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) 1996-2024, 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/tlist.h"
44 : : #include "parser/parse_clause.h"
45 : : #include "parser/parsetree.h"
46 : : #include "partitioning/partbounds.h"
47 : : #include "port/pg_bitutils.h"
48 : : #include "rewrite/rewriteManip.h"
49 : : #include "utils/lsyscache.h"
50 : :
51 : :
52 : : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
53 : : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
54 : : #define UNSAFE_HAS_SET_FUNC (1 << 1)
55 : : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
56 : : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
57 : : #define UNSAFE_TYPE_MISMATCH (1 << 4)
58 : :
59 : : /* results of subquery_is_pushdown_safe */
60 : : typedef struct pushdown_safety_info
61 : : {
62 : : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
63 : : * column is unsafe for qual pushdown, or 0 if
64 : : * no reason. */
65 : : bool unsafeVolatile; /* don't push down volatile quals */
66 : : bool unsafeLeaky; /* don't push down leaky quals */
67 : : } pushdown_safety_info;
68 : :
69 : : /* Return type for qual_is_pushdown_safe */
70 : : typedef enum pushdown_safe_type
71 : : {
72 : : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
73 : : PUSHDOWN_SAFE, /* safe to push qual into subquery */
74 : : PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
75 : : * run condition */
76 : : } pushdown_safe_type;
77 : :
78 : : /* These parameters are set by GUC */
79 : : bool enable_geqo = false; /* just in case GUC doesn't set it */
80 : : int geqo_threshold;
81 : : int min_parallel_table_scan_size;
82 : : int min_parallel_index_scan_size;
83 : :
84 : : /* Hook for plugins to get control in set_rel_pathlist() */
85 : : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
86 : :
87 : : /* Hook for plugins to replace standard_join_search() */
88 : : join_search_hook_type join_search_hook = NULL;
89 : :
90 : :
91 : : static void set_base_rel_consider_startup(PlannerInfo *root);
92 : : static void set_base_rel_sizes(PlannerInfo *root);
93 : : static void set_base_rel_pathlists(PlannerInfo *root);
94 : : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
95 : : Index rti, RangeTblEntry *rte);
96 : : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
97 : : Index rti, RangeTblEntry *rte);
98 : : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
99 : : RangeTblEntry *rte);
100 : : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
101 : : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
102 : : RangeTblEntry *rte);
103 : : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
104 : : RangeTblEntry *rte);
105 : : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
106 : : RangeTblEntry *rte);
107 : : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
108 : : RangeTblEntry *rte);
109 : : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
110 : : RangeTblEntry *rte);
111 : : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
112 : : RangeTblEntry *rte);
113 : : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
114 : : Index rti, RangeTblEntry *rte);
115 : : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
116 : : Index rti, RangeTblEntry *rte);
117 : : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
118 : : List *live_childrels,
119 : : List *all_child_pathkeys);
120 : : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
121 : : RelOptInfo *rel,
122 : : Relids required_outer);
123 : : static void accumulate_append_subpath(Path *path,
124 : : List **subpaths,
125 : : List **special_subpaths);
126 : : static Path *get_singleton_append_subpath(Path *path);
127 : : static void set_dummy_rel_pathlist(RelOptInfo *rel);
128 : : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
129 : : Index rti, RangeTblEntry *rte);
130 : : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
131 : : RangeTblEntry *rte);
132 : : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
133 : : RangeTblEntry *rte);
134 : : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
135 : : RangeTblEntry *rte);
136 : : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : : RangeTblEntry *rte);
138 : : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : : RangeTblEntry *rte);
140 : : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : : RangeTblEntry *rte);
142 : : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : : RangeTblEntry *rte);
144 : : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
145 : : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
146 : : pushdown_safety_info *safetyInfo);
147 : : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
148 : : pushdown_safety_info *safetyInfo);
149 : : static void check_output_expressions(Query *subquery,
150 : : pushdown_safety_info *safetyInfo);
151 : : static void compare_tlist_datatypes(List *tlist, List *colTypes,
152 : : pushdown_safety_info *safetyInfo);
153 : : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
154 : : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
155 : : RestrictInfo *rinfo,
156 : : pushdown_safety_info *safetyInfo);
157 : : static void subquery_push_qual(Query *subquery,
158 : : RangeTblEntry *rte, Index rti, Node *qual);
159 : : static void recurse_push_qual(Node *setOp, Query *topquery,
160 : : RangeTblEntry *rte, Index rti, Node *qual);
161 : : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
162 : : Bitmapset *extra_used_attrs);
163 : :
164 : :
165 : : /*
166 : : * make_one_rel
167 : : * Finds all possible access paths for executing a query, returning a
168 : : * single rel that represents the join of all base rels in the query.
169 : : */
170 : : RelOptInfo *
6690 tgl@sss.pgh.pa.us 171 :CBC 139535 : make_one_rel(PlannerInfo *root, List *joinlist)
172 : : {
173 : : RelOptInfo *rel;
174 : : Index rti;
175 : : double total_pages;
176 : :
177 : : /* Mark base rels as to whether we care about fast-start plans */
3238 178 : 139535 : set_base_rel_consider_startup(root);
179 : :
180 : : /*
181 : : * Compute size estimates and consider_parallel flags for each base rel.
182 : : */
4461 183 : 139535 : set_base_rel_sizes(root);
184 : :
185 : : /*
186 : : * We should now have size estimates for every actual table involved in
187 : : * the query, and we also know which if any have been deleted from the
188 : : * query by join removal, pruned by partition pruning, or eliminated by
189 : : * constraint exclusion. So we can now compute total_table_pages.
190 : : *
191 : : * Note that appendrels are not double-counted here, even though we don't
192 : : * bother to distinguish RelOptInfos for appendrel parents, because the
193 : : * parents will have pages = 0.
194 : : *
195 : : * XXX if a table is self-joined, we will count it once per appearance,
196 : : * which perhaps is the wrong thing ... but that's not completely clear,
197 : : * and detecting self-joins here is difficult, so ignore it for now.
198 : : */
1985 199 : 139521 : total_pages = 0;
200 [ + + ]: 422590 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
201 : : {
202 : 283069 : RelOptInfo *brel = root->simple_rel_array[rti];
203 : :
204 : : /* there may be empty slots corresponding to non-baserel RTEs */
205 [ + + ]: 283069 : if (brel == NULL)
206 : 65571 : continue;
207 : :
208 [ - + ]: 217498 : Assert(brel->relid == rti); /* sanity check on array */
209 : :
210 [ + + ]: 217498 : if (IS_DUMMY_REL(brel))
211 : 548 : continue;
212 : :
213 [ + + + - ]: 216950 : if (IS_SIMPLE_REL(brel))
214 : 216950 : total_pages += (double) brel->pages;
215 : : }
216 : 139521 : root->total_table_pages = total_pages;
217 : :
218 : : /*
219 : : * Generate access paths for each base rel.
220 : : */
8554 221 : 139521 : set_base_rel_pathlists(root);
222 : :
223 : : /*
224 : : * Generate access paths for the entire join tree.
225 : : */
6690 226 : 139521 : rel = make_rel_from_joinlist(root, joinlist);
227 : :
228 : : /*
229 : : * The result should join all and only the query's base + outer-join rels.
230 : : */
440 231 [ - + ]: 139520 : Assert(bms_equal(rel->relids, root->all_query_rels));
232 : :
4461 233 : 139520 : return rel;
234 : : }
235 : :
236 : : /*
237 : : * set_base_rel_consider_startup
238 : : * Set the consider_[param_]startup flags for each base-relation entry.
239 : : *
240 : : * For the moment, we only deal with consider_param_startup here; because the
241 : : * logic for consider_startup is pretty trivial and is the same for every base
242 : : * relation, we just let build_simple_rel() initialize that flag correctly to
243 : : * start with. If that logic ever gets more complicated it would probably
244 : : * be better to move it here.
245 : : */
246 : : static void
3238 247 : 139535 : set_base_rel_consider_startup(PlannerInfo *root)
248 : : {
249 : : /*
250 : : * Since parameterized paths can only be used on the inside of a nestloop
251 : : * join plan, there is usually little value in considering fast-start
252 : : * plans for them. However, for relations that are on the RHS of a SEMI
253 : : * or ANTI join, a fast-start plan can be useful because we're only going
254 : : * to care about fetching one tuple anyway.
255 : : *
256 : : * To minimize growth of planning time, we currently restrict this to
257 : : * cases where the RHS is a single base relation, not a join; there is no
258 : : * provision for consider_param_startup to get set at all on joinrels.
259 : : * Also we don't worry about appendrels. costsize.c's costing rules for
260 : : * nestloop semi/antijoins don't consider such cases either.
261 : : */
262 : : ListCell *lc;
263 : :
264 [ + + + + : 158532 : foreach(lc, root->join_info_list)
+ + ]
265 : : {
266 : 18997 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
267 : : int varno;
268 : :
269 [ + + + + : 21563 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
+ + ]
270 : 2566 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
271 : : {
272 : 2475 : RelOptInfo *rel = find_base_rel(root, varno);
273 : :
274 : 2475 : rel->consider_param_startup = true;
275 : : }
276 : : }
277 : 139535 : }
278 : :
279 : : /*
280 : : * set_base_rel_sizes
281 : : * Set the size estimates (rows and widths) for each base-relation entry.
282 : : * Also determine whether to consider parallel paths for base relations.
283 : : *
284 : : * We do this in a separate pass over the base rels so that rowcount
285 : : * estimates are available for parameterized path generation, and also so
286 : : * that each rel's consider_parallel flag is set correctly before we begin to
287 : : * generate paths.
288 : : */
289 : : static void
4461 290 : 139535 : set_base_rel_sizes(PlannerInfo *root)
291 : : {
292 : : Index rti;
293 : :
294 [ + + ]: 422605 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
295 : : {
296 : 283084 : RelOptInfo *rel = root->simple_rel_array[rti];
297 : : RangeTblEntry *rte;
298 : :
299 : : /* there may be empty slots corresponding to non-baserel RTEs */
300 [ + + ]: 283084 : if (rel == NULL)
301 : 65572 : continue;
302 : :
2489 303 [ - + ]: 217512 : Assert(rel->relid == rti); /* sanity check on array */
304 : :
305 : : /* ignore RTEs that are "other rels" */
4461 306 [ + + ]: 217512 : if (rel->reloptkind != RELOPT_BASEREL)
307 : 21369 : continue;
308 : :
3077 rhaas@postgresql.org 309 : 196143 : rte = root->simple_rte_array[rti];
310 : :
311 : : /*
312 : : * If parallelism is allowable for this query in general, see whether
313 : : * it's allowable for this rel in particular. We have to do this
314 : : * before set_rel_size(), because (a) if this rel is an inheritance
315 : : * parent, set_append_rel_size() will use and perhaps change the rel's
316 : : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
317 : : * goes ahead and makes paths immediately.
318 : : */
319 [ + + ]: 196143 : if (root->glob->parallelModeOK)
320 : 154369 : set_rel_consider_parallel(root, rel, rte);
321 : :
322 : 196143 : set_rel_size(root, rel, rti, rte);
323 : : }
10141 scrappy@hub.org 324 : 139521 : }
325 : :
326 : : /*
327 : : * set_base_rel_pathlists
328 : : * Finds all paths available for scanning each base-relation entry.
329 : : * Sequential scan and any available indices are considered.
330 : : * Each useful path is attached to its relation's 'pathlist' field.
331 : : */
332 : : static void
6888 tgl@sss.pgh.pa.us 333 : 139521 : set_base_rel_pathlists(PlannerInfo *root)
334 : : {
335 : : Index rti;
336 : :
6648 337 [ + + ]: 422590 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
338 : : {
339 : 283069 : RelOptInfo *rel = root->simple_rel_array[rti];
340 : :
341 : : /* there may be empty slots corresponding to non-baserel RTEs */
6887 342 [ + + ]: 283069 : if (rel == NULL)
343 : 65571 : continue;
344 : :
2489 345 [ - + ]: 217498 : Assert(rel->relid == rti); /* sanity check on array */
346 : :
347 : : /* ignore RTEs that are "other rels" */
6887 348 [ + + ]: 217498 : if (rel->reloptkind != RELOPT_BASEREL)
349 : 21369 : continue;
350 : :
6203 351 : 196129 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
352 : : }
6645 353 : 139521 : }
354 : :
355 : : /*
356 : : * set_rel_size
357 : : * Set size estimates for a base relation
358 : : */
359 : : static void
4461 360 : 217402 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
361 : : Index rti, RangeTblEntry *rte)
362 : : {
4586 363 [ + + + + ]: 413545 : if (rel->reloptkind == RELOPT_BASEREL &&
364 : 196143 : relation_excluded_by_constraints(root, rel, rte))
365 : : {
366 : : /*
367 : : * We proved we don't need to scan the rel via constraint exclusion,
368 : : * so set up a single dummy path for it. Here we only check this for
369 : : * regular baserels; if it's an otherrel, CE was already checked in
370 : : * set_append_rel_size().
371 : : *
372 : : * In this case, we go ahead and set up the relation's path right away
373 : : * instead of leaving it for set_rel_pathlist to do. This is because
374 : : * we don't have a convention for marking a rel as dummy except by
375 : : * assigning a dummy path to it.
376 : : */
377 : 234 : set_dummy_rel_pathlist(rel);
378 : : }
379 [ + + ]: 217168 : else if (rte->inh)
380 : : {
381 : : /* It's an "append relation", process accordingly */
4461 382 : 10128 : set_append_rel_size(root, rel, rti, rte);
383 : : }
384 : : else
385 : : {
4800 386 [ + + + + : 207040 : switch (rel->rtekind)
+ + + +
- ]
387 : : {
388 : 174602 : case RTE_RELATION:
389 [ + + ]: 174602 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
390 : : {
391 : : /* Foreign table */
4461 392 : 1179 : set_foreign_size(root, rel, rte);
393 : : }
2581 rhaas@postgresql.org 394 [ + + ]: 173423 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
395 : : {
396 : : /*
397 : : * We could get here if asked to scan a partitioned table
398 : : * with ONLY. In that case we shouldn't scan any of the
399 : : * partitions, so mark it as a dummy rel.
400 : : */
401 : 20 : set_dummy_rel_pathlist(rel);
402 : : }
3257 simon@2ndQuadrant.co 403 [ + + ]: 173403 : else if (rte->tablesample != NULL)
404 : : {
405 : : /* Sampled relation */
406 : 150 : set_tablesample_rel_size(root, rel, rte);
407 : : }
408 : : else
409 : : {
410 : : /* Plain relation */
4461 tgl@sss.pgh.pa.us 411 : 173253 : set_plain_rel_size(root, rel, rte);
412 : : }
4800 413 : 174588 : break;
414 : 3788 : case RTE_SUBQUERY:
415 : :
416 : : /*
417 : : * Subqueries don't support making a choice between
418 : : * parameterized and unparameterized paths, so just go ahead
419 : : * and build their paths immediately.
420 : : */
421 : 3788 : set_subquery_pathlist(root, rel, rti, rte);
422 : 3788 : break;
423 : 21525 : case RTE_FUNCTION:
4461 424 : 21525 : set_function_size_estimates(root, rel);
4800 425 : 21525 : break;
2594 alvherre@alvh.no-ip. 426 : 254 : case RTE_TABLEFUNC:
427 : 254 : set_tablefunc_size_estimates(root, rel);
428 : 254 : break;
4800 tgl@sss.pgh.pa.us 429 : 3858 : case RTE_VALUES:
4461 430 : 3858 : set_values_size_estimates(root, rel);
4800 431 : 3858 : break;
432 : 2009 : case RTE_CTE:
433 : :
434 : : /*
435 : : * CTEs don't support making a choice between parameterized
436 : : * and unparameterized paths, so just go ahead and build their
437 : : * paths immediately.
438 : : */
439 [ + + ]: 2009 : if (rte->self_reference)
440 : 406 : set_worktable_pathlist(root, rel, rte);
441 : : else
442 : 1603 : set_cte_pathlist(root, rel, rte);
443 : 2009 : break;
2571 kgrittn@postgresql.o 444 : 223 : case RTE_NAMEDTUPLESTORE:
445 : : /* Might as well just build the path immediately */
446 : 223 : set_namedtuplestore_pathlist(root, rel, rte);
447 : 223 : break;
1903 tgl@sss.pgh.pa.us 448 : 781 : case RTE_RESULT:
449 : : /* Might as well just build the path immediately */
450 : 781 : set_result_pathlist(root, rel, rte);
451 : 781 : break;
4800 tgl@sss.pgh.pa.us 452 :UBC 0 : default:
453 [ # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
454 : : break;
455 : : }
456 : : }
457 : :
458 : : /*
459 : : * We insist that all non-dummy rels have a nonzero rowcount estimate.
460 : : */
3185 tgl@sss.pgh.pa.us 461 [ + + - + ]:CBC 217387 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
4461 462 : 217387 : }
463 : :
464 : : /*
465 : : * set_rel_pathlist
466 : : * Build access paths for a base relation
467 : : */
468 : : static void
469 : 217411 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
470 : : Index rti, RangeTblEntry *rte)
471 : : {
472 [ + + ]: 217411 : if (IS_DUMMY_REL(rel))
473 : : {
474 : : /* We already proved the relation empty, so nothing more to do */
475 : : }
476 [ + + ]: 216932 : else if (rte->inh)
477 : : {
478 : : /* It's an "append relation", process accordingly */
479 : 9980 : set_append_rel_pathlist(root, rel, rti, rte);
480 : : }
481 : : else
482 : : {
483 [ + + + + : 206952 : switch (rel->rtekind)
+ + + +
- ]
484 : : {
485 : 174568 : case RTE_RELATION:
486 [ + + ]: 174568 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
487 : : {
488 : : /* Foreign table */
489 : 1177 : set_foreign_pathlist(root, rel, rte);
490 : : }
3257 simon@2ndQuadrant.co 491 [ + + ]: 173391 : else if (rte->tablesample != NULL)
492 : : {
493 : : /* Sampled relation */
494 : 150 : set_tablesample_rel_pathlist(root, rel, rte);
495 : : }
496 : : else
497 : : {
498 : : /* Plain relation */
4461 tgl@sss.pgh.pa.us 499 : 173241 : set_plain_rel_pathlist(root, rel, rte);
500 : : }
501 : 174568 : break;
502 : 3734 : case RTE_SUBQUERY:
503 : : /* Subquery --- fully handled during set_rel_size */
504 : 3734 : break;
505 : 21525 : case RTE_FUNCTION:
506 : : /* RangeFunction */
507 : 21525 : set_function_pathlist(root, rel, rte);
508 : 21525 : break;
2594 alvherre@alvh.no-ip. 509 : 254 : case RTE_TABLEFUNC:
510 : : /* Table Function */
511 : 254 : set_tablefunc_pathlist(root, rel, rte);
512 : 254 : break;
4461 tgl@sss.pgh.pa.us 513 : 3858 : case RTE_VALUES:
514 : : /* Values list */
515 : 3858 : set_values_pathlist(root, rel, rte);
516 : 3858 : break;
517 : 2009 : case RTE_CTE:
518 : : /* CTE reference --- fully handled during set_rel_size */
519 : 2009 : break;
2571 kgrittn@postgresql.o 520 : 223 : case RTE_NAMEDTUPLESTORE:
521 : : /* tuplestore reference --- fully handled during set_rel_size */
522 : 223 : break;
1903 tgl@sss.pgh.pa.us 523 : 781 : case RTE_RESULT:
524 : : /* simple Result --- fully handled during set_rel_size */
525 : 781 : break;
4461 tgl@sss.pgh.pa.us 526 :UBC 0 : default:
527 [ # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
528 : : break;
529 : : }
530 : : }
531 : :
532 : : /*
533 : : * Allow a plugin to editorialize on the set of Paths for this base
534 : : * relation. It could add new paths (such as CustomPaths) by calling
535 : : * add_path(), or add_partial_path() if parallel aware. It could also
536 : : * delete or modify paths added by the core code.
537 : : */
1891 tgl@sss.pgh.pa.us 538 [ - + ]:CBC 217411 : if (set_rel_pathlist_hook)
1891 tgl@sss.pgh.pa.us 539 :UBC 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
540 : :
541 : : /*
542 : : * If this is a baserel, we should normally consider gathering any partial
543 : : * paths we may have created for it. We have to do this after calling the
544 : : * set_rel_pathlist_hook, else it cannot add partial paths to be included
545 : : * here.
546 : : *
547 : : * However, if this is an inheritance child, skip it. Otherwise, we could
548 : : * end up with a very large number of gather nodes, each trying to grab
549 : : * its own pool of workers. Instead, we'll consider gathering partial
550 : : * paths for the parent appendrel.
551 : : *
552 : : * Also, if this is the topmost scan/join rel, we postpone gathering until
553 : : * the final scan/join targetlist is available (see grouping_planner).
554 : : */
2225 rhaas@postgresql.org 555 [ + + ]:CBC 217411 : if (rel->reloptkind == RELOPT_BASEREL &&
440 tgl@sss.pgh.pa.us 556 [ + + ]: 196129 : !bms_equal(rel->relids, root->all_query_rels))
1468 tomas.vondra@postgre 557 : 97259 : generate_useful_gather_paths(root, rel, false);
558 : :
559 : : /* Now find the cheapest of the paths for this rel */
3432 tgl@sss.pgh.pa.us 560 : 217411 : set_cheapest(rel);
561 : :
562 : : #ifdef OPTIMIZER_DEBUG
563 : : pprint(rel);
564 : : #endif
8554 565 : 217411 : }
566 : :
567 : : /*
568 : : * set_plain_rel_size
569 : : * Set size estimates for a plain relation (no subquery, no inheritance)
570 : : */
571 : : static void
4461 572 : 173253 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
573 : : {
574 : : /*
575 : : * Test any partial indexes of rel for applicability. We must do this
576 : : * first since partial unique indexes can affect size estimates.
577 : : */
2936 578 : 173253 : check_index_predicates(root, rel);
579 : :
580 : : /* Mark rel with estimated output rows, width, etc */
6203 581 : 173253 : set_baserel_size_estimates(root, rel);
4461 582 : 173241 : }
583 : :
584 : : /*
585 : : * If this relation could possibly be scanned from within a worker, then set
586 : : * its consider_parallel flag.
587 : : */
588 : : static void
3077 rhaas@postgresql.org 589 : 169576 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
590 : : RangeTblEntry *rte)
591 : : {
592 : : /*
593 : : * The flag has previously been initialized to false, so we can just
594 : : * return if it becomes clear that we can't safely set it.
595 : : */
2842 tgl@sss.pgh.pa.us 596 [ - + ]: 169576 : Assert(!rel->consider_parallel);
597 : :
598 : : /* Don't call this if parallelism is disallowed for the entire query. */
3077 rhaas@postgresql.org 599 [ - + ]: 169576 : Assert(root->glob->parallelModeOK);
600 : :
601 : : /* This should only be called for baserels and appendrel children. */
2568 602 [ + + - + ]: 169576 : Assert(IS_SIMPLE_REL(rel));
603 : :
604 : : /* Assorted checks based on rtekind. */
3077 605 [ + + - + : 169576 : switch (rte->rtekind)
+ + + + +
- ]
606 : : {
607 : 150823 : case RTE_RELATION:
608 : :
609 : : /*
610 : : * Currently, parallel workers can't access the leader's temporary
611 : : * tables. We could possibly relax this if we wrote all of its
612 : : * local buffers at the start of the query and made no changes
613 : : * thereafter (maybe we could allow hint bit changes), and if we
614 : : * taught the workers to read them. Writing a large number of
615 : : * temporary buffers could be expensive, though, and we don't have
616 : : * the rest of the necessary infrastructure right now anyway. So
617 : : * for now, bail out if we see a temporary table.
618 : : */
619 [ + + ]: 150823 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
620 : 3803 : return;
621 : :
622 : : /*
623 : : * Table sampling can be pushed down to workers if the sample
624 : : * function and its arguments are safe.
625 : : */
626 [ + + ]: 147020 : if (rte->tablesample != NULL)
627 : : {
2693 tgl@sss.pgh.pa.us 628 : 162 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
629 : :
3077 rhaas@postgresql.org 630 [ + + ]: 162 : if (proparallel != PROPARALLEL_SAFE)
631 : 18 : return;
2795 tgl@sss.pgh.pa.us 632 [ + + ]: 144 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
3077 rhaas@postgresql.org 633 : 6 : return;
634 : : }
635 : :
636 : : /*
637 : : * Ask FDWs whether they can support performing a ForeignScan
638 : : * within a worker. Most often, the answer will be no. For
639 : : * example, if the nature of the FDW is such that it opens a TCP
640 : : * connection with a remote server, each parallel worker would end
641 : : * up with a separate connection, and these connections might not
642 : : * be appropriately coordinated between workers and the leader.
643 : : */
2970 644 [ + + ]: 146996 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
645 : : {
646 [ - + ]: 762 : Assert(rel->fdwroutine);
647 [ + + ]: 762 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
648 : 730 : return;
649 [ - + ]: 32 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
2970 rhaas@postgresql.org 650 :UBC 0 : return;
651 : : }
652 : :
653 : : /*
654 : : * There are additional considerations for appendrels, which we'll
655 : : * deal with in set_append_rel_size and set_append_rel_pathlist.
656 : : * For now, just set consider_parallel based on the rel's own
657 : : * quals and targetlist.
658 : : */
3077 rhaas@postgresql.org 659 :CBC 146266 : break;
660 : :
661 : 3111 : case RTE_SUBQUERY:
662 : :
663 : : /*
664 : : * There's no intrinsic problem with scanning a subquery-in-FROM
665 : : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
666 : : * If the subquery doesn't happen to have any parallel-safe paths,
667 : : * then flagging it as consider_parallel won't change anything,
668 : : * but that's true for plain tables, too. We must set
669 : : * consider_parallel based on the rel's own quals and targetlist,
670 : : * so that if a subquery path is parallel-safe but the quals and
671 : : * projection we're sticking onto it are not, we correctly mark
672 : : * the SubqueryScanPath as not parallel-safe. (Note that
673 : : * set_subquery_pathlist() might push some of these quals down
674 : : * into the subquery itself, but that doesn't change anything.)
675 : : *
676 : : * We can't push sub-select containing LIMIT/OFFSET to workers as
677 : : * there is no guarantee that the row order will be fully
678 : : * deterministic, and applying LIMIT/OFFSET will lead to
679 : : * inconsistent results at the top-level. (In some cases, where
680 : : * the result is ordered, we could relax this restriction. But it
681 : : * doesn't currently seem worth expending extra effort to do so.)
682 : : */
683 : : {
2039 akapila@postgresql.o 684 : 3111 : Query *subquery = castNode(Query, rte->subquery);
685 : :
686 [ + + ]: 3111 : if (limit_needed(subquery))
687 : 223 : return;
688 : : }
2842 tgl@sss.pgh.pa.us 689 : 2888 : break;
690 : :
3077 rhaas@postgresql.org 691 :UBC 0 : case RTE_JOIN:
692 : : /* Shouldn't happen; we're only considering baserels here. */
693 : 0 : Assert(false);
694 : : return;
695 : :
3077 rhaas@postgresql.org 696 :CBC 11575 : case RTE_FUNCTION:
697 : : /* Check for parallel-restricted functions. */
2795 tgl@sss.pgh.pa.us 698 [ + + ]: 11575 : if (!is_parallel_safe(root, (Node *) rte->functions))
3077 rhaas@postgresql.org 699 : 5505 : return;
700 : 6070 : break;
701 : :
2594 alvherre@alvh.no-ip. 702 : 254 : case RTE_TABLEFUNC:
703 : : /* not parallel safe */
704 : 254 : return;
705 : :
3077 rhaas@postgresql.org 706 : 1375 : case RTE_VALUES:
707 : : /* Check for parallel-restricted functions. */
2795 tgl@sss.pgh.pa.us 708 [ + + ]: 1375 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
709 : 3 : return;
3077 rhaas@postgresql.org 710 : 1372 : break;
711 : :
712 : 1628 : case RTE_CTE:
713 : :
714 : : /*
715 : : * CTE tuplestores aren't shared among parallel workers, so we
716 : : * force all CTE scans to happen in the leader. Also, populating
717 : : * the CTE would require executing a subplan that's not available
718 : : * in the worker, might be parallel-restricted, and must get
719 : : * executed only once.
720 : : */
721 : 1628 : return;
722 : :
2571 kgrittn@postgresql.o 723 : 209 : case RTE_NAMEDTUPLESTORE:
724 : :
725 : : /*
726 : : * tuplestore cannot be shared, at least without more
727 : : * infrastructure to support that.
728 : : */
729 : 209 : return;
730 : :
1903 tgl@sss.pgh.pa.us 731 : 601 : case RTE_RESULT:
732 : : /* RESULT RTEs, in themselves, are no problem. */
733 : 601 : break;
734 : : }
735 : :
736 : : /*
737 : : * If there's anything in baserestrictinfo that's parallel-restricted, we
738 : : * give up on parallelizing access to this relation. We could consider
739 : : * instead postponing application of the restricted quals until we're
740 : : * above all the parallelism in the plan tree, but it's not clear that
741 : : * that would be a win in very many cases, and it might be tricky to make
742 : : * outer join clauses work correctly. It would likely break equivalence
743 : : * classes, too.
744 : : */
2795 745 [ + + ]: 157197 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
3077 rhaas@postgresql.org 746 : 10976 : return;
747 : :
748 : : /*
749 : : * Likewise, if the relation's outputs are not parallel-safe, give up.
750 : : * (Usually, they're just Vars, but sometimes they're not.)
751 : : */
2795 tgl@sss.pgh.pa.us 752 [ + + ]: 146221 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
2866 rhaas@postgresql.org 753 : 9 : return;
754 : :
755 : : /* We have a winner. */
3077 756 : 146212 : rel->consider_parallel = true;
757 : : }
758 : :
759 : : /*
760 : : * set_plain_rel_pathlist
761 : : * Build access paths for a plain relation (no subquery, no inheritance)
762 : : */
763 : : static void
4461 tgl@sss.pgh.pa.us 764 : 173241 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
765 : : {
766 : : Relids required_outer;
767 : :
768 : : /*
769 : : * We don't support pushing join clauses into the quals of a seqscan, but
770 : : * it could still have required parameterization due to LATERAL refs in
771 : : * its tlist.
772 : : */
4249 773 : 173241 : required_outer = rel->lateral_relids;
774 : :
775 : : /* Consider sequential scan */
3077 rhaas@postgresql.org 776 : 173241 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
777 : :
778 : : /* If appropriate, consider parallel sequential scan */
3007 779 [ + + + + ]: 173241 : if (rel->consider_parallel && required_outer == NULL)
2906 tgl@sss.pgh.pa.us 780 : 128810 : create_plain_partial_paths(root, rel);
781 : :
782 : : /* Consider index scans */
6929 783 : 173241 : create_index_paths(root, rel);
784 : :
785 : : /* Consider TID scans */
8554 786 : 173241 : create_tidscan_paths(root, rel);
787 : 173241 : }
788 : :
789 : : /*
790 : : * create_plain_partial_paths
791 : : * Build partial access paths for parallel scan of a plain relation
792 : : */
793 : : static void
2906 794 : 128810 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
795 : : {
796 : : int parallel_workers;
797 : :
2263 rhaas@postgresql.org 798 : 128810 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
799 : : max_parallel_workers_per_gather);
800 : :
801 : : /* If any limit was set to zero, the user doesn't want a parallel scan. */
2866 tgl@sss.pgh.pa.us 802 [ + + ]: 128810 : if (parallel_workers <= 0)
803 : 115775 : return;
804 : :
805 : : /* Add an unordered partial path based on a parallel sequential scan. */
rhaas@postgresql.org 806 : 13035 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
807 : : }
808 : :
809 : : /*
810 : : * set_tablesample_rel_size
811 : : * Set size estimates for a sampled relation
812 : : */
813 : : static void
3257 simon@2ndQuadrant.co 814 : 150 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
815 : : {
3186 tgl@sss.pgh.pa.us 816 : 150 : TableSampleClause *tsc = rte->tablesample;
817 : : TsmRoutine *tsm;
818 : : BlockNumber pages;
819 : : double tuples;
820 : :
821 : : /*
822 : : * Test any partial indexes of rel for applicability. We must do this
823 : : * first since partial unique indexes can affect size estimates.
824 : : */
2936 825 : 150 : check_index_predicates(root, rel);
826 : :
827 : : /*
828 : : * Call the sampling method's estimation function to estimate the number
829 : : * of pages it will read and the number of tuples it will return. (Note:
830 : : * we assume the function returns sane values.)
831 : : */
3186 832 : 150 : tsm = GetTsmRoutine(tsc->tsmhandler);
833 : 150 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
834 : : &pages, &tuples);
835 : :
836 : : /*
837 : : * For the moment, because we will only consider a SampleScan path for the
838 : : * rel, it's okay to just overwrite the pages and tuples estimates for the
839 : : * whole relation. If we ever consider multiple path types for sampled
840 : : * rels, we'll need more complication.
841 : : */
842 : 150 : rel->pages = pages;
843 : 150 : rel->tuples = tuples;
844 : :
845 : : /* Mark rel with estimated output rows, width, etc */
3257 simon@2ndQuadrant.co 846 : 150 : set_baserel_size_estimates(root, rel);
847 : 150 : }
848 : :
849 : : /*
850 : : * set_tablesample_rel_pathlist
851 : : * Build access paths for a sampled relation
852 : : */
853 : : static void
854 : 150 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
855 : : {
856 : : Relids required_outer;
857 : : Path *path;
858 : :
859 : : /*
860 : : * We don't support pushing join clauses into the quals of a samplescan,
861 : : * but it could still have required parameterization due to LATERAL refs
862 : : * in its tlist or TABLESAMPLE arguments.
863 : : */
864 : 150 : required_outer = rel->lateral_relids;
865 : :
866 : : /* Consider sampled scan */
867 : 150 : path = create_samplescan_path(root, rel, required_outer);
868 : :
869 : : /*
870 : : * If the sampling method does not support repeatable scans, we must avoid
871 : : * plans that would scan the rel multiple times. Ideally, we'd simply
872 : : * avoid putting the rel on the inside of a nestloop join; but adding such
873 : : * a consideration to the planner seems like a great deal of complication
874 : : * to support an uncommon usage of second-rate sampling methods. Instead,
875 : : * if there is a risk that the query might perform an unsafe join, just
876 : : * wrap the SampleScan in a Materialize node. We can check for joins by
877 : : * counting the membership of all_query_rels (note that this correctly
878 : : * counts inheritance trees as single rels). If we're inside a subquery,
879 : : * we can't easily check whether a join might occur in the outer query, so
880 : : * just assume one is possible.
881 : : *
882 : : * GetTsmRoutine is relatively expensive compared to the other tests here,
883 : : * so check repeatable_across_scans last, even though that's a bit odd.
884 : : */
3186 tgl@sss.pgh.pa.us 885 [ + + + + ]: 287 : if ((root->query_level > 1 ||
440 886 : 137 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
2489 887 [ + + ]: 46 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
888 : : {
3186 889 : 4 : path = (Path *) create_material_path(rel, path);
890 : : }
891 : :
892 : 150 : add_path(rel, path);
893 : :
894 : : /* For the moment, at least, there are no other paths to consider */
3257 simon@2ndQuadrant.co 895 : 150 : }
896 : :
897 : : /*
898 : : * set_foreign_size
899 : : * Set size estimates for a foreign table RTE
900 : : */
901 : : static void
4461 tgl@sss.pgh.pa.us 902 : 1179 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
903 : : {
904 : : /* Mark rel with estimated output rows, width, etc */
905 : 1179 : set_foreign_size_estimates(root, rel);
906 : :
907 : : /* Let FDW adjust the size estimates, if it can */
4419 908 : 1179 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
909 : :
910 : : /* ... but do not let it set the rows estimate to zero */
3185 911 : 1177 : rel->rows = clamp_row_est(rel->rows);
912 : :
913 : : /*
914 : : * Also, make sure rel->tuples is not insane relative to rel->rows.
915 : : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
916 : : * FDW doesn't do anything to replace that.
917 : : */
1381 918 [ + + ]: 1177 : rel->tuples = Max(rel->tuples, rel->rows);
4461 919 : 1177 : }
920 : :
921 : : /*
922 : : * set_foreign_pathlist
923 : : * Build access paths for a foreign table RTE
924 : : */
925 : : static void
926 : 1177 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
927 : : {
928 : : /* Call the FDW's GetForeignPaths function to generate path(s) */
4419 929 : 1177 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
4461 930 : 1177 : }
931 : :
932 : : /*
933 : : * set_append_rel_size
934 : : * Set size estimates for a simple "append relation"
935 : : *
936 : : * The passed-in rel and RTE represent the entire append relation. The
937 : : * relation's contents are computed by appending together the output of the
938 : : * individual member relations. Note that in the non-partitioned inheritance
939 : : * case, the first member relation is actually the same table as is mentioned
940 : : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
941 : : * a good thing because their outputs are not the same size.
942 : : */
943 : : static void
944 : 10128 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
945 : : Index rti, RangeTblEntry *rte)
946 : : {
8365 947 : 10128 : int parentRTindex = rti;
948 : : bool has_live_children;
949 : : double parent_rows;
950 : : double parent_size;
951 : : double *parent_attrsizes;
952 : : int nattrs;
953 : : ListCell *l;
954 : :
955 : : /* Guard against stack overflow due to overly deep inheritance tree. */
2404 rhaas@postgresql.org 956 : 10128 : check_stack_depth();
957 : :
2568 958 [ + + - + ]: 10128 : Assert(IS_SIMPLE_REL(rel));
959 : :
960 : : /*
961 : : * If this is a partitioned baserel, set the consider_partitionwise_join
962 : : * flag; currently, we only consider partitionwise joins with the baserel
963 : : * if its targetlist doesn't contain a whole-row Var.
964 : : */
2053 efujita@postgresql.o 965 [ + + ]: 10128 : if (enable_partitionwise_join &&
966 [ + + ]: 2035 : rel->reloptkind == RELOPT_BASEREL &&
967 [ + - ]: 1693 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
440 tgl@sss.pgh.pa.us 968 [ + + ]: 1693 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
2053 efujita@postgresql.o 969 : 1655 : rel->consider_partitionwise_join = true;
970 : :
971 : : /*
972 : : * Initialize to compute size estimates for whole append relation.
973 : : *
974 : : * We handle width estimates by weighting the widths of different child
975 : : * rels proportionally to their number of rows. This is sensible because
976 : : * the use of width estimates is mainly to compute the total relation
977 : : * "footprint" if we have to sort or hash it. To do this, we sum the
978 : : * total equivalent size (in "double" arithmetic) and then divide by the
979 : : * total rowcount estimate. This is done separately for the total rel
980 : : * width and each attribute.
981 : : *
982 : : * Note: if you consider changing this logic, beware that child rels could
983 : : * have zero rows and/or width, if they were excluded by constraints.
984 : : */
3185 tgl@sss.pgh.pa.us 985 : 10128 : has_live_children = false;
5770 986 : 10128 : parent_rows = 0;
987 : 10128 : parent_size = 0;
988 : 10128 : nattrs = rel->max_attr - rel->min_attr + 1;
989 : 10128 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
990 : :
6648 991 [ + + + + : 52258 : foreach(l, root->append_rel_list)
+ + ]
992 : : {
993 : 42131 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
994 : : int childRTindex;
995 : : RangeTblEntry *childRTE;
996 : : RelOptInfo *childrel;
997 : : List *childrinfos;
998 : : ListCell *parentvars;
999 : : ListCell *childvars;
1000 : : ListCell *lc;
1001 : :
1002 : : /* append_rel_list contains all append rels; ignore others */
1003 [ + + ]: 42131 : if (appinfo->parent_relid != parentRTindex)
1004 : 20941 : continue;
1005 : :
1006 : 21316 : childRTindex = appinfo->child_relid;
6203 1007 : 21316 : childRTE = root->simple_rte_array[childRTindex];
1008 : :
1009 : : /*
1010 : : * The child rel's RelOptInfo was already created during
1011 : : * add_other_rels_to_query.
1012 : : */
6417 1013 : 21316 : childrel = find_base_rel(root, childRTindex);
1014 [ - + ]: 21316 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1015 : :
1016 : : /* We may have already proven the child to be dummy. */
1842 1017 [ + + ]: 21316 : if (IS_DUMMY_REL(childrel))
1899 alvherre@alvh.no-ip. 1018 : 9 : continue;
1019 : :
1020 : : /*
1021 : : * We have to copy the parent's targetlist and quals to the child,
1022 : : * with appropriate substitution of variables. However, the
1023 : : * baserestrictinfo quals were already copied/substituted when the
1024 : : * child RelOptInfo was built. So we don't need any additional setup
1025 : : * before applying constraint exclusion.
1026 : : */
5857 tgl@sss.pgh.pa.us 1027 [ + + ]: 21307 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1028 : : {
1029 : : /*
1030 : : * This child need not be scanned, so we can omit it from the
1031 : : * appendrel.
1032 : : */
6168 1033 : 48 : set_dummy_rel_pathlist(childrel);
6203 1034 : 48 : continue;
1035 : : }
1036 : :
1037 : : /*
1038 : : * Constraint exclusion failed, so copy the parent's join quals and
1039 : : * targetlist to the child, with appropriate variable substitutions.
1040 : : *
1041 : : * We skip join quals that came from above outer joins that can null
1042 : : * this rel, since they would be of no value while generating paths
1043 : : * for the child. This saves some effort while processing the child
1044 : : * rel, and it also avoids an implementation restriction in
1045 : : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1046 : : */
399 1047 : 21259 : childrinfos = NIL;
1048 [ + + + + : 27577 : foreach(lc, rel->joininfo)
+ + ]
1049 : : {
1050 : 6318 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1051 : :
1052 [ + + ]: 6318 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1053 : 5157 : childrinfos = lappend(childrinfos,
1054 : 5157 : adjust_appendrel_attrs(root,
1055 : : (Node *) rinfo,
1056 : : 1, &appinfo));
1057 : : }
1058 : 21259 : childrel->joininfo = childrinfos;
1059 : :
1060 : : /*
1061 : : * Now for the child's targetlist.
1062 : : *
1063 : : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1064 : : * expressions, which otherwise would not occur in a rel's targetlist.
1065 : : * Code that might be looking at an appendrel child must cope with
1066 : : * such. (Normally, a rel's targetlist would only include Vars and
1067 : : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1068 : : * fields of childrel->reltarget; not clear if that would be useful.
1069 : : */
1910 efujita@postgresql.o 1070 : 42518 : childrel->reltarget->exprs = (List *)
1071 : 21259 : adjust_appendrel_attrs(root,
1072 : 21259 : (Node *) rel->reltarget->exprs,
1073 : : 1, &appinfo);
1074 : :
1075 : : /*
1076 : : * We have to make child entries in the EquivalenceClass data
1077 : : * structures as well. This is needed either if the parent
1078 : : * participates in some eclass joins (because we will want to consider
1079 : : * inner-indexscan joins on the individual children) or if the parent
1080 : : * has useful pathkeys (because we should try to build MergeAppend
1081 : : * paths that produce those sort orderings).
1082 : : */
1083 [ + + + + ]: 21259 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1084 : 10809 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1085 : 21259 : childrel->has_eclass_joins = rel->has_eclass_joins;
1086 : :
1087 : : /*
1088 : : * Note: we could compute appropriate attr_needed data for the child's
1089 : : * variables, by transforming the parent's attr_needed through the
1090 : : * translated_vars mapping. However, currently there's no need
1091 : : * because attr_needed is only examined for base relations not
1092 : : * otherrels. So we just leave the child's attr_needed empty.
1093 : : */
1094 : :
1095 : : /*
1096 : : * If we consider partitionwise joins with the parent rel, do the same
1097 : : * for partitioned child rels.
1098 : : *
1099 : : * Note: here we abuse the consider_partitionwise_join flag by setting
1100 : : * it for child rels that are not themselves partitioned. We do so to
1101 : : * tell try_partitionwise_join() that the child rel is sufficiently
1102 : : * valid to be used as a per-partition input, even if it later gets
1103 : : * proven to be dummy. (It's not usable until we've set up the
1104 : : * reltarget and EC entries, which we just did.)
1105 : : */
1106 [ + + ]: 21259 : if (rel->consider_partitionwise_join)
2053 1107 : 5452 : childrel->consider_partitionwise_join = true;
1108 : :
1109 : : /*
1110 : : * If parallelism is allowable for this query in general, see whether
1111 : : * it's allowable for this childrel in particular. But if we've
1112 : : * already decided the appendrel is not parallel-safe as a whole,
1113 : : * there's no point in considering parallelism for this child. For
1114 : : * consistency, do this before calling set_rel_size() for the child.
1115 : : */
2842 tgl@sss.pgh.pa.us 1116 [ + + + + ]: 21259 : if (root->glob->parallelModeOK && rel->consider_parallel)
1117 : 15207 : set_rel_consider_parallel(root, childrel, childRTE);
1118 : :
1119 : : /*
1120 : : * Compute the child's size.
1121 : : */
4461 1122 : 21259 : set_rel_size(root, childrel, childRTindex, childRTE);
1123 : :
1124 : : /*
1125 : : * It is possible that constraint exclusion detected a contradiction
1126 : : * within a child subquery, even though we didn't prove one above. If
1127 : : * so, we can skip this child.
1128 : : */
1129 [ + + ]: 21258 : if (IS_DUMMY_REL(childrel))
4586 1130 : 69 : continue;
1131 : :
1132 : : /* We have at least one live child. */
3185 1133 : 21189 : has_live_children = true;
1134 : :
1135 : : /*
1136 : : * If any live child is not parallel-safe, treat the whole appendrel
1137 : : * as not parallel-safe. In future we might be able to generate plans
1138 : : * in which some children are farmed out to workers while others are
1139 : : * not; but we don't have that today, so it's a waste to consider
1140 : : * partial paths anywhere in the appendrel unless it's all safe.
1141 : : * (Child rels visited before this one will be unmarked in
1142 : : * set_append_rel_pathlist().)
1143 : : */
2842 1144 [ + + ]: 21189 : if (!childrel->consider_parallel)
1145 : 6295 : rel->consider_parallel = false;
1146 : :
1147 : : /*
1148 : : * Accumulate size information from each live child.
1149 : : */
3185 1150 [ - + ]: 21189 : Assert(childrel->rows > 0);
1151 : :
1152 : 21189 : parent_rows += childrel->rows;
2953 1153 : 21189 : parent_size += childrel->reltarget->width * childrel->rows;
1154 : :
1155 : : /*
1156 : : * Accumulate per-column estimates too. We need not do anything for
1157 : : * PlaceHolderVars in the parent list. If child expression isn't a
1158 : : * Var, or we didn't record a width estimate for it, we have to fall
1159 : : * back on a datatype-based estimate.
1160 : : *
1161 : : * By construction, child's targetlist is 1-to-1 with parent's.
1162 : : */
1163 [ + + + + : 67533 : forboth(parentvars, rel->reltarget->exprs,
+ + + + +
+ + - +
+ ]
1164 : : childvars, childrel->reltarget->exprs)
1165 : : {
3185 1166 : 46344 : Var *parentvar = (Var *) lfirst(parentvars);
1167 : 46344 : Node *childvar = (Node *) lfirst(childvars);
1168 : :
1110 1169 [ + + + + ]: 46344 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1170 : : {
3185 1171 : 40462 : int pndx = parentvar->varattno - rel->min_attr;
1172 : 40462 : int32 child_width = 0;
1173 : :
1174 [ + + ]: 40462 : if (IsA(childvar, Var) &&
1175 [ + + ]: 39426 : ((Var *) childvar)->varno == childrel->relid)
1176 : : {
1177 : 39393 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1178 : :
1179 : 39393 : child_width = childrel->attr_widths[cndx];
1180 : : }
1181 [ + + ]: 40462 : if (child_width <= 0)
1182 : 1069 : child_width = get_typavgwidth(exprType(childvar),
1183 : : exprTypmod(childvar));
1184 [ - + ]: 40462 : Assert(child_width > 0);
1185 : 40462 : parent_attrsizes[pndx] += child_width * childrel->rows;
1186 : : }
1187 : : }
1188 : : }
1189 : :
1190 [ + + ]: 10127 : if (has_live_children)
1191 : : {
1192 : : /*
1193 : : * Save the finished size estimates.
1194 : : */
1195 : : int i;
1196 : :
1197 [ - + ]: 9980 : Assert(parent_rows > 0);
1198 : 9980 : rel->rows = parent_rows;
2953 1199 : 9980 : rel->reltarget->width = rint(parent_size / parent_rows);
5770 1200 [ + + ]: 101586 : for (i = 0; i < nattrs; i++)
1201 : 91606 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1202 : :
1203 : : /*
1204 : : * Set "raw tuples" count equal to "rows" for the appendrel; needed
1205 : : * because some places assume rel->tuples is valid for any baserel.
1206 : : */
3185 1207 : 9980 : rel->tuples = parent_rows;
1208 : :
1209 : : /*
1210 : : * Note that we leave rel->pages as zero; this is important to avoid
1211 : : * double-counting the appendrel tree in total_table_pages.
1212 : : */
1213 : : }
1214 : : else
1215 : : {
1216 : : /*
1217 : : * All children were excluded by constraints, so mark the whole
1218 : : * appendrel dummy. We must do this in this phase so that the rel's
1219 : : * dummy-ness is visible when we generate paths for other rels.
1220 : : */
1221 : 147 : set_dummy_rel_pathlist(rel);
1222 : : }
1223 : :
5770 1224 : 10127 : pfree(parent_attrsizes);
4461 1225 : 10127 : }
1226 : :
1227 : : /*
1228 : : * set_append_rel_pathlist
1229 : : * Build access paths for an "append relation"
1230 : : */
1231 : : static void
1232 : 9980 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1233 : : Index rti, RangeTblEntry *rte)
1234 : : {
1235 : 9980 : int parentRTindex = rti;
1236 : 9980 : List *live_childrels = NIL;
1237 : : ListCell *l;
1238 : :
1239 : : /*
1240 : : * Generate access paths for each member relation, and remember the
1241 : : * non-dummy children.
1242 : : */
1243 [ + - + + : 51903 : foreach(l, root->append_rel_list)
+ + ]
1244 : : {
1245 : 41923 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1246 : : int childRTindex;
1247 : : RangeTblEntry *childRTE;
1248 : : RelOptInfo *childrel;
1249 : :
1250 : : /* append_rel_list contains all append rels; ignore others */
1251 [ + + ]: 41923 : if (appinfo->parent_relid != parentRTindex)
1252 : 20641 : continue;
1253 : :
1254 : : /* Re-locate the child RTE and RelOptInfo */
1255 : 21282 : childRTindex = appinfo->child_relid;
1256 : 21282 : childRTE = root->simple_rte_array[childRTindex];
1257 : 21282 : childrel = root->simple_rel_array[childRTindex];
1258 : :
1259 : : /*
1260 : : * If set_append_rel_size() decided the parent appendrel was
1261 : : * parallel-unsafe at some point after visiting this child rel, we
1262 : : * need to propagate the unsafety marking down to the child, so that
1263 : : * we don't generate useless partial paths for it.
1264 : : */
2842 1265 [ + + ]: 21282 : if (!rel->consider_parallel)
1266 : 6348 : childrel->consider_parallel = false;
1267 : :
1268 : : /*
1269 : : * Compute the child's access paths.
1270 : : */
4461 1271 : 21282 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1272 : :
1273 : : /*
1274 : : * If child is dummy, ignore it.
1275 : : */
1276 [ + + ]: 21282 : if (IS_DUMMY_REL(childrel))
1277 : 93 : continue;
1278 : :
1279 : : /*
1280 : : * Child is live, so add it to the live_childrels list for use below.
1281 : : */
4264 1282 : 21189 : live_childrels = lappend(live_childrels, childrel);
1283 : : }
1284 : :
1285 : : /* Add paths to the append relation. */
2588 rhaas@postgresql.org 1286 : 9980 : add_paths_to_append_rel(root, rel, live_childrels);
1287 : 9980 : }
1288 : :
1289 : :
1290 : : /*
1291 : : * add_paths_to_append_rel
1292 : : * Generate paths for the given append relation given the set of non-dummy
1293 : : * child rels.
1294 : : *
1295 : : * The function collects all parameterizations and orderings supported by the
1296 : : * non-dummy children. For every such parameterization or ordering, it creates
1297 : : * an append path collecting one path from each non-dummy child with given
1298 : : * parameterization or ordering. Similarly it collects partial paths from
1299 : : * non-dummy children to create partial append paths.
1300 : : */
1301 : : void
1302 : 17261 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1303 : : List *live_childrels)
1304 : : {
1305 : 17261 : List *subpaths = NIL;
1306 : 17261 : bool subpaths_valid = true;
192 drowley@postgresql.o 1307 :GNC 17261 : List *startup_subpaths = NIL;
1308 : 17261 : bool startup_subpaths_valid = true;
2588 rhaas@postgresql.org 1309 :CBC 17261 : List *partial_subpaths = NIL;
2322 1310 : 17261 : List *pa_partial_subpaths = NIL;
1311 : 17261 : List *pa_nonpartial_subpaths = NIL;
2588 1312 : 17261 : bool partial_subpaths_valid = true;
1313 : : bool pa_subpaths_valid;
1314 : 17261 : List *all_child_pathkeys = NIL;
1315 : 17261 : List *all_child_outers = NIL;
1316 : : ListCell *l;
2322 1317 : 17261 : double partial_rows = -1;
1318 : :
1319 : : /* If appropriate, consider parallel append */
2125 akapila@postgresql.o 1320 [ + + + + ]: 17261 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1321 : :
1322 : : /*
1323 : : * For every non-dummy child, remember the cheapest path. Also, identify
1324 : : * all pathkeys (orderings) and parameterizations (required_outer sets)
1325 : : * available for the non-dummy member relations.
1326 : : */
2588 rhaas@postgresql.org 1327 [ + - + + : 52809 : foreach(l, live_childrels)
+ + ]
1328 : : {
1329 : 35548 : RelOptInfo *childrel = lfirst(l);
1330 : : ListCell *lcp;
2322 1331 : 35548 : Path *cheapest_partial_path = NULL;
1332 : :
1333 : : /*
1334 : : * If child has an unparameterized cheapest-total path, add that to
1335 : : * the unparameterized Append path we are constructing for the parent.
1336 : : * If not, there's no workable unparameterized path.
1337 : : *
1338 : : * With partitionwise aggregates, the child rel's pathlist may be
1339 : : * empty, so don't assume that a path exists here.
1340 : : */
2215 1341 [ + - ]: 35548 : if (childrel->pathlist != NIL &&
1342 [ + + ]: 35548 : childrel->cheapest_total_path->param_info == NULL)
2322 1343 : 35182 : accumulate_append_subpath(childrel->cheapest_total_path,
1344 : : &subpaths, NULL);
1345 : : else
4264 tgl@sss.pgh.pa.us 1346 : 366 : subpaths_valid = false;
1347 : :
1348 : : /*
1349 : : * When the planner is considering cheap startup plans, we'll also
1350 : : * collect all the cheapest_startup_paths (if set) and build an
1351 : : * AppendPath containing those as subpaths.
1352 : : */
187 drowley@postgresql.o 1353 [ + + + + ]:GNC 35548 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1354 : : {
1355 : : /* cheapest_startup_path must not be a parameterized path. */
1356 [ - + ]: 643 : Assert(childrel->cheapest_startup_path->param_info == NULL);
192 1357 : 643 : accumulate_append_subpath(childrel->cheapest_startup_path,
1358 : : &startup_subpaths,
1359 : : NULL);
1360 : : }
1361 : : else
1362 : 34905 : startup_subpaths_valid = false;
1363 : :
1364 : :
1365 : : /* Same idea, but for a partial plan. */
3007 rhaas@postgresql.org 1366 [ + + ]:CBC 35548 : if (childrel->partial_pathlist != NIL)
1367 : : {
2322 1368 : 23682 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1369 : 23682 : accumulate_append_subpath(cheapest_partial_path,
1370 : : &partial_subpaths, NULL);
1371 : : }
1372 : : else
3007 1373 : 11866 : partial_subpaths_valid = false;
1374 : :
1375 : : /*
1376 : : * Same idea, but for a parallel append mixing partial and non-partial
1377 : : * paths.
1378 : : */
2322 1379 [ + + ]: 35548 : if (pa_subpaths_valid)
1380 : : {
1381 : 25036 : Path *nppath = NULL;
1382 : :
1383 : : nppath =
1384 : 25036 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1385 : :
1386 [ + + + + ]: 25036 : if (cheapest_partial_path == NULL && nppath == NULL)
1387 : : {
1388 : : /* Neither a partial nor a parallel-safe path? Forget it. */
1389 : 225 : pa_subpaths_valid = false;
1390 : : }
1391 [ + + + + ]: 24811 : else if (nppath == NULL ||
1392 : 23457 : (cheapest_partial_path != NULL &&
1393 [ + + ]: 23457 : cheapest_partial_path->total_cost < nppath->total_cost))
1394 : : {
1395 : : /* Partial path is cheaper or the only option. */
1396 [ - + ]: 23339 : Assert(cheapest_partial_path != NULL);
1397 : 23339 : accumulate_append_subpath(cheapest_partial_path,
1398 : : &pa_partial_subpaths,
1399 : : &pa_nonpartial_subpaths);
1400 : : }
1401 : : else
1402 : : {
1403 : : /*
1404 : : * Either we've got only a non-partial path, or we think that
1405 : : * a single backend can execute the best non-partial path
1406 : : * faster than all the parallel backends working together can
1407 : : * execute the best partial path.
1408 : : *
1409 : : * It might make sense to be more aggressive here. Even if
1410 : : * the best non-partial path is more expensive than the best
1411 : : * partial path, it could still be better to choose the
1412 : : * non-partial path if there are several such paths that can
1413 : : * be given to different workers. For now, we don't try to
1414 : : * figure that out.
1415 : : */
1416 : 1472 : accumulate_append_subpath(nppath,
1417 : : &pa_nonpartial_subpaths,
1418 : : NULL);
1419 : : }
1420 : : }
1421 : :
1422 : : /*
1423 : : * Collect lists of all the available path orderings and
1424 : : * parameterizations for all the children. We use these as a
1425 : : * heuristic to indicate which sort orderings and parameterizations we
1426 : : * should build Append and MergeAppend paths for.
1427 : : */
4461 tgl@sss.pgh.pa.us 1428 [ + - + + : 81052 : foreach(lcp, childrel->pathlist)
+ + ]
1429 : : {
1430 : 45504 : Path *childpath = (Path *) lfirst(lcp);
1431 : 45504 : List *childkeys = childpath->pathkeys;
4378 1432 [ + + ]: 45504 : Relids childouter = PATH_REQ_OUTER(childpath);
1433 : :
1434 : : /* Unsorted paths don't contribute to pathkey list */
4461 1435 [ + + ]: 45504 : if (childkeys != NIL)
1436 : : {
1437 : : ListCell *lpk;
1438 : 10055 : bool found = false;
1439 : :
1440 : : /* Have we already seen this ordering? */
1441 [ + + + + : 10149 : foreach(lpk, all_child_pathkeys)
+ + ]
1442 : : {
1443 : 7265 : List *existing_pathkeys = (List *) lfirst(lpk);
1444 : :
1445 [ + + ]: 7265 : if (compare_pathkeys(existing_pathkeys,
1446 : : childkeys) == PATHKEYS_EQUAL)
1447 : : {
1448 : 7171 : found = true;
1449 : 7171 : break;
1450 : : }
1451 : : }
1452 [ + + ]: 10055 : if (!found)
1453 : : {
1454 : : /* No, so add it to all_child_pathkeys */
1455 : 2884 : all_child_pathkeys = lappend(all_child_pathkeys,
1456 : : childkeys);
1457 : : }
1458 : : }
1459 : :
1460 : : /* Unparameterized paths don't contribute to param-set list */
1461 [ + + ]: 45504 : if (childouter)
1462 : : {
1463 : : ListCell *lco;
1464 : 2940 : bool found = false;
1465 : :
1466 : : /* Have we already seen this param set? */
1467 [ + + + + : 3246 : foreach(lco, all_child_outers)
+ + ]
1468 : : {
4326 bruce@momjian.us 1469 : 2133 : Relids existing_outers = (Relids) lfirst(lco);
1470 : :
4461 tgl@sss.pgh.pa.us 1471 [ + + ]: 2133 : if (bms_equal(existing_outers, childouter))
1472 : : {
1473 : 1827 : found = true;
1474 : 1827 : break;
1475 : : }
1476 : : }
1477 [ + + ]: 2940 : if (!found)
1478 : : {
1479 : : /* No, so add it to all_child_outers */
1480 : 1113 : all_child_outers = lappend(all_child_outers,
1481 : : childouter);
1482 : : }
1483 : : }
1484 : : }
1485 : : }
1486 : :
1487 : : /*
1488 : : * If we found unparameterized paths for all children, build an unordered,
1489 : : * unparameterized Append path for the rel. (Note: this is correct even
1490 : : * if we have zero or one live subpath due to constraint exclusion.)
1491 : : */
4264 1492 [ + + ]: 17261 : if (subpaths_valid)
2199 alvherre@alvh.no-ip. 1493 : 17105 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1494 : : NIL, NULL, 0, false,
1495 : : -1));
1496 : :
1497 : : /* build an AppendPath for the cheap startup paths, if valid */
192 drowley@postgresql.o 1498 [ + + ]:GNC 17261 : if (startup_subpaths_valid)
1499 : 272 : add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1500 : : NIL, NIL, NULL, 0, false, -1));
1501 : :
1502 : : /*
1503 : : * Consider an append of unordered, unparameterized partial paths. Make
1504 : : * it parallel-aware if possible.
1505 : : */
1865 tgl@sss.pgh.pa.us 1506 [ + + + - ]:CBC 17261 : if (partial_subpaths_valid && partial_subpaths != NIL)
1507 : : {
1508 : : AppendPath *appendpath;
1509 : : ListCell *lc;
2866 rhaas@postgresql.org 1510 : 10648 : int parallel_workers = 0;
1511 : :
1512 : : /* Find the highest number of workers requested for any subpath. */
3007 1513 [ + - + + : 36366 : foreach(lc, partial_subpaths)
+ + ]
1514 : : {
1515 : 25718 : Path *path = lfirst(lc);
1516 : :
2866 1517 : 25718 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1518 : : }
1519 [ - + ]: 10648 : Assert(parallel_workers > 0);
1520 : :
1521 : : /*
1522 : : * If the use of parallel append is permitted, always request at least
1523 : : * log2(# of children) workers. We assume it can be useful to have
1524 : : * extra workers in this case because they will be spread out across
1525 : : * the children. The precise formula is just a guess, but we don't
1526 : : * want to end up with a radically different answer for a table with N
1527 : : * partitions vs. an unpartitioned table with the same data, so the
1528 : : * use of some kind of log-scaling here seems to make some sense.
1529 : : */
2322 1530 [ + + ]: 10648 : if (enable_parallel_append)
1531 : : {
1532 [ + + ]: 10624 : parallel_workers = Max(parallel_workers,
1533 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1534 : 10624 : parallel_workers = Min(parallel_workers,
1535 : : max_parallel_workers_per_gather);
1536 : : }
1537 [ - + ]: 10648 : Assert(parallel_workers > 0);
1538 : :
1539 : : /* Generate a partial append path. */
2199 alvherre@alvh.no-ip. 1540 : 10648 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1541 : : NIL, NULL, parallel_workers,
1542 : : enable_parallel_append,
1543 : : -1);
1544 : :
1545 : : /*
1546 : : * Make sure any subsequent partial paths use the same row count
1547 : : * estimate.
1548 : : */
2322 rhaas@postgresql.org 1549 : 10648 : partial_rows = appendpath->path.rows;
1550 : :
1551 : : /* Add the path. */
1552 : 10648 : add_partial_path(rel, (Path *) appendpath);
1553 : : }
1554 : :
1555 : : /*
1556 : : * Consider a parallel-aware append using a mix of partial and non-partial
1557 : : * paths. (This only makes sense if there's at least one child which has
1558 : : * a non-partial path that is substantially cheaper than any partial path;
1559 : : * otherwise, we should use the append path added in the previous step.)
1560 : : */
1561 [ + + + + ]: 17261 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1562 : : {
1563 : : AppendPath *appendpath;
1564 : : ListCell *lc;
1565 : 767 : int parallel_workers = 0;
1566 : :
1567 : : /*
1568 : : * Find the highest number of workers requested for any partial
1569 : : * subpath.
1570 : : */
1571 [ + + + + : 1209 : foreach(lc, pa_partial_subpaths)
+ + ]
1572 : : {
1573 : 442 : Path *path = lfirst(lc);
1574 : :
1575 : 442 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1576 : : }
1577 : :
1578 : : /*
1579 : : * Same formula here as above. It's even more important in this
1580 : : * instance because the non-partial paths won't contribute anything to
1581 : : * the planned number of parallel workers.
1582 : : */
1583 [ + - ]: 767 : parallel_workers = Max(parallel_workers,
1584 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1585 : 767 : parallel_workers = Min(parallel_workers,
1586 : : max_parallel_workers_per_gather);
1587 [ - + ]: 767 : Assert(parallel_workers > 0);
1588 : :
2199 alvherre@alvh.no-ip. 1589 : 767 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1590 : : pa_partial_subpaths,
1591 : : NIL, NULL, parallel_workers, true,
1592 : : partial_rows);
3007 rhaas@postgresql.org 1593 : 767 : add_partial_path(rel, (Path *) appendpath);
1594 : : }
1595 : :
1596 : : /*
1597 : : * Also build unparameterized ordered append paths based on the collected
1598 : : * list of child pathkeys.
1599 : : */
4264 tgl@sss.pgh.pa.us 1600 [ + + ]: 17261 : if (subpaths_valid)
1836 1601 : 17105 : generate_orderedappend_paths(root, rel, live_childrels,
1602 : : all_child_pathkeys);
1603 : :
1604 : : /*
1605 : : * Build Append paths for each parameterization seen among the child rels.
1606 : : * (This may look pretty expensive, but in most cases of practical
1607 : : * interest, the child rels will expose mostly the same parameterizations,
1608 : : * so that not that many cases actually get considered here.)
1609 : : *
1610 : : * The Append node itself cannot enforce quals, so all qual checking must
1611 : : * be done in the child paths. This means that to have a parameterized
1612 : : * Append path, we must have the exact same parameterization for each
1613 : : * child path; otherwise some children might be failing to check the
1614 : : * moved-down quals. To make them match up, we can try to increase the
1615 : : * parameterization of lesser-parameterized paths.
1616 : : */
4461 1617 [ + + + + : 18374 : foreach(l, all_child_outers)
+ + ]
1618 : : {
4326 bruce@momjian.us 1619 : 1113 : Relids required_outer = (Relids) lfirst(l);
1620 : : ListCell *lcr;
1621 : :
1622 : : /* Select the child paths for an Append with this parameterization */
4461 tgl@sss.pgh.pa.us 1623 : 1113 : subpaths = NIL;
4264 1624 : 1113 : subpaths_valid = true;
4461 1625 [ + - + + : 4095 : foreach(lcr, live_childrels)
+ + ]
1626 : : {
1627 : 2988 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1628 : : Path *subpath;
1629 : :
2215 rhaas@postgresql.org 1630 [ - + ]: 2988 : if (childrel->pathlist == NIL)
1631 : : {
1632 : : /* failed to make a suitable path for this child */
2215 rhaas@postgresql.org 1633 :UBC 0 : subpaths_valid = false;
1634 : 0 : break;
1635 : : }
1636 : :
3934 tgl@sss.pgh.pa.us 1637 :CBC 2988 : subpath = get_cheapest_parameterized_child_path(root,
1638 : : childrel,
1639 : : required_outer);
1640 [ + + ]: 2988 : if (subpath == NULL)
1641 : : {
1642 : : /* failed to make a suitable path for this child */
1643 : 6 : subpaths_valid = false;
1644 : 6 : break;
1645 : : }
1168 1646 : 2982 : accumulate_append_subpath(subpath, &subpaths, NULL);
1647 : : }
1648 : :
4264 1649 [ + + ]: 1113 : if (subpaths_valid)
4378 1650 : 1107 : add_path(rel, (Path *)
2199 alvherre@alvh.no-ip. 1651 : 1107 : create_append_path(root, rel, subpaths, NIL,
1652 : : NIL, required_outer, 0, false,
1653 : : -1));
1654 : : }
1655 : :
1656 : : /*
1657 : : * When there is only a single child relation, the Append path can inherit
1658 : : * any ordering available for the child rel's path, so that it's useful to
1659 : : * consider ordered partial paths. Above we only considered the cheapest
1660 : : * partial path for each child, but let's also make paths using any
1661 : : * partial paths that have pathkeys.
1662 : : */
1847 tgl@sss.pgh.pa.us 1663 [ + + ]: 17261 : if (list_length(live_childrels) == 1)
1664 : : {
1665 : 6950 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1666 : :
1667 : : /* skip the cheapest partial path, since we already used that above */
1259 drowley@postgresql.o 1668 [ + + + + : 7052 : for_each_from(l, childrel->partial_pathlist, 1)
+ + ]
1669 : : {
1847 tgl@sss.pgh.pa.us 1670 : 102 : Path *path = (Path *) lfirst(l);
1671 : : AppendPath *appendpath;
1672 : :
1673 : : /* skip paths with no pathkeys. */
1259 drowley@postgresql.o 1674 [ - + ]: 102 : if (path->pathkeys == NIL)
1847 tgl@sss.pgh.pa.us 1675 :UBC 0 : continue;
1676 : :
1847 tgl@sss.pgh.pa.us 1677 :CBC 102 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1678 : : NIL, NULL,
1679 : : path->parallel_workers, true,
1680 : : partial_rows);
1681 : 102 : add_partial_path(rel, (Path *) appendpath);
1682 : : }
1683 : : }
4461 1684 : 17261 : }
1685 : :
1686 : : /*
1687 : : * generate_orderedappend_paths
1688 : : * Generate ordered append paths for an append relation
1689 : : *
1690 : : * Usually we generate MergeAppend paths here, but there are some special
1691 : : * cases where we can generate simple Append paths, because the subpaths
1692 : : * can provide tuples in the required order already.
1693 : : *
1694 : : * We generate a path for each ordering (pathkey list) appearing in
1695 : : * all_child_pathkeys.
1696 : : *
1697 : : * We consider both cheapest-startup and cheapest-total cases, ie, for each
1698 : : * interesting ordering, collect all the cheapest startup subpaths and all the
1699 : : * cheapest total paths, and build a suitable path for each case.
1700 : : *
1701 : : * We don't currently generate any parameterized ordered paths here. While
1702 : : * it would not take much more code here to do so, it's very unclear that it
1703 : : * is worth the planning cycles to investigate such paths: there's little
1704 : : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1705 : : * that the current coding of add_path would reject such paths out of hand,
1706 : : * because add_path gives no credit for sort ordering of parameterized paths,
1707 : : * and a parameterized MergeAppend is going to be more expensive than the
1708 : : * corresponding parameterized Append path. If we ever try harder to support
1709 : : * parameterized mergejoin plans, it might be worth adding support for
1710 : : * parameterized paths here to feed such joins. (See notes in
1711 : : * optimizer/README for why that might not ever happen, though.)
1712 : : */
1713 : : static void
1836 1714 : 17105 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1715 : : List *live_childrels,
1716 : : List *all_child_pathkeys)
1717 : : {
1718 : : ListCell *lcp;
1719 : 17105 : List *partition_pathkeys = NIL;
1720 : 17105 : List *partition_pathkeys_desc = NIL;
1721 : 17105 : bool partition_pathkeys_partial = true;
1722 : 17105 : bool partition_pathkeys_desc_partial = true;
1723 : :
1724 : : /*
1725 : : * Some partitioned table setups may allow us to use an Append node
1726 : : * instead of a MergeAppend. This is possible in cases such as RANGE
1727 : : * partitioned tables where it's guaranteed that an earlier partition must
1728 : : * contain rows which come earlier in the sort order. To detect whether
1729 : : * this is relevant, build pathkey descriptions of the partition ordering,
1730 : : * for both forward and reverse scans.
1731 : : */
1732 [ + + + + : 30459 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
+ + + + ]
985 drowley@postgresql.o 1733 : 13354 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1734 : : {
1836 tgl@sss.pgh.pa.us 1735 : 10979 : partition_pathkeys = build_partition_pathkeys(root, rel,
1736 : : ForwardScanDirection,
1737 : : &partition_pathkeys_partial);
1738 : :
1739 : 10979 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1740 : : BackwardScanDirection,
1741 : : &partition_pathkeys_desc_partial);
1742 : :
1743 : : /*
1744 : : * You might think we should truncate_useless_pathkeys here, but
1745 : : * allowing partition keys which are a subset of the query's pathkeys
1746 : : * can often be useful. For example, consider a table partitioned by
1747 : : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1748 : : * paths that can produce the a, b, c ordering (perhaps via indexes on
1749 : : * (a, b, c)) then it works to consider the appendrel output as
1750 : : * ordered by a, b, c.
1751 : : */
1752 : : }
1753 : :
1754 : : /* Now consider each interesting sort ordering */
4461 1755 [ + + + + : 19959 : foreach(lcp, all_child_pathkeys)
+ + ]
1756 : : {
1757 : 2854 : List *pathkeys = (List *) lfirst(lcp);
4753 bruce@momjian.us 1758 : 2854 : List *startup_subpaths = NIL;
1759 : 2854 : List *total_subpaths = NIL;
823 tomas.vondra@postgre 1760 : 2854 : List *fractional_subpaths = NIL;
4753 bruce@momjian.us 1761 : 2854 : bool startup_neq_total = false;
1762 : : bool match_partition_order;
1763 : : bool match_partition_order_desc;
1764 : : int end_index;
1765 : : int first_index;
1766 : : int direction;
1767 : :
1768 : : /*
1769 : : * Determine if this sort ordering matches any partition pathkeys we
1770 : : * have, for both ascending and descending partition order. If the
1771 : : * partition pathkeys happen to be contained in pathkeys then it still
1772 : : * works, as described above, providing that the partition pathkeys
1773 : : * are complete and not just a prefix of the partition keys. (In such
1774 : : * cases we'll be relying on the child paths to have sorted the
1775 : : * lower-order columns of the required pathkeys.)
1776 : : */
1836 tgl@sss.pgh.pa.us 1777 : 2854 : match_partition_order =
1778 [ + + ]: 4649 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1779 [ + + + + ]: 1863 : (!partition_pathkeys_partial &&
1780 : 68 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1781 : :
1782 [ + + + + ]: 6363 : match_partition_order_desc = !match_partition_order &&
1783 : 1762 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1784 [ + + + + ]: 1767 : (!partition_pathkeys_desc_partial &&
1785 : 20 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1786 : :
1787 : : /*
1788 : : * When the required pathkeys match the reverse of the partition
1789 : : * order, we must build the list of paths in reverse starting with the
1790 : : * last matching partition first. We can get away without making any
1791 : : * special cases for this in the loop below by just looping backward
1792 : : * over the child relations in this case.
1793 : : */
419 drowley@postgresql.o 1794 [ + + ]: 2854 : if (match_partition_order_desc)
1795 : : {
1796 : : /* loop backward */
1797 : 21 : first_index = list_length(live_childrels) - 1;
1798 : 21 : end_index = -1;
1799 : 21 : direction = -1;
1800 : :
1801 : : /*
1802 : : * Set this to true to save us having to check for
1803 : : * match_partition_order_desc in the loop below.
1804 : : */
1805 : 21 : match_partition_order = true;
1806 : : }
1807 : : else
1808 : : {
1809 : : /* for all other case, loop forward */
1810 : 2833 : first_index = 0;
1811 : 2833 : end_index = list_length(live_childrels);
1812 : 2833 : direction = 1;
1813 : : }
1814 : :
1815 : : /* Select the child paths for this ordering... */
1816 [ + + ]: 10524 : for (int i = first_index; i != end_index; i += direction)
1817 : : {
1818 : 7670 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1819 : : Path *cheapest_startup,
1820 : : *cheapest_total,
823 tomas.vondra@postgre 1821 : 7670 : *cheapest_fractional = NULL;
1822 : :
1823 : : /* Locate the right paths, if they are available. */
1824 : : cheapest_startup =
4931 tgl@sss.pgh.pa.us 1825 : 7670 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1826 : : pathkeys,
1827 : : NULL,
1828 : : STARTUP_COST,
1829 : : false);
1830 : : cheapest_total =
1831 : 7670 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1832 : : pathkeys,
1833 : : NULL,
1834 : : TOTAL_COST,
1835 : : false);
1836 : :
1837 : : /*
1838 : : * If we can't find any paths with the right order just use the
1839 : : * cheapest-total path; we'll have to sort it later.
1840 : : */
4461 1841 [ + + - + ]: 7670 : if (cheapest_startup == NULL || cheapest_total == NULL)
1842 : : {
4378 1843 : 134 : cheapest_startup = cheapest_total =
1844 : : childrel->cheapest_total_path;
1845 : : /* Assert we do have an unparameterized path for this child */
4264 1846 [ - + ]: 134 : Assert(cheapest_total->param_info == NULL);
1847 : : }
1848 : :
1849 : : /*
1850 : : * When building a fractional path, determine a cheapest
1851 : : * fractional path for each child relation too. Looking at startup
1852 : : * and total costs is not enough, because the cheapest fractional
1853 : : * path may be dominated by two separate paths (one for startup,
1854 : : * one for total).
1855 : : *
1856 : : * When needed (building fractional path), determine the cheapest
1857 : : * fractional path too.
1858 : : */
823 tomas.vondra@postgre 1859 [ + + ]: 7670 : if (root->tuple_fraction > 0)
1860 : : {
703 tgl@sss.pgh.pa.us 1861 : 334 : double path_fraction = (1.0 / root->tuple_fraction);
1862 : :
1863 : : cheapest_fractional =
823 tomas.vondra@postgre 1864 : 334 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1865 : : pathkeys,
1866 : : NULL,
1867 : : path_fraction);
1868 : :
1869 : : /*
1870 : : * If we found no path with matching pathkeys, use the
1871 : : * cheapest total path instead.
1872 : : *
1873 : : * XXX We might consider partially sorted paths too (with an
1874 : : * incremental sort on top). But we'd have to build all the
1875 : : * incremental paths, do the costing etc.
1876 : : */
1877 [ + + ]: 334 : if (!cheapest_fractional)
1878 : 22 : cheapest_fractional = cheapest_total;
1879 : : }
1880 : :
1881 : : /*
1882 : : * Notice whether we actually have different paths for the
1883 : : * "cheapest" and "total" cases; frequently there will be no point
1884 : : * in two create_merge_append_path() calls.
1885 : : */
4931 tgl@sss.pgh.pa.us 1886 [ + + ]: 7670 : if (cheapest_startup != cheapest_total)
1887 : 24 : startup_neq_total = true;
1888 : :
1889 : : /*
1890 : : * Collect the appropriate child paths. The required logic varies
1891 : : * for the Append and MergeAppend cases.
1892 : : */
1836 1893 [ + + ]: 7670 : if (match_partition_order)
1894 : : {
1895 : : /*
1896 : : * We're going to make a plain Append path. We don't need
1897 : : * most of what accumulate_append_subpath would do, but we do
1898 : : * want to cut out child Appends or MergeAppends if they have
1899 : : * just a single subpath (and hence aren't doing anything
1900 : : * useful).
1901 : : */
1902 : 2989 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1903 : 2989 : cheapest_total = get_singleton_append_subpath(cheapest_total);
1904 : :
1905 : 2989 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1906 : 2989 : total_subpaths = lappend(total_subpaths, cheapest_total);
1907 : :
823 tomas.vondra@postgre 1908 [ + + ]: 2989 : if (cheapest_fractional)
1909 : : {
1910 : 60 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1911 : 60 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1912 : : }
1913 : : }
1914 : : else
1915 : : {
1916 : : /*
1917 : : * Otherwise, rely on accumulate_append_subpath to collect the
1918 : : * child paths for the MergeAppend.
1919 : : */
1836 tgl@sss.pgh.pa.us 1920 : 4681 : accumulate_append_subpath(cheapest_startup,
1921 : : &startup_subpaths, NULL);
1922 : 4681 : accumulate_append_subpath(cheapest_total,
1923 : : &total_subpaths, NULL);
1924 : :
823 tomas.vondra@postgre 1925 [ + + ]: 4681 : if (cheapest_fractional)
1926 : 274 : accumulate_append_subpath(cheapest_fractional,
1927 : : &fractional_subpaths, NULL);
1928 : : }
1929 : : }
1930 : :
1931 : : /* ... and build the Append or MergeAppend paths */
419 drowley@postgresql.o 1932 [ + + ]: 2854 : if (match_partition_order)
1933 : : {
1934 : : /* We only need Append */
1836 tgl@sss.pgh.pa.us 1935 : 1113 : add_path(rel, (Path *) create_append_path(root,
1936 : : rel,
1937 : : startup_subpaths,
1938 : : NIL,
1939 : : pathkeys,
1940 : : NULL,
1941 : : 0,
1942 : : false,
1943 : : -1));
1944 [ - + ]: 1113 : if (startup_neq_total)
1836 tgl@sss.pgh.pa.us 1945 :UBC 0 : add_path(rel, (Path *) create_append_path(root,
1946 : : rel,
1947 : : total_subpaths,
1948 : : NIL,
1949 : : pathkeys,
1950 : : NULL,
1951 : : 0,
1952 : : false,
1953 : : -1));
1954 : :
823 tomas.vondra@postgre 1955 [ + + ]:CBC 1113 : if (fractional_subpaths)
1956 : 30 : add_path(rel, (Path *) create_append_path(root,
1957 : : rel,
1958 : : fractional_subpaths,
1959 : : NIL,
1960 : : pathkeys,
1961 : : NULL,
1962 : : 0,
1963 : : false,
1964 : : -1));
1965 : : }
1966 : : else
1967 : : {
1968 : : /* We need MergeAppend */
4931 tgl@sss.pgh.pa.us 1969 : 1741 : add_path(rel, (Path *) create_merge_append_path(root,
1970 : : rel,
1971 : : startup_subpaths,
1972 : : pathkeys,
1973 : : NULL));
1836 1974 [ + + ]: 1741 : if (startup_neq_total)
1975 : 12 : add_path(rel, (Path *) create_merge_append_path(root,
1976 : : rel,
1977 : : total_subpaths,
1978 : : pathkeys,
1979 : : NULL));
1980 : :
823 tomas.vondra@postgre 1981 [ + + ]: 1741 : if (fractional_subpaths)
1982 : 98 : add_path(rel, (Path *) create_merge_append_path(root,
1983 : : rel,
1984 : : fractional_subpaths,
1985 : : pathkeys,
1986 : : NULL));
1987 : : }
1988 : : }
10141 scrappy@hub.org 1989 : 17105 : }
1990 : :
1991 : : /*
1992 : : * get_cheapest_parameterized_child_path
1993 : : * Get cheapest path for this relation that has exactly the requested
1994 : : * parameterization.
1995 : : *
1996 : : * Returns NULL if unable to create such a path.
1997 : : */
1998 : : static Path *
3934 tgl@sss.pgh.pa.us 1999 : 2988 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2000 : : Relids required_outer)
2001 : : {
2002 : : Path *cheapest;
2003 : : ListCell *lc;
2004 : :
2005 : : /*
2006 : : * Look up the cheapest existing path with no more than the needed
2007 : : * parameterization. If it has exactly the needed parameterization, we're
2008 : : * done.
2009 : : */
2010 : 2988 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2011 : : NIL,
2012 : : required_outer,
2013 : : TOTAL_COST,
2014 : : false);
2015 [ - + ]: 2988 : Assert(cheapest != NULL);
2016 [ + + + + ]: 2988 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2017 : 2842 : return cheapest;
2018 : :
2019 : : /*
2020 : : * Otherwise, we can "reparameterize" an existing path to match the given
2021 : : * parameterization, which effectively means pushing down additional
2022 : : * joinquals to be checked within the path's scan. However, some existing
2023 : : * paths might check the available joinquals already while others don't;
2024 : : * therefore, it's not clear which existing path will be cheapest after
2025 : : * reparameterization. We have to go through them all and find out.
2026 : : */
2027 : 146 : cheapest = NULL;
2028 [ + - + + : 506 : foreach(lc, rel->pathlist)
+ + ]
2029 : : {
2030 : 360 : Path *path = (Path *) lfirst(lc);
2031 : :
2032 : : /* Can't use it if it needs more than requested parameterization */
2033 [ + + + + ]: 360 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2034 : 12 : continue;
2035 : :
2036 : : /*
2037 : : * Reparameterization can only increase the path's cost, so if it's
2038 : : * already more expensive than the current cheapest, forget it.
2039 : : */
2040 [ + + + + ]: 540 : if (cheapest != NULL &&
2041 : 192 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2042 : 156 : continue;
2043 : :
2044 : : /* Reparameterize if needed, then recheck cost */
2045 [ + + + + ]: 192 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2046 : : {
2047 : 154 : path = reparameterize_path(root, path, required_outer, 1.0);
2048 [ + + ]: 154 : if (path == NULL)
2049 : 16 : continue; /* failed to reparameterize this one */
2050 [ + - - + ]: 138 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2051 : :
2052 [ - + - - ]: 138 : if (cheapest != NULL &&
3934 tgl@sss.pgh.pa.us 2053 :UBC 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2054 : 0 : continue;
2055 : : }
2056 : :
2057 : : /* We have a new best path */
3934 tgl@sss.pgh.pa.us 2058 :CBC 176 : cheapest = path;
2059 : : }
2060 : :
2061 : : /* Return the best path, or NULL if we found no suitable candidate */
2062 : 146 : return cheapest;
2063 : : }
2064 : :
2065 : : /*
2066 : : * accumulate_append_subpath
2067 : : * Add a subpath to the list being built for an Append or MergeAppend.
2068 : : *
2069 : : * It's possible that the child is itself an Append or MergeAppend path, in
2070 : : * which case we can "cut out the middleman" and just add its child paths to
2071 : : * our own list. (We don't try to do this earlier because we need to apply
2072 : : * both levels of transformation to the quals.)
2073 : : *
2074 : : * Note that if we omit a child MergeAppend in this way, we are effectively
2075 : : * omitting a sort step, which seems fine: if the parent is to be an Append,
2076 : : * its result would be unsorted anyway, while if the parent is to be a
2077 : : * MergeAppend, there's no point in a separate sort on a child.
2078 : : *
2079 : : * Normally, either path is a partial path and subpaths is a list of partial
2080 : : * paths, or else path is a non-partial plan and subpaths is a list of those.
2081 : : * However, if path is a parallel-aware Append, then we add its partial path
2082 : : * children to subpaths and the rest to special_subpaths. If the latter is
2083 : : * NULL, we don't flatten the path at all (unless it contains only partial
2084 : : * paths).
2085 : : */
2086 : : static void
1168 2087 : 96936 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2088 : : {
4931 2089 [ + + ]: 96936 : if (IsA(path, AppendPath))
2090 : : {
4753 bruce@momjian.us 2091 : 6940 : AppendPath *apath = (AppendPath *) path;
2092 : :
2322 rhaas@postgresql.org 2093 [ + + + + ]: 6940 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2094 : : {
1707 tgl@sss.pgh.pa.us 2095 : 6844 : *subpaths = list_concat(*subpaths, apath->subpaths);
2322 rhaas@postgresql.org 2096 : 6844 : return;
2097 : : }
2098 [ + + ]: 96 : else if (special_subpaths != NULL)
2099 : : {
2100 : : List *new_special_subpaths;
2101 : :
2102 : : /* Split Parallel Append into partial and non-partial subpaths */
2103 : 48 : *subpaths = list_concat(*subpaths,
2104 : 48 : list_copy_tail(apath->subpaths,
2105 : : apath->first_partial_path));
641 drowley@postgresql.o 2106 : 48 : new_special_subpaths = list_copy_head(apath->subpaths,
2107 : : apath->first_partial_path);
2322 rhaas@postgresql.org 2108 : 48 : *special_subpaths = list_concat(*special_subpaths,
2109 : : new_special_subpaths);
2286 2110 : 48 : return;
2111 : : }
2112 : : }
3670 tgl@sss.pgh.pa.us 2113 [ + + ]: 89996 : else if (IsA(path, MergeAppendPath))
2114 : : {
2115 : 310 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2116 : :
1707 2117 : 310 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2322 rhaas@postgresql.org 2118 : 310 : return;
2119 : : }
2120 : :
2121 : 89734 : *subpaths = lappend(*subpaths, path);
2122 : : }
2123 : :
2124 : : /*
2125 : : * get_singleton_append_subpath
2126 : : * Returns the single subpath of an Append/MergeAppend, or just
2127 : : * return 'path' if it's not a single sub-path Append/MergeAppend.
2128 : : *
2129 : : * Note: 'path' must not be a parallel-aware path.
2130 : : */
2131 : : static Path *
1836 tgl@sss.pgh.pa.us 2132 : 6038 : get_singleton_append_subpath(Path *path)
2133 : : {
2134 [ - + ]: 6038 : Assert(!path->parallel_aware);
2135 : :
2136 [ + + ]: 6038 : if (IsA(path, AppendPath))
2137 : : {
2138 : 170 : AppendPath *apath = (AppendPath *) path;
2139 : :
2140 [ + + ]: 170 : if (list_length(apath->subpaths) == 1)
2141 : 78 : return (Path *) linitial(apath->subpaths);
2142 : : }
2143 [ + + ]: 5868 : else if (IsA(path, MergeAppendPath))
2144 : : {
2145 : 126 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2146 : :
2147 [ - + ]: 126 : if (list_length(mpath->subpaths) == 1)
1836 tgl@sss.pgh.pa.us 2148 :UBC 0 : return (Path *) linitial(mpath->subpaths);
2149 : : }
2150 : :
1836 tgl@sss.pgh.pa.us 2151 :CBC 5960 : return path;
2152 : : }
2153 : :
2154 : : /*
2155 : : * set_dummy_rel_pathlist
2156 : : * Build a dummy path for a relation that's been excluded by constraints
2157 : : *
2158 : : * Rather than inventing a special "dummy" path type, we represent this as an
2159 : : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2160 : : *
2161 : : * (See also mark_dummy_rel, which does basically the same thing, but is
2162 : : * typically used to change a rel into dummy state after we already made
2163 : : * paths for it.)
2164 : : */
2165 : : static void
6168 2166 : 503 : set_dummy_rel_pathlist(RelOptInfo *rel)
2167 : : {
2168 : : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2169 : 503 : rel->rows = 0;
2953 2170 : 503 : rel->reltarget->width = 0;
2171 : :
2172 : : /* Discard any pre-existing paths; no further need for them */
4461 2173 : 503 : rel->pathlist = NIL;
3007 rhaas@postgresql.org 2174 : 503 : rel->partial_pathlist = NIL;
2175 : :
2176 : : /* Set up the dummy path */
1858 tgl@sss.pgh.pa.us 2177 : 503 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2178 : : NIL, rel->lateral_relids,
2179 : : 0, false, -1));
2180 : :
2181 : : /*
2182 : : * We set the cheapest-path fields immediately, just in case they were
2183 : : * pointing at some discarded path. This is redundant in current usage
2184 : : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2185 : : * for safety and consistency with mark_dummy_rel.
2186 : : */
6168 2187 : 503 : set_cheapest(rel);
2188 : 503 : }
2189 : :
2190 : : /*
2191 : : * find_window_run_conditions
2192 : : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2193 : : * function to determine the function's monotonic properties. We then
2194 : : * see if 'opexpr' can be used to short-circuit execution.
2195 : : *
2196 : : * For example row_number() over (order by ...) always produces a value one
2197 : : * higher than the previous. If someone has a window function in a subquery
2198 : : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2199 : : * as well stop processing the windowagg once the row number reaches 11. Here
2200 : : * we check if 'opexpr' might help us to stop doing needless extra processing
2201 : : * in WindowAgg nodes.
2202 : : *
2203 : : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2204 : : * its original purpose. This is set to false if the caller can assume that
2205 : : * the run condition will handle all of the required filtering.
2206 : : *
2207 : : * Returns true if 'opexpr' was found to be useful and was added to the
2208 : : * WindowClauses runCondition. We also set *keep_original accordingly and add
2209 : : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2210 : : * If the 'opexpr' cannot be used then we set *keep_original to true and
2211 : : * return false.
2212 : : */
2213 : : static bool
737 drowley@postgresql.o 2214 : 114 : find_window_run_conditions(Query *subquery, RangeTblEntry *rte, Index rti,
2215 : : AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2216 : : bool wfunc_left, bool *keep_original,
2217 : : Bitmapset **run_cond_attrs)
2218 : : {
2219 : : Oid prosupport;
2220 : : Expr *otherexpr;
2221 : : SupportRequestWFuncMonotonic req;
2222 : : SupportRequestWFuncMonotonic *res;
2223 : : WindowClause *wclause;
2224 : : List *opinfos;
2225 : : OpExpr *runopexpr;
2226 : : Oid runoperator;
2227 : : ListCell *lc;
2228 : :
2229 : 114 : *keep_original = true;
2230 : :
2231 [ - + ]: 114 : while (IsA(wfunc, RelabelType))
737 drowley@postgresql.o 2232 :UBC 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2233 : :
2234 : : /* we can only work with window functions */
737 drowley@postgresql.o 2235 [ + + ]:CBC 114 : if (!IsA(wfunc, WindowFunc))
2236 : 12 : return false;
2237 : :
2238 : : /* can't use it if there are subplans in the WindowFunc */
394 2239 [ + + ]: 102 : if (contain_subplans((Node *) wfunc))
2240 : 3 : return false;
2241 : :
737 2242 : 99 : prosupport = get_func_support(wfunc->winfnoid);
2243 : :
2244 : : /* Check if there's a support function for 'wfunc' */
2245 [ + + ]: 99 : if (!OidIsValid(prosupport))
2246 : 9 : return false;
2247 : :
2248 : : /* get the Expr from the other side of the OpExpr */
2249 [ + + ]: 90 : if (wfunc_left)
2250 : 78 : otherexpr = lsecond(opexpr->args);
2251 : : else
2252 : 12 : otherexpr = linitial(opexpr->args);
2253 : :
2254 : : /*
2255 : : * The value being compared must not change during the evaluation of the
2256 : : * window partition.
2257 : : */
2258 [ - + ]: 90 : if (!is_pseudo_constant_clause((Node *) otherexpr))
737 drowley@postgresql.o 2259 :UBC 0 : return false;
2260 : :
2261 : : /* find the window clause belonging to the window function */
737 drowley@postgresql.o 2262 :CBC 90 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2263 : 90 : wfunc->winref - 1);
2264 : :
2265 : 90 : req.type = T_SupportRequestWFuncMonotonic;
2266 : 90 : req.window_func = wfunc;
2267 : 90 : req.window_clause = wclause;
2268 : :
2269 : : /* call the support function */
2270 : : res = (SupportRequestWFuncMonotonic *)
2271 : 90 : DatumGetPointer(OidFunctionCall1(prosupport,
2272 : : PointerGetDatum(&req)));
2273 : :
2274 : : /*
2275 : : * Nothing to do if the function is neither monotonically increasing nor
2276 : : * monotonically decreasing.
2277 : : */
2278 [ + - - + ]: 90 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
737 drowley@postgresql.o 2279 :UBC 0 : return false;
2280 : :
737 drowley@postgresql.o 2281 :CBC 90 : runopexpr = NULL;
2282 : 90 : runoperator = InvalidOid;
2283 : 90 : opinfos = get_op_btree_interpretation(opexpr->opno);
2284 : :
2285 [ + - + - : 90 : foreach(lc, opinfos)
+ - ]
2286 : : {
2287 : 90 : OpBtreeInterpretation *opinfo = (OpBtreeInterpretation *) lfirst(lc);
2288 : 90 : int strategy = opinfo->strategy;
2289 : :
2290 : : /* handle < / <= */
2291 [ + + + + ]: 90 : if (strategy == BTLessStrategyNumber ||
2292 : : strategy == BTLessEqualStrategyNumber)
2293 : : {
2294 : : /*
2295 : : * < / <= is supported for monotonically increasing functions in
2296 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2297 : : * for monotonically decreasing functions.
2298 : : */
2299 [ + + + + ]: 66 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2300 [ + + + + ]: 9 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2301 : : {
2302 : 60 : *keep_original = false;
2303 : 60 : runopexpr = opexpr;
2304 : 60 : runoperator = opexpr->opno;
2305 : : }
2306 : 66 : break;
2307 : : }
2308 : : /* handle > / >= */
2309 [ + + + + ]: 24 : else if (strategy == BTGreaterStrategyNumber ||
2310 : : strategy == BTGreaterEqualStrategyNumber)
2311 : : {
2312 : : /*
2313 : : * > / >= is supported for monotonically decreasing functions in
2314 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2315 : : * for monotonically increasing functions.
2316 : : */
2317 [ + + - + ]: 9 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2318 [ + - + - ]: 6 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2319 : : {
2320 : 9 : *keep_original = false;
2321 : 9 : runopexpr = opexpr;
2322 : 9 : runoperator = opexpr->opno;
2323 : : }
2324 : 9 : break;
2325 : : }
2326 : : /* handle = */
2327 [ + - ]: 15 : else if (strategy == BTEqualStrategyNumber)
2328 : : {
2329 : : int16 newstrategy;
2330 : :
2331 : : /*
2332 : : * When both monotonically increasing and decreasing then the
2333 : : * return value of the window function will be the same each time.
2334 : : * We can simply use 'opexpr' as the run condition without
2335 : : * modifying it.
2336 : : */
2337 [ + + ]: 15 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2338 : : {
2339 : 3 : *keep_original = false;
2340 : 3 : runopexpr = opexpr;
618 2341 : 3 : runoperator = opexpr->opno;
737 2342 : 3 : break;
2343 : : }
2344 : :
2345 : : /*
2346 : : * When monotonically increasing we make a qual with <wfunc> <=
2347 : : * <value> or <value> >= <wfunc> in order to filter out values
2348 : : * which are above the value in the equality condition. For
2349 : : * monotonically decreasing functions we want to filter values
2350 : : * below the value in the equality condition.
2351 : : */
2352 [ + - ]: 12 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2353 [ + - ]: 12 : newstrategy = wfunc_left ? BTLessEqualStrategyNumber : BTGreaterEqualStrategyNumber;
2354 : : else
737 drowley@postgresql.o 2355 [ # # ]:UBC 0 : newstrategy = wfunc_left ? BTGreaterEqualStrategyNumber : BTLessEqualStrategyNumber;
2356 : :
2357 : : /* We must keep the original equality qual */
737 drowley@postgresql.o 2358 :CBC 12 : *keep_original = true;
2359 : 12 : runopexpr = opexpr;
2360 : :
2361 : : /* determine the operator to use for the runCondition qual */
2362 : 12 : runoperator = get_opfamily_member(opinfo->opfamily_id,
2363 : : opinfo->oplefttype,
2364 : : opinfo->oprighttype,
2365 : : newstrategy);
2366 : 12 : break;
2367 : : }
2368 : : }
2369 : :
2370 [ + + ]: 90 : if (runopexpr != NULL)
2371 : : {
2372 : : Expr *newexpr;
2373 : :
2374 : : /*
2375 : : * Build the qual required for the run condition keeping the
2376 : : * WindowFunc on the same side as it was originally.
2377 : : */
2378 [ + + ]: 84 : if (wfunc_left)
2379 : 75 : newexpr = make_opclause(runoperator,
2380 : : runopexpr->opresulttype,
2381 : 75 : runopexpr->opretset, (Expr *) wfunc,
2382 : : otherexpr, runopexpr->opcollid,
2383 : : runopexpr->inputcollid);
2384 : : else
2385 : 9 : newexpr = make_opclause(runoperator,
2386 : : runopexpr->opresulttype,
2387 : 9 : runopexpr->opretset,
2388 : : otherexpr, (Expr *) wfunc,
2389 : : runopexpr->opcollid,
2390 : : runopexpr->inputcollid);
2391 : :
2392 : 84 : wclause->runCondition = lappend(wclause->runCondition, newexpr);
2393 : :
2394 : : /* record that this attno was used in a run condition */
688 2395 : 84 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2396 : : attno - FirstLowInvalidHeapAttributeNumber);
737 2397 : 84 : return true;
2398 : : }
2399 : :
2400 : : /* unsupported OpExpr */
2401 : 6 : return false;
2402 : : }
2403 : :
2404 : : /*
2405 : : * check_and_push_window_quals
2406 : : * Check if 'clause' is a qual that can be pushed into a WindowFunc's
2407 : : * WindowClause as a 'runCondition' qual. These, when present, allow
2408 : : * some unnecessary work to be skipped during execution.
2409 : : *
2410 : : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2411 : : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2412 : : * window quals for.
2413 : : *
2414 : : * Returns true if the caller still must keep the original qual or false if
2415 : : * the caller can safely ignore the original qual because the WindowAgg node
2416 : : * will use the runCondition to stop returning tuples.
2417 : : */
2418 : : static bool
2419 : 120 : check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti,
2420 : : Node *clause, Bitmapset **run_cond_attrs)
2421 : : {
2422 : 120 : OpExpr *opexpr = (OpExpr *) clause;
2423 : 120 : bool keep_original = true;
2424 : : Var *var1;
2425 : : Var *var2;
2426 : :
2427 : : /* We're only able to use OpExprs with 2 operands */
2428 [ + + ]: 120 : if (!IsA(opexpr, OpExpr))
2429 : 9 : return true;
2430 : :
2431 [ - + ]: 111 : if (list_length(opexpr->args) != 2)
737 drowley@postgresql.o 2432 :UBC 0 : return true;
2433 : :
2434 : : /*
2435 : : * Currently, we restrict this optimization to strict OpExprs. The reason
2436 : : * for this is that during execution, once the runcondition becomes false,
2437 : : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2438 : : * function result values, we set them to NULL. Having only strict
2439 : : * OpExprs here ensures that we properly filter out the tuples with NULLs
2440 : : * in the top-level WindowAgg.
2441 : : */
494 drowley@postgresql.o 2442 :CBC 111 : set_opfuncid(opexpr);
2443 [ - + ]: 111 : if (!func_strict(opexpr->opfuncid))
494 drowley@postgresql.o 2444 :UBC 0 : return true;
2445 : :
2446 : : /*
2447 : : * Check for plain Vars that reference window functions in the subquery.
2448 : : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2449 : : * be used as part of the run condition.
2450 : : */
2451 : :
2452 : : /* Check the left side of the OpExpr */
737 drowley@postgresql.o 2453 :CBC 111 : var1 = linitial(opexpr->args);
2454 [ + + + - ]: 111 : if (IsA(var1, Var) && var1->varattno > 0)
2455 : : {
2456 : 93 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2457 : 93 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2458 : :
2459 [ + + ]: 93 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2460 : : opexpr, true, &keep_original,
2461 : : run_cond_attrs))
2462 : 75 : return keep_original;
2463 : : }
2464 : :
2465 : : /* and check the right side */
2466 : 36 : var2 = lsecond(opexpr->args);
2467 [ + + + - ]: 36 : if (IsA(var2, Var) && var2->varattno > 0)
2468 : : {
2469 : 21 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2470 : 21 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2471 : :
2472 [ + + ]: 21 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2473 : : opexpr, false, &keep_original,
2474 : : run_cond_attrs))
2475 : 9 : return keep_original;
2476 : : }
2477 : :
2478 : 27 : return true;
2479 : : }
2480 : :
2481 : : /*
2482 : : * set_subquery_pathlist
2483 : : * Generate SubqueryScan access paths for a subquery RTE
2484 : : *
2485 : : * We don't currently support generating parameterized paths for subqueries
2486 : : * by pushing join clauses down into them; it seems too expensive to re-plan
2487 : : * the subquery multiple times to consider different alternatives.
2488 : : * (XXX that could stand to be reconsidered, now that we use Paths.)
2489 : : * So the paths made here will be parameterized if the subquery contains
2490 : : * LATERAL references, otherwise not. As long as that's true, there's no need
2491 : : * for a separate set_subquery_size phase: just make the paths right away.
2492 : : */
2493 : : static void
6888 tgl@sss.pgh.pa.us 2494 : 3788 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2495 : : Index rti, RangeTblEntry *rte)
2496 : : {
6883 2497 : 3788 : Query *parse = root->parse;
8308 2498 : 3788 : Query *subquery = rte->subquery;
2499 : : bool trivial_pathtarget;
2500 : : Relids required_outer;
2501 : : pushdown_safety_info safetyInfo;
2502 : : double tuple_fraction;
2503 : : RelOptInfo *sub_final_rel;
688 drowley@postgresql.o 2504 : 3788 : Bitmapset *run_cond_attrs = NULL;
2505 : : ListCell *lc;
2506 : :
2507 : : /*
2508 : : * Must copy the Query so that planning doesn't mess up the RTE contents
2509 : : * (really really need to fix the planner to not scribble on its input,
2510 : : * someday ... but see remove_unused_subquery_outputs to start with).
2511 : : */
5514 tgl@sss.pgh.pa.us 2512 : 3788 : subquery = copyObject(subquery);
2513 : :
2514 : : /*
2515 : : * If it's a LATERAL subquery, it might contain some Vars of the current
2516 : : * query level, requiring it to be treated as parameterized, even though
2517 : : * we don't support pushing down join quals into subqueries.
2518 : : */
4249 2519 : 3788 : required_outer = rel->lateral_relids;
2520 : :
2521 : : /*
2522 : : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2523 : : * flags as needed while recursing. In particular, we need a workspace
2524 : : * for keeping track of the reasons why columns are unsafe to reference.
2525 : : * These reasons are stored in the bits inside unsafeFlags[i] when we
2526 : : * discover reasons that column i of the subquery is unsafe to be used in
2527 : : * a pushed-down qual.
2528 : : */
3579 2529 : 3788 : memset(&safetyInfo, 0, sizeof(safetyInfo));
394 drowley@postgresql.o 2530 : 3788 : safetyInfo.unsafeFlags = (unsigned char *)
2531 : 3788 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2532 : :
2533 : : /*
2534 : : * If the subquery has the "security_barrier" flag, it means the subquery
2535 : : * originated from a view that must enforce row-level security. Then we
2536 : : * must not push down quals that contain leaky functions. (Ideally this
2537 : : * would be checked inside subquery_is_pushdown_safe, but since we don't
2538 : : * currently pass the RTE to that function, we must do it here.)
2539 : : */
3579 tgl@sss.pgh.pa.us 2540 : 3788 : safetyInfo.unsafeLeaky = rte->security_barrier;
2541 : :
2542 : : /*
2543 : : * If there are any restriction clauses that have been attached to the
2544 : : * subquery relation, consider pushing them down to become WHERE or HAVING
2545 : : * quals of the subquery itself. This transformation is useful because it
2546 : : * may allow us to generate a better plan for the subquery than evaluating
2547 : : * all the subquery output rows and then filtering them.
2548 : : *
2549 : : * There are several cases where we cannot push down clauses. Restrictions
2550 : : * involving the subquery are checked by subquery_is_pushdown_safe().
2551 : : * Restrictions on individual clauses are checked by
2552 : : * qual_is_pushdown_safe(). Also, we don't want to push down
2553 : : * pseudoconstant clauses; better to have the gating node above the
2554 : : * subquery.
2555 : : *
2556 : : * Non-pushed-down clauses will get evaluated as qpquals of the
2557 : : * SubqueryScan node.
2558 : : *
2559 : : * XXX Are there any cases where we want to make a policy decision not to
2560 : : * push down a pushable qual, because it'd result in a worse plan?
2561 : : */
7899 2562 [ + + + + ]: 4436 : if (rel->baserestrictinfo != NIL &&
3579 2563 : 648 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2564 : : {
2565 : : /* OK to consider pushing down individual quals */
8308 2566 : 578 : List *upperrestrictlist = NIL;
2567 : : ListCell *l;
2568 : :
7263 neilc@samurai.com 2569 [ + - + + : 1344 : foreach(l, rel->baserestrictinfo)
+ + ]
2570 : : {
2571 : 766 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
737 drowley@postgresql.o 2572 : 766 : Node *clause = (Node *) rinfo->clause;
2573 : :
394 2574 [ + + ]: 766 : if (rinfo->pseudoconstant)
2575 : : {
2576 : 2 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2577 : 2 : continue;
2578 : : }
2579 : :
2580 [ + + + - ]: 764 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2581 : : {
2582 : 438 : case PUSHDOWN_SAFE:
2583 : : /* Push it down */
2584 : 438 : subquery_push_qual(subquery, rte, rti, clause);
2585 : 438 : break;
2586 : :
2587 : 120 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2588 : :
2589 : : /*
2590 : : * Since we can't push the qual down into the subquery,
2591 : : * check if it happens to reference a window function. If
2592 : : * so then it might be useful to use for the WindowAgg's
2593 : : * runCondition.
2594 : : */
2595 [ + - + + ]: 240 : if (!subquery->hasWindowFuncs ||
2596 : 120 : check_and_push_window_quals(subquery, rte, rti, clause,
2597 : : &run_cond_attrs))
2598 : : {
2599 : : /*
2600 : : * subquery has no window funcs or the clause is not a
2601 : : * suitable window run condition qual or it is, but
2602 : : * the original must also be kept in the upper query.
2603 : : */
2604 : 48 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2605 : : }
2606 : 120 : break;
2607 : :
2608 : 206 : case PUSHDOWN_UNSAFE:
737 2609 : 206 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
394 2610 : 206 : break;
2611 : : }
2612 : : }
8308 tgl@sss.pgh.pa.us 2613 : 578 : rel->baserestrictinfo = upperrestrictlist;
2614 : : /* We don't bother recomputing baserestrict_min_security */
2615 : : }
2616 : :
394 drowley@postgresql.o 2617 : 3788 : pfree(safetyInfo.unsafeFlags);
2618 : :
2619 : : /*
2620 : : * The upper query might not use all the subquery's output columns; if
2621 : : * not, we can simplify. Pass the attributes that were pushed down into
2622 : : * WindowAgg run conditions to ensure we don't accidentally think those
2623 : : * are unused.
2624 : : */
688 2625 : 3788 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2626 : :
2627 : : /*
2628 : : * We can safely pass the outer tuple_fraction down to the subquery if the
2629 : : * outer level has no joining, aggregation, or sorting to do. Otherwise
2630 : : * we'd better tell the subquery to plan for full retrieval. (XXX This
2631 : : * could probably be made more intelligent ...)
2632 : : */
6883 tgl@sss.pgh.pa.us 2633 [ + + ]: 3788 : if (parse->hasAggs ||
2634 [ + + ]: 3276 : parse->groupClause ||
3256 andres@anarazel.de 2635 [ + - ]: 3273 : parse->groupingSets ||
544 tgl@sss.pgh.pa.us 2636 [ + - ]: 3273 : root->hasHavingQual ||
6883 2637 [ + + ]: 3273 : parse->distinctClause ||
2638 [ + + + + ]: 5263 : parse->sortClause ||
187 tgl@sss.pgh.pa.us 2639 :GNC 2202 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
6883 tgl@sss.pgh.pa.us 2640 :CBC 2144 : tuple_fraction = 0.0; /* default case */
2641 : : else
2642 : 1644 : tuple_fraction = root->tuple_fraction;
2643 : :
2644 : : /* plan_params should not be in use in current query level */
4239 2645 [ - + ]: 3788 : Assert(root->plan_params == NIL);
2646 : :
2647 : : /* Generate a subroot and Paths for the subquery */
12 drowley@postgresql.o 2648 :GNC 3788 : rel->subroot = subquery_planner(root->glob, subquery, root, false,
2649 : : tuple_fraction, NULL);
2650 : :
2651 : : /* Isolate the params needed by this specific subplan */
4239 tgl@sss.pgh.pa.us 2652 :CBC 3788 : rel->subplan_params = root->plan_params;
2653 : 3788 : root->plan_params = NIL;
2654 : :
2655 : : /*
2656 : : * It's possible that constraint exclusion proved the subquery empty. If
2657 : : * so, it's desirable to produce an unadorned dummy path so that we will
2658 : : * recognize appropriate optimizations at this query level.
2659 : : */
2960 2660 : 3788 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2661 : :
2662 [ + + ]: 3788 : if (IS_DUMMY_REL(sub_final_rel))
2663 : : {
4586 2664 : 54 : set_dummy_rel_pathlist(rel);
2665 : 54 : return;
2666 : : }
2667 : :
2668 : : /*
2669 : : * Mark rel with estimated output rows, width, etc. Note that we have to
2670 : : * do this before generating outer-query paths, else cost_subqueryscan is
2671 : : * not happy.
2672 : : */
4607 2673 : 3734 : set_subquery_size_estimates(root, rel);
2674 : :
2675 : : /*
2676 : : * Also detect whether the reltarget is trivial, so that we can pass that
2677 : : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2678 : : * It's trivial if it fetches all the subplan output columns in order.
2679 : : */
635 2680 [ + + ]: 3734 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2681 : 1147 : trivial_pathtarget = false;
2682 : : else
2683 : : {
2684 : 2587 : trivial_pathtarget = true;
2685 [ + - + + : 7377 : foreach(lc, rel->reltarget->exprs)
+ + ]
2686 : : {
2687 : 4930 : Node *node = (Node *) lfirst(lc);
2688 : : Var *var;
2689 : :
2690 [ - + ]: 4930 : if (!IsA(node, Var))
2691 : : {
635 tgl@sss.pgh.pa.us 2692 :UBC 0 : trivial_pathtarget = false;
2693 : 0 : break;
2694 : : }
635 tgl@sss.pgh.pa.us 2695 :CBC 4930 : var = (Var *) node;
2696 [ + - ]: 4930 : if (var->varno != rti ||
2697 [ + + ]: 4930 : var->varattno != foreach_current_index(lc) + 1)
2698 : : {
2699 : 140 : trivial_pathtarget = false;
2700 : 140 : break;
2701 : : }
2702 : : }
2703 : : }
2704 : :
2705 : : /*
2706 : : * For each Path that subquery_planner produced, make a SubqueryScanPath
2707 : : * in the outer query.
2708 : : */
2960 2709 [ + - + + : 7791 : foreach(lc, sub_final_rel->pathlist)
+ + ]
2710 : : {
2711 : 4057 : Path *subpath = (Path *) lfirst(lc);
2712 : : List *pathkeys;
2713 : :
2714 : : /* Convert subpath's pathkeys to outer representation */
2715 : 4057 : pathkeys = convert_subquery_pathkeys(root,
2716 : : rel,
2717 : : subpath->pathkeys,
2718 : : make_tlist_from_pathtarget(subpath->pathtarget));
2719 : :
2720 : : /* Generate outer path using this subpath */
2721 : 4057 : add_path(rel, (Path *)
2722 : 4057 : create_subqueryscan_path(root, rel, subpath,
2723 : : trivial_pathtarget,
2724 : : pathkeys, required_outer));
2725 : : }
2726 : :
2727 : : /* If outer rel allows parallelism, do same for partial paths. */
2181 rhaas@postgresql.org 2728 [ + + + + ]: 3734 : if (rel->consider_parallel && bms_is_empty(required_outer))
2729 : : {
2730 : : /* If consider_parallel is false, there should be no partial paths. */
2731 [ + + - + ]: 1922 : Assert(sub_final_rel->consider_parallel ||
2732 : : sub_final_rel->partial_pathlist == NIL);
2733 : :
2734 : : /* Same for partial paths. */
2735 [ + + + + : 1943 : foreach(lc, sub_final_rel->partial_pathlist)
+ + ]
2736 : : {
2737 : 21 : Path *subpath = (Path *) lfirst(lc);
2738 : : List *pathkeys;
2739 : :
2740 : : /* Convert subpath's pathkeys to outer representation */
2741 : 21 : pathkeys = convert_subquery_pathkeys(root,
2742 : : rel,
2743 : : subpath->pathkeys,
2744 : : make_tlist_from_pathtarget(subpath->pathtarget));
2745 : :
2746 : : /* Generate outer path using this subpath */
2747 : 21 : add_partial_path(rel, (Path *)
2748 : 21 : create_subqueryscan_path(root, rel, subpath,
2749 : : trivial_pathtarget,
2750 : : pathkeys,
2751 : : required_outer));
2752 : : }
2753 : : }
2754 : : }
2755 : :
2756 : : /*
2757 : : * set_function_pathlist
2758 : : * Build the (single) access path for a function RTE
2759 : : */
2760 : : static void
6888 tgl@sss.pgh.pa.us 2761 : 21525 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2762 : : {
2763 : : Relids required_outer;
3797 2764 : 21525 : List *pathkeys = NIL;
2765 : :
2766 : : /*
2767 : : * We don't support pushing join clauses into the quals of a function
2768 : : * scan, but it could still have required parameterization due to LATERAL
2769 : : * refs in the function expression.
2770 : : */
4249 2771 : 21525 : required_outer = rel->lateral_relids;
2772 : :
2773 : : /*
2774 : : * The result is considered unordered unless ORDINALITY was used, in which
2775 : : * case it is ordered by the ordinal column (the last one). See if we
2776 : : * care, by checking for uses of that Var in equivalence classes.
2777 : : */
3797 2778 [ + + ]: 21525 : if (rte->funcordinality)
2779 : : {
2780 : 342 : AttrNumber ordattno = rel->max_attr;
2781 : 342 : Var *var = NULL;
2782 : : ListCell *lc;
2783 : :
2784 : : /*
2785 : : * Is there a Var for it in rel's targetlist? If not, the query did
2786 : : * not reference the ordinality column, or at least not in any way
2787 : : * that would be interesting for sorting.
2788 : : */
2953 2789 [ + - + + : 919 : foreach(lc, rel->reltarget->exprs)
+ + ]
2790 : : {
3797 2791 : 916 : Var *node = (Var *) lfirst(lc);
2792 : :
2793 : : /* checking varno/varlevelsup is just paranoia */
2794 [ + - ]: 916 : if (IsA(node, Var) &&
2795 [ + + ]: 916 : node->varattno == ordattno &&
2796 [ + - ]: 339 : node->varno == rel->relid &&
2797 [ + - ]: 339 : node->varlevelsup == 0)
2798 : : {
2799 : 339 : var = node;
2800 : 339 : break;
2801 : : }
2802 : : }
2803 : :
2804 : : /*
2805 : : * Try to build pathkeys for this Var with int8 sorting. We tell
2806 : : * build_expression_pathkey not to build any new equivalence class; if
2807 : : * the Var isn't already mentioned in some EC, it means that nothing
2808 : : * cares about the ordering.
2809 : : */
2810 [ + + ]: 342 : if (var)
2811 : 339 : pathkeys = build_expression_pathkey(root,
2812 : : (Expr *) var,
2813 : : Int8LessOperator,
2814 : : rel->relids,
2815 : : false);
2816 : : }
2817 : :
2818 : : /* Generate appropriate path */
2819 : 21525 : add_path(rel, create_functionscan_path(root, rel,
2820 : : pathkeys, required_outer));
8008 2821 : 21525 : }
2822 : :
2823 : : /*
2824 : : * set_values_pathlist
2825 : : * Build the (single) access path for a VALUES RTE
2826 : : */
2827 : : static void
6465 mail@joeconway.com 2828 : 3858 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2829 : : {
2830 : : Relids required_outer;
2831 : :
2832 : : /*
2833 : : * We don't support pushing join clauses into the quals of a values scan,
2834 : : * but it could still have required parameterization due to LATERAL refs
2835 : : * in the values expressions.
2836 : : */
4249 tgl@sss.pgh.pa.us 2837 : 3858 : required_outer = rel->lateral_relids;
2838 : :
2839 : : /* Generate appropriate path */
4263 2840 : 3858 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
6465 mail@joeconway.com 2841 : 3858 : }
2842 : :
2843 : : /*
2844 : : * set_tablefunc_pathlist
2845 : : * Build the (single) access path for a table func RTE
2846 : : */
2847 : : static void
2594 alvherre@alvh.no-ip. 2848 : 254 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2849 : : {
2850 : : Relids required_outer;
2851 : :
2852 : : /*
2853 : : * We don't support pushing join clauses into the quals of a tablefunc
2854 : : * scan, but it could still have required parameterization due to LATERAL
2855 : : * refs in the function expression.
2856 : : */
2857 : 254 : required_outer = rel->lateral_relids;
2858 : :
2859 : : /* Generate appropriate path */
2860 : 254 : add_path(rel, create_tablefuncscan_path(root, rel,
2861 : : required_outer));
2862 : 254 : }
2863 : :
2864 : : /*
2865 : : * set_cte_pathlist
2866 : : * Build the (single) access path for a non-self-reference CTE RTE
2867 : : *
2868 : : * There's no need for a separate set_cte_size phase, since we don't
2869 : : * support join-qual-parameterized paths for CTEs.
2870 : : */
2871 : : static void
5671 tgl@sss.pgh.pa.us 2872 : 1603 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2873 : : {
2874 : : Path *ctepath;
2875 : : Plan *cteplan;
2876 : : PlannerInfo *cteroot;
2877 : : Index levelsup;
2878 : : List *pathkeys;
2879 : : int ndx;
2880 : : ListCell *lc;
2881 : : int plan_id;
2882 : : Relids required_outer;
2883 : :
2884 : : /*
2885 : : * Find the referenced CTE, and locate the path and plan previously made
2886 : : * for it.
2887 : : */
2888 : 1603 : levelsup = rte->ctelevelsup;
2889 : 1603 : cteroot = root;
2890 [ + + ]: 2848 : while (levelsup-- > 0)
2891 : : {
2892 : 1245 : cteroot = cteroot->parent_root;
2893 [ - + ]: 1245 : if (!cteroot) /* shouldn't happen */
5671 tgl@sss.pgh.pa.us 2894 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2895 : : }
2896 : :
2897 : : /*
2898 : : * Note: cte_plan_ids can be shorter than cteList, if we are still working
2899 : : * on planning the CTEs (ie, this is a side-reference from another CTE).
2900 : : * So we mustn't use forboth here.
2901 : : */
5671 tgl@sss.pgh.pa.us 2902 :CBC 1603 : ndx = 0;
2903 [ + - + - : 2294 : foreach(lc, cteroot->parse->cteList)
+ - ]
2904 : : {
2905 : 2294 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
2906 : :
2907 [ + + ]: 2294 : if (strcmp(cte->ctename, rte->ctename) == 0)
2908 : 1603 : break;
2909 : 691 : ndx++;
2910 : : }
2911 [ - + ]: 1603 : if (lc == NULL) /* shouldn't happen */
5671 tgl@sss.pgh.pa.us 2912 [ # # ]:UBC 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5671 tgl@sss.pgh.pa.us 2913 [ - + ]:CBC 1603 : if (ndx >= list_length(cteroot->cte_plan_ids))
5671 tgl@sss.pgh.pa.us 2914 [ # # ]:UBC 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5671 tgl@sss.pgh.pa.us 2915 :CBC 1603 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
724 2916 [ - + ]: 1603 : if (plan_id <= 0)
724 tgl@sss.pgh.pa.us 2917 [ # # ]:UBC 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
2918 : :
19 tgl@sss.pgh.pa.us 2919 [ - + ]:GNC 1603 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
2920 : 1603 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
5671 tgl@sss.pgh.pa.us 2921 :CBC 1603 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2922 : :
2923 : : /* Mark rel with estimated output rows, width, etc */
2960 2924 : 1603 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
2925 : :
2926 : : /* Convert the ctepath's pathkeys to outer query's representation */
19 tgl@sss.pgh.pa.us 2927 :GNC 1603 : pathkeys = convert_subquery_pathkeys(root,
2928 : : rel,
2929 : : ctepath->pathkeys,
2930 : : cteplan->targetlist);
2931 : :
2932 : : /*
2933 : : * We don't support pushing join clauses into the quals of a CTE scan, but
2934 : : * it could still have required parameterization due to LATERAL refs in
2935 : : * its tlist.
2936 : : */
4249 tgl@sss.pgh.pa.us 2937 :CBC 1603 : required_outer = rel->lateral_relids;
2938 : :
2939 : : /* Generate appropriate path */
19 tgl@sss.pgh.pa.us 2940 :GNC 1603 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
5671 tgl@sss.pgh.pa.us 2941 :CBC 1603 : }
2942 : :
2943 : : /*
2944 : : * set_namedtuplestore_pathlist
2945 : : * Build the (single) access path for a named tuplestore RTE
2946 : : *
2947 : : * There's no need for a separate set_namedtuplestore_size phase, since we
2948 : : * don't support join-qual-parameterized paths for tuplestores.
2949 : : */
2950 : : static void
2571 kgrittn@postgresql.o 2951 : 223 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
2952 : : RangeTblEntry *rte)
2953 : : {
2954 : : Relids required_outer;
2955 : :
2956 : : /* Mark rel with estimated output rows, width, etc */
2957 : 223 : set_namedtuplestore_size_estimates(root, rel);
2958 : :
2959 : : /*
2960 : : * We don't support pushing join clauses into the quals of a tuplestore
2961 : : * scan, but it could still have required parameterization due to LATERAL
2962 : : * refs in its tlist.
2963 : : */
2964 : 223 : required_outer = rel->lateral_relids;
2965 : :
2966 : : /* Generate appropriate path */
2967 : 223 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2968 : 223 : }
2969 : :
2970 : : /*
2971 : : * set_result_pathlist
2972 : : * Build the (single) access path for an RTE_RESULT RTE
2973 : : *
2974 : : * There's no need for a separate set_result_size phase, since we
2975 : : * don't support join-qual-parameterized paths for these RTEs.
2976 : : */
2977 : : static void
1903 tgl@sss.pgh.pa.us 2978 : 781 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
2979 : : RangeTblEntry *rte)
2980 : : {
2981 : : Relids required_outer;
2982 : :
2983 : : /* Mark rel with estimated output rows, width, etc */
2984 : 781 : set_result_size_estimates(root, rel);
2985 : :
2986 : : /*
2987 : : * We don't support pushing join clauses into the quals of a Result scan,
2988 : : * but it could still have required parameterization due to LATERAL refs
2989 : : * in its tlist.
2990 : : */
2991 : 781 : required_outer = rel->lateral_relids;
2992 : :
2993 : : /* Generate appropriate path */
2994 : 781 : add_path(rel, create_resultscan_path(root, rel, required_outer));
2995 : 781 : }
2996 : :
2997 : : /*
2998 : : * set_worktable_pathlist
2999 : : * Build the (single) access path for a self-reference CTE RTE
3000 : : *
3001 : : * There's no need for a separate set_worktable_size phase, since we don't
3002 : : * support join-qual-parameterized paths for CTEs.
3003 : : */
3004 : : static void
5671 3005 : 406 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3006 : : {
3007 : : Path *ctepath;
3008 : : PlannerInfo *cteroot;
3009 : : Index levelsup;
3010 : : Relids required_outer;
3011 : :
3012 : : /*
3013 : : * We need to find the non-recursive term's path, which is in the plan
3014 : : * level that's processing the recursive UNION, which is one level *below*
3015 : : * where the CTE comes from.
3016 : : */
3017 : 406 : levelsup = rte->ctelevelsup;
3018 [ - + ]: 406 : if (levelsup == 0) /* shouldn't happen */
5671 tgl@sss.pgh.pa.us 3019 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5671 tgl@sss.pgh.pa.us 3020 :CBC 406 : levelsup--;
3021 : 406 : cteroot = root;
3022 [ + + ]: 922 : while (levelsup-- > 0)
3023 : : {
3024 : 516 : cteroot = cteroot->parent_root;
3025 [ - + ]: 516 : if (!cteroot) /* shouldn't happen */
5671 tgl@sss.pgh.pa.us 3026 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3027 : : }
2960 tgl@sss.pgh.pa.us 3028 :CBC 406 : ctepath = cteroot->non_recursive_path;
3029 [ - + ]: 406 : if (!ctepath) /* shouldn't happen */
2960 tgl@sss.pgh.pa.us 3030 [ # # ]:UBC 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3031 : :
3032 : : /* Mark rel with estimated output rows, width, etc */
2960 tgl@sss.pgh.pa.us 3033 :CBC 406 : set_cte_size_estimates(root, rel, ctepath->rows);
3034 : :
3035 : : /*
3036 : : * We don't support pushing join clauses into the quals of a worktable
3037 : : * scan, but it could still have required parameterization due to LATERAL
3038 : : * refs in its tlist. (I'm not sure this is actually possible given the
3039 : : * restrictions on recursive references, but it's easy enough to support.)
3040 : : */
4249 3041 : 406 : required_outer = rel->lateral_relids;
3042 : :
3043 : : /* Generate appropriate path */
3044 : 406 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
5671 3045 : 406 : }
3046 : :
3047 : : /*
3048 : : * generate_gather_paths
3049 : : * Generate parallel access paths for a relation by pushing a Gather or
3050 : : * Gather Merge on top of a partial path.
3051 : : *
3052 : : * This must not be called until after we're done creating all partial paths
3053 : : * for the specified relation. (Otherwise, add_partial_path might delete a
3054 : : * path that some GatherPath or GatherMergePath has a reference to.)
3055 : : *
3056 : : * If we're generating paths for a scan or join relation, override_rows will
3057 : : * be false, and we'll just use the relation's size estimate. When we're
3058 : : * being called for a partially-grouped or partially-distinct path, though, we
3059 : : * need to override the rowcount estimate. (It's not clear that the
3060 : : * particular value we're using here is actually best, but the underlying rel
3061 : : * has no estimate so we must do something.)
3062 : : */
3063 : : void
2239 rhaas@postgresql.org 3064 : 8065 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3065 : : {
3066 : : Path *cheapest_partial_path;
3067 : : Path *simple_gather_path;
3068 : : ListCell *lc;
3069 : : double rows;
3070 : 8065 : double *rowsp = NULL;
3071 : :
3072 : : /* If there are no partial paths, there's nothing to do here. */
3007 3073 [ - + ]: 8065 : if (rel->partial_pathlist == NIL)
3007 rhaas@postgresql.org 3074 :UBC 0 : return;
3075 : :
3076 : : /* Should we override the rel's rowcount estimate? */
2239 rhaas@postgresql.org 3077 [ + + ]:CBC 8065 : if (override_rows)
3078 : 869 : rowsp = &rows;
3079 : :
3080 : : /*
3081 : : * The output of Gather is always unsorted, so there's only one partial
3082 : : * path of interest: the cheapest one. That will be the one at the front
3083 : : * of partial_pathlist because of the way add_partial_path works.
3084 : : */
3007 3085 : 8065 : cheapest_partial_path = linitial(rel->partial_pathlist);
2239 3086 : 8065 : rows =
3087 : 8065 : cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
3088 : : simple_gather_path = (Path *)
2946 3089 : 8065 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3090 : : NULL, rowsp);
3007 3091 : 8065 : add_path(rel, simple_gather_path);
3092 : :
3093 : : /*
3094 : : * For each useful ordering, we can consider an order-preserving Gather
3095 : : * Merge.
3096 : : */
2524 bruce@momjian.us 3097 [ + - + + : 16779 : foreach(lc, rel->partial_pathlist)
+ + ]
3098 : : {
3099 : 8714 : Path *subpath = (Path *) lfirst(lc);
3100 : : GatherMergePath *path;
3101 : :
2593 rhaas@postgresql.org 3102 [ + + ]: 8714 : if (subpath->pathkeys == NIL)
3103 : 7902 : continue;
3104 : :
2239 3105 : 812 : rows = subpath->rows * subpath->parallel_workers;
2593 3106 : 812 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3107 : : subpath->pathkeys, NULL, rowsp);
3108 : 812 : add_path(rel, &path->path);
3109 : : }
3110 : : }
3111 : :
3112 : : /*
3113 : : * get_useful_pathkeys_for_relation
3114 : : * Determine which orderings of a relation might be useful.
3115 : : *
3116 : : * Getting data in sorted order can be useful either because the requested
3117 : : * order matches the final output ordering for the overall query we're
3118 : : * planning, or because it enables an efficient merge join. Here, we try
3119 : : * to figure out which pathkeys to consider.
3120 : : *
3121 : : * This allows us to do incremental sort on top of an index scan under a gather
3122 : : * merge node, i.e. parallelized.
3123 : : *
3124 : : * If the require_parallel_safe is true, we also require the expressions to
3125 : : * be parallel safe (which allows pushing the sort below Gather Merge).
3126 : : *
3127 : : * XXX At the moment this can only ever return a list with a single element,
3128 : : * because it looks at query_pathkeys only. So we might return the pathkeys
3129 : : * directly, but it seems plausible we'll want to consider other orderings
3130 : : * in the future. For example, we might want to consider pathkeys useful for
3131 : : * merge joins.
3132 : : */
3133 : : static List *
1210 tomas.vondra@postgre 3134 : 8065 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3135 : : bool require_parallel_safe)
3136 : : {
1468 3137 : 8065 : List *useful_pathkeys_list = NIL;
3138 : :
3139 : : /*
3140 : : * Considering query_pathkeys is always worth it, because it might allow
3141 : : * us to avoid a total sort when we have a partially presorted path
3142 : : * available or to push the total sort into the parallel portion of the
3143 : : * query.
3144 : : */
3145 [ + + ]: 8065 : if (root->query_pathkeys)
3146 : : {
3147 : : ListCell *lc;
1431 tgl@sss.pgh.pa.us 3148 : 3473 : int npathkeys = 0; /* useful pathkeys */
3149 : :
1468 tomas.vondra@postgre 3150 [ + - + + : 7458 : foreach(lc, root->query_pathkeys)
+ + ]
3151 : : {
3152 : 5011 : PathKey *pathkey = (PathKey *) lfirst(lc);
3153 : 5011 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3154 : :
3155 : : /*
3156 : : * We can only build a sort for pathkeys that contain a
3157 : : * safe-to-compute-early EC member computable from the current
3158 : : * relation's reltarget, so ignore the remainder of the list as
3159 : : * soon as we find a pathkey without such a member.
3160 : : *
3161 : : * It's still worthwhile to return any prefix of the pathkeys list
3162 : : * that meets this requirement, as we may be able to do an
3163 : : * incremental sort.
3164 : : *
3165 : : * If requested, ensure the sort expression is parallel-safe too.
3166 : : */
1090 tgl@sss.pgh.pa.us 3167 [ + + ]: 5011 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3168 : : require_parallel_safe))
1468 tomas.vondra@postgre 3169 : 1026 : break;
3170 : :
3171 : 3985 : npathkeys++;
3172 : : }
3173 : :
3174 : : /*
3175 : : * The whole query_pathkeys list matches, so append it directly, to
3176 : : * allow comparing pathkeys easily by comparing list pointer. If we
3177 : : * have to truncate the pathkeys, we gotta do a copy though.
3178 : : */
3179 [ + + ]: 3473 : if (npathkeys == list_length(root->query_pathkeys))
3180 : 2447 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3181 : 2447 : root->query_pathkeys);
3182 [ + + ]: 1026 : else if (npathkeys > 0)
3183 : 230 : useful_pathkeys_list = lappend(useful_pathkeys_list,
641 drowley@postgresql.o 3184 : 230 : list_copy_head(root->query_pathkeys,
3185 : : npathkeys));
3186 : : }
3187 : :
1468 tomas.vondra@postgre 3188 : 8065 : return useful_pathkeys_list;
3189 : : }
3190 : :
3191 : : /*
3192 : : * generate_useful_gather_paths
3193 : : * Generate parallel access paths for a relation by pushing a Gather or
3194 : : * Gather Merge on top of a partial path.
3195 : : *
3196 : : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3197 : : * paths (aiming to preserve the ordering), but also considers ordering that
3198 : : * might be useful for nodes above the gather merge node, and tries to add
3199 : : * a sort (regular or incremental) to provide that.
3200 : : */
3201 : : void
3202 : 257481 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3203 : : {
3204 : : ListCell *lc;
3205 : : double rows;
3206 : 257481 : double *rowsp = NULL;
3207 : 257481 : List *useful_pathkeys_list = NIL;
3208 : 257481 : Path *cheapest_partial_path = NULL;
3209 : :
3210 : : /* If there are no partial paths, there's nothing to do here. */
3211 [ + + ]: 257481 : if (rel->partial_pathlist == NIL)
3212 : 249416 : return;
3213 : :
3214 : : /* Should we override the rel's rowcount estimate? */
3215 [ + + ]: 8065 : if (override_rows)
3216 : 869 : rowsp = &rows;
3217 : :
3218 : : /* generate the regular gather (merge) paths */
3219 : 8065 : generate_gather_paths(root, rel, override_rows);
3220 : :
3221 : : /* consider incremental sort for interesting orderings */
1210 3222 : 8065 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3223 : :
3224 : : /* used for explicit (full) sort paths */
1468 3225 : 8065 : cheapest_partial_path = linitial(rel->partial_pathlist);
3226 : :
3227 : : /*
3228 : : * Consider sorted paths for each interesting ordering. We generate both
3229 : : * incremental and full sort.
3230 : : */
3231 [ + + + + : 10742 : foreach(lc, useful_pathkeys_list)
+ + ]
3232 : : {
3233 : 2677 : List *useful_pathkeys = lfirst(lc);
3234 : : ListCell *lc2;
3235 : : bool is_sorted;
3236 : : int presorted_keys;
3237 : :
3238 [ + - + + : 5946 : foreach(lc2, rel->partial_pathlist)
+ + ]
3239 : : {
3240 : 3269 : Path *subpath = (Path *) lfirst(lc2);
3241 : : GatherMergePath *path;
3242 : :
3243 : 3269 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3244 : : subpath->pathkeys,
3245 : : &presorted_keys);
3246 : :
3247 : : /*
3248 : : * We don't need to consider the case where a subpath is already
3249 : : * fully sorted because generate_gather_paths already creates a
3250 : : * gather merge path for every subpath that has pathkeys present.
3251 : : *
3252 : : * But since the subpath is already sorted, we know we don't need
3253 : : * to consider adding a sort (full or incremental) on top of it,
3254 : : * so we can continue here.
3255 : : */
3256 [ + + ]: 3269 : if (is_sorted)
3257 : 606 : continue;
3258 : :
3259 : : /*
3260 : : * Try at least sorting the cheapest path and also try
3261 : : * incrementally sorting any path which is partially sorted
3262 : : * already (no need to deal with paths which have presorted keys
3263 : : * when incremental sort is disabled unless it's the cheapest
3264 : : * input path).
3265 : : */
485 drowley@postgresql.o 3266 [ + + ]: 2663 : if (subpath != cheapest_partial_path &&
3267 [ + + + + ]: 99 : (presorted_keys == 0 || !enable_incremental_sort))
3268 : 27 : continue;
3269 : :
3270 : : /*
3271 : : * Consider regular sort for any path that's not presorted or if
3272 : : * incremental sort is disabled. We've no need to consider both
3273 : : * sort and incremental sort on the same path. We assume that
3274 : : * incremental sort is always faster when there are presorted
3275 : : * keys.
3276 : : *
3277 : : * This is not redundant with the gather paths created in
3278 : : * generate_gather_paths, because that doesn't generate ordered
3279 : : * output. Here we add an explicit sort to match the useful
3280 : : * ordering.
3281 : : */
3282 [ + + + + ]: 2636 : if (presorted_keys == 0 || !enable_incremental_sort)
3283 : : {
3284 : 2556 : subpath = (Path *) create_sort_path(root,
3285 : : rel,
3286 : : subpath,
3287 : : useful_pathkeys,
3288 : : -1.0);
3289 : 2556 : rows = subpath->rows * subpath->parallel_workers;
3290 : : }
3291 : : else
3292 : 80 : subpath = (Path *) create_incremental_sort_path(root,
3293 : : rel,
3294 : : subpath,
3295 : : useful_pathkeys,
3296 : : presorted_keys,
3297 : : -1);
3298 : 2636 : path = create_gather_merge_path(root, rel,
3299 : : subpath,
3300 : 2636 : rel->reltarget,
3301 : : subpath->pathkeys,
3302 : : NULL,
3303 : : rowsp);
3304 : :
3305 : 2636 : add_path(rel, &path->path);
3306 : : }
3307 : : }
3308 : : }
3309 : :
3310 : : /*
3311 : : * make_rel_from_joinlist
3312 : : * Build access paths using a "joinlist" to guide the join path search.
3313 : : *
3314 : : * See comments for deconstruct_jointree() for definition of the joinlist
3315 : : * data structure.
3316 : : */
3317 : : static RelOptInfo *
6690 tgl@sss.pgh.pa.us 3318 : 141162 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3319 : : {
3320 : : int levels_needed;
3321 : : List *initial_rels;
3322 : : ListCell *jl;
3323 : :
3324 : : /*
3325 : : * Count the number of child joinlist nodes. This is the depth of the
3326 : : * dynamic-programming algorithm we must employ to consider all ways of
3327 : : * joining the child nodes.
3328 : : */
3329 : 141162 : levels_needed = list_length(joinlist);
3330 : :
8598 3331 [ - + ]: 141162 : if (levels_needed <= 0)
8598 tgl@sss.pgh.pa.us 3332 :UBC 0 : return NULL; /* nothing to do? */
3333 : :
3334 : : /*
3335 : : * Construct a list of rels corresponding to the child joinlist nodes.
3336 : : * This may contain both base rels and rels constructed according to
3337 : : * sub-joinlists.
3338 : : */
6690 tgl@sss.pgh.pa.us 3339 :CBC 141162 : initial_rels = NIL;
3340 [ + - + + : 338932 : foreach(jl, joinlist)
+ + ]
3341 : : {
3342 : 197770 : Node *jlnode = (Node *) lfirst(jl);
3343 : : RelOptInfo *thisrel;
3344 : :
3345 [ + + ]: 197770 : if (IsA(jlnode, RangeTblRef))
3346 : : {
3347 : 196129 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3348 : :
3349 : 196129 : thisrel = find_base_rel(root, varno);
3350 : : }
3351 [ + - ]: 1641 : else if (IsA(jlnode, List))
3352 : : {
3353 : : /* Recurse to handle subproblem */
3354 : 1641 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3355 : : }
3356 : : else
3357 : : {
6690 tgl@sss.pgh.pa.us 3358 [ # # ]:UBC 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3359 : : (int) nodeTag(jlnode));
3360 : : thisrel = NULL; /* keep compiler quiet */
3361 : : }
3362 : :
6690 tgl@sss.pgh.pa.us 3363 :CBC 197770 : initial_rels = lappend(initial_rels, thisrel);
3364 : : }
3365 : :
8598 3366 [ + + ]: 141162 : if (levels_needed == 1)
3367 : : {
3368 : : /*
3369 : : * Single joinlist node, so we're done.
3370 : : */
7263 neilc@samurai.com 3371 : 100277 : return (RelOptInfo *) linitial(initial_rels);
3372 : : }
3373 : : else
3374 : : {
3375 : : /*
3376 : : * Consider the different orders in which we could join the rels,
3377 : : * using a plugin, GEQO, or the regular join search code.
3378 : : *
3379 : : * We put the initial_rels list into a PlannerInfo field because
3380 : : * has_legal_joinclause() needs to look at it (ugly :-().
3381 : : */
5938 tgl@sss.pgh.pa.us 3382 : 40885 : root->initial_rels = initial_rels;
3383 : :
6045 3384 [ - + ]: 40885 : if (join_search_hook)
6045 tgl@sss.pgh.pa.us 3385 :UBC 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
6045 tgl@sss.pgh.pa.us 3386 [ + - + + ]:CBC 40885 : else if (enable_geqo && levels_needed >= geqo_threshold)
8598 3387 : 3 : return geqo(root, levels_needed, initial_rels);
3388 : : else
6045 3389 : 40882 : return standard_join_search(root, levels_needed, initial_rels);
3390 : : }
3391 : : }
3392 : :
3393 : : /*
3394 : : * standard_join_search
3395 : : * Find possible joinpaths for a query by successively finding ways
3396 : : * to join component relations into join relations.
3397 : : *
3398 : : * 'levels_needed' is the number of iterations needed, ie, the number of
3399 : : * independent jointree items in the query. This is > 1.
3400 : : *
3401 : : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3402 : : * jointree item. These are the components to be joined together.
3403 : : * Note that levels_needed == list_length(initial_rels).
3404 : : *
3405 : : * Returns the final level of join relations, i.e., the relation that is
3406 : : * the result of joining all the original relations together.
3407 : : * At least one implementation path must be provided for this relation and
3408 : : * all required sub-relations.
3409 : : *
3410 : : * To support loadable plugins that modify planner behavior by changing the
3411 : : * join searching algorithm, we provide a hook variable that lets a plugin
3412 : : * replace or supplement this function. Any such hook must return the same
3413 : : * final join relation as the standard code would, but it might have a
3414 : : * different set of implementation paths attached, and only the sub-joinrels
3415 : : * needed for these paths need have been instantiated.
3416 : : *
3417 : : * Note to plugin authors: the functions invoked during standard_join_search()
3418 : : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3419 : : * than one join-order search, you'll probably need to save and restore the
3420 : : * original states of those data structures. See geqo_eval() for an example.
3421 : : */
3422 : : RelOptInfo *
3423 : 40882 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3424 : : {
3425 : : int lev;
3426 : : RelOptInfo *rel;
3427 : :
3428 : : /*
3429 : : * This function cannot be invoked recursively within any one planning
3430 : : * problem, so join_rel_level[] can't be in use already.
3431 : : */
5251 3432 [ - + ]: 40882 : Assert(root->join_rel_level == NULL);
3433 : :
3434 : : /*
3435 : : * We employ a simple "dynamic programming" algorithm: we first find all
3436 : : * ways to build joins of two jointree items, then all ways to build joins
3437 : : * of three items (from two-item joins and single items), then four-item
3438 : : * joins, and so on until we have considered all ways to join all the
3439 : : * items into one rel.
3440 : : *
3441 : : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3442 : : * set root->join_rel_level[1] to represent all the single-jointree-item
3443 : : * relations.
3444 : : */
3445 : 40882 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3446 : :
3447 : 40882 : root->join_rel_level[1] = initial_rels;
3448 : :
8833 3449 [ + + ]: 97476 : for (lev = 2; lev <= levels_needed; lev++)
3450 : : {
3451 : : ListCell *lc;
3452 : :
3453 : : /*
3454 : : * Determine all possible pairs of relations to be joined at this
3455 : : * level, and build paths for making each one from every available
3456 : : * pair of lower-level relations.
3457 : : */
5251 3458 : 56595 : join_search_one_level(root, lev);
3459 : :
3460 : : /*
3461 : : * Run generate_partitionwise_join_paths() and
3462 : : * generate_useful_gather_paths() for each just-processed joinrel. We
3463 : : * could not do this earlier because both regular and partial paths
3464 : : * can get added to a particular joinrel at multiple times within
3465 : : * join_search_one_level.
3466 : : *
3467 : : * After that, we're done creating paths for the joinrel, so run
3468 : : * set_cheapest().
3469 : : */
3470 [ + + + + : 143623 : foreach(lc, root->join_rel_level[lev])
+ + ]
3471 : : {
3472 : 87029 : rel = (RelOptInfo *) lfirst(lc);
3473 : :
3474 : : /* Create paths for partitionwise joins. */
2249 peter_e@gmx.net 3475 : 87029 : generate_partitionwise_join_paths(root, rel);
3476 : :
3477 : : /*
3478 : : * Except for the topmost scan/join rel, consider gathering
3479 : : * partial paths. We'll do the same for the topmost scan/join rel
3480 : : * once we know the final targetlist (see grouping_planner's and
3481 : : * its call to apply_scanjoin_target_to_paths).
3482 : : */
440 tgl@sss.pgh.pa.us 3483 [ + + ]: 87029 : if (!bms_equal(rel->relids, root->all_query_rels))
1468 tomas.vondra@postgre 3484 : 46382 : generate_useful_gather_paths(root, rel, false);
3485 : :
3486 : : /* Find and save the cheapest paths for this rel */
8825 tgl@sss.pgh.pa.us 3487 : 87029 : set_cheapest(rel);
3488 : :
3489 : : #ifdef OPTIMIZER_DEBUG
3490 : : pprint(rel);
3491 : : #endif
3492 : : }
3493 : : }
3494 : :
3495 : : /*
3496 : : * We should have a single rel at the final level.
3497 : : */
5251 3498 [ - + ]: 40881 : if (root->join_rel_level[levels_needed] == NIL)
7424 tgl@sss.pgh.pa.us 3499 [ # # ]:UBC 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
5251 tgl@sss.pgh.pa.us 3500 [ - + ]:CBC 40881 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3501 : :
3502 : 40881 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3503 : :
3504 : 40881 : root->join_rel_level = NULL;
3505 : :
8833 3506 : 40881 : return rel;
3507 : : }
3508 : :
3509 : : /*****************************************************************************
3510 : : * PUSHING QUALS DOWN INTO SUBQUERIES
3511 : : *****************************************************************************/
3512 : :
3513 : : /*
3514 : : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3515 : : *
3516 : : * subquery is the particular component query being checked. topquery
3517 : : * is the top component of a set-operations tree (the same Query if no
3518 : : * set-op is involved).
3519 : : *
3520 : : * Conditions checked here:
3521 : : *
3522 : : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3523 : : * since that could change the set of rows returned.
3524 : : *
3525 : : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3526 : : * quals into it, because that could change the results.
3527 : : *
3528 : : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3529 : : * This is because upper-level quals should semantically be evaluated only
3530 : : * once per distinct row, not once per original row, and if the qual is
3531 : : * volatile then extra evaluations could change the results. (This issue
3532 : : * does not apply to other forms of aggregation such as GROUP BY, because
3533 : : * when those are present we push into HAVING not WHERE, so that the quals
3534 : : * are still applied after aggregation.)
3535 : : *
3536 : : * 4. If the subquery contains window functions, we cannot push volatile quals
3537 : : * into it. The issue here is a bit different from DISTINCT: a volatile qual
3538 : : * might succeed for some rows of a window partition and fail for others,
3539 : : * thereby changing the partition contents and thus the window functions'
3540 : : * results for rows that remain.
3541 : : *
3542 : : * 5. If the subquery contains any set-returning functions in its targetlist,
3543 : : * we cannot push volatile quals into it. That would push them below the SRFs
3544 : : * and thereby change the number of times they are evaluated. Also, a
3545 : : * volatile qual could succeed for some SRF output rows and fail for others,
3546 : : * a behavior that cannot occur if it's evaluated before SRF expansion.
3547 : : *
3548 : : * 6. If the subquery has nonempty grouping sets, we cannot push down any
3549 : : * quals. The concern here is that a qual referencing a "constant" grouping
3550 : : * column could get constant-folded, which would be improper because the value
3551 : : * is potentially nullable by grouping-set expansion. This restriction could
3552 : : * be removed if we had a parsetree representation that shows that such
3553 : : * grouping columns are not really constant. (There are other ideas that
3554 : : * could be used to relax this restriction, but that's the approach most
3555 : : * likely to get taken in the future. Note that there's not much to be gained
3556 : : * so long as subquery_planner can't move HAVING clauses to WHERE within such
3557 : : * a subquery.)
3558 : : *
3559 : : * In addition, we make several checks on the subquery's output columns to see
3560 : : * if it is safe to reference them in pushed-down quals. If output column k
3561 : : * is found to be unsafe to reference, we set the reason for that inside
3562 : : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3563 : : * column k might not be referenced by some/all quals. The unsafeFlags[]
3564 : : * array will be consulted later by qual_is_pushdown_safe(). It's better to
3565 : : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3566 : : * because when the subquery involves set operations we have to check the
3567 : : * output expressions in each arm of the set op.
3568 : : *
3569 : : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3570 : : * we're effectively assuming that the quals cannot distinguish values that
3571 : : * the DISTINCT's equality operator sees as equal, yet there are many
3572 : : * counterexamples to that assumption. However use of such a qual with a
3573 : : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3574 : : * "equal" value will be chosen as the output value by the DISTINCT operation.
3575 : : * So we don't worry too much about that. Another objection is that if the
3576 : : * qual is expensive to evaluate, running it for each original row might cost
3577 : : * more than we save by eliminating rows before the DISTINCT step. But it
3578 : : * would be very hard to estimate that at this stage, and in practice pushdown
3579 : : * seldom seems to make things worse, so we ignore that problem too.
3580 : : *
3581 : : * Note: likewise, pushing quals into a subquery with window functions is a
3582 : : * bit dubious: the quals might remove some rows of a window partition while
3583 : : * leaving others, causing changes in the window functions' results for the
3584 : : * surviving rows. We insist that such a qual reference only partitioning
3585 : : * columns, but again that only protects us if the qual does not distinguish
3586 : : * values that the partitioning equality operator sees as equal. The risks
3587 : : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3588 : : * occurs and thus there is no theoretical problem with such a qual. But
3589 : : * we'll do this anyway because the potential performance benefits are very
3590 : : * large, and we've seen no field complaints about the longstanding comparable
3591 : : * behavior with DISTINCT.
3592 : : */
3593 : : static bool
7661 3594 : 722 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
3595 : : pushdown_safety_info *safetyInfo)
3596 : : {
3597 : : SetOperationStmt *topop;
3598 : :
3599 : : /* Check point 1 */
7694 3600 [ + + + + ]: 722 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
7899 3601 : 64 : return false;
3602 : :
3603 : : /* Check point 6 */
1331 3604 [ + + + + ]: 658 : if (subquery->groupClause && subquery->groupingSets)
3605 : 6 : return false;
3606 : :
3607 : : /* Check points 3, 4, and 5 */
2756 3608 [ + + ]: 652 : if (subquery->distinctClause ||
3609 [ + + ]: 616 : subquery->hasWindowFuncs ||
3610 [ + + ]: 487 : subquery->hasTargetSRFs)
3579 3611 : 254 : safetyInfo->unsafeVolatile = true;
3612 : :
3613 : : /*
3614 : : * If we're at a leaf query, check for unsafe expressions in its target
3615 : : * list, and mark any reasons why they're unsafe in unsafeFlags[].
3616 : : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3617 : : * so no need to check them.)
3618 : : */
3966 3619 [ + + ]: 652 : if (subquery->setOperations == NULL)
3579 3620 : 615 : check_output_expressions(subquery, safetyInfo);
3621 : :
3622 : : /* Are we at top level, or looking at a setop component? */
7899 3623 [ + + ]: 652 : if (subquery == topquery)
3624 : : {
3625 : : /* Top level, so check any component queries */
3626 [ + + ]: 578 : if (subquery->setOperations != NULL)
7661 3627 [ - + ]: 37 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3628 : : safetyInfo))
7899 tgl@sss.pgh.pa.us 3629 :UBC 0 : return false;
3630 : : }
3631 : : else
3632 : : {
3633 : : /* Setop component must not have more components (too weird) */
7899 tgl@sss.pgh.pa.us 3634 [ - + ]:CBC 74 : if (subquery->setOperations != NULL)
7899 tgl@sss.pgh.pa.us 3635 :UBC 0 : return false;
3636 : : /* Check whether setop component output types match top level */
2609 peter_e@gmx.net 3637 :CBC 74 : topop = castNode(SetOperationStmt, topquery->setOperations);
3638 [ - + ]: 74 : Assert(topop);
7661 tgl@sss.pgh.pa.us 3639 : 74 : compare_tlist_datatypes(subquery->targetList,
3640 : : topop->colTypes,
3641 : : safetyInfo);
3642 : : }
7899 3643 : 652 : return true;
3644 : : }
3645 : :
3646 : : /*
3647 : : * Helper routine to recurse through setOperations tree
3648 : : */
3649 : : static bool
7661 3650 : 111 : recurse_pushdown_safe(Node *setOp, Query *topquery,
3651 : : pushdown_safety_info *safetyInfo)
3652 : : {
7899 3653 [ + + ]: 111 : if (IsA(setOp, RangeTblRef))
3654 : : {
3655 : 74 : RangeTblRef *rtr = (RangeTblRef *) setOp;
3656 : 74 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3657 : 74 : Query *subquery = rte->subquery;
3658 : :
3659 [ - + ]: 74 : Assert(subquery != NULL);
3579 3660 : 74 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3661 : : }
7899 3662 [ + - ]: 37 : else if (IsA(setOp, SetOperationStmt))
3663 : : {
3664 : 37 : SetOperationStmt *op = (SetOperationStmt *) setOp;
3665 : :
3666 : : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
3667 [ - + ]: 37 : if (op->op == SETOP_EXCEPT)
7899 tgl@sss.pgh.pa.us 3668 :UBC 0 : return false;
3669 : : /* Else recurse */
3579 tgl@sss.pgh.pa.us 3670 [ - + ]:CBC 37 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
7899 tgl@sss.pgh.pa.us 3671 :UBC 0 : return false;
3579 tgl@sss.pgh.pa.us 3672 [ - + ]:CBC 37 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
7899 tgl@sss.pgh.pa.us 3673 :UBC 0 : return false;
3674 : : }
3675 : : else
3676 : : {
7569 3677 [ # # ]: 0 : elog(ERROR, "unrecognized node type: %d",
3678 : : (int) nodeTag(setOp));
3679 : : }
7899 tgl@sss.pgh.pa.us 3680 :CBC 37 : return true;
3681 : : }
3682 : :
3683 : : /*
3684 : : * check_output_expressions - check subquery's output expressions for safety
3685 : : *
3686 : : * There are several cases in which it's unsafe to push down an upper-level
3687 : : * qual if it references a particular output column of a subquery. We check
3688 : : * each output column of the subquery and set flags in unsafeFlags[k] when we
3689 : : * see that column is unsafe for a pushed-down qual to reference. The
3690 : : * conditions checked here are:
3691 : : *
3692 : : * 1. We must not push down any quals that refer to subselect outputs that
3693 : : * return sets, else we'd introduce functions-returning-sets into the
3694 : : * subquery's WHERE/HAVING quals.
3695 : : *
3696 : : * 2. We must not push down any quals that refer to subselect outputs that
3697 : : * contain volatile functions, for fear of introducing strange results due
3698 : : * to multiple evaluation of a volatile function.
3699 : : *
3700 : : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3701 : : * refer to non-DISTINCT output columns, because that could change the set
3702 : : * of rows returned. (This condition is vacuous for DISTINCT, because then
3703 : : * there are no non-DISTINCT output columns, so we needn't check. Note that
3704 : : * subquery_is_pushdown_safe already reported that we can't use volatile
3705 : : * quals if there's DISTINCT or DISTINCT ON.)
3706 : : *
3707 : : * 4. If the subquery has any window functions, we must not push down quals
3708 : : * that reference any output columns that are not listed in all the subquery's
3709 : : * window PARTITION BY clauses. We can push down quals that use only
3710 : : * partitioning columns because they should succeed or fail identically for
3711 : : * every row of any one window partition, and totally excluding some
3712 : : * partitions will not change a window function's results for remaining
3713 : : * partitions. (Again, this also requires nonvolatile quals, but
3714 : : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3715 : : * unsafe for this reason can still have WindowClause run conditions pushed
3716 : : * down.
3717 : : */
3718 : : static void
3579 3719 : 615 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
3720 : : {
3721 : : ListCell *lc;
3722 : :
3966 3723 [ + - + + : 4286 : foreach(lc, subquery->targetList)
+ + ]
3724 : : {
3725 : 3671 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
3726 : :
3727 [ + + ]: 3671 : if (tle->resjunk)
3728 : 72 : continue; /* ignore resjunk columns */
3729 : :
3730 : : /* Functions returning sets are unsafe (point 1) */
2770 3731 [ + + ]: 3599 : if (subquery->hasTargetSRFs &&
394 drowley@postgresql.o 3732 [ + - ]: 295 : (safetyInfo->unsafeFlags[tle->resno] &
3733 [ + + ]: 295 : UNSAFE_HAS_SET_FUNC) == 0 &&
2770 tgl@sss.pgh.pa.us 3734 : 295 : expression_returns_set((Node *) tle->expr))
3735 : : {
394 drowley@postgresql.o 3736 : 164 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3966 tgl@sss.pgh.pa.us 3737 : 164 : continue;
3738 : : }
3739 : :
3740 : : /* Volatile functions are unsafe (point 2) */
394 drowley@postgresql.o 3741 [ + + ]: 3435 : if ((safetyInfo->unsafeFlags[tle->resno] &
3742 [ + + ]: 3429 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3743 : 3429 : contain_volatile_functions((Node *) tle->expr))
3744 : : {
3745 : 39 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3966 tgl@sss.pgh.pa.us 3746 : 39 : continue;
3747 : : }
3748 : :
3749 : : /* If subquery uses DISTINCT ON, check point 3 */
3750 [ - + ]: 3396 : if (subquery->hasDistinctOn &&
394 drowley@postgresql.o 3751 [ # # ]:UBC 0 : (safetyInfo->unsafeFlags[tle->resno] &
3752 : 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3966 tgl@sss.pgh.pa.us 3753 [ # # ]: 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3754 : : {
3755 : : /* non-DISTINCT column, so mark it unsafe */
394 drowley@postgresql.o 3756 : 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3966 tgl@sss.pgh.pa.us 3757 : 0 : continue;
3758 : : }
3759 : :
3760 : : /* If subquery uses window functions, check point 4 */
3579 tgl@sss.pgh.pa.us 3761 [ + + ]:CBC 3396 : if (subquery->hasWindowFuncs &&
394 drowley@postgresql.o 3762 [ + - ]: 570 : (safetyInfo->unsafeFlags[tle->resno] &
3763 : 1092 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3579 tgl@sss.pgh.pa.us 3764 [ + + ]: 570 : !targetIsInAllPartitionLists(tle, subquery))
3765 : : {
3766 : : /* not present in all PARTITION BY clauses, so mark it unsafe */
394 drowley@postgresql.o 3767 : 522 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3579 tgl@sss.pgh.pa.us 3768 : 522 : continue;
3769 : : }
3770 : : }
3966 3771 : 615 : }
3772 : :
3773 : : /*
3774 : : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3775 : : * push quals into each component query, but the quals can only reference
3776 : : * subquery columns that suffer no type coercions in the set operation.
3777 : : * Otherwise there are possible semantic gotchas. So, we check the
3778 : : * component queries to see if any of them have output types different from
3779 : : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3780 : : * unsafeFlags[k] if column k has different type in any component.
3781 : : *
3782 : : * We don't have to care about typmods here: the only allowed difference
3783 : : * between set-op input and output typmods is input is a specific typmod
3784 : : * and output is -1, and that does not require a coercion.
3785 : : *
3786 : : * tlist is a subquery tlist.
3787 : : * colTypes is an OID list of the top-level setop's output column types.
3788 : : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
3789 : : */
3790 : : static void
7661 3791 : 74 : compare_tlist_datatypes(List *tlist, List *colTypes,
3792 : : pushdown_safety_info *safetyInfo)
3793 : : {
3794 : : ListCell *l;
7263 neilc@samurai.com 3795 : 74 : ListCell *colType = list_head(colTypes);
3796 : :
3797 [ + - + + : 240 : foreach(l, tlist)
+ + ]
3798 : : {
3799 : 166 : TargetEntry *tle = (TargetEntry *) lfirst(l);
3800 : :
6948 tgl@sss.pgh.pa.us 3801 [ - + ]: 166 : if (tle->resjunk)
7661 tgl@sss.pgh.pa.us 3802 :UBC 0 : continue; /* ignore resjunk columns */
7263 neilc@samurai.com 3803 [ - + ]:CBC 166 : if (colType == NULL)
7661 tgl@sss.pgh.pa.us 3804 [ # # ]:UBC 0 : elog(ERROR, "wrong number of tlist entries");
6948 tgl@sss.pgh.pa.us 3805 [ + + ]:CBC 166 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
394 drowley@postgresql.o 3806 : 14 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
1735 tgl@sss.pgh.pa.us 3807 : 166 : colType = lnext(colTypes, colType);
3808 : : }
7263 neilc@samurai.com 3809 [ - + ]: 74 : if (colType != NULL)
7661 tgl@sss.pgh.pa.us 3810 [ # # ]:UBC 0 : elog(ERROR, "wrong number of tlist entries");
7661 tgl@sss.pgh.pa.us 3811 :CBC 74 : }
3812 : :
3813 : : /*
3814 : : * targetIsInAllPartitionLists
3815 : : * True if the TargetEntry is listed in the PARTITION BY clause
3816 : : * of every window defined in the query.
3817 : : *
3818 : : * It would be safe to ignore windows not actually used by any window
3819 : : * function, but it's not easy to get that info at this stage; and it's
3820 : : * unlikely to be useful to spend any extra cycles getting it, since
3821 : : * unreferenced window definitions are probably infrequent in practice.
3822 : : */
3823 : : static bool
3579 3824 : 570 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
3825 : : {
3826 : : ListCell *lc;
3827 : :
3828 [ + - + + : 630 : foreach(lc, query->windowClause)
+ + ]
3829 : : {
3830 : 582 : WindowClause *wc = (WindowClause *) lfirst(lc);
3831 : :
3832 [ + + ]: 582 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
3833 : 522 : return false;
3834 : : }
3835 : 48 : return true;
3836 : : }
3837 : :
3838 : : /*
3839 : : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3840 : : *
3841 : : * rinfo is a restriction clause applying to the given subquery (whose RTE
3842 : : * has index rti in the parent query).
3843 : : *
3844 : : * Conditions checked here:
3845 : : *
3846 : : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3847 : : * unclear that it will work correctly: SubLinks will already have been
3848 : : * transformed into SubPlans in the qual, but not in the subquery). Note that
3849 : : * SubLinks that transform to initplans are safe, and will be accepted here
3850 : : * because what we'll see in the qual is just a Param referencing the initplan
3851 : : * output.
3852 : : *
3853 : : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3854 : : * functions.
3855 : : *
3856 : : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
3857 : : * functions that are passed Var nodes, and therefore might reveal values from
3858 : : * the subquery as side effects.
3859 : : *
3860 : : * 4. rinfo's clause must not refer to the whole-row output of the subquery
3861 : : * (since there is no easy way to name that within the subquery itself).
3862 : : *
3863 : : * 5. rinfo's clause must not refer to any subquery output columns that were
3864 : : * found to be unsafe to reference by subquery_is_pushdown_safe().
3865 : : */
3866 : : static pushdown_safe_type
1112 drowley@postgresql.o 3867 : 764 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
3868 : : pushdown_safety_info *safetyInfo)
3869 : : {
394 3870 : 764 : pushdown_safe_type safe = PUSHDOWN_SAFE;
1112 3871 : 764 : Node *qual = (Node *) rinfo->clause;
3872 : : List *vars;
3873 : : ListCell *vl;
3874 : :
3875 : : /* Refuse subselects (point 1) */
7694 tgl@sss.pgh.pa.us 3876 [ + + ]: 764 : if (contain_subplans(qual))
394 drowley@postgresql.o 3877 : 33 : return PUSHDOWN_UNSAFE;
3878 : :
3879 : : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3579 tgl@sss.pgh.pa.us 3880 [ + + + + ]: 1035 : if (safetyInfo->unsafeVolatile &&
1112 drowley@postgresql.o 3881 : 304 : contain_volatile_functions((Node *) rinfo))
394 3882 : 9 : return PUSHDOWN_UNSAFE;
3883 : :
3884 : : /* Refuse leaky quals if told to (point 3) */
3579 tgl@sss.pgh.pa.us 3885 [ + + + + ]: 863 : if (safetyInfo->unsafeLeaky &&
3275 sfrost@snowman.net 3886 : 141 : contain_leaked_vars(qual))
394 drowley@postgresql.o 3887 : 69 : return PUSHDOWN_UNSAFE;
3888 : :
3889 : : /*
3890 : : * Examine all Vars used in clause. Since it's a restriction clause, all
3891 : : * such Vars must refer to subselect output columns ... unless this is
3892 : : * part of a LATERAL subquery, in which case there could be lateral
3893 : : * references.
3894 : : *
3895 : : * By omitting the relevant flags, this also gives us a cheap sanity check
3896 : : * that no aggregates or window functions appear in the qual. Those would
3897 : : * be unsafe to push down, but at least for the moment we could never see
3898 : : * any in a qual anyhow.
3899 : : */
2957 tgl@sss.pgh.pa.us 3900 : 653 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
7694 3901 [ + + + + : 1265 : foreach(vl, vars)
+ + ]
3902 : : {
7559 bruce@momjian.us 3903 : 707 : Var *var = (Var *) lfirst(vl);
3904 : :
3905 : : /*
3906 : : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3907 : : * It's not clear whether a PHV could safely be pushed down, and even
3908 : : * less clear whether such a situation could arise in any cases of
3909 : : * practical interest anyway. So for the moment, just refuse to push
3910 : : * down.
3911 : : */
5654 tgl@sss.pgh.pa.us 3912 [ - + ]: 707 : if (!IsA(var, Var))
3913 : : {
394 drowley@postgresql.o 3914 :UBC 0 : safe = PUSHDOWN_UNSAFE;
5654 tgl@sss.pgh.pa.us 3915 : 0 : break;
3916 : : }
3917 : :
3918 : : /*
3919 : : * Punt if we find any lateral references. It would be safe to push
3920 : : * these down, but we'd have to convert them into outer references,
3921 : : * which subquery_push_qual lacks the infrastructure to do. The case
3922 : : * arises so seldom that it doesn't seem worth working hard on.
3923 : : */
1371 tgl@sss.pgh.pa.us 3924 [ + + ]:CBC 707 : if (var->varno != rti)
3925 : : {
394 drowley@postgresql.o 3926 : 6 : safe = PUSHDOWN_UNSAFE;
1371 tgl@sss.pgh.pa.us 3927 : 6 : break;
3928 : : }
3929 : :
3930 : : /* Subqueries have no system columns */
3966 3931 [ - + ]: 701 : Assert(var->varattno >= 0);
3932 : :
3933 : : /* Check point 4 */
6635 3934 [ - + ]: 701 : if (var->varattno == 0)
3935 : : {
394 drowley@postgresql.o 3936 :UBC 0 : safe = PUSHDOWN_UNSAFE;
6635 tgl@sss.pgh.pa.us 3937 : 0 : break;
3938 : : }
3939 : :
3940 : : /* Check point 5 */
394 drowley@postgresql.o 3941 [ + + ]:CBC 701 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
3942 : : {
3943 [ + + ]: 248 : if (safetyInfo->unsafeFlags[var->varattno] &
3944 : : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
3945 : : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
3946 : : {
3947 : 89 : safe = PUSHDOWN_UNSAFE;
3948 : 89 : break;
3949 : : }
3950 : : else
3951 : : {
3952 : : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
3953 : 159 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
3954 : : /* don't break, we might find another Var that's unsafe */
3955 : : }
3956 : : }
3957 : : }
3958 : :
7259 neilc@samurai.com 3959 : 653 : list_free(vars);
3960 : :
7694 tgl@sss.pgh.pa.us 3961 : 653 : return safe;
3962 : : }
3963 : :
3964 : : /*
3965 : : * subquery_push_qual - push down a qual that we have determined is safe
3966 : : */
3967 : : static void
6889 3968 : 488 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3969 : : {
7899 3970 [ + + ]: 488 : if (subquery->setOperations != NULL)
3971 : : {
3972 : : /* Recurse to push it separately to each component query */
7178 3973 : 25 : recurse_push_qual(subquery->setOperations, subquery,
3974 : : rte, rti, qual);
3975 : : }
3976 : : else
3977 : : {
3978 : : /*
3979 : : * We need to replace Vars in the qual (which must refer to outputs of
3980 : : * the subquery) with copies of the subquery's targetlist expressions.
3981 : : * Note that at this point, any uplevel Vars in the qual should have
3982 : : * been replaced with Params, so they need no work.
3983 : : *
3984 : : * This step also ensures that when we are pushing into a setop tree,
3985 : : * each component query gets its own copy of the qual.
3986 : : */
4175 3987 : 463 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
3988 : : subquery->targetList,
3989 : : REPLACEVARS_REPORT_ERROR, 0,
3990 : : &subquery->hasSubLinks);
3991 : :
3992 : : /*
3993 : : * Now attach the qual to the proper place: normally WHERE, but if the
3994 : : * subquery uses grouping or aggregation, put it in HAVING (since the
3995 : : * qual really refers to the group-result rows).
3996 : : */
3256 andres@anarazel.de 3997 [ + + + - : 463 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
+ - - + ]
6975 tgl@sss.pgh.pa.us 3998 : 81 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
3999 : : else
4000 : 382 : subquery->jointree->quals =
4001 : 382 : make_and_qual(subquery->jointree->quals, qual);
4002 : :
4003 : : /*
4004 : : * We need not change the subquery's hasAggs or hasSubLinks flags,
4005 : : * since we can't be pushing down any aggregates that weren't there
4006 : : * before, and we don't push down subselects at all.
4007 : : */
4008 : : }
7899 4009 : 488 : }
4010 : :
4011 : : /*
4012 : : * Helper routine to recurse through setOperations tree
4013 : : */
4014 : : static void
4015 : 75 : recurse_push_qual(Node *setOp, Query *topquery,
4016 : : RangeTblEntry *rte, Index rti, Node *qual)
4017 : : {
4018 [ + + ]: 75 : if (IsA(setOp, RangeTblRef))
4019 : : {
4020 : 50 : RangeTblRef *rtr = (RangeTblRef *) setOp;
7279 4021 : 50 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4022 : 50 : Query *subquery = subrte->subquery;
4023 : :
7899 4024 [ - + ]: 50 : Assert(subquery != NULL);
6889 4025 : 50 : subquery_push_qual(subquery, rte, rti, qual);
4026 : : }
7899 4027 [ + - ]: 25 : else if (IsA(setOp, SetOperationStmt))
4028 : : {
4029 : 25 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4030 : :
6889 4031 : 25 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4032 : 25 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4033 : : }
4034 : : else
4035 : : {
7569 tgl@sss.pgh.pa.us 4036 [ # # ]:UBC 0 : elog(ERROR, "unrecognized node type: %d",
4037 : : (int) nodeTag(setOp));
4038 : : }
7899 tgl@sss.pgh.pa.us 4039 :CBC 75 : }
4040 : :
4041 : : /*****************************************************************************
4042 : : * SIMPLIFYING SUBQUERY TARGETLISTS
4043 : : *****************************************************************************/
4044 : :
4045 : : /*
4046 : : * remove_unused_subquery_outputs
4047 : : * Remove subquery targetlist items we don't need
4048 : : *
4049 : : * It's possible, even likely, that the upper query does not read all the
4050 : : * output columns of the subquery. We can remove any such outputs that are
4051 : : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4052 : : * affect semantics otherwise (e.g., volatile functions can't be removed).
4053 : : * This is useful not only because we might be able to remove expensive-to-
4054 : : * compute expressions, but because deletion of output columns might allow
4055 : : * optimizations such as join removal to occur within the subquery.
4056 : : *
4057 : : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4058 : : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4059 : : * parameter is modified by the function, so callers must make a copy if they
4060 : : * need to use the passed in Bitmapset after calling this function.
4061 : : *
4062 : : * To avoid affecting column numbering in the targetlist, we don't physically
4063 : : * remove unused tlist entries, but rather replace their expressions with NULL
4064 : : * constants. This is implemented by modifying subquery->targetList.
4065 : : */
4066 : : static void
688 drowley@postgresql.o 4067 : 3788 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4068 : : Bitmapset *extra_used_attrs)
4069 : : {
4070 : : Bitmapset *attrs_used;
4071 : : ListCell *lc;
4072 : :
4073 : : /*
4074 : : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4075 : : * the current callers use the Bitmapset after calling this function.
4076 : : */
4077 : 3788 : attrs_used = extra_used_attrs;
4078 : :
4079 : : /*
4080 : : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4081 : : * could update all the child SELECTs' tlists, but it seems not worth the
4082 : : * trouble presently.
4083 : : */
3594 tgl@sss.pgh.pa.us 4084 [ + + ]: 3788 : if (subquery->setOperations)
4085 : 524 : return;
4086 : :
4087 : : /*
4088 : : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4089 : : * time: all its output columns must be used in the distinctClause.
4090 : : */
4091 [ + + + + ]: 3525 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4092 : 119 : return;
4093 : :
4094 : : /*
4095 : : * Collect a bitmap of all the output column numbers used by the upper
4096 : : * query.
4097 : : *
4098 : : * Add all the attributes needed for joins or final output. Note: we must
4099 : : * look at rel's targetlist, not the attr_needed data, because attr_needed
4100 : : * isn't computed for inheritance child rels, cf set_append_rel_size().
4101 : : * (XXX might be worth changing that sometime.)
4102 : : */
2953 4103 : 3406 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4104 : :
4105 : : /* Add all the attributes used by un-pushed-down restriction clauses. */
3594 4106 [ + + + + : 3744 : foreach(lc, rel->baserestrictinfo)
+ + ]
4107 : : {
4108 : 338 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4109 : :
4110 : 338 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4111 : : }
4112 : :
4113 : : /*
4114 : : * If there's a whole-row reference to the subquery, we can't remove
4115 : : * anything.
4116 : : */
4117 [ + + ]: 3406 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4118 : 142 : return;
4119 : :
4120 : : /*
4121 : : * Run through the tlist and zap entries we don't need. It's okay to
4122 : : * modify the tlist items in-place because set_subquery_pathlist made a
4123 : : * copy of the subquery.
4124 : : */
4125 [ + - + + : 13856 : foreach(lc, subquery->targetList)
+ + ]
4126 : : {
4127 : 10592 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4128 : 10592 : Node *texpr = (Node *) tle->expr;
4129 : :
4130 : : /*
4131 : : * If it has a sortgroupref number, it's used in some sort/group
4132 : : * clause so we'd better not remove it. Also, don't remove any
4133 : : * resjunk columns, since their reason for being has nothing to do
4134 : : * with anybody reading the subquery's output. (It's likely that
4135 : : * resjunk columns in a sub-SELECT would always have ressortgroupref
4136 : : * set, but even if they don't, it seems imprudent to remove them.)
4137 : : */
4138 [ + + - + ]: 10592 : if (tle->ressortgroupref || tle->resjunk)
4139 : 1175 : continue;
4140 : :
4141 : : /*
4142 : : * If it's used by the upper query, we can't remove it.
4143 : : */
4144 [ + + ]: 9417 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4145 : : attrs_used))
4146 : 6341 : continue;
4147 : :
4148 : : /*
4149 : : * If it contains a set-returning function, we can't remove it since
4150 : : * that could change the number of rows returned by the subquery.
4151 : : */
2770 4152 [ + + + + ]: 3542 : if (subquery->hasTargetSRFs &&
4153 : 466 : expression_returns_set(texpr))
3594 4154 : 338 : continue;
4155 : :
4156 : : /*
4157 : : * If it contains volatile functions, we daren't remove it for fear
4158 : : * that the user is expecting their side-effects to happen.
4159 : : */
4160 [ + + ]: 2738 : if (contain_volatile_functions(texpr))
4161 : 12 : continue;
4162 : :
4163 : : /*
4164 : : * OK, we don't need it. Replace the expression with a NULL constant.
4165 : : * Preserve the exposed type of the expression, in case something
4166 : : * looks at the rowtype of the subquery's result.
4167 : : */
4168 : 2726 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4169 : : exprTypmod(texpr),
4170 : : exprCollation(texpr));
4171 : : }
4172 : : }
4173 : :
4174 : : /*
4175 : : * create_partial_bitmap_paths
4176 : : * Build partial bitmap heap path for the relation
4177 : : */
4178 : : void
2594 rhaas@postgresql.org 4179 : 63083 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4180 : : Path *bitmapqual)
4181 : : {
4182 : : int parallel_workers;
4183 : : double pages_fetched;
4184 : :
4185 : : /* Compute heap pages for bitmap heap scan */
4186 : 63083 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4187 : : NULL, NULL);
4188 : :
2263 4189 : 63083 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4190 : : max_parallel_workers_per_gather);
4191 : :
2594 4192 [ + + ]: 63083 : if (parallel_workers <= 0)
4193 : 60996 : return;
4194 : :
4195 : 2087 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4196 : : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4197 : : }
4198 : :
4199 : : /*
4200 : : * Compute the number of parallel workers that should be used to scan a
4201 : : * relation. We compute the parallel workers based on the size of the heap to
4202 : : * be scanned and the size of the index to be scanned, then choose a minimum
4203 : : * of those.
4204 : : *
4205 : : * "heap_pages" is the number of pages from the table that we expect to scan, or
4206 : : * -1 if we don't expect to scan any.
4207 : : *
4208 : : * "index_pages" is the number of pages from the index that we expect to scan, or
4209 : : * -1 if we don't expect to scan any.
4210 : : *
4211 : : * "max_workers" is caller's limit on the number of workers. This typically
4212 : : * comes from a GUC.
4213 : : */
4214 : : int
2263 4215 : 318151 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4216 : : int max_workers)
4217 : : {
2615 4218 : 318151 : int parallel_workers = 0;
4219 : :
4220 : : /*
4221 : : * If the user has set the parallel_workers reloption, use that; otherwise
4222 : : * select a default number of workers.
4223 : : */
2643 4224 [ + + ]: 318151 : if (rel->rel_parallel_workers != -1)
4225 : 945 : parallel_workers = rel->rel_parallel_workers;
4226 : : else
4227 : : {
4228 : : /*
4229 : : * If the number of pages being scanned is insufficient to justify a
4230 : : * parallel scan, just return zero ... unless it's an inheritance
4231 : : * child. In that case, we want to generate a parallel path here
4232 : : * anyway. It might not be worthwhile just for this relation, but
4233 : : * when combined with all of its inheritance siblings it may well pay
4234 : : * off.
4235 : : */
2588 4236 [ + + + + ]: 317206 : if (rel->reloptkind == RELOPT_BASEREL &&
4237 [ + + + + ]: 298237 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
2489 tgl@sss.pgh.pa.us 4238 [ + + ]: 8635 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
2643 rhaas@postgresql.org 4239 : 297883 : return 0;
4240 : :
2588 4241 [ + + ]: 19323 : if (heap_pages >= 0)
4242 : : {
4243 : : int heap_parallel_threshold;
4244 : 18336 : int heap_parallel_workers = 1;
4245 : :
4246 : : /*
4247 : : * Select the number of workers based on the log of the size of
4248 : : * the relation. This probably needs to be a good deal more
4249 : : * sophisticated, but we need something here for now. Note that
4250 : : * the upper limit of the min_parallel_table_scan_size GUC is
4251 : : * chosen to prevent overflow here.
4252 : : */
2615 4253 : 18336 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4254 [ + + ]: 20720 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4255 : : {
4256 : 2384 : heap_parallel_workers++;
4257 : 2384 : heap_parallel_threshold *= 3;
4258 [ - + ]: 2384 : if (heap_parallel_threshold > INT_MAX / 3)
2615 rhaas@postgresql.org 4259 :UBC 0 : break; /* avoid overflow */
4260 : : }
4261 : :
2615 rhaas@postgresql.org 4262 :CBC 18336 : parallel_workers = heap_parallel_workers;
4263 : : }
4264 : :
2588 4265 [ + + ]: 19323 : if (index_pages >= 0)
4266 : : {
4267 : 4752 : int index_parallel_workers = 1;
4268 : : int index_parallel_threshold;
4269 : :
4270 : : /* same calculation as for heap_pages above */
2615 4271 : 4752 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4272 [ + + ]: 4890 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4273 : : {
4274 : 138 : index_parallel_workers++;
4275 : 138 : index_parallel_threshold *= 3;
4276 [ - + ]: 138 : if (index_parallel_threshold > INT_MAX / 3)
2615 rhaas@postgresql.org 4277 :UBC 0 : break; /* avoid overflow */
4278 : : }
4279 : :
2615 rhaas@postgresql.org 4280 [ + + ]:CBC 4752 : if (parallel_workers > 0)
4281 : 3765 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4282 : : else
4283 : 987 : parallel_workers = index_parallel_workers;
4284 : : }
4285 : : }
4286 : :
4287 : : /* In no case use more than caller supplied maximum number of workers */
2263 4288 : 20268 : parallel_workers = Min(parallel_workers, max_workers);
4289 : :
2643 4290 : 20268 : return parallel_workers;
4291 : : }
4292 : :
4293 : : /*
4294 : : * generate_partitionwise_join_paths
4295 : : * Create paths representing partitionwise join for given partitioned
4296 : : * join relation.
4297 : : *
4298 : : * This must not be called until after we are done adding paths for all
4299 : : * child-joins. Otherwise, add_path might delete a path to which some path
4300 : : * generated here has a reference.
4301 : : */
4302 : : void
2249 peter_e@gmx.net 4303 : 90826 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4304 : : {
2382 rhaas@postgresql.org 4305 : 90826 : List *live_children = NIL;
4306 : : int cnt_parts;
4307 : : int num_parts;
4308 : : RelOptInfo **part_rels;
4309 : :
4310 : : /* Handle only join relations here. */
4311 [ + + - + ]: 90826 : if (!IS_JOIN_REL(rel))
2382 rhaas@postgresql.org 4312 :UBC 0 : return;
4313 : :
4314 : : /* We've nothing to do if the relation is not partitioned. */
2260 rhaas@postgresql.org 4315 [ + + + + :CBC 90826 : if (!IS_PARTITIONED_REL(rel))
+ + + - -
+ ]
2382 4316 : 89998 : return;
4317 : :
4318 : : /* The relation should have consider_partitionwise_join set. */
2053 efujita@postgresql.o 4319 [ - + ]: 828 : Assert(rel->consider_partitionwise_join);
4320 : :
4321 : : /* Guard against stack overflow due to overly deep partition hierarchy. */
2382 rhaas@postgresql.org 4322 : 828 : check_stack_depth();
4323 : :
4324 : 828 : num_parts = rel->nparts;
4325 : 828 : part_rels = rel->part_rels;
4326 : :
4327 : : /* Collect non-dummy child-joins. */
4328 [ + + ]: 3103 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4329 : : {
4330 : 2275 : RelOptInfo *child_rel = part_rels[cnt_parts];
4331 : :
4332 : : /* If it's been pruned entirely, it's certainly dummy. */
1842 tgl@sss.pgh.pa.us 4333 [ + + ]: 2275 : if (child_rel == NULL)
4334 : 26 : continue;
4335 : :
4336 : : /* Make partitionwise join paths for this partitioned child-join. */
2249 peter_e@gmx.net 4337 : 2249 : generate_partitionwise_join_paths(root, child_rel);
4338 : :
4339 : : /* If we failed to make any path for this child, we must give up. */
497 tgl@sss.pgh.pa.us 4340 [ - + ]: 2249 : if (child_rel->pathlist == NIL)
4341 : : {
4342 : : /*
4343 : : * Mark the parent joinrel as unpartitioned so that later
4344 : : * functions treat it correctly.
4345 : : */
497 tgl@sss.pgh.pa.us 4346 :LBC (24) : rel->nparts = 0;
4347 : (24) : return;
4348 : : }
4349 : :
4350 : : /* Else, identify the cheapest path for it. */
1865 tgl@sss.pgh.pa.us 4351 :CBC 2249 : set_cheapest(child_rel);
4352 : :
4353 : : /* Dummy children need not be scanned, so ignore those. */
2382 rhaas@postgresql.org 4354 [ - + ]: 2249 : if (IS_DUMMY_REL(child_rel))
2382 rhaas@postgresql.org 4355 :UBC 0 : continue;
4356 : :
4357 : : #ifdef OPTIMIZER_DEBUG
4358 : : pprint(child_rel);
4359 : : #endif
4360 : :
2382 rhaas@postgresql.org 4361 :CBC 2249 : live_children = lappend(live_children, child_rel);
4362 : : }
4363 : :
4364 : : /* If all child-joins are dummy, parent join is also dummy. */
4365 [ - + ]: 828 : if (!live_children)
4366 : : {
2382 rhaas@postgresql.org 4367 :UBC 0 : mark_dummy_rel(rel);
4368 : 0 : return;
4369 : : }
4370 : :
4371 : : /* Build additional paths for this rel from child-join paths. */
2382 rhaas@postgresql.org 4372 :CBC 828 : add_paths_to_append_rel(root, rel, live_children);
4373 : 828 : list_free(live_children);
4374 : : }
|
