Age Owner Branch data TLA 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) 1996-2025, 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 : : static Path *get_singleton_append_subpath(Path *path);
133 : : static void set_dummy_rel_pathlist(RelOptInfo *rel);
134 : : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
135 : : Index rti, RangeTblEntry *rte);
136 : : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : : RangeTblEntry *rte);
138 : : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : : RangeTblEntry *rte);
140 : : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : : RangeTblEntry *rte);
142 : : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : : RangeTblEntry *rte);
144 : : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
145 : : RangeTblEntry *rte);
146 : : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
147 : : RangeTblEntry *rte);
148 : : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
149 : : RangeTblEntry *rte);
150 : : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
151 : : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
152 : : pushdown_safety_info *safetyInfo);
153 : : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
154 : : pushdown_safety_info *safetyInfo);
155 : : static void check_output_expressions(Query *subquery,
156 : : pushdown_safety_info *safetyInfo);
157 : : static void compare_tlist_datatypes(List *tlist, List *colTypes,
158 : : pushdown_safety_info *safetyInfo);
159 : : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
160 : : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
161 : : RestrictInfo *rinfo,
162 : : pushdown_safety_info *safetyInfo);
163 : : static void subquery_push_qual(Query *subquery,
164 : : RangeTblEntry *rte, Index rti, Node *qual);
165 : : static void recurse_push_qual(Node *setOp, Query *topquery,
166 : : RangeTblEntry *rte, Index rti, Node *qual);
167 : : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
168 : : Bitmapset *extra_used_attrs);
169 : :
170 : :
171 : : /*
172 : : * make_one_rel
173 : : * Finds all possible access paths for executing a query, returning a
174 : : * single rel that represents the join of all base rels in the query.
175 : : */
176 : : RelOptInfo *
7302 tgl@sss.pgh.pa.us 177 :CBC 171196 : make_one_rel(PlannerInfo *root, List *joinlist)
178 : : {
179 : : RelOptInfo *rel;
180 : : Index rti;
181 : : double total_pages;
182 : :
183 : : /* Mark base rels as to whether we care about fast-start plans */
3850 184 : 171196 : set_base_rel_consider_startup(root);
185 : :
186 : : /*
187 : : * Compute size estimates and consider_parallel flags for each base rel.
188 : : */
5073 189 : 171196 : set_base_rel_sizes(root);
190 : :
191 : : /*
192 : : * Build grouped relations for simple rels (i.e., base or "other" member
193 : : * relations) where possible.
194 : : */
70 rguo@postgresql.org 195 :GNC 171179 : setup_simple_grouped_rels(root);
196 : :
197 : : /*
198 : : * We should now have size estimates for every actual table involved in
199 : : * the query, and we also know which if any have been deleted from the
200 : : * query by join removal, pruned by partition pruning, or eliminated by
201 : : * constraint exclusion. So we can now compute total_table_pages.
202 : : *
203 : : * Note that appendrels are not double-counted here, even though we don't
204 : : * bother to distinguish RelOptInfos for appendrel parents, because the
205 : : * parents will have pages = 0.
206 : : *
207 : : * XXX if a table is self-joined, we will count it once per appearance,
208 : : * which perhaps is the wrong thing ... but that's not completely clear,
209 : : * and detecting self-joins here is difficult, so ignore it for now.
210 : : */
2597 tgl@sss.pgh.pa.us 211 :CBC 171179 : total_pages = 0;
212 [ + + ]: 524196 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
213 : : {
214 : 353017 : RelOptInfo *brel = root->simple_rel_array[rti];
215 : :
216 : : /* there may be empty slots corresponding to non-baserel RTEs */
217 [ + + ]: 353017 : if (brel == NULL)
218 : 81693 : continue;
219 : :
220 [ - + ]: 271324 : Assert(brel->relid == rti); /* sanity check on array */
221 : :
222 [ + + ]: 271324 : if (IS_DUMMY_REL(brel))
223 : 704 : continue;
224 : :
225 [ + + + - ]: 270620 : if (IS_SIMPLE_REL(brel))
226 : 270620 : total_pages += (double) brel->pages;
227 : : }
228 : 171179 : root->total_table_pages = total_pages;
229 : :
230 : : /*
231 : : * Generate access paths for each base rel.
232 : : */
9166 233 : 171179 : set_base_rel_pathlists(root);
234 : :
235 : : /*
236 : : * Generate access paths for the entire join tree.
237 : : */
7302 238 : 171179 : rel = make_rel_from_joinlist(root, joinlist);
239 : :
240 : : /*
241 : : * The result should join all and only the query's base + outer-join rels.
242 : : */
1052 243 [ - + ]: 171179 : Assert(bms_equal(rel->relids, root->all_query_rels));
244 : :
5073 245 : 171179 : return rel;
246 : : }
247 : :
248 : : /*
249 : : * set_base_rel_consider_startup
250 : : * Set the consider_[param_]startup flags for each base-relation entry.
251 : : *
252 : : * For the moment, we only deal with consider_param_startup here; because the
253 : : * logic for consider_startup is pretty trivial and is the same for every base
254 : : * relation, we just let build_simple_rel() initialize that flag correctly to
255 : : * start with. If that logic ever gets more complicated it would probably
256 : : * be better to move it here.
257 : : */
258 : : static void
3850 259 : 171196 : set_base_rel_consider_startup(PlannerInfo *root)
260 : : {
261 : : /*
262 : : * Since parameterized paths can only be used on the inside of a nestloop
263 : : * join plan, there is usually little value in considering fast-start
264 : : * plans for them. However, for relations that are on the RHS of a SEMI
265 : : * or ANTI join, a fast-start plan can be useful because we're only going
266 : : * to care about fetching one tuple anyway.
267 : : *
268 : : * To minimize growth of planning time, we currently restrict this to
269 : : * cases where the RHS is a single base relation, not a join; there is no
270 : : * provision for consider_param_startup to get set at all on joinrels.
271 : : * Also we don't worry about appendrels. costsize.c's costing rules for
272 : : * nestloop semi/antijoins don't consider such cases either.
273 : : */
274 : : ListCell *lc;
275 : :
276 [ + + + + : 192748 : foreach(lc, root->join_info_list)
+ + ]
277 : : {
278 : 21552 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
279 : : int varno;
280 : :
281 [ + + + + : 25902 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
+ + ]
282 : 4350 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
283 : : {
284 : 4250 : RelOptInfo *rel = find_base_rel(root, varno);
285 : :
286 : 4250 : rel->consider_param_startup = true;
287 : : }
288 : : }
289 : 171196 : }
290 : :
291 : : /*
292 : : * set_base_rel_sizes
293 : : * Set the size estimates (rows and widths) for each base-relation entry.
294 : : * Also determine whether to consider parallel paths for base relations.
295 : : *
296 : : * We do this in a separate pass over the base rels so that rowcount
297 : : * estimates are available for parameterized path generation, and also so
298 : : * that each rel's consider_parallel flag is set correctly before we begin to
299 : : * generate paths.
300 : : */
301 : : static void
5073 302 : 171196 : set_base_rel_sizes(PlannerInfo *root)
303 : : {
304 : : Index rti;
305 : :
306 [ + + ]: 524214 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
307 : : {
308 : 353035 : RelOptInfo *rel = root->simple_rel_array[rti];
309 : : RangeTblEntry *rte;
310 : :
311 : : /* there may be empty slots corresponding to non-baserel RTEs */
312 [ + + ]: 353035 : if (rel == NULL)
313 : 81694 : continue;
314 : :
3101 315 [ - + ]: 271341 : Assert(rel->relid == rti); /* sanity check on array */
316 : :
317 : : /* ignore RTEs that are "other rels" */
5073 318 [ + + ]: 271341 : if (rel->reloptkind != RELOPT_BASEREL)
319 : 29782 : continue;
320 : :
3689 rhaas@postgresql.org 321 : 241559 : rte = root->simple_rte_array[rti];
322 : :
323 : : /*
324 : : * If parallelism is allowable for this query in general, see whether
325 : : * it's allowable for this rel in particular. We have to do this
326 : : * before set_rel_size(), because (a) if this rel is an inheritance
327 : : * parent, set_append_rel_size() will use and perhaps change the rel's
328 : : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
329 : : * goes ahead and makes paths immediately.
330 : : */
331 [ + + ]: 241559 : if (root->glob->parallelModeOK)
332 : 188390 : set_rel_consider_parallel(root, rel, rte);
333 : :
334 : 241559 : set_rel_size(root, rel, rti, rte);
335 : : }
10753 scrappy@hub.org 336 : 171179 : }
337 : :
338 : : /*
339 : : * setup_simple_grouped_rels
340 : : * For each simple relation, build a grouped simple relation if eager
341 : : * aggregation is possible and if this relation can produce grouped paths.
342 : : */
343 : : static void
70 rguo@postgresql.org 344 :GNC 171179 : setup_simple_grouped_rels(PlannerInfo *root)
345 : : {
346 : : Index rti;
347 : :
348 : : /*
349 : : * If there are no aggregate expressions or grouping expressions, eager
350 : : * aggregation is not possible.
351 : : */
352 [ + + ]: 171179 : if (root->agg_clause_list == NIL ||
353 [ + + ]: 330 : root->group_expr_list == NIL)
354 : 170882 : return;
355 : :
356 [ + + ]: 2545 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
357 : : {
358 : 2248 : RelOptInfo *rel = root->simple_rel_array[rti];
359 : :
360 : : /* there may be empty slots corresponding to non-baserel RTEs */
361 [ + + ]: 2248 : if (rel == NULL)
362 : 712 : continue;
363 : :
364 [ - + ]: 1536 : Assert(rel->relid == rti); /* sanity check on array */
365 [ + + - + ]: 1536 : Assert(IS_SIMPLE_REL(rel)); /* sanity check on rel */
366 : :
367 : 1536 : (void) build_simple_grouped_rel(root, rel);
368 : : }
369 : : }
370 : :
371 : : /*
372 : : * set_base_rel_pathlists
373 : : * Finds all paths available for scanning each base-relation entry.
374 : : * Sequential scan and any available indices are considered.
375 : : * Each useful path is attached to its relation's 'pathlist' field.
376 : : */
377 : : static void
7500 tgl@sss.pgh.pa.us 378 :CBC 171179 : set_base_rel_pathlists(PlannerInfo *root)
379 : : {
380 : : Index rti;
381 : :
7260 382 [ + + ]: 524196 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
383 : : {
384 : 353017 : RelOptInfo *rel = root->simple_rel_array[rti];
385 : :
386 : : /* there may be empty slots corresponding to non-baserel RTEs */
7499 387 [ + + ]: 353017 : if (rel == NULL)
388 : 81693 : continue;
389 : :
3101 390 [ - + ]: 271324 : Assert(rel->relid == rti); /* sanity check on array */
391 : :
392 : : /* ignore RTEs that are "other rels" */
7499 393 [ + + ]: 271324 : if (rel->reloptkind != RELOPT_BASEREL)
394 : 29782 : continue;
395 : :
6815 396 : 241542 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
397 : : }
7257 398 : 171179 : }
399 : :
400 : : /*
401 : : * set_rel_size
402 : : * Set size estimates for a base relation
403 : : */
404 : : static void
5073 405 : 271186 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
406 : : Index rti, RangeTblEntry *rte)
407 : : {
5198 408 [ + + + + ]: 512745 : if (rel->reloptkind == RELOPT_BASEREL &&
409 : 241559 : relation_excluded_by_constraints(root, rel, rte))
410 : : {
411 : : /*
412 : : * We proved we don't need to scan the rel via constraint exclusion,
413 : : * so set up a single dummy path for it. Here we only check this for
414 : : * regular baserels; if it's an otherrel, CE was already checked in
415 : : * set_append_rel_size().
416 : : *
417 : : * In this case, we go ahead and set up the relation's path right away
418 : : * instead of leaving it for set_rel_pathlist to do. This is because
419 : : * we don't have a convention for marking a rel as dummy except by
420 : : * assigning a dummy path to it.
421 : : */
422 : 333 : set_dummy_rel_pathlist(rel);
423 : : }
424 [ + + ]: 270853 : else if (rte->inh)
425 : : {
426 : : /* It's an "append relation", process accordingly */
5073 427 : 13131 : set_append_rel_size(root, rel, rti, rte);
428 : : }
429 : : else
430 : : {
5412 431 [ + + + + : 257722 : switch (rel->rtekind)
+ + + +
- ]
432 : : {
433 : 211288 : case RTE_RELATION:
434 [ + + ]: 211288 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
435 : : {
436 : : /* Foreign table */
5073 437 : 1237 : set_foreign_size(root, rel, rte);
438 : : }
3193 rhaas@postgresql.org 439 [ + + ]: 210051 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
440 : : {
441 : : /*
442 : : * We could get here if asked to scan a partitioned table
443 : : * with ONLY. In that case we shouldn't scan any of the
444 : : * partitions, so mark it as a dummy rel.
445 : : */
446 : 20 : set_dummy_rel_pathlist(rel);
447 : : }
3869 simon@2ndQuadrant.co 448 [ + + ]: 210031 : else if (rte->tablesample != NULL)
449 : : {
450 : : /* Sampled relation */
451 : 153 : set_tablesample_rel_size(root, rel, rte);
452 : : }
453 : : else
454 : : {
455 : : /* Plain relation */
5073 tgl@sss.pgh.pa.us 456 : 209878 : set_plain_rel_size(root, rel, rte);
457 : : }
5412 458 : 211271 : break;
459 : 11713 : case RTE_SUBQUERY:
460 : :
461 : : /*
462 : : * Subqueries don't support making a choice between
463 : : * parameterized and unparameterized paths, so just go ahead
464 : : * and build their paths immediately.
465 : : */
466 : 11713 : set_subquery_pathlist(root, rel, rti, rte);
467 : 11713 : break;
468 : 25133 : case RTE_FUNCTION:
5073 469 : 25133 : set_function_size_estimates(root, rel);
5412 470 : 25133 : break;
3206 alvherre@alvh.no-ip. 471 : 311 : case RTE_TABLEFUNC:
472 : 311 : set_tablefunc_size_estimates(root, rel);
473 : 311 : break;
5412 tgl@sss.pgh.pa.us 474 : 4236 : case RTE_VALUES:
5073 475 : 4236 : set_values_size_estimates(root, rel);
5412 476 : 4236 : break;
477 : 2657 : case RTE_CTE:
478 : :
479 : : /*
480 : : * CTEs don't support making a choice between parameterized
481 : : * and unparameterized paths, so just go ahead and build their
482 : : * paths immediately.
483 : : */
484 [ + + ]: 2657 : if (rte->self_reference)
485 : 474 : set_worktable_pathlist(root, rel, rte);
486 : : else
487 : 2183 : set_cte_pathlist(root, rel, rte);
488 : 2657 : break;
3183 kgrittn@postgresql.o 489 : 241 : case RTE_NAMEDTUPLESTORE:
490 : : /* Might as well just build the path immediately */
491 : 241 : set_namedtuplestore_pathlist(root, rel, rte);
492 : 241 : break;
2515 tgl@sss.pgh.pa.us 493 : 2143 : case RTE_RESULT:
494 : : /* Might as well just build the path immediately */
495 : 2143 : set_result_pathlist(root, rel, rte);
496 : 2143 : break;
5412 tgl@sss.pgh.pa.us 497 :UBC 0 : default:
498 [ # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
499 : : break;
500 : : }
501 : : }
502 : :
503 : : /*
504 : : * We insist that all non-dummy rels have a nonzero rowcount estimate.
505 : : */
3797 tgl@sss.pgh.pa.us 506 [ + + - + ]:CBC 271168 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
5073 507 : 271168 : }
508 : :
509 : : /*
510 : : * set_rel_pathlist
511 : : * Build access paths for a base relation
512 : : */
513 : : static void
514 : 271237 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
515 : : Index rti, RangeTblEntry *rte)
516 : : {
517 [ + + ]: 271237 : if (IS_DUMMY_REL(rel))
518 : : {
519 : : /* We already proved the relation empty, so nothing more to do */
520 : : }
521 [ + + ]: 270602 : else if (rte->inh)
522 : : {
523 : : /* It's an "append relation", process accordingly */
524 : 12980 : set_append_rel_pathlist(root, rel, rti, rte);
525 : : }
526 : : else
527 : : {
528 [ + + + + : 257622 : switch (rel->rtekind)
+ + + +
- ]
529 : : {
530 : 211251 : case RTE_RELATION:
531 [ + + ]: 211251 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
532 : : {
533 : : /* Foreign table */
534 : 1235 : set_foreign_pathlist(root, rel, rte);
535 : : }
3869 simon@2ndQuadrant.co 536 [ + + ]: 210016 : else if (rte->tablesample != NULL)
537 : : {
538 : : /* Sampled relation */
539 : 153 : set_tablesample_rel_pathlist(root, rel, rte);
540 : : }
541 : : else
542 : : {
543 : : /* Plain relation */
5073 tgl@sss.pgh.pa.us 544 : 209863 : set_plain_rel_pathlist(root, rel, rte);
545 : : }
546 : 211251 : break;
547 : 11650 : case RTE_SUBQUERY:
548 : : /* Subquery --- fully handled during set_rel_size */
549 : 11650 : break;
550 : 25133 : case RTE_FUNCTION:
551 : : /* RangeFunction */
552 : 25133 : set_function_pathlist(root, rel, rte);
553 : 25133 : break;
3206 alvherre@alvh.no-ip. 554 : 311 : case RTE_TABLEFUNC:
555 : : /* Table Function */
556 : 311 : set_tablefunc_pathlist(root, rel, rte);
557 : 311 : break;
5073 tgl@sss.pgh.pa.us 558 : 4236 : case RTE_VALUES:
559 : : /* Values list */
560 : 4236 : set_values_pathlist(root, rel, rte);
561 : 4236 : break;
562 : 2657 : case RTE_CTE:
563 : : /* CTE reference --- fully handled during set_rel_size */
564 : 2657 : break;
3183 kgrittn@postgresql.o 565 : 241 : case RTE_NAMEDTUPLESTORE:
566 : : /* tuplestore reference --- fully handled during set_rel_size */
567 : 241 : break;
2515 tgl@sss.pgh.pa.us 568 : 2143 : case RTE_RESULT:
569 : : /* simple Result --- fully handled during set_rel_size */
570 : 2143 : break;
5073 tgl@sss.pgh.pa.us 571 :UBC 0 : default:
572 [ # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
573 : : break;
574 : : }
575 : : }
576 : :
577 : : /*
578 : : * Allow a plugin to editorialize on the set of Paths for this base
579 : : * relation. It could add new paths (such as CustomPaths) by calling
580 : : * add_path(), or add_partial_path() if parallel aware. It could also
581 : : * delete or modify paths added by the core code.
582 : : */
2503 tgl@sss.pgh.pa.us 583 [ - + ]:CBC 271237 : if (set_rel_pathlist_hook)
2503 tgl@sss.pgh.pa.us 584 :UBC 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
585 : :
586 : : /*
587 : : * If this is a baserel, we should normally consider gathering any partial
588 : : * paths we may have created for it. We have to do this after calling the
589 : : * set_rel_pathlist_hook, else it cannot add partial paths to be included
590 : : * here.
591 : : *
592 : : * However, if this is an inheritance child, skip it. Otherwise, we could
593 : : * end up with a very large number of gather nodes, each trying to grab
594 : : * its own pool of workers. Instead, we'll consider gathering partial
595 : : * paths for the parent appendrel.
596 : : *
597 : : * Also, if this is the topmost scan/join rel, we postpone gathering until
598 : : * the final scan/join targetlist is available (see grouping_planner).
599 : : */
2837 rhaas@postgresql.org 600 [ + + ]:CBC 271237 : if (rel->reloptkind == RELOPT_BASEREL &&
1052 tgl@sss.pgh.pa.us 601 [ + + ]: 241542 : !bms_equal(rel->relids, root->all_query_rels))
2080 tomas.vondra@postgre 602 : 121349 : generate_useful_gather_paths(root, rel, false);
603 : :
604 : : /* Now find the cheapest of the paths for this rel */
4044 tgl@sss.pgh.pa.us 605 : 271237 : set_cheapest(rel);
606 : :
607 : : /*
608 : : * If a grouped relation for this rel exists, build partial aggregation
609 : : * paths for it.
610 : : *
611 : : * Note that this can only happen after we've called set_cheapest() for
612 : : * this base rel, because we need its cheapest paths.
613 : : */
70 rguo@postgresql.org 614 :GNC 271237 : set_grouped_rel_pathlist(root, rel);
615 : :
616 : : #ifdef OPTIMIZER_DEBUG
617 : : pprint(rel);
618 : : #endif
9166 tgl@sss.pgh.pa.us 619 :CBC 271237 : }
620 : :
621 : : /*
622 : : * set_plain_rel_size
623 : : * Set size estimates for a plain relation (no subquery, no inheritance)
624 : : */
625 : : static void
5073 626 : 209878 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
627 : : {
628 : : /*
629 : : * Test any partial indexes of rel for applicability. We must do this
630 : : * first since partial unique indexes can affect size estimates.
631 : : */
3548 632 : 209878 : check_index_predicates(root, rel);
633 : :
634 : : /* Mark rel with estimated output rows, width, etc */
6815 635 : 209878 : set_baserel_size_estimates(root, rel);
5073 636 : 209863 : }
637 : :
638 : : /*
639 : : * If this relation could possibly be scanned from within a worker, then set
640 : : * its consider_parallel flag.
641 : : */
642 : : static void
3689 rhaas@postgresql.org 643 : 210638 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
644 : : RangeTblEntry *rte)
645 : : {
646 : : /*
647 : : * The flag has previously been initialized to false, so we can just
648 : : * return if it becomes clear that we can't safely set it.
649 : : */
3454 tgl@sss.pgh.pa.us 650 [ - + ]: 210638 : Assert(!rel->consider_parallel);
651 : :
652 : : /* Don't call this if parallelism is disallowed for the entire query. */
3689 rhaas@postgresql.org 653 [ - + ]: 210638 : Assert(root->glob->parallelModeOK);
654 : :
655 : : /* This should only be called for baserels and appendrel children. */
3180 656 [ + + - + ]: 210638 : Assert(IS_SIMPLE_REL(rel));
657 : :
658 : : /* Assorted checks based on rtekind. */
3689 659 [ + + - + : 210638 : switch (rte->rtekind)
+ + + + +
- - ]
660 : : {
661 : 181731 : case RTE_RELATION:
662 : :
663 : : /*
664 : : * Currently, parallel workers can't access the leader's temporary
665 : : * tables. We could possibly relax this if we wrote all of its
666 : : * local buffers at the start of the query and made no changes
667 : : * thereafter (maybe we could allow hint bit changes), and if we
668 : : * taught the workers to read them. Writing a large number of
669 : : * temporary buffers could be expensive, though, and we don't have
670 : : * the rest of the necessary infrastructure right now anyway. So
671 : : * for now, bail out if we see a temporary table.
672 : : */
673 [ + + ]: 181731 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
674 : 4305 : return;
675 : :
676 : : /*
677 : : * Table sampling can be pushed down to workers if the sample
678 : : * function and its arguments are safe.
679 : : */
680 [ + + ]: 177426 : if (rte->tablesample != NULL)
681 : : {
3305 tgl@sss.pgh.pa.us 682 : 165 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
683 : :
3689 rhaas@postgresql.org 684 [ + + ]: 165 : if (proparallel != PROPARALLEL_SAFE)
685 : 18 : return;
3407 tgl@sss.pgh.pa.us 686 [ + + ]: 147 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
3689 rhaas@postgresql.org 687 : 6 : return;
688 : : }
689 : :
690 : : /*
691 : : * Ask FDWs whether they can support performing a ForeignScan
692 : : * within a worker. Most often, the answer will be no. For
693 : : * example, if the nature of the FDW is such that it opens a TCP
694 : : * connection with a remote server, each parallel worker would end
695 : : * up with a separate connection, and these connections might not
696 : : * be appropriately coordinated between workers and the leader.
697 : : */
3582 698 [ + + ]: 177402 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
699 : : {
700 [ - + ]: 782 : Assert(rel->fdwroutine);
701 [ + + ]: 782 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
702 : 746 : return;
703 [ - + ]: 36 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
3582 rhaas@postgresql.org 704 :UBC 0 : return;
705 : : }
706 : :
707 : : /*
708 : : * There are additional considerations for appendrels, which we'll
709 : : * deal with in set_append_rel_size and set_append_rel_pathlist.
710 : : * For now, just set consider_parallel based on the rel's own
711 : : * quals and targetlist.
712 : : */
3689 rhaas@postgresql.org 713 :CBC 176656 : break;
714 : :
715 : 9855 : case RTE_SUBQUERY:
716 : :
717 : : /*
718 : : * There's no intrinsic problem with scanning a subquery-in-FROM
719 : : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
720 : : * If the subquery doesn't happen to have any parallel-safe paths,
721 : : * then flagging it as consider_parallel won't change anything,
722 : : * but that's true for plain tables, too. We must set
723 : : * consider_parallel based on the rel's own quals and targetlist,
724 : : * so that if a subquery path is parallel-safe but the quals and
725 : : * projection we're sticking onto it are not, we correctly mark
726 : : * the SubqueryScanPath as not parallel-safe. (Note that
727 : : * set_subquery_pathlist() might push some of these quals down
728 : : * into the subquery itself, but that doesn't change anything.)
729 : : *
730 : : * We can't push sub-select containing LIMIT/OFFSET to workers as
731 : : * there is no guarantee that the row order will be fully
732 : : * deterministic, and applying LIMIT/OFFSET will lead to
733 : : * inconsistent results at the top-level. (In some cases, where
734 : : * the result is ordered, we could relax this restriction. But it
735 : : * doesn't currently seem worth expending extra effort to do so.)
736 : : */
737 : : {
2651 akapila@postgresql.o 738 : 9855 : Query *subquery = castNode(Query, rte->subquery);
739 : :
740 [ + + ]: 9855 : if (limit_needed(subquery))
741 : 254 : return;
742 : : }
3454 tgl@sss.pgh.pa.us 743 : 9601 : break;
744 : :
3689 rhaas@postgresql.org 745 :UBC 0 : case RTE_JOIN:
746 : : /* Shouldn't happen; we're only considering baserels here. */
747 : 0 : Assert(false);
748 : : return;
749 : :
3689 rhaas@postgresql.org 750 :CBC 13005 : case RTE_FUNCTION:
751 : : /* Check for parallel-restricted functions. */
3407 tgl@sss.pgh.pa.us 752 [ + + ]: 13005 : if (!is_parallel_safe(root, (Node *) rte->functions))
3689 rhaas@postgresql.org 753 : 6328 : return;
754 : 6677 : break;
755 : :
3206 alvherre@alvh.no-ip. 756 : 311 : case RTE_TABLEFUNC:
757 : : /* not parallel safe */
758 : 311 : return;
759 : :
3689 rhaas@postgresql.org 760 : 1423 : case RTE_VALUES:
761 : : /* Check for parallel-restricted functions. */
3407 tgl@sss.pgh.pa.us 762 [ + + ]: 1423 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
763 : 6 : return;
3689 rhaas@postgresql.org 764 : 1417 : break;
765 : :
766 : 2164 : case RTE_CTE:
767 : :
768 : : /*
769 : : * CTE tuplestores aren't shared among parallel workers, so we
770 : : * force all CTE scans to happen in the leader. Also, populating
771 : : * the CTE would require executing a subplan that's not available
772 : : * in the worker, might be parallel-restricted, and must get
773 : : * executed only once.
774 : : */
775 : 2164 : return;
776 : :
3183 kgrittn@postgresql.o 777 : 227 : case RTE_NAMEDTUPLESTORE:
778 : :
779 : : /*
780 : : * tuplestore cannot be shared, at least without more
781 : : * infrastructure to support that.
782 : : */
783 : 227 : return;
784 : :
2515 tgl@sss.pgh.pa.us 785 : 1922 : case RTE_RESULT:
786 : : /* RESULT RTEs, in themselves, are no problem. */
787 : 1922 : break;
463 rguo@postgresql.org 788 :UBC 0 : case RTE_GROUP:
789 : : /* Shouldn't happen; we're only considering baserels here. */
790 : 0 : Assert(false);
791 : : return;
792 : : }
793 : :
794 : : /*
795 : : * If there's anything in baserestrictinfo that's parallel-restricted, we
796 : : * give up on parallelizing access to this relation. We could consider
797 : : * instead postponing application of the restricted quals until we're
798 : : * above all the parallelism in the plan tree, but it's not clear that
799 : : * that would be a win in very many cases, and it might be tricky to make
800 : : * outer join clauses work correctly. It would likely break equivalence
801 : : * classes, too.
802 : : */
3407 tgl@sss.pgh.pa.us 803 [ + + ]:CBC 196273 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
3689 rhaas@postgresql.org 804 : 13855 : return;
805 : :
806 : : /*
807 : : * Likewise, if the relation's outputs are not parallel-safe, give up.
808 : : * (Usually, they're just Vars, but sometimes they're not.)
809 : : */
3407 tgl@sss.pgh.pa.us 810 [ + + ]: 182418 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
3478 rhaas@postgresql.org 811 : 9 : return;
812 : :
813 : : /* We have a winner. */
3689 814 : 182409 : rel->consider_parallel = true;
815 : : }
816 : :
817 : : /*
818 : : * set_plain_rel_pathlist
819 : : * Build access paths for a plain relation (no subquery, no inheritance)
820 : : */
821 : : static void
5073 tgl@sss.pgh.pa.us 822 : 209863 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
823 : : {
824 : : Relids required_outer;
825 : :
826 : : /*
827 : : * We don't support pushing join clauses into the quals of a seqscan, but
828 : : * it could still have required parameterization due to LATERAL refs in
829 : : * its tlist.
830 : : */
4861 831 : 209863 : required_outer = rel->lateral_relids;
832 : :
833 : : /*
834 : : * Consider TID scans.
835 : : *
836 : : * If create_tidscan_paths returns true, then a TID scan path is forced.
837 : : * This happens when rel->baserestrictinfo contains CurrentOfExpr, because
838 : : * the executor can't handle any other type of path for such queries.
839 : : * Hence, we return without adding any other paths.
840 : : */
513 rhaas@postgresql.org 841 [ + + ]: 209863 : if (create_tidscan_paths(root, rel))
842 : 202 : return;
843 : :
844 : : /* Consider sequential scan */
3689 845 : 209661 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
846 : :
847 : : /* If appropriate, consider parallel sequential scan */
3619 848 [ + + + + ]: 209661 : if (rel->consider_parallel && required_outer == NULL)
3518 tgl@sss.pgh.pa.us 849 : 155787 : create_plain_partial_paths(root, rel);
850 : :
851 : : /* Consider index scans */
7541 852 : 209661 : create_index_paths(root, rel);
853 : : }
854 : :
855 : : /*
856 : : * create_plain_partial_paths
857 : : * Build partial access paths for parallel scan of a plain relation
858 : : */
859 : : static void
3518 860 : 155787 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
861 : : {
862 : : int parallel_workers;
863 : :
2875 rhaas@postgresql.org 864 : 155787 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
865 : : max_parallel_workers_per_gather);
866 : :
867 : : /* If any limit was set to zero, the user doesn't want a parallel scan. */
3478 tgl@sss.pgh.pa.us 868 [ + + ]: 155787 : if (parallel_workers <= 0)
869 : 141581 : return;
870 : :
871 : : /* Add an unordered partial path based on a parallel sequential scan. */
rhaas@postgresql.org 872 : 14206 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
873 : : }
874 : :
875 : : /*
876 : : * set_tablesample_rel_size
877 : : * Set size estimates for a sampled relation
878 : : */
879 : : static void
3869 simon@2ndQuadrant.co 880 : 153 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
881 : : {
3798 tgl@sss.pgh.pa.us 882 : 153 : TableSampleClause *tsc = rte->tablesample;
883 : : TsmRoutine *tsm;
884 : : BlockNumber pages;
885 : : double tuples;
886 : :
887 : : /*
888 : : * Test any partial indexes of rel for applicability. We must do this
889 : : * first since partial unique indexes can affect size estimates.
890 : : */
3548 891 : 153 : check_index_predicates(root, rel);
892 : :
893 : : /*
894 : : * Call the sampling method's estimation function to estimate the number
895 : : * of pages it will read and the number of tuples it will return. (Note:
896 : : * we assume the function returns sane values.)
897 : : */
3798 898 : 153 : tsm = GetTsmRoutine(tsc->tsmhandler);
899 : 153 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
900 : : &pages, &tuples);
901 : :
902 : : /*
903 : : * For the moment, because we will only consider a SampleScan path for the
904 : : * rel, it's okay to just overwrite the pages and tuples estimates for the
905 : : * whole relation. If we ever consider multiple path types for sampled
906 : : * rels, we'll need more complication.
907 : : */
908 : 153 : rel->pages = pages;
909 : 153 : rel->tuples = tuples;
910 : :
911 : : /* Mark rel with estimated output rows, width, etc */
3869 simon@2ndQuadrant.co 912 : 153 : set_baserel_size_estimates(root, rel);
913 : 153 : }
914 : :
915 : : /*
916 : : * set_tablesample_rel_pathlist
917 : : * Build access paths for a sampled relation
918 : : */
919 : : static void
920 : 153 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
921 : : {
922 : : Relids required_outer;
923 : : Path *path;
924 : :
925 : : /*
926 : : * We don't support pushing join clauses into the quals of a samplescan,
927 : : * but it could still have required parameterization due to LATERAL refs
928 : : * in its tlist or TABLESAMPLE arguments.
929 : : */
930 : 153 : required_outer = rel->lateral_relids;
931 : :
932 : : /* Consider sampled scan */
933 : 153 : path = create_samplescan_path(root, rel, required_outer);
934 : :
935 : : /*
936 : : * If the sampling method does not support repeatable scans, we must avoid
937 : : * plans that would scan the rel multiple times. Ideally, we'd simply
938 : : * avoid putting the rel on the inside of a nestloop join; but adding such
939 : : * a consideration to the planner seems like a great deal of complication
940 : : * to support an uncommon usage of second-rate sampling methods. Instead,
941 : : * if there is a risk that the query might perform an unsafe join, just
942 : : * wrap the SampleScan in a Materialize node. We can check for joins by
943 : : * counting the membership of all_query_rels (note that this correctly
944 : : * counts inheritance trees as single rels). If we're inside a subquery,
945 : : * we can't easily check whether a join might occur in the outer query, so
946 : : * just assume one is possible.
947 : : *
948 : : * GetTsmRoutine is relatively expensive compared to the other tests here,
949 : : * so check repeatable_across_scans last, even though that's a bit odd.
950 : : */
3798 tgl@sss.pgh.pa.us 951 [ + + + + ]: 293 : if ((root->query_level > 1 ||
1052 952 : 140 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
3101 953 [ + + ]: 49 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
954 : : {
3798 955 : 4 : path = (Path *) create_material_path(rel, path);
956 : : }
957 : :
958 : 153 : add_path(rel, path);
959 : :
960 : : /* For the moment, at least, there are no other paths to consider */
3869 simon@2ndQuadrant.co 961 : 153 : }
962 : :
963 : : /*
964 : : * set_foreign_size
965 : : * Set size estimates for a foreign table RTE
966 : : */
967 : : static void
5073 tgl@sss.pgh.pa.us 968 : 1237 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
969 : : {
970 : : /* Mark rel with estimated output rows, width, etc */
971 : 1237 : set_foreign_size_estimates(root, rel);
972 : :
973 : : /* Let FDW adjust the size estimates, if it can */
5031 974 : 1237 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
975 : :
976 : : /* ... but do not let it set the rows estimate to zero */
3797 977 : 1235 : rel->rows = clamp_row_est(rel->rows);
978 : :
979 : : /*
980 : : * Also, make sure rel->tuples is not insane relative to rel->rows.
981 : : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
982 : : * FDW doesn't do anything to replace that.
983 : : */
1993 984 [ + + ]: 1235 : rel->tuples = Max(rel->tuples, rel->rows);
5073 985 : 1235 : }
986 : :
987 : : /*
988 : : * set_foreign_pathlist
989 : : * Build access paths for a foreign table RTE
990 : : */
991 : : static void
992 : 1235 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
993 : : {
994 : : /* Call the FDW's GetForeignPaths function to generate path(s) */
5031 995 : 1235 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
5073 996 : 1235 : }
997 : :
998 : : /*
999 : : * set_append_rel_size
1000 : : * Set size estimates for a simple "append relation"
1001 : : *
1002 : : * The passed-in rel and RTE represent the entire append relation. The
1003 : : * relation's contents are computed by appending together the output of the
1004 : : * individual member relations. Note that in the non-partitioned inheritance
1005 : : * case, the first member relation is actually the same table as is mentioned
1006 : : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
1007 : : * a good thing because their outputs are not the same size.
1008 : : */
1009 : : static void
1010 : 13131 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
1011 : : Index rti, RangeTblEntry *rte)
1012 : : {
8977 1013 : 13131 : int parentRTindex = rti;
1014 : : bool has_live_children;
1015 : : double parent_tuples;
1016 : : double parent_rows;
1017 : : double parent_size;
1018 : : double *parent_attrsizes;
1019 : : int nattrs;
1020 : : ListCell *l;
1021 : :
1022 : : /* Guard against stack overflow due to overly deep inheritance tree. */
3016 rhaas@postgresql.org 1023 : 13131 : check_stack_depth();
1024 : :
3180 1025 [ + + - + ]: 13131 : Assert(IS_SIMPLE_REL(rel));
1026 : :
1027 : : /*
1028 : : * If this is a partitioned baserel, set the consider_partitionwise_join
1029 : : * flag; currently, we only consider partitionwise joins with the baserel
1030 : : * if its targetlist doesn't contain a whole-row Var.
1031 : : */
2665 efujita@postgresql.o 1032 [ + + ]: 13131 : if (enable_partitionwise_join &&
1033 [ + + ]: 2497 : rel->reloptkind == RELOPT_BASEREL &&
1034 [ + - ]: 1987 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
1052 tgl@sss.pgh.pa.us 1035 [ + + ]: 1987 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
2665 efujita@postgresql.o 1036 : 1949 : rel->consider_partitionwise_join = true;
1037 : :
1038 : : /*
1039 : : * Initialize to compute size estimates for whole append relation.
1040 : : *
1041 : : * We handle tuples estimates by setting "tuples" to the total number of
1042 : : * tuples accumulated from each live child, rather than using "rows".
1043 : : * Although an appendrel itself doesn't directly enforce any quals, its
1044 : : * child relations may. Therefore, setting "tuples" equal to "rows" for
1045 : : * an appendrel isn't always appropriate, and can lead to inaccurate cost
1046 : : * estimates. For example, when estimating the number of distinct values
1047 : : * from an appendrel, we would be unable to adjust the estimate based on
1048 : : * the restriction selectivity (see estimate_num_groups).
1049 : : *
1050 : : * We handle width estimates by weighting the widths of different child
1051 : : * rels proportionally to their number of rows. This is sensible because
1052 : : * the use of width estimates is mainly to compute the total relation
1053 : : * "footprint" if we have to sort or hash it. To do this, we sum the
1054 : : * total equivalent size (in "double" arithmetic) and then divide by the
1055 : : * total rowcount estimate. This is done separately for the total rel
1056 : : * width and each attribute.
1057 : : *
1058 : : * Note: if you consider changing this logic, beware that child rels could
1059 : : * have zero rows and/or width, if they were excluded by constraints.
1060 : : */
3797 tgl@sss.pgh.pa.us 1061 : 13131 : has_live_children = false;
303 rguo@postgresql.org 1062 : 13131 : parent_tuples = 0;
6382 tgl@sss.pgh.pa.us 1063 : 13131 : parent_rows = 0;
1064 : 13131 : parent_size = 0;
1065 : 13131 : nattrs = rel->max_attr - rel->min_attr + 1;
1066 : 13131 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1067 : :
7260 1068 [ + + + + : 68794 : foreach(l, root->append_rel_list)
+ + ]
1069 : : {
1070 : 55664 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1071 : : int childRTindex;
1072 : : RangeTblEntry *childRTE;
1073 : : RelOptInfo *childrel;
1074 : : List *childrinfos;
1075 : : ListCell *parentvars;
1076 : : ListCell *childvars;
1077 : : ListCell *lc;
1078 : :
1079 : : /* append_rel_list contains all append rels; ignore others */
1080 [ + + ]: 55664 : if (appinfo->parent_relid != parentRTindex)
1081 : 26106 : continue;
1082 : :
1083 : 29729 : childRTindex = appinfo->child_relid;
6815 1084 : 29729 : childRTE = root->simple_rte_array[childRTindex];
1085 : :
1086 : : /*
1087 : : * The child rel's RelOptInfo was already created during
1088 : : * add_other_rels_to_query.
1089 : : */
7029 1090 : 29729 : childrel = find_base_rel(root, childRTindex);
1091 [ - + ]: 29729 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1092 : :
1093 : : /* We may have already proven the child to be dummy. */
2454 1094 [ + + ]: 29729 : if (IS_DUMMY_REL(childrel))
2511 alvherre@alvh.no-ip. 1095 : 9 : continue;
1096 : :
1097 : : /*
1098 : : * We have to copy the parent's targetlist and quals to the child,
1099 : : * with appropriate substitution of variables. However, the
1100 : : * baserestrictinfo quals were already copied/substituted when the
1101 : : * child RelOptInfo was built. So we don't need any additional setup
1102 : : * before applying constraint exclusion.
1103 : : */
6469 tgl@sss.pgh.pa.us 1104 [ + + ]: 29720 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1105 : : {
1106 : : /*
1107 : : * This child need not be scanned, so we can omit it from the
1108 : : * appendrel.
1109 : : */
6780 1110 : 93 : set_dummy_rel_pathlist(childrel);
6815 1111 : 93 : continue;
1112 : : }
1113 : :
1114 : : /*
1115 : : * Constraint exclusion failed, so copy the parent's join quals and
1116 : : * targetlist to the child, with appropriate variable substitutions.
1117 : : *
1118 : : * We skip join quals that came from above outer joins that can null
1119 : : * this rel, since they would be of no value while generating paths
1120 : : * for the child. This saves some effort while processing the child
1121 : : * rel, and it also avoids an implementation restriction in
1122 : : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1123 : : */
1011 1124 : 29627 : childrinfos = NIL;
1125 [ + + + + : 36227 : foreach(lc, rel->joininfo)
+ + ]
1126 : : {
1127 : 6600 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1128 : :
1129 [ + + ]: 6600 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1130 : 5439 : childrinfos = lappend(childrinfos,
1131 : 5439 : adjust_appendrel_attrs(root,
1132 : : (Node *) rinfo,
1133 : : 1, &appinfo));
1134 : : }
1135 : 29627 : childrel->joininfo = childrinfos;
1136 : :
1137 : : /*
1138 : : * Now for the child's targetlist.
1139 : : *
1140 : : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1141 : : * expressions, which otherwise would not occur in a rel's targetlist.
1142 : : * Code that might be looking at an appendrel child must cope with
1143 : : * such. (Normally, a rel's targetlist would only include Vars and
1144 : : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1145 : : * fields of childrel->reltarget; not clear if that would be useful.
1146 : : */
2522 efujita@postgresql.o 1147 : 59254 : childrel->reltarget->exprs = (List *)
1148 : 29627 : adjust_appendrel_attrs(root,
1149 : 29627 : (Node *) rel->reltarget->exprs,
1150 : : 1, &appinfo);
1151 : :
1152 : : /*
1153 : : * We have to make child entries in the EquivalenceClass data
1154 : : * structures as well. This is needed either if the parent
1155 : : * participates in some eclass joins (because we will want to consider
1156 : : * inner-indexscan joins on the individual children) or if the parent
1157 : : * has useful pathkeys (because we should try to build MergeAppend
1158 : : * paths that produce those sort orderings).
1159 : : */
1160 [ + + + + ]: 29627 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1161 : 18228 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1162 : 29627 : childrel->has_eclass_joins = rel->has_eclass_joins;
1163 : :
1164 : : /*
1165 : : * Note: we could compute appropriate attr_needed data for the child's
1166 : : * variables, by transforming the parent's attr_needed through the
1167 : : * translated_vars mapping. However, currently there's no need
1168 : : * because attr_needed is only examined for base relations not
1169 : : * otherrels. So we just leave the child's attr_needed empty.
1170 : : */
1171 : :
1172 : : /*
1173 : : * If we consider partitionwise joins with the parent rel, do the same
1174 : : * for partitioned child rels.
1175 : : *
1176 : : * Note: here we abuse the consider_partitionwise_join flag by setting
1177 : : * it for child rels that are not themselves partitioned. We do so to
1178 : : * tell try_partitionwise_join() that the child rel is sufficiently
1179 : : * valid to be used as a per-partition input, even if it later gets
1180 : : * proven to be dummy. (It's not usable until we've set up the
1181 : : * reltarget and EC entries, which we just did.)
1182 : : */
1183 [ + + ]: 29627 : if (rel->consider_partitionwise_join)
2665 1184 : 6592 : childrel->consider_partitionwise_join = true;
1185 : :
1186 : : /*
1187 : : * If parallelism is allowable for this query in general, see whether
1188 : : * it's allowable for this childrel in particular. But if we've
1189 : : * already decided the appendrel is not parallel-safe as a whole,
1190 : : * there's no point in considering parallelism for this child. For
1191 : : * consistency, do this before calling set_rel_size() for the child.
1192 : : */
3454 tgl@sss.pgh.pa.us 1193 [ + + + + ]: 29627 : if (root->glob->parallelModeOK && rel->consider_parallel)
1194 : 22248 : set_rel_consider_parallel(root, childrel, childRTE);
1195 : :
1196 : : /*
1197 : : * Compute the child's size.
1198 : : */
5073 1199 : 29627 : set_rel_size(root, childrel, childRTindex, childRTE);
1200 : :
1201 : : /*
1202 : : * It is possible that constraint exclusion detected a contradiction
1203 : : * within a child subquery, even though we didn't prove one above. If
1204 : : * so, we can skip this child.
1205 : : */
1206 [ + + ]: 29626 : if (IS_DUMMY_REL(childrel))
5198 1207 : 69 : continue;
1208 : :
1209 : : /* We have at least one live child. */
3797 1210 : 29557 : has_live_children = true;
1211 : :
1212 : : /*
1213 : : * If any live child is not parallel-safe, treat the whole appendrel
1214 : : * as not parallel-safe. In future we might be able to generate plans
1215 : : * in which some children are farmed out to workers while others are
1216 : : * not; but we don't have that today, so it's a waste to consider
1217 : : * partial paths anywhere in the appendrel unless it's all safe.
1218 : : * (Child rels visited before this one will be unmarked in
1219 : : * set_append_rel_pathlist().)
1220 : : */
3454 1221 [ + + ]: 29557 : if (!childrel->consider_parallel)
1222 : 7737 : rel->consider_parallel = false;
1223 : :
1224 : : /*
1225 : : * Accumulate size information from each live child.
1226 : : */
3797 1227 [ - + ]: 29557 : Assert(childrel->rows > 0);
1228 : :
303 rguo@postgresql.org 1229 : 29557 : parent_tuples += childrel->tuples;
3797 tgl@sss.pgh.pa.us 1230 : 29557 : parent_rows += childrel->rows;
3565 1231 : 29557 : parent_size += childrel->reltarget->width * childrel->rows;
1232 : :
1233 : : /*
1234 : : * Accumulate per-column estimates too. We need not do anything for
1235 : : * PlaceHolderVars in the parent list. If child expression isn't a
1236 : : * Var, or we didn't record a width estimate for it, we have to fall
1237 : : * back on a datatype-based estimate.
1238 : : *
1239 : : * By construction, child's targetlist is 1-to-1 with parent's.
1240 : : */
1241 [ + + + + : 98044 : forboth(parentvars, rel->reltarget->exprs,
+ + + + +
+ + - +
+ ]
1242 : : childvars, childrel->reltarget->exprs)
1243 : : {
3797 1244 : 68487 : Var *parentvar = (Var *) lfirst(parentvars);
1245 : 68487 : Node *childvar = (Node *) lfirst(childvars);
1246 : :
1722 1247 [ + + + + ]: 68487 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1248 : : {
3797 1249 : 61997 : int pndx = parentvar->varattno - rel->min_attr;
1250 : 61997 : int32 child_width = 0;
1251 : :
1252 [ + + ]: 61997 : if (IsA(childvar, Var) &&
1253 [ + + ]: 59568 : ((Var *) childvar)->varno == childrel->relid)
1254 : : {
1255 : 59535 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1256 : :
1257 : 59535 : child_width = childrel->attr_widths[cndx];
1258 : : }
1259 [ + + ]: 61997 : if (child_width <= 0)
1260 : 2462 : child_width = get_typavgwidth(exprType(childvar),
1261 : : exprTypmod(childvar));
1262 [ - + ]: 61997 : Assert(child_width > 0);
1263 : 61997 : parent_attrsizes[pndx] += child_width * childrel->rows;
1264 : : }
1265 : : }
1266 : : }
1267 : :
1268 [ + + ]: 13130 : if (has_live_children)
1269 : : {
1270 : : /*
1271 : : * Save the finished size estimates.
1272 : : */
1273 : : int i;
1274 : :
1275 [ - + ]: 12980 : Assert(parent_rows > 0);
303 rguo@postgresql.org 1276 : 12980 : rel->tuples = parent_tuples;
3797 tgl@sss.pgh.pa.us 1277 : 12980 : rel->rows = parent_rows;
3565 1278 : 12980 : rel->reltarget->width = rint(parent_size / parent_rows);
6382 1279 [ + + ]: 120862 : for (i = 0; i < nattrs; i++)
1280 : 107882 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1281 : :
1282 : : /*
1283 : : * Note that we leave rel->pages as zero; this is important to avoid
1284 : : * double-counting the appendrel tree in total_table_pages.
1285 : : */
1286 : : }
1287 : : else
1288 : : {
1289 : : /*
1290 : : * All children were excluded by constraints, so mark the whole
1291 : : * appendrel dummy. We must do this in this phase so that the rel's
1292 : : * dummy-ness is visible when we generate paths for other rels.
1293 : : */
3797 1294 : 150 : set_dummy_rel_pathlist(rel);
1295 : : }
1296 : :
6382 1297 : 13130 : pfree(parent_attrsizes);
5073 1298 : 13130 : }
1299 : :
1300 : : /*
1301 : : * set_append_rel_pathlist
1302 : : * Build access paths for an "append relation"
1303 : : */
1304 : : static void
1305 : 12980 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1306 : : Index rti, RangeTblEntry *rte)
1307 : : {
1308 : 12980 : int parentRTindex = rti;
1309 : 12980 : List *live_childrels = NIL;
1310 : : ListCell *l;
1311 : :
1312 : : /*
1313 : : * Generate access paths for each member relation, and remember the
1314 : : * non-dummy children.
1315 : : */
1316 [ + - + + : 68436 : foreach(l, root->append_rel_list)
+ + ]
1317 : : {
1318 : 55456 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1319 : : int childRTindex;
1320 : : RangeTblEntry *childRTE;
1321 : : RelOptInfo *childrel;
1322 : :
1323 : : /* append_rel_list contains all append rels; ignore others */
1324 [ + + ]: 55456 : if (appinfo->parent_relid != parentRTindex)
1325 : 25761 : continue;
1326 : :
1327 : : /* Re-locate the child RTE and RelOptInfo */
1328 : 29695 : childRTindex = appinfo->child_relid;
1329 : 29695 : childRTE = root->simple_rte_array[childRTindex];
1330 : 29695 : childrel = root->simple_rel_array[childRTindex];
1331 : :
1332 : : /*
1333 : : * If set_append_rel_size() decided the parent appendrel was
1334 : : * parallel-unsafe at some point after visiting this child rel, we
1335 : : * need to propagate the unsafety marking down to the child, so that
1336 : : * we don't generate useless partial paths for it.
1337 : : */
3454 1338 [ + + ]: 29695 : if (!rel->consider_parallel)
1339 : 7835 : childrel->consider_parallel = false;
1340 : :
1341 : : /*
1342 : : * Compute the child's access paths.
1343 : : */
5073 1344 : 29695 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1345 : :
1346 : : /*
1347 : : * If child is dummy, ignore it.
1348 : : */
1349 [ + + ]: 29695 : if (IS_DUMMY_REL(childrel))
1350 : 138 : continue;
1351 : :
1352 : : /*
1353 : : * Child is live, so add it to the live_childrels list for use below.
1354 : : */
4876 1355 : 29557 : live_childrels = lappend(live_childrels, childrel);
1356 : : }
1357 : :
1358 : : /* Add paths to the append relation. */
3200 rhaas@postgresql.org 1359 : 12980 : add_paths_to_append_rel(root, rel, live_childrels);
1360 : 12980 : }
1361 : :
1362 : : /*
1363 : : * set_grouped_rel_pathlist
1364 : : * If a grouped relation for the given 'rel' exists, build partial
1365 : : * aggregation paths for it.
1366 : : */
1367 : : static void
70 rguo@postgresql.org 1368 :GNC 271237 : set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel)
1369 : : {
1370 : : RelOptInfo *grouped_rel;
1371 : :
1372 : : /*
1373 : : * If there are no aggregate expressions or grouping expressions, eager
1374 : : * aggregation is not possible.
1375 : : */
1376 [ + + ]: 271237 : if (root->agg_clause_list == NIL ||
1377 [ + + ]: 1668 : root->group_expr_list == NIL)
1378 : 269701 : return;
1379 : :
1380 : : /* Add paths to the grouped base relation if one exists. */
1381 : 1536 : grouped_rel = rel->grouped_rel;
1382 [ + + ]: 1536 : if (grouped_rel)
1383 : : {
1384 [ - + ]: 293 : Assert(IS_GROUPED_REL(grouped_rel));
1385 : :
1386 : 293 : generate_grouped_paths(root, grouped_rel, rel);
1387 : 293 : set_cheapest(grouped_rel);
1388 : : }
1389 : : }
1390 : :
1391 : :
1392 : : /*
1393 : : * add_paths_to_append_rel
1394 : : * Generate paths for the given append relation given the set of non-dummy
1395 : : * child rels.
1396 : : *
1397 : : * The function collects all parameterizations and orderings supported by the
1398 : : * non-dummy children. For every such parameterization or ordering, it creates
1399 : : * an append path collecting one path from each non-dummy child with given
1400 : : * parameterization or ordering. Similarly it collects partial paths from
1401 : : * non-dummy children to create partial append paths.
1402 : : */
1403 : : void
3200 rhaas@postgresql.org 1404 :CBC 23726 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1405 : : List *live_childrels)
1406 : : {
1407 : 23726 : List *subpaths = NIL;
1408 : 23726 : bool subpaths_valid = true;
804 drowley@postgresql.o 1409 : 23726 : List *startup_subpaths = NIL;
1410 : 23726 : bool startup_subpaths_valid = true;
3200 rhaas@postgresql.org 1411 : 23726 : List *partial_subpaths = NIL;
2934 1412 : 23726 : List *pa_partial_subpaths = NIL;
1413 : 23726 : List *pa_nonpartial_subpaths = NIL;
3200 1414 : 23726 : bool partial_subpaths_valid = true;
1415 : : bool pa_subpaths_valid;
1416 : 23726 : List *all_child_pathkeys = NIL;
1417 : 23726 : List *all_child_outers = NIL;
1418 : : ListCell *l;
2934 1419 : 23726 : double partial_rows = -1;
1420 : :
1421 : : /* If appropriate, consider parallel append */
2737 akapila@postgresql.o 1422 [ + + + + ]: 23726 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1423 : :
1424 : : /*
1425 : : * For every non-dummy child, remember the cheapest path. Also, identify
1426 : : * all pathkeys (orderings) and parameterizations (required_outer sets)
1427 : : * available for the non-dummy member relations.
1428 : : */
3200 rhaas@postgresql.org 1429 [ + - + + : 76160 : foreach(l, live_childrels)
+ + ]
1430 : : {
1431 : 52434 : RelOptInfo *childrel = lfirst(l);
1432 : : ListCell *lcp;
2934 1433 : 52434 : Path *cheapest_partial_path = NULL;
1434 : :
1435 : : /*
1436 : : * If child has an unparameterized cheapest-total path, add that to
1437 : : * the unparameterized Append path we are constructing for the parent.
1438 : : * If not, there's no workable unparameterized path.
1439 : : *
1440 : : * With partitionwise aggregates, the child rel's pathlist may be
1441 : : * empty, so don't assume that a path exists here.
1442 : : */
2827 1443 [ + - ]: 52434 : if (childrel->pathlist != NIL &&
1444 [ + + ]: 52434 : childrel->cheapest_total_path->param_info == NULL)
2934 1445 : 52068 : accumulate_append_subpath(childrel->cheapest_total_path,
1446 : : &subpaths, NULL);
1447 : : else
4876 tgl@sss.pgh.pa.us 1448 : 366 : subpaths_valid = false;
1449 : :
1450 : : /*
1451 : : * When the planner is considering cheap startup plans, we'll also
1452 : : * collect all the cheapest_startup_paths (if set) and build an
1453 : : * AppendPath containing those as subpaths.
1454 : : */
799 drowley@postgresql.o 1455 [ + + + + ]: 52434 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1456 : 865 : {
1457 : : Path *cheapest_path;
1458 : :
1459 : : /*
1460 : : * With an indication of how many tuples the query should provide,
1461 : : * the optimizer tries to choose the path optimal for that
1462 : : * specific number of tuples.
1463 : : */
282 akorotkov@postgresql 1464 [ + - ]: 865 : if (root->tuple_fraction > 0.0)
1465 : : cheapest_path =
1466 : 865 : get_cheapest_fractional_path(childrel,
1467 : : root->tuple_fraction);
1468 : : else
282 akorotkov@postgresql 1469 :UBC 0 : cheapest_path = childrel->cheapest_startup_path;
1470 : :
1471 : : /* cheapest_startup_path must not be a parameterized path. */
282 akorotkov@postgresql 1472 [ - + ]:CBC 865 : Assert(cheapest_path->param_info == NULL);
1473 : 865 : accumulate_append_subpath(cheapest_path,
1474 : : &startup_subpaths,
1475 : : NULL);
1476 : : }
1477 : : else
804 drowley@postgresql.o 1478 : 51569 : startup_subpaths_valid = false;
1479 : :
1480 : :
1481 : : /* Same idea, but for a partial plan. */
3619 rhaas@postgresql.org 1482 [ + + ]: 52434 : if (childrel->partial_pathlist != NIL)
1483 : : {
2934 1484 : 32605 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1485 : 32605 : accumulate_append_subpath(cheapest_partial_path,
1486 : : &partial_subpaths, NULL);
1487 : : }
1488 : : else
3619 1489 : 19829 : partial_subpaths_valid = false;
1490 : :
1491 : : /*
1492 : : * Same idea, but for a parallel append mixing partial and non-partial
1493 : : * paths.
1494 : : */
2934 1495 [ + + ]: 52434 : if (pa_subpaths_valid)
1496 : : {
1497 : 39666 : Path *nppath = NULL;
1498 : :
1499 : : nppath =
1500 : 39666 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1501 : :
1502 [ + + + + ]: 39666 : if (cheapest_partial_path == NULL && nppath == NULL)
1503 : : {
1504 : : /* Neither a partial nor a parallel-safe path? Forget it. */
1505 : 277 : pa_subpaths_valid = false;
1506 : : }
1507 [ + + + + ]: 39389 : else if (nppath == NULL ||
1508 : 32380 : (cheapest_partial_path != NULL &&
1509 [ + + ]: 32380 : cheapest_partial_path->total_cost < nppath->total_cost))
1510 : : {
1511 : : /* Partial path is cheaper or the only option. */
1512 [ - + ]: 32310 : Assert(cheapest_partial_path != NULL);
1513 : 32310 : accumulate_append_subpath(cheapest_partial_path,
1514 : : &pa_partial_subpaths,
1515 : : &pa_nonpartial_subpaths);
1516 : : }
1517 : : else
1518 : : {
1519 : : /*
1520 : : * Either we've got only a non-partial path, or we think that
1521 : : * a single backend can execute the best non-partial path
1522 : : * faster than all the parallel backends working together can
1523 : : * execute the best partial path.
1524 : : *
1525 : : * It might make sense to be more aggressive here. Even if
1526 : : * the best non-partial path is more expensive than the best
1527 : : * partial path, it could still be better to choose the
1528 : : * non-partial path if there are several such paths that can
1529 : : * be given to different workers. For now, we don't try to
1530 : : * figure that out.
1531 : : */
1532 : 7079 : accumulate_append_subpath(nppath,
1533 : : &pa_nonpartial_subpaths,
1534 : : NULL);
1535 : : }
1536 : : }
1537 : :
1538 : : /*
1539 : : * Collect lists of all the available path orderings and
1540 : : * parameterizations for all the children. We use these as a
1541 : : * heuristic to indicate which sort orderings and parameterizations we
1542 : : * should build Append and MergeAppend paths for.
1543 : : */
5073 tgl@sss.pgh.pa.us 1544 [ + - + + : 123480 : foreach(lcp, childrel->pathlist)
+ + ]
1545 : : {
1546 : 71046 : Path *childpath = (Path *) lfirst(lcp);
1547 : 71046 : List *childkeys = childpath->pathkeys;
4990 1548 [ + + ]: 71046 : Relids childouter = PATH_REQ_OUTER(childpath);
1549 : :
1550 : : /* Unsorted paths don't contribute to pathkey list */
5073 1551 [ + + ]: 71046 : if (childkeys != NIL)
1552 : : {
1553 : : ListCell *lpk;
1554 : 18439 : bool found = false;
1555 : :
1556 : : /* Have we already seen this ordering? */
1557 [ + + + + : 18554 : foreach(lpk, all_child_pathkeys)
+ + ]
1558 : : {
1559 : 12464 : List *existing_pathkeys = (List *) lfirst(lpk);
1560 : :
1561 [ + + ]: 12464 : if (compare_pathkeys(existing_pathkeys,
1562 : : childkeys) == PATHKEYS_EQUAL)
1563 : : {
1564 : 12349 : found = true;
1565 : 12349 : break;
1566 : : }
1567 : : }
1568 [ + + ]: 18439 : if (!found)
1569 : : {
1570 : : /* No, so add it to all_child_pathkeys */
1571 : 6090 : all_child_pathkeys = lappend(all_child_pathkeys,
1572 : : childkeys);
1573 : : }
1574 : : }
1575 : :
1576 : : /* Unparameterized paths don't contribute to param-set list */
1577 [ + + ]: 71046 : if (childouter)
1578 : : {
1579 : : ListCell *lco;
1580 : 3304 : bool found = false;
1581 : :
1582 : : /* Have we already seen this param set? */
1583 [ + + + + : 3664 : foreach(lco, all_child_outers)
+ + ]
1584 : : {
4938 bruce@momjian.us 1585 : 2405 : Relids existing_outers = (Relids) lfirst(lco);
1586 : :
5073 tgl@sss.pgh.pa.us 1587 [ + + ]: 2405 : if (bms_equal(existing_outers, childouter))
1588 : : {
1589 : 2045 : found = true;
1590 : 2045 : break;
1591 : : }
1592 : : }
1593 [ + + ]: 3304 : if (!found)
1594 : : {
1595 : : /* No, so add it to all_child_outers */
1596 : 1259 : all_child_outers = lappend(all_child_outers,
1597 : : childouter);
1598 : : }
1599 : : }
1600 : : }
1601 : : }
1602 : :
1603 : : /*
1604 : : * If we found unparameterized paths for all children, build an unordered,
1605 : : * unparameterized Append path for the rel. (Note: this is correct even
1606 : : * if we have zero or one live subpath due to constraint exclusion.)
1607 : : */
4876 1608 [ + + ]: 23726 : if (subpaths_valid)
2811 alvherre@alvh.no-ip. 1609 : 23570 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1610 : : NIL, NULL, 0, false,
1611 : : -1));
1612 : :
1613 : : /* build an AppendPath for the cheap startup paths, if valid */
804 drowley@postgresql.o 1614 [ + + ]: 23726 : if (startup_subpaths_valid)
1615 : 350 : add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1616 : : NIL, NIL, NULL, 0, false, -1));
1617 : :
1618 : : /*
1619 : : * Consider an append of unordered, unparameterized partial paths. Make
1620 : : * it parallel-aware if possible.
1621 : : */
2477 tgl@sss.pgh.pa.us 1622 [ + + + - ]: 23726 : if (partial_subpaths_valid && partial_subpaths != NIL)
1623 : : {
1624 : : AppendPath *appendpath;
1625 : : ListCell *lc;
3478 rhaas@postgresql.org 1626 : 14181 : int parallel_workers = 0;
1627 : :
1628 : : /* Find the highest number of workers requested for any subpath. */
3619 1629 [ + - + + : 48999 : foreach(lc, partial_subpaths)
+ + ]
1630 : : {
1631 : 34818 : Path *path = lfirst(lc);
1632 : :
3478 1633 : 34818 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1634 : : }
1635 [ - + ]: 14181 : Assert(parallel_workers > 0);
1636 : :
1637 : : /*
1638 : : * If the use of parallel append is permitted, always request at least
1639 : : * log2(# of children) workers. We assume it can be useful to have
1640 : : * extra workers in this case because they will be spread out across
1641 : : * the children. The precise formula is just a guess, but we don't
1642 : : * want to end up with a radically different answer for a table with N
1643 : : * partitions vs. an unpartitioned table with the same data, so the
1644 : : * use of some kind of log-scaling here seems to make some sense.
1645 : : */
2934 1646 [ + + ]: 14181 : if (enable_parallel_append)
1647 : : {
1648 [ + + ]: 14157 : parallel_workers = Max(parallel_workers,
1649 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1650 : 14157 : parallel_workers = Min(parallel_workers,
1651 : : max_parallel_workers_per_gather);
1652 : : }
1653 [ - + ]: 14181 : Assert(parallel_workers > 0);
1654 : :
1655 : : /* Generate a partial append path. */
2811 alvherre@alvh.no-ip. 1656 : 14181 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1657 : : NIL, NULL, parallel_workers,
1658 : : enable_parallel_append,
1659 : : -1);
1660 : :
1661 : : /*
1662 : : * Make sure any subsequent partial paths use the same row count
1663 : : * estimate.
1664 : : */
2934 rhaas@postgresql.org 1665 : 14181 : partial_rows = appendpath->path.rows;
1666 : :
1667 : : /* Add the path. */
1668 : 14181 : add_partial_path(rel, (Path *) appendpath);
1669 : : }
1670 : :
1671 : : /*
1672 : : * Consider a parallel-aware append using a mix of partial and non-partial
1673 : : * paths. (This only makes sense if there's at least one child which has
1674 : : * a non-partial path that is substantially cheaper than any partial path;
1675 : : * otherwise, we should use the append path added in the previous step.)
1676 : : */
1677 [ + + + + ]: 23726 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1678 : : {
1679 : : AppendPath *appendpath;
1680 : : ListCell *lc;
1681 : 2527 : int parallel_workers = 0;
1682 : :
1683 : : /*
1684 : : * Find the highest number of workers requested for any partial
1685 : : * subpath.
1686 : : */
1687 [ + + + + : 2978 : foreach(lc, pa_partial_subpaths)
+ + ]
1688 : : {
1689 : 451 : Path *path = lfirst(lc);
1690 : :
1691 : 451 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1692 : : }
1693 : :
1694 : : /*
1695 : : * Same formula here as above. It's even more important in this
1696 : : * instance because the non-partial paths won't contribute anything to
1697 : : * the planned number of parallel workers.
1698 : : */
1699 [ + - ]: 2527 : parallel_workers = Max(parallel_workers,
1700 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1701 : 2527 : parallel_workers = Min(parallel_workers,
1702 : : max_parallel_workers_per_gather);
1703 [ - + ]: 2527 : Assert(parallel_workers > 0);
1704 : :
2811 alvherre@alvh.no-ip. 1705 : 2527 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1706 : : pa_partial_subpaths,
1707 : : NIL, NULL, parallel_workers, true,
1708 : : partial_rows);
3619 rhaas@postgresql.org 1709 : 2527 : add_partial_path(rel, (Path *) appendpath);
1710 : : }
1711 : :
1712 : : /*
1713 : : * Also build unparameterized ordered append paths based on the collected
1714 : : * list of child pathkeys.
1715 : : */
4876 tgl@sss.pgh.pa.us 1716 [ + + ]: 23726 : if (subpaths_valid)
2448 1717 : 23570 : generate_orderedappend_paths(root, rel, live_childrels,
1718 : : all_child_pathkeys);
1719 : :
1720 : : /*
1721 : : * Build Append paths for each parameterization seen among the child rels.
1722 : : * (This may look pretty expensive, but in most cases of practical
1723 : : * interest, the child rels will expose mostly the same parameterizations,
1724 : : * so that not that many cases actually get considered here.)
1725 : : *
1726 : : * The Append node itself cannot enforce quals, so all qual checking must
1727 : : * be done in the child paths. This means that to have a parameterized
1728 : : * Append path, we must have the exact same parameterization for each
1729 : : * child path; otherwise some children might be failing to check the
1730 : : * moved-down quals. To make them match up, we can try to increase the
1731 : : * parameterization of lesser-parameterized paths.
1732 : : */
5073 1733 [ + + + + : 24985 : foreach(l, all_child_outers)
+ + ]
1734 : : {
4938 bruce@momjian.us 1735 : 1259 : Relids required_outer = (Relids) lfirst(l);
1736 : : ListCell *lcr;
1737 : :
1738 : : /* Select the child paths for an Append with this parameterization */
5073 tgl@sss.pgh.pa.us 1739 : 1259 : subpaths = NIL;
4876 1740 : 1259 : subpaths_valid = true;
5073 1741 [ + - + + : 4611 : foreach(lcr, live_childrels)
+ + ]
1742 : : {
1743 : 3358 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1744 : : Path *subpath;
1745 : :
2827 rhaas@postgresql.org 1746 [ - + ]: 3358 : if (childrel->pathlist == NIL)
1747 : : {
1748 : : /* failed to make a suitable path for this child */
2827 rhaas@postgresql.org 1749 :UBC 0 : subpaths_valid = false;
1750 : 0 : break;
1751 : : }
1752 : :
4546 tgl@sss.pgh.pa.us 1753 :CBC 3358 : subpath = get_cheapest_parameterized_child_path(root,
1754 : : childrel,
1755 : : required_outer);
1756 [ + + ]: 3358 : if (subpath == NULL)
1757 : : {
1758 : : /* failed to make a suitable path for this child */
1759 : 6 : subpaths_valid = false;
1760 : 6 : break;
1761 : : }
1780 1762 : 3352 : accumulate_append_subpath(subpath, &subpaths, NULL);
1763 : : }
1764 : :
4876 1765 [ + + ]: 1259 : if (subpaths_valid)
4990 1766 : 1253 : add_path(rel, (Path *)
2811 alvherre@alvh.no-ip. 1767 : 1253 : create_append_path(root, rel, subpaths, NIL,
1768 : : NIL, required_outer, 0, false,
1769 : : -1));
1770 : : }
1771 : :
1772 : : /*
1773 : : * When there is only a single child relation, the Append path can inherit
1774 : : * any ordering available for the child rel's path, so that it's useful to
1775 : : * consider ordered partial paths. Above we only considered the cheapest
1776 : : * partial path for each child, but let's also make paths using any
1777 : : * partial paths that have pathkeys.
1778 : : */
2459 tgl@sss.pgh.pa.us 1779 [ + + ]: 23726 : if (list_length(live_childrels) == 1)
1780 : : {
1781 : 7350 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1782 : :
1783 : : /* skip the cheapest partial path, since we already used that above */
1871 drowley@postgresql.o 1784 [ + + + + : 7452 : for_each_from(l, childrel->partial_pathlist, 1)
+ + ]
1785 : : {
2459 tgl@sss.pgh.pa.us 1786 : 102 : Path *path = (Path *) lfirst(l);
1787 : : AppendPath *appendpath;
1788 : :
1789 : : /* skip paths with no pathkeys. */
1871 drowley@postgresql.o 1790 [ - + ]: 102 : if (path->pathkeys == NIL)
2459 tgl@sss.pgh.pa.us 1791 :UBC 0 : continue;
1792 : :
2459 tgl@sss.pgh.pa.us 1793 :CBC 102 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1794 : : NIL, NULL,
1795 : : path->parallel_workers, true,
1796 : : partial_rows);
1797 : 102 : add_partial_path(rel, (Path *) appendpath);
1798 : : }
1799 : : }
5073 1800 : 23726 : }
1801 : :
1802 : : /*
1803 : : * generate_orderedappend_paths
1804 : : * Generate ordered append paths for an append relation
1805 : : *
1806 : : * Usually we generate MergeAppend paths here, but there are some special
1807 : : * cases where we can generate simple Append paths, because the subpaths
1808 : : * can provide tuples in the required order already.
1809 : : *
1810 : : * We generate a path for each ordering (pathkey list) appearing in
1811 : : * all_child_pathkeys.
1812 : : *
1813 : : * We consider the cheapest-startup and cheapest-total cases, and also the
1814 : : * cheapest-fractional case when not all tuples need to be retrieved. For each
1815 : : * interesting ordering, we collect all the cheapest startup subpaths, all the
1816 : : * cheapest total paths, and, if applicable, all the cheapest fractional paths,
1817 : : * and build a suitable path for each case.
1818 : : *
1819 : : * We don't currently generate any parameterized ordered paths here. While
1820 : : * it would not take much more code here to do so, it's very unclear that it
1821 : : * is worth the planning cycles to investigate such paths: there's little
1822 : : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1823 : : * that the current coding of add_path would reject such paths out of hand,
1824 : : * because add_path gives no credit for sort ordering of parameterized paths,
1825 : : * and a parameterized MergeAppend is going to be more expensive than the
1826 : : * corresponding parameterized Append path. If we ever try harder to support
1827 : : * parameterized mergejoin plans, it might be worth adding support for
1828 : : * parameterized paths here to feed such joins. (See notes in
1829 : : * optimizer/README for why that might not ever happen, though.)
1830 : : */
1831 : : static void
2448 1832 : 23570 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1833 : : List *live_childrels,
1834 : : List *all_child_pathkeys)
1835 : : {
1836 : : ListCell *lcp;
1837 : 23570 : List *partition_pathkeys = NIL;
1838 : 23570 : List *partition_pathkeys_desc = NIL;
1839 : 23570 : bool partition_pathkeys_partial = true;
1840 : 23570 : bool partition_pathkeys_desc_partial = true;
1841 : :
1842 : : /*
1843 : : * Some partitioned table setups may allow us to use an Append node
1844 : : * instead of a MergeAppend. This is possible in cases such as RANGE
1845 : : * partitioned tables where it's guaranteed that an earlier partition must
1846 : : * contain rows which come earlier in the sort order. To detect whether
1847 : : * this is relevant, build pathkey descriptions of the partition ordering,
1848 : : * for both forward and reverse scans.
1849 : : */
1850 [ + + + + : 38167 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
+ + + + ]
1597 drowley@postgresql.o 1851 : 14597 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1852 : : {
2448 tgl@sss.pgh.pa.us 1853 : 12159 : partition_pathkeys = build_partition_pathkeys(root, rel,
1854 : : ForwardScanDirection,
1855 : : &partition_pathkeys_partial);
1856 : :
1857 : 12159 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1858 : : BackwardScanDirection,
1859 : : &partition_pathkeys_desc_partial);
1860 : :
1861 : : /*
1862 : : * You might think we should truncate_useless_pathkeys here, but
1863 : : * allowing partition keys which are a subset of the query's pathkeys
1864 : : * can often be useful. For example, consider a table partitioned by
1865 : : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1866 : : * paths that can produce the a, b, c ordering (perhaps via indexes on
1867 : : * (a, b, c)) then it works to consider the appendrel output as
1868 : : * ordered by a, b, c.
1869 : : */
1870 : : }
1871 : :
1872 : : /* Now consider each interesting sort ordering */
5073 1873 [ + + + + : 29630 : foreach(lcp, all_child_pathkeys)
+ + ]
1874 : : {
1875 : 6060 : List *pathkeys = (List *) lfirst(lcp);
5365 bruce@momjian.us 1876 : 6060 : List *startup_subpaths = NIL;
1877 : 6060 : List *total_subpaths = NIL;
1435 tomas.vondra@postgre 1878 : 6060 : List *fractional_subpaths = NIL;
5365 bruce@momjian.us 1879 : 6060 : bool startup_neq_total = false;
42 rguo@postgresql.org 1880 :GNC 6060 : bool fraction_neq_total = false;
1881 : : bool match_partition_order;
1882 : : bool match_partition_order_desc;
1883 : : int end_index;
1884 : : int first_index;
1885 : : int direction;
1886 : :
1887 : : /*
1888 : : * Determine if this sort ordering matches any partition pathkeys we
1889 : : * have, for both ascending and descending partition order. If the
1890 : : * partition pathkeys happen to be contained in pathkeys then it still
1891 : : * works, as described above, providing that the partition pathkeys
1892 : : * are complete and not just a prefix of the partition keys. (In such
1893 : : * cases we'll be relying on the child paths to have sorted the
1894 : : * lower-order columns of the required pathkeys.)
1895 : : */
2448 tgl@sss.pgh.pa.us 1896 :CBC 6060 : match_partition_order =
1897 [ + + ]: 10980 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1898 [ + + + + ]: 5024 : (!partition_pathkeys_partial &&
1899 : 104 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1900 : :
1901 [ + + + + ]: 15774 : match_partition_order_desc = !match_partition_order &&
1902 : 4866 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1903 [ + + + + ]: 4880 : (!partition_pathkeys_desc_partial &&
1904 : 32 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1905 : :
1906 : : /*
1907 : : * When the required pathkeys match the reverse of the partition
1908 : : * order, we must build the list of paths in reverse starting with the
1909 : : * last matching partition first. We can get away without making any
1910 : : * special cases for this in the loop below by just looping backward
1911 : : * over the child relations in this case.
1912 : : */
1031 drowley@postgresql.o 1913 [ + + ]: 6060 : if (match_partition_order_desc)
1914 : : {
1915 : : /* loop backward */
1916 : 24 : first_index = list_length(live_childrels) - 1;
1917 : 24 : end_index = -1;
1918 : 24 : direction = -1;
1919 : :
1920 : : /*
1921 : : * Set this to true to save us having to check for
1922 : : * match_partition_order_desc in the loop below.
1923 : : */
1924 : 24 : match_partition_order = true;
1925 : : }
1926 : : else
1927 : : {
1928 : : /* for all other case, loop forward */
1929 : 6036 : first_index = 0;
1930 : 6036 : end_index = list_length(live_childrels);
1931 : 6036 : direction = 1;
1932 : : }
1933 : :
1934 : : /* Select the child paths for this ordering... */
1935 [ + + ]: 21753 : for (int i = first_index; i != end_index; i += direction)
1936 : : {
1937 : 15693 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1938 : : Path *cheapest_startup,
1939 : : *cheapest_total,
1435 tomas.vondra@postgre 1940 : 15693 : *cheapest_fractional = NULL;
1941 : :
1942 : : /* Locate the right paths, if they are available. */
1943 : : cheapest_startup =
5543 tgl@sss.pgh.pa.us 1944 : 15693 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1945 : : pathkeys,
1946 : : NULL,
1947 : : STARTUP_COST,
1948 : : false);
1949 : : cheapest_total =
1950 : 15693 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1951 : : pathkeys,
1952 : : NULL,
1953 : : TOTAL_COST,
1954 : : false);
1955 : :
1956 : : /*
1957 : : * If we can't find any paths with the right order just use the
1958 : : * cheapest-total path; we'll have to sort it later.
1959 : : */
5073 1960 [ + + - + ]: 15693 : if (cheapest_startup == NULL || cheapest_total == NULL)
1961 : : {
4990 1962 : 170 : cheapest_startup = cheapest_total =
1963 : : childrel->cheapest_total_path;
1964 : : /* Assert we do have an unparameterized path for this child */
4876 1965 [ - + ]: 170 : Assert(cheapest_total->param_info == NULL);
1966 : : }
1967 : :
1968 : : /*
1969 : : * When building a fractional path, determine a cheapest
1970 : : * fractional path for each child relation too. Looking at startup
1971 : : * and total costs is not enough, because the cheapest fractional
1972 : : * path may be dominated by two separate paths (one for startup,
1973 : : * one for total).
1974 : : *
1975 : : * When needed (building fractional path), determine the cheapest
1976 : : * fractional path too.
1977 : : */
1435 tomas.vondra@postgre 1978 [ + + ]: 15693 : if (root->tuple_fraction > 0)
1979 : : {
213 akorotkov@postgresql 1980 : 448 : double path_fraction = root->tuple_fraction;
1981 : :
1982 : : /*
1983 : : * We should not have a dummy child relation here. However,
1984 : : * we cannot use childrel->rows to compute the tuple fraction,
1985 : : * as childrel can be an upper relation with an unset row
1986 : : * estimate. Instead, we use the row estimate from the
1987 : : * cheapest_total path, which should already have been forced
1988 : : * to a sane value.
1989 : : */
42 rguo@postgresql.org 1990 [ - + ]: 448 : Assert(cheapest_total->rows > 0);
1991 : :
1992 : : /* Convert absolute limit to a path fraction */
213 akorotkov@postgresql 1993 [ + - ]: 448 : if (path_fraction >= 1.0)
42 rguo@postgresql.org 1994 : 448 : path_fraction /= cheapest_total->rows;
1995 : :
1996 : : cheapest_fractional =
1435 tomas.vondra@postgre 1997 : 448 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1998 : : pathkeys,
1999 : : NULL,
2000 : : path_fraction);
2001 : :
2002 : : /*
2003 : : * If we found no path with matching pathkeys, use the
2004 : : * cheapest total path instead.
2005 : : *
2006 : : * XXX We might consider partially sorted paths too (with an
2007 : : * incremental sort on top). But we'd have to build all the
2008 : : * incremental paths, do the costing etc.
2009 : : *
2010 : : * Also, notice whether we actually have different paths for
2011 : : * the "fractional" and "total" cases. This helps avoid
2012 : : * generating two identical ordered append paths.
2013 : : */
42 rguo@postgresql.org 2014 [ + + ]:GNC 448 : if (cheapest_fractional == NULL)
1435 tomas.vondra@postgre 2015 :CBC 22 : cheapest_fractional = cheapest_total;
42 rguo@postgresql.org 2016 [ - + ]:GNC 426 : else if (cheapest_fractional != cheapest_total)
42 rguo@postgresql.org 2017 :UNC 0 : fraction_neq_total = true;
2018 : : }
2019 : :
2020 : : /*
2021 : : * Notice whether we actually have different paths for the
2022 : : * "cheapest" and "total" cases. This helps avoid generating two
2023 : : * identical ordered append paths.
2024 : : */
5543 tgl@sss.pgh.pa.us 2025 [ + + ]:CBC 15693 : if (cheapest_startup != cheapest_total)
2026 : 48 : startup_neq_total = true;
2027 : :
2028 : : /*
2029 : : * Collect the appropriate child paths. The required logic varies
2030 : : * for the Append and MergeAppend cases.
2031 : : */
2448 2032 [ + + ]: 15693 : if (match_partition_order)
2033 : : {
2034 : : /*
2035 : : * We're going to make a plain Append path. We don't need
2036 : : * most of what accumulate_append_subpath would do, but we do
2037 : : * want to cut out child Appends or MergeAppends if they have
2038 : : * just a single subpath (and hence aren't doing anything
2039 : : * useful).
2040 : : */
2041 : 3244 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
2042 : 3244 : cheapest_total = get_singleton_append_subpath(cheapest_total);
2043 : :
2044 : 3244 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
2045 : 3244 : total_subpaths = lappend(total_subpaths, cheapest_total);
2046 : :
1435 tomas.vondra@postgre 2047 [ + + ]: 3244 : if (cheapest_fractional)
2048 : : {
2049 : 72 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
2050 : 72 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
2051 : : }
2052 : : }
2053 : : else
2054 : : {
2055 : : /*
2056 : : * Otherwise, rely on accumulate_append_subpath to collect the
2057 : : * child paths for the MergeAppend.
2058 : : */
2448 tgl@sss.pgh.pa.us 2059 : 12449 : accumulate_append_subpath(cheapest_startup,
2060 : : &startup_subpaths, NULL);
2061 : 12449 : accumulate_append_subpath(cheapest_total,
2062 : : &total_subpaths, NULL);
2063 : :
1435 tomas.vondra@postgre 2064 [ + + ]: 12449 : if (cheapest_fractional)
2065 : 376 : accumulate_append_subpath(cheapest_fractional,
2066 : : &fractional_subpaths, NULL);
2067 : : }
2068 : : }
2069 : :
2070 : : /* ... and build the Append or MergeAppend paths */
1031 drowley@postgresql.o 2071 [ + + ]: 6060 : if (match_partition_order)
2072 : : {
2073 : : /* We only need Append */
2448 tgl@sss.pgh.pa.us 2074 : 1218 : add_path(rel, (Path *) create_append_path(root,
2075 : : rel,
2076 : : startup_subpaths,
2077 : : NIL,
2078 : : pathkeys,
2079 : : NULL,
2080 : : 0,
2081 : : false,
2082 : : -1));
2083 [ - + ]: 1218 : if (startup_neq_total)
2448 tgl@sss.pgh.pa.us 2084 :UBC 0 : add_path(rel, (Path *) create_append_path(root,
2085 : : rel,
2086 : : total_subpaths,
2087 : : NIL,
2088 : : pathkeys,
2089 : : NULL,
2090 : : 0,
2091 : : false,
2092 : : -1));
2093 : :
42 rguo@postgresql.org 2094 [ + + - + ]:GNC 1218 : if (fractional_subpaths && fraction_neq_total)
1435 tomas.vondra@postgre 2095 :LBC (36) : add_path(rel, (Path *) create_append_path(root,
2096 : : rel,
2097 : : fractional_subpaths,
2098 : : NIL,
2099 : : pathkeys,
2100 : : NULL,
2101 : : 0,
2102 : : false,
2103 : : -1));
2104 : : }
2105 : : else
2106 : : {
2107 : : /* We need MergeAppend */
5543 tgl@sss.pgh.pa.us 2108 :CBC 4842 : add_path(rel, (Path *) create_merge_append_path(root,
2109 : : rel,
2110 : : startup_subpaths,
2111 : : pathkeys,
2112 : : NULL));
2448 2113 [ + + ]: 4842 : if (startup_neq_total)
2114 : 30 : add_path(rel, (Path *) create_merge_append_path(root,
2115 : : rel,
2116 : : total_subpaths,
2117 : : pathkeys,
2118 : : NULL));
2119 : :
42 rguo@postgresql.org 2120 [ + + - + ]:GNC 4842 : if (fractional_subpaths && fraction_neq_total)
1435 tomas.vondra@postgre 2121 :LBC (134) : add_path(rel, (Path *) create_merge_append_path(root,
2122 : : rel,
2123 : : fractional_subpaths,
2124 : : pathkeys,
2125 : : NULL));
2126 : : }
2127 : : }
10753 scrappy@hub.org 2128 :CBC 23570 : }
2129 : :
2130 : : /*
2131 : : * get_cheapest_parameterized_child_path
2132 : : * Get cheapest path for this relation that has exactly the requested
2133 : : * parameterization.
2134 : : *
2135 : : * Returns NULL if unable to create such a path.
2136 : : */
2137 : : static Path *
4546 tgl@sss.pgh.pa.us 2138 : 3358 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2139 : : Relids required_outer)
2140 : : {
2141 : : Path *cheapest;
2142 : : ListCell *lc;
2143 : :
2144 : : /*
2145 : : * Look up the cheapest existing path with no more than the needed
2146 : : * parameterization. If it has exactly the needed parameterization, we're
2147 : : * done.
2148 : : */
2149 : 3358 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2150 : : NIL,
2151 : : required_outer,
2152 : : TOTAL_COST,
2153 : : false);
2154 [ - + ]: 3358 : Assert(cheapest != NULL);
2155 [ + + + + ]: 3358 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2156 : 3188 : return cheapest;
2157 : :
2158 : : /*
2159 : : * Otherwise, we can "reparameterize" an existing path to match the given
2160 : : * parameterization, which effectively means pushing down additional
2161 : : * joinquals to be checked within the path's scan. However, some existing
2162 : : * paths might check the available joinquals already while others don't;
2163 : : * therefore, it's not clear which existing path will be cheapest after
2164 : : * reparameterization. We have to go through them all and find out.
2165 : : */
2166 : 170 : cheapest = NULL;
2167 [ + - + + : 590 : foreach(lc, rel->pathlist)
+ + ]
2168 : : {
2169 : 420 : Path *path = (Path *) lfirst(lc);
2170 : :
2171 : : /* Can't use it if it needs more than requested parameterization */
2172 [ + + + + ]: 420 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2173 : 12 : continue;
2174 : :
2175 : : /*
2176 : : * Reparameterization can only increase the path's cost, so if it's
2177 : : * already more expensive than the current cheapest, forget it.
2178 : : */
2179 [ + + + + ]: 636 : if (cheapest != NULL &&
2180 : 228 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2181 : 192 : continue;
2182 : :
2183 : : /* Reparameterize if needed, then recheck cost */
2184 [ + + + + ]: 216 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2185 : : {
2186 : 178 : path = reparameterize_path(root, path, required_outer, 1.0);
2187 [ + + ]: 178 : if (path == NULL)
2188 : 16 : continue; /* failed to reparameterize this one */
2189 [ + - - + ]: 162 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2190 : :
2191 [ - + - - ]: 162 : if (cheapest != NULL &&
4546 tgl@sss.pgh.pa.us 2192 :UBC 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2193 : 0 : continue;
2194 : : }
2195 : :
2196 : : /* We have a new best path */
4546 tgl@sss.pgh.pa.us 2197 :CBC 200 : cheapest = path;
2198 : : }
2199 : :
2200 : : /* Return the best path, or NULL if we found no suitable candidate */
2201 : 170 : return cheapest;
2202 : : }
2203 : :
2204 : : /*
2205 : : * accumulate_append_subpath
2206 : : * Add a subpath to the list being built for an Append or MergeAppend.
2207 : : *
2208 : : * It's possible that the child is itself an Append or MergeAppend path, in
2209 : : * which case we can "cut out the middleman" and just add its child paths to
2210 : : * our own list. (We don't try to do this earlier because we need to apply
2211 : : * both levels of transformation to the quals.)
2212 : : *
2213 : : * Note that if we omit a child MergeAppend in this way, we are effectively
2214 : : * omitting a sort step, which seems fine: if the parent is to be an Append,
2215 : : * its result would be unsorted anyway, while if the parent is to be a
2216 : : * MergeAppend, there's no point in a separate sort on a child.
2217 : : *
2218 : : * Normally, either path is a partial path and subpaths is a list of partial
2219 : : * paths, or else path is a non-partial plan and subpaths is a list of those.
2220 : : * However, if path is a parallel-aware Append, then we add its partial path
2221 : : * children to subpaths and the rest to special_subpaths. If the latter is
2222 : : * NULL, we don't flatten the path at all (unless it contains only partial
2223 : : * paths).
2224 : : */
2225 : : static void
1780 2226 : 153553 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2227 : : {
5543 2228 [ + + ]: 153553 : if (IsA(path, AppendPath))
2229 : : {
5365 bruce@momjian.us 2230 : 7810 : AppendPath *apath = (AppendPath *) path;
2231 : :
2934 rhaas@postgresql.org 2232 [ + + + + ]: 7810 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2233 : : {
2319 tgl@sss.pgh.pa.us 2234 : 7642 : *subpaths = list_concat(*subpaths, apath->subpaths);
2934 rhaas@postgresql.org 2235 : 7642 : return;
2236 : : }
2237 [ + + ]: 168 : else if (special_subpaths != NULL)
2238 : : {
2239 : : List *new_special_subpaths;
2240 : :
2241 : : /* Split Parallel Append into partial and non-partial subpaths */
2242 : 84 : *subpaths = list_concat(*subpaths,
2243 : 84 : list_copy_tail(apath->subpaths,
2244 : : apath->first_partial_path));
1253 drowley@postgresql.o 2245 : 84 : new_special_subpaths = list_copy_head(apath->subpaths,
2246 : : apath->first_partial_path);
2934 rhaas@postgresql.org 2247 : 84 : *special_subpaths = list_concat(*special_subpaths,
2248 : : new_special_subpaths);
2898 2249 : 84 : return;
2250 : : }
2251 : : }
4282 tgl@sss.pgh.pa.us 2252 [ + + ]: 145743 : else if (IsA(path, MergeAppendPath))
2253 : : {
2254 : 538 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2255 : :
2319 2256 : 538 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2934 rhaas@postgresql.org 2257 : 538 : return;
2258 : : }
2259 : :
2260 : 145289 : *subpaths = lappend(*subpaths, path);
2261 : : }
2262 : :
2263 : : /*
2264 : : * get_singleton_append_subpath
2265 : : * Returns the single subpath of an Append/MergeAppend, or just
2266 : : * return 'path' if it's not a single sub-path Append/MergeAppend.
2267 : : *
2268 : : * Note: 'path' must not be a parallel-aware path.
2269 : : */
2270 : : static Path *
2448 tgl@sss.pgh.pa.us 2271 : 6560 : get_singleton_append_subpath(Path *path)
2272 : : {
2273 [ - + ]: 6560 : Assert(!path->parallel_aware);
2274 : :
2275 [ + + ]: 6560 : if (IsA(path, AppendPath))
2276 : : {
2277 : 194 : AppendPath *apath = (AppendPath *) path;
2278 : :
2279 [ + + ]: 194 : if (list_length(apath->subpaths) == 1)
2280 : 96 : return (Path *) linitial(apath->subpaths);
2281 : : }
2282 [ + + ]: 6366 : else if (IsA(path, MergeAppendPath))
2283 : : {
2284 : 174 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2285 : :
2286 [ - + ]: 174 : if (list_length(mpath->subpaths) == 1)
2448 tgl@sss.pgh.pa.us 2287 :UBC 0 : return (Path *) linitial(mpath->subpaths);
2288 : : }
2289 : :
2448 tgl@sss.pgh.pa.us 2290 :CBC 6464 : return path;
2291 : : }
2292 : :
2293 : : /*
2294 : : * set_dummy_rel_pathlist
2295 : : * Build a dummy path for a relation that's been excluded by constraints
2296 : : *
2297 : : * Rather than inventing a special "dummy" path type, we represent this as an
2298 : : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2299 : : *
2300 : : * (See also mark_dummy_rel, which does basically the same thing, but is
2301 : : * typically used to change a rel into dummy state after we already made
2302 : : * paths for it.)
2303 : : */
2304 : : static void
6780 2305 : 659 : set_dummy_rel_pathlist(RelOptInfo *rel)
2306 : : {
2307 : : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2308 : 659 : rel->rows = 0;
3565 2309 : 659 : rel->reltarget->width = 0;
2310 : :
2311 : : /* Discard any pre-existing paths; no further need for them */
5073 2312 : 659 : rel->pathlist = NIL;
3619 rhaas@postgresql.org 2313 : 659 : rel->partial_pathlist = NIL;
2314 : :
2315 : : /* Set up the dummy path */
2470 tgl@sss.pgh.pa.us 2316 : 659 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2317 : : NIL, rel->lateral_relids,
2318 : : 0, false, -1));
2319 : :
2320 : : /*
2321 : : * We set the cheapest-path fields immediately, just in case they were
2322 : : * pointing at some discarded path. This is redundant in current usage
2323 : : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2324 : : * for safety and consistency with mark_dummy_rel.
2325 : : */
6780 2326 : 659 : set_cheapest(rel);
2327 : 659 : }
2328 : :
2329 : : /*
2330 : : * find_window_run_conditions
2331 : : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2332 : : * function to determine the function's monotonic properties. We then
2333 : : * see if 'opexpr' can be used to short-circuit execution.
2334 : : *
2335 : : * For example row_number() over (order by ...) always produces a value one
2336 : : * higher than the previous. If someone has a window function in a subquery
2337 : : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2338 : : * as well stop processing the windowagg once the row number reaches 11. Here
2339 : : * we check if 'opexpr' might help us to stop doing needless extra processing
2340 : : * in WindowAgg nodes.
2341 : : *
2342 : : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2343 : : * its original purpose. This is set to false if the caller can assume that
2344 : : * the run condition will handle all of the required filtering.
2345 : : *
2346 : : * Returns true if 'opexpr' was found to be useful and was added to the
2347 : : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2348 : : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2349 : : * If the 'opexpr' cannot be used then we set *keep_original to true and
2350 : : * return false.
2351 : : */
2352 : : static bool
78 drowley@postgresql.o 2353 :GNC 120 : find_window_run_conditions(Query *subquery, AttrNumber attno,
2354 : : WindowFunc *wfunc, OpExpr *opexpr, bool wfunc_left,
2355 : : bool *keep_original, Bitmapset **run_cond_attrs)
2356 : : {
2357 : : Oid prosupport;
2358 : : Expr *otherexpr;
2359 : : SupportRequestWFuncMonotonic req;
2360 : : SupportRequestWFuncMonotonic *res;
2361 : : WindowClause *wclause;
2362 : : List *opinfos;
2363 : : OpExpr *runopexpr;
2364 : : Oid runoperator;
2365 : : ListCell *lc;
2366 : :
1349 drowley@postgresql.o 2367 :CBC 120 : *keep_original = true;
2368 : :
2369 [ - + ]: 120 : while (IsA(wfunc, RelabelType))
1349 drowley@postgresql.o 2370 :UBC 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2371 : :
2372 : : /* we can only work with window functions */
1349 drowley@postgresql.o 2373 [ + + ]:CBC 120 : if (!IsA(wfunc, WindowFunc))
2374 : 12 : return false;
2375 : :
2376 : : /* can't use it if there are subplans in the WindowFunc */
1006 2377 [ + + ]: 108 : if (contain_subplans((Node *) wfunc))
2378 : 3 : return false;
2379 : :
1349 2380 : 105 : prosupport = get_func_support(wfunc->winfnoid);
2381 : :
2382 : : /* Check if there's a support function for 'wfunc' */
2383 [ + + ]: 105 : if (!OidIsValid(prosupport))
2384 : 9 : return false;
2385 : :
2386 : : /* get the Expr from the other side of the OpExpr */
2387 [ + + ]: 96 : if (wfunc_left)
2388 : 84 : otherexpr = lsecond(opexpr->args);
2389 : : else
2390 : 12 : otherexpr = linitial(opexpr->args);
2391 : :
2392 : : /*
2393 : : * The value being compared must not change during the evaluation of the
2394 : : * window partition.
2395 : : */
2396 [ - + ]: 96 : if (!is_pseudo_constant_clause((Node *) otherexpr))
1349 drowley@postgresql.o 2397 :UBC 0 : return false;
2398 : :
2399 : : /* find the window clause belonging to the window function */
1349 drowley@postgresql.o 2400 :CBC 96 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2401 : 96 : wfunc->winref - 1);
2402 : :
2403 : 96 : req.type = T_SupportRequestWFuncMonotonic;
2404 : 96 : req.window_func = wfunc;
2405 : 96 : req.window_clause = wclause;
2406 : :
2407 : : /* call the support function */
2408 : : res = (SupportRequestWFuncMonotonic *)
2409 : 96 : DatumGetPointer(OidFunctionCall1(prosupport,
2410 : : PointerGetDatum(&req)));
2411 : :
2412 : : /*
2413 : : * Nothing to do if the function is neither monotonically increasing nor
2414 : : * monotonically decreasing.
2415 : : */
2416 [ + - - + ]: 96 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
1349 drowley@postgresql.o 2417 :UBC 0 : return false;
2418 : :
1349 drowley@postgresql.o 2419 :CBC 96 : runopexpr = NULL;
2420 : 96 : runoperator = InvalidOid;
255 peter@eisentraut.org 2421 : 96 : opinfos = get_op_index_interpretation(opexpr->opno);
2422 : :
1349 drowley@postgresql.o 2423 [ + - + - : 96 : foreach(lc, opinfos)
+ - ]
2424 : : {
255 peter@eisentraut.org 2425 : 96 : OpIndexInterpretation *opinfo = (OpIndexInterpretation *) lfirst(lc);
2426 : 96 : CompareType cmptype = opinfo->cmptype;
2427 : :
2428 : : /* handle < / <= */
2429 [ + + + + ]: 96 : if (cmptype == COMPARE_LT || cmptype == COMPARE_LE)
2430 : : {
2431 : : /*
2432 : : * < / <= is supported for monotonically increasing functions in
2433 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2434 : : * for monotonically decreasing functions.
2435 : : */
1349 drowley@postgresql.o 2436 [ + + + + ]: 69 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2437 [ + + + + ]: 9 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2438 : : {
2439 : 63 : *keep_original = false;
2440 : 63 : runopexpr = opexpr;
2441 : 63 : runoperator = opexpr->opno;
2442 : : }
2443 : 69 : break;
2444 : : }
2445 : : /* handle > / >= */
255 peter@eisentraut.org 2446 [ + + + + ]: 27 : else if (cmptype == COMPARE_GT || cmptype == COMPARE_GE)
2447 : : {
2448 : : /*
2449 : : * > / >= is supported for monotonically decreasing functions in
2450 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2451 : : * for monotonically increasing functions.
2452 : : */
1349 drowley@postgresql.o 2453 [ + + - + ]: 9 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2454 [ + - + - ]: 6 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2455 : : {
2456 : 9 : *keep_original = false;
2457 : 9 : runopexpr = opexpr;
2458 : 9 : runoperator = opexpr->opno;
2459 : : }
2460 : 9 : break;
2461 : : }
2462 : : /* handle = */
255 peter@eisentraut.org 2463 [ + - ]: 18 : else if (cmptype == COMPARE_EQ)
2464 : : {
2465 : : CompareType newcmptype;
2466 : :
2467 : : /*
2468 : : * When both monotonically increasing and decreasing then the
2469 : : * return value of the window function will be the same each time.
2470 : : * We can simply use 'opexpr' as the run condition without
2471 : : * modifying it.
2472 : : */
1349 drowley@postgresql.o 2473 [ + + ]: 18 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2474 : : {
2475 : 3 : *keep_original = false;
2476 : 3 : runopexpr = opexpr;
1230 2477 : 3 : runoperator = opexpr->opno;
1349 2478 : 3 : break;
2479 : : }
2480 : :
2481 : : /*
2482 : : * When monotonically increasing we make a qual with <wfunc> <=
2483 : : * <value> or <value> >= <wfunc> in order to filter out values
2484 : : * which are above the value in the equality condition. For
2485 : : * monotonically decreasing functions we want to filter values
2486 : : * below the value in the equality condition.
2487 : : */
2488 [ + - ]: 15 : if (res->monotonic & MONOTONICFUNC_INCREASING)
255 peter@eisentraut.org 2489 [ + - ]: 15 : newcmptype = wfunc_left ? COMPARE_LE : COMPARE_GE;
2490 : : else
255 peter@eisentraut.org 2491 [ # # ]:UBC 0 : newcmptype = wfunc_left ? COMPARE_GE : COMPARE_LE;
2492 : :
2493 : : /* We must keep the original equality qual */
1349 drowley@postgresql.o 2494 :CBC 15 : *keep_original = true;
2495 : 15 : runopexpr = opexpr;
2496 : :
2497 : : /* determine the operator to use for the WindowFuncRunCondition */
255 peter@eisentraut.org 2498 : 15 : runoperator = get_opfamily_member_for_cmptype(opinfo->opfamily_id,
2499 : : opinfo->oplefttype,
2500 : : opinfo->oprighttype,
2501 : : newcmptype);
1349 drowley@postgresql.o 2502 : 15 : break;
2503 : : }
2504 : : }
2505 : :
2506 [ + + ]: 96 : if (runopexpr != NULL)
2507 : : {
2508 : : WindowFuncRunCondition *wfuncrc;
2509 : :
591 2510 : 90 : wfuncrc = makeNode(WindowFuncRunCondition);
2511 : 90 : wfuncrc->opno = runoperator;
2512 : 90 : wfuncrc->inputcollid = runopexpr->inputcollid;
2513 : 90 : wfuncrc->wfunc_left = wfunc_left;
2514 : 90 : wfuncrc->arg = copyObject(otherexpr);
2515 : :
2516 : 90 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2517 : :
2518 : : /* record that this attno was used in a run condition */
1300 2519 : 90 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2520 : : attno - FirstLowInvalidHeapAttributeNumber);
1349 2521 : 90 : return true;
2522 : : }
2523 : :
2524 : : /* unsupported OpExpr */
2525 : 6 : return false;
2526 : : }
2527 : :
2528 : : /*
2529 : : * check_and_push_window_quals
2530 : : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2531 : : * as a 'runCondition' qual. These, when present, allow some unnecessary
2532 : : * work to be skipped during execution.
2533 : : *
2534 : : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2535 : : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2536 : : * window quals for.
2537 : : *
2538 : : * Returns true if the caller still must keep the original qual or false if
2539 : : * the caller can safely ignore the original qual because the WindowAgg node
2540 : : * will use the runCondition to stop returning tuples.
2541 : : */
2542 : : static bool
78 drowley@postgresql.o 2543 :GNC 126 : check_and_push_window_quals(Query *subquery, Node *clause,
2544 : : Bitmapset **run_cond_attrs)
2545 : : {
1349 drowley@postgresql.o 2546 :CBC 126 : OpExpr *opexpr = (OpExpr *) clause;
2547 : 126 : bool keep_original = true;
2548 : : Var *var1;
2549 : : Var *var2;
2550 : :
2551 : : /* We're only able to use OpExprs with 2 operands */
2552 [ + + ]: 126 : if (!IsA(opexpr, OpExpr))
2553 : 9 : return true;
2554 : :
2555 [ - + ]: 117 : if (list_length(opexpr->args) != 2)
1349 drowley@postgresql.o 2556 :UBC 0 : return true;
2557 : :
2558 : : /*
2559 : : * Currently, we restrict this optimization to strict OpExprs. The reason
2560 : : * for this is that during execution, once the runcondition becomes false,
2561 : : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2562 : : * function result values, we set them to NULL. Having only strict
2563 : : * OpExprs here ensures that we properly filter out the tuples with NULLs
2564 : : * in the top-level WindowAgg.
2565 : : */
1106 drowley@postgresql.o 2566 :CBC 117 : set_opfuncid(opexpr);
2567 [ - + ]: 117 : if (!func_strict(opexpr->opfuncid))
1106 drowley@postgresql.o 2568 :UBC 0 : return true;
2569 : :
2570 : : /*
2571 : : * Check for plain Vars that reference window functions in the subquery.
2572 : : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2573 : : * be used as part of the run condition.
2574 : : */
2575 : :
2576 : : /* Check the left side of the OpExpr */
1349 drowley@postgresql.o 2577 :CBC 117 : var1 = linitial(opexpr->args);
2578 [ + + + - ]: 117 : if (IsA(var1, Var) && var1->varattno > 0)
2579 : : {
2580 : 99 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2581 : 99 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2582 : :
78 drowley@postgresql.o 2583 [ + + ]:GNC 99 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2584 : : true, &keep_original, run_cond_attrs))
1349 drowley@postgresql.o 2585 :CBC 81 : return keep_original;
2586 : : }
2587 : :
2588 : : /* and check the right side */
2589 : 36 : var2 = lsecond(opexpr->args);
2590 [ + + + - ]: 36 : if (IsA(var2, Var) && var2->varattno > 0)
2591 : : {
2592 : 21 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2593 : 21 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2594 : :
78 drowley@postgresql.o 2595 [ + + ]:GNC 21 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2596 : : false, &keep_original, run_cond_attrs))
1349 drowley@postgresql.o 2597 :CBC 9 : return keep_original;
2598 : : }
2599 : :
2600 : 27 : return true;
2601 : : }
2602 : :
2603 : : /*
2604 : : * set_subquery_pathlist
2605 : : * Generate SubqueryScan access paths for a subquery RTE
2606 : : *
2607 : : * We don't currently support generating parameterized paths for subqueries
2608 : : * by pushing join clauses down into them; it seems too expensive to re-plan
2609 : : * the subquery multiple times to consider different alternatives.
2610 : : * (XXX that could stand to be reconsidered, now that we use Paths.)
2611 : : * So the paths made here will be parameterized if the subquery contains
2612 : : * LATERAL references, otherwise not. As long as that's true, there's no need
2613 : : * for a separate set_subquery_size phase: just make the paths right away.
2614 : : */
2615 : : static void
7500 tgl@sss.pgh.pa.us 2616 : 11713 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2617 : : Index rti, RangeTblEntry *rte)
2618 : : {
7495 2619 : 11713 : Query *parse = root->parse;
8920 2620 : 11713 : Query *subquery = rte->subquery;
2621 : : bool trivial_pathtarget;
2622 : : Relids required_outer;
2623 : : pushdown_safety_info safetyInfo;
2624 : : double tuple_fraction;
2625 : : RelOptInfo *sub_final_rel;
1300 drowley@postgresql.o 2626 : 11713 : Bitmapset *run_cond_attrs = NULL;
2627 : : ListCell *lc;
2628 : : char *plan_name;
2629 : :
2630 : : /*
2631 : : * Must copy the Query so that planning doesn't mess up the RTE contents
2632 : : * (really really need to fix the planner to not scribble on its input,
2633 : : * someday ... but see remove_unused_subquery_outputs to start with).
2634 : : */
6126 tgl@sss.pgh.pa.us 2635 : 11713 : subquery = copyObject(subquery);
2636 : :
2637 : : /*
2638 : : * If it's a LATERAL subquery, it might contain some Vars of the current
2639 : : * query level, requiring it to be treated as parameterized, even though
2640 : : * we don't support pushing down join quals into subqueries.
2641 : : */
4861 2642 : 11713 : required_outer = rel->lateral_relids;
2643 : :
2644 : : /*
2645 : : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2646 : : * flags as needed while recursing. In particular, we need a workspace
2647 : : * for keeping track of the reasons why columns are unsafe to reference.
2648 : : * These reasons are stored in the bits inside unsafeFlags[i] when we
2649 : : * discover reasons that column i of the subquery is unsafe to be used in
2650 : : * a pushed-down qual.
2651 : : */
4191 2652 : 11713 : memset(&safetyInfo, 0, sizeof(safetyInfo));
1006 drowley@postgresql.o 2653 : 11713 : safetyInfo.unsafeFlags = (unsigned char *)
2654 : 11713 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2655 : :
2656 : : /*
2657 : : * If the subquery has the "security_barrier" flag, it means the subquery
2658 : : * originated from a view that must enforce row-level security. Then we
2659 : : * must not push down quals that contain leaky functions. (Ideally this
2660 : : * would be checked inside subquery_is_pushdown_safe, but since we don't
2661 : : * currently pass the RTE to that function, we must do it here.)
2662 : : */
4191 tgl@sss.pgh.pa.us 2663 : 11713 : safetyInfo.unsafeLeaky = rte->security_barrier;
2664 : :
2665 : : /*
2666 : : * If there are any restriction clauses that have been attached to the
2667 : : * subquery relation, consider pushing them down to become WHERE or HAVING
2668 : : * quals of the subquery itself. This transformation is useful because it
2669 : : * may allow us to generate a better plan for the subquery than evaluating
2670 : : * all the subquery output rows and then filtering them.
2671 : : *
2672 : : * There are several cases where we cannot push down clauses. Restrictions
2673 : : * involving the subquery are checked by subquery_is_pushdown_safe().
2674 : : * Restrictions on individual clauses are checked by
2675 : : * qual_is_pushdown_safe(). Also, we don't want to push down
2676 : : * pseudoconstant clauses; better to have the gating node above the
2677 : : * subquery.
2678 : : *
2679 : : * Non-pushed-down clauses will get evaluated as qpquals of the
2680 : : * SubqueryScan node.
2681 : : *
2682 : : * XXX Are there any cases where we want to make a policy decision not to
2683 : : * push down a pushable qual, because it'd result in a worse plan?
2684 : : */
8511 2685 [ + + + + ]: 12713 : if (rel->baserestrictinfo != NIL &&
4191 2686 : 1000 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2687 : : {
2688 : : /* OK to consider pushing down individual quals */
8920 2689 : 927 : List *upperrestrictlist = NIL;
2690 : : ListCell *l;
2691 : :
7875 neilc@samurai.com 2692 [ + - + + : 2330 : foreach(l, rel->baserestrictinfo)
+ + ]
2693 : : {
2694 : 1403 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
1349 drowley@postgresql.o 2695 : 1403 : Node *clause = (Node *) rinfo->clause;
2696 : :
1006 2697 [ + + ]: 1403 : if (rinfo->pseudoconstant)
2698 : : {
2699 : 2 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2700 : 2 : continue;
2701 : : }
2702 : :
2703 [ + + + - ]: 1401 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2704 : : {
2705 : 1048 : case PUSHDOWN_SAFE:
2706 : : /* Push it down */
2707 : 1048 : subquery_push_qual(subquery, rte, rti, clause);
2708 : 1048 : break;
2709 : :
2710 : 126 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2711 : :
2712 : : /*
2713 : : * Since we can't push the qual down into the subquery,
2714 : : * check if it happens to reference a window function. If
2715 : : * so then it might be useful to use for the WindowAgg's
2716 : : * runCondition.
2717 : : */
2718 [ + - + + ]: 252 : if (!subquery->hasWindowFuncs ||
78 drowley@postgresql.o 2719 :GNC 126 : check_and_push_window_quals(subquery, clause,
2720 : : &run_cond_attrs))
2721 : : {
2722 : : /*
2723 : : * subquery has no window funcs or the clause is not a
2724 : : * suitable window run condition qual or it is, but
2725 : : * the original must also be kept in the upper query.
2726 : : */
1006 drowley@postgresql.o 2727 :CBC 51 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2728 : : }
2729 : 126 : break;
2730 : :
2731 : 227 : case PUSHDOWN_UNSAFE:
1349 2732 : 227 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
1006 2733 : 227 : break;
2734 : : }
2735 : : }
8920 tgl@sss.pgh.pa.us 2736 : 927 : rel->baserestrictinfo = upperrestrictlist;
2737 : : /* We don't bother recomputing baserestrict_min_security */
2738 : : }
2739 : :
1006 drowley@postgresql.o 2740 : 11713 : pfree(safetyInfo.unsafeFlags);
2741 : :
2742 : : /*
2743 : : * The upper query might not use all the subquery's output columns; if
2744 : : * not, we can simplify. Pass the attributes that were pushed down into
2745 : : * WindowAgg run conditions to ensure we don't accidentally think those
2746 : : * are unused.
2747 : : */
1300 2748 : 11713 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2749 : :
2750 : : /*
2751 : : * We can safely pass the outer tuple_fraction down to the subquery if the
2752 : : * outer level has no joining, aggregation, or sorting to do. Otherwise
2753 : : * we'd better tell the subquery to plan for full retrieval. (XXX This
2754 : : * could probably be made more intelligent ...)
2755 : : */
7495 tgl@sss.pgh.pa.us 2756 [ + + ]: 11713 : if (parse->hasAggs ||
2757 [ + + ]: 8321 : parse->groupClause ||
3868 andres@anarazel.de 2758 [ + - ]: 8312 : parse->groupingSets ||
1156 tgl@sss.pgh.pa.us 2759 [ + - ]: 8312 : root->hasHavingQual ||
7495 2760 [ + + ]: 8312 : parse->distinctClause ||
2761 [ + + + + ]: 11271 : parse->sortClause ||
799 2762 : 3217 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
7495 2763 : 9174 : tuple_fraction = 0.0; /* default case */
2764 : : else
2765 : 2539 : tuple_fraction = root->tuple_fraction;
2766 : :
2767 : : /* plan_params should not be in use in current query level */
4851 2768 [ - + ]: 11713 : Assert(root->plan_params == NIL);
2769 : :
2770 : : /* Generate a subroot and Paths for the subquery */
71 rhaas@postgresql.org 2771 :GNC 11713 : plan_name = choose_plan_name(root->glob, rte->eref->aliasname, false);
2772 : 11713 : rel->subroot = subquery_planner(root->glob, subquery, plan_name,
2773 : : root, false, tuple_fraction, NULL);
2774 : :
2775 : : /* Isolate the params needed by this specific subplan */
4851 tgl@sss.pgh.pa.us 2776 :CBC 11713 : rel->subplan_params = root->plan_params;
2777 : 11713 : root->plan_params = NIL;
2778 : :
2779 : : /*
2780 : : * It's possible that constraint exclusion proved the subquery empty. If
2781 : : * so, it's desirable to produce an unadorned dummy path so that we will
2782 : : * recognize appropriate optimizations at this query level.
2783 : : */
3572 2784 : 11713 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2785 : :
2786 [ + + ]: 11713 : if (IS_DUMMY_REL(sub_final_rel))
2787 : : {
5198 2788 : 63 : set_dummy_rel_pathlist(rel);
2789 : 63 : return;
2790 : : }
2791 : :
2792 : : /*
2793 : : * Mark rel with estimated output rows, width, etc. Note that we have to
2794 : : * do this before generating outer-query paths, else cost_subqueryscan is
2795 : : * not happy.
2796 : : */
5219 2797 : 11650 : set_subquery_size_estimates(root, rel);
2798 : :
2799 : : /*
2800 : : * Also detect whether the reltarget is trivial, so that we can pass that
2801 : : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2802 : : * It's trivial if it fetches all the subplan output columns in order.
2803 : : */
1247 2804 [ + + ]: 11650 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2805 : 6707 : trivial_pathtarget = false;
2806 : : else
2807 : : {
2808 : 4943 : trivial_pathtarget = true;
2809 [ + + + + : 16100 : foreach(lc, rel->reltarget->exprs)
+ + ]
2810 : : {
2811 : 11306 : Node *node = (Node *) lfirst(lc);
2812 : : Var *var;
2813 : :
2814 [ - + ]: 11306 : if (!IsA(node, Var))
2815 : : {
1247 tgl@sss.pgh.pa.us 2816 :UBC 0 : trivial_pathtarget = false;
2817 : 0 : break;
2818 : : }
1247 tgl@sss.pgh.pa.us 2819 :CBC 11306 : var = (Var *) node;
2820 [ + - ]: 11306 : if (var->varno != rti ||
2821 [ + + ]: 11306 : var->varattno != foreach_current_index(lc) + 1)
2822 : : {
2823 : 149 : trivial_pathtarget = false;
2824 : 149 : break;
2825 : : }
2826 : : }
2827 : : }
2828 : :
2829 : : /*
2830 : : * For each Path that subquery_planner produced, make a SubqueryScanPath
2831 : : * in the outer query.
2832 : : */
3572 2833 [ + - + + : 24313 : foreach(lc, sub_final_rel->pathlist)
+ + ]
2834 : : {
2835 : 12663 : Path *subpath = (Path *) lfirst(lc);
2836 : : List *pathkeys;
2837 : :
2838 : : /* Convert subpath's pathkeys to outer representation */
2839 : 12663 : pathkeys = convert_subquery_pathkeys(root,
2840 : : rel,
2841 : : subpath->pathkeys,
2842 : : make_tlist_from_pathtarget(subpath->pathtarget));
2843 : :
2844 : : /* Generate outer path using this subpath */
2845 : 12663 : add_path(rel, (Path *)
2846 : 12663 : create_subqueryscan_path(root, rel, subpath,
2847 : : trivial_pathtarget,
2848 : : pathkeys, required_outer));
2849 : : }
2850 : :
2851 : : /* If outer rel allows parallelism, do same for partial paths. */
2793 rhaas@postgresql.org 2852 [ + + + + ]: 11650 : if (rel->consider_parallel && bms_is_empty(required_outer))
2853 : : {
2854 : : /* If consider_parallel is false, there should be no partial paths. */
2855 [ + + - + ]: 6738 : Assert(sub_final_rel->consider_parallel ||
2856 : : sub_final_rel->partial_pathlist == NIL);
2857 : :
2858 : : /* Same for partial paths. */
2859 [ + + + + : 6759 : foreach(lc, sub_final_rel->partial_pathlist)
+ + ]
2860 : : {
2861 : 21 : Path *subpath = (Path *) lfirst(lc);
2862 : : List *pathkeys;
2863 : :
2864 : : /* Convert subpath's pathkeys to outer representation */
2865 : 21 : pathkeys = convert_subquery_pathkeys(root,
2866 : : rel,
2867 : : subpath->pathkeys,
2868 : : make_tlist_from_pathtarget(subpath->pathtarget));
2869 : :
2870 : : /* Generate outer path using this subpath */
2871 : 21 : add_partial_path(rel, (Path *)
2872 : 21 : create_subqueryscan_path(root, rel, subpath,
2873 : : trivial_pathtarget,
2874 : : pathkeys,
2875 : : required_outer));
2876 : : }
2877 : : }
2878 : : }
2879 : :
2880 : : /*
2881 : : * set_function_pathlist
2882 : : * Build the (single) access path for a function RTE
2883 : : */
2884 : : static void
7500 tgl@sss.pgh.pa.us 2885 : 25133 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2886 : : {
2887 : : Relids required_outer;
4409 2888 : 25133 : List *pathkeys = NIL;
2889 : :
2890 : : /*
2891 : : * We don't support pushing join clauses into the quals of a function
2892 : : * scan, but it could still have required parameterization due to LATERAL
2893 : : * refs in the function expression.
2894 : : */
4861 2895 : 25133 : required_outer = rel->lateral_relids;
2896 : :
2897 : : /*
2898 : : * The result is considered unordered unless ORDINALITY was used, in which
2899 : : * case it is ordered by the ordinal column (the last one). See if we
2900 : : * care, by checking for uses of that Var in equivalence classes.
2901 : : */
4409 2902 [ + + ]: 25133 : if (rte->funcordinality)
2903 : : {
2904 : 479 : AttrNumber ordattno = rel->max_attr;
2905 : 479 : Var *var = NULL;
2906 : : ListCell *lc;
2907 : :
2908 : : /*
2909 : : * Is there a Var for it in rel's targetlist? If not, the query did
2910 : : * not reference the ordinality column, or at least not in any way
2911 : : * that would be interesting for sorting.
2912 : : */
3565 2913 [ + - + + : 1071 : foreach(lc, rel->reltarget->exprs)
+ + ]
2914 : : {
4409 2915 : 1068 : Var *node = (Var *) lfirst(lc);
2916 : :
2917 : : /* checking varno/varlevelsup is just paranoia */
2918 [ + - ]: 1068 : if (IsA(node, Var) &&
2919 [ + + ]: 1068 : node->varattno == ordattno &&
2920 [ + - ]: 476 : node->varno == rel->relid &&
2921 [ + - ]: 476 : node->varlevelsup == 0)
2922 : : {
2923 : 476 : var = node;
2924 : 476 : break;
2925 : : }
2926 : : }
2927 : :
2928 : : /*
2929 : : * Try to build pathkeys for this Var with int8 sorting. We tell
2930 : : * build_expression_pathkey not to build any new equivalence class; if
2931 : : * the Var isn't already mentioned in some EC, it means that nothing
2932 : : * cares about the ordering.
2933 : : */
2934 [ + + ]: 479 : if (var)
2935 : 476 : pathkeys = build_expression_pathkey(root,
2936 : : (Expr *) var,
2937 : : Int8LessOperator,
2938 : : rel->relids,
2939 : : false);
2940 : : }
2941 : :
2942 : : /* Generate appropriate path */
2943 : 25133 : add_path(rel, create_functionscan_path(root, rel,
2944 : : pathkeys, required_outer));
8620 2945 : 25133 : }
2946 : :
2947 : : /*
2948 : : * set_values_pathlist
2949 : : * Build the (single) access path for a VALUES RTE
2950 : : */
2951 : : static void
7077 mail@joeconway.com 2952 : 4236 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2953 : : {
2954 : : Relids required_outer;
2955 : :
2956 : : /*
2957 : : * We don't support pushing join clauses into the quals of a values scan,
2958 : : * but it could still have required parameterization due to LATERAL refs
2959 : : * in the values expressions.
2960 : : */
4861 tgl@sss.pgh.pa.us 2961 : 4236 : required_outer = rel->lateral_relids;
2962 : :
2963 : : /* Generate appropriate path */
4875 2964 : 4236 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
7077 mail@joeconway.com 2965 : 4236 : }
2966 : :
2967 : : /*
2968 : : * set_tablefunc_pathlist
2969 : : * Build the (single) access path for a table func RTE
2970 : : */
2971 : : static void
3206 alvherre@alvh.no-ip. 2972 : 311 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2973 : : {
2974 : : Relids required_outer;
2975 : :
2976 : : /*
2977 : : * We don't support pushing join clauses into the quals of a tablefunc
2978 : : * scan, but it could still have required parameterization due to LATERAL
2979 : : * refs in the function expression.
2980 : : */
2981 : 311 : required_outer = rel->lateral_relids;
2982 : :
2983 : : /* Generate appropriate path */
2984 : 311 : add_path(rel, create_tablefuncscan_path(root, rel,
2985 : : required_outer));
2986 : 311 : }
2987 : :
2988 : : /*
2989 : : * set_cte_pathlist
2990 : : * Build the (single) access path for a non-self-reference CTE RTE
2991 : : *
2992 : : * There's no need for a separate set_cte_size phase, since we don't
2993 : : * support join-qual-parameterized paths for CTEs.
2994 : : */
2995 : : static void
6283 tgl@sss.pgh.pa.us 2996 : 2183 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2997 : : {
2998 : : Path *ctepath;
2999 : : Plan *cteplan;
3000 : : PlannerInfo *cteroot;
3001 : : Index levelsup;
3002 : : List *pathkeys;
3003 : : int ndx;
3004 : : ListCell *lc;
3005 : : int plan_id;
3006 : : Relids required_outer;
3007 : :
3008 : : /*
3009 : : * Find the referenced CTE, and locate the path and plan previously made
3010 : : * for it.
3011 : : */
3012 : 2183 : levelsup = rte->ctelevelsup;
3013 : 2183 : cteroot = root;
3014 [ + + ]: 3814 : while (levelsup-- > 0)
3015 : : {
3016 : 1631 : cteroot = cteroot->parent_root;
3017 [ - + ]: 1631 : if (!cteroot) /* shouldn't happen */
6283 tgl@sss.pgh.pa.us 3018 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3019 : : }
3020 : :
3021 : : /*
3022 : : * Note: cte_plan_ids can be shorter than cteList, if we are still working
3023 : : * on planning the CTEs (ie, this is a side-reference from another CTE).
3024 : : * So we mustn't use forboth here.
3025 : : */
6283 tgl@sss.pgh.pa.us 3026 :CBC 2183 : ndx = 0;
3027 [ + - + - : 3021 : foreach(lc, cteroot->parse->cteList)
+ - ]
3028 : : {
3029 : 3021 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
3030 : :
3031 [ + + ]: 3021 : if (strcmp(cte->ctename, rte->ctename) == 0)
3032 : 2183 : break;
3033 : 838 : ndx++;
3034 : : }
3035 [ - + ]: 2183 : if (lc == NULL) /* shouldn't happen */
6283 tgl@sss.pgh.pa.us 3036 [ # # ]:UBC 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6283 tgl@sss.pgh.pa.us 3037 [ - + ]:CBC 2183 : if (ndx >= list_length(cteroot->cte_plan_ids))
6283 tgl@sss.pgh.pa.us 3038 [ # # ]:UBC 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6283 tgl@sss.pgh.pa.us 3039 :CBC 2183 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
1336 3040 [ - + ]: 2183 : if (plan_id <= 0)
1336 tgl@sss.pgh.pa.us 3041 [ # # ]:UBC 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
3042 : :
631 tgl@sss.pgh.pa.us 3043 [ - + ]:CBC 2183 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
3044 : 2183 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
6283 3045 : 2183 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
3046 : :
3047 : : /* Mark rel with estimated output rows, width, etc */
3572 3048 : 2183 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
3049 : :
3050 : : /* Convert the ctepath's pathkeys to outer query's representation */
631 3051 : 2183 : pathkeys = convert_subquery_pathkeys(root,
3052 : : rel,
3053 : : ctepath->pathkeys,
3054 : : cteplan->targetlist);
3055 : :
3056 : : /*
3057 : : * We don't support pushing join clauses into the quals of a CTE scan, but
3058 : : * it could still have required parameterization due to LATERAL refs in
3059 : : * its tlist.
3060 : : */
4861 3061 : 2183 : required_outer = rel->lateral_relids;
3062 : :
3063 : : /* Generate appropriate path */
631 3064 : 2183 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
6283 3065 : 2183 : }
3066 : :
3067 : : /*
3068 : : * set_namedtuplestore_pathlist
3069 : : * Build the (single) access path for a named tuplestore RTE
3070 : : *
3071 : : * There's no need for a separate set_namedtuplestore_size phase, since we
3072 : : * don't support join-qual-parameterized paths for tuplestores.
3073 : : */
3074 : : static void
3183 kgrittn@postgresql.o 3075 : 241 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
3076 : : RangeTblEntry *rte)
3077 : : {
3078 : : Relids required_outer;
3079 : :
3080 : : /* Mark rel with estimated output rows, width, etc */
3081 : 241 : set_namedtuplestore_size_estimates(root, rel);
3082 : :
3083 : : /*
3084 : : * We don't support pushing join clauses into the quals of a tuplestore
3085 : : * scan, but it could still have required parameterization due to LATERAL
3086 : : * refs in its tlist.
3087 : : */
3088 : 241 : required_outer = rel->lateral_relids;
3089 : :
3090 : : /* Generate appropriate path */
3091 : 241 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
3092 : 241 : }
3093 : :
3094 : : /*
3095 : : * set_result_pathlist
3096 : : * Build the (single) access path for an RTE_RESULT RTE
3097 : : *
3098 : : * There's no need for a separate set_result_size phase, since we
3099 : : * don't support join-qual-parameterized paths for these RTEs.
3100 : : */
3101 : : static void
2515 tgl@sss.pgh.pa.us 3102 : 2143 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
3103 : : RangeTblEntry *rte)
3104 : : {
3105 : : Relids required_outer;
3106 : :
3107 : : /* Mark rel with estimated output rows, width, etc */
3108 : 2143 : set_result_size_estimates(root, rel);
3109 : :
3110 : : /*
3111 : : * We don't support pushing join clauses into the quals of a Result scan,
3112 : : * but it could still have required parameterization due to LATERAL refs
3113 : : * in its tlist.
3114 : : */
3115 : 2143 : required_outer = rel->lateral_relids;
3116 : :
3117 : : /* Generate appropriate path */
3118 : 2143 : add_path(rel, create_resultscan_path(root, rel, required_outer));
3119 : 2143 : }
3120 : :
3121 : : /*
3122 : : * set_worktable_pathlist
3123 : : * Build the (single) access path for a self-reference CTE RTE
3124 : : *
3125 : : * There's no need for a separate set_worktable_size phase, since we don't
3126 : : * support join-qual-parameterized paths for CTEs.
3127 : : */
3128 : : static void
6283 3129 : 474 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3130 : : {
3131 : : Path *ctepath;
3132 : : PlannerInfo *cteroot;
3133 : : Index levelsup;
3134 : : Relids required_outer;
3135 : :
3136 : : /*
3137 : : * We need to find the non-recursive term's path, which is in the plan
3138 : : * level that's processing the recursive UNION, which is one level *below*
3139 : : * where the CTE comes from.
3140 : : */
3141 : 474 : levelsup = rte->ctelevelsup;
3142 [ - + ]: 474 : if (levelsup == 0) /* shouldn't happen */
6283 tgl@sss.pgh.pa.us 3143 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6283 tgl@sss.pgh.pa.us 3144 :CBC 474 : levelsup--;
3145 : 474 : cteroot = root;
3146 [ + + ]: 1162 : while (levelsup-- > 0)
3147 : : {
3148 : 688 : cteroot = cteroot->parent_root;
3149 [ - + ]: 688 : if (!cteroot) /* shouldn't happen */
6283 tgl@sss.pgh.pa.us 3150 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3151 : : }
3572 tgl@sss.pgh.pa.us 3152 :CBC 474 : ctepath = cteroot->non_recursive_path;
3153 [ - + ]: 474 : if (!ctepath) /* shouldn't happen */
3572 tgl@sss.pgh.pa.us 3154 [ # # ]:UBC 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3155 : :
3156 : : /* Mark rel with estimated output rows, width, etc */
3572 tgl@sss.pgh.pa.us 3157 :CBC 474 : set_cte_size_estimates(root, rel, ctepath->rows);
3158 : :
3159 : : /*
3160 : : * We don't support pushing join clauses into the quals of a worktable
3161 : : * scan, but it could still have required parameterization due to LATERAL
3162 : : * refs in its tlist. (I'm not sure this is actually possible given the
3163 : : * restrictions on recursive references, but it's easy enough to support.)
3164 : : */
4861 3165 : 474 : required_outer = rel->lateral_relids;
3166 : :
3167 : : /* Generate appropriate path */
3168 : 474 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
6283 3169 : 474 : }
3170 : :
3171 : : /*
3172 : : * generate_gather_paths
3173 : : * Generate parallel access paths for a relation by pushing a Gather or
3174 : : * Gather Merge on top of a partial path.
3175 : : *
3176 : : * This must not be called until after we're done creating all partial paths
3177 : : * for the specified relation. (Otherwise, add_partial_path might delete a
3178 : : * path that some GatherPath or GatherMergePath has a reference to.)
3179 : : *
3180 : : * If we're generating paths for a scan or join relation, override_rows will
3181 : : * be false, and we'll just use the relation's size estimate. When we're
3182 : : * being called for a partially-grouped or partially-distinct path, though, we
3183 : : * need to override the rowcount estimate. (It's not clear that the
3184 : : * particular value we're using here is actually best, but the underlying rel
3185 : : * has no estimate so we must do something.)
3186 : : */
3187 : : void
2851 rhaas@postgresql.org 3188 : 12714 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3189 : : {
3190 : : Path *cheapest_partial_path;
3191 : : Path *simple_gather_path;
3192 : : ListCell *lc;
3193 : : double rows;
3194 : 12714 : double *rowsp = NULL;
3195 : :
3196 : : /* If there are no partial paths, there's nothing to do here. */
3619 3197 [ - + ]: 12714 : if (rel->partial_pathlist == NIL)
3619 rhaas@postgresql.org 3198 :UBC 0 : return;
3199 : :
3200 : : /* Should we override the rel's rowcount estimate? */
2851 rhaas@postgresql.org 3201 [ + + ]:CBC 12714 : if (override_rows)
3202 : 3133 : rowsp = &rows;
3203 : :
3204 : : /*
3205 : : * The output of Gather is always unsorted, so there's only one partial
3206 : : * path of interest: the cheapest one. That will be the one at the front
3207 : : * of partial_pathlist because of the way add_partial_path works.
3208 : : */
3619 3209 : 12714 : cheapest_partial_path = linitial(rel->partial_pathlist);
512 rguo@postgresql.org 3210 : 12714 : rows = compute_gather_rows(cheapest_partial_path);
3211 : : simple_gather_path = (Path *)
3558 rhaas@postgresql.org 3212 : 12714 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3213 : : NULL, rowsp);
3619 3214 : 12714 : add_path(rel, simple_gather_path);
3215 : :
3216 : : /*
3217 : : * For each useful ordering, we can consider an order-preserving Gather
3218 : : * Merge.
3219 : : */
3136 bruce@momjian.us 3220 [ + - + + : 28482 : foreach(lc, rel->partial_pathlist)
+ + ]
3221 : : {
3222 : 15768 : Path *subpath = (Path *) lfirst(lc);
3223 : : GatherMergePath *path;
3224 : :
3205 rhaas@postgresql.org 3225 [ + + ]: 15768 : if (subpath->pathkeys == NIL)
3226 : 12333 : continue;
3227 : :
512 rguo@postgresql.org 3228 : 3435 : rows = compute_gather_rows(subpath);
3205 rhaas@postgresql.org 3229 : 3435 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3230 : : subpath->pathkeys, NULL, rowsp);
3231 : 3435 : add_path(rel, &path->path);
3232 : : }
3233 : : }
3234 : :
3235 : : /*
3236 : : * get_useful_pathkeys_for_relation
3237 : : * Determine which orderings of a relation might be useful.
3238 : : *
3239 : : * Getting data in sorted order can be useful either because the requested
3240 : : * order matches the final output ordering for the overall query we're
3241 : : * planning, or because it enables an efficient merge join. Here, we try
3242 : : * to figure out which pathkeys to consider.
3243 : : *
3244 : : * This allows us to do incremental sort on top of an index scan under a gather
3245 : : * merge node, i.e. parallelized.
3246 : : *
3247 : : * If the require_parallel_safe is true, we also require the expressions to
3248 : : * be parallel safe (which allows pushing the sort below Gather Merge).
3249 : : *
3250 : : * XXX At the moment this can only ever return a list with a single element,
3251 : : * because it looks at query_pathkeys only. So we might return the pathkeys
3252 : : * directly, but it seems plausible we'll want to consider other orderings
3253 : : * in the future. For example, we might want to consider pathkeys useful for
3254 : : * merge joins.
3255 : : */
3256 : : static List *
1822 tomas.vondra@postgre 3257 : 12714 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3258 : : bool require_parallel_safe)
3259 : : {
2080 3260 : 12714 : List *useful_pathkeys_list = NIL;
3261 : :
3262 : : /*
3263 : : * Considering query_pathkeys is always worth it, because it might allow
3264 : : * us to avoid a total sort when we have a partially presorted path
3265 : : * available or to push the total sort into the parallel portion of the
3266 : : * query.
3267 : : */
3268 [ + + ]: 12714 : if (root->query_pathkeys)
3269 : : {
3270 : : ListCell *lc;
2043 tgl@sss.pgh.pa.us 3271 : 7630 : int npathkeys = 0; /* useful pathkeys */
3272 : :
2080 tomas.vondra@postgre 3273 [ + - + + : 13399 : foreach(lc, root->query_pathkeys)
+ + ]
3274 : : {
3275 : 9700 : PathKey *pathkey = (PathKey *) lfirst(lc);
3276 : 9700 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3277 : :
3278 : : /*
3279 : : * We can only build a sort for pathkeys that contain a
3280 : : * safe-to-compute-early EC member computable from the current
3281 : : * relation's reltarget, so ignore the remainder of the list as
3282 : : * soon as we find a pathkey without such a member.
3283 : : *
3284 : : * It's still worthwhile to return any prefix of the pathkeys list
3285 : : * that meets this requirement, as we may be able to do an
3286 : : * incremental sort.
3287 : : *
3288 : : * If requested, ensure the sort expression is parallel-safe too.
3289 : : */
1702 tgl@sss.pgh.pa.us 3290 [ + + ]: 9700 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3291 : : require_parallel_safe))
2080 tomas.vondra@postgre 3292 : 3931 : break;
3293 : :
3294 : 5769 : npathkeys++;
3295 : : }
3296 : :
3297 : : /*
3298 : : * The whole query_pathkeys list matches, so append it directly, to
3299 : : * allow comparing pathkeys easily by comparing list pointer. If we
3300 : : * have to truncate the pathkeys, we gotta do a copy though.
3301 : : */
3302 [ + + ]: 7630 : if (npathkeys == list_length(root->query_pathkeys))
3303 : 3699 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3304 : 3699 : root->query_pathkeys);
3305 [ + + ]: 3931 : else if (npathkeys > 0)
3306 : 237 : useful_pathkeys_list = lappend(useful_pathkeys_list,
1253 drowley@postgresql.o 3307 : 237 : list_copy_head(root->query_pathkeys,
3308 : : npathkeys));
3309 : : }
3310 : :
2080 tomas.vondra@postgre 3311 : 12714 : return useful_pathkeys_list;
3312 : : }
3313 : :
3314 : : /*
3315 : : * generate_useful_gather_paths
3316 : : * Generate parallel access paths for a relation by pushing a Gather or
3317 : : * Gather Merge on top of a partial path.
3318 : : *
3319 : : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3320 : : * paths (aiming to preserve the ordering), but also considers ordering that
3321 : : * might be useful for nodes above the gather merge node, and tries to add
3322 : : * a sort (regular or incremental) to provide that.
3323 : : */
3324 : : void
3325 : 312372 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3326 : : {
3327 : : ListCell *lc;
3328 : : double rows;
3329 : 312372 : double *rowsp = NULL;
3330 : 312372 : List *useful_pathkeys_list = NIL;
3331 : 312372 : Path *cheapest_partial_path = NULL;
3332 : :
3333 : : /* If there are no partial paths, there's nothing to do here. */
3334 [ + + ]: 312372 : if (rel->partial_pathlist == NIL)
3335 : 299658 : return;
3336 : :
3337 : : /* Should we override the rel's rowcount estimate? */
3338 [ + + ]: 12714 : if (override_rows)
3339 : 3133 : rowsp = &rows;
3340 : :
3341 : : /* generate the regular gather (merge) paths */
3342 : 12714 : generate_gather_paths(root, rel, override_rows);
3343 : :
3344 : : /* consider incremental sort for interesting orderings */
1822 3345 : 12714 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3346 : :
3347 : : /* used for explicit (full) sort paths */
2080 3348 : 12714 : cheapest_partial_path = linitial(rel->partial_pathlist);
3349 : :
3350 : : /*
3351 : : * Consider sorted paths for each interesting ordering. We generate both
3352 : : * incremental and full sort.
3353 : : */
3354 [ + + + + : 16650 : foreach(lc, useful_pathkeys_list)
+ + ]
3355 : : {
3356 : 3936 : List *useful_pathkeys = lfirst(lc);
3357 : : ListCell *lc2;
3358 : : bool is_sorted;
3359 : : int presorted_keys;
3360 : :
3361 [ + - + + : 9283 : foreach(lc2, rel->partial_pathlist)
+ + ]
3362 : : {
3363 : 5347 : Path *subpath = (Path *) lfirst(lc2);
3364 : : GatherMergePath *path;
3365 : :
3366 : 5347 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3367 : : subpath->pathkeys,
3368 : : &presorted_keys);
3369 : :
3370 : : /*
3371 : : * We don't need to consider the case where a subpath is already
3372 : : * fully sorted because generate_gather_paths already creates a
3373 : : * gather merge path for every subpath that has pathkeys present.
3374 : : *
3375 : : * But since the subpath is already sorted, we know we don't need
3376 : : * to consider adding a sort (full or incremental) on top of it,
3377 : : * so we can continue here.
3378 : : */
3379 [ + + ]: 5347 : if (is_sorted)
3380 : 1490 : continue;
3381 : :
3382 : : /*
3383 : : * Try at least sorting the cheapest path and also try
3384 : : * incrementally sorting any path which is partially sorted
3385 : : * already (no need to deal with paths which have presorted keys
3386 : : * when incremental sort is disabled unless it's the cheapest
3387 : : * input path).
3388 : : */
1097 drowley@postgresql.o 3389 [ + + ]: 3857 : if (subpath != cheapest_partial_path &&
3390 [ + + + + ]: 189 : (presorted_keys == 0 || !enable_incremental_sort))
3391 : 51 : continue;
3392 : :
3393 : : /*
3394 : : * Consider regular sort for any path that's not presorted or if
3395 : : * incremental sort is disabled. We've no need to consider both
3396 : : * sort and incremental sort on the same path. We assume that
3397 : : * incremental sort is always faster when there are presorted
3398 : : * keys.
3399 : : *
3400 : : * This is not redundant with the gather paths created in
3401 : : * generate_gather_paths, because that doesn't generate ordered
3402 : : * output. Here we add an explicit sort to match the useful
3403 : : * ordering.
3404 : : */
3405 [ + + + + ]: 3806 : if (presorted_keys == 0 || !enable_incremental_sort)
3406 : : {
3407 : 3662 : subpath = (Path *) create_sort_path(root,
3408 : : rel,
3409 : : subpath,
3410 : : useful_pathkeys,
3411 : : -1.0);
3412 : : }
3413 : : else
3414 : 144 : subpath = (Path *) create_incremental_sort_path(root,
3415 : : rel,
3416 : : subpath,
3417 : : useful_pathkeys,
3418 : : presorted_keys,
3419 : : -1);
512 rguo@postgresql.org 3420 : 3806 : rows = compute_gather_rows(subpath);
1097 drowley@postgresql.o 3421 : 3806 : path = create_gather_merge_path(root, rel,
3422 : : subpath,
3423 : 3806 : rel->reltarget,
3424 : : subpath->pathkeys,
3425 : : NULL,
3426 : : rowsp);
3427 : :
3428 : 3806 : add_path(rel, &path->path);
3429 : : }
3430 : : }
3431 : : }
3432 : :
3433 : : /*
3434 : : * generate_grouped_paths
3435 : : * Generate paths for a grouped relation by adding sorted and hashed
3436 : : * partial aggregation paths on top of paths of the ungrouped relation.
3437 : : *
3438 : : * The information needed is provided by the RelAggInfo structure stored in
3439 : : * "grouped_rel".
3440 : : */
3441 : : void
70 rguo@postgresql.org 3442 :GNC 449 : generate_grouped_paths(PlannerInfo *root, RelOptInfo *grouped_rel,
3443 : : RelOptInfo *rel)
3444 : : {
3445 : 449 : RelAggInfo *agg_info = grouped_rel->agg_info;
3446 : : AggClauseCosts agg_costs;
3447 : : bool can_hash;
3448 : : bool can_sort;
3449 : 449 : Path *cheapest_total_path = NULL;
3450 : 449 : Path *cheapest_partial_path = NULL;
3451 : 449 : double dNumGroups = 0;
3452 : 449 : double dNumPartialGroups = 0;
3453 : 449 : List *group_pathkeys = NIL;
3454 : :
3455 [ - + ]: 449 : if (IS_DUMMY_REL(rel))
3456 : : {
70 rguo@postgresql.org 3457 :UNC 0 : mark_dummy_rel(grouped_rel);
3458 : 0 : return;
3459 : : }
3460 : :
3461 : : /*
3462 : : * We push partial aggregation only to the lowest possible level in the
3463 : : * join tree that is deemed useful.
3464 : : */
64 rguo@postgresql.org 3465 [ + - ]:GNC 449 : if (!bms_equal(agg_info->apply_agg_at, rel->relids) ||
70 3466 [ - + ]: 449 : !agg_info->agg_useful)
70 rguo@postgresql.org 3467 :UNC 0 : return;
3468 : :
70 rguo@postgresql.org 3469 [ + - + - :GNC 2694 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
3470 : 449 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL, &agg_costs);
3471 : :
3472 : : /*
3473 : : * Determine whether it's possible to perform sort-based implementations
3474 : : * of grouping, and generate the pathkeys that represent the grouping
3475 : : * requirements in that case.
3476 : : */
3477 : 449 : can_sort = grouping_is_sortable(agg_info->group_clauses);
3478 [ + - ]: 449 : if (can_sort)
3479 : : {
3480 : : RelOptInfo *top_grouped_rel;
3481 : : List *top_group_tlist;
3482 : :
3483 [ + + - + ]: 251 : top_grouped_rel = IS_OTHER_REL(rel) ?
3484 [ + + ]: 700 : rel->top_parent->grouped_rel : grouped_rel;
3485 : : top_group_tlist =
3486 : 449 : make_tlist_from_pathtarget(top_grouped_rel->agg_info->target);
3487 : :
3488 : : group_pathkeys =
3489 : 449 : make_pathkeys_for_sortclauses(root, agg_info->group_clauses,
3490 : : top_group_tlist);
3491 : : }
3492 : :
3493 : : /*
3494 : : * Determine whether we should consider hash-based implementations of
3495 : : * grouping.
3496 : : */
3497 [ - + ]: 449 : Assert(root->numOrderedAggs == 0);
3498 [ + - + - ]: 898 : can_hash = (agg_info->group_clauses != NIL &&
3499 : 449 : grouping_is_hashable(agg_info->group_clauses));
3500 : :
3501 : : /*
3502 : : * Consider whether we should generate partially aggregated non-partial
3503 : : * paths. We can only do this if we have a non-partial path.
3504 : : */
3505 [ + - ]: 449 : if (rel->pathlist != NIL)
3506 : : {
3507 : 449 : cheapest_total_path = rel->cheapest_total_path;
3508 [ - + ]: 449 : Assert(cheapest_total_path != NULL);
3509 : : }
3510 : :
3511 : : /*
3512 : : * If parallelism is possible for grouped_rel, then we should consider
3513 : : * generating partially-grouped partial paths. However, if the ungrouped
3514 : : * rel has no partial paths, then we can't.
3515 : : */
3516 [ + + + + ]: 449 : if (grouped_rel->consider_parallel && rel->partial_pathlist != NIL)
3517 : : {
3518 : 366 : cheapest_partial_path = linitial(rel->partial_pathlist);
3519 [ - + ]: 366 : Assert(cheapest_partial_path != NULL);
3520 : : }
3521 : :
3522 : : /* Estimate number of partial groups. */
3523 [ + - ]: 449 : if (cheapest_total_path != NULL)
3524 : 449 : dNumGroups = estimate_num_groups(root,
3525 : : agg_info->group_exprs,
3526 : : cheapest_total_path->rows,
3527 : : NULL, NULL);
3528 [ + + ]: 449 : if (cheapest_partial_path != NULL)
3529 : 366 : dNumPartialGroups = estimate_num_groups(root,
3530 : : agg_info->group_exprs,
3531 : : cheapest_partial_path->rows,
3532 : : NULL, NULL);
3533 : :
3534 [ + - + - ]: 449 : if (can_sort && cheapest_total_path != NULL)
3535 : : {
3536 : : ListCell *lc;
3537 : :
3538 : : /*
3539 : : * Use any available suitably-sorted path as input, and also consider
3540 : : * sorting the cheapest-total path and incremental sort on any paths
3541 : : * with presorted keys.
3542 : : *
3543 : : * To save planning time, we ignore parameterized input paths unless
3544 : : * they are the cheapest-total path.
3545 : : */
3546 [ + - + + : 1087 : foreach(lc, rel->pathlist)
+ + ]
3547 : : {
3548 : 638 : Path *input_path = (Path *) lfirst(lc);
3549 : : Path *path;
3550 : : bool is_sorted;
3551 : : int presorted_keys;
3552 : :
3553 : : /*
3554 : : * Ignore parameterized paths that are not the cheapest-total
3555 : : * path.
3556 : : */
3557 [ + + + - ]: 638 : if (input_path->param_info &&
3558 : : input_path != cheapest_total_path)
3559 : 15 : continue;
3560 : :
3561 : 635 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3562 : : input_path->pathkeys,
3563 : : &presorted_keys);
3564 : :
3565 : : /*
3566 : : * Ignore paths that are not suitably or partially sorted, unless
3567 : : * they are the cheapest total path (no need to deal with paths
3568 : : * which have presorted keys when incremental sort is disabled).
3569 : : */
3570 [ + + + + ]: 635 : if (!is_sorted && input_path != cheapest_total_path &&
3571 [ + + - + ]: 84 : (presorted_keys == 0 || !enable_incremental_sort))
3572 : 12 : continue;
3573 : :
3574 : : /*
3575 : : * Since the path originates from a non-grouped relation that is
3576 : : * not aware of eager aggregation, we must ensure that it provides
3577 : : * the correct input for partial aggregation.
3578 : : */
3579 : 623 : path = (Path *) create_projection_path(root,
3580 : : grouped_rel,
3581 : : input_path,
3582 : 623 : agg_info->agg_input);
3583 : :
3584 [ + + ]: 623 : if (!is_sorted)
3585 : : {
3586 : : /*
3587 : : * We've no need to consider both a sort and incremental sort.
3588 : : * We'll just do a sort if there are no presorted keys and an
3589 : : * incremental sort when there are presorted keys.
3590 : : */
3591 [ + + - + ]: 518 : if (presorted_keys == 0 || !enable_incremental_sort)
3592 : 446 : path = (Path *) create_sort_path(root,
3593 : : grouped_rel,
3594 : : path,
3595 : : group_pathkeys,
3596 : : -1.0);
3597 : : else
3598 : 72 : path = (Path *) create_incremental_sort_path(root,
3599 : : grouped_rel,
3600 : : path,
3601 : : group_pathkeys,
3602 : : presorted_keys,
3603 : : -1.0);
3604 : : }
3605 : :
3606 : : /*
3607 : : * qual is NIL because the HAVING clause cannot be evaluated until
3608 : : * the final value of the aggregate is known.
3609 : : */
3610 : 623 : path = (Path *) create_agg_path(root,
3611 : : grouped_rel,
3612 : : path,
3613 : 623 : agg_info->target,
3614 : : AGG_SORTED,
3615 : : AGGSPLIT_INITIAL_SERIAL,
3616 : : agg_info->group_clauses,
3617 : : NIL,
3618 : : &agg_costs,
3619 : : dNumGroups);
3620 : :
3621 : 623 : add_path(grouped_rel, path);
3622 : : }
3623 : : }
3624 : :
3625 [ + - + + ]: 449 : if (can_sort && cheapest_partial_path != NULL)
3626 : : {
3627 : : ListCell *lc;
3628 : :
3629 : : /* Similar to above logic, but for partial paths. */
3630 [ + - + + : 852 : foreach(lc, rel->partial_pathlist)
+ + ]
3631 : : {
3632 : 486 : Path *input_path = (Path *) lfirst(lc);
3633 : : Path *path;
3634 : : bool is_sorted;
3635 : : int presorted_keys;
3636 : :
3637 : 486 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3638 : : input_path->pathkeys,
3639 : : &presorted_keys);
3640 : :
3641 : : /*
3642 : : * Ignore paths that are not suitably or partially sorted, unless
3643 : : * they are the cheapest partial path (no need to deal with paths
3644 : : * which have presorted keys when incremental sort is disabled).
3645 : : */
3646 [ + + + + ]: 486 : if (!is_sorted && input_path != cheapest_partial_path &&
3647 [ + - - + ]: 48 : (presorted_keys == 0 || !enable_incremental_sort))
70 rguo@postgresql.org 3648 :UNC 0 : continue;
3649 : :
3650 : : /*
3651 : : * Since the path originates from a non-grouped relation that is
3652 : : * not aware of eager aggregation, we must ensure that it provides
3653 : : * the correct input for partial aggregation.
3654 : : */
70 rguo@postgresql.org 3655 :GNC 486 : path = (Path *) create_projection_path(root,
3656 : : grouped_rel,
3657 : : input_path,
3658 : 486 : agg_info->agg_input);
3659 : :
3660 [ + + ]: 486 : if (!is_sorted)
3661 : : {
3662 : : /*
3663 : : * We've no need to consider both a sort and incremental sort.
3664 : : * We'll just do a sort if there are no presorted keys and an
3665 : : * incremental sort when there are presorted keys.
3666 : : */
3667 [ + + - + ]: 414 : if (presorted_keys == 0 || !enable_incremental_sort)
3668 : 366 : path = (Path *) create_sort_path(root,
3669 : : grouped_rel,
3670 : : path,
3671 : : group_pathkeys,
3672 : : -1.0);
3673 : : else
3674 : 48 : path = (Path *) create_incremental_sort_path(root,
3675 : : grouped_rel,
3676 : : path,
3677 : : group_pathkeys,
3678 : : presorted_keys,
3679 : : -1.0);
3680 : : }
3681 : :
3682 : : /*
3683 : : * qual is NIL because the HAVING clause cannot be evaluated until
3684 : : * the final value of the aggregate is known.
3685 : : */
3686 : 486 : path = (Path *) create_agg_path(root,
3687 : : grouped_rel,
3688 : : path,
3689 : 486 : agg_info->target,
3690 : : AGG_SORTED,
3691 : : AGGSPLIT_INITIAL_SERIAL,
3692 : : agg_info->group_clauses,
3693 : : NIL,
3694 : : &agg_costs,
3695 : : dNumPartialGroups);
3696 : :
3697 : 486 : add_partial_path(grouped_rel, path);
3698 : : }
3699 : : }
3700 : :
3701 : : /*
3702 : : * Add a partially-grouped HashAgg Path where possible
3703 : : */
3704 [ + - + - ]: 449 : if (can_hash && cheapest_total_path != NULL)
3705 : : {
3706 : : Path *path;
3707 : :
3708 : : /*
3709 : : * Since the path originates from a non-grouped relation that is not
3710 : : * aware of eager aggregation, we must ensure that it provides the
3711 : : * correct input for partial aggregation.
3712 : : */
3713 : 449 : path = (Path *) create_projection_path(root,
3714 : : grouped_rel,
3715 : : cheapest_total_path,
3716 : 449 : agg_info->agg_input);
3717 : :
3718 : : /*
3719 : : * qual is NIL because the HAVING clause cannot be evaluated until the
3720 : : * final value of the aggregate is known.
3721 : : */
3722 : 449 : path = (Path *) create_agg_path(root,
3723 : : grouped_rel,
3724 : : path,
3725 : 449 : agg_info->target,
3726 : : AGG_HASHED,
3727 : : AGGSPLIT_INITIAL_SERIAL,
3728 : : agg_info->group_clauses,
3729 : : NIL,
3730 : : &agg_costs,
3731 : : dNumGroups);
3732 : :
3733 : 449 : add_path(grouped_rel, path);
3734 : : }
3735 : :
3736 : : /*
3737 : : * Now add a partially-grouped HashAgg partial Path where possible
3738 : : */
3739 [ + - + + ]: 449 : if (can_hash && cheapest_partial_path != NULL)
3740 : : {
3741 : : Path *path;
3742 : :
3743 : : /*
3744 : : * Since the path originates from a non-grouped relation that is not
3745 : : * aware of eager aggregation, we must ensure that it provides the
3746 : : * correct input for partial aggregation.
3747 : : */
3748 : 366 : path = (Path *) create_projection_path(root,
3749 : : grouped_rel,
3750 : : cheapest_partial_path,
3751 : 366 : agg_info->agg_input);
3752 : :
3753 : : /*
3754 : : * qual is NIL because the HAVING clause cannot be evaluated until the
3755 : : * final value of the aggregate is known.
3756 : : */
3757 : 366 : path = (Path *) create_agg_path(root,
3758 : : grouped_rel,
3759 : : path,
3760 : 366 : agg_info->target,
3761 : : AGG_HASHED,
3762 : : AGGSPLIT_INITIAL_SERIAL,
3763 : : agg_info->group_clauses,
3764 : : NIL,
3765 : : &agg_costs,
3766 : : dNumPartialGroups);
3767 : :
3768 : 366 : add_partial_path(grouped_rel, path);
3769 : : }
3770 : : }
3771 : :
3772 : : /*
3773 : : * make_rel_from_joinlist
3774 : : * Build access paths using a "joinlist" to guide the join path search.
3775 : : *
3776 : : * See comments for deconstruct_jointree() for definition of the joinlist
3777 : : * data structure.
3778 : : */
3779 : : static RelOptInfo *
7302 tgl@sss.pgh.pa.us 3780 :CBC 172877 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3781 : : {
3782 : : int levels_needed;
3783 : : List *initial_rels;
3784 : : ListCell *jl;
3785 : :
3786 : : /*
3787 : : * Count the number of child joinlist nodes. This is the depth of the
3788 : : * dynamic-programming algorithm we must employ to consider all ways of
3789 : : * joining the child nodes.
3790 : : */
3791 : 172877 : levels_needed = list_length(joinlist);
3792 : :
9210 3793 [ - + ]: 172877 : if (levels_needed <= 0)
9210 tgl@sss.pgh.pa.us 3794 :UBC 0 : return NULL; /* nothing to do? */
3795 : :
3796 : : /*
3797 : : * Construct a list of rels corresponding to the child joinlist nodes.
3798 : : * This may contain both base rels and rels constructed according to
3799 : : * sub-joinlists.
3800 : : */
7302 tgl@sss.pgh.pa.us 3801 :CBC 172877 : initial_rels = NIL;
3802 [ + - + + : 416117 : foreach(jl, joinlist)
+ + ]
3803 : : {
3804 : 243240 : Node *jlnode = (Node *) lfirst(jl);
3805 : : RelOptInfo *thisrel;
3806 : :
3807 [ + + ]: 243240 : if (IsA(jlnode, RangeTblRef))
3808 : : {
3809 : 241542 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3810 : :
3811 : 241542 : thisrel = find_base_rel(root, varno);
3812 : : }
3813 [ + - ]: 1698 : else if (IsA(jlnode, List))
3814 : : {
3815 : : /* Recurse to handle subproblem */
3816 : 1698 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3817 : : }
3818 : : else
3819 : : {
7302 tgl@sss.pgh.pa.us 3820 [ # # ]:UBC 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3821 : : (int) nodeTag(jlnode));
3822 : : thisrel = NULL; /* keep compiler quiet */
3823 : : }
3824 : :
7302 tgl@sss.pgh.pa.us 3825 :CBC 243240 : initial_rels = lappend(initial_rels, thisrel);
3826 : : }
3827 : :
9210 3828 [ + + ]: 172877 : if (levels_needed == 1)
3829 : : {
3830 : : /*
3831 : : * Single joinlist node, so we're done.
3832 : : */
7875 neilc@samurai.com 3833 : 121642 : return (RelOptInfo *) linitial(initial_rels);
3834 : : }
3835 : : else
3836 : : {
3837 : : /*
3838 : : * Consider the different orders in which we could join the rels,
3839 : : * using a plugin, GEQO, or the regular join search code.
3840 : : *
3841 : : * We put the initial_rels list into a PlannerInfo field because
3842 : : * has_legal_joinclause() needs to look at it (ugly :-().
3843 : : */
6550 tgl@sss.pgh.pa.us 3844 : 51235 : root->initial_rels = initial_rels;
3845 : :
6657 3846 [ - + ]: 51235 : if (join_search_hook)
6657 tgl@sss.pgh.pa.us 3847 :UBC 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
6657 tgl@sss.pgh.pa.us 3848 [ + - + + ]:CBC 51235 : else if (enable_geqo && levels_needed >= geqo_threshold)
9210 3849 : 21 : return geqo(root, levels_needed, initial_rels);
3850 : : else
6657 3851 : 51214 : return standard_join_search(root, levels_needed, initial_rels);
3852 : : }
3853 : : }
3854 : :
3855 : : /*
3856 : : * standard_join_search
3857 : : * Find possible joinpaths for a query by successively finding ways
3858 : : * to join component relations into join relations.
3859 : : *
3860 : : * 'levels_needed' is the number of iterations needed, ie, the number of
3861 : : * independent jointree items in the query. This is > 1.
3862 : : *
3863 : : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3864 : : * jointree item. These are the components to be joined together.
3865 : : * Note that levels_needed == list_length(initial_rels).
3866 : : *
3867 : : * Returns the final level of join relations, i.e., the relation that is
3868 : : * the result of joining all the original relations together.
3869 : : * At least one implementation path must be provided for this relation and
3870 : : * all required sub-relations.
3871 : : *
3872 : : * To support loadable plugins that modify planner behavior by changing the
3873 : : * join searching algorithm, we provide a hook variable that lets a plugin
3874 : : * replace or supplement this function. Any such hook must return the same
3875 : : * final join relation as the standard code would, but it might have a
3876 : : * different set of implementation paths attached, and only the sub-joinrels
3877 : : * needed for these paths need have been instantiated.
3878 : : *
3879 : : * Note to plugin authors: the functions invoked during standard_join_search()
3880 : : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3881 : : * than one join-order search, you'll probably need to save and restore the
3882 : : * original states of those data structures. See geqo_eval() for an example.
3883 : : */
3884 : : RelOptInfo *
3885 : 51214 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3886 : : {
3887 : : int lev;
3888 : : RelOptInfo *rel;
3889 : :
3890 : : /*
3891 : : * This function cannot be invoked recursively within any one planning
3892 : : * problem, so join_rel_level[] can't be in use already.
3893 : : */
5863 3894 [ - + ]: 51214 : Assert(root->join_rel_level == NULL);
3895 : :
3896 : : /*
3897 : : * We employ a simple "dynamic programming" algorithm: we first find all
3898 : : * ways to build joins of two jointree items, then all ways to build joins
3899 : : * of three items (from two-item joins and single items), then four-item
3900 : : * joins, and so on until we have considered all ways to join all the
3901 : : * items into one rel.
3902 : : *
3903 : : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3904 : : * set root->join_rel_level[1] to represent all the single-jointree-item
3905 : : * relations.
3906 : : */
3907 : 51214 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3908 : :
3909 : 51214 : root->join_rel_level[1] = initial_rels;
3910 : :
9445 3911 [ + + ]: 121547 : for (lev = 2; lev <= levels_needed; lev++)
3912 : : {
3913 : : ListCell *lc;
3914 : :
3915 : : /*
3916 : : * Determine all possible pairs of relations to be joined at this
3917 : : * level, and build paths for making each one from every available
3918 : : * pair of lower-level relations.
3919 : : */
5863 3920 : 70333 : join_search_one_level(root, lev);
3921 : :
3922 : : /*
3923 : : * Run generate_partitionwise_join_paths() and
3924 : : * generate_useful_gather_paths() for each just-processed joinrel. We
3925 : : * could not do this earlier because both regular and partial paths
3926 : : * can get added to a particular joinrel at multiple times within
3927 : : * join_search_one_level.
3928 : : *
3929 : : * After that, we're done creating paths for the joinrel, so run
3930 : : * set_cheapest().
3931 : : *
3932 : : * In addition, we also run generate_grouped_paths() for the grouped
3933 : : * relation of each just-processed joinrel, and run set_cheapest() for
3934 : : * the grouped relation afterwards.
3935 : : */
3936 [ + + + + : 178153 : foreach(lc, root->join_rel_level[lev])
+ + ]
3937 : : {
3938 : : bool is_top_rel;
3939 : :
3940 : 107820 : rel = (RelOptInfo *) lfirst(lc);
3941 : :
70 rguo@postgresql.org 3942 :GNC 107820 : is_top_rel = bms_equal(rel->relids, root->all_query_rels);
3943 : :
3944 : : /* Create paths for partitionwise joins. */
2861 peter_e@gmx.net 3945 :CBC 107820 : generate_partitionwise_join_paths(root, rel);
3946 : :
3947 : : /*
3948 : : * Except for the topmost scan/join rel, consider gathering
3949 : : * partial paths. We'll do the same for the topmost scan/join rel
3950 : : * once we know the final targetlist (see grouping_planner's and
3951 : : * its call to apply_scanjoin_target_to_paths).
3952 : : */
70 rguo@postgresql.org 3953 [ + + ]:GNC 107820 : if (!is_top_rel)
2080 tomas.vondra@postgre 3954 :CBC 56855 : generate_useful_gather_paths(root, rel, false);
3955 : :
3956 : : /* Find and save the cheapest paths for this rel */
9437 tgl@sss.pgh.pa.us 3957 : 107820 : set_cheapest(rel);
3958 : :
3959 : : /*
3960 : : * Except for the topmost scan/join rel, consider generating
3961 : : * partial aggregation paths for the grouped relation on top of
3962 : : * the paths of this rel. After that, we're done creating paths
3963 : : * for the grouped relation, so run set_cheapest().
3964 : : */
70 rguo@postgresql.org 3965 [ + + + + ]:GNC 107820 : if (rel->grouped_rel != NULL && !is_top_rel)
3966 : : {
3967 : 36 : RelOptInfo *grouped_rel = rel->grouped_rel;
3968 : :
3969 [ - + ]: 36 : Assert(IS_GROUPED_REL(grouped_rel));
3970 : :
3971 : 36 : generate_grouped_paths(root, grouped_rel, rel);
3972 : 36 : set_cheapest(grouped_rel);
3973 : : }
3974 : :
3975 : : #ifdef OPTIMIZER_DEBUG
3976 : : pprint(rel);
3977 : : #endif
3978 : : }
3979 : : }
3980 : :
3981 : : /*
3982 : : * We should have a single rel at the final level.
3983 : : */
5863 tgl@sss.pgh.pa.us 3984 [ - + ]:CBC 51214 : if (root->join_rel_level[levels_needed] == NIL)
8036 tgl@sss.pgh.pa.us 3985 [ # # ]:UBC 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
5863 tgl@sss.pgh.pa.us 3986 [ - + ]:CBC 51214 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3987 : :
3988 : 51214 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3989 : :
3990 : 51214 : root->join_rel_level = NULL;
3991 : :
9445 3992 : 51214 : return rel;
3993 : : }
3994 : :
3995 : : /*****************************************************************************
3996 : : * PUSHING QUALS DOWN INTO SUBQUERIES
3997 : : *****************************************************************************/
3998 : :
3999 : : /*
4000 : : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
4001 : : *
4002 : : * subquery is the particular component query being checked. topquery
4003 : : * is the top component of a set-operations tree (the same Query if no
4004 : : * set-op is involved).
4005 : : *
4006 : : * Conditions checked here:
4007 : : *
4008 : : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
4009 : : * since that could change the set of rows returned.
4010 : : *
4011 : : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
4012 : : * quals into it, because that could change the results.
4013 : : *
4014 : : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
4015 : : * This is because upper-level quals should semantically be evaluated only
4016 : : * once per distinct row, not once per original row, and if the qual is
4017 : : * volatile then extra evaluations could change the results. (This issue
4018 : : * does not apply to other forms of aggregation such as GROUP BY, because
4019 : : * when those are present we push into HAVING not WHERE, so that the quals
4020 : : * are still applied after aggregation.)
4021 : : *
4022 : : * 4. If the subquery contains window functions, we cannot push volatile quals
4023 : : * into it. The issue here is a bit different from DISTINCT: a volatile qual
4024 : : * might succeed for some rows of a window partition and fail for others,
4025 : : * thereby changing the partition contents and thus the window functions'
4026 : : * results for rows that remain.
4027 : : *
4028 : : * 5. If the subquery contains any set-returning functions in its targetlist,
4029 : : * we cannot push volatile quals into it. That would push them below the SRFs
4030 : : * and thereby change the number of times they are evaluated. Also, a
4031 : : * volatile qual could succeed for some SRF output rows and fail for others,
4032 : : * a behavior that cannot occur if it's evaluated before SRF expansion.
4033 : : *
4034 : : * 6. If the subquery has nonempty grouping sets, we cannot push down any
4035 : : * quals. The concern here is that a qual referencing a "constant" grouping
4036 : : * column could get constant-folded, which would be improper because the value
4037 : : * is potentially nullable by grouping-set expansion. This restriction could
4038 : : * be removed if we had a parsetree representation that shows that such
4039 : : * grouping columns are not really constant. (There are other ideas that
4040 : : * could be used to relax this restriction, but that's the approach most
4041 : : * likely to get taken in the future. Note that there's not much to be gained
4042 : : * so long as subquery_planner can't move HAVING clauses to WHERE within such
4043 : : * a subquery.)
4044 : : *
4045 : : * In addition, we make several checks on the subquery's output columns to see
4046 : : * if it is safe to reference them in pushed-down quals. If output column k
4047 : : * is found to be unsafe to reference, we set the reason for that inside
4048 : : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
4049 : : * column k might not be referenced by some/all quals. The unsafeFlags[]
4050 : : * array will be consulted later by qual_is_pushdown_safe(). It's better to
4051 : : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
4052 : : * because when the subquery involves set operations we have to check the
4053 : : * output expressions in each arm of the set op.
4054 : : *
4055 : : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
4056 : : * we're effectively assuming that the quals cannot distinguish values that
4057 : : * the DISTINCT's equality operator sees as equal, yet there are many
4058 : : * counterexamples to that assumption. However use of such a qual with a
4059 : : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
4060 : : * "equal" value will be chosen as the output value by the DISTINCT operation.
4061 : : * So we don't worry too much about that. Another objection is that if the
4062 : : * qual is expensive to evaluate, running it for each original row might cost
4063 : : * more than we save by eliminating rows before the DISTINCT step. But it
4064 : : * would be very hard to estimate that at this stage, and in practice pushdown
4065 : : * seldom seems to make things worse, so we ignore that problem too.
4066 : : *
4067 : : * Note: likewise, pushing quals into a subquery with window functions is a
4068 : : * bit dubious: the quals might remove some rows of a window partition while
4069 : : * leaving others, causing changes in the window functions' results for the
4070 : : * surviving rows. We insist that such a qual reference only partitioning
4071 : : * columns, but again that only protects us if the qual does not distinguish
4072 : : * values that the partitioning equality operator sees as equal. The risks
4073 : : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
4074 : : * occurs and thus there is no theoretical problem with such a qual. But
4075 : : * we'll do this anyway because the potential performance benefits are very
4076 : : * large, and we've seen no field complaints about the longstanding comparable
4077 : : * behavior with DISTINCT.
4078 : : */
4079 : : static bool
8273 4080 : 1188 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
4081 : : pushdown_safety_info *safetyInfo)
4082 : : {
4083 : : SetOperationStmt *topop;
4084 : :
4085 : : /* Check point 1 */
8306 4086 [ + + + + ]: 1188 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
8511 4087 : 67 : return false;
4088 : :
4089 : : /* Check point 6 */
1943 4090 [ + + + + ]: 1121 : if (subquery->groupClause && subquery->groupingSets)
4091 : 6 : return false;
4092 : :
4093 : : /* Check points 3, 4, and 5 */
3368 4094 [ + + ]: 1115 : if (subquery->distinctClause ||
4095 [ + + ]: 1073 : subquery->hasWindowFuncs ||
4096 [ + + ]: 940 : subquery->hasTargetSRFs)
4191 4097 : 273 : safetyInfo->unsafeVolatile = true;
4098 : :
4099 : : /*
4100 : : * If we're at a leaf query, check for unsafe expressions in its target
4101 : : * list, and mark any reasons why they're unsafe in unsafeFlags[].
4102 : : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
4103 : : * so no need to check them.)
4104 : : */
4578 4105 [ + + ]: 1115 : if (subquery->setOperations == NULL)
4191 4106 : 1021 : check_output_expressions(subquery, safetyInfo);
4107 : :
4108 : : /* Are we at top level, or looking at a setop component? */
8511 4109 [ + + ]: 1115 : if (subquery == topquery)
4110 : : {
4111 : : /* Top level, so check any component queries */
4112 [ + + ]: 927 : if (subquery->setOperations != NULL)
8273 4113 [ - + ]: 94 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
4114 : : safetyInfo))
8511 tgl@sss.pgh.pa.us 4115 :UBC 0 : return false;
4116 : : }
4117 : : else
4118 : : {
4119 : : /* Setop component must not have more components (too weird) */
8511 tgl@sss.pgh.pa.us 4120 [ - + ]:CBC 188 : if (subquery->setOperations != NULL)
8511 tgl@sss.pgh.pa.us 4121 :UBC 0 : return false;
4122 : : /* Check whether setop component output types match top level */
3221 peter_e@gmx.net 4123 :CBC 188 : topop = castNode(SetOperationStmt, topquery->setOperations);
4124 [ - + ]: 188 : Assert(topop);
8273 tgl@sss.pgh.pa.us 4125 : 188 : compare_tlist_datatypes(subquery->targetList,
4126 : : topop->colTypes,
4127 : : safetyInfo);
4128 : : }
8511 4129 : 1115 : return true;
4130 : : }
4131 : :
4132 : : /*
4133 : : * Helper routine to recurse through setOperations tree
4134 : : */
4135 : : static bool
8273 4136 : 282 : recurse_pushdown_safe(Node *setOp, Query *topquery,
4137 : : pushdown_safety_info *safetyInfo)
4138 : : {
8511 4139 [ + + ]: 282 : if (IsA(setOp, RangeTblRef))
4140 : : {
4141 : 188 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4142 : 188 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
4143 : 188 : Query *subquery = rte->subquery;
4144 : :
4145 [ - + ]: 188 : Assert(subquery != NULL);
4191 4146 : 188 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
4147 : : }
8511 4148 [ + - ]: 94 : else if (IsA(setOp, SetOperationStmt))
4149 : : {
4150 : 94 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4151 : :
4152 : : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
4153 [ - + ]: 94 : if (op->op == SETOP_EXCEPT)
8511 tgl@sss.pgh.pa.us 4154 :UBC 0 : return false;
4155 : : /* Else recurse */
4191 tgl@sss.pgh.pa.us 4156 [ - + ]:CBC 94 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
8511 tgl@sss.pgh.pa.us 4157 :UBC 0 : return false;
4191 tgl@sss.pgh.pa.us 4158 [ - + ]:CBC 94 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
8511 tgl@sss.pgh.pa.us 4159 :UBC 0 : return false;
4160 : : }
4161 : : else
4162 : : {
8181 4163 [ # # ]: 0 : elog(ERROR, "unrecognized node type: %d",
4164 : : (int) nodeTag(setOp));
4165 : : }
8511 tgl@sss.pgh.pa.us 4166 :CBC 94 : return true;
4167 : : }
4168 : :
4169 : : /*
4170 : : * check_output_expressions - check subquery's output expressions for safety
4171 : : *
4172 : : * There are several cases in which it's unsafe to push down an upper-level
4173 : : * qual if it references a particular output column of a subquery. We check
4174 : : * each output column of the subquery and set flags in unsafeFlags[k] when we
4175 : : * see that column is unsafe for a pushed-down qual to reference. The
4176 : : * conditions checked here are:
4177 : : *
4178 : : * 1. We must not push down any quals that refer to subselect outputs that
4179 : : * return sets, else we'd introduce functions-returning-sets into the
4180 : : * subquery's WHERE/HAVING quals.
4181 : : *
4182 : : * 2. We must not push down any quals that refer to subselect outputs that
4183 : : * contain volatile functions, for fear of introducing strange results due
4184 : : * to multiple evaluation of a volatile function.
4185 : : *
4186 : : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
4187 : : * refer to non-DISTINCT output columns, because that could change the set
4188 : : * of rows returned. (This condition is vacuous for DISTINCT, because then
4189 : : * there are no non-DISTINCT output columns, so we needn't check. Note that
4190 : : * subquery_is_pushdown_safe already reported that we can't use volatile
4191 : : * quals if there's DISTINCT or DISTINCT ON.)
4192 : : *
4193 : : * 4. If the subquery has any window functions, we must not push down quals
4194 : : * that reference any output columns that are not listed in all the subquery's
4195 : : * window PARTITION BY clauses. We can push down quals that use only
4196 : : * partitioning columns because they should succeed or fail identically for
4197 : : * every row of any one window partition, and totally excluding some
4198 : : * partitions will not change a window function's results for remaining
4199 : : * partitions. (Again, this also requires nonvolatile quals, but
4200 : : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
4201 : : * unsafe for this reason can still have WindowClause run conditions pushed
4202 : : * down.
4203 : : */
4204 : : static void
4191 4205 : 1021 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
4206 : : {
4207 : : ListCell *lc;
4208 : :
4578 4209 [ + - + + : 10721 : foreach(lc, subquery->targetList)
+ + ]
4210 : : {
4211 : 9700 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4212 : :
4213 [ + + ]: 9700 : if (tle->resjunk)
4214 : 67 : continue; /* ignore resjunk columns */
4215 : :
4216 : : /* Functions returning sets are unsafe (point 1) */
3382 4217 [ + + ]: 9633 : if (subquery->hasTargetSRFs &&
1006 drowley@postgresql.o 4218 [ + - ]: 334 : (safetyInfo->unsafeFlags[tle->resno] &
4219 [ + + ]: 334 : UNSAFE_HAS_SET_FUNC) == 0 &&
3382 tgl@sss.pgh.pa.us 4220 : 334 : expression_returns_set((Node *) tle->expr))
4221 : : {
1006 drowley@postgresql.o 4222 : 188 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
4578 tgl@sss.pgh.pa.us 4223 : 188 : continue;
4224 : : }
4225 : :
4226 : : /* Volatile functions are unsafe (point 2) */
1006 drowley@postgresql.o 4227 [ + + ]: 9445 : if ((safetyInfo->unsafeFlags[tle->resno] &
4228 [ + + ]: 9439 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
4229 : 9439 : contain_volatile_functions((Node *) tle->expr))
4230 : : {
4231 : 39 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
4578 tgl@sss.pgh.pa.us 4232 : 39 : continue;
4233 : : }
4234 : :
4235 : : /* If subquery uses DISTINCT ON, check point 3 */
4236 [ - + ]: 9406 : if (subquery->hasDistinctOn &&
1006 drowley@postgresql.o 4237 [ # # ]:UBC 0 : (safetyInfo->unsafeFlags[tle->resno] &
4238 : 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4578 tgl@sss.pgh.pa.us 4239 [ # # ]: 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
4240 : : {
4241 : : /* non-DISTINCT column, so mark it unsafe */
1006 drowley@postgresql.o 4242 : 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
4578 tgl@sss.pgh.pa.us 4243 : 0 : continue;
4244 : : }
4245 : :
4246 : : /* If subquery uses window functions, check point 4 */
4191 tgl@sss.pgh.pa.us 4247 [ + + ]:CBC 9406 : if (subquery->hasWindowFuncs &&
1006 drowley@postgresql.o 4248 [ + - ]: 545 : (safetyInfo->unsafeFlags[tle->resno] &
4249 : 1046 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4191 tgl@sss.pgh.pa.us 4250 [ + + ]: 545 : !targetIsInAllPartitionLists(tle, subquery))
4251 : : {
4252 : : /* not present in all PARTITION BY clauses, so mark it unsafe */
1006 drowley@postgresql.o 4253 : 501 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
4191 tgl@sss.pgh.pa.us 4254 : 501 : continue;
4255 : : }
4256 : : }
4578 4257 : 1021 : }
4258 : :
4259 : : /*
4260 : : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
4261 : : * push quals into each component query, but the quals can only reference
4262 : : * subquery columns that suffer no type coercions in the set operation.
4263 : : * Otherwise there are possible semantic gotchas. So, we check the
4264 : : * component queries to see if any of them have output types different from
4265 : : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
4266 : : * unsafeFlags[k] if column k has different type in any component.
4267 : : *
4268 : : * We don't have to care about typmods here: the only allowed difference
4269 : : * between set-op input and output typmods is input is a specific typmod
4270 : : * and output is -1, and that does not require a coercion.
4271 : : *
4272 : : * tlist is a subquery tlist.
4273 : : * colTypes is an OID list of the top-level setop's output column types.
4274 : : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
4275 : : */
4276 : : static void
8273 4277 : 188 : compare_tlist_datatypes(List *tlist, List *colTypes,
4278 : : pushdown_safety_info *safetyInfo)
4279 : : {
4280 : : ListCell *l;
7875 neilc@samurai.com 4281 : 188 : ListCell *colType = list_head(colTypes);
4282 : :
4283 [ + - + + : 588 : foreach(l, tlist)
+ + ]
4284 : : {
4285 : 400 : TargetEntry *tle = (TargetEntry *) lfirst(l);
4286 : :
7560 tgl@sss.pgh.pa.us 4287 [ - + ]: 400 : if (tle->resjunk)
8273 tgl@sss.pgh.pa.us 4288 :UBC 0 : continue; /* ignore resjunk columns */
7875 neilc@samurai.com 4289 [ - + ]:CBC 400 : if (colType == NULL)
8273 tgl@sss.pgh.pa.us 4290 [ # # ]:UBC 0 : elog(ERROR, "wrong number of tlist entries");
7560 tgl@sss.pgh.pa.us 4291 [ + + ]:CBC 400 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
1006 drowley@postgresql.o 4292 : 68 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
2347 tgl@sss.pgh.pa.us 4293 : 400 : colType = lnext(colTypes, colType);
4294 : : }
7875 neilc@samurai.com 4295 [ - + ]: 188 : if (colType != NULL)
8273 tgl@sss.pgh.pa.us 4296 [ # # ]:UBC 0 : elog(ERROR, "wrong number of tlist entries");
8273 tgl@sss.pgh.pa.us 4297 :CBC 188 : }
4298 : :
4299 : : /*
4300 : : * targetIsInAllPartitionLists
4301 : : * True if the TargetEntry is listed in the PARTITION BY clause
4302 : : * of every window defined in the query.
4303 : : *
4304 : : * It would be safe to ignore windows not actually used by any window
4305 : : * function, but it's not easy to get that info at this stage; and it's
4306 : : * unlikely to be useful to spend any extra cycles getting it, since
4307 : : * unreferenced window definitions are probably infrequent in practice.
4308 : : */
4309 : : static bool
4191 4310 : 545 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
4311 : : {
4312 : : ListCell *lc;
4313 : :
4314 [ + - + + : 601 : foreach(lc, query->windowClause)
+ + ]
4315 : : {
4316 : 557 : WindowClause *wc = (WindowClause *) lfirst(lc);
4317 : :
4318 [ + + ]: 557 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
4319 : 501 : return false;
4320 : : }
4321 : 44 : return true;
4322 : : }
4323 : :
4324 : : /*
4325 : : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
4326 : : *
4327 : : * rinfo is a restriction clause applying to the given subquery (whose RTE
4328 : : * has index rti in the parent query).
4329 : : *
4330 : : * Conditions checked here:
4331 : : *
4332 : : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
4333 : : * unclear that it will work correctly: SubLinks will already have been
4334 : : * transformed into SubPlans in the qual, but not in the subquery). Note that
4335 : : * SubLinks that transform to initplans are safe, and will be accepted here
4336 : : * because what we'll see in the qual is just a Param referencing the initplan
4337 : : * output.
4338 : : *
4339 : : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
4340 : : * functions.
4341 : : *
4342 : : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
4343 : : * functions that are passed Var nodes, and therefore might reveal values from
4344 : : * the subquery as side effects.
4345 : : *
4346 : : * 4. rinfo's clause must not refer to the whole-row output of the subquery
4347 : : * (since there is no easy way to name that within the subquery itself).
4348 : : *
4349 : : * 5. rinfo's clause must not refer to any subquery output columns that were
4350 : : * found to be unsafe to reference by subquery_is_pushdown_safe().
4351 : : */
4352 : : static pushdown_safe_type
1724 drowley@postgresql.o 4353 : 1401 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
4354 : : pushdown_safety_info *safetyInfo)
4355 : : {
1006 4356 : 1401 : pushdown_safe_type safe = PUSHDOWN_SAFE;
1724 4357 : 1401 : Node *qual = (Node *) rinfo->clause;
4358 : : List *vars;
4359 : : ListCell *vl;
4360 : :
4361 : : /* Refuse subselects (point 1) */
8306 tgl@sss.pgh.pa.us 4362 [ + + ]: 1401 : if (contain_subplans(qual))
1006 drowley@postgresql.o 4363 : 33 : return PUSHDOWN_UNSAFE;
4364 : :
4365 : : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
4191 tgl@sss.pgh.pa.us 4366 [ + + + + ]: 1694 : if (safetyInfo->unsafeVolatile &&
1724 drowley@postgresql.o 4367 : 326 : contain_volatile_functions((Node *) rinfo))
1006 4368 : 9 : return PUSHDOWN_UNSAFE;
4369 : :
4370 : : /* Refuse leaky quals if told to (point 3) */
4191 tgl@sss.pgh.pa.us 4371 [ + + + + ]: 1949 : if (safetyInfo->unsafeLeaky &&
3887 sfrost@snowman.net 4372 : 590 : contain_leaked_vars(qual))
1006 drowley@postgresql.o 4373 : 81 : return PUSHDOWN_UNSAFE;
4374 : :
4375 : : /*
4376 : : * Examine all Vars used in clause. Since it's a restriction clause, all
4377 : : * such Vars must refer to subselect output columns ... unless this is
4378 : : * part of a LATERAL subquery, in which case there could be lateral
4379 : : * references.
4380 : : *
4381 : : * By omitting the relevant flags, this also gives us a cheap sanity check
4382 : : * that no aggregates or window functions appear in the qual. Those would
4383 : : * be unsafe to push down, but at least for the moment we could never see
4384 : : * any in a qual anyhow.
4385 : : */
3569 tgl@sss.pgh.pa.us 4386 : 1278 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
8306 4387 [ + + + + : 2506 : foreach(vl, vars)
+ + ]
4388 : : {
8171 bruce@momjian.us 4389 : 1332 : Var *var = (Var *) lfirst(vl);
4390 : :
4391 : : /*
4392 : : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
4393 : : * It's not clear whether a PHV could safely be pushed down, and even
4394 : : * less clear whether such a situation could arise in any cases of
4395 : : * practical interest anyway. So for the moment, just refuse to push
4396 : : * down.
4397 : : */
6266 tgl@sss.pgh.pa.us 4398 [ - + ]: 1332 : if (!IsA(var, Var))
4399 : : {
1006 drowley@postgresql.o 4400 :UBC 0 : safe = PUSHDOWN_UNSAFE;
6266 tgl@sss.pgh.pa.us 4401 : 0 : break;
4402 : : }
4403 : :
4404 : : /*
4405 : : * Punt if we find any lateral references. It would be safe to push
4406 : : * these down, but we'd have to convert them into outer references,
4407 : : * which subquery_push_qual lacks the infrastructure to do. The case
4408 : : * arises so seldom that it doesn't seem worth working hard on.
4409 : : */
1983 tgl@sss.pgh.pa.us 4410 [ + + ]:CBC 1332 : if (var->varno != rti)
4411 : : {
1006 drowley@postgresql.o 4412 : 6 : safe = PUSHDOWN_UNSAFE;
1983 tgl@sss.pgh.pa.us 4413 : 6 : break;
4414 : : }
4415 : :
4416 : : /* Subqueries have no system columns */
4578 4417 [ - + ]: 1326 : Assert(var->varattno >= 0);
4418 : :
4419 : : /* Check point 4 */
7247 4420 [ - + ]: 1326 : if (var->varattno == 0)
4421 : : {
1006 drowley@postgresql.o 4422 :UBC 0 : safe = PUSHDOWN_UNSAFE;
7247 tgl@sss.pgh.pa.us 4423 : 0 : break;
4424 : : }
4425 : :
4426 : : /* Check point 5 */
1006 drowley@postgresql.o 4427 [ + + ]:CBC 1326 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
4428 : : {
4429 [ + + ]: 263 : if (safetyInfo->unsafeFlags[var->varattno] &
4430 : : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
4431 : : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
4432 : : {
4433 : 98 : safe = PUSHDOWN_UNSAFE;
4434 : 98 : break;
4435 : : }
4436 : : else
4437 : : {
4438 : : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
4439 : 165 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
4440 : : /* don't break, we might find another Var that's unsafe */
4441 : : }
4442 : : }
4443 : : }
4444 : :
7871 neilc@samurai.com 4445 : 1278 : list_free(vars);
4446 : :
8306 tgl@sss.pgh.pa.us 4447 : 1278 : return safe;
4448 : : }
4449 : :
4450 : : /*
4451 : : * subquery_push_qual - push down a qual that we have determined is safe
4452 : : */
4453 : : static void
7501 4454 : 1212 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
4455 : : {
8511 4456 [ + + ]: 1212 : if (subquery->setOperations != NULL)
4457 : : {
4458 : : /* Recurse to push it separately to each component query */
7790 4459 : 82 : recurse_push_qual(subquery->setOperations, subquery,
4460 : : rte, rti, qual);
4461 : : }
4462 : : else
4463 : : {
4464 : : /*
4465 : : * We need to replace Vars in the qual (which must refer to outputs of
4466 : : * the subquery) with copies of the subquery's targetlist expressions.
4467 : : * Note that at this point, any uplevel Vars in the qual should have
4468 : : * been replaced with Params, so they need no work.
4469 : : *
4470 : : * This step also ensures that when we are pushing into a setop tree,
4471 : : * each component query gets its own copy of the qual.
4472 : : */
4787 4473 : 1130 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
4474 : : subquery->targetList,
4475 : : subquery->resultRelation,
4476 : : REPLACEVARS_REPORT_ERROR, 0,
4477 : : &subquery->hasSubLinks);
4478 : :
4479 : : /*
4480 : : * Now attach the qual to the proper place: normally WHERE, but if the
4481 : : * subquery uses grouping or aggregation, put it in HAVING (since the
4482 : : * qual really refers to the group-result rows).
4483 : : */
3868 andres@anarazel.de 4484 [ + + + - : 1130 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
+ - - + ]
7587 tgl@sss.pgh.pa.us 4485 : 190 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
4486 : : else
4487 : 940 : subquery->jointree->quals =
4488 : 940 : make_and_qual(subquery->jointree->quals, qual);
4489 : :
4490 : : /*
4491 : : * We need not change the subquery's hasAggs or hasSubLinks flags,
4492 : : * since we can't be pushing down any aggregates that weren't there
4493 : : * before, and we don't push down subselects at all.
4494 : : */
4495 : : }
8511 4496 : 1212 : }
4497 : :
4498 : : /*
4499 : : * Helper routine to recurse through setOperations tree
4500 : : */
4501 : : static void
4502 : 246 : recurse_push_qual(Node *setOp, Query *topquery,
4503 : : RangeTblEntry *rte, Index rti, Node *qual)
4504 : : {
4505 [ + + ]: 246 : if (IsA(setOp, RangeTblRef))
4506 : : {
4507 : 164 : RangeTblRef *rtr = (RangeTblRef *) setOp;
7891 4508 : 164 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4509 : 164 : Query *subquery = subrte->subquery;
4510 : :
8511 4511 [ - + ]: 164 : Assert(subquery != NULL);
7501 4512 : 164 : subquery_push_qual(subquery, rte, rti, qual);
4513 : : }
8511 4514 [ + - ]: 82 : else if (IsA(setOp, SetOperationStmt))
4515 : : {
4516 : 82 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4517 : :
7501 4518 : 82 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4519 : 82 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4520 : : }
4521 : : else
4522 : : {
8181 tgl@sss.pgh.pa.us 4523 [ # # ]:UBC 0 : elog(ERROR, "unrecognized node type: %d",
4524 : : (int) nodeTag(setOp));
4525 : : }
8511 tgl@sss.pgh.pa.us 4526 :CBC 246 : }
4527 : :
4528 : : /*****************************************************************************
4529 : : * SIMPLIFYING SUBQUERY TARGETLISTS
4530 : : *****************************************************************************/
4531 : :
4532 : : /*
4533 : : * remove_unused_subquery_outputs
4534 : : * Remove subquery targetlist items we don't need
4535 : : *
4536 : : * It's possible, even likely, that the upper query does not read all the
4537 : : * output columns of the subquery. We can remove any such outputs that are
4538 : : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4539 : : * affect semantics otherwise (e.g., volatile functions can't be removed).
4540 : : * This is useful not only because we might be able to remove expensive-to-
4541 : : * compute expressions, but because deletion of output columns might allow
4542 : : * optimizations such as join removal to occur within the subquery.
4543 : : *
4544 : : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4545 : : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4546 : : * parameter is modified by the function, so callers must make a copy if they
4547 : : * need to use the passed in Bitmapset after calling this function.
4548 : : *
4549 : : * To avoid affecting column numbering in the targetlist, we don't physically
4550 : : * remove unused tlist entries, but rather replace their expressions with NULL
4551 : : * constants. This is implemented by modifying subquery->targetList.
4552 : : */
4553 : : static void
1300 drowley@postgresql.o 4554 : 11713 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4555 : : Bitmapset *extra_used_attrs)
4556 : : {
4557 : : Bitmapset *attrs_used;
4558 : : ListCell *lc;
4559 : :
4560 : : /*
4561 : : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4562 : : * the current callers use the Bitmapset after calling this function.
4563 : : */
4564 : 11713 : attrs_used = extra_used_attrs;
4565 : :
4566 : : /*
4567 : : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4568 : : * could update all the child SELECTs' tlists, but it seems not worth the
4569 : : * trouble presently.
4570 : : */
4206 tgl@sss.pgh.pa.us 4571 [ + + ]: 11713 : if (subquery->setOperations)
4572 : 1027 : return;
4573 : :
4574 : : /*
4575 : : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4576 : : * time: all its output columns must be used in the distinctClause.
4577 : : */
4578 [ + + + + ]: 11272 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4579 : 434 : return;
4580 : :
4581 : : /*
4582 : : * Collect a bitmap of all the output column numbers used by the upper
4583 : : * query.
4584 : : *
4585 : : * Add all the attributes needed for joins or final output. Note: we must
4586 : : * look at rel's targetlist, not the attr_needed data, because attr_needed
4587 : : * isn't computed for inheritance child rels, cf set_append_rel_size().
4588 : : * (XXX might be worth changing that sometime.)
4589 : : */
3565 4590 : 10838 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4591 : :
4592 : : /* Add all the attributes used by un-pushed-down restriction clauses. */
4206 4593 [ + + + + : 11206 : foreach(lc, rel->baserestrictinfo)
+ + ]
4594 : : {
4595 : 368 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4596 : :
4597 : 368 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4598 : : }
4599 : :
4600 : : /*
4601 : : * If there's a whole-row reference to the subquery, we can't remove
4602 : : * anything.
4603 : : */
4604 [ + + ]: 10838 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4605 : 152 : return;
4606 : :
4607 : : /*
4608 : : * Run through the tlist and zap entries we don't need. It's okay to
4609 : : * modify the tlist items in-place because set_subquery_pathlist made a
4610 : : * copy of the subquery.
4611 : : */
4612 [ + + + + : 54692 : foreach(lc, subquery->targetList)
+ + ]
4613 : : {
4614 : 44006 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4615 : 44006 : Node *texpr = (Node *) tle->expr;
4616 : :
4617 : : /*
4618 : : * If it has a sortgroupref number, it's used in some sort/group
4619 : : * clause so we'd better not remove it. Also, don't remove any
4620 : : * resjunk columns, since their reason for being has nothing to do
4621 : : * with anybody reading the subquery's output. (It's likely that
4622 : : * resjunk columns in a sub-SELECT would always have ressortgroupref
4623 : : * set, but even if they don't, it seems imprudent to remove them.)
4624 : : */
4625 [ + + - + ]: 44006 : if (tle->ressortgroupref || tle->resjunk)
4626 : 1469 : continue;
4627 : :
4628 : : /*
4629 : : * If it's used by the upper query, we can't remove it.
4630 : : */
4631 [ + + ]: 42537 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4632 : : attrs_used))
4633 : 24196 : continue;
4634 : :
4635 : : /*
4636 : : * If it contains a set-returning function, we can't remove it since
4637 : : * that could change the number of rows returned by the subquery.
4638 : : */
3382 4639 [ + + + + ]: 18869 : if (subquery->hasTargetSRFs &&
4640 : 528 : expression_returns_set(texpr))
4206 4641 : 394 : continue;
4642 : :
4643 : : /*
4644 : : * If it contains volatile functions, we daren't remove it for fear
4645 : : * that the user is expecting their side-effects to happen.
4646 : : */
4647 [ + + ]: 17947 : if (contain_volatile_functions(texpr))
4648 : 16 : continue;
4649 : :
4650 : : /*
4651 : : * OK, we don't need it. Replace the expression with a NULL constant.
4652 : : * Preserve the exposed type of the expression, in case something
4653 : : * looks at the rowtype of the subquery's result.
4654 : : */
4655 : 17931 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4656 : : exprTypmod(texpr),
4657 : : exprCollation(texpr));
4658 : : }
4659 : : }
4660 : :
4661 : : /*
4662 : : * create_partial_bitmap_paths
4663 : : * Build partial bitmap heap path for the relation
4664 : : */
4665 : : void
3206 rhaas@postgresql.org 4666 : 74373 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4667 : : Path *bitmapqual)
4668 : : {
4669 : : int parallel_workers;
4670 : : double pages_fetched;
4671 : :
4672 : : /* Compute heap pages for bitmap heap scan */
4673 : 74373 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4674 : : NULL, NULL);
4675 : :
2875 4676 : 74373 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4677 : : max_parallel_workers_per_gather);
4678 : :
3206 4679 [ + + ]: 74373 : if (parallel_workers <= 0)
4680 : 72229 : return;
4681 : :
4682 : 2144 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4683 : : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4684 : : }
4685 : :
4686 : : /*
4687 : : * Compute the number of parallel workers that should be used to scan a
4688 : : * relation. We compute the parallel workers based on the size of the heap to
4689 : : * be scanned and the size of the index to be scanned, then choose a minimum
4690 : : * of those.
4691 : : *
4692 : : * "heap_pages" is the number of pages from the table that we expect to scan, or
4693 : : * -1 if we don't expect to scan any.
4694 : : *
4695 : : * "index_pages" is the number of pages from the index that we expect to scan, or
4696 : : * -1 if we don't expect to scan any.
4697 : : *
4698 : : * "max_workers" is caller's limit on the number of workers. This typically
4699 : : * comes from a GUC.
4700 : : */
4701 : : int
2875 4702 : 389136 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4703 : : int max_workers)
4704 : : {
3227 4705 : 389136 : int parallel_workers = 0;
4706 : :
4707 : : /*
4708 : : * If the user has set the parallel_workers reloption, use that; otherwise
4709 : : * select a default number of workers.
4710 : : */
3255 4711 [ + + ]: 389136 : if (rel->rel_parallel_workers != -1)
4712 : 957 : parallel_workers = rel->rel_parallel_workers;
4713 : : else
4714 : : {
4715 : : /*
4716 : : * If the number of pages being scanned is insufficient to justify a
4717 : : * parallel scan, just return zero ... unless it's an inheritance
4718 : : * child. In that case, we want to generate a parallel path here
4719 : : * anyway. It might not be worthwhile just for this relation, but
4720 : : * when combined with all of its inheritance siblings it may well pay
4721 : : * off.
4722 : : */
3200 4723 [ + + + + ]: 388179 : if (rel->reloptkind == RELOPT_BASEREL &&
4724 [ + + + + ]: 367850 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
3101 tgl@sss.pgh.pa.us 4725 [ + + ]: 11333 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
3255 rhaas@postgresql.org 4726 : 367391 : return 0;
4727 : :
3200 4728 [ + + ]: 20788 : if (heap_pages >= 0)
4729 : : {
4730 : : int heap_parallel_threshold;
4731 : 19725 : int heap_parallel_workers = 1;
4732 : :
4733 : : /*
4734 : : * Select the number of workers based on the log of the size of
4735 : : * the relation. This probably needs to be a good deal more
4736 : : * sophisticated, but we need something here for now. Note that
4737 : : * the upper limit of the min_parallel_table_scan_size GUC is
4738 : : * chosen to prevent overflow here.
4739 : : */
3227 4740 : 19725 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4741 [ + + ]: 22419 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4742 : : {
4743 : 2694 : heap_parallel_workers++;
4744 : 2694 : heap_parallel_threshold *= 3;
4745 [ - + ]: 2694 : if (heap_parallel_threshold > INT_MAX / 3)
3227 rhaas@postgresql.org 4746 :UBC 0 : break; /* avoid overflow */
4747 : : }
4748 : :
3227 rhaas@postgresql.org 4749 :CBC 19725 : parallel_workers = heap_parallel_workers;
4750 : : }
4751 : :
3200 4752 [ + + ]: 20788 : if (index_pages >= 0)
4753 : : {
4754 : 4962 : int index_parallel_workers = 1;
4755 : : int index_parallel_threshold;
4756 : :
4757 : : /* same calculation as for heap_pages above */
3227 4758 : 4962 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4759 [ + + ]: 5100 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4760 : : {
4761 : 138 : index_parallel_workers++;
4762 : 138 : index_parallel_threshold *= 3;
4763 [ - + ]: 138 : if (index_parallel_threshold > INT_MAX / 3)
3227 rhaas@postgresql.org 4764 :UBC 0 : break; /* avoid overflow */
4765 : : }
4766 : :
3227 rhaas@postgresql.org 4767 [ + + ]:CBC 4962 : if (parallel_workers > 0)
4768 : 3899 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4769 : : else
4770 : 1063 : parallel_workers = index_parallel_workers;
4771 : : }
4772 : : }
4773 : :
4774 : : /* In no case use more than caller supplied maximum number of workers */
2875 4775 : 21745 : parallel_workers = Min(parallel_workers, max_workers);
4776 : :
3255 4777 : 21745 : return parallel_workers;
4778 : : }
4779 : :
4780 : : /*
4781 : : * generate_partitionwise_join_paths
4782 : : * Create paths representing partitionwise join for given partitioned
4783 : : * join relation.
4784 : : *
4785 : : * This must not be called until after we are done adding paths for all
4786 : : * child-joins. Otherwise, add_path might delete a path to which some path
4787 : : * generated here has a reference.
4788 : : */
4789 : : void
2861 peter_e@gmx.net 4790 : 120305 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4791 : : {
2994 rhaas@postgresql.org 4792 : 120305 : List *live_children = NIL;
4793 : : int cnt_parts;
4794 : : int num_parts;
4795 : : RelOptInfo **part_rels;
4796 : :
4797 : : /* Handle only join relations here. */
4798 [ + + - + ]: 120305 : if (!IS_JOIN_REL(rel))
2994 rhaas@postgresql.org 4799 :UBC 0 : return;
4800 : :
4801 : : /* We've nothing to do if the relation is not partitioned. */
2872 rhaas@postgresql.org 4802 [ + + + + :CBC 120305 : if (!IS_PARTITIONED_REL(rel))
+ + + - +
+ ]
2994 4803 : 116732 : return;
4804 : :
4805 : : /* The relation should have consider_partitionwise_join set. */
2665 efujita@postgresql.o 4806 [ - + ]: 3573 : Assert(rel->consider_partitionwise_join);
4807 : :
4808 : : /* Guard against stack overflow due to overly deep partition hierarchy. */
2994 rhaas@postgresql.org 4809 : 3573 : check_stack_depth();
4810 : :
4811 : 3573 : num_parts = rel->nparts;
4812 : 3573 : part_rels = rel->part_rels;
4813 : :
4814 : : /* Collect non-dummy child-joins. */
4815 [ + + ]: 12724 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4816 : : {
4817 : 9151 : RelOptInfo *child_rel = part_rels[cnt_parts];
4818 : :
4819 : : /* If it's been pruned entirely, it's certainly dummy. */
2454 tgl@sss.pgh.pa.us 4820 [ + + ]: 9151 : if (child_rel == NULL)
4821 : 32 : continue;
4822 : :
4823 : : /* Make partitionwise join paths for this partitioned child-join. */
2861 peter_e@gmx.net 4824 : 9119 : generate_partitionwise_join_paths(root, child_rel);
4825 : :
4826 : : /* If we failed to make any path for this child, we must give up. */
1109 tgl@sss.pgh.pa.us 4827 [ - + ]: 9119 : if (child_rel->pathlist == NIL)
4828 : : {
4829 : : /*
4830 : : * Mark the parent joinrel as unpartitioned so that later
4831 : : * functions treat it correctly.
4832 : : */
1109 tgl@sss.pgh.pa.us 4833 :UBC 0 : rel->nparts = 0;
4834 : 0 : return;
4835 : : }
4836 : :
4837 : : /* Else, identify the cheapest path for it. */
2477 tgl@sss.pgh.pa.us 4838 :CBC 9119 : set_cheapest(child_rel);
4839 : :
4840 : : /* Dummy children need not be scanned, so ignore those. */
2994 rhaas@postgresql.org 4841 [ - + ]: 9119 : if (IS_DUMMY_REL(child_rel))
2994 rhaas@postgresql.org 4842 :UBC 0 : continue;
4843 : :
4844 : : /*
4845 : : * Except for the topmost scan/join rel, consider generating partial
4846 : : * aggregation paths for the grouped relation on top of the paths of
4847 : : * this partitioned child-join. After that, we're done creating paths
4848 : : * for the grouped relation, so run set_cheapest().
4849 : : */
70 rguo@postgresql.org 4850 [ + + ]:GNC 9119 : if (child_rel->grouped_rel != NULL &&
4851 [ + + - + : 6438 : !bms_equal(IS_OTHER_REL(rel) ?
+ + ]
4852 : : rel->top_parent_relids : rel->relids,
4853 [ + - ]: 6438 : root->all_query_rels))
4854 : : {
4855 : 120 : RelOptInfo *grouped_rel = child_rel->grouped_rel;
4856 : :
4857 [ - + ]: 120 : Assert(IS_GROUPED_REL(grouped_rel));
4858 : :
4859 : 120 : generate_grouped_paths(root, grouped_rel, child_rel);
4860 : 120 : set_cheapest(grouped_rel);
4861 : : }
4862 : :
4863 : : #ifdef OPTIMIZER_DEBUG
4864 : : pprint(child_rel);
4865 : : #endif
4866 : :
2994 rhaas@postgresql.org 4867 :CBC 9119 : live_children = lappend(live_children, child_rel);
4868 : : }
4869 : :
4870 : : /* If all child-joins are dummy, parent join is also dummy. */
4871 [ - + ]: 3573 : if (!live_children)
4872 : : {
2994 rhaas@postgresql.org 4873 :UBC 0 : mark_dummy_rel(rel);
4874 : 0 : return;
4875 : : }
4876 : :
4877 : : /* Build additional paths for this rel from child-join paths. */
2994 rhaas@postgresql.org 4878 :CBC 3573 : add_paths_to_append_rel(root, rel, live_children);
4879 : 3573 : list_free(live_children);
4880 : : }
|
