoinRTEs = true;
950 [ + + ]: 71339 : if (IS_OUTER_JOIN(rte->jointype))
951 : 36295 : hasOuterJoins = true;
952 : 71339 : break;
953 : 146038 : case RTE_RESULT:
954 : 146038 : hasResultRTEs = true;
955 : 146038 : break;
956 : 4413 : case RTE_GROUP:
957 : : Assert(parse->hasGroupRTE);
958 : 4413 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
959 : 4413 : break;
960 : 470268 : default:
961 : : /* No work here for other RTE types */
962 : 470268 : break;
963 : : }
964 : :
965 [ + + ]: 692058 : if (rte->lateral)
966 : 8315 : root->hasLateralRTEs = true;
967 : :
968 : : /*
969 : : * We can also determine the maximum security level required for any
970 : : * securityQuals now. Addition of inheritance-child RTEs won't affect
971 : : * this, because child tables don't have their own securityQuals; see
972 : : * expand_single_inheritance_child().
973 : : */
974 [ + + ]: 692058 : if (rte->securityQuals)
975 [ - + ]: 2455 : root->qual_security_level = Max(root->qual_security_level,
976 : : list_length(rte->securityQuals));
977 : : }
978 : :
979 : : /*
980 : : * If we have now verified that the query target relation is
981 : : * non-inheriting, mark it as a leaf target.
982 : : */
983 [ + + ]: 394522 : if (parse->resultRelation)
984 : : {
985 : 62943 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
986 : :
987 [ + + ]: 62943 : if (!rte->inh)
988 : 60588 : root->leaf_result_relids =
989 : 60588 : bms_make_singleton(parse->resultRelation);
990 : : }
991 : :
992 : : /*
993 : : * This would be a convenient time to check access permissions for all
994 : : * relations mentioned in the query, since it would be better to fail now,
995 : : * before doing any detailed planning. However, for historical reasons,
996 : : * we leave this to be done at executor startup.
997 : : *
998 : : * Note, however, that we do need to check access permissions for any view
999 : : * relations mentioned in the query, in order to prevent information being
1000 : : * leaked by selectivity estimation functions, which only check view owner
1001 : : * permissions on underlying tables (see all_rows_selectable() and its
1002 : : * callers). This is a little ugly, because it means that access
1003 : : * permissions for views will be checked twice, which is another reason
1004 : : * why it would be better to do all the ACL checks here.
1005 : : */
1006 [ + - + + : 1085792 : foreach(l, parse->rtable)
+ + ]
1007 : : {
1008 : 691536 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1009 : :
1010 [ + + ]: 691536 : if (rte->perminfoindex != 0 &&
1011 [ + + ]: 372789 : rte->relkind == RELKIND_VIEW)
1012 : : {
1013 : : RTEPermissionInfo *perminfo;
1014 : : bool result;
1015 : :
1016 : 16330 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
1017 : 16330 : result = ExecCheckOneRelPerms(perminfo);
1018 [ + + ]: 16330 : if (!result)
1019 : 266 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
1020 : 266 : get_rel_name(perminfo->relid));
1021 : : }
1022 : : }
1023 : :
1024 : : /*
1025 : : * Preprocess RowMark information. We need to do this after subquery
1026 : : * pullup, so that all base relations are present.
1027 : : */
1028 : 394256 : preprocess_rowmarks(root);
1029 : :
1030 : : /*
1031 : : * Set hasHavingQual to remember if HAVING clause is present. Needed
1032 : : * because preprocess_expression will reduce a constant-true condition to
1033 : : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
1034 : : */
1035 : 394256 : root->hasHavingQual = (parse->havingQual != NULL);
1036 : :
1037 : : /*
1038 : : * Do expression preprocessing on targetlist and quals, as well as other
1039 : : * random expressions in the querytree. Note that we do not need to
1040 : : * handle sort/group expressions explicitly, because they are actually
1041 : : * part of the targetlist.
1042 : : */
1043 : 391587 : parse->targetList = (List *)
1044 : 394256 : preprocess_expression(root, (Node *) parse->targetList,
1045 : : EXPRKIND_TARGET);
1046 : :
1047 : 391587 : newWithCheckOptions = NIL;
1048 [ + + + + : 394155 : foreach(l, parse->withCheckOptions)
+ + ]
1049 : : {
1050 : 2568 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
1051 : :
1052 : 2568 : wco->qual = preprocess_expression(root, wco->qual,
1053 : : EXPRKIND_QUAL);
1054 [ + + ]: 2568 : if (wco->qual != NULL)
1055 : 2236 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
1056 : : }
1057 : 391587 : parse->withCheckOptions = newWithCheckOptions;
1058 : :
1059 : 391587 : parse->returningList = (List *)
1060 : 391587 : preprocess_expression(root, (Node *) parse->returningList,
1061 : : EXPRKIND_TARGET);
1062 : :
1063 : 391587 : preprocess_qual_conditions(root, (Node *) parse->jointree);
1064 : :
1065 : 391583 : parse->havingQual = preprocess_expression(root, parse->havingQual,
1066 : : EXPRKIND_QUAL);
1067 : :
1068 [ + + + + : 394139 : foreach(l, parse->windowClause)
+ + ]
1069 : : {
1070 : 2556 : WindowClause *wc = lfirst_node(WindowClause, l);
1071 : :
1072 : : /* partitionClause/orderClause are sort/group expressions */
1073 : 2556 : wc->startOffset = preprocess_expression(root, wc->startOffset,
1074 : : EXPRKIND_LIMIT);
1075 : 2556 : wc->endOffset = preprocess_expression(root, wc->endOffset,
1076 : : EXPRKIND_LIMIT);
1077 : : }
1078 : :
1079 : 391583 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
1080 : : EXPRKIND_LIMIT);
1081 : 391583 : parse->limitCount = preprocess_expression(root, parse->limitCount,
1082 : : EXPRKIND_LIMIT);
1083 : :
1084 [ + + ]: 391583 : if (parse->onConflict)
1085 : : {
1086 : 3580 : parse->onConflict->arbiterElems = (List *)
1087 : 1790 : preprocess_expression(root,
1088 : 1790 : (Node *) parse->onConflict->arbiterElems,
1089 : : EXPRKIND_ARBITER_ELEM);
1090 : 3580 : parse->onConflict->arbiterWhere =
1091 : 1790 : preprocess_expression(root,
1092 : 1790 : parse->onConflict->arbiterWhere,
1093 : : EXPRKIND_QUAL);
1094 : 3580 : parse->onConflict->onConflictSet = (List *)
1095 : 1790 : preprocess_expression(root,
1096 : 1790 : (Node *) parse->onConflict->onConflictSet,
1097 : : EXPRKIND_TARGET);
1098 : 1790 : parse->onConflict->onConflictWhere =
1099 : 1790 : preprocess_expression(root,
1100 : 1790 : parse->onConflict->onConflictWhere,
1101 : : EXPRKIND_QUAL);
1102 : : /* exclRelTlist contains only Vars, so no preprocessing needed */
1103 : : }
1104 : :
1105 [ + + ]: 391583 : if (parse->forPortionOf)
1106 : : {
1107 : 2094 : parse->forPortionOf->targetRange =
1108 : 1047 : preprocess_expression(root,
1109 : 1047 : parse->forPortionOf->targetRange,
1110 : : EXPRKIND_TARGET);
1111 [ + + ]: 1047 : if (contain_volatile_functions(parse->forPortionOf->targetRange))
1112 [ + - ]: 8 : ereport(ERROR,
1113 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1114 : : errmsg("FOR PORTION OF bounds cannot contain volatile functions")));
1115 : : }
1116 : :
1117 [ + + + + : 393863 : foreach(l, parse->mergeActionList)
+ + ]
1118 : : {
1119 : 2288 : MergeAction *action = (MergeAction *) lfirst(l);
1120 : :
1121 : 2288 : action->targetList = (List *)
1122 : 2288 : preprocess_expression(root,
1123 : 2288 : (Node *) action->targetList,
1124 : : EXPRKIND_TARGET);
1125 : 2288 : action->qual =
1126 : 2288 : preprocess_expression(root,
1127 : : (Node *) action->qual,
1128 : : EXPRKIND_QUAL);
1129 : : }
1130 : :
1131 : 391575 : parse->mergeJoinCondition =
1132 : 391575 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1133 : :
1134 : 391575 : root->append_rel_list = (List *)
1135 : 391575 : preprocess_expression(root, (Node *) root->append_rel_list,
1136 : : EXPRKIND_APPINFO);
1137 : :
1138 : : /* Also need to preprocess expressions within RTEs */
1139 [ + - + + : 1079755 : foreach(l, parse->rtable)
+ + ]
1140 : : {
1141 : 688188 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1142 : : int kind;
1143 : : ListCell *lcsq;
1144 : :
1145 [ + + ]: 688188 : if (rte->rtekind == RTE_RELATION)
1146 : : {
1147 [ + + ]: 359512 : if (rte->tablesample)
1148 : 180 : rte->tablesample = (TableSampleClause *)
1149 : 180 : preprocess_expression(root,
1150 : 180 : (Node *) rte->tablesample,
1151 : : EXPRKIND_TABLESAMPLE);
1152 : : }
1153 [ + + ]: 328676 : else if (rte->rtekind == RTE_SUBQUERY)
1154 : : {
1155 : : /*
1156 : : * We don't want to do all preprocessing yet on the subquery's
1157 : : * expressions, since that will happen when we plan it. But if it
1158 : : * contains any join aliases of our level, those have to get
1159 : : * expanded now, because planning of the subquery won't do it.
1160 : : * That's only possible if the subquery is LATERAL.
1161 : : */
1162 [ + + + + ]: 63284 : if (rte->lateral && root->hasJoinRTEs)
1163 : 1622 : rte->subquery = (Query *)
1164 : 1622 : flatten_join_alias_vars(root, root->parse,
1165 : 1622 : (Node *) rte->subquery);
1166 : : }
1167 [ + + ]: 265392 : else if (rte->rtekind == RTE_FUNCTION)
1168 : : {
1169 : : /* Preprocess the function expression(s) fully */
1170 [ + + ]: 34867 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
1171 : 34867 : rte->functions = (List *)
1172 : 34867 : preprocess_expression(root, (Node *) rte->functions, kind);
1173 : : }
1174 [ + + ]: 230525 : else if (rte->rtekind == RTE_TABLEFUNC)
1175 : : {
1176 : : /* Preprocess the function expression(s) fully */
1177 [ + + ]: 602 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1178 : 594 : rte->tablefunc = (TableFunc *)
1179 : 602 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1180 : : }
1181 [ + + ]: 229923 : else if (rte->rtekind == RTE_VALUES)
1182 : : {
1183 : : /* Preprocess the values lists fully */
1184 [ + + ]: 7003 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
1185 : 7003 : rte->values_lists = (List *)
1186 : 7003 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1187 : : }
1188 [ + + ]: 222920 : else if (rte->rtekind == RTE_GROUP)
1189 : : {
1190 : : /* Preprocess the groupexprs list fully */
1191 : 4413 : rte->groupexprs = (List *)
1192 : 4413 : preprocess_expression(root, (Node *) rte->groupexprs,
1193 : : EXPRKIND_GROUPEXPR);
1194 : : }
1195 : :
1196 : : /*
1197 : : * Process each element of the securityQuals list as if it were a
1198 : : * separate qual expression (as indeed it is). We need to do it this
1199 : : * way to get proper canonicalization of AND/OR structure. Note that
1200 : : * this converts each element into an implicit-AND sublist.
1201 : : */
1202 [ + + + + : 690985 : foreach(lcsq, rte->securityQuals)
+ + ]
1203 : : {
1204 : 2805 : lfirst(lcsq) = preprocess_expression(root,
1205 : 2805 : (Node *) lfirst(lcsq),
1206 : : EXPRKIND_QUAL);
1207 : : }
1208 : : }
1209 : :
1210 : : /*
1211 : : * Now that we are done preprocessing expressions, and in particular done
1212 : : * flattening join alias variables, get rid of the joinaliasvars lists.
1213 : : * They no longer match what expressions in the rest of the tree look
1214 : : * like, because we have not preprocessed expressions in those lists (and
1215 : : * do not want to; for example, expanding a SubLink there would result in
1216 : : * a useless unreferenced subplan). Leaving them in place simply creates
1217 : : * a hazard for later scans of the tree. We could try to prevent that by
1218 : : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1219 : : * but that doesn't sound very reliable.
1220 : : */
1221 [ + + ]: 391567 : if (root->hasJoinRTEs)
1222 : : {
1223 [ + - + + : 248831 : foreach(l, parse->rtable)
+ + ]
1224 : : {
1225 : 203948 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1226 : :
1227 : 203948 : rte->joinaliasvars = NIL;
1228 : : }
1229 : : }
1230 : :
1231 : : /*
1232 : : * Before we flatten GROUP Vars, identify HAVING clauses whose equality
1233 : : * semantics disagree with the GROUP BY's. See find_having_conflicts.
1234 : : */
1235 [ + + ]: 391567 : if (parse->hasGroupRTE)
1236 : 4413 : havingPushdownConflicts = find_having_conflicts(parse,
1237 : 4413 : root->group_rtindex);
1238 : : else
1239 : 387154 : havingPushdownConflicts = NULL;
1240 : :
1241 : : /*
1242 : : * Replace any Vars in the subquery's targetlist and havingQual that
1243 : : * reference GROUP outputs with the underlying grouping expressions.
1244 : : *
1245 : : * Note that we need to perform this replacement after we've preprocessed
1246 : : * the grouping expressions. This is to ensure that there is only one
1247 : : * instance of SubPlan for each SubLink contained within the grouping
1248 : : * expressions.
1249 : : */
1250 [ + + ]: 391567 : if (parse->hasGroupRTE)
1251 : : {
1252 : 4413 : parse->targetList = (List *)
1253 : 4413 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1254 : 4413 : parse->havingQual =
1255 : 4413 : flatten_group_exprs(root, root->parse, parse->havingQual);
1256 : : }
1257 : :
1258 : : /* Constant-folding might have removed all set-returning functions */
1259 [ + + ]: 391567 : if (parse->hasTargetSRFs)
1260 : 10208 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1261 : :
1262 : : /*
1263 : : * If we have grouping sets, expand the groupingSets tree of this query to
1264 : : * a flat list of grouping sets. We need to do this before optimizing
1265 : : * HAVING, since we can't easily tell if there's an empty grouping set
1266 : : * until we have this representation.
1267 : : */
1268 [ + + ]: 391567 : if (parse->groupingSets)
1269 : : {
1270 : 915 : parse->groupingSets =
1271 : 915 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1272 : : }
1273 : :
1274 : : /*
1275 : : * In some cases we may want to transfer a HAVING clause into WHERE. We
1276 : : * cannot do so if the HAVING clause contains aggregates (obviously) or
1277 : : * volatile functions (since a HAVING clause is supposed to be executed
1278 : : * only once per group). We also can't do this if there are any grouping
1279 : : * sets and the clause references any columns that are nullable by the
1280 : : * grouping sets; the nulled values of those columns are not available
1281 : : * before the grouping step. (The test on groupClause might seem wrong,
1282 : : * but it's okay: it's just an optimization to avoid running pull_varnos
1283 : : * when there cannot be any Vars in the HAVING clause.)
1284 : : *
1285 : : * We also cannot do this for HAVING clauses that conflict with GROUP BY
1286 : : * on collation or operator family. Both kinds of conflict are detected
1287 : : * before flatten_group_exprs (see find_having_conflicts above) and
1288 : : * recorded in the havingPushdownConflicts bitmapset. The bitmapset
1289 : : * indexes remain valid here because flatten_group_exprs uses
1290 : : * expression_tree_mutator, which preserves the list length and ordering
1291 : : * of havingQual.
1292 : : *
1293 : : * Also, it may be that the clause is so expensive to execute that we're
1294 : : * better off doing it only once per group, despite the loss of
1295 : : * selectivity. This is hard to estimate short of doing the entire
1296 : : * planning process twice, so we use a heuristic: clauses containing
1297 : : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1298 : : * clause into WHERE, in hopes of eliminating tuples before aggregation
1299 : : * instead of after.
1300 : : *
1301 : : * If the query has no empty grouping set then we can simply move such a
1302 : : * clause into WHERE; any group that fails the clause will not be in the
1303 : : * output because none of its tuples will reach the grouping or
1304 : : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1305 : : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1306 : : * clause must be degenerate (variable-free), so we can copy it into WHERE
1307 : : * so that query_planner() can use it in a gating Result node. (This could
1308 : : * be done better, but it seems not worth optimizing.)
1309 : : *
1310 : : * Note that a HAVING clause may contain expressions that are not fully
1311 : : * preprocessed. This can happen if these expressions are part of
1312 : : * grouping items. In such cases, they are replaced with GROUP Vars in
1313 : : * the parser and then replaced back after we're done with expression
1314 : : * preprocessing on havingQual. This is not an issue if the clause
1315 : : * remains in HAVING, because these expressions will be matched to lower
1316 : : * target items in setrefs.c. However, if the clause is moved or copied
1317 : : * into WHERE, we need to ensure that these expressions are fully
1318 : : * preprocessed.
1319 : : *
1320 : : * Note that both havingQual and parse->jointree->quals are in
1321 : : * implicitly-ANDed-list form at this point, even though they are declared
1322 : : * as Node *.
1323 : : */
1324 : 391567 : newHaving = NIL;
1325 : 391567 : havingIdx = 0;
1326 [ + + + + : 392877 : foreach(l, (List *) parse->havingQual)
+ + ]
1327 : : {
1328 : 1310 : Node *havingclause = (Node *) lfirst(l);
1329 : :
1330 [ + + + - ]: 1941 : if (contain_agg_clause(havingclause) ||
1331 [ + - ]: 1262 : contain_volatile_functions(havingclause) ||
1332 [ + + ]: 1262 : contain_subplans(havingclause) ||
1333 : 631 : bms_is_member(havingIdx, havingPushdownConflicts) ||
1334 [ + + + + : 611 : (parse->groupClause && parse->groupingSets &&
+ + ]
1335 : 100 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1336 : : {
1337 : : /* keep it in HAVING */
1338 : 859 : newHaving = lappend(newHaving, havingclause);
1339 : : }
1340 [ + + ]: 451 : else if (parse->groupClause &&
1341 [ + + ]: 421 : (parse->groupingSets == NIL ||
1342 [ + + ]: 40 : (List *) linitial(parse->groupingSets) != NIL))
1343 : 411 : {
1344 : : /* There is GROUP BY, but no empty grouping set */
1345 : : Node *whereclause;
1346 : :
1347 : : /* Preprocess the HAVING clause fully */
1348 : 411 : whereclause = preprocess_expression(root, havingclause,
1349 : : EXPRKIND_QUAL);
1350 : : /* ... and move it to WHERE */
1351 : 411 : parse->jointree->quals = (Node *)
1352 : 411 : list_concat((List *) parse->jointree->quals,
1353 : : (List *) whereclause);
1354 : : }
1355 : : else
1356 : : {
1357 : : /* There is an empty grouping set (perhaps implicitly) */
1358 : : Node *whereclause;
1359 : :
1360 : : /* Preprocess the HAVING clause fully */
1361 : 40 : whereclause = preprocess_expression(root, copyObject(havingclause),
1362 : : EXPRKIND_QUAL);
1363 : : /* ... and put a copy in WHERE */
1364 : 80 : parse->jointree->quals = (Node *)
1365 : 40 : list_concat((List *) parse->jointree->quals,
1366 : : (List *) whereclause);
1367 : : /* ... and also keep it in HAVING */
1368 : 40 : newHaving = lappend(newHaving, havingclause);
1369 : : }
1370 : :
1371 : 1310 : havingIdx++;
1372 : : }
1373 : 391567 : parse->havingQual = (Node *) newHaving;
1374 : :
1375 : : /*
1376 : : * If we have any outer joins, try to reduce them to plain inner joins.
1377 : : * This step is most easily done after we've done expression
1378 : : * preprocessing.
1379 : : */
1380 [ + + ]: 391567 : if (hasOuterJoins)
1381 : 25224 : reduce_outer_joins(root);
1382 : :
1383 : : /*
1384 : : * If we have any RTE_RESULT relations, see if they can be deleted from
1385 : : * the jointree. We also rely on this processing to flatten single-child
1386 : : * FromExprs underneath outer joins. This step is most effectively done
1387 : : * after we've done expression preprocessing and outer join reduction.
1388 : : */
1389 [ + + + + ]: 391567 : if (hasResultRTEs || hasOuterJoins)
1390 : 167399 : remove_useless_result_rtes(root);
1391 : :
1392 : : /*
1393 : : * Do the main planning.
1394 : : */
1395 : 391567 : grouping_planner(root, tuple_fraction, setops);
1396 : :
1397 : : /*
1398 : : * Capture the set of outer-level param IDs we have access to, for use in
1399 : : * extParam/allParam calculations later.
1400 : : */
1401 : 391520 : SS_identify_outer_params(root);
1402 : :
1403 : : /*
1404 : : * If any initPlans were created in this query level, adjust the surviving
1405 : : * Paths' costs and parallel-safety flags to account for them. The
1406 : : * initPlans won't actually get attached to the plan tree till
1407 : : * create_plan() runs, but we must include their effects now.
1408 : : */
1409 : 391520 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1410 : 391520 : SS_charge_for_initplans(root, final_rel);
1411 : :
1412 : : /*
1413 : : * Make sure we've identified the cheapest Path for the final rel. (By
1414 : : * doing this here not in grouping_planner, we include initPlan costs in
1415 : : * the decision, though it's unlikely that will change anything.)
1416 : : */
1417 : 391520 : set_cheapest(final_rel);
1418 : :
1419 : 391520 : return root;
1420 : : }
1421 : :
1422 : : /*
1423 : : * preprocess_expression
1424 : : * Do subquery_planner's preprocessing work for an expression,
1425 : : * which can be a targetlist, a WHERE clause (including JOIN/ON
1426 : : * conditions), a HAVING clause, or a few other things.
1427 : : */
1428 : : static Node *
1429 : 3301136 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1430 : : {
1431 : : /*
1432 : : * Fall out quickly if expression is empty. This occurs often enough to
1433 : : * be worth checking. Note that null->null is the correct conversion for
1434 : : * implicit-AND result format, too.
1435 : : */
1436 [ + + ]: 3301136 : if (expr == NULL)
1437 : 2595820 : return NULL;
1438 : :
1439 : : /*
1440 : : * If the query has any join RTEs, replace join alias variables with
1441 : : * base-relation variables. We must do this first, since any expressions
1442 : : * we may extract from the joinaliasvars lists have not been preprocessed.
1443 : : * For example, if we did this after sublink processing, sublinks expanded
1444 : : * out from join aliases would not get processed. But we can skip this in
1445 : : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1446 : : * they can't contain any Vars of the current query level.
1447 : : */
1448 [ + + + + ]: 705316 : if (root->hasJoinRTEs &&
1449 [ + + + - ]: 320112 : !(kind == EXPRKIND_RTFUNC ||
1450 [ + + ]: 159907 : kind == EXPRKIND_VALUES ||
1451 : : kind == EXPRKIND_TABLESAMPLE ||
1452 : : kind == EXPRKIND_TABLEFUNC))
1453 : 159892 : expr = flatten_join_alias_vars(root, root->parse, expr);
1454 : :
1455 : : /*
1456 : : * Simplify constant expressions. For function RTEs, this was already
1457 : : * done by preprocess_function_rtes. (But note we must do it again for
1458 : : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1459 : : * un-simplified subexpressions inserted by flattening of subqueries or
1460 : : * join alias variables.)
1461 : : *
1462 : : * Note: an essential effect of this is to convert named-argument function
1463 : : * calls to positional notation and insert the current actual values of
1464 : : * any default arguments for functions. To ensure that happens, we *must*
1465 : : * process all expressions here. Previous PG versions sometimes skipped
1466 : : * const-simplification if it didn't seem worth the trouble, but we can't
1467 : : * do that anymore.
1468 : : *
1469 : : * Note: this also flattens nested AND and OR expressions into N-argument
1470 : : * form. All processing of a qual expression after this point must be
1471 : : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1472 : : * with AND directly under AND, nor OR directly under OR.
1473 : : */
1474 [ + + ]: 705316 : if (kind != EXPRKIND_RTFUNC)
1475 : 675033 : expr = eval_const_expressions(root, expr);
1476 : :
1477 : : /*
1478 : : * If it's a qual or havingQual, canonicalize it.
1479 : : */
1480 [ + + ]: 702635 : if (kind == EXPRKIND_QUAL)
1481 : : {
1482 : 258495 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1483 : :
1484 : : #ifdef OPTIMIZER_DEBUG
1485 : : printf("After canonicalize_qual()\n");
1486 : : pprint(expr);
1487 : : #endif
1488 : : }
1489 : :
1490 : : /*
1491 : : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1492 : : * hashfuncid of any that might execute more quickly by using hash lookups
1493 : : * instead of a linear search.
1494 : : */
1495 [ + + + + ]: 702635 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1496 : : {
1497 : 644104 : convert_saop_to_hashed_saop(expr);
1498 : : }
1499 : :
1500 : : /* Expand SubLinks to SubPlans */
1501 [ + + ]: 702635 : if (root->parse->hasSubLinks)
1502 : 88739 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1503 : :
1504 : : /*
1505 : : * XXX do not insert anything here unless you have grokked the comments in
1506 : : * SS_replace_correlation_vars ...
1507 : : */
1508 : :
1509 : : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1510 [ + + ]: 702635 : if (root->query_level > 1)
1511 : 143458 : expr = SS_replace_correlation_vars(root, expr);
1512 : :
1513 : : /*
1514 : : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1515 : : * don't want to do this before eval_const_expressions, since the latter
1516 : : * would be unable to simplify a top-level AND correctly. Also,
1517 : : * SS_process_sublinks expects explicit-AND format.)
1518 : : */
1519 [ + + ]: 702635 : if (kind == EXPRKIND_QUAL)
1520 : 258495 : expr = (Node *) make_ands_implicit((Expr *) expr);
1521 : :
1522 : 702635 : return expr;
1523 : : }
1524 : :
1525 : : /*
1526 : : * preprocess_qual_conditions
1527 : : * Recursively scan the query's jointree and do subquery_planner's
1528 : : * preprocessing work on each qual condition found therein.
1529 : : */
1530 : : static void
1531 : 991242 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1532 : : {
1533 [ - + ]: 991242 : if (jtnode == NULL)
1534 : 0 : return;
1535 [ + + ]: 991242 : if (IsA(jtnode, RangeTblRef))
1536 : : {
1537 : : /* nothing to do here */
1538 : : }
1539 [ + + ]: 486535 : else if (IsA(jtnode, FromExpr))
1540 : : {
1541 : 406478 : FromExpr *f = (FromExpr *) jtnode;
1542 : : ListCell *l;
1543 : :
1544 [ + + + + : 846019 : foreach(l, f->fromlist)
+ + ]
1545 : 439541 : preprocess_qual_conditions(root, lfirst(l));
1546 : :
1547 : 406478 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1548 : : }
1549 [ + - ]: 80057 : else if (IsA(jtnode, JoinExpr))
1550 : : {
1551 : 80057 : JoinExpr *j = (JoinExpr *) jtnode;
1552 : :
1553 : 80057 : preprocess_qual_conditions(root, j->larg);
1554 : 80057 : preprocess_qual_conditions(root, j->rarg);
1555 : :
1556 : 80057 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1557 : : }
1558 : : else
1559 [ # # ]: 0 : elog(ERROR, "unrecognized node type: %d",
1560 : : (int) nodeTag(jtnode));
1561 : : }
1562 : :
1563 : : /*
1564 : : * find_having_conflicts
1565 : : * Identify HAVING clauses that must not be moved to WHERE because they
1566 : : * apply a different equivalence relation than GROUP BY. Pushing such a
1567 : : * clause to WHERE would filter individual rows before grouping happens,
1568 : : * eliminating rows that GROUP BY would have merged into a single group
1569 : : * and thereby changing aggregate results.
1570 : : *
1571 : : * The actual walking is done by expression_has_grouping_conflict; see that
1572 : : * function for the kinds of conflict it looks for. We just iterate over
1573 : : * havingQual and supply a HAVING-specific callback that identifies GROUP
1574 : : * Vars.
1575 : : *
1576 : : * This must be called before flatten_group_exprs, while the HAVING clause
1577 : : * still contains GROUP Vars (Vars referencing RTE_GROUP). These GROUP Vars
1578 : : * carry the GROUP BY collation as their varcollid and let us recover the
1579 : : * grouping eqop via varattno. After flattening, those Vars are replaced by
1580 : : * the underlying expressions, and matching back to grouping expressions is
1581 : : * much harder.
1582 : : *
1583 : : * Returns a Bitmapset of zero-based indexes into the havingQual list for
1584 : : * clauses that conflict and must stay in HAVING.
1585 : : */
1586 : : static Bitmapset *
1587 : 4413 : find_having_conflicts(Query *parse, Index group_rtindex)
1588 : : {
1589 : 4413 : Bitmapset *result = NULL;
1590 : : having_grouping_ctx ctx;
1591 : : int idx;
1592 : :
1593 [ + + ]: 4413 : if (parse->havingQual == NULL)
1594 : 3535 : return NULL;
1595 : :
1596 : 878 : ctx.parse = parse;
1597 : 878 : ctx.group_rtindex = group_rtindex;
1598 : :
1599 : 878 : idx = 0;
1600 [ + - + + : 2773 : foreach_ptr(Node, clause, (List *) parse->havingQual)
+ + ]
1601 : : {
1602 [ + + ]: 1017 : if (expression_has_grouping_conflict(clause, having_var_grouping_eqop,
1603 : : &ctx))
1604 : 120 : result = bms_add_member(result, idx);
1605 : 1017 : idx++;
1606 : : }
1607 : :
1608 : 878 : return result;
1609 : : }
1610 : :
1611 : : /*
1612 : : * having_var_grouping_eqop
1613 : : * grouping_eqop_callback for find_having_conflicts.
1614 : : *
1615 : : * Returns the GROUP BY equality operator for 'var' if it references the
1616 : : * query's RTE_GROUP, or InvalidOid otherwise.
1617 : : */
1618 : : static Oid
1619 : 968 : having_var_grouping_eqop(Var *var, void *context)
1620 : : {
1621 : 968 : having_grouping_ctx *ctx = (having_grouping_ctx *) context;
1622 : :
1623 [ + + - + ]: 968 : if (var->varno != ctx->group_rtindex || var->varlevelsup != 0)
1624 : 384 : return InvalidOid;
1625 : :
1626 : 584 : return group_var_eqop(ctx->parse, var);
1627 : : }
1628 : :
1629 : : /*
1630 : : * group_var_eqop
1631 : : * Return the equality operator that GROUP BY uses for the given GROUP Var.
1632 : : *
1633 : : * A GROUP Var's varattno is its 1-based position in the RTE_GROUP's groupexprs
1634 : : * list, which addRangeTableEntryForGroup built by iterating parse->groupClause
1635 : : * and including every SortGroupClause whose TLE was present in the targetlist.
1636 : : * Replay that traversal here to recover the SortGroupClause for the given
1637 : : * varattno.
1638 : : */
1639 : : static Oid
1640 : 584 : group_var_eqop(Query *parse, Var *var)
1641 : : {
1642 : 584 : int counter = 0;
1643 : :
1644 : : Assert(var->varlevelsup == 0);
1645 : :
1646 [ + - + - : 926 : foreach_node(SortGroupClause, sgc, parse->groupClause)
+ - ]
1647 : : {
1648 [ - + ]: 926 : if (get_sortgroupclause_tle(sgc, parse->targetList) == NULL)
1649 : 0 : continue;
1650 [ + + ]: 926 : if (++counter == var->varattno)
1651 : 584 : return sgc->eqop;
1652 : : }
1653 : :
1654 [ # # ]: 0 : elog(ERROR, "could not find GROUP clause for GROUP Var attno %d",
1655 : : var->varattno);
1656 : : return InvalidOid; /* keep compiler quiet */
1657 : : }
1658 : :
1659 : : /*
1660 : : * preprocess_phv_expression
1661 : : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1662 : : *
1663 : : * If a LATERAL subquery references an output of another subquery, and that
1664 : : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1665 : : * join, then we'll push the PlaceHolderVar expression down into the subquery
1666 : : * and later pull it back up during find_lateral_references, which runs after
1667 : : * subquery_planner has preprocessed all the expressions that were in the
1668 : : * current query level to start with. So we need to preprocess it then.
1669 : : */
1670 : : Expr *
1671 : 75 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1672 : : {
1673 : 75 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1674 : : }
1675 : :
1676 : : /*--------------------
1677 : : * grouping_planner
1678 : : * Perform planning steps related to grouping, aggregation, etc.
1679 : : *
1680 : : * This function adds all required top-level processing to the scan/join
1681 : : * Path(s) produced by query_planner.
1682 : : *
1683 : : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1684 : : * tuple_fraction is interpreted as follows:
1685 : : * 0: expect all tuples to be retrieved (normal case)
1686 : : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1687 : : * from the plan to be retrieved
1688 : : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1689 : : * expected to be retrieved (ie, a LIMIT specification).
1690 : : * setops is used for set operation subqueries to provide the subquery with
1691 : : * the context in which it's being used so that Paths correctly sorted for the
1692 : : * set operation can be generated. NULL when not planning a set operation
1693 : : * child, or when a child of a set op that isn't interested in sorted input.
1694 : : *
1695 : : * Returns nothing; the useful output is in the Paths we attach to the
1696 : : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1697 : : * root->processed_tlist contains the final processed targetlist.
1698 : : *
1699 : : * Note that we have not done set_cheapest() on the final rel; it's convenient
1700 : : * to leave this to the caller.
1701 : : *--------------------
1702 : : */
1703 : : static void
1704 : 391567 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1705 : : SetOperationStmt *setops)
1706 : : {
1707 : 391567 : Query *parse = root->parse;
1708 : 391567 : int64 offset_est = 0;
1709 : 391567 : int64 count_est = 0;
1710 : 391567 : double limit_tuples = -1.0;
1711 : 391567 : bool have_postponed_srfs = false;
1712 : : PathTarget *final_target;
1713 : : List *final_targets;
1714 : : List *final_targets_contain_srfs;
1715 : : bool final_target_parallel_safe;
1716 : : RelOptInfo *current_rel;
1717 : : RelOptInfo *final_rel;
1718 : : FinalPathExtraData extra;
1719 : : ListCell *lc;
1720 : :
1721 : : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
1722 [ + + + + ]: 391567 : if (parse->limitCount || parse->limitOffset)
1723 : : {
1724 : 3737 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1725 : : &offset_est, &count_est);
1726 : :
1727 : : /*
1728 : : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1729 : : * estimate the effects of using a bounded sort.
1730 : : */
1731 [ + + + + ]: 3737 : if (count_est > 0 && offset_est >= 0)
1732 : 3255 : limit_tuples = (double) count_est + (double) offset_est;
1733 : : }
1734 : :
1735 : : /* Make tuple_fraction accessible to lower-level routines */
1736 : 391567 : root->tuple_fraction = tuple_fraction;
1737 : :
1738 [ + + ]: 391567 : if (parse->setOperations)
1739 : : {
1740 : : /*
1741 : : * Construct Paths for set operations. The results will not need any
1742 : : * work except perhaps a top-level sort and/or LIMIT. Note that any
1743 : : * special work for recursive unions is the responsibility of
1744 : : * plan_set_operations.
1745 : : */
1746 : 4932 : current_rel = plan_set_operations(root);
1747 : :
1748 : : /*
1749 : : * We should not need to call preprocess_targetlist, since we must be
1750 : : * in a SELECT query node. Instead, use the processed_tlist returned
1751 : : * by plan_set_operations (since this tells whether it returned any
1752 : : * resjunk columns!), and transfer any sort key information from the
1753 : : * original tlist.
1754 : : */
1755 : : Assert(parse->commandType == CMD_SELECT);
1756 : :
1757 : : /* for safety, copy processed_tlist instead of modifying in-place */
1758 : 4928 : root->processed_tlist =
1759 : 4928 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1760 : : parse->targetList);
1761 : :
1762 : : /* Also extract the PathTarget form of the setop result tlist */
1763 : 4928 : final_target = current_rel->cheapest_total_path->pathtarget;
1764 : :
1765 : : /* And check whether it's parallel safe */
1766 : : final_target_parallel_safe =
1767 : 4928 : is_parallel_safe(root, (Node *) final_target->exprs);
1768 : :
1769 : : /* The setop result tlist couldn't contain any SRFs */
1770 : : Assert(!parse->hasTargetSRFs);
1771 : 4928 : final_targets = final_targets_contain_srfs = NIL;
1772 : :
1773 : : /*
1774 : : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1775 : : * checked already, but let's make sure).
1776 : : */
1777 [ - + ]: 4928 : if (parse->rowMarks)
1778 [ # # ]: 0 : ereport(ERROR,
1779 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1780 : : /*------
1781 : : translator: %s is a SQL row locking clause such as FOR UPDATE */
1782 : : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1783 : : LCS_asString(linitial_node(RowMarkClause,
1784 : : parse->rowMarks)->strength))));
1785 : :
1786 : : /*
1787 : : * Calculate pathkeys that represent result ordering requirements
1788 : : */
1789 : : Assert(parse->distinctClause == NIL);
1790 : 4928 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1791 : : parse->sortClause,
1792 : : root->processed_tlist);
1793 : : }
1794 : : else
1795 : : {
1796 : : /* No set operations, do regular planning */
1797 : : PathTarget *sort_input_target;
1798 : : List *sort_input_targets;
1799 : : List *sort_input_targets_contain_srfs;
1800 : : bool sort_input_target_parallel_safe;
1801 : : PathTarget *grouping_target;
1802 : : List *grouping_targets;
1803 : : List *grouping_targets_contain_srfs;
1804 : : bool grouping_target_parallel_safe;
1805 : : PathTarget *scanjoin_target;
1806 : : List *scanjoin_targets;
1807 : : List *scanjoin_targets_contain_srfs;
1808 : : bool scanjoin_target_parallel_safe;
1809 : : bool scanjoin_target_same_exprs;
1810 : : bool have_grouping;
1811 : 386635 : WindowFuncLists *wflists = NULL;
1812 : 386635 : List *activeWindows = NIL;
1813 : 386635 : grouping_sets_data *gset_data = NULL;
1814 : : standard_qp_extra qp_extra;
1815 : :
1816 : : /* A recursive query should always have setOperations */
1817 : : Assert(!root->hasRecursion);
1818 : :
1819 : : /* Preprocess grouping sets and GROUP BY clause, if any */
1820 [ + + ]: 386635 : if (parse->groupingSets)
1821 : : {
1822 : 915 : gset_data = preprocess_grouping_sets(root);
1823 : : }
1824 [ + + ]: 385720 : else if (parse->groupClause)
1825 : : {
1826 : : /* Preprocess regular GROUP BY clause, if any */
1827 : 3573 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1828 : : }
1829 : :
1830 : : /*
1831 : : * Preprocess targetlist. Note that much of the remaining planning
1832 : : * work will be done with the PathTarget representation of tlists, but
1833 : : * we must also maintain the full representation of the final tlist so
1834 : : * that we can transfer its decoration (resnames etc) to the topmost
1835 : : * tlist of the finished Plan. This is kept in processed_tlist.
1836 : : */
1837 : 386631 : preprocess_targetlist(root);
1838 : :
1839 : : /*
1840 : : * Mark all the aggregates with resolved aggtranstypes, and detect
1841 : : * aggregates that are duplicates or can share transition state. We
1842 : : * must do this before slicing and dicing the tlist into various
1843 : : * pathtargets, else some copies of the Aggref nodes might escape
1844 : : * being marked.
1845 : : */
1846 [ + + ]: 386631 : if (parse->hasAggs)
1847 : : {
1848 : 33345 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1849 : 33345 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1850 : : }
1851 : :
1852 : : /*
1853 : : * Locate any window functions in the tlist. (We don't need to look
1854 : : * anywhere else, since expressions used in ORDER BY will be in there
1855 : : * too.) Note that they could all have been eliminated by constant
1856 : : * folding, in which case we don't need to do any more work.
1857 : : */
1858 [ + + ]: 386631 : if (parse->hasWindowFuncs)
1859 : : {
1860 : 2336 : wflists = find_window_functions((Node *) root->processed_tlist,
1861 : 2336 : list_length(parse->windowClause));
1862 [ + + ]: 2336 : if (wflists->numWindowFuncs > 0)
1863 : : {
1864 : : /*
1865 : : * See if any modifications can be made to each WindowClause
1866 : : * to allow the executor to execute the WindowFuncs more
1867 : : * quickly.
1868 : : */
1869 : 2331 : optimize_window_clauses(root, wflists);
1870 : :
1871 : : /* Extract the list of windows actually in use. */
1872 : 2331 : activeWindows = select_active_windows(root, wflists);
1873 : :
1874 : : /* Make sure they all have names, for EXPLAIN's use. */
1875 : 2331 : name_active_windows(activeWindows);
1876 : : }
1877 : : else
1878 : 5 : parse->hasWindowFuncs = false;
1879 : : }
1880 : :
1881 : : /*
1882 : : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1883 : : * adding logic between here and the query_planner() call. Anything
1884 : : * that is needed in MIN/MAX-optimizable cases will have to be
1885 : : * duplicated in planagg.c.
1886 : : */
1887 [ + + ]: 386631 : if (parse->hasAggs)
1888 : 33345 : preprocess_minmax_aggregates(root);
1889 : :
1890 : : /*
1891 : : * Figure out whether there's a hard limit on the number of rows that
1892 : : * query_planner's result subplan needs to return. Even if we know a
1893 : : * hard limit overall, it doesn't apply if the query has any
1894 : : * grouping/aggregation operations, or SRFs in the tlist.
1895 : : */
1896 [ + + ]: 386631 : if (parse->groupClause ||
1897 [ + + ]: 382222 : parse->groupingSets ||
1898 [ + + ]: 382147 : parse->distinctClause ||
1899 [ + + ]: 380161 : parse->hasAggs ||
1900 [ + + ]: 350607 : parse->hasWindowFuncs ||
1901 [ + + ]: 348397 : parse->hasTargetSRFs ||
1902 [ + + ]: 338518 : root->hasHavingQual)
1903 : 48133 : root->limit_tuples = -1.0;
1904 : : else
1905 : 338498 : root->limit_tuples = limit_tuples;
1906 : :
1907 : : /* Set up data needed by standard_qp_callback */
1908 : 386631 : qp_extra.activeWindows = activeWindows;
1909 : 386631 : qp_extra.gset_data = gset_data;
1910 : :
1911 : : /*
1912 : : * If we're a subquery for a set operation, store the SetOperationStmt
1913 : : * in qp_extra.
1914 : : */
1915 : 386631 : qp_extra.setop = setops;
1916 : :
1917 : : /*
1918 : : * Generate the best unsorted and presorted paths for the scan/join
1919 : : * portion of this Query, ie the processing represented by the
1920 : : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1921 : : * We also generate (in standard_qp_callback) pathkey representations
1922 : : * of the query's sort clause, distinct clause, etc.
1923 : : */
1924 : 386631 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1925 : :
1926 : : /*
1927 : : * Convert the query's result tlist into PathTarget format.
1928 : : *
1929 : : * Note: this cannot be done before query_planner() has performed
1930 : : * appendrel expansion, because that might add resjunk entries to
1931 : : * root->processed_tlist. Waiting till afterwards is also helpful
1932 : : * because the target width estimates can use per-Var width numbers
1933 : : * that were obtained within query_planner().
1934 : : */
1935 : 386596 : final_target = create_pathtarget(root, root->processed_tlist);
1936 : : final_target_parallel_safe =
1937 : 386596 : is_parallel_safe(root, (Node *) final_target->exprs);
1938 : :
1939 : : /*
1940 : : * If ORDER BY was given, consider whether we should use a post-sort
1941 : : * projection, and compute the adjusted target for preceding steps if
1942 : : * so.
1943 : : */
1944 [ + + ]: 386596 : if (parse->sortClause)
1945 : : {
1946 : 57783 : sort_input_target = make_sort_input_target(root,
1947 : : final_target,
1948 : : &have_postponed_srfs);
1949 : : sort_input_target_parallel_safe =
1950 : 57783 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1951 : : }
1952 : : else
1953 : : {
1954 : 328813 : sort_input_target = final_target;
1955 : 328813 : sort_input_target_parallel_safe = final_target_parallel_safe;
1956 : : }
1957 : :
1958 : : /*
1959 : : * If we have window functions to deal with, the output from any
1960 : : * grouping step needs to be what the window functions want;
1961 : : * otherwise, it should be sort_input_target.
1962 : : */
1963 [ + + ]: 386596 : if (activeWindows)
1964 : : {
1965 : 2331 : grouping_target = make_window_input_target(root,
1966 : : final_target,
1967 : : activeWindows);
1968 : : grouping_target_parallel_safe =
1969 : 2331 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1970 : : }
1971 : : else
1972 : : {
1973 : 384265 : grouping_target = sort_input_target;
1974 : 384265 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1975 : : }
1976 : :
1977 : : /*
1978 : : * If we have grouping or aggregation to do, the topmost scan/join
1979 : : * plan node must emit what the grouping step wants; otherwise, it
1980 : : * should emit grouping_target.
1981 : : */
1982 [ + + ]: 382187 : have_grouping = (parse->groupClause || parse->groupingSets ||
1983 [ + + + + : 768783 : parse->hasAggs || root->hasHavingQual);
+ + ]
1984 [ + + ]: 386596 : if (have_grouping)
1985 : : {
1986 : 34089 : scanjoin_target = make_group_input_target(root, final_target);
1987 : : scanjoin_target_parallel_safe =
1988 : 34089 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1989 : : }
1990 : : else
1991 : : {
1992 : 352507 : scanjoin_target = grouping_target;
1993 : 352507 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1994 : : }
1995 : :
1996 : : /*
1997 : : * If there are any SRFs in the targetlist, we must separate each of
1998 : : * these PathTargets into SRF-computing and SRF-free targets. Replace
1999 : : * each of the named targets with a SRF-free version, and remember the
2000 : : * list of additional projection steps we need to add afterwards.
2001 : : */
2002 [ + + ]: 386596 : if (parse->hasTargetSRFs)
2003 : : {
2004 : : /* final_target doesn't recompute any SRFs in sort_input_target */
2005 : 10208 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
2006 : : &final_targets,
2007 : : &final_targets_contain_srfs);
2008 : 10208 : final_target = linitial_node(PathTarget, final_targets);
2009 : : Assert(!linitial_int(final_targets_contain_srfs));
2010 : : /* likewise for sort_input_target vs. grouping_target */
2011 : 10208 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
2012 : : &sort_input_targets,
2013 : : &sort_input_targets_contain_srfs);
2014 : 10208 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
2015 : : Assert(!linitial_int(sort_input_targets_contain_srfs));
2016 : : /* likewise for grouping_target vs. scanjoin_target */
2017 : 10208 : split_pathtarget_at_srfs_grouping(root,
2018 : : grouping_target, scanjoin_target,
2019 : : &grouping_targets,
2020 : : &grouping_targets_contain_srfs);
2021 : 10208 : grouping_target = linitial_node(PathTarget, grouping_targets);
2022 : : Assert(!linitial_int(grouping_targets_contain_srfs));
2023 : : /* scanjoin_target will not have any SRFs precomputed for it */
2024 : 10208 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
2025 : : &scanjoin_targets,
2026 : : &scanjoin_targets_contain_srfs);
2027 : 10208 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
2028 : : Assert(!linitial_int(scanjoin_targets_contain_srfs));
2029 : : }
2030 : : else
2031 : : {
2032 : : /* initialize lists; for most of these, dummy values are OK */
2033 : 376388 : final_targets = final_targets_contain_srfs = NIL;
2034 : 376388 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
2035 : 376388 : grouping_targets = grouping_targets_contain_srfs = NIL;
2036 : 376388 : scanjoin_targets = list_make1(scanjoin_target);
2037 : 376388 : scanjoin_targets_contain_srfs = NIL;
2038 : : }
2039 : :
2040 : : /* Apply scan/join target. */
2041 : 386596 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
2042 [ + + + + ]: 386596 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
2043 : 386596 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
2044 : : scanjoin_targets_contain_srfs,
2045 : : scanjoin_target_parallel_safe,
2046 : : scanjoin_target_same_exprs);
2047 : :
2048 : : /*
2049 : : * Save the various upper-rel PathTargets we just computed into
2050 : : * root->upper_targets[]. The core code doesn't use this, but it
2051 : : * provides a convenient place for extensions to get at the info. For
2052 : : * consistency, we save all the intermediate targets, even though some
2053 : : * of the corresponding upperrels might not be needed for this query.
2054 : : */
2055 : 386596 : root->upper_targets[UPPERREL_FINAL] = final_target;
2056 : 386596 : root->upper_targets[UPPERREL_ORDERED] = final_target;
2057 : 386596 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
2058 : 386596 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
2059 : 386596 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
2060 : 386596 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
2061 : :
2062 : : /*
2063 : : * If we have grouping and/or aggregation, consider ways to implement
2064 : : * that. We build a new upperrel representing the output of this
2065 : : * phase.
2066 : : */
2067 [ + + ]: 386596 : if (have_grouping)
2068 : : {
2069 : 34089 : current_rel = create_grouping_paths(root,
2070 : : current_rel,
2071 : : grouping_target,
2072 : : grouping_target_parallel_safe,
2073 : : gset_data);
2074 : : /* Fix things up if grouping_target contains SRFs */
2075 [ + + ]: 34085 : if (parse->hasTargetSRFs)
2076 : 294 : adjust_paths_for_srfs(root, current_rel,
2077 : : grouping_targets,
2078 : : grouping_targets_contain_srfs);
2079 : : }
2080 : :
2081 : : /*
2082 : : * If we have window functions, consider ways to implement those. We
2083 : : * build a new upperrel representing the output of this phase.
2084 : : */
2085 [ + + ]: 386592 : if (activeWindows)
2086 : : {
2087 : 2331 : current_rel = create_window_paths(root,
2088 : : current_rel,
2089 : : grouping_target,
2090 : : sort_input_target,
2091 : : sort_input_target_parallel_safe,
2092 : : wflists,
2093 : : activeWindows);
2094 : : /* Fix things up if sort_input_target contains SRFs */
2095 [ + + ]: 2331 : if (parse->hasTargetSRFs)
2096 : 5 : adjust_paths_for_srfs(root, current_rel,
2097 : : sort_input_targets,
2098 : : sort_input_targets_contain_srfs);
2099 : : }
2100 : :
2101 : : /*
2102 : : * If there is a DISTINCT clause, consider ways to implement that. We
2103 : : * build a new upperrel representing the output of this phase.
2104 : : */
2105 [ + + ]: 386592 : if (parse->distinctClause)
2106 : : {
2107 : 2013 : current_rel = create_distinct_paths(root,
2108 : : current_rel,
2109 : : sort_input_target);
2110 : : }
2111 : : } /* end of if (setOperations) */
2112 : :
2113 : : /*
2114 : : * If ORDER BY was given, consider ways to implement that, and generate a
2115 : : * new upperrel containing only paths that emit the correct ordering and
2116 : : * project the correct final_target. We can apply the original
2117 : : * limit_tuples limit in sort costing here, but only if there are no
2118 : : * postponed SRFs.
2119 : : */
2120 [ + + ]: 391520 : if (parse->sortClause)
2121 : : {
2122 [ + + ]: 60982 : current_rel = create_ordered_paths(root,
2123 : : current_rel,
2124 : : final_target,
2125 : : final_target_parallel_safe,
2126 : : have_postponed_srfs ? -1.0 :
2127 : : limit_tuples);
2128 : : /* Fix things up if final_target contains SRFs */
2129 [ + + ]: 60982 : if (parse->hasTargetSRFs)
2130 : 182 : adjust_paths_for_srfs(root, current_rel,
2131 : : final_targets,
2132 : : final_targets_contain_srfs);
2133 : : }
2134 : :
2135 : : /*
2136 : : * Now we are prepared to build the final-output upperrel.
2137 : : */
2138 : 391520 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
2139 : :
2140 : : /*
2141 : : * If the input rel is marked consider_parallel and there's nothing that's
2142 : : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
2143 : : * consider_parallel as well. Note that if the query has rowMarks or is
2144 : : * not a SELECT, consider_parallel will be false for every relation in the
2145 : : * query.
2146 : : */
2147 [ + + + + ]: 527139 : if (current_rel->consider_parallel &&
2148 [ + + ]: 271218 : is_parallel_safe(root, parse->limitOffset) &&
2149 : 135599 : is_parallel_safe(root, parse->limitCount))
2150 : 135594 : final_rel->consider_parallel = true;
2151 : :
2152 : : /*
2153 : : * If the current_rel belongs to a single FDW, so does the final_rel.
2154 : : */
2155 : 391520 : final_rel->serverid = current_rel->serverid;
2156 : 391520 : final_rel->userid = current_rel->userid;
2157 : 391520 : final_rel->useridiscurrent = current_rel->useridiscurrent;
2158 : 391520 : final_rel->fdwroutine = current_rel->fdwroutine;
2159 : :
2160 : : /*
2161 : : * Generate paths for the final_rel. Insert all surviving paths, with
2162 : : * LockRows, Limit, and/or ModifyTable steps added if needed.
2163 : : */
2164 [ + - + + : 799669 : foreach(lc, current_rel->pathlist)
+ + ]
2165 : : {
2166 : 408149 : Path *path = (Path *) lfirst(lc);
2167 : :
2168 : : /*
2169 : : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
2170 : : * (Note: we intentionally test parse->rowMarks not root->rowMarks
2171 : : * here. If there are only non-locking rowmarks, they should be
2172 : : * handled by the ModifyTable node instead. However, root->rowMarks
2173 : : * is what goes into the LockRows node.)
2174 : : */
2175 [ + + ]: 408149 : if (parse->rowMarks)
2176 : : {
2177 : 6819 : path = (Path *) create_lockrows_path(root, final_rel, path,
2178 : : root->rowMarks,
2179 : : assign_special_exec_param(root));
2180 : : }
2181 : :
2182 : : /*
2183 : : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
2184 : : */
2185 [ + + ]: 408149 : if (limit_needed(parse))
2186 : : {
2187 : 4208 : path = (Path *) create_limit_path(root, final_rel, path,
2188 : : parse->limitOffset,
2189 : : parse->limitCount,
2190 : : parse->limitOption,
2191 : : offset_est, count_est);
2192 : : }
2193 : :
2194 : : /*
2195 : : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
2196 : : */
2197 [ + + ]: 408149 : if (parse->commandType != CMD_SELECT)
2198 : : {
2199 : : Index rootRelation;
2200 : 62573 : List *resultRelations = NIL;
2201 : 62573 : List *updateColnosLists = NIL;
2202 : 62573 : List *withCheckOptionLists = NIL;
2203 : 62573 : List *returningLists = NIL;
2204 : 62573 : List *mergeActionLists = NIL;
2205 : 62573 : List *mergeJoinConditions = NIL;
2206 : : List *rowMarks;
2207 : :
2208 [ + + ]: 62573 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
2209 : : {
2210 : : /* Inherited UPDATE/DELETE/MERGE */
2211 : 2327 : RelOptInfo *top_result_rel = find_base_rel(root,
2212 : : parse->resultRelation);
2213 : 2327 : int resultRelation = -1;
2214 : :
2215 : : /* Pass the root result rel forward to the executor. */
2216 : 2327 : rootRelation = parse->resultRelation;
2217 : :
2218 : : /* Add only leaf children to ModifyTable. */
2219 : 6821 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
2220 [ + + ]: 6821 : resultRelation)) >= 0)
2221 : : {
2222 : 4494 : RelOptInfo *this_result_rel = find_base_rel(root,
2223 : : resultRelation);
2224 : :
2225 : : /*
2226 : : * Also exclude any leaf rels that have turned dummy since
2227 : : * being added to the list, for example, by being excluded
2228 : : * by constraint exclusion.
2229 : : */
2230 [ + + ]: 4494 : if (IS_DUMMY_REL(this_result_rel))
2231 : 153 : continue;
2232 : :
2233 : : /* Build per-target-rel lists needed by ModifyTable */
2234 : 4341 : resultRelations = lappend_int(resultRelations,
2235 : : resultRelation);
2236 [ + + ]: 4341 : if (parse->commandType == CMD_UPDATE)
2237 : : {
2238 : 2968 : List *update_colnos = root->update_colnos;
2239 : :
2240 [ + - ]: 2968 : if (this_result_rel != top_result_rel)
2241 : : update_colnos =
2242 : 2968 : adjust_inherited_attnums_multilevel(root,
2243 : : update_colnos,
2244 : : this_result_rel->relid,
2245 : : top_result_rel->relid);
2246 : 2968 : updateColnosLists = lappend(updateColnosLists,
2247 : : update_colnos);
2248 : : }
2249 [ + + ]: 4341 : if (parse->withCheckOptions)
2250 : : {
2251 : 416 : List *withCheckOptions = parse->withCheckOptions;
2252 : :
2253 [ + - ]: 416 : if (this_result_rel != top_result_rel)
2254 : : withCheckOptions = (List *)
2255 : 416 : adjust_appendrel_attrs_multilevel(root,
2256 : : (Node *) withCheckOptions,
2257 : : this_result_rel,
2258 : : top_result_rel);
2259 : 416 : withCheckOptionLists = lappend(withCheckOptionLists,
2260 : : withCheckOptions);
2261 : : }
2262 [ + + ]: 4341 : if (parse->returningList)
2263 : : {
2264 : 669 : List *returningList = parse->returningList;
2265 : :
2266 [ + - ]: 669 : if (this_result_rel != top_result_rel)
2267 : : returningList = (List *)
2268 : 669 : adjust_appendrel_attrs_multilevel(root,
2269 : : (Node *) returningList,
2270 : : this_result_rel,
2271 : : top_result_rel);
2272 : 669 : returningLists = lappend(returningLists,
2273 : : returningList);
2274 : : }
2275 [ + + ]: 4341 : if (parse->mergeActionList)
2276 : : {
2277 : : ListCell *l;
2278 : 427 : List *mergeActionList = NIL;
2279 : :
2280 : : /*
2281 : : * Copy MergeActions and translate stuff that
2282 : : * references attribute numbers.
2283 : : */
2284 [ + - + + : 1340 : foreach(l, parse->mergeActionList)
+ + ]
2285 : : {
2286 : 913 : MergeAction *action = lfirst(l),
2287 : 913 : *leaf_action = copyObject(action);
2288 : :
2289 : 913 : leaf_action->qual =
2290 : 913 : adjust_appendrel_attrs_multilevel(root,
2291 : : (Node *) action->qual,
2292 : : this_result_rel,
2293 : : top_result_rel);
2294 : 913 : leaf_action->targetList = (List *)
2295 : 913 : adjust_appendrel_attrs_multilevel(root,
2296 : 913 : (Node *) action->targetList,
2297 : : this_result_rel,
2298 : : top_result_rel);
2299 [ + + ]: 913 : if (leaf_action->commandType == CMD_UPDATE)
2300 : 497 : leaf_action->updateColnos =
2301 : 497 : adjust_inherited_attnums_multilevel(root,
2302 : : action->updateColnos,
2303 : : this_result_rel->relid,
2304 : : top_result_rel->relid);
2305 : 913 : mergeActionList = lappend(mergeActionList,
2306 : : leaf_action);
2307 : : }
2308 : :
2309 : 427 : mergeActionLists = lappend(mergeActionLists,
2310 : : mergeActionList);
2311 : : }
2312 [ + + ]: 4341 : if (parse->commandType == CMD_MERGE)
2313 : : {
2314 : 427 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2315 : :
2316 [ + - ]: 427 : if (this_result_rel != top_result_rel)
2317 : : mergeJoinCondition =
2318 : 427 : adjust_appendrel_attrs_multilevel(root,
2319 : : mergeJoinCondition,
2320 : : this_result_rel,
2321 : : top_result_rel);
2322 : 427 : mergeJoinConditions = lappend(mergeJoinConditions,
2323 : : mergeJoinCondition);
2324 : : }
2325 : : }
2326 : :
2327 [ + + ]: 2327 : if (resultRelations == NIL)
2328 : : {
2329 : : /*
2330 : : * We managed to exclude every child rel, so generate a
2331 : : * dummy one-relation plan using info for the top target
2332 : : * rel (even though that may not be a leaf target).
2333 : : * Although it's clear that no data will be updated or
2334 : : * deleted, we still need to have a ModifyTable node so
2335 : : * that any statement triggers will be executed. (This
2336 : : * could be cleaner if we fixed nodeModifyTable.c to allow
2337 : : * zero target relations, but that probably wouldn't be a
2338 : : * net win.)
2339 : : */
2340 : 28 : resultRelations = list_make1_int(parse->resultRelation);
2341 [ + + ]: 28 : if (parse->commandType == CMD_UPDATE)
2342 : 26 : updateColnosLists = list_make1(root->update_colnos);
2343 [ - + ]: 28 : if (parse->withCheckOptions)
2344 : 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2345 [ + + ]: 28 : if (parse->returningList)
2346 : 15 : returningLists = list_make1(parse->returningList);
2347 [ + + ]: 28 : if (parse->mergeActionList)
2348 : 1 : mergeActionLists = list_make1(parse->mergeActionList);
2349 [ + + ]: 28 : if (parse->commandType == CMD_MERGE)
2350 : 1 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2351 : : }
2352 : : }
2353 : : else
2354 : : {
2355 : : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
2356 : 60246 : rootRelation = 0; /* there's no separate root rel */
2357 : 60246 : resultRelations = list_make1_int(parse->resultRelation);
2358 [ + + ]: 60246 : if (parse->commandType == CMD_UPDATE)
2359 : 8849 : updateColnosLists = list_make1(root->update_colnos);
2360 [ + + ]: 60246 : if (parse->withCheckOptions)
2361 : 889 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2362 [ + + ]: 60246 : if (parse->returningList)
2363 : 2216 : returningLists = list_make1(parse->returningList);
2364 [ + + ]: 60246 : if (parse->mergeActionList)
2365 : 1289 : mergeActionLists = list_make1(parse->mergeActionList);
2366 [ + + ]: 60246 : if (parse->commandType == CMD_MERGE)
2367 : 1289 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2368 : : }
2369 : :
2370 : : /*
2371 : : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2372 : : * will have dealt with fetching non-locked marked rows, else we
2373 : : * need to have ModifyTable do that.
2374 : : */
2375 [ - + ]: 62573 : if (parse->rowMarks)
2376 : 0 : rowMarks = NIL;
2377 : : else
2378 : 62573 : rowMarks = root->rowMarks;
2379 : :
2380 : : path = (Path *)
2381 : 62573 : create_modifytable_path(root, final_rel,
2382 : : path,
2383 : : parse->commandType,
2384 : 62573 : parse->canSetTag,
2385 : 62573 : parse->resultRelation,
2386 : : rootRelation,
2387 : : resultRelations,
2388 : : updateColnosLists,
2389 : : withCheckOptionLists,
2390 : : returningLists,
2391 : : rowMarks,
2392 : : parse->onConflict,
2393 : : mergeActionLists,
2394 : : mergeJoinConditions,
2395 : : parse->forPortionOf,
2396 : : assign_special_exec_param(root));
2397 : : }
2398 : :
2399 : : /* And shove it into final_rel */
2400 : 408149 : add_path(final_rel, path);
2401 : : }
2402 : :
2403 : : /*
2404 : : * Generate partial paths for final_rel, too, if outer query levels might
2405 : : * be able to make use of them.
2406 : : */
2407 [ + + + + ]: 391520 : if (final_rel->consider_parallel && root->query_level > 1 &&
2408 [ + + ]: 23310 : !limit_needed(parse))
2409 : : {
2410 : : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2411 [ + + + + : 23244 : foreach(lc, current_rel->partial_pathlist)
+ + ]
2412 : : {
2413 : 112 : Path *partial_path = (Path *) lfirst(lc);
2414 : :
2415 : 112 : add_partial_path(final_rel, partial_path);
2416 : : }
2417 : : }
2418 : :
2419 : 391520 : extra.limit_needed = limit_needed(parse);
2420 : 391520 : extra.limit_tuples = limit_tuples;
2421 : 391520 : extra.count_est = count_est;
2422 : 391520 : extra.offset_est = offset_est;
2423 : :
2424 : : /*
2425 : : * If there is an FDW that's responsible for all baserels of the query,
2426 : : * let it consider adding ForeignPaths.
2427 : : */
2428 [ + + ]: 391520 : if (final_rel->fdwroutine &&
2429 [ + + ]: 662 : final_rel->fdwroutine->GetForeignUpperPaths)
2430 : 625 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2431 : : current_rel, final_rel,
2432 : : &extra);
2433 : :
2434 : : /* Let extensions possibly add some more paths */
2435 [ - + ]: 391520 : if (create_upper_paths_hook)
2436 : 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2437 : : current_rel, final_rel, &extra);
2438 : :
2439 : : /* Note: currently, we leave it to callers to do set_cheapest() */
2440 : 391520 : }
2441 : :
2442 : : /*
2443 : : * Do preprocessing for groupingSets clause and related data.
2444 : : *
2445 : : * We expect that parse->groupingSets has already been expanded into a flat
2446 : : * list of grouping sets (that is, just integer Lists of ressortgroupref
2447 : : * numbers) by expand_grouping_sets(). This function handles the preliminary
2448 : : * steps of organizing the grouping sets into lists of rollups, and preparing
2449 : : * annotations which will later be filled in with size estimates.
2450 : : */
2451 : : static grouping_sets_data *
2452 : 915 : preprocess_grouping_sets(PlannerInfo *root)
2453 : : {
2454 : 915 : Query *parse = root->parse;
2455 : : List *sets;
2456 : 915 : int maxref = 0;
2457 : : ListCell *lc_set;
2458 : 915 : grouping_sets_data *gd = palloc0_object(grouping_sets_data);
2459 : :
2460 : : /*
2461 : : * We don't currently make any attempt to optimize the groupClause when
2462 : : * there are grouping sets, so just duplicate it in processed_groupClause.
2463 : : */
2464 : 915 : root->processed_groupClause = parse->groupClause;
2465 : :
2466 : : /* Detect unhashable and unsortable grouping expressions */
2467 : 915 : gd->any_hashable = false;
2468 : 915 : gd->unhashable_refs = NULL;
2469 : 915 : gd->unsortable_refs = NULL;
2470 : 915 : gd->unsortable_sets = NIL;
2471 : :
2472 [ + + ]: 915 : if (parse->groupClause)
2473 : : {
2474 : : ListCell *lc;
2475 : :
2476 [ + - + + : 2651 : foreach(lc, parse->groupClause)
+ + ]
2477 : : {
2478 : 1811 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2479 : 1811 : Index ref = gc->tleSortGroupRef;
2480 : :
2481 [ + + ]: 1811 : if (ref > maxref)
2482 : 1771 : maxref = ref;
2483 : :
2484 [ + + ]: 1811 : if (!gc->hashable)
2485 : 24 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2486 : :
2487 [ + + ]: 1811 : if (!OidIsValid(gc->sortop))
2488 : 33 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2489 : : }
2490 : : }
2491 : :
2492 : : /* Allocate workspace array for remapping */
2493 : 915 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2494 : :
2495 : : /*
2496 : : * If we have any unsortable sets, we must extract them before trying to
2497 : : * prepare rollups. Unsortable sets don't go through
2498 : : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2499 : : * here.
2500 : : */
2501 [ + + ]: 915 : if (!bms_is_empty(gd->unsortable_refs))
2502 : : {
2503 : 33 : List *sortable_sets = NIL;
2504 : : ListCell *lc;
2505 : :
2506 [ + - + + : 99 : foreach(lc, parse->groupingSets)
+ + ]
2507 : : {
2508 : 70 : List *gset = (List *) lfirst(lc);
2509 : :
2510 [ + + ]: 70 : if (bms_overlap_list(gd->unsortable_refs, gset))
2511 : : {
2512 : 38 : GroupingSetData *gs = makeNode(GroupingSetData);
2513 : :
2514 : 38 : gs->set = gset;
2515 : 38 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2516 : :
2517 : : /*
2518 : : * We must enforce here that an unsortable set is hashable;
2519 : : * later code assumes this. Parse analysis only checks that
2520 : : * every individual column is either hashable or sortable.
2521 : : *
2522 : : * Note that passing this test doesn't guarantee we can
2523 : : * generate a plan; there might be other showstoppers.
2524 : : */
2525 [ + + ]: 38 : if (bms_overlap_list(gd->unhashable_refs, gset))
2526 [ + - ]: 4 : ereport(ERROR,
2527 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2528 : : errmsg("could not implement GROUP BY"),
2529 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2530 : : }
2531 : : else
2532 : 32 : sortable_sets = lappend(sortable_sets, gset);
2533 : : }
2534 : :
2535 [ + + ]: 29 : if (sortable_sets)
2536 : 24 : sets = extract_rollup_sets(sortable_sets);
2537 : : else
2538 : 5 : sets = NIL;
2539 : : }
2540 : : else
2541 : 882 : sets = extract_rollup_sets(parse->groupingSets);
2542 : :
2543 [ + + + + : 2396 : foreach(lc_set, sets)
+ + ]
2544 : : {
2545 : 1485 : List *current_sets = (List *) lfirst(lc_set);
2546 : 1485 : RollupData *rollup = makeNode(RollupData);
2547 : : GroupingSetData *gs;
2548 : :
2549 : : /*
2550 : : * Reorder the current list of grouping sets into correct prefix
2551 : : * order. If only one aggregation pass is needed, try to make the
2552 : : * list match the ORDER BY clause; if more than one pass is needed, we
2553 : : * don't bother with that.
2554 : : *
2555 : : * Note that this reorders the sets from smallest-member-first to
2556 : : * largest-member-first, and applies the GroupingSetData annotations,
2557 : : * though the data will be filled in later.
2558 : : */
2559 [ + + ]: 1485 : current_sets = reorder_grouping_sets(current_sets,
2560 : 1485 : (list_length(sets) == 1
2561 : : ? parse->sortClause
2562 : : : NIL));
2563 : :
2564 : : /*
2565 : : * Get the initial (and therefore largest) grouping set.
2566 : : */
2567 : 1485 : gs = linitial_node(GroupingSetData, current_sets);
2568 : :
2569 : : /*
2570 : : * Order the groupClause appropriately. If the first grouping set is
2571 : : * empty, then the groupClause must also be empty; otherwise we have
2572 : : * to force the groupClause to match that grouping set's order.
2573 : : *
2574 : : * (The first grouping set can be empty even though parse->groupClause
2575 : : * is not empty only if all non-empty grouping sets are unsortable.
2576 : : * The groupClauses for hashed grouping sets are built later on.)
2577 : : */
2578 [ + + ]: 1485 : if (gs->set)
2579 : 1410 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2580 : : else
2581 : 75 : rollup->groupClause = NIL;
2582 : :
2583 : : /*
2584 : : * Is it hashable? We pretend empty sets are hashable even though we
2585 : : * actually force them not to be hashed later. But don't bother if
2586 : : * there's nothing but empty sets (since in that case we can't hash
2587 : : * anything).
2588 : : */
2589 [ + + ]: 1485 : if (gs->set &&
2590 [ + + ]: 1410 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2591 : : {
2592 : 1390 : rollup->hashable = true;
2593 : 1390 : gd->any_hashable = true;
2594 : : }
2595 : :
2596 : : /*
2597 : : * Now that we've pinned down an order for the groupClause for this
2598 : : * list of grouping sets, we need to remap the entries in the grouping
2599 : : * sets from sortgrouprefs to plain indices (0-based) into the
2600 : : * groupClause for this collection of grouping sets. We keep the
2601 : : * original form for later use, though.
2602 : : */
2603 : 1485 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2604 : : current_sets,
2605 : : gd->tleref_to_colnum_map);
2606 : 1485 : rollup->gsets_data = current_sets;
2607 : :
2608 : 1485 : gd->rollups = lappend(gd->rollups, rollup);
2609 : : }
2610 : :
2611 [ + + ]: 911 : if (gd->unsortable_sets)
2612 : : {
2613 : : /*
2614 : : * We have not yet pinned down a groupclause for this, but we will
2615 : : * need index-based lists for estimation purposes. Construct
2616 : : * hash_sets_idx based on the entire original groupclause for now.
2617 : : */
2618 : 29 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2619 : : gd->unsortable_sets,
2620 : : gd->tleref_to_colnum_map);
2621 : 29 : gd->any_hashable = true;
2622 : : }
2623 : :
2624 : 911 : return gd;
2625 : : }
2626 : :
2627 : : /*
2628 : : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2629 : : * (without annotation) mapped to indexes into the given groupclause.
2630 : : */
2631 : : static List *
2632 : 4157 : remap_to_groupclause_idx(List *groupClause,
2633 : : List *gsets,
2634 : : int *tleref_to_colnum_map)
2635 : : {
2636 : 4157 : int ref = 0;
2637 : 4157 : List *result = NIL;
2638 : : ListCell *lc;
2639 : :
2640 [ + + + + : 9871 : foreach(lc, groupClause)
+ + ]
2641 : : {
2642 : 5714 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2643 : :
2644 : 5714 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2645 : : }
2646 : :
2647 [ + - + + : 9464 : foreach(lc, gsets)
+ + ]
2648 : : {
2649 : 5307 : List *set = NIL;
2650 : : ListCell *lc2;
2651 : 5307 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2652 : :
2653 [ + + + + : 11730 : foreach(lc2, gs->set)
+ + ]
2654 : : {
2655 : 6423 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2656 : : }
2657 : :
2658 : 5307 : result = lappend(result, set);
2659 : : }
2660 : :
2661 : 4157 : return result;
2662 : : }
2663 : :
2664 : :
2665 : : /*
2666 : : * preprocess_rowmarks - set up PlanRowMarks if needed
2667 : : */
2668 : : static void
2669 : 394256 : preprocess_rowmarks(PlannerInfo *root)
2670 : : {
2671 : 394256 : Query *parse = root->parse;
2672 : : Bitmapset *rels;
2673 : : List *prowmarks;
2674 : : ListCell *l;
2675 : : int i;
2676 : :
2677 [ + + ]: 394256 : if (parse->rowMarks)
2678 : : {
2679 : : /*
2680 : : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2681 : : * grouping, since grouping renders a reference to individual tuple
2682 : : * CTIDs invalid. This is also checked at parse time, but that's
2683 : : * insufficient because of rule substitution, query pullup, etc.
2684 : : */
2685 : 6572 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2686 : : parse->rowMarks)->strength);
2687 : : }
2688 : : else
2689 : : {
2690 : : /*
2691 : : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2692 : : * UPDATE/SHARE.
2693 : : */
2694 [ + + ]: 387684 : if (parse->commandType != CMD_UPDATE &&
2695 [ + + ]: 377211 : parse->commandType != CMD_DELETE &&
2696 [ + + ]: 373456 : parse->commandType != CMD_MERGE)
2697 : 371972 : return;
2698 : : }
2699 : :
2700 : : /*
2701 : : * We need to have rowmarks for all base relations except the target. We
2702 : : * make a bitmapset of all base rels and then remove the items we don't
2703 : : * need or have FOR [KEY] UPDATE/SHARE marks for.
2704 : : */
2705 : 22284 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
2706 [ + + ]: 22284 : if (parse->resultRelation)
2707 : 15712 : rels = bms_del_member(rels, parse->resultRelation);
2708 : :
2709 : : /*
2710 : : * Convert RowMarkClauses to PlanRowMark representation.
2711 : : */
2712 : 22284 : prowmarks = NIL;
2713 [ + + + + : 28991 : foreach(l, parse->rowMarks)
+ + ]
2714 : : {
2715 : 6707 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
2716 : 6707 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2717 : : PlanRowMark *newrc;
2718 : :
2719 : : /*
2720 : : * Currently, it is syntactically impossible to have FOR UPDATE et al
2721 : : * applied to an update/delete target rel. If that ever becomes
2722 : : * possible, we should drop the target from the PlanRowMark list.
2723 : : */
2724 : : Assert(rc->rti != parse->resultRelation);
2725 : :
2726 : : /*
2727 : : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2728 : : * can't support true locking. Subqueries that got flattened into the
2729 : : * main query should be ignored completely. Any that didn't will get
2730 : : * ROW_MARK_COPY items in the next loop.
2731 : : */
2732 [ + + ]: 6707 : if (rte->rtekind != RTE_RELATION)
2733 : 46 : continue;
2734 : :
2735 : 6661 : rels = bms_del_member(rels, rc->rti);
2736 : :
2737 : 6661 : newrc = makeNode(PlanRowMark);
2738 : 6661 : newrc->rti = newrc->prti = rc->rti;
2739 : 6661 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2740 : 6661 : newrc->markType = select_rowmark_type(rte, rc->strength);
2741 : 6661 : newrc->allMarkTypes = (1 << newrc->markType);
2742 : 6661 : newrc->strength = rc->strength;
2743 : 6661 : newrc->waitPolicy = rc->waitPolicy;
2744 : 6661 : newrc->isParent = false;
2745 : :
2746 : 6661 : prowmarks = lappend(prowmarks, newrc);
2747 : : }
2748 : :
2749 : : /*
2750 : : * Now, add rowmarks for any non-target, non-locked base relations.
2751 : : */
2752 : 22284 : i = 0;
2753 [ + - + + : 53188 : foreach(l, parse->rtable)
+ + ]
2754 : : {
2755 : 30904 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2756 : : PlanRowMark *newrc;
2757 : :
2758 : 30904 : i++;
2759 [ + + ]: 30904 : if (!bms_is_member(i, rels))
2760 : 27951 : continue;
2761 : :
2762 : 2953 : newrc = makeNode(PlanRowMark);
2763 : 2953 : newrc->rti = newrc->prti = i;
2764 : 2953 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
2765 : 2953 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
2766 : 2953 : newrc->allMarkTypes = (1 << newrc->markType);
2767 : 2953 : newrc->strength = LCS_NONE;
2768 : 2953 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
2769 : 2953 : newrc->isParent = false;
2770 : :
2771 : 2953 : prowmarks = lappend(prowmarks, newrc);
2772 : : }
2773 : :
2774 : 22284 : root->rowMarks = prowmarks;
2775 : : }
2776 : :
2777 : : /*
2778 : : * Select RowMarkType to use for a given table
2779 : : */
2780 : : RowMarkType
2781 : 11508 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2782 : : {
2783 [ + + ]: 11508 : if (rte->rtekind != RTE_RELATION)
2784 : : {
2785 : : /* If it's not a table at all, use ROW_MARK_COPY */
2786 : 1114 : return ROW_MARK_COPY;
2787 : : }
2788 [ + + ]: 10394 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2789 : : {
2790 : : /* Let the FDW select the rowmark type, if it wants to */
2791 : 114 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2792 : :
2793 [ - + ]: 114 : if (fdwroutine->GetForeignRowMarkType != NULL)
2794 : 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2795 : : /* Otherwise, use ROW_MARK_COPY by default */
2796 : 114 : return ROW_MARK_COPY;
2797 : : }
2798 : : else
2799 : : {
2800 : : /* Regular table, apply the appropriate lock type */
2801 [ + + + + : 10280 : switch (strength)
+ - ]
2802 : : {
2803 : 2048 : case LCS_NONE:
2804 : :
2805 : : /*
2806 : : * We don't need a tuple lock, only the ability to re-fetch
2807 : : * the row.
2808 : : */
2809 : 2048 : return ROW_MARK_REFERENCE;
2810 : : break;
2811 : 7057 : case LCS_FORKEYSHARE:
2812 : 7057 : return ROW_MARK_KEYSHARE;
2813 : : break;
2814 : 201 : case LCS_FORSHARE:
2815 : 201 : return ROW_MARK_SHARE;
2816 : : break;
2817 : 41 : case LCS_FORNOKEYUPDATE:
2818 : 41 : return ROW_MARK_NOKEYEXCLUSIVE;
2819 : : break;
2820 : 933 : case LCS_FORUPDATE:
2821 : 933 : return ROW_MARK_EXCLUSIVE;
2822 : : break;
2823 : : }
2824 [ # # ]: 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2825 : : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2826 : : }
2827 : : }
2828 : :
2829 : : /*
2830 : : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2831 : : *
2832 : : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2833 : : * results back in *count_est and *offset_est. These variables are set to
2834 : : * 0 if the corresponding clause is not present, and -1 if it's present
2835 : : * but we couldn't estimate the value for it. (The "0" convention is OK
2836 : : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2837 : : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2838 : : * usual practice of never estimating less than one row.) These values will
2839 : : * be passed to create_limit_path, which see if you change this code.
2840 : : *
2841 : : * The return value is the suitably adjusted tuple_fraction to use for
2842 : : * planning the query. This adjustment is not overridable, since it reflects
2843 : : * plan actions that grouping_planner() will certainly take, not assumptions
2844 : : * about context.
2845 : : */
2846 : : static double
2847 : 3737 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2848 : : int64 *offset_est, int64 *count_est)
2849 : : {
2850 : 3737 : Query *parse = root->parse;
2851 : : Node *est;
2852 : : double limit_fraction;
2853 : :
2854 : : /* Should not be called unless LIMIT or OFFSET */
2855 : : Assert(parse->limitCount || parse->limitOffset);
2856 : :
2857 : : /*
2858 : : * Try to obtain the clause values. We use estimate_expression_value
2859 : : * primarily because it can sometimes do something useful with Params.
2860 : : */
2861 [ + + ]: 3737 : if (parse->limitCount)
2862 : : {
2863 : 3275 : est = estimate_expression_value(root, parse->limitCount);
2864 [ + - + + ]: 3275 : if (est && IsA(est, Const))
2865 : : {
2866 [ - + ]: 3270 : if (((Const *) est)->constisnull)
2867 : : {
2868 : : /* NULL indicates LIMIT ALL, ie, no limit */
2869 : 0 : *count_est = 0; /* treat as not present */
2870 : : }
2871 : : else
2872 : : {
2873 : 3270 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
2874 [ + + ]: 3270 : if (*count_est <= 0)
2875 : 190 : *count_est = 1; /* force to at least 1 */
2876 : : }
2877 : : }
2878 : : else
2879 : 5 : *count_est = -1; /* can't estimate */
2880 : : }
2881 : : else
2882 : 462 : *count_est = 0; /* not present */
2883 : :
2884 [ + + ]: 3737 : if (parse->limitOffset)
2885 : : {
2886 : 672 : est = estimate_expression_value(root, parse->limitOffset);
2887 [ + - + + ]: 672 : if (est && IsA(est, Const))
2888 : : {
2889 [ - + ]: 652 : if (((Const *) est)->constisnull)
2890 : : {
2891 : : /* Treat NULL as no offset; the executor will too */
2892 : 0 : *offset_est = 0; /* treat as not present */
2893 : : }
2894 : : else
2895 : : {
2896 : 652 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
2897 [ - + ]: 652 : if (*offset_est < 0)
2898 : 0 : *offset_est = 0; /* treat as not present */
2899 : : }
2900 : : }
2901 : : else
2902 : 20 : *offset_est = -1; /* can't estimate */
2903 : : }
2904 : : else
2905 : 3065 : *offset_est = 0; /* not present */
2906 : :
2907 [ + + ]: 3737 : if (*count_est != 0)
2908 : : {
2909 : : /*
2910 : : * A LIMIT clause limits the absolute number of tuples returned.
2911 : : * However, if it's not a constant LIMIT then we have to guess; for
2912 : : * lack of a better idea, assume 10% of the plan's result is wanted.
2913 : : */
2914 [ + + + + ]: 3275 : if (*count_est < 0 || *offset_est < 0)
2915 : : {
2916 : : /* LIMIT or OFFSET is an expression ... punt ... */
2917 : 20 : limit_fraction = 0.10;
2918 : : }
2919 : : else
2920 : : {
2921 : : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2922 : 3255 : limit_fraction = (double) *count_est + (double) *offset_est;
2923 : : }
2924 : :
2925 : : /*
2926 : : * If we have absolute limits from both caller and LIMIT, use the
2927 : : * smaller value; likewise if they are both fractional. If one is
2928 : : * fractional and the other absolute, we can't easily determine which
2929 : : * is smaller, but we use the heuristic that the absolute will usually
2930 : : * be smaller.
2931 : : */
2932 [ + + ]: 3275 : if (tuple_fraction >= 1.0)
2933 : : {
2934 [ + - ]: 5 : if (limit_fraction >= 1.0)
2935 : : {
2936 : : /* both absolute */
2937 [ - + ]: 5 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2938 : : }
2939 : : else
2940 : : {
2941 : : /* caller absolute, limit fractional; use caller's value */
2942 : : }
2943 : : }
2944 [ + + ]: 3270 : else if (tuple_fraction > 0.0)
2945 : : {
2946 [ + - ]: 82 : if (limit_fraction >= 1.0)
2947 : : {
2948 : : /* caller fractional, limit absolute; use limit */
2949 : 82 : tuple_fraction = limit_fraction;
2950 : : }
2951 : : else
2952 : : {
2953 : : /* both fractional */
2954 [ # # ]: 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2955 : : }
2956 : : }
2957 : : else
2958 : : {
2959 : : /* no info from caller, just use limit */
2960 : 3188 : tuple_fraction = limit_fraction;
2961 : : }
2962 : : }
2963 [ + + + + ]: 462 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2964 : : {
2965 : : /*
2966 : : * We have an OFFSET but no LIMIT. This acts entirely differently
2967 : : * from the LIMIT case: here, we need to increase rather than decrease
2968 : : * the caller's tuple_fraction, because the OFFSET acts to cause more
2969 : : * tuples to be fetched instead of fewer. This only matters if we got
2970 : : * a tuple_fraction > 0, however.
2971 : : *
2972 : : * As above, use 10% if OFFSET is present but unestimatable.
2973 : : */
2974 [ - + ]: 12 : if (*offset_est < 0)
2975 : 0 : limit_fraction = 0.10;
2976 : : else
2977 : 12 : limit_fraction = (double) *offset_est;
2978 : :
2979 : : /*
2980 : : * If we have absolute counts from both caller and OFFSET, add them
2981 : : * together; likewise if they are both fractional. If one is
2982 : : * fractional and the other absolute, we want to take the larger, and
2983 : : * we heuristically assume that's the fractional one.
2984 : : */
2985 [ - + ]: 12 : if (tuple_fraction >= 1.0)
2986 : : {
2987 [ # # ]: 0 : if (limit_fraction >= 1.0)
2988 : : {
2989 : : /* both absolute, so add them together */
2990 : 0 : tuple_fraction += limit_fraction;
2991 : : }
2992 : : else
2993 : : {
2994 : : /* caller absolute, limit fractional; use limit */
2995 : 0 : tuple_fraction = limit_fraction;
2996 : : }
2997 : : }
2998 : : else
2999 : : {
3000 [ - + ]: 12 : if (limit_fraction >= 1.0)
3001 : : {
3002 : : /* caller fractional, limit absolute; use caller's value */
3003 : : }
3004 : : else
3005 : : {
3006 : : /* both fractional, so add them together */
3007 : 0 : tuple_fraction += limit_fraction;
3008 [ # # ]: 0 : if (tuple_fraction >= 1.0)
3009 : 0 : tuple_fraction = 0.0; /* assume fetch all */
3010 : : }
3011 : : }
3012 : : }
3013 : :
3014 : 3737 : return tuple_fraction;
3015 : : }
3016 : :
3017 : : /*
3018 : : * limit_needed - do we actually need a Limit plan node?
3019 : : *
3020 : : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
3021 : : * a Limit node. This is worth checking for because "OFFSET 0" is a common
3022 : : * locution for an optimization fence. (Because other places in the planner
3023 : : * merely check whether parse->limitOffset isn't NULL, it will still work as
3024 : : * an optimization fence --- we're just suppressing unnecessary run-time
3025 : : * overhead.)
3026 : : *
3027 : : * This might look like it could be merged into preprocess_limit, but there's
3028 : : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
3029 : : * in preprocess_limit it's good enough to consider estimated values.
3030 : : */
3031 : : bool
3032 : 838785 : limit_needed(Query *parse)
3033 : : {
3034 : : Node *node;
3035 : :
3036 : 838785 : node = parse->limitCount;
3037 [ + + ]: 838785 : if (node)
3038 : : {
3039 [ + + ]: 7801 : if (IsA(node, Const))
3040 : : {
3041 : : /* NULL indicates LIMIT ALL, ie, no limit */
3042 [ + - ]: 7580 : if (!((Const *) node)->constisnull)
3043 : 7580 : return true; /* LIMIT with a constant value */
3044 : : }
3045 : : else
3046 : 221 : return true; /* non-constant LIMIT */
3047 : : }
3048 : :
3049 : 830984 : node = parse->limitOffset;
3050 [ + + ]: 830984 : if (node)
3051 : : {
3052 [ + + ]: 1403 : if (IsA(node, Const))
3053 : : {
3054 : : /* Treat NULL as no offset; the executor would too */
3055 [ + - ]: 1125 : if (!((Const *) node)->constisnull)
3056 : : {
3057 : 1125 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
3058 : :
3059 [ + + ]: 1125 : if (offset != 0)
3060 : 107 : return true; /* OFFSET with a nonzero value */
3061 : : }
3062 : : }
3063 : : else
3064 : 278 : return true; /* non-constant OFFSET */
3065 : : }
3066 : :
3067 : 830599 : return false; /* don't need a Limit plan node */
3068 : : }
3069 : :
3070 : : /*
3071 : : * preprocess_groupclause - do preparatory work on GROUP BY clause
3072 : : *
3073 : : * The idea here is to adjust the ordering of the GROUP BY elements
3074 : : * (which in itself is semantically insignificant) to match ORDER BY,
3075 : : * thereby allowing a single sort operation to both implement the ORDER BY
3076 : : * requirement and set up for a Unique step that implements GROUP BY.
3077 : : * We also consider partial match between GROUP BY and ORDER BY elements,
3078 : : * which could allow to implement ORDER BY using the incremental sort.
3079 : : *
3080 : : * We also consider other orderings of the GROUP BY elements, which could
3081 : : * match the sort ordering of other possible plans (eg an indexscan) and
3082 : : * thereby reduce cost. This is implemented during the generation of grouping
3083 : : * paths. See get_useful_group_keys_orderings() for details.
3084 : : *
3085 : : * Note: we need no comparable processing of the distinctClause because
3086 : : * the parser already enforced that that matches ORDER BY.
3087 : : *
3088 : : * Note: we return a fresh List, but its elements are the same
3089 : : * SortGroupClauses appearing in parse->groupClause. This is important
3090 : : * because later processing may modify the processed_groupClause list.
3091 : : *
3092 : : * For grouping sets, the order of items is instead forced to agree with that
3093 : : * of the grouping set (and items not in the grouping set are skipped). The
3094 : : * work of sorting the order of grouping set elements to match the ORDER BY if
3095 : : * possible is done elsewhere.
3096 : : */
3097 : : static List *
3098 : 7626 : preprocess_groupclause(PlannerInfo *root, List *force)
3099 : : {
3100 : 7626 : Query *parse = root->parse;
3101 : 7626 : List *new_groupclause = NIL;
3102 : : ListCell *sl;
3103 : : ListCell *gl;
3104 : :
3105 : : /* For grouping sets, we need to force the ordering */
3106 [ + + ]: 7626 : if (force)
3107 : : {
3108 [ + - + + : 9704 : foreach(sl, force)
+ + ]
3109 : : {
3110 : 5651 : Index ref = lfirst_int(sl);
3111 : 5651 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
3112 : :
3113 : 5651 : new_groupclause = lappend(new_groupclause, cl);
3114 : : }
3115 : :
3116 : 4053 : return new_groupclause;
3117 : : }
3118 : :
3119 : : /* If no ORDER BY, nothing useful to do here */
3120 [ + + ]: 3573 : if (parse->sortClause == NIL)
3121 : 2044 : return list_copy(parse->groupClause);
3122 : :
3123 : : /*
3124 : : * Scan the ORDER BY clause and construct a list of matching GROUP BY
3125 : : * items, but only as far as we can make a matching prefix.
3126 : : *
3127 : : * This code assumes that the sortClause contains no duplicate items.
3128 : : */
3129 [ + - + + : 3005 : foreach(sl, parse->sortClause)
+ + ]
3130 : : {
3131 : 1997 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
3132 : :
3133 [ + - + + : 2909 : foreach(gl, parse->groupClause)
+ + ]
3134 : : {
3135 : 2388 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3136 : :
3137 [ + + ]: 2388 : if (equal(gc, sc))
3138 : : {
3139 : 1476 : new_groupclause = lappend(new_groupclause, gc);
3140 : 1476 : break;
3141 : : }
3142 : : }
3143 [ + + ]: 1997 : if (gl == NULL)
3144 : 521 : break; /* no match, so stop scanning */
3145 : : }
3146 : :
3147 : :
3148 : : /* If no match at all, no point in reordering GROUP BY */
3149 [ + + ]: 1529 : if (new_groupclause == NIL)
3150 : 233 : return list_copy(parse->groupClause);
3151 : :
3152 : : /*
3153 : : * Add any remaining GROUP BY items to the new list. We don't require a
3154 : : * complete match, because even partial match allows ORDER BY to be
3155 : : * implemented using incremental sort. Also, give up if there are any
3156 : : * non-sortable GROUP BY items, since then there's no hope anyway.
3157 : : */
3158 [ + - + + : 2909 : foreach(gl, parse->groupClause)
+ + ]
3159 : : {
3160 : 1613 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
3161 : :
3162 [ + + ]: 1613 : if (list_member_ptr(new_groupclause, gc))
3163 : 1476 : continue; /* it matched an ORDER BY item */
3164 [ - + ]: 137 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
3165 : 0 : return list_copy(parse->groupClause);
3166 : 137 : new_groupclause = lappend(new_groupclause, gc);
3167 : : }
3168 : :
3169 : : /* Success --- install the rearranged GROUP BY list */
3170 : : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3171 : 1296 : return new_groupclause;
3172 : : }
3173 : :
3174 : : /*
3175 : : * Extract lists of grouping sets that can be implemented using a single
3176 : : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
3177 : : *
3178 : : * Input must be sorted with smallest sets first. Result has each sublist
3179 : : * sorted with smallest sets first.
3180 : : *
3181 : : * We want to produce the absolute minimum possible number of lists here to
3182 : : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
3183 : : * of finding the minimal partition of a partially-ordered set into chains
3184 : : * (which is what we need, taking the list of grouping sets as a poset ordered
3185 : : * by set inclusion) can be mapped to the problem of finding the maximum
3186 : : * cardinality matching on a bipartite graph, which is solvable in polynomial
3187 : : * time with a worst case of no worse than O(n^2.5) and usually much
3188 : : * better. Since our N is at most 4096, we don't need to consider fallbacks to
3189 : : * heuristic or approximate methods. (Planning time for a 12-d cube is under
3190 : : * half a second on my modest system even with optimization off and assertions
3191 : : * on.)
3192 : : */
3193 : : static List *
3194 : 906 : extract_rollup_sets(List *groupingSets)
3195 : : {
3196 : 906 : int num_sets_raw = list_length(groupingSets);
3197 : 906 : int num_empty = 0;
3198 : 906 : int num_sets = 0; /* distinct sets */
3199 : 906 : int num_chains = 0;
3200 : 906 : List *result = NIL;
3201 : : List **results;
3202 : : List **orig_sets;
3203 : : Bitmapset **set_masks;
3204 : : int *chains;
3205 : : short **adjacency;
3206 : : short *adjacency_buf;
3207 : : BipartiteMatchState *state;
3208 : : int i;
3209 : : int j;
3210 : : int j_size;
3211 : 906 : ListCell *lc1 = list_head(groupingSets);
3212 : : ListCell *lc;
3213 : :
3214 : : /*
3215 : : * Start by stripping out empty sets. The algorithm doesn't require this,
3216 : : * but the planner currently needs all empty sets to be returned in the
3217 : : * first list, so we strip them here and add them back after.
3218 : : */
3219 [ + + + + ]: 1506 : while (lc1 && lfirst(lc1) == NIL)
3220 : : {
3221 : 600 : ++num_empty;
3222 : 600 : lc1 = lnext(groupingSets, lc1);
3223 : : }
3224 : :
3225 : : /* bail out now if it turns out that all we had were empty sets. */
3226 [ + + ]: 906 : if (!lc1)
3227 : 75 : return list_make1(groupingSets);
3228 : :
3229 : : /*----------
3230 : : * We don't strictly need to remove duplicate sets here, but if we don't,
3231 : : * they tend to become scattered through the result, which is a bit
3232 : : * confusing (and irritating if we ever decide to optimize them out).
3233 : : * So we remove them here and add them back after.
3234 : : *
3235 : : * For each non-duplicate set, we fill in the following:
3236 : : *
3237 : : * orig_sets[i] = list of the original set lists
3238 : : * set_masks[i] = bitmapset for testing inclusion
3239 : : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
3240 : : *
3241 : : * chains[i] will be the result group this set is assigned to.
3242 : : *
3243 : : * We index all of these from 1 rather than 0 because it is convenient
3244 : : * to leave 0 free for the NIL node in the graph algorithm.
3245 : : *----------
3246 : : */
3247 : 831 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3248 : 831 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3249 : 831 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3250 : 831 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3251 : :
3252 : 831 : j_size = 0;
3253 : 831 : j = 0;
3254 : 831 : i = 1;
3255 : :
3256 [ + - + + : 2861 : for_each_cell(lc, groupingSets, lc1)
+ + ]
3257 : : {
3258 : 2030 : List *candidate = (List *) lfirst(lc);
3259 : 2030 : Bitmapset *candidate_set = NULL;
3260 : : ListCell *lc2;
3261 : 2030 : int dup_of = 0;
3262 : :
3263 [ + - + + : 4827 : foreach(lc2, candidate)
+ + ]
3264 : : {
3265 : 2797 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3266 : : }
3267 : :
3268 : : /* we can only be a dup if we're the same length as a previous set */
3269 [ + + ]: 2030 : if (j_size == list_length(candidate))
3270 : : {
3271 : : int k;
3272 : :
3273 [ + + ]: 1828 : for (k = j; k < i; ++k)
3274 : : {
3275 [ + + ]: 1169 : if (bms_equal(set_masks[k], candidate_set))
3276 : : {
3277 : 129 : dup_of = k;
3278 : 129 : break;
3279 : : }
3280 : : }
3281 : : }
3282 [ + - ]: 1242 : else if (j_size < list_length(candidate))
3283 : : {
3284 : 1242 : j_size = list_length(candidate);
3285 : 1242 : j = i;
3286 : : }
3287 : :
3288 [ + + ]: 2030 : if (dup_of > 0)
3289 : : {
3290 : 129 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3291 : 129 : bms_free(candidate_set);
3292 : : }
3293 : : else
3294 : : {
3295 : : int k;
3296 : 1901 : int n_adj = 0;
3297 : :
3298 : 1901 : orig_sets[i] = list_make1(candidate);
3299 : 1901 : set_masks[i] = candidate_set;
3300 : :
3301 : : /* fill in adjacency list; no need to compare equal-size sets */
3302 : :
3303 [ + + ]: 2973 : for (k = j - 1; k > 0; --k)
3304 : : {
3305 [ + + ]: 1072 : if (bms_is_subset(set_masks[k], candidate_set))
3306 : 937 : adjacency_buf[++n_adj] = k;
3307 : : }
3308 : :
3309 [ + + ]: 1901 : if (n_adj > 0)
3310 : : {
3311 : 513 : adjacency_buf[0] = n_adj;
3312 : 513 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3313 : 513 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3314 : : }
3315 : : else
3316 : 1388 : adjacency[i] = NULL;
3317 : :
3318 : 1901 : ++i;
3319 : : }
3320 : : }
3321 : :
3322 : 831 : num_sets = i - 1;
3323 : :
3324 : : /*
3325 : : * Apply the graph matching algorithm to do the work.
3326 : : */
3327 : 831 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3328 : :
3329 : : /*
3330 : : * Now, the state->pair* fields have the info we need to assign sets to
3331 : : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3332 : : * pair_vu[v] = u (both will be true, but we check both so that we can do
3333 : : * it in one pass)
3334 : : */
3335 : 831 : chains = palloc0((num_sets + 1) * sizeof(int));
3336 : :
3337 [ + + ]: 2732 : for (i = 1; i <= num_sets; ++i)
3338 : : {
3339 : 1901 : int u = state->pair_vu[i];
3340 : 1901 : int v = state->pair_uv[i];
3341 : :
3342 [ + + - + ]: 1901 : if (u > 0 && u < i)
3343 : 0 : chains[i] = chains[u];
3344 [ + + + - ]: 1901 : else if (v > 0 && v < i)
3345 : 491 : chains[i] = chains[v];
3346 : : else
3347 : 1410 : chains[i] = ++num_chains;
3348 : : }
3349 : :
3350 : : /* build result lists. */
3351 : 831 : results = palloc0((num_chains + 1) * sizeof(List *));
3352 : :
3353 [ + + ]: 2732 : for (i = 1; i <= num_sets; ++i)
3354 : : {
3355 : 1901 : int c = chains[i];
3356 : :
3357 : : Assert(c > 0);
3358 : :
3359 : 1901 : results[c] = list_concat(results[c], orig_sets[i]);
3360 : : }
3361 : :
3362 : : /* push any empty sets back on the first list. */
3363 [ + + ]: 1306 : while (num_empty-- > 0)
3364 : 475 : results[1] = lcons(NIL, results[1]);
3365 : :
3366 : : /* make result list */
3367 [ + + ]: 2241 : for (i = 1; i <= num_chains; ++i)
3368 : 1410 : result = lappend(result, results[i]);
3369 : :
3370 : : /*
3371 : : * Free all the things.
3372 : : *
3373 : : * (This is over-fussy for small sets but for large sets we could have
3374 : : * tied up a nontrivial amount of memory.)
3375 : : */
3376 : 831 : BipartiteMatchFree(state);
3377 : 831 : pfree(results);
3378 : 831 : pfree(chains);
3379 [ + + ]: 2732 : for (i = 1; i <= num_sets; ++i)
3380 [ + + ]: 1901 : if (adjacency[i])
3381 : 513 : pfree(adjacency[i]);
3382 : 831 : pfree(adjacency);
3383 : 831 : pfree(adjacency_buf);
3384 : 831 : pfree(orig_sets);
3385 [ + + ]: 2732 : for (i = 1; i <= num_sets; ++i)
3386 : 1901 : bms_free(set_masks[i]);
3387 : 831 : pfree(set_masks);
3388 : :
3389 : 831 : return result;
3390 : : }
3391 : :
3392 : : /*
3393 : : * Reorder the elements of a list of grouping sets such that they have correct
3394 : : * prefix relationships. Also inserts the GroupingSetData annotations.
3395 : : *
3396 : : * The input must be ordered with smallest sets first; the result is returned
3397 : : * with largest sets first. Note that the result shares no list substructure
3398 : : * with the input, so it's safe for the caller to modify it later.
3399 : : *
3400 : : * If we're passed in a sortclause, we follow its order of columns to the
3401 : : * extent possible, to minimize the chance that we add unnecessary sorts.
3402 : : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3403 : : * gets implemented in one pass.)
3404 : : */
3405 : : static List *
3406 : 1485 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3407 : : {
3408 : : ListCell *lc;
3409 : 1485 : List *previous = NIL;
3410 : 1485 : List *result = NIL;
3411 : :
3412 [ + - + + : 4115 : foreach(lc, groupingSets)
+ + ]
3413 : : {
3414 : 2630 : List *candidate = (List *) lfirst(lc);
3415 : 2630 : List *new_elems = list_difference_int(candidate, previous);
3416 : 2630 : GroupingSetData *gs = makeNode(GroupingSetData);
3417 : :
3418 [ + + + + ]: 2774 : while (list_length(sortclause) > list_length(previous) &&
3419 : : new_elems != NIL)
3420 : : {
3421 : 244 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3422 : 244 : int ref = sc->tleSortGroupRef;
3423 : :
3424 [ + + ]: 244 : if (list_member_int(new_elems, ref))
3425 : : {
3426 : 144 : previous = lappend_int(previous, ref);
3427 : 144 : new_elems = list_delete_int(new_elems, ref);
3428 : : }
3429 : : else
3430 : : {
3431 : : /* diverged from the sortclause; give up on it */
3432 : 100 : sortclause = NIL;
3433 : 100 : break;
3434 : : }
3435 : : }
3436 : :
3437 : 2630 : previous = list_concat(previous, new_elems);
3438 : :
3439 : 2630 : gs->set = list_copy(previous);
3440 : 2630 : result = lcons(gs, result);
3441 : : }
3442 : :
3443 : 1485 : list_free(previous);
3444 : :
3445 : 1485 : return result;
3446 : : }
3447 : :
3448 : : /*
3449 : : * has_volatile_pathkey
3450 : : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3451 : : * containing a volatile function. Otherwise returns false.
3452 : : */
3453 : : static bool
3454 : 2035 : has_volatile_pathkey(List *keys)
3455 : : {
3456 : : ListCell *lc;
3457 : :
3458 [ + + + + : 4192 : foreach(lc, keys)
+ + ]
3459 : : {
3460 : 2170 : PathKey *pathkey = lfirst_node(PathKey, lc);
3461 : :
3462 [ + + ]: 2170 : if (pathkey->pk_eclass->ec_has_volatile)
3463 : 13 : return true;
3464 : : }
3465 : :
3466 : 2022 : return false;
3467 : : }
3468 : :
3469 : : /*
3470 : : * adjust_group_pathkeys_for_groupagg
3471 : : * Add pathkeys to root->group_pathkeys to reflect the best set of
3472 : : * pre-ordered input for ordered aggregates.
3473 : : *
3474 : : * We define "best" as the pathkeys that suit the largest number of
3475 : : * aggregate functions. We find these by looking at the first ORDER BY /
3476 : : * DISTINCT aggregate and take the pathkeys for that before searching for
3477 : : * other aggregates that require the same or a more strict variation of the
3478 : : * same pathkeys. We then repeat that process for any remaining aggregates
3479 : : * with different pathkeys and if we find another set of pathkeys that suits a
3480 : : * larger number of aggregates then we select those pathkeys instead.
3481 : : *
3482 : : * When the best pathkeys are found we also mark each Aggref that can use
3483 : : * those pathkeys as aggpresorted = true.
3484 : : *
3485 : : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3486 : : * volatile functions, we never make use of these pathkeys. We want to ensure
3487 : : * that sorts using volatile functions are done independently in each Aggref
3488 : : * rather than once at the query level. If we were to allow this then Aggrefs
3489 : : * with compatible sort orders would all transition their rows in the same
3490 : : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3491 : : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3492 : : * better pathkeys to sort on, then the volatile function Aggrefs would be
3493 : : * left to perform their sorts individually. To avoid this inconsistent
3494 : : * behavior which could make Aggref results depend on what other Aggrefs the
3495 : : * query contains, we always force Aggrefs with volatile functions to perform
3496 : : * their own sorts.
3497 : : */
3498 : : static void
3499 : 1679 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3500 : : {
3501 : 1679 : List *grouppathkeys = root->group_pathkeys;
3502 : : List *bestpathkeys;
3503 : : Bitmapset *bestaggs;
3504 : : Bitmapset *unprocessed_aggs;
3505 : : ListCell *lc;
3506 : : int i;
3507 : :
3508 : : /* Shouldn't be here if there are grouping sets */
3509 : : Assert(root->parse->groupingSets == NIL);
3510 : : /* Shouldn't be here unless there are some ordered aggregates */
3511 : : Assert(root->numOrderedAggs > 0);
3512 : :
3513 : : /* Do nothing if disabled */
3514 [ + + ]: 1679 : if (!enable_presorted_aggregate)
3515 : 5 : return;
3516 : :
3517 : : /*
3518 : : * Make a first pass over all AggInfos to collect a Bitmapset containing
3519 : : * the indexes of all AggInfos to be processed below.
3520 : : */
3521 : 1674 : unprocessed_aggs = NULL;
3522 [ + - + + : 3954 : foreach(lc, root->agginfos)
+ + ]
3523 : : {
3524 : 2280 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3525 : 2280 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3526 : :
3527 [ + + ]: 2280 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3528 : 214 : continue;
3529 : :
3530 : : /* Skip unless there's a DISTINCT or ORDER BY clause */
3531 [ + + + + ]: 2066 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3532 : 268 : continue;
3533 : :
3534 : : /* Additional safety checks are needed if there's a FILTER clause */
3535 [ + + ]: 1798 : if (aggref->aggfilter != NULL)
3536 : : {
3537 : : ListCell *lc2;
3538 : 43 : bool allow_presort = true;
3539 : :
3540 : : /*
3541 : : * When the Aggref has a FILTER clause, it's possible that the
3542 : : * filter removes rows that cannot be sorted because the
3543 : : * expression to sort by results in an error during its
3544 : : * evaluation. This is a problem for presorting as that happens
3545 : : * before the FILTER, whereas without presorting, the Aggregate
3546 : : * node will apply the FILTER *before* sorting. So that we never
3547 : : * try to sort anything that might error, here we aim to skip over
3548 : : * any Aggrefs with arguments with expressions which, when
3549 : : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3550 : : * may be more cases that should be allowed, but more thought
3551 : : * needs to be given. Err on the side of caution.
3552 : : */
3553 [ + - + + : 83 : foreach(lc2, aggref->args)
+ + ]
3554 : : {
3555 : 58 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3556 : 58 : Expr *expr = tle->expr;
3557 : :
3558 [ + + ]: 68 : while (IsA(expr, RelabelType))
3559 : 10 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3560 : :
3561 : : /* Common case, Vars and Consts are ok */
3562 [ + + + + ]: 58 : if (IsA(expr, Var) || IsA(expr, Const))
3563 : 40 : continue;
3564 : :
3565 : : /* Unsupported. Don't try to presort for this Aggref */
3566 : 18 : allow_presort = false;
3567 : 18 : break;
3568 : : }
3569 : :
3570 : : /* Skip unsupported Aggrefs */
3571 [ + + ]: 43 : if (!allow_presort)
3572 : 18 : continue;
3573 : : }
3574 : :
3575 : 1780 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3576 : : foreach_current_index(lc));
3577 : : }
3578 : :
3579 : : /*
3580 : : * Now process all the unprocessed_aggs to find the best set of pathkeys
3581 : : * for the given set of aggregates.
3582 : : *
3583 : : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3584 : : * this during the first loop using the pathkeys for the very first
3585 : : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3586 : : * a more strict set of compatible pathkeys. Once the outer loop is
3587 : : * complete, we mark off all the aggregates with compatible pathkeys then
3588 : : * remove those from the unprocessed_aggs and repeat the process to try to
3589 : : * find another set of pathkeys that are suitable for a larger number of
3590 : : * aggregates. The outer loop will stop when there are not enough
3591 : : * unprocessed aggregates for it to be possible to find a set of pathkeys
3592 : : * to suit a larger number of aggregates.
3593 : : */
3594 : 1674 : bestpathkeys = NIL;
3595 : 1674 : bestaggs = NULL;
3596 [ + + ]: 3299 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3597 : : {
3598 : 1625 : Bitmapset *aggindexes = NULL;
3599 : 1625 : List *currpathkeys = NIL;
3600 : :
3601 : 1625 : i = -1;
3602 [ + + ]: 3660 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3603 : : {
3604 : 2035 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3605 : 2035 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3606 : : List *sortlist;
3607 : : List *pathkeys;
3608 : :
3609 [ + + ]: 2035 : if (aggref->aggdistinct != NIL)
3610 : 599 : sortlist = aggref->aggdistinct;
3611 : : else
3612 : 1436 : sortlist = aggref->aggorder;
3613 : :
3614 : 2035 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3615 : : aggref->args);
3616 : :
3617 : : /*
3618 : : * Ignore Aggrefs which have volatile functions in their ORDER BY
3619 : : * or DISTINCT clause.
3620 : : */
3621 [ + + ]: 2035 : if (has_volatile_pathkey(pathkeys))
3622 : : {
3623 : 13 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3624 : 13 : continue;
3625 : : }
3626 : :
3627 : : /*
3628 : : * When not set yet, take the pathkeys from the first unprocessed
3629 : : * aggregate.
3630 : : */
3631 [ + + ]: 2022 : if (currpathkeys == NIL)
3632 : : {
3633 : 1622 : currpathkeys = pathkeys;
3634 : :
3635 : : /* include the GROUP BY pathkeys, if they exist */
3636 [ + + ]: 1622 : if (grouppathkeys != NIL)
3637 : 236 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3638 : : currpathkeys);
3639 : :
3640 : : /* record that we found pathkeys for this aggregate */
3641 : 1622 : aggindexes = bms_add_member(aggindexes, i);
3642 : : }
3643 : : else
3644 : : {
3645 : : /* now look for a stronger set of matching pathkeys */
3646 : :
3647 : : /* include the GROUP BY pathkeys, if they exist */
3648 [ + + ]: 400 : if (grouppathkeys != NIL)
3649 : 260 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3650 : : pathkeys);
3651 : :
3652 : : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3653 [ + + + - ]: 400 : switch (compare_pathkeys(currpathkeys, pathkeys))
3654 : : {
3655 : 10 : case PATHKEYS_BETTER2:
3656 : : /* 'pathkeys' are stronger, use these ones instead */
3657 : 10 : currpathkeys = pathkeys;
3658 : : pg_fallthrough;
3659 : :
3660 : 75 : case PATHKEYS_BETTER1:
3661 : : /* 'pathkeys' are less strict */
3662 : : pg_fallthrough;
3663 : :
3664 : : case PATHKEYS_EQUAL:
3665 : : /* mark this aggregate as covered by 'currpathkeys' */
3666 : 75 : aggindexes = bms_add_member(aggindexes, i);
3667 : 75 : break;
3668 : :
3669 : 325 : case PATHKEYS_DIFFERENT:
3670 : 325 : break;
3671 : : }
3672 : : }
3673 : : }
3674 : :
3675 : : /* remove the aggregates that we've just processed */
3676 : 1625 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3677 : :
3678 : : /*
3679 : : * If this pass included more aggregates than the previous best then
3680 : : * use these ones as the best set.
3681 : : */
3682 [ + + ]: 1625 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3683 : : {
3684 : 1537 : bestaggs = aggindexes;
3685 : 1537 : bestpathkeys = currpathkeys;
3686 : : }
3687 : : }
3688 : :
3689 : : /*
3690 : : * If we found any ordered aggregates, update root->group_pathkeys to add
3691 : : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3692 : : * the original GROUP BY pathkeys already.
3693 : : */
3694 [ + + ]: 1674 : if (bestpathkeys != NIL)
3695 : 1487 : root->group_pathkeys = bestpathkeys;
3696 : :
3697 : : /*
3698 : : * Now that we've found the best set of aggregates we can set the
3699 : : * presorted flag to indicate to the executor that it needn't bother
3700 : : * performing a sort for these Aggrefs. We're able to do this now as
3701 : : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3702 : : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3703 : : * of ordered aggregates.
3704 : : */
3705 : 1674 : i = -1;
3706 [ + + ]: 3261 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3707 : : {
3708 : 1587 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3709 : :
3710 [ + - + + : 3189 : foreach(lc, agginfo->aggrefs)
+ + ]
3711 : : {
3712 : 1602 : Aggref *aggref = lfirst_node(Aggref, lc);
3713 : :
3714 : 1602 : aggref->aggpresorted = true;
3715 : : }
3716 : : }
3717 : : }
3718 : :
3719 : : /*
3720 : : * Compute query_pathkeys and other pathkeys during plan generation
3721 : : */
3722 : : static void
3723 : 386619 : standard_qp_callback(PlannerInfo *root, void *extra)
3724 : : {
3725 : 386619 : Query *parse = root->parse;
3726 : 386619 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
3727 : 386619 : List *tlist = root->processed_tlist;
3728 : 386619 : List *activeWindows = qp_extra->activeWindows;
3729 : :
3730 : : /*
3731 : : * Calculate pathkeys that represent grouping/ordering and/or ordered
3732 : : * aggregate requirements.
3733 : : */
3734 [ + + ]: 386619 : if (qp_extra->gset_data)
3735 : : {
3736 : : /*
3737 : : * With grouping sets, just use the first RollupData's groupClause. We
3738 : : * don't make any effort to optimize grouping clauses when there are
3739 : : * grouping sets, nor can we combine aggregate ordering keys with
3740 : : * grouping.
3741 : : */
3742 : 911 : List *rollups = qp_extra->gset_data->rollups;
3743 [ + + ]: 911 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3744 : :
3745 [ + - ]: 911 : if (grouping_is_sortable(groupClause))
3746 : : {
3747 : : bool sortable;
3748 : :
3749 : : /*
3750 : : * The groupClause is logically below the grouping step. So if
3751 : : * there is an RTE entry for the grouping step, we need to remove
3752 : : * its RT index from the sort expressions before we make PathKeys
3753 : : * for them.
3754 : : */
3755 : 911 : root->group_pathkeys =
3756 : 911 : make_pathkeys_for_sortclauses_extended(root,
3757 : : &groupClause,
3758 : : tlist,
3759 : : false,
3760 : 911 : parse->hasGroupRTE,
3761 : : &sortable,
3762 : : false);
3763 : : Assert(sortable);
3764 : 911 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3765 : : }
3766 : : else
3767 : : {
3768 : 0 : root->group_pathkeys = NIL;
3769 : 0 : root->num_groupby_pathkeys = 0;
3770 : : }
3771 : : }
3772 [ + + + + ]: 385708 : else if (parse->groupClause || root->numOrderedAggs > 0)
3773 : 5054 : {
3774 : : /*
3775 : : * With a plain GROUP BY list, we can remove any grouping items that
3776 : : * are proven redundant by EquivalenceClass processing. For example,
3777 : : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3778 : : * especially common cases, but they're nearly free to detect. Note
3779 : : * that we remove redundant items from processed_groupClause but not
3780 : : * the original parse->groupClause.
3781 : : */
3782 : : bool sortable;
3783 : :
3784 : : /*
3785 : : * Convert group clauses into pathkeys. Set the ec_sortref field of
3786 : : * EquivalenceClass'es if it's not set yet.
3787 : : */
3788 : 5054 : root->group_pathkeys =
3789 : 5054 : make_pathkeys_for_sortclauses_extended(root,
3790 : : &root->processed_groupClause,
3791 : : tlist,
3792 : : true,
3793 : : false,
3794 : : &sortable,
3795 : : true);
3796 [ - + ]: 5054 : if (!sortable)
3797 : : {
3798 : : /* Can't sort; no point in considering aggregate ordering either */
3799 : 0 : root->group_pathkeys = NIL;
3800 : 0 : root->num_groupby_pathkeys = 0;
3801 : : }
3802 : : else
3803 : : {
3804 : 5054 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3805 : : /* If we have ordered aggs, consider adding onto group_pathkeys */
3806 [ + + ]: 5054 : if (root->numOrderedAggs > 0)
3807 : 1679 : adjust_group_pathkeys_for_groupagg(root);
3808 : : }
3809 : : }
3810 : : else
3811 : : {
3812 : 380654 : root->group_pathkeys = NIL;
3813 : 380654 : root->num_groupby_pathkeys = 0;
3814 : : }
3815 : :
3816 : : /* We consider only the first (bottom) window in pathkeys logic */
3817 [ + + ]: 386619 : if (activeWindows != NIL)
3818 : : {
3819 : 2331 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3820 : :
3821 : 2331 : root->window_pathkeys = make_pathkeys_for_window(root,
3822 : : wc,
3823 : : tlist);
3824 : : }
3825 : : else
3826 : 384288 : root->window_pathkeys = NIL;
3827 : :
3828 : : /*
3829 : : * As with GROUP BY, we can discard any DISTINCT items that are proven
3830 : : * redundant by EquivalenceClass processing. The non-redundant list is
3831 : : * kept in root->processed_distinctClause, leaving the original
3832 : : * parse->distinctClause alone.
3833 : : */
3834 [ + + ]: 386619 : if (parse->distinctClause)
3835 : : {
3836 : : bool sortable;
3837 : :
3838 : : /* Make a copy since pathkey processing can modify the list */
3839 : 2013 : root->processed_distinctClause = list_copy(parse->distinctClause);
3840 : 2013 : root->distinct_pathkeys =
3841 : 2013 : make_pathkeys_for_sortclauses_extended(root,
3842 : : &root->processed_distinctClause,
3843 : : tlist,
3844 : : true,
3845 : : false,
3846 : : &sortable,
3847 : : false);
3848 [ + + ]: 2013 : if (!sortable)
3849 : 5 : root->distinct_pathkeys = NIL;
3850 : : }
3851 : : else
3852 : 384606 : root->distinct_pathkeys = NIL;
3853 : :
3854 : 386619 : root->sort_pathkeys =
3855 : 386619 : make_pathkeys_for_sortclauses(root,
3856 : : parse->sortClause,
3857 : : tlist);
3858 : :
3859 : : /* setting setop_pathkeys might be useful to the union planner */
3860 [ + + ]: 386619 : if (qp_extra->setop != NULL)
3861 : : {
3862 : : List *groupClauses;
3863 : : bool sortable;
3864 : :
3865 : 10679 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3866 : :
3867 : 10679 : root->setop_pathkeys =
3868 : 10679 : make_pathkeys_for_sortclauses_extended(root,
3869 : : &groupClauses,
3870 : : tlist,
3871 : : false,
3872 : : false,
3873 : : &sortable,
3874 : : false);
3875 [ + + ]: 10679 : if (!sortable)
3876 : 120 : root->setop_pathkeys = NIL;
3877 : : }
3878 : : else
3879 : 375940 : root->setop_pathkeys = NIL;
3880 : :
3881 : : /*
3882 : : * Figure out whether we want a sorted result from query_planner.
3883 : : *
3884 : : * If we have a sortable GROUP BY clause, then we want a result sorted
3885 : : * properly for grouping. Otherwise, if we have window functions to
3886 : : * evaluate, we try to sort for the first window. Otherwise, if there's a
3887 : : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3888 : : * we try to produce output that's sufficiently well sorted for the
3889 : : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3890 : : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3891 : : * for a set operation which can benefit from presorted results and have a
3892 : : * sortable targetlist, we want to sort by the target list.
3893 : : *
3894 : : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3895 : : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3896 : : * that might just leave us failing to exploit an available sort order at
3897 : : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3898 : : * much easier, since we know that the parser ensured that one is a
3899 : : * superset of the other.
3900 : : */
3901 [ + + ]: 386619 : if (root->group_pathkeys)
3902 : 5558 : root->query_pathkeys = root->group_pathkeys;
3903 [ + + ]: 381061 : else if (root->window_pathkeys)
3904 : 1871 : root->query_pathkeys = root->window_pathkeys;
3905 [ + + ]: 758380 : else if (list_length(root->distinct_pathkeys) >
3906 : 379190 : list_length(root->sort_pathkeys))
3907 : 1602 : root->query_pathkeys = root->distinct_pathkeys;
3908 [ + + ]: 377588 : else if (root->sort_pathkeys)
3909 : 55514 : root->query_pathkeys = root->sort_pathkeys;
3910 [ + + ]: 322074 : else if (root->setop_pathkeys != NIL)
3911 : 9671 : root->query_pathkeys = root->setop_pathkeys;
3912 : : else
3913 : 312403 : root->query_pathkeys = NIL;
3914 : 386619 : }
3915 : :
3916 : : /*
3917 : : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3918 : : *
3919 : : * path_rows: number of output rows from scan/join step
3920 : : * gd: grouping sets data including list of grouping sets and their clauses
3921 : : * target_list: target list containing group clause references
3922 : : *
3923 : : * If doing grouping sets, we also annotate the gsets data with the estimates
3924 : : * for each set and each individual rollup list, with a view to later
3925 : : * determining whether some combination of them could be hashed instead.
3926 : : */
3927 : : static double
3928 : 41087 : get_number_of_groups(PlannerInfo *root,
3929 : : double path_rows,
3930 : : grouping_sets_data *gd,
3931 : : List *target_list)
3932 : : {
3933 : 41087 : Query *parse = root->parse;
3934 : : double dNumGroups;
3935 : :
3936 [ + + ]: 41087 : if (parse->groupClause)
3937 : : {
3938 : : List *groupExprs;
3939 : :
3940 [ + + ]: 8830 : if (parse->groupingSets)
3941 : : {
3942 : : /* Add up the estimates for each grouping set */
3943 : : ListCell *lc;
3944 : :
3945 : : Assert(gd); /* keep Coverity happy */
3946 : :
3947 : 836 : dNumGroups = 0;
3948 : :
3949 [ + + + + : 2246 : foreach(lc, gd->rollups)
+ + ]
3950 : : {
3951 : 1410 : RollupData *rollup = lfirst_node(RollupData, lc);
3952 : : ListCell *lc2;
3953 : : ListCell *lc3;
3954 : :
3955 : 1410 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3956 : : target_list);
3957 : :
3958 : 1410 : rollup->numGroups = 0.0;
3959 : :
3960 [ + - + + : 3915 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
+ - + + +
+ + - +
+ ]
3961 : : {
3962 : 2505 : List *gset = (List *) lfirst(lc2);
3963 : 2505 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3964 : 2505 : double numGroups = estimate_num_groups(root,
3965 : : groupExprs,
3966 : : path_rows,
3967 : : &gset,
3968 : : NULL);
3969 : :
3970 : 2505 : gs->numGroups = numGroups;
3971 : 2505 : rollup->numGroups += numGroups;
3972 : : }
3973 : :
3974 : 1410 : dNumGroups += rollup->numGroups;
3975 : : }
3976 : :
3977 [ + + ]: 836 : if (gd->hash_sets_idx)
3978 : : {
3979 : : ListCell *lc2;
3980 : :
3981 : 29 : gd->dNumHashGroups = 0;
3982 : :
3983 : 29 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3984 : : target_list);
3985 : :
3986 [ + - + + : 63 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
+ - + + +
+ + - +
+ ]
3987 : : {
3988 : 34 : List *gset = (List *) lfirst(lc);
3989 : 34 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3990 : 34 : double numGroups = estimate_num_groups(root,
3991 : : groupExprs,
3992 : : path_rows,
3993 : : &gset,
3994 : : NULL);
3995 : :
3996 : 34 : gs->numGroups = numGroups;
3997 : 34 : gd->dNumHashGroups += numGroups;
3998 : : }
3999 : :
4000 : 29 : dNumGroups += gd->dNumHashGroups;
4001 : : }
4002 : : }
4003 : : else
4004 : : {
4005 : : /* Plain GROUP BY -- estimate based on optimized groupClause */
4006 : 7994 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
4007 : : target_list);
4008 : :
4009 : 7994 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
4010 : : NULL, NULL);
4011 : : }
4012 : : }
4013 [ + + ]: 32257 : else if (parse->groupingSets)
4014 : : {
4015 : : /* Empty grouping sets ... one result row for each one */
4016 : 50 : dNumGroups = list_length(parse->groupingSets);
4017 : : }
4018 [ - + - - ]: 32207 : else if (parse->hasAggs || root->hasHavingQual)
4019 : : {
4020 : : /* Plain aggregation, one result row */
4021 : 32207 : dNumGroups = 1;
4022 : : }
4023 : : else
4024 : : {
4025 : : /* Not grouping */
4026 : 0 : dNumGroups = 1;
4027 : : }
4028 : :
4029 : 41087 : return dNumGroups;
4030 : : }
4031 : :
4032 : : /*
4033 : : * create_grouping_paths
4034 : : *
4035 : : * Build a new upperrel containing Paths for grouping and/or aggregation.
4036 : : * Along the way, we also build an upperrel for Paths which are partially
4037 : : * grouped and/or aggregated. A partially grouped and/or aggregated path
4038 : : * needs a FinalizeAggregate node to complete the aggregation. Currently,
4039 : : * the only partially grouped paths we build are also partial paths; that
4040 : : * is, they need a Gather and then a FinalizeAggregate.
4041 : : *
4042 : : * input_rel: contains the source-data Paths
4043 : : * target: the pathtarget for the result Paths to compute
4044 : : * gd: grouping sets data including list of grouping sets and their clauses
4045 : : *
4046 : : * Note: all Paths in input_rel are expected to return the target computed
4047 : : * by make_group_input_target.
4048 : : */
4049 : : static RelOptInfo *
4050 : 34089 : create_grouping_paths(PlannerInfo *root,
4051 : : RelOptInfo *input_rel,
4052 : : PathTarget *target,
4053 : : bool target_parallel_safe,
4054 : : grouping_sets_data *gd)
4055 : : {
4056 : 34089 : Query *parse = root->parse;
4057 : : RelOptInfo *grouped_rel;
4058 : : RelOptInfo *partially_grouped_rel;
4059 : : AggClauseCosts agg_costs;
4060 : :
4061 [ + - + - : 204534 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
4062 : 34089 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
4063 : :
4064 : : /*
4065 : : * Create grouping relation to hold fully aggregated grouping and/or
4066 : : * aggregation paths.
4067 : : */
4068 : 34089 : grouped_rel = make_grouping_rel(root, input_rel, target,
4069 : : target_parallel_safe, parse->havingQual);
4070 : :
4071 : : /*
4072 : : * Create either paths for a degenerate grouping or paths for ordinary
4073 : : * grouping, as appropriate.
4074 : : */
4075 [ + + ]: 34089 : if (is_degenerate_grouping(root))
4076 : 45 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
4077 : : else
4078 : : {
4079 : 34044 : int flags = 0;
4080 : : GroupPathExtraData extra;
4081 : :
4082 : : /*
4083 : : * Determine whether it's possible to perform sort-based
4084 : : * implementations of grouping. (Note that if processed_groupClause
4085 : : * is empty, grouping_is_sortable() is trivially true, and all the
4086 : : * pathkeys_contained_in() tests will succeed too, so that we'll
4087 : : * consider every surviving input path.)
4088 : : *
4089 : : * If we have grouping sets, we might be able to sort some but not all
4090 : : * of them; in this case, we need can_sort to be true as long as we
4091 : : * must consider any sorted-input plan.
4092 : : */
4093 [ + + + + ]: 34044 : if ((gd && gd->rollups != NIL)
4094 [ + + ]: 33163 : || grouping_is_sortable(root->processed_groupClause))
4095 : 34039 : flags |= GROUPING_CAN_USE_SORT;
4096 : :
4097 : : /*
4098 : : * Determine whether we should consider hash-based implementations of
4099 : : * grouping.
4100 : : *
4101 : : * Hashed aggregation only applies if we're grouping. If we have
4102 : : * grouping sets, some groups might be hashable but others not; in
4103 : : * this case we set can_hash true as long as there is nothing globally
4104 : : * preventing us from hashing (and we should therefore consider plans
4105 : : * with hashes).
4106 : : *
4107 : : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
4108 : : * BY aggregates. (Doing so would imply storing *all* the input
4109 : : * values in the hash table, and/or running many sorts in parallel,
4110 : : * either of which seems like a certain loser.) We similarly don't
4111 : : * support ordered-set aggregates in hashed aggregation, but that case
4112 : : * is also included in the numOrderedAggs count.
4113 : : *
4114 : : * Note: grouping_is_hashable() is much more expensive to check than
4115 : : * the other gating conditions, so we want to do it last.
4116 : : */
4117 [ + + ]: 34044 : if ((parse->groupClause != NIL &&
4118 [ + + + + : 7784 : root->numOrderedAggs == 0 &&
+ + ]
4119 : 3375 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
4120 : 4179 : flags |= GROUPING_CAN_USE_HASH;
4121 : :
4122 : : /*
4123 : : * Determine whether partial aggregation is possible.
4124 : : */
4125 [ + + ]: 34044 : if (can_partial_agg(root))
4126 : 30101 : flags |= GROUPING_CAN_PARTIAL_AGG;
4127 : :
4128 : 34044 : extra.flags = flags;
4129 : 34044 : extra.target_parallel_safe = target_parallel_safe;
4130 : 34044 : extra.havingQual = parse->havingQual;
4131 : 34044 : extra.targetList = parse->targetList;
4132 : 34044 : extra.partial_costs_set = false;
4133 : :
4134 : : /*
4135 : : * Determine whether partitionwise aggregation is in theory possible.
4136 : : * It can be disabled by the user, and for now, we don't try to
4137 : : * support grouping sets. create_ordinary_grouping_paths() will check
4138 : : * additional conditions, such as whether input_rel is partitioned.
4139 : : */
4140 [ + + + + ]: 34044 : if (enable_partitionwise_aggregate && !parse->groupingSets)
4141 : 580 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
4142 : : else
4143 : 33464 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
4144 : :
4145 : 34044 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
4146 : : &agg_costs, gd, &extra,
4147 : : &partially_grouped_rel);
4148 : : }
4149 : :
4150 : 34085 : set_cheapest(grouped_rel);
4151 : 34085 : return grouped_rel;
4152 : : }
4153 : :
4154 : : /*
4155 : : * make_grouping_rel
4156 : : *
4157 : : * Create a new grouping rel and set basic properties.
4158 : : *
4159 : : * input_rel represents the underlying scan/join relation.
4160 : : * target is the output expected from the grouping relation.
4161 : : */
4162 : : static RelOptInfo *
4163 : 35884 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
4164 : : PathTarget *target, bool target_parallel_safe,
4165 : : Node *havingQual)
4166 : : {
4167 : : RelOptInfo *grouped_rel;
4168 : :
4169 [ + + + + : 35884 : if (IS_OTHER_REL(input_rel))
- + ]
4170 : : {
4171 : 1795 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
4172 : : input_rel->relids);
4173 : 1795 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
4174 : : }
4175 : : else
4176 : : {
4177 : : /*
4178 : : * By tradition, the relids set for the main grouping relation is
4179 : : * NULL. (This could be changed, but might require adjustments
4180 : : * elsewhere.)
4181 : : */
4182 : 34089 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
4183 : : }
4184 : :
4185 : : /* Set target. */
4186 : 35884 : grouped_rel->reltarget = target;
4187 : :
4188 : : /*
4189 : : * If the input relation is not parallel-safe, then the grouped relation
4190 : : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
4191 : : * target list and HAVING quals are parallel-safe.
4192 : : */
4193 [ + + + + : 58955 : if (input_rel->consider_parallel && target_parallel_safe &&
+ + ]
4194 : 23071 : is_parallel_safe(root, havingQual))
4195 : 23046 : grouped_rel->consider_parallel = true;
4196 : :
4197 : : /* Assume that the same path generation strategies are allowed */
4198 : 35884 : grouped_rel->pgs_mask = input_rel->pgs_mask;
4199 : :
4200 : : /*
4201 : : * If the input rel belongs to a single FDW, so does the grouped rel.
4202 : : */
4203 : 35884 : grouped_rel->serverid = input_rel->serverid;
4204 : 35884 : grouped_rel->userid = input_rel->userid;
4205 : 35884 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
4206 : 35884 : grouped_rel->fdwroutine = input_rel->fdwroutine;
4207 : :
4208 : 35884 : return grouped_rel;
4209 : : }
4210 : :
4211 : : /*
4212 : : * is_degenerate_grouping
4213 : : *
4214 : : * A degenerate grouping is one in which the query has a HAVING qual and/or
4215 : : * grouping sets, but no aggregates and no GROUP BY (which implies that the
4216 : : * grouping sets are all empty).
4217 : : */
4218 : : static bool
4219 : 34089 : is_degenerate_grouping(PlannerInfo *root)
4220 : : {
4221 : 34089 : Query *parse = root->parse;
4222 : :
4223 [ + + ]: 32908 : return (root->hasHavingQual || parse->groupingSets) &&
4224 [ + + + + : 66997 : !parse->hasAggs && parse->groupClause == NIL;
+ + ]
4225 : : }
4226 : :
4227 : : /*
4228 : : * create_degenerate_grouping_paths
4229 : : *
4230 : : * When the grouping is degenerate (see is_degenerate_grouping), we are
4231 : : * supposed to emit either zero or one row for each grouping set depending on
4232 : : * whether HAVING succeeds. Furthermore, there cannot be any variables in
4233 : : * either HAVING or the targetlist, so we actually do not need the FROM table
4234 : : * at all! We can just throw away the plan-so-far and generate a Result node.
4235 : : * This is a sufficiently unusual corner case that it's not worth contorting
4236 : : * the structure of this module to avoid having to generate the earlier paths
4237 : : * in the first place.
4238 : : */
4239 : : static void
4240 : 45 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4241 : : RelOptInfo *grouped_rel)
4242 : : {
4243 : 45 : Query *parse = root->parse;
4244 : : int nrows;
4245 : : Path *path;
4246 : :
4247 : 45 : nrows = list_length(parse->groupingSets);
4248 [ + + ]: 45 : if (nrows > 1)
4249 : : {
4250 : : /*
4251 : : * Doesn't seem worthwhile writing code to cons up a generate_series
4252 : : * or a values scan to emit multiple rows. Instead just make N clones
4253 : : * and append them. (With a volatile HAVING clause, this means you
4254 : : * might get between 0 and N output rows. Offhand I think that's
4255 : : * desired.)
4256 : : */
4257 : 10 : AppendPathInput append = {0};
4258 : :
4259 [ + + ]: 30 : while (--nrows >= 0)
4260 : : {
4261 : : path = (Path *)
4262 : 20 : create_group_result_path(root, grouped_rel,
4263 : 20 : grouped_rel->reltarget,
4264 : 20 : (List *) parse->havingQual);
4265 : 20 : append.subpaths = lappend(append.subpaths, path);
4266 : : }
4267 : : path = (Path *)
4268 : 10 : create_append_path(root,
4269 : : grouped_rel,
4270 : : append,
4271 : : NIL,
4272 : : NULL,
4273 : : 0,
4274 : : false,
4275 : : -1);
4276 : : }
4277 : : else
4278 : : {
4279 : : /* No grouping sets, or just one, so one output row */
4280 : : path = (Path *)
4281 : 35 : create_group_result_path(root, grouped_rel,
4282 : 35 : grouped_rel->reltarget,
4283 : 35 : (List *) parse->havingQual);
4284 : : }
4285 : :
4286 : 45 : add_path(grouped_rel, path);
4287 : 45 : }
4288 : :
4289 : : /*
4290 : : * create_ordinary_grouping_paths
4291 : : *
4292 : : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4293 : : *
4294 : : * We need to consider sorted and hashed aggregation in the same function,
4295 : : * because otherwise (1) it would be harder to throw an appropriate error
4296 : : * message if neither way works, and (2) we should not allow hashtable size
4297 : : * considerations to dissuade us from using hashing if sorting is not possible.
4298 : : *
4299 : : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4300 : : * function creates, or to NULL if it doesn't create one.
4301 : : */
4302 : : static void
4303 : 35839 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4304 : : RelOptInfo *grouped_rel,
4305 : : const AggClauseCosts *agg_costs,
4306 : : grouping_sets_data *gd,
4307 : : GroupPathExtraData *extra,
4308 : : RelOptInfo **partially_grouped_rel_p)
4309 : : {
4310 : 35839 : RelOptInfo *partially_grouped_rel = NULL;
4311 : 35839 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4312 : :
4313 : : /*
4314 : : * If this is the topmost grouping relation or if the parent relation is
4315 : : * doing some form of partitionwise aggregation, then we may be able to do
4316 : : * it at this level also. However, if the input relation is not
4317 : : * partitioned, partitionwise aggregate is impossible.
4318 : : */
4319 [ + + ]: 35839 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
4320 [ + + + - : 2375 : IS_PARTITIONED_REL(input_rel))
+ + + - +
+ ]
4321 : : {
4322 : : /*
4323 : : * If this is the topmost relation or if the parent relation is doing
4324 : : * full partitionwise aggregation, then we can do full partitionwise
4325 : : * aggregation provided that the GROUP BY clause contains all of the
4326 : : * partitioning columns at this level and the collation used by GROUP
4327 : : * BY matches the partitioning collation. Otherwise, we can do at
4328 : : * most partial partitionwise aggregation. But if partial aggregation
4329 : : * is not supported in general then we can't use it for partitionwise
4330 : : * aggregation either.
4331 : : *
4332 : : * Check parse->groupClause not processed_groupClause, because it's
4333 : : * okay if some of the partitioning columns were proved redundant.
4334 : : */
4335 [ + + + + ]: 1360 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4336 : 640 : group_by_has_partkey(input_rel, extra->targetList,
4337 : 640 : root->parse->groupClause))
4338 : 404 : patype = PARTITIONWISE_AGGREGATE_FULL;
4339 [ + + ]: 316 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4340 : 281 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4341 : : else
4342 : 35 : patype = PARTITIONWISE_AGGREGATE_NONE;
4343 : : }
4344 : :
4345 : : /*
4346 : : * Before generating paths for grouped_rel, we first generate any possible
4347 : : * partially grouped paths; that way, later code can easily consider both
4348 : : * parallel and non-parallel approaches to grouping.
4349 : : */
4350 [ + + ]: 35839 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4351 : : {
4352 : : bool force_rel_creation;
4353 : :
4354 : : /*
4355 : : * If we're doing partitionwise aggregation at this level, force
4356 : : * creation of a partially_grouped_rel so we can add partitionwise
4357 : : * paths to it.
4358 : : */
4359 : 31836 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4360 : :
4361 : : partially_grouped_rel =
4362 : 31836 : create_partial_grouping_paths(root,
4363 : : grouped_rel,
4364 : : input_rel,
4365 : : gd,
4366 : : extra,
4367 : : force_rel_creation);
4368 : : }
4369 : :
4370 : : /* Set out parameter. */
4371 : 35839 : *partially_grouped_rel_p = partially_grouped_rel;
4372 : :
4373 : : /* Apply partitionwise aggregation technique, if possible. */
4374 [ + + ]: 35839 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4375 : 685 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4376 : : partially_grouped_rel, agg_costs,
4377 : : gd, patype, extra);
4378 : :
4379 : : /* If we are doing partial aggregation only, return. */
4380 [ + + ]: 35839 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4381 : : {
4382 : : Assert(partially_grouped_rel);
4383 : :
4384 [ + - ]: 713 : if (partially_grouped_rel->pathlist)
4385 : 713 : set_cheapest(partially_grouped_rel);
4386 : :
4387 : 713 : return;
4388 : : }
4389 : :
4390 : : /* Gather any partially grouped partial paths. */
4391 [ + + + + ]: 35126 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
4392 : 2325 : gather_grouping_paths(root, partially_grouped_rel);
4393 : :
4394 : : /* Now choose the best path(s) for partially_grouped_rel. */
4395 [ + + + - ]: 35126 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4396 : 2503 : set_cheapest(partially_grouped_rel);
4397 : :
4398 : : /* Build final grouping paths */
4399 : 35126 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4400 : : partially_grouped_rel, agg_costs, gd,
4401 : : extra);
4402 : :
4403 : : /* Give a helpful error if we failed to find any implementation */
4404 [ + + ]: 35126 : if (grouped_rel->pathlist == NIL)
4405 [ + - ]: 4 : ereport(ERROR,
4406 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4407 : : errmsg("could not implement GROUP BY"),
4408 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4409 : :
4410 : : /*
4411 : : * If there is an FDW that's responsible for all baserels of the query,
4412 : : * let it consider adding ForeignPaths.
4413 : : */
4414 [ + + ]: 35122 : if (grouped_rel->fdwroutine &&
4415 [ + + ]: 172 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4416 : 171 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4417 : : input_rel, grouped_rel,
4418 : : extra);
4419 : :
4420 : : /* Let extensions possibly add some more paths */
4421 [ - + ]: 35122 : if (create_upper_paths_hook)
4422 : 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4423 : : input_rel, grouped_rel,
4424 : : extra);
4425 : : }
4426 : :
4427 : : /*
4428 : : * For a given input path, consider the possible ways of doing grouping sets on
4429 : : * it, by combinations of hashing and sorting. This can be called multiple
4430 : : * times, so it's important that it not scribble on input. No result is
4431 : : * returned, but any generated paths are added to grouped_rel.
4432 : : */
4433 : : static void
4434 : 1782 : consider_groupingsets_paths(PlannerInfo *root,
4435 : : RelOptInfo *grouped_rel,
4436 : : Path *path,
4437 : : bool is_sorted,
4438 : : bool can_hash,
4439 : : grouping_sets_data *gd,
4440 : : const AggClauseCosts *agg_costs,
4441 : : double dNumGroups)
4442 : : {
4443 : 1782 : Query *parse = root->parse;
4444 : 1782 : Size hash_mem_limit = get_hash_memory_limit();
4445 : :
4446 : : /*
4447 : : * If we're not being offered sorted input, then only consider plans that
4448 : : * can be done entirely by hashing.
4449 : : *
4450 : : * We can hash everything if it looks like it'll fit in hash_mem. But if
4451 : : * the input is actually sorted despite not being advertised as such, we
4452 : : * prefer to make use of that in order to use less memory.
4453 : : *
4454 : : * If none of the grouping sets are sortable, then ignore the hash_mem
4455 : : * limit and generate a path anyway, since otherwise we'll just fail.
4456 : : */
4457 [ + + ]: 1782 : if (!is_sorted)
4458 : : {
4459 : 807 : List *new_rollups = NIL;
4460 : 807 : RollupData *unhashed_rollup = NULL;
4461 : : List *sets_data;
4462 : 807 : List *empty_sets_data = NIL;
4463 : 807 : List *empty_sets = NIL;
4464 : : ListCell *lc;
4465 : 807 : ListCell *l_start = list_head(gd->rollups);
4466 : 807 : AggStrategy strat = AGG_HASHED;
4467 : : double hashsize;
4468 : 807 : double exclude_groups = 0.0;
4469 : :
4470 : : Assert(can_hash);
4471 : :
4472 : : /*
4473 : : * If the input is coincidentally sorted usefully (which can happen
4474 : : * even if is_sorted is false, since that only means that our caller
4475 : : * has set up the sorting for us), then save some hashtable space by
4476 : : * making use of that. But we need to watch out for degenerate cases:
4477 : : *
4478 : : * 1) If there are any empty grouping sets, then group_pathkeys might
4479 : : * be NIL if all non-empty grouping sets are unsortable. In this case,
4480 : : * there will be a rollup containing only empty groups, and the
4481 : : * pathkeys_contained_in test is vacuously true; this is ok.
4482 : : *
4483 : : * XXX: the above relies on the fact that group_pathkeys is generated
4484 : : * from the first rollup. If we add the ability to consider multiple
4485 : : * sort orders for grouping input, this assumption might fail.
4486 : : *
4487 : : * 2) If there are no empty sets and only unsortable sets, then the
4488 : : * rollups list will be empty (and thus l_start == NULL), and
4489 : : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4490 : : * pathkeys_contained_in test doesn't cause us to crash.
4491 : : */
4492 [ + + + + ]: 1609 : if (l_start != NULL &&
4493 : 802 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4494 : : {
4495 : 40 : unhashed_rollup = lfirst_node(RollupData, l_start);
4496 : 40 : exclude_groups = unhashed_rollup->numGroups;
4497 : 40 : l_start = lnext(gd->rollups, l_start);
4498 : : }
4499 : :
4500 : 807 : hashsize = estimate_hashagg_tablesize(root,
4501 : : path,
4502 : : agg_costs,
4503 : : dNumGroups - exclude_groups);
4504 : :
4505 : : /*
4506 : : * gd->rollups is empty if we have only unsortable columns to work
4507 : : * with. Override hash_mem in that case; otherwise, we'll rely on the
4508 : : * sorted-input case to generate usable mixed paths.
4509 : : */
4510 [ + + + - ]: 807 : if (hashsize > hash_mem_limit && gd->rollups)
4511 : 15 : return; /* nope, won't fit */
4512 : :
4513 : : /*
4514 : : * We need to burst the existing rollups list into individual grouping
4515 : : * sets and recompute a groupClause for each set.
4516 : : */
4517 : 792 : sets_data = list_copy(gd->unsortable_sets);
4518 : :
4519 [ + + + + : 2013 : for_each_cell(lc, gd->rollups, l_start)
+ + ]
4520 : : {
4521 : 1241 : RollupData *rollup = lfirst_node(RollupData, lc);
4522 : :
4523 : : /*
4524 : : * If we find an unhashable rollup that's not been skipped by the
4525 : : * "actually sorted" check above, we can't cope; we'd need sorted
4526 : : * input (with a different sort order) but we can't get that here.
4527 : : * So bail out; we'll get a valid path from the is_sorted case
4528 : : * instead.
4529 : : *
4530 : : * The mere presence of empty grouping sets doesn't make a rollup
4531 : : * unhashable (see preprocess_grouping_sets), we handle those
4532 : : * specially below.
4533 : : */
4534 [ + + ]: 1241 : if (!rollup->hashable)
4535 : 20 : return;
4536 : :
4537 : 1221 : sets_data = list_concat(sets_data, rollup->gsets_data);
4538 : : }
4539 [ + - + + : 3053 : foreach(lc, sets_data)
+ + ]
4540 : : {
4541 : 2281 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4542 : 2281 : List *gset = gs->set;
4543 : : RollupData *rollup;
4544 : :
4545 [ + + ]: 2281 : if (gset == NIL)
4546 : : {
4547 : : /* Empty grouping sets can't be hashed. */
4548 : 445 : empty_sets_data = lappend(empty_sets_data, gs);
4549 : 445 : empty_sets = lappend(empty_sets, NIL);
4550 : : }
4551 : : else
4552 : : {
4553 : 1836 : rollup = makeNode(RollupData);
4554 : :
4555 : 1836 : rollup->groupClause = preprocess_groupclause(root, gset);
4556 : 1836 : rollup->gsets_data = list_make1(gs);
4557 : 1836 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4558 : : rollup->gsets_data,
4559 : : gd->tleref_to_colnum_map);
4560 : 1836 : rollup->numGroups = gs->numGroups;
4561 : 1836 : rollup->hashable = true;
4562 : 1836 : rollup->is_hashed = true;
4563 : 1836 : new_rollups = lappend(new_rollups, rollup);
4564 : : }
4565 : : }
4566 : :
4567 : : /*
4568 : : * If we didn't find anything nonempty to hash, then bail. We'll
4569 : : * generate a path from the is_sorted case.
4570 : : */
4571 [ - + ]: 772 : if (new_rollups == NIL)
4572 : 0 : return;
4573 : :
4574 : : /*
4575 : : * If there were empty grouping sets they should have been in the
4576 : : * first rollup.
4577 : : */
4578 : : Assert(!unhashed_rollup || !empty_sets);
4579 : :
4580 [ + + ]: 772 : if (unhashed_rollup)
4581 : : {
4582 : 40 : new_rollups = lappend(new_rollups, unhashed_rollup);
4583 : 40 : strat = AGG_MIXED;
4584 : : }
4585 [ + + ]: 732 : else if (empty_sets)
4586 : : {
4587 : 405 : RollupData *rollup = makeNode(RollupData);
4588 : :
4589 : 405 : rollup->groupClause = NIL;
4590 : 405 : rollup->gsets_data = empty_sets_data;
4591 : 405 : rollup->gsets = empty_sets;
4592 : 405 : rollup->numGroups = list_length(empty_sets);
4593 : 405 : rollup->hashable = false;
4594 : 405 : rollup->is_hashed = false;
4595 : 405 : new_rollups = lappend(new_rollups, rollup);
4596 : 405 : strat = AGG_MIXED;
4597 : : }
4598 : :
4599 : 772 : add_path(grouped_rel, (Path *)
4600 : 772 : create_groupingsets_path(root,
4601 : : grouped_rel,
4602 : : path,
4603 : 772 : (List *) parse->havingQual,
4604 : : strat,
4605 : : new_rollups,
4606 : : agg_costs));
4607 : 772 : return;
4608 : : }
4609 : :
4610 : : /*
4611 : : * If we have sorted input but nothing we can do with it, bail.
4612 : : */
4613 [ - + ]: 975 : if (gd->rollups == NIL)
4614 : 0 : return;
4615 : :
4616 : : /*
4617 : : * Given sorted input, we try and make two paths: one sorted and one mixed
4618 : : * sort/hash. (We need to try both because hashagg might be disabled, or
4619 : : * some columns might not be sortable.)
4620 : : *
4621 : : * can_hash is passed in as false if some obstacle elsewhere (such as
4622 : : * ordered aggs) means that we shouldn't consider hashing at all.
4623 : : */
4624 [ + + + - ]: 975 : if (can_hash && gd->any_hashable)
4625 : : {
4626 : 896 : List *rollups = NIL;
4627 : 896 : List *hash_sets = list_copy(gd->unsortable_sets);
4628 : 896 : double availspace = hash_mem_limit;
4629 : : ListCell *lc;
4630 : :
4631 : : /*
4632 : : * Account first for space needed for groups we can't sort at all.
4633 : : */
4634 : 896 : availspace -= estimate_hashagg_tablesize(root,
4635 : : path,
4636 : : agg_costs,
4637 : : gd->dNumHashGroups);
4638 : :
4639 [ + - + + ]: 896 : if (availspace > 0 && list_length(gd->rollups) > 1)
4640 : : {
4641 : : double scale;
4642 : 501 : int num_rollups = list_length(gd->rollups);
4643 : : int k_capacity;
4644 : 501 : int *k_weights = palloc(num_rollups * sizeof(int));
4645 : 501 : Bitmapset *hash_items = NULL;
4646 : : int i;
4647 : :
4648 : : /*
4649 : : * We treat this as a knapsack problem: the knapsack capacity
4650 : : * represents hash_mem, the item weights are the estimated memory
4651 : : * usage of the hashtables needed to implement a single rollup,
4652 : : * and we really ought to use the cost saving as the item value;
4653 : : * however, currently the costs assigned to sort nodes don't
4654 : : * reflect the comparison costs well, and so we treat all items as
4655 : : * of equal value (each rollup we hash instead saves us one sort).
4656 : : *
4657 : : * To use the discrete knapsack, we need to scale the values to a
4658 : : * reasonably small bounded range. We choose to allow a 5% error
4659 : : * margin; we have no more than 4096 rollups in the worst possible
4660 : : * case, which with a 5% error margin will require a bit over 42MB
4661 : : * of workspace. (Anyone wanting to plan queries that complex had
4662 : : * better have the memory for it. In more reasonable cases, with
4663 : : * no more than a couple of dozen rollups, the memory usage will
4664 : : * be negligible.)
4665 : : *
4666 : : * k_capacity is naturally bounded, but we clamp the values for
4667 : : * scale and weight (below) to avoid overflows or underflows (or
4668 : : * uselessly trying to use a scale factor less than 1 byte).
4669 : : */
4670 [ + - ]: 501 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4671 : 501 : k_capacity = (int) floor(availspace / scale);
4672 : :
4673 : : /*
4674 : : * We leave the first rollup out of consideration since it's the
4675 : : * one that matches the input sort order. We assign indexes "i"
4676 : : * to only those entries considered for hashing; the second loop,
4677 : : * below, must use the same condition.
4678 : : */
4679 : 501 : i = 0;
4680 [ + - + + : 1222 : for_each_from(lc, gd->rollups, 1)
+ + ]
4681 : : {
4682 : 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4683 : :
4684 [ + - ]: 721 : if (rollup->hashable)
4685 : : {
4686 : 721 : double sz = estimate_hashagg_tablesize(root,
4687 : : path,
4688 : : agg_costs,
4689 : : rollup->numGroups);
4690 : :
4691 : : /*
4692 : : * If sz is enormous, but hash_mem (and hence scale) is
4693 : : * small, avoid integer overflow here.
4694 : : */
4695 [ + + ]: 721 : k_weights[i] = (int) Min(floor(sz / scale),
4696 : : k_capacity + 1.0);
4697 : 721 : ++i;
4698 : : }
4699 : : }
4700 : :
4701 : : /*
4702 : : * Apply knapsack algorithm; compute the set of items which
4703 : : * maximizes the value stored (in this case the number of sorts
4704 : : * saved) while keeping the total size (approximately) within
4705 : : * capacity.
4706 : : */
4707 [ + - ]: 501 : if (i > 0)
4708 : 501 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4709 : :
4710 [ + - ]: 501 : if (!bms_is_empty(hash_items))
4711 : : {
4712 : 501 : rollups = list_make1(linitial(gd->rollups));
4713 : :
4714 : 501 : i = 0;
4715 [ + - + + : 1222 : for_each_from(lc, gd->rollups, 1)
+ + ]
4716 : : {
4717 : 721 : RollupData *rollup = lfirst_node(RollupData, lc);
4718 : :
4719 [ + - ]: 721 : if (rollup->hashable)
4720 : : {
4721 [ + + ]: 721 : if (bms_is_member(i, hash_items))
4722 : 691 : hash_sets = list_concat(hash_sets,
4723 : 691 : rollup->gsets_data);
4724 : : else
4725 : 30 : rollups = lappend(rollups, rollup);
4726 : 721 : ++i;
4727 : : }
4728 : : else
4729 : 0 : rollups = lappend(rollups, rollup);
4730 : : }
4731 : : }
4732 : : }
4733 : :
4734 [ + + + + ]: 896 : if (!rollups && hash_sets)
4735 : 20 : rollups = list_copy(gd->rollups);
4736 : :
4737 [ + + + + : 1703 : foreach(lc, hash_sets)
+ + ]
4738 : : {
4739 : 807 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
4740 : 807 : RollupData *rollup = makeNode(RollupData);
4741 : :
4742 : : Assert(gs->set != NIL);
4743 : :
4744 : 807 : rollup->groupClause = preprocess_groupclause(root, gs->set);
4745 : 807 : rollup->gsets_data = list_make1(gs);
4746 : 807 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4747 : : rollup->gsets_data,
4748 : : gd->tleref_to_colnum_map);
4749 : 807 : rollup->numGroups = gs->numGroups;
4750 : 807 : rollup->hashable = true;
4751 : 807 : rollup->is_hashed = true;
4752 : 807 : rollups = lcons(rollup, rollups);
4753 : : }
4754 : :
4755 [ + + ]: 896 : if (rollups)
4756 : : {
4757 : 521 : add_path(grouped_rel, (Path *)
4758 : 521 : create_groupingsets_path(root,
4759 : : grouped_rel,
4760 : : path,
4761 : 521 : (List *) parse->havingQual,
4762 : : AGG_MIXED,
4763 : : rollups,
4764 : : agg_costs));
4765 : : }
4766 : : }
4767 : :
4768 : : /*
4769 : : * Now try the simple sorted case.
4770 : : */
4771 [ + + ]: 975 : if (!gd->unsortable_sets)
4772 : 951 : add_path(grouped_rel, (Path *)
4773 : 951 : create_groupingsets_path(root,
4774 : : grouped_rel,
4775 : : path,
4776 : 951 : (List *) parse->havingQual,
4777 : : AGG_SORTED,
4778 : : gd->rollups,
4779 : : agg_costs));
4780 : : }
4781 : :
4782 : : /*
4783 : : * create_window_paths
4784 : : *
4785 : : * Build a new upperrel containing Paths for window-function evaluation.
4786 : : *
4787 : : * input_rel: contains the source-data Paths
4788 : : * input_target: result of make_window_input_target
4789 : : * output_target: what the topmost WindowAggPath should return
4790 : : * wflists: result of find_window_functions
4791 : : * activeWindows: result of select_active_windows
4792 : : *
4793 : : * Note: all Paths in input_rel are expected to return input_target.
4794 : : */
4795 : : static RelOptInfo *
4796 : 2331 : create_window_paths(PlannerInfo *root,
4797 : : RelOptInfo *input_rel,
4798 : : PathTarget *input_target,
4799 : : PathTarget *output_target,
4800 : : bool output_target_parallel_safe,
4801 : : WindowFuncLists *wflists,
4802 : : List *activeWindows)
4803 : : {
4804 : : RelOptInfo *window_rel;
4805 : : ListCell *lc;
4806 : :
4807 : : /* For now, do all work in the (WINDOW, NULL) upperrel */
4808 : 2331 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4809 : :
4810 : : /*
4811 : : * If the input relation is not parallel-safe, then the window relation
4812 : : * can't be parallel-safe, either. Otherwise, we need to examine the
4813 : : * target list and active windows for non-parallel-safe constructs.
4814 : : */
4815 [ + + - + : 2331 : if (input_rel->consider_parallel && output_target_parallel_safe &&
- - ]
4816 : 0 : is_parallel_safe(root, (Node *) activeWindows))
4817 : 0 : window_rel->consider_parallel = true;
4818 : :
4819 : : /*
4820 : : * If the input rel belongs to a single FDW, so does the window rel.
4821 : : */
4822 : 2331 : window_rel->serverid = input_rel->serverid;
4823 : 2331 : window_rel->userid = input_rel->userid;
4824 : 2331 : window_rel->useridiscurrent = input_rel->useridiscurrent;
4825 : 2331 : window_rel->fdwroutine = input_rel->fdwroutine;
4826 : :
4827 : : /*
4828 : : * Consider computing window functions starting from the existing
4829 : : * cheapest-total path (which will likely require a sort) as well as any
4830 : : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4831 : : */
4832 [ + - + + : 4918 : foreach(lc, input_rel->pathlist)
+ + ]
4833 : : {
4834 : 2587 : Path *path = (Path *) lfirst(lc);
4835 : : int presorted_keys;
4836 : :
4837 [ + + + + ]: 2843 : if (path == input_rel->cheapest_total_path ||
4838 : 256 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4839 : 115 : &presorted_keys) ||
4840 [ + + ]: 115 : presorted_keys > 0)
4841 : 2492 : create_one_window_path(root,
4842 : : window_rel,
4843 : : path,
4844 : : input_target,
4845 : : output_target,
4846 : : wflists,
4847 : : activeWindows);
4848 : : }
4849 : :
4850 : : /*
4851 : : * If there is an FDW that's responsible for all baserels of the query,
4852 : : * let it consider adding ForeignPaths.
4853 : : */
4854 [ + + ]: 2331 : if (window_rel->fdwroutine &&
4855 [ + - ]: 6 : window_rel->fdwroutine->GetForeignUpperPaths)
4856 : 6 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4857 : : input_rel, window_rel,
4858 : : NULL);
4859 : :
4860 : : /* Let extensions possibly add some more paths */
4861 [ - + ]: 2331 : if (create_upper_paths_hook)
4862 : 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4863 : : input_rel, window_rel, NULL);
4864 : :
4865 : : /* Now choose the best path(s) */
4866 : 2331 : set_cheapest(window_rel);
4867 : :
4868 : 2331 : return window_rel;
4869 : : }
4870 : :
4871 : : /*
4872 : : * Stack window-function implementation steps atop the given Path, and
4873 : : * add the result to window_rel.
4874 : : *
4875 : : * window_rel: upperrel to contain result
4876 : : * path: input Path to use (must return input_target)
4877 : : * input_target: result of make_window_input_target
4878 : : * output_target: what the topmost WindowAggPath should return
4879 : : * wflists: result of find_window_functions
4880 : : * activeWindows: result of select_active_windows
4881 : : */
4882 : : static void
4883 : 2492 : create_one_window_path(PlannerInfo *root,
4884 : : RelOptInfo *window_rel,
4885 : : Path *path,
4886 : : PathTarget *input_target,
4887 : : PathTarget *output_target,
4888 : : WindowFuncLists *wflists,
4889 : : List *activeWindows)
4890 : : {
4891 : : PathTarget *window_target;
4892 : : ListCell *l;
4893 : 2492 : List *topqual = NIL;
4894 : :
4895 : : /*
4896 : : * Since each window clause could require a different sort order, we stack
4897 : : * up a WindowAgg node for each clause, with sort steps between them as
4898 : : * needed. (We assume that select_active_windows chose a good order for
4899 : : * executing the clauses in.)
4900 : : *
4901 : : * input_target should contain all Vars and Aggs needed for the result.
4902 : : * (In some cases we wouldn't need to propagate all of these all the way
4903 : : * to the top, since they might only be needed as inputs to WindowFuncs.
4904 : : * It's probably not worth trying to optimize that though.) It must also
4905 : : * contain all window partitioning and sorting expressions, to ensure
4906 : : * they're computed only once at the bottom of the stack (that's critical
4907 : : * for volatile functions). As we climb up the stack, we'll add outputs
4908 : : * for the WindowFuncs computed at each level.
4909 : : */
4910 : 2492 : window_target = input_target;
4911 : :
4912 [ + - + + : 5149 : foreach(l, activeWindows)
+ + ]
4913 : : {
4914 : 2657 : WindowClause *wc = lfirst_node(WindowClause, l);
4915 : : List *window_pathkeys;
4916 : 2657 : List *runcondition = NIL;
4917 : : int presorted_keys;
4918 : : bool is_sorted;
4919 : : bool topwindow;
4920 : : ListCell *lc2;
4921 : :
4922 : 2657 : window_pathkeys = make_pathkeys_for_window(root,
4923 : : wc,
4924 : : root->processed_tlist);
4925 : :
4926 : 2657 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4927 : : path->pathkeys,
4928 : : &presorted_keys);
4929 : :
4930 : : /* Sort if necessary */
4931 [ + + ]: 2657 : if (!is_sorted)
4932 : : {
4933 : : /*
4934 : : * No presorted keys or incremental sort disabled, just perform a
4935 : : * complete sort.
4936 : : */
4937 [ + + - + ]: 1929 : if (presorted_keys == 0 || !enable_incremental_sort)
4938 : 1879 : path = (Path *) create_sort_path(root, window_rel,
4939 : : path,
4940 : : window_pathkeys,
4941 : : -1.0);
4942 : : else
4943 : : {
4944 : : /*
4945 : : * Since we have presorted keys and incremental sort is
4946 : : * enabled, just use incremental sort.
4947 : : */
4948 : 50 : path = (Path *) create_incremental_sort_path(root,
4949 : : window_rel,
4950 : : path,
4951 : : window_pathkeys,
4952 : : presorted_keys,
4953 : : -1.0);
4954 : : }
4955 : : }
4956 : :
4957 [ + + ]: 2657 : if (lnext(activeWindows, l))
4958 : : {
4959 : : /*
4960 : : * Add the current WindowFuncs to the output target for this
4961 : : * intermediate WindowAggPath. We must copy window_target to
4962 : : * avoid changing the previous path's target.
4963 : : *
4964 : : * Note: a WindowFunc adds nothing to the target's eval costs; but
4965 : : * we do need to account for the increase in tlist width.
4966 : : */
4967 : 165 : int64 tuple_width = window_target->width;
4968 : :
4969 : 165 : window_target = copy_pathtarget(window_target);
4970 [ + - + + : 390 : foreach(lc2, wflists->windowFuncs[wc->winref])
+ + ]
4971 : : {
4972 : 225 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4973 : :
4974 : 225 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
4975 : 225 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4976 : : }
4977 : 165 : window_target->width = clamp_width_est(tuple_width);
4978 : : }
4979 : : else
4980 : : {
4981 : : /* Install the goal target in the topmost WindowAgg */
4982 : 2492 : window_target = output_target;
4983 : : }
4984 : :
4985 : : /* mark the final item in the list as the top-level window */
4986 : 2657 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4987 : :
4988 : : /*
4989 : : * Collect the WindowFuncRunConditions from each WindowFunc and
4990 : : * convert them into OpExprs
4991 : : */
4992 [ + - + + : 6047 : foreach(lc2, wflists->windowFuncs[wc->winref])
+ + ]
4993 : : {
4994 : : ListCell *lc3;
4995 : 3390 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4996 : :
4997 [ + + + + : 3560 : foreach(lc3, wfunc->runCondition)
+ + ]
4998 : : {
4999 : 170 : WindowFuncRunCondition *wfuncrc =
5000 : : lfirst_node(WindowFuncRunCondition, lc3);
5001 : : Expr *opexpr;
5002 : : Expr *leftop;
5003 : : Expr *rightop;
5004 : :
5005 [ + + ]: 170 : if (wfuncrc->wfunc_left)
5006 : : {
5007 : 155 : leftop = (Expr *) copyObject(wfunc);
5008 : 155 : rightop = copyObject(wfuncrc->arg);
5009 : : }
5010 : : else
5011 : : {
5012 : 15 : leftop = copyObject(wfuncrc->arg);
5013 : 15 : rightop = (Expr *) copyObject(wfunc);
5014 : : }
5015 : :
5016 : 170 : opexpr = make_opclause(wfuncrc->opno,
5017 : : BOOLOID,
5018 : : false,
5019 : : leftop,
5020 : : rightop,
5021 : : InvalidOid,
5022 : : wfuncrc->inputcollid);
5023 : :
5024 : 170 : runcondition = lappend(runcondition, opexpr);
5025 : :
5026 [ + + ]: 170 : if (!topwindow)
5027 : 20 : topqual = lappend(topqual, opexpr);
5028 : : }
5029 : : }
5030 : :
5031 : : path = (Path *)
5032 [ + + ]: 2657 : create_windowagg_path(root, window_rel, path, window_target,
5033 : 2657 : wflists->windowFuncs[wc->winref],
5034 : : runcondition, wc,
5035 : : topwindow ? topqual : NIL, topwindow);
5036 : : }
5037 : :
5038 : 2492 : add_path(window_rel, path);
5039 : 2492 : }
5040 : :
5041 : : /*
5042 : : * create_distinct_paths
5043 : : *
5044 : : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
5045 : : *
5046 : : * input_rel: contains the source-data Paths
5047 : : * target: the pathtarget for the result Paths to compute
5048 : : *
5049 : : * Note: input paths should already compute the desired pathtarget, since
5050 : : * Sort/Unique won't project anything.
5051 : : */
5052 : : static RelOptInfo *
5053 : 2013 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5054 : : PathTarget *target)
5055 : : {
5056 : : RelOptInfo *distinct_rel;
5057 : :
5058 : : /* For now, do all work in the (DISTINCT, NULL) upperrel */
5059 : 2013 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
5060 : :
5061 : : /*
5062 : : * We don't compute anything at this level, so distinct_rel will be
5063 : : * parallel-safe if the input rel is parallel-safe. In particular, if
5064 : : * there is a DISTINCT ON (...) clause, any path for the input_rel will
5065 : : * output those expressions, and will not be parallel-safe unless those
5066 : : * expressions are parallel-safe.
5067 : : */
5068 : 2013 : distinct_rel->consider_parallel = input_rel->consider_parallel;
5069 : :
5070 : : /*
5071 : : * If the input rel belongs to a single FDW, so does the distinct_rel.
5072 : : */
5073 : 2013 : distinct_rel->serverid = input_rel->serverid;
5074 : 2013 : distinct_rel->userid = input_rel->userid;
5075 : 2013 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5076 : 2013 : distinct_rel->fdwroutine = input_rel->fdwroutine;
5077 : :
5078 : : /* build distinct paths based on input_rel's pathlist */
5079 : 2013 : create_final_distinct_paths(root, input_rel, distinct_rel);
5080 : :
5081 : : /* now build distinct paths based on input_rel's partial_pathlist */
5082 : 2013 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
5083 : :
5084 : : /* Give a helpful error if we failed to create any paths */
5085 [ - + ]: 2013 : if (distinct_rel->pathlist == NIL)
5086 [ # # ]: 0 : ereport(ERROR,
5087 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
5088 : : errmsg("could not implement DISTINCT"),
5089 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
5090 : :
5091 : : /*
5092 : : * If there is an FDW that's responsible for all baserels of the query,
5093 : : * let it consider adding ForeignPaths.
5094 : : */
5095 [ + + ]: 2013 : if (distinct_rel->fdwroutine &&
5096 [ + - ]: 8 : distinct_rel->fdwroutine->GetForeignUpperPaths)
5097 : 8 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5098 : : UPPERREL_DISTINCT,
5099 : : input_rel,
5100 : : distinct_rel,
5101 : : NULL);
5102 : :
5103 : : /* Let extensions possibly add some more paths */
5104 [ - + ]: 2013 : if (create_upper_paths_hook)
5105 : 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
5106 : : distinct_rel, NULL);
5107 : :
5108 : : /* Now choose the best path(s) */
5109 : 2013 : set_cheapest(distinct_rel);
5110 : :
5111 : 2013 : return distinct_rel;
5112 : : }
5113 : :
5114 : : /*
5115 : : * create_partial_distinct_paths
5116 : : *
5117 : : * Process 'input_rel' partial paths and add unique/aggregate paths to the
5118 : : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
5119 : : * paths on top and add a final unique/aggregate path to remove any duplicate
5120 : : * produced from combining rows from parallel workers.
5121 : : */
5122 : : static void
5123 : 2013 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5124 : : RelOptInfo *final_distinct_rel,
5125 : : PathTarget *target)
5126 : : {
5127 : : RelOptInfo *partial_distinct_rel;
5128 : : Query *parse;
5129 : : List *distinctExprs;
5130 : : double numDistinctRows;
5131 : : Path *cheapest_partial_path;
5132 : : ListCell *lc;
5133 : :
5134 : : /* nothing to do when there are no partial paths in the input rel */
5135 [ + + + + ]: 2013 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
5136 : 1923 : return;
5137 : :
5138 : 90 : parse = root->parse;
5139 : :
5140 : : /* can't do parallel DISTINCT ON */
5141 [ - + ]: 90 : if (parse->hasDistinctOn)
5142 : 0 : return;
5143 : :
5144 : 90 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
5145 : : NULL);
5146 : 90 : partial_distinct_rel->reltarget = target;
5147 : 90 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
5148 : :
5149 : : /*
5150 : : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
5151 : : */
5152 : 90 : partial_distinct_rel->serverid = input_rel->serverid;
5153 : 90 : partial_distinct_rel->userid = input_rel->userid;
5154 : 90 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
5155 : 90 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
5156 : :
5157 : 90 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5158 : :
5159 : 90 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5160 : : parse->targetList);
5161 : :
5162 : : /* estimate how many distinct rows we'll get from each worker */
5163 : 90 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5164 : : cheapest_partial_path->rows,
5165 : : NULL, NULL);
5166 : :
5167 : : /*
5168 : : * Try sorting the cheapest path and incrementally sorting any paths with
5169 : : * presorted keys and put a unique paths atop of those. We'll also
5170 : : * attempt to reorder the required pathkeys to match the input path's
5171 : : * pathkeys as much as possible, in hopes of avoiding a possible need to
5172 : : * re-sort.
5173 : : */
5174 [ + - ]: 90 : if (grouping_is_sortable(root->processed_distinctClause))
5175 : : {
5176 [ + - + + : 193 : foreach(lc, input_rel->partial_pathlist)
+ + ]
5177 : : {
5178 : 103 : Path *input_path = (Path *) lfirst(lc);
5179 : : Path *sorted_path;
5180 : 103 : List *useful_pathkeys_list = NIL;
5181 : :
5182 : : useful_pathkeys_list =
5183 : 103 : get_useful_pathkeys_for_distinct(root,
5184 : : root->distinct_pathkeys,
5185 : : input_path->pathkeys);
5186 : : Assert(list_length(useful_pathkeys_list) > 0);
5187 : :
5188 [ + - + + : 319 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
+ + ]
5189 : : {
5190 : 113 : sorted_path = make_ordered_path(root,
5191 : : partial_distinct_rel,
5192 : : input_path,
5193 : : cheapest_partial_path,
5194 : : useful_pathkeys,
5195 : : -1.0);
5196 : :
5197 [ + + ]: 113 : if (sorted_path == NULL)
5198 : 8 : continue;
5199 : :
5200 : : /*
5201 : : * An empty distinct_pathkeys means all tuples have the same
5202 : : * value for the DISTINCT clause. See
5203 : : * create_final_distinct_paths()
5204 : : */
5205 [ + + ]: 105 : if (root->distinct_pathkeys == NIL)
5206 : : {
5207 : : Node *limitCount;
5208 : :
5209 : 5 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5210 : : sizeof(int64),
5211 : : Int64GetDatum(1), false,
5212 : : true);
5213 : :
5214 : : /*
5215 : : * Apply a LimitPath onto the partial path to restrict the
5216 : : * tuples from each worker to 1.
5217 : : * create_final_distinct_paths will need to apply an
5218 : : * additional LimitPath to restrict this to a single row
5219 : : * after the Gather node. If the query already has a
5220 : : * LIMIT clause, then we could end up with three Limit
5221 : : * nodes in the final plan. Consolidating the top two of
5222 : : * these could be done, but does not seem worth troubling
5223 : : * over.
5224 : : */
5225 : 5 : add_partial_path(partial_distinct_rel, (Path *)
5226 : 5 : create_limit_path(root, partial_distinct_rel,
5227 : : sorted_path,
5228 : : NULL,
5229 : : limitCount,
5230 : : LIMIT_OPTION_COUNT,
5231 : : 0, 1));
5232 : : }
5233 : : else
5234 : : {
5235 : 100 : add_partial_path(partial_distinct_rel, (Path *)
5236 : 100 : create_unique_path(root, partial_distinct_rel,
5237 : : sorted_path,
5238 : 100 : list_length(root->distinct_pathkeys),
5239 : : numDistinctRows));
5240 : : }
5241 : : }
5242 : : }
5243 : : }
5244 : :
5245 : : /*
5246 : : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5247 : : * we're not on the hook to ensure we do our best to create at least one
5248 : : * path here, we treat enable_hashagg as a hard off-switch rather than the
5249 : : * slightly softer variant in create_final_distinct_paths.
5250 : : */
5251 [ + + + - ]: 90 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5252 : : {
5253 : 65 : add_partial_path(partial_distinct_rel, (Path *)
5254 : 65 : create_agg_path(root,
5255 : : partial_distinct_rel,
5256 : : cheapest_partial_path,
5257 : : cheapest_partial_path->pathtarget,
5258 : : AGG_HASHED,
5259 : : AGGSPLIT_SIMPLE,
5260 : : root->processed_distinctClause,
5261 : : NIL,
5262 : : NULL,
5263 : : numDistinctRows));
5264 : : }
5265 : :
5266 : : /*
5267 : : * If there is an FDW that's responsible for all baserels of the query,
5268 : : * let it consider adding ForeignPaths.
5269 : : */
5270 [ - + ]: 90 : if (partial_distinct_rel->fdwroutine &&
5271 [ # # ]: 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5272 : 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5273 : : UPPERREL_PARTIAL_DISTINCT,
5274 : : input_rel,
5275 : : partial_distinct_rel,
5276 : : NULL);
5277 : :
5278 : : /* Let extensions possibly add some more partial paths */
5279 [ - + ]: 90 : if (create_upper_paths_hook)
5280 : 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5281 : : input_rel, partial_distinct_rel, NULL);
5282 : :
5283 [ + - ]: 90 : if (partial_distinct_rel->partial_pathlist != NIL)
5284 : : {
5285 : 90 : generate_useful_gather_paths(root, partial_distinct_rel, true);
5286 : 90 : set_cheapest(partial_distinct_rel);
5287 : :
5288 : : /*
5289 : : * Finally, create paths to distinctify the final result. This step
5290 : : * is needed to remove any duplicates due to combining rows from
5291 : : * parallel workers.
5292 : : */
5293 : 90 : create_final_distinct_paths(root, partial_distinct_rel,
5294 : : final_distinct_rel);
5295 : : }
5296 : : }
5297 : :
5298 : : /*
5299 : : * create_final_distinct_paths
5300 : : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5301 : : *
5302 : : * input_rel: contains the source-data paths
5303 : : * distinct_rel: destination relation for storing created paths
5304 : : */
5305 : : static RelOptInfo *
5306 : 2103 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5307 : : RelOptInfo *distinct_rel)
5308 : : {
5309 : 2103 : Query *parse = root->parse;
5310 : 2103 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5311 : : double numDistinctRows;
5312 : : bool allow_hash;
5313 : :
5314 : : /* Estimate number of distinct rows there will be */
5315 [ + + + - : 2103 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
+ + ]
5316 [ - + ]: 2044 : root->hasHavingQual)
5317 : : {
5318 : : /*
5319 : : * If there was grouping or aggregation, use the number of input rows
5320 : : * as the estimated number of DISTINCT rows (ie, assume the input is
5321 : : * already mostly unique).
5322 : : */
5323 : 59 : numDistinctRows = cheapest_input_path->rows;
5324 : : }
5325 : : else
5326 : : {
5327 : : /*
5328 : : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5329 : : */
5330 : : List *distinctExprs;
5331 : :
5332 : 2044 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5333 : : parse->targetList);
5334 : 2044 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5335 : : cheapest_input_path->rows,
5336 : : NULL, NULL);
5337 : : }
5338 : :
5339 : : /*
5340 : : * Consider sort-based implementations of DISTINCT, if possible.
5341 : : */
5342 [ + + ]: 2103 : if (grouping_is_sortable(root->processed_distinctClause))
5343 : : {
5344 : : /*
5345 : : * Firstly, if we have any adequately-presorted paths, just stick a
5346 : : * Unique node on those. We also, consider doing an explicit sort of
5347 : : * the cheapest input path and Unique'ing that. If any paths have
5348 : : * presorted keys then we'll create an incremental sort atop of those
5349 : : * before adding a unique node on the top. We'll also attempt to
5350 : : * reorder the required pathkeys to match the input path's pathkeys as
5351 : : * much as possible, in hopes of avoiding a possible need to re-sort.
5352 : : *
5353 : : * When we have DISTINCT ON, we must sort by the more rigorous of
5354 : : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5355 : : * Also, if we do have to do an explicit sort, we might as well use
5356 : : * the more rigorous ordering to avoid a second sort later. (Note
5357 : : * that the parser will have ensured that one clause is a prefix of
5358 : : * the other.)
5359 : : */
5360 : : List *needed_pathkeys;
5361 : : ListCell *lc;
5362 [ + + ]: 2098 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5363 : :
5364 [ + + + + ]: 2326 : if (parse->hasDistinctOn &&
5365 : 228 : list_length(root->distinct_pathkeys) <
5366 : 228 : list_length(root->sort_pathkeys))
5367 : 93 : needed_pathkeys = root->sort_pathkeys;
5368 : : else
5369 : 2005 : needed_pathkeys = root->distinct_pathkeys;
5370 : :
5371 [ + - + + : 5522 : foreach(lc, input_rel->pathlist)
+ + ]
5372 : : {
5373 : 3424 : Path *input_path = (Path *) lfirst(lc);
5374 : : Path *sorted_path;
5375 : 3424 : List *useful_pathkeys_list = NIL;
5376 : :
5377 : : useful_pathkeys_list =
5378 : 3424 : get_useful_pathkeys_for_distinct(root,
5379 : : needed_pathkeys,
5380 : : input_path->pathkeys);
5381 : : Assert(list_length(useful_pathkeys_list) > 0);
5382 : :
5383 [ + - + + : 10787 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
+ + ]
5384 : : {
5385 : 3939 : sorted_path = make_ordered_path(root,
5386 : : distinct_rel,
5387 : : input_path,
5388 : : cheapest_input_path,
5389 : : useful_pathkeys,
5390 : : limittuples);
5391 : :
5392 [ + + ]: 3939 : if (sorted_path == NULL)
5393 : 546 : continue;
5394 : :
5395 : : /*
5396 : : * distinct_pathkeys may have become empty if all of the
5397 : : * pathkeys were determined to be redundant. If all of the
5398 : : * pathkeys are redundant then each DISTINCT target must only
5399 : : * allow a single value, therefore all resulting tuples must
5400 : : * be identical (or at least indistinguishable by an equality
5401 : : * check). We can uniquify these tuples simply by just taking
5402 : : * the first tuple. All we do here is add a path to do "LIMIT
5403 : : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5404 : : * still have a non-NIL sort_pathkeys list, so we must still
5405 : : * only do this with paths which are correctly sorted by
5406 : : * sort_pathkeys.
5407 : : */
5408 [ + + ]: 3393 : if (root->distinct_pathkeys == NIL)
5409 : : {
5410 : : Node *limitCount;
5411 : :
5412 : 116 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5413 : : sizeof(int64),
5414 : : Int64GetDatum(1), false,
5415 : : true);
5416 : :
5417 : : /*
5418 : : * If the query already has a LIMIT clause, then we could
5419 : : * end up with a duplicate LimitPath in the final plan.
5420 : : * That does not seem worth troubling over too much.
5421 : : */
5422 : 116 : add_path(distinct_rel, (Path *)
5423 : 116 : create_limit_path(root, distinct_rel, sorted_path,
5424 : : NULL, limitCount,
5425 : : LIMIT_OPTION_COUNT, 0, 1));
5426 : : }
5427 : : else
5428 : : {
5429 : 3277 : add_path(distinct_rel, (Path *)
5430 : 3277 : create_unique_path(root, distinct_rel,
5431 : : sorted_path,
5432 : 3277 : list_length(root->distinct_pathkeys),
5433 : : numDistinctRows));
5434 : : }
5435 : : }
5436 : : }
5437 : : }
5438 : :
5439 : : /*
5440 : : * Consider hash-based implementations of DISTINCT, if possible.
5441 : : *
5442 : : * If we were not able to make any other types of path, we *must* hash or
5443 : : * die trying. If we do have other choices, there are two things that
5444 : : * should prevent selection of hashing: if the query uses DISTINCT ON
5445 : : * (because it won't really have the expected behavior if we hash), or if
5446 : : * enable_hashagg is off.
5447 : : *
5448 : : * Note: grouping_is_hashable() is much more expensive to check than the
5449 : : * other gating conditions, so we want to do it last.
5450 : : */
5451 [ + + ]: 2103 : if (distinct_rel->pathlist == NIL)
5452 : 5 : allow_hash = true; /* we have no alternatives */
5453 [ + + + + ]: 2098 : else if (parse->hasDistinctOn || !enable_hashagg)
5454 : 353 : allow_hash = false; /* policy-based decision not to hash */
5455 : : else
5456 : 1745 : allow_hash = true; /* default */
5457 : :
5458 [ + + + - ]: 2103 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5459 : : {
5460 : : /* Generate hashed aggregate path --- no sort needed */
5461 : 1750 : add_path(distinct_rel, (Path *)
5462 : 1750 : create_agg_path(root,
5463 : : distinct_rel,
5464 : : cheapest_input_path,
5465 : : cheapest_input_path->pathtarget,
5466 : : AGG_HASHED,
5467 : : AGGSPLIT_SIMPLE,
5468 : : root->processed_distinctClause,
5469 : : NIL,
5470 : : NULL,
5471 : : numDistinctRows));
5472 : : }
5473 : :
5474 : 2103 : return distinct_rel;
5475 : : }
5476 : :
5477 : : /*
5478 : : * get_useful_pathkeys_for_distinct
5479 : : * Get useful orderings of pathkeys for distinctClause by reordering
5480 : : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5481 : : *
5482 : : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5483 : : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5484 : : */
5485 : : static List *
5486 : 3527 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5487 : : List *path_pathkeys)
5488 : : {
5489 : 3527 : List *useful_pathkeys_list = NIL;
5490 : 3527 : List *useful_pathkeys = NIL;
5491 : :
5492 : : /* always include the given 'needed_pathkeys' */
5493 : 3527 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5494 : : needed_pathkeys);
5495 : :
5496 [ - + ]: 3527 : if (!enable_distinct_reordering)
5497 : 0 : return useful_pathkeys_list;
5498 : :
5499 : : /*
5500 : : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5501 : : * that match 'needed_pathkeys', but only up to the longest matching
5502 : : * prefix.
5503 : : *
5504 : : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5505 : : * list matches initial distinctClause pathkeys; otherwise, it won't have
5506 : : * the desired behavior.
5507 : : */
5508 [ + + + + : 8774 : foreach_node(PathKey, pathkey, path_pathkeys)
+ + ]
5509 : : {
5510 : : /*
5511 : : * The PathKey nodes are canonical, so they can be checked for
5512 : : * equality by simple pointer comparison.
5513 : : */
5514 [ + + ]: 1742 : if (!list_member_ptr(needed_pathkeys, pathkey))
5515 : 7 : break;
5516 [ + + ]: 1735 : if (root->parse->hasDistinctOn &&
5517 [ + + ]: 140 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5518 : 15 : break;
5519 : :
5520 : 1720 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5521 : : }
5522 : :
5523 : : /* If no match at all, no point in reordering needed_pathkeys */
5524 [ + + ]: 3527 : if (useful_pathkeys == NIL)
5525 : 2055 : return useful_pathkeys_list;
5526 : :
5527 : : /*
5528 : : * If not full match, the resulting pathkey list is not useful without
5529 : : * incremental sort.
5530 : : */
5531 [ + + ]: 1472 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5532 [ + + ]: 939 : !enable_incremental_sort)
5533 : 48 : return useful_pathkeys_list;
5534 : :
5535 : : /* Append the remaining PathKey nodes in needed_pathkeys */
5536 : 1424 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5537 : : needed_pathkeys);
5538 : :
5539 : : /*
5540 : : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5541 : : * just drop it.
5542 : : */
5543 [ + + ]: 1424 : if (compare_pathkeys(needed_pathkeys,
5544 : : useful_pathkeys) == PATHKEYS_EQUAL)
5545 : 899 : return useful_pathkeys_list;
5546 : :
5547 : 525 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5548 : : useful_pathkeys);
5549 : :
5550 : 525 : return useful_pathkeys_list;
5551 : : }
5552 : :
5553 : : /*
5554 : : * create_ordered_paths
5555 : : *
5556 : : * Build a new upperrel containing Paths for ORDER BY evaluation.
5557 : : *
5558 : : * All paths in the result must satisfy the ORDER BY ordering.
5559 : : * The only new paths we need consider are an explicit full sort
5560 : : * and incremental sort on the cheapest-total existing path.
5561 : : *
5562 : : * input_rel: contains the source-data Paths
5563 : : * target: the output tlist the result Paths must emit
5564 : : * limit_tuples: estimated bound on the number of output tuples,
5565 : : * or -1 if no LIMIT or couldn't estimate
5566 : : *
5567 : : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5568 : : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5569 : : */
5570 : : static RelOptInfo *
5571 : 60982 : create_ordered_paths(PlannerInfo *root,
5572 : : RelOptInfo *input_rel,
5573 : : PathTarget *target,
5574 : : bool target_parallel_safe,
5575 : : double limit_tuples)
5576 : : {
5577 : 60982 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5578 : : RelOptInfo *ordered_rel;
5579 : : ListCell *lc;
5580 : :
5581 : : /* For now, do all work in the (ORDERED, NULL) upperrel */
5582 : 60982 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5583 : :
5584 : : /*
5585 : : * If the input relation is not parallel-safe, then the ordered relation
5586 : : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5587 : : * target list is parallel-safe.
5588 : : */
5589 [ + + + + ]: 60982 : if (input_rel->consider_parallel && target_parallel_safe)
5590 : 42826 : ordered_rel->consider_parallel = true;
5591 : :
5592 : : /* Assume that the same path generation strategies are allowed. */
5593 : 60982 : ordered_rel->pgs_mask = input_rel->pgs_mask;
5594 : :
5595 : : /*
5596 : : * If the input rel belongs to a single FDW, so does the ordered_rel.
5597 : : */
5598 : 60982 : ordered_rel->serverid = input_rel->serverid;
5599 : 60982 : ordered_rel->userid = input_rel->userid;
5600 : 60982 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
5601 : 60982 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5602 : :
5603 [ + - + + : 155134 : foreach(lc, input_rel->pathlist)
+ + ]
5604 : : {
5605 : 94152 : Path *input_path = (Path *) lfirst(lc);
5606 : : Path *sorted_path;
5607 : : bool is_sorted;
5608 : : int presorted_keys;
5609 : :
5610 : 94152 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5611 : : input_path->pathkeys, &presorted_keys);
5612 : :
5613 [ + + ]: 94152 : if (is_sorted)
5614 : 36426 : sorted_path = input_path;
5615 : : else
5616 : : {
5617 : : /*
5618 : : * Try at least sorting the cheapest path and also try
5619 : : * incrementally sorting any path which is partially sorted
5620 : : * already (no need to deal with paths which have presorted keys
5621 : : * when incremental sort is disabled unless it's the cheapest
5622 : : * input path).
5623 : : */
5624 [ + + ]: 57726 : if (input_path != cheapest_input_path &&
5625 [ + + + + ]: 4352 : (presorted_keys == 0 || !enable_incremental_sort))
5626 : 1614 : continue;
5627 : :
5628 : : /*
5629 : : * We've no need to consider both a sort and incremental sort.
5630 : : * We'll just do a sort if there are no presorted keys and an
5631 : : * incremental sort when there are presorted keys.
5632 : : */
5633 [ + + + + ]: 56112 : if (presorted_keys == 0 || !enable_incremental_sort)
5634 : 52688 : sorted_path = (Path *) create_sort_path(root,
5635 : : ordered_rel,
5636 : : input_path,
5637 : : root->sort_pathkeys,
5638 : : limit_tuples);
5639 : : else
5640 : 3424 : sorted_path = (Path *) create_incremental_sort_path(root,
5641 : : ordered_rel,
5642 : : input_path,
5643 : : root->sort_pathkeys,
5644 : : presorted_keys,
5645 : : limit_tuples);
5646 : : }
5647 : :
5648 : : /*
5649 : : * If the pathtarget of the result path has different expressions from
5650 : : * the target to be applied, a projection step is needed.
5651 : : */
5652 [ + + ]: 92538 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5653 : 303 : sorted_path = apply_projection_to_path(root, ordered_rel,
5654 : : sorted_path, target);
5655 : :
5656 : 92538 : add_path(ordered_rel, sorted_path);
5657 : : }
5658 : :
5659 : : /*
5660 : : * generate_gather_paths() will have already generated a simple Gather
5661 : : * path for the best parallel path, if any, and the loop above will have
5662 : : * considered sorting it. Similarly, generate_gather_paths() will also
5663 : : * have generated order-preserving Gather Merge plans which can be used
5664 : : * without sorting if they happen to match the sort_pathkeys, and the loop
5665 : : * above will have handled those as well. However, there's one more
5666 : : * possibility: it may make sense to sort the cheapest partial path or
5667 : : * incrementally sort any partial path that is partially sorted according
5668 : : * to the required output order and then use Gather Merge.
5669 : : */
5670 [ + + + + ]: 60982 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5671 [ + + ]: 42706 : input_rel->partial_pathlist != NIL)
5672 : : {
5673 : : Path *cheapest_partial_path;
5674 : :
5675 : 2378 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5676 : :
5677 [ + - + + : 5311 : foreach(lc, input_rel->partial_pathlist)
+ + ]
5678 : : {
5679 : 2933 : Path *input_path = (Path *) lfirst(lc);
5680 : : Path *sorted_path;
5681 : : bool is_sorted;
5682 : : int presorted_keys;
5683 : : double total_groups;
5684 : :
5685 : 2933 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5686 : : input_path->pathkeys,
5687 : : &presorted_keys);
5688 : :
5689 [ + + ]: 2933 : if (is_sorted)
5690 : 476 : continue;
5691 : :
5692 : : /*
5693 : : * Try at least sorting the cheapest path and also try
5694 : : * incrementally sorting any path which is partially sorted
5695 : : * already (no need to deal with paths which have presorted keys
5696 : : * when incremental sort is disabled unless it's the cheapest
5697 : : * partial path).
5698 : : */
5699 [ + + ]: 2457 : if (input_path != cheapest_partial_path &&
5700 [ + - - + ]: 101 : (presorted_keys == 0 || !enable_incremental_sort))
5701 : 0 : continue;
5702 : :
5703 : : /*
5704 : : * We've no need to consider both a sort and incremental sort.
5705 : : * We'll just do a sort if there are no presorted keys and an
5706 : : * incremental sort when there are presorted keys.
5707 : : */
5708 [ + + + + ]: 2457 : if (presorted_keys == 0 || !enable_incremental_sort)
5709 : 2341 : sorted_path = (Path *) create_sort_path(root,
5710 : : ordered_rel,
5711 : : input_path,
5712 : : root->sort_pathkeys,
5713 : : limit_tuples);
5714 : : else
5715 : 116 : sorted_path = (Path *) create_incremental_sort_path(root,
5716 : : ordered_rel,
5717 : : input_path,
5718 : : root->sort_pathkeys,
5719 : : presorted_keys,
5720 : : limit_tuples);
5721 : 2457 : total_groups = compute_gather_rows(sorted_path);
5722 : : sorted_path = (Path *)
5723 : 2457 : create_gather_merge_path(root, ordered_rel,
5724 : : sorted_path,
5725 : : sorted_path->pathtarget,
5726 : : root->sort_pathkeys, NULL,
5727 : : &total_groups);
5728 : :
5729 : : /*
5730 : : * If the pathtarget of the result path has different expressions
5731 : : * from the target to be applied, a projection step is needed.
5732 : : */
5733 [ + + ]: 2457 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
5734 : 5 : sorted_path = apply_projection_to_path(root, ordered_rel,
5735 : : sorted_path, target);
5736 : :
5737 : 2457 : add_path(ordered_rel, sorted_path);
5738 : : }
5739 : : }
5740 : :
5741 : : /*
5742 : : * If there is an FDW that's responsible for all baserels of the query,
5743 : : * let it consider adding ForeignPaths.
5744 : : */
5745 [ + + ]: 60982 : if (ordered_rel->fdwroutine &&
5746 [ + + ]: 197 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5747 : 189 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5748 : : input_rel, ordered_rel,
5749 : : NULL);
5750 : :
5751 : : /* Let extensions possibly add some more paths */
5752 [ - + ]: 60982 : if (create_upper_paths_hook)
5753 : 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5754 : : input_rel, ordered_rel, NULL);
5755 : :
5756 : : /*
5757 : : * No need to bother with set_cheapest here; grouping_planner does not
5758 : : * need us to do it.
5759 : : */
5760 : : Assert(ordered_rel->pathlist != NIL);
5761 : :
5762 : 60982 : return ordered_rel;
5763 : : }
5764 : :
5765 : :
5766 : : /*
5767 : : * make_group_input_target
5768 : : * Generate appropriate PathTarget for initial input to grouping nodes.
5769 : : *
5770 : : * If there is grouping or aggregation, the scan/join subplan cannot emit
5771 : : * the query's final targetlist; for example, it certainly can't emit any
5772 : : * aggregate function calls. This routine generates the correct target
5773 : : * for the scan/join subplan.
5774 : : *
5775 : : * The query target list passed from the parser already contains entries
5776 : : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5777 : : * for variables used only in HAVING clauses; so we need to add those
5778 : : * variables to the subplan target list. Also, we flatten all expressions
5779 : : * except GROUP BY items into their component variables; other expressions
5780 : : * will be computed by the upper plan nodes rather than by the subplan.
5781 : : * For example, given a query like
5782 : : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5783 : : * we want to pass this targetlist to the subplan:
5784 : : * a+b,c,d
5785 : : * where the a+b target will be used by the Sort/Group steps, and the
5786 : : * other targets will be used for computing the final results.
5787 : : *
5788 : : * 'final_target' is the query's final target list (in PathTarget form)
5789 : : *
5790 : : * The result is the PathTarget to be computed by the Paths returned from
5791 : : * query_planner().
5792 : : */
5793 : : static PathTarget *
5794 : 34089 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5795 : : {
5796 : 34089 : Query *parse = root->parse;
5797 : : PathTarget *input_target;
5798 : : List *non_group_cols;
5799 : : List *non_group_vars;
5800 : : int i;
5801 : : ListCell *lc;
5802 : :
5803 : : /*
5804 : : * We must build a target containing all grouping columns, plus any other
5805 : : * Vars mentioned in the query's targetlist and HAVING qual.
5806 : : */
5807 : 34089 : input_target = create_empty_pathtarget();
5808 : 34089 : non_group_cols = NIL;
5809 : :
5810 : 34089 : i = 0;
5811 [ + + + + : 86009 : foreach(lc, final_target->exprs)
+ + ]
5812 : : {
5813 : 51920 : Expr *expr = (Expr *) lfirst(lc);
5814 [ + - ]: 51920 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5815 : :
5816 [ + + + + : 60197 : if (sgref && root->processed_groupClause &&
+ + ]
5817 : 8277 : get_sortgroupref_clause_noerr(sgref,
5818 : : root->processed_groupClause) != NULL)
5819 : : {
5820 : : /*
5821 : : * It's a grouping column, so add it to the input target as-is.
5822 : : *
5823 : : * Note that the target is logically below the grouping step. So
5824 : : * with grouping sets we need to remove the RT index of the
5825 : : * grouping step if there is any from the target expression.
5826 : : */
5827 [ + - + + ]: 6756 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5828 : : {
5829 : : Assert(root->group_rtindex > 0);
5830 : : expr = (Expr *)
5831 : 1803 : remove_nulling_relids((Node *) expr,
5832 : 1803 : bms_make_singleton(root->group_rtindex),
5833 : : NULL);
5834 : : }
5835 : 6756 : add_column_to_pathtarget(input_target, expr, sgref);
5836 : : }
5837 : : else
5838 : : {
5839 : : /*
5840 : : * Non-grouping column, so just remember the expression for later
5841 : : * call to pull_var_clause.
5842 : : */
5843 : 45164 : non_group_cols = lappend(non_group_cols, expr);
5844 : : }
5845 : :
5846 : 51920 : i++;
5847 : : }
5848 : :
5849 : : /*
5850 : : * If there's a HAVING clause, we'll need the Vars it uses, too.
5851 : : */
5852 [ + + ]: 34089 : if (parse->havingQual)
5853 : 880 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5854 : :
5855 : : /*
5856 : : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5857 : : * add them to the input target if not already present. (A Var used
5858 : : * directly as a GROUP BY item will be present already.) Note this
5859 : : * includes Vars used in resjunk items, so we are covering the needs of
5860 : : * ORDER BY and window specifications. Vars used within Aggrefs and
5861 : : * WindowFuncs will be pulled out here, too.
5862 : : *
5863 : : * Note that the target is logically below the grouping step. So with
5864 : : * grouping sets we need to remove the RT index of the grouping step if
5865 : : * there is any from the non-group Vars.
5866 : : */
5867 : 34089 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5868 : : PVC_RECURSE_AGGREGATES |
5869 : : PVC_RECURSE_WINDOWFUNCS |
5870 : : PVC_INCLUDE_PLACEHOLDERS);
5871 [ + + + + ]: 34089 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5872 : : {
5873 : : Assert(root->group_rtindex > 0);
5874 : : non_group_vars = (List *)
5875 : 836 : remove_nulling_relids((Node *) non_group_vars,
5876 : 836 : bms_make_singleton(root->group_rtindex),
5877 : : NULL);
5878 : : }
5879 : 34089 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5880 : :
5881 : : /* clean up cruft */
5882 : 34089 : list_free(non_group_vars);
5883 : 34089 : list_free(non_group_cols);
5884 : :
5885 : : /* XXX this causes some redundant cost calculation ... */
5886 : 34089 : return set_pathtarget_cost_width(root, input_target);
5887 : : }
5888 : :
5889 : : /*
5890 : : * make_partial_grouping_target
5891 : : * Generate appropriate PathTarget for output of partial aggregate
5892 : : * (or partial grouping, if there are no aggregates) nodes.
5893 : : *
5894 : : * A partial aggregation node needs to emit all the same aggregates that
5895 : : * a regular aggregation node would, plus any aggregates used in HAVING;
5896 : : * except that the Aggref nodes should be marked as partial aggregates.
5897 : : *
5898 : : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5899 : : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5900 : : * these would be Vars that are grouped by or used in grouping expressions.)
5901 : : *
5902 : : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5903 : : * havingQual represents the HAVING clause.
5904 : : */
5905 : : static PathTarget *
5906 : 3216 : make_partial_grouping_target(PlannerInfo *root,
5907 : : PathTarget *grouping_target,
5908 : : Node *havingQual)
5909 : : {
5910 : : PathTarget *partial_target;
5911 : : List *non_group_cols;
5912 : : List *non_group_exprs;
5913 : : int i;
5914 : : ListCell *lc;
5915 : :
5916 : 3216 : partial_target = create_empty_pathtarget();
5917 : 3216 : non_group_cols = NIL;
5918 : :
5919 : 3216 : i = 0;
5920 [ + - + + : 10593 : foreach(lc, grouping_target->exprs)
+ + ]
5921 : : {
5922 : 7377 : Expr *expr = (Expr *) lfirst(lc);
5923 [ + - ]: 7377 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5924 : :
5925 [ + + + + : 11319 : if (sgref && root->processed_groupClause &&
+ + ]
5926 : 3942 : get_sortgroupref_clause_noerr(sgref,
5927 : : root->processed_groupClause) != NULL)
5928 : : {
5929 : : /*
5930 : : * It's a grouping column, so add it to the partial_target as-is.
5931 : : * (This allows the upper agg step to repeat the grouping calcs.)
5932 : : */
5933 : 2340 : add_column_to_pathtarget(partial_target, expr, sgref);
5934 : : }
5935 : : else
5936 : : {
5937 : : /*
5938 : : * Non-grouping column, so just remember the expression for later
5939 : : * call to pull_var_clause.
5940 : : */
5941 : 5037 : non_group_cols = lappend(non_group_cols, expr);
5942 : : }
5943 : :
5944 : 7377 : i++;
5945 : : }
5946 : :
5947 : : /*
5948 : : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5949 : : */
5950 [ + + ]: 3216 : if (havingQual)
5951 : 729 : non_group_cols = lappend(non_group_cols, havingQual);
5952 : :
5953 : : /*
5954 : : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5955 : : * non-group cols (plus HAVING), and add them to the partial_target if not
5956 : : * already present. (An expression used directly as a GROUP BY item will
5957 : : * be present already.) Note this includes Vars used in resjunk items, so
5958 : : * we are covering the needs of ORDER BY and window specifications.
5959 : : */
5960 : 3216 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5961 : : PVC_INCLUDE_AGGREGATES |
5962 : : PVC_RECURSE_WINDOWFUNCS |
5963 : : PVC_INCLUDE_PLACEHOLDERS);
5964 : :
5965 : 3216 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5966 : :
5967 : : /*
5968 : : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5969 : : * are at the top level of the target list, so we can just scan the list
5970 : : * rather than recursing through the expression trees.
5971 : : */
5972 [ + - + + : 11107 : foreach(lc, partial_target->exprs)
+ + ]
5973 : : {
5974 : 7891 : Aggref *aggref = (Aggref *) lfirst(lc);
5975 : :
5976 [ + + ]: 7891 : if (IsA(aggref, Aggref))
5977 : : {
5978 : : Aggref *newaggref;
5979 : :
5980 : : /*
5981 : : * We shouldn't need to copy the substructure of the Aggref node,
5982 : : * but flat-copy the node itself to avoid damaging other trees.
5983 : : */
5984 : 5526 : newaggref = makeNode(Aggref);
5985 : 5526 : memcpy(newaggref, aggref, sizeof(Aggref));
5986 : :
5987 : : /* For now, assume serialization is required */
5988 : 5526 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5989 : :
5990 : 5526 : lfirst(lc) = newaggref;
5991 : : }
5992 : : }
5993 : :
5994 : : /* clean up cruft */
5995 : 3216 : list_free(non_group_exprs);
5996 : 3216 : list_free(non_group_cols);
5997 : :
5998 : : /* XXX this causes some redundant cost calculation ... */
5999 : 3216 : return set_pathtarget_cost_width(root, partial_target);
6000 : : }
6001 : :
6002 : : /*
6003 : : * mark_partial_aggref
6004 : : * Adjust an Aggref to make it represent a partial-aggregation step.
6005 : : *
6006 : : * The Aggref node is modified in-place; caller must do any copying required.
6007 : : */
6008 : : void
6009 : 15395 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
6010 : : {
6011 : : /* aggtranstype should be computed by this point */
6012 : : Assert(OidIsValid(agg->aggtranstype));
6013 : : /* ... but aggsplit should still be as the parser left it */
6014 : : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
6015 : :
6016 : : /* Mark the Aggref with the intended partial-aggregation mode */
6017 : 15395 : agg->aggsplit = aggsplit;
6018 : :
6019 : : /*
6020 : : * Adjust result type if needed. Normally, a partial aggregate returns
6021 : : * the aggregate's transition type; but if that's INTERNAL and we're
6022 : : * serializing, it returns BYTEA instead.
6023 : : */
6024 [ + + ]: 15395 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
6025 : : {
6026 [ + + + - ]: 13480 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
6027 : 261 : agg->aggtype = BYTEAOID;
6028 : : else
6029 : 13219 : agg->aggtype = agg->aggtranstype;
6030 : : }
6031 : 15395 : }
6032 : :
6033 : : /*
6034 : : * postprocess_setop_tlist
6035 : : * Fix up targetlist returned by plan_set_operations().
6036 : : *
6037 : : * We need to transpose sort key info from the orig_tlist into new_tlist.
6038 : : * NOTE: this would not be good enough if we supported resjunk sort keys
6039 : : * for results of set operations --- then, we'd need to project a whole
6040 : : * new tlist to evaluate the resjunk columns. For now, just ereport if we
6041 : : * find any resjunk columns in orig_tlist.
6042 : : */
6043 : : static List *
6044 : 4928 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
6045 : : {
6046 : : ListCell *l;
6047 : 4928 : ListCell *orig_tlist_item = list_head(orig_tlist);
6048 : :
6049 [ + + + + : 18835 : foreach(l, new_tlist)
+ + ]
6050 : : {
6051 : 13907 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
6052 : : TargetEntry *orig_tle;
6053 : :
6054 : : /* ignore resjunk columns in setop result */
6055 [ - + ]: 13907 : if (new_tle->resjunk)
6056 : 0 : continue;
6057 : :
6058 : : Assert(orig_tlist_item != NULL);
6059 : 13907 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
6060 : 13907 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
6061 [ - + ]: 13907 : if (orig_tle->resjunk) /* should not happen */
6062 [ # # ]: 0 : elog(ERROR, "resjunk output columns are not implemented");
6063 : : Assert(new_tle->resno == orig_tle->resno);
6064 : 13907 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
6065 : : }
6066 [ - + ]: 4928 : if (orig_tlist_item != NULL)
6067 [ # # ]: 0 : elog(ERROR, "resjunk output columns are not implemented");
6068 : 4928 : return new_tlist;
6069 : : }
6070 : :
6071 : : /*
6072 : : * optimize_window_clauses
6073 : : * Call each WindowFunc's prosupport function to see if we're able to
6074 : : * make any adjustments to any of the WindowClause's so that the executor
6075 : : * can execute the window functions in a more optimal way.
6076 : : *
6077 : : * Currently we only allow adjustments to the WindowClause's frameOptions. We
6078 : : * may allow more things to be done here in the future.
6079 : : */
6080 : : static void
6081 : 2331 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
6082 : : {
6083 : 2331 : List *windowClause = root->parse->windowClause;
6084 : : ListCell *lc;
6085 : :
6086 [ + - + + : 4877 : foreach(lc, windowClause)
+ + ]
6087 : : {
6088 : 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
6089 : : ListCell *lc2;
6090 : 2546 : int optimizedFrameOptions = 0;
6091 : :
6092 : : Assert(wc->winref <= wflists->maxWinRef);
6093 : :
6094 : : /* skip any WindowClauses that have no WindowFuncs */
6095 [ + + ]: 2546 : if (wflists->windowFuncs[wc->winref] == NIL)
6096 : 20 : continue;
6097 : :
6098 [ + - + + : 3088 : foreach(lc2, wflists->windowFuncs[wc->winref])
+ + ]
6099 : : {
6100 : : SupportRequestOptimizeWindowClause req;
6101 : : SupportRequestOptimizeWindowClause *res;
6102 : 2561 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
6103 : : Oid prosupport;
6104 : :
6105 : 2561 : prosupport = get_func_support(wfunc->winfnoid);
6106 : :
6107 : : /* Check if there's a support function for 'wfunc' */
6108 [ + + ]: 2561 : if (!OidIsValid(prosupport))
6109 : 1999 : break; /* can't optimize this WindowClause */
6110 : :
6111 : 834 : req.type = T_SupportRequestOptimizeWindowClause;
6112 : 834 : req.window_clause = wc;
6113 : 834 : req.window_func = wfunc;
6114 : 834 : req.frameOptions = wc->frameOptions;
6115 : :
6116 : : /* call the support function */
6117 : : res = (SupportRequestOptimizeWindowClause *)
6118 : 834 : DatumGetPointer(OidFunctionCall1(prosupport,
6119 : : PointerGetDatum(&req)));
6120 : :
6121 : : /*
6122 : : * Skip to next WindowClause if the support function does not
6123 : : * support this request type.
6124 : : */
6125 [ + + ]: 834 : if (res == NULL)
6126 : 272 : break;
6127 : :
6128 : : /*
6129 : : * Save these frameOptions for the first WindowFunc for this
6130 : : * WindowClause.
6131 : : */
6132 [ + + ]: 562 : if (foreach_current_index(lc2) == 0)
6133 : 542 : optimizedFrameOptions = res->frameOptions;
6134 : :
6135 : : /*
6136 : : * On subsequent WindowFuncs, if the frameOptions are not the same
6137 : : * then we're unable to optimize the frameOptions for this
6138 : : * WindowClause.
6139 : : */
6140 [ - + ]: 20 : else if (optimizedFrameOptions != res->frameOptions)
6141 : 0 : break; /* skip to the next WindowClause, if any */
6142 : : }
6143 : :
6144 : : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
6145 [ + + + - ]: 2526 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
6146 : : {
6147 : : ListCell *lc3;
6148 : :
6149 : : /* apply the new frame options */
6150 : 527 : wc->frameOptions = optimizedFrameOptions;
6151 : :
6152 : : /*
6153 : : * We now check to see if changing the frameOptions has caused
6154 : : * this WindowClause to be a duplicate of some other WindowClause.
6155 : : * This can only happen if we have multiple WindowClauses, so
6156 : : * don't bother if there's only 1.
6157 : : */
6158 [ + + ]: 527 : if (list_length(windowClause) == 1)
6159 : 442 : continue;
6160 : :
6161 : : /*
6162 : : * Do the duplicate check and reuse the existing WindowClause if
6163 : : * we find a duplicate.
6164 : : */
6165 [ + - + + : 220 : foreach(lc3, windowClause)
+ + ]
6166 : : {
6167 : 165 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
6168 : :
6169 : : /* skip over the WindowClause we're currently editing */
6170 [ + + ]: 165 : if (existing_wc == wc)
6171 : 55 : continue;
6172 : :
6173 : : /*
6174 : : * Perform the same duplicate check that is done in
6175 : : * transformWindowFuncCall.
6176 : : */
6177 [ + - + + ]: 220 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
6178 : 110 : equal(wc->orderClause, existing_wc->orderClause) &&
6179 [ + + + - ]: 100 : wc->frameOptions == existing_wc->frameOptions &&
6180 [ + - ]: 60 : equal(wc->startOffset, existing_wc->startOffset) &&
6181 : 30 : equal(wc->endOffset, existing_wc->endOffset))
6182 : : {
6183 : : ListCell *lc4;
6184 : :
6185 : : /*
6186 : : * Now move each WindowFunc in 'wc' into 'existing_wc'.
6187 : : * This required adjusting each WindowFunc's winref and
6188 : : * moving the WindowFuncs in 'wc' to the list of
6189 : : * WindowFuncs in 'existing_wc'.
6190 : : */
6191 [ + - + + : 65 : foreach(lc4, wflists->windowFuncs[wc->winref])
+ + ]
6192 : : {
6193 : 35 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
6194 : :
6195 : 35 : wfunc->winref = existing_wc->winref;
6196 : : }
6197 : :
6198 : : /* move list items */
6199 : 60 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
6200 : 30 : wflists->windowFuncs[wc->winref]);
6201 : 30 : wflists->windowFuncs[wc->winref] = NIL;
6202 : :
6203 : : /*
6204 : : * transformWindowFuncCall() should have made sure there
6205 : : * are no other duplicates, so we needn't bother looking
6206 : : * any further.
6207 : : */
6208 : 30 : break;
6209 : : }
6210 : : }
6211 : : }
6212 : : }
6213 : 2331 : }
6214 : :
6215 : : /*
6216 : : * select_active_windows
6217 : : * Create a list of the "active" window clauses (ie, those referenced
6218 : : * by non-deleted WindowFuncs) in the order they are to be executed.
6219 : : */
6220 : : static List *
6221 : 2331 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
6222 : : {
6223 : 2331 : List *windowClause = root->parse->windowClause;
6224 : 2331 : List *result = NIL;
6225 : : ListCell *lc;
6226 : 2331 : int nActive = 0;
6227 : 2331 : WindowClauseSortData *actives = palloc_array(WindowClauseSortData,
6228 : : list_length(windowClause));
6229 : :
6230 : : /* First, construct an array of the active windows */
6231 [ + - + + : 4877 : foreach(lc, windowClause)
+ + ]
6232 : : {
6233 : 2546 : WindowClause *wc = lfirst_node(WindowClause, lc);
6234 : :
6235 : : /* It's only active if wflists shows some related WindowFuncs */
6236 : : Assert(wc->winref <= wflists->maxWinRef);
6237 [ + + ]: 2546 : if (wflists->windowFuncs[wc->winref] == NIL)
6238 : 50 : continue;
6239 : :
6240 : 2496 : actives[nActive].wc = wc; /* original clause */
6241 : :
6242 : : /*
6243 : : * For sorting, we want the list of partition keys followed by the
6244 : : * list of sort keys. But pathkeys construction will remove duplicates
6245 : : * between the two, so we can as well (even though we can't detect all
6246 : : * of the duplicates, since some may come from ECs - that might mean
6247 : : * we miss optimization chances here). We must, however, ensure that
6248 : : * the order of entries is preserved with respect to the ones we do
6249 : : * keep.
6250 : : *
6251 : : * partitionClause and orderClause had their own duplicates removed in
6252 : : * parse analysis, so we're only concerned here with removing
6253 : : * orderClause entries that also appear in partitionClause.
6254 : : */
6255 : 4992 : actives[nActive].uniqueOrder =
6256 : 2496 : list_concat_unique(list_copy(wc->partitionClause),
6257 : 2496 : wc->orderClause);
6258 : 2496 : nActive++;
6259 : : }
6260 : :
6261 : : /*
6262 : : * Sort active windows by their partitioning/ordering clauses, ignoring
6263 : : * any framing clauses, so that the windows that need the same sorting are
6264 : : * adjacent in the list. When we come to generate paths, this will avoid
6265 : : * inserting additional Sort nodes.
6266 : : *
6267 : : * This is how we implement a specific requirement from the SQL standard,
6268 : : * which says that when two or more windows are order-equivalent (i.e.
6269 : : * have matching partition and order clauses, even if their names or
6270 : : * framing clauses differ), then all peer rows must be presented in the
6271 : : * same order in all of them. If we allowed multiple sort nodes for such
6272 : : * cases, we'd risk having the peer rows end up in different orders in
6273 : : * equivalent windows due to sort instability. (See General Rule 4 of
6274 : : * <window clause> in SQL2008 - SQL2016.)
6275 : : *
6276 : : * Additionally, if the entire list of clauses of one window is a prefix
6277 : : * of another, put first the window with stronger sorting requirements.
6278 : : * This way we will first sort for stronger window, and won't have to sort
6279 : : * again for the weaker one.
6280 : : */
6281 : 2331 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6282 : :
6283 : : /* build ordered list of the original WindowClause nodes */
6284 [ + + ]: 4827 : for (int i = 0; i < nActive; i++)
6285 : 2496 : result = lappend(result, actives[i].wc);
6286 : :
6287 : 2331 : pfree(actives);
6288 : :
6289 : 2331 : return result;
6290 : : }
6291 : :
6292 : : /*
6293 : : * name_active_windows
6294 : : * Ensure all active windows have unique names.
6295 : : *
6296 : : * The parser will have checked that user-assigned window names are unique
6297 : : * within the Query. Here we assign made-up names to any unnamed
6298 : : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6299 : : * at parse time, because it'd mess up decompilation of views.)
6300 : : *
6301 : : * activeWindows: result of select_active_windows
6302 : : */
6303 : : static void
6304 : 2331 : name_active_windows(List *activeWindows)
6305 : : {
6306 : 2331 : int next_n = 1;
6307 : : char newname[16];
6308 : : ListCell *lc;
6309 : :
6310 [ + - + + : 4827 : foreach(lc, activeWindows)
+ + ]
6311 : : {
6312 : 2496 : WindowClause *wc = lfirst_node(WindowClause, lc);
6313 : :
6314 : : /* Nothing to do if it has a name already. */
6315 [ + + ]: 2496 : if (wc->name)
6316 : 480 : continue;
6317 : :
6318 : : /* Select a name not currently present in the list. */
6319 : : for (;;)
6320 : 5 : {
6321 : : ListCell *lc2;
6322 : :
6323 : 2021 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6324 [ + - + + : 4352 : foreach(lc2, activeWindows)
+ + ]
6325 : : {
6326 : 2336 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6327 : :
6328 [ + + + + ]: 2336 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6329 : 5 : break; /* matched */
6330 : : }
6331 [ + + ]: 2021 : if (lc2 == NULL)
6332 : 2016 : break; /* reached the end with no match */
6333 : : }
6334 : 2016 : wc->name = pstrdup(newname);
6335 : : }
6336 : 2331 : }
6337 : :
6338 : : /*
6339 : : * common_prefix_cmp
6340 : : * QSort comparison function for WindowClauseSortData
6341 : : *
6342 : : * Sort the windows by the required sorting clauses. First, compare the sort
6343 : : * clauses themselves. Second, if one window's clauses are a prefix of another
6344 : : * one's clauses, put the window with more sort clauses first.
6345 : : *
6346 : : * We purposefully sort by the highest tleSortGroupRef first. Since
6347 : : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6348 : : * and because here we sort the lowest tleSortGroupRefs last, if a
6349 : : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6350 : : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6351 : : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6352 : : * reducing the total number of sorts required for the query.
6353 : : */
6354 : : static int
6355 : 180 : common_prefix_cmp(const void *a, const void *b)
6356 : : {
6357 : 180 : const WindowClauseSortData *wcsa = a;
6358 : 180 : const WindowClauseSortData *wcsb = b;
6359 : : ListCell *item_a;
6360 : : ListCell *item_b;
6361 : :
6362 [ + + + + : 315 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
+ + + + +
+ + + +
+ ]
6363 : : {
6364 : 230 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6365 : 230 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6366 : :
6367 [ + + ]: 230 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6368 : 95 : return -1;
6369 [ + + ]: 220 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6370 : 55 : return 1;
6371 [ + + ]: 165 : else if (sca->sortop > scb->sortop)
6372 : 10 : return -1;
6373 [ + + ]: 155 : else if (sca->sortop < scb->sortop)
6374 : 20 : return 1;
6375 [ - + - - ]: 135 : else if (sca->nulls_first && !scb->nulls_first)
6376 : 0 : return -1;
6377 [ + - - + ]: 135 : else if (!sca->nulls_first && scb->nulls_first)
6378 : 0 : return 1;
6379 : : /* no need to compare eqop, since it is fully determined by sortop */
6380 : : }
6381 : :
6382 [ + + ]: 85 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6383 : 5 : return -1;
6384 [ + + ]: 80 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6385 : 25 : return 1;
6386 : :
6387 : 55 : return 0;
6388 : : }
6389 : :
6390 : : /*
6391 : : * make_window_input_target
6392 : : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6393 : : *
6394 : : * When the query has window functions, this function computes the desired
6395 : : * target to be computed by the node just below the first WindowAgg.
6396 : : * This tlist must contain all values needed to evaluate the window functions,
6397 : : * compute the final target list, and perform any required final sort step.
6398 : : * If multiple WindowAggs are needed, each intermediate one adds its window
6399 : : * function results onto this base tlist; only the topmost WindowAgg computes
6400 : : * the actual desired target list.
6401 : : *
6402 : : * This function is much like make_group_input_target, though not quite enough
6403 : : * like it to share code. As in that function, we flatten most expressions
6404 : : * into their component variables. But we do not want to flatten window
6405 : : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6406 : : * evaluations of them, which would be bad (possibly even resulting in
6407 : : * inconsistent answers, if they contain volatile functions).
6408 : : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6409 : : * make_group_input_target, because we may no longer have access to the
6410 : : * individual Vars in them.
6411 : : *
6412 : : * Another key difference from make_group_input_target is that we don't
6413 : : * flatten Aggref expressions, since those are to be computed below the
6414 : : * window functions and just referenced like Vars above that.
6415 : : *
6416 : : * 'final_target' is the query's final target list (in PathTarget form)
6417 : : * 'activeWindows' is the list of active windows previously identified by
6418 : : * select_active_windows.
6419 : : *
6420 : : * The result is the PathTarget to be computed by the plan node immediately
6421 : : * below the first WindowAgg node.
6422 : : */
6423 : : static PathTarget *
6424 : 2331 : make_window_input_target(PlannerInfo *root,
6425 : : PathTarget *final_target,
6426 : : List *activeWindows)
6427 : : {
6428 : : PathTarget *input_target;
6429 : : Bitmapset *sgrefs;
6430 : : List *flattenable_cols;
6431 : : List *flattenable_vars;
6432 : : int i;
6433 : : ListCell *lc;
6434 : :
6435 : : Assert(root->parse->hasWindowFuncs);
6436 : :
6437 : : /*
6438 : : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6439 : : * into a bitmapset for convenient reference below.
6440 : : */
6441 : 2331 : sgrefs = NULL;
6442 [ + - + + : 4827 : foreach(lc, activeWindows)
+ + ]
6443 : : {
6444 : 2496 : WindowClause *wc = lfirst_node(WindowClause, lc);
6445 : : ListCell *lc2;
6446 : :
6447 [ + + + + : 3138 : foreach(lc2, wc->partitionClause)
+ + ]
6448 : : {
6449 : 642 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6450 : :
6451 : 642 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6452 : : }
6453 [ + + + + : 4485 : foreach(lc2, wc->orderClause)
+ + ]
6454 : : {
6455 : 1989 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6456 : :
6457 : 1989 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6458 : : }
6459 : : }
6460 : :
6461 : : /* Add in sortgroupref numbers of GROUP BY clauses, too */
6462 [ + + + + : 2492 : foreach(lc, root->processed_groupClause)
+ + ]
6463 : : {
6464 : 161 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6465 : :
6466 : 161 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6467 : : }
6468 : :
6469 : : /*
6470 : : * Construct a target containing all the non-flattenable targetlist items,
6471 : : * and save aside the others for a moment.
6472 : : */
6473 : 2331 : input_target = create_empty_pathtarget();
6474 : 2331 : flattenable_cols = NIL;
6475 : :
6476 : 2331 : i = 0;
6477 [ + - + + : 9802 : foreach(lc, final_target->exprs)
+ + ]
6478 : : {
6479 : 7471 : Expr *expr = (Expr *) lfirst(lc);
6480 [ + - ]: 7471 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6481 : :
6482 : : /*
6483 : : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6484 : : * that such items can't contain window functions, so it's okay to
6485 : : * compute them below the WindowAgg nodes.)
6486 : : */
6487 [ + + + + ]: 7471 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6488 : : {
6489 : : /*
6490 : : * Don't want to deconstruct this value, so add it to the input
6491 : : * target as-is.
6492 : : */
6493 : 2488 : add_column_to_pathtarget(input_target, expr, sgref);
6494 : : }
6495 : : else
6496 : : {
6497 : : /*
6498 : : * Column is to be flattened, so just remember the expression for
6499 : : * later call to pull_var_clause.
6500 : : */
6501 : 4983 : flattenable_cols = lappend(flattenable_cols, expr);
6502 : : }
6503 : :
6504 : 7471 : i++;
6505 : : }
6506 : :
6507 : : /*
6508 : : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6509 : : * add them to the input target if not already present. (Some might be
6510 : : * there already because they're used directly as window/group clauses.)
6511 : : *
6512 : : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6513 : : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6514 : : * at higher levels. On the other hand, we should recurse into
6515 : : * WindowFuncs to make sure their input expressions are available.
6516 : : */
6517 : 2331 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6518 : : PVC_INCLUDE_AGGREGATES |
6519 : : PVC_RECURSE_WINDOWFUNCS |
6520 : : PVC_INCLUDE_PLACEHOLDERS);
6521 : 2331 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6522 : :
6523 : : /* clean up cruft */
6524 : 2331 : list_free(flattenable_vars);
6525 : 2331 : list_free(flattenable_cols);
6526 : :
6527 : : /* XXX this causes some redundant cost calculation ... */
6528 : 2331 : return set_pathtarget_cost_width(root, input_target);
6529 : : }
6530 : :
6531 : : /*
6532 : : * make_pathkeys_for_window
6533 : : * Create a pathkeys list describing the required input ordering
6534 : : * for the given WindowClause.
6535 : : *
6536 : : * Modifies wc's partitionClause to remove any clauses which are deemed
6537 : : * redundant by the pathkey logic.
6538 : : *
6539 : : * The required ordering is first the PARTITION keys, then the ORDER keys.
6540 : : * In the future we might try to implement windowing using hashing, in which
6541 : : * case the ordering could be relaxed, but for now we always sort.
6542 : : */
6543 : : static List *
6544 : 4988 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6545 : : List *tlist)
6546 : : {
6547 : 4988 : List *window_pathkeys = NIL;
6548 : :
6549 : : /* Throw error if can't sort */
6550 [ - + ]: 4988 : if (!grouping_is_sortable(wc->partitionClause))
6551 [ # # ]: 0 : ereport(ERROR,
6552 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6553 : : errmsg("could not implement window PARTITION BY"),
6554 : : errdetail("Window partitioning columns must be of sortable datatypes.")));
6555 [ - + ]: 4988 : if (!grouping_is_sortable(wc->orderClause))
6556 [ # # ]: 0 : ereport(ERROR,
6557 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6558 : : errmsg("could not implement window ORDER BY"),
6559 : : errdetail("Window ordering columns must be of sortable datatypes.")));
6560 : :
6561 : : /*
6562 : : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6563 : : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6564 : : */
6565 [ + + ]: 4988 : if (wc->partitionClause != NIL)
6566 : : {
6567 : : bool sortable;
6568 : :
6569 : 1130 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6570 : : &wc->partitionClause,
6571 : : tlist,
6572 : : true,
6573 : : false,
6574 : : &sortable,
6575 : : false);
6576 : :
6577 : : Assert(sortable);
6578 : : }
6579 : :
6580 : : /*
6581 : : * In principle, we could also consider removing redundant ORDER BY items
6582 : : * too as doing so does not alter the result of peer row checks done by
6583 : : * the executor. However, we must *not* remove the ordering column for
6584 : : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6585 : : * if it's known to be equal to some partitioning column.
6586 : : */
6587 [ + + ]: 4988 : if (wc->orderClause != NIL)
6588 : : {
6589 : : List *orderby_pathkeys;
6590 : :
6591 : 3876 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6592 : : wc->orderClause,
6593 : : tlist);
6594 : :
6595 : : /* Okay, make the combined pathkeys */
6596 [ + + ]: 3876 : if (window_pathkeys != NIL)
6597 : 763 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6598 : : else
6599 : 3113 : window_pathkeys = orderby_pathkeys;
6600 : : }
6601 : :
6602 : 4988 : return window_pathkeys;
6603 : : }
6604 : :
6605 : : /*
6606 : : * make_sort_input_target
6607 : : * Generate appropriate PathTarget for initial input to Sort step.
6608 : : *
6609 : : * If the query has ORDER BY, this function chooses the target to be computed
6610 : : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6611 : : * project) steps. This might or might not be identical to the query's final
6612 : : * output target.
6613 : : *
6614 : : * The main argument for keeping the sort-input tlist the same as the final
6615 : : * is that we avoid a separate projection node (which will be needed if
6616 : : * they're different, because Sort can't project). However, there are also
6617 : : * advantages to postponing tlist evaluation till after the Sort: it ensures
6618 : : * a consistent order of evaluation for any volatile functions in the tlist,
6619 : : * and if there's also a LIMIT, we can stop the query without ever computing
6620 : : * tlist functions for later rows, which is beneficial for both volatile and
6621 : : * expensive functions.
6622 : : *
6623 : : * Our current policy is to postpone volatile expressions till after the sort
6624 : : * unconditionally (assuming that that's possible, ie they are in plain tlist
6625 : : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6626 : : * postpone set-returning expressions, because running them beforehand would
6627 : : * bloat the sort dataset, and because it might cause unexpected output order
6628 : : * if the sort isn't stable. However there's a constraint on that: all SRFs
6629 : : * in the tlist should be evaluated at the same plan step, so that they can
6630 : : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6631 : : * mustn't postpone any SRFs. (Note that in principle that policy should
6632 : : * probably get applied to the group/window input targetlists too, but we
6633 : : * have not done that historically.) Lastly, expensive expressions are
6634 : : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6635 : : * partial evaluation of the query is possible (if neither is true, we expect
6636 : : * to have to evaluate the expressions for every row anyway), or if there are
6637 : : * any volatile or set-returning expressions (since once we've put in a
6638 : : * projection at all, it won't cost any more to postpone more stuff).
6639 : : *
6640 : : * Another issue that could potentially be considered here is that
6641 : : * evaluating tlist expressions could result in data that's either wider
6642 : : * or narrower than the input Vars, thus changing the volume of data that
6643 : : * has to go through the Sort. However, we usually have only a very bad
6644 : : * idea of the output width of any expression more complex than a Var,
6645 : : * so for now it seems too risky to try to optimize on that basis.
6646 : : *
6647 : : * Note that if we do produce a modified sort-input target, and then the
6648 : : * query ends up not using an explicit Sort, no particular harm is done:
6649 : : * we'll initially use the modified target for the preceding path nodes,
6650 : : * but then change them to the final target with apply_projection_to_path.
6651 : : * Moreover, in such a case the guarantees about evaluation order of
6652 : : * volatile functions still hold, since the rows are sorted already.
6653 : : *
6654 : : * This function has some things in common with make_group_input_target and
6655 : : * make_window_input_target, though the detailed rules for what to do are
6656 : : * different. We never flatten/postpone any grouping or ordering columns;
6657 : : * those are needed before the sort. If we do flatten a particular
6658 : : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6659 : : * computed earlier.
6660 : : *
6661 : : * 'final_target' is the query's final target list (in PathTarget form)
6662 : : * 'have_postponed_srfs' is an output argument, see below
6663 : : *
6664 : : * The result is the PathTarget to be computed by the plan node immediately
6665 : : * below the Sort step (and the Distinct step, if any). This will be
6666 : : * exactly final_target if we decide a projection step wouldn't be helpful.
6667 : : *
6668 : : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6669 : : * any set-returning functions to after the Sort.
6670 : : */
6671 : : static PathTarget *
6672 : 57783 : make_sort_input_target(PlannerInfo *root,
6673 : : PathTarget *final_target,
6674 : : bool *have_postponed_srfs)
6675 : : {
6676 : 57783 : Query *parse = root->parse;
6677 : : PathTarget *input_target;
6678 : : int ncols;
6679 : : bool *col_is_srf;
6680 : : bool *postpone_col;
6681 : : bool have_srf;
6682 : : bool have_volatile;
6683 : : bool have_expensive;
6684 : : bool have_srf_sortcols;
6685 : : bool postpone_srfs;
6686 : : List *postponable_cols;
6687 : : List *postponable_vars;
6688 : : int i;
6689 : : ListCell *lc;
6690 : :
6691 : : /* Shouldn't get here unless query has ORDER BY */
6692 : : Assert(parse->sortClause);
6693 : :
6694 : 57783 : *have_postponed_srfs = false; /* default result */
6695 : :
6696 : : /* Inspect tlist and collect per-column information */
6697 : 57783 : ncols = list_length(final_target->exprs);
6698 : 57783 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
6699 : 57783 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
6700 : 57783 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6701 : :
6702 : 57783 : i = 0;
6703 [ + - + + : 329744 : foreach(lc, final_target->exprs)
+ + ]
6704 : : {
6705 : 271961 : Expr *expr = (Expr *) lfirst(lc);
6706 : :
6707 : : /*
6708 : : * If the column has a sortgroupref, assume it has to be evaluated
6709 : : * before sorting. Generally such columns would be ORDER BY, GROUP
6710 : : * BY, etc targets. One exception is columns that were removed from
6711 : : * GROUP BY by remove_useless_groupby_columns() ... but those would
6712 : : * only be Vars anyway. There don't seem to be any cases where it
6713 : : * would be worth the trouble to double-check.
6714 : : */
6715 [ + - + + ]: 271961 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6716 : : {
6717 : : /*
6718 : : * Check for SRF or volatile functions. Check the SRF case first
6719 : : * because we must know whether we have any postponed SRFs.
6720 : : */
6721 [ + + + + ]: 190586 : if (parse->hasTargetSRFs &&
6722 : 180 : expression_returns_set((Node *) expr))
6723 : : {
6724 : : /* We'll decide below whether these are postponable */
6725 : 80 : col_is_srf[i] = true;
6726 : 80 : have_srf = true;
6727 : : }
6728 [ + + ]: 190326 : else if (contain_volatile_functions((Node *) expr))
6729 : : {
6730 : : /* Unconditionally postpone */
6731 : 195 : postpone_col[i] = true;
6732 : 195 : have_volatile = true;
6733 : : }
6734 : : else
6735 : : {
6736 : : /*
6737 : : * Else check the cost. XXX it's annoying to have to do this
6738 : : * when set_pathtarget_cost_width() just did it. Refactor to
6739 : : * allow sharing the work?
6740 : : */
6741 : : QualCost cost;
6742 : :
6743 : 190131 : cost_qual_eval_node(&cost, (Node *) expr, root);
6744 : :
6745 : : /*
6746 : : * We arbitrarily define "expensive" as "more than 10X
6747 : : * cpu_operator_cost". Note this will take in any PL function
6748 : : * with default cost.
6749 : : */
6750 [ + + ]: 190131 : if (cost.per_tuple > 10 * cpu_operator_cost)
6751 : : {
6752 : 13930 : postpone_col[i] = true;
6753 : 13930 : have_expensive = true;
6754 : : }
6755 : : }
6756 : : }
6757 : : else
6758 : : {
6759 : : /* For sortgroupref cols, just check if any contain SRFs */
6760 [ + + ]: 81555 : if (!have_srf_sortcols &&
6761 [ + + + + ]: 81812 : parse->hasTargetSRFs &&
6762 : 277 : expression_returns_set((Node *) expr))
6763 : 122 : have_srf_sortcols = true;
6764 : : }
6765 : :
6766 : 271961 : i++;
6767 : : }
6768 : :
6769 : : /*
6770 : : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6771 : : */
6772 [ + + + + ]: 57783 : postpone_srfs = (have_srf && !have_srf_sortcols);
6773 : :
6774 : : /*
6775 : : * If we don't need a post-sort projection, just return final_target.
6776 : : */
6777 [ + + + + ]: 57783 : if (!(postpone_srfs || have_volatile ||
6778 [ + + ]: 57540 : (have_expensive &&
6779 [ + + + - ]: 8386 : (parse->limitCount || root->tuple_fraction > 0))))
6780 : 57510 : return final_target;
6781 : :
6782 : : /*
6783 : : * Report whether the post-sort projection will contain set-returning
6784 : : * functions. This is important because it affects whether the Sort can
6785 : : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6786 : : * to return.
6787 : : */
6788 : 273 : *have_postponed_srfs = postpone_srfs;
6789 : :
6790 : : /*
6791 : : * Construct the sort-input target, taking all non-postponable columns and
6792 : : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6793 : : * the postponable ones.
6794 : : */
6795 : 273 : input_target = create_empty_pathtarget();
6796 : 273 : postponable_cols = NIL;
6797 : :
6798 : 273 : i = 0;
6799 [ + - + + : 1895 : foreach(lc, final_target->exprs)
+ + ]
6800 : : {
6801 : 1622 : Expr *expr = (Expr *) lfirst(lc);
6802 : :
6803 [ + + + + : 1622 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
+ + ]
6804 : 325 : postponable_cols = lappend(postponable_cols, expr);
6805 : : else
6806 : 1297 : add_column_to_pathtarget(input_target, expr,
6807 [ + - ]: 1297 : get_pathtarget_sortgroupref(final_target, i));
6808 : :
6809 : 1622 : i++;
6810 : : }
6811 : :
6812 : : /*
6813 : : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6814 : : * postponable columns, and add them to the sort-input target if not
6815 : : * already present. (Some might be there already.) We mustn't
6816 : : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6817 : : * would be unable to recompute them.
6818 : : */
6819 : 273 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6820 : : PVC_INCLUDE_AGGREGATES |
6821 : : PVC_INCLUDE_WINDOWFUNCS |
6822 : : PVC_INCLUDE_PLACEHOLDERS);
6823 : 273 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6824 : :
6825 : : /* clean up cruft */
6826 : 273 : list_free(postponable_vars);
6827 : 273 : list_free(postponable_cols);
6828 : :
6829 : : /* XXX this represents even more redundant cost calculation ... */
6830 : 273 : return set_pathtarget_cost_width(root, input_target);
6831 : : }
6832 : :
6833 : : /*
6834 : : * get_cheapest_fractional_path
6835 : : * Find the cheapest path for retrieving a specified fraction of all
6836 : : * the tuples expected to be returned by the given relation.
6837 : : *
6838 : : * Do not consider parameterized paths. If the caller needs a path for upper
6839 : : * rel, it can't have parameterized paths. If the caller needs an append
6840 : : * subpath, it could become limited by the treatment of similar
6841 : : * parameterization of all the subpaths.
6842 : : *
6843 : : * We interpret tuple_fraction the same way as grouping_planner.
6844 : : *
6845 : : * We assume set_cheapest() has been run on the given rel.
6846 : : */
6847 : : Path *
6848 : 359725 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6849 : : {
6850 : 359725 : Path *best_path = rel->cheapest_total_path;
6851 : : ListCell *l;
6852 : :
6853 : : /* If all tuples will be retrieved, just return the cheapest-total path */
6854 [ + + ]: 359725 : if (tuple_fraction <= 0.0)
6855 : 352770 : return best_path;
6856 : :
6857 : : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
6858 [ + + + + ]: 6955 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
6859 : 2941 : tuple_fraction /= best_path->rows;
6860 : :
6861 [ + - + + : 17799 : foreach(l, rel->pathlist)
+ + ]
6862 : : {
6863 : 10844 : Path *path = (Path *) lfirst(l);
6864 : :
6865 [ + + ]: 10844 : if (path->param_info)
6866 : 166 : continue;
6867 : :
6868 [ + + + + ]: 14401 : if (path == rel->cheapest_total_path ||
6869 : 3723 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
6870 : 10356 : continue;
6871 : :
6872 : 322 : best_path = path;
6873 : : }
6874 : :
6875 : 6955 : return best_path;
6876 : : }
6877 : :
6878 : : /*
6879 : : * adjust_paths_for_srfs
6880 : : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6881 : : *
6882 : : * The executor can only handle set-returning functions that appear at the
6883 : : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6884 : : * that are not at top level, we need to split up the evaluation into multiple
6885 : : * plan levels in which each level satisfies this constraint. This function
6886 : : * modifies each Path of an upperrel that (might) compute any SRFs in its
6887 : : * output tlist to insert appropriate projection steps.
6888 : : *
6889 : : * The given targets and targets_contain_srfs lists are from
6890 : : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6891 : : * target in targets.
6892 : : */
6893 : : static void
6894 : 10689 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6895 : : List *targets, List *targets_contain_srfs)
6896 : : {
6897 : : ListCell *lc;
6898 : :
6899 : : Assert(list_length(targets) == list_length(targets_contain_srfs));
6900 : : Assert(!linitial_int(targets_contain_srfs));
6901 : :
6902 : : /* If no SRFs appear at this plan level, nothing to do */
6903 [ + + ]: 10689 : if (list_length(targets) == 1)
6904 : 481 : return;
6905 : :
6906 : : /*
6907 : : * Stack SRF-evaluation nodes atop each path for the rel.
6908 : : *
6909 : : * In principle we should re-run set_cheapest() here to identify the
6910 : : * cheapest path, but it seems unlikely that adding the same tlist eval
6911 : : * costs to all the paths would change that, so we don't bother. Instead,
6912 : : * just assume that the cheapest-startup and cheapest-total paths remain
6913 : : * so. (There should be no parameterized paths anymore, so we needn't
6914 : : * worry about updating cheapest_parameterized_paths.)
6915 : : */
6916 [ + - + + : 20446 : foreach(lc, rel->pathlist)
+ + ]
6917 : : {
6918 : 10238 : Path *subpath = (Path *) lfirst(lc);
6919 : 10238 : Path *newpath = subpath;
6920 : : ListCell *lc1,
6921 : : *lc2;
6922 : :
6923 : : Assert(subpath->param_info == NULL);
6924 [ + - + + : 31499 : forboth(lc1, targets, lc2, targets_contain_srfs)
+ - + + +
+ + - +
+ ]
6925 : : {
6926 : 21261 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6927 : 21261 : bool contains_srfs = (bool) lfirst_int(lc2);
6928 : :
6929 : : /* If this level doesn't contain SRFs, do regular projection */
6930 [ + + ]: 21261 : if (contains_srfs)
6931 : 10288 : newpath = (Path *) create_set_projection_path(root,
6932 : : rel,
6933 : : newpath,
6934 : : thistarget);
6935 : : else
6936 : 10973 : newpath = (Path *) apply_projection_to_path(root,
6937 : : rel,
6938 : : newpath,
6939 : : thistarget);
6940 : : }
6941 : 10238 : lfirst(lc) = newpath;
6942 [ + + ]: 10238 : if (subpath == rel->cheapest_startup_path)
6943 : 239 : rel->cheapest_startup_path = newpath;
6944 [ + + ]: 10238 : if (subpath == rel->cheapest_total_path)
6945 : 239 : rel->cheapest_total_path = newpath;
6946 : : }
6947 : :
6948 : : /* Likewise for partial paths, if any */
6949 [ + + + + : 10223 : foreach(lc, rel->partial_pathlist)
+ + ]
6950 : : {
6951 : 15 : Path *subpath = (Path *) lfirst(lc);
6952 : 15 : Path *newpath = subpath;
6953 : : ListCell *lc1,
6954 : : *lc2;
6955 : :
6956 : : Assert(subpath->param_info == NULL);
6957 [ + - + + : 60 : forboth(lc1, targets, lc2, targets_contain_srfs)
+ - + + +
+ + - +
+ ]
6958 : : {
6959 : 45 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6960 : 45 : bool contains_srfs = (bool) lfirst_int(lc2);
6961 : :
6962 : : /* If this level doesn't contain SRFs, do regular projection */
6963 [ + + ]: 45 : if (contains_srfs)
6964 : 15 : newpath = (Path *) create_set_projection_path(root,
6965 : : rel,
6966 : : newpath,
6967 : : thistarget);
6968 : : else
6969 : : {
6970 : : /* avoid apply_projection_to_path, in case of multiple refs */
6971 : 30 : newpath = (Path *) create_projection_path(root,
6972 : : rel,
6973 : : newpath,
6974 : : thistarget);
6975 : : }
6976 : : }
6977 : 15 : lfirst(lc) = newpath;
6978 : : }
6979 : : }
6980 : :
6981 : : /*
6982 : : * expression_planner
6983 : : * Perform planner's transformations on a standalone expression.
6984 : : *
6985 : : * Various utility commands need to evaluate expressions that are not part
6986 : : * of a plannable query. They can do so using the executor's regular
6987 : : * expression-execution machinery, but first the expression has to be fed
6988 : : * through here to transform it from parser output to something executable.
6989 : : *
6990 : : * Currently, we disallow sublinks in standalone expressions, so there's no
6991 : : * real "planning" involved here. (That might not always be true though.)
6992 : : * What we must do is run eval_const_expressions to ensure that any function
6993 : : * calls are converted to positional notation and function default arguments
6994 : : * get inserted. The fact that constant subexpressions get simplified is a
6995 : : * side-effect that is useful when the expression will get evaluated more than
6996 : : * once. Also, we must fix operator function IDs.
6997 : : *
6998 : : * This does not return any information about dependencies of the expression.
6999 : : * Hence callers should use the results only for the duration of the current
7000 : : * query. Callers that would like to cache the results for longer should use
7001 : : * expression_planner_with_deps, probably via the plancache.
7002 : : *
7003 : : * Note: this must not make any damaging changes to the passed-in expression
7004 : : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
7005 : : * we first do an expression_tree_mutator-based walk, what is returned will
7006 : : * be a new node tree.) The result is constructed in the current memory
7007 : : * context; beware that this can leak a lot of additional stuff there, too.
7008 : : */
7009 : : Expr *
7010 : 137464 : expression_planner(Expr *expr)
7011 : : {
7012 : : Node *result;
7013 : :
7014 : : /*
7015 : : * Convert named-argument function calls, insert default arguments and
7016 : : * simplify constant subexprs
7017 : : */
7018 : 137464 : result = eval_const_expressions(NULL, (Node *) expr);
7019 : :
7020 : : /* Fill in opfuncid values if missing */
7021 : 137452 : fix_opfuncids(result);
7022 : :
7023 : 137452 : return (Expr *) result;
7024 : : }
7025 : :
7026 : : /*
7027 : : * expression_planner_with_deps
7028 : : * Perform planner's transformations on a standalone expression,
7029 : : * returning expression dependency information along with the result.
7030 : : *
7031 : : * This is identical to expression_planner() except that it also returns
7032 : : * information about possible dependencies of the expression, ie identities of
7033 : : * objects whose definitions affect the result. As in a PlannedStmt, these
7034 : : * are expressed as a list of relation Oids and a list of PlanInvalItems.
7035 : : */
7036 : : Expr *
7037 : 264 : expression_planner_with_deps(Expr *expr,
7038 : : List **relationOids,
7039 : : List **invalItems)
7040 : : {
7041 : : Node *result;
7042 : : PlannerGlobal glob;
7043 : : PlannerInfo root;
7044 : :
7045 : : /* Make up dummy planner state so we can use setrefs machinery */
7046 [ + - + - : 7656 : MemSet(&glob, 0, sizeof(glob));
+ - + - +
+ ]
7047 : 264 : glob.type = T_PlannerGlobal;
7048 : 264 : glob.relationOids = NIL;
7049 : 264 : glob.invalItems = NIL;
7050 : :
7051 [ + - + - : 24816 : MemSet(&root, 0, sizeof(root));
+ - + - +
+ ]
7052 : 264 : root.type = T_PlannerInfo;
7053 : 264 : root.glob = &glob;
7054 : :
7055 : : /*
7056 : : * Convert named-argument function calls, insert default arguments and
7057 : : * simplify constant subexprs. Collect identities of inlined functions
7058 : : * and elided domains, too.
7059 : : */
7060 : 264 : result = eval_const_expressions(&root, (Node *) expr);
7061 : :
7062 : : /* Fill in opfuncid values if missing */
7063 : 264 : fix_opfuncids(result);
7064 : :
7065 : : /*
7066 : : * Now walk the finished expression to find anything else we ought to
7067 : : * record as an expression dependency.
7068 : : */
7069 : 264 : (void) extract_query_dependencies_walker(result, &root);
7070 : :
7071 : 264 : *relationOids = glob.relationOids;
7072 : 264 : *invalItems = glob.invalItems;
7073 : :
7074 : 264 : return (Expr *) result;
7075 : : }
7076 : :
7077 : :
7078 : : /*
7079 : : * plan_cluster_use_sort
7080 : : * Use the planner to decide how CLUSTER should implement sorting
7081 : : *
7082 : : * tableOid is the OID of a table to be clustered on its index indexOid
7083 : : * (which is already known to be a btree index). Decide whether it's
7084 : : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
7085 : : * Return true to use sorting, false to use an indexscan.
7086 : : *
7087 : : * Note: caller had better already hold some type of lock on the table.
7088 : : */
7089 : : bool
7090 : 144 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
7091 : : {
7092 : : PlannerInfo *root;
7093 : : Query *query;
7094 : : PlannerGlobal *glob;
7095 : : RangeTblEntry *rte;
7096 : : RelOptInfo *rel;
7097 : : IndexOptInfo *indexInfo;
7098 : : QualCost indexExprCost;
7099 : : Cost comparisonCost;
7100 : : Path *seqScanPath;
7101 : : Path seqScanAndSortPath;
7102 : : IndexPath *indexScanPath;
7103 : : ListCell *lc;
7104 : :
7105 : : /* We can short-circuit the cost comparison if indexscans are disabled */
7106 [ + + ]: 144 : if (!enable_indexscan)
7107 : 20 : return true; /* use sort */
7108 : :
7109 : : /* Set up mostly-dummy planner state */
7110 : 124 : query = makeNode(Query);
7111 : 124 : query->commandType = CMD_SELECT;
7112 : :
7113 : 124 : glob = makeNode(PlannerGlobal);
7114 : :
7115 : 124 : root = makeNode(PlannerInfo);
7116 : 124 : root->parse = query;
7117 : 124 : root->glob = glob;
7118 : 124 : root->query_level = 1;
7119 : 124 : root->planner_cxt = CurrentMemoryContext;
7120 : 124 : root->wt_param_id = -1;
7121 : 124 : root->join_domains = list_make1(makeNode(JoinDomain));
7122 : :
7123 : : /* Build a minimal RTE for the rel */
7124 : 124 : rte = makeNode(RangeTblEntry);
7125 : 124 : rte->rtekind = RTE_RELATION;
7126 : 124 : rte->relid = tableOid;
7127 : 124 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7128 : 124 : rte->rellockmode = AccessShareLock;
7129 : 124 : rte->lateral = false;
7130 : 124 : rte->inh = false;
7131 : 124 : rte->inFromCl = true;
7132 : 124 : query->rtable = list_make1(rte);
7133 : 124 : addRTEPermissionInfo(&query->rteperminfos, rte);
7134 : :
7135 : : /* Set up RTE/RelOptInfo arrays */
7136 : 124 : setup_simple_rel_arrays(root);
7137 : :
7138 : : /* Build RelOptInfo */
7139 : 124 : rel = build_simple_rel(root, 1, NULL);
7140 : :
7141 : : /* Locate IndexOptInfo for the target index */
7142 : 124 : indexInfo = NULL;
7143 [ + - + - : 154 : foreach(lc, rel->indexlist)
+ - ]
7144 : : {
7145 : 154 : indexInfo = lfirst_node(IndexOptInfo, lc);
7146 [ + + ]: 154 : if (indexInfo->indexoid == indexOid)
7147 : 124 : break;
7148 : : }
7149 : :
7150 : : /*
7151 : : * It's possible that get_relation_info did not generate an IndexOptInfo
7152 : : * for the desired index; this could happen if it's not yet reached its
7153 : : * indcheckxmin usability horizon, or if it's a system index and we're
7154 : : * ignoring system indexes. In such cases we should tell CLUSTER to not
7155 : : * trust the index contents but use seqscan-and-sort.
7156 : : */
7157 [ - + ]: 124 : if (lc == NULL) /* not in the list? */
7158 : 0 : return true; /* use sort */
7159 : :
7160 : : /*
7161 : : * Rather than doing all the pushups that would be needed to use
7162 : : * set_baserel_size_estimates, just do a quick hack for rows and width.
7163 : : */
7164 : 124 : rel->rows = rel->tuples;
7165 : 124 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
7166 : :
7167 : 124 : root->total_table_pages = rel->pages;
7168 : :
7169 : : /*
7170 : : * Determine eval cost of the index expressions, if any. We need to
7171 : : * charge twice that amount for each tuple comparison that happens during
7172 : : * the sort, since tuplesort.c will have to re-evaluate the index
7173 : : * expressions each time. (XXX that's pretty inefficient...)
7174 : : */
7175 : 124 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
7176 : 124 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
7177 : :
7178 : : /* Estimate the cost of seq scan + sort */
7179 : 124 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
7180 : 124 : cost_sort(&seqScanAndSortPath, root, NIL,
7181 : : seqScanPath->disabled_nodes,
7182 : 124 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
7183 : : comparisonCost, maintenance_work_mem, -1.0);
7184 : :
7185 : : /* Estimate the cost of index scan */
7186 : 124 : indexScanPath = create_index_path(root, indexInfo,
7187 : : NIL, NIL, NIL, NIL,
7188 : : ForwardScanDirection, false,
7189 : : NULL, 1.0, false);
7190 : :
7191 : 124 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
7192 : : }
7193 : :
7194 : : /*
7195 : : * plan_create_index_workers
7196 : : * Use the planner to decide how many parallel worker processes
7197 : : * CREATE INDEX should request for use
7198 : : *
7199 : : * tableOid is the table on which the index is to be built. indexOid is the
7200 : : * OID of an index to be created or reindexed (which must be an index with
7201 : : * support for parallel builds - currently btree, GIN, or BRIN).
7202 : : *
7203 : : * Return value is the number of parallel worker processes to request. It
7204 : : * may be unsafe to proceed if this is 0. Note that this does not include the
7205 : : * leader participating as a worker (value is always a number of parallel
7206 : : * worker processes).
7207 : : *
7208 : : * Note: caller had better already hold some type of lock on the table and
7209 : : * index.
7210 : : */
7211 : : int
7212 : 23393 : plan_create_index_workers(Oid tableOid, Oid indexOid)
7213 : : {
7214 : : PlannerInfo *root;
7215 : : Query *query;
7216 : : PlannerGlobal *glob;
7217 : : RangeTblEntry *rte;
7218 : : Relation heap;
7219 : : Relation index;
7220 : : RelOptInfo *rel;
7221 : : int parallel_workers;
7222 : : BlockNumber heap_blocks;
7223 : : double reltuples;
7224 : : double allvisfrac;
7225 : :
7226 : : /*
7227 : : * We don't allow performing parallel operation in standalone backend or
7228 : : * when parallelism is disabled.
7229 : : */
7230 [ + + + + ]: 23393 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
7231 : 301 : return 0;
7232 : :
7233 : : /* Set up largely-dummy planner state */
7234 : 23092 : query = makeNode(Query);
7235 : 23092 : query->commandType = CMD_SELECT;
7236 : :
7237 : 23092 : glob = makeNode(PlannerGlobal);
7238 : :
7239 : 23092 : root = makeNode(PlannerInfo);
7240 : 23092 : root->parse = query;
7241 : 23092 : root->glob = glob;
7242 : 23092 : root->query_level = 1;
7243 : 23092 : root->planner_cxt = CurrentMemoryContext;
7244 : 23092 : root->wt_param_id = -1;
7245 : 23092 : root->join_domains = list_make1(makeNode(JoinDomain));
7246 : :
7247 : : /*
7248 : : * Build a minimal RTE.
7249 : : *
7250 : : * Mark the RTE with inh = true. This is a kludge to prevent
7251 : : * get_relation_info() from fetching index info, which is necessary
7252 : : * because it does not expect that any IndexOptInfo is currently
7253 : : * undergoing REINDEX.
7254 : : */
7255 : 23092 : rte = makeNode(RangeTblEntry);
7256 : 23092 : rte->rtekind = RTE_RELATION;
7257 : 23092 : rte->relid = tableOid;
7258 : 23092 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
7259 : 23092 : rte->rellockmode = AccessShareLock;
7260 : 23092 : rte->lateral = false;
7261 : 23092 : rte->inh = true;
7262 : 23092 : rte->inFromCl = true;
7263 : 23092 : query->rtable = list_make1(rte);
7264 : 23092 : addRTEPermissionInfo(&query->rteperminfos, rte);
7265 : :
7266 : : /* Set up RTE/RelOptInfo arrays */
7267 : 23092 : setup_simple_rel_arrays(root);
7268 : :
7269 : : /* Build RelOptInfo */
7270 : 23092 : rel = build_simple_rel(root, 1, NULL);
7271 : :
7272 : : /* Rels are assumed already locked by the caller */
7273 : 23092 : heap = table_open(tableOid, NoLock);
7274 : 23092 : index = index_open(indexOid, NoLock);
7275 : :
7276 : : /*
7277 : : * Determine if it's safe to proceed.
7278 : : *
7279 : : * Currently, parallel workers can't access the leader's temporary tables.
7280 : : * Furthermore, any index predicate or index expressions must be parallel
7281 : : * safe.
7282 : : */
7283 [ + + ]: 23092 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
7284 [ + + ]: 21685 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7285 [ - + ]: 21595 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7286 : : {
7287 : 1497 : parallel_workers = 0;
7288 : 1497 : goto done;
7289 : : }
7290 : :
7291 : : /*
7292 : : * If parallel_workers storage parameter is set for the table, accept that
7293 : : * as the number of parallel worker processes to launch (though still cap
7294 : : * at max_parallel_maintenance_workers). Note that we deliberately do not
7295 : : * consider any other factor when parallel_workers is set. (e.g., memory
7296 : : * use by workers.)
7297 : : */
7298 [ + + ]: 21595 : if (rel->rel_parallel_workers != -1)
7299 : : {
7300 : 59 : parallel_workers = Min(rel->rel_parallel_workers,
7301 : : max_parallel_maintenance_workers);
7302 : 59 : goto done;
7303 : : }
7304 : :
7305 : : /*
7306 : : * Estimate heap relation size ourselves, since rel->pages cannot be
7307 : : * trusted (heap RTE was marked as inheritance parent)
7308 : : */
7309 : 21536 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7310 : :
7311 : : /*
7312 : : * Determine number of workers to scan the heap relation using generic
7313 : : * model
7314 : : */
7315 : 21536 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7316 : : max_parallel_maintenance_workers);
7317 : :
7318 : : /*
7319 : : * Cap workers based on available maintenance_work_mem as needed.
7320 : : *
7321 : : * Note that each tuplesort participant receives an even share of the
7322 : : * total maintenance_work_mem budget. Aim to leave participants
7323 : : * (including the leader as a participant) with no less than 32MB of
7324 : : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7325 : : * immediately past the threshold of being capable of launching a single
7326 : : * parallel worker to sort.
7327 : : */
7328 [ + + ]: 21645 : while (parallel_workers > 0 &&
7329 [ + + ]: 219 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
7330 : 109 : parallel_workers--;
7331 : :
7332 : 21536 : done:
7333 : 23092 : index_close(index, NoLock);
7334 : 23092 : table_close(heap, NoLock);
7335 : :
7336 : 23092 : return parallel_workers;
7337 : : }
7338 : :
7339 : : /*
7340 : : * add_paths_to_grouping_rel
7341 : : *
7342 : : * Add non-partial paths to grouping relation.
7343 : : */
7344 : : static void
7345 : 35126 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7346 : : RelOptInfo *grouped_rel,
7347 : : RelOptInfo *partially_grouped_rel,
7348 : : const AggClauseCosts *agg_costs,
7349 : : grouping_sets_data *gd,
7350 : : GroupPathExtraData *extra)
7351 : : {
7352 : 35126 : Query *parse = root->parse;
7353 : 35126 : Path *cheapest_path = input_rel->cheapest_total_path;
7354 : 35126 : Path *cheapest_partially_grouped_path = NULL;
7355 : : ListCell *lc;
7356 : 35126 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7357 : 35126 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7358 : 35126 : List *havingQual = (List *) extra->havingQual;
7359 : 35126 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7360 : 35126 : double dNumGroups = 0;
7361 : 35126 : double dNumFinalGroups = 0;
7362 : :
7363 : : /*
7364 : : * Estimate number of groups for non-split aggregation.
7365 : : */
7366 : 35126 : dNumGroups = get_number_of_groups(root,
7367 : : cheapest_path->rows,
7368 : : gd,
7369 : : extra->targetList);
7370 : :
7371 [ + + + - ]: 35126 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7372 : : {
7373 : 2503 : cheapest_partially_grouped_path =
7374 : : partially_grouped_rel->cheapest_total_path;
7375 : :
7376 : : /*
7377 : : * Estimate number of groups for final phase of partial aggregation.
7378 : : */
7379 : : dNumFinalGroups =
7380 : 2503 : get_number_of_groups(root,
7381 : : cheapest_partially_grouped_path->rows,
7382 : : gd,
7383 : : extra->targetList);
7384 : : }
7385 : :
7386 [ + + ]: 35126 : if (can_sort)
7387 : : {
7388 : : /*
7389 : : * Use any available suitably-sorted path as input, and also consider
7390 : : * sorting the cheapest-total path and incremental sort on any paths
7391 : : * with presorted keys.
7392 : : */
7393 [ + - + + : 72851 : foreach(lc, input_rel->pathlist)
+ + ]
7394 : : {
7395 : : ListCell *lc2;
7396 : 37730 : Path *path = (Path *) lfirst(lc);
7397 : 37730 : Path *path_save = path;
7398 : 37730 : List *pathkey_orderings = NIL;
7399 : :
7400 : : /* generate alternative group orderings that might be useful */
7401 : 37730 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7402 : :
7403 : : Assert(list_length(pathkey_orderings) > 0);
7404 : :
7405 [ + - + + : 75592 : foreach(lc2, pathkey_orderings)
+ + ]
7406 : : {
7407 : 37862 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7408 : :
7409 : : /* restore the path (we replace it in the loop) */
7410 : 37862 : path = path_save;
7411 : :
7412 : 37862 : path = make_ordered_path(root,
7413 : : grouped_rel,
7414 : : path,
7415 : : cheapest_path,
7416 : : info->pathkeys,
7417 : : -1.0);
7418 [ + + ]: 37862 : if (path == NULL)
7419 : 321 : continue;
7420 : :
7421 : : /* Now decide what to stick atop it */
7422 [ + + ]: 37541 : if (parse->groupingSets)
7423 : : {
7424 : 975 : consider_groupingsets_paths(root, grouped_rel,
7425 : : path, true, can_hash,
7426 : : gd, agg_costs, dNumGroups);
7427 : : }
7428 [ + + ]: 36566 : else if (parse->hasAggs)
7429 : : {
7430 : : /*
7431 : : * We have aggregation, possibly with plain GROUP BY. Make
7432 : : * an AggPath.
7433 : : */
7434 : 35874 : add_path(grouped_rel, (Path *)
7435 : 35874 : create_agg_path(root,
7436 : : grouped_rel,
7437 : : path,
7438 : 35874 : grouped_rel->reltarget,
7439 : 35874 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7440 : : AGGSPLIT_SIMPLE,
7441 : : info->clauses,
7442 : : havingQual,
7443 : : agg_costs,
7444 : : dNumGroups));
7445 : : }
7446 [ + - ]: 692 : else if (parse->groupClause)
7447 : : {
7448 : : /*
7449 : : * We have GROUP BY without aggregation or grouping sets.
7450 : : * Make a GroupPath.
7451 : : */
7452 : 692 : add_path(grouped_rel, (Path *)
7453 : 692 : create_group_path(root,
7454 : : grouped_rel,
7455 : : path,
7456 : : info->clauses,
7457 : : havingQual,
7458 : : dNumGroups));
7459 : : }
7460 : : else
7461 : : {
7462 : : /* Other cases should have been handled above */
7463 : : Assert(false);
7464 : : }
7465 : : }
7466 : : }
7467 : :
7468 : : /*
7469 : : * Instead of operating directly on the input relation, we can
7470 : : * consider finalizing a partially aggregated path.
7471 : : */
7472 [ + + ]: 35121 : if (partially_grouped_rel != NULL)
7473 : : {
7474 [ + - + + : 6452 : foreach(lc, partially_grouped_rel->pathlist)
+ + ]
7475 : : {
7476 : : ListCell *lc2;
7477 : 3949 : Path *path = (Path *) lfirst(lc);
7478 : 3949 : Path *path_save = path;
7479 : 3949 : List *pathkey_orderings = NIL;
7480 : :
7481 : : /* generate alternative group orderings that might be useful */
7482 : 3949 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7483 : :
7484 : : Assert(list_length(pathkey_orderings) > 0);
7485 : :
7486 : : /* process all potentially interesting grouping reorderings */
7487 [ + - + + : 7898 : foreach(lc2, pathkey_orderings)
+ + ]
7488 : : {
7489 : 3949 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7490 : :
7491 : : /* restore the path (we replace it in the loop) */
7492 : 3949 : path = path_save;
7493 : :
7494 : 3949 : path = make_ordered_path(root,
7495 : : grouped_rel,
7496 : : path,
7497 : : cheapest_partially_grouped_path,
7498 : : info->pathkeys,
7499 : : -1.0);
7500 : :
7501 [ + + ]: 3949 : if (path == NULL)
7502 : 172 : continue;
7503 : :
7504 [ + + ]: 3777 : if (parse->hasAggs)
7505 : 3575 : add_path(grouped_rel, (Path *)
7506 : 3575 : create_agg_path(root,
7507 : : grouped_rel,
7508 : : path,
7509 : 3575 : grouped_rel->reltarget,
7510 : 3575 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7511 : : AGGSPLIT_FINAL_DESERIAL,
7512 : : info->clauses,
7513 : : havingQual,
7514 : : agg_final_costs,
7515 : : dNumFinalGroups));
7516 : : else
7517 : 202 : add_path(grouped_rel, (Path *)
7518 : 202 : create_group_path(root,
7519 : : grouped_rel,
7520 : : path,
7521 : : info->clauses,
7522 : : havingQual,
7523 : : dNumFinalGroups));
7524 : :
7525 : : }
7526 : : }
7527 : : }
7528 : : }
7529 : :
7530 [ + + ]: 35126 : if (can_hash)
7531 : : {
7532 [ + + ]: 5201 : if (parse->groupingSets)
7533 : : {
7534 : : /*
7535 : : * Try for a hash-only groupingsets path over unsorted input.
7536 : : */
7537 : 807 : consider_groupingsets_paths(root, grouped_rel,
7538 : : cheapest_path, false, true,
7539 : : gd, agg_costs, dNumGroups);
7540 : : }
7541 : : else
7542 : : {
7543 : : /*
7544 : : * Generate a HashAgg Path. We just need an Agg over the
7545 : : * cheapest-total input path, since input order won't matter.
7546 : : */
7547 : 4394 : add_path(grouped_rel, (Path *)
7548 : 4394 : create_agg_path(root, grouped_rel,
7549 : : cheapest_path,
7550 : 4394 : grouped_rel->reltarget,
7551 : : AGG_HASHED,
7552 : : AGGSPLIT_SIMPLE,
7553 : : root->processed_groupClause,
7554 : : havingQual,
7555 : : agg_costs,
7556 : : dNumGroups));
7557 : : }
7558 : :
7559 : : /*
7560 : : * Generate a Finalize HashAgg Path atop of the cheapest partially
7561 : : * grouped path, assuming there is one
7562 : : */
7563 [ + + + - ]: 5201 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7564 : : {
7565 : 1192 : add_path(grouped_rel, (Path *)
7566 : 1192 : create_agg_path(root,
7567 : : grouped_rel,
7568 : : cheapest_partially_grouped_path,
7569 : 1192 : grouped_rel->reltarget,
7570 : : AGG_HASHED,
7571 : : AGGSPLIT_FINAL_DESERIAL,
7572 : : root->processed_groupClause,
7573 : : havingQual,
7574 : : agg_final_costs,
7575 : : dNumFinalGroups));
7576 : : }
7577 : : }
7578 : :
7579 : : /*
7580 : : * When partitionwise aggregate is used, we might have fully aggregated
7581 : : * paths in the partial pathlist, because add_paths_to_append_rel() will
7582 : : * consider a path for grouped_rel consisting of a Parallel Append of
7583 : : * non-partial paths from each child.
7584 : : */
7585 [ + + ]: 35126 : if (grouped_rel->partial_pathlist != NIL)
7586 : 265 : gather_grouping_paths(root, grouped_rel);
7587 : 35126 : }
7588 : :
7589 : : /*
7590 : : * create_partial_grouping_paths
7591 : : *
7592 : : * Create a new upper relation representing the result of partial aggregation
7593 : : * and populate it with appropriate paths. Note that we don't finalize the
7594 : : * lists of paths here, so the caller can add additional partial or non-partial
7595 : : * paths and must afterward call gather_grouping_paths and set_cheapest on
7596 : : * the returned upper relation.
7597 : : *
7598 : : * All paths for this new upper relation -- both partial and non-partial --
7599 : : * have been partially aggregated but require a subsequent FinalizeAggregate
7600 : : * step.
7601 : : *
7602 : : * NB: This function is allowed to return NULL if it determines that there is
7603 : : * no real need to create a new RelOptInfo.
7604 : : */
7605 : : static RelOptInfo *
7606 : 31836 : create_partial_grouping_paths(PlannerInfo *root,
7607 : : RelOptInfo *grouped_rel,
7608 : : RelOptInfo *input_rel,
7609 : : grouping_sets_data *gd,
7610 : : GroupPathExtraData *extra,
7611 : : bool force_rel_creation)
7612 : : {
7613 : 31836 : Query *parse = root->parse;
7614 : : RelOptInfo *partially_grouped_rel;
7615 : 31836 : RelOptInfo *eager_agg_rel = NULL;
7616 : 31836 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7617 : 31836 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7618 : 31836 : Path *cheapest_partial_path = NULL;
7619 : 31836 : Path *cheapest_total_path = NULL;
7620 : 31836 : double dNumPartialGroups = 0;
7621 : 31836 : double dNumPartialPartialGroups = 0;
7622 : : ListCell *lc;
7623 : 31836 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7624 : 31836 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7625 : :
7626 : : /*
7627 : : * Check whether any partially aggregated paths have been generated
7628 : : * through eager aggregation.
7629 : : */
7630 [ + + ]: 31836 : if (input_rel->grouped_rel &&
7631 [ + - ]: 797 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7632 [ + + ]: 797 : input_rel->grouped_rel->pathlist != NIL)
7633 : 737 : eager_agg_rel = input_rel->grouped_rel;
7634 : :
7635 : : /*
7636 : : * Consider whether we should generate partially aggregated non-partial
7637 : : * paths. We can only do this if we have a non-partial path, and only if
7638 : : * the parent of the input rel is performing partial partitionwise
7639 : : * aggregation. (Note that extra->patype is the type of partitionwise
7640 : : * aggregation being used at the parent level, not this level.)
7641 : : */
7642 [ + - ]: 31836 : if (input_rel->pathlist != NIL &&
7643 [ + + ]: 31836 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7644 : 713 : cheapest_total_path = input_rel->cheapest_total_path;
7645 : :
7646 : : /*
7647 : : * If parallelism is possible for grouped_rel, then we should consider
7648 : : * generating partially-grouped partial paths. However, if the input rel
7649 : : * has no partial paths, then we can't.
7650 : : */
7651 [ + + + + ]: 31836 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7652 : 2745 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7653 : :
7654 : : /*
7655 : : * If we can't partially aggregate partial paths, and we can't partially
7656 : : * aggregate non-partial paths, and no partially aggregated paths were
7657 : : * generated by eager aggregation, then don't bother creating the new
7658 : : * RelOptInfo at all, unless the caller specified force_rel_creation.
7659 : : */
7660 [ + + + + ]: 31836 : if (cheapest_total_path == NULL &&
7661 [ + + ]: 28798 : cheapest_partial_path == NULL &&
7662 : 28701 : eager_agg_rel == NULL &&
7663 [ + + ]: 28701 : !force_rel_creation)
7664 : 28620 : return NULL;
7665 : :
7666 : : /*
7667 : : * Build a new upper relation to represent the result of partially
7668 : : * aggregating the rows from the input relation.
7669 : : */
7670 : 3216 : partially_grouped_rel = fetch_upper_rel(root,
7671 : : UPPERREL_PARTIAL_GROUP_AGG,
7672 : : grouped_rel->relids);
7673 : 3216 : partially_grouped_rel->consider_parallel =
7674 : 3216 : grouped_rel->consider_parallel;
7675 : 3216 : partially_grouped_rel->pgs_mask = grouped_rel->pgs_mask;
7676 : 3216 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
7677 : 3216 : partially_grouped_rel->serverid = grouped_rel->serverid;
7678 : 3216 : partially_grouped_rel->userid = grouped_rel->userid;
7679 : 3216 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7680 : 3216 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7681 : :
7682 : : /*
7683 : : * Build target list for partial aggregate paths. These paths cannot just
7684 : : * emit the same tlist as regular aggregate paths, because (1) we must
7685 : : * include Vars and Aggrefs needed in HAVING, which might not appear in
7686 : : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7687 : : */
7688 : 3216 : partially_grouped_rel->reltarget =
7689 : 3216 : make_partial_grouping_target(root, grouped_rel->reltarget,
7690 : : extra->havingQual);
7691 : :
7692 [ + + ]: 3216 : if (!extra->partial_costs_set)
7693 : : {
7694 : : /*
7695 : : * Collect statistics about aggregates for estimating costs of
7696 : : * performing aggregation in parallel.
7697 : : */
7698 [ + - + - : 11598 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
7699 [ + - + - : 11598 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
7700 [ + + ]: 1933 : if (parse->hasAggs)
7701 : : {
7702 : : /* partial phase */
7703 : 1824 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7704 : : agg_partial_costs);
7705 : :
7706 : : /* final phase */
7707 : 1824 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7708 : : agg_final_costs);
7709 : : }
7710 : :
7711 : 1933 : extra->partial_costs_set = true;
7712 : : }
7713 : :
7714 : : /* Estimate number of partial groups. */
7715 [ + + ]: 3216 : if (cheapest_total_path != NULL)
7716 : : dNumPartialGroups =
7717 : 713 : get_number_of_groups(root,
7718 : : cheapest_total_path->rows,
7719 : : gd,
7720 : : extra->targetList);
7721 [ + + ]: 3216 : if (cheapest_partial_path != NULL)
7722 : : dNumPartialPartialGroups =
7723 : 2745 : get_number_of_groups(root,
7724 : : cheapest_partial_path->rows,
7725 : : gd,
7726 : : extra->targetList);
7727 : :
7728 [ + - + + ]: 3216 : if (can_sort && cheapest_total_path != NULL)
7729 : : {
7730 : : /* This should have been checked previously */
7731 : : Assert(parse->hasAggs || parse->groupClause);
7732 : :
7733 : : /*
7734 : : * Use any available suitably-sorted path as input, and also consider
7735 : : * sorting the cheapest partial path.
7736 : : */
7737 [ + - + + : 1426 : foreach(lc, input_rel->pathlist)
+ + ]
7738 : : {
7739 : : ListCell *lc2;
7740 : 713 : Path *path = (Path *) lfirst(lc);
7741 : 713 : Path *path_save = path;
7742 : 713 : List *pathkey_orderings = NIL;
7743 : :
7744 : : /* generate alternative group orderings that might be useful */
7745 : 713 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7746 : :
7747 : : Assert(list_length(pathkey_orderings) > 0);
7748 : :
7749 : : /* process all potentially interesting grouping reorderings */
7750 [ + - + + : 1426 : foreach(lc2, pathkey_orderings)
+ + ]
7751 : : {
7752 : 713 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7753 : :
7754 : : /* restore the path (we replace it in the loop) */
7755 : 713 : path = path_save;
7756 : :
7757 : 713 : path = make_ordered_path(root,
7758 : : partially_grouped_rel,
7759 : : path,
7760 : : cheapest_total_path,
7761 : : info->pathkeys,
7762 : : -1.0);
7763 : :
7764 [ - + ]: 713 : if (path == NULL)
7765 : 0 : continue;
7766 : :
7767 [ + + ]: 713 : if (parse->hasAggs)
7768 : 653 : add_path(partially_grouped_rel, (Path *)
7769 : 653 : create_agg_path(root,
7770 : : partially_grouped_rel,
7771 : : path,
7772 : 653 : partially_grouped_rel->reltarget,
7773 : 653 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7774 : : AGGSPLIT_INITIAL_SERIAL,
7775 : : info->clauses,
7776 : : NIL,
7777 : : agg_partial_costs,
7778 : : dNumPartialGroups));
7779 : : else
7780 : 60 : add_path(partially_grouped_rel, (Path *)
7781 : 60 : create_group_path(root,
7782 : : partially_grouped_rel,
7783 : : path,
7784 : : info->clauses,
7785 : : NIL,
7786 : : dNumPartialGroups));
7787 : : }
7788 : : }
7789 : : }
7790 : :
7791 [ + - + + ]: 3216 : if (can_sort && cheapest_partial_path != NULL)
7792 : : {
7793 : : /* Similar to above logic, but for partial paths. */
7794 [ + - + + : 5865 : foreach(lc, input_rel->partial_pathlist)
+ + ]
7795 : : {
7796 : : ListCell *lc2;
7797 : 3120 : Path *path = (Path *) lfirst(lc);
7798 : 3120 : Path *path_save = path;
7799 : 3120 : List *pathkey_orderings = NIL;
7800 : :
7801 : : /* generate alternative group orderings that might be useful */
7802 : 3120 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7803 : :
7804 : : Assert(list_length(pathkey_orderings) > 0);
7805 : :
7806 : : /* process all potentially interesting grouping reorderings */
7807 [ + - + + : 6240 : foreach(lc2, pathkey_orderings)
+ + ]
7808 : : {
7809 : 3120 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7810 : :
7811 : :
7812 : : /* restore the path (we replace it in the loop) */
7813 : 3120 : path = path_save;
7814 : :
7815 : 3120 : path = make_ordered_path(root,
7816 : : partially_grouped_rel,
7817 : : path,
7818 : : cheapest_partial_path,
7819 : : info->pathkeys,
7820 : : -1.0);
7821 : :
7822 [ + + ]: 3120 : if (path == NULL)
7823 : 5 : continue;
7824 : :
7825 [ + + ]: 3115 : if (parse->hasAggs)
7826 : 3016 : add_partial_path(partially_grouped_rel, (Path *)
7827 : 3016 : create_agg_path(root,
7828 : : partially_grouped_rel,
7829 : : path,
7830 : 3016 : partially_grouped_rel->reltarget,
7831 : 3016 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7832 : : AGGSPLIT_INITIAL_SERIAL,
7833 : : info->clauses,
7834 : : NIL,
7835 : : agg_partial_costs,
7836 : : dNumPartialPartialGroups));
7837 : : else
7838 : 99 : add_partial_path(partially_grouped_rel, (Path *)
7839 : 99 : create_group_path(root,
7840 : : partially_grouped_rel,
7841 : : path,
7842 : : info->clauses,
7843 : : NIL,
7844 : : dNumPartialPartialGroups));
7845 : : }
7846 : : }
7847 : : }
7848 : :
7849 : : /*
7850 : : * Add a partially-grouped HashAgg Path where possible
7851 : : */
7852 [ + + + + ]: 3216 : if (can_hash && cheapest_total_path != NULL)
7853 : : {
7854 : : /* Checked above */
7855 : : Assert(parse->hasAggs || parse->groupClause);
7856 : :
7857 : 713 : add_path(partially_grouped_rel, (Path *)
7858 : 713 : create_agg_path(root,
7859 : : partially_grouped_rel,
7860 : : cheapest_total_path,
7861 : 713 : partially_grouped_rel->reltarget,
7862 : : AGG_HASHED,
7863 : : AGGSPLIT_INITIAL_SERIAL,
7864 : : root->processed_groupClause,
7865 : : NIL,
7866 : : agg_partial_costs,
7867 : : dNumPartialGroups));
7868 : : }
7869 : :
7870 : : /*
7871 : : * Now add a partially-grouped HashAgg partial Path where possible
7872 : : */
7873 [ + + + + ]: 3216 : if (can_hash && cheapest_partial_path != NULL)
7874 : : {
7875 : 1434 : add_partial_path(partially_grouped_rel, (Path *)
7876 : 1434 : create_agg_path(root,
7877 : : partially_grouped_rel,
7878 : : cheapest_partial_path,
7879 : 1434 : partially_grouped_rel->reltarget,
7880 : : AGG_HASHED,
7881 : : AGGSPLIT_INITIAL_SERIAL,
7882 : : root->processed_groupClause,
7883 : : NIL,
7884 : : agg_partial_costs,
7885 : : dNumPartialPartialGroups));
7886 : : }
7887 : :
7888 : : /*
7889 : : * Add any partially aggregated paths generated by eager aggregation to
7890 : : * the new upper relation after applying projection steps as needed.
7891 : : */
7892 [ + + ]: 3216 : if (eager_agg_rel)
7893 : : {
7894 : : /* Add the paths */
7895 [ + - + + : 1924 : foreach(lc, eager_agg_rel->pathlist)
+ + ]
7896 : : {
7897 : 1187 : Path *path = (Path *) lfirst(lc);
7898 : :
7899 : : /* Shouldn't have any parameterized paths anymore */
7900 : : Assert(path->param_info == NULL);
7901 : :
7902 : 1187 : path = (Path *) create_projection_path(root,
7903 : : partially_grouped_rel,
7904 : : path,
7905 : 1187 : partially_grouped_rel->reltarget);
7906 : :
7907 : 1187 : add_path(partially_grouped_rel, path);
7908 : : }
7909 : :
7910 : : /*
7911 : : * Likewise add the partial paths, but only if parallelism is possible
7912 : : * for partially_grouped_rel.
7913 : : */
7914 [ + + ]: 737 : if (partially_grouped_rel->consider_parallel)
7915 : : {
7916 [ + + + + : 1700 : foreach(lc, eager_agg_rel->partial_pathlist)
+ + ]
7917 : : {
7918 : 1010 : Path *path = (Path *) lfirst(lc);
7919 : :
7920 : : /* Shouldn't have any parameterized paths anymore */
7921 : : Assert(path->param_info == NULL);
7922 : :
7923 : 1010 : path = (Path *) create_projection_path(root,
7924 : : partially_grouped_rel,
7925 : : path,
7926 : 1010 : partially_grouped_rel->reltarget);
7927 : :
7928 : 1010 : add_partial_path(partially_grouped_rel, path);
7929 : : }
7930 : : }
7931 : : }
7932 : :
7933 : : /*
7934 : : * If there is an FDW that's responsible for all baserels of the query,
7935 : : * let it consider adding partially grouped ForeignPaths.
7936 : : */
7937 [ + + ]: 3216 : if (partially_grouped_rel->fdwroutine &&
7938 [ + - ]: 3 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7939 : : {
7940 : 3 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7941 : :
7942 : 3 : fdwroutine->GetForeignUpperPaths(root,
7943 : : UPPERREL_PARTIAL_GROUP_AGG,
7944 : : input_rel, partially_grouped_rel,
7945 : : extra);
7946 : : }
7947 : :
7948 : 3216 : return partially_grouped_rel;
7949 : : }
7950 : :
7951 : : /*
7952 : : * make_ordered_path
7953 : : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7954 : : * return NULL if it doesn't make sense to generate an ordered path in
7955 : : * this case.
7956 : : */
7957 : : static Path *
7958 : 49696 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7959 : : Path *cheapest_path, List *pathkeys, double limit_tuples)
7960 : : {
7961 : : bool is_sorted;
7962 : : int presorted_keys;
7963 : :
7964 : 49696 : is_sorted = pathkeys_count_contained_in(pathkeys,
7965 : : path->pathkeys,
7966 : : &presorted_keys);
7967 : :
7968 [ + + ]: 49696 : if (!is_sorted)
7969 : : {
7970 : : /*
7971 : : * Try at least sorting the cheapest path and also try incrementally
7972 : : * sorting any path which is partially sorted already (no need to deal
7973 : : * with paths which have presorted keys when incremental sort is
7974 : : * disabled unless it's the cheapest input path).
7975 : : */
7976 [ + + ]: 13358 : if (path != cheapest_path &&
7977 [ + + + + ]: 2217 : (presorted_keys == 0 || !enable_incremental_sort))
7978 : 1052 : return NULL;
7979 : :
7980 : : /*
7981 : : * We've no need to consider both a sort and incremental sort. We'll
7982 : : * just do a sort if there are no presorted keys and an incremental
7983 : : * sort when there are presorted keys.
7984 : : */
7985 [ + + + + ]: 12306 : if (presorted_keys == 0 || !enable_incremental_sort)
7986 : 10995 : path = (Path *) create_sort_path(root,
7987 : : rel,
7988 : : path,
7989 : : pathkeys,
7990 : : limit_tuples);
7991 : : else
7992 : 1311 : path = (Path *) create_incremental_sort_path(root,
7993 : : rel,
7994 : : path,
7995 : : pathkeys,
7996 : : presorted_keys,
7997 : : limit_tuples);
7998 : : }
7999 : :
8000 : 48644 : return path;
8001 : : }
8002 : :
8003 : : /*
8004 : : * Generate Gather and Gather Merge paths for a grouping relation or partial
8005 : : * grouping relation.
8006 : : *
8007 : : * generate_useful_gather_paths does most of the work, but we also consider a
8008 : : * special case: we could try sorting the data by the group_pathkeys and then
8009 : : * applying Gather Merge.
8010 : : *
8011 : : * NB: This function shouldn't be used for anything other than a grouped or
8012 : : * partially grouped relation not only because of the fact that it explicitly
8013 : : * references group_pathkeys but we pass "true" as the third argument to
8014 : : * generate_useful_gather_paths().
8015 : : */
8016 : : static void
8017 : 2590 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
8018 : : {
8019 : : ListCell *lc;
8020 : : Path *cheapest_partial_path;
8021 : : List *groupby_pathkeys;
8022 : :
8023 : : /*
8024 : : * This occurs after any partial aggregation has taken place, so trim off
8025 : : * any pathkeys added for ORDER BY / DISTINCT aggregates.
8026 : : */
8027 [ + + ]: 2590 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
8028 : 15 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
8029 : : root->num_groupby_pathkeys);
8030 : : else
8031 : 2575 : groupby_pathkeys = root->group_pathkeys;
8032 : :
8033 : : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
8034 : 2590 : generate_useful_gather_paths(root, rel, true);
8035 : :
8036 : 2590 : cheapest_partial_path = linitial(rel->partial_pathlist);
8037 : :
8038 : : /* XXX Shouldn't this also consider the group-key-reordering? */
8039 [ + - + + : 6152 : foreach(lc, rel->partial_pathlist)
+ + ]
8040 : : {
8041 : 3562 : Path *path = (Path *) lfirst(lc);
8042 : : bool is_sorted;
8043 : : int presorted_keys;
8044 : : double total_groups;
8045 : :
8046 : 3562 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
8047 : : path->pathkeys,
8048 : : &presorted_keys);
8049 : :
8050 [ + + ]: 3562 : if (is_sorted)
8051 : 2329 : continue;
8052 : :
8053 : : /*
8054 : : * Try at least sorting the cheapest path and also try incrementally
8055 : : * sorting any path which is partially sorted already (no need to deal
8056 : : * with paths which have presorted keys when incremental sort is
8057 : : * disabled unless it's the cheapest input path).
8058 : : */
8059 [ - + ]: 1233 : if (path != cheapest_partial_path &&
8060 [ # # # # ]: 0 : (presorted_keys == 0 || !enable_incremental_sort))
8061 : 0 : continue;
8062 : :
8063 : : /*
8064 : : * We've no need to consider both a sort and incremental sort. We'll
8065 : : * just do a sort if there are no presorted keys and an incremental
8066 : : * sort when there are presorted keys.
8067 : : */
8068 [ - + - - ]: 1233 : if (presorted_keys == 0 || !enable_incremental_sort)
8069 : 1233 : path = (Path *) create_sort_path(root, rel, path,
8070 : : groupby_pathkeys,
8071 : : -1.0);
8072 : : else
8073 : 0 : path = (Path *) create_incremental_sort_path(root,
8074 : : rel,
8075 : : path,
8076 : : groupby_pathkeys,
8077 : : presorted_keys,
8078 : : -1.0);
8079 : 1233 : total_groups = compute_gather_rows(path);
8080 : : path = (Path *)
8081 : 1233 : create_gather_merge_path(root,
8082 : : rel,
8083 : : path,
8084 : 1233 : rel->reltarget,
8085 : : groupby_pathkeys,
8086 : : NULL,
8087 : : &total_groups);
8088 : :
8089 : 1233 : add_path(rel, path);
8090 : : }
8091 : 2590 : }
8092 : :
8093 : : /*
8094 : : * can_partial_agg
8095 : : *
8096 : : * Determines whether or not partial grouping and/or aggregation is possible.
8097 : : * Returns true when possible, false otherwise.
8098 : : */
8099 : : static bool
8100 : 34044 : can_partial_agg(PlannerInfo *root)
8101 : : {
8102 : 34044 : Query *parse = root->parse;
8103 : :
8104 [ + + - + ]: 34044 : if (!parse->hasAggs && parse->groupClause == NIL)
8105 : : {
8106 : : /*
8107 : : * We don't know how to do parallel aggregation unless we have either
8108 : : * some aggregates or a grouping clause.
8109 : : */
8110 : 0 : return false;
8111 : : }
8112 [ + + ]: 34044 : else if (parse->groupingSets)
8113 : : {
8114 : : /* We don't know how to do grouping sets in parallel. */
8115 : 886 : return false;
8116 : : }
8117 [ + + + + ]: 33158 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
8118 : : {
8119 : : /* Insufficient support for partial mode. */
8120 : 3057 : return false;
8121 : : }
8122 : :
8123 : : /* Everything looks good. */
8124 : 30101 : return true;
8125 : : }
8126 : :
8127 : : /*
8128 : : * apply_scanjoin_target_to_paths
8129 : : *
8130 : : * Adjust the final scan/join relation, and recursively all of its children,
8131 : : * to generate the final scan/join target. It would be more correct to model
8132 : : * this as a separate planning step with a new RelOptInfo at the toplevel and
8133 : : * for each child relation, but doing it this way is noticeably cheaper.
8134 : : * Maybe that problem can be solved at some point, but for now we do this.
8135 : : *
8136 : : * If tlist_same_exprs is true, then the scan/join target to be applied has
8137 : : * the same expressions as the existing reltarget, so we need only insert the
8138 : : * appropriate sortgroupref information. By avoiding the creation of
8139 : : * projection paths we save effort both immediately and at plan creation time.
8140 : : */
8141 : : static void
8142 : 406688 : apply_scanjoin_target_to_paths(PlannerInfo *root,
8143 : : RelOptInfo *rel,
8144 : : List *scanjoin_targets,
8145 : : List *scanjoin_targets_contain_srfs,
8146 : : bool scanjoin_target_parallel_safe,
8147 : : bool tlist_same_exprs)
8148 : : {
8149 [ + + + + : 406688 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
+ + + - +
+ ]
8150 : : PathTarget *scanjoin_target;
8151 : : ListCell *lc;
8152 : :
8153 : : /* This recurses, so be paranoid. */
8154 : 406688 : check_stack_depth();
8155 : :
8156 : : /*
8157 : : * If the rel only has Append and MergeAppend paths, we want to drop its
8158 : : * existing paths and generate new ones. This function would still be
8159 : : * correct if we kept the existing paths: we'd modify them to generate the
8160 : : * correct target above the partitioning Append, and then they'd compete
8161 : : * on cost with paths generating the target below the Append. However, in
8162 : : * our current cost model the latter way is always the same or cheaper
8163 : : * cost, so modifying the existing paths would just be useless work.
8164 : : * Moreover, when the cost is the same, varying roundoff errors might
8165 : : * sometimes allow an existing path to be picked, resulting in undesirable
8166 : : * cross-platform plan variations. So we drop old paths and thereby force
8167 : : * the work to be done below the Append.
8168 : : *
8169 : : * However, there are several cases when this optimization is not safe. If
8170 : : * the rel isn't partitioned, then none of the paths will be Append or
8171 : : * MergeAppend paths, so we should definitely not do this. If it is
8172 : : * partitioned but is a joinrel, it may have Append and MergeAppend paths,
8173 : : * but it can also have join paths that we can't afford to discard.
8174 : : *
8175 : : * Some care is needed, because we have to allow
8176 : : * generate_useful_gather_paths to see the old partial paths in the next
8177 : : * stanza. Hence, zap the main pathlist here, then allow
8178 : : * generate_useful_gather_paths to add path(s) to the main list, and
8179 : : * finally zap the partial pathlist.
8180 : : */
8181 [ + + + + : 406688 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
+ + ]
8182 : 8751 : rel->pathlist = NIL;
8183 : :
8184 : : /*
8185 : : * If the scan/join target is not parallel-safe, partial paths cannot
8186 : : * generate it.
8187 : : */
8188 [ + + ]: 406688 : if (!scanjoin_target_parallel_safe)
8189 : : {
8190 : : /*
8191 : : * Since we can't generate the final scan/join target in parallel
8192 : : * workers, this is our last opportunity to use any partial paths that
8193 : : * exist; so build Gather path(s) that use them and emit whatever the
8194 : : * current reltarget is. We don't do this in the case where the
8195 : : * target is parallel-safe, since we will be able to generate superior
8196 : : * paths by doing it after the final scan/join target has been
8197 : : * applied.
8198 : : */
8199 : 52943 : generate_useful_gather_paths(root, rel, false);
8200 : :
8201 : : /* Can't use parallel query above this level. */
8202 : 52943 : rel->partial_pathlist = NIL;
8203 : 52943 : rel->consider_parallel = false;
8204 : : }
8205 : :
8206 : : /* Finish dropping old paths for a partitioned rel, per comment above */
8207 [ + + + + : 406688 : if (rel_is_partitioned && IS_SIMPLE_REL(rel))
+ + ]
8208 : 8751 : rel->partial_pathlist = NIL;
8209 : :
8210 : : /* Extract SRF-free scan/join target. */
8211 : 406688 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
8212 : :
8213 : : /*
8214 : : * Apply the SRF-free scan/join target to each existing path.
8215 : : *
8216 : : * If the tlist exprs are the same, we can just inject the sortgroupref
8217 : : * information into the existing pathtargets. Otherwise, replace each
8218 : : * path with a projection path that generates the SRF-free scan/join
8219 : : * target. This can't change the ordering of paths within rel->pathlist,
8220 : : * so we just modify the list in place.
8221 : : */
8222 [ + + + + : 848629 : foreach(lc, rel->pathlist)
+ + ]
8223 : : {
8224 : 441941 : Path *subpath = (Path *) lfirst(lc);
8225 : :
8226 : : /* Shouldn't have any parameterized paths anymore */
8227 : : Assert(subpath->param_info == NULL);
8228 : :
8229 [ + + ]: 441941 : if (tlist_same_exprs)
8230 : 160621 : subpath->pathtarget->sortgrouprefs =
8231 : 160621 : scanjoin_target->sortgrouprefs;
8232 : : else
8233 : : {
8234 : : Path *newpath;
8235 : :
8236 : 281320 : newpath = (Path *) create_projection_path(root, rel, subpath,
8237 : : scanjoin_target);
8238 : 281320 : lfirst(lc) = newpath;
8239 : : }
8240 : : }
8241 : :
8242 : : /* Likewise adjust the targets for any partial paths. */
8243 [ + + + + : 426937 : foreach(lc, rel->partial_pathlist)
+ + ]
8244 : : {
8245 : 20249 : Path *subpath = (Path *) lfirst(lc);
8246 : :
8247 : : /* Shouldn't have any parameterized paths anymore */
8248 : : Assert(subpath->param_info == NULL);
8249 : :
8250 [ + + ]: 20249 : if (tlist_same_exprs)
8251 : 16116 : subpath->pathtarget->sortgrouprefs =
8252 : 16116 : scanjoin_target->sortgrouprefs;
8253 : : else
8254 : : {
8255 : : Path *newpath;
8256 : :
8257 : 4133 : newpath = (Path *) create_projection_path(root, rel, subpath,
8258 : : scanjoin_target);
8259 : 4133 : lfirst(lc) = newpath;
8260 : : }
8261 : : }
8262 : :
8263 : : /*
8264 : : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8265 : : * atop each existing path. (Note that this function doesn't look at the
8266 : : * cheapest-path fields, which is a good thing because they're bogus right
8267 : : * now.)
8268 : : */
8269 [ + + ]: 406688 : if (root->parse->hasTargetSRFs)
8270 : 10208 : adjust_paths_for_srfs(root, rel,
8271 : : scanjoin_targets,
8272 : : scanjoin_targets_contain_srfs);
8273 : :
8274 : : /*
8275 : : * Update the rel's target to be the final (with SRFs) scan/join target.
8276 : : * This now matches the actual output of all the paths, and we might get
8277 : : * confused in createplan.c if they don't agree. We must do this now so
8278 : : * that any append paths made in the next part will use the correct
8279 : : * pathtarget (cf. create_append_path).
8280 : : *
8281 : : * Note that this is also necessary if GetForeignUpperPaths() gets called
8282 : : * on the final scan/join relation or on any of its children, since the
8283 : : * FDW might look at the rel's target to create ForeignPaths.
8284 : : */
8285 : 406688 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8286 : :
8287 : : /*
8288 : : * If the relation is partitioned, recursively apply the scan/join target
8289 : : * to all partitions, and generate brand-new Append paths in which the
8290 : : * scan/join target is computed below the Append rather than above it.
8291 : : * Since Append is not projection-capable, that might save a separate
8292 : : * Result node, and it also is important for partitionwise aggregate.
8293 : : */
8294 [ + + ]: 406688 : if (rel_is_partitioned)
8295 : : {
8296 : 10054 : List *live_children = NIL;
8297 : : int i;
8298 : :
8299 : : /* Adjust each partition. */
8300 : 10054 : i = -1;
8301 [ + + ]: 30186 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8302 : : {
8303 : 20132 : RelOptInfo *child_rel = rel->part_rels[i];
8304 : : AppendRelInfo **appinfos;
8305 : : int nappinfos;
8306 : 20132 : List *child_scanjoin_targets = NIL;
8307 : :
8308 : : Assert(child_rel != NULL);
8309 : :
8310 : : /* Dummy children can be ignored. */
8311 [ + + ]: 20132 : if (IS_DUMMY_REL(child_rel))
8312 : 40 : continue;
8313 : :
8314 : : /* Translate scan/join targets for this child. */
8315 : 20092 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8316 : : &nappinfos);
8317 [ + - + + : 40184 : foreach(lc, scanjoin_targets)
+ + ]
8318 : : {
8319 : 20092 : PathTarget *target = lfirst_node(PathTarget, lc);
8320 : :
8321 : 20092 : target = copy_pathtarget(target);
8322 : 20092 : target->exprs = (List *)
8323 : 20092 : adjust_appendrel_attrs(root,
8324 : 20092 : (Node *) target->exprs,
8325 : : nappinfos, appinfos);
8326 : 20092 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8327 : : target);
8328 : : }
8329 : 20092 : pfree(appinfos);
8330 : :
8331 : : /* Recursion does the real work. */
8332 : 20092 : apply_scanjoin_target_to_paths(root, child_rel,
8333 : : child_scanjoin_targets,
8334 : : scanjoin_targets_contain_srfs,
8335 : : scanjoin_target_parallel_safe,
8336 : : tlist_same_exprs);
8337 : :
8338 : : /* Save non-dummy children for Append paths. */
8339 [ + - ]: 20092 : if (!IS_DUMMY_REL(child_rel))
8340 : 20092 : live_children = lappend(live_children, child_rel);
8341 : : }
8342 : :
8343 : : /* Build new paths for this relation by appending child paths. */
8344 : 10054 : add_paths_to_append_rel(root, rel, live_children);
8345 : : }
8346 : :
8347 : : /*
8348 : : * Consider generating Gather or Gather Merge paths. We must only do this
8349 : : * if the relation is parallel safe, and we don't do it for child rels to
8350 : : * avoid creating multiple Gather nodes within the same plan. We must do
8351 : : * this after all paths have been generated and before set_cheapest, since
8352 : : * one of the generated paths may turn out to be the cheapest one.
8353 : : */
8354 [ + + + + : 406688 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
+ + + - ]
8355 : 136372 : generate_useful_gather_paths(root, rel, false);
8356 : :
8357 : : /*
8358 : : * Reassess which paths are the cheapest, now that we've potentially added
8359 : : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8360 : : * this relation.
8361 : : */
8362 : 406688 : set_cheapest(rel);
8363 : 406688 : }
8364 : :
8365 : : /*
8366 : : * create_partitionwise_grouping_paths
8367 : : *
8368 : : * If the partition keys of input relation are part of the GROUP BY clause, all
8369 : : * the rows belonging to a given group come from a single partition. This
8370 : : * allows aggregation/grouping over a partitioned relation to be broken down
8371 : : * into aggregation/grouping on each partition. This should be no worse, and
8372 : : * often better, than the normal approach.
8373 : : *
8374 : : * However, if the GROUP BY clause does not contain all the partition keys,
8375 : : * rows from a given group may be spread across multiple partitions. In that
8376 : : * case, we perform partial aggregation for each group, append the results,
8377 : : * and then finalize aggregation. This is less certain to win than the
8378 : : * previous case. It may win if the PartialAggregate stage greatly reduces
8379 : : * the number of groups, because fewer rows will pass through the Append node.
8380 : : * It may lose if we have lots of small groups.
8381 : : */
8382 : : static void
8383 : 685 : create_partitionwise_grouping_paths(PlannerInfo *root,
8384 : : RelOptInfo *input_rel,
8385 : : RelOptInfo *grouped_rel,
8386 : : RelOptInfo *partially_grouped_rel,
8387 : : const AggClauseCosts *agg_costs,
8388 : : grouping_sets_data *gd,
8389 : : PartitionwiseAggregateType patype,
8390 : : GroupPathExtraData *extra)
8391 : : {
8392 : 685 : List *grouped_live_children = NIL;
8393 : 685 : List *partially_grouped_live_children = NIL;
8394 : 685 : PathTarget *target = grouped_rel->reltarget;
8395 : 685 : bool partial_grouping_valid = true;
8396 : : int i;
8397 : :
8398 : : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8399 : : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8400 : : partially_grouped_rel != NULL);
8401 : :
8402 : : /* Add paths for partitionwise aggregation/grouping. */
8403 : 685 : i = -1;
8404 [ + + ]: 2480 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8405 : : {
8406 : 1795 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8407 : : PathTarget *child_target;
8408 : : AppendRelInfo **appinfos;
8409 : : int nappinfos;
8410 : : GroupPathExtraData child_extra;
8411 : : RelOptInfo *child_grouped_rel;
8412 : : RelOptInfo *child_partially_grouped_rel;
8413 : :
8414 : : Assert(child_input_rel != NULL);
8415 : :
8416 : : /* Dummy children can be ignored. */
8417 [ - + ]: 1795 : if (IS_DUMMY_REL(child_input_rel))
8418 : 0 : continue;
8419 : :
8420 : 1795 : child_target = copy_pathtarget(target);
8421 : :
8422 : : /*
8423 : : * Copy the given "extra" structure as is and then override the
8424 : : * members specific to this child.
8425 : : */
8426 : 1795 : memcpy(&child_extra, extra, sizeof(child_extra));
8427 : :
8428 : 1795 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8429 : : &nappinfos);
8430 : :
8431 : 1795 : child_target->exprs = (List *)
8432 : 1795 : adjust_appendrel_attrs(root,
8433 : 1795 : (Node *) target->exprs,
8434 : : nappinfos, appinfos);
8435 : :
8436 : : /* Translate havingQual and targetList. */
8437 : 1795 : child_extra.havingQual = (Node *)
8438 : : adjust_appendrel_attrs(root,
8439 : : extra->havingQual,
8440 : : nappinfos, appinfos);
8441 : 1795 : child_extra.targetList = (List *)
8442 : 1795 : adjust_appendrel_attrs(root,
8443 : 1795 : (Node *) extra->targetList,
8444 : : nappinfos, appinfos);
8445 : :
8446 : : /*
8447 : : * extra->patype was the value computed for our parent rel; patype is
8448 : : * the value for this relation. For the child, our value is its
8449 : : * parent rel's value.
8450 : : */
8451 : 1795 : child_extra.patype = patype;
8452 : :
8453 : : /*
8454 : : * Create grouping relation to hold fully aggregated grouping and/or
8455 : : * aggregation paths for the child.
8456 : : */
8457 : 1795 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8458 : : child_target,
8459 : 1795 : extra->target_parallel_safe,
8460 : : child_extra.havingQual);
8461 : :
8462 : : /* Create grouping paths for this child relation. */
8463 : 1795 : create_ordinary_grouping_paths(root, child_input_rel,
8464 : : child_grouped_rel,
8465 : : agg_costs, gd, &child_extra,
8466 : : &child_partially_grouped_rel);
8467 : :
8468 [ + + ]: 1795 : if (child_partially_grouped_rel)
8469 : : {
8470 : : partially_grouped_live_children =
8471 : 1283 : lappend(partially_grouped_live_children,
8472 : : child_partially_grouped_rel);
8473 : : }
8474 : : else
8475 : 512 : partial_grouping_valid = false;
8476 : :
8477 [ + + ]: 1795 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8478 : : {
8479 : 1082 : set_cheapest(child_grouped_rel);
8480 : 1082 : grouped_live_children = lappend(grouped_live_children,
8481 : : child_grouped_rel);
8482 : : }
8483 : :
8484 : 1795 : pfree(appinfos);
8485 : : }
8486 : :
8487 : : /*
8488 : : * Try to create append paths for partially grouped children. For full
8489 : : * partitionwise aggregation, we might have paths in the partial_pathlist
8490 : : * if parallel aggregation is possible. For partial partitionwise
8491 : : * aggregation, we may have paths in both pathlist and partial_pathlist.
8492 : : *
8493 : : * NB: We must have a partially grouped path for every child in order to
8494 : : * generate a partially grouped path for this relation.
8495 : : */
8496 [ + + + + ]: 685 : if (partially_grouped_rel && partial_grouping_valid)
8497 : : {
8498 : : Assert(partially_grouped_live_children != NIL);
8499 : :
8500 : 501 : add_paths_to_append_rel(root, partially_grouped_rel,
8501 : : partially_grouped_live_children);
8502 : : }
8503 : :
8504 : : /* If possible, create append paths for fully grouped children. */
8505 [ + + ]: 685 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8506 : : {
8507 : : Assert(grouped_live_children != NIL);
8508 : :
8509 : 404 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8510 : : }
8511 : 685 : }
8512 : :
8513 : : /*
8514 : : * group_by_has_partkey
8515 : : *
8516 : : * Returns true if all the partition keys of the given relation are part of
8517 : : * the GROUP BY clauses, including having matching collation, false otherwise.
8518 : : */
8519 : : static bool
8520 : 640 : group_by_has_partkey(RelOptInfo *input_rel,
8521 : : List *targetList,
8522 : : List *groupClause)
8523 : : {
8524 : 640 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8525 : 640 : int cnt = 0;
8526 : : int partnatts;
8527 : :
8528 : : /* Input relation should be partitioned. */
8529 : : Assert(input_rel->part_scheme);
8530 : :
8531 : : /* Rule out early, if there are no partition keys present. */
8532 [ - + ]: 640 : if (!input_rel->partexprs)
8533 : 0 : return false;
8534 : :
8535 : 640 : partnatts = input_rel->part_scheme->partnatts;
8536 : :
8537 [ + + ]: 1074 : for (cnt = 0; cnt < partnatts; cnt++)
8538 : : {
8539 : 670 : List *partexprs = input_rel->partexprs[cnt];
8540 : : ListCell *lc;
8541 : 670 : bool found = false;
8542 : :
8543 [ + + + + : 1001 : foreach(lc, partexprs)
+ + ]
8544 : : {
8545 : : ListCell *lg;
8546 : 775 : Expr *partexpr = lfirst(lc);
8547 : 775 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8548 : :
8549 [ + - + + : 1206 : foreach(lg, groupexprs)
+ + ]
8550 : : {
8551 : 875 : Expr *groupexpr = lfirst(lg);
8552 : 875 : Oid groupcoll = exprCollation((Node *) groupexpr);
8553 : :
8554 : : /*
8555 : : * Note: we can assume there is at most one RelabelType node;
8556 : : * eval_const_expressions() will have simplified if more than
8557 : : * one.
8558 : : */
8559 [ + + ]: 875 : if (IsA(groupexpr, RelabelType))
8560 : 20 : groupexpr = ((RelabelType *) groupexpr)->arg;
8561 : :
8562 [ + + ]: 875 : if (equal(groupexpr, partexpr))
8563 : : {
8564 : : /*
8565 : : * Reject a match if the grouping collation does not match
8566 : : * the partitioning collation.
8567 : : */
8568 [ + + + - : 444 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
+ + ]
8569 : : partcoll != groupcoll)
8570 : 10 : return false;
8571 : :
8572 : 434 : found = true;
8573 : 434 : break;
8574 : : }
8575 : : }
8576 : :
8577 [ + + ]: 765 : if (found)
8578 : 434 : break;
8579 : : }
8580 : :
8581 : : /*
8582 : : * If none of the partition key expressions match with any of the
8583 : : * GROUP BY expression, return false.
8584 : : */
8585 [ + + ]: 660 : if (!found)
8586 : 226 : return false;
8587 : : }
8588 : :
8589 : 404 : return true;
8590 : : }
8591 : :
8592 : : /*
8593 : : * generate_setop_child_grouplist
8594 : : * Build a SortGroupClause list defining the sort/grouping properties
8595 : : * of the child of a set operation.
8596 : : *
8597 : : * This is similar to generate_setop_grouplist() but differs as the setop
8598 : : * child query's targetlist entries may already have a tleSortGroupRef
8599 : : * assigned for other purposes, such as GROUP BYs. Here we keep the
8600 : : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8601 : : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8602 : : * any of the columns in the targetlist don't match to the setop's colTypes
8603 : : * then we return an empty list. This may leave some TLEs with unreferenced
8604 : : * ressortgroupref markings, but that's harmless.
8605 : : */
8606 : : static List *
8607 : 10679 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8608 : : {
8609 : 10679 : List *grouplist = copyObject(op->groupClauses);
8610 : : ListCell *lg;
8611 : : ListCell *lt;
8612 : : ListCell *ct;
8613 : :
8614 : 10679 : lg = list_head(grouplist);
8615 : 10679 : ct = list_head(op->colTypes);
8616 [ + + + + : 41153 : foreach(lt, targetlist)
+ + ]
8617 : : {
8618 : 30719 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8619 : : SortGroupClause *sgc;
8620 : : Oid coltype;
8621 : :
8622 : : /* resjunk columns could have sortgrouprefs. Leave these alone */
8623 [ - + ]: 30719 : if (tle->resjunk)
8624 : 0 : continue;
8625 : :
8626 : : /*
8627 : : * We expect every non-resjunk target to have a SortGroupClause and
8628 : : * colTypes.
8629 : : */
8630 : : Assert(lg != NULL);
8631 : : Assert(ct != NULL);
8632 : 30719 : sgc = (SortGroupClause *) lfirst(lg);
8633 : 30719 : coltype = lfirst_oid(ct);
8634 : :
8635 : : /* reject if target type isn't the same as the setop target type */
8636 [ + + ]: 30719 : if (coltype != exprType((Node *) tle->expr))
8637 : 245 : return NIL;
8638 : :
8639 : 30474 : lg = lnext(grouplist, lg);
8640 : 30474 : ct = lnext(op->colTypes, ct);
8641 : :
8642 : : /* assign a tleSortGroupRef, or reuse the existing one */
8643 : 30474 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8644 : : }
8645 : :
8646 : : Assert(lg == NULL);
8647 : : Assert(ct == NULL);
8648 : :
8649 : 10434 : return grouplist;
8650 : : }
8651 : :
8652 : : /*
8653 : : * create_unique_paths
8654 : : * Build a new RelOptInfo containing Paths that represent elimination of
8655 : : * distinct rows from the input data. Distinct-ness is defined according to
8656 : : * the needs of the semijoin represented by sjinfo. If it is not possible
8657 : : * to identify how to make the data unique, NULL is returned.
8658 : : *
8659 : : * If used at all, this is likely to be called repeatedly on the same rel,
8660 : : * so we cache the result.
8661 : : */
8662 : : RelOptInfo *
8663 : 6944 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8664 : : {
8665 : : RelOptInfo *unique_rel;
8666 : 6944 : List *sortPathkeys = NIL;
8667 : 6944 : List *groupClause = NIL;
8668 : : MemoryContext oldcontext;
8669 : :
8670 : : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8671 : : Assert(sjinfo->jointype == JOIN_SEMI);
8672 : : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8673 : :
8674 : : /* If result already cached, return it */
8675 [ + + ]: 6944 : if (rel->unique_rel)
8676 : 1206 : return rel->unique_rel;
8677 : :
8678 : : /* If it's not possible to unique-ify, return NULL */
8679 [ + + + - ]: 5738 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8680 : 104 : return NULL;
8681 : :
8682 : : /*
8683 : : * Punt if this is a child relation and we failed to build a unique-ified
8684 : : * relation for its parent. This can happen if all the RHS columns were
8685 : : * found to be equated to constants when unique-ifying the parent table,
8686 : : * leaving no columns to unique-ify.
8687 : : */
8688 [ + + + + : 5634 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
- + + + ]
8689 : 10 : return NULL;
8690 : :
8691 : : /*
8692 : : * When called during GEQO join planning, we are in a short-lived memory
8693 : : * context. We must make sure that the unique rel and any subsidiary data
8694 : : * structures created for a baserel survive the GEQO cycle, else the
8695 : : * baserel is trashed for future GEQO cycles. On the other hand, when we
8696 : : * are creating those for a joinrel during GEQO, we don't want them to
8697 : : * clutter the main planning context. Upshot is that the best solution is
8698 : : * to explicitly allocate memory in the same context the given RelOptInfo
8699 : : * is in.
8700 : : */
8701 : 5624 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8702 : :
8703 : 5624 : unique_rel = makeNode(RelOptInfo);
8704 : 5624 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8705 : :
8706 : : /*
8707 : : * clear path info
8708 : : */
8709 : 5624 : unique_rel->pathlist = NIL;
8710 : 5624 : unique_rel->ppilist = NIL;
8711 : 5624 : unique_rel->partial_pathlist = NIL;
8712 : 5624 : unique_rel->cheapest_startup_path = NULL;
8713 : 5624 : unique_rel->cheapest_total_path = NULL;
8714 : 5624 : unique_rel->cheapest_parameterized_paths = NIL;
8715 : :
8716 : : /*
8717 : : * Build the target list for the unique rel. We also build the pathkeys
8718 : : * that represent the ordering requirements for the sort-based
8719 : : * implementation, and the list of SortGroupClause nodes that represent
8720 : : * the columns to be grouped on for the hash-based implementation.
8721 : : *
8722 : : * For a child rel, we can construct these fields from those of its
8723 : : * parent.
8724 : : */
8725 [ + + + + : 5624 : if (IS_OTHER_REL(rel))
- + ]
8726 : 360 : {
8727 : : PathTarget *child_unique_target;
8728 : : PathTarget *parent_unique_target;
8729 : :
8730 : 360 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8731 : :
8732 : 360 : child_unique_target = copy_pathtarget(parent_unique_target);
8733 : :
8734 : : /* Translate the target expressions */
8735 : 360 : child_unique_target->exprs = (List *)
8736 : 360 : adjust_appendrel_attrs_multilevel(root,
8737 : 360 : (Node *) parent_unique_target->exprs,
8738 : : rel,
8739 : 360 : rel->top_parent);
8740 : :
8741 : 360 : unique_rel->reltarget = child_unique_target;
8742 : :
8743 : 360 : sortPathkeys = rel->top_parent->unique_pathkeys;
8744 : 360 : groupClause = rel->top_parent->unique_groupclause;
8745 : : }
8746 : : else
8747 : : {
8748 : : List *newtlist;
8749 : : int nextresno;
8750 : 5264 : List *sortList = NIL;
8751 : : ListCell *lc1;
8752 : : ListCell *lc2;
8753 : :
8754 : : /*
8755 : : * The values we are supposed to unique-ify may be expressions in the
8756 : : * variables of the input rel's targetlist. We have to add any such
8757 : : * expressions to the unique rel's targetlist.
8758 : : *
8759 : : * To complicate matters, some of the values to be unique-ified may be
8760 : : * known redundant by the EquivalenceClass machinery (e.g., because
8761 : : * they have been equated to constants). There is no need to compare
8762 : : * such values during unique-ification, and indeed we had better not
8763 : : * try because the Vars involved may not have propagated as high as
8764 : : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8765 : : * detect such cases, which is a tad inefficient but it doesn't seem
8766 : : * worth building specialized infrastructure for this.
8767 : : */
8768 : 5264 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8769 : 5264 : nextresno = list_length(newtlist) + 1;
8770 : :
8771 [ + - + + : 10729 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
+ - + + +
+ + - +
+ ]
8772 : : {
8773 : 5465 : Expr *uniqexpr = lfirst(lc1);
8774 : 5465 : Oid in_oper = lfirst_oid(lc2);
8775 : : Oid sortop;
8776 : : TargetEntry *tle;
8777 : 5465 : bool made_tle = false;
8778 : :
8779 : 5465 : tle = tlist_member(uniqexpr, newtlist);
8780 [ + + ]: 5465 : if (!tle)
8781 : : {
8782 : 2752 : tle = makeTargetEntry(uniqexpr,
8783 : : nextresno,
8784 : : NULL,
8785 : : false);
8786 : 2752 : newtlist = lappend(newtlist, tle);
8787 : 2752 : nextresno++;
8788 : 2752 : made_tle = true;
8789 : : }
8790 : :
8791 : : /*
8792 : : * Try to build an ORDER BY list to sort the input compatibly. We
8793 : : * do this for each sortable clause even when the clauses are not
8794 : : * all sortable, so that we can detect clauses that are redundant
8795 : : * according to the pathkey machinery.
8796 : : */
8797 : 5465 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8798 [ + - ]: 5465 : if (OidIsValid(sortop))
8799 : : {
8800 : : Oid eqop;
8801 : : SortGroupClause *sortcl;
8802 : :
8803 : : /*
8804 : : * The Unique node will need equality operators. Normally
8805 : : * these are the same as the IN clause operators, but if those
8806 : : * are cross-type operators then the equality operators are
8807 : : * the ones for the IN clause operators' RHS datatype.
8808 : : */
8809 : 5465 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8810 [ - + ]: 5465 : if (!OidIsValid(eqop)) /* shouldn't happen */
8811 [ # # ]: 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8812 : : sortop);
8813 : :
8814 : 5465 : sortcl = makeNode(SortGroupClause);
8815 : 5465 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8816 : 5465 : sortcl->eqop = eqop;
8817 : 5465 : sortcl->sortop = sortop;
8818 : 5465 : sortcl->reverse_sort = false;
8819 : 5465 : sortcl->nulls_first = false;
8820 : 5465 : sortcl->hashable = false; /* no need to make this accurate */
8821 : 5465 : sortList = lappend(sortList, sortcl);
8822 : :
8823 : : /*
8824 : : * At each step, convert the SortGroupClause list to pathkey
8825 : : * form. If the just-added SortGroupClause is redundant, the
8826 : : * result will be shorter than the SortGroupClause list.
8827 : : */
8828 : 5465 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8829 : : newtlist);
8830 [ + + ]: 5465 : if (list_length(sortPathkeys) != list_length(sortList))
8831 : : {
8832 : : /* Drop the redundant SortGroupClause */
8833 : 1710 : sortList = list_delete_last(sortList);
8834 : : Assert(list_length(sortPathkeys) == list_length(sortList));
8835 : : /* Undo tlist addition, if we made one */
8836 [ + + ]: 1710 : if (made_tle)
8837 : : {
8838 : 10 : newtlist = list_delete_last(newtlist);
8839 : 10 : nextresno--;
8840 : : }
8841 : : /* We need not consider this clause for hashing, either */
8842 : 1710 : continue;
8843 : : }
8844 : : }
8845 [ # # ]: 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8846 [ # # ]: 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8847 : : in_oper);
8848 : :
8849 [ + - ]: 3755 : if (sjinfo->semi_can_hash)
8850 : : {
8851 : : /* Create a GROUP BY list for the Agg node to use */
8852 : : Oid eq_oper;
8853 : : SortGroupClause *groupcl;
8854 : :
8855 : : /*
8856 : : * Get the hashable equality operators for the Agg node to
8857 : : * use. Normally these are the same as the IN clause
8858 : : * operators, but if those are cross-type operators then the
8859 : : * equality operators are the ones for the IN clause
8860 : : * operators' RHS datatype.
8861 : : */
8862 [ - + ]: 3755 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
8863 [ # # ]: 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8864 : : in_oper);
8865 : :
8866 : 3755 : groupcl = makeNode(SortGroupClause);
8867 : 3755 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8868 : 3755 : groupcl->eqop = eq_oper;
8869 : 3755 : groupcl->sortop = sortop;
8870 : 3755 : groupcl->reverse_sort = false;
8871 : 3755 : groupcl->nulls_first = false;
8872 : 3755 : groupcl->hashable = true;
8873 : 3755 : groupClause = lappend(groupClause, groupcl);
8874 : : }
8875 : : }
8876 : :
8877 : : /*
8878 : : * Done building the sortPathkeys and groupClause. But the
8879 : : * sortPathkeys are bogus if not all the clauses were sortable.
8880 : : */
8881 [ - + ]: 5264 : if (!sjinfo->semi_can_btree)
8882 : 0 : sortPathkeys = NIL;
8883 : :
8884 : : /*
8885 : : * It can happen that all the RHS columns are equated to constants.
8886 : : * We'd have to do something special to unique-ify in that case, and
8887 : : * it's such an unlikely-in-the-real-world case that it's not worth
8888 : : * the effort. So just punt if we found no columns to unique-ify.
8889 : : */
8890 [ + + + - ]: 5264 : if (sortPathkeys == NIL && groupClause == NIL)
8891 : : {
8892 : 1625 : MemoryContextSwitchTo(oldcontext);
8893 : 1625 : return NULL;
8894 : : }
8895 : :
8896 : : /* Convert the required targetlist back to PathTarget form */
8897 : 3639 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8898 : : }
8899 : :
8900 : : /* build unique paths based on input rel's pathlist */
8901 : 3999 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8902 : : sjinfo, unique_rel);
8903 : :
8904 : : /* build unique paths based on input rel's partial_pathlist */
8905 : 3999 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8906 : : sjinfo, unique_rel);
8907 : :
8908 : : /* Now choose the best path(s) */
8909 : 3999 : set_cheapest(unique_rel);
8910 : :
8911 : : /*
8912 : : * There shouldn't be any partial paths for the unique relation;
8913 : : * otherwise, we won't be able to properly guarantee uniqueness.
8914 : : */
8915 : : Assert(unique_rel->partial_pathlist == NIL);
8916 : :
8917 : : /* Cache the result */
8918 : 3999 : rel->unique_rel = unique_rel;
8919 : 3999 : rel->unique_pathkeys = sortPathkeys;
8920 : 3999 : rel->unique_groupclause = groupClause;
8921 : :
8922 : 3999 : MemoryContextSwitchTo(oldcontext);
8923 : :
8924 : 3999 : return unique_rel;
8925 : : }
8926 : :
8927 : : /*
8928 : : * create_final_unique_paths
8929 : : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8930 : : */
8931 : : static void
8932 : 7131 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8933 : : List *sortPathkeys, List *groupClause,
8934 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8935 : : {
8936 : 7131 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8937 : :
8938 : : /* Estimate number of output rows */
8939 : 7131 : unique_rel->rows = estimate_num_groups(root,
8940 : : sjinfo->semi_rhs_exprs,
8941 : : cheapest_input_path->rows,
8942 : : NULL,
8943 : : NULL);
8944 : :
8945 : : /* Consider sort-based implementations, if possible. */
8946 [ + - ]: 7131 : if (sjinfo->semi_can_btree)
8947 : : {
8948 : : ListCell *lc;
8949 : :
8950 : : /*
8951 : : * Use any available suitably-sorted path as input, and also consider
8952 : : * sorting the cheapest-total path and incremental sort on any paths
8953 : : * with presorted keys.
8954 : : *
8955 : : * To save planning time, we ignore parameterized input paths unless
8956 : : * they are the cheapest-total path.
8957 : : */
8958 [ + - + + : 15540 : foreach(lc, input_rel->pathlist)
+ + ]
8959 : : {
8960 : 8409 : Path *input_path = (Path *) lfirst(lc);
8961 : : Path *path;
8962 : : bool is_sorted;
8963 : : int presorted_keys;
8964 : :
8965 : : /*
8966 : : * Ignore parameterized paths that are not the cheapest-total
8967 : : * path.
8968 : : */
8969 [ + + + + ]: 8409 : if (input_path->param_info &&
8970 : : input_path != cheapest_input_path)
8971 : 740 : continue;
8972 : :
8973 : 7705 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8974 : : input_path->pathkeys,
8975 : : &presorted_keys);
8976 : :
8977 : : /*
8978 : : * Ignore paths that are not suitably or partially sorted, unless
8979 : : * they are the cheapest total path (no need to deal with paths
8980 : : * which have presorted keys when incremental sort is disabled).
8981 : : */
8982 [ + + + + ]: 7705 : if (!is_sorted && input_path != cheapest_input_path &&
8983 [ + + - + ]: 76 : (presorted_keys == 0 || !enable_incremental_sort))
8984 : 36 : continue;
8985 : :
8986 : : /*
8987 : : * Make a separate ProjectionPath in case we need a Result node.
8988 : : */
8989 : 7669 : path = (Path *) create_projection_path(root,
8990 : : unique_rel,
8991 : : input_path,
8992 : 7669 : unique_rel->reltarget);
8993 : :
8994 [ + + ]: 7669 : if (!is_sorted)
8995 : : {
8996 : : /*
8997 : : * We've no need to consider both a sort and incremental sort.
8998 : : * We'll just do a sort if there are no presorted keys and an
8999 : : * incremental sort when there are presorted keys.
9000 : : */
9001 [ + + - + ]: 3982 : if (presorted_keys == 0 || !enable_incremental_sort)
9002 : 3942 : path = (Path *) create_sort_path(root,
9003 : : unique_rel,
9004 : : path,
9005 : : sortPathkeys,
9006 : : -1.0);
9007 : : else
9008 : 40 : path = (Path *) create_incremental_sort_path(root,
9009 : : unique_rel,
9010 : : path,
9011 : : sortPathkeys,
9012 : : presorted_keys,
9013 : : -1.0);
9014 : : }
9015 : :
9016 : 7669 : path = (Path *) create_unique_path(root, unique_rel, path,
9017 : : list_length(sortPathkeys),
9018 : : unique_rel->rows);
9019 : :
9020 : 7669 : add_path(unique_rel, path);
9021 : : }
9022 : : }
9023 : :
9024 : : /* Consider hash-based implementation, if possible. */
9025 [ + - ]: 7131 : if (sjinfo->semi_can_hash)
9026 : : {
9027 : : Path *path;
9028 : :
9029 : : /*
9030 : : * Make a separate ProjectionPath in case we need a Result node.
9031 : : */
9032 : 7131 : path = (Path *) create_projection_path(root,
9033 : : unique_rel,
9034 : : cheapest_input_path,
9035 : 7131 : unique_rel->reltarget);
9036 : :
9037 : 7131 : path = (Path *) create_agg_path(root,
9038 : : unique_rel,
9039 : : path,
9040 : : cheapest_input_path->pathtarget,
9041 : : AGG_HASHED,
9042 : : AGGSPLIT_SIMPLE,
9043 : : groupClause,
9044 : : NIL,
9045 : : NULL,
9046 : : unique_rel->rows);
9047 : :
9048 : 7131 : add_path(unique_rel, path);
9049 : : }
9050 : 7131 : }
9051 : :
9052 : : /*
9053 : : * create_partial_unique_paths
9054 : : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
9055 : : */
9056 : : static void
9057 : 3999 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
9058 : : List *sortPathkeys, List *groupClause,
9059 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
9060 : : {
9061 : : RelOptInfo *partial_unique_rel;
9062 : : Path *cheapest_partial_path;
9063 : :
9064 : : /* nothing to do when there are no partial paths in the input rel */
9065 [ + + + + ]: 3999 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
9066 : 867 : return;
9067 : :
9068 : : /*
9069 : : * nothing to do if there's anything in the targetlist that's
9070 : : * parallel-restricted.
9071 : : */
9072 [ - + ]: 3132 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
9073 : 0 : return;
9074 : :
9075 : 3132 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
9076 : :
9077 : 3132 : partial_unique_rel = makeNode(RelOptInfo);
9078 : 3132 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
9079 : :
9080 : : /*
9081 : : * clear path info
9082 : : */
9083 : 3132 : partial_unique_rel->pathlist = NIL;
9084 : 3132 : partial_unique_rel->ppilist = NIL;
9085 : 3132 : partial_unique_rel->partial_pathlist = NIL;
9086 : 3132 : partial_unique_rel->cheapest_startup_path = NULL;
9087 : 3132 : partial_unique_rel->cheapest_total_path = NULL;
9088 : 3132 : partial_unique_rel->cheapest_parameterized_paths = NIL;
9089 : :
9090 : : /* Estimate number of output rows */
9091 : 3132 : partial_unique_rel->rows = estimate_num_groups(root,
9092 : : sjinfo->semi_rhs_exprs,
9093 : : cheapest_partial_path->rows,
9094 : : NULL,
9095 : : NULL);
9096 : 3132 : partial_unique_rel->reltarget = unique_rel->reltarget;
9097 : :
9098 : : /* Consider sort-based implementations, if possible. */
9099 [ + - ]: 3132 : if (sjinfo->semi_can_btree)
9100 : : {
9101 : : ListCell *lc;
9102 : :
9103 : : /*
9104 : : * Use any available suitably-sorted path as input, and also consider
9105 : : * sorting the cheapest partial path and incremental sort on any paths
9106 : : * with presorted keys.
9107 : : */
9108 [ + - + + : 6518 : foreach(lc, input_rel->partial_pathlist)
+ + ]
9109 : : {
9110 : 3386 : Path *input_path = (Path *) lfirst(lc);
9111 : : Path *path;
9112 : : bool is_sorted;
9113 : : int presorted_keys;
9114 : :
9115 : 3386 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
9116 : : input_path->pathkeys,
9117 : : &presorted_keys);
9118 : :
9119 : : /*
9120 : : * Ignore paths that are not suitably or partially sorted, unless
9121 : : * they are the cheapest partial path (no need to deal with paths
9122 : : * which have presorted keys when incremental sort is disabled).
9123 : : */
9124 [ + + - + ]: 3386 : if (!is_sorted && input_path != cheapest_partial_path &&
9125 [ # # # # ]: 0 : (presorted_keys == 0 || !enable_incremental_sort))
9126 : 0 : continue;
9127 : :
9128 : : /*
9129 : : * Make a separate ProjectionPath in case we need a Result node.
9130 : : */
9131 : 3386 : path = (Path *) create_projection_path(root,
9132 : : partial_unique_rel,
9133 : : input_path,
9134 : 3386 : partial_unique_rel->reltarget);
9135 : :
9136 [ + + ]: 3386 : if (!is_sorted)
9137 : : {
9138 : : /*
9139 : : * We've no need to consider both a sort and incremental sort.
9140 : : * We'll just do a sort if there are no presorted keys and an
9141 : : * incremental sort when there are presorted keys.
9142 : : */
9143 [ - + - - ]: 3092 : if (presorted_keys == 0 || !enable_incremental_sort)
9144 : 3092 : path = (Path *) create_sort_path(root,
9145 : : partial_unique_rel,
9146 : : path,
9147 : : sortPathkeys,
9148 : : -1.0);
9149 : : else
9150 : 0 : path = (Path *) create_incremental_sort_path(root,
9151 : : partial_unique_rel,
9152 : : path,
9153 : : sortPathkeys,
9154 : : presorted_keys,
9155 : : -1.0);
9156 : : }
9157 : :
9158 : 3386 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
9159 : : list_length(sortPathkeys),
9160 : : partial_unique_rel->rows);
9161 : :
9162 : 3386 : add_partial_path(partial_unique_rel, path);
9163 : : }
9164 : : }
9165 : :
9166 : : /* Consider hash-based implementation, if possible. */
9167 [ + - ]: 3132 : if (sjinfo->semi_can_hash)
9168 : : {
9169 : : Path *path;
9170 : :
9171 : : /*
9172 : : * Make a separate ProjectionPath in case we need a Result node.
9173 : : */
9174 : 3132 : path = (Path *) create_projection_path(root,
9175 : : partial_unique_rel,
9176 : : cheapest_partial_path,
9177 : 3132 : partial_unique_rel->reltarget);
9178 : :
9179 : 3132 : path = (Path *) create_agg_path(root,
9180 : : partial_unique_rel,
9181 : : path,
9182 : : cheapest_partial_path->pathtarget,
9183 : : AGG_HASHED,
9184 : : AGGSPLIT_SIMPLE,
9185 : : groupClause,
9186 : : NIL,
9187 : : NULL,
9188 : : partial_unique_rel->rows);
9189 : :
9190 : 3132 : add_partial_path(partial_unique_rel, path);
9191 : : }
9192 : :
9193 [ + - ]: 3132 : if (partial_unique_rel->partial_pathlist != NIL)
9194 : : {
9195 : 3132 : generate_useful_gather_paths(root, partial_unique_rel, true);
9196 : 3132 : set_cheapest(partial_unique_rel);
9197 : :
9198 : : /*
9199 : : * Finally, create paths to unique-ify the final result. This step is
9200 : : * needed to remove any duplicates due to combining rows from parallel
9201 : : * workers.
9202 : : */
9203 : 3132 : create_final_unique_paths(root, partial_unique_rel,
9204 : : sortPathkeys, groupClause,
9205 : : sjinfo, unique_rel);
9206 : : }
9207 : : }
9208 : :
9209 : : /*
9210 : : * Choose a unique name for some subroot.
9211 : : *
9212 : : * Modifies glob->subplanNames to track names already used.
9213 : : */
9214 : : char *
9215 : 63058 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
9216 : : {
9217 : : unsigned n;
9218 : :
9219 : : /*
9220 : : * If a numeric suffix is not required, then search the list of
9221 : : * previously-assigned names for a match. If none is found, then we can
9222 : : * use the provided name without modification.
9223 : : */
9224 [ + + ]: 63058 : if (!always_number)
9225 : : {
9226 : 18802 : bool found = false;
9227 : :
9228 [ + + + + : 45686 : foreach_ptr(char, subplan_name, glob->subplanNames)
+ + ]
9229 : : {
9230 [ + + ]: 11624 : if (strcmp(subplan_name, name) == 0)
9231 : : {
9232 : 3542 : found = true;
9233 : 3542 : break;
9234 : : }
9235 : : }
9236 : :
9237 [ + + ]: 18802 : if (!found)
9238 : : {
9239 : : /* pstrdup here is just to avoid cast-away-const */
9240 : 15260 : char *chosen_name = pstrdup(name);
9241 : :
9242 : 15260 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
9243 : 15260 : return chosen_name;
9244 : : }
9245 : : }
9246 : :
9247 : : /*
9248 : : * If a numeric suffix is required or if the un-suffixed name is already
9249 : : * in use, then loop until we find a positive integer that produces a
9250 : : * novel name.
9251 : : */
9252 : 47798 : for (n = 1; true; ++n)
9253 : 37850 : {
9254 : 85648 : char *proposed_name = psprintf("%s_%u", name, n);
9255 : 85648 : bool found = false;
9256 : :
9257 [ + + + + : 306607 : foreach_ptr(char, subplan_name, glob->subplanNames)
+ + ]
9258 : : {
9259 [ + + ]: 173161 : if (strcmp(subplan_name, proposed_name) == 0)
9260 : : {
9261 : 37850 : found = true;
9262 : 37850 : break;
9263 : : }
9264 : : }
9265 : :
9266 [ + + ]: 85648 : if (!found)
9267 : : {
9268 : 47798 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9269 : 47798 : return proposed_name;
9270 : : }
9271 : :
9272 : 37850 : pfree(proposed_name);
9273 : : }
9274 : : }
|
