Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * allpaths.c
4 : : * Routines to find possible search paths for processing a query
5 : : *
6 : : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/optimizer/path/allpaths.c
12 : : *
13 : : *-------------------------------------------------------------------------
14 : : */
15 : :
16 : : #include "postgres.h"
17 : :
18 : : #include <limits.h>
19 : : #include <math.h>
20 : :
21 : : #include "access/sysattr.h"
22 : : #include "access/tsmapi.h"
23 : : #include "catalog/pg_class.h"
24 : : #include "catalog/pg_operator.h"
25 : : #include "catalog/pg_proc.h"
26 : : #include "foreign/fdwapi.h"
27 : : #include "miscadmin.h"
28 : : #include "nodes/makefuncs.h"
29 : : #include "nodes/nodeFuncs.h"
30 : : #include "nodes/supportnodes.h"
31 : : #ifdef OPTIMIZER_DEBUG
32 : : #include "nodes/print.h"
33 : : #endif
34 : : #include "optimizer/appendinfo.h"
35 : : #include "optimizer/clauses.h"
36 : : #include "optimizer/cost.h"
37 : : #include "optimizer/geqo.h"
38 : : #include "optimizer/optimizer.h"
39 : : #include "optimizer/pathnode.h"
40 : : #include "optimizer/paths.h"
41 : : #include "optimizer/plancat.h"
42 : : #include "optimizer/planner.h"
43 : : #include "optimizer/prep.h"
44 : : #include "optimizer/tlist.h"
45 : : #include "parser/parse_clause.h"
46 : : #include "parser/parsetree.h"
47 : : #include "partitioning/partbounds.h"
48 : : #include "port/pg_bitutils.h"
49 : : #include "rewrite/rewriteManip.h"
50 : : #include "utils/lsyscache.h"
51 : : #include "utils/selfuncs.h"
52 : :
53 : :
54 : : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
55 : : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
56 : : #define UNSAFE_HAS_SET_FUNC (1 << 1)
57 : : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
58 : : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
59 : : #define UNSAFE_TYPE_MISMATCH (1 << 4)
60 : :
61 : : /* results of subquery_is_pushdown_safe */
62 : : typedef struct pushdown_safety_info
63 : : {
64 : : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
65 : : * column is unsafe for qual pushdown, or 0 if
66 : : * no reason. */
67 : : bool unsafeVolatile; /* don't push down volatile quals */
68 : : bool unsafeLeaky; /* don't push down leaky quals */
69 : : } pushdown_safety_info;
70 : :
71 : : /* Return type for qual_is_pushdown_safe */
72 : : typedef enum pushdown_safe_type
73 : : {
74 : : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
75 : : PUSHDOWN_SAFE, /* safe to push qual into subquery */
76 : : PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
77 : : * run condition */
78 : : } pushdown_safe_type;
79 : :
80 : : /* These parameters are set by GUC */
81 : : bool enable_geqo = false; /* just in case GUC doesn't set it */
82 : : bool enable_eager_aggregate = true;
83 : : int geqo_threshold;
84 : : double min_eager_agg_group_size;
85 : : int min_parallel_table_scan_size;
86 : : int min_parallel_index_scan_size;
87 : :
88 : : /* Hook for plugins to get control in set_rel_pathlist() */
89 : : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
90 : :
91 : : /* Hook for plugins to replace standard_join_search() */
92 : : join_search_hook_type join_search_hook = NULL;
93 : :
94 : :
95 : : static void set_base_rel_consider_startup(PlannerInfo *root);
96 : : static void set_base_rel_sizes(PlannerInfo *root);
97 : : static void setup_simple_grouped_rels(PlannerInfo *root);
98 : : static void set_base_rel_pathlists(PlannerInfo *root);
99 : : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
100 : : Index rti, RangeTblEntry *rte);
101 : : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
102 : : Index rti, RangeTblEntry *rte);
103 : : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
104 : : RangeTblEntry *rte);
105 : : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
106 : : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
107 : : RangeTblEntry *rte);
108 : : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
109 : : RangeTblEntry *rte);
110 : : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
111 : : RangeTblEntry *rte);
112 : : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
113 : : RangeTblEntry *rte);
114 : : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
115 : : RangeTblEntry *rte);
116 : : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
117 : : RangeTblEntry *rte);
118 : : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
119 : : Index rti, RangeTblEntry *rte);
120 : : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
121 : : Index rti, RangeTblEntry *rte);
122 : : static void set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel);
123 : : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
124 : : List *live_childrels,
125 : : List *all_child_pathkeys);
126 : : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
127 : : RelOptInfo *rel,
128 : : Relids required_outer);
129 : : static void accumulate_append_subpath(Path *path,
130 : : List **subpaths,
131 : : List **special_subpaths,
132 : : List **child_append_relid_sets);
133 : : static Path *get_singleton_append_subpath(Path *path,
134 : : List **child_append_relid_sets);
135 : : static void set_dummy_rel_pathlist(RelOptInfo *rel);
136 : : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : : Index rti, RangeTblEntry *rte);
138 : : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : : RangeTblEntry *rte);
140 : : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : : RangeTblEntry *rte);
142 : : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : : RangeTblEntry *rte);
144 : : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
145 : : RangeTblEntry *rte);
146 : : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
147 : : RangeTblEntry *rte);
148 : : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
149 : : RangeTblEntry *rte);
150 : : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
151 : : RangeTblEntry *rte);
152 : : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
153 : : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
154 : : pushdown_safety_info *safetyInfo);
155 : : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
156 : : pushdown_safety_info *safetyInfo);
157 : : static void check_output_expressions(Query *subquery,
158 : : pushdown_safety_info *safetyInfo);
159 : : static void compare_tlist_datatypes(List *tlist, List *colTypes,
160 : : pushdown_safety_info *safetyInfo);
161 : : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
162 : : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
163 : : RestrictInfo *rinfo,
164 : : pushdown_safety_info *safetyInfo);
165 : : static void subquery_push_qual(Query *subquery,
166 : : RangeTblEntry *rte, Index rti, Node *qual);
167 : : static void recurse_push_qual(Node *setOp, Query *topquery,
168 : : RangeTblEntry *rte, Index rti, Node *qual);
169 : : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
170 : : Bitmapset *extra_used_attrs);
171 : :
172 : :
173 : : /*
174 : : * make_one_rel
175 : : * Finds all possible access paths for executing a query, returning a
176 : : * single rel that represents the join of all base rels in the query.
177 : : */
178 : : RelOptInfo *
179 : 33896 : make_one_rel(PlannerInfo *root, List *joinlist)
180 : : {
181 : 33896 : RelOptInfo *rel;
182 : 33896 : Index rti;
183 : 33896 : double total_pages;
184 : :
185 : : /* Mark base rels as to whether we care about fast-start plans */
186 : 33896 : set_base_rel_consider_startup(root);
187 : :
188 : : /*
189 : : * Compute size estimates and consider_parallel flags for each base rel.
190 : : */
191 : 33896 : set_base_rel_sizes(root);
192 : :
193 : : /*
194 : : * Build grouped relations for simple rels (i.e., base or "other" member
195 : : * relations) where possible.
196 : : */
197 : 33896 : setup_simple_grouped_rels(root);
198 : :
199 : : /*
200 : : * We should now have size estimates for every actual table involved in
201 : : * the query, and we also know which if any have been deleted from the
202 : : * query by join removal, pruned by partition pruning, or eliminated by
203 : : * constraint exclusion. So we can now compute total_table_pages.
204 : : *
205 : : * Note that appendrels are not double-counted here, even though we don't
206 : : * bother to distinguish RelOptInfos for appendrel parents, because the
207 : : * parents will have pages = 0.
208 : : *
209 : : * XXX if a table is self-joined, we will count it once per appearance,
210 : : * which perhaps is the wrong thing ... but that's not completely clear,
211 : : * and detecting self-joins here is difficult, so ignore it for now.
212 : : */
213 : 33896 : total_pages = 0;
214 [ + + ]: 104507 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
215 : : {
216 : 70611 : RelOptInfo *brel = root->simple_rel_array[rti];
217 : :
218 : : /* there may be empty slots corresponding to non-baserel RTEs */
219 [ + + ]: 70611 : if (brel == NULL)
220 : 16382 : continue;
221 : :
222 [ + - ]: 54229 : Assert(brel->relid == rti); /* sanity check on array */
223 : :
224 [ + + ]: 54229 : if (IS_DUMMY_REL(brel))
225 : 224 : continue;
226 : :
227 [ + + + - ]: 54005 : if (IS_SIMPLE_REL(brel))
228 : 54005 : total_pages += (double) brel->pages;
229 [ - + + ]: 70611 : }
230 : 33896 : root->total_table_pages = total_pages;
231 : :
232 : : /*
233 : : * Generate access paths for each base rel.
234 : : */
235 : 33896 : set_base_rel_pathlists(root);
236 : :
237 : : /*
238 : : * Generate access paths for the entire join tree.
239 : : */
240 : 33896 : rel = make_rel_from_joinlist(root, joinlist);
241 : :
242 : : /*
243 : : * The result should join all and only the query's base + outer-join rels.
244 : : */
245 [ + - ]: 33896 : Assert(bms_equal(rel->relids, root->all_query_rels));
246 : :
247 : 67792 : return rel;
248 : 33896 : }
249 : :
250 : : /*
251 : : * set_base_rel_consider_startup
252 : : * Set the consider_[param_]startup flags for each base-relation entry.
253 : : *
254 : : * For the moment, we only deal with consider_param_startup here; because the
255 : : * logic for consider_startup is pretty trivial and is the same for every base
256 : : * relation, we just let build_simple_rel() initialize that flag correctly to
257 : : * start with. If that logic ever gets more complicated it would probably
258 : : * be better to move it here.
259 : : */
260 : : static void
261 : 33901 : set_base_rel_consider_startup(PlannerInfo *root)
262 : : {
263 : : /*
264 : : * Since parameterized paths can only be used on the inside of a nestloop
265 : : * join plan, there is usually little value in considering fast-start
266 : : * plans for them. However, for relations that are on the RHS of a SEMI
267 : : * or ANTI join, a fast-start plan can be useful because we're only going
268 : : * to care about fetching one tuple anyway.
269 : : *
270 : : * To minimize growth of planning time, we currently restrict this to
271 : : * cases where the RHS is a single base relation, not a join; there is no
272 : : * provision for consider_param_startup to get set at all on joinrels.
273 : : * Also we don't worry about appendrels. costsize.c's costing rules for
274 : : * nestloop semi/antijoins don't consider such cases either.
275 : : */
276 : 33901 : ListCell *lc;
277 : :
278 [ + + + + : 38056 : foreach(lc, root->join_info_list)
+ + ]
279 : : {
280 : 4155 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
281 : 4155 : int varno;
282 : :
283 [ + + + + ]: 4155 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
284 : 4155 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
285 : : {
286 : 1066 : RelOptInfo *rel = find_base_rel(root, varno);
287 : :
288 : 1066 : rel->consider_param_startup = true;
289 : 1066 : }
290 : 4155 : }
291 : 33901 : }
292 : :
293 : : /*
294 : : * set_base_rel_sizes
295 : : * Set the size estimates (rows and widths) for each base-relation entry.
296 : : * Also determine whether to consider parallel paths for base relations.
297 : : *
298 : : * We do this in a separate pass over the base rels so that rowcount
299 : : * estimates are available for parameterized path generation, and also so
300 : : * that each rel's consider_parallel flag is set correctly before we begin to
301 : : * generate paths.
302 : : */
303 : : static void
304 : 33901 : set_base_rel_sizes(PlannerInfo *root)
305 : : {
306 : 33901 : Index rti;
307 : :
308 [ + + ]: 104517 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
309 : : {
310 : 70616 : RelOptInfo *rel = root->simple_rel_array[rti];
311 : 70616 : RangeTblEntry *rte;
312 : :
313 : : /* there may be empty slots corresponding to non-baserel RTEs */
314 [ + + ]: 70616 : if (rel == NULL)
315 : 16382 : continue;
316 : :
317 [ + - ]: 54234 : Assert(rel->relid == rti); /* sanity check on array */
318 : :
319 : : /* ignore RTEs that are "other rels" */
320 [ + + ]: 54234 : if (rel->reloptkind != RELOPT_BASEREL)
321 : 7800 : continue;
322 : :
323 : 46434 : rte = root->simple_rte_array[rti];
324 : :
325 : : /*
326 : : * If parallelism is allowable for this query in general, see whether
327 : : * it's allowable for this rel in particular. We have to do this
328 : : * before set_rel_size(), because (a) if this rel is an inheritance
329 : : * parent, set_append_rel_size() will use and perhaps change the rel's
330 : : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
331 : : * goes ahead and makes paths immediately.
332 : : */
333 [ + + ]: 46434 : if (root->glob->parallelModeOK)
334 : 38410 : set_rel_consider_parallel(root, rel, rte);
335 : :
336 : 46434 : set_rel_size(root, rel, rti, rte);
337 [ - + + ]: 70616 : }
338 : 33901 : }
339 : :
340 : : /*
341 : : * setup_simple_grouped_rels
342 : : * For each simple relation, build a grouped simple relation if eager
343 : : * aggregation is possible and if this relation can produce grouped paths.
344 : : */
345 : : static void
346 : 33896 : setup_simple_grouped_rels(PlannerInfo *root)
347 : : {
348 : 33896 : Index rti;
349 : :
350 : : /*
351 : : * If there are no aggregate expressions or grouping expressions, eager
352 : : * aggregation is not possible.
353 : : */
354 [ + + + + ]: 33896 : if (root->agg_clause_list == NIL ||
355 : 101 : root->group_expr_list == NIL)
356 : 33804 : return;
357 : :
358 [ + + ]: 812 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
359 : : {
360 : 720 : RelOptInfo *rel = root->simple_rel_array[rti];
361 : :
362 : : /* there may be empty slots corresponding to non-baserel RTEs */
363 [ + + ]: 720 : if (rel == NULL)
364 : 222 : continue;
365 : :
366 [ + - ]: 498 : Assert(rel->relid == rti); /* sanity check on array */
367 [ + + + - ]: 498 : Assert(IS_SIMPLE_REL(rel)); /* sanity check on rel */
368 : :
369 : 498 : (void) build_simple_grouped_rel(root, rel);
370 [ + + ]: 720 : }
371 : 33896 : }
372 : :
373 : : /*
374 : : * set_base_rel_pathlists
375 : : * Finds all paths available for scanning each base-relation entry.
376 : : * Sequential scan and any available indices are considered.
377 : : * Each useful path is attached to its relation's 'pathlist' field.
378 : : */
379 : : static void
380 : 33896 : set_base_rel_pathlists(PlannerInfo *root)
381 : : {
382 : 33896 : Index rti;
383 : :
384 [ + + ]: 104507 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
385 : : {
386 : 70611 : RelOptInfo *rel = root->simple_rel_array[rti];
387 : :
388 : : /* there may be empty slots corresponding to non-baserel RTEs */
389 [ + + ]: 70611 : if (rel == NULL)
390 : 16382 : continue;
391 : :
392 [ + - ]: 54229 : Assert(rel->relid == rti); /* sanity check on array */
393 : :
394 : : /* ignore RTEs that are "other rels" */
395 [ + + ]: 54229 : if (rel->reloptkind != RELOPT_BASEREL)
396 : 7800 : continue;
397 : :
398 : 46429 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
399 [ - + + ]: 70611 : }
400 : 33896 : }
401 : :
402 : : /*
403 : : * set_rel_size
404 : : * Set size estimates for a base relation
405 : : */
406 : : static void
407 : 54182 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
408 : : Index rti, RangeTblEntry *rte)
409 : : {
410 [ + + + + ]: 54182 : if (rel->reloptkind == RELOPT_BASEREL &&
411 : 46434 : relation_excluded_by_constraints(root, rel, rte))
412 : : {
413 : : /*
414 : : * We proved we don't need to scan the rel via constraint exclusion,
415 : : * so set up a single dummy path for it. Here we only check this for
416 : : * regular baserels; if it's an otherrel, CE was already checked in
417 : : * set_append_rel_size().
418 : : *
419 : : * In this case, we go ahead and set up the relation's path right away
420 : : * instead of leaving it for set_rel_pathlist to do. This is because
421 : : * we don't have a convention for marking a rel as dummy except by
422 : : * assigning a dummy path to it.
423 : : */
424 : 111 : set_dummy_rel_pathlist(rel);
425 : 111 : }
426 [ + + ]: 54071 : else if (rte->inh)
427 : : {
428 : : /* It's an "append relation", process accordingly */
429 : 3448 : set_append_rel_size(root, rel, rti, rte);
430 : 3448 : }
431 : : else
432 : : {
433 [ + + + + : 50623 : switch (rel->rtekind)
+ + + +
- ]
434 : : {
435 : : case RTE_RELATION:
436 [ - + ]: 43245 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
437 : : {
438 : : /* Foreign table */
439 : 0 : set_foreign_size(root, rel, rte);
440 : 0 : }
441 [ + + ]: 43245 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
442 : : {
443 : : /*
444 : : * We could get here if asked to scan a partitioned table
445 : : * with ONLY. In that case we shouldn't scan any of the
446 : : * partitions, so mark it as a dummy rel.
447 : : */
448 : 3 : set_dummy_rel_pathlist(rel);
449 : 3 : }
450 [ + + ]: 43242 : else if (rte->tablesample != NULL)
451 : : {
452 : : /* Sampled relation */
453 : 45 : set_tablesample_rel_size(root, rel, rte);
454 : 45 : }
455 : : else
456 : : {
457 : : /* Plain relation */
458 : 43197 : set_plain_rel_size(root, rel, rte);
459 : : }
460 : 43245 : break;
461 : : case RTE_SUBQUERY:
462 : :
463 : : /*
464 : : * Subqueries don't support making a choice between
465 : : * parameterized and unparameterized paths, so just go ahead
466 : : * and build their paths immediately.
467 : : */
468 : 1480 : set_subquery_pathlist(root, rel, rti, rte);
469 : 1480 : break;
470 : : case RTE_FUNCTION:
471 : 3642 : set_function_size_estimates(root, rel);
472 : 3642 : break;
473 : : case RTE_TABLEFUNC:
474 : 103 : set_tablefunc_size_estimates(root, rel);
475 : 103 : break;
476 : : case RTE_VALUES:
477 : 1114 : set_values_size_estimates(root, rel);
478 : 1114 : break;
479 : : case RTE_CTE:
480 : :
481 : : /*
482 : : * CTEs don't support making a choice between parameterized
483 : : * and unparameterized paths, so just go ahead and build their
484 : : * paths immediately.
485 : : */
486 [ + + ]: 286 : if (rte->self_reference)
487 : 74 : set_worktable_pathlist(root, rel, rte);
488 : : else
489 : 212 : set_cte_pathlist(root, rel, rte);
490 : 286 : break;
491 : : case RTE_NAMEDTUPLESTORE:
492 : : /* Might as well just build the path immediately */
493 : 77 : set_namedtuplestore_pathlist(root, rel, rte);
494 : 77 : break;
495 : : case RTE_RESULT:
496 : : /* Might as well just build the path immediately */
497 : 676 : set_result_pathlist(root, rel, rte);
498 : 676 : break;
499 : : default:
500 [ # # # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
501 : 0 : break;
502 : : }
503 : : }
504 : :
505 : : /*
506 : : * We insist that all non-dummy rels have a nonzero rowcount estimate.
507 : : */
508 [ + + - + ]: 54182 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
509 : 54182 : }
510 : :
511 : : /*
512 : : * set_rel_pathlist
513 : : * Build access paths for a base relation
514 : : */
515 : : static void
516 : 54200 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
517 : : Index rti, RangeTblEntry *rte)
518 : : {
519 [ + + ]: 54200 : if (IS_DUMMY_REL(rel))
520 : : {
521 : : /* We already proved the relation empty, so nothing more to do */
522 : 201 : }
523 [ + + ]: 53999 : else if (rte->inh)
524 : : {
525 : : /* It's an "append relation", process accordingly */
526 : 3405 : set_append_rel_pathlist(root, rel, rti, rte);
527 : 3405 : }
528 : : else
529 : : {
530 [ + + + + : 50594 : switch (rel->rtekind)
+ - ]
531 : : {
532 : : case RTE_RELATION:
533 [ - + ]: 43237 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
534 : : {
535 : : /* Foreign table */
536 : 0 : set_foreign_pathlist(root, rel, rte);
537 : 0 : }
538 [ + + ]: 43237 : else if (rte->tablesample != NULL)
539 : : {
540 : : /* Sampled relation */
541 : 45 : set_tablesample_rel_pathlist(root, rel, rte);
542 : 45 : }
543 : : else
544 : : {
545 : : /* Plain relation */
546 : 43192 : set_plain_rel_pathlist(root, rel, rte);
547 : : }
548 : 43237 : break;
549 : : case RTE_SUBQUERY:
550 : : /* Subquery --- fully handled during set_rel_size */
551 : : break;
552 : : case RTE_FUNCTION:
553 : : /* RangeFunction */
554 : 3642 : set_function_pathlist(root, rel, rte);
555 : 3642 : break;
556 : : case RTE_TABLEFUNC:
557 : : /* Table Function */
558 : 103 : set_tablefunc_pathlist(root, rel, rte);
559 : 103 : break;
560 : : case RTE_VALUES:
561 : : /* Values list */
562 : 1114 : set_values_pathlist(root, rel, rte);
563 : 1114 : break;
564 : : case RTE_CTE:
565 : : /* CTE reference --- fully handled during set_rel_size */
566 : : break;
567 : : case RTE_NAMEDTUPLESTORE:
568 : : /* tuplestore reference --- fully handled during set_rel_size */
569 : : break;
570 : : case RTE_RESULT:
571 : : /* simple Result --- fully handled during set_rel_size */
572 : : break;
573 : : default:
574 [ # # # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
575 : 0 : break;
576 : : }
577 : : }
578 : :
579 : : /*
580 : : * Allow a plugin to editorialize on the set of Paths for this base
581 : : * relation. It could add new paths (such as CustomPaths) by calling
582 : : * add_path(), or add_partial_path() if parallel aware. It could also
583 : : * delete or modify paths added by the core code.
584 : : */
585 [ + - ]: 54200 : if (set_rel_pathlist_hook)
586 : 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
587 : :
588 : : /*
589 : : * If this is a baserel, we should normally consider gathering any partial
590 : : * paths we may have created for it. We have to do this after calling the
591 : : * set_rel_pathlist_hook, else it cannot add partial paths to be included
592 : : * here.
593 : : *
594 : : * However, if this is an inheritance child, skip it. Otherwise, we could
595 : : * end up with a very large number of gather nodes, each trying to grab
596 : : * its own pool of workers. Instead, we'll consider gathering partial
597 : : * paths for the parent appendrel.
598 : : *
599 : : * Also, if this is the topmost scan/join rel, we postpone gathering until
600 : : * the final scan/join targetlist is available (see grouping_planner).
601 : : */
602 [ + + + + ]: 54200 : if (rel->reloptkind == RELOPT_BASEREL &&
603 : 46429 : !bms_equal(rel->relids, root->all_query_rels))
604 : 21882 : generate_useful_gather_paths(root, rel, false);
605 : :
606 : : /* Now find the cheapest of the paths for this rel */
607 : 54200 : set_cheapest(rel);
608 : :
609 : : /*
610 : : * If a grouped relation for this rel exists, build partial aggregation
611 : : * paths for it.
612 : : *
613 : : * Note that this can only happen after we've called set_cheapest() for
614 : : * this base rel, because we need its cheapest paths.
615 : : */
616 : 54200 : set_grouped_rel_pathlist(root, rel);
617 : :
618 : : #ifdef OPTIMIZER_DEBUG
619 : : pprint(rel);
620 : : #endif
621 : 54200 : }
622 : :
623 : : /*
624 : : * set_plain_rel_size
625 : : * Set size estimates for a plain relation (no subquery, no inheritance)
626 : : */
627 : : static void
628 : 43197 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
629 : : {
630 : : /*
631 : : * Test any partial indexes of rel for applicability. We must do this
632 : : * first since partial unique indexes can affect size estimates.
633 : : */
634 : 43197 : check_index_predicates(root, rel);
635 : :
636 : : /* Mark rel with estimated output rows, width, etc */
637 : 43197 : set_baserel_size_estimates(root, rel);
638 : 43197 : }
639 : :
640 : : /*
641 : : * If this relation could possibly be scanned from within a worker, then set
642 : : * its consider_parallel flag.
643 : : */
644 : : static void
645 : 44079 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
646 : : RangeTblEntry *rte)
647 : : {
648 : : /*
649 : : * The flag has previously been initialized to false, so we can just
650 : : * return if it becomes clear that we can't safely set it.
651 : : */
652 [ + - ]: 44079 : Assert(!rel->consider_parallel);
653 : :
654 : : /* Don't call this if parallelism is disallowed for the entire query. */
655 [ + - ]: 44079 : Assert(root->glob->parallelModeOK);
656 : :
657 : : /* This should only be called for baserels and appendrel children. */
658 [ + + + - ]: 44079 : Assert(IS_SIMPLE_REL(rel));
659 : :
660 : : /* Assorted checks based on rtekind. */
661 [ - + + + : 44079 : switch (rte->rtekind)
+ + + + +
- ]
662 : : {
663 : : case RTE_RELATION:
664 : :
665 : : /*
666 : : * Currently, parallel workers can't access the leader's temporary
667 : : * tables. We could possibly relax this if we wrote all of its
668 : : * local buffers at the start of the query and made no changes
669 : : * thereafter (maybe we could allow hint bit changes), and if we
670 : : * taught the workers to read them. Writing a large number of
671 : : * temporary buffers could be expensive, though, and we don't have
672 : : * the rest of the necessary infrastructure right now anyway. So
673 : : * for now, bail out if we see a temporary table.
674 : : */
675 [ + + ]: 38416 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
676 : 1365 : return;
677 : :
678 : : /*
679 : : * Table sampling can be pushed down to workers if the sample
680 : : * function and its arguments are safe.
681 : : */
682 [ + + ]: 37051 : if (rte->tablesample != NULL)
683 : : {
684 : 49 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
685 : :
686 [ - + ]: 49 : if (proparallel != PROPARALLEL_SAFE)
687 : 0 : return;
688 [ + + ]: 49 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
689 : 2 : return;
690 [ + + ]: 49 : }
691 : :
692 : : /*
693 : : * Ask FDWs whether they can support performing a ForeignScan
694 : : * within a worker. Most often, the answer will be no. For
695 : : * example, if the nature of the FDW is such that it opens a TCP
696 : : * connection with a remote server, each parallel worker would end
697 : : * up with a separate connection, and these connections might not
698 : : * be appropriately coordinated between workers and the leader.
699 : : */
700 [ + - ]: 37049 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
701 : : {
702 [ # # ]: 0 : Assert(rel->fdwroutine);
703 [ # # ]: 0 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
704 : 0 : return;
705 [ # # ]: 0 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
706 : 0 : return;
707 : 0 : }
708 : :
709 : : /*
710 : : * There are additional considerations for appendrels, which we'll
711 : : * deal with in set_append_rel_size and set_append_rel_pathlist.
712 : : * For now, just set consider_parallel based on the rel's own
713 : : * quals and targetlist.
714 : : */
715 : 37049 : break;
716 : :
717 : : case RTE_SUBQUERY:
718 : :
719 : : /*
720 : : * There's no intrinsic problem with scanning a subquery-in-FROM
721 : : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
722 : : * If the subquery doesn't happen to have any parallel-safe paths,
723 : : * then flagging it as consider_parallel won't change anything,
724 : : * but that's true for plain tables, too. We must set
725 : : * consider_parallel based on the rel's own quals and targetlist,
726 : : * so that if a subquery path is parallel-safe but the quals and
727 : : * projection we're sticking onto it are not, we correctly mark
728 : : * the SubqueryScanPath as not parallel-safe. (Note that
729 : : * set_subquery_pathlist() might push some of these quals down
730 : : * into the subquery itself, but that doesn't change anything.)
731 : : *
732 : : * We can't push sub-select containing LIMIT/OFFSET to workers as
733 : : * there is no guarantee that the row order will be fully
734 : : * deterministic, and applying LIMIT/OFFSET will lead to
735 : : * inconsistent results at the top-level. (In some cases, where
736 : : * the result is ordered, we could relax this restriction. But it
737 : : * doesn't currently seem worth expending extra effort to do so.)
738 : : */
739 : : {
740 : 1801 : Query *subquery = castNode(Query, rte->subquery);
741 : :
742 [ + + ]: 1801 : if (limit_needed(subquery))
743 : 84 : return;
744 [ + + ]: 1801 : }
745 : 1717 : break;
746 : :
747 : : case RTE_JOIN:
748 : : /* Shouldn't happen; we're only considering baserels here. */
749 : 0 : Assert(false);
750 : 0 : return;
751 : :
752 : : case RTE_FUNCTION:
753 : : /* Check for parallel-restricted functions. */
754 [ + + ]: 2406 : if (!is_parallel_safe(root, (Node *) rte->functions))
755 : 1097 : return;
756 : 1309 : break;
757 : :
758 : : case RTE_TABLEFUNC:
759 : : /* not parallel safe */
760 : 103 : return;
761 : :
762 : : case RTE_VALUES:
763 : : /* Check for parallel-restricted functions. */
764 [ + + ]: 459 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
765 : 2 : return;
766 : 457 : break;
767 : :
768 : : case RTE_CTE:
769 : :
770 : : /*
771 : : * CTE tuplestores aren't shared among parallel workers, so we
772 : : * force all CTE scans to happen in the leader. Also, populating
773 : : * the CTE would require executing a subplan that's not available
774 : : * in the worker, might be parallel-restricted, and must get
775 : : * executed only once.
776 : : */
777 : 202 : return;
778 : :
779 : : case RTE_NAMEDTUPLESTORE:
780 : :
781 : : /*
782 : : * tuplestore cannot be shared, at least without more
783 : : * infrastructure to support that.
784 : : */
785 : 73 : return;
786 : :
787 : : case RTE_RESULT:
788 : : /* RESULT RTEs, in themselves, are no problem. */
789 : : break;
790 : : case RTE_GROUP:
791 : : /* Shouldn't happen; we're only considering baserels here. */
792 : 0 : Assert(false);
793 : 0 : return;
794 : : }
795 : :
796 : : /*
797 : : * If there's anything in baserestrictinfo that's parallel-restricted, we
798 : : * give up on parallelizing access to this relation. We could consider
799 : : * instead postponing application of the restricted quals until we're
800 : : * above all the parallelism in the plan tree, but it's not clear that
801 : : * that would be a win in very many cases, and it might be tricky to make
802 : : * outer join clauses work correctly. It would likely break equivalence
803 : : * classes, too.
804 : : */
805 [ + + ]: 41151 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
806 : 3275 : return;
807 : :
808 : : /*
809 : : * Likewise, if the relation's outputs are not parallel-safe, give up.
810 : : * (Usually, they're just Vars, but sometimes they're not.)
811 : : */
812 [ + + ]: 37876 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
813 : 3 : return;
814 : :
815 : : /* We have a winner. */
816 : 37873 : rel->consider_parallel = true;
817 : 44079 : }
818 : :
819 : : /*
820 : : * set_plain_rel_pathlist
821 : : * Build access paths for a plain relation (no subquery, no inheritance)
822 : : */
823 : : static void
824 : 43192 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
825 : : {
826 : 43192 : Relids required_outer;
827 : :
828 : : /*
829 : : * We don't support pushing join clauses into the quals of a seqscan, but
830 : : * it could still have required parameterization due to LATERAL refs in
831 : : * its tlist.
832 : : */
833 : 43192 : required_outer = rel->lateral_relids;
834 : :
835 : : /*
836 : : * Consider TID scans.
837 : : *
838 : : * If create_tidscan_paths returns true, then a TID scan path is forced.
839 : : * This happens when rel->baserestrictinfo contains CurrentOfExpr, because
840 : : * the executor can't handle any other type of path for such queries.
841 : : * Hence, we return without adding any other paths.
842 : : */
843 [ + + ]: 43192 : if (create_tidscan_paths(root, rel))
844 : 66 : return;
845 : :
846 : : /* Consider sequential scan */
847 : 43126 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
848 : :
849 : : /* If appropriate, consider parallel sequential scan */
850 [ + + + + ]: 43126 : if (rel->consider_parallel && required_outer == NULL)
851 : 31908 : create_plain_partial_paths(root, rel);
852 : :
853 : : /* Consider index scans */
854 : 43126 : create_index_paths(root, rel);
855 [ - + ]: 43192 : }
856 : :
857 : : /*
858 : : * create_plain_partial_paths
859 : : * Build partial access paths for parallel scan of a plain relation
860 : : */
861 : : static void
862 : 31908 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
863 : : {
864 : 31908 : int parallel_workers;
865 : :
866 : 63816 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
867 : 31908 : max_parallel_workers_per_gather);
868 : :
869 : : /* If any limit was set to zero, the user doesn't want a parallel scan. */
870 [ + + ]: 31908 : if (parallel_workers <= 0)
871 : 27702 : return;
872 : :
873 : : /* Add an unordered partial path based on a parallel sequential scan. */
874 : 4206 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
875 [ - + ]: 31908 : }
876 : :
877 : : /*
878 : : * set_tablesample_rel_size
879 : : * Set size estimates for a sampled relation
880 : : */
881 : : static void
882 : 45 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
883 : : {
884 : 45 : TableSampleClause *tsc = rte->tablesample;
885 : 45 : TsmRoutine *tsm;
886 : 45 : BlockNumber pages;
887 : 45 : double tuples;
888 : :
889 : : /*
890 : : * Test any partial indexes of rel for applicability. We must do this
891 : : * first since partial unique indexes can affect size estimates.
892 : : */
893 : 45 : check_index_predicates(root, rel);
894 : :
895 : : /*
896 : : * Call the sampling method's estimation function to estimate the number
897 : : * of pages it will read and the number of tuples it will return. (Note:
898 : : * we assume the function returns sane values.)
899 : : */
900 : 45 : tsm = GetTsmRoutine(tsc->tsmhandler);
901 : 45 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
902 : : &pages, &tuples);
903 : :
904 : : /*
905 : : * For the moment, because we will only consider a SampleScan path for the
906 : : * rel, it's okay to just overwrite the pages and tuples estimates for the
907 : : * whole relation. If we ever consider multiple path types for sampled
908 : : * rels, we'll need more complication.
909 : : */
910 : 45 : rel->pages = pages;
911 : 45 : rel->tuples = tuples;
912 : :
913 : : /* Mark rel with estimated output rows, width, etc */
914 : 45 : set_baserel_size_estimates(root, rel);
915 : 45 : }
916 : :
917 : : /*
918 : : * set_tablesample_rel_pathlist
919 : : * Build access paths for a sampled relation
920 : : */
921 : : static void
922 : 45 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
923 : : {
924 : 45 : Relids required_outer;
925 : 45 : Path *path;
926 : :
927 : : /*
928 : : * We don't support pushing join clauses into the quals of a samplescan,
929 : : * but it could still have required parameterization due to LATERAL refs
930 : : * in its tlist or TABLESAMPLE arguments.
931 : : */
932 : 45 : required_outer = rel->lateral_relids;
933 : :
934 : : /* Consider sampled scan */
935 : 45 : path = create_samplescan_path(root, rel, required_outer);
936 : :
937 : : /*
938 : : * If the sampling method does not support repeatable scans, we must avoid
939 : : * plans that would scan the rel multiple times. Ideally, we'd simply
940 : : * avoid putting the rel on the inside of a nestloop join; but adding such
941 : : * a consideration to the planner seems like a great deal of complication
942 : : * to support an uncommon usage of second-rate sampling methods. Instead,
943 : : * if there is a risk that the query might perform an unsafe join, just
944 : : * wrap the SampleScan in a Materialize node. We can check for joins by
945 : : * counting the membership of all_query_rels (note that this correctly
946 : : * counts inheritance trees as single rels). If we're inside a subquery,
947 : : * we can't easily check whether a join might occur in the outer query, so
948 : : * just assume one is possible.
949 : : *
950 : : * GetTsmRoutine is relatively expensive compared to the other tests here,
951 : : * so check repeatable_across_scans last, even though that's a bit odd.
952 : : */
953 [ + + ]: 45 : if ((root->query_level > 1 ||
954 [ + - ]: 45 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
955 : 45 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
956 : : {
957 : 0 : path = (Path *) create_material_path(rel, path, true);
958 : 0 : }
959 : :
960 : 45 : add_path(rel, path);
961 : :
962 : : /* For the moment, at least, there are no other paths to consider */
963 : 45 : }
964 : :
965 : : /*
966 : : * set_foreign_size
967 : : * Set size estimates for a foreign table RTE
968 : : */
969 : : static void
970 : 0 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
971 : : {
972 : : /* Mark rel with estimated output rows, width, etc */
973 : 0 : set_foreign_size_estimates(root, rel);
974 : :
975 : : /* Let FDW adjust the size estimates, if it can */
976 : 0 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
977 : :
978 : : /* ... but do not let it set the rows estimate to zero */
979 : 0 : rel->rows = clamp_row_est(rel->rows);
980 : :
981 : : /*
982 : : * Also, make sure rel->tuples is not insane relative to rel->rows.
983 : : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
984 : : * FDW doesn't do anything to replace that.
985 : : */
986 [ # # ]: 0 : rel->tuples = Max(rel->tuples, rel->rows);
987 : 0 : }
988 : :
989 : : /*
990 : : * set_foreign_pathlist
991 : : * Build access paths for a foreign table RTE
992 : : */
993 : : static void
994 : 0 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
995 : : {
996 : : /* Call the FDW's GetForeignPaths function to generate path(s) */
997 : 0 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
998 : 0 : }
999 : :
1000 : : /*
1001 : : * set_append_rel_size
1002 : : * Set size estimates for a simple "append relation"
1003 : : *
1004 : : * The passed-in rel and RTE represent the entire append relation. The
1005 : : * relation's contents are computed by appending together the output of the
1006 : : * individual member relations. Note that in the non-partitioned inheritance
1007 : : * case, the first member relation is actually the same table as is mentioned
1008 : : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
1009 : : * a good thing because their outputs are not the same size.
1010 : : */
1011 : : static void
1012 : 3448 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
1013 : : Index rti, RangeTblEntry *rte)
1014 : : {
1015 : 3448 : int parentRTindex = rti;
1016 : 3448 : bool has_live_children;
1017 : 3448 : double parent_tuples;
1018 : 3448 : double parent_rows;
1019 : 3448 : double parent_size;
1020 : 3448 : double *parent_attrsizes;
1021 : 3448 : int nattrs;
1022 : 3448 : ListCell *l;
1023 : :
1024 : : /* Guard against stack overflow due to overly deep inheritance tree. */
1025 : 3448 : check_stack_depth();
1026 : :
1027 [ + + + - ]: 3448 : Assert(IS_SIMPLE_REL(rel));
1028 : :
1029 : : /*
1030 : : * If this is a partitioned baserel, set the consider_partitionwise_join
1031 : : * flag; currently, we only consider partitionwise joins with the baserel
1032 : : * if its targetlist doesn't contain a whole-row Var.
1033 : : */
1034 [ + + ]: 3448 : if (enable_partitionwise_join &&
1035 [ + + ]: 841 : rel->reloptkind == RELOPT_BASEREL &&
1036 [ + - + + ]: 671 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
1037 : 671 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
1038 : 661 : rel->consider_partitionwise_join = true;
1039 : :
1040 : : /*
1041 : : * Initialize to compute size estimates for whole append relation.
1042 : : *
1043 : : * We handle tuples estimates by setting "tuples" to the total number of
1044 : : * tuples accumulated from each live child, rather than using "rows".
1045 : : * Although an appendrel itself doesn't directly enforce any quals, its
1046 : : * child relations may. Therefore, setting "tuples" equal to "rows" for
1047 : : * an appendrel isn't always appropriate, and can lead to inaccurate cost
1048 : : * estimates. For example, when estimating the number of distinct values
1049 : : * from an appendrel, we would be unable to adjust the estimate based on
1050 : : * the restriction selectivity (see estimate_num_groups).
1051 : : *
1052 : : * We handle width estimates by weighting the widths of different child
1053 : : * rels proportionally to their number of rows. This is sensible because
1054 : : * the use of width estimates is mainly to compute the total relation
1055 : : * "footprint" if we have to sort or hash it. To do this, we sum the
1056 : : * total equivalent size (in "double" arithmetic) and then divide by the
1057 : : * total rowcount estimate. This is done separately for the total rel
1058 : : * width and each attribute.
1059 : : *
1060 : : * Note: if you consider changing this logic, beware that child rels could
1061 : : * have zero rows and/or width, if they were excluded by constraints.
1062 : : */
1063 : 3448 : has_live_children = false;
1064 : 3448 : parent_tuples = 0;
1065 : 3448 : parent_rows = 0;
1066 : 3448 : parent_size = 0;
1067 : 3448 : nattrs = rel->max_attr - rel->min_attr + 1;
1068 : 3448 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1069 : :
1070 [ + + + + : 19893 : foreach(l, root->append_rel_list)
+ + ]
1071 : : {
1072 : 16445 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1073 : 16445 : int childRTindex;
1074 : 16445 : RangeTblEntry *childRTE;
1075 : 16445 : RelOptInfo *childrel;
1076 : 16445 : List *childrinfos;
1077 : 16445 : ListCell *parentvars;
1078 : 16445 : ListCell *childvars;
1079 : 16445 : ListCell *lc;
1080 : :
1081 : : /* append_rel_list contains all append rels; ignore others */
1082 [ + + ]: 16445 : if (appinfo->parent_relid != parentRTindex)
1083 : 8663 : continue;
1084 : :
1085 : 7782 : childRTindex = appinfo->child_relid;
1086 : 7782 : childRTE = root->simple_rte_array[childRTindex];
1087 : :
1088 : : /*
1089 : : * The child rel's RelOptInfo was already created during
1090 : : * add_other_rels_to_query.
1091 : : */
1092 : 7782 : childrel = find_base_rel(root, childRTindex);
1093 [ - + ]: 7782 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1094 : :
1095 : : /* We may have already proven the child to be dummy. */
1096 [ + + ]: 7782 : if (IS_DUMMY_REL(childrel))
1097 : 3 : continue;
1098 : :
1099 : : /*
1100 : : * We have to copy the parent's targetlist and quals to the child,
1101 : : * with appropriate substitution of variables. However, the
1102 : : * baserestrictinfo quals were already copied/substituted when the
1103 : : * child RelOptInfo was built. So we don't need any additional setup
1104 : : * before applying constraint exclusion.
1105 : : */
1106 [ + + ]: 7779 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1107 : : {
1108 : : /*
1109 : : * This child need not be scanned, so we can omit it from the
1110 : : * appendrel.
1111 : : */
1112 : 31 : set_dummy_rel_pathlist(childrel);
1113 : 31 : continue;
1114 : : }
1115 : :
1116 : : /*
1117 : : * Constraint exclusion failed, so copy the parent's join quals and
1118 : : * targetlist to the child, with appropriate variable substitutions.
1119 : : *
1120 : : * We skip join quals that came from above outer joins that can null
1121 : : * this rel, since they would be of no value while generating paths
1122 : : * for the child. This saves some effort while processing the child
1123 : : * rel, and it also avoids an implementation restriction in
1124 : : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1125 : : */
1126 : 7748 : childrinfos = NIL;
1127 [ + + + + : 9900 : foreach(lc, rel->joininfo)
+ + ]
1128 : : {
1129 : 2152 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1130 : :
1131 [ + + ]: 2152 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1132 : 3530 : childrinfos = lappend(childrinfos,
1133 : 3530 : adjust_appendrel_attrs(root,
1134 : 1765 : (Node *) rinfo,
1135 : : 1, &appinfo));
1136 : 2152 : }
1137 : 7748 : childrel->joininfo = childrinfos;
1138 : :
1139 : : /*
1140 : : * Now for the child's targetlist.
1141 : : *
1142 : : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1143 : : * expressions, which otherwise would not occur in a rel's targetlist.
1144 : : * Code that might be looking at an appendrel child must cope with
1145 : : * such. (Normally, a rel's targetlist would only include Vars and
1146 : : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1147 : : * fields of childrel->reltarget; not clear if that would be useful.
1148 : : */
1149 : 7748 : childrel->reltarget->exprs = (List *)
1150 : 15496 : adjust_appendrel_attrs(root,
1151 : 7748 : (Node *) rel->reltarget->exprs,
1152 : : 1, &appinfo);
1153 : :
1154 : : /*
1155 : : * We have to make child entries in the EquivalenceClass data
1156 : : * structures as well. This is needed either if the parent
1157 : : * participates in some eclass joins (because we will want to consider
1158 : : * inner-indexscan joins on the individual children) or if the parent
1159 : : * has useful pathkeys (because we should try to build MergeAppend
1160 : : * paths that produce those sort orderings).
1161 : : */
1162 [ + + + + ]: 7748 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1163 : 4921 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1164 : 7748 : childrel->has_eclass_joins = rel->has_eclass_joins;
1165 : :
1166 : : /*
1167 : : * Note: we could compute appropriate attr_needed data for the child's
1168 : : * variables, by transforming the parent's attr_needed through the
1169 : : * translated_vars mapping. However, currently there's no need
1170 : : * because attr_needed is only examined for base relations not
1171 : : * otherrels. So we just leave the child's attr_needed empty.
1172 : : */
1173 : :
1174 : : /*
1175 : : * If we consider partitionwise joins with the parent rel, do the same
1176 : : * for partitioned child rels.
1177 : : *
1178 : : * Note: here we abuse the consider_partitionwise_join flag by setting
1179 : : * it for child rels that are not themselves partitioned. We do so to
1180 : : * tell try_partitionwise_join() that the child rel is sufficiently
1181 : : * valid to be used as a per-partition input, even if it later gets
1182 : : * proven to be dummy. (It's not usable until we've set up the
1183 : : * reltarget and EC entries, which we just did.)
1184 : : */
1185 [ + + ]: 7748 : if (rel->consider_partitionwise_join)
1186 : 2237 : childrel->consider_partitionwise_join = true;
1187 : :
1188 : : /*
1189 : : * If parallelism is allowable for this query in general, see whether
1190 : : * it's allowable for this childrel in particular. But if we've
1191 : : * already decided the appendrel is not parallel-safe as a whole,
1192 : : * there's no point in considering parallelism for this child. For
1193 : : * consistency, do this before calling set_rel_size() for the child.
1194 : : */
1195 [ + + + + ]: 7748 : if (root->glob->parallelModeOK && rel->consider_parallel)
1196 : 5669 : set_rel_consider_parallel(root, childrel, childRTE);
1197 : :
1198 : : /*
1199 : : * Compute the child's size.
1200 : : */
1201 : 7748 : set_rel_size(root, childrel, childRTindex, childRTE);
1202 : :
1203 : : /*
1204 : : * It is possible that constraint exclusion detected a contradiction
1205 : : * within a child subquery, even though we didn't prove one above. If
1206 : : * so, we can skip this child.
1207 : : */
1208 [ + + ]: 7748 : if (IS_DUMMY_REL(childrel))
1209 : 23 : continue;
1210 : :
1211 : : /* We have at least one live child. */
1212 : 7725 : has_live_children = true;
1213 : :
1214 : : /*
1215 : : * If any live child is not parallel-safe, treat the whole appendrel
1216 : : * as not parallel-safe. In future we might be able to generate plans
1217 : : * in which some children are farmed out to workers while others are
1218 : : * not; but we don't have that today, so it's a waste to consider
1219 : : * partial paths anywhere in the appendrel unless it's all safe.
1220 : : * (Child rels visited before this one will be unmarked in
1221 : : * set_append_rel_pathlist().)
1222 : : */
1223 [ + + ]: 7725 : if (!childrel->consider_parallel)
1224 : 2095 : rel->consider_parallel = false;
1225 : :
1226 : : /*
1227 : : * Accumulate size information from each live child.
1228 : : */
1229 [ - + ]: 7725 : Assert(childrel->rows > 0);
1230 : :
1231 : 7725 : parent_tuples += childrel->tuples;
1232 : 7725 : parent_rows += childrel->rows;
1233 : 7725 : parent_size += childrel->reltarget->width * childrel->rows;
1234 : :
1235 : : /*
1236 : : * Accumulate per-column estimates too. We need not do anything for
1237 : : * PlaceHolderVars in the parent list. If child expression isn't a
1238 : : * Var, or we didn't record a width estimate for it, we have to fall
1239 : : * back on a datatype-based estimate.
1240 : : *
1241 : : * By construction, child's targetlist is 1-to-1 with parent's.
1242 : : */
1243 [ + + + + : 24024 : forboth(parentvars, rel->reltarget->exprs,
+ + + + +
+ + + ]
1244 : : childvars, childrel->reltarget->exprs)
1245 : : {
1246 : 16299 : Var *parentvar = (Var *) lfirst(parentvars);
1247 : 16299 : Node *childvar = (Node *) lfirst(childvars);
1248 : :
1249 [ + + + + ]: 16299 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1250 : : {
1251 : 14495 : int pndx = parentvar->varattno - rel->min_attr;
1252 : 14495 : int32 child_width = 0;
1253 : :
1254 [ + + + + ]: 14495 : if (IsA(childvar, Var) &&
1255 : 13715 : ((Var *) childvar)->varno == childrel->relid)
1256 : : {
1257 : 13704 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1258 : :
1259 : 13704 : child_width = childrel->attr_widths[cndx];
1260 : 13704 : }
1261 [ + + ]: 14495 : if (child_width <= 0)
1262 : 1582 : child_width = get_typavgwidth(exprType(childvar),
1263 : 791 : exprTypmod(childvar));
1264 [ - + ]: 14495 : Assert(child_width > 0);
1265 : 14495 : parent_attrsizes[pndx] += child_width * childrel->rows;
1266 : 14495 : }
1267 : 16299 : }
1268 [ - + + ]: 16445 : }
1269 : :
1270 [ + + ]: 3448 : if (has_live_children)
1271 : : {
1272 : : /*
1273 : : * Save the finished size estimates.
1274 : : */
1275 : 3405 : int i;
1276 : :
1277 [ + - ]: 3405 : Assert(parent_rows > 0);
1278 : 3405 : rel->tuples = parent_tuples;
1279 : 3405 : rel->rows = parent_rows;
1280 : 3405 : rel->reltarget->width = rint(parent_size / parent_rows);
1281 [ + + ]: 31129 : for (i = 0; i < nattrs; i++)
1282 : 27724 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1283 : :
1284 : : /*
1285 : : * Note that we leave rel->pages as zero; this is important to avoid
1286 : : * double-counting the appendrel tree in total_table_pages.
1287 : : */
1288 : 3405 : }
1289 : : else
1290 : : {
1291 : : /*
1292 : : * All children were excluded by constraints, so mark the whole
1293 : : * appendrel dummy. We must do this in this phase so that the rel's
1294 : : * dummy-ness is visible when we generate paths for other rels.
1295 : : */
1296 : 43 : set_dummy_rel_pathlist(rel);
1297 : : }
1298 : :
1299 : 3448 : pfree(parent_attrsizes);
1300 : 3448 : }
1301 : :
1302 : : /*
1303 : : * set_append_rel_pathlist
1304 : : * Build access paths for an "append relation"
1305 : : */
1306 : : static void
1307 : 3405 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1308 : : Index rti, RangeTblEntry *rte)
1309 : : {
1310 : 3405 : int parentRTindex = rti;
1311 : 3405 : List *live_childrels = NIL;
1312 : 3405 : ListCell *l;
1313 : :
1314 : : /*
1315 : : * Generate access paths for each member relation, and remember the
1316 : : * non-dummy children.
1317 : : */
1318 [ + - + + : 19781 : foreach(l, root->append_rel_list)
+ + ]
1319 : : {
1320 : 16376 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1321 : 16376 : int childRTindex;
1322 : 16376 : RangeTblEntry *childRTE;
1323 : 16376 : RelOptInfo *childrel;
1324 : :
1325 : : /* append_rel_list contains all append rels; ignore others */
1326 [ + + ]: 16376 : if (appinfo->parent_relid != parentRTindex)
1327 : 8605 : continue;
1328 : :
1329 : : /* Re-locate the child RTE and RelOptInfo */
1330 : 7771 : childRTindex = appinfo->child_relid;
1331 : 7771 : childRTE = root->simple_rte_array[childRTindex];
1332 : 7771 : childrel = root->simple_rel_array[childRTindex];
1333 : :
1334 : : /*
1335 : : * If set_append_rel_size() decided the parent appendrel was
1336 : : * parallel-unsafe at some point after visiting this child rel, we
1337 : : * need to propagate the unsafety marking down to the child, so that
1338 : : * we don't generate useless partial paths for it.
1339 : : */
1340 [ + + ]: 7771 : if (!rel->consider_parallel)
1341 : 2118 : childrel->consider_parallel = false;
1342 : :
1343 : : /*
1344 : : * Compute the child's access paths.
1345 : : */
1346 : 7771 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1347 : :
1348 : : /*
1349 : : * If child is dummy, ignore it.
1350 : : */
1351 [ + + ]: 7771 : if (IS_DUMMY_REL(childrel))
1352 : 46 : continue;
1353 : :
1354 : : /*
1355 : : * Child is live, so add it to the live_childrels list for use below.
1356 : : */
1357 : 7725 : live_childrels = lappend(live_childrels, childrel);
1358 [ - + + ]: 16376 : }
1359 : :
1360 : : /* Add paths to the append relation. */
1361 : 3405 : add_paths_to_append_rel(root, rel, live_childrels);
1362 : 3405 : }
1363 : :
1364 : : /*
1365 : : * set_grouped_rel_pathlist
1366 : : * If a grouped relation for the given 'rel' exists, build partial
1367 : : * aggregation paths for it.
1368 : : */
1369 : : static void
1370 : 54200 : set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel)
1371 : : {
1372 : 54200 : RelOptInfo *grouped_rel;
1373 : :
1374 : : /*
1375 : : * If there are no aggregate expressions or grouping expressions, eager
1376 : : * aggregation is not possible.
1377 : : */
1378 [ + + + + ]: 54200 : if (root->agg_clause_list == NIL ||
1379 : 534 : root->group_expr_list == NIL)
1380 : 53702 : return;
1381 : :
1382 : : /* Add paths to the grouped base relation if one exists. */
1383 : 498 : grouped_rel = rel->grouped_rel;
1384 [ + + ]: 498 : if (grouped_rel)
1385 : : {
1386 [ + - ]: 97 : Assert(IS_GROUPED_REL(grouped_rel));
1387 : :
1388 : 97 : generate_grouped_paths(root, grouped_rel, rel);
1389 : 97 : set_cheapest(grouped_rel);
1390 : 97 : }
1391 [ - + ]: 54200 : }
1392 : :
1393 : :
1394 : : /*
1395 : : * add_paths_to_append_rel
1396 : : * Generate paths for the given append relation given the set of non-dummy
1397 : : * child rels.
1398 : : *
1399 : : * The function collects all parameterizations and orderings supported by the
1400 : : * non-dummy children. For every such parameterization or ordering, it creates
1401 : : * an append path collecting one path from each non-dummy child with given
1402 : : * parameterization or ordering. Similarly it collects partial paths from
1403 : : * non-dummy children to create partial append paths.
1404 : : */
1405 : : void
1406 : 6378 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1407 : : List *live_childrels)
1408 : : {
1409 : 6378 : AppendPathInput unparameterized = {0};
1410 : 6378 : AppendPathInput startup = {0};
1411 : 6378 : AppendPathInput partial_only = {0};
1412 : 6378 : AppendPathInput parallel_append = {0};
1413 : 6378 : bool unparameterized_valid = true;
1414 : 6378 : bool startup_valid = true;
1415 : 6378 : bool partial_only_valid = true;
1416 : 6378 : bool parallel_append_valid = true;
1417 : 6378 : List *all_child_pathkeys = NIL;
1418 : 6378 : List *all_child_outers = NIL;
1419 : 6378 : ListCell *l;
1420 : 6378 : double partial_rows = -1;
1421 : :
1422 : : /* If appropriate, consider parallel append */
1423 [ + + ]: 6378 : parallel_append_valid = enable_parallel_append && rel->consider_parallel;
1424 : :
1425 : : /*
1426 : : * For every non-dummy child, remember the cheapest path. Also, identify
1427 : : * all pathkeys (orderings) and parameterizations (required_outer sets)
1428 : : * available for the non-dummy member relations.
1429 : : */
1430 [ + - + + : 21043 : foreach(l, live_childrels)
+ + ]
1431 : : {
1432 : 14665 : RelOptInfo *childrel = lfirst(l);
1433 : 14665 : ListCell *lcp;
1434 : 14665 : Path *cheapest_partial_path = NULL;
1435 : :
1436 : : /*
1437 : : * If child has an unparameterized cheapest-total path, add that to
1438 : : * the unparameterized Append path we are constructing for the parent.
1439 : : * If not, there's no workable unparameterized path.
1440 : : *
1441 : : * With partitionwise aggregates, the child rel's pathlist may be
1442 : : * empty, so don't assume that a path exists here.
1443 : : */
1444 [ + - + + ]: 14665 : if (childrel->pathlist != NIL &&
1445 : 14665 : childrel->cheapest_total_path->param_info == NULL)
1446 : 29086 : accumulate_append_subpath(childrel->cheapest_total_path,
1447 : 14543 : &unparameterized.subpaths, NULL, &unparameterized.child_append_relid_sets);
1448 : : else
1449 : 122 : unparameterized_valid = false;
1450 : :
1451 : : /*
1452 : : * When the planner is considering cheap startup plans, we'll also
1453 : : * collect all the cheapest_startup_paths (if set) and build an
1454 : : * AppendPath containing those as subpaths.
1455 : : */
1456 [ + + + + ]: 14665 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1457 : : {
1458 : 265 : Path *cheapest_path;
1459 : :
1460 : : /*
1461 : : * With an indication of how many tuples the query should provide,
1462 : : * the optimizer tries to choose the path optimal for that
1463 : : * specific number of tuples.
1464 : : */
1465 [ + - ]: 265 : if (root->tuple_fraction > 0.0)
1466 : 265 : cheapest_path =
1467 : 530 : get_cheapest_fractional_path(childrel,
1468 : 265 : root->tuple_fraction);
1469 : : else
1470 : 0 : cheapest_path = childrel->cheapest_startup_path;
1471 : :
1472 : : /* cheapest_startup_path must not be a parameterized path. */
1473 [ + - ]: 265 : Assert(cheapest_path->param_info == NULL);
1474 : 530 : accumulate_append_subpath(cheapest_path,
1475 : 265 : &startup.subpaths,
1476 : : NULL,
1477 : 265 : &startup.child_append_relid_sets);
1478 : 265 : }
1479 : : else
1480 : 14400 : startup_valid = false;
1481 : :
1482 : :
1483 : : /* Same idea, but for a partial plan. */
1484 [ + + ]: 14665 : if (childrel->partial_pathlist != NIL)
1485 : : {
1486 : 9774 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1487 : 19548 : accumulate_append_subpath(cheapest_partial_path,
1488 : 9774 : &partial_only.partial_subpaths, NULL,
1489 : 9774 : &partial_only.child_append_relid_sets);
1490 : 9774 : }
1491 : : else
1492 : 4891 : partial_only_valid = false;
1493 : :
1494 : : /*
1495 : : * Same idea, but for a parallel append mixing partial and non-partial
1496 : : * paths.
1497 : : */
1498 [ + + ]: 14665 : if (parallel_append_valid)
1499 : : {
1500 : 11139 : Path *nppath = NULL;
1501 : :
1502 : 11139 : nppath =
1503 : 11139 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1504 : :
1505 [ + + + + ]: 11139 : if (cheapest_partial_path == NULL && nppath == NULL)
1506 : : {
1507 : : /* Neither a partial nor a parallel-safe path? Forget it. */
1508 : 61 : parallel_append_valid = false;
1509 : 61 : }
1510 [ + + + + ]: 20777 : else if (nppath == NULL ||
1511 [ + + ]: 11040 : (cheapest_partial_path != NULL &&
1512 : 9699 : cheapest_partial_path->total_cost < nppath->total_cost))
1513 : : {
1514 : : /* Partial path is cheaper or the only option. */
1515 [ + - ]: 9677 : Assert(cheapest_partial_path != NULL);
1516 : 19354 : accumulate_append_subpath(cheapest_partial_path,
1517 : 9677 : ¶llel_append.partial_subpaths,
1518 : 9677 : ¶llel_append.subpaths,
1519 : 9677 : ¶llel_append.child_append_relid_sets);
1520 : 9677 : }
1521 : : else
1522 : : {
1523 : : /*
1524 : : * Either we've got only a non-partial path, or we think that
1525 : : * a single backend can execute the best non-partial path
1526 : : * faster than all the parallel backends working together can
1527 : : * execute the best partial path.
1528 : : *
1529 : : * It might make sense to be more aggressive here. Even if
1530 : : * the best non-partial path is more expensive than the best
1531 : : * partial path, it could still be better to choose the
1532 : : * non-partial path if there are several such paths that can
1533 : : * be given to different workers. For now, we don't try to
1534 : : * figure that out.
1535 : : */
1536 : 2802 : accumulate_append_subpath(nppath,
1537 : 1401 : ¶llel_append.subpaths,
1538 : : NULL,
1539 : 1401 : ¶llel_append.child_append_relid_sets);
1540 : : }
1541 : 11139 : }
1542 : :
1543 : : /*
1544 : : * Collect lists of all the available path orderings and
1545 : : * parameterizations for all the children. We use these as a
1546 : : * heuristic to indicate which sort orderings and parameterizations we
1547 : : * should build Append and MergeAppend paths for.
1548 : : */
1549 [ + - + + : 35616 : foreach(lcp, childrel->pathlist)
+ + ]
1550 : : {
1551 : 20951 : Path *childpath = (Path *) lfirst(lcp);
1552 : 20951 : List *childkeys = childpath->pathkeys;
1553 [ + + ]: 20951 : Relids childouter = PATH_REQ_OUTER(childpath);
1554 : :
1555 : : /* Unsorted paths don't contribute to pathkey list */
1556 [ + + ]: 20951 : if (childkeys != NIL)
1557 : : {
1558 : 6222 : ListCell *lpk;
1559 : 6222 : bool found = false;
1560 : :
1561 : : /* Have we already seen this ordering? */
1562 [ + + + + : 10474 : foreach(lpk, all_child_pathkeys)
+ + ]
1563 : : {
1564 : 4252 : List *existing_pathkeys = (List *) lfirst(lpk);
1565 : :
1566 : 8504 : if (compare_pathkeys(existing_pathkeys,
1567 [ + + + + ]: 8504 : childkeys) == PATHKEYS_EQUAL)
1568 : : {
1569 : 4219 : found = true;
1570 : 4219 : break;
1571 : : }
1572 [ + + ]: 4252 : }
1573 [ + + ]: 6222 : if (!found)
1574 : : {
1575 : : /* No, so add it to all_child_pathkeys */
1576 : 4006 : all_child_pathkeys = lappend(all_child_pathkeys,
1577 : 2003 : childkeys);
1578 : 2003 : }
1579 : 6222 : }
1580 : :
1581 : : /* Unparameterized paths don't contribute to param-set list */
1582 [ + + ]: 20951 : if (childouter)
1583 : : {
1584 : 1174 : ListCell *lco;
1585 : 1174 : bool found = false;
1586 : :
1587 : : /* Have we already seen this param set? */
1588 [ + + + + : 2083 : foreach(lco, all_child_outers)
+ + ]
1589 : : {
1590 : 909 : Relids existing_outers = (Relids) lfirst(lco);
1591 : :
1592 [ + + ]: 909 : if (bms_equal(existing_outers, childouter))
1593 : : {
1594 : 737 : found = true;
1595 : 737 : break;
1596 : : }
1597 [ + + ]: 909 : }
1598 [ + + ]: 1174 : if (!found)
1599 : : {
1600 : : /* No, so add it to all_child_outers */
1601 : 874 : all_child_outers = lappend(all_child_outers,
1602 : 437 : childouter);
1603 : 437 : }
1604 : 1174 : }
1605 : 20951 : }
1606 : 14665 : }
1607 : :
1608 : : /*
1609 : : * If we found unparameterized paths for all children, build an unordered,
1610 : : * unparameterized Append path for the rel. (Note: this is correct even
1611 : : * if we have zero or one live subpath due to constraint exclusion.)
1612 : : */
1613 [ + + ]: 6378 : if (unparameterized_valid)
1614 : 6326 : add_path(rel, (Path *) create_append_path(root, rel, unparameterized,
1615 : : NIL, NULL, 0, false,
1616 : : -1));
1617 : :
1618 : : /* build an AppendPath for the cheap startup paths, if valid */
1619 [ + + ]: 6378 : if (startup_valid)
1620 : 105 : add_path(rel, (Path *) create_append_path(root, rel, startup,
1621 : : NIL, NULL, 0, false, -1));
1622 : :
1623 : : /*
1624 : : * Consider an append of unordered, unparameterized partial paths. Make
1625 : : * it parallel-aware if possible.
1626 : : */
1627 [ + + - + ]: 6378 : if (partial_only_valid && partial_only.partial_subpaths != NIL)
1628 : : {
1629 : 3797 : AppendPath *appendpath;
1630 : 3797 : ListCell *lc;
1631 : 3797 : int parallel_workers = 0;
1632 : :
1633 : : /* Find the highest number of workers requested for any subpath. */
1634 [ + - + + : 14311 : foreach(lc, partial_only.partial_subpaths)
+ + ]
1635 : : {
1636 : 10514 : Path *path = lfirst(lc);
1637 : :
1638 [ + + ]: 10514 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1639 : 10514 : }
1640 [ + - ]: 3797 : Assert(parallel_workers > 0);
1641 : :
1642 : : /*
1643 : : * If the use of parallel append is permitted, always request at least
1644 : : * log2(# of children) workers. We assume it can be useful to have
1645 : : * extra workers in this case because they will be spread out across
1646 : : * the children. The precise formula is just a guess, but we don't
1647 : : * want to end up with a radically different answer for a table with N
1648 : : * partitions vs. an unpartitioned table with the same data, so the
1649 : : * use of some kind of log-scaling here seems to make some sense.
1650 : : */
1651 [ + + ]: 3797 : if (enable_parallel_append)
1652 : : {
1653 [ + + ]: 3789 : parallel_workers = Max(parallel_workers,
1654 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1655 [ + + ]: 3789 : parallel_workers = Min(parallel_workers,
1656 : : max_parallel_workers_per_gather);
1657 : 3789 : }
1658 [ + - ]: 3797 : Assert(parallel_workers > 0);
1659 : :
1660 : : /* Generate a partial append path. */
1661 : 7594 : appendpath = create_append_path(root, rel, partial_only,
1662 : 3797 : NIL, NULL, parallel_workers,
1663 : 3797 : enable_parallel_append,
1664 : : -1);
1665 : :
1666 : : /*
1667 : : * Make sure any subsequent partial paths use the same row count
1668 : : * estimate.
1669 : : */
1670 : 3797 : partial_rows = appendpath->path.rows;
1671 : :
1672 : : /* Add the path. */
1673 : 3797 : add_partial_path(rel, (Path *) appendpath);
1674 : 3797 : }
1675 : :
1676 : : /*
1677 : : * Consider a parallel-aware append using a mix of partial and non-partial
1678 : : * paths. (This only makes sense if there's at least one child which has
1679 : : * a non-partial path that is substantially cheaper than any partial path;
1680 : : * otherwise, we should use the append path added in the previous step.)
1681 : : */
1682 [ + + + + ]: 6378 : if (parallel_append_valid && parallel_append.subpaths != NIL)
1683 : : {
1684 : 712 : AppendPath *appendpath;
1685 : 712 : ListCell *lc;
1686 : 712 : int parallel_workers = 0;
1687 : :
1688 : : /*
1689 : : * Find the highest number of workers requested for any partial
1690 : : * subpath.
1691 : : */
1692 [ + + + + : 858 : foreach(lc, parallel_append.partial_subpaths)
+ + ]
1693 : : {
1694 : 146 : Path *path = lfirst(lc);
1695 : :
1696 [ - + ]: 146 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1697 : 146 : }
1698 : :
1699 : : /*
1700 : : * Same formula here as above. It's even more important in this
1701 : : * instance because the non-partial paths won't contribute anything to
1702 : : * the planned number of parallel workers.
1703 : : */
1704 [ - + ]: 712 : parallel_workers = Max(parallel_workers,
1705 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1706 [ + + ]: 712 : parallel_workers = Min(parallel_workers,
1707 : : max_parallel_workers_per_gather);
1708 [ + - ]: 712 : Assert(parallel_workers > 0);
1709 : :
1710 : 1424 : appendpath = create_append_path(root, rel, parallel_append,
1711 : 712 : NIL, NULL, parallel_workers, true,
1712 : 712 : partial_rows);
1713 : 712 : add_partial_path(rel, (Path *) appendpath);
1714 : 712 : }
1715 : :
1716 : : /*
1717 : : * Also build unparameterized ordered append paths based on the collected
1718 : : * list of child pathkeys.
1719 : : */
1720 [ + + ]: 6378 : if (unparameterized_valid)
1721 : 12652 : generate_orderedappend_paths(root, rel, live_childrels,
1722 : 6326 : all_child_pathkeys);
1723 : :
1724 : : /*
1725 : : * Build Append paths for each parameterization seen among the child rels.
1726 : : * (This may look pretty expensive, but in most cases of practical
1727 : : * interest, the child rels will expose mostly the same parameterizations,
1728 : : * so that not that many cases actually get considered here.)
1729 : : *
1730 : : * The Append node itself cannot enforce quals, so all qual checking must
1731 : : * be done in the child paths. This means that to have a parameterized
1732 : : * Append path, we must have the exact same parameterization for each
1733 : : * child path; otherwise some children might be failing to check the
1734 : : * moved-down quals. To make them match up, we can try to increase the
1735 : : * parameterization of lesser-parameterized paths.
1736 : : */
1737 [ + + + + : 6815 : foreach(l, all_child_outers)
+ + ]
1738 : : {
1739 : 437 : Relids required_outer = (Relids) lfirst(l);
1740 : 437 : ListCell *lcr;
1741 : 437 : AppendPathInput parameterized = {0};
1742 : 437 : bool parameterized_valid = true;
1743 : :
1744 : : /* Select the child paths for an Append with this parameterization */
1745 [ + - + + : 1627 : foreach(lcr, live_childrels)
+ + ]
1746 : : {
1747 : 1190 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1748 : 1190 : Path *subpath;
1749 : :
1750 [ + - ]: 1190 : if (childrel->pathlist == NIL)
1751 : : {
1752 : : /* failed to make a suitable path for this child */
1753 : 0 : parameterized_valid = false;
1754 : 0 : break;
1755 : : }
1756 : :
1757 : 2380 : subpath = get_cheapest_parameterized_child_path(root,
1758 : 1190 : childrel,
1759 : 1190 : required_outer);
1760 [ + - ]: 1190 : if (subpath == NULL)
1761 : : {
1762 : : /* failed to make a suitable path for this child */
1763 : 0 : parameterized_valid = false;
1764 : 0 : break;
1765 : : }
1766 : 2380 : accumulate_append_subpath(subpath, ¶meterized.subpaths, NULL,
1767 : 1190 : ¶meterized.child_append_relid_sets);
1768 [ - + ]: 1190 : }
1769 : :
1770 [ - + ]: 437 : if (parameterized_valid)
1771 : 874 : add_path(rel, (Path *)
1772 : 874 : create_append_path(root, rel, parameterized,
1773 : 437 : NIL, required_outer, 0, false,
1774 : : -1));
1775 : 437 : }
1776 : :
1777 : : /*
1778 : : * When there is only a single child relation, the Append path can inherit
1779 : : * any ordering available for the child rel's path, so that it's useful to
1780 : : * consider ordered partial paths. Above we only considered the cheapest
1781 : : * partial path for each child, but let's also make paths using any
1782 : : * partial paths that have pathkeys.
1783 : : */
1784 [ + + ]: 6378 : if (list_length(live_childrels) == 1)
1785 : : {
1786 : 1364 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1787 : :
1788 : : /* skip the cheapest partial path, since we already used that above */
1789 [ + + + + : 1394 : for_each_from(l, childrel->partial_pathlist, 1)
+ + ]
1790 : : {
1791 : 30 : Path *path = (Path *) lfirst(l);
1792 : 30 : AppendPath *appendpath;
1793 : 30 : AppendPathInput append = {0};
1794 : :
1795 : 30 : append.partial_subpaths = list_make1(path);
1796 : 30 : append.child_append_relid_sets = list_make1(rel->relids);
1797 : :
1798 : : /* skip paths with no pathkeys. */
1799 [ + - ]: 30 : if (path->pathkeys == NIL)
1800 : 0 : continue;
1801 : :
1802 : 60 : appendpath = create_append_path(root, rel, append, NIL, NULL,
1803 : 30 : path->parallel_workers, true,
1804 : 30 : partial_rows);
1805 : 30 : add_partial_path(rel, (Path *) appendpath);
1806 [ - - + ]: 30 : }
1807 : 1364 : }
1808 : 6378 : }
1809 : :
1810 : : /*
1811 : : * generate_orderedappend_paths
1812 : : * Generate ordered append paths for an append relation
1813 : : *
1814 : : * Usually we generate MergeAppend paths here, but there are some special
1815 : : * cases where we can generate simple Append paths, because the subpaths
1816 : : * can provide tuples in the required order already.
1817 : : *
1818 : : * We generate a path for each ordering (pathkey list) appearing in
1819 : : * all_child_pathkeys.
1820 : : *
1821 : : * We consider the cheapest-startup and cheapest-total cases, and also the
1822 : : * cheapest-fractional case when not all tuples need to be retrieved. For each
1823 : : * interesting ordering, we collect all the cheapest startup subpaths, all the
1824 : : * cheapest total paths, and, if applicable, all the cheapest fractional paths,
1825 : : * and build a suitable path for each case.
1826 : : *
1827 : : * We don't currently generate any parameterized ordered paths here. While
1828 : : * it would not take much more code here to do so, it's very unclear that it
1829 : : * is worth the planning cycles to investigate such paths: there's little
1830 : : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1831 : : * that the current coding of add_path would reject such paths out of hand,
1832 : : * because add_path gives no credit for sort ordering of parameterized paths,
1833 : : * and a parameterized MergeAppend is going to be more expensive than the
1834 : : * corresponding parameterized Append path. If we ever try harder to support
1835 : : * parameterized mergejoin plans, it might be worth adding support for
1836 : : * parameterized paths here to feed such joins. (See notes in
1837 : : * optimizer/README for why that might not ever happen, though.)
1838 : : */
1839 : : static void
1840 : 6326 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1841 : : List *live_childrels,
1842 : : List *all_child_pathkeys)
1843 : : {
1844 : 6326 : ListCell *lcp;
1845 : 6326 : List *partition_pathkeys = NIL;
1846 : 6326 : List *partition_pathkeys_desc = NIL;
1847 : 6326 : bool partition_pathkeys_partial = true;
1848 : 6326 : bool partition_pathkeys_desc_partial = true;
1849 : :
1850 : : /*
1851 : : * Some partitioned table setups may allow us to use an Append node
1852 : : * instead of a MergeAppend. This is possible in cases such as RANGE
1853 : : * partitioned tables where it's guaranteed that an earlier partition must
1854 : : * contain rows which come earlier in the sort order. To detect whether
1855 : : * this is relevant, build pathkey descriptions of the partition ordering,
1856 : : * for both forward and reverse scans.
1857 : : */
1858 [ + + + + : 6326 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
+ + ]
1859 : 5058 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1860 : : {
1861 : 2829 : partition_pathkeys = build_partition_pathkeys(root, rel,
1862 : : ForwardScanDirection,
1863 : : &partition_pathkeys_partial);
1864 : :
1865 : 2829 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1866 : : BackwardScanDirection,
1867 : : &partition_pathkeys_desc_partial);
1868 : :
1869 : : /*
1870 : : * You might think we should truncate_useless_pathkeys here, but
1871 : : * allowing partition keys which are a subset of the query's pathkeys
1872 : : * can often be useful. For example, consider a table partitioned by
1873 : : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1874 : : * paths that can produce the a, b, c ordering (perhaps via indexes on
1875 : : * (a, b, c)) then it works to consider the appendrel output as
1876 : : * ordered by a, b, c.
1877 : : */
1878 : 2829 : }
1879 : :
1880 : : /* Now consider each interesting sort ordering */
1881 [ + + + + : 8319 : foreach(lcp, all_child_pathkeys)
+ + ]
1882 : : {
1883 : 1993 : List *pathkeys = (List *) lfirst(lcp);
1884 : 1993 : AppendPathInput startup = {0};
1885 : 1993 : AppendPathInput total = {0};
1886 : 1993 : AppendPathInput fractional = {0};
1887 : 1993 : bool startup_neq_total = false;
1888 : 1993 : bool fraction_neq_total = false;
1889 : 1993 : bool match_partition_order;
1890 : 1993 : bool match_partition_order_desc;
1891 : 1993 : int end_index;
1892 : 1993 : int first_index;
1893 : 1993 : int direction;
1894 : :
1895 : : /*
1896 : : * Determine if this sort ordering matches any partition pathkeys we
1897 : : * have, for both ascending and descending partition order. If the
1898 : : * partition pathkeys happen to be contained in pathkeys then it still
1899 : : * works, as described above, providing that the partition pathkeys
1900 : : * are complete and not just a prefix of the partition keys. (In such
1901 : : * cases we'll be relying on the child paths to have sorted the
1902 : : * lower-order columns of the required pathkeys.)
1903 : : */
1904 : 1993 : match_partition_order =
1905 [ + + ]: 3627 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1906 [ + + ]: 1634 : (!partition_pathkeys_partial &&
1907 : 30 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1908 : :
1909 [ + + ]: 3611 : match_partition_order_desc = !match_partition_order &&
1910 [ + + ]: 3230 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1911 [ + + ]: 1612 : (!partition_pathkeys_desc_partial &&
1912 : 8 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1913 : :
1914 : : /*
1915 : : * When the required pathkeys match the reverse of the partition
1916 : : * order, we must build the list of paths in reverse starting with the
1917 : : * last matching partition first. We can get away without making any
1918 : : * special cases for this in the loop below by just looping backward
1919 : : * over the child relations in this case.
1920 : : */
1921 [ + + ]: 1993 : if (match_partition_order_desc)
1922 : : {
1923 : : /* loop backward */
1924 : 8 : first_index = list_length(live_childrels) - 1;
1925 : 8 : end_index = -1;
1926 : 8 : direction = -1;
1927 : :
1928 : : /*
1929 : : * Set this to true to save us having to check for
1930 : : * match_partition_order_desc in the loop below.
1931 : : */
1932 : 8 : match_partition_order = true;
1933 : 8 : }
1934 : : else
1935 : : {
1936 : : /* for all other case, loop forward */
1937 : 1985 : first_index = 0;
1938 : 1985 : end_index = list_length(live_childrels);
1939 : 1985 : direction = 1;
1940 : : }
1941 : :
1942 : : /* Select the child paths for this ordering... */
1943 [ + + ]: 7207 : for (int i = first_index; i != end_index; i += direction)
1944 : : {
1945 : 5214 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1946 : 5214 : Path *cheapest_startup,
1947 : : *cheapest_total,
1948 : 5214 : *cheapest_fractional = NULL;
1949 : :
1950 : : /* Locate the right paths, if they are available. */
1951 : 5214 : cheapest_startup =
1952 : 10428 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1953 : 5214 : pathkeys,
1954 : : NULL,
1955 : : STARTUP_COST,
1956 : : false);
1957 : 5214 : cheapest_total =
1958 : 10428 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1959 : 5214 : pathkeys,
1960 : : NULL,
1961 : : TOTAL_COST,
1962 : : false);
1963 : :
1964 : : /*
1965 : : * If we can't find any paths with the right order just use the
1966 : : * cheapest-total path; we'll have to sort it later.
1967 : : */
1968 [ + + + - ]: 5214 : if (cheapest_startup == NULL || cheapest_total == NULL)
1969 : : {
1970 : 50 : cheapest_startup = cheapest_total =
1971 : 50 : childrel->cheapest_total_path;
1972 : : /* Assert we do have an unparameterized path for this child */
1973 [ - + ]: 50 : Assert(cheapest_total->param_info == NULL);
1974 : 50 : }
1975 : :
1976 : : /*
1977 : : * When building a fractional path, determine a cheapest
1978 : : * fractional path for each child relation too. Looking at startup
1979 : : * and total costs is not enough, because the cheapest fractional
1980 : : * path may be dominated by two separate paths (one for startup,
1981 : : * one for total).
1982 : : *
1983 : : * When needed (building fractional path), determine the cheapest
1984 : : * fractional path too.
1985 : : */
1986 [ + + ]: 5214 : if (root->tuple_fraction > 0)
1987 : : {
1988 : 144 : double path_fraction = root->tuple_fraction;
1989 : :
1990 : : /*
1991 : : * We should not have a dummy child relation here. However,
1992 : : * we cannot use childrel->rows to compute the tuple fraction,
1993 : : * as childrel can be an upper relation with an unset row
1994 : : * estimate. Instead, we use the row estimate from the
1995 : : * cheapest_total path, which should already have been forced
1996 : : * to a sane value.
1997 : : */
1998 [ - + ]: 144 : Assert(cheapest_total->rows > 0);
1999 : :
2000 : : /* Convert absolute limit to a path fraction */
2001 [ - + ]: 144 : if (path_fraction >= 1.0)
2002 : 144 : path_fraction /= cheapest_total->rows;
2003 : :
2004 : 144 : cheapest_fractional =
2005 : 288 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
2006 : 144 : pathkeys,
2007 : : NULL,
2008 : 144 : path_fraction);
2009 : :
2010 : : /*
2011 : : * If we found no path with matching pathkeys, use the
2012 : : * cheapest total path instead.
2013 : : *
2014 : : * XXX We might consider partially sorted paths too (with an
2015 : : * incremental sort on top). But we'd have to build all the
2016 : : * incremental paths, do the costing etc.
2017 : : *
2018 : : * Also, notice whether we actually have different paths for
2019 : : * the "fractional" and "total" cases. This helps avoid
2020 : : * generating two identical ordered append paths.
2021 : : */
2022 [ + + ]: 144 : if (cheapest_fractional == NULL)
2023 : 6 : cheapest_fractional = cheapest_total;
2024 [ - + ]: 138 : else if (cheapest_fractional != cheapest_total)
2025 : 0 : fraction_neq_total = true;
2026 : 144 : }
2027 : :
2028 : : /*
2029 : : * Notice whether we actually have different paths for the
2030 : : * "cheapest" and "total" cases. This helps avoid generating two
2031 : : * identical ordered append paths.
2032 : : */
2033 [ + + ]: 5214 : if (cheapest_startup != cheapest_total)
2034 : 16 : startup_neq_total = true;
2035 : :
2036 : : /*
2037 : : * Collect the appropriate child paths. The required logic varies
2038 : : * for the Append and MergeAppend cases.
2039 : : */
2040 [ + + ]: 5214 : if (match_partition_order)
2041 : : {
2042 : : /*
2043 : : * We're going to make a plain Append path. We don't need
2044 : : * most of what accumulate_append_subpath would do, but we do
2045 : : * want to cut out child Appends or MergeAppends if they have
2046 : : * just a single subpath (and hence aren't doing anything
2047 : : * useful).
2048 : : */
2049 : 1056 : cheapest_startup =
2050 : 2112 : get_singleton_append_subpath(cheapest_startup,
2051 : 1056 : &startup.child_append_relid_sets);
2052 : 1056 : cheapest_total =
2053 : 2112 : get_singleton_append_subpath(cheapest_total,
2054 : 1056 : &total.child_append_relid_sets);
2055 : :
2056 : 1056 : startup.subpaths = lappend(startup.subpaths, cheapest_startup);
2057 : 1056 : total.subpaths = lappend(total.subpaths, cheapest_total);
2058 : :
2059 [ + + ]: 1056 : if (cheapest_fractional)
2060 : : {
2061 : 24 : cheapest_fractional =
2062 : 48 : get_singleton_append_subpath(cheapest_fractional,
2063 : 24 : &fractional.child_append_relid_sets);
2064 : 24 : fractional.subpaths =
2065 : 24 : lappend(fractional.subpaths, cheapest_fractional);
2066 : 24 : }
2067 : 1056 : }
2068 : : else
2069 : : {
2070 : : /*
2071 : : * Otherwise, rely on accumulate_append_subpath to collect the
2072 : : * child paths for the MergeAppend.
2073 : : */
2074 : 8316 : accumulate_append_subpath(cheapest_startup,
2075 : 4158 : &startup.subpaths, NULL,
2076 : 4158 : &startup.child_append_relid_sets);
2077 : 8316 : accumulate_append_subpath(cheapest_total,
2078 : 4158 : &total.subpaths, NULL,
2079 : 4158 : &total.child_append_relid_sets);
2080 : :
2081 [ + + ]: 4158 : if (cheapest_fractional)
2082 : 240 : accumulate_append_subpath(cheapest_fractional,
2083 : 120 : &fractional.subpaths, NULL,
2084 : 120 : &fractional.child_append_relid_sets);
2085 : : }
2086 : 5214 : }
2087 : :
2088 : : /* ... and build the Append or MergeAppend paths */
2089 [ + + ]: 1993 : if (match_partition_order)
2090 : : {
2091 : : /* We only need Append */
2092 : 766 : add_path(rel, (Path *) create_append_path(root,
2093 : 383 : rel,
2094 : : startup,
2095 : 383 : pathkeys,
2096 : : NULL,
2097 : : 0,
2098 : : false,
2099 : : -1));
2100 [ + - ]: 383 : if (startup_neq_total)
2101 : 0 : add_path(rel, (Path *) create_append_path(root,
2102 : 0 : rel,
2103 : : total,
2104 : 0 : pathkeys,
2105 : : NULL,
2106 : : 0,
2107 : : false,
2108 : : -1));
2109 : :
2110 [ + + + - ]: 383 : if (fractional.subpaths && fraction_neq_total)
2111 : 0 : add_path(rel, (Path *) create_append_path(root,
2112 : 0 : rel,
2113 : : fractional,
2114 : 0 : pathkeys,
2115 : : NULL,
2116 : : 0,
2117 : : false,
2118 : : -1));
2119 : 383 : }
2120 : : else
2121 : : {
2122 : : /* We need MergeAppend */
2123 : 3220 : add_path(rel, (Path *) create_merge_append_path(root,
2124 : 1610 : rel,
2125 : 1610 : startup.subpaths,
2126 : 1610 : startup.child_append_relid_sets,
2127 : 1610 : pathkeys,
2128 : : NULL));
2129 [ + + ]: 1610 : if (startup_neq_total)
2130 : 20 : add_path(rel, (Path *) create_merge_append_path(root,
2131 : 10 : rel,
2132 : 10 : total.subpaths,
2133 : 10 : total.child_append_relid_sets,
2134 : 10 : pathkeys,
2135 : : NULL));
2136 : :
2137 [ + + + - ]: 1610 : if (fractional.subpaths && fraction_neq_total)
2138 : 0 : add_path(rel, (Path *) create_merge_append_path(root,
2139 : 0 : rel,
2140 : 0 : fractional.subpaths,
2141 : 0 : fractional.child_append_relid_sets,
2142 : 0 : pathkeys,
2143 : : NULL));
2144 : : }
2145 : 1993 : }
2146 : 6326 : }
2147 : :
2148 : : /*
2149 : : * get_cheapest_parameterized_child_path
2150 : : * Get cheapest path for this relation that has exactly the requested
2151 : : * parameterization.
2152 : : *
2153 : : * Returns NULL if unable to create such a path.
2154 : : */
2155 : : static Path *
2156 : 1190 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2157 : : Relids required_outer)
2158 : : {
2159 : 1190 : Path *cheapest;
2160 : 1190 : ListCell *lc;
2161 : :
2162 : : /*
2163 : : * Look up the cheapest existing path with no more than the needed
2164 : : * parameterization. If it has exactly the needed parameterization, we're
2165 : : * done.
2166 : : */
2167 : 2380 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2168 : : NIL,
2169 : 1190 : required_outer,
2170 : : TOTAL_COST,
2171 : : false);
2172 [ + - ]: 1190 : Assert(cheapest != NULL);
2173 [ + + + + ]: 1190 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2174 : 1137 : return cheapest;
2175 : :
2176 : : /*
2177 : : * Otherwise, we can "reparameterize" an existing path to match the given
2178 : : * parameterization, which effectively means pushing down additional
2179 : : * joinquals to be checked within the path's scan. However, some existing
2180 : : * paths might check the available joinquals already while others don't;
2181 : : * therefore, it's not clear which existing path will be cheapest after
2182 : : * reparameterization. We have to go through them all and find out.
2183 : : */
2184 : 53 : cheapest = NULL;
2185 [ + - + + : 184 : foreach(lc, rel->pathlist)
+ + ]
2186 : : {
2187 : 131 : Path *path = (Path *) lfirst(lc);
2188 : :
2189 : : /* Can't use it if it needs more than requested parameterization */
2190 [ + + + + ]: 131 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2191 : 4 : continue;
2192 : :
2193 : : /*
2194 : : * Reparameterization can only increase the path's cost, so if it's
2195 : : * already more expensive than the current cheapest, forget it.
2196 : : */
2197 [ + + + + ]: 127 : if (cheapest != NULL &&
2198 : 74 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2199 : 62 : continue;
2200 : :
2201 : : /* Reparameterize if needed, then recheck cost */
2202 [ + + + + ]: 65 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2203 : : {
2204 : 53 : path = reparameterize_path(root, path, required_outer, 1.0);
2205 [ + - ]: 53 : if (path == NULL)
2206 : 0 : continue; /* failed to reparameterize this one */
2207 [ + - - + ]: 53 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2208 : :
2209 [ - + # # ]: 53 : if (cheapest != NULL &&
2210 : 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2211 : 0 : continue;
2212 : 53 : }
2213 : :
2214 : : /* We have a new best path */
2215 : 65 : cheapest = path;
2216 [ - + + ]: 131 : }
2217 : :
2218 : : /* Return the best path, or NULL if we found no suitable candidate */
2219 : 53 : return cheapest;
2220 : 1190 : }
2221 : :
2222 : : /*
2223 : : * accumulate_append_subpath
2224 : : * Add a subpath to the list being built for an Append or MergeAppend.
2225 : : *
2226 : : * It's possible that the child is itself an Append or MergeAppend path, in
2227 : : * which case we can "cut out the middleman" and just add its child paths to
2228 : : * our own list. (We don't try to do this earlier because we need to apply
2229 : : * both levels of transformation to the quals.)
2230 : : *
2231 : : * Note that if we omit a child MergeAppend in this way, we are effectively
2232 : : * omitting a sort step, which seems fine: if the parent is to be an Append,
2233 : : * its result would be unsorted anyway, while if the parent is to be a
2234 : : * MergeAppend, there's no point in a separate sort on a child.
2235 : : *
2236 : : * Normally, either path is a partial path and subpaths is a list of partial
2237 : : * paths, or else path is a non-partial plan and subpaths is a list of those.
2238 : : * However, if path is a parallel-aware Append, then we add its partial path
2239 : : * children to subpaths and the rest to special_subpaths. If the latter is
2240 : : * NULL, we don't flatten the path at all (unless it contains only partial
2241 : : * paths).
2242 : : */
2243 : : static void
2244 : 45286 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths,
2245 : : List **child_append_relid_sets)
2246 : : {
2247 [ + + ]: 45286 : if (IsA(path, AppendPath))
2248 : : {
2249 : 2570 : AppendPath *apath = (AppendPath *) path;
2250 : :
2251 [ + + + + ]: 2570 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2252 : : {
2253 : 2514 : *subpaths = list_concat(*subpaths, apath->subpaths);
2254 : 2514 : *child_append_relid_sets =
2255 : 2514 : lappend(*child_append_relid_sets, path->parent->relids);
2256 : 2514 : *child_append_relid_sets =
2257 : 5028 : list_concat(*child_append_relid_sets,
2258 : 2514 : apath->child_append_relid_sets);
2259 : 2514 : return;
2260 : : }
2261 [ + + ]: 56 : else if (special_subpaths != NULL)
2262 : : {
2263 : 28 : List *new_special_subpaths;
2264 : :
2265 : : /* Split Parallel Append into partial and non-partial subpaths */
2266 : 56 : *subpaths = list_concat(*subpaths,
2267 : 56 : list_copy_tail(apath->subpaths,
2268 : 28 : apath->first_partial_path));
2269 : 56 : new_special_subpaths = list_copy_head(apath->subpaths,
2270 : 28 : apath->first_partial_path);
2271 : 56 : *special_subpaths = list_concat(*special_subpaths,
2272 : 28 : new_special_subpaths);
2273 : 28 : *child_append_relid_sets =
2274 : 28 : lappend(*child_append_relid_sets, path->parent->relids);
2275 : 28 : *child_append_relid_sets =
2276 : 56 : list_concat(*child_append_relid_sets,
2277 : 28 : apath->child_append_relid_sets);
2278 : : return;
2279 : 28 : }
2280 [ - + + ]: 2570 : }
2281 [ + + ]: 42716 : else if (IsA(path, MergeAppendPath))
2282 : : {
2283 : 176 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2284 : :
2285 : 176 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2286 : 176 : *child_append_relid_sets =
2287 : 176 : lappend(*child_append_relid_sets, path->parent->relids);
2288 : 176 : *child_append_relid_sets =
2289 : 352 : list_concat(*child_append_relid_sets,
2290 : 176 : mpath->child_append_relid_sets);
2291 : : return;
2292 : 176 : }
2293 : :
2294 : 42568 : *subpaths = lappend(*subpaths, path);
2295 : 45286 : }
2296 : :
2297 : : /*
2298 : : * get_singleton_append_subpath
2299 : : * Returns the single subpath of an Append/MergeAppend, or just
2300 : : * return 'path' if it's not a single sub-path Append/MergeAppend.
2301 : : *
2302 : : * As a side effect, whenever we return a single subpath rather than the
2303 : : * original path, add the relid set for the original path to
2304 : : * child_append_relid_sets, so that those relids don't entirely disappear
2305 : : * from the final plan.
2306 : : *
2307 : : * Note: 'path' must not be a parallel-aware path.
2308 : : */
2309 : : static Path *
2310 : 2136 : get_singleton_append_subpath(Path *path, List **child_append_relid_sets)
2311 : : {
2312 [ + - ]: 2136 : Assert(!path->parallel_aware);
2313 : :
2314 [ + + ]: 2136 : if (IsA(path, AppendPath))
2315 : : {
2316 : 52 : AppendPath *apath = (AppendPath *) path;
2317 : :
2318 [ + + ]: 52 : if (list_length(apath->subpaths) == 1)
2319 : : {
2320 : 26 : *child_append_relid_sets =
2321 : 26 : lappend(*child_append_relid_sets, path->parent->relids);
2322 : 26 : return (Path *) linitial(apath->subpaths);
2323 : : }
2324 [ + + ]: 52 : }
2325 [ + + ]: 2084 : else if (IsA(path, MergeAppendPath))
2326 : : {
2327 : 58 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2328 : :
2329 [ - + ]: 58 : if (list_length(mpath->subpaths) == 1)
2330 : : {
2331 : 0 : *child_append_relid_sets =
2332 : 0 : lappend(*child_append_relid_sets, path->parent->relids);
2333 : 0 : return (Path *) linitial(mpath->subpaths);
2334 : : }
2335 [ - + ]: 58 : }
2336 : :
2337 : 2110 : return path;
2338 : 2136 : }
2339 : :
2340 : : /*
2341 : : * set_dummy_rel_pathlist
2342 : : * Build a dummy path for a relation that's been excluded by constraints
2343 : : *
2344 : : * Rather than inventing a special "dummy" path type, we represent this as an
2345 : : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2346 : : *
2347 : : * (See also mark_dummy_rel, which does basically the same thing, but is
2348 : : * typically used to change a rel into dummy state after we already made
2349 : : * paths for it.)
2350 : : */
2351 : : static void
2352 : 209 : set_dummy_rel_pathlist(RelOptInfo *rel)
2353 : : {
2354 : 209 : AppendPathInput in = {0};
2355 : :
2356 : : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2357 : 209 : rel->rows = 0;
2358 : 209 : rel->reltarget->width = 0;
2359 : :
2360 : : /* Discard any pre-existing paths; no further need for them */
2361 : 209 : rel->pathlist = NIL;
2362 : 209 : rel->partial_pathlist = NIL;
2363 : :
2364 : : /* Set up the dummy path */
2365 : 418 : add_path(rel, (Path *) create_append_path(NULL, rel, in,
2366 : 209 : NIL, rel->lateral_relids,
2367 : : 0, false, -1));
2368 : :
2369 : : /*
2370 : : * We set the cheapest-path fields immediately, just in case they were
2371 : : * pointing at some discarded path. This is redundant in current usage
2372 : : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2373 : : * for safety and consistency with mark_dummy_rel.
2374 : : */
2375 : 209 : set_cheapest(rel);
2376 : 209 : }
2377 : :
2378 : : /*
2379 : : * find_window_run_conditions
2380 : : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2381 : : * function to determine the function's monotonic properties. We then
2382 : : * see if 'opexpr' can be used to short-circuit execution.
2383 : : *
2384 : : * For example row_number() over (order by ...) always produces a value one
2385 : : * higher than the previous. If someone has a window function in a subquery
2386 : : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2387 : : * as well stop processing the windowagg once the row number reaches 11. Here
2388 : : * we check if 'opexpr' might help us to stop doing needless extra processing
2389 : : * in WindowAgg nodes.
2390 : : *
2391 : : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2392 : : * its original purpose. This is set to false if the caller can assume that
2393 : : * the run condition will handle all of the required filtering.
2394 : : *
2395 : : * Returns true if 'opexpr' was found to be useful and was added to the
2396 : : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2397 : : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2398 : : * If the 'opexpr' cannot be used then we set *keep_original to true and
2399 : : * return false.
2400 : : */
2401 : : static bool
2402 : 40 : find_window_run_conditions(Query *subquery, AttrNumber attno,
2403 : : WindowFunc *wfunc, OpExpr *opexpr, bool wfunc_left,
2404 : : bool *keep_original, Bitmapset **run_cond_attrs)
2405 : : {
2406 : 40 : Oid prosupport;
2407 : 40 : Expr *otherexpr;
2408 : 40 : SupportRequestWFuncMonotonic req;
2409 : 40 : SupportRequestWFuncMonotonic *res;
2410 : 40 : WindowClause *wclause;
2411 : 40 : List *opinfos;
2412 : 40 : OpExpr *runopexpr;
2413 : 40 : Oid runoperator;
2414 : 40 : ListCell *lc;
2415 : :
2416 : 40 : *keep_original = true;
2417 : :
2418 [ - + ]: 40 : while (IsA(wfunc, RelabelType))
2419 : 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2420 : :
2421 : : /* we can only work with window functions */
2422 [ + + ]: 40 : if (!IsA(wfunc, WindowFunc))
2423 : 4 : return false;
2424 : :
2425 : : /* can't use it if there are subplans in the WindowFunc */
2426 [ + + ]: 36 : if (contain_subplans((Node *) wfunc))
2427 : 1 : return false;
2428 : :
2429 : 35 : prosupport = get_func_support(wfunc->winfnoid);
2430 : :
2431 : : /* Check if there's a support function for 'wfunc' */
2432 [ + + ]: 35 : if (!OidIsValid(prosupport))
2433 : 3 : return false;
2434 : :
2435 : : /* get the Expr from the other side of the OpExpr */
2436 [ + + ]: 32 : if (wfunc_left)
2437 : 28 : otherexpr = lsecond(opexpr->args);
2438 : : else
2439 : 4 : otherexpr = linitial(opexpr->args);
2440 : :
2441 : : /*
2442 : : * The value being compared must not change during the evaluation of the
2443 : : * window partition.
2444 : : */
2445 [ + - ]: 32 : if (!is_pseudo_constant_clause((Node *) otherexpr))
2446 : 0 : return false;
2447 : :
2448 : : /* find the window clause belonging to the window function */
2449 : 64 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2450 : 32 : wfunc->winref - 1);
2451 : :
2452 : 32 : req.type = T_SupportRequestWFuncMonotonic;
2453 : 32 : req.window_func = wfunc;
2454 : 32 : req.window_clause = wclause;
2455 : :
2456 : : /* call the support function */
2457 : 32 : res = (SupportRequestWFuncMonotonic *)
2458 : 32 : DatumGetPointer(OidFunctionCall1(prosupport,
2459 : : PointerGetDatum(&req)));
2460 : :
2461 : : /*
2462 : : * Nothing to do if the function is neither monotonically increasing nor
2463 : : * monotonically decreasing.
2464 : : */
2465 [ + - - + ]: 32 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
2466 : 0 : return false;
2467 : :
2468 : 32 : runopexpr = NULL;
2469 : 32 : runoperator = InvalidOid;
2470 : 32 : opinfos = get_op_index_interpretation(opexpr->opno);
2471 : :
2472 [ + - - + : 64 : foreach(lc, opinfos)
+ - ]
2473 : : {
2474 : 32 : OpIndexInterpretation *opinfo = (OpIndexInterpretation *) lfirst(lc);
2475 : 32 : CompareType cmptype = opinfo->cmptype;
2476 : :
2477 : : /* handle < / <= */
2478 [ + + + + ]: 32 : if (cmptype == COMPARE_LT || cmptype == COMPARE_LE)
2479 : : {
2480 : : /*
2481 : : * < / <= is supported for monotonically increasing functions in
2482 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2483 : : * for monotonically decreasing functions.
2484 : : */
2485 [ + + + + ]: 25 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2486 [ + + ]: 4 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2487 : : {
2488 : 21 : *keep_original = false;
2489 : 21 : runopexpr = opexpr;
2490 : 21 : runoperator = opexpr->opno;
2491 : 21 : }
2492 : 23 : break;
2493 : : }
2494 : : /* handle > / >= */
2495 [ + + + + ]: 9 : else if (cmptype == COMPARE_GT || cmptype == COMPARE_GE)
2496 : : {
2497 : : /*
2498 : : * > / >= is supported for monotonically decreasing functions in
2499 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2500 : : * for monotonically increasing functions.
2501 : : */
2502 [ + + + - ]: 5 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2503 [ + - ]: 2 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2504 : : {
2505 : 3 : *keep_original = false;
2506 : 3 : runopexpr = opexpr;
2507 : 3 : runoperator = opexpr->opno;
2508 : 3 : }
2509 : 3 : break;
2510 : : }
2511 : : /* handle = */
2512 [ + - ]: 6 : else if (cmptype == COMPARE_EQ)
2513 : : {
2514 : 6 : CompareType newcmptype;
2515 : :
2516 : : /*
2517 : : * When both monotonically increasing and decreasing then the
2518 : : * return value of the window function will be the same each time.
2519 : : * We can simply use 'opexpr' as the run condition without
2520 : : * modifying it.
2521 : : */
2522 [ + + ]: 6 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2523 : : {
2524 : 1 : *keep_original = false;
2525 : 1 : runopexpr = opexpr;
2526 : 1 : runoperator = opexpr->opno;
2527 : 1 : break;
2528 : : }
2529 : :
2530 : : /*
2531 : : * When monotonically increasing we make a qual with <wfunc> <=
2532 : : * <value> or <value> >= <wfunc> in order to filter out values
2533 : : * which are above the value in the equality condition. For
2534 : : * monotonically decreasing functions we want to filter values
2535 : : * below the value in the equality condition.
2536 : : */
2537 [ + - ]: 5 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2538 : 5 : newcmptype = wfunc_left ? COMPARE_LE : COMPARE_GE;
2539 : : else
2540 : 0 : newcmptype = wfunc_left ? COMPARE_GE : COMPARE_LE;
2541 : :
2542 : : /* We must keep the original equality qual */
2543 : 5 : *keep_original = true;
2544 : 5 : runopexpr = opexpr;
2545 : :
2546 : : /* determine the operator to use for the WindowFuncRunCondition */
2547 : 10 : runoperator = get_opfamily_member_for_cmptype(opinfo->opfamily_id,
2548 : 5 : opinfo->oplefttype,
2549 : 5 : opinfo->oprighttype,
2550 : 5 : newcmptype);
2551 : 5 : break;
2552 : 6 : }
2553 [ + - ]: 32 : }
2554 : :
2555 [ + + ]: 32 : if (runopexpr != NULL)
2556 : : {
2557 : 30 : WindowFuncRunCondition *wfuncrc;
2558 : :
2559 : 30 : wfuncrc = makeNode(WindowFuncRunCondition);
2560 : 30 : wfuncrc->opno = runoperator;
2561 : 30 : wfuncrc->inputcollid = runopexpr->inputcollid;
2562 : 30 : wfuncrc->wfunc_left = wfunc_left;
2563 : 30 : wfuncrc->arg = copyObject(otherexpr);
2564 : :
2565 : 30 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2566 : :
2567 : : /* record that this attno was used in a run condition */
2568 : 60 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2569 : 30 : attno - FirstLowInvalidHeapAttributeNumber);
2570 : 30 : return true;
2571 : 30 : }
2572 : :
2573 : : /* unsupported OpExpr */
2574 : 2 : return false;
2575 : 40 : }
2576 : :
2577 : : /*
2578 : : * check_and_push_window_quals
2579 : : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2580 : : * as a 'runCondition' qual. These, when present, allow some unnecessary
2581 : : * work to be skipped during execution.
2582 : : *
2583 : : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2584 : : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2585 : : * window quals for.
2586 : : *
2587 : : * Returns true if the caller still must keep the original qual or false if
2588 : : * the caller can safely ignore the original qual because the WindowAgg node
2589 : : * will use the runCondition to stop returning tuples.
2590 : : */
2591 : : static bool
2592 : 42 : check_and_push_window_quals(Query *subquery, Node *clause,
2593 : : Bitmapset **run_cond_attrs)
2594 : : {
2595 : 42 : OpExpr *opexpr = (OpExpr *) clause;
2596 : 42 : bool keep_original = true;
2597 : 42 : Var *var1;
2598 : 42 : Var *var2;
2599 : :
2600 : : /* We're only able to use OpExprs with 2 operands */
2601 [ + + ]: 42 : if (!IsA(opexpr, OpExpr))
2602 : 3 : return true;
2603 : :
2604 [ - + ]: 39 : if (list_length(opexpr->args) != 2)
2605 : 0 : return true;
2606 : :
2607 : : /*
2608 : : * Currently, we restrict this optimization to strict OpExprs. The reason
2609 : : * for this is that during execution, once the runcondition becomes false,
2610 : : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2611 : : * function result values, we set them to NULL. Having only strict
2612 : : * OpExprs here ensures that we properly filter out the tuples with NULLs
2613 : : * in the top-level WindowAgg.
2614 : : */
2615 : 39 : set_opfuncid(opexpr);
2616 [ + - ]: 39 : if (!func_strict(opexpr->opfuncid))
2617 : 0 : return true;
2618 : :
2619 : : /*
2620 : : * Check for plain Vars that reference window functions in the subquery.
2621 : : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2622 : : * be used as part of the run condition.
2623 : : */
2624 : :
2625 : : /* Check the left side of the OpExpr */
2626 : 39 : var1 = linitial(opexpr->args);
2627 [ + + - + ]: 39 : if (IsA(var1, Var) && var1->varattno > 0)
2628 : : {
2629 : 33 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2630 : 33 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2631 : :
2632 [ + + + + ]: 66 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2633 : 33 : true, &keep_original, run_cond_attrs))
2634 : 27 : return keep_original;
2635 [ + + ]: 33 : }
2636 : :
2637 : : /* and check the right side */
2638 : 12 : var2 = lsecond(opexpr->args);
2639 [ + + - + ]: 12 : if (IsA(var2, Var) && var2->varattno > 0)
2640 : : {
2641 : 7 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2642 : 7 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2643 : :
2644 [ + + + + ]: 14 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2645 : 7 : false, &keep_original, run_cond_attrs))
2646 : 3 : return keep_original;
2647 [ + + ]: 7 : }
2648 : :
2649 : 9 : return true;
2650 : 42 : }
2651 : :
2652 : : /*
2653 : : * set_subquery_pathlist
2654 : : * Generate SubqueryScan access paths for a subquery RTE
2655 : : *
2656 : : * We don't currently support generating parameterized paths for subqueries
2657 : : * by pushing join clauses down into them; it seems too expensive to re-plan
2658 : : * the subquery multiple times to consider different alternatives.
2659 : : * (XXX that could stand to be reconsidered, now that we use Paths.)
2660 : : * So the paths made here will be parameterized if the subquery contains
2661 : : * LATERAL references, otherwise not. As long as that's true, there's no need
2662 : : * for a separate set_subquery_size phase: just make the paths right away.
2663 : : */
2664 : : static void
2665 : 1480 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2666 : : Index rti, RangeTblEntry *rte)
2667 : : {
2668 : 1480 : Query *parse = root->parse;
2669 : 1480 : Query *subquery = rte->subquery;
2670 : 1480 : bool trivial_pathtarget;
2671 : 1480 : Relids required_outer;
2672 : 1480 : pushdown_safety_info safetyInfo;
2673 : 1480 : double tuple_fraction;
2674 : 1480 : RelOptInfo *sub_final_rel;
2675 : 1480 : Bitmapset *run_cond_attrs = NULL;
2676 : 1480 : ListCell *lc;
2677 : 1480 : char *plan_name;
2678 : :
2679 : : /*
2680 : : * Must copy the Query so that planning doesn't mess up the RTE contents
2681 : : * (really really need to fix the planner to not scribble on its input,
2682 : : * someday ... but see remove_unused_subquery_outputs to start with).
2683 : : */
2684 : 1480 : subquery = copyObject(subquery);
2685 : :
2686 : : /*
2687 : : * If it's a LATERAL subquery, it might contain some Vars of the current
2688 : : * query level, requiring it to be treated as parameterized, even though
2689 : : * we don't support pushing down join quals into subqueries.
2690 : : */
2691 : 1480 : required_outer = rel->lateral_relids;
2692 : :
2693 : : /*
2694 : : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2695 : : * flags as needed while recursing. In particular, we need a workspace
2696 : : * for keeping track of the reasons why columns are unsafe to reference.
2697 : : * These reasons are stored in the bits inside unsafeFlags[i] when we
2698 : : * discover reasons that column i of the subquery is unsafe to be used in
2699 : : * a pushed-down qual.
2700 : : */
2701 : 1480 : memset(&safetyInfo, 0, sizeof(safetyInfo));
2702 : 1480 : safetyInfo.unsafeFlags = (unsigned char *)
2703 : 1480 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2704 : :
2705 : : /*
2706 : : * If the subquery has the "security_barrier" flag, it means the subquery
2707 : : * originated from a view that must enforce row-level security. Then we
2708 : : * must not push down quals that contain leaky functions. (Ideally this
2709 : : * would be checked inside subquery_is_pushdown_safe, but since we don't
2710 : : * currently pass the RTE to that function, we must do it here.)
2711 : : */
2712 : 1480 : safetyInfo.unsafeLeaky = rte->security_barrier;
2713 : :
2714 : : /*
2715 : : * If there are any restriction clauses that have been attached to the
2716 : : * subquery relation, consider pushing them down to become WHERE or HAVING
2717 : : * quals of the subquery itself. This transformation is useful because it
2718 : : * may allow us to generate a better plan for the subquery than evaluating
2719 : : * all the subquery output rows and then filtering them.
2720 : : *
2721 : : * There are several cases where we cannot push down clauses. Restrictions
2722 : : * involving the subquery are checked by subquery_is_pushdown_safe().
2723 : : * Restrictions on individual clauses are checked by
2724 : : * qual_is_pushdown_safe(). Also, we don't want to push down
2725 : : * pseudoconstant clauses; better to have the gating node above the
2726 : : * subquery.
2727 : : *
2728 : : * Non-pushed-down clauses will get evaluated as qpquals of the
2729 : : * SubqueryScan node.
2730 : : *
2731 : : * XXX Are there any cases where we want to make a policy decision not to
2732 : : * push down a pushable qual, because it'd result in a worse plan?
2733 : : */
2734 [ + + + + ]: 1480 : if (rel->baserestrictinfo != NIL &&
2735 : 290 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2736 : : {
2737 : : /* OK to consider pushing down individual quals */
2738 : 266 : List *upperrestrictlist = NIL;
2739 : 266 : ListCell *l;
2740 : :
2741 [ + - + + : 699 : foreach(l, rel->baserestrictinfo)
+ + ]
2742 : : {
2743 : 433 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
2744 : 433 : Node *clause = (Node *) rinfo->clause;
2745 : :
2746 [ - + ]: 433 : if (rinfo->pseudoconstant)
2747 : : {
2748 : 0 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2749 : 0 : continue;
2750 : : }
2751 : :
2752 [ + - + + ]: 433 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2753 : : {
2754 : : case PUSHDOWN_SAFE:
2755 : : /* Push it down */
2756 : 310 : subquery_push_qual(subquery, rte, rti, clause);
2757 : 310 : break;
2758 : :
2759 : : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2760 : :
2761 : : /*
2762 : : * Since we can't push the qual down into the subquery,
2763 : : * check if it happens to reference a window function. If
2764 : : * so then it might be useful to use for the WindowAgg's
2765 : : * runCondition.
2766 : : */
2767 [ + - + + ]: 42 : if (!subquery->hasWindowFuncs ||
2768 : 42 : check_and_push_window_quals(subquery, clause,
2769 : : &run_cond_attrs))
2770 : : {
2771 : : /*
2772 : : * subquery has no window funcs or the clause is not a
2773 : : * suitable window run condition qual or it is, but
2774 : : * the original must also be kept in the upper query.
2775 : : */
2776 : 17 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2777 : 17 : }
2778 : 42 : break;
2779 : :
2780 : : case PUSHDOWN_UNSAFE:
2781 : 81 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2782 : 81 : break;
2783 : : }
2784 [ - + ]: 433 : }
2785 : 266 : rel->baserestrictinfo = upperrestrictlist;
2786 : : /* We don't bother recomputing baserestrict_min_security */
2787 : 266 : }
2788 : :
2789 : 1480 : pfree(safetyInfo.unsafeFlags);
2790 : :
2791 : : /*
2792 : : * The upper query might not use all the subquery's output columns; if
2793 : : * not, we can simplify. Pass the attributes that were pushed down into
2794 : : * WindowAgg run conditions to ensure we don't accidentally think those
2795 : : * are unused.
2796 : : */
2797 : 1480 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2798 : :
2799 : : /*
2800 : : * We can safely pass the outer tuple_fraction down to the subquery if the
2801 : : * outer level has no joining, aggregation, or sorting to do. Otherwise
2802 : : * we'd better tell the subquery to plan for full retrieval. (XXX This
2803 : : * could probably be made more intelligent ...)
2804 : : */
2805 [ + + ]: 1480 : if (parse->hasAggs ||
2806 [ + + ]: 1300 : parse->groupClause ||
2807 [ + - ]: 1297 : parse->groupingSets ||
2808 [ + - ]: 1297 : root->hasHavingQual ||
2809 [ + + ]: 1297 : parse->distinctClause ||
2810 [ + + + + ]: 1263 : parse->sortClause ||
2811 : 582 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
2812 : 1102 : tuple_fraction = 0.0; /* default case */
2813 : : else
2814 : 378 : tuple_fraction = root->tuple_fraction;
2815 : :
2816 : : /* plan_params should not be in use in current query level */
2817 [ + - ]: 1480 : Assert(root->plan_params == NIL);
2818 : :
2819 : : /* Generate a subroot and Paths for the subquery */
2820 : 1480 : plan_name = choose_plan_name(root->glob, rte->eref->aliasname, false);
2821 : 2960 : rel->subroot = subquery_planner(root->glob, subquery, plan_name,
2822 : 1480 : root, false, tuple_fraction, NULL);
2823 : :
2824 : : /* Isolate the params needed by this specific subplan */
2825 : 1480 : rel->subplan_params = root->plan_params;
2826 : 1480 : root->plan_params = NIL;
2827 : :
2828 : : /*
2829 : : * It's possible that constraint exclusion proved the subquery empty. If
2830 : : * so, it's desirable to produce an unadorned dummy path so that we will
2831 : : * recognize appropriate optimizations at this query level.
2832 : : */
2833 : 1480 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2834 : :
2835 [ + + ]: 1480 : if (IS_DUMMY_REL(sub_final_rel))
2836 : : {
2837 : 21 : set_dummy_rel_pathlist(rel);
2838 : 21 : return;
2839 : : }
2840 : :
2841 : : /*
2842 : : * Mark rel with estimated output rows, width, etc. Note that we have to
2843 : : * do this before generating outer-query paths, else cost_subqueryscan is
2844 : : * not happy.
2845 : : */
2846 : 1459 : set_subquery_size_estimates(root, rel);
2847 : :
2848 : : /*
2849 : : * Also detect whether the reltarget is trivial, so that we can pass that
2850 : : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2851 : : * It's trivial if it fetches all the subplan output columns in order.
2852 : : */
2853 [ + + ]: 1459 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2854 : 358 : trivial_pathtarget = false;
2855 : : else
2856 : : {
2857 : 1101 : trivial_pathtarget = true;
2858 [ + - + + : 3266 : foreach(lc, rel->reltarget->exprs)
+ + ]
2859 : : {
2860 : 2165 : Node *node = (Node *) lfirst(lc);
2861 : 2165 : Var *var;
2862 : :
2863 [ + - ]: 2165 : if (!IsA(node, Var))
2864 : : {
2865 : 0 : trivial_pathtarget = false;
2866 : 0 : break;
2867 : : }
2868 : 2165 : var = (Var *) node;
2869 [ + - + + ]: 2165 : if (var->varno != rti ||
2870 : 2165 : var->varattno != foreach_current_index(lc) + 1)
2871 : : {
2872 : 49 : trivial_pathtarget = false;
2873 : 49 : break;
2874 : : }
2875 [ + + ]: 2165 : }
2876 : : }
2877 : :
2878 : : /*
2879 : : * For each Path that subquery_planner produced, make a SubqueryScanPath
2880 : : * in the outer query.
2881 : : */
2882 [ + - + + : 3026 : foreach(lc, sub_final_rel->pathlist)
+ + ]
2883 : : {
2884 : 1567 : Path *subpath = (Path *) lfirst(lc);
2885 : 1567 : List *pathkeys;
2886 : :
2887 : : /* Convert subpath's pathkeys to outer representation */
2888 : 3134 : pathkeys = convert_subquery_pathkeys(root,
2889 : 1567 : rel,
2890 : 1567 : subpath->pathkeys,
2891 : 1567 : make_tlist_from_pathtarget(subpath->pathtarget));
2892 : :
2893 : : /* Generate outer path using this subpath */
2894 : 3134 : add_path(rel, (Path *)
2895 : 3134 : create_subqueryscan_path(root, rel, subpath,
2896 : 1567 : trivial_pathtarget,
2897 : 1567 : pathkeys, required_outer));
2898 : 1567 : }
2899 : :
2900 : : /* If outer rel allows parallelism, do same for partial paths. */
2901 [ + + + + ]: 1459 : if (rel->consider_parallel && bms_is_empty(required_outer))
2902 : : {
2903 : : /* If consider_parallel is false, there should be no partial paths. */
2904 [ + + + - ]: 997 : Assert(sub_final_rel->consider_parallel ||
2905 : : sub_final_rel->partial_pathlist == NIL);
2906 : :
2907 : : /* Same for partial paths. */
2908 [ + + + + : 1004 : foreach(lc, sub_final_rel->partial_pathlist)
+ + ]
2909 : : {
2910 : 7 : Path *subpath = (Path *) lfirst(lc);
2911 : 7 : List *pathkeys;
2912 : :
2913 : : /* Convert subpath's pathkeys to outer representation */
2914 : 14 : pathkeys = convert_subquery_pathkeys(root,
2915 : 7 : rel,
2916 : 7 : subpath->pathkeys,
2917 : 7 : make_tlist_from_pathtarget(subpath->pathtarget));
2918 : :
2919 : : /* Generate outer path using this subpath */
2920 : 14 : add_partial_path(rel, (Path *)
2921 : 14 : create_subqueryscan_path(root, rel, subpath,
2922 : 7 : trivial_pathtarget,
2923 : 7 : pathkeys,
2924 : 7 : required_outer));
2925 : 7 : }
2926 : 997 : }
2927 : 1480 : }
2928 : :
2929 : : /*
2930 : : * set_function_pathlist
2931 : : * Build the (single) access path for a function RTE
2932 : : */
2933 : : static void
2934 : 3642 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2935 : : {
2936 : 3642 : Relids required_outer;
2937 : 3642 : List *pathkeys = NIL;
2938 : :
2939 : : /*
2940 : : * We don't support pushing join clauses into the quals of a function
2941 : : * scan, but it could still have required parameterization due to LATERAL
2942 : : * refs in the function expression.
2943 : : */
2944 : 3642 : required_outer = rel->lateral_relids;
2945 : :
2946 : : /*
2947 : : * The result is considered unordered unless ORDINALITY was used, in which
2948 : : * case it is ordered by the ordinal column (the last one). See if we
2949 : : * care, by checking for uses of that Var in equivalence classes.
2950 : : */
2951 [ + + ]: 3642 : if (rte->funcordinality)
2952 : : {
2953 : 126 : AttrNumber ordattno = rel->max_attr;
2954 : 126 : Var *var = NULL;
2955 : 126 : ListCell *lc;
2956 : :
2957 : : /*
2958 : : * Is there a Var for it in rel's targetlist? If not, the query did
2959 : : * not reference the ordinality column, or at least not in any way
2960 : : * that would be interesting for sorting.
2961 : : */
2962 [ + - + + : 448 : foreach(lc, rel->reltarget->exprs)
+ + ]
2963 : : {
2964 : 322 : Var *node = (Var *) lfirst(lc);
2965 : :
2966 : : /* checking varno/varlevelsup is just paranoia */
2967 [ + - ]: 322 : if (IsA(node, Var) &&
2968 [ + + ]: 322 : node->varattno == ordattno &&
2969 [ + - - + ]: 125 : node->varno == rel->relid &&
2970 : 125 : node->varlevelsup == 0)
2971 : : {
2972 : 125 : var = node;
2973 : 125 : break;
2974 : : }
2975 [ + + ]: 322 : }
2976 : :
2977 : : /*
2978 : : * Try to build pathkeys for this Var with int8 sorting. We tell
2979 : : * build_expression_pathkey not to build any new equivalence class; if
2980 : : * the Var isn't already mentioned in some EC, it means that nothing
2981 : : * cares about the ordering.
2982 : : */
2983 [ + + ]: 126 : if (var)
2984 : 250 : pathkeys = build_expression_pathkey(root,
2985 : 125 : (Expr *) var,
2986 : : Int8LessOperator,
2987 : 125 : rel->relids,
2988 : : false);
2989 : 126 : }
2990 : :
2991 : : /* Generate appropriate path */
2992 : 7284 : add_path(rel, create_functionscan_path(root, rel,
2993 : 3642 : pathkeys, required_outer));
2994 : 3642 : }
2995 : :
2996 : : /*
2997 : : * set_values_pathlist
2998 : : * Build the (single) access path for a VALUES RTE
2999 : : */
3000 : : static void
3001 : 1114 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3002 : : {
3003 : 1114 : Relids required_outer;
3004 : :
3005 : : /*
3006 : : * We don't support pushing join clauses into the quals of a values scan,
3007 : : * but it could still have required parameterization due to LATERAL refs
3008 : : * in the values expressions.
3009 : : */
3010 : 1114 : required_outer = rel->lateral_relids;
3011 : :
3012 : : /* Generate appropriate path */
3013 : 1114 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
3014 : 1114 : }
3015 : :
3016 : : /*
3017 : : * set_tablefunc_pathlist
3018 : : * Build the (single) access path for a table func RTE
3019 : : */
3020 : : static void
3021 : 103 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3022 : : {
3023 : 103 : Relids required_outer;
3024 : :
3025 : : /*
3026 : : * We don't support pushing join clauses into the quals of a tablefunc
3027 : : * scan, but it could still have required parameterization due to LATERAL
3028 : : * refs in the function expression.
3029 : : */
3030 : 103 : required_outer = rel->lateral_relids;
3031 : :
3032 : : /* Generate appropriate path */
3033 : 206 : add_path(rel, create_tablefuncscan_path(root, rel,
3034 : 103 : required_outer));
3035 : 103 : }
3036 : :
3037 : : /*
3038 : : * set_cte_pathlist
3039 : : * Build the (single) access path for a non-self-reference CTE RTE
3040 : : *
3041 : : * There's no need for a separate set_cte_size phase, since we don't
3042 : : * support join-qual-parameterized paths for CTEs.
3043 : : */
3044 : : static void
3045 : 212 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3046 : : {
3047 : 212 : Path *ctepath;
3048 : 212 : Plan *cteplan;
3049 : 212 : PlannerInfo *cteroot;
3050 : 212 : Index levelsup;
3051 : 212 : List *pathkeys;
3052 : 212 : int ndx;
3053 : 212 : ListCell *lc;
3054 : 212 : int plan_id;
3055 : 212 : Relids required_outer;
3056 : :
3057 : : /*
3058 : : * Find the referenced CTE, and locate the path and plan previously made
3059 : : * for it.
3060 : : */
3061 : 212 : levelsup = rte->ctelevelsup;
3062 : 212 : cteroot = root;
3063 [ + + ]: 280 : while (levelsup-- > 0)
3064 : : {
3065 : 68 : cteroot = cteroot->parent_root;
3066 [ + - ]: 68 : if (!cteroot) /* shouldn't happen */
3067 [ # # # # ]: 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3068 : : }
3069 : :
3070 : : /*
3071 : : * Note: cte_plan_ids can be shorter than cteList, if we are still working
3072 : : * on planning the CTEs (ie, this is a side-reference from another CTE).
3073 : : * So we mustn't use forboth here.
3074 : : */
3075 : 212 : ndx = 0;
3076 [ + - - + : 446 : foreach(lc, cteroot->parse->cteList)
+ - ]
3077 : : {
3078 : 234 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
3079 : :
3080 [ + + ]: 234 : if (strcmp(cte->ctename, rte->ctename) == 0)
3081 : 212 : break;
3082 : 22 : ndx++;
3083 [ + + ]: 234 : }
3084 [ + - ]: 212 : if (lc == NULL) /* shouldn't happen */
3085 [ # # # # ]: 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
3086 [ + - ]: 212 : if (ndx >= list_length(cteroot->cte_plan_ids))
3087 [ # # # # ]: 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
3088 : 212 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
3089 [ + - ]: 212 : if (plan_id <= 0)
3090 [ # # # # ]: 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
3091 : :
3092 [ + - ]: 212 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
3093 : 212 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
3094 : 212 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
3095 : :
3096 : : /* Mark rel with estimated output rows, width, etc */
3097 : 212 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
3098 : :
3099 : : /* Convert the ctepath's pathkeys to outer query's representation */
3100 : 424 : pathkeys = convert_subquery_pathkeys(root,
3101 : 212 : rel,
3102 : 212 : ctepath->pathkeys,
3103 : 212 : cteplan->targetlist);
3104 : :
3105 : : /*
3106 : : * We don't support pushing join clauses into the quals of a CTE scan, but
3107 : : * it could still have required parameterization due to LATERAL refs in
3108 : : * its tlist.
3109 : : */
3110 : 212 : required_outer = rel->lateral_relids;
3111 : :
3112 : : /* Generate appropriate path */
3113 : 212 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
3114 : 212 : }
3115 : :
3116 : : /*
3117 : : * set_namedtuplestore_pathlist
3118 : : * Build the (single) access path for a named tuplestore RTE
3119 : : *
3120 : : * There's no need for a separate set_namedtuplestore_size phase, since we
3121 : : * don't support join-qual-parameterized paths for tuplestores.
3122 : : */
3123 : : static void
3124 : 77 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
3125 : : RangeTblEntry *rte)
3126 : : {
3127 : 77 : Relids required_outer;
3128 : :
3129 : : /* Mark rel with estimated output rows, width, etc */
3130 : 77 : set_namedtuplestore_size_estimates(root, rel);
3131 : :
3132 : : /*
3133 : : * We don't support pushing join clauses into the quals of a tuplestore
3134 : : * scan, but it could still have required parameterization due to LATERAL
3135 : : * refs in its tlist.
3136 : : */
3137 : 77 : required_outer = rel->lateral_relids;
3138 : :
3139 : : /* Generate appropriate path */
3140 : 77 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
3141 : 77 : }
3142 : :
3143 : : /*
3144 : : * set_result_pathlist
3145 : : * Build the (single) access path for an RTE_RESULT RTE
3146 : : *
3147 : : * There's no need for a separate set_result_size phase, since we
3148 : : * don't support join-qual-parameterized paths for these RTEs.
3149 : : */
3150 : : static void
3151 : 676 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
3152 : : RangeTblEntry *rte)
3153 : : {
3154 : 676 : Relids required_outer;
3155 : :
3156 : : /* Mark rel with estimated output rows, width, etc */
3157 : 676 : set_result_size_estimates(root, rel);
3158 : :
3159 : : /*
3160 : : * We don't support pushing join clauses into the quals of a Result scan,
3161 : : * but it could still have required parameterization due to LATERAL refs
3162 : : * in its tlist.
3163 : : */
3164 : 676 : required_outer = rel->lateral_relids;
3165 : :
3166 : : /* Generate appropriate path */
3167 : 676 : add_path(rel, create_resultscan_path(root, rel, required_outer));
3168 : 676 : }
3169 : :
3170 : : /*
3171 : : * set_worktable_pathlist
3172 : : * Build the (single) access path for a self-reference CTE RTE
3173 : : *
3174 : : * There's no need for a separate set_worktable_size phase, since we don't
3175 : : * support join-qual-parameterized paths for CTEs.
3176 : : */
3177 : : static void
3178 : 74 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3179 : : {
3180 : 74 : Path *ctepath;
3181 : 74 : PlannerInfo *cteroot;
3182 : 74 : Index levelsup;
3183 : 74 : Relids required_outer;
3184 : :
3185 : : /*
3186 : : * We need to find the non-recursive term's path, which is in the plan
3187 : : * level that's processing the recursive UNION, which is one level *below*
3188 : : * where the CTE comes from.
3189 : : */
3190 : 74 : levelsup = rte->ctelevelsup;
3191 [ + - ]: 74 : if (levelsup == 0) /* shouldn't happen */
3192 [ # # # # ]: 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3193 : 74 : levelsup--;
3194 : 74 : cteroot = root;
3195 [ + + ]: 158 : while (levelsup-- > 0)
3196 : : {
3197 : 84 : cteroot = cteroot->parent_root;
3198 [ + - ]: 84 : if (!cteroot) /* shouldn't happen */
3199 [ # # # # ]: 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3200 : : }
3201 : 74 : ctepath = cteroot->non_recursive_path;
3202 [ + - ]: 74 : if (!ctepath) /* shouldn't happen */
3203 [ # # # # ]: 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3204 : :
3205 : : /* Mark rel with estimated output rows, width, etc */
3206 : 74 : set_cte_size_estimates(root, rel, ctepath->rows);
3207 : :
3208 : : /*
3209 : : * We don't support pushing join clauses into the quals of a worktable
3210 : : * scan, but it could still have required parameterization due to LATERAL
3211 : : * refs in its tlist. (I'm not sure this is actually possible given the
3212 : : * restrictions on recursive references, but it's easy enough to support.)
3213 : : */
3214 : 74 : required_outer = rel->lateral_relids;
3215 : :
3216 : : /* Generate appropriate path */
3217 : 74 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
3218 : 74 : }
3219 : :
3220 : : /*
3221 : : * generate_gather_paths
3222 : : * Generate parallel access paths for a relation by pushing a Gather or
3223 : : * Gather Merge on top of a partial path.
3224 : : *
3225 : : * This must not be called until after we're done creating all partial paths
3226 : : * for the specified relation. (Otherwise, add_partial_path might delete a
3227 : : * path that some GatherPath or GatherMergePath has a reference to.)
3228 : : *
3229 : : * If we're generating paths for a scan or join relation, override_rows will
3230 : : * be false, and we'll just use the relation's size estimate. When we're
3231 : : * being called for a partially-grouped or partially-distinct path, though, we
3232 : : * need to override the rowcount estimate. (It's not clear that the
3233 : : * particular value we're using here is actually best, but the underlying rel
3234 : : * has no estimate so we must do something.)
3235 : : */
3236 : : void
3237 : 3644 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3238 : : {
3239 : 3644 : Path *cheapest_partial_path;
3240 : 3644 : Path *simple_gather_path;
3241 : 3644 : ListCell *lc;
3242 : 3644 : double rows;
3243 : 3644 : double *rowsp = NULL;
3244 : :
3245 : : /* If there are no partial paths, there's nothing to do here. */
3246 [ + - ]: 3644 : if (rel->partial_pathlist == NIL)
3247 : 0 : return;
3248 : :
3249 : : /* Should we override the rel's rowcount estimate? */
3250 [ + + ]: 3644 : if (override_rows)
3251 : 1027 : rowsp = &rows;
3252 : :
3253 : : /*
3254 : : * The output of Gather is always unsorted, so there's only one partial
3255 : : * path of interest: the cheapest one. That will be the one at the front
3256 : : * of partial_pathlist because of the way add_partial_path works.
3257 : : */
3258 : 3644 : cheapest_partial_path = linitial(rel->partial_pathlist);
3259 : 3644 : rows = compute_gather_rows(cheapest_partial_path);
3260 : 3644 : simple_gather_path = (Path *)
3261 : 7288 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3262 : 3644 : NULL, rowsp);
3263 : 3644 : add_path(rel, simple_gather_path);
3264 : :
3265 : : /*
3266 : : * For each useful ordering, we can consider an order-preserving Gather
3267 : : * Merge.
3268 : : */
3269 [ + - + + : 8304 : foreach(lc, rel->partial_pathlist)
+ + ]
3270 : : {
3271 : 4660 : Path *subpath = (Path *) lfirst(lc);
3272 : 4660 : GatherMergePath *path;
3273 : :
3274 [ + + ]: 4660 : if (subpath->pathkeys == NIL)
3275 : 3534 : continue;
3276 : :
3277 : 1126 : rows = compute_gather_rows(subpath);
3278 : 2252 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3279 : 1126 : subpath->pathkeys, NULL, rowsp);
3280 : 1126 : add_path(rel, &path->path);
3281 [ + + ]: 4660 : }
3282 : 3644 : }
3283 : :
3284 : : /*
3285 : : * get_useful_pathkeys_for_relation
3286 : : * Determine which orderings of a relation might be useful.
3287 : : *
3288 : : * Getting data in sorted order can be useful either because the requested
3289 : : * order matches the final output ordering for the overall query we're
3290 : : * planning, or because it enables an efficient merge join. Here, we try
3291 : : * to figure out which pathkeys to consider.
3292 : : *
3293 : : * This allows us to do incremental sort on top of an index scan under a gather
3294 : : * merge node, i.e. parallelized.
3295 : : *
3296 : : * If the require_parallel_safe is true, we also require the expressions to
3297 : : * be parallel safe (which allows pushing the sort below Gather Merge).
3298 : : *
3299 : : * XXX At the moment this can only ever return a list with a single element,
3300 : : * because it looks at query_pathkeys only. So we might return the pathkeys
3301 : : * directly, but it seems plausible we'll want to consider other orderings
3302 : : * in the future. For example, we might want to consider pathkeys useful for
3303 : : * merge joins.
3304 : : */
3305 : : static List *
3306 : 3644 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3307 : : bool require_parallel_safe)
3308 : : {
3309 : 3644 : List *useful_pathkeys_list = NIL;
3310 : :
3311 : : /*
3312 : : * Considering query_pathkeys is always worth it, because it might allow
3313 : : * us to avoid a total sort when we have a partially presorted path
3314 : : * available or to push the total sort into the parallel portion of the
3315 : : * query.
3316 : : */
3317 [ + + ]: 3644 : if (root->query_pathkeys)
3318 : : {
3319 : 2464 : ListCell *lc;
3320 : 2464 : int npathkeys = 0; /* useful pathkeys */
3321 : :
3322 [ + - + + : 5473 : foreach(lc, root->query_pathkeys)
+ + ]
3323 : : {
3324 : 3009 : PathKey *pathkey = (PathKey *) lfirst(lc);
3325 : 3009 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3326 : :
3327 : : /*
3328 : : * We can only build a sort for pathkeys that contain a
3329 : : * safe-to-compute-early EC member computable from the current
3330 : : * relation's reltarget, so ignore the remainder of the list as
3331 : : * soon as we find a pathkey without such a member.
3332 : : *
3333 : : * It's still worthwhile to return any prefix of the pathkeys list
3334 : : * that meets this requirement, as we may be able to do an
3335 : : * incremental sort.
3336 : : *
3337 : : * If requested, ensure the sort expression is parallel-safe too.
3338 : : */
3339 [ + + + + ]: 6018 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3340 : 3009 : require_parallel_safe))
3341 : 1289 : break;
3342 : :
3343 : 1720 : npathkeys++;
3344 [ + + ]: 3009 : }
3345 : :
3346 : : /*
3347 : : * The whole query_pathkeys list matches, so append it directly, to
3348 : : * allow comparing pathkeys easily by comparing list pointer. If we
3349 : : * have to truncate the pathkeys, we gotta do a copy though.
3350 : : */
3351 [ + + ]: 2464 : if (npathkeys == list_length(root->query_pathkeys))
3352 : 2350 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3353 : 1175 : root->query_pathkeys);
3354 [ + + ]: 1289 : else if (npathkeys > 0)
3355 : 158 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3356 : 158 : list_copy_head(root->query_pathkeys,
3357 : 79 : npathkeys));
3358 : 2464 : }
3359 : :
3360 : 7288 : return useful_pathkeys_list;
3361 : 3644 : }
3362 : :
3363 : : /*
3364 : : * generate_useful_gather_paths
3365 : : * Generate parallel access paths for a relation by pushing a Gather or
3366 : : * Gather Merge on top of a partial path.
3367 : : *
3368 : : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3369 : : * paths (aiming to preserve the ordering), but also considers ordering that
3370 : : * might be useful for nodes above the gather merge node, and tries to add
3371 : : * a sort (regular or incremental) to provide that.
3372 : : */
3373 : : void
3374 : 56633 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3375 : : {
3376 : 56633 : ListCell *lc;
3377 : 56633 : double rows;
3378 : 56633 : double *rowsp = NULL;
3379 : 56633 : List *useful_pathkeys_list = NIL;
3380 : 56633 : Path *cheapest_partial_path = NULL;
3381 : :
3382 : : /* If there are no partial paths, there's nothing to do here. */
3383 [ + + ]: 56633 : if (rel->partial_pathlist == NIL)
3384 : 52989 : return;
3385 : :
3386 : : /* Should we override the rel's rowcount estimate? */
3387 [ + + ]: 3644 : if (override_rows)
3388 : 1027 : rowsp = &rows;
3389 : :
3390 : : /* generate the regular gather (merge) paths */
3391 : 3644 : generate_gather_paths(root, rel, override_rows);
3392 : :
3393 : : /* consider incremental sort for interesting orderings */
3394 : 3644 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3395 : :
3396 : : /* used for explicit (full) sort paths */
3397 : 3644 : cheapest_partial_path = linitial(rel->partial_pathlist);
3398 : :
3399 : : /*
3400 : : * Consider sorted paths for each interesting ordering. We generate both
3401 : : * incremental and full sort.
3402 : : */
3403 [ + + + + : 4898 : foreach(lc, useful_pathkeys_list)
+ + ]
3404 : : {
3405 : 1254 : List *useful_pathkeys = lfirst(lc);
3406 : 1254 : ListCell *lc2;
3407 : 1254 : bool is_sorted;
3408 : 1254 : int presorted_keys;
3409 : :
3410 [ + - + + : 2987 : foreach(lc2, rel->partial_pathlist)
+ + ]
3411 : : {
3412 : 1733 : Path *subpath = (Path *) lfirst(lc2);
3413 : 1733 : GatherMergePath *path;
3414 : :
3415 : 3466 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3416 : 1733 : subpath->pathkeys,
3417 : : &presorted_keys);
3418 : :
3419 : : /*
3420 : : * We don't need to consider the case where a subpath is already
3421 : : * fully sorted because generate_gather_paths already creates a
3422 : : * gather merge path for every subpath that has pathkeys present.
3423 : : *
3424 : : * But since the subpath is already sorted, we know we don't need
3425 : : * to consider adding a sort (full or incremental) on top of it,
3426 : : * so we can continue here.
3427 : : */
3428 [ + + ]: 1733 : if (is_sorted)
3429 : 505 : continue;
3430 : :
3431 : : /*
3432 : : * Try at least sorting the cheapest path and also try
3433 : : * incrementally sorting any path which is partially sorted
3434 : : * already (no need to deal with paths which have presorted keys
3435 : : * when incremental sort is disabled unless it's the cheapest
3436 : : * input path).
3437 : : */
3438 [ + + + + ]: 1275 : if (subpath != cheapest_partial_path &&
3439 [ + + ]: 63 : (presorted_keys == 0 || !enable_incremental_sort))
3440 : 17 : continue;
3441 : :
3442 : : /*
3443 : : * Consider regular sort for any path that's not presorted or if
3444 : : * incremental sort is disabled. We've no need to consider both
3445 : : * sort and incremental sort on the same path. We assume that
3446 : : * incremental sort is always faster when there are presorted
3447 : : * keys.
3448 : : *
3449 : : * This is not redundant with the gather paths created in
3450 : : * generate_gather_paths, because that doesn't generate ordered
3451 : : * output. Here we add an explicit sort to match the useful
3452 : : * ordering.
3453 : : */
3454 [ + + + + ]: 1211 : if (presorted_keys == 0 || !enable_incremental_sort)
3455 : : {
3456 : 2326 : subpath = (Path *) create_sort_path(root,
3457 : 1163 : rel,
3458 : 1163 : subpath,
3459 : 1163 : useful_pathkeys,
3460 : : -1.0);
3461 : 1163 : }
3462 : : else
3463 : 96 : subpath = (Path *) create_incremental_sort_path(root,
3464 : 48 : rel,
3465 : 48 : subpath,
3466 : 48 : useful_pathkeys,
3467 : 48 : presorted_keys,
3468 : : -1);
3469 : 1211 : rows = compute_gather_rows(subpath);
3470 : 2422 : path = create_gather_merge_path(root, rel,
3471 : 1211 : subpath,
3472 : 1211 : rel->reltarget,
3473 : 1211 : subpath->pathkeys,
3474 : : NULL,
3475 : 1211 : rowsp);
3476 : :
3477 : 1211 : add_path(rel, &path->path);
3478 [ + + ]: 1733 : }
3479 : 1254 : }
3480 : 56633 : }
3481 : :
3482 : : /*
3483 : : * generate_grouped_paths
3484 : : * Generate paths for a grouped relation by adding sorted and hashed
3485 : : * partial aggregation paths on top of paths of the ungrouped relation.
3486 : : *
3487 : : * The information needed is provided by the RelAggInfo structure stored in
3488 : : * "grouped_rel".
3489 : : */
3490 : : void
3491 : 149 : generate_grouped_paths(PlannerInfo *root, RelOptInfo *grouped_rel,
3492 : : RelOptInfo *rel)
3493 : : {
3494 : 149 : RelAggInfo *agg_info = grouped_rel->agg_info;
3495 : 149 : AggClauseCosts agg_costs;
3496 : 149 : bool can_hash;
3497 : 149 : bool can_sort;
3498 : 149 : Path *cheapest_total_path = NULL;
3499 : 149 : Path *cheapest_partial_path = NULL;
3500 : 149 : double dNumGroups = 0;
3501 : 149 : double dNumPartialGroups = 0;
3502 : 149 : List *group_pathkeys = NIL;
3503 : :
3504 [ - + ]: 149 : if (IS_DUMMY_REL(rel))
3505 : : {
3506 : 0 : mark_dummy_rel(grouped_rel);
3507 : 0 : return;
3508 : : }
3509 : :
3510 : : /*
3511 : : * We push partial aggregation only to the lowest possible level in the
3512 : : * join tree that is deemed useful.
3513 : : */
3514 [ + - - + ]: 149 : if (!bms_equal(agg_info->apply_agg_at, rel->relids) ||
3515 : 149 : !agg_info->agg_useful)
3516 : 0 : return;
3517 : :
3518 [ + - + - : 894 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
+ - - + +
+ ]
3519 : 149 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL, &agg_costs);
3520 : :
3521 : : /*
3522 : : * Determine whether it's possible to perform sort-based implementations
3523 : : * of grouping, and generate the pathkeys that represent the grouping
3524 : : * requirements in that case.
3525 : : */
3526 : 149 : can_sort = grouping_is_sortable(agg_info->group_clauses);
3527 [ - + ]: 149 : if (can_sort)
3528 : : {
3529 : 149 : RelOptInfo *top_grouped_rel;
3530 : 149 : List *top_group_tlist;
3531 : :
3532 [ + + + + : 149 : top_grouped_rel = IS_OTHER_REL(rel) ?
- + ]
3533 : 149 : rel->top_parent->grouped_rel : grouped_rel;
3534 : 149 : top_group_tlist =
3535 : 149 : make_tlist_from_pathtarget(top_grouped_rel->agg_info->target);
3536 : :
3537 : 149 : group_pathkeys =
3538 : 298 : make_pathkeys_for_sortclauses(root, agg_info->group_clauses,
3539 : 149 : top_group_tlist);
3540 : 149 : }
3541 : :
3542 : : /*
3543 : : * Determine whether we should consider hash-based implementations of
3544 : : * grouping.
3545 : : */
3546 [ + - ]: 149 : Assert(root->numOrderedAggs == 0);
3547 [ - + ]: 298 : can_hash = (agg_info->group_clauses != NIL &&
3548 : 149 : grouping_is_hashable(agg_info->group_clauses));
3549 : :
3550 : : /*
3551 : : * Consider whether we should generate partially aggregated non-partial
3552 : : * paths. We can only do this if we have a non-partial path.
3553 : : */
3554 [ - + ]: 149 : if (rel->pathlist != NIL)
3555 : : {
3556 : 149 : cheapest_total_path = rel->cheapest_total_path;
3557 [ + - ]: 149 : Assert(cheapest_total_path != NULL);
3558 : 149 : }
3559 : :
3560 : : /*
3561 : : * If parallelism is possible for grouped_rel, then we should consider
3562 : : * generating partially-grouped partial paths. However, if the ungrouped
3563 : : * rel has no partial paths, then we can't.
3564 : : */
3565 [ + + + + ]: 149 : if (grouped_rel->consider_parallel && rel->partial_pathlist != NIL)
3566 : : {
3567 : 122 : cheapest_partial_path = linitial(rel->partial_pathlist);
3568 [ + - ]: 122 : Assert(cheapest_partial_path != NULL);
3569 : 122 : }
3570 : :
3571 : : /* Estimate number of partial groups. */
3572 [ - + ]: 149 : if (cheapest_total_path != NULL)
3573 : 298 : dNumGroups = estimate_num_groups(root,
3574 : 149 : agg_info->group_exprs,
3575 : 149 : cheapest_total_path->rows,
3576 : : NULL, NULL);
3577 [ + + ]: 149 : if (cheapest_partial_path != NULL)
3578 : 244 : dNumPartialGroups = estimate_num_groups(root,
3579 : 122 : agg_info->group_exprs,
3580 : 122 : cheapest_partial_path->rows,
3581 : : NULL, NULL);
3582 : :
3583 [ + - - + ]: 149 : if (can_sort && cheapest_total_path != NULL)
3584 : : {
3585 : 149 : ListCell *lc;
3586 : :
3587 : : /*
3588 : : * Use any available suitably-sorted path as input, and also consider
3589 : : * sorting the cheapest-total path and incremental sort on any paths
3590 : : * with presorted keys.
3591 : : *
3592 : : * To save planning time, we ignore parameterized input paths unless
3593 : : * they are the cheapest-total path.
3594 : : */
3595 [ + - + + : 361 : foreach(lc, rel->pathlist)
+ + ]
3596 : : {
3597 : 212 : Path *input_path = (Path *) lfirst(lc);
3598 : 212 : Path *path;
3599 : 212 : bool is_sorted;
3600 : 212 : int presorted_keys;
3601 : :
3602 : : /*
3603 : : * Ignore parameterized paths that are not the cheapest-total
3604 : : * path.
3605 : : */
3606 [ + + - + ]: 212 : if (input_path->param_info &&
3607 : 1 : input_path != cheapest_total_path)
3608 : 1 : continue;
3609 : :
3610 : 422 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3611 : 211 : input_path->pathkeys,
3612 : : &presorted_keys);
3613 : :
3614 : : /*
3615 : : * Ignore paths that are not suitably or partially sorted, unless
3616 : : * they are the cheapest total path (no need to deal with paths
3617 : : * which have presorted keys when incremental sort is disabled).
3618 : : */
3619 [ + + + + : 235 : if (!is_sorted && input_path != cheapest_total_path &&
+ - ]
3620 [ + + ]: 28 : (presorted_keys == 0 || !enable_incremental_sort))
3621 : 4 : continue;
3622 : :
3623 : : /*
3624 : : * Since the path originates from a non-grouped relation that is
3625 : : * not aware of eager aggregation, we must ensure that it provides
3626 : : * the correct input for partial aggregation.
3627 : : */
3628 : 414 : path = (Path *) create_projection_path(root,
3629 : 207 : grouped_rel,
3630 : 207 : input_path,
3631 : 207 : agg_info->agg_input);
3632 : :
3633 [ + + ]: 207 : if (!is_sorted)
3634 : : {
3635 : : /*
3636 : : * We've no need to consider both a sort and incremental sort.
3637 : : * We'll just do a sort if there are no presorted keys and an
3638 : : * incremental sort when there are presorted keys.
3639 : : */
3640 [ + + - + ]: 172 : if (presorted_keys == 0 || !enable_incremental_sort)
3641 : 296 : path = (Path *) create_sort_path(root,
3642 : 148 : grouped_rel,
3643 : 148 : path,
3644 : 148 : group_pathkeys,
3645 : : -1.0);
3646 : : else
3647 : 48 : path = (Path *) create_incremental_sort_path(root,
3648 : 24 : grouped_rel,
3649 : 24 : path,
3650 : 24 : group_pathkeys,
3651 : 24 : presorted_keys,
3652 : : -1.0);
3653 : 172 : }
3654 : :
3655 : : /*
3656 : : * qual is NIL because the HAVING clause cannot be evaluated until
3657 : : * the final value of the aggregate is known.
3658 : : */
3659 : 414 : path = (Path *) create_agg_path(root,
3660 : 207 : grouped_rel,
3661 : 207 : path,
3662 : 207 : agg_info->target,
3663 : : AGG_SORTED,
3664 : : AGGSPLIT_INITIAL_SERIAL,
3665 : 207 : agg_info->group_clauses,
3666 : : NIL,
3667 : : &agg_costs,
3668 : 207 : dNumGroups);
3669 : :
3670 : 207 : add_path(grouped_rel, path);
3671 [ + + ]: 212 : }
3672 : 149 : }
3673 : :
3674 [ + - + + ]: 149 : if (can_sort && cheapest_partial_path != NULL)
3675 : : {
3676 : 122 : ListCell *lc;
3677 : :
3678 : : /* Similar to above logic, but for partial paths. */
3679 [ + - + + : 284 : foreach(lc, rel->partial_pathlist)
+ + ]
3680 : : {
3681 : 162 : Path *input_path = (Path *) lfirst(lc);
3682 : 162 : Path *path;
3683 : 162 : bool is_sorted;
3684 : 162 : int presorted_keys;
3685 : :
3686 : 324 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3687 : 162 : input_path->pathkeys,
3688 : : &presorted_keys);
3689 : :
3690 : : /*
3691 : : * Ignore paths that are not suitably or partially sorted, unless
3692 : : * they are the cheapest partial path (no need to deal with paths
3693 : : * which have presorted keys when incremental sort is disabled).
3694 : : */
3695 [ + + + + : 178 : if (!is_sorted && input_path != cheapest_partial_path &&
+ - ]
3696 [ + - ]: 16 : (presorted_keys == 0 || !enable_incremental_sort))
3697 : 0 : continue;
3698 : :
3699 : : /*
3700 : : * Since the path originates from a non-grouped relation that is
3701 : : * not aware of eager aggregation, we must ensure that it provides
3702 : : * the correct input for partial aggregation.
3703 : : */
3704 : 324 : path = (Path *) create_projection_path(root,
3705 : 162 : grouped_rel,
3706 : 162 : input_path,
3707 : 162 : agg_info->agg_input);
3708 : :
3709 [ + + ]: 162 : if (!is_sorted)
3710 : : {
3711 : : /*
3712 : : * We've no need to consider both a sort and incremental sort.
3713 : : * We'll just do a sort if there are no presorted keys and an
3714 : : * incremental sort when there are presorted keys.
3715 : : */
3716 [ + + - + ]: 138 : if (presorted_keys == 0 || !enable_incremental_sort)
3717 : 244 : path = (Path *) create_sort_path(root,
3718 : 122 : grouped_rel,
3719 : 122 : path,
3720 : 122 : group_pathkeys,
3721 : : -1.0);
3722 : : else
3723 : 32 : path = (Path *) create_incremental_sort_path(root,
3724 : 16 : grouped_rel,
3725 : 16 : path,
3726 : 16 : group_pathkeys,
3727 : 16 : presorted_keys,
3728 : : -1.0);
3729 : 138 : }
3730 : :
3731 : : /*
3732 : : * qual is NIL because the HAVING clause cannot be evaluated until
3733 : : * the final value of the aggregate is known.
3734 : : */
3735 : 324 : path = (Path *) create_agg_path(root,
3736 : 162 : grouped_rel,
3737 : 162 : path,
3738 : 162 : agg_info->target,
3739 : : AGG_SORTED,
3740 : : AGGSPLIT_INITIAL_SERIAL,
3741 : 162 : agg_info->group_clauses,
3742 : : NIL,
3743 : : &agg_costs,
3744 : 162 : dNumPartialGroups);
3745 : :
3746 : 162 : add_partial_path(grouped_rel, path);
3747 [ - + ]: 162 : }
3748 : 122 : }
3749 : :
3750 : : /*
3751 : : * Add a partially-grouped HashAgg Path where possible
3752 : : */
3753 [ + - - + ]: 149 : if (can_hash && cheapest_total_path != NULL)
3754 : : {
3755 : 149 : Path *path;
3756 : :
3757 : : /*
3758 : : * Since the path originates from a non-grouped relation that is not
3759 : : * aware of eager aggregation, we must ensure that it provides the
3760 : : * correct input for partial aggregation.
3761 : : */
3762 : 298 : path = (Path *) create_projection_path(root,
3763 : 149 : grouped_rel,
3764 : 149 : cheapest_total_path,
3765 : 149 : agg_info->agg_input);
3766 : :
3767 : : /*
3768 : : * qual is NIL because the HAVING clause cannot be evaluated until the
3769 : : * final value of the aggregate is known.
3770 : : */
3771 : 298 : path = (Path *) create_agg_path(root,
3772 : 149 : grouped_rel,
3773 : 149 : path,
3774 : 149 : agg_info->target,
3775 : : AGG_HASHED,
3776 : : AGGSPLIT_INITIAL_SERIAL,
3777 : 149 : agg_info->group_clauses,
3778 : : NIL,
3779 : : &agg_costs,
3780 : 149 : dNumGroups);
3781 : :
3782 : 149 : add_path(grouped_rel, path);
3783 : 149 : }
3784 : :
3785 : : /*
3786 : : * Now add a partially-grouped HashAgg partial Path where possible
3787 : : */
3788 [ + - + + ]: 149 : if (can_hash && cheapest_partial_path != NULL)
3789 : : {
3790 : 122 : Path *path;
3791 : :
3792 : : /*
3793 : : * Since the path originates from a non-grouped relation that is not
3794 : : * aware of eager aggregation, we must ensure that it provides the
3795 : : * correct input for partial aggregation.
3796 : : */
3797 : 244 : path = (Path *) create_projection_path(root,
3798 : 122 : grouped_rel,
3799 : 122 : cheapest_partial_path,
3800 : 122 : agg_info->agg_input);
3801 : :
3802 : : /*
3803 : : * qual is NIL because the HAVING clause cannot be evaluated until the
3804 : : * final value of the aggregate is known.
3805 : : */
3806 : 244 : path = (Path *) create_agg_path(root,
3807 : 122 : grouped_rel,
3808 : 122 : path,
3809 : 122 : agg_info->target,
3810 : : AGG_HASHED,
3811 : : AGGSPLIT_INITIAL_SERIAL,
3812 : 122 : agg_info->group_clauses,
3813 : : NIL,
3814 : : &agg_costs,
3815 : 122 : dNumPartialGroups);
3816 : :
3817 : 122 : add_partial_path(grouped_rel, path);
3818 : 122 : }
3819 : 149 : }
3820 : :
3821 : : /*
3822 : : * make_rel_from_joinlist
3823 : : * Build access paths using a "joinlist" to guide the join path search.
3824 : : *
3825 : : * See comments for deconstruct_jointree() for definition of the joinlist
3826 : : * data structure.
3827 : : */
3828 : : static RelOptInfo *
3829 : 34368 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3830 : : {
3831 : 34368 : int levels_needed;
3832 : 34368 : List *initial_rels;
3833 : 34368 : ListCell *jl;
3834 : :
3835 : : /*
3836 : : * Count the number of child joinlist nodes. This is the depth of the
3837 : : * dynamic-programming algorithm we must employ to consider all ways of
3838 : : * joining the child nodes.
3839 : : */
3840 : 34368 : levels_needed = list_length(joinlist);
3841 : :
3842 [ - + ]: 34368 : if (levels_needed <= 0)
3843 : 0 : return NULL; /* nothing to do? */
3844 : :
3845 : : /*
3846 : : * Construct a list of rels corresponding to the child joinlist nodes.
3847 : : * This may contain both base rels and rels constructed according to
3848 : : * sub-joinlists.
3849 : : */
3850 : 34368 : initial_rels = NIL;
3851 [ + - + + : 81269 : foreach(jl, joinlist)
+ + ]
3852 : : {
3853 : 46901 : Node *jlnode = (Node *) lfirst(jl);
3854 : 46901 : RelOptInfo *thisrel;
3855 : :
3856 [ + + ]: 46901 : if (IsA(jlnode, RangeTblRef))
3857 : : {
3858 : 46429 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3859 : :
3860 : 46429 : thisrel = find_base_rel(root, varno);
3861 : 46429 : }
3862 [ + - ]: 472 : else if (IsA(jlnode, List))
3863 : : {
3864 : : /* Recurse to handle subproblem */
3865 : 472 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3866 : 472 : }
3867 : : else
3868 : : {
3869 [ # # # # ]: 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3870 : : (int) nodeTag(jlnode));
3871 : 0 : thisrel = NULL; /* keep compiler quiet */
3872 : : }
3873 : :
3874 : 46901 : initial_rels = lappend(initial_rels, thisrel);
3875 : 46901 : }
3876 : :
3877 [ + + ]: 34368 : if (levels_needed == 1)
3878 : : {
3879 : : /*
3880 : : * Single joinlist node, so we're done.
3881 : : */
3882 : 24941 : return (RelOptInfo *) linitial(initial_rels);
3883 : : }
3884 : : else
3885 : : {
3886 : : /*
3887 : : * Consider the different orders in which we could join the rels,
3888 : : * using a plugin, GEQO, or the regular join search code.
3889 : : *
3890 : : * We put the initial_rels list into a PlannerInfo field because
3891 : : * has_legal_joinclause() needs to look at it (ugly :-().
3892 : : */
3893 : 9427 : root->initial_rels = initial_rels;
3894 : :
3895 [ - + ]: 9427 : if (join_search_hook)
3896 : 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
3897 [ + - + + ]: 9427 : else if (enable_geqo && levels_needed >= geqo_threshold)
3898 : 7 : return geqo(root, levels_needed, initial_rels);
3899 : : else
3900 : 9420 : return standard_join_search(root, levels_needed, initial_rels);
3901 : : }
3902 : 34368 : }
3903 : :
3904 : : /*
3905 : : * standard_join_search
3906 : : * Find possible joinpaths for a query by successively finding ways
3907 : : * to join component relations into join relations.
3908 : : *
3909 : : * 'levels_needed' is the number of iterations needed, ie, the number of
3910 : : * independent jointree items in the query. This is > 1.
3911 : : *
3912 : : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3913 : : * jointree item. These are the components to be joined together.
3914 : : * Note that levels_needed == list_length(initial_rels).
3915 : : *
3916 : : * Returns the final level of join relations, i.e., the relation that is
3917 : : * the result of joining all the original relations together.
3918 : : * At least one implementation path must be provided for this relation and
3919 : : * all required sub-relations.
3920 : : *
3921 : : * To support loadable plugins that modify planner behavior by changing the
3922 : : * join searching algorithm, we provide a hook variable that lets a plugin
3923 : : * replace or supplement this function. Any such hook must return the same
3924 : : * final join relation as the standard code would, but it might have a
3925 : : * different set of implementation paths attached, and only the sub-joinrels
3926 : : * needed for these paths need have been instantiated.
3927 : : *
3928 : : * Note to plugin authors: the functions invoked during standard_join_search()
3929 : : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3930 : : * than one join-order search, you'll probably need to save and restore the
3931 : : * original states of those data structures. See geqo_eval() for an example.
3932 : : */
3933 : : RelOptInfo *
3934 : 9420 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3935 : : {
3936 : 9420 : int lev;
3937 : 9420 : RelOptInfo *rel;
3938 : :
3939 : : /*
3940 : : * This function cannot be invoked recursively within any one planning
3941 : : * problem, so join_rel_level[] can't be in use already.
3942 : : */
3943 [ + - ]: 9420 : Assert(root->join_rel_level == NULL);
3944 : :
3945 : : /*
3946 : : * We employ a simple "dynamic programming" algorithm: we first find all
3947 : : * ways to build joins of two jointree items, then all ways to build joins
3948 : : * of three items (from two-item joins and single items), then four-item
3949 : : * joins, and so on until we have considered all ways to join all the
3950 : : * items into one rel.
3951 : : *
3952 : : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3953 : : * set root->join_rel_level[1] to represent all the single-jointree-item
3954 : : * relations.
3955 : : */
3956 : 9420 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3957 : :
3958 : 9420 : root->join_rel_level[1] = initial_rels;
3959 : :
3960 [ + + ]: 21943 : for (lev = 2; lev <= levels_needed; lev++)
3961 : : {
3962 : 12523 : ListCell *lc;
3963 : :
3964 : : /*
3965 : : * Determine all possible pairs of relations to be joined at this
3966 : : * level, and build paths for making each one from every available
3967 : : * pair of lower-level relations.
3968 : : */
3969 : 12523 : join_search_one_level(root, lev);
3970 : :
3971 : : /*
3972 : : * Run generate_partitionwise_join_paths() and
3973 : : * generate_useful_gather_paths() for each just-processed joinrel. We
3974 : : * could not do this earlier because both regular and partial paths
3975 : : * can get added to a particular joinrel at multiple times within
3976 : : * join_search_one_level.
3977 : : *
3978 : : * After that, we're done creating paths for the joinrel, so run
3979 : : * set_cheapest().
3980 : : *
3981 : : * In addition, we also run generate_grouped_paths() for the grouped
3982 : : * relation of each just-processed joinrel, and run set_cheapest() for
3983 : : * the grouped relation afterwards.
3984 : : */
3985 [ + + + + : 30192 : foreach(lc, root->join_rel_level[lev])
+ + ]
3986 : : {
3987 : 17669 : bool is_top_rel;
3988 : :
3989 : 17669 : rel = (RelOptInfo *) lfirst(lc);
3990 : :
3991 : 17669 : is_top_rel = bms_equal(rel->relids, root->all_query_rels);
3992 : :
3993 : : /* Create paths for partitionwise joins. */
3994 : 17669 : generate_partitionwise_join_paths(root, rel);
3995 : :
3996 : : /*
3997 : : * Except for the topmost scan/join rel, consider gathering
3998 : : * partial paths. We'll do the same for the topmost scan/join rel
3999 : : * once we know the final targetlist (see grouping_planner's and
4000 : : * its call to apply_scanjoin_target_to_paths).
4001 : : */
4002 [ + + ]: 17669 : if (!is_top_rel)
4003 : 8327 : generate_useful_gather_paths(root, rel, false);
4004 : :
4005 : : /* Find and save the cheapest paths for this rel */
4006 : 17669 : set_cheapest(rel);
4007 : :
4008 : : /*
4009 : : * Except for the topmost scan/join rel, consider generating
4010 : : * partial aggregation paths for the grouped relation on top of
4011 : : * the paths of this rel. After that, we're done creating paths
4012 : : * for the grouped relation, so run set_cheapest().
4013 : : */
4014 [ + + + + ]: 17669 : if (rel->grouped_rel != NULL && !is_top_rel)
4015 : : {
4016 : 12 : RelOptInfo *grouped_rel = rel->grouped_rel;
4017 : :
4018 [ - + ]: 12 : Assert(IS_GROUPED_REL(grouped_rel));
4019 : :
4020 : 12 : generate_grouped_paths(root, grouped_rel, rel);
4021 : 12 : set_cheapest(grouped_rel);
4022 : 12 : }
4023 : :
4024 : : #ifdef OPTIMIZER_DEBUG
4025 : : pprint(rel);
4026 : : #endif
4027 : 17669 : }
4028 : 12523 : }
4029 : :
4030 : : /*
4031 : : * We should have a single rel at the final level.
4032 : : */
4033 [ + - ]: 9420 : if (root->join_rel_level[levels_needed] == NIL)
4034 [ # # # # ]: 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
4035 [ + - ]: 9420 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
4036 : :
4037 : 9420 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
4038 : :
4039 : 9420 : root->join_rel_level = NULL;
4040 : :
4041 : 18840 : return rel;
4042 : 9420 : }
4043 : :
4044 : : /*****************************************************************************
4045 : : * PUSHING QUALS DOWN INTO SUBQUERIES
4046 : : *****************************************************************************/
4047 : :
4048 : : /*
4049 : : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
4050 : : *
4051 : : * subquery is the particular component query being checked. topquery
4052 : : * is the top component of a set-operations tree (the same Query if no
4053 : : * set-op is involved).
4054 : : *
4055 : : * Conditions checked here:
4056 : : *
4057 : : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
4058 : : * since that could change the set of rows returned.
4059 : : *
4060 : : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
4061 : : * quals into it, because that could change the results.
4062 : : *
4063 : : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
4064 : : * This is because upper-level quals should semantically be evaluated only
4065 : : * once per distinct row, not once per original row, and if the qual is
4066 : : * volatile then extra evaluations could change the results. (This issue
4067 : : * does not apply to other forms of aggregation such as GROUP BY, because
4068 : : * when those are present we push into HAVING not WHERE, so that the quals
4069 : : * are still applied after aggregation.)
4070 : : *
4071 : : * 4. If the subquery contains window functions, we cannot push volatile quals
4072 : : * into it. The issue here is a bit different from DISTINCT: a volatile qual
4073 : : * might succeed for some rows of a window partition and fail for others,
4074 : : * thereby changing the partition contents and thus the window functions'
4075 : : * results for rows that remain.
4076 : : *
4077 : : * 5. If the subquery contains any set-returning functions in its targetlist,
4078 : : * we cannot push volatile quals into it. That would push them below the SRFs
4079 : : * and thereby change the number of times they are evaluated. Also, a
4080 : : * volatile qual could succeed for some SRF output rows and fail for others,
4081 : : * a behavior that cannot occur if it's evaluated before SRF expansion.
4082 : : *
4083 : : * 6. If the subquery has nonempty grouping sets, we cannot push down any
4084 : : * quals. The concern here is that a qual referencing a "constant" grouping
4085 : : * column could get constant-folded, which would be improper because the value
4086 : : * is potentially nullable by grouping-set expansion. This restriction could
4087 : : * be removed if we had a parsetree representation that shows that such
4088 : : * grouping columns are not really constant. (There are other ideas that
4089 : : * could be used to relax this restriction, but that's the approach most
4090 : : * likely to get taken in the future. Note that there's not much to be gained
4091 : : * so long as subquery_planner can't move HAVING clauses to WHERE within such
4092 : : * a subquery.)
4093 : : *
4094 : : * In addition, we make several checks on the subquery's output columns to see
4095 : : * if it is safe to reference them in pushed-down quals. If output column k
4096 : : * is found to be unsafe to reference, we set the reason for that inside
4097 : : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
4098 : : * column k might not be referenced by some/all quals. The unsafeFlags[]
4099 : : * array will be consulted later by qual_is_pushdown_safe(). It's better to
4100 : : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
4101 : : * because when the subquery involves set operations we have to check the
4102 : : * output expressions in each arm of the set op.
4103 : : *
4104 : : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
4105 : : * we're effectively assuming that the quals cannot distinguish values that
4106 : : * the DISTINCT's equality operator sees as equal, yet there are many
4107 : : * counterexamples to that assumption. However use of such a qual with a
4108 : : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
4109 : : * "equal" value will be chosen as the output value by the DISTINCT operation.
4110 : : * So we don't worry too much about that. Another objection is that if the
4111 : : * qual is expensive to evaluate, running it for each original row might cost
4112 : : * more than we save by eliminating rows before the DISTINCT step. But it
4113 : : * would be very hard to estimate that at this stage, and in practice pushdown
4114 : : * seldom seems to make things worse, so we ignore that problem too.
4115 : : *
4116 : : * Note: likewise, pushing quals into a subquery with window functions is a
4117 : : * bit dubious: the quals might remove some rows of a window partition while
4118 : : * leaving others, causing changes in the window functions' results for the
4119 : : * surviving rows. We insist that such a qual reference only partitioning
4120 : : * columns, but again that only protects us if the qual does not distinguish
4121 : : * values that the partitioning equality operator sees as equal. The risks
4122 : : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
4123 : : * occurs and thus there is no theoretical problem with such a qual. But
4124 : : * we'll do this anyway because the potential performance benefits are very
4125 : : * large, and we've seen no field complaints about the longstanding comparable
4126 : : * behavior with DISTINCT.
4127 : : */
4128 : : static bool
4129 : 318 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
4130 : : pushdown_safety_info *safetyInfo)
4131 : : {
4132 : 318 : SetOperationStmt *topop;
4133 : :
4134 : : /* Check point 1 */
4135 [ + + + + ]: 318 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
4136 : 22 : return false;
4137 : :
4138 : : /* Check point 6 */
4139 [ + + + + ]: 296 : if (subquery->groupClause && subquery->groupingSets)
4140 : 2 : return false;
4141 : :
4142 : : /* Check points 3, 4, and 5 */
4143 [ + + ]: 294 : if (subquery->distinctClause ||
4144 [ + + + + ]: 280 : subquery->hasWindowFuncs ||
4145 : 236 : subquery->hasTargetSRFs)
4146 : 94 : safetyInfo->unsafeVolatile = true;
4147 : :
4148 : : /*
4149 : : * If we're at a leaf query, check for unsafe expressions in its target
4150 : : * list, and mark any reasons why they're unsafe in unsafeFlags[].
4151 : : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
4152 : : * so no need to check them.)
4153 : : */
4154 [ + + ]: 294 : if (subquery->setOperations == NULL)
4155 : 280 : check_output_expressions(subquery, safetyInfo);
4156 : :
4157 : : /* Are we at top level, or looking at a setop component? */
4158 [ + + ]: 294 : if (subquery == topquery)
4159 : : {
4160 : : /* Top level, so check any component queries */
4161 [ + + ]: 266 : if (subquery->setOperations != NULL)
4162 [ + - + - ]: 28 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
4163 : 14 : safetyInfo))
4164 : 0 : return false;
4165 : 266 : }
4166 : : else
4167 : : {
4168 : : /* Setop component must not have more components (too weird) */
4169 [ - + ]: 28 : if (subquery->setOperations != NULL)
4170 : 0 : return false;
4171 : : /* Check whether setop component output types match top level */
4172 : 28 : topop = castNode(SetOperationStmt, topquery->setOperations);
4173 [ + - ]: 28 : Assert(topop);
4174 : 56 : compare_tlist_datatypes(subquery->targetList,
4175 : 28 : topop->colTypes,
4176 : 28 : safetyInfo);
4177 : : }
4178 : 294 : return true;
4179 : 318 : }
4180 : :
4181 : : /*
4182 : : * Helper routine to recurse through setOperations tree
4183 : : */
4184 : : static bool
4185 : 42 : recurse_pushdown_safe(Node *setOp, Query *topquery,
4186 : : pushdown_safety_info *safetyInfo)
4187 : : {
4188 [ + + ]: 42 : if (IsA(setOp, RangeTblRef))
4189 : : {
4190 : 28 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4191 : 28 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
4192 : 28 : Query *subquery = rte->subquery;
4193 : :
4194 [ + - ]: 28 : Assert(subquery != NULL);
4195 : 28 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
4196 : 28 : }
4197 [ + - ]: 14 : else if (IsA(setOp, SetOperationStmt))
4198 : : {
4199 : 14 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4200 : :
4201 : : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
4202 [ - + ]: 14 : if (op->op == SETOP_EXCEPT)
4203 : 0 : return false;
4204 : : /* Else recurse */
4205 [ + - ]: 14 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
4206 : 0 : return false;
4207 [ + - ]: 14 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
4208 : 0 : return false;
4209 [ - - + ]: 14 : }
4210 : : else
4211 : : {
4212 [ # # # # ]: 0 : elog(ERROR, "unrecognized node type: %d",
4213 : : (int) nodeTag(setOp));
4214 : : }
4215 : 14 : return true;
4216 : 42 : }
4217 : :
4218 : : /*
4219 : : * check_output_expressions - check subquery's output expressions for safety
4220 : : *
4221 : : * There are several cases in which it's unsafe to push down an upper-level
4222 : : * qual if it references a particular output column of a subquery. We check
4223 : : * each output column of the subquery and set flags in unsafeFlags[k] when we
4224 : : * see that column is unsafe for a pushed-down qual to reference. The
4225 : : * conditions checked here are:
4226 : : *
4227 : : * 1. We must not push down any quals that refer to subselect outputs that
4228 : : * return sets, else we'd introduce functions-returning-sets into the
4229 : : * subquery's WHERE/HAVING quals.
4230 : : *
4231 : : * 2. We must not push down any quals that refer to subselect outputs that
4232 : : * contain volatile functions, for fear of introducing strange results due
4233 : : * to multiple evaluation of a volatile function.
4234 : : *
4235 : : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
4236 : : * refer to non-DISTINCT output columns, because that could change the set
4237 : : * of rows returned. (This condition is vacuous for DISTINCT, because then
4238 : : * there are no non-DISTINCT output columns, so we needn't check. Note that
4239 : : * subquery_is_pushdown_safe already reported that we can't use volatile
4240 : : * quals if there's DISTINCT or DISTINCT ON.)
4241 : : *
4242 : : * 4. If the subquery has any window functions, we must not push down quals
4243 : : * that reference any output columns that are not listed in all the subquery's
4244 : : * window PARTITION BY clauses. We can push down quals that use only
4245 : : * partitioning columns because they should succeed or fail identically for
4246 : : * every row of any one window partition, and totally excluding some
4247 : : * partitions will not change a window function's results for remaining
4248 : : * partitions. (Again, this also requires nonvolatile quals, but
4249 : : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
4250 : : * unsafe for this reason can still have WindowClause run conditions pushed
4251 : : * down.
4252 : : */
4253 : : static void
4254 : 280 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
4255 : : {
4256 : 280 : List *flattened_targetList = subquery->targetList;
4257 : 280 : ListCell *lc;
4258 : :
4259 : : /*
4260 : : * We must be careful with grouping Vars and join alias Vars in the
4261 : : * subquery's outputs, as they hide the underlying expressions.
4262 : : *
4263 : : * We need to expand grouping Vars to their underlying expressions (the
4264 : : * grouping clauses) because the grouping expressions themselves might be
4265 : : * volatile or set-returning. However, we do not need to expand join
4266 : : * alias Vars, as their underlying structure does not introduce volatile
4267 : : * or set-returning functions at the current level.
4268 : : *
4269 : : * In neither case do we need to recursively examine the Vars contained in
4270 : : * these underlying expressions. Even if they reference outputs from
4271 : : * lower-level subqueries (at any depth), those references are guaranteed
4272 : : * not to expand to volatile or set-returning functions, because
4273 : : * subqueries containing such functions in their targetlists are never
4274 : : * pulled up.
4275 : : */
4276 [ + + ]: 280 : if (subquery->hasGroupRTE)
4277 : : {
4278 : 46 : flattened_targetList = (List *)
4279 : 46 : flatten_group_exprs(NULL, subquery, (Node *) subquery->targetList);
4280 : 46 : }
4281 : :
4282 [ + - + + : 2986 : foreach(lc, flattened_targetList)
+ + ]
4283 : : {
4284 : 2706 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4285 : :
4286 [ + + ]: 2706 : if (tle->resjunk)
4287 : 21 : continue; /* ignore resjunk columns */
4288 : :
4289 : : /* Functions returning sets are unsafe (point 1) */
4290 [ + + ]: 2685 : if (subquery->hasTargetSRFs &&
4291 : 118 : (safetyInfo->unsafeFlags[tle->resno] &
4292 [ + - + + ]: 118 : UNSAFE_HAS_SET_FUNC) == 0 &&
4293 : 118 : expression_returns_set((Node *) tle->expr))
4294 : : {
4295 : 68 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
4296 : 68 : continue;
4297 : : }
4298 : :
4299 : : /* Volatile functions are unsafe (point 2) */
4300 : 2617 : if ((safetyInfo->unsafeFlags[tle->resno] &
4301 [ + + + + ]: 2617 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
4302 : 2615 : contain_volatile_functions((Node *) tle->expr))
4303 : : {
4304 : 15 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
4305 : 15 : continue;
4306 : : }
4307 : :
4308 : : /* If subquery uses DISTINCT ON, check point 3 */
4309 [ - + ]: 2602 : if (subquery->hasDistinctOn &&
4310 : 0 : (safetyInfo->unsafeFlags[tle->resno] &
4311 [ # # # # ]: 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4312 : 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
4313 : : {
4314 : : /* non-DISTINCT column, so mark it unsafe */
4315 : 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
4316 : 0 : continue;
4317 : : }
4318 : :
4319 : : /* If subquery uses window functions, check point 4 */
4320 [ + + ]: 2602 : if (subquery->hasWindowFuncs &&
4321 : 176 : (safetyInfo->unsafeFlags[tle->resno] &
4322 [ + - + + ]: 176 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4323 : 176 : !targetIsInAllPartitionLists(tle, subquery))
4324 : : {
4325 : : /* not present in all PARTITION BY clauses, so mark it unsafe */
4326 : 162 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
4327 : 162 : continue;
4328 : : }
4329 [ - + + ]: 2706 : }
4330 : 280 : }
4331 : :
4332 : : /*
4333 : : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
4334 : : * push quals into each component query, but the quals can only reference
4335 : : * subquery columns that suffer no type coercions in the set operation.
4336 : : * Otherwise there are possible semantic gotchas. So, we check the
4337 : : * component queries to see if any of them have output types different from
4338 : : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
4339 : : * unsafeFlags[k] if column k has different type in any component.
4340 : : *
4341 : : * We don't have to care about typmods here: the only allowed difference
4342 : : * between set-op input and output typmods is input is a specific typmod
4343 : : * and output is -1, and that does not require a coercion.
4344 : : *
4345 : : * tlist is a subquery tlist.
4346 : : * colTypes is an OID list of the top-level setop's output column types.
4347 : : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
4348 : : */
4349 : : static void
4350 : 28 : compare_tlist_datatypes(List *tlist, List *colTypes,
4351 : : pushdown_safety_info *safetyInfo)
4352 : : {
4353 : 28 : ListCell *l;
4354 : 28 : ListCell *colType = list_head(colTypes);
4355 : :
4356 [ + - + + : 84 : foreach(l, tlist)
+ + ]
4357 : : {
4358 : 56 : TargetEntry *tle = (TargetEntry *) lfirst(l);
4359 : :
4360 [ - + ]: 56 : if (tle->resjunk)
4361 : 0 : continue; /* ignore resjunk columns */
4362 [ + - ]: 56 : if (colType == NULL)
4363 [ # # # # ]: 0 : elog(ERROR, "wrong number of tlist entries");
4364 [ + + ]: 56 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
4365 : 6 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
4366 : 56 : colType = lnext(colTypes, colType);
4367 [ - - + ]: 56 : }
4368 [ + - ]: 28 : if (colType != NULL)
4369 [ # # # # ]: 0 : elog(ERROR, "wrong number of tlist entries");
4370 : 28 : }
4371 : :
4372 : : /*
4373 : : * targetIsInAllPartitionLists
4374 : : * True if the TargetEntry is listed in the PARTITION BY clause
4375 : : * of every window defined in the query.
4376 : : *
4377 : : * It would be safe to ignore windows not actually used by any window
4378 : : * function, but it's not easy to get that info at this stage; and it's
4379 : : * unlikely to be useful to spend any extra cycles getting it, since
4380 : : * unreferenced window definitions are probably infrequent in practice.
4381 : : */
4382 : : static bool
4383 : 176 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
4384 : : {
4385 : 176 : ListCell *lc;
4386 : :
4387 [ + - + + : 356 : foreach(lc, query->windowClause)
+ + + + ]
4388 : : {
4389 : 180 : WindowClause *wc = (WindowClause *) lfirst(lc);
4390 : :
4391 [ + + ]: 180 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
4392 : 162 : return false;
4393 [ + + ]: 180 : }
4394 : 14 : return true;
4395 : 176 : }
4396 : :
4397 : : /*
4398 : : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
4399 : : *
4400 : : * rinfo is a restriction clause applying to the given subquery (whose RTE
4401 : : * has index rti in the parent query).
4402 : : *
4403 : : * Conditions checked here:
4404 : : *
4405 : : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
4406 : : * unclear that it will work correctly: SubLinks will already have been
4407 : : * transformed into SubPlans in the qual, but not in the subquery). Note that
4408 : : * SubLinks that transform to initplans are safe, and will be accepted here
4409 : : * because what we'll see in the qual is just a Param referencing the initplan
4410 : : * output.
4411 : : *
4412 : : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
4413 : : * functions.
4414 : : *
4415 : : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
4416 : : * functions that are passed Var nodes, and therefore might reveal values from
4417 : : * the subquery as side effects.
4418 : : *
4419 : : * 4. rinfo's clause must not refer to the whole-row output of the subquery
4420 : : * (since there is no easy way to name that within the subquery itself).
4421 : : *
4422 : : * 5. rinfo's clause must not refer to any subquery output columns that were
4423 : : * found to be unsafe to reference by subquery_is_pushdown_safe().
4424 : : */
4425 : : static pushdown_safe_type
4426 : 433 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
4427 : : pushdown_safety_info *safetyInfo)
4428 : : {
4429 : 433 : pushdown_safe_type safe = PUSHDOWN_SAFE;
4430 : 433 : Node *qual = (Node *) rinfo->clause;
4431 : 433 : List *vars;
4432 : 433 : ListCell *vl;
4433 : :
4434 : : /* Refuse subselects (point 1) */
4435 [ + + ]: 433 : if (contain_subplans(qual))
4436 : 11 : return PUSHDOWN_UNSAFE;
4437 : :
4438 : : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
4439 [ + + + + ]: 422 : if (safetyInfo->unsafeVolatile &&
4440 : 111 : contain_volatile_functions((Node *) rinfo))
4441 : 3 : return PUSHDOWN_UNSAFE;
4442 : :
4443 : : /* Refuse leaky quals if told to (point 3) */
4444 [ + + + + ]: 419 : if (safetyInfo->unsafeLeaky &&
4445 : 181 : contain_leaked_vars(qual))
4446 : 27 : return PUSHDOWN_UNSAFE;
4447 : :
4448 : : /*
4449 : : * Examine all Vars used in clause. Since it's a restriction clause, all
4450 : : * such Vars must refer to subselect output columns ... unless this is
4451 : : * part of a LATERAL subquery, in which case there could be lateral
4452 : : * references.
4453 : : *
4454 : : * By omitting the relevant flags, this also gives us a cheap sanity check
4455 : : * that no aggregates or window functions appear in the qual. Those would
4456 : : * be unsafe to push down, but at least for the moment we could never see
4457 : : * any in a qual anyhow.
4458 : : */
4459 : 392 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
4460 [ + + + + : 802 : foreach(vl, vars)
+ + ]
4461 : : {
4462 : 410 : Var *var = (Var *) lfirst(vl);
4463 : :
4464 : : /*
4465 : : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
4466 : : * It's not clear whether a PHV could safely be pushed down, and even
4467 : : * less clear whether such a situation could arise in any cases of
4468 : : * practical interest anyway. So for the moment, just refuse to push
4469 : : * down.
4470 : : */
4471 [ + - ]: 410 : if (!IsA(var, Var))
4472 : : {
4473 : 0 : safe = PUSHDOWN_UNSAFE;
4474 : 0 : break;
4475 : : }
4476 : :
4477 : : /*
4478 : : * Punt if we find any lateral references. It would be safe to push
4479 : : * these down, but we'd have to convert them into outer references,
4480 : : * which subquery_push_qual lacks the infrastructure to do. The case
4481 : : * arises so seldom that it doesn't seem worth working hard on.
4482 : : */
4483 [ + + ]: 410 : if (var->varno != rti)
4484 : : {
4485 : 2 : safe = PUSHDOWN_UNSAFE;
4486 : 2 : break;
4487 : : }
4488 : :
4489 : : /* Subqueries have no system columns */
4490 [ - + ]: 408 : Assert(var->varattno >= 0);
4491 : :
4492 : : /* Check point 4 */
4493 [ + - ]: 408 : if (var->varattno == 0)
4494 : : {
4495 : 0 : safe = PUSHDOWN_UNSAFE;
4496 : 0 : break;
4497 : : }
4498 : :
4499 : : /* Check point 5 */
4500 [ + + ]: 408 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
4501 : : {
4502 [ + + ]: 93 : if (safetyInfo->unsafeFlags[var->varattno] &
4503 : : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
4504 : : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
4505 : : {
4506 : 38 : safe = PUSHDOWN_UNSAFE;
4507 : 38 : break;
4508 : : }
4509 : : else
4510 : : {
4511 : : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
4512 : 55 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
4513 : : /* don't break, we might find another Var that's unsafe */
4514 : : }
4515 : 55 : }
4516 [ + + ]: 410 : }
4517 : :
4518 : 392 : list_free(vars);
4519 : :
4520 : 392 : return safe;
4521 : 433 : }
4522 : :
4523 : : /*
4524 : : * subquery_push_qual - push down a qual that we have determined is safe
4525 : : */
4526 : : static void
4527 : 330 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
4528 : : {
4529 [ + + ]: 330 : if (subquery->setOperations != NULL)
4530 : : {
4531 : : /* Recurse to push it separately to each component query */
4532 : 20 : recurse_push_qual(subquery->setOperations, subquery,
4533 : 10 : rte, rti, qual);
4534 : 10 : }
4535 : : else
4536 : : {
4537 : : /*
4538 : : * We need to replace Vars in the qual (which must refer to outputs of
4539 : : * the subquery) with copies of the subquery's targetlist expressions.
4540 : : * Note that at this point, any uplevel Vars in the qual should have
4541 : : * been replaced with Params, so they need no work.
4542 : : *
4543 : : * This step also ensures that when we are pushing into a setop tree,
4544 : : * each component query gets its own copy of the qual.
4545 : : */
4546 : 640 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
4547 : 320 : subquery->targetList,
4548 : 320 : subquery->resultRelation,
4549 : : REPLACEVARS_REPORT_ERROR, 0,
4550 : 320 : &subquery->hasSubLinks);
4551 : :
4552 : : /*
4553 : : * Now attach the qual to the proper place: normally WHERE, but if the
4554 : : * subquery uses grouping or aggregation, put it in HAVING (since the
4555 : : * qual really refers to the group-result rows).
4556 : : */
4557 [ + + + - : 320 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
+ - - + ]
4558 : 58 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
4559 : : else
4560 : 262 : subquery->jointree->quals =
4561 : 262 : make_and_qual(subquery->jointree->quals, qual);
4562 : :
4563 : : /*
4564 : : * We need not change the subquery's hasAggs or hasSubLinks flags,
4565 : : * since we can't be pushing down any aggregates that weren't there
4566 : : * before, and we don't push down subselects at all.
4567 : : */
4568 : : }
4569 : 330 : }
4570 : :
4571 : : /*
4572 : : * Helper routine to recurse through setOperations tree
4573 : : */
4574 : : static void
4575 : 30 : recurse_push_qual(Node *setOp, Query *topquery,
4576 : : RangeTblEntry *rte, Index rti, Node *qual)
4577 : : {
4578 [ + + ]: 30 : if (IsA(setOp, RangeTblRef))
4579 : : {
4580 : 20 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4581 : 20 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4582 : 20 : Query *subquery = subrte->subquery;
4583 : :
4584 [ + - ]: 20 : Assert(subquery != NULL);
4585 : 20 : subquery_push_qual(subquery, rte, rti, qual);
4586 : 20 : }
4587 [ + - ]: 10 : else if (IsA(setOp, SetOperationStmt))
4588 : : {
4589 : 10 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4590 : :
4591 : 10 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4592 : 10 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4593 : 10 : }
4594 : : else
4595 : : {
4596 [ # # # # ]: 0 : elog(ERROR, "unrecognized node type: %d",
4597 : : (int) nodeTag(setOp));
4598 : : }
4599 : 30 : }
4600 : :
4601 : : /*****************************************************************************
4602 : : * SIMPLIFYING SUBQUERY TARGETLISTS
4603 : : *****************************************************************************/
4604 : :
4605 : : /*
4606 : : * remove_unused_subquery_outputs
4607 : : * Remove subquery targetlist items we don't need
4608 : : *
4609 : : * It's possible, even likely, that the upper query does not read all the
4610 : : * output columns of the subquery. We can remove any such outputs that are
4611 : : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4612 : : * affect semantics otherwise (e.g., volatile functions can't be removed).
4613 : : * This is useful not only because we might be able to remove expensive-to-
4614 : : * compute expressions, but because deletion of output columns might allow
4615 : : * optimizations such as join removal to occur within the subquery.
4616 : : *
4617 : : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4618 : : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4619 : : * parameter is modified by the function, so callers must make a copy if they
4620 : : * need to use the passed in Bitmapset after calling this function.
4621 : : *
4622 : : * To avoid affecting column numbering in the targetlist, we don't physically
4623 : : * remove unused tlist entries, but rather replace their expressions with NULL
4624 : : * constants. This is implemented by modifying subquery->targetList.
4625 : : */
4626 : : static void
4627 : 1480 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4628 : : Bitmapset *extra_used_attrs)
4629 : : {
4630 : 1480 : Bitmapset *attrs_used;
4631 : 1480 : ListCell *lc;
4632 : :
4633 : : /*
4634 : : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4635 : : * the current callers use the Bitmapset after calling this function.
4636 : : */
4637 : 1480 : attrs_used = extra_used_attrs;
4638 : :
4639 : : /*
4640 : : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4641 : : * could update all the child SELECTs' tlists, but it seems not worth the
4642 : : * trouble presently.
4643 : : */
4644 [ + + ]: 1480 : if (subquery->setOperations)
4645 : 48 : return;
4646 : :
4647 : : /*
4648 : : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4649 : : * time: all its output columns must be used in the distinctClause.
4650 : : */
4651 [ + + + + ]: 1432 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4652 : 34 : return;
4653 : :
4654 : : /*
4655 : : * Collect a bitmap of all the output column numbers used by the upper
4656 : : * query.
4657 : : *
4658 : : * Add all the attributes needed for joins or final output. Note: we must
4659 : : * look at rel's targetlist, not the attr_needed data, because attr_needed
4660 : : * isn't computed for inheritance child rels, cf set_append_rel_size().
4661 : : * (XXX might be worth changing that sometime.)
4662 : : */
4663 : 1398 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4664 : :
4665 : : /* Add all the attributes used by un-pushed-down restriction clauses. */
4666 [ + + + + : 1525 : foreach(lc, rel->baserestrictinfo)
+ + ]
4667 : : {
4668 : 127 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4669 : :
4670 : 127 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4671 : 127 : }
4672 : :
4673 : : /*
4674 : : * If there's a whole-row reference to the subquery, we can't remove
4675 : : * anything.
4676 : : */
4677 [ + + ]: 1398 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4678 : 45 : return;
4679 : :
4680 : : /*
4681 : : * Run through the tlist and zap entries we don't need. It's okay to
4682 : : * modify the tlist items in-place because set_subquery_pathlist made a
4683 : : * copy of the subquery.
4684 : : */
4685 [ + - + + : 6165 : foreach(lc, subquery->targetList)
+ + ]
4686 : : {
4687 : 4812 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4688 : 4812 : Node *texpr = (Node *) tle->expr;
4689 : :
4690 : : /*
4691 : : * If it has a sortgroupref number, it's used in some sort/group
4692 : : * clause so we'd better not remove it. Also, don't remove any
4693 : : * resjunk columns, since their reason for being has nothing to do
4694 : : * with anybody reading the subquery's output. (It's likely that
4695 : : * resjunk columns in a sub-SELECT would always have ressortgroupref
4696 : : * set, but even if they don't, it seems imprudent to remove them.)
4697 : : */
4698 [ + + - + ]: 4812 : if (tle->ressortgroupref || tle->resjunk)
4699 : 440 : continue;
4700 : :
4701 : : /*
4702 : : * If it's used by the upper query, we can't remove it.
4703 : : */
4704 [ + + + + ]: 8744 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4705 : 4372 : attrs_used))
4706 : 2767 : continue;
4707 : :
4708 : : /*
4709 : : * If it contains a set-returning function, we can't remove it since
4710 : : * that could change the number of rows returned by the subquery.
4711 : : */
4712 [ + + + + ]: 1605 : if (subquery->hasTargetSRFs &&
4713 : 57 : expression_returns_set(texpr))
4714 : 29 : continue;
4715 : :
4716 : : /*
4717 : : * If it contains volatile functions, we daren't remove it for fear
4718 : : * that the user is expecting their side-effects to happen.
4719 : : */
4720 [ + + ]: 1576 : if (contain_volatile_functions(texpr))
4721 : 5 : continue;
4722 : :
4723 : : /*
4724 : : * OK, we don't need it. Replace the expression with a NULL constant.
4725 : : * Preserve the exposed type of the expression, in case something
4726 : : * looks at the rowtype of the subquery's result.
4727 : : */
4728 : 3142 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4729 : 1571 : exprTypmod(texpr),
4730 : 1571 : exprCollation(texpr));
4731 [ + + ]: 4812 : }
4732 : 1480 : }
4733 : :
4734 : : /*
4735 : : * create_partial_bitmap_paths
4736 : : * Build partial bitmap heap path for the relation
4737 : : */
4738 : : void
4739 : 13074 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4740 : : Path *bitmapqual)
4741 : : {
4742 : 13074 : int parallel_workers;
4743 : 13074 : double pages_fetched;
4744 : :
4745 : : /* Compute heap pages for bitmap heap scan */
4746 : 13074 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4747 : : NULL, NULL);
4748 : :
4749 : 26148 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4750 : 13074 : max_parallel_workers_per_gather);
4751 : :
4752 [ + + ]: 13074 : if (parallel_workers <= 0)
4753 : 12783 : return;
4754 : :
4755 : 582 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4756 : 291 : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4757 [ - + ]: 13074 : }
4758 : :
4759 : : /*
4760 : : * Compute the number of parallel workers that should be used to scan a
4761 : : * relation. We compute the parallel workers based on the size of the heap to
4762 : : * be scanned and the size of the index to be scanned, then choose a minimum
4763 : : * of those.
4764 : : *
4765 : : * "heap_pages" is the number of pages from the table that we expect to scan, or
4766 : : * -1 if we don't expect to scan any.
4767 : : *
4768 : : * "index_pages" is the number of pages from the index that we expect to scan, or
4769 : : * -1 if we don't expect to scan any.
4770 : : *
4771 : : * "max_workers" is caller's limit on the number of workers. This typically
4772 : : * comes from a GUC.
4773 : : */
4774 : : int
4775 : 73600 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4776 : : int max_workers)
4777 : : {
4778 : 73600 : int parallel_workers = 0;
4779 : :
4780 : : /*
4781 : : * If the user has set the parallel_workers reloption, use that; otherwise
4782 : : * select a default number of workers.
4783 : : */
4784 [ + + ]: 73600 : if (rel->rel_parallel_workers != -1)
4785 : 2317 : parallel_workers = rel->rel_parallel_workers;
4786 : : else
4787 : : {
4788 : : /*
4789 : : * If the number of pages being scanned is insufficient to justify a
4790 : : * parallel scan, just return zero ... unless it's an inheritance
4791 : : * child. In that case, we want to generate a parallel path here
4792 : : * anyway. It might not be worthwhile just for this relation, but
4793 : : * when combined with all of its inheritance siblings it may well pay
4794 : : * off.
4795 : : */
4796 [ + + ]: 73685 : if (rel->reloptkind == RELOPT_BASEREL &&
4797 [ + + + + ]: 65918 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
4798 [ + + ]: 2575 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
4799 : 65745 : return 0;
4800 : :
4801 [ + + ]: 5538 : if (heap_pages >= 0)
4802 : : {
4803 : 5210 : int heap_parallel_threshold;
4804 : 5210 : int heap_parallel_workers = 1;
4805 : :
4806 : : /*
4807 : : * Select the number of workers based on the log of the size of
4808 : : * the relation. This probably needs to be a good deal more
4809 : : * sophisticated, but we need something here for now. Note that
4810 : : * the upper limit of the min_parallel_table_scan_size GUC is
4811 : : * chosen to prevent overflow here.
4812 : : */
4813 [ + + ]: 5210 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4814 [ + + ]: 6178 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4815 : : {
4816 : 968 : heap_parallel_workers++;
4817 : 968 : heap_parallel_threshold *= 3;
4818 [ - + ]: 968 : if (heap_parallel_threshold > INT_MAX / 3)
4819 : 0 : break; /* avoid overflow */
4820 : : }
4821 : :
4822 : 5210 : parallel_workers = heap_parallel_workers;
4823 : 5210 : }
4824 : :
4825 [ + + ]: 5538 : if (index_pages >= 0)
4826 : : {
4827 : 1211 : int index_parallel_workers = 1;
4828 : 1211 : int index_parallel_threshold;
4829 : :
4830 : : /* same calculation as for heap_pages above */
4831 [ + + ]: 1211 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4832 [ + + ]: 1257 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4833 : : {
4834 : 46 : index_parallel_workers++;
4835 : 46 : index_parallel_threshold *= 3;
4836 [ - + ]: 46 : if (index_parallel_threshold > INT_MAX / 3)
4837 : 0 : break; /* avoid overflow */
4838 : : }
4839 : :
4840 [ + + ]: 1211 : if (parallel_workers > 0)
4841 [ - + ]: 883 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4842 : : else
4843 : 328 : parallel_workers = index_parallel_workers;
4844 : 1211 : }
4845 : : }
4846 : :
4847 : : /* In no case use more than caller supplied maximum number of workers */
4848 [ + + ]: 7855 : parallel_workers = Min(parallel_workers, max_workers);
4849 : :
4850 : 7855 : return parallel_workers;
4851 : 73600 : }
4852 : :
4853 : : /*
4854 : : * generate_partitionwise_join_paths
4855 : : * Create paths representing partitionwise join for given partitioned
4856 : : * join relation.
4857 : : *
4858 : : * This must not be called until after we are done adding paths for all
4859 : : * child-joins. Otherwise, add_path might delete a path to which some path
4860 : : * generated here has a reference.
4861 : : */
4862 : : void
4863 : 21890 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4864 : : {
4865 : 21890 : List *live_children = NIL;
4866 : 21890 : int cnt_parts;
4867 : 21890 : int num_parts;
4868 : 21890 : RelOptInfo **part_rels;
4869 : :
4870 : : /* Handle only join relations here. */
4871 [ + + + - ]: 21890 : if (!IS_JOIN_REL(rel))
4872 : 0 : return;
4873 : :
4874 : : /* We've nothing to do if the relation is not partitioned. */
4875 [ + + + + : 21890 : if (!IS_PARTITIONED_REL(rel))
+ + + - +
+ ]
4876 : 20681 : return;
4877 : :
4878 : : /* The relation should have consider_partitionwise_join set. */
4879 [ + - ]: 1209 : Assert(rel->consider_partitionwise_join);
4880 : :
4881 : : /* Guard against stack overflow due to overly deep partition hierarchy. */
4882 : 1209 : check_stack_depth();
4883 : :
4884 : 1209 : num_parts = rel->nparts;
4885 : 1209 : part_rels = rel->part_rels;
4886 : :
4887 : : /* Collect non-dummy child-joins. */
4888 [ + + ]: 4318 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4889 : : {
4890 : 3109 : RelOptInfo *child_rel = part_rels[cnt_parts];
4891 : :
4892 : : /* If it's been pruned entirely, it's certainly dummy. */
4893 [ + + ]: 3109 : if (child_rel == NULL)
4894 : 10 : continue;
4895 : :
4896 : : /* Make partitionwise join paths for this partitioned child-join. */
4897 : 3099 : generate_partitionwise_join_paths(root, child_rel);
4898 : :
4899 : : /* If we failed to make any path for this child, we must give up. */
4900 [ + - ]: 3099 : if (child_rel->pathlist == NIL)
4901 : : {
4902 : : /*
4903 : : * Mark the parent joinrel as unpartitioned so that later
4904 : : * functions treat it correctly.
4905 : : */
4906 : 0 : rel->nparts = 0;
4907 : 0 : return;
4908 : : }
4909 : :
4910 : : /* Else, identify the cheapest path for it. */
4911 : 3099 : set_cheapest(child_rel);
4912 : :
4913 : : /* Dummy children need not be scanned, so ignore those. */
4914 [ - + ]: 3099 : if (IS_DUMMY_REL(child_rel))
4915 : 0 : continue;
4916 : :
4917 : : /*
4918 : : * Except for the topmost scan/join rel, consider generating partial
4919 : : * aggregation paths for the grouped relation on top of the paths of
4920 : : * this partitioned child-join. After that, we're done creating paths
4921 : : * for the grouped relation, so run set_cheapest().
4922 : : */
4923 [ + + + + ]: 5245 : if (child_rel->grouped_rel != NULL &&
4924 [ + - + + : 2146 : !bms_equal(IS_OTHER_REL(rel) ?
- + ]
4925 : 2146 : rel->top_parent_relids : rel->relids,
4926 : 2146 : root->all_query_rels))
4927 : : {
4928 : 40 : RelOptInfo *grouped_rel = child_rel->grouped_rel;
4929 : :
4930 [ - + ]: 40 : Assert(IS_GROUPED_REL(grouped_rel));
4931 : :
4932 : 40 : generate_grouped_paths(root, grouped_rel, child_rel);
4933 : 40 : set_cheapest(grouped_rel);
4934 : 40 : }
4935 : :
4936 : : #ifdef OPTIMIZER_DEBUG
4937 : : pprint(child_rel);
4938 : : #endif
4939 : :
4940 : 3099 : live_children = lappend(live_children, child_rel);
4941 [ + - + ]: 3109 : }
4942 : :
4943 : : /* If all child-joins are dummy, parent join is also dummy. */
4944 [ + - ]: 1209 : if (!live_children)
4945 : : {
4946 : 0 : mark_dummy_rel(rel);
4947 : 0 : return;
4948 : : }
4949 : :
4950 : : /* Build additional paths for this rel from child-join paths. */
4951 : 1209 : add_paths_to_append_rel(root, rel, live_children);
4952 : 1209 : list_free(live_children);
4953 [ - + ]: 21890 : }
|
