Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * planner.c
4 : : * The query optimizer external interface.
5 : : *
6 : : * Portions Copyright (c) 1996-2025, 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 : : /* Local functions */
138 : : static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
139 : : static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
140 : : static void grouping_planner(PlannerInfo *root, double tuple_fraction,
141 : : SetOperationStmt *setops);
142 : : static grouping_sets_data *preprocess_grouping_sets(PlannerInfo *root);
143 : : static List *remap_to_groupclause_idx(List *groupClause, List *gsets,
144 : : int *tleref_to_colnum_map);
145 : : static void preprocess_rowmarks(PlannerInfo *root);
146 : : static double preprocess_limit(PlannerInfo *root,
147 : : double tuple_fraction,
148 : : int64 *offset_est, int64 *count_est);
149 : : static List *preprocess_groupclause(PlannerInfo *root, List *force);
150 : : static List *extract_rollup_sets(List *groupingSets);
151 : : static List *reorder_grouping_sets(List *groupingSets, List *sortclause);
152 : : static void standard_qp_callback(PlannerInfo *root, void *extra);
153 : : static double get_number_of_groups(PlannerInfo *root,
154 : : double path_rows,
155 : : grouping_sets_data *gd,
156 : : List *target_list);
157 : : static RelOptInfo *create_grouping_paths(PlannerInfo *root,
158 : : RelOptInfo *input_rel,
159 : : PathTarget *target,
160 : : bool target_parallel_safe,
161 : : grouping_sets_data *gd);
162 : : static bool is_degenerate_grouping(PlannerInfo *root);
163 : : static void create_degenerate_grouping_paths(PlannerInfo *root,
164 : : RelOptInfo *input_rel,
165 : : RelOptInfo *grouped_rel);
166 : : static RelOptInfo *make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
167 : : PathTarget *target, bool target_parallel_safe,
168 : : Node *havingQual);
169 : : static void create_ordinary_grouping_paths(PlannerInfo *root,
170 : : RelOptInfo *input_rel,
171 : : RelOptInfo *grouped_rel,
172 : : const AggClauseCosts *agg_costs,
173 : : grouping_sets_data *gd,
174 : : GroupPathExtraData *extra,
175 : : RelOptInfo **partially_grouped_rel_p);
176 : : static void consider_groupingsets_paths(PlannerInfo *root,
177 : : RelOptInfo *grouped_rel,
178 : : Path *path,
179 : : bool is_sorted,
180 : : bool can_hash,
181 : : grouping_sets_data *gd,
182 : : const AggClauseCosts *agg_costs,
183 : : double dNumGroups);
184 : : static RelOptInfo *create_window_paths(PlannerInfo *root,
185 : : RelOptInfo *input_rel,
186 : : PathTarget *input_target,
187 : : PathTarget *output_target,
188 : : bool output_target_parallel_safe,
189 : : WindowFuncLists *wflists,
190 : : List *activeWindows);
191 : : static void create_one_window_path(PlannerInfo *root,
192 : : RelOptInfo *window_rel,
193 : : Path *path,
194 : : PathTarget *input_target,
195 : : PathTarget *output_target,
196 : : WindowFuncLists *wflists,
197 : : List *activeWindows);
198 : : static RelOptInfo *create_distinct_paths(PlannerInfo *root,
199 : : RelOptInfo *input_rel,
200 : : PathTarget *target);
201 : : static void create_partial_distinct_paths(PlannerInfo *root,
202 : : RelOptInfo *input_rel,
203 : : RelOptInfo *final_distinct_rel,
204 : : PathTarget *target);
205 : : static RelOptInfo *create_final_distinct_paths(PlannerInfo *root,
206 : : RelOptInfo *input_rel,
207 : : RelOptInfo *distinct_rel);
208 : : static List *get_useful_pathkeys_for_distinct(PlannerInfo *root,
209 : : List *needed_pathkeys,
210 : : List *path_pathkeys);
211 : : static RelOptInfo *create_ordered_paths(PlannerInfo *root,
212 : : RelOptInfo *input_rel,
213 : : PathTarget *target,
214 : : bool target_parallel_safe,
215 : : double limit_tuples);
216 : : static PathTarget *make_group_input_target(PlannerInfo *root,
217 : : PathTarget *final_target);
218 : : static PathTarget *make_partial_grouping_target(PlannerInfo *root,
219 : : PathTarget *grouping_target,
220 : : Node *havingQual);
221 : : static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
222 : : static void optimize_window_clauses(PlannerInfo *root,
223 : : WindowFuncLists *wflists);
224 : : static List *select_active_windows(PlannerInfo *root, WindowFuncLists *wflists);
225 : : static void name_active_windows(List *activeWindows);
226 : : static PathTarget *make_window_input_target(PlannerInfo *root,
227 : : PathTarget *final_target,
228 : : List *activeWindows);
229 : : static List *make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
230 : : List *tlist);
231 : : static PathTarget *make_sort_input_target(PlannerInfo *root,
232 : : PathTarget *final_target,
233 : : bool *have_postponed_srfs);
234 : : static void adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
235 : : List *targets, List *targets_contain_srfs);
236 : : static void add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
237 : : RelOptInfo *grouped_rel,
238 : : RelOptInfo *partially_grouped_rel,
239 : : const AggClauseCosts *agg_costs,
240 : : grouping_sets_data *gd,
241 : : GroupPathExtraData *extra);
242 : : static RelOptInfo *create_partial_grouping_paths(PlannerInfo *root,
243 : : RelOptInfo *grouped_rel,
244 : : RelOptInfo *input_rel,
245 : : grouping_sets_data *gd,
246 : : GroupPathExtraData *extra,
247 : : bool force_rel_creation);
248 : : static Path *make_ordered_path(PlannerInfo *root,
249 : : RelOptInfo *rel,
250 : : Path *path,
251 : : Path *cheapest_path,
252 : : List *pathkeys,
253 : : double limit_tuples);
254 : : static void gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel);
255 : : static bool can_partial_agg(PlannerInfo *root);
256 : : static void apply_scanjoin_target_to_paths(PlannerInfo *root,
257 : : RelOptInfo *rel,
258 : : List *scanjoin_targets,
259 : : List *scanjoin_targets_contain_srfs,
260 : : bool scanjoin_target_parallel_safe,
261 : : bool tlist_same_exprs);
262 : : static void create_partitionwise_grouping_paths(PlannerInfo *root,
263 : : RelOptInfo *input_rel,
264 : : RelOptInfo *grouped_rel,
265 : : RelOptInfo *partially_grouped_rel,
266 : : const AggClauseCosts *agg_costs,
267 : : grouping_sets_data *gd,
268 : : PartitionwiseAggregateType patype,
269 : : GroupPathExtraData *extra);
270 : : static bool group_by_has_partkey(RelOptInfo *input_rel,
271 : : List *targetList,
272 : : List *groupClause);
273 : : static int common_prefix_cmp(const void *a, const void *b);
274 : : static List *generate_setop_child_grouplist(SetOperationStmt *op,
275 : : List *targetlist);
276 : : static void create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
277 : : List *sortPathkeys, List *groupClause,
278 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
279 : : static void create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
280 : : List *sortPathkeys, List *groupClause,
281 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel);
282 : :
283 : :
284 : : /*****************************************************************************
285 : : *
286 : : * Query optimizer entry point
287 : : *
288 : : * Inputs:
289 : : * parse: an analyzed-and-rewritten query tree for an optimizable statement
290 : : * query_string: source text for the query tree (used for error reports)
291 : : * cursorOptions: bitmask of CURSOR_OPT_XXX flags, see parsenodes.h
292 : : * boundParams: passed-in parameter values, or NULL if none
293 : : * es: ExplainState if being called from EXPLAIN, else NULL
294 : : *
295 : : * The result is a PlannedStmt tree.
296 : : *
297 : : * PARAM_EXTERN Param nodes within the parse tree can be replaced by Consts
298 : : * using values from boundParams, if those values are marked PARAM_FLAG_CONST.
299 : : * Parameter values not so marked are still relied on for estimation purposes.
300 : : *
301 : : * The ExplainState pointer is not currently used by the core planner, but it
302 : : * is passed through to some planner hooks so that they can report information
303 : : * back to EXPLAIN extension hooks.
304 : : *
305 : : * To support loadable plugins that monitor or modify planner behavior,
306 : : * we provide a hook variable that lets a plugin get control before and
307 : : * after the standard planning process. The plugin would normally call
308 : : * standard_planner().
309 : : *
310 : : * Note to plugin authors: standard_planner() scribbles on its Query input,
311 : : * so you'd better copy that data structure if you want to plan more than once.
312 : : *
313 : : *****************************************************************************/
314 : : PlannedStmt *
2037 fujii@postgresql.org 315 :CBC 223125 : planner(Query *parse, const char *query_string, int cursorOptions,
316 : : ParamListInfo boundParams, ExplainState *es)
317 : : {
318 : : PlannedStmt *result;
319 : :
6730 tgl@sss.pgh.pa.us 320 [ + + ]: 223125 : if (planner_hook)
19 rhaas@postgresql.org 321 :GNC 48823 : result = (*planner_hook) (parse, query_string, cursorOptions,
322 : : boundParams, es);
323 : : else
324 : 174302 : result = standard_planner(parse, query_string, cursorOptions,
325 : : boundParams, es);
326 : :
217 michael@paquier.xyz 327 :CBC 220718 : pgstat_report_plan_id(result->planId, false);
328 : :
6730 tgl@sss.pgh.pa.us 329 : 220718 : return result;
330 : : }
331 : :
332 : : PlannedStmt *
2037 fujii@postgresql.org 333 : 223125 : standard_planner(Query *parse, const char *query_string, int cursorOptions,
334 : : ParamListInfo boundParams, ExplainState *es)
335 : : {
336 : : PlannedStmt *result;
337 : : PlannerGlobal *glob;
338 : : double tuple_fraction;
339 : : PlannerInfo *root;
340 : : RelOptInfo *final_rel;
341 : : Path *best_path;
342 : : Plan *top_plan;
343 : : ListCell *lp,
344 : : *lr;
345 : :
346 : : /*
347 : : * Set up global state for this planner invocation. This data is needed
348 : : * across all levels of sub-Query that might exist in the given command,
349 : : * so we keep it in a separate struct that's linked to by each per-Query
350 : : * PlannerInfo.
351 : : */
6825 tgl@sss.pgh.pa.us 352 : 223125 : glob = makeNode(PlannerGlobal);
353 : :
354 : 223125 : glob->boundParams = boundParams;
6822 355 : 223125 : glob->subplans = NIL;
580 356 : 223125 : glob->subpaths = NIL;
5168 357 : 223125 : glob->subroots = NIL;
6817 358 : 223125 : glob->rewindPlanIDs = NULL;
6822 359 : 223125 : glob->finalrtable = NIL;
166 rguo@postgresql.org 360 : 223125 : glob->allRelids = NULL;
361 : 223125 : glob->prunableRelids = NULL;
1056 alvherre@alvh.no-ip. 362 : 223125 : glob->finalrteperminfos = NIL;
5859 tgl@sss.pgh.pa.us 363 : 223125 : glob->finalrowmarks = NIL;
5358 364 : 223125 : glob->resultRelations = NIL;
2147 365 : 223125 : glob->appendRelations = NIL;
166 rguo@postgresql.org 366 : 223125 : glob->partPruneInfos = NIL;
6591 tgl@sss.pgh.pa.us 367 : 223125 : glob->relationOids = NIL;
6257 368 : 223125 : glob->invalItems = NIL;
2905 rhaas@postgresql.org 369 : 223125 : glob->paramExecTypes = NIL;
6215 tgl@sss.pgh.pa.us 370 : 223125 : glob->lastPHId = 0;
5374 371 : 223125 : glob->lastRowMarkId = 0;
3682 rhaas@postgresql.org 372 : 223125 : glob->lastPlanNodeId = 0;
6612 tgl@sss.pgh.pa.us 373 : 223125 : glob->transientPlan = false;
3391 374 : 223125 : glob->dependsOnRole = false;
166 rguo@postgresql.org 375 : 223125 : glob->partition_directory = NULL;
97 rguo@postgresql.org 376 :GNC 223125 : glob->rel_notnullatts_hash = NULL;
377 : :
378 : : /*
379 : : * Assess whether it's feasible to use parallel mode for this query. We
380 : : * can't do this in a standalone backend, or if the command will try to
381 : : * modify any data, or if this is a cursor operation, or if GUCs are set
382 : : * to values that don't permit parallelism, or if parallel-unsafe
383 : : * functions are present in the query tree.
384 : : *
385 : : * (Note that we do allow CREATE TABLE AS, SELECT INTO, and CREATE
386 : : * MATERIALIZED VIEW to use parallel plans, but this is safe only because
387 : : * the command is writing into a completely new table which workers won't
388 : : * be able to see. If the workers could see the table, the fact that
389 : : * group locking would cause them to ignore the leader's heavyweight GIN
390 : : * page locks would make this unsafe. We'll have to fix that somehow if
391 : : * we want to allow parallel inserts in general; updates and deletes have
392 : : * additional problems especially around combo CIDs.)
393 : : *
394 : : * For now, we don't try to use parallel mode if we're running inside a
395 : : * parallel worker. We might eventually be able to relax this
396 : : * restriction, but for now it seems best not to have parallel workers
397 : : * trying to create their own parallel workers.
398 : : */
3356 tgl@sss.pgh.pa.us 399 [ + + + + ]:CBC 223125 : if ((cursorOptions & CURSOR_OPT_PARALLEL_OK) != 0 &&
400 : 208929 : IsUnderPostmaster &&
1678 akapila@postgresql.o 401 [ + + ]: 208929 : parse->commandType == CMD_SELECT &&
3356 tgl@sss.pgh.pa.us 402 [ + + ]: 169957 : !parse->hasModifyingCTE &&
403 [ + + ]: 169885 : max_parallel_workers_per_gather > 0 &&
2418 tmunro@postgresql.or 404 [ + + ]: 169575 : !IsParallelWorker())
405 : : {
406 : : /* all the cheap tests pass, so scan the query tree */
1678 akapila@postgresql.o 407 : 169551 : glob->maxParallelHazard = max_parallel_hazard(parse);
3356 tgl@sss.pgh.pa.us 408 : 169551 : glob->parallelModeOK = (glob->maxParallelHazard != PROPARALLEL_UNSAFE);
409 : : }
410 : : else
411 : : {
412 : : /* skip the query tree scan, just assume it's unsafe */
413 : 53574 : glob->maxParallelHazard = PROPARALLEL_UNSAFE;
414 : 53574 : glob->parallelModeOK = false;
415 : : }
416 : :
417 : : /*
418 : : * glob->parallelModeNeeded is normally set to false here and changed to
419 : : * true during plan creation if a Gather or Gather Merge plan is actually
420 : : * created (cf. create_gather_plan, create_gather_merge_plan).
421 : : *
422 : : * However, if debug_parallel_query = on or debug_parallel_query =
423 : : * regress, then we impose parallel mode whenever it's safe to do so, even
424 : : * if the final plan doesn't use parallelism. It's not safe to do so if
425 : : * the query contains anything parallel-unsafe; parallelModeOK will be
426 : : * false in that case. Note that parallelModeOK can't change after this
427 : : * point. Otherwise, everything in the query is either parallel-safe or
428 : : * parallel-restricted, and in either case it should be OK to impose
429 : : * parallel-mode restrictions. If that ends up breaking something, then
430 : : * either some function the user included in the query is incorrectly
431 : : * labeled as parallel-safe or parallel-restricted when in reality it's
432 : : * parallel-unsafe, or else the query planner itself has a bug.
433 : : */
3405 rhaas@postgresql.org 434 [ + + ]: 368412 : glob->parallelModeNeeded = glob->parallelModeOK &&
985 drowley@postgresql.o 435 [ + + ]: 145287 : (debug_parallel_query != DEBUG_PARALLEL_OFF);
436 : :
437 : : /* Determine what fraction of the plan is likely to be scanned */
6769 tgl@sss.pgh.pa.us 438 [ + + ]: 223125 : if (cursorOptions & CURSOR_OPT_FAST_PLAN)
439 : : {
440 : : /*
441 : : * We have no real idea how many tuples the user will ultimately FETCH
442 : : * from a cursor, but it is often the case that he doesn't want 'em
443 : : * all, or would prefer a fast-start plan anyway so that he can
444 : : * process some of the tuples sooner. Use a GUC parameter to decide
445 : : * what fraction to optimize for.
446 : : */
6387 447 : 2312 : tuple_fraction = cursor_tuple_fraction;
448 : :
449 : : /*
450 : : * We document cursor_tuple_fraction as simply being a fraction, which
451 : : * means the edge cases 0 and 1 have to be treated specially here. We
452 : : * convert 1 to 0 ("all the tuples") and 0 to a very small fraction.
453 : : */
454 [ - + ]: 2312 : if (tuple_fraction >= 1.0)
6387 tgl@sss.pgh.pa.us 455 :UBC 0 : tuple_fraction = 0.0;
6387 tgl@sss.pgh.pa.us 456 [ - + ]:CBC 2312 : else if (tuple_fraction <= 0.0)
6387 tgl@sss.pgh.pa.us 457 :UBC 0 : tuple_fraction = 1e-10;
458 : : }
459 : : else
460 : : {
461 : : /* Default assumption is we need all the tuples */
8267 tgl@sss.pgh.pa.us 462 :CBC 220813 : tuple_fraction = 0.0;
463 : : }
464 : :
465 : : /* Allow plugins to take control after we've initialized "glob" */
19 rhaas@postgresql.org 466 [ - + ]:GNC 223125 : if (planner_setup_hook)
19 rhaas@postgresql.org 467 :UNC 0 : (*planner_setup_hook) (glob, parse, query_string, &tuple_fraction, es);
468 : :
469 : : /* primary planning entry point (may recurse for subqueries) */
20 rhaas@postgresql.org 470 :GNC 223125 : root = subquery_planner(glob, parse, NULL, NULL, false, tuple_fraction,
471 : : NULL);
472 : :
473 : : /* Select best Path and turn it into a Plan */
3521 tgl@sss.pgh.pa.us 474 :CBC 220916 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
475 : 220916 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
476 : :
477 : 220916 : top_plan = create_plan(root, best_path);
478 : :
479 : : /*
480 : : * If creating a plan for a scrollable cursor, make sure it can run
481 : : * backwards on demand. Add a Material node at the top at need.
482 : : */
6769 483 [ + + ]: 220718 : if (cursorOptions & CURSOR_OPT_SCROLL)
484 : : {
6824 485 [ + + ]: 133 : if (!ExecSupportsBackwardScan(top_plan))
3189 486 : 16 : top_plan = materialize_finished_plan(top_plan);
487 : : }
488 : :
489 : : /*
490 : : * Optionally add a Gather node for testing purposes, provided this is
491 : : * actually a safe thing to do.
492 : : *
493 : : * We can add Gather even when top_plan has parallel-safe initPlans, but
494 : : * then we have to move the initPlans to the Gather node because of
495 : : * SS_finalize_plan's limitations. That would cause cosmetic breakage of
496 : : * regression tests when debug_parallel_query = regress, because initPlans
497 : : * that would normally appear on the top_plan move to the Gather, causing
498 : : * them to disappear from EXPLAIN output. That doesn't seem worth kluging
499 : : * EXPLAIN to hide, so skip it when debug_parallel_query = regress.
500 : : */
837 501 [ + + ]: 220718 : if (debug_parallel_query != DEBUG_PARALLEL_OFF &&
502 [ + + ]: 97 : top_plan->parallel_safe &&
503 [ - + ]: 64 : (top_plan->initPlan == NIL ||
837 tgl@sss.pgh.pa.us 504 [ # # ]:UBC 0 : debug_parallel_query != DEBUG_PARALLEL_REGRESS))
505 : : {
3550 rhaas@postgresql.org 506 :CBC 64 : Gather *gather = makeNode(Gather);
507 : : Cost initplan_cost;
508 : : bool unsafe_initplans;
509 : :
510 : 64 : gather->plan.targetlist = top_plan->targetlist;
511 : 64 : gather->plan.qual = NIL;
512 : 64 : gather->plan.lefttree = top_plan;
513 : 64 : gather->plan.righttree = NULL;
514 : 64 : gather->num_workers = 1;
515 : 64 : gather->single_copy = true;
985 drowley@postgresql.o 516 : 64 : gather->invisible = (debug_parallel_query == DEBUG_PARALLEL_REGRESS);
517 : :
518 : : /* Transfer any initPlans to the new top node */
837 tgl@sss.pgh.pa.us 519 : 64 : gather->plan.initPlan = top_plan->initPlan;
520 : 64 : top_plan->initPlan = NIL;
521 : :
522 : : /*
523 : : * Since this Gather has no parallel-aware descendants to signal to,
524 : : * we don't need a rescan Param.
525 : : */
2980 526 : 64 : gather->rescan_param = -1;
527 : :
528 : : /*
529 : : * Ideally we'd use cost_gather here, but setting up dummy path data
530 : : * to satisfy it doesn't seem much cleaner than knowing what it does.
531 : : */
3403 532 : 64 : gather->plan.startup_cost = top_plan->startup_cost +
533 : : parallel_setup_cost;
534 : 64 : gather->plan.total_cost = top_plan->total_cost +
535 : 64 : parallel_setup_cost + parallel_tuple_cost * top_plan->plan_rows;
536 : 64 : gather->plan.plan_rows = top_plan->plan_rows;
537 : 64 : gather->plan.plan_width = top_plan->plan_width;
538 : 64 : gather->plan.parallel_aware = false;
3120 539 : 64 : gather->plan.parallel_safe = false;
540 : :
541 : : /*
542 : : * Delete the initplans' cost from top_plan. We needn't add it to the
543 : : * Gather node, since the above coding already included it there.
544 : : */
837 545 : 64 : SS_compute_initplan_cost(gather->plan.initPlan,
546 : : &initplan_cost, &unsafe_initplans);
547 : 64 : top_plan->startup_cost -= initplan_cost;
548 : 64 : top_plan->total_cost -= initplan_cost;
549 : :
550 : : /* use parallel mode for parallel plans. */
3550 rhaas@postgresql.org 551 : 64 : root->glob->parallelModeNeeded = true;
552 : :
553 : 64 : top_plan = &gather->plan;
554 : : }
555 : :
556 : : /*
557 : : * If any Params were generated, run through the plan tree and compute
558 : : * each plan node's extParam/allParam sets. Ideally we'd merge this into
559 : : * set_plan_references' tree traversal, but for now it has to be separate
560 : : * because we need to visit subplans before not after main plan.
561 : : */
2905 562 [ + + ]: 220718 : if (glob->paramExecTypes != NIL)
563 : : {
3730 tgl@sss.pgh.pa.us 564 [ - + ]: 74530 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
565 [ + + + + : 96523 : forboth(lp, glob->subplans, lr, glob->subroots)
+ + + + +
+ + - +
+ ]
566 : : {
567 : 21993 : Plan *subplan = (Plan *) lfirst(lp);
2974 568 : 21993 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
569 : :
3730 570 : 21993 : SS_finalize_plan(subroot, subplan);
571 : : }
572 : 74530 : SS_finalize_plan(root, top_plan);
573 : : }
574 : :
575 : : /* final cleanup of the plan */
6822 576 [ - + ]: 220718 : Assert(glob->finalrtable == NIL);
1056 alvherre@alvh.no-ip. 577 [ - + ]: 220718 : Assert(glob->finalrteperminfos == NIL);
5859 tgl@sss.pgh.pa.us 578 [ - + ]: 220718 : Assert(glob->finalrowmarks == NIL);
5358 579 [ - + ]: 220718 : Assert(glob->resultRelations == NIL);
2147 580 [ - + ]: 220718 : Assert(glob->appendRelations == NIL);
5168 581 : 220718 : top_plan = set_plan_references(root, top_plan);
582 : : /* ... and the subplans (both regular subplans and initplans) */
583 [ - + ]: 220718 : Assert(list_length(glob->subplans) == list_length(glob->subroots));
584 [ + + + + : 242711 : forboth(lp, glob->subplans, lr, glob->subroots)
+ + + + +
+ + - +
+ ]
585 : : {
6556 bruce@momjian.us 586 : 21993 : Plan *subplan = (Plan *) lfirst(lp);
2974 tgl@sss.pgh.pa.us 587 : 21993 : PlannerInfo *subroot = lfirst_node(PlannerInfo, lr);
588 : :
5168 589 : 21993 : lfirst(lp) = set_plan_references(subroot, subplan);
590 : : }
591 : :
592 : : /* build the PlannedStmt result */
6824 593 : 220718 : result = makeNode(PlannedStmt);
594 : :
595 : 220718 : result->commandType = parse->commandType;
4962 596 : 220718 : result->queryId = parse->queryId;
88 michael@paquier.xyz 597 :GNC 220718 : result->planOrigin = PLAN_STMT_STANDARD;
5861 tgl@sss.pgh.pa.us 598 :CBC 220718 : result->hasReturning = (parse->returningList != NIL);
5358 599 : 220718 : result->hasModifyingCTE = parse->hasModifyingCTE;
6824 600 : 220718 : result->canSetTag = parse->canSetTag;
6612 601 : 220718 : result->transientPlan = glob->transientPlan;
3391 602 : 220718 : result->dependsOnRole = glob->dependsOnRole;
603 : 220718 : result->parallelModeNeeded = glob->parallelModeNeeded;
6824 604 : 220718 : result->planTree = top_plan;
270 amitlan@postgresql.o 605 : 220718 : result->partPruneInfos = glob->partPruneInfos;
6822 tgl@sss.pgh.pa.us 606 : 220718 : result->rtable = glob->finalrtable;
262 amitlan@postgresql.o 607 : 441436 : result->unprunableRelids = bms_difference(glob->allRelids,
608 : 220718 : glob->prunableRelids);
1056 alvherre@alvh.no-ip. 609 : 220718 : result->permInfos = glob->finalrteperminfos;
5358 tgl@sss.pgh.pa.us 610 : 220718 : result->resultRelations = glob->resultRelations;
2147 611 : 220718 : result->appendRelations = glob->appendRelations;
6822 612 : 220718 : result->subplans = glob->subplans;
6817 613 : 220718 : result->rewindPlanIDs = glob->rewindPlanIDs;
5859 614 : 220718 : result->rowMarks = glob->finalrowmarks;
6591 615 : 220718 : result->relationOids = glob->relationOids;
6257 616 : 220718 : result->invalItems = glob->invalItems;
2905 rhaas@postgresql.org 617 : 220718 : result->paramExecTypes = glob->paramExecTypes;
618 : : /* utilityStmt should be null, but we might as well copy it */
3208 tgl@sss.pgh.pa.us 619 : 220718 : result->utilityStmt = parse->utilityStmt;
620 : 220718 : result->stmt_location = parse->stmt_location;
621 : 220718 : result->stmt_len = parse->stmt_len;
622 : :
2776 andres@anarazel.de 623 : 220718 : result->jitFlags = PGJIT_NONE;
624 [ + + + - ]: 220718 : if (jit_enabled && jit_above_cost >= 0 &&
625 [ + + ]: 220362 : top_plan->total_cost > jit_above_cost)
626 : : {
627 : 475 : result->jitFlags |= PGJIT_PERFORM;
628 : :
629 : : /*
630 : : * Decide how much effort should be put into generating better code.
631 : : */
632 [ + - ]: 475 : if (jit_optimize_above_cost >= 0 &&
633 [ + + ]: 475 : top_plan->total_cost > jit_optimize_above_cost)
634 : 218 : result->jitFlags |= PGJIT_OPT3;
2770 635 [ + - ]: 475 : if (jit_inline_above_cost >= 0 &&
636 [ + + ]: 475 : top_plan->total_cost > jit_inline_above_cost)
637 : 218 : result->jitFlags |= PGJIT_INLINE;
638 : :
639 : : /*
640 : : * Decide which operations should be JITed.
641 : : */
2778 642 [ + - ]: 475 : if (jit_expressions)
643 : 475 : result->jitFlags |= PGJIT_EXPR;
2772 644 [ + - ]: 475 : if (jit_tuple_deforming)
645 : 475 : result->jitFlags |= PGJIT_DEFORM;
646 : : }
647 : :
648 : : /* Allow plugins to take control before we discard "glob" */
19 rhaas@postgresql.org 649 [ - + ]:GNC 220718 : if (planner_shutdown_hook)
19 rhaas@postgresql.org 650 :UNC 0 : (*planner_shutdown_hook) (glob, parse, query_string, result);
651 : :
2426 rhaas@postgresql.org 652 [ + + ]:CBC 220718 : if (glob->partition_directory != NULL)
653 : 5881 : DestroyPartitionDirectory(glob->partition_directory);
654 : :
6824 tgl@sss.pgh.pa.us 655 : 220718 : return result;
656 : : }
657 : :
658 : :
659 : : /*--------------------
660 : : * subquery_planner
661 : : * Invokes the planner on a subquery. We recurse to here for each
662 : : * sub-SELECT found in the query tree.
663 : : *
664 : : * glob is the global state for the current planner run.
665 : : * parse is the querytree produced by the parser & rewriter.
666 : : * plan_name is the name to assign to this subplan (NULL at the top level).
667 : : * parent_root is the immediate parent Query's info (NULL at the top level).
668 : : * hasRecursion is true if this is a recursive WITH query.
669 : : * tuple_fraction is the fraction of tuples we expect will be retrieved.
670 : : * tuple_fraction is interpreted as explained for grouping_planner, below.
671 : : * setops is used for set operation subqueries to provide the subquery with
672 : : * the context in which it's being used so that Paths correctly sorted for the
673 : : * set operation can be generated. NULL when not planning a set operation
674 : : * child, or when a child of a set op that isn't interested in sorted input.
675 : : *
676 : : * Basically, this routine does the stuff that should only be done once
677 : : * per Query object. It then calls grouping_planner. At one time,
678 : : * grouping_planner could be invoked recursively on the same Query object;
679 : : * that's not currently true, but we keep the separation between the two
680 : : * routines anyway, in case we need it again someday.
681 : : *
682 : : * subquery_planner will be called recursively to handle sub-Query nodes
683 : : * found within the query's expressions and rangetable.
684 : : *
685 : : * Returns the PlannerInfo struct ("root") that contains all data generated
686 : : * while planning the subquery. In particular, the Path(s) attached to
687 : : * the (UPPERREL_FINAL, NULL) upperrel represent our conclusions about the
688 : : * cheapest way(s) to implement the query. The top level will select the
689 : : * best Path and pass it through createplan.c to produce a finished Plan.
690 : : *--------------------
691 : : */
692 : : PlannerInfo *
20 rhaas@postgresql.org 693 :GNC 261760 : subquery_planner(PlannerGlobal *glob, Query *parse, char *plan_name,
694 : : PlannerInfo *parent_root, bool hasRecursion,
695 : : double tuple_fraction, SetOperationStmt *setops)
696 : : {
697 : : PlannerInfo *root;
698 : : List *newWithCheckOptions;
699 : : List *newHaving;
700 : : bool hasOuterJoins;
701 : : bool hasResultRTEs;
702 : : RelOptInfo *final_rel;
703 : : ListCell *l;
704 : :
705 : : /* Create a PlannerInfo data structure for this subquery */
7449 tgl@sss.pgh.pa.us 706 :CBC 261760 : root = makeNode(PlannerInfo);
707 : 261760 : root->parse = parse;
6825 708 : 261760 : root->glob = glob;
6232 709 [ + + ]: 261760 : root->query_level = parent_root ? parent_root->query_level + 1 : 1;
20 rhaas@postgresql.org 710 :GNC 261760 : root->plan_name = plan_name;
6232 tgl@sss.pgh.pa.us 711 :CBC 261760 : root->parent_root = parent_root;
4800 712 : 261760 : root->plan_params = NIL;
3730 713 : 261760 : root->outer_params = NULL;
6855 714 : 261760 : root->planner_cxt = CurrentMemoryContext;
6825 715 : 261760 : root->init_plans = NIL;
6232 716 : 261760 : root->cte_plan_ids = NIL;
4149 717 : 261760 : root->multiexpr_params = NIL;
1001 718 : 261760 : root->join_domains = NIL;
6855 719 : 261760 : root->eq_classes = NIL;
2290 drowley@postgresql.o 720 : 261760 : root->ec_merging_done = false;
1001 tgl@sss.pgh.pa.us 721 : 261760 : root->last_rinfo_serial = 0;
1671 722 : 261760 : root->all_result_relids =
723 [ + + ]: 261760 : parse->resultRelation ? bms_make_singleton(parse->resultRelation) : NULL;
724 : 261760 : root->leaf_result_relids = NULL; /* we'll find out leaf-ness later */
7209 725 : 261760 : root->append_rel_list = NIL;
1671 726 : 261760 : root->row_identity_vars = NIL;
5845 727 : 261760 : root->rowMarks = NIL;
3521 728 : 261760 : memset(root->upper_rels, 0, sizeof(root->upper_rels));
3514 729 : 261760 : memset(root->upper_targets, 0, sizeof(root->upper_targets));
1013 730 : 261760 : root->processed_groupClause = NIL;
731 : 261760 : root->processed_distinctClause = NIL;
3521 732 : 261760 : root->processed_tlist = NIL;
1671 733 : 261760 : root->update_colnos = NIL;
3817 andres@anarazel.de 734 : 261760 : root->grouping_map = NULL;
3521 tgl@sss.pgh.pa.us 735 : 261760 : root->minmax_aggs = NIL;
3204 736 : 261760 : root->qual_security_level = 0;
1856 737 : 261760 : root->hasPseudoConstantQuals = false;
738 : 261760 : root->hasAlternativeSubPlans = false;
1167 739 : 261760 : root->placeholdersFrozen = false;
6232 740 : 261760 : root->hasRecursion = hasRecursion;
68 rhaas@postgresql.org 741 :GNC 261760 : root->assumeReplanning = false;
6232 tgl@sss.pgh.pa.us 742 [ + + ]:CBC 261760 : if (hasRecursion)
2481 743 : 470 : root->wt_param_id = assign_special_exec_param(root);
744 : : else
6232 745 : 261290 : root->wt_param_id = -1;
3521 746 : 261760 : root->non_recursive_path = NULL;
747 : :
748 : : /*
749 : : * Create the top-level join domain. This won't have valid contents until
750 : : * deconstruct_jointree fills it in, but the node needs to exist before
751 : : * that so we can build EquivalenceClasses referencing it.
752 : : */
1001 753 : 261760 : root->join_domains = list_make1(makeNode(JoinDomain));
754 : :
755 : : /*
756 : : * If there is a WITH list, process each WITH query and either convert it
757 : : * to RTE_SUBQUERY RTE(s) or build an initplan SubPlan structure for it.
758 : : */
6232 759 [ + + ]: 261760 : if (parse->cteList)
760 : 1444 : SS_process_ctes(root);
761 : :
762 : : /*
763 : : * If it's a MERGE command, transform the joinlist as appropriate.
764 : : */
1309 alvherre@alvh.no-ip. 765 : 261757 : transform_MERGE_to_join(parse);
766 : :
767 : : /*
768 : : * Scan the rangetable for relation RTEs and retrieve the necessary
769 : : * catalog information for each relation. Using this information, clear
770 : : * the inh flag for any relation that has no children, collect not-null
771 : : * attribute numbers for any relation that has column not-null
772 : : * constraints, and expand virtual generated columns for any relation that
773 : : * contains them. Note that this step does not descend into sublinks and
774 : : * subqueries; if we pull up any sublinks or subqueries below, their
775 : : * relation RTEs are processed just before pulling them up.
776 : : */
97 rguo@postgresql.org 777 :GNC 261757 : parse = root->parse = preprocess_relation_rtes(root);
778 : :
779 : : /*
780 : : * If the FROM clause is empty, replace it with a dummy RTE_RESULT RTE, so
781 : : * that we don't need so many special cases to deal with that situation.
782 : : */
2464 tgl@sss.pgh.pa.us 783 :CBC 261757 : replace_empty_jointree(parse);
784 : :
785 : : /*
786 : : * Look for ANY and EXISTS SubLinks in WHERE and JOIN/ON clauses, and try
787 : : * to transform them into joins. Note that this step does not descend
788 : : * into subqueries; if we pull up any subqueries below, their SubLinks are
789 : : * processed just before pulling them up.
790 : : */
8316 791 [ + + ]: 261757 : if (parse->hasSubLinks)
6280 792 : 18347 : pull_up_sublinks(root);
793 : :
794 : : /*
795 : : * Scan the rangetable for function RTEs, do const-simplification on them,
796 : : * and then inline them if possible (producing subqueries that might get
797 : : * pulled up next). Recursion issues here are handled in the same way as
798 : : * for SubLinks.
799 : : */
2279 800 : 261757 : preprocess_function_rtes(root);
801 : :
802 : : /*
803 : : * Check to see if any subqueries in the jointree can be merged into this
804 : : * query.
805 : : */
3883 806 : 261754 : pull_up_subqueries(root);
807 : :
808 : : /*
809 : : * If this is a simple UNION ALL query, flatten it into an appendrel. We
810 : : * do this now because it requires applying pull_up_subqueries to the leaf
811 : : * queries of the UNION ALL, which weren't touched above because they
812 : : * weren't referenced by the jointree (they will be after we do this).
813 : : */
5467 814 [ + + ]: 261751 : if (parse->setOperations)
815 : 3465 : flatten_simple_union_all(root);
816 : :
817 : : /*
818 : : * Survey the rangetable to see what kinds of entries are present. We can
819 : : * skip some later processing if relevant SQL features are not used; for
820 : : * example if there are no JOIN RTEs we can avoid the expense of doing
821 : : * flatten_join_alias_vars(). This must be done after we have finished
822 : : * adding rangetable entries, of course. (Note: actually, processing of
823 : : * inherited or partitioned rels can cause RTEs for their child tables to
824 : : * get added later; but those must all be RTE_RELATION entries, so they
825 : : * don't invalidate the conclusions drawn here.)
826 : : */
7449 827 : 261751 : root->hasJoinRTEs = false;
4810 828 : 261751 : root->hasLateralRTEs = false;
412 rguo@postgresql.org 829 : 261751 : root->group_rtindex = 0;
6283 tgl@sss.pgh.pa.us 830 : 261751 : hasOuterJoins = false;
2464 831 : 261751 : hasResultRTEs = false;
7824 neilc@samurai.com 832 [ + - + + : 715320 : foreach(l, parse->rtable)
+ + ]
833 : : {
2974 tgl@sss.pgh.pa.us 834 : 453569 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
835 : :
2403 836 [ + + + + : 453569 : switch (rte->rtekind)
+ ]
837 : : {
838 : 46357 : case RTE_JOIN:
839 : 46357 : root->hasJoinRTEs = true;
840 [ + + ]: 46357 : if (IS_OUTER_JOIN(rte->jointype))
841 : 23807 : hasOuterJoins = true;
842 : 46357 : break;
843 : 98876 : case RTE_RESULT:
844 : 98876 : hasResultRTEs = true;
845 : 98876 : break;
412 rguo@postgresql.org 846 : 2391 : case RTE_GROUP:
847 [ - + ]: 2391 : Assert(parse->hasGroupRTE);
848 : 2391 : root->group_rtindex = list_cell_number(parse->rtable, l) + 1;
849 : 2391 : break;
2403 tgl@sss.pgh.pa.us 850 : 305945 : default:
851 : : /* No work here for other RTE types */
852 : 305945 : break;
853 : : }
854 : :
4810 855 [ + + ]: 453569 : if (rte->lateral)
856 : 5482 : root->hasLateralRTEs = true;
857 : :
858 : : /*
859 : : * We can also determine the maximum security level required for any
860 : : * securityQuals now. Addition of inheritance-child RTEs won't affect
861 : : * this, because child tables don't have their own securityQuals; see
862 : : * expand_single_inheritance_child().
863 : : */
2402 864 [ + + ]: 453569 : if (rte->securityQuals)
865 [ - + ]: 1311 : root->qual_security_level = Max(root->qual_security_level,
866 : : list_length(rte->securityQuals));
867 : : }
868 : :
869 : : /*
870 : : * If we have now verified that the query target relation is
871 : : * non-inheriting, mark it as a leaf target.
872 : : */
1671 873 [ + + ]: 261751 : if (parse->resultRelation)
874 : : {
875 : 42417 : RangeTblEntry *rte = rt_fetch(parse->resultRelation, parse->rtable);
876 : :
877 [ + + ]: 42417 : if (!rte->inh)
878 : 40965 : root->leaf_result_relids =
879 : 40965 : bms_make_singleton(parse->resultRelation);
880 : : }
881 : :
882 : : /*
883 : : * This would be a convenient time to check access permissions for all
884 : : * relations mentioned in the query, since it would be better to fail now,
885 : : * before doing any detailed planning. However, for historical reasons,
886 : : * we leave this to be done at executor startup.
887 : : *
888 : : * Note, however, that we do need to check access permissions for any view
889 : : * relations mentioned in the query, in order to prevent information being
890 : : * leaked by selectivity estimation functions, which only check view owner
891 : : * permissions on underlying tables (see all_rows_selectable() and its
892 : : * callers). This is a little ugly, because it means that access
893 : : * permissions for views will be checked twice, which is another reason
894 : : * why it would be better to do all the ACL checks here.
895 : : */
77 dean.a.rasheed@gmail 896 [ + - + + : 714758 : foreach(l, parse->rtable)
+ + ]
897 : : {
898 : 453201 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
899 : :
900 [ + + ]: 453201 : if (rte->perminfoindex != 0 &&
901 [ + + ]: 243905 : rte->relkind == RELKIND_VIEW)
902 : : {
903 : : RTEPermissionInfo *perminfo;
904 : : bool result;
905 : :
906 : 10373 : perminfo = getRTEPermissionInfo(parse->rteperminfos, rte);
907 : 10373 : result = ExecCheckOneRelPerms(perminfo);
908 [ + + ]: 10373 : if (!result)
909 : 194 : aclcheck_error(ACLCHECK_NO_PRIV, OBJECT_VIEW,
910 : 194 : get_rel_name(perminfo->relid));
911 : : }
912 : : }
913 : :
914 : : /*
915 : : * Preprocess RowMark information. We need to do this after subquery
916 : : * pullup, so that all base relations are present.
917 : : */
5845 tgl@sss.pgh.pa.us 918 : 261557 : preprocess_rowmarks(root);
919 : :
920 : : /*
921 : : * Set hasHavingQual to remember if HAVING clause is present. Needed
922 : : * because preprocess_expression will reduce a constant-true condition to
923 : : * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
924 : : */
7449 925 : 261557 : root->hasHavingQual = (parse->havingQual != NULL);
926 : :
927 : : /*
928 : : * Do expression preprocessing on targetlist and quals, as well as other
929 : : * random expressions in the querytree. Note that we do not need to
930 : : * handle sort/group expressions explicitly, because they are actually
931 : : * part of the targetlist.
932 : : */
9351 933 : 259584 : parse->targetList = (List *)
7449 934 : 261557 : preprocess_expression(root, (Node *) parse->targetList,
935 : : EXPRKIND_TARGET);
936 : :
4484 sfrost@snowman.net 937 : 259584 : newWithCheckOptions = NIL;
938 [ + + + + : 260828 : foreach(l, parse->withCheckOptions)
+ + ]
939 : : {
2974 tgl@sss.pgh.pa.us 940 : 1244 : WithCheckOption *wco = lfirst_node(WithCheckOption, l);
941 : :
4484 sfrost@snowman.net 942 : 1244 : wco->qual = preprocess_expression(root, wco->qual,
943 : : EXPRKIND_QUAL);
944 [ + + ]: 1244 : if (wco->qual != NULL)
945 : 1044 : newWithCheckOptions = lappend(newWithCheckOptions, wco);
946 : : }
947 : 259584 : parse->withCheckOptions = newWithCheckOptions;
948 : :
7016 tgl@sss.pgh.pa.us 949 : 259584 : parse->returningList = (List *)
950 : 259584 : preprocess_expression(root, (Node *) parse->returningList,
951 : : EXPRKIND_TARGET);
952 : :
7449 953 : 259584 : preprocess_qual_conditions(root, (Node *) parse->jointree);
954 : :
955 : 259584 : parse->havingQual = preprocess_expression(root, parse->havingQual,
956 : : EXPRKIND_QUAL);
957 : :
5736 958 [ + + + + : 260995 : foreach(l, parse->windowClause)
+ + ]
959 : : {
2974 960 : 1411 : WindowClause *wc = lfirst_node(WindowClause, l);
961 : :
962 : : /* partitionClause/orderClause are sort/group expressions */
5736 963 : 1411 : wc->startOffset = preprocess_expression(root, wc->startOffset,
964 : : EXPRKIND_LIMIT);
965 : 1411 : wc->endOffset = preprocess_expression(root, wc->endOffset,
966 : : EXPRKIND_LIMIT);
967 : : }
968 : :
7449 969 : 259584 : parse->limitOffset = preprocess_expression(root, parse->limitOffset,
970 : : EXPRKIND_LIMIT);
971 : 259584 : parse->limitCount = preprocess_expression(root, parse->limitCount,
972 : : EXPRKIND_LIMIT);
973 : :
3825 andres@anarazel.de 974 [ + + ]: 259584 : if (parse->onConflict)
975 : : {
3456 tgl@sss.pgh.pa.us 976 : 1820 : parse->onConflict->arbiterElems = (List *)
977 : 910 : preprocess_expression(root,
978 : 910 : (Node *) parse->onConflict->arbiterElems,
979 : : EXPRKIND_ARBITER_ELEM);
980 : 1820 : parse->onConflict->arbiterWhere =
981 : 910 : preprocess_expression(root,
982 : 910 : parse->onConflict->arbiterWhere,
983 : : EXPRKIND_QUAL);
3825 andres@anarazel.de 984 : 1820 : parse->onConflict->onConflictSet = (List *)
3456 tgl@sss.pgh.pa.us 985 : 910 : preprocess_expression(root,
986 : 910 : (Node *) parse->onConflict->onConflictSet,
987 : : EXPRKIND_TARGET);
3825 andres@anarazel.de 988 : 910 : parse->onConflict->onConflictWhere =
3456 tgl@sss.pgh.pa.us 989 : 910 : preprocess_expression(root,
990 : 910 : parse->onConflict->onConflictWhere,
991 : : EXPRKIND_QUAL);
992 : : /* exclRelTlist contains only Vars, so no preprocessing needed */
993 : : }
994 : :
1309 alvherre@alvh.no-ip. 995 [ + + + + : 260982 : foreach(l, parse->mergeActionList)
+ + ]
996 : : {
997 : 1398 : MergeAction *action = (MergeAction *) lfirst(l);
998 : :
999 : 1398 : action->targetList = (List *)
1000 : 1398 : preprocess_expression(root,
1001 : 1398 : (Node *) action->targetList,
1002 : : EXPRKIND_TARGET);
1003 : 1398 : action->qual =
1004 : 1398 : preprocess_expression(root,
1005 : : (Node *) action->qual,
1006 : : EXPRKIND_QUAL);
1007 : : }
1008 : :
576 dean.a.rasheed@gmail 1009 : 259584 : parse->mergeJoinCondition =
1010 : 259584 : preprocess_expression(root, parse->mergeJoinCondition, EXPRKIND_QUAL);
1011 : :
7209 tgl@sss.pgh.pa.us 1012 : 259584 : root->append_rel_list = (List *)
1013 : 259584 : preprocess_expression(root, (Node *) root->append_rel_list,
1014 : : EXPRKIND_APPINFO);
1015 : :
1016 : : /* Also need to preprocess expressions within RTEs */
7824 neilc@samurai.com 1017 [ + - + + : 710320 : foreach(l, parse->rtable)
+ + ]
1018 : : {
2974 tgl@sss.pgh.pa.us 1019 : 450736 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1020 : : int kind;
1021 : : ListCell *lcsq;
1022 : :
3818 simon@2ndQuadrant.co 1023 [ + + ]: 450736 : if (rte->rtekind == RTE_RELATION)
1024 : : {
1025 [ + + ]: 235623 : if (rte->tablesample)
3747 tgl@sss.pgh.pa.us 1026 : 114 : rte->tablesample = (TableSampleClause *)
1027 : 114 : preprocess_expression(root,
1028 : 114 : (Node *) rte->tablesample,
1029 : : EXPRKIND_TABLESAMPLE);
1030 : : }
3818 simon@2ndQuadrant.co 1031 [ + + ]: 215113 : else if (rte->rtekind == RTE_SUBQUERY)
1032 : : {
1033 : : /*
1034 : : * We don't want to do all preprocessing yet on the subquery's
1035 : : * expressions, since that will happen when we plan it. But if it
1036 : : * contains any join aliases of our level, those have to get
1037 : : * expanded now, because planning of the subquery won't do it.
1038 : : * That's only possible if the subquery is LATERAL.
1039 : : */
4805 tgl@sss.pgh.pa.us 1040 [ + + + + ]: 37718 : if (rte->lateral && root->hasJoinRTEs)
1041 : 718 : rte->subquery = (Query *)
1001 1042 : 718 : flatten_join_alias_vars(root, root->parse,
2463 1043 : 718 : (Node *) rte->subquery);
1044 : : }
4805 1045 [ + + ]: 177395 : else if (rte->rtekind == RTE_FUNCTION)
1046 : : {
1047 : : /* Preprocess the function expression(s) fully */
1048 [ + + ]: 24600 : kind = rte->lateral ? EXPRKIND_RTFUNC_LATERAL : EXPRKIND_RTFUNC;
3155 alvherre@alvh.no-ip. 1049 : 24600 : rte->functions = (List *)
1050 : 24600 : preprocess_expression(root, (Node *) rte->functions, kind);
1051 : : }
1052 [ + + ]: 152795 : else if (rte->rtekind == RTE_TABLEFUNC)
1053 : : {
1054 : : /* Preprocess the function expression(s) fully */
1055 [ + + ]: 311 : kind = rte->lateral ? EXPRKIND_TABLEFUNC_LATERAL : EXPRKIND_TABLEFUNC;
1056 : 311 : rte->tablefunc = (TableFunc *)
1057 : 311 : preprocess_expression(root, (Node *) rte->tablefunc, kind);
1058 : : }
7026 mail@joeconway.com 1059 [ + + ]: 152484 : else if (rte->rtekind == RTE_VALUES)
1060 : : {
1061 : : /* Preprocess the values lists fully */
4805 tgl@sss.pgh.pa.us 1062 [ + + ]: 4131 : kind = rte->lateral ? EXPRKIND_VALUES_LATERAL : EXPRKIND_VALUES;
7026 mail@joeconway.com 1063 : 4131 : rte->values_lists = (List *)
4805 tgl@sss.pgh.pa.us 1064 : 4131 : preprocess_expression(root, (Node *) rte->values_lists, kind);
1065 : : }
412 rguo@postgresql.org 1066 [ + + ]: 148353 : else if (rte->rtekind == RTE_GROUP)
1067 : : {
1068 : : /* Preprocess the groupexprs list fully */
1069 : 2391 : rte->groupexprs = (List *)
1070 : 2391 : preprocess_expression(root, (Node *) rte->groupexprs,
1071 : : EXPRKIND_GROUPEXPR);
1072 : : }
1073 : :
1074 : : /*
1075 : : * Process each element of the securityQuals list as if it were a
1076 : : * separate qual expression (as indeed it is). We need to do it this
1077 : : * way to get proper canonicalization of AND/OR structure. Note that
1078 : : * this converts each element into an implicit-AND sublist.
1079 : : */
3204 tgl@sss.pgh.pa.us 1080 [ + + + + : 452220 : foreach(lcsq, rte->securityQuals)
+ + ]
1081 : : {
1082 : 1484 : lfirst(lcsq) = preprocess_expression(root,
1083 : 1484 : (Node *) lfirst(lcsq),
1084 : : EXPRKIND_QUAL);
1085 : : }
1086 : : }
1087 : :
1088 : : /*
1089 : : * Now that we are done preprocessing expressions, and in particular done
1090 : : * flattening join alias variables, get rid of the joinaliasvars lists.
1091 : : * They no longer match what expressions in the rest of the tree look
1092 : : * like, because we have not preprocessed expressions in those lists (and
1093 : : * do not want to; for example, expanding a SubLink there would result in
1094 : : * a useless unreferenced subplan). Leaving them in place simply creates
1095 : : * a hazard for later scans of the tree. We could try to prevent that by
1096 : : * using QTW_IGNORE_JOINALIASES in every tree scan done after this point,
1097 : : * but that doesn't sound very reliable.
1098 : : */
2925 1099 [ + + ]: 259584 : if (root->hasJoinRTEs)
1100 : : {
1101 [ + - + + : 160669 : foreach(l, parse->rtable)
+ + ]
1102 : : {
1103 : 132161 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
1104 : :
1105 : 132161 : rte->joinaliasvars = NIL;
1106 : : }
1107 : : }
1108 : :
1109 : : /*
1110 : : * Replace any Vars in the subquery's targetlist and havingQual that
1111 : : * reference GROUP outputs with the underlying grouping expressions.
1112 : : *
1113 : : * Note that we need to perform this replacement after we've preprocessed
1114 : : * the grouping expressions. This is to ensure that there is only one
1115 : : * instance of SubPlan for each SubLink contained within the grouping
1116 : : * expressions.
1117 : : */
412 rguo@postgresql.org 1118 [ + + ]: 259584 : if (parse->hasGroupRTE)
1119 : : {
1120 : 2391 : parse->targetList = (List *)
1121 : 2391 : flatten_group_exprs(root, root->parse, (Node *) parse->targetList);
1122 : 2391 : parse->havingQual =
1123 : 2391 : flatten_group_exprs(root, root->parse, parse->havingQual);
1124 : : }
1125 : :
1126 : : /* Constant-folding might have removed all set-returning functions */
1127 [ + + ]: 259584 : if (parse->hasTargetSRFs)
1128 : 5868 : parse->hasTargetSRFs = expression_returns_set((Node *) parse->targetList);
1129 : :
1130 : : /*
1131 : : * If we have grouping sets, expand the groupingSets tree of this query to
1132 : : * a flat list of grouping sets. We need to do this before optimizing
1133 : : * HAVING, since we can't easily tell if there's an empty grouping set
1134 : : * until we have this representation.
1135 : : */
6 1136 [ + + ]: 259584 : if (parse->groupingSets)
1137 : : {
1138 : 469 : parse->groupingSets =
1139 : 469 : expand_grouping_sets(parse->groupingSets, parse->groupDistinct, -1);
1140 : : }
1141 : :
1142 : : /*
1143 : : * In some cases we may want to transfer a HAVING clause into WHERE. We
1144 : : * cannot do so if the HAVING clause contains aggregates (obviously) or
1145 : : * volatile functions (since a HAVING clause is supposed to be executed
1146 : : * only once per group). We also can't do this if there are any grouping
1147 : : * sets and the clause references any columns that are nullable by the
1148 : : * grouping sets; the nulled values of those columns are not available
1149 : : * before the grouping step. (The test on groupClause might seem wrong,
1150 : : * but it's okay: it's just an optimization to avoid running pull_varnos
1151 : : * when there cannot be any Vars in the HAVING clause.)
1152 : : *
1153 : : * Also, it may be that the clause is so expensive to execute that we're
1154 : : * better off doing it only once per group, despite the loss of
1155 : : * selectivity. This is hard to estimate short of doing the entire
1156 : : * planning process twice, so we use a heuristic: clauses containing
1157 : : * subplans are left in HAVING. Otherwise, we move or copy the HAVING
1158 : : * clause into WHERE, in hopes of eliminating tuples before aggregation
1159 : : * instead of after.
1160 : : *
1161 : : * If the query has no empty grouping set then we can simply move such a
1162 : : * clause into WHERE; any group that fails the clause will not be in the
1163 : : * output because none of its tuples will reach the grouping or
1164 : : * aggregation stage. Otherwise we have to keep the clause in HAVING to
1165 : : * ensure that we don't emit a bogus aggregated row. But then the HAVING
1166 : : * clause must be degenerate (variable-free), so we can copy it into WHERE
1167 : : * so that query_planner() can use it in a gating Result node. (This could
1168 : : * be done better, but it seems not worth optimizing.)
1169 : : *
1170 : : * Note that a HAVING clause may contain expressions that are not fully
1171 : : * preprocessed. This can happen if these expressions are part of
1172 : : * grouping items. In such cases, they are replaced with GROUP Vars in
1173 : : * the parser and then replaced back after we're done with expression
1174 : : * preprocessing on havingQual. This is not an issue if the clause
1175 : : * remains in HAVING, because these expressions will be matched to lower
1176 : : * target items in setrefs.c. However, if the clause is moved or copied
1177 : : * into WHERE, we need to ensure that these expressions are fully
1178 : : * preprocessed.
1179 : : *
1180 : : * Note that both havingQual and parse->jointree->quals are in
1181 : : * implicitly-ANDed-list form at this point, even though they are declared
1182 : : * as Node *.
1183 : : */
9048 tgl@sss.pgh.pa.us 1184 : 259584 : newHaving = NIL;
7824 neilc@samurai.com 1185 [ + + + + : 260215 : foreach(l, (List *) parse->havingQual)
+ + ]
1186 : : {
1187 : 631 : Node *havingclause = (Node *) lfirst(l);
1188 : :
383 rguo@postgresql.org 1189 [ + + + - ]: 829 : if (contain_agg_clause(havingclause) ||
7536 tgl@sss.pgh.pa.us 1190 [ + - ]: 396 : contain_volatile_functions(havingclause) ||
383 rguo@postgresql.org 1191 : 198 : contain_subplans(havingclause) ||
1192 [ + + + + : 258 : (parse->groupClause && parse->groupingSets &&
+ + ]
1193 : 60 : bms_is_member(root->group_rtindex, pull_varnos(root, havingclause))))
1194 : : {
1195 : : /* keep it in HAVING */
9048 tgl@sss.pgh.pa.us 1196 : 469 : newHaving = lappend(newHaving, havingclause);
1197 : : }
6 rguo@postgresql.org 1198 [ + + ]: 162 : else if (parse->groupClause &&
1199 [ + + ]: 144 : (parse->groupingSets == NIL ||
1200 [ + + ]: 24 : (List *) linitial(parse->groupingSets) != NIL))
7536 tgl@sss.pgh.pa.us 1201 : 138 : {
1202 : : /* There is GROUP BY, but no empty grouping set */
1203 : : Node *whereclause;
1204 : :
1205 : : /* Preprocess the HAVING clause fully */
412 rguo@postgresql.org 1206 : 138 : whereclause = preprocess_expression(root, havingclause,
1207 : : EXPRKIND_QUAL);
1208 : : /* ... and move it to WHERE */
9048 tgl@sss.pgh.pa.us 1209 : 138 : parse->jointree->quals = (Node *)
412 rguo@postgresql.org 1210 : 138 : list_concat((List *) parse->jointree->quals,
1211 : : (List *) whereclause);
1212 : : }
1213 : : else
1214 : : {
1215 : : /* There is an empty grouping set (perhaps implicitly) */
1216 : : Node *whereclause;
1217 : :
1218 : : /* Preprocess the HAVING clause fully */
1219 : 24 : whereclause = preprocess_expression(root, copyObject(havingclause),
1220 : : EXPRKIND_QUAL);
1221 : : /* ... and put a copy in WHERE */
7536 tgl@sss.pgh.pa.us 1222 : 48 : parse->jointree->quals = (Node *)
412 rguo@postgresql.org 1223 : 24 : list_concat((List *) parse->jointree->quals,
1224 : : (List *) whereclause);
1225 : : /* ... and also keep it in HAVING */
7536 tgl@sss.pgh.pa.us 1226 : 24 : newHaving = lappend(newHaving, havingclause);
1227 : : }
1228 : : }
9048 1229 : 259584 : parse->havingQual = (Node *) newHaving;
1230 : :
1231 : : /*
1232 : : * If we have any outer joins, try to reduce them to plain inner joins.
1233 : : * This step is most easily done after we've done expression
1234 : : * preprocessing.
1235 : : */
6283 1236 [ + + ]: 259584 : if (hasOuterJoins)
7449 1237 : 16622 : reduce_outer_joins(root);
1238 : :
1239 : : /*
1240 : : * If we have any RTE_RESULT relations, see if they can be deleted from
1241 : : * the jointree. We also rely on this processing to flatten single-child
1242 : : * FromExprs underneath outer joins. This step is most effectively done
1243 : : * after we've done expression preprocessing and outer join reduction.
1244 : : */
1001 1245 [ + + + + ]: 259584 : if (hasResultRTEs || hasOuterJoins)
2464 1246 : 112764 : remove_useless_result_rtes(root);
1247 : :
1248 : : /*
1249 : : * Do the main planning.
1250 : : */
524 rhaas@postgresql.org 1251 : 259584 : grouping_planner(root, tuple_fraction, setops);
1252 : :
1253 : : /*
1254 : : * Capture the set of outer-level param IDs we have access to, for use in
1255 : : * extParam/allParam calculations later.
1256 : : */
3730 tgl@sss.pgh.pa.us 1257 : 259548 : SS_identify_outer_params(root);
1258 : :
1259 : : /*
1260 : : * If any initPlans were created in this query level, adjust the surviving
1261 : : * Paths' costs and parallel-safety flags to account for them. The
1262 : : * initPlans won't actually get attached to the plan tree till
1263 : : * create_plan() runs, but we must include their effects now.
1264 : : */
3521 1265 : 259548 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1266 : 259548 : SS_charge_for_initplans(root, final_rel);
1267 : :
1268 : : /*
1269 : : * Make sure we've identified the cheapest Path for the final rel. (By
1270 : : * doing this here not in grouping_planner, we include initPlan costs in
1271 : : * the decision, though it's unlikely that will change anything.)
1272 : : */
1273 : 259548 : set_cheapest(final_rel);
1274 : :
1275 : 259548 : return root;
1276 : : }
1277 : :
1278 : : /*
1279 : : * preprocess_expression
1280 : : * Do subquery_planner's preprocessing work for an expression,
1281 : : * which can be a targetlist, a WHERE clause (including JOIN/ON
1282 : : * conditions), a HAVING clause, or a few other things.
1283 : : */
1284 : : static Node *
7449 1285 : 2180284 : preprocess_expression(PlannerInfo *root, Node *expr, int kind)
1286 : : {
1287 : : /*
1288 : : * Fall out quickly if expression is empty. This occurs often enough to
1289 : : * be worth checking. Note that null->null is the correct conversion for
1290 : : * implicit-AND result format, too.
1291 : : */
7455 1292 [ + + ]: 2180284 : if (expr == NULL)
1293 : 1718559 : return NULL;
1294 : :
1295 : : /*
1296 : : * If the query has any join RTEs, replace join alias variables with
1297 : : * base-relation variables. We must do this first, since any expressions
1298 : : * we may extract from the joinaliasvars lists have not been preprocessed.
1299 : : * For example, if we did this after sublink processing, sublinks expanded
1300 : : * out from join aliases would not get processed. But we can skip this in
1301 : : * non-lateral RTE functions, VALUES lists, and TABLESAMPLE clauses, since
1302 : : * they can't contain any Vars of the current query level.
1303 : : */
4805 1304 [ + + + + ]: 461725 : if (root->hasJoinRTEs &&
3747 1305 [ + + + - ]: 203937 : !(kind == EXPRKIND_RTFUNC ||
1306 [ + + ]: 101882 : kind == EXPRKIND_VALUES ||
1307 : : kind == EXPRKIND_TABLESAMPLE ||
1308 : : kind == EXPRKIND_TABLEFUNC))
1001 1309 : 101873 : expr = flatten_join_alias_vars(root, root->parse, expr);
1310 : :
1311 : : /*
1312 : : * Simplify constant expressions. For function RTEs, this was already
1313 : : * done by preprocess_function_rtes. (But note we must do it again for
1314 : : * EXPRKIND_RTFUNC_LATERAL, because those might by now contain
1315 : : * un-simplified subexpressions inserted by flattening of subqueries or
1316 : : * join alias variables.)
1317 : : *
1318 : : * Note: an essential effect of this is to convert named-argument function
1319 : : * calls to positional notation and insert the current actual values of
1320 : : * any default arguments for functions. To ensure that happens, we *must*
1321 : : * process all expressions here. Previous PG versions sometimes skipped
1322 : : * const-simplification if it didn't seem worth the trouble, but we can't
1323 : : * do that anymore.
1324 : : *
1325 : : * Note: this also flattens nested AND and OR expressions into N-argument
1326 : : * form. All processing of a qual expression after this point must be
1327 : : * careful to maintain AND/OR flatness --- that is, do not generate a tree
1328 : : * with AND directly under AND, nor OR directly under OR.
1329 : : */
1474 1330 [ + + ]: 461725 : if (kind != EXPRKIND_RTFUNC)
2279 1331 : 441427 : expr = eval_const_expressions(root, expr);
1332 : :
1333 : : /*
1334 : : * If it's a qual or havingQual, canonicalize it.
1335 : : */
8316 1336 [ + + ]: 459752 : if (kind == EXPRKIND_QUAL)
1337 : : {
2787 1338 : 167972 : expr = (Node *) canonicalize_qual((Expr *) expr, false);
1339 : :
1340 : : #ifdef OPTIMIZER_DEBUG
1341 : : printf("After canonicalize_qual()\n");
1342 : : pprint(expr);
1343 : : #endif
1344 : : }
1345 : :
1346 : : /*
1347 : : * Check for ANY ScalarArrayOpExpr with Const arrays and set the
1348 : : * hashfuncid of any that might execute more quickly by using hash lookups
1349 : : * instead of a linear search.
1350 : : */
1663 drowley@postgresql.o 1351 [ + + + + ]: 459752 : if (kind == EXPRKIND_QUAL || kind == EXPRKIND_TARGET)
1352 : : {
1353 : 421192 : convert_saop_to_hashed_saop(expr);
1354 : : }
1355 : :
1356 : : /* Expand SubLinks to SubPlans */
7449 tgl@sss.pgh.pa.us 1357 [ + + ]: 459752 : if (root->parse->hasSubLinks)
6825 1358 : 52605 : expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
1359 : :
1360 : : /*
1361 : : * XXX do not insert anything here unless you have grokked the comments in
1362 : : * SS_replace_correlation_vars ...
1363 : : */
1364 : :
1365 : : /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
1366 [ + + ]: 459752 : if (root->query_level > 1)
1367 : 87582 : expr = SS_replace_correlation_vars(root, expr);
1368 : :
1369 : : /*
1370 : : * If it's a qual or havingQual, convert it to implicit-AND format. (We
1371 : : * don't want to do this before eval_const_expressions, since the latter
1372 : : * would be unable to simplify a top-level AND correctly. Also,
1373 : : * SS_process_sublinks expects explicit-AND format.)
1374 : : */
7959 1375 [ + + ]: 459752 : if (kind == EXPRKIND_QUAL)
1376 : 167972 : expr = (Node *) make_ands_implicit((Expr *) expr);
1377 : :
9159 1378 : 459752 : return expr;
1379 : : }
1380 : :
1381 : : /*
1382 : : * preprocess_qual_conditions
1383 : : * Recursively scan the query's jointree and do subquery_planner's
1384 : : * preprocessing work on each qual condition found therein.
1385 : : */
1386 : : static void
7449 1387 : 649390 : preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
1388 : : {
9159 1389 [ - + ]: 649390 : if (jtnode == NULL)
9159 tgl@sss.pgh.pa.us 1390 :UBC 0 : return;
9159 tgl@sss.pgh.pa.us 1391 [ + + ]:CBC 649390 : if (IsA(jtnode, RangeTblRef))
1392 : : {
1393 : : /* nothing to do here */
1394 : : }
1395 [ + + ]: 317483 : else if (IsA(jtnode, FromExpr))
1396 : : {
1397 : 267289 : FromExpr *f = (FromExpr *) jtnode;
1398 : : ListCell *l;
1399 : :
1400 [ + + + + : 556707 : foreach(l, f->fromlist)
+ + ]
7449 1401 : 289418 : preprocess_qual_conditions(root, lfirst(l));
1402 : :
1403 : 267289 : f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
1404 : : }
9159 1405 [ + - ]: 50194 : else if (IsA(jtnode, JoinExpr))
1406 : : {
1407 : 50194 : JoinExpr *j = (JoinExpr *) jtnode;
1408 : :
7449 1409 : 50194 : preprocess_qual_conditions(root, j->larg);
1410 : 50194 : preprocess_qual_conditions(root, j->rarg);
1411 : :
1412 : 50194 : j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
1413 : : }
1414 : : else
8130 tgl@sss.pgh.pa.us 1415 [ # # ]:UBC 0 : elog(ERROR, "unrecognized node type: %d",
1416 : : (int) nodeTag(jtnode));
1417 : : }
1418 : :
1419 : : /*
1420 : : * preprocess_phv_expression
1421 : : * Do preprocessing on a PlaceHolderVar expression that's been pulled up.
1422 : : *
1423 : : * If a LATERAL subquery references an output of another subquery, and that
1424 : : * output must be wrapped in a PlaceHolderVar because of an intermediate outer
1425 : : * join, then we'll push the PlaceHolderVar expression down into the subquery
1426 : : * and later pull it back up during find_lateral_references, which runs after
1427 : : * subquery_planner has preprocessed all the expressions that were in the
1428 : : * current query level to start with. So we need to preprocess it then.
1429 : : */
1430 : : Expr *
4810 tgl@sss.pgh.pa.us 1431 :CBC 45 : preprocess_phv_expression(PlannerInfo *root, Expr *expr)
1432 : : {
1433 : 45 : return (Expr *) preprocess_expression(root, (Node *) expr, EXPRKIND_PHV);
1434 : : }
1435 : :
1436 : : /*--------------------
1437 : : * grouping_planner
1438 : : * Perform planning steps related to grouping, aggregation, etc.
1439 : : *
1440 : : * This function adds all required top-level processing to the scan/join
1441 : : * Path(s) produced by query_planner.
1442 : : *
1443 : : * tuple_fraction is the fraction of tuples we expect will be retrieved.
1444 : : * tuple_fraction is interpreted as follows:
1445 : : * 0: expect all tuples to be retrieved (normal case)
1446 : : * 0 < tuple_fraction < 1: expect the given fraction of tuples available
1447 : : * from the plan to be retrieved
1448 : : * tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
1449 : : * expected to be retrieved (ie, a LIMIT specification).
1450 : : * setops is used for set operation subqueries to provide the subquery with
1451 : : * the context in which it's being used so that Paths correctly sorted for the
1452 : : * set operation can be generated. NULL when not planning a set operation
1453 : : * child, or when a child of a set op that isn't interested in sorted input.
1454 : : *
1455 : : * Returns nothing; the useful output is in the Paths we attach to the
1456 : : * (UPPERREL_FINAL, NULL) upperrel in *root. In addition,
1457 : : * root->processed_tlist contains the final processed targetlist.
1458 : : *
1459 : : * Note that we have not done set_cheapest() on the final rel; it's convenient
1460 : : * to leave this to the caller.
1461 : : *--------------------
1462 : : */
1463 : : static void
524 rhaas@postgresql.org 1464 : 259584 : grouping_planner(PlannerInfo *root, double tuple_fraction,
1465 : : SetOperationStmt *setops)
1466 : : {
7449 tgl@sss.pgh.pa.us 1467 : 259584 : Query *parse = root->parse;
7033 bruce@momjian.us 1468 : 259584 : int64 offset_est = 0;
1469 : 259584 : int64 count_est = 0;
6751 tgl@sss.pgh.pa.us 1470 : 259584 : double limit_tuples = -1.0;
3517 1471 : 259584 : bool have_postponed_srfs = false;
1472 : : PathTarget *final_target;
1473 : : List *final_targets;
1474 : : List *final_targets_contain_srfs;
1475 : : bool final_target_parallel_safe;
1476 : : RelOptInfo *current_rel;
1477 : : RelOptInfo *final_rel;
1478 : : FinalPathExtraData extra;
1479 : : ListCell *lc;
1480 : :
1481 : : /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
7375 1482 [ + + + + ]: 259584 : if (parse->limitCount || parse->limitOffset)
1483 : : {
1484 : 2499 : tuple_fraction = preprocess_limit(root, tuple_fraction,
1485 : : &offset_est, &count_est);
1486 : :
1487 : : /*
1488 : : * If we have a known LIMIT, and don't have an unknown OFFSET, we can
1489 : : * estimate the effects of using a bounded sort.
1490 : : */
6751 1491 [ + + + + ]: 2499 : if (count_est > 0 && offset_est >= 0)
1492 : 2224 : limit_tuples = (double) count_est + (double) offset_est;
1493 : : }
1494 : :
1495 : : /* Make tuple_fraction accessible to lower-level routines */
3521 1496 : 259584 : root->tuple_fraction = tuple_fraction;
1497 : :
9153 1498 [ + + ]: 259584 : if (parse->setOperations)
1499 : : {
1500 : : /*
1501 : : * Construct Paths for set operations. The results will not need any
1502 : : * work except perhaps a top-level sort and/or LIMIT. Note that any
1503 : : * special work for recursive unions is the responsibility of
1504 : : * plan_set_operations.
1505 : : */
3521 1506 : 3076 : current_rel = plan_set_operations(root);
1507 : :
1508 : : /*
1509 : : * We should not need to call preprocess_targetlist, since we must be
1510 : : * in a SELECT query node. Instead, use the processed_tlist returned
1511 : : * by plan_set_operations (since this tells whether it returned any
1512 : : * resjunk columns!), and transfer any sort key information from the
1513 : : * original tlist.
1514 : : */
9153 1515 [ - + ]: 3073 : Assert(parse->commandType == CMD_SELECT);
1516 : :
1517 : : /* for safety, copy processed_tlist instead of modifying in-place */
2406 1518 : 3073 : root->processed_tlist =
1519 : 3073 : postprocess_setop_tlist(copyObject(root->processed_tlist),
1520 : : parse->targetList);
1521 : :
1522 : : /* Also extract the PathTarget form of the setop result tlist */
3517 1523 : 3073 : final_target = current_rel->cheapest_total_path->pathtarget;
1524 : :
1525 : : /* And check whether it's parallel safe */
1526 : : final_target_parallel_safe =
2790 rhaas@postgresql.org 1527 : 3073 : is_parallel_safe(root, (Node *) final_target->exprs);
1528 : :
1529 : : /* The setop result tlist couldn't contain any SRFs */
3204 andres@anarazel.de 1530 [ - + ]: 3073 : Assert(!parse->hasTargetSRFs);
1531 : 3073 : final_targets = final_targets_contain_srfs = NIL;
1532 : :
1533 : : /*
1534 : : * Can't handle FOR [KEY] UPDATE/SHARE here (parser should have
1535 : : * checked already, but let's make sure).
1536 : : */
9091 tgl@sss.pgh.pa.us 1537 [ - + ]: 3073 : if (parse->rowMarks)
8130 tgl@sss.pgh.pa.us 1538 [ # # ]:UBC 0 : ereport(ERROR,
1539 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1540 : : /*------
1541 : : translator: %s is a SQL row locking clause such as FOR UPDATE */
1542 : : errmsg("%s is not allowed with UNION/INTERSECT/EXCEPT",
1543 : : LCS_asString(linitial_node(RowMarkClause,
1544 : : parse->rowMarks)->strength))));
1545 : :
1546 : : /*
1547 : : * Calculate pathkeys that represent result ordering requirements
1548 : : */
6297 tgl@sss.pgh.pa.us 1549 [ - + ]:CBC 3073 : Assert(parse->distinctClause == NIL);
6292 1550 : 3073 : root->sort_pathkeys = make_pathkeys_for_sortclauses(root,
1551 : : parse->sortClause,
1552 : : root->processed_tlist);
1553 : : }
1554 : : else
1555 : : {
1556 : : /* No set operations, do regular planning */
1557 : : PathTarget *sort_input_target;
1558 : : List *sort_input_targets;
1559 : : List *sort_input_targets_contain_srfs;
1560 : : bool sort_input_target_parallel_safe;
1561 : : PathTarget *grouping_target;
1562 : : List *grouping_targets;
1563 : : List *grouping_targets_contain_srfs;
1564 : : bool grouping_target_parallel_safe;
1565 : : PathTarget *scanjoin_target;
1566 : : List *scanjoin_targets;
1567 : : List *scanjoin_targets_contain_srfs;
1568 : : bool scanjoin_target_parallel_safe;
1569 : : bool scanjoin_target_same_exprs;
1570 : : bool have_grouping;
6147 1571 : 256508 : WindowFuncLists *wflists = NULL;
1572 : 256508 : List *activeWindows = NIL;
3136 rhodiumtoad@postgres 1573 : 256508 : grouping_sets_data *gset_data = NULL;
1574 : : standard_qp_extra qp_extra;
1575 : :
1576 : : /* A recursive query should always have setOperations */
6232 tgl@sss.pgh.pa.us 1577 [ - + ]: 256508 : Assert(!root->hasRecursion);
1578 : :
1579 : : /* Preprocess grouping sets and GROUP BY clause, if any */
3817 andres@anarazel.de 1580 [ + + ]: 256508 : if (parse->groupingSets)
1581 : : {
3136 rhodiumtoad@postgres 1582 : 469 : gset_data = preprocess_grouping_sets(root);
1583 : : }
1013 tgl@sss.pgh.pa.us 1584 [ + + ]: 256039 : else if (parse->groupClause)
1585 : : {
1586 : : /* Preprocess regular GROUP BY clause, if any */
508 akorotkov@postgresql 1587 : 1955 : root->processed_groupClause = preprocess_groupclause(root, NIL);
1588 : : }
1589 : :
1590 : : /*
1591 : : * Preprocess targetlist. Note that much of the remaining planning
1592 : : * work will be done with the PathTarget representation of tlists, but
1593 : : * we must also maintain the full representation of the final tlist so
1594 : : * that we can transfer its decoration (resnames etc) to the topmost
1595 : : * tlist of the finished Plan. This is kept in processed_tlist.
1596 : : */
1671 tgl@sss.pgh.pa.us 1597 : 256505 : preprocess_targetlist(root);
1598 : :
1599 : : /*
1600 : : * Mark all the aggregates with resolved aggtranstypes, and detect
1601 : : * aggregates that are duplicates or can share transition state. We
1602 : : * must do this before slicing and dicing the tlist into various
1603 : : * pathtargets, else some copies of the Aggref nodes might escape
1604 : : * being marked.
1605 : : */
3404 1606 [ + + ]: 256505 : if (parse->hasAggs)
1607 : : {
1798 heikki.linnakangas@i 1608 : 18341 : preprocess_aggrefs(root, (Node *) root->processed_tlist);
1609 : 18341 : preprocess_aggrefs(root, (Node *) parse->havingQual);
1610 : : }
1611 : :
1612 : : /*
1613 : : * Locate any window functions in the tlist. (We don't need to look
1614 : : * anywhere else, since expressions used in ORDER BY will be in there
1615 : : * too.) Note that they could all have been eliminated by constant
1616 : : * folding, in which case we don't need to do any more work.
1617 : : */
6147 tgl@sss.pgh.pa.us 1618 [ + + ]: 256505 : if (parse->hasWindowFuncs)
1619 : : {
2406 1620 : 1285 : wflists = find_window_functions((Node *) root->processed_tlist,
6147 1621 : 1285 : list_length(parse->windowClause));
1622 [ + + ]: 1285 : if (wflists->numWindowFuncs > 0)
1623 : : {
1624 : : /*
1625 : : * See if any modifications can be made to each WindowClause
1626 : : * to allow the executor to execute the WindowFuncs more
1627 : : * quickly.
1628 : : */
1039 drowley@postgresql.o 1629 : 1282 : optimize_window_clauses(root, wflists);
1630 : :
1631 : : /* Extract the list of windows actually in use. */
6147 tgl@sss.pgh.pa.us 1632 : 1282 : activeWindows = select_active_windows(root, wflists);
1633 : :
1634 : : /* Make sure they all have names, for EXPLAIN's use. */
230 1635 : 1282 : name_active_windows(activeWindows);
1636 : : }
1637 : : else
6147 1638 : 3 : parse->hasWindowFuncs = false;
1639 : : }
1640 : :
1641 : : /*
1642 : : * Preprocess MIN/MAX aggregates, if any. Note: be careful about
1643 : : * adding logic between here and the query_planner() call. Anything
1644 : : * that is needed in MIN/MAX-optimizable cases will have to be
1645 : : * duplicated in planagg.c.
1646 : : */
5471 1647 [ + + ]: 256505 : if (parse->hasAggs)
2406 1648 : 18341 : preprocess_minmax_aggregates(root);
1649 : :
1650 : : /*
1651 : : * Figure out whether there's a hard limit on the number of rows that
1652 : : * query_planner's result subplan needs to return. Even if we know a
1653 : : * hard limit overall, it doesn't apply if the query has any
1654 : : * grouping/aggregation operations, or SRFs in the tlist.
1655 : : */
5457 1656 [ + + ]: 256505 : if (parse->groupClause ||
3817 andres@anarazel.de 1657 [ + + ]: 254117 : parse->groupingSets ||
5457 tgl@sss.pgh.pa.us 1658 [ + + ]: 254084 : parse->distinctClause ||
1659 [ + + ]: 252622 : parse->hasAggs ||
1660 [ + + ]: 236387 : parse->hasWindowFuncs ||
3331 1661 [ + + ]: 235177 : parse->hasTargetSRFs ||
5457 1662 [ + + ]: 229545 : root->hasHavingQual)
4466 1663 : 26972 : root->limit_tuples = -1.0;
1664 : : else
1665 : 229533 : root->limit_tuples = limit_tuples;
1666 : :
1667 : : /* Set up data needed by standard_qp_callback */
4564 1668 : 256505 : qp_extra.activeWindows = activeWindows;
1013 1669 : 256505 : qp_extra.gset_data = gset_data;
1670 : :
1671 : : /*
1672 : : * If we're a subquery for a set operation, store the SetOperationStmt
1673 : : * in qp_extra.
1674 : : */
524 rhaas@postgresql.org 1675 : 256505 : qp_extra.setop = setops;
1676 : :
1677 : : /*
1678 : : * Generate the best unsorted and presorted paths for the scan/join
1679 : : * portion of this Query, ie the processing represented by the
1680 : : * FROM/WHERE clauses. (Note there may not be any presorted paths.)
1681 : : * We also generate (in standard_qp_callback) pathkey representations
1682 : : * of the query's sort clause, distinct clause, etc.
1683 : : */
2406 tgl@sss.pgh.pa.us 1684 : 256505 : current_rel = query_planner(root, standard_qp_callback, &qp_extra);
1685 : :
1686 : : /*
1687 : : * Convert the query's result tlist into PathTarget format.
1688 : : *
1689 : : * Note: this cannot be done before query_planner() has performed
1690 : : * appendrel expansion, because that might add resjunk entries to
1691 : : * root->processed_tlist. Waiting till afterwards is also helpful
1692 : : * because the target width estimates can use per-Var width numbers
1693 : : * that were obtained within query_planner().
1694 : : */
1695 : 256478 : final_target = create_pathtarget(root, root->processed_tlist);
1696 : : final_target_parallel_safe =
2790 rhaas@postgresql.org 1697 : 256478 : is_parallel_safe(root, (Node *) final_target->exprs);
1698 : :
1699 : : /*
1700 : : * If ORDER BY was given, consider whether we should use a post-sort
1701 : : * projection, and compute the adjusted target for preceding steps if
1702 : : * so.
1703 : : */
3517 tgl@sss.pgh.pa.us 1704 [ + + ]: 256478 : if (parse->sortClause)
1705 : : {
1706 : 35149 : sort_input_target = make_sort_input_target(root,
1707 : : final_target,
1708 : : &have_postponed_srfs);
1709 : : sort_input_target_parallel_safe =
2790 rhaas@postgresql.org 1710 : 35149 : is_parallel_safe(root, (Node *) sort_input_target->exprs);
1711 : : }
1712 : : else
1713 : : {
3517 tgl@sss.pgh.pa.us 1714 : 221329 : sort_input_target = final_target;
2790 rhaas@postgresql.org 1715 : 221329 : sort_input_target_parallel_safe = final_target_parallel_safe;
1716 : : }
1717 : :
1718 : : /*
1719 : : * If we have window functions to deal with, the output from any
1720 : : * grouping step needs to be what the window functions want;
1721 : : * otherwise, it should be sort_input_target.
1722 : : */
3519 tgl@sss.pgh.pa.us 1723 [ + + ]: 256478 : if (activeWindows)
1724 : : {
1725 : 1282 : grouping_target = make_window_input_target(root,
1726 : : final_target,
1727 : : activeWindows);
1728 : : grouping_target_parallel_safe =
2790 rhaas@postgresql.org 1729 : 1282 : is_parallel_safe(root, (Node *) grouping_target->exprs);
1730 : : }
1731 : : else
1732 : : {
3517 tgl@sss.pgh.pa.us 1733 : 255196 : grouping_target = sort_input_target;
2790 rhaas@postgresql.org 1734 : 255196 : grouping_target_parallel_safe = sort_input_target_parallel_safe;
1735 : : }
1736 : :
1737 : : /*
1738 : : * If we have grouping or aggregation to do, the topmost scan/join
1739 : : * plan node must emit what the grouping step wants; otherwise, it
1740 : : * should emit grouping_target.
1741 : : */
3519 tgl@sss.pgh.pa.us 1742 [ + + ]: 254090 : have_grouping = (parse->groupClause || parse->groupingSets ||
1743 [ + + + + : 510568 : parse->hasAggs || root->hasHavingQual);
+ + ]
1744 [ + + ]: 256478 : if (have_grouping)
1745 : : {
3517 1746 : 18687 : scanjoin_target = make_group_input_target(root, final_target);
1747 : : scanjoin_target_parallel_safe =
2421 efujita@postgresql.o 1748 : 18687 : is_parallel_safe(root, (Node *) scanjoin_target->exprs);
1749 : : }
1750 : : else
1751 : : {
3519 tgl@sss.pgh.pa.us 1752 : 237791 : scanjoin_target = grouping_target;
2790 rhaas@postgresql.org 1753 : 237791 : scanjoin_target_parallel_safe = grouping_target_parallel_safe;
1754 : : }
1755 : :
1756 : : /*
1757 : : * If there are any SRFs in the targetlist, we must separate each of
1758 : : * these PathTargets into SRF-computing and SRF-free targets. Replace
1759 : : * each of the named targets with a SRF-free version, and remember the
1760 : : * list of additional projection steps we need to add afterwards.
1761 : : */
3204 andres@anarazel.de 1762 [ + + ]: 256478 : if (parse->hasTargetSRFs)
1763 : : {
1764 : : /* final_target doesn't recompute any SRFs in sort_input_target */
1765 : 5868 : split_pathtarget_at_srfs(root, final_target, sort_input_target,
1766 : : &final_targets,
1767 : : &final_targets_contain_srfs);
2974 tgl@sss.pgh.pa.us 1768 : 5868 : final_target = linitial_node(PathTarget, final_targets);
3204 andres@anarazel.de 1769 [ - + ]: 5868 : Assert(!linitial_int(final_targets_contain_srfs));
1770 : : /* likewise for sort_input_target vs. grouping_target */
1771 : 5868 : split_pathtarget_at_srfs(root, sort_input_target, grouping_target,
1772 : : &sort_input_targets,
1773 : : &sort_input_targets_contain_srfs);
2974 tgl@sss.pgh.pa.us 1774 : 5868 : sort_input_target = linitial_node(PathTarget, sort_input_targets);
3204 andres@anarazel.de 1775 [ - + ]: 5868 : Assert(!linitial_int(sort_input_targets_contain_srfs));
1776 : : /* likewise for grouping_target vs. scanjoin_target */
1777 : 5868 : split_pathtarget_at_srfs(root, grouping_target, scanjoin_target,
1778 : : &grouping_targets,
1779 : : &grouping_targets_contain_srfs);
2974 tgl@sss.pgh.pa.us 1780 : 5868 : grouping_target = linitial_node(PathTarget, grouping_targets);
3204 andres@anarazel.de 1781 [ - + ]: 5868 : Assert(!linitial_int(grouping_targets_contain_srfs));
1782 : : /* scanjoin_target will not have any SRFs precomputed for it */
1783 : 5868 : split_pathtarget_at_srfs(root, scanjoin_target, NULL,
1784 : : &scanjoin_targets,
1785 : : &scanjoin_targets_contain_srfs);
2974 tgl@sss.pgh.pa.us 1786 : 5868 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
3204 andres@anarazel.de 1787 [ - + ]: 5868 : Assert(!linitial_int(scanjoin_targets_contain_srfs));
1788 : : }
1789 : : else
1790 : : {
1791 : : /* initialize lists; for most of these, dummy values are OK */
1792 : 250610 : final_targets = final_targets_contain_srfs = NIL;
1793 : 250610 : sort_input_targets = sort_input_targets_contain_srfs = NIL;
1794 : 250610 : grouping_targets = grouping_targets_contain_srfs = NIL;
2769 rhaas@postgresql.org 1795 : 250610 : scanjoin_targets = list_make1(scanjoin_target);
1796 : 250610 : scanjoin_targets_contain_srfs = NIL;
1797 : : }
1798 : :
1799 : : /* Apply scan/join target. */
1800 : 256478 : scanjoin_target_same_exprs = list_length(scanjoin_targets) == 1
1801 [ + + + + ]: 256478 : && equal(scanjoin_target->exprs, current_rel->reltarget->exprs);
1802 : 256478 : apply_scanjoin_target_to_paths(root, current_rel, scanjoin_targets,
1803 : : scanjoin_targets_contain_srfs,
1804 : : scanjoin_target_parallel_safe,
1805 : : scanjoin_target_same_exprs);
1806 : :
1807 : : /*
1808 : : * Save the various upper-rel PathTargets we just computed into
1809 : : * root->upper_targets[]. The core code doesn't use this, but it
1810 : : * provides a convenient place for extensions to get at the info. For
1811 : : * consistency, we save all the intermediate targets, even though some
1812 : : * of the corresponding upperrels might not be needed for this query.
1813 : : */
3514 tgl@sss.pgh.pa.us 1814 : 256478 : root->upper_targets[UPPERREL_FINAL] = final_target;
2443 efujita@postgresql.o 1815 : 256478 : root->upper_targets[UPPERREL_ORDERED] = final_target;
1816 : 256478 : root->upper_targets[UPPERREL_DISTINCT] = sort_input_target;
628 drowley@postgresql.o 1817 : 256478 : root->upper_targets[UPPERREL_PARTIAL_DISTINCT] = sort_input_target;
3514 tgl@sss.pgh.pa.us 1818 : 256478 : root->upper_targets[UPPERREL_WINDOW] = sort_input_target;
1819 : 256478 : root->upper_targets[UPPERREL_GROUP_AGG] = grouping_target;
1820 : :
1821 : : /*
1822 : : * If we have grouping and/or aggregation, consider ways to implement
1823 : : * that. We build a new upperrel representing the output of this
1824 : : * phase.
1825 : : */
3519 1826 [ + + ]: 256478 : if (have_grouping)
1827 : : {
3521 1828 : 18687 : current_rel = create_grouping_paths(root,
1829 : : current_rel,
1830 : : grouping_target,
1831 : : grouping_target_parallel_safe,
1832 : : gset_data);
1833 : : /* Fix things up if grouping_target contains SRFs */
3204 andres@anarazel.de 1834 [ + + ]: 18684 : if (parse->hasTargetSRFs)
1835 : 215 : adjust_paths_for_srfs(root, current_rel,
1836 : : grouping_targets,
1837 : : grouping_targets_contain_srfs);
1838 : : }
1839 : :
1840 : : /*
1841 : : * If we have window functions, consider ways to implement those. We
1842 : : * build a new upperrel representing the output of this phase.
1843 : : */
3521 tgl@sss.pgh.pa.us 1844 [ + + ]: 256475 : if (activeWindows)
1845 : : {
1846 : 1282 : current_rel = create_window_paths(root,
1847 : : current_rel,
1848 : : grouping_target,
1849 : : sort_input_target,
1850 : : sort_input_target_parallel_safe,
1851 : : wflists,
1852 : : activeWindows);
1853 : : /* Fix things up if sort_input_target contains SRFs */
3204 andres@anarazel.de 1854 [ + + ]: 1282 : if (parse->hasTargetSRFs)
1855 : 6 : adjust_paths_for_srfs(root, current_rel,
1856 : : sort_input_targets,
1857 : : sort_input_targets_contain_srfs);
1858 : : }
1859 : :
1860 : : /*
1861 : : * If there is a DISTINCT clause, consider ways to implement that. We
1862 : : * build a new upperrel representing the output of this phase.
1863 : : */
3521 tgl@sss.pgh.pa.us 1864 [ + + ]: 256475 : if (parse->distinctClause)
1865 : : {
1866 : 1479 : current_rel = create_distinct_paths(root,
1867 : : current_rel,
1868 : : sort_input_target);
1869 : : }
1870 : : } /* end of if (setOperations) */
1871 : :
1872 : : /*
1873 : : * If ORDER BY was given, consider ways to implement that, and generate a
1874 : : * new upperrel containing only paths that emit the correct ordering and
1875 : : * project the correct final_target. We can apply the original
1876 : : * limit_tuples limit in sort costing here, but only if there are no
1877 : : * postponed SRFs.
1878 : : */
1879 [ + + ]: 259548 : if (parse->sortClause)
1880 : : {
1881 [ + + ]: 37097 : current_rel = create_ordered_paths(root,
1882 : : current_rel,
1883 : : final_target,
1884 : : final_target_parallel_safe,
1885 : : have_postponed_srfs ? -1.0 :
1886 : : limit_tuples);
1887 : : /* Fix things up if final_target contains SRFs */
3204 andres@anarazel.de 1888 [ + + ]: 37097 : if (parse->hasTargetSRFs)
1889 : 98 : adjust_paths_for_srfs(root, current_rel,
1890 : : final_targets,
1891 : : final_targets_contain_srfs);
1892 : : }
1893 : :
1894 : : /*
1895 : : * Now we are prepared to build the final-output upperrel.
1896 : : */
3521 tgl@sss.pgh.pa.us 1897 : 259548 : final_rel = fetch_upper_rel(root, UPPERREL_FINAL, NULL);
1898 : :
1899 : : /*
1900 : : * If the input rel is marked consider_parallel and there's nothing that's
1901 : : * not parallel-safe in the LIMIT clause, then the final_rel can be marked
1902 : : * consider_parallel as well. Note that if the query has rowMarks or is
1903 : : * not a SELECT, consider_parallel will be false for every relation in the
1904 : : * query.
1905 : : */
3405 rhaas@postgresql.org 1906 [ + + + + ]: 347701 : if (current_rel->consider_parallel &&
3356 tgl@sss.pgh.pa.us 1907 [ + + ]: 176294 : is_parallel_safe(root, parse->limitOffset) &&
1908 : 88141 : is_parallel_safe(root, parse->limitCount))
3405 rhaas@postgresql.org 1909 : 88138 : final_rel->consider_parallel = true;
1910 : :
1911 : : /*
1912 : : * If the current_rel belongs to a single FDW, so does the final_rel.
1913 : : */
tgl@sss.pgh.pa.us 1914 : 259548 : final_rel->serverid = current_rel->serverid;
3391 1915 : 259548 : final_rel->userid = current_rel->userid;
1916 : 259548 : final_rel->useridiscurrent = current_rel->useridiscurrent;
3405 1917 : 259548 : final_rel->fdwroutine = current_rel->fdwroutine;
1918 : :
1919 : : /*
1920 : : * Generate paths for the final_rel. Insert all surviving paths, with
1921 : : * LockRows, Limit, and/or ModifyTable steps added if needed.
1922 : : */
3521 1923 [ + - + + : 529288 : foreach(lc, current_rel->pathlist)
+ + ]
1924 : : {
1925 : 269740 : Path *path = (Path *) lfirst(lc);
1926 : :
1927 : : /*
1928 : : * If there is a FOR [KEY] UPDATE/SHARE clause, add the LockRows node.
1929 : : * (Note: we intentionally test parse->rowMarks not root->rowMarks
1930 : : * here. If there are only non-locking rowmarks, they should be
1931 : : * handled by the ModifyTable node instead. However, root->rowMarks
1932 : : * is what goes into the LockRows node.)
1933 : : */
1934 [ + + ]: 269740 : if (parse->rowMarks)
1935 : : {
1936 : 4183 : path = (Path *) create_lockrows_path(root, final_rel, path,
1937 : : root->rowMarks,
1938 : : assign_special_exec_param(root));
1939 : : }
1940 : :
1941 : : /*
1942 : : * If there is a LIMIT/OFFSET clause, add the LIMIT node.
1943 : : */
1944 [ + + ]: 269740 : if (limit_needed(parse))
1945 : : {
1946 : 2956 : path = (Path *) create_limit_path(root, final_rel, path,
1947 : : parse->limitOffset,
1948 : : parse->limitCount,
1949 : : parse->limitOption,
1950 : : offset_est, count_est);
1951 : : }
1952 : :
1953 : : /*
1954 : : * If this is an INSERT/UPDATE/DELETE/MERGE, add the ModifyTable node.
1955 : : */
1671 1956 [ + + ]: 269740 : if (parse->commandType != CMD_SELECT)
1957 : : {
1958 : : Index rootRelation;
1959 : 42152 : List *resultRelations = NIL;
1960 : 42152 : List *updateColnosLists = NIL;
1961 : 42152 : List *withCheckOptionLists = NIL;
1962 : 42152 : List *returningLists = NIL;
1309 alvherre@alvh.no-ip. 1963 : 42152 : List *mergeActionLists = NIL;
576 dean.a.rasheed@gmail 1964 : 42152 : List *mergeJoinConditions = NIL;
1965 : : List *rowMarks;
1966 : :
1671 tgl@sss.pgh.pa.us 1967 [ + + ]: 42152 : if (bms_membership(root->all_result_relids) == BMS_MULTIPLE)
1968 : : {
1969 : : /* Inherited UPDATE/DELETE/MERGE */
1970 : 1437 : RelOptInfo *top_result_rel = find_base_rel(root,
1971 : : parse->resultRelation);
1972 : 1437 : int resultRelation = -1;
1973 : :
1974 : : /* Pass the root result rel forward to the executor. */
734 1975 : 1437 : rootRelation = parse->resultRelation;
1976 : :
1977 : : /* Add only leaf children to ModifyTable. */
1671 1978 : 4193 : while ((resultRelation = bms_next_member(root->leaf_result_relids,
1979 [ + + ]: 4193 : resultRelation)) >= 0)
1980 : : {
1981 : 2756 : RelOptInfo *this_result_rel = find_base_rel(root,
1982 : : resultRelation);
1983 : :
1984 : : /*
1985 : : * Also exclude any leaf rels that have turned dummy since
1986 : : * being added to the list, for example, by being excluded
1987 : : * by constraint exclusion.
1988 : : */
1989 [ + + ]: 2756 : if (IS_DUMMY_REL(this_result_rel))
1990 : 87 : continue;
1991 : :
1992 : : /* Build per-target-rel lists needed by ModifyTable */
1993 : 2669 : resultRelations = lappend_int(resultRelations,
1994 : : resultRelation);
1995 [ + + ]: 2669 : if (parse->commandType == CMD_UPDATE)
1996 : : {
1997 : 1833 : List *update_colnos = root->update_colnos;
1998 : :
1999 [ + - ]: 1833 : if (this_result_rel != top_result_rel)
2000 : : update_colnos =
2001 : 1833 : adjust_inherited_attnums_multilevel(root,
2002 : : update_colnos,
2003 : : this_result_rel->relid,
2004 : : top_result_rel->relid);
2005 : 1833 : updateColnosLists = lappend(updateColnosLists,
2006 : : update_colnos);
2007 : : }
2008 [ + + ]: 2669 : if (parse->withCheckOptions)
2009 : : {
2010 : 252 : List *withCheckOptions = parse->withCheckOptions;
2011 : :
2012 [ + - ]: 252 : if (this_result_rel != top_result_rel)
2013 : : withCheckOptions = (List *)
2014 : 252 : adjust_appendrel_attrs_multilevel(root,
2015 : : (Node *) withCheckOptions,
2016 : : this_result_rel,
2017 : : top_result_rel);
2018 : 252 : withCheckOptionLists = lappend(withCheckOptionLists,
2019 : : withCheckOptions);
2020 : : }
2021 [ + + ]: 2669 : if (parse->returningList)
2022 : : {
2023 : 420 : List *returningList = parse->returningList;
2024 : :
2025 [ + - ]: 420 : if (this_result_rel != top_result_rel)
2026 : : returningList = (List *)
2027 : 420 : adjust_appendrel_attrs_multilevel(root,
2028 : : (Node *) returningList,
2029 : : this_result_rel,
2030 : : top_result_rel);
2031 : 420 : returningLists = lappend(returningLists,
2032 : : returningList);
2033 : : }
1309 alvherre@alvh.no-ip. 2034 [ + + ]: 2669 : if (parse->mergeActionList)
2035 : : {
2036 : : ListCell *l;
2037 : 267 : List *mergeActionList = NIL;
2038 : :
2039 : : /*
2040 : : * Copy MergeActions and translate stuff that
2041 : : * references attribute numbers.
2042 : : */
2043 [ + - + + : 831 : foreach(l, parse->mergeActionList)
+ + ]
2044 : : {
2045 : 564 : MergeAction *action = lfirst(l),
2046 : 564 : *leaf_action = copyObject(action);
2047 : :
2048 : 564 : leaf_action->qual =
2049 : 564 : adjust_appendrel_attrs_multilevel(root,
2050 : : (Node *) action->qual,
2051 : : this_result_rel,
2052 : : top_result_rel);
2053 : 564 : leaf_action->targetList = (List *)
2054 : 564 : adjust_appendrel_attrs_multilevel(root,
2055 : 564 : (Node *) action->targetList,
2056 : : this_result_rel,
2057 : : top_result_rel);
2058 [ + + ]: 564 : if (leaf_action->commandType == CMD_UPDATE)
2059 : 314 : leaf_action->updateColnos =
2060 : 314 : adjust_inherited_attnums_multilevel(root,
2061 : : action->updateColnos,
2062 : : this_result_rel->relid,
2063 : : top_result_rel->relid);
2064 : 564 : mergeActionList = lappend(mergeActionList,
2065 : : leaf_action);
2066 : : }
2067 : :
2068 : 267 : mergeActionLists = lappend(mergeActionLists,
2069 : : mergeActionList);
2070 : : }
576 dean.a.rasheed@gmail 2071 [ + + ]: 2669 : if (parse->commandType == CMD_MERGE)
2072 : : {
2073 : 267 : Node *mergeJoinCondition = parse->mergeJoinCondition;
2074 : :
2075 [ + - ]: 267 : if (this_result_rel != top_result_rel)
2076 : : mergeJoinCondition =
2077 : 267 : adjust_appendrel_attrs_multilevel(root,
2078 : : mergeJoinCondition,
2079 : : this_result_rel,
2080 : : top_result_rel);
2081 : 267 : mergeJoinConditions = lappend(mergeJoinConditions,
2082 : : mergeJoinCondition);
2083 : : }
2084 : : }
2085 : :
1671 tgl@sss.pgh.pa.us 2086 [ + + ]: 1437 : if (resultRelations == NIL)
2087 : : {
2088 : : /*
2089 : : * We managed to exclude every child rel, so generate a
2090 : : * dummy one-relation plan using info for the top target
2091 : : * rel (even though that may not be a leaf target).
2092 : : * Although it's clear that no data will be updated or
2093 : : * deleted, we still need to have a ModifyTable node so
2094 : : * that any statement triggers will be executed. (This
2095 : : * could be cleaner if we fixed nodeModifyTable.c to allow
2096 : : * zero target relations, but that probably wouldn't be a
2097 : : * net win.)
2098 : : */
2099 : 15 : resultRelations = list_make1_int(parse->resultRelation);
2100 [ + - ]: 15 : if (parse->commandType == CMD_UPDATE)
2101 : 15 : updateColnosLists = list_make1(root->update_colnos);
2102 [ - + ]: 15 : if (parse->withCheckOptions)
1671 tgl@sss.pgh.pa.us 2103 :UBC 0 : withCheckOptionLists = list_make1(parse->withCheckOptions);
1671 tgl@sss.pgh.pa.us 2104 [ + + ]:CBC 15 : if (parse->returningList)
2105 : 9 : returningLists = list_make1(parse->returningList);
1309 alvherre@alvh.no-ip. 2106 [ - + ]: 15 : if (parse->mergeActionList)
1309 alvherre@alvh.no-ip. 2107 :UBC 0 : mergeActionLists = list_make1(parse->mergeActionList);
576 dean.a.rasheed@gmail 2108 [ - + ]:CBC 15 : if (parse->commandType == CMD_MERGE)
576 dean.a.rasheed@gmail 2109 :UBC 0 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2110 : : }
2111 : : }
2112 : : else
2113 : : {
2114 : : /* Single-relation INSERT/UPDATE/DELETE/MERGE. */
734 tgl@sss.pgh.pa.us 2115 :CBC 40715 : rootRelation = 0; /* there's no separate root rel */
1671 2116 : 40715 : resultRelations = list_make1_int(parse->resultRelation);
2117 [ + + ]: 40715 : if (parse->commandType == CMD_UPDATE)
2118 : 6046 : updateColnosLists = list_make1(root->update_colnos);
2119 [ + + ]: 40715 : if (parse->withCheckOptions)
2120 : 463 : withCheckOptionLists = list_make1(parse->withCheckOptions);
2121 [ + + ]: 40715 : if (parse->returningList)
2122 : 1218 : returningLists = list_make1(parse->returningList);
1309 alvherre@alvh.no-ip. 2123 [ + + ]: 40715 : if (parse->mergeActionList)
2124 : 771 : mergeActionLists = list_make1(parse->mergeActionList);
576 dean.a.rasheed@gmail 2125 [ + + ]: 40715 : if (parse->commandType == CMD_MERGE)
2126 : 771 : mergeJoinConditions = list_make1(parse->mergeJoinCondition);
2127 : : }
2128 : :
2129 : : /*
2130 : : * If there was a FOR [KEY] UPDATE/SHARE clause, the LockRows node
2131 : : * will have dealt with fetching non-locked marked rows, else we
2132 : : * need to have ModifyTable do that.
2133 : : */
3521 tgl@sss.pgh.pa.us 2134 [ - + ]: 42152 : if (parse->rowMarks)
3521 tgl@sss.pgh.pa.us 2135 :UBC 0 : rowMarks = NIL;
2136 : : else
3521 tgl@sss.pgh.pa.us 2137 :CBC 42152 : rowMarks = root->rowMarks;
2138 : :
2139 : : path = (Path *)
2140 : 42152 : create_modifytable_path(root, final_rel,
2141 : : path,
2142 : : parse->commandType,
2143 : 42152 : parse->canSetTag,
2144 : 42152 : parse->resultRelation,
2145 : : rootRelation,
2146 : : resultRelations,
2147 : : updateColnosLists,
2148 : : withCheckOptionLists,
2149 : : returningLists,
2150 : : rowMarks,
2151 : : parse->onConflict,
2152 : : mergeActionLists,
2153 : : mergeJoinConditions,
2154 : : assign_special_exec_param(root));
2155 : : }
2156 : :
2157 : : /* And shove it into final_rel */
2158 : 269740 : add_path(final_rel, path);
2159 : : }
2160 : :
2161 : : /*
2162 : : * Generate partial paths for final_rel, too, if outer query levels might
2163 : : * be able to make use of them.
2164 : : */
2785 rhaas@postgresql.org 2165 [ + + + + ]: 259548 : if (final_rel->consider_parallel && root->query_level > 1 &&
2166 [ + + ]: 15174 : !limit_needed(parse))
2167 : : {
2168 [ + - - + ]: 15082 : Assert(!parse->rowMarks && parse->commandType == CMD_SELECT);
2169 [ + + + + : 15136 : foreach(lc, current_rel->partial_pathlist)
+ + ]
2170 : : {
2171 : 54 : Path *partial_path = (Path *) lfirst(lc);
2172 : :
2173 : 54 : add_partial_path(final_rel, partial_path);
2174 : : }
2175 : : }
2176 : :
2400 efujita@postgresql.o 2177 : 259548 : extra.limit_needed = limit_needed(parse);
2178 : 259548 : extra.limit_tuples = limit_tuples;
2179 : 259548 : extra.count_est = count_est;
2180 : 259548 : extra.offset_est = offset_est;
2181 : :
2182 : : /*
2183 : : * If there is an FDW that's responsible for all baserels of the query,
2184 : : * let it consider adding ForeignPaths.
2185 : : */
3405 tgl@sss.pgh.pa.us 2186 [ + + ]: 259548 : if (final_rel->fdwroutine &&
2187 [ + + ]: 635 : final_rel->fdwroutine->GetForeignUpperPaths)
2188 : 601 : final_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_FINAL,
2189 : : current_rel, final_rel,
2190 : : &extra);
2191 : :
2192 : : /* Let extensions possibly add some more paths */
3485 2193 [ - + ]: 259548 : if (create_upper_paths_hook)
3485 tgl@sss.pgh.pa.us 2194 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_FINAL,
2195 : : current_rel, final_rel, &extra);
2196 : :
2197 : : /* Note: currently, we leave it to callers to do set_cheapest() */
3521 tgl@sss.pgh.pa.us 2198 :CBC 259548 : }
2199 : :
2200 : : /*
2201 : : * Do preprocessing for groupingSets clause and related data.
2202 : : *
2203 : : * We expect that parse->groupingSets has already been expanded into a flat
2204 : : * list of grouping sets (that is, just integer Lists of ressortgroupref
2205 : : * numbers) by expand_grouping_sets(). This function handles the preliminary
2206 : : * steps of organizing the grouping sets into lists of rollups, and preparing
2207 : : * annotations which will later be filled in with size estimates.
2208 : : */
2209 : : static grouping_sets_data *
3136 rhodiumtoad@postgres 2210 : 469 : preprocess_grouping_sets(PlannerInfo *root)
2211 : : {
2212 : 469 : Query *parse = root->parse;
2213 : : List *sets;
2214 : 469 : int maxref = 0;
2215 : : ListCell *lc_set;
2216 : 469 : grouping_sets_data *gd = palloc0(sizeof(grouping_sets_data));
2217 : :
2218 : : /*
2219 : : * We don't currently make any attempt to optimize the groupClause when
2220 : : * there are grouping sets, so just duplicate it in processed_groupClause.
2221 : : */
1013 tgl@sss.pgh.pa.us 2222 : 469 : root->processed_groupClause = parse->groupClause;
2223 : :
2224 : : /* Detect unhashable and unsortable grouping expressions */
6 rguo@postgresql.org 2225 : 469 : gd->any_hashable = false;
2226 : 469 : gd->unhashable_refs = NULL;
2227 : 469 : gd->unsortable_refs = NULL;
2228 : 469 : gd->unsortable_sets = NIL;
2229 : :
3136 rhodiumtoad@postgres 2230 [ + + ]: 469 : if (parse->groupClause)
2231 : : {
2232 : : ListCell *lc;
2233 : :
2234 [ + - + + : 1390 : foreach(lc, parse->groupClause)
+ + ]
2235 : : {
2974 tgl@sss.pgh.pa.us 2236 : 954 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
3136 rhodiumtoad@postgres 2237 : 954 : Index ref = gc->tleSortGroupRef;
2238 : :
2239 [ + + ]: 954 : if (ref > maxref)
2240 : 936 : maxref = ref;
2241 : :
2242 [ + + ]: 954 : if (!gc->hashable)
2243 : 15 : gd->unhashable_refs = bms_add_member(gd->unhashable_refs, ref);
2244 : :
2245 [ + + ]: 954 : if (!OidIsValid(gc->sortop))
2246 : 21 : gd->unsortable_refs = bms_add_member(gd->unsortable_refs, ref);
2247 : : }
2248 : : }
2249 : :
2250 : : /* Allocate workspace array for remapping */
2251 : 469 : gd->tleref_to_colnum_map = (int *) palloc((maxref + 1) * sizeof(int));
2252 : :
2253 : : /*
2254 : : * If we have any unsortable sets, we must extract them before trying to
2255 : : * prepare rollups. Unsortable sets don't go through
2256 : : * reorder_grouping_sets, so we must apply the GroupingSetData annotation
2257 : : * here.
2258 : : */
2259 [ + + ]: 469 : if (!bms_is_empty(gd->unsortable_refs))
2260 : : {
2261 : 21 : List *sortable_sets = NIL;
2262 : : ListCell *lc;
2263 : :
2264 [ + - + + : 63 : foreach(lc, parse->groupingSets)
+ + ]
2265 : : {
2974 tgl@sss.pgh.pa.us 2266 : 45 : List *gset = (List *) lfirst(lc);
2267 : :
3136 rhodiumtoad@postgres 2268 [ + + ]: 45 : if (bms_overlap_list(gd->unsortable_refs, gset))
2269 : : {
2270 : 24 : GroupingSetData *gs = makeNode(GroupingSetData);
2271 : :
2272 : 24 : gs->set = gset;
2273 : 24 : gd->unsortable_sets = lappend(gd->unsortable_sets, gs);
2274 : :
2275 : : /*
2276 : : * We must enforce here that an unsortable set is hashable;
2277 : : * later code assumes this. Parse analysis only checks that
2278 : : * every individual column is either hashable or sortable.
2279 : : *
2280 : : * Note that passing this test doesn't guarantee we can
2281 : : * generate a plan; there might be other showstoppers.
2282 : : */
2283 [ + + ]: 24 : if (bms_overlap_list(gd->unhashable_refs, gset))
2284 [ + - ]: 3 : ereport(ERROR,
2285 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2286 : : errmsg("could not implement GROUP BY"),
2287 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
2288 : : }
2289 : : else
2290 : 21 : sortable_sets = lappend(sortable_sets, gset);
2291 : : }
2292 : :
2293 [ + + ]: 18 : if (sortable_sets)
2294 : 15 : sets = extract_rollup_sets(sortable_sets);
2295 : : else
2296 : 3 : sets = NIL;
2297 : : }
2298 : : else
2299 : 448 : sets = extract_rollup_sets(parse->groupingSets);
2300 : :
2301 [ + + + + : 1221 : foreach(lc_set, sets)
+ + ]
2302 : : {
2303 : 755 : List *current_sets = (List *) lfirst(lc_set);
2304 : 755 : RollupData *rollup = makeNode(RollupData);
2305 : : GroupingSetData *gs;
2306 : :
2307 : : /*
2308 : : * Reorder the current list of grouping sets into correct prefix
2309 : : * order. If only one aggregation pass is needed, try to make the
2310 : : * list match the ORDER BY clause; if more than one pass is needed, we
2311 : : * don't bother with that.
2312 : : *
2313 : : * Note that this reorders the sets from smallest-member-first to
2314 : : * largest-member-first, and applies the GroupingSetData annotations,
2315 : : * though the data will be filled in later.
2316 : : */
2317 [ + + ]: 755 : current_sets = reorder_grouping_sets(current_sets,
2318 : 755 : (list_length(sets) == 1
2319 : : ? parse->sortClause
2320 : : : NIL));
2321 : :
2322 : : /*
2323 : : * Get the initial (and therefore largest) grouping set.
2324 : : */
2974 tgl@sss.pgh.pa.us 2325 : 755 : gs = linitial_node(GroupingSetData, current_sets);
2326 : :
2327 : : /*
2328 : : * Order the groupClause appropriately. If the first grouping set is
2329 : : * empty, then the groupClause must also be empty; otherwise we have
2330 : : * to force the groupClause to match that grouping set's order.
2331 : : *
2332 : : * (The first grouping set can be empty even though parse->groupClause
2333 : : * is not empty only if all non-empty grouping sets are unsortable.
2334 : : * The groupClauses for hashed grouping sets are built later on.)
2335 : : */
3136 rhodiumtoad@postgres 2336 [ + + ]: 755 : if (gs->set)
508 akorotkov@postgresql 2337 : 722 : rollup->groupClause = preprocess_groupclause(root, gs->set);
2338 : : else
3136 rhodiumtoad@postgres 2339 : 33 : rollup->groupClause = NIL;
2340 : :
2341 : : /*
2342 : : * Is it hashable? We pretend empty sets are hashable even though we
2343 : : * actually force them not to be hashed later. But don't bother if
2344 : : * there's nothing but empty sets (since in that case we can't hash
2345 : : * anything).
2346 : : */
2347 [ + + ]: 755 : if (gs->set &&
2348 [ + + ]: 722 : !bms_overlap_list(gd->unhashable_refs, gs->set))
2349 : : {
2350 : 710 : rollup->hashable = true;
2351 : 710 : gd->any_hashable = true;
2352 : : }
2353 : :
2354 : : /*
2355 : : * Now that we've pinned down an order for the groupClause for this
2356 : : * list of grouping sets, we need to remap the entries in the grouping
2357 : : * sets from sortgrouprefs to plain indices (0-based) into the
2358 : : * groupClause for this collection of grouping sets. We keep the
2359 : : * original form for later use, though.
2360 : : */
2361 : 755 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
2362 : : current_sets,
2363 : : gd->tleref_to_colnum_map);
2364 : 755 : rollup->gsets_data = current_sets;
2365 : :
2366 : 755 : gd->rollups = lappend(gd->rollups, rollup);
2367 : : }
2368 : :
2369 [ + + ]: 466 : if (gd->unsortable_sets)
2370 : : {
2371 : : /*
2372 : : * We have not yet pinned down a groupclause for this, but we will
2373 : : * need index-based lists for estimation purposes. Construct
2374 : : * hash_sets_idx based on the entire original groupclause for now.
2375 : : */
2376 : 18 : gd->hash_sets_idx = remap_to_groupclause_idx(parse->groupClause,
2377 : : gd->unsortable_sets,
2378 : : gd->tleref_to_colnum_map);
2379 : 18 : gd->any_hashable = true;
2380 : : }
2381 : :
2382 : 466 : return gd;
2383 : : }
2384 : :
2385 : : /*
2386 : : * Given a groupclause and a list of GroupingSetData, return equivalent sets
2387 : : * (without annotation) mapped to indexes into the given groupclause.
2388 : : */
2389 : : static List *
2390 : 2172 : remap_to_groupclause_idx(List *groupClause,
2391 : : List *gsets,
2392 : : int *tleref_to_colnum_map)
2393 : : {
2394 : 2172 : int ref = 0;
2395 : 2172 : List *result = NIL;
2396 : : ListCell *lc;
2397 : :
2398 [ + + + + : 5276 : foreach(lc, groupClause)
+ + ]
2399 : : {
2974 tgl@sss.pgh.pa.us 2400 : 3104 : SortGroupClause *gc = lfirst_node(SortGroupClause, lc);
2401 : :
3136 rhodiumtoad@postgres 2402 : 3104 : tleref_to_colnum_map[gc->tleSortGroupRef] = ref++;
2403 : : }
2404 : :
2405 [ + - + + : 5007 : foreach(lc, gsets)
+ + ]
2406 : : {
2407 : 2835 : List *set = NIL;
2408 : : ListCell *lc2;
2974 tgl@sss.pgh.pa.us 2409 : 2835 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
2410 : :
3136 rhodiumtoad@postgres 2411 [ + + + + : 6364 : foreach(lc2, gs->set)
+ + ]
2412 : : {
2413 : 3529 : set = lappend_int(set, tleref_to_colnum_map[lfirst_int(lc2)]);
2414 : : }
2415 : :
2416 : 2835 : result = lappend(result, set);
2417 : : }
2418 : :
2419 : 2172 : return result;
2420 : : }
2421 : :
2422 : :
2423 : : /*
2424 : : * preprocess_rowmarks - set up PlanRowMarks if needed
2425 : : */
2426 : : static void
5845 tgl@sss.pgh.pa.us 2427 : 261557 : preprocess_rowmarks(PlannerInfo *root)
2428 : : {
2429 : 261557 : Query *parse = root->parse;
2430 : : Bitmapset *rels;
2431 : : List *prowmarks;
2432 : : ListCell *l;
2433 : : int i;
2434 : :
2435 [ + + ]: 261557 : if (parse->rowMarks)
2436 : : {
2437 : : /*
2438 : : * We've got trouble if FOR [KEY] UPDATE/SHARE appears inside
2439 : : * grouping, since grouping renders a reference to individual tuple
2440 : : * CTIDs invalid. This is also checked at parse time, but that's
2441 : : * insufficient because of rule substitution, query pullup, etc.
2442 : : */
2974 2443 : 3939 : CheckSelectLocking(parse, linitial_node(RowMarkClause,
2444 : : parse->rowMarks)->strength);
2445 : : }
2446 : : else
2447 : : {
2448 : : /*
2449 : : * We only need rowmarks for UPDATE, DELETE, MERGE, or FOR [KEY]
2450 : : * UPDATE/SHARE.
2451 : : */
5845 2452 [ + + ]: 257618 : if (parse->commandType != CMD_UPDATE &&
758 dean.a.rasheed@gmail 2453 [ + + ]: 250566 : parse->commandType != CMD_DELETE &&
2454 [ + + ]: 248399 : parse->commandType != CMD_MERGE)
5845 tgl@sss.pgh.pa.us 2455 : 247505 : return;
2456 : : }
2457 : :
2458 : : /*
2459 : : * We need to have rowmarks for all base relations except the target. We
2460 : : * make a bitmapset of all base rels and then remove the items we don't
2461 : : * need or have FOR [KEY] UPDATE/SHARE marks for.
2462 : : */
1001 2463 : 14052 : rels = get_relids_in_jointree((Node *) parse->jointree, false, false);
5845 2464 [ + + ]: 14052 : if (parse->resultRelation)
2465 : 10113 : rels = bms_del_member(rels, parse->resultRelation);
2466 : :
2467 : : /*
2468 : : * Convert RowMarkClauses to PlanRowMark representation.
2469 : : */
2470 : 14052 : prowmarks = NIL;
2471 [ + + + + : 18106 : foreach(l, parse->rowMarks)
+ + ]
2472 : : {
2974 2473 : 4054 : RowMarkClause *rc = lfirst_node(RowMarkClause, l);
5843 2474 : 4054 : RangeTblEntry *rte = rt_fetch(rc->rti, parse->rtable);
2475 : : PlanRowMark *newrc;
2476 : :
2477 : : /*
2478 : : * Currently, it is syntactically impossible to have FOR UPDATE et al
2479 : : * applied to an update/delete target rel. If that ever becomes
2480 : : * possible, we should drop the target from the PlanRowMark list.
2481 : : */
5845 2482 [ - + ]: 4054 : Assert(rc->rti != parse->resultRelation);
2483 : :
2484 : : /*
2485 : : * Ignore RowMarkClauses for subqueries; they aren't real tables and
2486 : : * can't support true locking. Subqueries that got flattened into the
2487 : : * main query should be ignored completely. Any that didn't will get
2488 : : * ROW_MARK_COPY items in the next loop.
2489 : : */
5843 2490 [ + + ]: 4054 : if (rte->rtekind != RTE_RELATION)
2491 : 30 : continue;
2492 : :
5845 2493 : 4024 : rels = bms_del_member(rels, rc->rti);
2494 : :
5843 2495 : 4024 : newrc = makeNode(PlanRowMark);
5845 2496 : 4024 : newrc->rti = newrc->prti = rc->rti;
5374 2497 : 4024 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
3872 2498 : 4024 : newrc->markType = select_rowmark_type(rte, rc->strength);
3879 2499 : 4024 : newrc->allMarkTypes = (1 << newrc->markType);
2500 : 4024 : newrc->strength = rc->strength;
4038 alvherre@alvh.no-ip. 2501 : 4024 : newrc->waitPolicy = rc->waitPolicy;
5845 tgl@sss.pgh.pa.us 2502 : 4024 : newrc->isParent = false;
2503 : :
2504 : 4024 : prowmarks = lappend(prowmarks, newrc);
2505 : : }
2506 : :
2507 : : /*
2508 : : * Now, add rowmarks for any non-target, non-locked base relations.
2509 : : */
2510 : 14052 : i = 0;
2511 [ + - + + : 33443 : foreach(l, parse->rtable)
+ + ]
2512 : : {
2974 2513 : 19391 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, l);
2514 : : PlanRowMark *newrc;
2515 : :
5845 2516 : 19391 : i++;
2517 [ + + ]: 19391 : if (!bms_is_member(i, rels))
2518 : 17552 : continue;
2519 : :
2520 : 1839 : newrc = makeNode(PlanRowMark);
2521 : 1839 : newrc->rti = newrc->prti = i;
5374 2522 : 1839 : newrc->rowmarkId = ++(root->glob->lastRowMarkId);
3872 2523 : 1839 : newrc->markType = select_rowmark_type(rte, LCS_NONE);
3879 2524 : 1839 : newrc->allMarkTypes = (1 << newrc->markType);
2525 : 1839 : newrc->strength = LCS_NONE;
3050 2526 : 1839 : newrc->waitPolicy = LockWaitBlock; /* doesn't matter */
5845 2527 : 1839 : newrc->isParent = false;
2528 : :
2529 : 1839 : prowmarks = lappend(prowmarks, newrc);
2530 : : }
2531 : :
2532 : 14052 : root->rowMarks = prowmarks;
2533 : : }
2534 : :
2535 : : /*
2536 : : * Select RowMarkType to use for a given table
2537 : : */
2538 : : RowMarkType
3872 2539 : 7035 : select_rowmark_type(RangeTblEntry *rte, LockClauseStrength strength)
2540 : : {
2541 [ + + ]: 7035 : if (rte->rtekind != RTE_RELATION)
2542 : : {
2543 : : /* If it's not a table at all, use ROW_MARK_COPY */
2544 : 702 : return ROW_MARK_COPY;
2545 : : }
2546 [ + + ]: 6333 : else if (rte->relkind == RELKIND_FOREIGN_TABLE)
2547 : : {
2548 : : /* Let the FDW select the rowmark type, if it wants to */
3821 2549 : 106 : FdwRoutine *fdwroutine = GetFdwRoutineByRelId(rte->relid);
2550 : :
2551 [ - + ]: 106 : if (fdwroutine->GetForeignRowMarkType != NULL)
3821 tgl@sss.pgh.pa.us 2552 :UBC 0 : return fdwroutine->GetForeignRowMarkType(rte, strength);
2553 : : /* Otherwise, use ROW_MARK_COPY by default */
3872 tgl@sss.pgh.pa.us 2554 :CBC 106 : return ROW_MARK_COPY;
2555 : : }
2556 : : else
2557 : : {
2558 : : /* Regular table, apply the appropriate lock type */
2559 [ + + + + : 6227 : switch (strength)
+ - ]
2560 : : {
2561 : 1252 : case LCS_NONE:
2562 : :
2563 : : /*
2564 : : * We don't need a tuple lock, only the ability to re-fetch
2565 : : * the row.
2566 : : */
2567 : 1252 : return ROW_MARK_REFERENCE;
2568 : : break;
2569 : 4033 : case LCS_FORKEYSHARE:
2570 : 4033 : return ROW_MARK_KEYSHARE;
2571 : : break;
2572 : 150 : case LCS_FORSHARE:
2573 : 150 : return ROW_MARK_SHARE;
2574 : : break;
2575 : 36 : case LCS_FORNOKEYUPDATE:
2576 : 36 : return ROW_MARK_NOKEYEXCLUSIVE;
2577 : : break;
2578 : 756 : case LCS_FORUPDATE:
2579 : 756 : return ROW_MARK_EXCLUSIVE;
2580 : : break;
2581 : : }
3872 tgl@sss.pgh.pa.us 2582 [ # # ]:UBC 0 : elog(ERROR, "unrecognized LockClauseStrength %d", (int) strength);
2583 : : return ROW_MARK_EXCLUSIVE; /* keep compiler quiet */
2584 : : }
2585 : : }
2586 : :
2587 : : /*
2588 : : * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
2589 : : *
2590 : : * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
2591 : : * results back in *count_est and *offset_est. These variables are set to
2592 : : * 0 if the corresponding clause is not present, and -1 if it's present
2593 : : * but we couldn't estimate the value for it. (The "0" convention is OK
2594 : : * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
2595 : : * LIMIT 0 as though it were LIMIT 1. But this is in line with the planner's
2596 : : * usual practice of never estimating less than one row.) These values will
2597 : : * be passed to create_limit_path, which see if you change this code.
2598 : : *
2599 : : * The return value is the suitably adjusted tuple_fraction to use for
2600 : : * planning the query. This adjustment is not overridable, since it reflects
2601 : : * plan actions that grouping_planner() will certainly take, not assumptions
2602 : : * about context.
2603 : : */
2604 : : static double
7375 tgl@sss.pgh.pa.us 2605 :CBC 2499 : preprocess_limit(PlannerInfo *root, double tuple_fraction,
2606 : : int64 *offset_est, int64 *count_est)
2607 : : {
7444 2608 : 2499 : Query *parse = root->parse;
2609 : : Node *est;
2610 : : double limit_fraction;
2611 : :
2612 : : /* Should not be called unless LIMIT or OFFSET */
7375 2613 [ + + - + ]: 2499 : Assert(parse->limitCount || parse->limitOffset);
2614 : :
2615 : : /*
2616 : : * Try to obtain the clause values. We use estimate_expression_value
2617 : : * primarily because it can sometimes do something useful with Params.
2618 : : */
2619 [ + + ]: 2499 : if (parse->limitCount)
2620 : : {
6825 2621 : 2236 : est = estimate_expression_value(root, parse->limitCount);
7375 2622 [ + - + + ]: 2236 : if (est && IsA(est, Const))
2623 : : {
2624 [ - + ]: 2233 : if (((Const *) est)->constisnull)
2625 : : {
2626 : : /* NULL indicates LIMIT ALL, ie, no limit */
7317 bruce@momjian.us 2627 :UBC 0 : *count_est = 0; /* treat as not present */
2628 : : }
2629 : : else
2630 : : {
7033 bruce@momjian.us 2631 :CBC 2233 : *count_est = DatumGetInt64(((Const *) est)->constvalue);
7375 tgl@sss.pgh.pa.us 2632 [ + + ]: 2233 : if (*count_est <= 0)
3050 2633 : 75 : *count_est = 1; /* force to at least 1 */
2634 : : }
2635 : : }
2636 : : else
7375 2637 : 3 : *count_est = -1; /* can't estimate */
2638 : : }
2639 : : else
2640 : 263 : *count_est = 0; /* not present */
2641 : :
2642 [ + + ]: 2499 : if (parse->limitOffset)
2643 : : {
6825 2644 : 449 : est = estimate_expression_value(root, parse->limitOffset);
7375 2645 [ + - + + ]: 449 : if (est && IsA(est, Const))
2646 : : {
2647 [ - + ]: 437 : if (((Const *) est)->constisnull)
2648 : : {
2649 : : /* Treat NULL as no offset; the executor will too */
7317 bruce@momjian.us 2650 :UBC 0 : *offset_est = 0; /* treat as not present */
2651 : : }
2652 : : else
2653 : : {
7033 bruce@momjian.us 2654 :CBC 437 : *offset_est = DatumGetInt64(((Const *) est)->constvalue);
7375 tgl@sss.pgh.pa.us 2655 [ - + ]: 437 : if (*offset_est < 0)
4115 tgl@sss.pgh.pa.us 2656 :UBC 0 : *offset_est = 0; /* treat as not present */
2657 : : }
2658 : : }
2659 : : else
7375 tgl@sss.pgh.pa.us 2660 :CBC 12 : *offset_est = -1; /* can't estimate */
2661 : : }
2662 : : else
2663 : 2050 : *offset_est = 0; /* not present */
2664 : :
2665 [ + + ]: 2499 : if (*count_est != 0)
2666 : : {
2667 : : /*
2668 : : * A LIMIT clause limits the absolute number of tuples returned.
2669 : : * However, if it's not a constant LIMIT then we have to guess; for
2670 : : * lack of a better idea, assume 10% of the plan's result is wanted.
2671 : : */
2672 [ + + + + ]: 2236 : if (*count_est < 0 || *offset_est < 0)
2673 : : {
2674 : : /* LIMIT or OFFSET is an expression ... punt ... */
2675 : 12 : limit_fraction = 0.10;
2676 : : }
2677 : : else
2678 : : {
2679 : : /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
2680 : 2224 : limit_fraction = (double) *count_est + (double) *offset_est;
2681 : : }
2682 : :
2683 : : /*
2684 : : * If we have absolute limits from both caller and LIMIT, use the
2685 : : * smaller value; likewise if they are both fractional. If one is
2686 : : * fractional and the other absolute, we can't easily determine which
2687 : : * is smaller, but we use the heuristic that the absolute will usually
2688 : : * be smaller.
2689 : : */
7444 2690 [ + + ]: 2236 : if (tuple_fraction >= 1.0)
2691 : : {
2692 [ + - ]: 3 : if (limit_fraction >= 1.0)
2693 : : {
2694 : : /* both absolute */
2695 [ - + ]: 3 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2696 : : }
2697 : : else
2698 : : {
2699 : : /* caller absolute, limit fractional; use caller's value */
2700 : : }
2701 : : }
2702 [ + + ]: 2233 : else if (tuple_fraction > 0.0)
2703 : : {
2704 [ + - ]: 74 : if (limit_fraction >= 1.0)
2705 : : {
2706 : : /* caller fractional, limit absolute; use limit */
7375 2707 : 74 : tuple_fraction = limit_fraction;
2708 : : }
2709 : : else
2710 : : {
2711 : : /* both fractional */
7375 tgl@sss.pgh.pa.us 2712 [ # # ]:UBC 0 : tuple_fraction = Min(tuple_fraction, limit_fraction);
2713 : : }
2714 : : }
2715 : : else
2716 : : {
2717 : : /* no info from caller, just use limit */
7444 tgl@sss.pgh.pa.us 2718 :CBC 2159 : tuple_fraction = limit_fraction;
2719 : : }
2720 : : }
7375 2721 [ + + + + ]: 263 : else if (*offset_est != 0 && tuple_fraction > 0.0)
2722 : : {
2723 : : /*
2724 : : * We have an OFFSET but no LIMIT. This acts entirely differently
2725 : : * from the LIMIT case: here, we need to increase rather than decrease
2726 : : * the caller's tuple_fraction, because the OFFSET acts to cause more
2727 : : * tuples to be fetched instead of fewer. This only matters if we got
2728 : : * a tuple_fraction > 0, however.
2729 : : *
2730 : : * As above, use 10% if OFFSET is present but unestimatable.
2731 : : */
2732 [ - + ]: 7 : if (*offset_est < 0)
7375 tgl@sss.pgh.pa.us 2733 :UBC 0 : limit_fraction = 0.10;
2734 : : else
7375 tgl@sss.pgh.pa.us 2735 :CBC 7 : limit_fraction = (double) *offset_est;
2736 : :
2737 : : /*
2738 : : * If we have absolute counts from both caller and OFFSET, add them
2739 : : * together; likewise if they are both fractional. If one is
2740 : : * fractional and the other absolute, we want to take the larger, and
2741 : : * we heuristically assume that's the fractional one.
2742 : : */
2743 [ - + ]: 7 : if (tuple_fraction >= 1.0)
2744 : : {
7375 tgl@sss.pgh.pa.us 2745 [ # # ]:UBC 0 : if (limit_fraction >= 1.0)
2746 : : {
2747 : : /* both absolute, so add them together */
2748 : 0 : tuple_fraction += limit_fraction;
2749 : : }
2750 : : else
2751 : : {
2752 : : /* caller absolute, limit fractional; use limit */
2753 : 0 : tuple_fraction = limit_fraction;
2754 : : }
2755 : : }
2756 : : else
2757 : : {
7375 tgl@sss.pgh.pa.us 2758 [ - + ]:CBC 7 : if (limit_fraction >= 1.0)
2759 : : {
2760 : : /* caller fractional, limit absolute; use caller's value */
2761 : : }
2762 : : else
2763 : : {
2764 : : /* both fractional, so add them together */
7375 tgl@sss.pgh.pa.us 2765 :UBC 0 : tuple_fraction += limit_fraction;
2766 [ # # ]: 0 : if (tuple_fraction >= 1.0)
3050 2767 : 0 : tuple_fraction = 0.0; /* assume fetch all */
2768 : : }
2769 : : }
2770 : : }
2771 : :
7444 tgl@sss.pgh.pa.us 2772 :CBC 2499 : return tuple_fraction;
2773 : : }
2774 : :
2775 : : /*
2776 : : * limit_needed - do we actually need a Limit plan node?
2777 : : *
2778 : : * If we have constant-zero OFFSET and constant-null LIMIT, we can skip adding
2779 : : * a Limit node. This is worth checking for because "OFFSET 0" is a common
2780 : : * locution for an optimization fence. (Because other places in the planner
2781 : : * merely check whether parse->limitOffset isn't NULL, it will still work as
2782 : : * an optimization fence --- we're just suppressing unnecessary run-time
2783 : : * overhead.)
2784 : : *
2785 : : * This might look like it could be merged into preprocess_limit, but there's
2786 : : * a key distinction: here we need hard constants in OFFSET/LIMIT, whereas
2787 : : * in preprocess_limit it's good enough to consider estimated values.
2788 : : */
2789 : : bool
4610 2790 : 554082 : limit_needed(Query *parse)
2791 : : {
2792 : : Node *node;
2793 : :
2794 : 554082 : node = parse->limitCount;
2795 [ + + ]: 554082 : if (node)
2796 : : {
2797 [ + + ]: 5350 : if (IsA(node, Const))
2798 : : {
2799 : : /* NULL indicates LIMIT ALL, ie, no limit */
2800 [ + - ]: 5232 : if (!((Const *) node)->constisnull)
2801 : 5232 : return true; /* LIMIT with a constant value */
2802 : : }
2803 : : else
2804 : 118 : return true; /* non-constant LIMIT */
2805 : : }
2806 : :
2807 : 548732 : node = parse->limitOffset;
2808 [ + + ]: 548732 : if (node)
2809 : : {
2810 [ + + ]: 765 : if (IsA(node, Const))
2811 : : {
2812 : : /* Treat NULL as no offset; the executor would too */
2813 [ + - ]: 611 : if (!((Const *) node)->constisnull)
2814 : : {
4534 bruce@momjian.us 2815 : 611 : int64 offset = DatumGetInt64(((Const *) node)->constvalue);
2816 : :
4115 tgl@sss.pgh.pa.us 2817 [ + + ]: 611 : if (offset != 0)
2818 : 71 : return true; /* OFFSET with a nonzero value */
2819 : : }
2820 : : }
2821 : : else
4610 2822 : 154 : return true; /* non-constant OFFSET */
2823 : : }
2824 : :
2825 : 548507 : return false; /* don't need a Limit plan node */
2826 : : }
2827 : :
2828 : : /*
2829 : : * preprocess_groupclause - do preparatory work on GROUP BY clause
2830 : : *
2831 : : * The idea here is to adjust the ordering of the GROUP BY elements
2832 : : * (which in itself is semantically insignificant) to match ORDER BY,
2833 : : * thereby allowing a single sort operation to both implement the ORDER BY
2834 : : * requirement and set up for a Unique step that implements GROUP BY.
2835 : : * We also consider partial match between GROUP BY and ORDER BY elements,
2836 : : * which could allow to implement ORDER BY using the incremental sort.
2837 : : *
2838 : : * We also consider other orderings of the GROUP BY elements, which could
2839 : : * match the sort ordering of other possible plans (eg an indexscan) and
2840 : : * thereby reduce cost. This is implemented during the generation of grouping
2841 : : * paths. See get_useful_group_keys_orderings() for details.
2842 : : *
2843 : : * Note: we need no comparable processing of the distinctClause because
2844 : : * the parser already enforced that that matches ORDER BY.
2845 : : *
2846 : : * Note: we return a fresh List, but its elements are the same
2847 : : * SortGroupClauses appearing in parse->groupClause. This is important
2848 : : * because later processing may modify the processed_groupClause list.
2849 : : *
2850 : : * For grouping sets, the order of items is instead forced to agree with that
2851 : : * of the grouping set (and items not in the grouping set are skipped). The
2852 : : * work of sorting the order of grouping set elements to match the ORDER BY if
2853 : : * possible is done elsewhere.
2854 : : */
2855 : : static List *
508 akorotkov@postgresql 2856 : 4076 : preprocess_groupclause(PlannerInfo *root, List *force)
2857 : : {
6295 tgl@sss.pgh.pa.us 2858 : 4076 : Query *parse = root->parse;
3817 andres@anarazel.de 2859 : 4076 : List *new_groupclause = NIL;
2860 : : ListCell *sl;
2861 : : ListCell *gl;
2862 : :
2863 : : /* For grouping sets, we need to force the ordering */
508 akorotkov@postgresql 2864 [ + + ]: 4076 : if (force)
2865 : : {
2866 [ + - + + : 5186 : foreach(sl, force)
+ + ]
2867 : : {
2868 : 3065 : Index ref = lfirst_int(sl);
2869 : 3065 : SortGroupClause *cl = get_sortgroupref_clause(ref, parse->groupClause);
2870 : :
2871 : 3065 : new_groupclause = lappend(new_groupclause, cl);
2872 : : }
2873 : :
2874 : 2121 : return new_groupclause;
2875 : : }
2876 : :
2877 : : /* If no ORDER BY, nothing useful to do here */
2878 [ + + ]: 1955 : if (parse->sortClause == NIL)
2879 : 1054 : return list_copy(parse->groupClause);
2880 : :
2881 : : /*
2882 : : * Scan the ORDER BY clause and construct a list of matching GROUP BY
2883 : : * items, but only as far as we can make a matching prefix.
2884 : : *
2885 : : * This code assumes that the sortClause contains no duplicate items.
2886 : : */
2887 [ + - + + : 1755 : foreach(sl, parse->sortClause)
+ + ]
2888 : : {
2889 : 1177 : SortGroupClause *sc = lfirst_node(SortGroupClause, sl);
2890 : :
2891 [ + - + + : 1725 : foreach(gl, parse->groupClause)
+ + ]
2892 : : {
2893 : 1402 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2894 : :
2895 [ + + ]: 1402 : if (equal(gc, sc))
2896 : : {
2897 : 854 : new_groupclause = lappend(new_groupclause, gc);
2898 : 854 : break;
2899 : : }
2900 : : }
2901 [ + + ]: 1177 : if (gl == NULL)
2902 : 323 : break; /* no match, so stop scanning */
2903 : : }
2904 : :
2905 : :
2906 : : /* If no match at all, no point in reordering GROUP BY */
2907 [ + + ]: 901 : if (new_groupclause == NIL)
2908 : 149 : return list_copy(parse->groupClause);
2909 : :
2910 : : /*
2911 : : * Add any remaining GROUP BY items to the new list. We don't require a
2912 : : * complete match, because even partial match allows ORDER BY to be
2913 : : * implemented using incremental sort. Also, give up if there are any
2914 : : * non-sortable GROUP BY items, since then there's no hope anyway.
2915 : : */
2916 [ + - + + : 1689 : foreach(gl, parse->groupClause)
+ + ]
2917 : : {
2918 : 937 : SortGroupClause *gc = lfirst_node(SortGroupClause, gl);
2919 : :
2920 [ + + ]: 937 : if (list_member_ptr(new_groupclause, gc))
2921 : 854 : continue; /* it matched an ORDER BY item */
2922 [ - + ]: 83 : if (!OidIsValid(gc->sortop)) /* give up, GROUP BY can't be sorted */
508 akorotkov@postgresql 2923 :UBC 0 : return list_copy(parse->groupClause);
508 akorotkov@postgresql 2924 :CBC 83 : new_groupclause = lappend(new_groupclause, gc);
2925 : : }
2926 : :
2927 : : /* Success --- install the rearranged GROUP BY list */
2928 [ - + ]: 752 : Assert(list_length(parse->groupClause) == list_length(new_groupclause));
3817 andres@anarazel.de 2929 : 752 : return new_groupclause;
2930 : : }
2931 : :
2932 : : /*
2933 : : * Extract lists of grouping sets that can be implemented using a single
2934 : : * rollup-type aggregate pass each. Returns a list of lists of grouping sets.
2935 : : *
2936 : : * Input must be sorted with smallest sets first. Result has each sublist
2937 : : * sorted with smallest sets first.
2938 : : *
2939 : : * We want to produce the absolute minimum possible number of lists here to
2940 : : * avoid excess sorts. Fortunately, there is an algorithm for this; the problem
2941 : : * of finding the minimal partition of a partially-ordered set into chains
2942 : : * (which is what we need, taking the list of grouping sets as a poset ordered
2943 : : * by set inclusion) can be mapped to the problem of finding the maximum
2944 : : * cardinality matching on a bipartite graph, which is solvable in polynomial
2945 : : * time with a worst case of no worse than O(n^2.5) and usually much
2946 : : * better. Since our N is at most 4096, we don't need to consider fallbacks to
2947 : : * heuristic or approximate methods. (Planning time for a 12-d cube is under
2948 : : * half a second on my modest system even with optimization off and assertions
2949 : : * on.)
2950 : : */
2951 : : static List *
2952 : 463 : extract_rollup_sets(List *groupingSets)
2953 : : {
2954 : 463 : int num_sets_raw = list_length(groupingSets);
2955 : 463 : int num_empty = 0;
3810 bruce@momjian.us 2956 : 463 : int num_sets = 0; /* distinct sets */
3817 andres@anarazel.de 2957 : 463 : int num_chains = 0;
2958 : 463 : List *result = NIL;
2959 : : List **results;
2960 : : List **orig_sets;
2961 : : Bitmapset **set_masks;
2962 : : int *chains;
2963 : : short **adjacency;
2964 : : short *adjacency_buf;
2965 : : BipartiteMatchState *state;
2966 : : int i;
2967 : : int j;
2968 : : int j_size;
2969 : 463 : ListCell *lc1 = list_head(groupingSets);
2970 : : ListCell *lc;
2971 : :
2972 : : /*
2973 : : * Start by stripping out empty sets. The algorithm doesn't require this,
2974 : : * but the planner currently needs all empty sets to be returned in the
2975 : : * first list, so we strip them here and add them back after.
2976 : : */
2977 [ + + + + ]: 786 : while (lc1 && lfirst(lc1) == NIL)
2978 : : {
2979 : 323 : ++num_empty;
2296 tgl@sss.pgh.pa.us 2980 : 323 : lc1 = lnext(groupingSets, lc1);
2981 : : }
2982 : :
2983 : : /* bail out now if it turns out that all we had were empty sets. */
3817 andres@anarazel.de 2984 [ + + ]: 463 : if (!lc1)
2985 : 33 : return list_make1(groupingSets);
2986 : :
2987 : : /*----------
2988 : : * We don't strictly need to remove duplicate sets here, but if we don't,
2989 : : * they tend to become scattered through the result, which is a bit
2990 : : * confusing (and irritating if we ever decide to optimize them out).
2991 : : * So we remove them here and add them back after.
2992 : : *
2993 : : * For each non-duplicate set, we fill in the following:
2994 : : *
2995 : : * orig_sets[i] = list of the original set lists
2996 : : * set_masks[i] = bitmapset for testing inclusion
2997 : : * adjacency[i] = array [n, v1, v2, ... vn] of adjacency indices
2998 : : *
2999 : : * chains[i] will be the result group this set is assigned to.
3000 : : *
3001 : : * We index all of these from 1 rather than 0 because it is convenient
3002 : : * to leave 0 free for the NIL node in the graph algorithm.
3003 : : *----------
3004 : : */
3810 bruce@momjian.us 3005 : 430 : orig_sets = palloc0((num_sets_raw + 1) * sizeof(List *));
3817 andres@anarazel.de 3006 : 430 : set_masks = palloc0((num_sets_raw + 1) * sizeof(Bitmapset *));
3007 : 430 : adjacency = palloc0((num_sets_raw + 1) * sizeof(short *));
3008 : 430 : adjacency_buf = palloc((num_sets_raw + 1) * sizeof(short));
3009 : :
3010 : 430 : j_size = 0;
3011 : 430 : j = 0;
3012 : 430 : i = 1;
3013 : :
2296 tgl@sss.pgh.pa.us 3014 [ + - + + : 1522 : for_each_cell(lc, groupingSets, lc1)
+ + ]
3015 : : {
2974 3016 : 1092 : List *candidate = (List *) lfirst(lc);
3817 andres@anarazel.de 3017 : 1092 : Bitmapset *candidate_set = NULL;
3018 : : ListCell *lc2;
3019 : 1092 : int dup_of = 0;
3020 : :
3021 [ + - + + : 2637 : foreach(lc2, candidate)
+ + ]
3022 : : {
3023 : 1545 : candidate_set = bms_add_member(candidate_set, lfirst_int(lc2));
3024 : : }
3025 : :
3026 : : /* we can only be a dup if we're the same length as a previous set */
3027 [ + + ]: 1092 : if (j_size == list_length(candidate))
3028 : : {
3029 : : int k;
3030 : :
3031 [ + + ]: 964 : for (k = j; k < i; ++k)
3032 : : {
3033 [ + + ]: 624 : if (bms_equal(set_masks[k], candidate_set))
3034 : : {
3035 : 79 : dup_of = k;
3036 : 79 : break;
3037 : : }
3038 : : }
3039 : : }
3040 [ + - ]: 673 : else if (j_size < list_length(candidate))
3041 : : {
3042 : 673 : j_size = list_length(candidate);
3043 : 673 : j = i;
3044 : : }
3045 : :
3046 [ + + ]: 1092 : if (dup_of > 0)
3047 : : {
3048 : 79 : orig_sets[dup_of] = lappend(orig_sets[dup_of], candidate);
3049 : 79 : bms_free(candidate_set);
3050 : : }
3051 : : else
3052 : : {
3053 : : int k;
3810 bruce@momjian.us 3054 : 1013 : int n_adj = 0;
3055 : :
3817 andres@anarazel.de 3056 : 1013 : orig_sets[i] = list_make1(candidate);
3057 : 1013 : set_masks[i] = candidate_set;
3058 : :
3059 : : /* fill in adjacency list; no need to compare equal-size sets */
3060 : :
3061 [ + + ]: 1655 : for (k = j - 1; k > 0; --k)
3062 : : {
3063 [ + + ]: 642 : if (bms_is_subset(set_masks[k], candidate_set))
3064 : 561 : adjacency_buf[++n_adj] = k;
3065 : : }
3066 : :
3067 [ + + ]: 1013 : if (n_adj > 0)
3068 : : {
3069 : 305 : adjacency_buf[0] = n_adj;
3070 : 305 : adjacency[i] = palloc((n_adj + 1) * sizeof(short));
3071 : 305 : memcpy(adjacency[i], adjacency_buf, (n_adj + 1) * sizeof(short));
3072 : : }
3073 : : else
3074 : 708 : adjacency[i] = NULL;
3075 : :
3076 : 1013 : ++i;
3077 : : }
3078 : : }
3079 : :
3080 : 430 : num_sets = i - 1;
3081 : :
3082 : : /*
3083 : : * Apply the graph matching algorithm to do the work.
3084 : : */
3085 : 430 : state = BipartiteMatch(num_sets, num_sets, adjacency);
3086 : :
3087 : : /*
3088 : : * Now, the state->pair* fields have the info we need to assign sets to
3089 : : * chains. Two sets (u,v) belong to the same chain if pair_uv[u] = v or
3090 : : * pair_vu[v] = u (both will be true, but we check both so that we can do
3091 : : * it in one pass)
3092 : : */
3093 : 430 : chains = palloc0((num_sets + 1) * sizeof(int));
3094 : :
3095 [ + + ]: 1443 : for (i = 1; i <= num_sets; ++i)
3096 : : {
3810 bruce@momjian.us 3097 : 1013 : int u = state->pair_vu[i];
3098 : 1013 : int v = state->pair_uv[i];
3099 : :
3817 andres@anarazel.de 3100 [ + + - + ]: 1013 : if (u > 0 && u < i)
3817 andres@anarazel.de 3101 :UBC 0 : chains[i] = chains[u];
3817 andres@anarazel.de 3102 [ + + + - ]:CBC 1013 : else if (v > 0 && v < i)
3103 : 291 : chains[i] = chains[v];
3104 : : else
3105 : 722 : chains[i] = ++num_chains;
3106 : : }
3107 : :
3108 : : /* build result lists. */
3810 bruce@momjian.us 3109 : 430 : results = palloc0((num_chains + 1) * sizeof(List *));
3110 : :
3817 andres@anarazel.de 3111 [ + + ]: 1443 : for (i = 1; i <= num_sets; ++i)
3112 : : {
3810 bruce@momjian.us 3113 : 1013 : int c = chains[i];
3114 : :
3817 andres@anarazel.de 3115 [ - + ]: 1013 : Assert(c > 0);
3116 : :
3117 : 1013 : results[c] = list_concat(results[c], orig_sets[i]);
3118 : : }
3119 : :
3120 : : /* push any empty sets back on the first list. */
3121 [ + + ]: 696 : while (num_empty-- > 0)
3122 : 266 : results[1] = lcons(NIL, results[1]);
3123 : :
3124 : : /* make result list */
3125 [ + + ]: 1152 : for (i = 1; i <= num_chains; ++i)
3126 : 722 : result = lappend(result, results[i]);
3127 : :
3128 : : /*
3129 : : * Free all the things.
3130 : : *
3131 : : * (This is over-fussy for small sets but for large sets we could have
3132 : : * tied up a nontrivial amount of memory.)
3133 : : */
3134 : 430 : BipartiteMatchFree(state);
3135 : 430 : pfree(results);
3136 : 430 : pfree(chains);
3137 [ + + ]: 1443 : for (i = 1; i <= num_sets; ++i)
3138 [ + + ]: 1013 : if (adjacency[i])
3139 : 305 : pfree(adjacency[i]);
3140 : 430 : pfree(adjacency);
3141 : 430 : pfree(adjacency_buf);
3142 : 430 : pfree(orig_sets);
3143 [ + + ]: 1443 : for (i = 1; i <= num_sets; ++i)
3144 : 1013 : bms_free(set_masks[i]);
3145 : 430 : pfree(set_masks);
3146 : :
3147 : 430 : return result;
3148 : : }
3149 : :
3150 : : /*
3151 : : * Reorder the elements of a list of grouping sets such that they have correct
3152 : : * prefix relationships. Also inserts the GroupingSetData annotations.
3153 : : *
3154 : : * The input must be ordered with smallest sets first; the result is returned
3155 : : * with largest sets first. Note that the result shares no list substructure
3156 : : * with the input, so it's safe for the caller to modify it later.
3157 : : *
3158 : : * If we're passed in a sortclause, we follow its order of columns to the
3159 : : * extent possible, to minimize the chance that we add unnecessary sorts.
3160 : : * (We're trying here to ensure that GROUPING SETS ((a,b,c),(c)) ORDER BY c,b,a
3161 : : * gets implemented in one pass.)
3162 : : */
3163 : : static List *
1133 pg@bowt.ie 3164 : 755 : reorder_grouping_sets(List *groupingSets, List *sortclause)
3165 : : {
3166 : : ListCell *lc;
3817 andres@anarazel.de 3167 : 755 : List *previous = NIL;
3168 : 755 : List *result = NIL;
3169 : :
1133 pg@bowt.ie 3170 [ + - + + : 2170 : foreach(lc, groupingSets)
+ + ]
3171 : : {
2974 tgl@sss.pgh.pa.us 3172 : 1415 : List *candidate = (List *) lfirst(lc);
3810 bruce@momjian.us 3173 : 1415 : List *new_elems = list_difference_int(candidate, previous);
3136 rhodiumtoad@postgres 3174 : 1415 : GroupingSetData *gs = makeNode(GroupingSetData);
3175 : :
2311 3176 [ + + + + ]: 1497 : while (list_length(sortclause) > list_length(previous) &&
3177 : : new_elems != NIL)
3178 : : {
3179 : 136 : SortGroupClause *sc = list_nth(sortclause, list_length(previous));
3180 : 136 : int ref = sc->tleSortGroupRef;
3181 : :
3182 [ + + ]: 136 : if (list_member_int(new_elems, ref))
3183 : : {
3184 : 82 : previous = lappend_int(previous, ref);
3185 : 82 : new_elems = list_delete_int(new_elems, ref);
3186 : : }
3187 : : else
3188 : : {
3189 : : /* diverged from the sortclause; give up on it */
3190 : 54 : sortclause = NIL;
3191 : 54 : break;
3192 : : }
3193 : : }
3194 : :
3195 : 1415 : previous = list_concat(previous, new_elems);
3196 : :
3136 3197 : 1415 : gs->set = list_copy(previous);
3198 : 1415 : result = lcons(gs, result);
3199 : : }
3200 : :
3817 andres@anarazel.de 3201 : 755 : list_free(previous);
3202 : :
3203 : 755 : return result;
3204 : : }
3205 : :
3206 : : /*
3207 : : * has_volatile_pathkey
3208 : : * Returns true if any PathKey in 'keys' has an EquivalenceClass
3209 : : * containing a volatile function. Otherwise returns false.
3210 : : */
3211 : : static bool
1014 drowley@postgresql.o 3212 : 1421 : has_volatile_pathkey(List *keys)
3213 : : {
3214 : : ListCell *lc;
3215 : :
3216 [ + + + + : 2914 : foreach(lc, keys)
+ + ]
3217 : : {
3218 : 1502 : PathKey *pathkey = lfirst_node(PathKey, lc);
3219 : :
3220 [ + + ]: 1502 : if (pathkey->pk_eclass->ec_has_volatile)
3221 : 9 : return true;
3222 : : }
3223 : :
3224 : 1412 : return false;
3225 : : }
3226 : :
3227 : : /*
3228 : : * adjust_group_pathkeys_for_groupagg
3229 : : * Add pathkeys to root->group_pathkeys to reflect the best set of
3230 : : * pre-ordered input for ordered aggregates.
3231 : : *
3232 : : * We define "best" as the pathkeys that suit the largest number of
3233 : : * aggregate functions. We find these by looking at the first ORDER BY /
3234 : : * DISTINCT aggregate and take the pathkeys for that before searching for
3235 : : * other aggregates that require the same or a more strict variation of the
3236 : : * same pathkeys. We then repeat that process for any remaining aggregates
3237 : : * with different pathkeys and if we find another set of pathkeys that suits a
3238 : : * larger number of aggregates then we select those pathkeys instead.
3239 : : *
3240 : : * When the best pathkeys are found we also mark each Aggref that can use
3241 : : * those pathkeys as aggpresorted = true.
3242 : : *
3243 : : * Note: When an aggregate function's ORDER BY / DISTINCT clause contains any
3244 : : * volatile functions, we never make use of these pathkeys. We want to ensure
3245 : : * that sorts using volatile functions are done independently in each Aggref
3246 : : * rather than once at the query level. If we were to allow this then Aggrefs
3247 : : * with compatible sort orders would all transition their rows in the same
3248 : : * order if those pathkeys were deemed to be the best pathkeys to sort on.
3249 : : * Whereas, if some other set of Aggref's pathkeys happened to be deemed
3250 : : * better pathkeys to sort on, then the volatile function Aggrefs would be
3251 : : * left to perform their sorts individually. To avoid this inconsistent
3252 : : * behavior which could make Aggref results depend on what other Aggrefs the
3253 : : * query contains, we always force Aggrefs with volatile functions to perform
3254 : : * their own sorts.
3255 : : */
3256 : : static void
1013 tgl@sss.pgh.pa.us 3257 : 1223 : adjust_group_pathkeys_for_groupagg(PlannerInfo *root)
3258 : : {
3259 : 1223 : List *grouppathkeys = root->group_pathkeys;
3260 : : List *bestpathkeys;
3261 : : Bitmapset *bestaggs;
3262 : : Bitmapset *unprocessed_aggs;
3263 : : ListCell *lc;
3264 : : int i;
3265 : :
3266 : : /* Shouldn't be here if there are grouping sets */
3267 [ - + ]: 1223 : Assert(root->parse->groupingSets == NIL);
3268 : : /* Shouldn't be here unless there are some ordered aggregates */
3269 [ - + ]: 1223 : Assert(root->numOrderedAggs > 0);
3270 : :
3271 : : /* Do nothing if disabled */
3272 [ + + ]: 1223 : if (!enable_presorted_aggregate)
3273 : 3 : return;
3274 : :
3275 : : /*
3276 : : * Make a first pass over all AggInfos to collect a Bitmapset containing
3277 : : * the indexes of all AggInfos to be processed below.
3278 : : */
1182 drowley@postgresql.o 3279 : 1220 : unprocessed_aggs = NULL;
3280 [ + - + + : 2782 : foreach(lc, root->agginfos)
+ + ]
3281 : : {
3282 : 1562 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
3283 : 1562 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3284 : :
3285 [ + + ]: 1562 : if (AGGKIND_IS_ORDERED_SET(aggref->aggkind))
3286 : 132 : continue;
3287 : :
3288 : : /* Skip unless there's a DISTINCT or ORDER BY clause */
190 3289 [ + + + + ]: 1430 : if (aggref->aggdistinct == NIL && aggref->aggorder == NIL)
3290 : 150 : continue;
3291 : :
3292 : : /* Additional safety checks are needed if there's a FILTER clause */
3293 [ + + ]: 1280 : if (aggref->aggfilter != NULL)
3294 : : {
3295 : : ListCell *lc2;
3296 : 27 : bool allow_presort = true;
3297 : :
3298 : : /*
3299 : : * When the Aggref has a FILTER clause, it's possible that the
3300 : : * filter removes rows that cannot be sorted because the
3301 : : * expression to sort by results in an error during its
3302 : : * evaluation. This is a problem for presorting as that happens
3303 : : * before the FILTER, whereas without presorting, the Aggregate
3304 : : * node will apply the FILTER *before* sorting. So that we never
3305 : : * try to sort anything that might error, here we aim to skip over
3306 : : * any Aggrefs with arguments with expressions which, when
3307 : : * evaluated, could cause an ERROR. Vars and Consts are ok. There
3308 : : * may be more cases that should be allowed, but more thought
3309 : : * needs to be given. Err on the side of caution.
3310 : : */
3311 [ + - + + : 51 : foreach(lc2, aggref->args)
+ + ]
3312 : : {
3313 : 36 : TargetEntry *tle = (TargetEntry *) lfirst(lc2);
3314 : 36 : Expr *expr = tle->expr;
3315 : :
3316 [ + + ]: 42 : while (IsA(expr, RelabelType))
3317 : 6 : expr = (Expr *) (castNode(RelabelType, expr))->arg;
3318 : :
3319 : : /* Common case, Vars and Consts are ok */
3320 [ + + + + ]: 36 : if (IsA(expr, Var) || IsA(expr, Const))
3321 : 24 : continue;
3322 : :
3323 : : /* Unsupported. Don't try to presort for this Aggref */
3324 : 12 : allow_presort = false;
3325 : 12 : break;
3326 : : }
3327 : :
3328 : : /* Skip unsupported Aggrefs */
3329 [ + + ]: 27 : if (!allow_presort)
3330 : 12 : continue;
3331 : : }
3332 : :
3333 : 1268 : unprocessed_aggs = bms_add_member(unprocessed_aggs,
3334 : : foreach_current_index(lc));
3335 : : }
3336 : :
3337 : : /*
3338 : : * Now process all the unprocessed_aggs to find the best set of pathkeys
3339 : : * for the given set of aggregates.
3340 : : *
3341 : : * On the first outer loop here 'bestaggs' will be empty. We'll populate
3342 : : * this during the first loop using the pathkeys for the very first
3343 : : * AggInfo then taking any stronger pathkeys from any other AggInfos with
3344 : : * a more strict set of compatible pathkeys. Once the outer loop is
3345 : : * complete, we mark off all the aggregates with compatible pathkeys then
3346 : : * remove those from the unprocessed_aggs and repeat the process to try to
3347 : : * find another set of pathkeys that are suitable for a larger number of
3348 : : * aggregates. The outer loop will stop when there are not enough
3349 : : * unprocessed aggregates for it to be possible to find a set of pathkeys
3350 : : * to suit a larger number of aggregates.
3351 : : */
1182 3352 : 1220 : bestpathkeys = NIL;
3353 : 1220 : bestaggs = NULL;
3354 [ + + ]: 2407 : while (bms_num_members(unprocessed_aggs) > bms_num_members(bestaggs))
3355 : : {
3356 : 1187 : Bitmapset *aggindexes = NULL;
3357 : 1187 : List *currpathkeys = NIL;
3358 : :
3359 : 1187 : i = -1;
3360 [ + + ]: 2608 : while ((i = bms_next_member(unprocessed_aggs, i)) >= 0)
3361 : : {
3362 : 1421 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3363 : 1421 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
3364 : : List *sortlist;
3365 : : List *pathkeys;
3366 : :
3367 [ + + ]: 1421 : if (aggref->aggdistinct != NIL)
3368 : 359 : sortlist = aggref->aggdistinct;
3369 : : else
3370 : 1062 : sortlist = aggref->aggorder;
3371 : :
1014 3372 : 1421 : pathkeys = make_pathkeys_for_sortclauses(root, sortlist,
3373 : : aggref->args);
3374 : :
3375 : : /*
3376 : : * Ignore Aggrefs which have volatile functions in their ORDER BY
3377 : : * or DISTINCT clause.
3378 : : */
3379 [ + + ]: 1421 : if (has_volatile_pathkey(pathkeys))
3380 : : {
3381 : 9 : unprocessed_aggs = bms_del_member(unprocessed_aggs, i);
3382 : 9 : continue;
3383 : : }
3384 : :
3385 : : /*
3386 : : * When not set yet, take the pathkeys from the first unprocessed
3387 : : * aggregate.
3388 : : */
1182 3389 [ + + ]: 1412 : if (currpathkeys == NIL)
3390 : : {
1014 3391 : 1184 : currpathkeys = pathkeys;
3392 : :
3393 : : /* include the GROUP BY pathkeys, if they exist */
1182 3394 [ + + ]: 1184 : if (grouppathkeys != NIL)
3395 : 138 : currpathkeys = append_pathkeys(list_copy(grouppathkeys),
3396 : : currpathkeys);
3397 : :
3398 : : /* record that we found pathkeys for this aggregate */
3399 : 1184 : aggindexes = bms_add_member(aggindexes, i);
3400 : : }
3401 : : else
3402 : : {
3403 : : /* now look for a stronger set of matching pathkeys */
3404 : :
3405 : : /* include the GROUP BY pathkeys, if they exist */
3406 [ + + ]: 228 : if (grouppathkeys != NIL)
3407 : 144 : pathkeys = append_pathkeys(list_copy(grouppathkeys),
3408 : : pathkeys);
3409 : :
3410 : : /* are 'pathkeys' compatible or better than 'currpathkeys'? */
3411 [ + + + - ]: 228 : switch (compare_pathkeys(currpathkeys, pathkeys))
3412 : : {
3413 : 6 : case PATHKEYS_BETTER2:
3414 : : /* 'pathkeys' are stronger, use these ones instead */
3415 : 6 : currpathkeys = pathkeys;
3416 : : /* FALLTHROUGH */
3417 : :
3418 : 33 : case PATHKEYS_BETTER1:
3419 : : /* 'pathkeys' are less strict */
3420 : : /* FALLTHROUGH */
3421 : :
3422 : : case PATHKEYS_EQUAL:
3423 : : /* mark this aggregate as covered by 'currpathkeys' */
3424 : 33 : aggindexes = bms_add_member(aggindexes, i);
3425 : 33 : break;
3426 : :
3427 : 195 : case PATHKEYS_DIFFERENT:
3428 : 195 : break;
3429 : : }
3430 : : }
3431 : : }
3432 : :
3433 : : /* remove the aggregates that we've just processed */
3434 : 1187 : unprocessed_aggs = bms_del_members(unprocessed_aggs, aggindexes);
3435 : :
3436 : : /*
3437 : : * If this pass included more aggregates than the previous best then
3438 : : * use these ones as the best set.
3439 : : */
3440 [ + + ]: 1187 : if (bms_num_members(aggindexes) > bms_num_members(bestaggs))
3441 : : {
3442 : 1133 : bestaggs = aggindexes;
3443 : 1133 : bestpathkeys = currpathkeys;
3444 : : }
3445 : : }
3446 : :
3447 : : /*
3448 : : * If we found any ordered aggregates, update root->group_pathkeys to add
3449 : : * the best set of aggregate pathkeys. Note that bestpathkeys includes
3450 : : * the original GROUP BY pathkeys already.
3451 : : */
1013 tgl@sss.pgh.pa.us 3452 [ + + ]: 1220 : if (bestpathkeys != NIL)
3453 : 1103 : root->group_pathkeys = bestpathkeys;
3454 : :
3455 : : /*
3456 : : * Now that we've found the best set of aggregates we can set the
3457 : : * presorted flag to indicate to the executor that it needn't bother
3458 : : * performing a sort for these Aggrefs. We're able to do this now as
3459 : : * there's no chance of a Hash Aggregate plan as create_grouping_paths
3460 : : * will not mark the GROUP BY as GROUPING_CAN_USE_HASH due to the presence
3461 : : * of ordered aggregates.
3462 : : */
1182 drowley@postgresql.o 3463 : 1220 : i = -1;
3464 [ + + ]: 2371 : while ((i = bms_next_member(bestaggs, i)) >= 0)
3465 : : {
3466 : 1151 : AggInfo *agginfo = list_nth_node(AggInfo, root->agginfos, i);
3467 : :
3468 [ + - + + : 2311 : foreach(lc, agginfo->aggrefs)
+ + ]
3469 : : {
3470 : 1160 : Aggref *aggref = lfirst_node(Aggref, lc);
3471 : :
3472 : 1160 : aggref->aggpresorted = true;
3473 : : }
3474 : : }
3475 : : }
3476 : :
3477 : : /*
3478 : : * Compute query_pathkeys and other pathkeys during plan generation
3479 : : */
3480 : : static void
4564 tgl@sss.pgh.pa.us 3481 : 256496 : standard_qp_callback(PlannerInfo *root, void *extra)
3482 : : {
3483 : 256496 : Query *parse = root->parse;
3484 : 256496 : standard_qp_extra *qp_extra = (standard_qp_extra *) extra;
2406 3485 : 256496 : List *tlist = root->processed_tlist;
4564 3486 : 256496 : List *activeWindows = qp_extra->activeWindows;
3487 : :
3488 : : /*
3489 : : * Calculate pathkeys that represent grouping/ordering and/or ordered
3490 : : * aggregate requirements.
3491 : : */
1013 3492 [ + + ]: 256496 : if (qp_extra->gset_data)
3493 : : {
3494 : : /*
3495 : : * With grouping sets, just use the first RollupData's groupClause. We
3496 : : * don't make any effort to optimize grouping clauses when there are
3497 : : * grouping sets, nor can we combine aggregate ordering keys with
3498 : : * grouping.
3499 : : */
3500 : 466 : List *rollups = qp_extra->gset_data->rollups;
3501 [ + + ]: 466 : List *groupClause = (rollups ? linitial_node(RollupData, rollups)->groupClause : NIL);
3502 : :
3503 [ + - ]: 466 : if (grouping_is_sortable(groupClause))
3504 : : {
3505 : : bool sortable;
3506 : :
3507 : : /*
3508 : : * The groupClause is logically below the grouping step. So if
3509 : : * there is an RTE entry for the grouping step, we need to remove
3510 : : * its RT index from the sort expressions before we make PathKeys
3511 : : * for them.
3512 : : */
412 rguo@postgresql.org 3513 : 466 : root->group_pathkeys =
3514 : 466 : make_pathkeys_for_sortclauses_extended(root,
3515 : : &groupClause,
3516 : : tlist,
3517 : : false,
3518 : 466 : parse->hasGroupRTE,
3519 : : &sortable,
3520 : : false);
3521 [ - + ]: 466 : Assert(sortable);
1013 tgl@sss.pgh.pa.us 3522 : 466 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3523 : : }
3524 : : else
3525 : : {
1013 tgl@sss.pgh.pa.us 3526 :UBC 0 : root->group_pathkeys = NIL;
3527 : 0 : root->num_groupby_pathkeys = 0;
3528 : : }
3529 : : }
1013 tgl@sss.pgh.pa.us 3530 [ + + + + ]:CBC 256030 : else if (parse->groupClause || root->numOrderedAggs > 0)
3531 : 3056 : {
3532 : : /*
3533 : : * With a plain GROUP BY list, we can remove any grouping items that
3534 : : * are proven redundant by EquivalenceClass processing. For example,
3535 : : * we can remove y given "WHERE x = y GROUP BY x, y". These aren't
3536 : : * especially common cases, but they're nearly free to detect. Note
3537 : : * that we remove redundant items from processed_groupClause but not
3538 : : * the original parse->groupClause.
3539 : : */
3540 : : bool sortable;
3541 : :
3542 : : /*
3543 : : * Convert group clauses into pathkeys. Set the ec_sortref field of
3544 : : * EquivalenceClass'es if it's not set yet.
3545 : : */
3546 : 3056 : root->group_pathkeys =
3547 : 3056 : make_pathkeys_for_sortclauses_extended(root,
3548 : : &root->processed_groupClause,
3549 : : tlist,
3550 : : true,
3551 : : false,
3552 : : &sortable,
3553 : : true);
3554 [ - + ]: 3056 : if (!sortable)
3555 : : {
3556 : : /* Can't sort; no point in considering aggregate ordering either */
1013 tgl@sss.pgh.pa.us 3557 :UBC 0 : root->group_pathkeys = NIL;
3558 : 0 : root->num_groupby_pathkeys = 0;
3559 : : }
3560 : : else
3561 : : {
1013 tgl@sss.pgh.pa.us 3562 :CBC 3056 : root->num_groupby_pathkeys = list_length(root->group_pathkeys);
3563 : : /* If we have ordered aggs, consider adding onto group_pathkeys */
3564 [ + + ]: 3056 : if (root->numOrderedAggs > 0)
3565 : 1223 : adjust_group_pathkeys_for_groupagg(root);
3566 : : }
3567 : : }
3568 : : else
3569 : : {
4564 3570 : 252974 : root->group_pathkeys = NIL;
1182 drowley@postgresql.o 3571 : 252974 : root->num_groupby_pathkeys = 0;
3572 : : }
3573 : :
3574 : : /* We consider only the first (bottom) window in pathkeys logic */
4564 tgl@sss.pgh.pa.us 3575 [ + + ]: 256496 : if (activeWindows != NIL)
3576 : : {
2974 3577 : 1282 : WindowClause *wc = linitial_node(WindowClause, activeWindows);
3578 : :
4564 3579 : 1282 : root->window_pathkeys = make_pathkeys_for_window(root,
3580 : : wc,
3581 : : tlist);
3582 : : }
3583 : : else
3584 : 255214 : root->window_pathkeys = NIL;
3585 : :
3586 : : /*
3587 : : * As with GROUP BY, we can discard any DISTINCT items that are proven
3588 : : * redundant by EquivalenceClass processing. The non-redundant list is
3589 : : * kept in root->processed_distinctClause, leaving the original
3590 : : * parse->distinctClause alone.
3591 : : */
1013 3592 [ + + ]: 256496 : if (parse->distinctClause)
3593 : : {
3594 : : bool sortable;
3595 : :
3596 : : /* Make a copy since pathkey processing can modify the list */
3597 : 1479 : root->processed_distinctClause = list_copy(parse->distinctClause);
4564 3598 : 1479 : root->distinct_pathkeys =
1013 3599 : 1479 : make_pathkeys_for_sortclauses_extended(root,
3600 : : &root->processed_distinctClause,
3601 : : tlist,
3602 : : true,
3603 : : false,
3604 : : &sortable,
3605 : : false);
3606 [ + + ]: 1479 : if (!sortable)
3607 : 3 : root->distinct_pathkeys = NIL;
3608 : : }
3609 : : else
4564 3610 : 255017 : root->distinct_pathkeys = NIL;
3611 : :
3612 : 256496 : root->sort_pathkeys =
3613 : 256496 : make_pathkeys_for_sortclauses(root,
3614 : : parse->sortClause,
3615 : : tlist);
3616 : :
3617 : : /* setting setop_pathkeys might be useful to the union planner */
312 3618 [ + + ]: 256496 : if (qp_extra->setop != NULL)
3619 : : {
3620 : : List *groupClauses;
3621 : : bool sortable;
3622 : :
524 rhaas@postgresql.org 3623 : 6244 : groupClauses = generate_setop_child_grouplist(qp_extra->setop, tlist);
3624 : :
3625 : 6244 : root->setop_pathkeys =
3626 : 6244 : make_pathkeys_for_sortclauses_extended(root,
3627 : : &groupClauses,
3628 : : tlist,
3629 : : false,
3630 : : false,
3631 : : &sortable,
3632 : : false);
3633 [ + + ]: 6244 : if (!sortable)
3634 : 102 : root->setop_pathkeys = NIL;
3635 : : }
3636 : : else
3637 : 250252 : root->setop_pathkeys = NIL;
3638 : :
3639 : : /*
3640 : : * Figure out whether we want a sorted result from query_planner.
3641 : : *
3642 : : * If we have a sortable GROUP BY clause, then we want a result sorted
3643 : : * properly for grouping. Otherwise, if we have window functions to
3644 : : * evaluate, we try to sort for the first window. Otherwise, if there's a
3645 : : * sortable DISTINCT clause that's more rigorous than the ORDER BY clause,
3646 : : * we try to produce output that's sufficiently well sorted for the
3647 : : * DISTINCT. Otherwise, if there is an ORDER BY clause, we want to sort
3648 : : * by the ORDER BY clause. Otherwise, if we're a subquery being planned
3649 : : * for a set operation which can benefit from presorted results and have a
3650 : : * sortable targetlist, we want to sort by the target list.
3651 : : *
3652 : : * Note: if we have both ORDER BY and GROUP BY, and ORDER BY is a superset
3653 : : * of GROUP BY, it would be tempting to request sort by ORDER BY --- but
3654 : : * that might just leave us failing to exploit an available sort order at
3655 : : * all. Needs more thought. The choice for DISTINCT versus ORDER BY is
3656 : : * much easier, since we know that the parser ensured that one is a
3657 : : * superset of the other.
3658 : : */
4564 tgl@sss.pgh.pa.us 3659 [ + + ]: 256496 : if (root->group_pathkeys)
3660 : 3332 : root->query_pathkeys = root->group_pathkeys;
3661 [ + + ]: 253164 : else if (root->window_pathkeys)
3662 : 1049 : root->query_pathkeys = root->window_pathkeys;
3663 [ + + ]: 504230 : else if (list_length(root->distinct_pathkeys) >
3664 : 252115 : list_length(root->sort_pathkeys))
3665 : 1237 : root->query_pathkeys = root->distinct_pathkeys;
3666 [ + + ]: 250878 : else if (root->sort_pathkeys)
3667 : 33783 : root->query_pathkeys = root->sort_pathkeys;
524 rhaas@postgresql.org 3668 [ + + ]: 217095 : else if (root->setop_pathkeys != NIL)
3669 : 5546 : root->query_pathkeys = root->setop_pathkeys;
3670 : : else
4564 tgl@sss.pgh.pa.us 3671 : 211549 : root->query_pathkeys = NIL;
3672 : 256496 : }
3673 : :
3674 : : /*
3675 : : * Estimate number of groups produced by grouping clauses (1 if not grouping)
3676 : : *
3677 : : * path_rows: number of output rows from scan/join step
3678 : : * gd: grouping sets data including list of grouping sets and their clauses
3679 : : * target_list: target list containing group clause references
3680 : : *
3681 : : * If doing grouping sets, we also annotate the gsets data with the estimates
3682 : : * for each set and each individual rollup list, with a view to later
3683 : : * determining whether some combination of them could be hashed instead.
3684 : : */
3685 : : static double
3521 3686 : 22117 : get_number_of_groups(PlannerInfo *root,
3687 : : double path_rows,
3688 : : grouping_sets_data *gd,
3689 : : List *target_list)
3690 : : {
5738 3691 : 22117 : Query *parse = root->parse;
3692 : : double dNumGroups;
3693 : :
3521 3694 [ + + ]: 22117 : if (parse->groupClause)
3695 : : {
3696 : : List *groupExprs;
3697 : :
3698 [ + + ]: 5033 : if (parse->groupingSets)
3699 : : {
3700 : : /* Add up the estimates for each grouping set */
3701 : : ListCell *lc;
3702 : :
3085 bruce@momjian.us 3703 [ - + ]: 433 : Assert(gd); /* keep Coverity happy */
3704 : :
3521 tgl@sss.pgh.pa.us 3705 : 433 : dNumGroups = 0;
3706 : :
3136 rhodiumtoad@postgres 3707 [ + + + + : 1155 : foreach(lc, gd->rollups)
+ + ]
3708 : : {
2974 tgl@sss.pgh.pa.us 3709 : 722 : RollupData *rollup = lfirst_node(RollupData, lc);
3710 : : ListCell *lc2;
3711 : : ListCell *lc3;
3712 : :
3136 rhodiumtoad@postgres 3713 : 722 : groupExprs = get_sortgrouplist_exprs(rollup->groupClause,
3714 : : target_list);
3715 : :
3716 : 722 : rollup->numGroups = 0.0;
3717 : :
1118 drowley@postgresql.o 3718 [ + - + + : 2080 : forboth(lc2, rollup->gsets, lc3, rollup->gsets_data)
+ - + + +
+ + - +
+ ]
3719 : : {
3720 : 1358 : List *gset = (List *) lfirst(lc2);
3721 : 1358 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc3);
3136 rhodiumtoad@postgres 3722 : 1358 : double numGroups = estimate_num_groups(root,
3723 : : groupExprs,
3724 : : path_rows,
3725 : : &gset,
3726 : : NULL);
3727 : :
3728 : 1358 : gs->numGroups = numGroups;
3729 : 1358 : rollup->numGroups += numGroups;
3730 : : }
3731 : :
3732 : 722 : dNumGroups += rollup->numGroups;
3733 : : }
3734 : :
3735 [ + + ]: 433 : if (gd->hash_sets_idx)
3736 : : {
3737 : : ListCell *lc2;
3738 : :
3739 : 18 : gd->dNumHashGroups = 0;
3740 : :
3741 : 18 : groupExprs = get_sortgrouplist_exprs(parse->groupClause,
3742 : : target_list);
3743 : :
3744 [ + - + + : 39 : forboth(lc, gd->hash_sets_idx, lc2, gd->unsortable_sets)
+ - + + +
+ + - +
+ ]
3745 : : {
3746 : 21 : List *gset = (List *) lfirst(lc);
2974 tgl@sss.pgh.pa.us 3747 : 21 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc2);
3136 rhodiumtoad@postgres 3748 : 21 : double numGroups = estimate_num_groups(root,
3749 : : groupExprs,
3750 : : path_rows,
3751 : : &gset,
3752 : : NULL);
3753 : :
3754 : 21 : gs->numGroups = numGroups;
3755 : 21 : gd->dNumHashGroups += numGroups;
3756 : : }
3757 : :
3758 : 18 : dNumGroups += gd->dNumHashGroups;
3759 : : }
3760 : : }
3761 : : else
3762 : : {
3763 : : /* Plain GROUP BY -- estimate based on optimized groupClause */
1013 tgl@sss.pgh.pa.us 3764 : 4600 : groupExprs = get_sortgrouplist_exprs(root->processed_groupClause,
3765 : : target_list);
3766 : :
3521 3767 : 4600 : dNumGroups = estimate_num_groups(root, groupExprs, path_rows,
3768 : : NULL, NULL);
3769 : : }
3770 : : }
3771 [ + + ]: 17084 : else if (parse->groupingSets)
3772 : : {
3773 : : /* Empty grouping sets ... one result row for each one */
3774 : 30 : dNumGroups = list_length(parse->groupingSets);
3775 : : }
3776 [ - + - - ]: 17054 : else if (parse->hasAggs || root->hasHavingQual)
3777 : : {
3778 : : /* Plain aggregation, one result row */
3779 : 17054 : dNumGroups = 1;
3780 : : }
3781 : : else
3782 : : {
3783 : : /* Not grouping */
3521 tgl@sss.pgh.pa.us 3784 :UBC 0 : dNumGroups = 1;
3785 : : }
3786 : :
3521 tgl@sss.pgh.pa.us 3787 :CBC 22117 : return dNumGroups;
3788 : : }
3789 : :
3790 : : /*
3791 : : * create_grouping_paths
3792 : : *
3793 : : * Build a new upperrel containing Paths for grouping and/or aggregation.
3794 : : * Along the way, we also build an upperrel for Paths which are partially
3795 : : * grouped and/or aggregated. A partially grouped and/or aggregated path
3796 : : * needs a FinalizeAggregate node to complete the aggregation. Currently,
3797 : : * the only partially grouped paths we build are also partial paths; that
3798 : : * is, they need a Gather and then a FinalizeAggregate.
3799 : : *
3800 : : * input_rel: contains the source-data Paths
3801 : : * target: the pathtarget for the result Paths to compute
3802 : : * gd: grouping sets data including list of grouping sets and their clauses
3803 : : *
3804 : : * Note: all Paths in input_rel are expected to return the target computed
3805 : : * by make_group_input_target.
3806 : : */
3807 : : static RelOptInfo *
3808 : 18687 : create_grouping_paths(PlannerInfo *root,
3809 : : RelOptInfo *input_rel,
3810 : : PathTarget *target,
3811 : : bool target_parallel_safe,
3812 : : grouping_sets_data *gd)
3813 : : {
3814 : 18687 : Query *parse = root->parse;
3815 : : RelOptInfo *grouped_rel;
3816 : : RelOptInfo *partially_grouped_rel;
3817 : : AggClauseCosts agg_costs;
3818 : :
1798 heikki.linnakangas@i 3819 [ + - + - : 112122 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
3820 : 18687 : get_agg_clause_costs(root, AGGSPLIT_SIMPLE, &agg_costs);
3821 : :
3822 : : /*
3823 : : * Create grouping relation to hold fully aggregated grouping and/or
3824 : : * aggregation paths.
3825 : : */
2776 rhaas@postgresql.org 3826 : 18687 : grouped_rel = make_grouping_rel(root, input_rel, target,
3827 : : target_parallel_safe, parse->havingQual);
3828 : :
3829 : : /*
3830 : : * Create either paths for a degenerate grouping or paths for ordinary
3831 : : * grouping, as appropriate.
3832 : : */
2783 3833 [ + + ]: 18687 : if (is_degenerate_grouping(root))
2778 3834 : 15 : create_degenerate_grouping_paths(root, input_rel, grouped_rel);
3835 : : else
3836 : : {
3837 : 18672 : int flags = 0;
3838 : : GroupPathExtraData extra;
3839 : :
3840 : : /*
3841 : : * Determine whether it's possible to perform sort-based
3842 : : * implementations of grouping. (Note that if processed_groupClause
3843 : : * is empty, grouping_is_sortable() is trivially true, and all the
3844 : : * pathkeys_contained_in() tests will succeed too, so that we'll
3845 : : * consider every surviving input path.)
3846 : : *
3847 : : * If we have grouping sets, we might be able to sort some but not all
3848 : : * of them; in this case, we need can_sort to be true as long as we
3849 : : * must consider any sorted-input plan.
3850 : : */
3851 [ + + + + ]: 18672 : if ((gd && gd->rollups != NIL)
1013 tgl@sss.pgh.pa.us 3852 [ + + ]: 18212 : || grouping_is_sortable(root->processed_groupClause))
2778 rhaas@postgresql.org 3853 : 18669 : flags |= GROUPING_CAN_USE_SORT;
3854 : :
3855 : : /*
3856 : : * Determine whether we should consider hash-based implementations of
3857 : : * grouping.
3858 : : *
3859 : : * Hashed aggregation only applies if we're grouping. If we have
3860 : : * grouping sets, some groups might be hashable but others not; in
3861 : : * this case we set can_hash true as long as there is nothing globally
3862 : : * preventing us from hashing (and we should therefore consider plans
3863 : : * with hashes).
3864 : : *
3865 : : * Executor doesn't support hashed aggregation with DISTINCT or ORDER
3866 : : * BY aggregates. (Doing so would imply storing *all* the input
3867 : : * values in the hash table, and/or running many sorts in parallel,
3868 : : * either of which seems like a certain loser.) We similarly don't
3869 : : * support ordered-set aggregates in hashed aggregation, but that case
3870 : : * is also included in the numOrderedAggs count.
3871 : : *
3872 : : * Note: grouping_is_hashable() is much more expensive to check than
3873 : : * the other gating conditions, so we want to do it last.
3874 : : */
3875 [ + + ]: 18672 : if ((parse->groupClause != NIL &&
1798 heikki.linnakangas@i 3876 [ + + + + : 4636 : root->numOrderedAggs == 0 &&
+ + ]
1013 tgl@sss.pgh.pa.us 3877 : 2248 : (gd ? gd->any_hashable : grouping_is_hashable(root->processed_groupClause))))
2778 rhaas@postgresql.org 3878 : 2246 : flags |= GROUPING_CAN_USE_HASH;
3879 : :
3880 : : /*
3881 : : * Determine whether partial aggregation is possible.
3882 : : */
1798 heikki.linnakangas@i 3883 [ + + ]: 18672 : if (can_partial_agg(root))
2778 rhaas@postgresql.org 3884 : 16259 : flags |= GROUPING_CAN_PARTIAL_AGG;
3885 : :
2776 3886 : 18672 : extra.flags = flags;
3887 : 18672 : extra.target_parallel_safe = target_parallel_safe;
3888 : 18672 : extra.havingQual = parse->havingQual;
3889 : 18672 : extra.targetList = parse->targetList;
3890 : 18672 : extra.partial_costs_set = false;
3891 : :
3892 : : /*
3893 : : * Determine whether partitionwise aggregation is in theory possible.
3894 : : * It can be disabled by the user, and for now, we don't try to
3895 : : * support grouping sets. create_ordinary_grouping_paths() will check
3896 : : * additional conditions, such as whether input_rel is partitioned.
3897 : : */
3898 [ + + + + ]: 18672 : if (enable_partitionwise_aggregate && !parse->groupingSets)
3899 : 344 : extra.patype = PARTITIONWISE_AGGREGATE_FULL;
3900 : : else
3901 : 18328 : extra.patype = PARTITIONWISE_AGGREGATE_NONE;
3902 : :
2778 3903 : 18672 : create_ordinary_grouping_paths(root, input_rel, grouped_rel,
3904 : : &agg_costs, gd, &extra,
3905 : : &partially_grouped_rel);
3906 : : }
3907 : :
2783 3908 : 18684 : set_cheapest(grouped_rel);
3909 : 18684 : return grouped_rel;
3910 : : }
3911 : :
3912 : : /*
3913 : : * make_grouping_rel
3914 : : *
3915 : : * Create a new grouping rel and set basic properties.
3916 : : *
3917 : : * input_rel represents the underlying scan/join relation.
3918 : : * target is the output expected from the grouping relation.
3919 : : */
3920 : : static RelOptInfo *
2776 3921 : 19752 : make_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
3922 : : PathTarget *target, bool target_parallel_safe,
3923 : : Node *havingQual)
3924 : : {
3925 : : RelOptInfo *grouped_rel;
3926 : :
3927 [ + + + + : 19752 : if (IS_OTHER_REL(input_rel))
- + ]
3928 : : {
3929 : 1065 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG,
3930 : : input_rel->relids);
3931 : 1065 : grouped_rel->reloptkind = RELOPT_OTHER_UPPER_REL;
3932 : : }
3933 : : else
3934 : : {
3935 : : /*
3936 : : * By tradition, the relids set for the main grouping relation is
3937 : : * NULL. (This could be changed, but might require adjustments
3938 : : * elsewhere.)
3939 : : */
3940 : 18687 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
3941 : : }
3942 : :
3943 : : /* Set target. */
3944 : 19752 : grouped_rel->reltarget = target;
3945 : :
3946 : : /*
3947 : : * If the input relation is not parallel-safe, then the grouped relation
3948 : : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
3949 : : * target list and HAVING quals are parallel-safe.
3950 : : */
3951 [ + + + + : 33311 : if (input_rel->consider_parallel && target_parallel_safe &&
+ + ]
3952 : 13559 : is_parallel_safe(root, (Node *) havingQual))
3953 : 13550 : grouped_rel->consider_parallel = true;
3954 : :
3955 : : /*
3956 : : * If the input rel belongs to a single FDW, so does the grouped rel.
3957 : : */
3958 : 19752 : grouped_rel->serverid = input_rel->serverid;
3959 : 19752 : grouped_rel->userid = input_rel->userid;
3960 : 19752 : grouped_rel->useridiscurrent = input_rel->useridiscurrent;
3961 : 19752 : grouped_rel->fdwroutine = input_rel->fdwroutine;
3962 : :
3963 : 19752 : return grouped_rel;
3964 : : }
3965 : :
3966 : : /*
3967 : : * is_degenerate_grouping
3968 : : *
3969 : : * A degenerate grouping is one in which the query has a HAVING qual and/or
3970 : : * grouping sets, but no aggregates and no GROUP BY (which implies that the
3971 : : * grouping sets are all empty).
3972 : : */
3973 : : static bool
2783 3974 : 18687 : is_degenerate_grouping(PlannerInfo *root)
3975 : : {
3976 : 18687 : Query *parse = root->parse;
3977 : :
3978 [ + + ]: 18125 : return (root->hasHavingQual || parse->groupingSets) &&
3979 [ + + + + : 36812 : !parse->hasAggs && parse->groupClause == NIL;
+ + ]
3980 : : }
3981 : :
3982 : : /*
3983 : : * create_degenerate_grouping_paths
3984 : : *
3985 : : * When the grouping is degenerate (see is_degenerate_grouping), we are
3986 : : * supposed to emit either zero or one row for each grouping set depending on
3987 : : * whether HAVING succeeds. Furthermore, there cannot be any variables in
3988 : : * either HAVING or the targetlist, so we actually do not need the FROM table
3989 : : * at all! We can just throw away the plan-so-far and generate a Result node.
3990 : : * This is a sufficiently unusual corner case that it's not worth contorting
3991 : : * the structure of this module to avoid having to generate the earlier paths
3992 : : * in the first place.
3993 : : */
3994 : : static void
3995 : 15 : create_degenerate_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
3996 : : RelOptInfo *grouped_rel)
3997 : : {
3998 : 15 : Query *parse = root->parse;
3999 : : int nrows;
4000 : : Path *path;
4001 : :
4002 : 15 : nrows = list_length(parse->groupingSets);
4003 [ - + ]: 15 : if (nrows > 1)
4004 : : {
4005 : : /*
4006 : : * Doesn't seem worthwhile writing code to cons up a generate_series
4007 : : * or a values scan to emit multiple rows. Instead just make N clones
4008 : : * and append them. (With a volatile HAVING clause, this means you
4009 : : * might get between 0 and N output rows. Offhand I think that's
4010 : : * desired.)
4011 : : */
2783 rhaas@postgresql.org 4012 :UBC 0 : List *paths = NIL;
4013 : :
4014 [ # # ]: 0 : while (--nrows >= 0)
4015 : : {
4016 : : path = (Path *)
2464 tgl@sss.pgh.pa.us 4017 : 0 : create_group_result_path(root, grouped_rel,
4018 : 0 : grouped_rel->reltarget,
4019 : 0 : (List *) parse->havingQual);
2783 rhaas@postgresql.org 4020 : 0 : paths = lappend(paths, path);
4021 : : }
4022 : : path = (Path *)
2760 alvherre@alvh.no-ip. 4023 : 0 : create_append_path(root,
4024 : : grouped_rel,
4025 : : paths,
4026 : : NIL,
4027 : : NIL,
4028 : : NULL,
4029 : : 0,
4030 : : false,
4031 : : -1);
4032 : : }
4033 : : else
4034 : : {
4035 : : /* No grouping sets, or just one, so one output row */
4036 : : path = (Path *)
2464 tgl@sss.pgh.pa.us 4037 :CBC 15 : create_group_result_path(root, grouped_rel,
4038 : 15 : grouped_rel->reltarget,
4039 : 15 : (List *) parse->havingQual);
4040 : : }
4041 : :
2783 rhaas@postgresql.org 4042 : 15 : add_path(grouped_rel, path);
4043 : 15 : }
4044 : :
4045 : : /*
4046 : : * create_ordinary_grouping_paths
4047 : : *
4048 : : * Create grouping paths for the ordinary (that is, non-degenerate) case.
4049 : : *
4050 : : * We need to consider sorted and hashed aggregation in the same function,
4051 : : * because otherwise (1) it would be harder to throw an appropriate error
4052 : : * message if neither way works, and (2) we should not allow hashtable size
4053 : : * considerations to dissuade us from using hashing if sorting is not possible.
4054 : : *
4055 : : * *partially_grouped_rel_p will be set to the partially grouped rel which this
4056 : : * function creates, or to NULL if it doesn't create one.
4057 : : */
4058 : : static void
4059 : 19737 : create_ordinary_grouping_paths(PlannerInfo *root, RelOptInfo *input_rel,
4060 : : RelOptInfo *grouped_rel,
4061 : : const AggClauseCosts *agg_costs,
4062 : : grouping_sets_data *gd,
4063 : : GroupPathExtraData *extra,
4064 : : RelOptInfo **partially_grouped_rel_p)
4065 : : {
2778 4066 : 19737 : RelOptInfo *partially_grouped_rel = NULL;
2776 4067 : 19737 : PartitionwiseAggregateType patype = PARTITIONWISE_AGGREGATE_NONE;
4068 : :
4069 : : /*
4070 : : * If this is the topmost grouping relation or if the parent relation is
4071 : : * doing some form of partitionwise aggregation, then we may be able to do
4072 : : * it at this level also. However, if the input relation is not
4073 : : * partitioned, partitionwise aggregate is impossible.
4074 : : */
4075 [ + + ]: 19737 : if (extra->patype != PARTITIONWISE_AGGREGATE_NONE &&
2426 tgl@sss.pgh.pa.us 4076 [ + + + - : 1409 : IS_PARTITIONED_REL(input_rel))
+ + + - +
+ ]
4077 : : {
4078 : : /*
4079 : : * If this is the topmost relation or if the parent relation is doing
4080 : : * full partitionwise aggregation, then we can do full partitionwise
4081 : : * aggregation provided that the GROUP BY clause contains all of the
4082 : : * partitioning columns at this level and the collation used by GROUP
4083 : : * BY matches the partitioning collation. Otherwise, we can do at
4084 : : * most partial partitionwise aggregation. But if partial aggregation
4085 : : * is not supported in general then we can't use it for partitionwise
4086 : : * aggregation either.
4087 : : *
4088 : : * Check parse->groupClause not processed_groupClause, because it's
4089 : : * okay if some of the partitioning columns were proved redundant.
4090 : : */
2776 rhaas@postgresql.org 4091 [ + + + + ]: 808 : if (extra->patype == PARTITIONWISE_AGGREGATE_FULL &&
4092 : 380 : group_by_has_partkey(input_rel, extra->targetList,
4093 : 380 : root->parse->groupClause))
4094 : 238 : patype = PARTITIONWISE_AGGREGATE_FULL;
4095 [ + + ]: 190 : else if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4096 : 169 : patype = PARTITIONWISE_AGGREGATE_PARTIAL;
4097 : : else
4098 : 21 : patype = PARTITIONWISE_AGGREGATE_NONE;
4099 : : }
4100 : :
4101 : : /*
4102 : : * Before generating paths for grouped_rel, we first generate any possible
4103 : : * partially grouped paths; that way, later code can easily consider both
4104 : : * parallel and non-parallel approaches to grouping.
4105 : : */
4106 [ + + ]: 19737 : if ((extra->flags & GROUPING_CAN_PARTIAL_AGG) != 0)
4107 : : {
4108 : : bool force_rel_creation;
4109 : :
4110 : : /*
4111 : : * If we're doing partitionwise aggregation at this level, force
4112 : : * creation of a partially_grouped_rel so we can add partitionwise
4113 : : * paths to it.
4114 : : */
4115 : 17288 : force_rel_creation = (patype == PARTITIONWISE_AGGREGATE_PARTIAL);
4116 : :
4117 : : partially_grouped_rel =
2778 4118 : 17288 : create_partial_grouping_paths(root,
4119 : : grouped_rel,
4120 : : input_rel,
4121 : : gd,
4122 : : extra,
4123 : : force_rel_creation);
4124 : : }
4125 : :
4126 : : /* Set out parameter. */
2776 4127 : 19737 : *partially_grouped_rel_p = partially_grouped_rel;
4128 : :
4129 : : /* Apply partitionwise aggregation technique, if possible. */
4130 [ + + ]: 19737 : if (patype != PARTITIONWISE_AGGREGATE_NONE)
4131 : 407 : create_partitionwise_grouping_paths(root, input_rel, grouped_rel,
4132 : : partially_grouped_rel, agg_costs,
4133 : : gd, patype, extra);
4134 : :
4135 : : /* If we are doing partial aggregation only, return. */
4136 [ + + ]: 19737 : if (extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
4137 : : {
4138 [ - + ]: 429 : Assert(partially_grouped_rel);
4139 : :
4140 [ + - ]: 429 : if (partially_grouped_rel->pathlist)
4141 : 429 : set_cheapest(partially_grouped_rel);
4142 : :
4143 : 429 : return;
4144 : : }
4145 : :
4146 : : /* Gather any partially grouped partial paths. */
4147 [ + + + + ]: 19308 : if (partially_grouped_rel && partially_grouped_rel->partial_pathlist)
2778 4148 : 1007 : gather_grouping_paths(root, partially_grouped_rel);
4149 : :
4150 : : /* Now choose the best path(s) for partially_grouped_rel. */
19 rguo@postgresql.org 4151 [ + + + - ]:GNC 19308 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
4152 : 1121 : set_cheapest(partially_grouped_rel);
4153 : :
4154 : : /* Build final grouping paths */
2778 rhaas@postgresql.org 4155 :CBC 19308 : add_paths_to_grouping_rel(root, input_rel, grouped_rel,
4156 : : partially_grouped_rel, agg_costs, gd,
4157 : : extra);
4158 : :
4159 : : /* Give a helpful error if we failed to find any implementation */
2831 4160 [ + + ]: 19308 : if (grouped_rel->pathlist == NIL)
4161 [ + - ]: 3 : ereport(ERROR,
4162 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4163 : : errmsg("could not implement GROUP BY"),
4164 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4165 : :
4166 : : /*
4167 : : * If there is an FDW that's responsible for all baserels of the query,
4168 : : * let it consider adding ForeignPaths.
4169 : : */
4170 [ + + ]: 19305 : if (grouped_rel->fdwroutine &&
4171 [ + - ]: 168 : grouped_rel->fdwroutine->GetForeignUpperPaths)
4172 : 168 : grouped_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_GROUP_AGG,
4173 : : input_rel, grouped_rel,
4174 : : extra);
4175 : :
4176 : : /* Let extensions possibly add some more paths */
4177 [ - + ]: 19305 : if (create_upper_paths_hook)
2831 rhaas@postgresql.org 4178 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_GROUP_AGG,
4179 : : input_rel, grouped_rel,
4180 : : extra);
4181 : : }
4182 : :
4183 : : /*
4184 : : * For a given input path, consider the possible ways of doing grouping sets on
4185 : : * it, by combinations of hashing and sorting. This can be called multiple
4186 : : * times, so it's important that it not scribble on input. No result is
4187 : : * returned, but any generated paths are added to grouped_rel.
4188 : : */
4189 : : static void
2831 rhaas@postgresql.org 4190 :CBC 911 : consider_groupingsets_paths(PlannerInfo *root,
4191 : : RelOptInfo *grouped_rel,
4192 : : Path *path,
4193 : : bool is_sorted,
4194 : : bool can_hash,
4195 : : grouping_sets_data *gd,
4196 : : const AggClauseCosts *agg_costs,
4197 : : double dNumGroups)
4198 : : {
4199 : 911 : Query *parse = root->parse;
1555 tgl@sss.pgh.pa.us 4200 : 911 : Size hash_mem_limit = get_hash_memory_limit();
4201 : :
4202 : : /*
4203 : : * If we're not being offered sorted input, then only consider plans that
4204 : : * can be done entirely by hashing.
4205 : : *
4206 : : * We can hash everything if it looks like it'll fit in hash_mem. But if
4207 : : * the input is actually sorted despite not being advertised as such, we
4208 : : * prefer to make use of that in order to use less memory.
4209 : : *
4210 : : * If none of the grouping sets are sortable, then ignore the hash_mem
4211 : : * limit and generate a path anyway, since otherwise we'll just fail.
4212 : : */
2831 rhaas@postgresql.org 4213 [ + + ]: 911 : if (!is_sorted)
4214 : : {
4215 : 415 : List *new_rollups = NIL;
4216 : 415 : RollupData *unhashed_rollup = NULL;
4217 : : List *sets_data;
4218 : 415 : List *empty_sets_data = NIL;
4219 : 415 : List *empty_sets = NIL;
4220 : : ListCell *lc;
4221 : 415 : ListCell *l_start = list_head(gd->rollups);
4222 : 415 : AggStrategy strat = AGG_HASHED;
4223 : : double hashsize;
4224 : 415 : double exclude_groups = 0.0;
4225 : :
4226 [ - + ]: 415 : Assert(can_hash);
4227 : :
4228 : : /*
4229 : : * If the input is coincidentally sorted usefully (which can happen
4230 : : * even if is_sorted is false, since that only means that our caller
4231 : : * has set up the sorting for us), then save some hashtable space by
4232 : : * making use of that. But we need to watch out for degenerate cases:
4233 : : *
4234 : : * 1) If there are any empty grouping sets, then group_pathkeys might
4235 : : * be NIL if all non-empty grouping sets are unsortable. In this case,
4236 : : * there will be a rollup containing only empty groups, and the
4237 : : * pathkeys_contained_in test is vacuously true; this is ok.
4238 : : *
4239 : : * XXX: the above relies on the fact that group_pathkeys is generated
4240 : : * from the first rollup. If we add the ability to consider multiple
4241 : : * sort orders for grouping input, this assumption might fail.
4242 : : *
4243 : : * 2) If there are no empty sets and only unsortable sets, then the
4244 : : * rollups list will be empty (and thus l_start == NULL), and
4245 : : * group_pathkeys will be NIL; we must ensure that the vacuously-true
4246 : : * pathkeys_contained_in test doesn't cause us to crash.
4247 : : */
2777 rhodiumtoad@postgres 4248 [ + + + + ]: 827 : if (l_start != NULL &&
4249 : 412 : pathkeys_contained_in(root->group_pathkeys, path->pathkeys))
4250 : : {
2831 rhaas@postgresql.org 4251 : 6 : unhashed_rollup = lfirst_node(RollupData, l_start);
4252 : 6 : exclude_groups = unhashed_rollup->numGroups;
2296 tgl@sss.pgh.pa.us 4253 : 6 : l_start = lnext(gd->rollups, l_start);
4254 : : }
4255 : :
1798 heikki.linnakangas@i 4256 : 415 : hashsize = estimate_hashagg_tablesize(root,
4257 : : path,
4258 : : agg_costs,
4259 : : dNumGroups - exclude_groups);
4260 : :
4261 : : /*
4262 : : * gd->rollups is empty if we have only unsortable columns to work
4263 : : * with. Override hash_mem in that case; otherwise, we'll rely on the
4264 : : * sorted-input case to generate usable mixed paths.
4265 : : */
1555 tgl@sss.pgh.pa.us 4266 [ + + + - ]: 415 : if (hashsize > hash_mem_limit && gd->rollups)
2831 rhaas@postgresql.org 4267 : 9 : return; /* nope, won't fit */
4268 : :
4269 : : /*
4270 : : * We need to burst the existing rollups list into individual grouping
4271 : : * sets and recompute a groupClause for each set.
4272 : : */
4273 : 406 : sets_data = list_copy(gd->unsortable_sets);
4274 : :
2296 tgl@sss.pgh.pa.us 4275 [ + + + + : 1032 : for_each_cell(lc, gd->rollups, l_start)
+ + ]
4276 : : {
2831 rhaas@postgresql.org 4277 : 638 : RollupData *rollup = lfirst_node(RollupData, lc);
4278 : :
4279 : : /*
4280 : : * If we find an unhashable rollup that's not been skipped by the
4281 : : * "actually sorted" check above, we can't cope; we'd need sorted
4282 : : * input (with a different sort order) but we can't get that here.
4283 : : * So bail out; we'll get a valid path from the is_sorted case
4284 : : * instead.
4285 : : *
4286 : : * The mere presence of empty grouping sets doesn't make a rollup
4287 : : * unhashable (see preprocess_grouping_sets), we handle those
4288 : : * specially below.
4289 : : */
3136 rhodiumtoad@postgres 4290 [ + + ]: 638 : if (!rollup->hashable)
4291 : 12 : return;
4292 : :
2268 tgl@sss.pgh.pa.us 4293 : 626 : sets_data = list_concat(sets_data, rollup->gsets_data);
4294 : : }
3136 rhodiumtoad@postgres 4295 [ + - + + : 1635 : foreach(lc, sets_data)
+ + ]
4296 : : {
2974 tgl@sss.pgh.pa.us 4297 : 1241 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
3136 rhodiumtoad@postgres 4298 : 1241 : List *gset = gs->set;
4299 : : RollupData *rollup;
4300 : :
4301 [ + + ]: 1241 : if (gset == NIL)
4302 : : {
4303 : : /* Empty grouping sets can't be hashed. */
4304 : 248 : empty_sets_data = lappend(empty_sets_data, gs);
4305 : 248 : empty_sets = lappend(empty_sets, NIL);
4306 : : }
4307 : : else
4308 : : {
4309 : 993 : rollup = makeNode(RollupData);
4310 : :
508 akorotkov@postgresql 4311 : 993 : rollup->groupClause = preprocess_groupclause(root, gset);
3136 rhodiumtoad@postgres 4312 : 993 : rollup->gsets_data = list_make1(gs);
4313 : 993 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4314 : : rollup->gsets_data,
4315 : : gd->tleref_to_colnum_map);
4316 : 993 : rollup->numGroups = gs->numGroups;
4317 : 993 : rollup->hashable = true;
4318 : 993 : rollup->is_hashed = true;
4319 : 993 : new_rollups = lappend(new_rollups, rollup);
4320 : : }
4321 : : }
4322 : :
4323 : : /*
4324 : : * If we didn't find anything nonempty to hash, then bail. We'll
4325 : : * generate a path from the is_sorted case.
4326 : : */
4327 [ - + ]: 394 : if (new_rollups == NIL)
3136 rhodiumtoad@postgres 4328 :UBC 0 : return;
4329 : :
4330 : : /*
4331 : : * If there were empty grouping sets they should have been in the
4332 : : * first rollup.
4333 : : */
3136 rhodiumtoad@postgres 4334 [ + + - + ]:CBC 394 : Assert(!unhashed_rollup || !empty_sets);
4335 : :
4336 [ + + ]: 394 : if (unhashed_rollup)
4337 : : {
4338 : 6 : new_rollups = lappend(new_rollups, unhashed_rollup);
4339 : 6 : strat = AGG_MIXED;
4340 : : }
4341 [ + + ]: 388 : else if (empty_sets)
4342 : : {
4343 : 224 : RollupData *rollup = makeNode(RollupData);
4344 : :
4345 : 224 : rollup->groupClause = NIL;
4346 : 224 : rollup->gsets_data = empty_sets_data;
4347 : 224 : rollup->gsets = empty_sets;
4348 : 224 : rollup->numGroups = list_length(empty_sets);
4349 : 224 : rollup->hashable = false;
4350 : 224 : rollup->is_hashed = false;
4351 : 224 : new_rollups = lappend(new_rollups, rollup);
4352 : 224 : strat = AGG_MIXED;
4353 : : }
4354 : :
4355 : 394 : add_path(grouped_rel, (Path *)
4356 : 394 : create_groupingsets_path(root,
4357 : : grouped_rel,
4358 : : path,
4359 : 394 : (List *) parse->havingQual,
4360 : : strat,
4361 : : new_rollups,
4362 : : agg_costs));
4363 : 394 : return;
4364 : : }
4365 : :
4366 : : /*
4367 : : * If we have sorted input but nothing we can do with it, bail.
4368 : : */
1167 tgl@sss.pgh.pa.us 4369 [ - + ]: 496 : if (gd->rollups == NIL)
3136 rhodiumtoad@postgres 4370 :UBC 0 : return;
4371 : :
4372 : : /*
4373 : : * Given sorted input, we try and make two paths: one sorted and one mixed
4374 : : * sort/hash. (We need to try both because hashagg might be disabled, or
4375 : : * some columns might not be sortable.)
4376 : : *
4377 : : * can_hash is passed in as false if some obstacle elsewhere (such as
4378 : : * ordered aggs) means that we shouldn't consider hashing at all.
4379 : : */
3136 rhodiumtoad@postgres 4380 [ + + + - ]:CBC 496 : if (can_hash && gd->any_hashable)
4381 : : {
4382 : 448 : List *rollups = NIL;
4383 : 448 : List *hash_sets = list_copy(gd->unsortable_sets);
1555 tgl@sss.pgh.pa.us 4384 : 448 : double availspace = hash_mem_limit;
4385 : : ListCell *lc;
4386 : :
4387 : : /*
4388 : : * Account first for space needed for groups we can't sort at all.
4389 : : */
1798 heikki.linnakangas@i 4390 : 448 : availspace -= estimate_hashagg_tablesize(root,
4391 : : path,
4392 : : agg_costs,
4393 : : gd->dNumHashGroups);
4394 : :
3136 rhodiumtoad@postgres 4395 [ + - + + ]: 448 : if (availspace > 0 && list_length(gd->rollups) > 1)
4396 : : {
4397 : : double scale;
4398 : 228 : int num_rollups = list_length(gd->rollups);
4399 : : int k_capacity;
4400 : 228 : int *k_weights = palloc(num_rollups * sizeof(int));
4401 : 228 : Bitmapset *hash_items = NULL;
4402 : : int i;
4403 : :
4404 : : /*
4405 : : * We treat this as a knapsack problem: the knapsack capacity
4406 : : * represents hash_mem, the item weights are the estimated memory
4407 : : * usage of the hashtables needed to implement a single rollup,
4408 : : * and we really ought to use the cost saving as the item value;
4409 : : * however, currently the costs assigned to sort nodes don't
4410 : : * reflect the comparison costs well, and so we treat all items as
4411 : : * of equal value (each rollup we hash instead saves us one sort).
4412 : : *
4413 : : * To use the discrete knapsack, we need to scale the values to a
4414 : : * reasonably small bounded range. We choose to allow a 5% error
4415 : : * margin; we have no more than 4096 rollups in the worst possible
4416 : : * case, which with a 5% error margin will require a bit over 42MB
4417 : : * of workspace. (Anyone wanting to plan queries that complex had
4418 : : * better have the memory for it. In more reasonable cases, with
4419 : : * no more than a couple of dozen rollups, the memory usage will
4420 : : * be negligible.)
4421 : : *
4422 : : * k_capacity is naturally bounded, but we clamp the values for
4423 : : * scale and weight (below) to avoid overflows or underflows (or
4424 : : * uselessly trying to use a scale factor less than 1 byte).
4425 : : */
4426 [ + - ]: 228 : scale = Max(availspace / (20.0 * num_rollups), 1.0);
4427 : 228 : k_capacity = (int) floor(availspace / scale);
4428 : :
4429 : : /*
4430 : : * We leave the first rollup out of consideration since it's the
4431 : : * one that matches the input sort order. We assign indexes "i"
4432 : : * to only those entries considered for hashing; the second loop,
4433 : : * below, must use the same condition.
4434 : : */
4435 : 228 : i = 0;
1855 tgl@sss.pgh.pa.us 4436 [ + - + + : 582 : for_each_from(lc, gd->rollups, 1)
+ + ]
4437 : : {
2974 4438 : 354 : RollupData *rollup = lfirst_node(RollupData, lc);
4439 : :
3136 rhodiumtoad@postgres 4440 [ + - ]: 354 : if (rollup->hashable)
4441 : : {
1798 heikki.linnakangas@i 4442 : 354 : double sz = estimate_hashagg_tablesize(root,
4443 : : path,
4444 : : agg_costs,
4445 : : rollup->numGroups);
4446 : :
4447 : : /*
4448 : : * If sz is enormous, but hash_mem (and hence scale) is
4449 : : * small, avoid integer overflow here.
4450 : : */
3136 rhodiumtoad@postgres 4451 [ + + ]: 354 : k_weights[i] = (int) Min(floor(sz / scale),
4452 : : k_capacity + 1.0);
4453 : 354 : ++i;
4454 : : }
4455 : : }
4456 : :
4457 : : /*
4458 : : * Apply knapsack algorithm; compute the set of items which
4459 : : * maximizes the value stored (in this case the number of sorts
4460 : : * saved) while keeping the total size (approximately) within
4461 : : * capacity.
4462 : : */
4463 [ + - ]: 228 : if (i > 0)
4464 : 228 : hash_items = DiscreteKnapsack(k_capacity, i, k_weights, NULL);
4465 : :
4466 [ + - ]: 228 : if (!bms_is_empty(hash_items))
4467 : : {
4468 : 228 : rollups = list_make1(linitial(gd->rollups));
4469 : :
4470 : 228 : i = 0;
1855 tgl@sss.pgh.pa.us 4471 [ + - + + : 582 : for_each_from(lc, gd->rollups, 1)
+ + ]
4472 : : {
2974 4473 : 354 : RollupData *rollup = lfirst_node(RollupData, lc);
4474 : :
3136 rhodiumtoad@postgres 4475 [ + - ]: 354 : if (rollup->hashable)
4476 : : {
4477 [ + + ]: 354 : if (bms_is_member(i, hash_items))
4478 : 336 : hash_sets = list_concat(hash_sets,
2268 tgl@sss.pgh.pa.us 4479 : 336 : rollup->gsets_data);
4480 : : else
3136 rhodiumtoad@postgres 4481 : 18 : rollups = lappend(rollups, rollup);
4482 : 354 : ++i;
4483 : : }
4484 : : else
3136 rhodiumtoad@postgres 4485 :UBC 0 : rollups = lappend(rollups, rollup);
4486 : : }
4487 : : }
4488 : : }
4489 : :
3136 rhodiumtoad@postgres 4490 [ + + + + ]:CBC 448 : if (!rollups && hash_sets)
4491 : 12 : rollups = list_copy(gd->rollups);
4492 : :
4493 [ + + + + : 854 : foreach(lc, hash_sets)
+ + ]
4494 : : {
2974 tgl@sss.pgh.pa.us 4495 : 406 : GroupingSetData *gs = lfirst_node(GroupingSetData, lc);
3136 rhodiumtoad@postgres 4496 : 406 : RollupData *rollup = makeNode(RollupData);
4497 : :
4498 [ - + ]: 406 : Assert(gs->set != NIL);
4499 : :
508 akorotkov@postgresql 4500 : 406 : rollup->groupClause = preprocess_groupclause(root, gs->set);
3136 rhodiumtoad@postgres 4501 : 406 : rollup->gsets_data = list_make1(gs);
4502 : 406 : rollup->gsets = remap_to_groupclause_idx(rollup->groupClause,
4503 : : rollup->gsets_data,
4504 : : gd->tleref_to_colnum_map);
4505 : 406 : rollup->numGroups = gs->numGroups;
4506 : 406 : rollup->hashable = true;
4507 : 406 : rollup->is_hashed = true;
4508 : 406 : rollups = lcons(rollup, rollups);
4509 : : }
4510 : :
4511 [ + + ]: 448 : if (rollups)
4512 : : {
4513 : 240 : add_path(grouped_rel, (Path *)
4514 : 240 : create_groupingsets_path(root,
4515 : : grouped_rel,
4516 : : path,
4517 : 240 : (List *) parse->havingQual,
4518 : : AGG_MIXED,
4519 : : rollups,
4520 : : agg_costs));
4521 : : }
4522 : : }
4523 : :
4524 : : /*
4525 : : * Now try the simple sorted case.
4526 : : */
4527 [ + + ]: 496 : if (!gd->unsortable_sets)
4528 : 481 : add_path(grouped_rel, (Path *)
4529 : 481 : create_groupingsets_path(root,
4530 : : grouped_rel,
4531 : : path,
4532 : 481 : (List *) parse->havingQual,
4533 : : AGG_SORTED,
4534 : : gd->rollups,
4535 : : agg_costs));
4536 : : }
4537 : :
4538 : : /*
4539 : : * create_window_paths
4540 : : *
4541 : : * Build a new upperrel containing Paths for window-function evaluation.
4542 : : *
4543 : : * input_rel: contains the source-data Paths
4544 : : * input_target: result of make_window_input_target
4545 : : * output_target: what the topmost WindowAggPath should return
4546 : : * wflists: result of find_window_functions
4547 : : * activeWindows: result of select_active_windows
4548 : : *
4549 : : * Note: all Paths in input_rel are expected to return input_target.
4550 : : */
4551 : : static RelOptInfo *
3521 tgl@sss.pgh.pa.us 4552 : 1282 : create_window_paths(PlannerInfo *root,
4553 : : RelOptInfo *input_rel,
4554 : : PathTarget *input_target,
4555 : : PathTarget *output_target,
4556 : : bool output_target_parallel_safe,
4557 : : WindowFuncLists *wflists,
4558 : : List *activeWindows)
4559 : : {
4560 : : RelOptInfo *window_rel;
4561 : : ListCell *lc;
4562 : :
4563 : : /* For now, do all work in the (WINDOW, NULL) upperrel */
4564 : 1282 : window_rel = fetch_upper_rel(root, UPPERREL_WINDOW, NULL);
4565 : :
4566 : : /*
4567 : : * If the input relation is not parallel-safe, then the window relation
4568 : : * can't be parallel-safe, either. Otherwise, we need to examine the
4569 : : * target list and active windows for non-parallel-safe constructs.
4570 : : */
2790 rhaas@postgresql.org 4571 [ + + - + : 1282 : if (input_rel->consider_parallel && output_target_parallel_safe &&
- - ]
3356 tgl@sss.pgh.pa.us 4572 :UBC 0 : is_parallel_safe(root, (Node *) activeWindows))
3405 rhaas@postgresql.org 4573 : 0 : window_rel->consider_parallel = true;
4574 : :
4575 : : /*
4576 : : * If the input rel belongs to a single FDW, so does the window rel.
4577 : : */
3405 tgl@sss.pgh.pa.us 4578 :CBC 1282 : window_rel->serverid = input_rel->serverid;
3391 4579 : 1282 : window_rel->userid = input_rel->userid;
4580 : 1282 : window_rel->useridiscurrent = input_rel->useridiscurrent;
3405 4581 : 1282 : window_rel->fdwroutine = input_rel->fdwroutine;
4582 : :
4583 : : /*
4584 : : * Consider computing window functions starting from the existing
4585 : : * cheapest-total path (which will likely require a sort) as well as any
4586 : : * existing paths that satisfy or partially satisfy root->window_pathkeys.
4587 : : */
3521 4588 [ + - + + : 2734 : foreach(lc, input_rel->pathlist)
+ + ]
4589 : : {
4590 : 1452 : Path *path = (Path *) lfirst(lc);
4591 : : int presorted_keys;
4592 : :
4593 [ + + + + ]: 1622 : if (path == input_rel->cheapest_total_path ||
1868 drowley@postgresql.o 4594 : 170 : pathkeys_count_contained_in(root->window_pathkeys, path->pathkeys,
4595 : 76 : &presorted_keys) ||
4596 [ + + ]: 76 : presorted_keys > 0)
3521 tgl@sss.pgh.pa.us 4597 : 1389 : create_one_window_path(root,
4598 : : window_rel,
4599 : : path,
4600 : : input_target,
4601 : : output_target,
4602 : : wflists,
4603 : : activeWindows);
4604 : : }
4605 : :
4606 : : /*
4607 : : * If there is an FDW that's responsible for all baserels of the query,
4608 : : * let it consider adding ForeignPaths.
4609 : : */
3405 4610 [ + + ]: 1282 : if (window_rel->fdwroutine &&
4611 [ + - ]: 6 : window_rel->fdwroutine->GetForeignUpperPaths)
4612 : 6 : window_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_WINDOW,
4613 : : input_rel, window_rel,
4614 : : NULL);
4615 : :
4616 : : /* Let extensions possibly add some more paths */
3485 4617 [ - + ]: 1282 : if (create_upper_paths_hook)
3485 tgl@sss.pgh.pa.us 4618 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_WINDOW,
4619 : : input_rel, window_rel, NULL);
4620 : :
4621 : : /* Now choose the best path(s) */
3521 tgl@sss.pgh.pa.us 4622 :CBC 1282 : set_cheapest(window_rel);
4623 : :
4624 : 1282 : return window_rel;
4625 : : }
4626 : :
4627 : : /*
4628 : : * Stack window-function implementation steps atop the given Path, and
4629 : : * add the result to window_rel.
4630 : : *
4631 : : * window_rel: upperrel to contain result
4632 : : * path: input Path to use (must return input_target)
4633 : : * input_target: result of make_window_input_target
4634 : : * output_target: what the topmost WindowAggPath should return
4635 : : * wflists: result of find_window_functions
4636 : : * activeWindows: result of select_active_windows
4637 : : */
4638 : : static void
4639 : 1389 : create_one_window_path(PlannerInfo *root,
4640 : : RelOptInfo *window_rel,
4641 : : Path *path,
4642 : : PathTarget *input_target,
4643 : : PathTarget *output_target,
4644 : : WindowFuncLists *wflists,
4645 : : List *activeWindows)
4646 : : {
4647 : : PathTarget *window_target;
4648 : : ListCell *l;
1298 drowley@postgresql.o 4649 : 1389 : List *topqual = NIL;
4650 : :
4651 : : /*
4652 : : * Since each window clause could require a different sort order, we stack
4653 : : * up a WindowAgg node for each clause, with sort steps between them as
4654 : : * needed. (We assume that select_active_windows chose a good order for
4655 : : * executing the clauses in.)
4656 : : *
4657 : : * input_target should contain all Vars and Aggs needed for the result.
4658 : : * (In some cases we wouldn't need to propagate all of these all the way
4659 : : * to the top, since they might only be needed as inputs to WindowFuncs.
4660 : : * It's probably not worth trying to optimize that though.) It must also
4661 : : * contain all window partitioning and sorting expressions, to ensure
4662 : : * they're computed only once at the bottom of the stack (that's critical
4663 : : * for volatile functions). As we climb up the stack, we'll add outputs
4664 : : * for the WindowFuncs computed at each level.
4665 : : */
3519 tgl@sss.pgh.pa.us 4666 : 1389 : window_target = input_target;
4667 : :
3521 4668 [ + - + + : 2871 : foreach(l, activeWindows)
+ + ]
4669 : : {
2974 4670 : 1482 : WindowClause *wc = lfirst_node(WindowClause, l);
4671 : : List *window_pathkeys;
540 drowley@postgresql.o 4672 : 1482 : List *runcondition = NIL;
4673 : : int presorted_keys;
4674 : : bool is_sorted;
4675 : : bool topwindow;
4676 : : ListCell *lc2;
4677 : :
3521 tgl@sss.pgh.pa.us 4678 : 1482 : window_pathkeys = make_pathkeys_for_window(root,
4679 : : wc,
4680 : : root->processed_tlist);
4681 : :
1868 drowley@postgresql.o 4682 : 1482 : is_sorted = pathkeys_count_contained_in(window_pathkeys,
4683 : : path->pathkeys,
4684 : : &presorted_keys);
4685 : :
4686 : : /* Sort if necessary */
4687 [ + + ]: 1482 : if (!is_sorted)
4688 : : {
4689 : : /*
4690 : : * No presorted keys or incremental sort disabled, just perform a
4691 : : * complete sort.
4692 : : */
4693 [ + + - + ]: 1086 : if (presorted_keys == 0 || !enable_incremental_sort)
4694 : 1055 : path = (Path *) create_sort_path(root, window_rel,
4695 : : path,
4696 : : window_pathkeys,
4697 : : -1.0);
4698 : : else
4699 : : {
4700 : : /*
4701 : : * Since we have presorted keys and incremental sort is
4702 : : * enabled, just use incremental sort.
4703 : : */
4704 : 31 : path = (Path *) create_incremental_sort_path(root,
4705 : : window_rel,
4706 : : path,
4707 : : window_pathkeys,
4708 : : presorted_keys,
4709 : : -1.0);
4710 : : }
4711 : : }
4712 : :
2296 tgl@sss.pgh.pa.us 4713 [ + + ]: 1482 : if (lnext(activeWindows, l))
4714 : : {
4715 : : /*
4716 : : * Add the current WindowFuncs to the output target for this
4717 : : * intermediate WindowAggPath. We must copy window_target to
4718 : : * avoid changing the previous path's target.
4719 : : *
4720 : : * Note: a WindowFunc adds nothing to the target's eval costs; but
4721 : : * we do need to account for the increase in tlist width.
4722 : : */
678 4723 : 93 : int64 tuple_width = window_target->width;
4724 : :
3519 4725 : 93 : window_target = copy_pathtarget(window_target);
4726 [ + - + + : 219 : foreach(lc2, wflists->windowFuncs[wc->winref])
+ + ]
4727 : : {
3122 4728 : 126 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4729 : :
3519 4730 : 126 : add_column_to_pathtarget(window_target, (Expr *) wfunc, 0);
678 4731 : 126 : tuple_width += get_typavgwidth(wfunc->wintype, -1);
4732 : : }
4733 : 93 : window_target->width = clamp_width_est(tuple_width);
4734 : : }
4735 : : else
4736 : : {
4737 : : /* Install the goal target in the topmost WindowAgg */
3519 4738 : 1389 : window_target = output_target;
4739 : : }
4740 : :
4741 : : /* mark the final item in the list as the top-level window */
1298 drowley@postgresql.o 4742 : 1482 : topwindow = foreach_current_index(l) == list_length(activeWindows) - 1;
4743 : :
4744 : : /*
4745 : : * Collect the WindowFuncRunConditions from each WindowFunc and
4746 : : * convert them into OpExprs
4747 : : */
540 4748 [ + - + + : 3399 : foreach(lc2, wflists->windowFuncs[wc->winref])
+ + ]
4749 : : {
4750 : : ListCell *lc3;
4751 : 1917 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
4752 : :
4753 [ + + + + : 2007 : foreach(lc3, wfunc->runCondition)
+ + ]
4754 : : {
4755 : : WindowFuncRunCondition *wfuncrc =
4756 : 90 : lfirst_node(WindowFuncRunCondition, lc3);
4757 : : Expr *opexpr;
4758 : : Expr *leftop;
4759 : : Expr *rightop;
4760 : :
4761 [ + + ]: 90 : if (wfuncrc->wfunc_left)
4762 : : {
4763 : 81 : leftop = (Expr *) copyObject(wfunc);
4764 : 81 : rightop = copyObject(wfuncrc->arg);
4765 : : }
4766 : : else
4767 : : {
4768 : 9 : leftop = copyObject(wfuncrc->arg);
4769 : 9 : rightop = (Expr *) copyObject(wfunc);
4770 : : }
4771 : :
4772 : 90 : opexpr = make_opclause(wfuncrc->opno,
4773 : : BOOLOID,
4774 : : false,
4775 : : leftop,
4776 : : rightop,
4777 : : InvalidOid,
4778 : : wfuncrc->inputcollid);
4779 : :
4780 : 90 : runcondition = lappend(runcondition, opexpr);
4781 : :
4782 [ + + ]: 90 : if (!topwindow)
4783 : 12 : topqual = lappend(topqual, opexpr);
4784 : : }
4785 : : }
4786 : :
4787 : : path = (Path *)
3519 tgl@sss.pgh.pa.us 4788 [ + + ]: 1482 : create_windowagg_path(root, window_rel, path, window_target,
3521 4789 : 1482 : wflists->windowFuncs[wc->winref],
4790 : : runcondition, wc,
4791 : : topwindow ? topqual : NIL, topwindow);
4792 : : }
4793 : :
4794 : 1389 : add_path(window_rel, path);
7505 4795 : 1389 : }
4796 : :
4797 : : /*
4798 : : * create_distinct_paths
4799 : : *
4800 : : * Build a new upperrel containing Paths for SELECT DISTINCT evaluation.
4801 : : *
4802 : : * input_rel: contains the source-data Paths
4803 : : * target: the pathtarget for the result Paths to compute
4804 : : *
4805 : : * Note: input paths should already compute the desired pathtarget, since
4806 : : * Sort/Unique won't project anything.
4807 : : */
4808 : : static RelOptInfo *
628 drowley@postgresql.o 4809 : 1479 : create_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4810 : : PathTarget *target)
4811 : : {
4812 : : RelOptInfo *distinct_rel;
4813 : :
4814 : : /* For now, do all work in the (DISTINCT, NULL) upperrel */
3521 tgl@sss.pgh.pa.us 4815 : 1479 : distinct_rel = fetch_upper_rel(root, UPPERREL_DISTINCT, NULL);
4816 : :
4817 : : /*
4818 : : * We don't compute anything at this level, so distinct_rel will be
4819 : : * parallel-safe if the input rel is parallel-safe. In particular, if
4820 : : * there is a DISTINCT ON (...) clause, any path for the input_rel will
4821 : : * output those expressions, and will not be parallel-safe unless those
4822 : : * expressions are parallel-safe.
4823 : : */
3405 rhaas@postgresql.org 4824 : 1479 : distinct_rel->consider_parallel = input_rel->consider_parallel;
4825 : :
4826 : : /*
4827 : : * If the input rel belongs to a single FDW, so does the distinct_rel.
4828 : : */
tgl@sss.pgh.pa.us 4829 : 1479 : distinct_rel->serverid = input_rel->serverid;
3391 4830 : 1479 : distinct_rel->userid = input_rel->userid;
4831 : 1479 : distinct_rel->useridiscurrent = input_rel->useridiscurrent;
3405 4832 : 1479 : distinct_rel->fdwroutine = input_rel->fdwroutine;
4833 : :
4834 : : /* build distinct paths based on input_rel's pathlist */
1527 drowley@postgresql.o 4835 : 1479 : create_final_distinct_paths(root, input_rel, distinct_rel);
4836 : :
4837 : : /* now build distinct paths based on input_rel's partial_pathlist */
628 4838 : 1479 : create_partial_distinct_paths(root, input_rel, distinct_rel, target);
4839 : :
4840 : : /* Give a helpful error if we failed to create any paths */
1527 4841 [ - + ]: 1479 : if (distinct_rel->pathlist == NIL)
1527 drowley@postgresql.o 4842 [ # # ]:UBC 0 : ereport(ERROR,
4843 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
4844 : : errmsg("could not implement DISTINCT"),
4845 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
4846 : :
4847 : : /*
4848 : : * If there is an FDW that's responsible for all baserels of the query,
4849 : : * let it consider adding ForeignPaths.
4850 : : */
1527 drowley@postgresql.o 4851 [ + + ]:CBC 1479 : if (distinct_rel->fdwroutine &&
4852 [ + - ]: 8 : distinct_rel->fdwroutine->GetForeignUpperPaths)
4853 : 8 : distinct_rel->fdwroutine->GetForeignUpperPaths(root,
4854 : : UPPERREL_DISTINCT,
4855 : : input_rel,
4856 : : distinct_rel,
4857 : : NULL);
4858 : :
4859 : : /* Let extensions possibly add some more paths */
4860 [ - + ]: 1479 : if (create_upper_paths_hook)
1527 drowley@postgresql.o 4861 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_DISTINCT, input_rel,
4862 : : distinct_rel, NULL);
4863 : :
4864 : : /* Now choose the best path(s) */
1527 drowley@postgresql.o 4865 :CBC 1479 : set_cheapest(distinct_rel);
4866 : :
4867 : 1479 : return distinct_rel;
4868 : : }
4869 : :
4870 : : /*
4871 : : * create_partial_distinct_paths
4872 : : *
4873 : : * Process 'input_rel' partial paths and add unique/aggregate paths to the
4874 : : * UPPERREL_PARTIAL_DISTINCT rel. For paths created, add Gather/GatherMerge
4875 : : * paths on top and add a final unique/aggregate path to remove any duplicate
4876 : : * produced from combining rows from parallel workers.
4877 : : */
4878 : : static void
4879 : 1479 : create_partial_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
4880 : : RelOptInfo *final_distinct_rel,
4881 : : PathTarget *target)
4882 : : {
4883 : : RelOptInfo *partial_distinct_rel;
4884 : : Query *parse;
4885 : : List *distinctExprs;
4886 : : double numDistinctRows;
4887 : : Path *cheapest_partial_path;
4888 : : ListCell *lc;
4889 : :
4890 : : /* nothing to do when there are no partial paths in the input rel */
4891 [ + + + + ]: 1479 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
4892 : 1425 : return;
4893 : :
4894 : 54 : parse = root->parse;
4895 : :
4896 : : /* can't do parallel DISTINCT ON */
4897 [ - + ]: 54 : if (parse->hasDistinctOn)
1527 drowley@postgresql.o 4898 :UBC 0 : return;
4899 : :
1527 drowley@postgresql.o 4900 :CBC 54 : partial_distinct_rel = fetch_upper_rel(root, UPPERREL_PARTIAL_DISTINCT,
4901 : : NULL);
628 4902 : 54 : partial_distinct_rel->reltarget = target;
1527 4903 : 54 : partial_distinct_rel->consider_parallel = input_rel->consider_parallel;
4904 : :
4905 : : /*
4906 : : * If input_rel belongs to a single FDW, so does the partial_distinct_rel.
4907 : : */
4908 : 54 : partial_distinct_rel->serverid = input_rel->serverid;
4909 : 54 : partial_distinct_rel->userid = input_rel->userid;
4910 : 54 : partial_distinct_rel->useridiscurrent = input_rel->useridiscurrent;
4911 : 54 : partial_distinct_rel->fdwroutine = input_rel->fdwroutine;
4912 : :
4913 : 54 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
4914 : :
1013 tgl@sss.pgh.pa.us 4915 : 54 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
4916 : : parse->targetList);
4917 : :
4918 : : /* estimate how many distinct rows we'll get from each worker */
1527 drowley@postgresql.o 4919 : 54 : numDistinctRows = estimate_num_groups(root, distinctExprs,
4920 : : cheapest_partial_path->rows,
4921 : : NULL, NULL);
4922 : :
4923 : : /*
4924 : : * Try sorting the cheapest path and incrementally sorting any paths with
4925 : : * presorted keys and put a unique paths atop of those. We'll also
4926 : : * attempt to reorder the required pathkeys to match the input path's
4927 : : * pathkeys as much as possible, in hopes of avoiding a possible need to
4928 : : * re-sort.
4929 : : */
1013 tgl@sss.pgh.pa.us 4930 [ + - ]: 54 : if (grouping_is_sortable(root->processed_distinctClause))
4931 : : {
1527 drowley@postgresql.o 4932 [ + - + + : 117 : foreach(lc, input_rel->partial_pathlist)
+ + ]
4933 : : {
1020 4934 : 63 : Path *input_path = (Path *) lfirst(lc);
4935 : : Path *sorted_path;
335 rguo@postgresql.org 4936 : 63 : List *useful_pathkeys_list = NIL;
4937 : :
4938 : : useful_pathkeys_list =
4939 : 63 : get_useful_pathkeys_for_distinct(root,
4940 : : root->distinct_pathkeys,
4941 : : input_path->pathkeys);
4942 [ - + ]: 63 : Assert(list_length(useful_pathkeys_list) > 0);
4943 : :
4944 [ + - + + : 195 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
+ + ]
4945 : : {
4946 : 69 : sorted_path = make_ordered_path(root,
4947 : : partial_distinct_rel,
4948 : : input_path,
4949 : : cheapest_partial_path,
4950 : : useful_pathkeys,
4951 : : -1.0);
4952 : :
4953 [ + + ]: 69 : if (sorted_path == NULL)
1020 drowley@postgresql.o 4954 : 6 : continue;
4955 : :
4956 : : /*
4957 : : * An empty distinct_pathkeys means all tuples have the same
4958 : : * value for the DISTINCT clause. See
4959 : : * create_final_distinct_paths()
4960 : : */
335 rguo@postgresql.org 4961 [ + + ]: 63 : if (root->distinct_pathkeys == NIL)
4962 : : {
4963 : : Node *limitCount;
4964 : :
4965 : 3 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
4966 : : sizeof(int64),
4967 : : Int64GetDatum(1), false,
4968 : : true);
4969 : :
4970 : : /*
4971 : : * Apply a LimitPath onto the partial path to restrict the
4972 : : * tuples from each worker to 1.
4973 : : * create_final_distinct_paths will need to apply an
4974 : : * additional LimitPath to restrict this to a single row
4975 : : * after the Gather node. If the query already has a
4976 : : * LIMIT clause, then we could end up with three Limit
4977 : : * nodes in the final plan. Consolidating the top two of
4978 : : * these could be done, but does not seem worth troubling
4979 : : * over.
4980 : : */
4981 : 3 : add_partial_path(partial_distinct_rel, (Path *)
4982 : 3 : create_limit_path(root, partial_distinct_rel,
4983 : : sorted_path,
4984 : : NULL,
4985 : : limitCount,
4986 : : LIMIT_OPTION_COUNT,
4987 : : 0, 1));
4988 : : }
4989 : : else
4990 : : {
4991 : 60 : add_partial_path(partial_distinct_rel, (Path *)
69 rguo@postgresql.org 4992 :GNC 60 : create_unique_path(root, partial_distinct_rel,
4993 : : sorted_path,
4994 : 60 : list_length(root->distinct_pathkeys),
4995 : : numDistinctRows));
4996 : : }
4997 : : }
4998 : : }
4999 : : }
5000 : :
5001 : : /*
5002 : : * Now try hash aggregate paths, if enabled and hashing is possible. Since
5003 : : * we're not on the hook to ensure we do our best to create at least one
5004 : : * path here, we treat enable_hashagg as a hard off-switch rather than the
5005 : : * slightly softer variant in create_final_distinct_paths.
5006 : : */
1013 tgl@sss.pgh.pa.us 5007 [ + + + - ]:CBC 54 : if (enable_hashagg && grouping_is_hashable(root->processed_distinctClause))
5008 : : {
1527 drowley@postgresql.o 5009 : 39 : add_partial_path(partial_distinct_rel, (Path *)
5010 : 39 : create_agg_path(root,
5011 : : partial_distinct_rel,
5012 : : cheapest_partial_path,
5013 : : cheapest_partial_path->pathtarget,
5014 : : AGG_HASHED,
5015 : : AGGSPLIT_SIMPLE,
5016 : : root->processed_distinctClause,
5017 : : NIL,
5018 : : NULL,
5019 : : numDistinctRows));
5020 : : }
5021 : :
5022 : : /*
5023 : : * If there is an FDW that's responsible for all baserels of the query,
5024 : : * let it consider adding ForeignPaths.
5025 : : */
5026 [ - + ]: 54 : if (partial_distinct_rel->fdwroutine &&
1527 drowley@postgresql.o 5027 [ # # ]:UBC 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths)
5028 : 0 : partial_distinct_rel->fdwroutine->GetForeignUpperPaths(root,
5029 : : UPPERREL_PARTIAL_DISTINCT,
5030 : : input_rel,
5031 : : partial_distinct_rel,
5032 : : NULL);
5033 : :
5034 : : /* Let extensions possibly add some more partial paths */
1527 drowley@postgresql.o 5035 [ - + ]:CBC 54 : if (create_upper_paths_hook)
1527 drowley@postgresql.o 5036 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_PARTIAL_DISTINCT,
5037 : : input_rel, partial_distinct_rel, NULL);
5038 : :
1527 drowley@postgresql.o 5039 [ + - ]:CBC 54 : if (partial_distinct_rel->partial_pathlist != NIL)
5040 : : {
632 5041 : 54 : generate_useful_gather_paths(root, partial_distinct_rel, true);
1527 5042 : 54 : set_cheapest(partial_distinct_rel);
5043 : :
5044 : : /*
5045 : : * Finally, create paths to distinctify the final result. This step
5046 : : * is needed to remove any duplicates due to combining rows from
5047 : : * parallel workers.
5048 : : */
5049 : 54 : create_final_distinct_paths(root, partial_distinct_rel,
5050 : : final_distinct_rel);
5051 : : }
5052 : : }
5053 : :
5054 : : /*
5055 : : * create_final_distinct_paths
5056 : : * Create distinct paths in 'distinct_rel' based on 'input_rel' pathlist
5057 : : *
5058 : : * input_rel: contains the source-data paths
5059 : : * distinct_rel: destination relation for storing created paths
5060 : : */
5061 : : static RelOptInfo *
5062 : 1533 : create_final_distinct_paths(PlannerInfo *root, RelOptInfo *input_rel,
5063 : : RelOptInfo *distinct_rel)
5064 : : {
5065 : 1533 : Query *parse = root->parse;
5066 : 1533 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5067 : : double numDistinctRows;
5068 : : bool allow_hash;
5069 : :
5070 : : /* Estimate number of distinct rows there will be */
3521 tgl@sss.pgh.pa.us 5071 [ + + + - : 1533 : if (parse->groupClause || parse->groupingSets || parse->hasAggs ||
+ + ]
5072 [ - + ]: 1496 : root->hasHavingQual)
5073 : : {
5074 : : /*
5075 : : * If there was grouping or aggregation, use the number of input rows
5076 : : * as the estimated number of DISTINCT rows (ie, assume the input is
5077 : : * already mostly unique).
5078 : : */
5079 : 37 : numDistinctRows = cheapest_input_path->rows;
5080 : : }
5081 : : else
5082 : : {
5083 : : /*
5084 : : * Otherwise, the UNIQUE filter has effects comparable to GROUP BY.
5085 : : */
5086 : : List *distinctExprs;
5087 : :
1013 5088 : 1496 : distinctExprs = get_sortgrouplist_exprs(root->processed_distinctClause,
5089 : : parse->targetList);
3521 5090 : 1496 : numDistinctRows = estimate_num_groups(root, distinctExprs,
5091 : : cheapest_input_path->rows,
5092 : : NULL, NULL);
5093 : : }
5094 : :
5095 : : /*
5096 : : * Consider sort-based implementations of DISTINCT, if possible.
5097 : : */
1013 5098 [ + + ]: 1533 : if (grouping_is_sortable(root->processed_distinctClause))
5099 : : {
5100 : : /*
5101 : : * Firstly, if we have any adequately-presorted paths, just stick a
5102 : : * Unique node on those. We also, consider doing an explicit sort of
5103 : : * the cheapest input path and Unique'ing that. If any paths have
5104 : : * presorted keys then we'll create an incremental sort atop of those
5105 : : * before adding a unique node on the top. We'll also attempt to
5106 : : * reorder the required pathkeys to match the input path's pathkeys as
5107 : : * much as possible, in hopes of avoiding a possible need to re-sort.
5108 : : *
5109 : : * When we have DISTINCT ON, we must sort by the more rigorous of
5110 : : * DISTINCT and ORDER BY, else it won't have the desired behavior.
5111 : : * Also, if we do have to do an explicit sort, we might as well use
5112 : : * the more rigorous ordering to avoid a second sort later. (Note
5113 : : * that the parser will have ensured that one clause is a prefix of
5114 : : * the other.)
5115 : : */
5116 : : List *needed_pathkeys;
5117 : : ListCell *lc;
1020 drowley@postgresql.o 5118 [ + + ]: 1530 : double limittuples = root->distinct_pathkeys == NIL ? 1.0 : -1.0;
5119 : :
3521 tgl@sss.pgh.pa.us 5120 [ + + + + ]: 1654 : if (parse->hasDistinctOn &&
5121 : 124 : list_length(root->distinct_pathkeys) <
5122 : 124 : list_length(root->sort_pathkeys))
5123 : 27 : needed_pathkeys = root->sort_pathkeys;
5124 : : else
5125 : 1503 : needed_pathkeys = root->distinct_pathkeys;
5126 : :
5127 [ + - + + : 4130 : foreach(lc, input_rel->pathlist)
+ + ]
5128 : : {
1020 drowley@postgresql.o 5129 : 2600 : Path *input_path = (Path *) lfirst(lc);
5130 : : Path *sorted_path;
335 rguo@postgresql.org 5131 : 2600 : List *useful_pathkeys_list = NIL;
5132 : :
5133 : : useful_pathkeys_list =
5134 : 2600 : get_useful_pathkeys_for_distinct(root,
5135 : : needed_pathkeys,
5136 : : input_path->pathkeys);
5137 [ - + ]: 2600 : Assert(list_length(useful_pathkeys_list) > 0);
5138 : :
5139 [ + - + + : 8200 : foreach_node(List, useful_pathkeys, useful_pathkeys_list)
+ + ]
5140 : : {
5141 : 3000 : sorted_path = make_ordered_path(root,
5142 : : distinct_rel,
5143 : : input_path,
5144 : : cheapest_input_path,
5145 : : useful_pathkeys,
5146 : : limittuples);
5147 : :
5148 [ + + ]: 3000 : if (sorted_path == NULL)
1020 drowley@postgresql.o 5149 : 429 : continue;
5150 : :
5151 : : /*
5152 : : * distinct_pathkeys may have become empty if all of the
5153 : : * pathkeys were determined to be redundant. If all of the
5154 : : * pathkeys are redundant then each DISTINCT target must only
5155 : : * allow a single value, therefore all resulting tuples must
5156 : : * be identical (or at least indistinguishable by an equality
5157 : : * check). We can uniquify these tuples simply by just taking
5158 : : * the first tuple. All we do here is add a path to do "LIMIT
5159 : : * 1" atop of 'sorted_path'. When doing a DISTINCT ON we may
5160 : : * still have a non-NIL sort_pathkeys list, so we must still
5161 : : * only do this with paths which are correctly sorted by
5162 : : * sort_pathkeys.
5163 : : */
335 rguo@postgresql.org 5164 [ + + ]: 2571 : if (root->distinct_pathkeys == NIL)
5165 : : {
5166 : : Node *limitCount;
5167 : :
5168 : 69 : limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
5169 : : sizeof(int64),
5170 : : Int64GetDatum(1), false,
5171 : : true);
5172 : :
5173 : : /*
5174 : : * If the query already has a LIMIT clause, then we could
5175 : : * end up with a duplicate LimitPath in the final plan.
5176 : : * That does not seem worth troubling over too much.
5177 : : */
5178 : 69 : add_path(distinct_rel, (Path *)
5179 : 69 : create_limit_path(root, distinct_rel, sorted_path,
5180 : : NULL, limitCount,
5181 : : LIMIT_OPTION_COUNT, 0, 1));
5182 : : }
5183 : : else
5184 : : {
5185 : 2502 : add_path(distinct_rel, (Path *)
69 rguo@postgresql.org 5186 :GNC 2502 : create_unique_path(root, distinct_rel,
5187 : : sorted_path,
5188 : 2502 : list_length(root->distinct_pathkeys),
5189 : : numDistinctRows));
5190 : : }
5191 : : }
5192 : : }
5193 : : }
5194 : :
5195 : : /*
5196 : : * Consider hash-based implementations of DISTINCT, if possible.
5197 : : *
5198 : : * If we were not able to make any other types of path, we *must* hash or
5199 : : * die trying. If we do have other choices, there are two things that
5200 : : * should prevent selection of hashing: if the query uses DISTINCT ON
5201 : : * (because it won't really have the expected behavior if we hash), or if
5202 : : * enable_hashagg is off.
5203 : : *
5204 : : * Note: grouping_is_hashable() is much more expensive to check than the
5205 : : * other gating conditions, so we want to do it last.
5206 : : */
3521 tgl@sss.pgh.pa.us 5207 [ + + ]:CBC 1533 : if (distinct_rel->pathlist == NIL)
5208 : 3 : allow_hash = true; /* we have no alternatives */
5209 [ + + + + ]: 1530 : else if (parse->hasDistinctOn || !enable_hashagg)
5210 : 199 : allow_hash = false; /* policy-based decision not to hash */
5211 : : else
1918 pg@bowt.ie 5212 : 1331 : allow_hash = true; /* default */
5213 : :
1013 tgl@sss.pgh.pa.us 5214 [ + + + - ]: 1533 : if (allow_hash && grouping_is_hashable(root->processed_distinctClause))
5215 : : {
5216 : : /* Generate hashed aggregate path --- no sort needed */
3521 5217 : 1334 : add_path(distinct_rel, (Path *)
5218 : 1334 : create_agg_path(root,
5219 : : distinct_rel,
5220 : : cheapest_input_path,
5221 : : cheapest_input_path->pathtarget,
5222 : : AGG_HASHED,
5223 : : AGGSPLIT_SIMPLE,
5224 : : root->processed_distinctClause,
5225 : : NIL,
5226 : : NULL,
5227 : : numDistinctRows));
5228 : : }
5229 : :
5230 : 1533 : return distinct_rel;
5231 : : }
5232 : :
5233 : : /*
5234 : : * get_useful_pathkeys_for_distinct
5235 : : * Get useful orderings of pathkeys for distinctClause by reordering
5236 : : * 'needed_pathkeys' to match the given 'path_pathkeys' as much as possible.
5237 : : *
5238 : : * This returns a list of pathkeys that can be useful for DISTINCT or DISTINCT
5239 : : * ON clause. For convenience, it always includes the given 'needed_pathkeys'.
5240 : : */
5241 : : static List *
335 rguo@postgresql.org 5242 : 2663 : get_useful_pathkeys_for_distinct(PlannerInfo *root, List *needed_pathkeys,
5243 : : List *path_pathkeys)
5244 : : {
5245 : 2663 : List *useful_pathkeys_list = NIL;
5246 : 2663 : List *useful_pathkeys = NIL;
5247 : :
5248 : : /* always include the given 'needed_pathkeys' */
5249 : 2663 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5250 : : needed_pathkeys);
5251 : :
5252 [ - + ]: 2663 : if (!enable_distinct_reordering)
335 rguo@postgresql.org 5253 :UBC 0 : return useful_pathkeys_list;
5254 : :
5255 : : /*
5256 : : * Scan the given 'path_pathkeys' and construct a list of PathKey nodes
5257 : : * that match 'needed_pathkeys', but only up to the longest matching
5258 : : * prefix.
5259 : : *
5260 : : * When we have DISTINCT ON, we must ensure that the resulting pathkey
5261 : : * list matches initial distinctClause pathkeys; otherwise, it won't have
5262 : : * the desired behavior.
5263 : : */
335 rguo@postgresql.org 5264 [ + + + + :CBC 6598 : foreach_node(PathKey, pathkey, path_pathkeys)
+ + ]
5265 : : {
5266 : : /*
5267 : : * The PathKey nodes are canonical, so they can be checked for
5268 : : * equality by simple pointer comparison.
5269 : : */
5270 [ + + ]: 1286 : if (!list_member_ptr(needed_pathkeys, pathkey))
5271 : 5 : break;
5272 [ + + ]: 1281 : if (root->parse->hasDistinctOn &&
5273 [ + + ]: 100 : !list_member_ptr(root->distinct_pathkeys, pathkey))
5274 : 9 : break;
5275 : :
5276 : 1272 : useful_pathkeys = lappend(useful_pathkeys, pathkey);
5277 : : }
5278 : :
5279 : : /* If no match at all, no point in reordering needed_pathkeys */
5280 [ + + ]: 2663 : if (useful_pathkeys == NIL)
5281 : 1523 : return useful_pathkeys_list;
5282 : :
5283 : : /*
5284 : : * If not full match, the resulting pathkey list is not useful without
5285 : : * incremental sort.
5286 : : */
5287 [ + + ]: 1140 : if (list_length(useful_pathkeys) < list_length(needed_pathkeys) &&
5288 [ + + ]: 760 : !enable_incremental_sort)
5289 : 30 : return useful_pathkeys_list;
5290 : :
5291 : : /* Append the remaining PathKey nodes in needed_pathkeys */
5292 : 1110 : useful_pathkeys = list_concat_unique_ptr(useful_pathkeys,
5293 : : needed_pathkeys);
5294 : :
5295 : : /*
5296 : : * If the resulting pathkey list is the same as the 'needed_pathkeys',
5297 : : * just drop it.
5298 : : */
5299 [ + + ]: 1110 : if (compare_pathkeys(needed_pathkeys,
5300 : : useful_pathkeys) == PATHKEYS_EQUAL)
5301 : 704 : return useful_pathkeys_list;
5302 : :
5303 : 406 : useful_pathkeys_list = lappend(useful_pathkeys_list,
5304 : : useful_pathkeys);
5305 : :
5306 : 406 : return useful_pathkeys_list;
5307 : : }
5308 : :
5309 : : /*
5310 : : * create_ordered_paths
5311 : : *
5312 : : * Build a new upperrel containing Paths for ORDER BY evaluation.
5313 : : *
5314 : : * All paths in the result must satisfy the ORDER BY ordering.
5315 : : * The only new paths we need consider are an explicit full sort
5316 : : * and incremental sort on the cheapest-total existing path.
5317 : : *
5318 : : * input_rel: contains the source-data Paths
5319 : : * target: the output tlist the result Paths must emit
5320 : : * limit_tuples: estimated bound on the number of output tuples,
5321 : : * or -1 if no LIMIT or couldn't estimate
5322 : : *
5323 : : * XXX This only looks at sort_pathkeys. I wonder if it needs to look at the
5324 : : * other pathkeys (grouping, ...) like generate_useful_gather_paths.
5325 : : */
5326 : : static RelOptInfo *
3521 tgl@sss.pgh.pa.us 5327 : 37097 : create_ordered_paths(PlannerInfo *root,
5328 : : RelOptInfo *input_rel,
5329 : : PathTarget *target,
5330 : : bool target_parallel_safe,
5331 : : double limit_tuples)
5332 : : {
5333 : 37097 : Path *cheapest_input_path = input_rel->cheapest_total_path;
5334 : : RelOptInfo *ordered_rel;
5335 : : ListCell *lc;
5336 : :
5337 : : /* For now, do all work in the (ORDERED, NULL) upperrel */
5338 : 37097 : ordered_rel = fetch_upper_rel(root, UPPERREL_ORDERED, NULL);
5339 : :
5340 : : /*
5341 : : * If the input relation is not parallel-safe, then the ordered relation
5342 : : * can't be parallel-safe, either. Otherwise, it's parallel-safe if the
5343 : : * target list is parallel-safe.
5344 : : */
2790 rhaas@postgresql.org 5345 [ + + + + ]: 37097 : if (input_rel->consider_parallel && target_parallel_safe)
3405 5346 : 25842 : ordered_rel->consider_parallel = true;
5347 : :
5348 : : /*
5349 : : * If the input rel belongs to a single FDW, so does the ordered_rel.
5350 : : */
tgl@sss.pgh.pa.us 5351 : 37097 : ordered_rel->serverid = input_rel->serverid;
3391 5352 : 37097 : ordered_rel->userid = input_rel->userid;
5353 : 37097 : ordered_rel->useridiscurrent = input_rel->useridiscurrent;
3405 5354 : 37097 : ordered_rel->fdwroutine = input_rel->fdwroutine;
5355 : :
3521 5356 [ + - + + : 93537 : foreach(lc, input_rel->pathlist)
+ + ]
5357 : : {
2030 tomas.vondra@postgre 5358 : 56440 : Path *input_path = (Path *) lfirst(lc);
5359 : : Path *sorted_path;
5360 : : bool is_sorted;
5361 : : int presorted_keys;
5362 : :
5363 : 56440 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5364 : : input_path->pathkeys, &presorted_keys);
5365 : :
5366 [ + + ]: 56440 : if (is_sorted)
1046 drowley@postgresql.o 5367 : 21315 : sorted_path = input_path;
5368 : : else
5369 : : {
5370 : : /*
5371 : : * Try at least sorting the cheapest path and also try
5372 : : * incrementally sorting any path which is partially sorted
5373 : : * already (no need to deal with paths which have presorted keys
5374 : : * when incremental sort is disabled unless it's the cheapest
5375 : : * input path).
5376 : : */
5377 [ + + ]: 35125 : if (input_path != cheapest_input_path &&
5378 [ + + + + ]: 3009 : (presorted_keys == 0 || !enable_incremental_sort))
5379 : 981 : continue;
5380 : :
5381 : : /*
5382 : : * We've no need to consider both a sort and incremental sort.
5383 : : * We'll just do a sort if there are no presorted keys and an
5384 : : * incremental sort when there are presorted keys.
5385 : : */
5386 [ + + + + ]: 34144 : if (presorted_keys == 0 || !enable_incremental_sort)
2030 tomas.vondra@postgre 5387 : 31786 : sorted_path = (Path *) create_sort_path(root,
5388 : : ordered_rel,
5389 : : input_path,
5390 : : root->sort_pathkeys,
5391 : : limit_tuples);
5392 : : else
1046 drowley@postgresql.o 5393 : 2358 : sorted_path = (Path *) create_incremental_sort_path(root,
5394 : : ordered_rel,
5395 : : input_path,
5396 : : root->sort_pathkeys,
5397 : : presorted_keys,
5398 : : limit_tuples);
5399 : : }
5400 : :
5401 : : /*
5402 : : * If the pathtarget of the result path has different expressions from
5403 : : * the target to be applied, a projection step is needed.
5404 : : */
418 rguo@postgresql.org 5405 [ + + ]: 55459 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
1046 drowley@postgresql.o 5406 : 147 : sorted_path = apply_projection_to_path(root, ordered_rel,
5407 : : sorted_path, target);
5408 : :
5409 : 55459 : add_path(ordered_rel, sorted_path);
5410 : : }
5411 : :
5412 : : /*
5413 : : * generate_gather_paths() will have already generated a simple Gather
5414 : : * path for the best parallel path, if any, and the loop above will have
5415 : : * considered sorting it. Similarly, generate_gather_paths() will also
5416 : : * have generated order-preserving Gather Merge plans which can be used
5417 : : * without sorting if they happen to match the sort_pathkeys, and the loop
5418 : : * above will have handled those as well. However, there's one more
5419 : : * possibility: it may make sense to sort the cheapest partial path or
5420 : : * incrementally sort any partial path that is partially sorted according
5421 : : * to the required output order and then use Gather Merge.
5422 : : */
3154 rhaas@postgresql.org 5423 [ + + + + ]: 37097 : if (ordered_rel->consider_parallel && root->sort_pathkeys != NIL &&
5424 [ + + ]: 25740 : input_rel->partial_pathlist != NIL)
5425 : : {
5426 : : Path *cheapest_partial_path;
5427 : :
5428 : 1347 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
5429 : :
635 drowley@postgresql.o 5430 [ + - + + : 2797 : foreach(lc, input_rel->partial_pathlist)
+ + ]
5431 : : {
5432 : 1450 : Path *input_path = (Path *) lfirst(lc);
5433 : : Path *sorted_path;
5434 : : bool is_sorted;
5435 : : int presorted_keys;
5436 : : double total_groups;
5437 : :
5438 : 1450 : is_sorted = pathkeys_count_contained_in(root->sort_pathkeys,
5439 : : input_path->pathkeys,
5440 : : &presorted_keys);
5441 : :
5442 [ + + ]: 1450 : if (is_sorted)
5443 : 91 : continue;
5444 : :
5445 : : /*
5446 : : * Try at least sorting the cheapest path and also try
5447 : : * incrementally sorting any path which is partially sorted
5448 : : * already (no need to deal with paths which have presorted keys
5449 : : * when incremental sort is disabled unless it's the cheapest
5450 : : * partial path).
5451 : : */
5452 [ + + ]: 1359 : if (input_path != cheapest_partial_path &&
5453 [ + - - + ]: 21 : (presorted_keys == 0 || !enable_incremental_sort))
635 drowley@postgresql.o 5454 :UBC 0 : continue;
5455 : :
5456 : : /*
5457 : : * We've no need to consider both a sort and incremental sort.
5458 : : * We'll just do a sort if there are no presorted keys and an
5459 : : * incremental sort when there are presorted keys.
5460 : : */
635 drowley@postgresql.o 5461 [ + + + + ]:CBC 1359 : if (presorted_keys == 0 || !enable_incremental_sort)
5462 : 1329 : sorted_path = (Path *) create_sort_path(root,
5463 : : ordered_rel,
5464 : : input_path,
5465 : : root->sort_pathkeys,
5466 : : limit_tuples);
5467 : : else
2029 tomas.vondra@postgre 5468 : 30 : sorted_path = (Path *) create_incremental_sort_path(root,
5469 : : ordered_rel,
5470 : : input_path,
5471 : : root->sort_pathkeys,
5472 : : presorted_keys,
5473 : : limit_tuples);
461 rguo@postgresql.org 5474 : 1359 : total_groups = compute_gather_rows(sorted_path);
5475 : : sorted_path = (Path *)
635 drowley@postgresql.o 5476 : 1359 : create_gather_merge_path(root, ordered_rel,
5477 : : sorted_path,
5478 : : sorted_path->pathtarget,
5479 : : root->sort_pathkeys, NULL,
5480 : : &total_groups);
5481 : :
5482 : : /*
5483 : : * If the pathtarget of the result path has different expressions
5484 : : * from the target to be applied, a projection step is needed.
5485 : : */
418 rguo@postgresql.org 5486 [ + + ]: 1359 : if (!equal(sorted_path->pathtarget->exprs, target->exprs))
635 drowley@postgresql.o 5487 : 3 : sorted_path = apply_projection_to_path(root, ordered_rel,
5488 : : sorted_path, target);
5489 : :
5490 : 1359 : add_path(ordered_rel, sorted_path);
5491 : : }
5492 : : }
5493 : :
5494 : : /*
5495 : : * If there is an FDW that's responsible for all baserels of the query,
5496 : : * let it consider adding ForeignPaths.
5497 : : */
3405 tgl@sss.pgh.pa.us 5498 [ + + ]: 37097 : if (ordered_rel->fdwroutine &&
5499 [ + + ]: 192 : ordered_rel->fdwroutine->GetForeignUpperPaths)
5500 : 185 : ordered_rel->fdwroutine->GetForeignUpperPaths(root, UPPERREL_ORDERED,
5501 : : input_rel, ordered_rel,
5502 : : NULL);
5503 : :
5504 : : /* Let extensions possibly add some more paths */
3485 5505 [ - + ]: 37097 : if (create_upper_paths_hook)
3485 tgl@sss.pgh.pa.us 5506 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_ORDERED,
5507 : : input_rel, ordered_rel, NULL);
5508 : :
5509 : : /*
5510 : : * No need to bother with set_cheapest here; grouping_planner does not
5511 : : * need us to do it.
5512 : : */
3521 tgl@sss.pgh.pa.us 5513 [ - + ]:CBC 37097 : Assert(ordered_rel->pathlist != NIL);
5514 : :
5515 : 37097 : return ordered_rel;
5516 : : }
5517 : :
5518 : :
5519 : : /*
5520 : : * make_group_input_target
5521 : : * Generate appropriate PathTarget for initial input to grouping nodes.
5522 : : *
5523 : : * If there is grouping or aggregation, the scan/join subplan cannot emit
5524 : : * the query's final targetlist; for example, it certainly can't emit any
5525 : : * aggregate function calls. This routine generates the correct target
5526 : : * for the scan/join subplan.
5527 : : *
5528 : : * The query target list passed from the parser already contains entries
5529 : : * for all ORDER BY and GROUP BY expressions, but it will not have entries
5530 : : * for variables used only in HAVING clauses; so we need to add those
5531 : : * variables to the subplan target list. Also, we flatten all expressions
5532 : : * except GROUP BY items into their component variables; other expressions
5533 : : * will be computed by the upper plan nodes rather than by the subplan.
5534 : : * For example, given a query like
5535 : : * SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
5536 : : * we want to pass this targetlist to the subplan:
5537 : : * a+b,c,d
5538 : : * where the a+b target will be used by the Sort/Group steps, and the
5539 : : * other targets will be used for computing the final results.
5540 : : *
5541 : : * 'final_target' is the query's final target list (in PathTarget form)
5542 : : *
5543 : : * The result is the PathTarget to be computed by the Paths returned from
5544 : : * query_planner().
5545 : : */
5546 : : static PathTarget *
3517 5547 : 18687 : make_group_input_target(PlannerInfo *root, PathTarget *final_target)
5548 : : {
7449 5549 : 18687 : Query *parse = root->parse;
5550 : : PathTarget *input_target;
5551 : : List *non_group_cols;
5552 : : List *non_group_vars;
5553 : : int i;
5554 : : ListCell *lc;
5555 : :
5556 : : /*
5557 : : * We must build a target containing all grouping columns, plus any other
5558 : : * Vars mentioned in the query's targetlist and HAVING qual.
5559 : : */
3517 5560 : 18687 : input_target = create_empty_pathtarget();
5217 5561 : 18687 : non_group_cols = NIL;
5562 : :
3517 5563 : 18687 : i = 0;
5564 [ + + + + : 46283 : foreach(lc, final_target->exprs)
+ + ]
5565 : : {
5566 : 27596 : Expr *expr = (Expr *) lfirst(lc);
3423 5567 [ + - ]: 27596 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
5568 : :
1013 5569 [ + + + + : 32163 : if (sgref && root->processed_groupClause &&
+ + ]
5570 : 4567 : get_sortgroupref_clause_noerr(sgref,
5571 : : root->processed_groupClause) != NULL)
5572 : : {
5573 : : /*
5574 : : * It's a grouping column, so add it to the input target as-is.
5575 : : *
5576 : : * Note that the target is logically below the grouping step. So
5577 : : * with grouping sets we need to remove the RT index of the
5578 : : * grouping step if there is any from the target expression.
5579 : : */
412 rguo@postgresql.org 5580 [ + - + + ]: 3689 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5581 : : {
5582 [ - + ]: 948 : Assert(root->group_rtindex > 0);
5583 : : expr = (Expr *)
5584 : 948 : remove_nulling_relids((Node *) expr,
5585 : 948 : bms_make_singleton(root->group_rtindex),
5586 : : NULL);
5587 : : }
3517 tgl@sss.pgh.pa.us 5588 : 3689 : add_column_to_pathtarget(input_target, expr, sgref);
5589 : : }
5590 : : else
5591 : : {
5592 : : /*
5593 : : * Non-grouping column, so just remember the expression for later
5594 : : * call to pull_var_clause.
5595 : : */
5596 : 23907 : non_group_cols = lappend(non_group_cols, expr);
5597 : : }
5598 : :
5599 : 27596 : i++;
5600 : : }
5601 : :
5602 : : /*
5603 : : * If there's a HAVING clause, we'll need the Vars it uses, too.
5604 : : */
5217 5605 [ + + ]: 18687 : if (parse->havingQual)
5606 : 478 : non_group_cols = lappend(non_group_cols, parse->havingQual);
5607 : :
5608 : : /*
5609 : : * Pull out all the Vars mentioned in non-group cols (plus HAVING), and
5610 : : * add them to the input target if not already present. (A Var used
5611 : : * directly as a GROUP BY item will be present already.) Note this
5612 : : * includes Vars used in resjunk items, so we are covering the needs of
5613 : : * ORDER BY and window specifications. Vars used within Aggrefs and
5614 : : * WindowFuncs will be pulled out here, too.
5615 : : *
5616 : : * Note that the target is logically below the grouping step. So with
5617 : : * grouping sets we need to remove the RT index of the grouping step if
5618 : : * there is any from the non-group Vars.
5619 : : */
5620 : 18687 : non_group_vars = pull_var_clause((Node *) non_group_cols,
5621 : : PVC_RECURSE_AGGREGATES |
5622 : : PVC_RECURSE_WINDOWFUNCS |
5623 : : PVC_INCLUDE_PLACEHOLDERS);
412 rguo@postgresql.org 5624 [ + + + + ]: 18687 : if (parse->hasGroupRTE && parse->groupingSets != NIL)
5625 : : {
5626 [ - + ]: 433 : Assert(root->group_rtindex > 0);
5627 : : non_group_vars = (List *)
5628 : 433 : remove_nulling_relids((Node *) non_group_vars,
5629 : 433 : bms_make_singleton(root->group_rtindex),
5630 : : NULL);
5631 : : }
3517 tgl@sss.pgh.pa.us 5632 : 18687 : add_new_columns_to_pathtarget(input_target, non_group_vars);
5633 : :
5634 : : /* clean up cruft */
5217 5635 : 18687 : list_free(non_group_vars);
5636 : 18687 : list_free(non_group_cols);
5637 : :
5638 : : /* XXX this causes some redundant cost calculation ... */
3517 5639 : 18687 : return set_pathtarget_cost_width(root, input_target);
5640 : : }
5641 : :
5642 : : /*
5643 : : * make_partial_grouping_target
5644 : : * Generate appropriate PathTarget for output of partial aggregate
5645 : : * (or partial grouping, if there are no aggregates) nodes.
5646 : : *
5647 : : * A partial aggregation node needs to emit all the same aggregates that
5648 : : * a regular aggregation node would, plus any aggregates used in HAVING;
5649 : : * except that the Aggref nodes should be marked as partial aggregates.
5650 : : *
5651 : : * In addition, we'd better emit any Vars and PlaceHolderVars that are
5652 : : * used outside of Aggrefs in the aggregation tlist and HAVING. (Presumably,
5653 : : * these would be Vars that are grouped by or used in grouping expressions.)
5654 : : *
5655 : : * grouping_target is the tlist to be emitted by the topmost aggregation step.
5656 : : * havingQual represents the HAVING clause.
5657 : : */
5658 : : static PathTarget *
2783 rhaas@postgresql.org 5659 : 1550 : make_partial_grouping_target(PlannerInfo *root,
5660 : : PathTarget *grouping_target,
5661 : : Node *havingQual)
5662 : : {
5663 : : PathTarget *partial_target;
5664 : : List *non_group_cols;
5665 : : List *non_group_exprs;
5666 : : int i;
5667 : : ListCell *lc;
5668 : :
3410 tgl@sss.pgh.pa.us 5669 : 1550 : partial_target = create_empty_pathtarget();
3507 rhaas@postgresql.org 5670 : 1550 : non_group_cols = NIL;
5671 : :
5672 : 1550 : i = 0;
3410 tgl@sss.pgh.pa.us 5673 [ + - + + : 5615 : foreach(lc, grouping_target->exprs)
+ + ]
5674 : : {
3507 rhaas@postgresql.org 5675 : 4065 : Expr *expr = (Expr *) lfirst(lc);
3410 tgl@sss.pgh.pa.us 5676 [ + - ]: 4065 : Index sgref = get_pathtarget_sortgroupref(grouping_target, i);
5677 : :
1013 5678 [ + + + + : 6447 : if (sgref && root->processed_groupClause &&
+ + ]
5679 : 2382 : get_sortgroupref_clause_noerr(sgref,
5680 : : root->processed_groupClause) != NULL)
5681 : : {
5682 : : /*
5683 : : * It's a grouping column, so add it to the partial_target as-is.
5684 : : * (This allows the upper agg step to repeat the grouping calcs.)
5685 : : */
3410 5686 : 1411 : add_column_to_pathtarget(partial_target, expr, sgref);
5687 : : }
5688 : : else
5689 : : {
5690 : : /*
5691 : : * Non-grouping column, so just remember the expression for later
5692 : : * call to pull_var_clause.
5693 : : */
3507 rhaas@postgresql.org 5694 : 2654 : non_group_cols = lappend(non_group_cols, expr);
5695 : : }
5696 : :
5697 : 4065 : i++;
5698 : : }
5699 : :
5700 : : /*
5701 : : * If there's a HAVING clause, we'll need the Vars/Aggrefs it uses, too.
5702 : : */
2783 5703 [ + + ]: 1550 : if (havingQual)
5704 : 439 : non_group_cols = lappend(non_group_cols, havingQual);
5705 : :
5706 : : /*
5707 : : * Pull out all the Vars, PlaceHolderVars, and Aggrefs mentioned in
5708 : : * non-group cols (plus HAVING), and add them to the partial_target if not
5709 : : * already present. (An expression used directly as a GROUP BY item will
5710 : : * be present already.) Note this includes Vars used in resjunk items, so
5711 : : * we are covering the needs of ORDER BY and window specifications.
5712 : : */
3507 5713 : 1550 : non_group_exprs = pull_var_clause((Node *) non_group_cols,
5714 : : PVC_INCLUDE_AGGREGATES |
5715 : : PVC_RECURSE_WINDOWFUNCS |
5716 : : PVC_INCLUDE_PLACEHOLDERS);
5717 : :
3410 tgl@sss.pgh.pa.us 5718 : 1550 : add_new_columns_to_pathtarget(partial_target, non_group_exprs);
5719 : :
5720 : : /*
5721 : : * Adjust Aggrefs to put them in partial mode. At this point all Aggrefs
5722 : : * are at the top level of the target list, so we can just scan the list
5723 : : * rather than recursing through the expression trees.
5724 : : */
5725 [ + - + + : 5913 : foreach(lc, partial_target->exprs)
+ + ]
5726 : : {
5727 : 4363 : Aggref *aggref = (Aggref *) lfirst(lc);
5728 : :
5729 [ + + ]: 4363 : if (IsA(aggref, Aggref))
5730 : : {
5731 : : Aggref *newaggref;
5732 : :
5733 : : /*
5734 : : * We shouldn't need to copy the substructure of the Aggref node,
5735 : : * but flat-copy the node itself to avoid damaging other trees.
5736 : : */
5737 : 2937 : newaggref = makeNode(Aggref);
5738 : 2937 : memcpy(newaggref, aggref, sizeof(Aggref));
5739 : :
5740 : : /* For now, assume serialization is required */
5741 : 2937 : mark_partial_aggref(newaggref, AGGSPLIT_INITIAL_SERIAL);
5742 : :
5743 : 2937 : lfirst(lc) = newaggref;
5744 : : }
5745 : : }
5746 : :
5747 : : /* clean up cruft */
3507 rhaas@postgresql.org 5748 : 1550 : list_free(non_group_exprs);
5749 : 1550 : list_free(non_group_cols);
5750 : :
5751 : : /* XXX this causes some redundant cost calculation ... */
3410 tgl@sss.pgh.pa.us 5752 : 1550 : return set_pathtarget_cost_width(root, partial_target);
5753 : : }
5754 : :
5755 : : /*
5756 : : * mark_partial_aggref
5757 : : * Adjust an Aggref to make it represent a partial-aggregation step.
5758 : : *
5759 : : * The Aggref node is modified in-place; caller must do any copying required.
5760 : : */
5761 : : void
5762 : 8856 : mark_partial_aggref(Aggref *agg, AggSplit aggsplit)
5763 : : {
5764 : : /* aggtranstype should be computed by this point */
5765 [ - + ]: 8856 : Assert(OidIsValid(agg->aggtranstype));
5766 : : /* ... but aggsplit should still be as the parser left it */
5767 [ - + ]: 8856 : Assert(agg->aggsplit == AGGSPLIT_SIMPLE);
5768 : :
5769 : : /* Mark the Aggref with the intended partial-aggregation mode */
5770 : 8856 : agg->aggsplit = aggsplit;
5771 : :
5772 : : /*
5773 : : * Adjust result type if needed. Normally, a partial aggregate returns
5774 : : * the aggregate's transition type; but if that's INTERNAL and we're
5775 : : * serializing, it returns BYTEA instead.
5776 : : */
5777 [ + + ]: 8856 : if (DO_AGGSPLIT_SKIPFINAL(aggsplit))
5778 : : {
5779 [ + + + - ]: 7730 : if (agg->aggtranstype == INTERNALOID && DO_AGGSPLIT_SERIALIZE(aggsplit))
5780 : 157 : agg->aggtype = BYTEAOID;
5781 : : else
5782 : 7573 : agg->aggtype = agg->aggtranstype;
5783 : : }
3507 rhaas@postgresql.org 5784 : 8856 : }
5785 : :
5786 : : /*
5787 : : * postprocess_setop_tlist
5788 : : * Fix up targetlist returned by plan_set_operations().
5789 : : *
5790 : : * We need to transpose sort key info from the orig_tlist into new_tlist.
5791 : : * NOTE: this would not be good enough if we supported resjunk sort keys
5792 : : * for results of set operations --- then, we'd need to project a whole
5793 : : * new tlist to evaluate the resjunk columns. For now, just ereport if we
5794 : : * find any resjunk columns in orig_tlist.
5795 : : */
5796 : : static List *
9118 tgl@sss.pgh.pa.us 5797 : 3073 : postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
5798 : : {
5799 : : ListCell *l;
7824 neilc@samurai.com 5800 : 3073 : ListCell *orig_tlist_item = list_head(orig_tlist);
5801 : :
9118 tgl@sss.pgh.pa.us 5802 [ + + + + : 11775 : foreach(l, new_tlist)
+ + ]
5803 : : {
2974 5804 : 8702 : TargetEntry *new_tle = lfirst_node(TargetEntry, l);
5805 : : TargetEntry *orig_tle;
5806 : :
5807 : : /* ignore resjunk columns in setop result */
7509 5808 [ - + ]: 8702 : if (new_tle->resjunk)
9118 tgl@sss.pgh.pa.us 5809 :UBC 0 : continue;
5810 : :
7824 neilc@samurai.com 5811 [ - + ]:CBC 8702 : Assert(orig_tlist_item != NULL);
2974 tgl@sss.pgh.pa.us 5812 : 8702 : orig_tle = lfirst_node(TargetEntry, orig_tlist_item);
2296 5813 : 8702 : orig_tlist_item = lnext(orig_tlist, orig_tlist_item);
7317 bruce@momjian.us 5814 [ - + ]: 8702 : if (orig_tle->resjunk) /* should not happen */
8130 tgl@sss.pgh.pa.us 5815 [ # # ]:UBC 0 : elog(ERROR, "resjunk output columns are not implemented");
7509 tgl@sss.pgh.pa.us 5816 [ - + ]:CBC 8702 : Assert(new_tle->resno == orig_tle->resno);
5817 : 8702 : new_tle->ressortgroupref = orig_tle->ressortgroupref;
5818 : : }
7824 neilc@samurai.com 5819 [ - + ]: 3073 : if (orig_tlist_item != NULL)
8130 tgl@sss.pgh.pa.us 5820 [ # # ]:UBC 0 : elog(ERROR, "resjunk output columns are not implemented");
9118 tgl@sss.pgh.pa.us 5821 :CBC 3073 : return new_tlist;
5822 : : }
5823 : :
5824 : : /*
5825 : : * optimize_window_clauses
5826 : : * Call each WindowFunc's prosupport function to see if we're able to
5827 : : * make any adjustments to any of the WindowClause's so that the executor
5828 : : * can execute the window functions in a more optimal way.
5829 : : *
5830 : : * Currently we only allow adjustments to the WindowClause's frameOptions. We
5831 : : * may allow more things to be done here in the future.
5832 : : */
5833 : : static void
1039 drowley@postgresql.o 5834 : 1282 : optimize_window_clauses(PlannerInfo *root, WindowFuncLists *wflists)
5835 : : {
5836 : 1282 : List *windowClause = root->parse->windowClause;
5837 : : ListCell *lc;
5838 : :
5839 [ + - + + : 2687 : foreach(lc, windowClause)
+ + ]
5840 : : {
5841 : 1405 : WindowClause *wc = lfirst_node(WindowClause, lc);
5842 : : ListCell *lc2;
5843 : 1405 : int optimizedFrameOptions = 0;
5844 : :
5845 [ - + ]: 1405 : Assert(wc->winref <= wflists->maxWinRef);
5846 : :
5847 : : /* skip any WindowClauses that have no WindowFuncs */
5848 [ + + ]: 1405 : if (wflists->windowFuncs[wc->winref] == NIL)
5849 : 12 : continue;
5850 : :
5851 [ + - + + : 1720 : foreach(lc2, wflists->windowFuncs[wc->winref])
+ + ]
5852 : : {
5853 : : SupportRequestOptimizeWindowClause req;
5854 : : SupportRequestOptimizeWindowClause *res;
5855 : 1414 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc2);
5856 : : Oid prosupport;
5857 : :
5858 : 1414 : prosupport = get_func_support(wfunc->winfnoid);
5859 : :
5860 : : /* Check if there's a support function for 'wfunc' */
5861 [ + + ]: 1414 : if (!OidIsValid(prosupport))
5862 : 1087 : break; /* can't optimize this WindowClause */
5863 : :
5864 : 440 : req.type = T_SupportRequestOptimizeWindowClause;
5865 : 440 : req.window_clause = wc;
5866 : 440 : req.window_func = wfunc;
5867 : 440 : req.frameOptions = wc->frameOptions;
5868 : :
5869 : : /* call the support function */
5870 : : res = (SupportRequestOptimizeWindowClause *)
5871 : 440 : DatumGetPointer(OidFunctionCall1(prosupport,
5872 : : PointerGetDatum(&req)));
5873 : :
5874 : : /*
5875 : : * Skip to next WindowClause if the support function does not
5876 : : * support this request type.
5877 : : */
5878 [ + + ]: 440 : if (res == NULL)
5879 : 113 : break;
5880 : :
5881 : : /*
5882 : : * Save these frameOptions for the first WindowFunc for this
5883 : : * WindowClause.
5884 : : */
5885 [ + + ]: 327 : if (foreach_current_index(lc2) == 0)
5886 : 315 : optimizedFrameOptions = res->frameOptions;
5887 : :
5888 : : /*
5889 : : * On subsequent WindowFuncs, if the frameOptions are not the same
5890 : : * then we're unable to optimize the frameOptions for this
5891 : : * WindowClause.
5892 : : */
5893 [ - + ]: 12 : else if (optimizedFrameOptions != res->frameOptions)
1039 drowley@postgresql.o 5894 :UBC 0 : break; /* skip to the next WindowClause, if any */
5895 : : }
5896 : :
5897 : : /* adjust the frameOptions if all WindowFunc's agree that it's ok */
1039 drowley@postgresql.o 5898 [ + + + - ]:CBC 1393 : if (lc2 == NULL && wc->frameOptions != optimizedFrameOptions)
5899 : : {
5900 : : ListCell *lc3;
5901 : :
5902 : : /* apply the new frame options */
5903 : 306 : wc->frameOptions = optimizedFrameOptions;
5904 : :
5905 : : /*
5906 : : * We now check to see if changing the frameOptions has caused
5907 : : * this WindowClause to be a duplicate of some other WindowClause.
5908 : : * This can only happen if we have multiple WindowClauses, so
5909 : : * don't bother if there's only 1.
5910 : : */
5911 [ + + ]: 306 : if (list_length(windowClause) == 1)
5912 : 261 : continue;
5913 : :
5914 : : /*
5915 : : * Do the duplicate check and reuse the existing WindowClause if
5916 : : * we find a duplicate.
5917 : : */
5918 [ + - + + : 114 : foreach(lc3, windowClause)
+ + ]
5919 : : {
5920 : 87 : WindowClause *existing_wc = lfirst_node(WindowClause, lc3);
5921 : :
5922 : : /* skip over the WindowClause we're currently editing */
5923 [ + + ]: 87 : if (existing_wc == wc)
5924 : 27 : continue;
5925 : :
5926 : : /*
5927 : : * Perform the same duplicate check that is done in
5928 : : * transformWindowFuncCall.
5929 : : */
5930 [ + - + + ]: 120 : if (equal(wc->partitionClause, existing_wc->partitionClause) &&
5931 : 60 : equal(wc->orderClause, existing_wc->orderClause) &&
5932 [ + + + - ]: 60 : wc->frameOptions == existing_wc->frameOptions &&
5933 [ + - ]: 36 : equal(wc->startOffset, existing_wc->startOffset) &&
5934 : 18 : equal(wc->endOffset, existing_wc->endOffset))
5935 : : {
5936 : : ListCell *lc4;
5937 : :
5938 : : /*
5939 : : * Now move each WindowFunc in 'wc' into 'existing_wc'.
5940 : : * This required adjusting each WindowFunc's winref and
5941 : : * moving the WindowFuncs in 'wc' to the list of
5942 : : * WindowFuncs in 'existing_wc'.
5943 : : */
5944 [ + - + + : 39 : foreach(lc4, wflists->windowFuncs[wc->winref])
+ + ]
5945 : : {
5946 : 21 : WindowFunc *wfunc = lfirst_node(WindowFunc, lc4);
5947 : :
5948 : 21 : wfunc->winref = existing_wc->winref;
5949 : : }
5950 : :
5951 : : /* move list items */
5952 : 36 : wflists->windowFuncs[existing_wc->winref] = list_concat(wflists->windowFuncs[existing_wc->winref],
5953 : 18 : wflists->windowFuncs[wc->winref]);
5954 : 18 : wflists->windowFuncs[wc->winref] = NIL;
5955 : :
5956 : : /*
5957 : : * transformWindowFuncCall() should have made sure there
5958 : : * are no other duplicates, so we needn't bother looking
5959 : : * any further.
5960 : : */
5961 : 18 : break;
5962 : : }
5963 : : }
5964 : : }
5965 : : }
5966 : 1282 : }
5967 : :
5968 : : /*
5969 : : * select_active_windows
5970 : : * Create a list of the "active" window clauses (ie, those referenced
5971 : : * by non-deleted WindowFuncs) in the order they are to be executed.
5972 : : */
5973 : : static List *
6147 tgl@sss.pgh.pa.us 5974 : 1282 : select_active_windows(PlannerInfo *root, WindowFuncLists *wflists)
5975 : : {
2600 rhodiumtoad@postgres 5976 : 1282 : List *windowClause = root->parse->windowClause;
5977 : 1282 : List *result = NIL;
5978 : : ListCell *lc;
5979 : 1282 : int nActive = 0;
5980 : 1282 : WindowClauseSortData *actives = palloc(sizeof(WindowClauseSortData)
5981 : 1282 : * list_length(windowClause));
5982 : :
5983 : : /* First, construct an array of the active windows */
5984 [ + - + + : 2687 : foreach(lc, windowClause)
+ + ]
5985 : : {
2974 tgl@sss.pgh.pa.us 5986 : 1405 : WindowClause *wc = lfirst_node(WindowClause, lc);
5987 : :
5988 : : /* It's only active if wflists shows some related WindowFuncs */
6147 5989 [ - + ]: 1405 : Assert(wc->winref <= wflists->maxWinRef);
2600 rhodiumtoad@postgres 5990 [ + + ]: 1405 : if (wflists->windowFuncs[wc->winref] == NIL)
5991 : 30 : continue;
5992 : :
5993 : 1375 : actives[nActive].wc = wc; /* original clause */
5994 : :
5995 : : /*
5996 : : * For sorting, we want the list of partition keys followed by the
5997 : : * list of sort keys. But pathkeys construction will remove duplicates
5998 : : * between the two, so we can as well (even though we can't detect all
5999 : : * of the duplicates, since some may come from ECs - that might mean
6000 : : * we miss optimization chances here). We must, however, ensure that
6001 : : * the order of entries is preserved with respect to the ones we do
6002 : : * keep.
6003 : : *
6004 : : * partitionClause and orderClause had their own duplicates removed in
6005 : : * parse analysis, so we're only concerned here with removing
6006 : : * orderClause entries that also appear in partitionClause.
6007 : : */
6008 : 2750 : actives[nActive].uniqueOrder =
6009 : 1375 : list_concat_unique(list_copy(wc->partitionClause),
6010 : 1375 : wc->orderClause);
6011 : 1375 : nActive++;
6012 : : }
6013 : :
6014 : : /*
6015 : : * Sort active windows by their partitioning/ordering clauses, ignoring
6016 : : * any framing clauses, so that the windows that need the same sorting are
6017 : : * adjacent in the list. When we come to generate paths, this will avoid
6018 : : * inserting additional Sort nodes.
6019 : : *
6020 : : * This is how we implement a specific requirement from the SQL standard,
6021 : : * which says that when two or more windows are order-equivalent (i.e.
6022 : : * have matching partition and order clauses, even if their names or
6023 : : * framing clauses differ), then all peer rows must be presented in the
6024 : : * same order in all of them. If we allowed multiple sort nodes for such
6025 : : * cases, we'd risk having the peer rows end up in different orders in
6026 : : * equivalent windows due to sort instability. (See General Rule 4 of
6027 : : * <window clause> in SQL2008 - SQL2016.)
6028 : : *
6029 : : * Additionally, if the entire list of clauses of one window is a prefix
6030 : : * of another, put first the window with stronger sorting requirements.
6031 : : * This way we will first sort for stronger window, and won't have to sort
6032 : : * again for the weaker one.
6033 : : */
6034 : 1282 : qsort(actives, nActive, sizeof(WindowClauseSortData), common_prefix_cmp);
6035 : :
6036 : : /* build ordered list of the original WindowClause nodes */
6037 [ + + ]: 2657 : for (int i = 0; i < nActive; i++)
6038 : 1375 : result = lappend(result, actives[i].wc);
6039 : :
6040 : 1282 : pfree(actives);
6041 : :
6042 : 1282 : return result;
6043 : : }
6044 : :
6045 : : /*
6046 : : * name_active_windows
6047 : : * Ensure all active windows have unique names.
6048 : : *
6049 : : * The parser will have checked that user-assigned window names are unique
6050 : : * within the Query. Here we assign made-up names to any unnamed
6051 : : * WindowClauses for the benefit of EXPLAIN. (We don't want to do this
6052 : : * at parse time, because it'd mess up decompilation of views.)
6053 : : *
6054 : : * activeWindows: result of select_active_windows
6055 : : */
6056 : : static void
230 tgl@sss.pgh.pa.us 6057 : 1282 : name_active_windows(List *activeWindows)
6058 : : {
6059 : 1282 : int next_n = 1;
6060 : : char newname[16];
6061 : : ListCell *lc;
6062 : :
6063 [ + - + + : 2657 : foreach(lc, activeWindows)
+ + ]
6064 : : {
6065 : 1375 : WindowClause *wc = lfirst_node(WindowClause, lc);
6066 : :
6067 : : /* Nothing to do if it has a name already. */
6068 [ + + ]: 1375 : if (wc->name)
6069 : 288 : continue;
6070 : :
6071 : : /* Select a name not currently present in the list. */
6072 : : for (;;)
6073 : 3 : {
6074 : : ListCell *lc2;
6075 : :
6076 : 1090 : snprintf(newname, sizeof(newname), "w%d", next_n++);
6077 [ + - + + : 2354 : foreach(lc2, activeWindows)
+ + ]
6078 : : {
6079 : 1267 : WindowClause *wc2 = lfirst_node(WindowClause, lc2);
6080 : :
6081 [ + + + + ]: 1267 : if (wc2->name && strcmp(wc2->name, newname) == 0)
6082 : 3 : break; /* matched */
6083 : : }
6084 [ + + ]: 1090 : if (lc2 == NULL)
6085 : 1087 : break; /* reached the end with no match */
6086 : : }
6087 : 1087 : wc->name = pstrdup(newname);
6088 : : }
6089 : 1282 : }
6090 : :
6091 : : /*
6092 : : * common_prefix_cmp
6093 : : * QSort comparison function for WindowClauseSortData
6094 : : *
6095 : : * Sort the windows by the required sorting clauses. First, compare the sort
6096 : : * clauses themselves. Second, if one window's clauses are a prefix of another
6097 : : * one's clauses, put the window with more sort clauses first.
6098 : : *
6099 : : * We purposefully sort by the highest tleSortGroupRef first. Since
6100 : : * tleSortGroupRefs are assigned for the query's DISTINCT and ORDER BY first
6101 : : * and because here we sort the lowest tleSortGroupRefs last, if a
6102 : : * WindowClause is sharing a tleSortGroupRef with the query's DISTINCT or
6103 : : * ORDER BY clause, this makes it more likely that the final WindowAgg will
6104 : : * provide presorted input for the query's DISTINCT or ORDER BY clause, thus
6105 : : * reducing the total number of sorts required for the query.
6106 : : */
6107 : : static int
2600 rhodiumtoad@postgres 6108 : 102 : common_prefix_cmp(const void *a, const void *b)
6109 : : {
6110 : 102 : const WindowClauseSortData *wcsa = a;
6111 : 102 : const WindowClauseSortData *wcsb = b;
6112 : : ListCell *item_a;
6113 : : ListCell *item_b;
6114 : :
6115 [ + + + + : 183 : forboth(item_a, wcsa->uniqueOrder, item_b, wcsb->uniqueOrder)
+ + + + +
+ + + +
+ ]
6116 : : {
6117 : 132 : SortGroupClause *sca = lfirst_node(SortGroupClause, item_a);
6118 : 132 : SortGroupClause *scb = lfirst_node(SortGroupClause, item_b);
6119 : :
6120 [ + + ]: 132 : if (sca->tleSortGroupRef > scb->tleSortGroupRef)
6121 : 51 : return -1;
6122 [ + + ]: 126 : else if (sca->tleSortGroupRef < scb->tleSortGroupRef)
6123 : 33 : return 1;
6124 [ - + ]: 93 : else if (sca->sortop > scb->sortop)
2600 rhodiumtoad@postgres 6125 :UBC 0 : return -1;
2600 rhodiumtoad@postgres 6126 [ + + ]:CBC 93 : else if (sca->sortop < scb->sortop)
6127 : 12 : return 1;
6128 [ - + - - ]: 81 : else if (sca->nulls_first && !scb->nulls_first)
2600 rhodiumtoad@postgres 6129 :UBC 0 : return -1;
2600 rhodiumtoad@postgres 6130 [ + - - + ]:CBC 81 : else if (!sca->nulls_first && scb->nulls_first)
2600 rhodiumtoad@postgres 6131 :UBC 0 : return 1;
6132 : : /* no need to compare eqop, since it is fully determined by sortop */
6133 : : }
6134 : :
2600 rhodiumtoad@postgres 6135 [ + + ]:CBC 51 : if (list_length(wcsa->uniqueOrder) > list_length(wcsb->uniqueOrder))
6136 : 3 : return -1;
6137 [ + + ]: 48 : else if (list_length(wcsa->uniqueOrder) < list_length(wcsb->uniqueOrder))
6138 : 15 : return 1;
6139 : :
6140 : 33 : return 0;
6141 : : }
6142 : :
6143 : : /*
6144 : : * make_window_input_target
6145 : : * Generate appropriate PathTarget for initial input to WindowAgg nodes.
6146 : : *
6147 : : * When the query has window functions, this function computes the desired
6148 : : * target to be computed by the node just below the first WindowAgg.
6149 : : * This tlist must contain all values needed to evaluate the window functions,
6150 : : * compute the final target list, and perform any required final sort step.
6151 : : * If multiple WindowAggs are needed, each intermediate one adds its window
6152 : : * function results onto this base tlist; only the topmost WindowAgg computes
6153 : : * the actual desired target list.
6154 : : *
6155 : : * This function is much like make_group_input_target, though not quite enough
6156 : : * like it to share code. As in that function, we flatten most expressions
6157 : : * into their component variables. But we do not want to flatten window
6158 : : * PARTITION BY/ORDER BY clauses, since that might result in multiple
6159 : : * evaluations of them, which would be bad (possibly even resulting in
6160 : : * inconsistent answers, if they contain volatile functions).
6161 : : * Also, we must not flatten GROUP BY clauses that were left unflattened by
6162 : : * make_group_input_target, because we may no longer have access to the
6163 : : * individual Vars in them.
6164 : : *
6165 : : * Another key difference from make_group_input_target is that we don't
6166 : : * flatten Aggref expressions, since those are to be computed below the
6167 : : * window functions and just referenced like Vars above that.
6168 : : *
6169 : : * 'final_target' is the query's final target list (in PathTarget form)
6170 : : * 'activeWindows' is the list of active windows previously identified by
6171 : : * select_active_windows.
6172 : : *
6173 : : * The result is the PathTarget to be computed by the plan node immediately
6174 : : * below the first WindowAgg node.
6175 : : */
6176 : : static PathTarget *
3519 tgl@sss.pgh.pa.us 6177 : 1282 : make_window_input_target(PlannerInfo *root,
6178 : : PathTarget *final_target,
6179 : : List *activeWindows)
6180 : : {
6181 : : PathTarget *input_target;
6182 : : Bitmapset *sgrefs;
6183 : : List *flattenable_cols;
6184 : : List *flattenable_vars;
6185 : : int i;
6186 : : ListCell *lc;
6187 : :
1013 6188 [ - + ]: 1282 : Assert(root->parse->hasWindowFuncs);
6189 : :
6190 : : /*
6191 : : * Collect the sortgroupref numbers of window PARTITION/ORDER BY clauses
6192 : : * into a bitmapset for convenient reference below.
6193 : : */
4792 6194 : 1282 : sgrefs = NULL;
6055 6195 [ + - + + : 2657 : foreach(lc, activeWindows)
+ + ]
6196 : : {
2974 6197 : 1375 : WindowClause *wc = lfirst_node(WindowClause, lc);
6198 : : ListCell *lc2;
6199 : :
6055 6200 [ + + + + : 1753 : foreach(lc2, wc->partitionClause)
+ + ]
6201 : : {
2974 6202 : 378 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6203 : :
6055 6204 : 378 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6205 : : }
6206 [ + + + + : 2505 : foreach(lc2, wc->orderClause)
+ + ]
6207 : : {
2974 6208 : 1130 : SortGroupClause *sortcl = lfirst_node(SortGroupClause, lc2);
6209 : :
6055 6210 : 1130 : sgrefs = bms_add_member(sgrefs, sortcl->tleSortGroupRef);
6211 : : }
6212 : : }
6213 : :
6214 : : /* Add in sortgroupref numbers of GROUP BY clauses, too */
1013 6215 [ + + + + : 1378 : foreach(lc, root->processed_groupClause)
+ + ]
6216 : : {
2974 6217 : 96 : SortGroupClause *grpcl = lfirst_node(SortGroupClause, lc);
6218 : :
4792 6219 : 96 : sgrefs = bms_add_member(sgrefs, grpcl->tleSortGroupRef);
6220 : : }
6221 : :
6222 : : /*
6223 : : * Construct a target containing all the non-flattenable targetlist items,
6224 : : * and save aside the others for a moment.
6225 : : */
3517 6226 : 1282 : input_target = create_empty_pathtarget();
4792 6227 : 1282 : flattenable_cols = NIL;
6228 : :
3517 6229 : 1282 : i = 0;
6230 [ + - + + : 5446 : foreach(lc, final_target->exprs)
+ + ]
6231 : : {
6232 : 4164 : Expr *expr = (Expr *) lfirst(lc);
3423 6233 [ + - ]: 4164 : Index sgref = get_pathtarget_sortgroupref(final_target, i);
6234 : :
6235 : : /*
6236 : : * Don't want to deconstruct window clauses or GROUP BY items. (Note
6237 : : * that such items can't contain window functions, so it's okay to
6238 : : * compute them below the WindowAgg nodes.)
6239 : : */
3517 6240 [ + + + + ]: 4164 : if (sgref != 0 && bms_is_member(sgref, sgrefs))
6241 : : {
6242 : : /*
6243 : : * Don't want to deconstruct this value, so add it to the input
6244 : : * target as-is.
6245 : : */
6246 : 1426 : add_column_to_pathtarget(input_target, expr, sgref);
6247 : : }
6248 : : else
6249 : : {
6250 : : /*
6251 : : * Column is to be flattened, so just remember the expression for
6252 : : * later call to pull_var_clause.
6253 : : */
6254 : 2738 : flattenable_cols = lappend(flattenable_cols, expr);
6255 : : }
6256 : :
6257 : 4164 : i++;
6258 : : }
6259 : :
6260 : : /*
6261 : : * Pull out all the Vars and Aggrefs mentioned in flattenable columns, and
6262 : : * add them to the input target if not already present. (Some might be
6263 : : * there already because they're used directly as window/group clauses.)
6264 : : *
6265 : : * Note: it's essential to use PVC_INCLUDE_AGGREGATES here, so that any
6266 : : * Aggrefs are placed in the Agg node's tlist and not left to be computed
6267 : : * at higher levels. On the other hand, we should recurse into
6268 : : * WindowFuncs to make sure their input expressions are available.
6269 : : */
4792 6270 : 1282 : flattenable_vars = pull_var_clause((Node *) flattenable_cols,
6271 : : PVC_INCLUDE_AGGREGATES |
6272 : : PVC_RECURSE_WINDOWFUNCS |
6273 : : PVC_INCLUDE_PLACEHOLDERS);
3517 6274 : 1282 : add_new_columns_to_pathtarget(input_target, flattenable_vars);
6275 : :
6276 : : /* clean up cruft */
4792 6277 : 1282 : list_free(flattenable_vars);
6278 : 1282 : list_free(flattenable_cols);
6279 : :
6280 : : /* XXX this causes some redundant cost calculation ... */
3517 6281 : 1282 : return set_pathtarget_cost_width(root, input_target);
6282 : : }
6283 : :
6284 : : /*
6285 : : * make_pathkeys_for_window
6286 : : * Create a pathkeys list describing the required input ordering
6287 : : * for the given WindowClause.
6288 : : *
6289 : : * Modifies wc's partitionClause to remove any clauses which are deemed
6290 : : * redundant by the pathkey logic.
6291 : : *
6292 : : * The required ordering is first the PARTITION keys, then the ORDER keys.
6293 : : * In the future we might try to implement windowing using hashing, in which
6294 : : * case the ordering could be relaxed, but for now we always sort.
6295 : : */
6296 : : static List *
6147 6297 : 2764 : make_pathkeys_for_window(PlannerInfo *root, WindowClause *wc,
6298 : : List *tlist)
6299 : : {
847 drowley@postgresql.o 6300 : 2764 : List *window_pathkeys = NIL;
6301 : :
6302 : : /* Throw error if can't sort */
6147 tgl@sss.pgh.pa.us 6303 [ - + ]: 2764 : if (!grouping_is_sortable(wc->partitionClause))
6147 tgl@sss.pgh.pa.us 6304 [ # # ]:UBC 0 : ereport(ERROR,
6305 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6306 : : errmsg("could not implement window PARTITION BY"),
6307 : : errdetail("Window partitioning columns must be of sortable datatypes.")));
6147 tgl@sss.pgh.pa.us 6308 [ - + ]:CBC 2764 : if (!grouping_is_sortable(wc->orderClause))
6147 tgl@sss.pgh.pa.us 6309 [ # # ]:UBC 0 : ereport(ERROR,
6310 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
6311 : : errmsg("could not implement window ORDER BY"),
6312 : : errdetail("Window ordering columns must be of sortable datatypes.")));
6313 : :
6314 : : /*
6315 : : * First fetch the pathkeys for the PARTITION BY clause. We can safely
6316 : : * remove any clauses from the wc->partitionClause for redundant pathkeys.
6317 : : */
847 drowley@postgresql.o 6318 [ + + ]:CBC 2764 : if (wc->partitionClause != NIL)
6319 : : {
6320 : : bool sortable;
6321 : :
6322 : 660 : window_pathkeys = make_pathkeys_for_sortclauses_extended(root,
6323 : : &wc->partitionClause,
6324 : : tlist,
6325 : : true,
6326 : : false,
6327 : : &sortable,
6328 : : false);
6329 : :
6330 [ - + ]: 660 : Assert(sortable);
6331 : : }
6332 : :
6333 : : /*
6334 : : * In principle, we could also consider removing redundant ORDER BY items
6335 : : * too as doing so does not alter the result of peer row checks done by
6336 : : * the executor. However, we must *not* remove the ordering column for
6337 : : * RANGE OFFSET cases, as the executor needs that for in_range tests even
6338 : : * if it's known to be equal to some partitioning column.
6339 : : */
6340 [ + + ]: 2764 : if (wc->orderClause != NIL)
6341 : : {
6342 : : List *orderby_pathkeys;
6343 : :
6344 : 2213 : orderby_pathkeys = make_pathkeys_for_sortclauses(root,
6345 : : wc->orderClause,
6346 : : tlist);
6347 : :
6348 : : /* Okay, make the combined pathkeys */
6349 [ + + ]: 2213 : if (window_pathkeys != NIL)
6350 : 473 : window_pathkeys = append_pathkeys(window_pathkeys, orderby_pathkeys);
6351 : : else
6352 : 1740 : window_pathkeys = orderby_pathkeys;
6353 : : }
6354 : :
6147 tgl@sss.pgh.pa.us 6355 : 2764 : return window_pathkeys;
6356 : : }
6357 : :
6358 : : /*
6359 : : * make_sort_input_target
6360 : : * Generate appropriate PathTarget for initial input to Sort step.
6361 : : *
6362 : : * If the query has ORDER BY, this function chooses the target to be computed
6363 : : * by the node just below the Sort (and DISTINCT, if any, since Unique can't
6364 : : * project) steps. This might or might not be identical to the query's final
6365 : : * output target.
6366 : : *
6367 : : * The main argument for keeping the sort-input tlist the same as the final
6368 : : * is that we avoid a separate projection node (which will be needed if
6369 : : * they're different, because Sort can't project). However, there are also
6370 : : * advantages to postponing tlist evaluation till after the Sort: it ensures
6371 : : * a consistent order of evaluation for any volatile functions in the tlist,
6372 : : * and if there's also a LIMIT, we can stop the query without ever computing
6373 : : * tlist functions for later rows, which is beneficial for both volatile and
6374 : : * expensive functions.
6375 : : *
6376 : : * Our current policy is to postpone volatile expressions till after the sort
6377 : : * unconditionally (assuming that that's possible, ie they are in plain tlist
6378 : : * columns and not ORDER BY/GROUP BY/DISTINCT columns). We also prefer to
6379 : : * postpone set-returning expressions, because running them beforehand would
6380 : : * bloat the sort dataset, and because it might cause unexpected output order
6381 : : * if the sort isn't stable. However there's a constraint on that: all SRFs
6382 : : * in the tlist should be evaluated at the same plan step, so that they can
6383 : : * run in sync in nodeProjectSet. So if any SRFs are in sort columns, we
6384 : : * mustn't postpone any SRFs. (Note that in principle that policy should
6385 : : * probably get applied to the group/window input targetlists too, but we
6386 : : * have not done that historically.) Lastly, expensive expressions are
6387 : : * postponed if there is a LIMIT, or if root->tuple_fraction shows that
6388 : : * partial evaluation of the query is possible (if neither is true, we expect
6389 : : * to have to evaluate the expressions for every row anyway), or if there are
6390 : : * any volatile or set-returning expressions (since once we've put in a
6391 : : * projection at all, it won't cost any more to postpone more stuff).
6392 : : *
6393 : : * Another issue that could potentially be considered here is that
6394 : : * evaluating tlist expressions could result in data that's either wider
6395 : : * or narrower than the input Vars, thus changing the volume of data that
6396 : : * has to go through the Sort. However, we usually have only a very bad
6397 : : * idea of the output width of any expression more complex than a Var,
6398 : : * so for now it seems too risky to try to optimize on that basis.
6399 : : *
6400 : : * Note that if we do produce a modified sort-input target, and then the
6401 : : * query ends up not using an explicit Sort, no particular harm is done:
6402 : : * we'll initially use the modified target for the preceding path nodes,
6403 : : * but then change them to the final target with apply_projection_to_path.
6404 : : * Moreover, in such a case the guarantees about evaluation order of
6405 : : * volatile functions still hold, since the rows are sorted already.
6406 : : *
6407 : : * This function has some things in common with make_group_input_target and
6408 : : * make_window_input_target, though the detailed rules for what to do are
6409 : : * different. We never flatten/postpone any grouping or ordering columns;
6410 : : * those are needed before the sort. If we do flatten a particular
6411 : : * expression, we leave Aggref and WindowFunc nodes alone, since those were
6412 : : * computed earlier.
6413 : : *
6414 : : * 'final_target' is the query's final target list (in PathTarget form)
6415 : : * 'have_postponed_srfs' is an output argument, see below
6416 : : *
6417 : : * The result is the PathTarget to be computed by the plan node immediately
6418 : : * below the Sort step (and the Distinct step, if any). This will be
6419 : : * exactly final_target if we decide a projection step wouldn't be helpful.
6420 : : *
6421 : : * In addition, *have_postponed_srfs is set to true if we choose to postpone
6422 : : * any set-returning functions to after the Sort.
6423 : : */
6424 : : static PathTarget *
3517 6425 : 35149 : make_sort_input_target(PlannerInfo *root,
6426 : : PathTarget *final_target,
6427 : : bool *have_postponed_srfs)
6428 : : {
6429 : 35149 : Query *parse = root->parse;
6430 : : PathTarget *input_target;
6431 : : int ncols;
6432 : : bool *col_is_srf;
6433 : : bool *postpone_col;
6434 : : bool have_srf;
6435 : : bool have_volatile;
6436 : : bool have_expensive;
6437 : : bool have_srf_sortcols;
6438 : : bool postpone_srfs;
6439 : : List *postponable_cols;
6440 : : List *postponable_vars;
6441 : : int i;
6442 : : ListCell *lc;
6443 : :
6444 : : /* Shouldn't get here unless query has ORDER BY */
6445 [ - + ]: 35149 : Assert(parse->sortClause);
6446 : :
3050 6447 : 35149 : *have_postponed_srfs = false; /* default result */
6448 : :
6449 : : /* Inspect tlist and collect per-column information */
3517 6450 : 35149 : ncols = list_length(final_target->exprs);
3503 6451 : 35149 : col_is_srf = (bool *) palloc0(ncols * sizeof(bool));
3517 6452 : 35149 : postpone_col = (bool *) palloc0(ncols * sizeof(bool));
3503 6453 : 35149 : have_srf = have_volatile = have_expensive = have_srf_sortcols = false;
6454 : :
3517 6455 : 35149 : i = 0;
6456 [ + - + + : 212496 : foreach(lc, final_target->exprs)
+ + ]
6457 : : {
6458 : 177347 : Expr *expr = (Expr *) lfirst(lc);
6459 : :
6460 : : /*
6461 : : * If the column has a sortgroupref, assume it has to be evaluated
6462 : : * before sorting. Generally such columns would be ORDER BY, GROUP
6463 : : * BY, etc targets. One exception is columns that were removed from
6464 : : * GROUP BY by remove_useless_groupby_columns() ... but those would
6465 : : * only be Vars anyway. There don't seem to be any cases where it
6466 : : * would be worth the trouble to double-check.
6467 : : */
3423 6468 [ + - + + ]: 177347 : if (get_pathtarget_sortgroupref(final_target, i) == 0)
6469 : : {
6470 : : /*
6471 : : * Check for SRF or volatile functions. Check the SRF case first
6472 : : * because we must know whether we have any postponed SRFs.
6473 : : */
3331 6474 [ + + + + ]: 127898 : if (parse->hasTargetSRFs &&
6475 : 108 : expression_returns_set((Node *) expr))
6476 : : {
6477 : : /* We'll decide below whether these are postponable */
3503 6478 : 48 : col_is_srf[i] = true;
3517 6479 : 48 : have_srf = true;
6480 : : }
6481 [ + + ]: 127742 : else if (contain_volatile_functions((Node *) expr))
6482 : : {
6483 : : /* Unconditionally postpone */
6484 : 74 : postpone_col[i] = true;
6485 : 74 : have_volatile = true;
6486 : : }
6487 : : else
6488 : : {
6489 : : /*
6490 : : * Else check the cost. XXX it's annoying to have to do this
6491 : : * when set_pathtarget_cost_width() just did it. Refactor to
6492 : : * allow sharing the work?
6493 : : */
6494 : : QualCost cost;
6495 : :
6496 : 127668 : cost_qual_eval_node(&cost, (Node *) expr, root);
6497 : :
6498 : : /*
6499 : : * We arbitrarily define "expensive" as "more than 10X
6500 : : * cpu_operator_cost". Note this will take in any PL function
6501 : : * with default cost.
6502 : : */
6503 [ + + ]: 127668 : if (cost.per_tuple > 10 * cpu_operator_cost)
6504 : : {
6505 : 8281 : postpone_col[i] = true;
6506 : 8281 : have_expensive = true;
6507 : : }
6508 : : }
6509 : : }
6510 : : else
6511 : : {
6512 : : /* For sortgroupref cols, just check if any contain SRFs */
3503 6513 [ + - ]: 49557 : if (!have_srf_sortcols &&
3331 6514 [ + + + + ]: 49712 : parse->hasTargetSRFs &&
3503 6515 : 155 : expression_returns_set((Node *) expr))
6516 : 62 : have_srf_sortcols = true;
6517 : : }
6518 : :
3517 6519 : 177347 : i++;
6520 : : }
6521 : :
6522 : : /*
6523 : : * We can postpone SRFs if we have some but none are in sortgroupref cols.
6524 : : */
3503 6525 [ + + + + ]: 35149 : postpone_srfs = (have_srf && !have_srf_sortcols);
6526 : :
6527 : : /*
6528 : : * If we don't need a post-sort projection, just return final_target.
6529 : : */
6530 [ + + + + ]: 35149 : if (!(postpone_srfs || have_volatile ||
3517 6531 [ + + ]: 35047 : (have_expensive &&
6532 [ + + + - ]: 4867 : (parse->limitCount || root->tuple_fraction > 0))))
6533 : 35029 : return final_target;
6534 : :
6535 : : /*
6536 : : * Report whether the post-sort projection will contain set-returning
6537 : : * functions. This is important because it affects whether the Sort can
6538 : : * rely on the query's LIMIT (if any) to bound the number of rows it needs
6539 : : * to return.
6540 : : */
3503 6541 : 120 : *have_postponed_srfs = postpone_srfs;
6542 : :
6543 : : /*
6544 : : * Construct the sort-input target, taking all non-postponable columns and
6545 : : * then adding Vars, PlaceHolderVars, Aggrefs, and WindowFuncs found in
6546 : : * the postponable ones.
6547 : : */
3517 6548 : 120 : input_target = create_empty_pathtarget();
6549 : 120 : postponable_cols = NIL;
6550 : :
6551 : 120 : i = 0;
6552 [ + - + + : 995 : foreach(lc, final_target->exprs)
+ + ]
6553 : : {
6554 : 875 : Expr *expr = (Expr *) lfirst(lc);
6555 : :
3503 6556 [ + + + + : 875 : if (postpone_col[i] || (postpone_srfs && col_is_srf[i]))
+ + ]
3517 6557 : 149 : postponable_cols = lappend(postponable_cols, expr);
6558 : : else
6559 : 726 : add_column_to_pathtarget(input_target, expr,
3050 6560 [ + - ]: 726 : get_pathtarget_sortgroupref(final_target, i));
6561 : :
3517 6562 : 875 : i++;
6563 : : }
6564 : :
6565 : : /*
6566 : : * Pull out all the Vars, Aggrefs, and WindowFuncs mentioned in
6567 : : * postponable columns, and add them to the sort-input target if not
6568 : : * already present. (Some might be there already.) We mustn't
6569 : : * deconstruct Aggrefs or WindowFuncs here, since the projection node
6570 : : * would be unable to recompute them.
6571 : : */
6572 : 120 : postponable_vars = pull_var_clause((Node *) postponable_cols,
6573 : : PVC_INCLUDE_AGGREGATES |
6574 : : PVC_INCLUDE_WINDOWFUNCS |
6575 : : PVC_INCLUDE_PLACEHOLDERS);
6576 : 120 : add_new_columns_to_pathtarget(input_target, postponable_vars);
6577 : :
6578 : : /* clean up cruft */
6579 : 120 : list_free(postponable_vars);
6580 : 120 : list_free(postponable_cols);
6581 : :
6582 : : /* XXX this represents even more redundant cost calculation ... */
6583 : 120 : return set_pathtarget_cost_width(root, input_target);
6584 : : }
6585 : :
6586 : : /*
6587 : : * get_cheapest_fractional_path
6588 : : * Find the cheapest path for retrieving a specified fraction of all
6589 : : * the tuples expected to be returned by the given relation.
6590 : : *
6591 : : * Do not consider parameterized paths. If the caller needs a path for upper
6592 : : * rel, it can't have parameterized paths. If the caller needs an append
6593 : : * subpath, it could become limited by the treatment of similar
6594 : : * parameterization of all the subpaths.
6595 : : *
6596 : : * We interpret tuple_fraction the same way as grouping_planner.
6597 : : *
6598 : : * We assume set_cheapest() has been run on the given rel.
6599 : : */
6600 : : Path *
3521 6601 : 241421 : get_cheapest_fractional_path(RelOptInfo *rel, double tuple_fraction)
6602 : : {
6603 : 241421 : Path *best_path = rel->cheapest_total_path;
6604 : : ListCell *l;
6605 : :
6606 : : /* If all tuples will be retrieved, just return the cheapest-total path */
6607 [ + + ]: 241421 : if (tuple_fraction <= 0.0)
6608 : 236783 : return best_path;
6609 : :
6610 : : /* Convert absolute # of tuples to a fraction; no need to clamp to 0..1 */
3502 6611 [ + + + + ]: 4638 : if (tuple_fraction >= 1.0 && best_path->rows > 0)
3521 6612 : 1883 : tuple_fraction /= best_path->rows;
6613 : :
6614 [ + - + + : 12072 : foreach(l, rel->pathlist)
+ + ]
6615 : : {
6616 : 7434 : Path *path = (Path *) lfirst(l);
6617 : :
231 akorotkov@postgresql 6618 [ + + ]: 7434 : if (path->param_info)
6619 : 100 : continue;
6620 : :
3521 tgl@sss.pgh.pa.us 6621 [ + + + + ]: 10030 : if (path == rel->cheapest_total_path ||
3050 6622 : 2696 : compare_fractional_path_costs(best_path, path, tuple_fraction) <= 0)
3521 6623 : 7077 : continue;
6624 : :
6625 : 257 : best_path = path;
6626 : : }
6627 : :
6628 : 4638 : return best_path;
6629 : : }
6630 : :
6631 : : /*
6632 : : * adjust_paths_for_srfs
6633 : : * Fix up the Paths of the given upperrel to handle tSRFs properly.
6634 : : *
6635 : : * The executor can only handle set-returning functions that appear at the
6636 : : * top level of the targetlist of a ProjectSet plan node. If we have any SRFs
6637 : : * that are not at top level, we need to split up the evaluation into multiple
6638 : : * plan levels in which each level satisfies this constraint. This function
6639 : : * modifies each Path of an upperrel that (might) compute any SRFs in its
6640 : : * output tlist to insert appropriate projection steps.
6641 : : *
6642 : : * The given targets and targets_contain_srfs lists are from
6643 : : * split_pathtarget_at_srfs(). We assume the existing Paths emit the first
6644 : : * target in targets.
6645 : : */
6646 : : static void
3204 andres@anarazel.de 6647 : 6187 : adjust_paths_for_srfs(PlannerInfo *root, RelOptInfo *rel,
6648 : : List *targets, List *targets_contain_srfs)
6649 : : {
6650 : : ListCell *lc;
6651 : :
6652 [ - + ]: 6187 : Assert(list_length(targets) == list_length(targets_contain_srfs));
6653 [ - + ]: 6187 : Assert(!linitial_int(targets_contain_srfs));
6654 : :
6655 : : /* If no SRFs appear at this plan level, nothing to do */
6656 [ + + ]: 6187 : if (list_length(targets) == 1)
6657 : 319 : return;
6658 : :
6659 : : /*
6660 : : * Stack SRF-evaluation nodes atop each path for the rel.
6661 : : *
6662 : : * In principle we should re-run set_cheapest() here to identify the
6663 : : * cheapest path, but it seems unlikely that adding the same tlist eval
6664 : : * costs to all the paths would change that, so we don't bother. Instead,
6665 : : * just assume that the cheapest-startup and cheapest-total paths remain
6666 : : * so. (There should be no parameterized paths anymore, so we needn't
6667 : : * worry about updating cheapest_parameterized_paths.)
6668 : : */
6669 [ + - + + : 11749 : foreach(lc, rel->pathlist)
+ + ]
6670 : : {
6671 : 5881 : Path *subpath = (Path *) lfirst(lc);
6672 : 5881 : Path *newpath = subpath;
6673 : : ListCell *lc1,
6674 : : *lc2;
6675 : :
6676 [ - + ]: 5881 : Assert(subpath->param_info == NULL);
6677 [ + - + + : 18237 : forboth(lc1, targets, lc2, targets_contain_srfs)
+ - + + +
+ + - +
+ ]
6678 : : {
2974 tgl@sss.pgh.pa.us 6679 : 12356 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
3204 andres@anarazel.de 6680 : 12356 : bool contains_srfs = (bool) lfirst_int(lc2);
6681 : :
6682 : : /* If this level doesn't contain SRFs, do regular projection */
6683 [ + + ]: 12356 : if (contains_srfs)
6684 : 5911 : newpath = (Path *) create_set_projection_path(root,
6685 : : rel,
6686 : : newpath,
6687 : : thistarget);
6688 : : else
6689 : 6445 : newpath = (Path *) apply_projection_to_path(root,
6690 : : rel,
6691 : : newpath,
6692 : : thistarget);
6693 : : }
6694 : 5881 : lfirst(lc) = newpath;
6695 [ + + ]: 5881 : if (subpath == rel->cheapest_startup_path)
6696 : 191 : rel->cheapest_startup_path = newpath;
6697 [ + + ]: 5881 : if (subpath == rel->cheapest_total_path)
6698 : 191 : rel->cheapest_total_path = newpath;
6699 : : }
6700 : :
6701 : : /* Likewise for partial paths, if any */
2831 rhaas@postgresql.org 6702 [ + + + + : 5871 : foreach(lc, rel->partial_pathlist)
+ + ]
6703 : : {
6704 : 3 : Path *subpath = (Path *) lfirst(lc);
6705 : 3 : Path *newpath = subpath;
6706 : : ListCell *lc1,
6707 : : *lc2;
6708 : :
6709 [ - + ]: 3 : Assert(subpath->param_info == NULL);
6710 [ + - + + : 12 : forboth(lc1, targets, lc2, targets_contain_srfs)
+ - + + +
+ + - +
+ ]
6711 : : {
6712 : 9 : PathTarget *thistarget = lfirst_node(PathTarget, lc1);
6713 : 9 : bool contains_srfs = (bool) lfirst_int(lc2);
6714 : :
6715 : : /* If this level doesn't contain SRFs, do regular projection */
6716 [ + + ]: 9 : if (contains_srfs)
6717 : 3 : newpath = (Path *) create_set_projection_path(root,
6718 : : rel,
6719 : : newpath,
6720 : : thistarget);
6721 : : else
6722 : : {
6723 : : /* avoid apply_projection_to_path, in case of multiple refs */
6724 : 6 : newpath = (Path *) create_projection_path(root,
6725 : : rel,
6726 : : newpath,
6727 : : thistarget);
6728 : : }
6729 : : }
6730 : 3 : lfirst(lc) = newpath;
6731 : : }
6732 : : }
6733 : :
6734 : : /*
6735 : : * expression_planner
6736 : : * Perform planner's transformations on a standalone expression.
6737 : : *
6738 : : * Various utility commands need to evaluate expressions that are not part
6739 : : * of a plannable query. They can do so using the executor's regular
6740 : : * expression-execution machinery, but first the expression has to be fed
6741 : : * through here to transform it from parser output to something executable.
6742 : : *
6743 : : * Currently, we disallow sublinks in standalone expressions, so there's no
6744 : : * real "planning" involved here. (That might not always be true though.)
6745 : : * What we must do is run eval_const_expressions to ensure that any function
6746 : : * calls are converted to positional notation and function default arguments
6747 : : * get inserted. The fact that constant subexpressions get simplified is a
6748 : : * side-effect that is useful when the expression will get evaluated more than
6749 : : * once. Also, we must fix operator function IDs.
6750 : : *
6751 : : * This does not return any information about dependencies of the expression.
6752 : : * Hence callers should use the results only for the duration of the current
6753 : : * query. Callers that would like to cache the results for longer should use
6754 : : * expression_planner_with_deps, probably via the plancache.
6755 : : *
6756 : : * Note: this must not make any damaging changes to the passed-in expression
6757 : : * tree. (It would actually be okay to apply fix_opfuncids to it, but since
6758 : : * we first do an expression_tree_mutator-based walk, what is returned will
6759 : : * be a new node tree.) The result is constructed in the current memory
6760 : : * context; beware that this can leak a lot of additional stuff there, too.
6761 : : */
6762 : : Expr *
6763 : 119452 : expression_planner(Expr *expr)
6764 : : {
6765 : : Node *result;
6766 : :
6767 : : /*
6768 : : * Convert named-argument function calls, insert default arguments and
6769 : : * simplify constant subexprs
6770 : : */
6771 : 119452 : result = eval_const_expressions(NULL, (Node *) expr);
6772 : :
6773 : : /* Fill in opfuncid values if missing */
6774 : 119443 : fix_opfuncids(result);
6775 : :
6776 : 119443 : return (Expr *) result;
6777 : : }
6778 : :
6779 : : /*
6780 : : * expression_planner_with_deps
6781 : : * Perform planner's transformations on a standalone expression,
6782 : : * returning expression dependency information along with the result.
6783 : : *
6784 : : * This is identical to expression_planner() except that it also returns
6785 : : * information about possible dependencies of the expression, ie identities of
6786 : : * objects whose definitions affect the result. As in a PlannedStmt, these
6787 : : * are expressed as a list of relation Oids and a list of PlanInvalItems.
6788 : : */
6789 : : Expr *
2510 tgl@sss.pgh.pa.us 6790 : 183 : expression_planner_with_deps(Expr *expr,
6791 : : List **relationOids,
6792 : : List **invalItems)
6793 : : {
6794 : : Node *result;
6795 : : PlannerGlobal glob;
6796 : : PlannerInfo root;
6797 : :
6798 : : /* Make up dummy planner state so we can use setrefs machinery */
6799 [ + - + - : 4758 : MemSet(&glob, 0, sizeof(glob));
+ - + - +
+ ]
6800 : 183 : glob.type = T_PlannerGlobal;
6801 : 183 : glob.relationOids = NIL;
6802 : 183 : glob.invalItems = NIL;
6803 : :
6804 [ + - + - : 17019 : MemSet(&root, 0, sizeof(root));
+ - + - +
+ ]
6805 : 183 : root.type = T_PlannerInfo;
6806 : 183 : root.glob = &glob;
6807 : :
6808 : : /*
6809 : : * Convert named-argument function calls, insert default arguments and
6810 : : * simplify constant subexprs. Collect identities of inlined functions
6811 : : * and elided domains, too.
6812 : : */
6813 : 183 : result = eval_const_expressions(&root, (Node *) expr);
6814 : :
6815 : : /* Fill in opfuncid values if missing */
6816 : 183 : fix_opfuncids(result);
6817 : :
6818 : : /*
6819 : : * Now walk the finished expression to find anything else we ought to
6820 : : * record as an expression dependency.
6821 : : */
6822 : 183 : (void) extract_query_dependencies_walker(result, &root);
6823 : :
6824 : 183 : *relationOids = glob.relationOids;
6825 : 183 : *invalItems = glob.invalItems;
6826 : :
6827 : 183 : return (Expr *) result;
6828 : : }
6829 : :
6830 : :
6831 : : /*
6832 : : * plan_cluster_use_sort
6833 : : * Use the planner to decide how CLUSTER should implement sorting
6834 : : *
6835 : : * tableOid is the OID of a table to be clustered on its index indexOid
6836 : : * (which is already known to be a btree index). Decide whether it's
6837 : : * cheaper to do an indexscan or a seqscan-plus-sort to execute the CLUSTER.
6838 : : * Return true to use sorting, false to use an indexscan.
6839 : : *
6840 : : * Note: caller had better already hold some type of lock on the table.
6841 : : */
6842 : : bool
2831 rhaas@postgresql.org 6843 : 96 : plan_cluster_use_sort(Oid tableOid, Oid indexOid)
6844 : : {
6845 : : PlannerInfo *root;
6846 : : Query *query;
6847 : : PlannerGlobal *glob;
6848 : : RangeTblEntry *rte;
6849 : : RelOptInfo *rel;
6850 : : IndexOptInfo *indexInfo;
6851 : : QualCost indexExprCost;
6852 : : Cost comparisonCost;
6853 : : Path *seqScanPath;
6854 : : Path seqScanAndSortPath;
6855 : : IndexPath *indexScanPath;
6856 : : ListCell *lc;
6857 : :
6858 : : /* We can short-circuit the cost comparison if indexscans are disabled */
6859 [ + + ]: 96 : if (!enable_indexscan)
6860 : 15 : return true; /* use sort */
6861 : :
6862 : : /* Set up mostly-dummy planner state */
6863 : 81 : query = makeNode(Query);
6864 : 81 : query->commandType = CMD_SELECT;
6865 : :
6866 : 81 : glob = makeNode(PlannerGlobal);
6867 : :
6868 : 81 : root = makeNode(PlannerInfo);
6869 : 81 : root->parse = query;
6870 : 81 : root->glob = glob;
6871 : 81 : root->query_level = 1;
6872 : 81 : root->planner_cxt = CurrentMemoryContext;
6873 : 81 : root->wt_param_id = -1;
1001 tgl@sss.pgh.pa.us 6874 : 81 : root->join_domains = list_make1(makeNode(JoinDomain));
6875 : :
6876 : : /* Build a minimal RTE for the rel */
2831 rhaas@postgresql.org 6877 : 81 : rte = makeNode(RangeTblEntry);
6878 : 81 : rte->rtekind = RTE_RELATION;
6879 : 81 : rte->relid = tableOid;
6880 : 81 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
2584 tgl@sss.pgh.pa.us 6881 : 81 : rte->rellockmode = AccessShareLock;
2831 rhaas@postgresql.org 6882 : 81 : rte->lateral = false;
6883 : 81 : rte->inh = false;
6884 : 81 : rte->inFromCl = true;
6885 : 81 : query->rtable = list_make1(rte);
1056 alvherre@alvh.no-ip. 6886 : 81 : addRTEPermissionInfo(&query->rteperminfos, rte);
6887 : :
6888 : : /* Set up RTE/RelOptInfo arrays */
2831 rhaas@postgresql.org 6889 : 81 : setup_simple_rel_arrays(root);
6890 : :
6891 : : /* Build RelOptInfo */
6892 : 81 : rel = build_simple_rel(root, 1, NULL);
6893 : :
6894 : : /* Locate IndexOptInfo for the target index */
6895 : 81 : indexInfo = NULL;
6896 [ + - + - : 100 : foreach(lc, rel->indexlist)
+ - ]
6897 : : {
6898 : 100 : indexInfo = lfirst_node(IndexOptInfo, lc);
6899 [ + + ]: 100 : if (indexInfo->indexoid == indexOid)
6900 : 81 : break;
6901 : : }
6902 : :
6903 : : /*
6904 : : * It's possible that get_relation_info did not generate an IndexOptInfo
6905 : : * for the desired index; this could happen if it's not yet reached its
6906 : : * indcheckxmin usability horizon, or if it's a system index and we're
6907 : : * ignoring system indexes. In such cases we should tell CLUSTER to not
6908 : : * trust the index contents but use seqscan-and-sort.
6909 : : */
6910 [ - + ]: 81 : if (lc == NULL) /* not in the list? */
2831 rhaas@postgresql.org 6911 :UBC 0 : return true; /* use sort */
6912 : :
6913 : : /*
6914 : : * Rather than doing all the pushups that would be needed to use
6915 : : * set_baserel_size_estimates, just do a quick hack for rows and width.
6916 : : */
2831 rhaas@postgresql.org 6917 :CBC 81 : rel->rows = rel->tuples;
6918 : 81 : rel->reltarget->width = get_relation_data_width(tableOid, NULL);
6919 : :
6920 : 81 : root->total_table_pages = rel->pages;
6921 : :
6922 : : /*
6923 : : * Determine eval cost of the index expressions, if any. We need to
6924 : : * charge twice that amount for each tuple comparison that happens during
6925 : : * the sort, since tuplesort.c will have to re-evaluate the index
6926 : : * expressions each time. (XXX that's pretty inefficient...)
6927 : : */
6928 : 81 : cost_qual_eval(&indexExprCost, indexInfo->indexprs, root);
6929 : 81 : comparisonCost = 2.0 * (indexExprCost.startup + indexExprCost.per_tuple);
6930 : :
6931 : : /* Estimate the cost of seq scan + sort */
6932 : 81 : seqScanPath = create_seqscan_path(root, rel, NULL, 0);
6933 : 81 : cost_sort(&seqScanAndSortPath, root, NIL,
6934 : : seqScanPath->disabled_nodes,
6935 : 81 : seqScanPath->total_cost, rel->tuples, rel->reltarget->width,
6936 : : comparisonCost, maintenance_work_mem, -1.0);
6937 : :
6938 : : /* Estimate the cost of index scan */
6939 : 81 : indexScanPath = create_index_path(root, indexInfo,
6940 : : NIL, NIL, NIL, NIL,
6941 : : ForwardScanDirection, false,
6942 : : NULL, 1.0, false);
6943 : :
6944 : 81 : return (seqScanAndSortPath.total_cost < indexScanPath->path.total_cost);
6945 : : }
6946 : :
6947 : : /*
6948 : : * plan_create_index_workers
6949 : : * Use the planner to decide how many parallel worker processes
6950 : : * CREATE INDEX should request for use
6951 : : *
6952 : : * tableOid is the table on which the index is to be built. indexOid is the
6953 : : * OID of an index to be created or reindexed (which must be an index with
6954 : : * support for parallel builds - currently btree, GIN, or BRIN).
6955 : : *
6956 : : * Return value is the number of parallel worker processes to request. It
6957 : : * may be unsafe to proceed if this is 0. Note that this does not include the
6958 : : * leader participating as a worker (value is always a number of parallel
6959 : : * worker processes).
6960 : : *
6961 : : * Note: caller had better already hold some type of lock on the table and
6962 : : * index.
6963 : : */
6964 : : int
2824 6965 : 17538 : plan_create_index_workers(Oid tableOid, Oid indexOid)
6966 : : {
6967 : : PlannerInfo *root;
6968 : : Query *query;
6969 : : PlannerGlobal *glob;
6970 : : RangeTblEntry *rte;
6971 : : Relation heap;
6972 : : Relation index;
6973 : : RelOptInfo *rel;
6974 : : int parallel_workers;
6975 : : BlockNumber heap_blocks;
6976 : : double reltuples;
6977 : : double allvisfrac;
6978 : :
6979 : : /*
6980 : : * We don't allow performing parallel operation in standalone backend or
6981 : : * when parallelism is disabled.
6982 : : */
1792 tgl@sss.pgh.pa.us 6983 [ + + + + ]: 17538 : if (!IsUnderPostmaster || max_parallel_maintenance_workers == 0)
2824 rhaas@postgresql.org 6984 : 253 : return 0;
6985 : :
6986 : : /* Set up largely-dummy planner state */
6987 : 17285 : query = makeNode(Query);
6988 : 17285 : query->commandType = CMD_SELECT;
6989 : :
6990 : 17285 : glob = makeNode(PlannerGlobal);
6991 : :
6992 : 17285 : root = makeNode(PlannerInfo);
6993 : 17285 : root->parse = query;
6994 : 17285 : root->glob = glob;
6995 : 17285 : root->query_level = 1;
6996 : 17285 : root->planner_cxt = CurrentMemoryContext;
6997 : 17285 : root->wt_param_id = -1;
1001 tgl@sss.pgh.pa.us 6998 : 17285 : root->join_domains = list_make1(makeNode(JoinDomain));
6999 : :
7000 : : /*
7001 : : * Build a minimal RTE.
7002 : : *
7003 : : * Mark the RTE with inh = true. This is a kludge to prevent
7004 : : * get_relation_info() from fetching index info, which is necessary
7005 : : * because it does not expect that any IndexOptInfo is currently
7006 : : * undergoing REINDEX.
7007 : : */
2824 rhaas@postgresql.org 7008 : 17285 : rte = makeNode(RangeTblEntry);
7009 : 17285 : rte->rtekind = RTE_RELATION;
7010 : 17285 : rte->relid = tableOid;
7011 : 17285 : rte->relkind = RELKIND_RELATION; /* Don't be too picky. */
2584 tgl@sss.pgh.pa.us 7012 : 17285 : rte->rellockmode = AccessShareLock;
2824 rhaas@postgresql.org 7013 : 17285 : rte->lateral = false;
7014 : 17285 : rte->inh = true;
7015 : 17285 : rte->inFromCl = true;
7016 : 17285 : query->rtable = list_make1(rte);
1056 alvherre@alvh.no-ip. 7017 : 17285 : addRTEPermissionInfo(&query->rteperminfos, rte);
7018 : :
7019 : : /* Set up RTE/RelOptInfo arrays */
2824 rhaas@postgresql.org 7020 : 17285 : setup_simple_rel_arrays(root);
7021 : :
7022 : : /* Build RelOptInfo */
7023 : 17285 : rel = build_simple_rel(root, 1, NULL);
7024 : :
7025 : : /* Rels are assumed already locked by the caller */
2471 andres@anarazel.de 7026 : 17285 : heap = table_open(tableOid, NoLock);
2824 rhaas@postgresql.org 7027 : 17285 : index = index_open(indexOid, NoLock);
7028 : :
7029 : : /*
7030 : : * Determine if it's safe to proceed.
7031 : : *
7032 : : * Currently, parallel workers can't access the leader's temporary tables.
7033 : : * Furthermore, any index predicate or index expressions must be parallel
7034 : : * safe.
7035 : : */
7036 [ + + ]: 17285 : if (heap->rd_rel->relpersistence == RELPERSISTENCE_TEMP ||
599 michael@paquier.xyz 7037 [ + + ]: 16246 : !is_parallel_safe(root, (Node *) RelationGetIndexExpressions(index)) ||
7038 [ - + ]: 16186 : !is_parallel_safe(root, (Node *) RelationGetIndexPredicate(index)))
7039 : : {
2824 rhaas@postgresql.org 7040 : 1099 : parallel_workers = 0;
7041 : 1099 : goto done;
7042 : : }
7043 : :
7044 : : /*
7045 : : * If parallel_workers storage parameter is set for the table, accept that
7046 : : * as the number of parallel worker processes to launch (though still cap
7047 : : * at max_parallel_maintenance_workers). Note that we deliberately do not
7048 : : * consider any other factor when parallel_workers is set. (e.g., memory
7049 : : * use by workers.)
7050 : : */
7051 [ + + ]: 16186 : if (rel->rel_parallel_workers != -1)
7052 : : {
7053 : 7 : parallel_workers = Min(rel->rel_parallel_workers,
7054 : : max_parallel_maintenance_workers);
7055 : 7 : goto done;
7056 : : }
7057 : :
7058 : : /*
7059 : : * Estimate heap relation size ourselves, since rel->pages cannot be
7060 : : * trusted (heap RTE was marked as inheritance parent)
7061 : : */
7062 : 16179 : estimate_rel_size(heap, NULL, &heap_blocks, &reltuples, &allvisfrac);
7063 : :
7064 : : /*
7065 : : * Determine number of workers to scan the heap relation using generic
7066 : : * model
7067 : : */
7068 : 16179 : parallel_workers = compute_parallel_worker(rel, heap_blocks, -1,
7069 : : max_parallel_maintenance_workers);
7070 : :
7071 : : /*
7072 : : * Cap workers based on available maintenance_work_mem as needed.
7073 : : *
7074 : : * Note that each tuplesort participant receives an even share of the
7075 : : * total maintenance_work_mem budget. Aim to leave participants
7076 : : * (including the leader as a participant) with no less than 32MB of
7077 : : * memory. This leaves cases where maintenance_work_mem is set to 64MB
7078 : : * immediately past the threshold of being capable of launching a single
7079 : : * parallel worker to sort.
7080 : : */
7081 [ + + ]: 16257 : while (parallel_workers > 0 &&
269 tgl@sss.pgh.pa.us 7082 [ + + ]: 157 : maintenance_work_mem / (parallel_workers + 1) < 32 * 1024)
2824 rhaas@postgresql.org 7083 : 78 : parallel_workers--;
7084 : :
7085 : 16179 : done:
7086 : 17285 : index_close(index, NoLock);
2471 andres@anarazel.de 7087 : 17285 : table_close(heap, NoLock);
7088 : :
2824 rhaas@postgresql.org 7089 : 17285 : return parallel_workers;
7090 : : }
7091 : :
7092 : : /*
7093 : : * add_paths_to_grouping_rel
7094 : : *
7095 : : * Add non-partial paths to grouping relation.
7096 : : */
7097 : : static void
2831 7098 : 19308 : add_paths_to_grouping_rel(PlannerInfo *root, RelOptInfo *input_rel,
7099 : : RelOptInfo *grouped_rel,
7100 : : RelOptInfo *partially_grouped_rel,
7101 : : const AggClauseCosts *agg_costs,
7102 : : grouping_sets_data *gd,
7103 : : GroupPathExtraData *extra)
7104 : : {
7105 : 19308 : Query *parse = root->parse;
7106 : 19308 : Path *cheapest_path = input_rel->cheapest_total_path;
19 rguo@postgresql.org 7107 :GNC 19308 : Path *cheapest_partially_grouped_path = NULL;
7108 : : ListCell *lc;
2776 rhaas@postgresql.org 7109 :CBC 19308 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7110 : 19308 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7111 : 19308 : List *havingQual = (List *) extra->havingQual;
7112 : 19308 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
19 rguo@postgresql.org 7113 :GNC 19308 : double dNumGroups = 0;
7114 : 19308 : double dNumFinalGroups = 0;
7115 : :
7116 : : /*
7117 : : * Estimate number of groups for non-split aggregation.
7118 : : */
7119 : 19308 : dNumGroups = get_number_of_groups(root,
7120 : : cheapest_path->rows,
7121 : : gd,
7122 : : extra->targetList);
7123 : :
7124 [ + + + - ]: 19308 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7125 : : {
7126 : 1121 : cheapest_partially_grouped_path =
7127 : : partially_grouped_rel->cheapest_total_path;
7128 : :
7129 : : /*
7130 : : * Estimate number of groups for final phase of partial aggregation.
7131 : : */
7132 : : dNumFinalGroups =
7133 : 1121 : get_number_of_groups(root,
7134 : : cheapest_partially_grouped_path->rows,
7135 : : gd,
7136 : : extra->targetList);
7137 : : }
7138 : :
2831 rhaas@postgresql.org 7139 [ + + ]:CBC 19308 : if (can_sort)
7140 : : {
7141 : : /*
7142 : : * Use any available suitably-sorted path as input, and also consider
7143 : : * sorting the cheapest-total path and incremental sort on any paths
7144 : : * with presorted keys.
7145 : : */
7146 [ + - + + : 40247 : foreach(lc, input_rel->pathlist)
+ + ]
7147 : : {
7148 : : ListCell *lc2;
7149 : 20942 : Path *path = (Path *) lfirst(lc);
645 akorotkov@postgresql 7150 : 20942 : Path *path_save = path;
7151 : 20942 : List *pathkey_orderings = NIL;
7152 : :
7153 : : /* generate alternative group orderings that might be useful */
7154 : 20942 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7155 : :
7156 [ - + ]: 20942 : Assert(list_length(pathkey_orderings) > 0);
7157 : :
7158 [ + - + + : 41956 : foreach(lc2, pathkey_orderings)
+ + ]
7159 : : {
508 7160 : 21014 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7161 : :
7162 : : /* restore the path (we replace it in the loop) */
645 7163 : 21014 : path = path_save;
7164 : :
7165 : 21014 : path = make_ordered_path(root,
7166 : : grouped_rel,
7167 : : path,
7168 : : cheapest_path,
7169 : : info->pathkeys,
7170 : : -1.0);
7171 [ + + ]: 21014 : if (path == NULL)
7172 : 202 : continue;
7173 : :
7174 : : /* Now decide what to stick atop it */
7175 [ + + ]: 20812 : if (parse->groupingSets)
7176 : : {
7177 : 496 : consider_groupingsets_paths(root, grouped_rel,
7178 : : path, true, can_hash,
7179 : : gd, agg_costs, dNumGroups);
7180 : : }
7181 [ + + ]: 20316 : else if (parse->hasAggs)
7182 : : {
7183 : : /*
7184 : : * We have aggregation, possibly with plain GROUP BY. Make
7185 : : * an AggPath.
7186 : : */
1120 tgl@sss.pgh.pa.us 7187 : 19924 : add_path(grouped_rel, (Path *)
7188 : 19924 : create_agg_path(root,
7189 : : grouped_rel,
7190 : : path,
7191 : 19924 : grouped_rel->reltarget,
7192 : 19924 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7193 : : AGGSPLIT_SIMPLE,
7194 : : info->clauses,
7195 : : havingQual,
7196 : : agg_costs,
7197 : : dNumGroups));
7198 : : }
645 akorotkov@postgresql 7199 [ + - ]: 392 : else if (parse->groupClause)
7200 : : {
7201 : : /*
7202 : : * We have GROUP BY without aggregation or grouping sets.
7203 : : * Make a GroupPath.
7204 : : */
1120 tgl@sss.pgh.pa.us 7205 : 392 : add_path(grouped_rel, (Path *)
7206 : 392 : create_group_path(root,
7207 : : grouped_rel,
7208 : : path,
7209 : : info->clauses,
7210 : : havingQual,
7211 : : dNumGroups));
7212 : : }
7213 : : else
7214 : : {
7215 : : /* Other cases should have been handled above */
645 akorotkov@postgresql 7216 :UBC 0 : Assert(false);
7217 : : }
7218 : : }
7219 : : }
7220 : :
7221 : : /*
7222 : : * Instead of operating directly on the input relation, we can
7223 : : * consider finalizing a partially aggregated path.
7224 : : */
645 akorotkov@postgresql 7225 [ + + ]:CBC 19305 : if (partially_grouped_rel != NULL)
7226 : : {
7227 [ + - + + : 3083 : foreach(lc, partially_grouped_rel->pathlist)
+ + ]
7228 : : {
7229 : : ListCell *lc2;
7230 : 1962 : Path *path = (Path *) lfirst(lc);
7231 : 1962 : Path *path_save = path;
7232 : 1962 : List *pathkey_orderings = NIL;
7233 : :
7234 : : /* generate alternative group orderings that might be useful */
7235 : 1962 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7236 : :
7237 [ - + ]: 1962 : Assert(list_length(pathkey_orderings) > 0);
7238 : :
7239 : : /* process all potentially interesting grouping reorderings */
7240 [ + - + + : 3924 : foreach(lc2, pathkey_orderings)
+ + ]
7241 : : {
508 7242 : 1962 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7243 : :
7244 : : /* restore the path (we replace it in the loop) */
645 7245 : 1962 : path = path_save;
7246 : :
7247 : 1962 : path = make_ordered_path(root,
7248 : : grouped_rel,
7249 : : path,
7250 : : cheapest_partially_grouped_path,
7251 : : info->pathkeys,
7252 : : -1.0);
7253 : :
7254 [ + + ]: 1962 : if (path == NULL)
7255 : 102 : continue;
7256 : :
7257 [ + + ]: 1860 : if (parse->hasAggs)
7258 : 1736 : add_path(grouped_rel, (Path *)
7259 : 1736 : create_agg_path(root,
7260 : : grouped_rel,
7261 : : path,
7262 : 1736 : grouped_rel->reltarget,
7263 : 1736 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7264 : : AGGSPLIT_FINAL_DESERIAL,
7265 : : info->clauses,
7266 : : havingQual,
7267 : : agg_final_costs,
7268 : : dNumFinalGroups));
7269 : : else
7270 : 124 : add_path(grouped_rel, (Path *)
7271 : 124 : create_group_path(root,
7272 : : grouped_rel,
7273 : : path,
7274 : : info->clauses,
7275 : : havingQual,
7276 : : dNumFinalGroups));
7277 : :
7278 : : }
7279 : : }
7280 : : }
7281 : : }
7282 : :
2831 rhaas@postgresql.org 7283 [ + + ]: 19308 : if (can_hash)
7284 : : {
7285 [ + + ]: 2846 : if (parse->groupingSets)
7286 : : {
7287 : : /*
7288 : : * Try for a hash-only groupingsets path over unsorted input.
7289 : : */
7290 : 415 : consider_groupingsets_paths(root, grouped_rel,
7291 : : cheapest_path, false, true,
7292 : : gd, agg_costs, dNumGroups);
7293 : : }
7294 : : else
7295 : : {
7296 : : /*
7297 : : * Generate a HashAgg Path. We just need an Agg over the
7298 : : * cheapest-total input path, since input order won't matter.
7299 : : */
1918 pg@bowt.ie 7300 : 2431 : add_path(grouped_rel, (Path *)
7301 : 2431 : create_agg_path(root, grouped_rel,
7302 : : cheapest_path,
7303 : 2431 : grouped_rel->reltarget,
7304 : : AGG_HASHED,
7305 : : AGGSPLIT_SIMPLE,
7306 : : root->processed_groupClause,
7307 : : havingQual,
7308 : : agg_costs,
7309 : : dNumGroups));
7310 : : }
7311 : :
7312 : : /*
7313 : : * Generate a Finalize HashAgg Path atop of the cheapest partially
7314 : : * grouped path, assuming there is one
7315 : : */
2778 rhaas@postgresql.org 7316 [ + + + - ]: 2846 : if (partially_grouped_rel && partially_grouped_rel->pathlist)
7317 : : {
1918 pg@bowt.ie 7318 : 721 : add_path(grouped_rel, (Path *)
7319 : 721 : create_agg_path(root,
7320 : : grouped_rel,
7321 : : cheapest_partially_grouped_path,
7322 : 721 : grouped_rel->reltarget,
7323 : : AGG_HASHED,
7324 : : AGGSPLIT_FINAL_DESERIAL,
7325 : : root->processed_groupClause,
7326 : : havingQual,
7327 : : agg_final_costs,
7328 : : dNumFinalGroups));
7329 : : }
7330 : : }
7331 : :
7332 : : /*
7333 : : * When partitionwise aggregate is used, we might have fully aggregated
7334 : : * paths in the partial pathlist, because add_paths_to_append_rel() will
7335 : : * consider a path for grouped_rel consisting of a Parallel Append of
7336 : : * non-partial paths from each child.
7337 : : */
2776 rhaas@postgresql.org 7338 [ + + ]: 19308 : if (grouped_rel->partial_pathlist != NIL)
7339 : 159 : gather_grouping_paths(root, grouped_rel);
5499 tgl@sss.pgh.pa.us 7340 : 19308 : }
7341 : :
7342 : : /*
7343 : : * create_partial_grouping_paths
7344 : : *
7345 : : * Create a new upper relation representing the result of partial aggregation
7346 : : * and populate it with appropriate paths. Note that we don't finalize the
7347 : : * lists of paths here, so the caller can add additional partial or non-partial
7348 : : * paths and must afterward call gather_grouping_paths and set_cheapest on
7349 : : * the returned upper relation.
7350 : : *
7351 : : * All paths for this new upper relation -- both partial and non-partial --
7352 : : * have been partially aggregated but require a subsequent FinalizeAggregate
7353 : : * step.
7354 : : *
7355 : : * NB: This function is allowed to return NULL if it determines that there is
7356 : : * no real need to create a new RelOptInfo.
7357 : : */
7358 : : static RelOptInfo *
2778 rhaas@postgresql.org 7359 : 17288 : create_partial_grouping_paths(PlannerInfo *root,
7360 : : RelOptInfo *grouped_rel,
7361 : : RelOptInfo *input_rel,
7362 : : grouping_sets_data *gd,
7363 : : GroupPathExtraData *extra,
7364 : : bool force_rel_creation)
7365 : : {
2831 7366 : 17288 : Query *parse = root->parse;
7367 : : RelOptInfo *partially_grouped_rel;
19 rguo@postgresql.org 7368 :GNC 17288 : RelOptInfo *eager_agg_rel = NULL;
2776 rhaas@postgresql.org 7369 :CBC 17288 : AggClauseCosts *agg_partial_costs = &extra->agg_partial_costs;
7370 : 17288 : AggClauseCosts *agg_final_costs = &extra->agg_final_costs;
7371 : 17288 : Path *cheapest_partial_path = NULL;
7372 : 17288 : Path *cheapest_total_path = NULL;
2831 7373 : 17288 : double dNumPartialGroups = 0;
2776 7374 : 17288 : double dNumPartialPartialGroups = 0;
7375 : : ListCell *lc;
7376 : 17288 : bool can_hash = (extra->flags & GROUPING_CAN_USE_HASH) != 0;
7377 : 17288 : bool can_sort = (extra->flags & GROUPING_CAN_USE_SORT) != 0;
7378 : :
7379 : : /*
7380 : : * Check whether any partially aggregated paths have been generated
7381 : : * through eager aggregation.
7382 : : */
19 rguo@postgresql.org 7383 [ + + ]:GNC 17288 : if (input_rel->grouped_rel &&
7384 [ + - ]: 479 : !IS_DUMMY_REL(input_rel->grouped_rel) &&
7385 [ + + ]: 479 : input_rel->grouped_rel->pathlist != NIL)
7386 : 449 : eager_agg_rel = input_rel->grouped_rel;
7387 : :
7388 : : /*
7389 : : * Consider whether we should generate partially aggregated non-partial
7390 : : * paths. We can only do this if we have a non-partial path, and only if
7391 : : * the parent of the input rel is performing partial partitionwise
7392 : : * aggregation. (Note that extra->patype is the type of partitionwise
7393 : : * aggregation being used at the parent level, not this level.)
7394 : : */
2776 rhaas@postgresql.org 7395 [ + - ]:CBC 17288 : if (input_rel->pathlist != NIL &&
7396 [ + + ]: 17288 : extra->patype == PARTITIONWISE_AGGREGATE_PARTIAL)
7397 : 429 : cheapest_total_path = input_rel->cheapest_total_path;
7398 : :
7399 : : /*
7400 : : * If parallelism is possible for grouped_rel, then we should consider
7401 : : * generating partially-grouped partial paths. However, if the input rel
7402 : : * has no partial paths, then we can't.
7403 : : */
7404 [ + + + + ]: 17288 : if (grouped_rel->consider_parallel && input_rel->partial_pathlist != NIL)
7405 : 1259 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
7406 : :
7407 : : /*
7408 : : * If we can't partially aggregate partial paths, and we can't partially
7409 : : * aggregate non-partial paths, and no partially aggregated paths were
7410 : : * generated by eager aggregation, then don't bother creating the new
7411 : : * RelOptInfo at all, unless the caller specified force_rel_creation.
7412 : : */
7413 [ + + + + ]: 17288 : if (cheapest_total_path == NULL &&
7414 [ + + ]: 15852 : cheapest_partial_path == NULL &&
19 rguo@postgresql.org 7415 :GNC 15787 : eager_agg_rel == NULL &&
2776 rhaas@postgresql.org 7416 [ + + ]:CBC 15787 : !force_rel_creation)
7417 : 15738 : return NULL;
7418 : :
7419 : : /*
7420 : : * Build a new upper relation to represent the result of partially
7421 : : * aggregating the rows from the input relation.
7422 : : */
2778 7423 : 1550 : partially_grouped_rel = fetch_upper_rel(root,
7424 : : UPPERREL_PARTIAL_GROUP_AGG,
7425 : : grouped_rel->relids);
7426 : 1550 : partially_grouped_rel->consider_parallel =
7427 : 1550 : grouped_rel->consider_parallel;
2776 7428 : 1550 : partially_grouped_rel->reloptkind = grouped_rel->reloptkind;
2778 7429 : 1550 : partially_grouped_rel->serverid = grouped_rel->serverid;
7430 : 1550 : partially_grouped_rel->userid = grouped_rel->userid;
7431 : 1550 : partially_grouped_rel->useridiscurrent = grouped_rel->useridiscurrent;
7432 : 1550 : partially_grouped_rel->fdwroutine = grouped_rel->fdwroutine;
7433 : :
7434 : : /*
7435 : : * Build target list for partial aggregate paths. These paths cannot just
7436 : : * emit the same tlist as regular aggregate paths, because (1) we must
7437 : : * include Vars and Aggrefs needed in HAVING, which might not appear in
7438 : : * the result tlist, and (2) the Aggrefs must be set in partial mode.
7439 : : */
7440 : 1550 : partially_grouped_rel->reltarget =
7441 : 1550 : make_partial_grouping_target(root, grouped_rel->reltarget,
7442 : : extra->havingQual);
7443 : :
2776 7444 [ + + ]: 1550 : if (!extra->partial_costs_set)
7445 : : {
7446 : : /*
7447 : : * Collect statistics about aggregates for estimating costs of
7448 : : * performing aggregation in parallel.
7449 : : */
7450 [ + - + - : 4674 : MemSet(agg_partial_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
7451 [ + - + - : 4674 : MemSet(agg_final_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
7452 [ + + ]: 779 : if (parse->hasAggs)
7453 : : {
7454 : : /* partial phase */
1798 heikki.linnakangas@i 7455 : 712 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL,
7456 : : agg_partial_costs);
7457 : :
7458 : : /* final phase */
7459 : 712 : get_agg_clause_costs(root, AGGSPLIT_FINAL_DESERIAL,
7460 : : agg_final_costs);
7461 : : }
7462 : :
2776 rhaas@postgresql.org 7463 : 779 : extra->partial_costs_set = true;
7464 : : }
7465 : :
7466 : : /* Estimate number of partial groups. */
7467 [ + + ]: 1550 : if (cheapest_total_path != NULL)
7468 : : dNumPartialGroups =
7469 : 429 : get_number_of_groups(root,
7470 : : cheapest_total_path->rows,
7471 : : gd,
7472 : : extra->targetList);
7473 [ + + ]: 1550 : if (cheapest_partial_path != NULL)
7474 : : dNumPartialPartialGroups =
7475 : 1259 : get_number_of_groups(root,
7476 : : cheapest_partial_path->rows,
7477 : : gd,
7478 : : extra->targetList);
7479 : :
7480 [ + - + + ]: 1550 : if (can_sort && cheapest_total_path != NULL)
7481 : : {
7482 : : /* This should have been checked previously */
2831 7483 [ + + - + ]: 429 : Assert(parse->hasAggs || parse->groupClause);
7484 : :
7485 : : /*
7486 : : * Use any available suitably-sorted path as input, and also consider
7487 : : * sorting the cheapest partial path.
7488 : : */
2776 7489 [ + - + + : 858 : foreach(lc, input_rel->pathlist)
+ + ]
7490 : : {
7491 : : ListCell *lc2;
7492 : 429 : Path *path = (Path *) lfirst(lc);
645 akorotkov@postgresql 7493 : 429 : Path *path_save = path;
7494 : 429 : List *pathkey_orderings = NIL;
7495 : :
7496 : : /* generate alternative group orderings that might be useful */
7497 : 429 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7498 : :
7499 [ - + ]: 429 : Assert(list_length(pathkey_orderings) > 0);
7500 : :
7501 : : /* process all potentially interesting grouping reorderings */
7502 [ + - + + : 858 : foreach(lc2, pathkey_orderings)
+ + ]
7503 : : {
508 7504 : 429 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7505 : :
7506 : : /* restore the path (we replace it in the loop) */
645 7507 : 429 : path = path_save;
7508 : :
7509 : 429 : path = make_ordered_path(root,
7510 : : partially_grouped_rel,
7511 : : path,
7512 : : cheapest_total_path,
7513 : : info->pathkeys,
7514 : : -1.0);
7515 : :
7516 [ - + ]: 429 : if (path == NULL)
645 akorotkov@postgresql 7517 :UBC 0 : continue;
7518 : :
645 akorotkov@postgresql 7519 [ + + ]:CBC 429 : if (parse->hasAggs)
7520 : 393 : add_path(partially_grouped_rel, (Path *)
7521 : 393 : create_agg_path(root,
7522 : : partially_grouped_rel,
7523 : : path,
7524 : 393 : partially_grouped_rel->reltarget,
7525 : 393 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7526 : : AGGSPLIT_INITIAL_SERIAL,
7527 : : info->clauses,
7528 : : NIL,
7529 : : agg_partial_costs,
7530 : : dNumPartialGroups));
7531 : : else
7532 : 36 : add_path(partially_grouped_rel, (Path *)
7533 : 36 : create_group_path(root,
7534 : : partially_grouped_rel,
7535 : : path,
7536 : : info->clauses,
7537 : : NIL,
7538 : : dNumPartialGroups));
7539 : : }
7540 : : }
7541 : : }
7542 : :
2776 rhaas@postgresql.org 7543 [ + - + + ]: 1550 : if (can_sort && cheapest_partial_path != NULL)
7544 : : {
7545 : : /* Similar to above logic, but for partial paths. */
2831 7546 [ + - + + : 2686 : foreach(lc, input_rel->partial_pathlist)
+ + ]
7547 : : {
7548 : : ListCell *lc2;
7549 : 1427 : Path *path = (Path *) lfirst(lc);
645 akorotkov@postgresql 7550 : 1427 : Path *path_save = path;
7551 : 1427 : List *pathkey_orderings = NIL;
7552 : :
7553 : : /* generate alternative group orderings that might be useful */
7554 : 1427 : pathkey_orderings = get_useful_group_keys_orderings(root, path);
7555 : :
7556 [ - + ]: 1427 : Assert(list_length(pathkey_orderings) > 0);
7557 : :
7558 : : /* process all potentially interesting grouping reorderings */
7559 [ + - + + : 2854 : foreach(lc2, pathkey_orderings)
+ + ]
7560 : : {
508 7561 : 1427 : GroupByOrdering *info = (GroupByOrdering *) lfirst(lc2);
7562 : :
7563 : :
7564 : : /* restore the path (we replace it in the loop) */
645 7565 : 1427 : path = path_save;
7566 : :
7567 : 1427 : path = make_ordered_path(root,
7568 : : partially_grouped_rel,
7569 : : path,
7570 : : cheapest_partial_path,
7571 : : info->pathkeys,
7572 : : -1.0);
7573 : :
7574 [ + + ]: 1427 : if (path == NULL)
7575 : 3 : continue;
7576 : :
7577 [ + + ]: 1424 : if (parse->hasAggs)
7578 : 1363 : add_partial_path(partially_grouped_rel, (Path *)
7579 : 1363 : create_agg_path(root,
7580 : : partially_grouped_rel,
7581 : : path,
7582 : 1363 : partially_grouped_rel->reltarget,
7583 : 1363 : parse->groupClause ? AGG_SORTED : AGG_PLAIN,
7584 : : AGGSPLIT_INITIAL_SERIAL,
7585 : : info->clauses,
7586 : : NIL,
7587 : : agg_partial_costs,
7588 : : dNumPartialPartialGroups));
7589 : : else
7590 : 61 : add_partial_path(partially_grouped_rel, (Path *)
7591 : 61 : create_group_path(root,
7592 : : partially_grouped_rel,
7593 : : path,
7594 : : info->clauses,
7595 : : NIL,
7596 : : dNumPartialPartialGroups));
7597 : : }
7598 : : }
7599 : : }
7600 : :
7601 : : /*
7602 : : * Add a partially-grouped HashAgg Path where possible
7603 : : */
2776 rhaas@postgresql.org 7604 [ + + + + ]: 1550 : if (can_hash && cheapest_total_path != NULL)
7605 : : {
7606 : : /* Checked above */
2831 7607 [ + + - + ]: 429 : Assert(parse->hasAggs || parse->groupClause);
7608 : :
1918 pg@bowt.ie 7609 : 429 : add_path(partially_grouped_rel, (Path *)
7610 : 429 : create_agg_path(root,
7611 : : partially_grouped_rel,
7612 : : cheapest_total_path,
7613 : 429 : partially_grouped_rel->reltarget,
7614 : : AGG_HASHED,
7615 : : AGGSPLIT_INITIAL_SERIAL,
7616 : : root->processed_groupClause,
7617 : : NIL,
7618 : : agg_partial_costs,
7619 : : dNumPartialGroups));
7620 : : }
7621 : :
7622 : : /*
7623 : : * Now add a partially-grouped HashAgg partial Path where possible
7624 : : */
2776 rhaas@postgresql.org 7625 [ + + + + ]: 1550 : if (can_hash && cheapest_partial_path != NULL)
7626 : : {
1918 pg@bowt.ie 7627 : 859 : add_partial_path(partially_grouped_rel, (Path *)
7628 : 859 : create_agg_path(root,
7629 : : partially_grouped_rel,
7630 : : cheapest_partial_path,
7631 : 859 : partially_grouped_rel->reltarget,
7632 : : AGG_HASHED,
7633 : : AGGSPLIT_INITIAL_SERIAL,
7634 : : root->processed_groupClause,
7635 : : NIL,
7636 : : agg_partial_costs,
7637 : : dNumPartialPartialGroups));
7638 : : }
7639 : :
7640 : : /*
7641 : : * Add any partially aggregated paths generated by eager aggregation to
7642 : : * the new upper relation after applying projection steps as needed.
7643 : : */
19 rguo@postgresql.org 7644 [ + + ]:GNC 1550 : if (eager_agg_rel)
7645 : : {
7646 : : /* Add the paths */
7647 [ + - + + : 1174 : foreach(lc, eager_agg_rel->pathlist)
+ + ]
7648 : : {
7649 : 725 : Path *path = (Path *) lfirst(lc);
7650 : :
7651 : : /* Shouldn't have any parameterized paths anymore */
7652 [ - + ]: 725 : Assert(path->param_info == NULL);
7653 : :
7654 : 725 : path = (Path *) create_projection_path(root,
7655 : : partially_grouped_rel,
7656 : : path,
7657 : 725 : partially_grouped_rel->reltarget);
7658 : :
7659 : 725 : add_path(partially_grouped_rel, path);
7660 : : }
7661 : :
7662 : : /*
7663 : : * Likewise add the partial paths, but only if parallelism is possible
7664 : : * for partially_grouped_rel.
7665 : : */
7666 [ + + ]: 449 : if (partially_grouped_rel->consider_parallel)
7667 : : {
7668 [ + + + + : 1014 : foreach(lc, eager_agg_rel->partial_pathlist)
+ + ]
7669 : : {
7670 : 606 : Path *path = (Path *) lfirst(lc);
7671 : :
7672 : : /* Shouldn't have any parameterized paths anymore */
7673 [ - + ]: 606 : Assert(path->param_info == NULL);
7674 : :
7675 : 606 : path = (Path *) create_projection_path(root,
7676 : : partially_grouped_rel,
7677 : : path,
7678 : 606 : partially_grouped_rel->reltarget);
7679 : :
7680 : 606 : add_partial_path(partially_grouped_rel, path);
7681 : : }
7682 : : }
7683 : : }
7684 : :
7685 : : /*
7686 : : * If there is an FDW that's responsible for all baserels of the query,
7687 : : * let it consider adding partially grouped ForeignPaths.
7688 : : */
2800 rhaas@postgresql.org 7689 [ + + ]:CBC 1550 : if (partially_grouped_rel->fdwroutine &&
7690 [ + - ]: 3 : partially_grouped_rel->fdwroutine->GetForeignUpperPaths)
7691 : : {
7692 : 3 : FdwRoutine *fdwroutine = partially_grouped_rel->fdwroutine;
7693 : :
7694 : 3 : fdwroutine->GetForeignUpperPaths(root,
7695 : : UPPERREL_PARTIAL_GROUP_AGG,
7696 : : input_rel, partially_grouped_rel,
7697 : : extra);
7698 : : }
7699 : :
2778 7700 : 1550 : return partially_grouped_rel;
7701 : : }
7702 : :
7703 : : /*
7704 : : * make_ordered_path
7705 : : * Return a path ordered by 'pathkeys' based on the given 'path'. May
7706 : : * return NULL if it doesn't make sense to generate an ordered path in
7707 : : * this case.
7708 : : */
7709 : : static Path *
335 rguo@postgresql.org 7710 : 27901 : make_ordered_path(PlannerInfo *root, RelOptInfo *rel, Path *path,
7711 : : Path *cheapest_path, List *pathkeys, double limit_tuples)
7712 : : {
7713 : : bool is_sorted;
7714 : : int presorted_keys;
7715 : :
7716 : 27901 : is_sorted = pathkeys_count_contained_in(pathkeys,
7717 : : path->pathkeys,
7718 : : &presorted_keys);
7719 : :
7720 [ + + ]: 27901 : if (!is_sorted)
7721 : : {
7722 : : /*
7723 : : * Try at least sorting the cheapest path and also try incrementally
7724 : : * sorting any path which is partially sorted already (no need to deal
7725 : : * with paths which have presorted keys when incremental sort is
7726 : : * disabled unless it's the cheapest input path).
7727 : : */
7728 [ + + ]: 8296 : if (path != cheapest_path &&
7729 [ + + + + ]: 1560 : (presorted_keys == 0 || !enable_incremental_sort))
7730 : 742 : return NULL;
7731 : :
7732 : : /*
7733 : : * We've no need to consider both a sort and incremental sort. We'll
7734 : : * just do a sort if there are no presorted keys and an incremental
7735 : : * sort when there are presorted keys.
7736 : : */
7737 [ + + + + ]: 7554 : if (presorted_keys == 0 || !enable_incremental_sort)
7738 : 6649 : path = (Path *) create_sort_path(root,
7739 : : rel,
7740 : : path,
7741 : : pathkeys,
7742 : : limit_tuples);
7743 : : else
7744 : 905 : path = (Path *) create_incremental_sort_path(root,
7745 : : rel,
7746 : : path,
7747 : : pathkeys,
7748 : : presorted_keys,
7749 : : limit_tuples);
7750 : : }
7751 : :
7752 : 27159 : return path;
7753 : : }
7754 : :
7755 : : /*
7756 : : * Generate Gather and Gather Merge paths for a grouping relation or partial
7757 : : * grouping relation.
7758 : : *
7759 : : * generate_useful_gather_paths does most of the work, but we also consider a
7760 : : * special case: we could try sorting the data by the group_pathkeys and then
7761 : : * applying Gather Merge.
7762 : : *
7763 : : * NB: This function shouldn't be used for anything other than a grouped or
7764 : : * partially grouped relation not only because of the fact that it explicitly
7765 : : * references group_pathkeys but we pass "true" as the third argument to
7766 : : * generate_useful_gather_paths().
7767 : : */
7768 : : static void
2778 rhaas@postgresql.org 7769 : 1166 : gather_grouping_paths(PlannerInfo *root, RelOptInfo *rel)
7770 : : {
7771 : : ListCell *lc;
7772 : : Path *cheapest_partial_path;
7773 : : List *groupby_pathkeys;
7774 : :
7775 : : /*
7776 : : * This occurs after any partial aggregation has taken place, so trim off
7777 : : * any pathkeys added for ORDER BY / DISTINCT aggregates.
7778 : : */
591 drowley@postgresql.o 7779 [ + + ]: 1166 : if (list_length(root->group_pathkeys) > root->num_groupby_pathkeys)
7780 : 9 : groupby_pathkeys = list_copy_head(root->group_pathkeys,
7781 : : root->num_groupby_pathkeys);
7782 : : else
7783 : 1157 : groupby_pathkeys = root->group_pathkeys;
7784 : :
7785 : : /* Try Gather for unordered paths and Gather Merge for ordered ones. */
2029 tomas.vondra@postgre 7786 : 1166 : generate_useful_gather_paths(root, rel, true);
7787 : :
2778 rhaas@postgresql.org 7788 : 1166 : cheapest_partial_path = linitial(rel->partial_pathlist);
7789 : :
7790 : : /* XXX Shouldn't this also consider the group-key-reordering? */
2029 tomas.vondra@postgre 7791 [ + - + + : 2899 : foreach(lc, rel->partial_pathlist)
+ + ]
7792 : : {
7793 : 1733 : Path *path = (Path *) lfirst(lc);
7794 : : bool is_sorted;
7795 : : int presorted_keys;
7796 : : double total_groups;
7797 : :
591 drowley@postgresql.o 7798 : 1733 : is_sorted = pathkeys_count_contained_in(groupby_pathkeys,
7799 : : path->pathkeys,
7800 : : &presorted_keys);
7801 : :
2029 tomas.vondra@postgre 7802 [ + + ]: 1733 : if (is_sorted)
7803 : 1001 : continue;
7804 : :
7805 : : /*
7806 : : * Try at least sorting the cheapest path and also try incrementally
7807 : : * sorting any path which is partially sorted already (no need to deal
7808 : : * with paths which have presorted keys when incremental sort is
7809 : : * disabled unless it's the cheapest input path).
7810 : : */
635 drowley@postgresql.o 7811 [ - + ]: 732 : if (path != cheapest_partial_path &&
635 drowley@postgresql.o 7812 [ # # # # ]:UBC 0 : (presorted_keys == 0 || !enable_incremental_sort))
2029 tomas.vondra@postgre 7813 : 0 : continue;
7814 : :
7815 : : /*
7816 : : * We've no need to consider both a sort and incremental sort. We'll
7817 : : * just do a sort if there are no presorted keys and an incremental
7818 : : * sort when there are presorted keys.
7819 : : */
635 drowley@postgresql.o 7820 [ - + - - ]:CBC 732 : if (presorted_keys == 0 || !enable_incremental_sort)
7821 : 732 : path = (Path *) create_sort_path(root, rel, path,
7822 : : groupby_pathkeys,
7823 : : -1.0);
7824 : : else
635 drowley@postgresql.o 7825 :UBC 0 : path = (Path *) create_incremental_sort_path(root,
7826 : : rel,
7827 : : path,
7828 : : groupby_pathkeys,
7829 : : presorted_keys,
7830 : : -1.0);
461 rguo@postgresql.org 7831 :CBC 732 : total_groups = compute_gather_rows(path);
7832 : : path = (Path *)
2029 tomas.vondra@postgre 7833 : 732 : create_gather_merge_path(root,
7834 : : rel,
7835 : : path,
7836 : 732 : rel->reltarget,
7837 : : groupby_pathkeys,
7838 : : NULL,
7839 : : &total_groups);
7840 : :
7841 : 732 : add_path(rel, path);
7842 : : }
3142 rhaas@postgresql.org 7843 : 1166 : }
7844 : :
7845 : : /*
7846 : : * can_partial_agg
7847 : : *
7848 : : * Determines whether or not partial grouping and/or aggregation is possible.
7849 : : * Returns true when possible, false otherwise.
7850 : : */
7851 : : static bool
1798 heikki.linnakangas@i 7852 : 18672 : can_partial_agg(PlannerInfo *root)
7853 : : {
2831 rhaas@postgresql.org 7854 : 18672 : Query *parse = root->parse;
7855 : :
2778 7856 [ + + - + ]: 18672 : if (!parse->hasAggs && parse->groupClause == NIL)
7857 : : {
7858 : : /*
7859 : : * We don't know how to do parallel aggregation unless we have either
7860 : : * some aggregates or a grouping clause.
7861 : : */
2831 rhaas@postgresql.org 7862 :UBC 0 : return false;
7863 : : }
2831 rhaas@postgresql.org 7864 [ + + ]:CBC 18672 : else if (parse->groupingSets)
7865 : : {
7866 : : /* We don't know how to do grouping sets in parallel. */
7867 : 463 : return false;
7868 : : }
1798 heikki.linnakangas@i 7869 [ + + + + ]: 18209 : else if (root->hasNonPartialAggs || root->hasNonSerialAggs)
7870 : : {
7871 : : /* Insufficient support for partial mode. */
2831 rhaas@postgresql.org 7872 : 1950 : return false;
7873 : : }
7874 : :
7875 : : /* Everything looks good. */
7876 : 16259 : return true;
7877 : : }
7878 : :
7879 : : /*
7880 : : * apply_scanjoin_target_to_paths
7881 : : *
7882 : : * Adjust the final scan/join relation, and recursively all of its children,
7883 : : * to generate the final scan/join target. It would be more correct to model
7884 : : * this as a separate planning step with a new RelOptInfo at the toplevel and
7885 : : * for each child relation, but doing it this way is noticeably cheaper.
7886 : : * Maybe that problem can be solved at some point, but for now we do this.
7887 : : *
7888 : : * If tlist_same_exprs is true, then the scan/join target to be applied has
7889 : : * the same expressions as the existing reltarget, so we need only insert the
7890 : : * appropriate sortgroupref information. By avoiding the creation of
7891 : : * projection paths we save effort both immediately and at plan creation time.
7892 : : */
7893 : : static void
2776 7894 : 268315 : apply_scanjoin_target_to_paths(PlannerInfo *root,
7895 : : RelOptInfo *rel,
7896 : : List *scanjoin_targets,
7897 : : List *scanjoin_targets_contain_srfs,
7898 : : bool scanjoin_target_parallel_safe,
7899 : : bool tlist_same_exprs)
7900 : : {
2426 tgl@sss.pgh.pa.us 7901 [ + + + + : 268315 : bool rel_is_partitioned = IS_PARTITIONED_REL(rel);
+ + + - +
+ ]
7902 : : PathTarget *scanjoin_target;
7903 : : ListCell *lc;
7904 : :
7905 : : /* This recurses, so be paranoid. */
2769 rhaas@postgresql.org 7906 : 268315 : check_stack_depth();
7907 : :
7908 : : /*
7909 : : * If the rel is partitioned, we want to drop its existing paths and
7910 : : * generate new ones. This function would still be correct if we kept the
7911 : : * existing paths: we'd modify them to generate the correct target above
7912 : : * the partitioning Append, and then they'd compete on cost with paths
7913 : : * generating the target below the Append. However, in our current cost
7914 : : * model the latter way is always the same or cheaper cost, so modifying
7915 : : * the existing paths would just be useless work. Moreover, when the cost
7916 : : * is the same, varying roundoff errors might sometimes allow an existing
7917 : : * path to be picked, resulting in undesirable cross-platform plan
7918 : : * variations. So we drop old paths and thereby force the work to be done
7919 : : * below the Append, except in the case of a non-parallel-safe target.
7920 : : *
7921 : : * Some care is needed, because we have to allow
7922 : : * generate_useful_gather_paths to see the old partial paths in the next
7923 : : * stanza. Hence, zap the main pathlist here, then allow
7924 : : * generate_useful_gather_paths to add path(s) to the main list, and
7925 : : * finally zap the partial pathlist.
7926 : : */
2426 tgl@sss.pgh.pa.us 7927 [ + + ]: 268315 : if (rel_is_partitioned)
7928 : 6450 : rel->pathlist = NIL;
7929 : :
7930 : : /*
7931 : : * If the scan/join target is not parallel-safe, partial paths cannot
7932 : : * generate it.
7933 : : */
2769 rhaas@postgresql.org 7934 [ + + ]: 268315 : if (!scanjoin_target_parallel_safe)
7935 : : {
7936 : : /*
7937 : : * Since we can't generate the final scan/join target in parallel
7938 : : * workers, this is our last opportunity to use any partial paths that
7939 : : * exist; so build Gather path(s) that use them and emit whatever the
7940 : : * current reltarget is. We don't do this in the case where the
7941 : : * target is parallel-safe, since we will be able to generate superior
7942 : : * paths by doing it after the final scan/join target has been
7943 : : * applied.
7944 : : */
2029 tomas.vondra@postgre 7945 : 39367 : generate_useful_gather_paths(root, rel, false);
7946 : :
7947 : : /* Can't use parallel query above this level. */
2769 rhaas@postgresql.org 7948 : 39367 : rel->partial_pathlist = NIL;
7949 : 39367 : rel->consider_parallel = false;
7950 : : }
7951 : :
7952 : : /* Finish dropping old paths for a partitioned rel, per comment above */
2426 tgl@sss.pgh.pa.us 7953 [ + + ]: 268315 : if (rel_is_partitioned)
2769 rhaas@postgresql.org 7954 : 6450 : rel->partial_pathlist = NIL;
7955 : :
7956 : : /* Extract SRF-free scan/join target. */
7957 : 268315 : scanjoin_target = linitial_node(PathTarget, scanjoin_targets);
7958 : :
7959 : : /*
7960 : : * Apply the SRF-free scan/join target to each existing path.
7961 : : *
7962 : : * If the tlist exprs are the same, we can just inject the sortgroupref
7963 : : * information into the existing pathtargets. Otherwise, replace each
7964 : : * path with a projection path that generates the SRF-free scan/join
7965 : : * target. This can't change the ordering of paths within rel->pathlist,
7966 : : * so we just modify the list in place.
7967 : : */
2776 7968 [ + + + + : 556381 : foreach(lc, rel->pathlist)
+ + ]
7969 : : {
7970 : 288066 : Path *subpath = (Path *) lfirst(lc);
7971 : :
7972 : : /* Shouldn't have any parameterized paths anymore */
7973 [ - + ]: 288066 : Assert(subpath->param_info == NULL);
7974 : :
2769 7975 [ + + ]: 288066 : if (tlist_same_exprs)
7976 : 102349 : subpath->pathtarget->sortgrouprefs =
7977 : 102349 : scanjoin_target->sortgrouprefs;
7978 : : else
7979 : : {
7980 : : Path *newpath;
7981 : :
2776 7982 : 185717 : newpath = (Path *) create_projection_path(root, rel, subpath,
7983 : : scanjoin_target);
7984 : 185717 : lfirst(lc) = newpath;
7985 : : }
7986 : : }
7987 : :
7988 : : /* Likewise adjust the targets for any partial paths. */
2769 7989 [ + + + + : 278916 : foreach(lc, rel->partial_pathlist)
+ + ]
7990 : : {
7991 : 10601 : Path *subpath = (Path *) lfirst(lc);
7992 : :
7993 : : /* Shouldn't have any parameterized paths anymore */
7994 [ - + ]: 10601 : Assert(subpath->param_info == NULL);
7995 : :
7996 [ + + ]: 10601 : if (tlist_same_exprs)
7997 : 8631 : subpath->pathtarget->sortgrouprefs =
7998 : 8631 : scanjoin_target->sortgrouprefs;
7999 : : else
8000 : : {
8001 : : Path *newpath;
8002 : :
2426 tgl@sss.pgh.pa.us 8003 : 1970 : newpath = (Path *) create_projection_path(root, rel, subpath,
8004 : : scanjoin_target);
2776 rhaas@postgresql.org 8005 : 1970 : lfirst(lc) = newpath;
8006 : : }
8007 : : }
8008 : :
8009 : : /*
8010 : : * Now, if final scan/join target contains SRFs, insert ProjectSetPath(s)
8011 : : * atop each existing path. (Note that this function doesn't look at the
8012 : : * cheapest-path fields, which is a good thing because they're bogus right
8013 : : * now.)
8014 : : */
2769 8015 [ + + ]: 268315 : if (root->parse->hasTargetSRFs)
8016 : 5868 : adjust_paths_for_srfs(root, rel,
8017 : : scanjoin_targets,
8018 : : scanjoin_targets_contain_srfs);
8019 : :
8020 : : /*
8021 : : * Update the rel's target to be the final (with SRFs) scan/join target.
8022 : : * This now matches the actual output of all the paths, and we might get
8023 : : * confused in createplan.c if they don't agree. We must do this now so
8024 : : * that any append paths made in the next part will use the correct
8025 : : * pathtarget (cf. create_append_path).
8026 : : *
8027 : : * Note that this is also necessary if GetForeignUpperPaths() gets called
8028 : : * on the final scan/join relation or on any of its children, since the
8029 : : * FDW might look at the rel's target to create ForeignPaths.
8030 : : */
2426 tgl@sss.pgh.pa.us 8031 : 268315 : rel->reltarget = llast_node(PathTarget, scanjoin_targets);
8032 : :
8033 : : /*
8034 : : * If the relation is partitioned, recursively apply the scan/join target
8035 : : * to all partitions, and generate brand-new Append paths in which the
8036 : : * scan/join target is computed below the Append rather than above it.
8037 : : * Since Append is not projection-capable, that might save a separate
8038 : : * Result node, and it also is important for partitionwise aggregate.
8039 : : */
8040 [ + + ]: 268315 : if (rel_is_partitioned)
8041 : : {
2769 rhaas@postgresql.org 8042 : 6450 : List *live_children = NIL;
8043 : : int i;
8044 : :
8045 : : /* Adjust each partition. */
1546 drowley@postgresql.o 8046 : 6450 : i = -1;
8047 [ + + ]: 18307 : while ((i = bms_next_member(rel->live_parts, i)) >= 0)
8048 : : {
8049 : 11857 : RelOptInfo *child_rel = rel->part_rels[i];
8050 : : AppendRelInfo **appinfos;
8051 : : int nappinfos;
2769 rhaas@postgresql.org 8052 : 11857 : List *child_scanjoin_targets = NIL;
8053 : :
1546 drowley@postgresql.o 8054 [ - + ]: 11857 : Assert(child_rel != NULL);
8055 : :
8056 : : /* Dummy children can be ignored. */
8057 [ + + ]: 11857 : if (IS_DUMMY_REL(child_rel))
2403 tgl@sss.pgh.pa.us 8058 : 21 : continue;
8059 : :
8060 : : /* Translate scan/join targets for this child. */
2769 rhaas@postgresql.org 8061 : 11836 : appinfos = find_appinfos_by_relids(root, child_rel->relids,
8062 : : &nappinfos);
8063 [ + - + + : 23672 : foreach(lc, scanjoin_targets)
+ + ]
8064 : : {
8065 : 11836 : PathTarget *target = lfirst_node(PathTarget, lc);
8066 : :
8067 : 11836 : target = copy_pathtarget(target);
8068 : 11836 : target->exprs = (List *)
8069 : 11836 : adjust_appendrel_attrs(root,
8070 : 11836 : (Node *) target->exprs,
8071 : : nappinfos, appinfos);
8072 : 11836 : child_scanjoin_targets = lappend(child_scanjoin_targets,
8073 : : target);
8074 : : }
8075 : 11836 : pfree(appinfos);
8076 : :
8077 : : /* Recursion does the real work. */
8078 : 11836 : apply_scanjoin_target_to_paths(root, child_rel,
8079 : : child_scanjoin_targets,
8080 : : scanjoin_targets_contain_srfs,
8081 : : scanjoin_target_parallel_safe,
8082 : : tlist_same_exprs);
8083 : :
8084 : : /* Save non-dummy children for Append paths. */
8085 [ + - ]: 11836 : if (!IS_DUMMY_REL(child_rel))
8086 : 11836 : live_children = lappend(live_children, child_rel);
8087 : : }
8088 : :
8089 : : /* Build new paths for this relation by appending child paths. */
2426 tgl@sss.pgh.pa.us 8090 : 6450 : add_paths_to_append_rel(root, rel, live_children);
8091 : : }
8092 : :
8093 : : /*
8094 : : * Consider generating Gather or Gather Merge paths. We must only do this
8095 : : * if the relation is parallel safe, and we don't do it for child rels to
8096 : : * avoid creating multiple Gather nodes within the same plan. We must do
8097 : : * this after all paths have been generated and before set_cheapest, since
8098 : : * one of the generated paths may turn out to be the cheapest one.
8099 : : */
2769 rhaas@postgresql.org 8100 [ + + + + : 268315 : if (rel->consider_parallel && !IS_OTHER_REL(rel))
+ + + - ]
2029 tomas.vondra@postgre 8101 : 88400 : generate_useful_gather_paths(root, rel, false);
8102 : :
8103 : : /*
8104 : : * Reassess which paths are the cheapest, now that we've potentially added
8105 : : * new Gather (or Gather Merge) and/or Append (or MergeAppend) paths to
8106 : : * this relation.
8107 : : */
2769 rhaas@postgresql.org 8108 : 268315 : set_cheapest(rel);
2776 8109 : 268315 : }
8110 : :
8111 : : /*
8112 : : * create_partitionwise_grouping_paths
8113 : : *
8114 : : * If the partition keys of input relation are part of the GROUP BY clause, all
8115 : : * the rows belonging to a given group come from a single partition. This
8116 : : * allows aggregation/grouping over a partitioned relation to be broken down
8117 : : * into aggregation/grouping on each partition. This should be no worse, and
8118 : : * often better, than the normal approach.
8119 : : *
8120 : : * However, if the GROUP BY clause does not contain all the partition keys,
8121 : : * rows from a given group may be spread across multiple partitions. In that
8122 : : * case, we perform partial aggregation for each group, append the results,
8123 : : * and then finalize aggregation. This is less certain to win than the
8124 : : * previous case. It may win if the PartialAggregate stage greatly reduces
8125 : : * the number of groups, because fewer rows will pass through the Append node.
8126 : : * It may lose if we have lots of small groups.
8127 : : */
8128 : : static void
8129 : 407 : create_partitionwise_grouping_paths(PlannerInfo *root,
8130 : : RelOptInfo *input_rel,
8131 : : RelOptInfo *grouped_rel,
8132 : : RelOptInfo *partially_grouped_rel,
8133 : : const AggClauseCosts *agg_costs,
8134 : : grouping_sets_data *gd,
8135 : : PartitionwiseAggregateType patype,
8136 : : GroupPathExtraData *extra)
8137 : : {
8138 : 407 : List *grouped_live_children = NIL;
8139 : 407 : List *partially_grouped_live_children = NIL;
2769 8140 : 407 : PathTarget *target = grouped_rel->reltarget;
2684 8141 : 407 : bool partial_grouping_valid = true;
8142 : : int i;
8143 : :
2776 8144 [ - + ]: 407 : Assert(patype != PARTITIONWISE_AGGREGATE_NONE);
8145 [ + + - + ]: 407 : Assert(patype != PARTITIONWISE_AGGREGATE_PARTIAL ||
8146 : : partially_grouped_rel != NULL);
8147 : :
8148 : : /* Add paths for partitionwise aggregation/grouping. */
1546 drowley@postgresql.o 8149 : 407 : i = -1;
8150 [ + + ]: 1472 : while ((i = bms_next_member(input_rel->live_parts, i)) >= 0)
8151 : : {
8152 : 1065 : RelOptInfo *child_input_rel = input_rel->part_rels[i];
8153 : : PathTarget *child_target;
8154 : : AppendRelInfo **appinfos;
8155 : : int nappinfos;
8156 : : GroupPathExtraData child_extra;
8157 : : RelOptInfo *child_grouped_rel;
8158 : : RelOptInfo *child_partially_grouped_rel;
8159 : :
8160 [ - + ]: 1065 : Assert(child_input_rel != NULL);
8161 : :
8162 : : /* Dummy children can be ignored. */
8163 [ - + ]: 1065 : if (IS_DUMMY_REL(child_input_rel))
2403 tgl@sss.pgh.pa.us 8164 :UBC 0 : continue;
8165 : :
1546 drowley@postgresql.o 8166 :CBC 1065 : child_target = copy_pathtarget(target);
8167 : :
8168 : : /*
8169 : : * Copy the given "extra" structure as is and then override the
8170 : : * members specific to this child.
8171 : : */
2776 rhaas@postgresql.org 8172 : 1065 : memcpy(&child_extra, extra, sizeof(child_extra));
8173 : :
8174 : 1065 : appinfos = find_appinfos_by_relids(root, child_input_rel->relids,
8175 : : &nappinfos);
8176 : :
8177 : 1065 : child_target->exprs = (List *)
8178 : 1065 : adjust_appendrel_attrs(root,
8179 : 1065 : (Node *) target->exprs,
8180 : : nappinfos, appinfos);
8181 : :
8182 : : /* Translate havingQual and targetList. */
8183 : 1065 : child_extra.havingQual = (Node *)
8184 : : adjust_appendrel_attrs(root,
8185 : : extra->havingQual,
8186 : : nappinfos, appinfos);
8187 : 1065 : child_extra.targetList = (List *)
8188 : 1065 : adjust_appendrel_attrs(root,
8189 : 1065 : (Node *) extra->targetList,
8190 : : nappinfos, appinfos);
8191 : :
8192 : : /*
8193 : : * extra->patype was the value computed for our parent rel; patype is
8194 : : * the value for this relation. For the child, our value is its
8195 : : * parent rel's value.
8196 : : */
8197 : 1065 : child_extra.patype = patype;
8198 : :
8199 : : /*
8200 : : * Create grouping relation to hold fully aggregated grouping and/or
8201 : : * aggregation paths for the child.
8202 : : */
8203 : 1065 : child_grouped_rel = make_grouping_rel(root, child_input_rel,
8204 : : child_target,
8205 : 1065 : extra->target_parallel_safe,
8206 : : child_extra.havingQual);
8207 : :
8208 : : /* Create grouping paths for this child relation. */
8209 : 1065 : create_ordinary_grouping_paths(root, child_input_rel,
8210 : : child_grouped_rel,
8211 : : agg_costs, gd, &child_extra,
8212 : : &child_partially_grouped_rel);
8213 : :
8214 [ + + ]: 1065 : if (child_partially_grouped_rel)
8215 : : {
8216 : : partially_grouped_live_children =
8217 : 771 : lappend(partially_grouped_live_children,
8218 : : child_partially_grouped_rel);
8219 : : }
8220 : : else
2684 8221 : 294 : partial_grouping_valid = false;
8222 : :
2776 8223 [ + + ]: 1065 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8224 : : {
8225 : 636 : set_cheapest(child_grouped_rel);
8226 : 636 : grouped_live_children = lappend(grouped_live_children,
8227 : : child_grouped_rel);
8228 : : }
8229 : :
8230 : 1065 : pfree(appinfos);
8231 : : }
8232 : :
8233 : : /*
8234 : : * Try to create append paths for partially grouped children. For full
8235 : : * partitionwise aggregation, we might have paths in the partial_pathlist
8236 : : * if parallel aggregation is possible. For partial partitionwise
8237 : : * aggregation, we may have paths in both pathlist and partial_pathlist.
8238 : : *
8239 : : * NB: We must have a partially grouped path for every child in order to
8240 : : * generate a partially grouped path for this relation.
8241 : : */
2684 8242 [ + + + + ]: 407 : if (partially_grouped_rel && partial_grouping_valid)
8243 : : {
8244 [ - + ]: 301 : Assert(partially_grouped_live_children != NIL);
8245 : :
2776 8246 : 301 : add_paths_to_append_rel(root, partially_grouped_rel,
8247 : : partially_grouped_live_children);
8248 : : }
8249 : :
8250 : : /* If possible, create append paths for fully grouped children. */
8251 [ + + ]: 407 : if (patype == PARTITIONWISE_AGGREGATE_FULL)
8252 : : {
2684 8253 [ - + ]: 238 : Assert(grouped_live_children != NIL);
8254 : :
2776 8255 : 238 : add_paths_to_append_rel(root, grouped_rel, grouped_live_children);
8256 : : }
8257 : 407 : }
8258 : :
8259 : : /*
8260 : : * group_by_has_partkey
8261 : : *
8262 : : * Returns true if all the partition keys of the given relation are part of
8263 : : * the GROUP BY clauses, including having matching collation, false otherwise.
8264 : : */
8265 : : static bool
8266 : 380 : group_by_has_partkey(RelOptInfo *input_rel,
8267 : : List *targetList,
8268 : : List *groupClause)
8269 : : {
8270 : 380 : List *groupexprs = get_sortgrouplist_exprs(groupClause, targetList);
8271 : 380 : int cnt = 0;
8272 : : int partnatts;
8273 : :
8274 : : /* Input relation should be partitioned. */
8275 [ - + ]: 380 : Assert(input_rel->part_scheme);
8276 : :
8277 : : /* Rule out early, if there are no partition keys present. */
8278 [ - + ]: 380 : if (!input_rel->partexprs)
2776 rhaas@postgresql.org 8279 :UBC 0 : return false;
8280 : :
2776 rhaas@postgresql.org 8281 :CBC 380 : partnatts = input_rel->part_scheme->partnatts;
8282 : :
8283 [ + + ]: 636 : for (cnt = 0; cnt < partnatts; cnt++)
8284 : : {
8285 : 398 : List *partexprs = input_rel->partexprs[cnt];
8286 : : ListCell *lc;
8287 : 398 : bool found = false;
8288 : :
8289 [ + + + + : 597 : foreach(lc, partexprs)
+ + ]
8290 : : {
8291 : : ListCell *lg;
8292 : 461 : Expr *partexpr = lfirst(lc);
353 amitlan@postgresql.o 8293 : 461 : Oid partcoll = input_rel->part_scheme->partcollation[cnt];
8294 : :
8295 [ + - + + : 720 : foreach(lg, groupexprs)
+ + ]
8296 : : {
8297 : 521 : Expr *groupexpr = lfirst(lg);
8298 : 521 : Oid groupcoll = exprCollation((Node *) groupexpr);
8299 : :
8300 : : /*
8301 : : * Note: we can assume there is at most one RelabelType node;
8302 : : * eval_const_expressions() will have simplified if more than
8303 : : * one.
8304 : : */
8305 [ + + ]: 521 : if (IsA(groupexpr, RelabelType))
8306 : 12 : groupexpr = ((RelabelType *) groupexpr)->arg;
8307 : :
8308 [ + + ]: 521 : if (equal(groupexpr, partexpr))
8309 : : {
8310 : : /*
8311 : : * Reject a match if the grouping collation does not match
8312 : : * the partitioning collation.
8313 : : */
8314 [ + + + - : 262 : if (OidIsValid(partcoll) && OidIsValid(groupcoll) &&
+ + ]
8315 : : partcoll != groupcoll)
8316 : 6 : return false;
8317 : :
8318 : 256 : found = true;
8319 : 256 : break;
8320 : : }
8321 : : }
8322 : :
8323 [ + + ]: 455 : if (found)
8324 : 256 : break;
8325 : : }
8326 : :
8327 : : /*
8328 : : * If none of the partition key expressions match with any of the
8329 : : * GROUP BY expression, return false.
8330 : : */
2776 rhaas@postgresql.org 8331 [ + + ]: 392 : if (!found)
8332 : 136 : return false;
8333 : : }
8334 : :
8335 : 238 : return true;
8336 : : }
8337 : :
8338 : : /*
8339 : : * generate_setop_child_grouplist
8340 : : * Build a SortGroupClause list defining the sort/grouping properties
8341 : : * of the child of a set operation.
8342 : : *
8343 : : * This is similar to generate_setop_grouplist() but differs as the setop
8344 : : * child query's targetlist entries may already have a tleSortGroupRef
8345 : : * assigned for other purposes, such as GROUP BYs. Here we keep the
8346 : : * SortGroupClause list in the same order as 'op' groupClauses and just adjust
8347 : : * the tleSortGroupRef to reference the TargetEntry's 'ressortgroupref'. If
8348 : : * any of the columns in the targetlist don't match to the setop's colTypes
8349 : : * then we return an empty list. This may leave some TLEs with unreferenced
8350 : : * ressortgroupref markings, but that's harmless.
8351 : : */
8352 : : static List *
524 8353 : 6244 : generate_setop_child_grouplist(SetOperationStmt *op, List *targetlist)
8354 : : {
8355 : 6244 : List *grouplist = copyObject(op->groupClauses);
8356 : : ListCell *lg;
8357 : : ListCell *lt;
8358 : : ListCell *ct;
8359 : :
8360 : 6244 : lg = list_head(grouplist);
290 drowley@postgresql.o 8361 : 6244 : ct = list_head(op->colTypes);
524 rhaas@postgresql.org 8362 [ + + + + : 24045 : foreach(lt, targetlist)
+ + ]
8363 : : {
8364 : 18008 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
8365 : : SortGroupClause *sgc;
8366 : : Oid coltype;
8367 : :
8368 : : /* resjunk columns could have sortgrouprefs. Leave these alone */
8369 [ - + ]: 18008 : if (tle->resjunk)
524 rhaas@postgresql.org 8370 :UBC 0 : continue;
8371 : :
8372 : : /*
8373 : : * We expect every non-resjunk target to have a SortGroupClause and
8374 : : * colTypes.
8375 : : */
524 rhaas@postgresql.org 8376 [ - + ]:CBC 18008 : Assert(lg != NULL);
290 drowley@postgresql.o 8377 [ - + ]: 18008 : Assert(ct != NULL);
524 rhaas@postgresql.org 8378 : 18008 : sgc = (SortGroupClause *) lfirst(lg);
290 drowley@postgresql.o 8379 : 18008 : coltype = lfirst_oid(ct);
8380 : :
8381 : : /* reject if target type isn't the same as the setop target type */
8382 [ + + ]: 18008 : if (coltype != exprType((Node *) tle->expr))
8383 : 207 : return NIL;
8384 : :
524 rhaas@postgresql.org 8385 : 17801 : lg = lnext(grouplist, lg);
290 drowley@postgresql.o 8386 : 17801 : ct = lnext(op->colTypes, ct);
8387 : :
8388 : : /* assign a tleSortGroupRef, or reuse the existing one */
524 rhaas@postgresql.org 8389 : 17801 : sgc->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
8390 : : }
8391 : :
8392 [ - + ]: 6037 : Assert(lg == NULL);
290 drowley@postgresql.o 8393 [ - + ]: 6037 : Assert(ct == NULL);
8394 : :
524 rhaas@postgresql.org 8395 : 6037 : return grouplist;
8396 : : }
8397 : :
8398 : : /*
8399 : : * create_unique_paths
8400 : : * Build a new RelOptInfo containing Paths that represent elimination of
8401 : : * distinct rows from the input data. Distinct-ness is defined according to
8402 : : * the needs of the semijoin represented by sjinfo. If it is not possible
8403 : : * to identify how to make the data unique, NULL is returned.
8404 : : *
8405 : : * If used at all, this is likely to be called repeatedly on the same rel,
8406 : : * so we cache the result.
8407 : : */
8408 : : RelOptInfo *
69 rguo@postgresql.org 8409 :GNC 4388 : create_unique_paths(PlannerInfo *root, RelOptInfo *rel, SpecialJoinInfo *sjinfo)
8410 : : {
8411 : : RelOptInfo *unique_rel;
8412 : 4388 : List *sortPathkeys = NIL;
8413 : 4388 : List *groupClause = NIL;
8414 : : MemoryContext oldcontext;
8415 : :
8416 : : /* Caller made a mistake if SpecialJoinInfo is the wrong one */
8417 [ - + ]: 4388 : Assert(sjinfo->jointype == JOIN_SEMI);
8418 [ - + ]: 4388 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
8419 : :
8420 : : /* If result already cached, return it */
8421 [ + + ]: 4388 : if (rel->unique_rel)
8422 : 910 : return rel->unique_rel;
8423 : :
8424 : : /* If it's not possible to unique-ify, return NULL */
8425 [ + + + - ]: 3478 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
8426 : 66 : return NULL;
8427 : :
8428 : : /*
8429 : : * Punt if this is a child relation and we failed to build a unique-ified
8430 : : * relation for its parent. This can happen if all the RHS columns were
8431 : : * found to be equated to constants when unique-ifying the parent table,
8432 : : * leaving no columns to unique-ify.
8433 : : */
59 8434 [ + + + + : 3412 : if (IS_OTHER_REL(rel) && rel->top_parent->unique_rel == NULL)
- + + + ]
8435 : 6 : return NULL;
8436 : :
8437 : : /*
8438 : : * When called during GEQO join planning, we are in a short-lived memory
8439 : : * context. We must make sure that the unique rel and any subsidiary data
8440 : : * structures created for a baserel survive the GEQO cycle, else the
8441 : : * baserel is trashed for future GEQO cycles. On the other hand, when we
8442 : : * are creating those for a joinrel during GEQO, we don't want them to
8443 : : * clutter the main planning context. Upshot is that the best solution is
8444 : : * to explicitly allocate memory in the same context the given RelOptInfo
8445 : : * is in.
8446 : : */
69 8447 : 3406 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
8448 : :
8449 : 3406 : unique_rel = makeNode(RelOptInfo);
8450 : 3406 : memcpy(unique_rel, rel, sizeof(RelOptInfo));
8451 : :
8452 : : /*
8453 : : * clear path info
8454 : : */
8455 : 3406 : unique_rel->pathlist = NIL;
8456 : 3406 : unique_rel->ppilist = NIL;
8457 : 3406 : unique_rel->partial_pathlist = NIL;
8458 : 3406 : unique_rel->cheapest_startup_path = NULL;
8459 : 3406 : unique_rel->cheapest_total_path = NULL;
8460 : 3406 : unique_rel->cheapest_parameterized_paths = NIL;
8461 : :
8462 : : /*
8463 : : * Build the target list for the unique rel. We also build the pathkeys
8464 : : * that represent the ordering requirements for the sort-based
8465 : : * implementation, and the list of SortGroupClause nodes that represent
8466 : : * the columns to be grouped on for the hash-based implementation.
8467 : : *
8468 : : * For a child rel, we can construct these fields from those of its
8469 : : * parent.
8470 : : */
8471 [ + + + + : 3406 : if (IS_OTHER_REL(rel))
- + ]
8472 : 216 : {
8473 : : PathTarget *child_unique_target;
8474 : : PathTarget *parent_unique_target;
8475 : :
8476 : 216 : parent_unique_target = rel->top_parent->unique_rel->reltarget;
8477 : :
8478 : 216 : child_unique_target = copy_pathtarget(parent_unique_target);
8479 : :
8480 : : /* Translate the target expressions */
8481 : 216 : child_unique_target->exprs = (List *)
8482 : 216 : adjust_appendrel_attrs_multilevel(root,
8483 : 216 : (Node *) parent_unique_target->exprs,
8484 : : rel,
8485 : 216 : rel->top_parent);
8486 : :
8487 : 216 : unique_rel->reltarget = child_unique_target;
8488 : :
8489 : 216 : sortPathkeys = rel->top_parent->unique_pathkeys;
8490 : 216 : groupClause = rel->top_parent->unique_groupclause;
8491 : : }
8492 : : else
8493 : : {
8494 : : List *newtlist;
8495 : : int nextresno;
8496 : 3190 : List *sortList = NIL;
8497 : : ListCell *lc1;
8498 : : ListCell *lc2;
8499 : :
8500 : : /*
8501 : : * The values we are supposed to unique-ify may be expressions in the
8502 : : * variables of the input rel's targetlist. We have to add any such
8503 : : * expressions to the unique rel's targetlist.
8504 : : *
8505 : : * To complicate matters, some of the values to be unique-ified may be
8506 : : * known redundant by the EquivalenceClass machinery (e.g., because
8507 : : * they have been equated to constants). There is no need to compare
8508 : : * such values during unique-ification, and indeed we had better not
8509 : : * try because the Vars involved may not have propagated as high as
8510 : : * the semijoin's level. We use make_pathkeys_for_sortclauses to
8511 : : * detect such cases, which is a tad inefficient but it doesn't seem
8512 : : * worth building specialized infrastructure for this.
8513 : : */
8514 : 3190 : newtlist = make_tlist_from_pathtarget(rel->reltarget);
8515 : 3190 : nextresno = list_length(newtlist) + 1;
8516 : :
8517 [ + - + + : 6497 : forboth(lc1, sjinfo->semi_rhs_exprs, lc2, sjinfo->semi_operators)
+ - + + +
+ + - +
+ ]
8518 : : {
8519 : 3307 : Expr *uniqexpr = lfirst(lc1);
8520 : 3307 : Oid in_oper = lfirst_oid(lc2);
8521 : : Oid sortop;
8522 : : TargetEntry *tle;
60 tgl@sss.pgh.pa.us 8523 : 3307 : bool made_tle = false;
8524 : :
69 rguo@postgresql.org 8525 : 3307 : tle = tlist_member(uniqexpr, newtlist);
8526 [ + + ]: 3307 : if (!tle)
8527 : : {
8528 : 1592 : tle = makeTargetEntry((Expr *) uniqexpr,
8529 : : nextresno,
8530 : : NULL,
8531 : : false);
8532 : 1592 : newtlist = lappend(newtlist, tle);
8533 : 1592 : nextresno++;
60 tgl@sss.pgh.pa.us 8534 : 1592 : made_tle = true;
8535 : : }
8536 : :
8537 : : /*
8538 : : * Try to build an ORDER BY list to sort the input compatibly. We
8539 : : * do this for each sortable clause even when the clauses are not
8540 : : * all sortable, so that we can detect clauses that are redundant
8541 : : * according to the pathkey machinery.
8542 : : */
8543 : 3307 : sortop = get_ordering_op_for_equality_op(in_oper, false);
8544 [ + - ]: 3307 : if (OidIsValid(sortop))
8545 : : {
8546 : : Oid eqop;
8547 : : SortGroupClause *sortcl;
8548 : :
8549 : : /*
8550 : : * The Unique node will need equality operators. Normally
8551 : : * these are the same as the IN clause operators, but if those
8552 : : * are cross-type operators then the equality operators are
8553 : : * the ones for the IN clause operators' RHS datatype.
8554 : : */
69 rguo@postgresql.org 8555 : 3307 : eqop = get_equality_op_for_ordering_op(sortop, NULL);
8556 [ - + ]: 3307 : if (!OidIsValid(eqop)) /* shouldn't happen */
69 rguo@postgresql.org 8557 [ # # ]:UNC 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
8558 : : sortop);
8559 : :
69 rguo@postgresql.org 8560 :GNC 3307 : sortcl = makeNode(SortGroupClause);
8561 : 3307 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8562 : 3307 : sortcl->eqop = eqop;
8563 : 3307 : sortcl->sortop = sortop;
8564 : 3307 : sortcl->reverse_sort = false;
8565 : 3307 : sortcl->nulls_first = false;
8566 : 3307 : sortcl->hashable = false; /* no need to make this accurate */
8567 : 3307 : sortList = lappend(sortList, sortcl);
8568 : :
8569 : : /*
8570 : : * At each step, convert the SortGroupClause list to pathkey
8571 : : * form. If the just-added SortGroupClause is redundant, the
8572 : : * result will be shorter than the SortGroupClause list.
8573 : : */
60 tgl@sss.pgh.pa.us 8574 : 3307 : sortPathkeys = make_pathkeys_for_sortclauses(root, sortList,
8575 : : newtlist);
8576 [ + + ]: 3307 : if (list_length(sortPathkeys) != list_length(sortList))
8577 : : {
8578 : : /* Drop the redundant SortGroupClause */
8579 : 1026 : sortList = list_delete_last(sortList);
8580 [ - + ]: 1026 : Assert(list_length(sortPathkeys) == list_length(sortList));
8581 : : /* Undo tlist addition, if we made one */
8582 [ + + ]: 1026 : if (made_tle)
8583 : : {
8584 : 6 : newtlist = list_delete_last(newtlist);
8585 : 6 : nextresno--;
8586 : : }
8587 : : /* We need not consider this clause for hashing, either */
8588 : 1026 : continue;
8589 : : }
8590 : : }
60 tgl@sss.pgh.pa.us 8591 [ # # ]:UNC 0 : else if (sjinfo->semi_can_btree) /* shouldn't happen */
8592 [ # # ]: 0 : elog(ERROR, "could not find ordering operator for equality operator %u",
8593 : : in_oper);
8594 : :
69 rguo@postgresql.org 8595 [ + - ]:GNC 2281 : if (sjinfo->semi_can_hash)
8596 : : {
8597 : : /* Create a GROUP BY list for the Agg node to use */
8598 : : Oid eq_oper;
8599 : : SortGroupClause *groupcl;
8600 : :
8601 : : /*
8602 : : * Get the hashable equality operators for the Agg node to
8603 : : * use. Normally these are the same as the IN clause
8604 : : * operators, but if those are cross-type operators then the
8605 : : * equality operators are the ones for the IN clause
8606 : : * operators' RHS datatype.
8607 : : */
8608 [ - + ]: 2281 : if (!get_compatible_hash_operators(in_oper, NULL, &eq_oper))
69 rguo@postgresql.org 8609 [ # # ]:UNC 0 : elog(ERROR, "could not find compatible hash operator for operator %u",
8610 : : in_oper);
8611 : :
69 rguo@postgresql.org 8612 :GNC 2281 : groupcl = makeNode(SortGroupClause);
8613 : 2281 : groupcl->tleSortGroupRef = assignSortGroupRef(tle, newtlist);
8614 : 2281 : groupcl->eqop = eq_oper;
8615 : 2281 : groupcl->sortop = sortop;
8616 : 2281 : groupcl->reverse_sort = false;
8617 : 2281 : groupcl->nulls_first = false;
8618 : 2281 : groupcl->hashable = true;
8619 : 2281 : groupClause = lappend(groupClause, groupcl);
8620 : : }
8621 : : }
8622 : :
8623 : : /*
8624 : : * Done building the sortPathkeys and groupClause. But the
8625 : : * sortPathkeys are bogus if not all the clauses were sortable.
8626 : : */
60 tgl@sss.pgh.pa.us 8627 [ - + ]: 3190 : if (!sjinfo->semi_can_btree)
60 tgl@sss.pgh.pa.us 8628 :UNC 0 : sortPathkeys = NIL;
8629 : :
8630 : : /*
8631 : : * It can happen that all the RHS columns are equated to constants.
8632 : : * We'd have to do something special to unique-ify in that case, and
8633 : : * it's such an unlikely-in-the-real-world case that it's not worth
8634 : : * the effort. So just punt if we found no columns to unique-ify.
8635 : : */
60 tgl@sss.pgh.pa.us 8636 [ + + + - ]:GNC 3190 : if (sortPathkeys == NIL && groupClause == NIL)
8637 : : {
8638 : 975 : MemoryContextSwitchTo(oldcontext);
8639 : 975 : return NULL;
8640 : : }
8641 : :
8642 : : /* Convert the required targetlist back to PathTarget form */
69 rguo@postgresql.org 8643 : 2215 : unique_rel->reltarget = create_pathtarget(root, newtlist);
8644 : : }
8645 : :
8646 : : /* build unique paths based on input rel's pathlist */
8647 : 2431 : create_final_unique_paths(root, rel, sortPathkeys, groupClause,
8648 : : sjinfo, unique_rel);
8649 : :
8650 : : /* build unique paths based on input rel's partial_pathlist */
8651 : 2431 : create_partial_unique_paths(root, rel, sortPathkeys, groupClause,
8652 : : sjinfo, unique_rel);
8653 : :
8654 : : /* Now choose the best path(s) */
8655 : 2431 : set_cheapest(unique_rel);
8656 : :
8657 : : /*
8658 : : * There shouldn't be any partial paths for the unique relation;
8659 : : * otherwise, we won't be able to properly guarantee uniqueness.
8660 : : */
8661 [ - + ]: 2431 : Assert(unique_rel->partial_pathlist == NIL);
8662 : :
8663 : : /* Cache the result */
8664 : 2431 : rel->unique_rel = unique_rel;
8665 : 2431 : rel->unique_pathkeys = sortPathkeys;
8666 : 2431 : rel->unique_groupclause = groupClause;
8667 : :
8668 : 2431 : MemoryContextSwitchTo(oldcontext);
8669 : :
8670 : 2431 : return unique_rel;
8671 : : }
8672 : :
8673 : : /*
8674 : : * create_final_unique_paths
8675 : : * Create unique paths in 'unique_rel' based on 'input_rel' pathlist
8676 : : */
8677 : : static void
8678 : 4248 : create_final_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8679 : : List *sortPathkeys, List *groupClause,
8680 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8681 : : {
8682 : 4248 : Path *cheapest_input_path = input_rel->cheapest_total_path;
8683 : :
8684 : : /* Estimate number of output rows */
8685 : 4248 : unique_rel->rows = estimate_num_groups(root,
8686 : : sjinfo->semi_rhs_exprs,
8687 : : cheapest_input_path->rows,
8688 : : NULL,
8689 : : NULL);
8690 : :
8691 : : /* Consider sort-based implementations, if possible. */
8692 [ + - ]: 4248 : if (sjinfo->semi_can_btree)
8693 : : {
8694 : : ListCell *lc;
8695 : :
8696 : : /*
8697 : : * Use any available suitably-sorted path as input, and also consider
8698 : : * sorting the cheapest-total path and incremental sort on any paths
8699 : : * with presorted keys.
8700 : : *
8701 : : * To save planning time, we ignore parameterized input paths unless
8702 : : * they are the cheapest-total path.
8703 : : */
8704 [ + - + + : 9274 : foreach(lc, input_rel->pathlist)
+ + ]
8705 : : {
8706 : 5026 : Path *input_path = (Path *) lfirst(lc);
8707 : : Path *path;
8708 : : bool is_sorted;
8709 : : int presorted_keys;
8710 : :
8711 : : /*
8712 : : * Ignore parameterized paths that are not the cheapest-total
8713 : : * path.
8714 : : */
8715 [ + + + + ]: 5026 : if (input_path->param_info &&
8716 : : input_path != cheapest_input_path)
8717 : 458 : continue;
8718 : :
8719 : 4590 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8720 : : input_path->pathkeys,
8721 : : &presorted_keys);
8722 : :
8723 : : /*
8724 : : * Ignore paths that are not suitably or partially sorted, unless
8725 : : * they are the cheapest total path (no need to deal with paths
8726 : : * which have presorted keys when incremental sort is disabled).
8727 : : */
8728 [ + + + + ]: 4590 : if (!is_sorted && input_path != cheapest_input_path &&
8729 [ + + - + ]: 46 : (presorted_keys == 0 || !enable_incremental_sort))
8730 : 22 : continue;
8731 : :
8732 : : /*
8733 : : * Make a separate ProjectionPath in case we need a Result node.
8734 : : */
8735 : 4568 : path = (Path *) create_projection_path(root,
8736 : : unique_rel,
8737 : : input_path,
8738 : 4568 : unique_rel->reltarget);
8739 : :
8740 [ + + ]: 4568 : if (!is_sorted)
8741 : : {
8742 : : /*
8743 : : * We've no need to consider both a sort and incremental sort.
8744 : : * We'll just do a sort if there are no presorted keys and an
8745 : : * incremental sort when there are presorted keys.
8746 : : */
8747 [ + + - + ]: 2421 : if (presorted_keys == 0 || !enable_incremental_sort)
8748 : 2397 : path = (Path *) create_sort_path(root,
8749 : : unique_rel,
8750 : : path,
8751 : : sortPathkeys,
8752 : : -1.0);
8753 : : else
8754 : 24 : path = (Path *) create_incremental_sort_path(root,
8755 : : unique_rel,
8756 : : path,
8757 : : sortPathkeys,
8758 : : presorted_keys,
8759 : : -1.0);
8760 : : }
8761 : :
8762 : 4568 : path = (Path *) create_unique_path(root, unique_rel, path,
8763 : : list_length(sortPathkeys),
8764 : : unique_rel->rows);
8765 : :
8766 : 4568 : add_path(unique_rel, path);
8767 : : }
8768 : : }
8769 : :
8770 : : /* Consider hash-based implementation, if possible. */
8771 [ + - ]: 4248 : if (sjinfo->semi_can_hash)
8772 : : {
8773 : : Path *path;
8774 : :
8775 : : /*
8776 : : * Make a separate ProjectionPath in case we need a Result node.
8777 : : */
8778 : 4248 : path = (Path *) create_projection_path(root,
8779 : : unique_rel,
8780 : : cheapest_input_path,
8781 : 4248 : unique_rel->reltarget);
8782 : :
8783 : 4248 : path = (Path *) create_agg_path(root,
8784 : : unique_rel,
8785 : : path,
8786 : : cheapest_input_path->pathtarget,
8787 : : AGG_HASHED,
8788 : : AGGSPLIT_SIMPLE,
8789 : : groupClause,
8790 : : NIL,
8791 : : NULL,
8792 : : unique_rel->rows);
8793 : :
8794 : 4248 : add_path(unique_rel, path);
8795 : : }
8796 : 4248 : }
8797 : :
8798 : : /*
8799 : : * create_partial_unique_paths
8800 : : * Create unique paths in 'unique_rel' based on 'input_rel' partial_pathlist
8801 : : */
8802 : : static void
8803 : 2431 : create_partial_unique_paths(PlannerInfo *root, RelOptInfo *input_rel,
8804 : : List *sortPathkeys, List *groupClause,
8805 : : SpecialJoinInfo *sjinfo, RelOptInfo *unique_rel)
8806 : : {
8807 : : RelOptInfo *partial_unique_rel;
8808 : : Path *cheapest_partial_path;
8809 : :
8810 : : /* nothing to do when there are no partial paths in the input rel */
8811 [ + + + + ]: 2431 : if (!input_rel->consider_parallel || input_rel->partial_pathlist == NIL)
8812 : 614 : return;
8813 : :
8814 : : /*
8815 : : * nothing to do if there's anything in the targetlist that's
8816 : : * parallel-restricted.
8817 : : */
8818 [ - + ]: 1817 : if (!is_parallel_safe(root, (Node *) unique_rel->reltarget->exprs))
69 rguo@postgresql.org 8819 :UNC 0 : return;
8820 : :
69 rguo@postgresql.org 8821 :GNC 1817 : cheapest_partial_path = linitial(input_rel->partial_pathlist);
8822 : :
8823 : 1817 : partial_unique_rel = makeNode(RelOptInfo);
8824 : 1817 : memcpy(partial_unique_rel, input_rel, sizeof(RelOptInfo));
8825 : :
8826 : : /*
8827 : : * clear path info
8828 : : */
8829 : 1817 : partial_unique_rel->pathlist = NIL;
8830 : 1817 : partial_unique_rel->ppilist = NIL;
8831 : 1817 : partial_unique_rel->partial_pathlist = NIL;
8832 : 1817 : partial_unique_rel->cheapest_startup_path = NULL;
8833 : 1817 : partial_unique_rel->cheapest_total_path = NULL;
8834 : 1817 : partial_unique_rel->cheapest_parameterized_paths = NIL;
8835 : :
8836 : : /* Estimate number of output rows */
8837 : 1817 : partial_unique_rel->rows = estimate_num_groups(root,
8838 : : sjinfo->semi_rhs_exprs,
8839 : : cheapest_partial_path->rows,
8840 : : NULL,
8841 : : NULL);
8842 : 1817 : partial_unique_rel->reltarget = unique_rel->reltarget;
8843 : :
8844 : : /* Consider sort-based implementations, if possible. */
8845 [ + - ]: 1817 : if (sjinfo->semi_can_btree)
8846 : : {
8847 : : ListCell *lc;
8848 : :
8849 : : /*
8850 : : * Use any available suitably-sorted path as input, and also consider
8851 : : * sorting the cheapest partial path and incremental sort on any paths
8852 : : * with presorted keys.
8853 : : */
8854 [ + - + + : 3790 : foreach(lc, input_rel->partial_pathlist)
+ + ]
8855 : : {
8856 : 1973 : Path *input_path = (Path *) lfirst(lc);
8857 : : Path *path;
8858 : : bool is_sorted;
8859 : : int presorted_keys;
8860 : :
8861 : 1973 : is_sorted = pathkeys_count_contained_in(sortPathkeys,
8862 : : input_path->pathkeys,
8863 : : &presorted_keys);
8864 : :
8865 : : /*
8866 : : * Ignore paths that are not suitably or partially sorted, unless
8867 : : * they are the cheapest partial path (no need to deal with paths
8868 : : * which have presorted keys when incremental sort is disabled).
8869 : : */
8870 [ + + - + ]: 1973 : if (!is_sorted && input_path != cheapest_partial_path &&
69 rguo@postgresql.org 8871 [ # # # # ]:UNC 0 : (presorted_keys == 0 || !enable_incremental_sort))
8872 : 0 : continue;
8873 : :
8874 : : /*
8875 : : * Make a separate ProjectionPath in case we need a Result node.
8876 : : */
69 rguo@postgresql.org 8877 :GNC 1973 : path = (Path *) create_projection_path(root,
8878 : : partial_unique_rel,
8879 : : input_path,
8880 : 1973 : partial_unique_rel->reltarget);
8881 : :
8882 [ + + ]: 1973 : if (!is_sorted)
8883 : : {
8884 : : /*
8885 : : * We've no need to consider both a sort and incremental sort.
8886 : : * We'll just do a sort if there are no presorted keys and an
8887 : : * incremental sort when there are presorted keys.
8888 : : */
8889 [ - + - - ]: 1793 : if (presorted_keys == 0 || !enable_incremental_sort)
8890 : 1793 : path = (Path *) create_sort_path(root,
8891 : : partial_unique_rel,
8892 : : path,
8893 : : sortPathkeys,
8894 : : -1.0);
8895 : : else
69 rguo@postgresql.org 8896 :UNC 0 : path = (Path *) create_incremental_sort_path(root,
8897 : : partial_unique_rel,
8898 : : path,
8899 : : sortPathkeys,
8900 : : presorted_keys,
8901 : : -1.0);
8902 : : }
8903 : :
69 rguo@postgresql.org 8904 :GNC 1973 : path = (Path *) create_unique_path(root, partial_unique_rel, path,
8905 : : list_length(sortPathkeys),
8906 : : partial_unique_rel->rows);
8907 : :
8908 : 1973 : add_partial_path(partial_unique_rel, path);
8909 : : }
8910 : : }
8911 : :
8912 : : /* Consider hash-based implementation, if possible. */
8913 [ + - ]: 1817 : if (sjinfo->semi_can_hash)
8914 : : {
8915 : : Path *path;
8916 : :
8917 : : /*
8918 : : * Make a separate ProjectionPath in case we need a Result node.
8919 : : */
8920 : 1817 : path = (Path *) create_projection_path(root,
8921 : : partial_unique_rel,
8922 : : cheapest_partial_path,
8923 : 1817 : partial_unique_rel->reltarget);
8924 : :
8925 : 1817 : path = (Path *) create_agg_path(root,
8926 : : partial_unique_rel,
8927 : : path,
8928 : : cheapest_partial_path->pathtarget,
8929 : : AGG_HASHED,
8930 : : AGGSPLIT_SIMPLE,
8931 : : groupClause,
8932 : : NIL,
8933 : : NULL,
8934 : : partial_unique_rel->rows);
8935 : :
8936 : 1817 : add_partial_path(partial_unique_rel, path);
8937 : : }
8938 : :
8939 [ + - ]: 1817 : if (partial_unique_rel->partial_pathlist != NIL)
8940 : : {
8941 : 1817 : generate_useful_gather_paths(root, partial_unique_rel, true);
8942 : 1817 : set_cheapest(partial_unique_rel);
8943 : :
8944 : : /*
8945 : : * Finally, create paths to unique-ify the final result. This step is
8946 : : * needed to remove any duplicates due to combining rows from parallel
8947 : : * workers.
8948 : : */
8949 : 1817 : create_final_unique_paths(root, partial_unique_rel,
8950 : : sortPathkeys, groupClause,
8951 : : sjinfo, unique_rel);
8952 : : }
8953 : : }
8954 : :
8955 : : /*
8956 : : * Choose a unique name for some subroot.
8957 : : *
8958 : : * Modifies glob->subplanNames to track names already used.
8959 : : */
8960 : : char *
20 rhaas@postgresql.org 8961 : 39168 : choose_plan_name(PlannerGlobal *glob, const char *name, bool always_number)
8962 : : {
8963 : : unsigned n;
8964 : :
8965 : : /*
8966 : : * If a numeric suffix is not required, then search the list of
8967 : : * previously-assigned names for a match. If none is found, then we can
8968 : : * use the provided name without modification.
8969 : : */
8970 [ + + ]: 39168 : if (!always_number)
8971 : : {
8972 : 10010 : bool found = false;
8973 : :
8974 [ + + + + : 24943 : foreach_ptr(char, subplan_name, glob->subplanNames)
+ + ]
8975 : : {
8976 [ + + ]: 7782 : if (strcmp(subplan_name, name) == 0)
8977 : : {
8978 : 2859 : found = true;
8979 : 2859 : break;
8980 : : }
8981 : : }
8982 : :
8983 [ + + ]: 10010 : if (!found)
8984 : : {
8985 : : /* pstrdup here is just to avoid cast-away-const */
8986 : 7151 : char *chosen_name = pstrdup(name);
8987 : :
8988 : 7151 : glob->subplanNames = lappend(glob->subplanNames, chosen_name);
8989 : 7151 : return chosen_name;
8990 : : }
8991 : : }
8992 : :
8993 : : /*
8994 : : * If a numeric suffix is required or if the un-suffixed name is already
8995 : : * in use, then loop until we find a positive integer that produces a
8996 : : * novel name.
8997 : : */
8998 : 32017 : for (n = 1; true; ++n)
8999 : 27649 : {
9000 : 59666 : char *proposed_name = psprintf("%s_%u", name, n);
9001 : 59666 : bool found = false;
9002 : :
9003 [ + + + + : 228137 : foreach_ptr(char, subplan_name, glob->subplanNames)
+ + ]
9004 : : {
9005 [ + + ]: 136454 : if (strcmp(subplan_name, proposed_name) == 0)
9006 : : {
9007 : 27649 : found = true;
9008 : 27649 : break;
9009 : : }
9010 : : }
9011 : :
9012 [ + + ]: 59666 : if (!found)
9013 : : {
9014 : 32017 : glob->subplanNames = lappend(glob->subplanNames, proposed_name);
9015 : 32017 : return proposed_name;
9016 : : }
9017 : :
9018 : 27649 : pfree(proposed_name);
9019 : : }
9020 : : }
|
