Age Owner 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) 1996-2023, 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/inherit.h"
39 : #include "optimizer/optimizer.h"
40 : #include "optimizer/pathnode.h"
41 : #include "optimizer/paths.h"
42 : #include "optimizer/plancat.h"
43 : #include "optimizer/planner.h"
44 : #include "optimizer/restrictinfo.h"
45 : #include "optimizer/tlist.h"
46 : #include "parser/parse_clause.h"
47 : #include "parser/parsetree.h"
48 : #include "partitioning/partbounds.h"
49 : #include "partitioning/partprune.h"
50 : #include "port/pg_bitutils.h"
51 : #include "rewrite/rewriteManip.h"
52 : #include "utils/lsyscache.h"
53 :
54 :
55 : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
56 : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
57 : #define UNSAFE_HAS_SET_FUNC (1 << 1)
58 : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
59 : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
60 : #define UNSAFE_TYPE_MISMATCH (1 << 4)
61 :
62 : /* results of subquery_is_pushdown_safe */
63 : typedef struct pushdown_safety_info
64 : {
65 : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
66 : * column is unsafe for qual pushdown, or 0 if
67 : * no reason. */
68 : bool unsafeVolatile; /* don't push down volatile quals */
69 : bool unsafeLeaky; /* don't push down leaky quals */
70 : } pushdown_safety_info;
71 :
72 : /* Return type for qual_is_pushdown_safe */
73 : typedef enum pushdown_safe_type
74 : {
75 : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
76 : PUSHDOWN_SAFE, /* safe to push qual into subquery */
77 : PUSHDOWN_WINDOWCLAUSE_RUNCOND /* unsafe, but may work as WindowClause
78 : * run condition */
79 : } pushdown_safe_type;
80 :
81 : /* These parameters are set by GUC */
82 : bool enable_geqo = false; /* just in case GUC doesn't set it */
83 : int geqo_threshold;
84 : int min_parallel_table_scan_size;
85 : int min_parallel_index_scan_size;
86 :
87 : /* Hook for plugins to get control in set_rel_pathlist() */
88 : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
89 :
90 : /* Hook for plugins to replace standard_join_search() */
91 : join_search_hook_type join_search_hook = NULL;
92 :
93 :
94 : static void set_base_rel_consider_startup(PlannerInfo *root);
95 : static void set_base_rel_sizes(PlannerInfo *root);
96 : static void set_base_rel_pathlists(PlannerInfo *root);
97 : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
98 : Index rti, RangeTblEntry *rte);
99 : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
100 : Index rti, RangeTblEntry *rte);
101 : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
102 : RangeTblEntry *rte);
103 : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
104 : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
105 : RangeTblEntry *rte);
106 : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
107 : RangeTblEntry *rte);
108 : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
109 : RangeTblEntry *rte);
110 : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
111 : RangeTblEntry *rte);
112 : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
113 : RangeTblEntry *rte);
114 : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
115 : RangeTblEntry *rte);
116 : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
117 : Index rti, RangeTblEntry *rte);
118 : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
119 : Index rti, RangeTblEntry *rte);
120 : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
121 : List *live_childrels,
122 : List *all_child_pathkeys);
123 : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
124 : RelOptInfo *rel,
125 : Relids required_outer);
126 : static void accumulate_append_subpath(Path *path,
127 : List **subpaths,
128 : List **special_subpaths);
129 : static Path *get_singleton_append_subpath(Path *path);
130 : static void set_dummy_rel_pathlist(RelOptInfo *rel);
131 : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
132 : Index rti, RangeTblEntry *rte);
133 : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
134 : RangeTblEntry *rte);
135 : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
136 : RangeTblEntry *rte);
137 : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
138 : RangeTblEntry *rte);
139 : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
140 : RangeTblEntry *rte);
141 : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
142 : RangeTblEntry *rte);
143 : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
144 : RangeTblEntry *rte);
145 : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
146 : RangeTblEntry *rte);
147 : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
148 : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
149 : pushdown_safety_info *safetyInfo);
150 : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
151 : pushdown_safety_info *safetyInfo);
152 : static void check_output_expressions(Query *subquery,
153 : pushdown_safety_info *safetyInfo);
154 : static void compare_tlist_datatypes(List *tlist, List *colTypes,
155 : pushdown_safety_info *safetyInfo);
156 : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
157 : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
158 : RestrictInfo *rinfo,
159 : pushdown_safety_info *safetyInfo);
160 : static void subquery_push_qual(Query *subquery,
161 : RangeTblEntry *rte, Index rti, Node *qual);
162 : static void recurse_push_qual(Node *setOp, Query *topquery,
163 : RangeTblEntry *rte, Index rti, Node *qual);
164 : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
165 : Bitmapset *extra_used_attrs);
166 :
167 :
168 : /*
169 : * make_one_rel
170 : * Finds all possible access paths for executing a query, returning a
171 : * single rel that represents the join of all base rels in the query.
172 : */
173 : RelOptInfo *
6319 tgl 174 GIC 128142 : make_one_rel(PlannerInfo *root, List *joinlist)
9770 scrappy 175 ECB : {
176 : RelOptInfo *rel;
177 : Index rti;
178 : double total_pages;
179 :
180 : /* Mark base rels as to whether we care about fast-start plans */
2867 tgl 181 GIC 128142 : set_base_rel_consider_startup(root);
2867 tgl 182 ECB :
8654 183 : /*
184 : * Compute size estimates and consider_parallel flags for each base rel.
185 : */
4090 tgl 186 GIC 128142 : set_base_rel_sizes(root);
187 :
1614 tgl 188 ECB : /*
189 : * We should now have size estimates for every actual table involved in
190 : * the query, and we also know which if any have been deleted from the
191 : * query by join removal, pruned by partition pruning, or eliminated by
192 : * constraint exclusion. So we can now compute total_table_pages.
193 : *
194 : * Note that appendrels are not double-counted here, even though we don't
195 : * bother to distinguish RelOptInfos for appendrel parents, because the
196 : * parents will have pages = 0.
197 : *
198 : * XXX if a table is self-joined, we will count it once per appearance,
199 : * which perhaps is the wrong thing ... but that's not completely clear,
200 : * and detecting self-joins here is difficult, so ignore it for now.
201 : */
1614 tgl 202 GIC 128130 : total_pages = 0;
203 384559 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
1614 tgl 204 ECB : {
1614 tgl 205 GIC 256429 : RelOptInfo *brel = root->simple_rel_array[rti];
206 :
207 : /* there may be empty slots corresponding to non-baserel RTEs */
208 256429 : if (brel == NULL)
209 60721 : continue;
1614 tgl 210 ECB :
1614 tgl 211 GIC 195708 : Assert(brel->relid == rti); /* sanity check on array */
212 :
213 195708 : if (IS_DUMMY_REL(brel))
214 482 : continue;
1614 tgl 215 ECB :
1614 tgl 216 GIC 195226 : if (IS_SIMPLE_REL(brel))
1614 tgl 217 CBC 195226 : total_pages += (double) brel->pages;
218 : }
1614 tgl 219 GIC 128130 : root->total_table_pages = total_pages;
220 :
221 : /*
222 : * Generate access paths for each base rel.
223 : */
8183 224 128130 : set_base_rel_pathlists(root);
225 :
226 : /*
227 : * Generate access paths for the entire join tree.
228 : */
6319 229 128130 : rel = make_rel_from_joinlist(root, joinlist);
230 :
8227 tgl 231 ECB : /*
232 : * The result should join all and only the query's base + outer-join rels.
233 : */
69 tgl 234 GNC 128130 : Assert(bms_equal(rel->relids, root->all_query_rels));
235 :
4090 tgl 236 GIC 128130 : return rel;
237 : }
238 :
239 : /*
240 : * set_base_rel_consider_startup
241 : * Set the consider_[param_]startup flags for each base-relation entry.
242 : *
243 : * For the moment, we only deal with consider_param_startup here; because the
244 : * logic for consider_startup is pretty trivial and is the same for every base
245 : * relation, we just let build_simple_rel() initialize that flag correctly to
246 : * start with. If that logic ever gets more complicated it would probably
247 : * be better to move it here.
2867 tgl 248 ECB : */
249 : static void
2867 tgl 250 CBC 128142 : set_base_rel_consider_startup(PlannerInfo *root)
251 : {
252 : /*
2867 tgl 253 ECB : * Since parameterized paths can only be used on the inside of a nestloop
254 : * join plan, there is usually little value in considering fast-start
255 : * plans for them. However, for relations that are on the RHS of a SEMI
256 : * or ANTI join, a fast-start plan can be useful because we're only going
257 : * to care about fetching one tuple anyway.
258 : *
259 : * To minimize growth of planning time, we currently restrict this to
260 : * cases where the RHS is a single base relation, not a join; there is no
261 : * provision for consider_param_startup to get set at all on joinrels.
262 : * Also we don't worry about appendrels. costsize.c's costing rules for
263 : * nestloop semi/antijoins don't consider such cases either.
264 : */
265 : ListCell *lc;
266 :
2867 tgl 267 GIC 144307 : foreach(lc, root->join_info_list)
268 : {
269 16165 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
270 : int varno;
271 :
272 18921 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
273 2756 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
2867 tgl 274 ECB : {
2867 tgl 275 GIC 2693 : RelOptInfo *rel = find_base_rel(root, varno);
276 :
277 2693 : rel->consider_param_startup = true;
2867 tgl 278 ECB : }
279 : }
2867 tgl 280 CBC 128142 : }
281 :
282 : /*
283 : * set_base_rel_sizes
4090 tgl 284 ECB : * Set the size estimates (rows and widths) for each base-relation entry.
2495 rhaas 285 : * Also determine whether to consider parallel paths for base relations.
286 : *
4090 tgl 287 : * We do this in a separate pass over the base rels so that rowcount
288 : * estimates are available for parameterized path generation, and also so
289 : * that each rel's consider_parallel flag is set correctly before we begin to
2706 rhaas 290 : * generate paths.
4090 tgl 291 : */
292 : static void
4090 tgl 293 CBC 128142 : set_base_rel_sizes(PlannerInfo *root)
294 : {
295 : Index rti;
296 :
4090 tgl 297 GIC 384571 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
298 : {
299 256441 : RelOptInfo *rel = root->simple_rel_array[rti];
300 : RangeTblEntry *rte;
301 :
302 : /* there may be empty slots corresponding to non-baserel RTEs */
4090 tgl 303 CBC 256441 : if (rel == NULL)
304 60721 : continue;
305 :
2118 306 195720 : Assert(rel->relid == rti); /* sanity check on array */
307 :
4090 tgl 308 ECB : /* ignore RTEs that are "other rels" */
4090 tgl 309 GIC 195720 : if (rel->reloptkind != RELOPT_BASEREL)
310 19589 : continue;
311 :
2706 rhaas 312 176131 : rte = root->simple_rte_array[rti];
313 :
314 : /*
315 : * If parallelism is allowable for this query in general, see whether
316 : * it's allowable for this rel in particular. We have to do this
2471 tgl 317 ECB : * before set_rel_size(), because (a) if this rel is an inheritance
318 : * parent, set_append_rel_size() will use and perhaps change the rel's
319 : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
320 : * goes ahead and makes paths immediately.
2706 rhaas 321 : */
2706 rhaas 322 GIC 176131 : if (root->glob->parallelModeOK)
2706 rhaas 323 CBC 129841 : set_rel_consider_parallel(root, rel, rte);
324 :
2706 rhaas 325 GIC 176131 : set_rel_size(root, rel, rti, rte);
4090 tgl 326 ECB : }
9770 scrappy 327 CBC 128130 : }
328 :
9345 bruce 329 ECB : /*
330 : * set_base_rel_pathlists
331 : * Finds all paths available for scanning each base-relation entry.
8462 tgl 332 : * Sequential scan and any available indices are considered.
333 : * Each useful path is attached to its relation's 'pathlist' field.
334 : */
9770 scrappy 335 : static void
6517 tgl 336 GIC 128130 : set_base_rel_pathlists(PlannerInfo *root)
9770 scrappy 337 ECB : {
338 : Index rti;
339 :
6277 tgl 340 GIC 384559 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
341 : {
342 256429 : RelOptInfo *rel = root->simple_rel_array[rti];
343 :
6516 tgl 344 ECB : /* there may be empty slots corresponding to non-baserel RTEs */
6516 tgl 345 GIC 256429 : if (rel == NULL)
346 60721 : continue;
6516 tgl 347 ECB :
2118 tgl 348 CBC 195708 : Assert(rel->relid == rti); /* sanity check on array */
349 :
350 : /* ignore RTEs that are "other rels" */
6516 tgl 351 GIC 195708 : if (rel->reloptkind != RELOPT_BASEREL)
352 19589 : continue;
353 :
5832 354 176119 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
355 : }
6274 356 128130 : }
357 :
358 : /*
359 : * set_rel_size
360 : * Set size estimates for a base relation
6274 tgl 361 ECB : */
362 : static void
4090 tgl 363 CBC 195621 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
364 : Index rti, RangeTblEntry *rte)
365 : {
4215 366 371752 : if (rel->reloptkind == RELOPT_BASEREL &&
4215 tgl 367 GIC 176131 : relation_excluded_by_constraints(root, rel, rte))
368 : {
369 : /*
4215 tgl 370 ECB : * We proved we don't need to scan the rel via constraint exclusion,
371 : * so set up a single dummy path for it. Here we only check this for
372 : * regular baserels; if it's an otherrel, CE was already checked in
2907 373 : * set_append_rel_size().
374 : *
375 : * In this case, we go ahead and set up the relation's path right away
4090 376 : * instead of leaving it for set_rel_pathlist to do. This is because
377 : * we don't have a convention for marking a rel as dummy except by
378 : * assigning a dummy path to it.
379 : */
4215 tgl 380 GIC 168 : set_dummy_rel_pathlist(rel);
381 : }
382 195453 : else if (rte->inh)
383 : {
384 : /* It's an "append relation", process accordingly */
4090 tgl 385 CBC 9313 : set_append_rel_size(root, rel, rti, rte);
386 : }
6274 tgl 387 ECB : else
388 : {
4429 tgl 389 GIC 186140 : switch (rel->rtekind)
4431 tgl 390 ECB : {
4429 tgl 391 GIC 158608 : case RTE_RELATION:
392 158608 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
393 : {
394 : /* Foreign table */
4090 tgl 395 CBC 1099 : set_foreign_size(root, rel, rte);
396 : }
2210 rhaas 397 157509 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
2210 rhaas 398 ECB : {
399 : /*
400 : * We could get here if asked to scan a partitioned table
401 : * with ONLY. In that case we shouldn't scan any of the
402 : * partitions, so mark it as a dummy rel.
403 : */
2210 rhaas 404 GIC 20 : set_dummy_rel_pathlist(rel);
2210 rhaas 405 ECB : }
2886 simon 406 CBC 157489 : else if (rte->tablesample != NULL)
2886 simon 407 ECB : {
408 : /* Sampled relation */
2886 simon 409 CBC 126 : set_tablesample_rel_size(root, rel, rte);
2886 simon 410 ECB : }
4429 tgl 411 : else
412 : {
413 : /* Plain relation */
4090 tgl 414 CBC 157363 : set_plain_rel_size(root, rel, rte);
4429 tgl 415 ECB : }
4429 tgl 416 CBC 158596 : break;
4429 tgl 417 GIC 3695 : case RTE_SUBQUERY:
418 :
419 : /*
420 : * Subqueries don't support making a choice between
421 : * parameterized and unparameterized paths, so just go ahead
422 : * and build their paths immediately.
4090 tgl 423 ECB : */
4429 tgl 424 CBC 3695 : set_subquery_pathlist(root, rel, rti, rte);
4429 tgl 425 GIC 3695 : break;
4429 tgl 426 CBC 17699 : case RTE_FUNCTION:
4090 427 17699 : set_function_size_estimates(root, rel);
4429 428 17699 : break;
2223 alvherre 429 GIC 108 : case RTE_TABLEFUNC:
2223 alvherre 430 CBC 108 : set_tablefunc_size_estimates(root, rel);
431 108 : break;
4429 tgl 432 3553 : case RTE_VALUES:
4090 tgl 433 GIC 3553 : set_values_size_estimates(root, rel);
4429 tgl 434 CBC 3553 : break;
435 1597 : case RTE_CTE:
3955 bruce 436 EUB :
4090 tgl 437 : /*
438 : * CTEs don't support making a choice between parameterized
439 : * and unparameterized paths, so just go ahead and build their
440 : * paths immediately.
441 : */
4429 tgl 442 GIC 1597 : if (rte->self_reference)
443 357 : set_worktable_pathlist(root, rel, rte);
444 : else
4429 tgl 445 CBC 1240 : set_cte_pathlist(root, rel, rte);
446 1597 : break;
2200 kgrittn 447 GIC 219 : case RTE_NAMEDTUPLESTORE:
448 : /* Might as well just build the path immediately */
449 219 : set_namedtuplestore_pathlist(root, rel, rte);
450 219 : break;
1532 tgl 451 661 : case RTE_RESULT:
452 : /* Might as well just build the path immediately */
1532 tgl 453 CBC 661 : set_result_pathlist(root, rel, rte);
1532 tgl 454 GIC 661 : break;
4429 tgl 455 UIC 0 : default:
4429 tgl 456 LBC 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
457 : break;
458 : }
459 : }
2814 tgl 460 ECB :
461 : /*
462 : * We insist that all non-dummy rels have a nonzero rowcount estimate.
463 : */
2814 tgl 464 GIC 195609 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
4090 465 195609 : }
466 :
4090 tgl 467 ECB : /*
468 : * set_rel_pathlist
469 : * Build access paths for a base relation
470 : */
471 : static void
4090 tgl 472 GIC 195621 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
4090 tgl 473 ECB : Index rti, RangeTblEntry *rte)
474 : {
4090 tgl 475 CBC 195621 : if (IS_DUMMY_REL(rel))
476 : {
477 : /* We already proved the relation empty, so nothing more to do */
4090 tgl 478 ECB : }
4090 tgl 479 GIC 195208 : else if (rte->inh)
480 : {
481 : /* It's an "append relation", process accordingly */
482 9154 : set_append_rel_pathlist(root, rel, rti, rte);
4090 tgl 483 ECB : }
484 : else
485 : {
4090 tgl 486 CBC 186054 : switch (rel->rtekind)
487 : {
488 158576 : case RTE_RELATION:
489 158576 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
490 : {
4090 tgl 491 ECB : /* Foreign table */
4090 tgl 492 CBC 1099 : set_foreign_pathlist(root, rel, rte);
4090 tgl 493 ECB : }
2886 simon 494 GIC 157477 : else if (rte->tablesample != NULL)
2886 simon 495 ECB : {
2815 tgl 496 : /* Sampled relation */
2886 simon 497 CBC 126 : set_tablesample_rel_pathlist(root, rel, rte);
498 : }
4090 tgl 499 ECB : else
500 : {
501 : /* Plain relation */
4090 tgl 502 GIC 157351 : set_plain_rel_pathlist(root, rel, rte);
4090 tgl 503 ECB : }
4090 tgl 504 CBC 158576 : break;
4090 tgl 505 GIC 3641 : case RTE_SUBQUERY:
4090 tgl 506 ECB : /* Subquery --- fully handled during set_rel_size */
4090 tgl 507 CBC 3641 : break;
4090 tgl 508 GIC 17699 : case RTE_FUNCTION:
4090 tgl 509 ECB : /* RangeFunction */
4090 tgl 510 GBC 17699 : set_function_pathlist(root, rel, rte);
511 17699 : break;
2223 alvherre 512 GIC 108 : case RTE_TABLEFUNC:
513 : /* Table Function */
514 108 : set_tablefunc_pathlist(root, rel, rte);
515 108 : break;
4090 tgl 516 3553 : case RTE_VALUES:
517 : /* Values list */
518 3553 : set_values_pathlist(root, rel, rte);
519 3553 : break;
520 1597 : case RTE_CTE:
521 : /* CTE reference --- fully handled during set_rel_size */
4090 tgl 522 CBC 1597 : break;
2200 kgrittn 523 GBC 219 : case RTE_NAMEDTUPLESTORE:
524 : /* tuplestore reference --- fully handled during set_rel_size */
2200 kgrittn 525 GIC 219 : break;
1532 tgl 526 661 : case RTE_RESULT:
527 : /* simple Result --- fully handled during set_rel_size */
528 661 : break;
4090 tgl 529 UIC 0 : default:
530 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
531 : break;
532 : }
533 : }
534 :
535 : /*
536 : * Allow a plugin to editorialize on the set of Paths for this base
537 : * relation. It could add new paths (such as CustomPaths) by calling
538 : * add_path(), or add_partial_path() if parallel aware. It could also
1520 tgl 539 ECB : * delete or modify paths added by the core code.
540 : */
1520 tgl 541 CBC 195621 : if (set_rel_pathlist_hook)
1520 tgl 542 UIC 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
543 :
2535 tgl 544 ECB : /*
545 : * If this is a baserel, we should normally consider gathering any partial
546 : * paths we may have created for it. We have to do this after calling the
547 : * set_rel_pathlist_hook, else it cannot add partial paths to be included
548 : * here.
1854 rhaas 549 : *
550 : * However, if this is an inheritance child, skip it. Otherwise, we could
551 : * end up with a very large number of gather nodes, each trying to grab
552 : * its own pool of workers. Instead, we'll consider gathering partial
553 : * paths for the parent appendrel.
554 : *
555 : * Also, if this is the topmost scan/join rel, we postpone gathering until
253 tgl 556 : * the final scan/join targetlist is available (see grouping_planner).
557 : */
1854 rhaas 558 GIC 195621 : if (rel->reloptkind == RELOPT_BASEREL &&
69 tgl 559 GNC 176119 : !bms_equal(rel->relids, root->all_query_rels))
1097 tomas.vondra 560 GIC 82284 : generate_useful_gather_paths(root, rel, false);
561 :
3061 tgl 562 ECB : /* Now find the cheapest of the paths for this rel */
3061 tgl 563 GIC 195621 : set_cheapest(rel);
564 :
7843 tgl 565 ECB : #ifdef OPTIMIZER_DEBUG
6274 566 : debug_print_rel(root, rel);
567 : #endif
8183 tgl 568 GIC 195621 : }
569 :
570 : /*
571 : * set_plain_rel_size
572 : * Set size estimates for a plain relation (no subquery, no inheritance)
8183 tgl 573 ECB : */
574 : static void
4090 tgl 575 GIC 157363 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
576 : {
577 : /*
578 : * Test any partial indexes of rel for applicability. We must do this
579 : * first since partial unique indexes can affect size estimates.
5166 tgl 580 ECB : */
2565 tgl 581 GIC 157363 : check_index_predicates(root, rel);
582 :
5832 tgl 583 ECB : /* Mark rel with estimated output rows, width, etc */
5832 tgl 584 GIC 157363 : set_baserel_size_estimates(root, rel);
4090 585 157351 : }
5832 tgl 586 ECB :
587 : /*
588 : * If this relation could possibly be scanned from within a worker, then set
2471 589 : * its consider_parallel flag.
590 : */
2706 rhaas 591 : static void
2706 rhaas 592 GIC 143695 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
593 : RangeTblEntry *rte)
594 : {
595 : /*
596 : * The flag has previously been initialized to false, so we can just
597 : * return if it becomes clear that we can't safely set it.
598 : */
2471 tgl 599 143695 : Assert(!rel->consider_parallel);
600 :
601 : /* Don't call this if parallelism is disallowed for the entire query. */
2706 rhaas 602 143695 : Assert(root->glob->parallelModeOK);
2706 rhaas 603 ECB :
2471 tgl 604 : /* This should only be called for baserels and appendrel children. */
2197 rhaas 605 GIC 143695 : Assert(IS_SIMPLE_REL(rel));
606 :
607 : /* Assorted checks based on rtekind. */
2706 608 143695 : switch (rte->rtekind)
609 : {
2706 rhaas 610 CBC 128265 : case RTE_RELATION:
611 :
2706 rhaas 612 ECB : /*
613 : * Currently, parallel workers can't access the leader's temporary
1158 akapila 614 : * tables. We could possibly relax this if we wrote all of its
2706 rhaas 615 : * local buffers at the start of the query and made no changes
616 : * thereafter (maybe we could allow hint bit changes), and if we
617 : * taught the workers to read them. Writing a large number of
618 : * temporary buffers could be expensive, though, and we don't have
619 : * the rest of the necessary infrastructure right now anyway. So
620 : * for now, bail out if we see a temporary table.
621 : */
2706 rhaas 622 GIC 128265 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
623 3630 : return;
624 :
625 : /*
626 : * Table sampling can be pushed down to workers if the sample
627 : * function and its arguments are safe.
2706 rhaas 628 ECB : */
2706 rhaas 629 GIC 124635 : if (rte->tablesample != NULL)
2706 rhaas 630 ECB : {
2322 tgl 631 CBC 126 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
2706 rhaas 632 ECB :
2706 rhaas 633 CBC 126 : if (proparallel != PROPARALLEL_SAFE)
2706 rhaas 634 GBC 18 : return;
2424 tgl 635 GIC 108 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
2706 rhaas 636 6 : return;
637 : }
638 :
639 : /*
640 : * Ask FDWs whether they can support performing a ForeignScan
641 : * within a worker. Most often, the answer will be no. For
642 : * example, if the nature of the FDW is such that it opens a TCP
2599 rhaas 643 ECB : * connection with a remote server, each parallel worker would end
644 : * up with a separate connection, and these connections might not
645 : * be appropriately coordinated between workers and the leader.
646 : */
2599 rhaas 647 GIC 124611 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
648 : {
649 687 : Assert(rel->fdwroutine);
650 687 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
651 655 : return;
652 32 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
2599 rhaas 653 UIC 0 : return;
654 : }
655 :
656 : /*
657 : * There are additional considerations for appendrels, which we'll
658 : * deal with in set_append_rel_size and set_append_rel_pathlist.
659 : * For now, just set consider_parallel based on the rel's own
660 : * quals and targetlist.
661 : */
2706 rhaas 662 GIC 123956 : break;
663 :
664 2718 : case RTE_SUBQUERY:
665 :
666 : /*
667 : * There's no intrinsic problem with scanning a subquery-in-FROM
2471 tgl 668 ECB : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
669 : * If the subquery doesn't happen to have any parallel-safe paths,
670 : * then flagging it as consider_parallel won't change anything,
671 : * but that's true for plain tables, too. We must set
672 : * consider_parallel based on the rel's own quals and targetlist,
673 : * so that if a subquery path is parallel-safe but the quals and
674 : * projection we're sticking onto it are not, we correctly mark
2471 tgl 675 EUB : * the SubqueryScanPath as not parallel-safe. (Note that
676 : * set_subquery_pathlist() might push some of these quals down
677 : * into the subquery itself, but that doesn't change anything.)
678 : *
679 : * We can't push sub-select containing LIMIT/OFFSET to workers as
1668 akapila 680 ECB : * there is no guarantee that the row order will be fully
681 : * deterministic, and applying LIMIT/OFFSET will lead to
682 : * inconsistent results at the top-level. (In some cases, where
683 : * the result is ordered, we could relax this restriction. But it
684 : * doesn't currently seem worth expending extra effort to do so.)
685 : */
686 : {
1668 akapila 687 GIC 2718 : Query *subquery = castNode(Query, rte->subquery);
1668 akapila 688 ECB :
1668 akapila 689 GIC 2718 : if (limit_needed(subquery))
1668 akapila 690 CBC 196 : return;
691 : }
2471 tgl 692 2522 : break;
2706 rhaas 693 ECB :
2706 rhaas 694 LBC 0 : case RTE_JOIN:
695 : /* Shouldn't happen; we're only considering baserels here. */
696 0 : Assert(false);
697 : return;
698 :
2706 rhaas 699 GIC 9250 : case RTE_FUNCTION:
700 : /* Check for parallel-restricted functions. */
2424 tgl 701 9250 : if (!is_parallel_safe(root, (Node *) rte->functions))
2706 rhaas 702 4608 : return;
703 4642 : break;
704 :
2223 alvherre 705 CBC 108 : case RTE_TABLEFUNC:
706 : /* not parallel safe */
707 108 : return;
708 :
2706 rhaas 709 GIC 1312 : case RTE_VALUES:
710 : /* Check for parallel-restricted functions. */
2424 tgl 711 1312 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
712 3 : return;
2706 rhaas 713 CBC 1309 : break;
714 :
715 1300 : case RTE_CTE:
716 :
2706 rhaas 717 ECB : /*
718 : * CTE tuplestores aren't shared among parallel workers, so we
719 : * force all CTE scans to happen in the leader. Also, populating
720 : * the CTE would require executing a subplan that's not available
721 : * in the worker, might be parallel-restricted, and must get
722 : * executed only once.
723 : */
2706 rhaas 724 GIC 1300 : return;
725 :
2200 kgrittn 726 205 : case RTE_NAMEDTUPLESTORE:
727 :
728 : /*
2200 kgrittn 729 ECB : * tuplestore cannot be shared, at least without more
730 : * infrastructure to support that.
731 : */
2200 kgrittn 732 GIC 205 : return;
733 :
1532 tgl 734 537 : case RTE_RESULT:
735 : /* RESULT RTEs, in themselves, are no problem. */
1532 tgl 736 CBC 537 : break;
2706 rhaas 737 ECB : }
738 :
739 : /*
2495 740 : * If there's anything in baserestrictinfo that's parallel-restricted, we
741 : * give up on parallelizing access to this relation. We could consider
742 : * instead postponing application of the restricted quals until we're
743 : * above all the parallelism in the plan tree, but it's not clear that
744 : * that would be a win in very many cases, and it might be tricky to make
745 : * outer join clauses work correctly. It would likely break equivalence
746 : * classes, too.
747 : */
2424 tgl 748 CBC 132966 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
2706 rhaas 749 GIC 9304 : return;
750 :
751 : /*
752 : * Likewise, if the relation's outputs are not parallel-safe, give up.
753 : * (Usually, they're just Vars, but sometimes they're not.)
754 : */
2424 tgl 755 123662 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
2495 rhaas 756 9 : return;
2495 rhaas 757 ECB :
758 : /* We have a winner. */
2706 rhaas 759 GIC 123653 : rel->consider_parallel = true;
2706 rhaas 760 ECB : }
761 :
762 : /*
4090 tgl 763 : * set_plain_rel_pathlist
764 : * Build access paths for a plain relation (no subquery, no inheritance)
765 : */
766 : static void
4090 tgl 767 CBC 157351 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
768 : {
769 : Relids required_outer;
3878 tgl 770 ECB :
771 : /*
772 : * We don't support pushing join clauses into the quals of a seqscan, but
773 : * it could still have required parameterization due to LATERAL refs in
774 : * its tlist.
775 : */
3878 tgl 776 GIC 157351 : required_outer = rel->lateral_relids;
777 :
8183 tgl 778 ECB : /* Consider sequential scan */
2706 rhaas 779 GIC 157351 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
780 :
781 : /* If appropriate, consider parallel sequential scan */
2636 rhaas 782 CBC 157351 : if (rel->consider_parallel && required_outer == NULL)
2535 tgl 783 GIC 108759 : create_plain_partial_paths(root, rel);
784 :
785 : /* Consider index scans */
6558 tgl 786 CBC 157351 : create_index_paths(root, rel);
6558 tgl 787 ECB :
788 : /* Consider TID scans */
8183 tgl 789 GIC 157351 : create_tidscan_paths(root, rel);
8183 tgl 790 CBC 157351 : }
791 :
792 : /*
793 : * create_plain_partial_paths
794 : * Build partial access paths for parallel scan of a plain relation
795 : */
796 : static void
2535 tgl 797 GIC 108759 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
2636 rhaas 798 ECB : {
799 : int parallel_workers;
800 :
1892 rhaas 801 GIC 108759 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
802 : max_parallel_workers_per_gather);
803 :
804 : /* If any limit was set to zero, the user doesn't want a parallel scan. */
2495 tgl 805 108759 : if (parallel_workers <= 0)
806 96746 : return;
807 :
808 : /* Add an unordered partial path based on a parallel sequential scan. */
2495 rhaas 809 CBC 12013 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
810 : }
811 :
812 : /*
813 : * set_tablesample_rel_size
814 : * Set size estimates for a sampled relation
815 : */
2886 simon 816 ECB : static void
2886 simon 817 CBC 126 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
818 : {
2815 tgl 819 GIC 126 : TableSampleClause *tsc = rte->tablesample;
820 : TsmRoutine *tsm;
821 : BlockNumber pages;
822 : double tuples;
823 :
824 : /*
825 : * Test any partial indexes of rel for applicability. We must do this
2815 tgl 826 ECB : * first since partial unique indexes can affect size estimates.
827 : */
2565 tgl 828 GIC 126 : check_index_predicates(root, rel);
829 :
2815 tgl 830 ECB : /*
831 : * Call the sampling method's estimation function to estimate the number
832 : * of pages it will read and the number of tuples it will return. (Note:
833 : * we assume the function returns sane values.)
834 : */
2815 tgl 835 GIC 126 : tsm = GetTsmRoutine(tsc->tsmhandler);
836 126 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
837 : &pages, &tuples);
2815 tgl 838 ECB :
839 : /*
840 : * For the moment, because we will only consider a SampleScan path for the
841 : * rel, it's okay to just overwrite the pages and tuples estimates for the
842 : * whole relation. If we ever consider multiple path types for sampled
843 : * rels, we'll need more complication.
844 : */
2815 tgl 845 GIC 126 : rel->pages = pages;
846 126 : rel->tuples = tuples;
847 :
2886 simon 848 ECB : /* Mark rel with estimated output rows, width, etc */
2886 simon 849 GIC 126 : set_baserel_size_estimates(root, rel);
850 126 : }
2886 simon 851 ECB :
852 : /*
853 : * set_tablesample_rel_pathlist
854 : * Build access paths for a sampled relation
855 : */
856 : static void
2886 simon 857 GIC 126 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
858 : {
859 : Relids required_outer;
860 : Path *path;
861 :
862 : /*
863 : * We don't support pushing join clauses into the quals of a samplescan,
864 : * but it could still have required parameterization due to LATERAL refs
865 : * in its tlist or TABLESAMPLE arguments.
866 : */
867 126 : required_outer = rel->lateral_relids;
868 :
2815 tgl 869 ECB : /* Consider sampled scan */
2886 simon 870 CBC 126 : path = create_samplescan_path(root, rel, required_outer);
2815 tgl 871 ECB :
872 : /*
873 : * If the sampling method does not support repeatable scans, we must avoid
874 : * plans that would scan the rel multiple times. Ideally, we'd simply
875 : * avoid putting the rel on the inside of a nestloop join; but adding such
876 : * a consideration to the planner seems like a great deal of complication
877 : * to support an uncommon usage of second-rate sampling methods. Instead,
878 : * if there is a risk that the query might perform an unsafe join, just
879 : * wrap the SampleScan in a Materialize node. We can check for joins by
880 : * counting the membership of all_query_rels (note that this correctly
881 : * counts inheritance trees as single rels). If we're inside a subquery,
882 : * we can't easily check whether a join might occur in the outer query, so
883 : * just assume one is possible.
884 : *
885 : * GetTsmRoutine is relatively expensive compared to the other tests here,
886 : * so check repeatable_across_scans last, even though that's a bit odd.
887 : */
2815 tgl 888 GIC 239 : if ((root->query_level > 1 ||
69 tgl 889 GNC 113 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
2118 tgl 890 GIC 22 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
891 : {
2815 tgl 892 CBC 4 : path = (Path *) create_material_path(rel, path);
893 : }
894 :
895 126 : add_path(rel, path);
896 :
897 : /* For the moment, at least, there are no other paths to consider */
2886 simon 898 GIC 126 : }
899 :
900 : /*
901 : * set_foreign_size
4090 tgl 902 ECB : * Set size estimates for a foreign table RTE
903 : */
904 : static void
4090 tgl 905 GIC 1099 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
906 : {
907 : /* Mark rel with estimated output rows, width, etc */
908 1099 : set_foreign_size_estimates(root, rel);
909 :
4048 tgl 910 ECB : /* Let FDW adjust the size estimates, if it can */
4048 tgl 911 GIC 1099 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
912 :
2814 tgl 913 ECB : /* ... but do not let it set the rows estimate to zero */
2814 tgl 914 CBC 1099 : rel->rows = clamp_row_est(rel->rows);
915 :
916 : /*
917 : * Also, make sure rel->tuples is not insane relative to rel->rows.
918 : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
919 : * FDW doesn't do anything to replace that.
920 : */
1010 tgl 921 GIC 1099 : rel->tuples = Max(rel->tuples, rel->rows);
4090 922 1099 : }
923 :
924 : /*
925 : * set_foreign_pathlist
926 : * Build access paths for a foreign table RTE
927 : */
4090 tgl 928 ECB : static void
4090 tgl 929 GIC 1099 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
930 : {
4048 tgl 931 ECB : /* Call the FDW's GetForeignPaths function to generate path(s) */
4048 tgl 932 GIC 1099 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
4090 933 1099 : }
934 :
935 : /*
936 : * set_append_rel_size
937 : * Set size estimates for a simple "append relation"
938 : *
939 : * The passed-in rel and RTE represent the entire append relation. The
2117 rhaas 940 ECB : * relation's contents are computed by appending together the output of the
941 : * individual member relations. Note that in the non-partitioned inheritance
942 : * case, the first member relation is actually the same table as is mentioned
943 : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
944 : * a good thing because their outputs are not the same size.
945 : */
946 : static void
4090 tgl 947 GIC 9313 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
948 : Index rti, RangeTblEntry *rte)
8183 tgl 949 ECB : {
7994 tgl 950 CBC 9313 : int parentRTindex = rti;
2814 tgl 951 ECB : bool has_live_children;
5399 952 : double parent_rows;
953 : double parent_size;
954 : double *parent_attrsizes;
955 : int nattrs;
956 : ListCell *l;
957 :
958 : /* Guard against stack overflow due to overly deep inheritance tree. */
2033 rhaas 959 GIC 9313 : check_stack_depth();
960 :
2197 961 9313 : Assert(IS_SIMPLE_REL(rel));
962 :
963 : /*
964 : * If this is a partitioned baserel, set the consider_partitionwise_join
965 : * flag; currently, we only consider partitionwise joins with the baserel
966 : * if its targetlist doesn't contain a whole-row Var.
967 : */
1682 efujita 968 9313 : if (enable_partitionwise_join &&
1682 efujita 969 CBC 1951 : rel->reloptkind == RELOPT_BASEREL &&
970 1645 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
69 tgl 971 GNC 1645 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
1682 efujita 972 CBC 1607 : rel->consider_partitionwise_join = true;
1682 efujita 973 ECB :
974 : /*
5399 tgl 975 : * Initialize to compute size estimates for whole append relation.
976 : *
5050 bruce 977 : * We handle width estimates by weighting the widths of different child
978 : * rels proportionally to their number of rows. This is sensible because
979 : * the use of width estimates is mainly to compute the total relation
980 : * "footprint" if we have to sort or hash it. To do this, we sum the
981 : * total equivalent size (in "double" arithmetic) and then divide by the
982 : * total rowcount estimate. This is done separately for the total rel
983 : * width and each attribute.
984 : *
985 : * Note: if you consider changing this logic, beware that child rels could
986 : * have zero rows and/or width, if they were excluded by constraints.
8183 tgl 987 : */
2814 tgl 988 CBC 9313 : has_live_children = false;
5399 tgl 989 GIC 9313 : parent_rows = 0;
5399 tgl 990 CBC 9313 : parent_size = 0;
991 9313 : nattrs = rel->max_attr - rel->min_attr + 1;
5399 tgl 992 GIC 9313 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
993 :
6277 994 48784 : foreach(l, root->append_rel_list)
995 : {
996 39471 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
6277 tgl 997 ECB : int childRTindex;
5832 998 : RangeTblEntry *childRTE;
999 : RelOptInfo *childrel;
1000 : List *childrinfos;
1001 : ListCell *parentvars;
6892 neilc 1002 : ListCell *childvars;
1003 : ListCell *lc;
8183 tgl 1004 :
1005 : /* append_rel_list contains all append rels; ignore others */
6277 tgl 1006 GIC 39471 : if (appinfo->parent_relid != parentRTindex)
1007 20050 : continue;
1008 :
1009 19535 : childRTindex = appinfo->child_relid;
5832 1010 19535 : childRTE = root->simple_rte_array[childRTindex];
1011 :
1012 : /*
6046 tgl 1013 ECB : * The child rel's RelOptInfo was already created during
1014 : * add_other_rels_to_query.
1015 : */
6046 tgl 1016 GIC 19535 : childrel = find_base_rel(root, childRTindex);
1017 19535 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1018 :
1471 tgl 1019 ECB : /* We may have already proven the child to be dummy. */
1471 tgl 1020 CBC 19535 : if (IS_DUMMY_REL(childrel))
1528 alvherre 1021 GIC 3 : continue;
1022 :
1023 : /*
1024 : * We have to copy the parent's targetlist and quals to the child,
1025 : * with appropriate substitution of variables. However, the
1026 : * baserestrictinfo quals were already copied/substituted when the
1027 : * child RelOptInfo was built. So we don't need any additional setup
1028 : * before applying constraint exclusion.
1029 : */
5486 tgl 1030 19532 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1031 : {
1032 : /*
5797 tgl 1033 ECB : * This child need not be scanned, so we can omit it from the
4090 1034 : * appendrel.
1035 : */
5797 tgl 1036 CBC 42 : set_dummy_rel_pathlist(childrel);
5832 tgl 1037 GIC 42 : continue;
5832 tgl 1038 ECB : }
1039 :
1539 efujita 1040 : /*
1041 : * Constraint exclusion failed, so copy the parent's join quals and
1042 : * targetlist to the child, with appropriate variable substitutions.
1043 : *
1044 : * We skip join quals that came from above outer joins that can null
1045 : * this rel, since they would be of no value while generating paths
1046 : * for the child. This saves some effort while processing the child
1047 : * rel, and it also avoids an implementation restriction in
1048 : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1049 : */
28 tgl 1050 GNC 19490 : childrinfos = NIL;
1051 25598 : foreach(lc, rel->joininfo)
1052 : {
1053 6108 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1054 :
1055 6108 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1056 4965 : childrinfos = lappend(childrinfos,
1057 4965 : adjust_appendrel_attrs(root,
1058 : (Node *) rinfo,
1059 : 1, &appinfo));
1060 : }
1061 19490 : childrel->joininfo = childrinfos;
1062 :
1063 : /*
1064 : * Now for the child's targetlist.
1065 : *
1539 efujita 1066 ECB : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1067 : * expressions, which otherwise would not occur in a rel's targetlist.
1068 : * Code that might be looking at an appendrel child must cope with
1069 : * such. (Normally, a rel's targetlist would only include Vars and
1070 : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1071 : * fields of childrel->reltarget; not clear if that would be useful.
1072 : */
1539 efujita 1073 GIC 38980 : childrel->reltarget->exprs = (List *)
1539 efujita 1074 CBC 19490 : adjust_appendrel_attrs(root,
1075 19490 : (Node *) rel->reltarget->exprs,
1539 efujita 1076 ECB : 1, &appinfo);
1077 :
1078 : /*
1079 : * We have to make child entries in the EquivalenceClass data
1080 : * structures as well. This is needed either if the parent
1081 : * participates in some eclass joins (because we will want to consider
1082 : * inner-indexscan joins on the individual children) or if the parent
1083 : * has useful pathkeys (because we should try to build MergeAppend
1084 : * paths that produce those sort orderings).
1085 : */
1539 efujita 1086 GIC 19490 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1539 efujita 1087 CBC 10098 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1088 19490 : childrel->has_eclass_joins = rel->has_eclass_joins;
8183 tgl 1089 ECB :
1090 : /*
1091 : * Note: we could compute appropriate attr_needed data for the child's
1092 : * variables, by transforming the parent's attr_needed through the
1093 : * translated_vars mapping. However, currently there's no need
1094 : * because attr_needed is only examined for base relations not
1095 : * otherrels. So we just leave the child's attr_needed empty.
1096 : */
1097 :
1098 : /*
1099 : * If we consider partitionwise joins with the parent rel, do the same
1100 : * for partitioned child rels.
1101 : *
1102 : * Note: here we abuse the consider_partitionwise_join flag by setting
1103 : * it for child rels that are not themselves partitioned. We do so to
1104 : * tell try_partitionwise_join() that the child rel is sufficiently
1105 : * valid to be used as a per-partition input, even if it later gets
1106 : * proven to be dummy. (It's not usable until we've set up the
1107 : * reltarget and EC entries, which we just did.)
1108 : */
1539 efujita 1109 GIC 19490 : if (rel->consider_partitionwise_join)
1682 efujita 1110 CBC 5182 : childrel->consider_partitionwise_join = true;
1682 efujita 1111 ECB :
1112 : /*
1113 : * If parallelism is allowable for this query in general, see whether
1114 : * it's allowable for this childrel in particular. But if we've
1115 : * already decided the appendrel is not parallel-safe as a whole,
1116 : * there's no point in considering parallelism for this child. For
1117 : * consistency, do this before calling set_rel_size() for the child.
1118 : */
2471 tgl 1119 GIC 19490 : if (root->glob->parallelModeOK && rel->consider_parallel)
2471 tgl 1120 CBC 13854 : set_rel_consider_parallel(root, childrel, childRTE);
2471 tgl 1121 ECB :
1122 : /*
1123 : * Compute the child's size.
1124 : */
4090 tgl 1125 GIC 19490 : set_rel_size(root, childrel, childRTindex, childRTE);
8183 tgl 1126 ECB :
1127 : /*
1128 : * It is possible that constraint exclusion detected a contradiction
1129 : * within a child subquery, even though we didn't prove one above. If
1130 : * so, we can skip this child.
1131 : */
4090 tgl 1132 GIC 19490 : if (IS_DUMMY_REL(childrel))
4215 tgl 1133 CBC 69 : continue;
4215 tgl 1134 ECB :
1135 : /* We have at least one live child. */
2814 tgl 1136 GIC 19421 : has_live_children = true;
2814 tgl 1137 ECB :
1138 : /*
1139 : * If any live child is not parallel-safe, treat the whole appendrel
1140 : * as not parallel-safe. In future we might be able to generate plans
1141 : * in which some children are farmed out to workers while others are
1142 : * not; but we don't have that today, so it's a waste to consider
1143 : * partial paths anywhere in the appendrel unless it's all safe.
1144 : * (Child rels visited before this one will be unmarked in
1145 : * set_append_rel_pathlist().)
1146 : */
2471 tgl 1147 GIC 19421 : if (!childrel->consider_parallel)
2471 tgl 1148 CBC 5814 : rel->consider_parallel = false;
2471 tgl 1149 ECB :
1150 : /*
1151 : * Accumulate size information from each live child.
1152 : */
2814 tgl 1153 GIC 19421 : Assert(childrel->rows > 0);
2814 tgl 1154 ECB :
2814 tgl 1155 GIC 19421 : parent_rows += childrel->rows;
2582 tgl 1156 CBC 19421 : parent_size += childrel->reltarget->width * childrel->rows;
2814 tgl 1157 ECB :
1158 : /*
1159 : * Accumulate per-column estimates too. We need not do anything for
1160 : * PlaceHolderVars in the parent list. If child expression isn't a
1161 : * Var, or we didn't record a width estimate for it, we have to fall
1162 : * back on a datatype-based estimate.
1163 : *
1164 : * By construction, child's targetlist is 1-to-1 with parent's.
1165 : */
2582 tgl 1166 GIC 61185 : forboth(parentvars, rel->reltarget->exprs,
2582 tgl 1167 ECB : childvars, childrel->reltarget->exprs)
1168 : {
2814 tgl 1169 GIC 41764 : Var *parentvar = (Var *) lfirst(parentvars);
2814 tgl 1170 CBC 41764 : Node *childvar = (Node *) lfirst(childvars);
7224 tgl 1171 ECB :
739 tgl 1172 GIC 41764 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
6882 tgl 1173 ECB : {
2814 tgl 1174 GIC 36316 : int pndx = parentvar->varattno - rel->min_attr;
2814 tgl 1175 CBC 36316 : int32 child_width = 0;
4247 tgl 1176 ECB :
2814 tgl 1177 GIC 36316 : if (IsA(childvar, Var) &&
2814 tgl 1178 CBC 35327 : ((Var *) childvar)->varno == childrel->relid)
5399 tgl 1179 ECB : {
2814 tgl 1180 GIC 35297 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
2814 tgl 1181 ECB :
2814 tgl 1182 GIC 35297 : child_width = childrel->attr_widths[cndx];
5399 tgl 1183 ECB : }
2814 tgl 1184 GIC 36316 : if (child_width <= 0)
2814 tgl 1185 CBC 1019 : child_width = get_typavgwidth(exprType(childvar),
2814 tgl 1186 ECB : exprTypmod(childvar));
2814 tgl 1187 GIC 36316 : Assert(child_width > 0);
2814 tgl 1188 CBC 36316 : parent_attrsizes[pndx] += child_width * childrel->rows;
6882 tgl 1189 ECB : }
1190 : }
1191 : }
1192 :
2814 tgl 1193 GIC 9313 : if (has_live_children)
5399 tgl 1194 ECB : {
1195 : /*
1196 : * Save the finished size estimates.
1197 : */
1198 : int i;
1199 :
2814 tgl 1200 GIC 9154 : Assert(parent_rows > 0);
2814 tgl 1201 CBC 9154 : rel->rows = parent_rows;
2582 1202 9154 : rel->reltarget->width = rint(parent_size / parent_rows);
5399 1203 93532 : for (i = 0; i < nattrs; i++)
1204 84378 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
2814 tgl 1205 ECB :
1206 : /*
1207 : * Set "raw tuples" count equal to "rows" for the appendrel; needed
1208 : * because some places assume rel->tuples is valid for any baserel.
1209 : */
2814 tgl 1210 GIC 9154 : rel->tuples = parent_rows;
1614 tgl 1211 ECB :
1212 : /*
1213 : * Note that we leave rel->pages as zero; this is important to avoid
1214 : * double-counting the appendrel tree in total_table_pages.
1215 : */
1216 : }
1217 : else
1218 : {
1219 : /*
1220 : * All children were excluded by constraints, so mark the whole
1221 : * appendrel dummy. We must do this in this phase so that the rel's
1222 : * dummy-ness is visible when we generate paths for other rels.
1223 : */
2814 tgl 1224 GIC 159 : set_dummy_rel_pathlist(rel);
2814 tgl 1225 ECB : }
1226 :
5399 tgl 1227 GIC 9313 : pfree(parent_attrsizes);
4090 tgl 1228 CBC 9313 : }
4090 tgl 1229 ECB :
1230 : /*
1231 : * set_append_rel_pathlist
1232 : * Build access paths for an "append relation"
1233 : */
1234 : static void
4090 tgl 1235 GIC 9154 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
4090 tgl 1236 ECB : Index rti, RangeTblEntry *rte)
1237 : {
4090 tgl 1238 GIC 9154 : int parentRTindex = rti;
4090 tgl 1239 CBC 9154 : List *live_childrels = NIL;
4090 tgl 1240 ECB : ListCell *l;
1241 :
1242 : /*
1243 : * Generate access paths for each member relation, and remember the
1244 : * non-dummy children.
1245 : */
4090 tgl 1246 GIC 48361 : foreach(l, root->append_rel_list)
4090 tgl 1247 ECB : {
4090 tgl 1248 GIC 39207 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
4090 tgl 1249 ECB : int childRTindex;
1250 : RangeTblEntry *childRTE;
1251 : RelOptInfo *childrel;
1252 :
1253 : /* append_rel_list contains all append rels; ignore others */
4090 tgl 1254 GIC 39207 : if (appinfo->parent_relid != parentRTindex)
4090 tgl 1255 CBC 19705 : continue;
4090 tgl 1256 ECB :
1257 : /* Re-locate the child RTE and RelOptInfo */
4090 tgl 1258 GIC 19502 : childRTindex = appinfo->child_relid;
4090 tgl 1259 CBC 19502 : childRTE = root->simple_rte_array[childRTindex];
1260 19502 : childrel = root->simple_rel_array[childRTindex];
4090 tgl 1261 ECB :
1262 : /*
1263 : * If set_append_rel_size() decided the parent appendrel was
1264 : * parallel-unsafe at some point after visiting this child rel, we
1265 : * need to propagate the unsafety marking down to the child, so that
1266 : * we don't generate useless partial paths for it.
1267 : */
2471 tgl 1268 GIC 19502 : if (!rel->consider_parallel)
2471 tgl 1269 CBC 5860 : childrel->consider_parallel = false;
2471 tgl 1270 ECB :
1271 : /*
1272 : * Compute the child's access paths.
1273 : */
4090 tgl 1274 GIC 19502 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
4090 tgl 1275 ECB :
1276 : /*
1277 : * If child is dummy, ignore it.
1278 : */
4090 tgl 1279 GIC 19502 : if (IS_DUMMY_REL(childrel))
4090 tgl 1280 CBC 81 : continue;
4090 tgl 1281 ECB :
1282 : /*
1283 : * Child is live, so add it to the live_childrels list for use below.
1284 : */
3893 tgl 1285 GIC 19421 : live_childrels = lappend(live_childrels, childrel);
2217 rhaas 1286 ECB : }
1287 :
1288 : /* Add paths to the append relation. */
2217 rhaas 1289 GIC 9154 : add_paths_to_append_rel(root, rel, live_childrels);
2217 rhaas 1290 CBC 9154 : }
2217 rhaas 1291 ECB :
1292 :
1293 : /*
1294 : * add_paths_to_append_rel
1295 : * Generate paths for the given append relation given the set of non-dummy
1296 : * child rels.
1297 : *
1298 : * The function collects all parameterizations and orderings supported by the
1299 : * non-dummy children. For every such parameterization or ordering, it creates
1300 : * an append path collecting one path from each non-dummy child with given
1301 : * parameterization or ordering. Similarly it collects partial paths from
1302 : * non-dummy children to create partial append paths.
1303 : */
1304 : void
2217 rhaas 1305 GIC 15788 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
2217 rhaas 1306 ECB : List *live_childrels)
1307 : {
2217 rhaas 1308 GIC 15788 : List *subpaths = NIL;
2217 rhaas 1309 CBC 15788 : bool subpaths_valid = true;
1310 15788 : List *partial_subpaths = NIL;
1951 1311 15788 : List *pa_partial_subpaths = NIL;
1312 15788 : List *pa_nonpartial_subpaths = NIL;
2217 1313 15788 : bool partial_subpaths_valid = true;
1754 akapila 1314 ECB : bool pa_subpaths_valid;
2217 rhaas 1315 GIC 15788 : List *all_child_pathkeys = NIL;
2217 rhaas 1316 CBC 15788 : List *all_child_outers = NIL;
2217 rhaas 1317 ECB : ListCell *l;
1951 rhaas 1318 GIC 15788 : double partial_rows = -1;
2210 rhaas 1319 ECB :
1320 : /* If appropriate, consider parallel append */
1754 akapila 1321 GIC 15788 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1754 akapila 1322 ECB :
1323 : /*
1324 : * For every non-dummy child, remember the cheapest path. Also, identify
1325 : * all pathkeys (orderings) and parameterizations (required_outer sets)
1326 : * available for the non-dummy member relations.
1327 : */
2217 rhaas 1328 GIC 48226 : foreach(l, live_childrels)
2217 rhaas 1329 ECB : {
2217 rhaas 1330 GIC 32438 : RelOptInfo *childrel = lfirst(l);
2217 rhaas 1331 ECB : ListCell *lcp;
1951 rhaas 1332 GIC 32438 : Path *cheapest_partial_path = NULL;
3897 tgl 1333 ECB :
1334 : /*
1335 : * If child has an unparameterized cheapest-total path, add that to
1336 : * the unparameterized Append path we are constructing for the parent.
1337 : * If not, there's no workable unparameterized path.
1338 : *
1339 : * With partitionwise aggregates, the child rel's pathlist may be
1340 : * empty, so don't assume that a path exists here.
1341 : */
1844 rhaas 1342 GIC 32438 : if (childrel->pathlist != NIL &&
1844 rhaas 1343 CBC 32438 : childrel->cheapest_total_path->param_info == NULL)
1951 1344 32186 : accumulate_append_subpath(childrel->cheapest_total_path,
797 tgl 1345 ECB : &subpaths, NULL);
1346 : else
3893 tgl 1347 GIC 252 : subpaths_valid = false;
4090 tgl 1348 ECB :
1349 : /* Same idea, but for a partial plan. */
2636 rhaas 1350 GIC 32438 : if (childrel->partial_pathlist != NIL)
1951 rhaas 1351 ECB : {
1951 rhaas 1352 GIC 21542 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1951 rhaas 1353 CBC 21542 : accumulate_append_subpath(cheapest_partial_path,
797 tgl 1354 ECB : &partial_subpaths, NULL);
1355 : }
1356 : else
2636 rhaas 1357 GIC 10896 : partial_subpaths_valid = false;
2636 rhaas 1358 ECB :
1359 : /*
1360 : * Same idea, but for a parallel append mixing partial and non-partial
1361 : * paths.
1362 : */
1951 rhaas 1363 GIC 32438 : if (pa_subpaths_valid)
1951 rhaas 1364 ECB : {
1951 rhaas 1365 GIC 22533 : Path *nppath = NULL;
1951 rhaas 1366 ECB :
1367 : nppath =
1951 rhaas 1368 GIC 22533 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1951 rhaas 1369 ECB :
1951 rhaas 1370 GIC 22533 : if (cheapest_partial_path == NULL && nppath == NULL)
1951 rhaas 1371 ECB : {
1372 : /* Neither a partial nor a parallel-safe path? Forget it. */
1951 rhaas 1373 GIC 374 : pa_subpaths_valid = false;
1951 rhaas 1374 ECB : }
1951 rhaas 1375 GIC 22159 : else if (nppath == NULL ||
1951 rhaas 1376 CBC 21317 : (cheapest_partial_path != NULL &&
1377 21317 : cheapest_partial_path->total_cost < nppath->total_cost))
1951 rhaas 1378 ECB : {
1379 : /* Partial path is cheaper or the only option. */
1951 rhaas 1380 GIC 21210 : Assert(cheapest_partial_path != NULL);
1951 rhaas 1381 CBC 21210 : accumulate_append_subpath(cheapest_partial_path,
1951 rhaas 1382 ECB : &pa_partial_subpaths,
1383 : &pa_nonpartial_subpaths);
1384 : }
1385 : else
1386 : {
1387 : /*
1388 : * Either we've got only a non-partial path, or we think that
1389 : * a single backend can execute the best non-partial path
1390 : * faster than all the parallel backends working together can
1391 : * execute the best partial path.
1392 : *
1393 : * It might make sense to be more aggressive here. Even if
1394 : * the best non-partial path is more expensive than the best
1395 : * partial path, it could still be better to choose the
1396 : * non-partial path if there are several such paths that can
1397 : * be given to different workers. For now, we don't try to
1398 : * figure that out.
1399 : */
1951 rhaas 1400 GIC 949 : accumulate_append_subpath(nppath,
1951 rhaas 1401 ECB : &pa_nonpartial_subpaths,
1402 : NULL);
1403 : }
1404 : }
1405 :
1406 : /*
1407 : * Collect lists of all the available path orderings and
1408 : * parameterizations for all the children. We use these as a
1409 : * heuristic to indicate which sort orderings and parameterizations we
1410 : * should build Append and MergeAppend paths for.
1411 : */
4090 tgl 1412 GIC 74670 : foreach(lcp, childrel->pathlist)
4090 tgl 1413 ECB : {
4090 tgl 1414 GIC 42232 : Path *childpath = (Path *) lfirst(lcp);
4090 tgl 1415 CBC 42232 : List *childkeys = childpath->pathkeys;
4007 1416 42232 : Relids childouter = PATH_REQ_OUTER(childpath);
4090 tgl 1417 ECB :
1418 : /* Unsorted paths don't contribute to pathkey list */
4090 tgl 1419 GIC 42232 : if (childkeys != NIL)
4090 tgl 1420 ECB : {
1421 : ListCell *lpk;
4090 tgl 1422 GIC 9841 : bool found = false;
4090 tgl 1423 ECB :
1424 : /* Have we already seen this ordering? */
4090 tgl 1425 GIC 9935 : foreach(lpk, all_child_pathkeys)
4090 tgl 1426 ECB : {
4090 tgl 1427 GIC 7107 : List *existing_pathkeys = (List *) lfirst(lpk);
4090 tgl 1428 ECB :
4090 tgl 1429 GIC 7107 : if (compare_pathkeys(existing_pathkeys,
4090 tgl 1430 ECB : childkeys) == PATHKEYS_EQUAL)
1431 : {
4090 tgl 1432 GIC 7013 : found = true;
4090 tgl 1433 CBC 7013 : break;
4090 tgl 1434 ECB : }
1435 : }
4090 tgl 1436 GIC 9841 : if (!found)
4090 tgl 1437 ECB : {
1438 : /* No, so add it to all_child_pathkeys */
4090 tgl 1439 GIC 2828 : all_child_pathkeys = lappend(all_child_pathkeys,
4090 tgl 1440 ECB : childkeys);
1441 : }
1442 : }
1443 :
1444 : /* Unparameterized paths don't contribute to param-set list */
4090 tgl 1445 GIC 42232 : if (childouter)
4090 tgl 1446 ECB : {
1447 : ListCell *lco;
4090 tgl 1448 GIC 2772 : bool found = false;
4090 tgl 1449 ECB :
1450 : /* Have we already seen this param set? */
4090 tgl 1451 GIC 3078 : foreach(lco, all_child_outers)
4090 tgl 1452 ECB : {
3955 bruce 1453 GIC 2028 : Relids existing_outers = (Relids) lfirst(lco);
4090 tgl 1454 ECB :
4090 tgl 1455 GIC 2028 : if (bms_equal(existing_outers, childouter))
4090 tgl 1456 ECB : {
4090 tgl 1457 GIC 1722 : found = true;
4090 tgl 1458 CBC 1722 : break;
4090 tgl 1459 ECB : }
1460 : }
4090 tgl 1461 GIC 2772 : if (!found)
4090 tgl 1462 ECB : {
1463 : /* No, so add it to all_child_outers */
4090 tgl 1464 GIC 1050 : all_child_outers = lappend(all_child_outers,
4090 tgl 1465 ECB : childouter);
1466 : }
1467 : }
1468 : }
1469 : }
1470 :
1471 : /*
1472 : * If we found unparameterized paths for all children, build an unordered,
1473 : * unparameterized Append path for the rel. (Note: this is correct even
1474 : * if we have zero or one live subpath due to constraint exclusion.)
1475 : */
3893 tgl 1476 GIC 15788 : if (subpaths_valid)
1828 alvherre 1477 CBC 15677 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1465 tgl 1478 ECB : NIL, NULL, 0, false,
1479 : -1));
1480 :
1481 : /*
1482 : * Consider an append of unordered, unparameterized partial paths. Make
1483 : * it parallel-aware if possible.
1484 : */
1494 tgl 1485 GIC 15788 : if (partial_subpaths_valid && partial_subpaths != NIL)
2636 rhaas 1486 ECB : {
1487 : AppendPath *appendpath;
1488 : ListCell *lc;
2495 rhaas 1489 GIC 9716 : int parallel_workers = 0;
2636 rhaas 1490 ECB :
1491 : /* Find the highest number of workers requested for any subpath. */
2636 rhaas 1492 GIC 33216 : foreach(lc, partial_subpaths)
2636 rhaas 1493 ECB : {
2636 rhaas 1494 GIC 23500 : Path *path = lfirst(lc);
2636 rhaas 1495 ECB :
2495 rhaas 1496 GIC 23500 : parallel_workers = Max(parallel_workers, path->parallel_workers);
2636 rhaas 1497 ECB : }
2495 rhaas 1498 GIC 9716 : Assert(parallel_workers > 0);
2636 rhaas 1499 ECB :
1500 : /*
1501 : * If the use of parallel append is permitted, always request at least
1502 : * log2(# of children) workers. We assume it can be useful to have
1503 : * extra workers in this case because they will be spread out across
1504 : * the children. The precise formula is just a guess, but we don't
1505 : * want to end up with a radically different answer for a table with N
1506 : * partitions vs. an unpartitioned table with the same data, so the
1507 : * use of some kind of log-scaling here seems to make some sense.
1508 : */
1951 rhaas 1509 GIC 9716 : if (enable_parallel_append)
1951 rhaas 1510 ECB : {
1951 rhaas 1511 GIC 9692 : parallel_workers = Max(parallel_workers,
1512 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1513 9692 : parallel_workers = Min(parallel_workers,
1951 rhaas 1514 ECB : max_parallel_workers_per_gather);
1515 : }
1951 rhaas 1516 GIC 9716 : Assert(parallel_workers > 0);
1951 rhaas 1517 ECB :
1518 : /* Generate a partial append path. */
1828 alvherre 1519 GIC 9716 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1465 tgl 1520 ECB : NIL, NULL, parallel_workers,
1521 : enable_parallel_append,
1522 : -1);
1523 :
1524 : /*
1525 : * Make sure any subsequent partial paths use the same row count
1526 : * estimate.
1527 : */
1951 rhaas 1528 GIC 9716 : partial_rows = appendpath->path.rows;
1951 rhaas 1529 ECB :
1530 : /* Add the path. */
1951 rhaas 1531 GIC 9716 : add_partial_path(rel, (Path *) appendpath);
1951 rhaas 1532 ECB : }
1533 :
1534 : /*
1535 : * Consider a parallel-aware append using a mix of partial and non-partial
1536 : * paths. (This only makes sense if there's at least one child which has
1537 : * a non-partial path that is substantially cheaper than any partial path;
1538 : * otherwise, we should use the append path added in the previous step.)
1539 : */
1951 rhaas 1540 GIC 15788 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1951 rhaas 1541 ECB : {
1542 : AppendPath *appendpath;
1543 : ListCell *lc;
1951 rhaas 1544 GIC 500 : int parallel_workers = 0;
1951 rhaas 1545 ECB :
1546 : /*
1547 : * Find the highest number of workers requested for any partial
1548 : * subpath.
1549 : */
1951 rhaas 1550 GIC 930 : foreach(lc, pa_partial_subpaths)
1951 rhaas 1551 ECB : {
1951 rhaas 1552 GIC 430 : Path *path = lfirst(lc);
1951 rhaas 1553 ECB :
1951 rhaas 1554 GIC 430 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1951 rhaas 1555 ECB : }
1556 :
1557 : /*
1558 : * Same formula here as above. It's even more important in this
1559 : * instance because the non-partial paths won't contribute anything to
1560 : * the planned number of parallel workers.
1561 : */
1951 rhaas 1562 GIC 500 : parallel_workers = Max(parallel_workers,
1563 : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1564 500 : parallel_workers = Min(parallel_workers,
1951 rhaas 1565 ECB : max_parallel_workers_per_gather);
1951 rhaas 1566 GIC 500 : Assert(parallel_workers > 0);
1951 rhaas 1567 ECB :
1828 alvherre 1568 GIC 500 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1951 rhaas 1569 ECB : pa_partial_subpaths,
1570 : NIL, NULL, parallel_workers, true,
1571 : partial_rows);
2636 rhaas 1572 GIC 500 : add_partial_path(rel, (Path *) appendpath);
2636 rhaas 1573 ECB : }
1574 :
1575 : /*
1576 : * Also build unparameterized ordered append paths based on the collected
1577 : * list of child pathkeys.
1578 : */
3893 tgl 1579 GIC 15788 : if (subpaths_valid)
1465 tgl 1580 CBC 15677 : generate_orderedappend_paths(root, rel, live_childrels,
797 tgl 1581 ECB : all_child_pathkeys);
1582 :
1583 : /*
1584 : * Build Append paths for each parameterization seen among the child rels.
1585 : * (This may look pretty expensive, but in most cases of practical
1586 : * interest, the child rels will expose mostly the same parameterizations,
1587 : * so that not that many cases actually get considered here.)
1588 : *
1589 : * The Append node itself cannot enforce quals, so all qual checking must
1590 : * be done in the child paths. This means that to have a parameterized
1591 : * Append path, we must have the exact same parameterization for each
1592 : * child path; otherwise some children might be failing to check the
1593 : * moved-down quals. To make them match up, we can try to increase the
1594 : * parameterization of lesser-parameterized paths.
1595 : */
4090 tgl 1596 GIC 16838 : foreach(l, all_child_outers)
4090 tgl 1597 ECB : {
3955 bruce 1598 GIC 1050 : Relids required_outer = (Relids) lfirst(l);
4090 tgl 1599 ECB : ListCell *lcr;
1600 :
1601 : /* Select the child paths for an Append with this parameterization */
4090 tgl 1602 GIC 1050 : subpaths = NIL;
3893 tgl 1603 CBC 1050 : subpaths_valid = true;
4090 1604 3864 : foreach(lcr, live_childrels)
4090 tgl 1605 ECB : {
4090 tgl 1606 GIC 2820 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
3563 tgl 1607 ECB : Path *subpath;
1608 :
1844 rhaas 1609 GIC 2820 : if (childrel->pathlist == NIL)
1844 rhaas 1610 ECB : {
1611 : /* failed to make a suitable path for this child */
1844 rhaas 1612 UIC 0 : subpaths_valid = false;
1844 rhaas 1613 UBC 0 : break;
1844 rhaas 1614 EUB : }
1615 :
3563 tgl 1616 GIC 2820 : subpath = get_cheapest_parameterized_child_path(root,
3563 tgl 1617 ECB : childrel,
1618 : required_outer);
3563 tgl 1619 GIC 2820 : if (subpath == NULL)
4007 tgl 1620 ECB : {
1621 : /* failed to make a suitable path for this child */
3563 tgl 1622 GIC 6 : subpaths_valid = false;
3563 tgl 1623 CBC 6 : break;
4007 tgl 1624 ECB : }
797 tgl 1625 GIC 2814 : accumulate_append_subpath(subpath, &subpaths, NULL);
4090 tgl 1626 ECB : }
1627 :
3893 tgl 1628 GIC 1050 : if (subpaths_valid)
4007 tgl 1629 CBC 1044 : add_path(rel, (Path *)
1828 alvherre 1630 1044 : create_append_path(root, rel, subpaths, NIL,
1465 tgl 1631 ECB : NIL, required_outer, 0, false,
1632 : -1));
1633 : }
1634 :
1635 : /*
1636 : * When there is only a single child relation, the Append path can inherit
1637 : * any ordering available for the child rel's path, so that it's useful to
1638 : * consider ordered partial paths. Above we only considered the cheapest
1639 : * partial path for each child, but let's also make paths using any
1640 : * partial paths that have pathkeys.
1641 : */
1476 tgl 1642 GIC 15788 : if (list_length(live_childrels) == 1)
1476 tgl 1643 ECB : {
1476 tgl 1644 GIC 6353 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1476 tgl 1645 ECB :
1646 : /* skip the cheapest partial path, since we already used that above */
888 drowley 1647 GIC 6455 : for_each_from(l, childrel->partial_pathlist, 1)
1476 tgl 1648 ECB : {
1476 tgl 1649 GIC 102 : Path *path = (Path *) lfirst(l);
1476 tgl 1650 ECB : AppendPath *appendpath;
1651 :
1652 : /* skip paths with no pathkeys. */
888 drowley 1653 GIC 102 : if (path->pathkeys == NIL)
1476 tgl 1654 LBC 0 : continue;
1476 tgl 1655 EUB :
1476 tgl 1656 GIC 102 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1465 tgl 1657 ECB : NIL, NULL,
1658 : path->parallel_workers, true,
1659 : partial_rows);
1476 tgl 1660 GIC 102 : add_partial_path(rel, (Path *) appendpath);
1476 tgl 1661 ECB : }
1662 : }
4090 tgl 1663 GIC 15788 : }
4090 tgl 1664 ECB :
1665 : /*
1666 : * generate_orderedappend_paths
1667 : * Generate ordered append paths for an append relation
1668 : *
1669 : * Usually we generate MergeAppend paths here, but there are some special
1670 : * cases where we can generate simple Append paths, because the subpaths
1671 : * can provide tuples in the required order already.
1672 : *
1673 : * We generate a path for each ordering (pathkey list) appearing in
1674 : * all_child_pathkeys.
1675 : *
1676 : * We consider both cheapest-startup and cheapest-total cases, ie, for each
1677 : * interesting ordering, collect all the cheapest startup subpaths and all the
1678 : * cheapest total paths, and build a suitable path for each case.
1679 : *
1680 : * We don't currently generate any parameterized ordered paths here. While
1681 : * it would not take much more code here to do so, it's very unclear that it
1682 : * is worth the planning cycles to investigate such paths: there's little
1683 : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1684 : * that the current coding of add_path would reject such paths out of hand,
1685 : * because add_path gives no credit for sort ordering of parameterized paths,
1686 : * and a parameterized MergeAppend is going to be more expensive than the
1687 : * corresponding parameterized Append path. If we ever try harder to support
1688 : * parameterized mergejoin plans, it might be worth adding support for
1689 : * parameterized paths here to feed such joins. (See notes in
1690 : * optimizer/README for why that might not ever happen, though.)
1691 : */
1692 : static void
1465 tgl 1693 GIC 15677 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1465 tgl 1694 ECB : List *live_childrels,
1695 : List *all_child_pathkeys)
1696 : {
1697 : ListCell *lcp;
1465 tgl 1698 GIC 15677 : List *partition_pathkeys = NIL;
1465 tgl 1699 CBC 15677 : List *partition_pathkeys_desc = NIL;
1700 15677 : bool partition_pathkeys_partial = true;
1701 15677 : bool partition_pathkeys_desc_partial = true;
1465 tgl 1702 ECB :
1703 : /*
1704 : * Some partitioned table setups may allow us to use an Append node
1705 : * instead of a MergeAppend. This is possible in cases such as RANGE
1706 : * partitioned tables where it's guaranteed that an earlier partition must
1707 : * contain rows which come earlier in the sort order. To detect whether
1708 : * this is relevant, build pathkey descriptions of the partition ordering,
1709 : * for both forward and reverse scans.
1710 : */
1465 tgl 1711 GIC 27862 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
614 drowley 1712 CBC 12185 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1465 tgl 1713 ECB : {
1465 tgl 1714 GIC 10517 : partition_pathkeys = build_partition_pathkeys(root, rel,
1465 tgl 1715 ECB : ForwardScanDirection,
1716 : &partition_pathkeys_partial);
1717 :
1465 tgl 1718 GIC 10517 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1465 tgl 1719 ECB : BackwardScanDirection,
1720 : &partition_pathkeys_desc_partial);
1721 :
1722 : /*
1723 : * You might think we should truncate_useless_pathkeys here, but
1724 : * allowing partition keys which are a subset of the query's pathkeys
1725 : * can often be useful. For example, consider a table partitioned by
1726 : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1727 : * paths that can produce the a, b, c ordering (perhaps via indexes on
1728 : * (a, b, c)) then it works to consider the appendrel output as
1729 : * ordered by a, b, c.
1730 : */
1731 : }
1732 :
1733 : /* Now consider each interesting sort ordering */
4090 tgl 1734 GIC 18487 : foreach(lcp, all_child_pathkeys)
4560 tgl 1735 ECB : {
4090 tgl 1736 GIC 2810 : List *pathkeys = (List *) lfirst(lcp);
4382 bruce 1737 CBC 2810 : List *startup_subpaths = NIL;
1738 2810 : List *total_subpaths = NIL;
452 tomas.vondra 1739 2810 : List *fractional_subpaths = NIL;
4382 bruce 1740 2810 : bool startup_neq_total = false;
1741 : bool match_partition_order;
1742 : bool match_partition_order_desc;
1743 : int end_index;
1744 : int first_index;
1745 : int direction;
1746 :
1747 : /*
1748 : * Determine if this sort ordering matches any partition pathkeys we
1749 : * have, for both ascending and descending partition order. If the
1750 : * partition pathkeys happen to be contained in pathkeys then it still
1751 : * works, as described above, providing that the partition pathkeys
1752 : * are complete and not just a prefix of the partition keys. (In such
1753 : * cases we'll be relying on the child paths to have sorted the
1754 : * lower-order columns of the required pathkeys.)
1755 : */
1465 tgl 1756 GIC 2810 : match_partition_order =
1757 4594 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1758 1852 : (!partition_pathkeys_partial &&
1465 tgl 1759 CBC 68 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1465 tgl 1760 ECB :
1465 tgl 1761 CBC 6297 : match_partition_order_desc = !match_partition_order &&
1762 1751 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1465 tgl 1763 GIC 1756 : (!partition_pathkeys_desc_partial &&
1465 tgl 1764 CBC 20 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
4560 tgl 1765 ECB :
1766 : /*
1767 : * When the required pathkeys match the reverse of the partition
1768 : * order, we must build the list of paths in reverse starting with the
1769 : * last matching partition first. We can get away without making any
1770 : * special cases for this in the loop below by just looping backward
1771 : * over the child relations in this case.
1772 : */
48 drowley 1773 GNC 2810 : if (match_partition_order_desc)
1774 : {
1775 : /* loop backward */
1776 21 : first_index = list_length(live_childrels) - 1;
1777 21 : end_index = -1;
1778 21 : direction = -1;
1779 :
1780 : /*
1781 : * Set this to true to save us having to check for
1782 : * match_partition_order_desc in the loop below.
1783 : */
1784 21 : match_partition_order = true;
1785 : }
1786 : else
1787 : {
1788 : /* for all other case, loop forward */
1789 2789 : first_index = 0;
1790 2789 : end_index = list_length(live_childrels);
1791 2789 : direction = 1;
1792 : }
1793 :
1794 : /* Select the child paths for this ordering... */
1795 10353 : for (int i = first_index; i != end_index; i += direction)
1796 : {
1797 7543 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1798 : Path *cheapest_startup,
1799 : *cheapest_total,
452 tomas.vondra 1800 GIC 7543 : *cheapest_fractional = NULL;
1801 :
1802 : /* Locate the right paths, if they are available. */
1803 : cheapest_startup =
4560 tgl 1804 CBC 7543 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1805 : pathkeys,
1806 : NULL,
2224 rhaas 1807 ECB : STARTUP_COST,
1808 : false);
4560 tgl 1809 : cheapest_total =
4560 tgl 1810 GIC 7543 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1811 : pathkeys,
1812 : NULL,
1813 : TOTAL_COST,
1814 : false);
4560 tgl 1815 ECB :
1816 : /*
1817 : * If we can't find any paths with the right order just use the
1818 : * cheapest-total path; we'll have to sort it later.
1819 : */
4090 tgl 1820 CBC 7543 : if (cheapest_startup == NULL || cheapest_total == NULL)
4090 tgl 1821 ECB : {
4007 tgl 1822 CBC 131 : cheapest_startup = cheapest_total =
1823 : childrel->cheapest_total_path;
1824 : /* Assert we do have an unparameterized path for this child */
3893 tgl 1825 GIC 131 : Assert(cheapest_total->param_info == NULL);
4090 tgl 1826 ECB : }
1827 :
452 tomas.vondra 1828 : /*
1829 : * When building a fractional path, determine a cheapest
1830 : * fractional path for each child relation too. Looking at startup
332 tgl 1831 : * and total costs is not enough, because the cheapest fractional
1832 : * path may be dominated by two separate paths (one for startup,
1833 : * one for total).
1834 : *
452 tomas.vondra 1835 : * When needed (building fractional path), determine the cheapest
1836 : * fractional path too.
1837 : */
452 tomas.vondra 1838 GIC 7543 : if (root->tuple_fraction > 0)
1839 : {
332 tgl 1840 334 : double path_fraction = (1.0 / root->tuple_fraction);
452 tomas.vondra 1841 ECB :
1842 : cheapest_fractional =
452 tomas.vondra 1843 GIC 334 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1844 : pathkeys,
1845 : NULL,
1846 : path_fraction);
1847 :
1848 : /*
1849 : * If we found no path with matching pathkeys, use the
1850 : * cheapest total path instead.
452 tomas.vondra 1851 ECB : *
1852 : * XXX We might consider partially sorted paths too (with an
1853 : * incremental sort on top). But we'd have to build all the
1854 : * incremental paths, do the costing etc.
1855 : */
452 tomas.vondra 1856 CBC 334 : if (!cheapest_fractional)
452 tomas.vondra 1857 GIC 22 : cheapest_fractional = cheapest_total;
1858 : }
1859 :
1860 : /*
1861 : * Notice whether we actually have different paths for the
1862 : * "cheapest" and "total" cases; frequently there will be no point
1863 : * in two create_merge_append_path() calls.
1864 : */
4560 tgl 1865 7543 : if (cheapest_startup != cheapest_total)
1866 24 : startup_neq_total = true;
1867 :
1868 : /*
1465 tgl 1869 ECB : * Collect the appropriate child paths. The required logic varies
1870 : * for the Append and MergeAppend cases.
1871 : */
1465 tgl 1872 GIC 7543 : if (match_partition_order)
1873 : {
1465 tgl 1874 ECB : /*
1875 : * We're going to make a plain Append path. We don't need
1876 : * most of what accumulate_append_subpath would do, but we do
1877 : * want to cut out child Appends or MergeAppends if they have
1878 : * just a single subpath (and hence aren't doing anything
1879 : * useful).
1880 : */
1465 tgl 1881 GIC 2890 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
1882 2890 : cheapest_total = get_singleton_append_subpath(cheapest_total);
1883 :
1884 2890 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
1885 2890 : total_subpaths = lappend(total_subpaths, cheapest_total);
1886 :
452 tomas.vondra 1887 CBC 2890 : if (cheapest_fractional)
452 tomas.vondra 1888 ECB : {
452 tomas.vondra 1889 GIC 60 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
1890 60 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
1891 : }
1892 : }
1465 tgl 1893 ECB : else
1894 : {
1895 : /*
1896 : * Otherwise, rely on accumulate_append_subpath to collect the
1897 : * child paths for the MergeAppend.
1898 : */
1465 tgl 1899 CBC 4653 : accumulate_append_subpath(cheapest_startup,
1900 : &startup_subpaths, NULL);
1901 4653 : accumulate_append_subpath(cheapest_total,
797 tgl 1902 ECB : &total_subpaths, NULL);
1903 :
452 tomas.vondra 1904 GIC 4653 : if (cheapest_fractional)
1905 274 : accumulate_append_subpath(cheapest_fractional,
1906 : &fractional_subpaths, NULL);
1907 : }
1908 : }
1909 :
1910 : /* ... and build the Append or MergeAppend paths */
48 drowley 1911 GNC 2810 : if (match_partition_order)
1912 : {
1465 tgl 1913 ECB : /* We only need Append */
1465 tgl 1914 GIC 1080 : add_path(rel, (Path *) create_append_path(root,
1915 : rel,
1465 tgl 1916 ECB : startup_subpaths,
1917 : NIL,
1918 : pathkeys,
1919 : NULL,
1920 : 0,
1921 : false,
1922 : -1));
1465 tgl 1923 CBC 1080 : if (startup_neq_total)
1465 tgl 1924 UIC 0 : add_path(rel, (Path *) create_append_path(root,
1925 : rel,
1465 tgl 1926 ECB : total_subpaths,
1927 : NIL,
1928 : pathkeys,
1929 : NULL,
1930 : 0,
1931 : false,
1932 : -1));
1933 :
452 tomas.vondra 1934 GIC 1080 : if (fractional_subpaths)
452 tomas.vondra 1935 CBC 30 : add_path(rel, (Path *) create_append_path(root,
452 tomas.vondra 1936 EUB : rel,
1937 : fractional_subpaths,
1938 : NIL,
1939 : pathkeys,
1940 : NULL,
1941 : 0,
1942 : false,
1943 : -1));
1944 : }
1945 : else
1465 tgl 1946 ECB : {
1947 : /* We need MergeAppend */
4560 tgl 1948 GIC 1730 : add_path(rel, (Path *) create_merge_append_path(root,
1949 : rel,
1950 : startup_subpaths,
1951 : pathkeys,
1952 : NULL));
1465 1953 1730 : if (startup_neq_total)
1954 12 : add_path(rel, (Path *) create_merge_append_path(root,
1955 : rel,
1956 : total_subpaths,
1957 : pathkeys,
1958 : NULL));
1959 :
452 tomas.vondra 1960 CBC 1730 : if (fractional_subpaths)
452 tomas.vondra 1961 GIC 98 : add_path(rel, (Path *) create_merge_append_path(root,
1962 : rel,
1963 : fractional_subpaths,
1964 : pathkeys,
452 tomas.vondra 1965 ECB : NULL));
1465 tgl 1966 : }
1967 : }
9770 scrappy 1968 GIC 15677 : }
1969 :
1970 : /*
1971 : * get_cheapest_parameterized_child_path
3563 tgl 1972 ECB : * Get cheapest path for this relation that has exactly the requested
1973 : * parameterization.
1974 : *
1975 : * Returns NULL if unable to create such a path.
1976 : */
1977 : static Path *
3563 tgl 1978 GIC 2820 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
1979 : Relids required_outer)
3563 tgl 1980 ECB : {
1981 : Path *cheapest;
1982 : ListCell *lc;
1983 :
1984 : /*
1985 : * Look up the cheapest existing path with no more than the needed
1986 : * parameterization. If it has exactly the needed parameterization, we're
1987 : * done.
1988 : */
3563 tgl 1989 GIC 2820 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
3563 tgl 1990 ECB : NIL,
1991 : required_outer,
1992 : TOTAL_COST,
1993 : false);
3563 tgl 1994 GIC 2820 : Assert(cheapest != NULL);
1995 2820 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
1996 2674 : return cheapest;
1997 :
1998 : /*
1999 : * Otherwise, we can "reparameterize" an existing path to match the given
2000 : * parameterization, which effectively means pushing down additional
3563 tgl 2001 ECB : * joinquals to be checked within the path's scan. However, some existing
2002 : * paths might check the available joinquals already while others don't;
2003 : * therefore, it's not clear which existing path will be cheapest after
2004 : * reparameterization. We have to go through them all and find out.
2005 : */
3563 tgl 2006 CBC 146 : cheapest = NULL;
2007 506 : foreach(lc, rel->pathlist)
3563 tgl 2008 ECB : {
3563 tgl 2009 GIC 360 : Path *path = (Path *) lfirst(lc);
2010 :
2011 : /* Can't use it if it needs more than requested parameterization */
2012 360 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2013 12 : continue;
2014 :
2015 : /*
2016 : * Reparameterization can only increase the path's cost, so if it's
2017 : * already more expensive than the current cheapest, forget it.
3563 tgl 2018 ECB : */
3563 tgl 2019 CBC 540 : if (cheapest != NULL &&
3563 tgl 2020 GIC 192 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
3563 tgl 2021 CBC 156 : continue;
2022 :
2023 : /* Reparameterize if needed, then recheck cost */
2024 192 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
3563 tgl 2025 ECB : {
3563 tgl 2026 GIC 154 : path = reparameterize_path(root, path, required_outer, 1.0);
2027 154 : if (path == NULL)
2028 16 : continue; /* failed to reparameterize this one */
2029 138 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2030 :
3563 tgl 2031 CBC 138 : if (cheapest != NULL &&
3563 tgl 2032 LBC 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2033 0 : continue;
2034 : }
2035 :
3563 tgl 2036 ECB : /* We have a new best path */
3563 tgl 2037 GIC 176 : cheapest = path;
3563 tgl 2038 ECB : }
2039 :
2040 : /* Return the best path, or NULL if we found no suitable candidate */
3563 tgl 2041 CBC 146 : return cheapest;
2042 : }
3563 tgl 2043 ECB :
4560 tgl 2044 EUB : /*
2045 : * accumulate_append_subpath
2046 : * Add a subpath to the list being built for an Append or MergeAppend.
2047 : *
2048 : * It's possible that the child is itself an Append or MergeAppend path, in
3299 tgl 2049 ECB : * which case we can "cut out the middleman" and just add its child paths to
2050 : * our own list. (We don't try to do this earlier because we need to apply
2051 : * both levels of transformation to the quals.)
2052 : *
2053 : * Note that if we omit a child MergeAppend in this way, we are effectively
2054 : * omitting a sort step, which seems fine: if the parent is to be an Append,
2055 : * its result would be unsorted anyway, while if the parent is to be a
2056 : * MergeAppend, there's no point in a separate sort on a child.
2057 : *
2058 : * Normally, either path is a partial path and subpaths is a list of partial
2059 : * paths, or else path is a non-partial plan and subpaths is a list of those.
2060 : * However, if path is a parallel-aware Append, then we add its partial path
2061 : * children to subpaths and the rest to special_subpaths. If the latter is
2062 : * NULL, we don't flatten the path at all (unless it contains only partial
2063 : * paths).
2064 : */
2065 : static void
797 tgl 2066 GIC 88281 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2067 : {
4560 2068 88281 : if (IsA(path, AppendPath))
2069 : {
4382 bruce 2070 6649 : AppendPath *apath = (AppendPath *) path;
2071 :
1951 rhaas 2072 6649 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2073 : {
1336 tgl 2074 6553 : *subpaths = list_concat(*subpaths, apath->subpaths);
1951 rhaas 2075 6553 : return;
2076 : }
2077 96 : else if (special_subpaths != NULL)
1951 rhaas 2078 ECB : {
2079 : List *new_special_subpaths;
2080 :
2081 : /* Split Parallel Append into partial and non-partial subpaths */
1951 rhaas 2082 CBC 48 : *subpaths = list_concat(*subpaths,
1951 rhaas 2083 GIC 48 : list_copy_tail(apath->subpaths,
1951 rhaas 2084 ECB : apath->first_partial_path));
270 drowley 2085 GNC 48 : new_special_subpaths = list_copy_head(apath->subpaths,
2086 : apath->first_partial_path);
1951 rhaas 2087 GIC 48 : *special_subpaths = list_concat(*special_subpaths,
1951 rhaas 2088 ECB : new_special_subpaths);
1915 rhaas 2089 GIC 48 : return;
2090 : }
2091 : }
3299 tgl 2092 81632 : else if (IsA(path, MergeAppendPath))
3299 tgl 2093 ECB : {
3299 tgl 2094 CBC 310 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2095 :
1336 2096 310 : *subpaths = list_concat(*subpaths, mpath->subpaths);
1951 rhaas 2097 GIC 310 : return;
3299 tgl 2098 ECB : }
2099 :
1951 rhaas 2100 CBC 81370 : *subpaths = lappend(*subpaths, path);
2101 : }
2102 :
1465 tgl 2103 ECB : /*
2104 : * get_singleton_append_subpath
2105 : * Returns the single subpath of an Append/MergeAppend, or just
2106 : * return 'path' if it's not a single sub-path Append/MergeAppend.
2107 : *
2108 : * Note: 'path' must not be a parallel-aware path.
2109 : */
2110 : static Path *
1465 tgl 2111 CBC 5840 : get_singleton_append_subpath(Path *path)
2112 : {
1465 tgl 2113 GIC 5840 : Assert(!path->parallel_aware);
2114 :
2115 5840 : if (IsA(path, AppendPath))
2116 : {
2117 170 : AppendPath *apath = (AppendPath *) path;
2118 :
2119 170 : if (list_length(apath->subpaths) == 1)
2120 78 : return (Path *) linitial(apath->subpaths);
2121 : }
1465 tgl 2122 CBC 5670 : else if (IsA(path, MergeAppendPath))
2123 : {
2124 126 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2125 :
2126 126 : if (list_length(mpath->subpaths) == 1)
1465 tgl 2127 UIC 0 : return (Path *) linitial(mpath->subpaths);
1465 tgl 2128 ECB : }
2129 :
1465 tgl 2130 CBC 5762 : return path;
1465 tgl 2131 ECB : }
2132 :
5797 2133 : /*
2134 : * set_dummy_rel_pathlist
2135 : * Build a dummy path for a relation that's been excluded by constraints
2136 : *
2137 : * Rather than inventing a special "dummy" path type, we represent this as an
1494 tgl 2138 EUB : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2139 : *
2140 : * (See also mark_dummy_rel, which does basically the same thing, but is
1494 tgl 2141 ECB : * typically used to change a rel into dummy state after we already made
2142 : * paths for it.)
2143 : */
2144 : static void
5797 tgl 2145 GIC 443 : set_dummy_rel_pathlist(RelOptInfo *rel)
2146 : {
2147 : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2148 443 : rel->rows = 0;
2582 2149 443 : rel->reltarget->width = 0;
2150 :
2151 : /* Discard any pre-existing paths; no further need for them */
4090 2152 443 : rel->pathlist = NIL;
2636 rhaas 2153 443 : rel->partial_pathlist = NIL;
2154 :
2155 : /* Set up the dummy path */
1487 tgl 2156 CBC 443 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2157 : NIL, rel->lateral_relids,
2158 : 0, false, -1));
5797 tgl 2159 ECB :
3061 2160 : /*
2161 : * We set the cheapest-path fields immediately, just in case they were
2162 : * pointing at some discarded path. This is redundant when we're called
1494 2163 : * from set_rel_size(), but not when called from elsewhere, and doing it
2164 : * twice is harmless anyway.
2165 : */
5797 tgl 2166 GIC 443 : set_cheapest(rel);
5797 tgl 2167 CBC 443 : }
2168 :
2169 : /* quick-and-dirty test to see if any joining is needed */
2170 : static bool
6512 tgl 2171 GIC 2245 : has_multiple_baserels(PlannerInfo *root)
2172 : {
2173 2245 : int num_base_rels = 0;
2174 : Index rti;
2175 :
6277 2176 6020 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
6512 tgl 2177 ECB : {
6277 tgl 2178 CBC 4272 : RelOptInfo *brel = root->simple_rel_array[rti];
2179 :
6512 tgl 2180 GIC 4272 : if (brel == NULL)
2181 983 : continue;
6512 tgl 2182 ECB :
2183 : /* ignore RTEs that are "other rels" */
6512 tgl 2184 CBC 3289 : if (brel->reloptkind == RELOPT_BASEREL)
6512 tgl 2185 GIC 2742 : if (++num_base_rels > 1)
2186 497 : return true;
6512 tgl 2187 ECB : }
6512 tgl 2188 GIC 1748 : return false;
6512 tgl 2189 ECB : }
2190 :
366 drowley 2191 : /*
2192 : * find_window_run_conditions
2193 : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2194 : * function to determine the function's monotonic properties. We then
2195 : * see if 'opexpr' can be used to short-circuit execution.
2196 : *
2197 : * For example row_number() over (order by ...) always produces a value one
2198 : * higher than the previous. If someone has a window function in a subquery
2199 : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2200 : * as well stop processing the windowagg once the row number reaches 11. Here
2201 : * we check if 'opexpr' might help us to stop doing needless extra processing
2202 : * in WindowAgg nodes.
2203 : *
2204 : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2205 : * its original purpose. This is set to false if the caller can assume that
2206 : * the run condition will handle all of the required filtering.
2207 : *
2208 : * Returns true if 'opexpr' was found to be useful and was added to the
2209 : * WindowClauses runCondition. We also set *keep_original accordingly and add
2210 : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2211 : * If the 'opexpr' cannot be used then we set *keep_original to true and
2212 : * return false.
2213 : */
2214 : static bool
366 drowley 2215 GIC 111 : find_window_run_conditions(Query *subquery, RangeTblEntry *rte, Index rti,
2216 : AttrNumber attno, WindowFunc *wfunc, OpExpr *opexpr,
2217 : bool wfunc_left, bool *keep_original,
2218 : Bitmapset **run_cond_attrs)
2219 : {
2220 : Oid prosupport;
2221 : Expr *otherexpr;
2222 : SupportRequestWFuncMonotonic req;
2223 : SupportRequestWFuncMonotonic *res;
2224 : WindowClause *wclause;
2225 : List *opinfos;
366 drowley 2226 ECB : OpExpr *runopexpr;
2227 : Oid runoperator;
2228 : ListCell *lc;
2229 :
366 drowley 2230 GIC 111 : *keep_original = true;
2231 :
2232 111 : while (IsA(wfunc, RelabelType))
366 drowley 2233 UIC 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2234 :
2235 : /* we can only work with window functions */
366 drowley 2236 GIC 111 : if (!IsA(wfunc, WindowFunc))
2237 12 : return false;
2238 :
2239 : /* can't use it if there are subplans in the WindowFunc */
23 2240 99 : if (contain_subplans((Node *) wfunc))
23 drowley 2241 CBC 3 : return false;
2242 :
366 2243 96 : prosupport = get_func_support(wfunc->winfnoid);
366 drowley 2244 EUB :
2245 : /* Check if there's a support function for 'wfunc' */
366 drowley 2246 GIC 96 : if (!OidIsValid(prosupport))
366 drowley 2247 CBC 9 : return false;
366 drowley 2248 ECB :
2249 : /* get the Expr from the other side of the OpExpr */
366 drowley 2250 GIC 87 : if (wfunc_left)
366 drowley 2251 CBC 75 : otherexpr = lsecond(opexpr->args);
366 drowley 2252 ECB : else
366 drowley 2253 GIC 12 : otherexpr = linitial(opexpr->args);
366 drowley 2254 ECB :
2255 : /*
2256 : * The value being compared must not change during the evaluation of the
2257 : * window partition.
2258 : */
366 drowley 2259 GIC 87 : if (!is_pseudo_constant_clause((Node *) otherexpr))
366 drowley 2260 UIC 0 : return false;
366 drowley 2261 ECB :
2262 : /* find the window clause belonging to the window function */
366 drowley 2263 GIC 87 : wclause = (WindowClause *) list_nth(subquery->windowClause,
366 drowley 2264 CBC 87 : wfunc->winref - 1);
2265 :
366 drowley 2266 GIC 87 : req.type = T_SupportRequestWFuncMonotonic;
2267 87 : req.window_func = wfunc;
2268 87 : req.window_clause = wclause;
2269 :
366 drowley 2270 ECB : /* call the support function */
366 drowley 2271 EUB : res = (SupportRequestWFuncMonotonic *)
366 drowley 2272 GIC 87 : DatumGetPointer(OidFunctionCall1(prosupport,
2273 : PointerGetDatum(&req)));
366 drowley 2274 ECB :
2275 : /*
2276 : * Nothing to do if the function is neither monotonically increasing nor
2277 : * monotonically decreasing.
2278 : */
366 drowley 2279 CBC 87 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
366 drowley 2280 UIC 0 : return false;
2281 :
366 drowley 2282 GIC 87 : runopexpr = NULL;
366 drowley 2283 CBC 87 : runoperator = InvalidOid;
366 drowley 2284 GIC 87 : opinfos = get_op_btree_interpretation(opexpr->opno);
2285 :
2286 87 : foreach(lc, opinfos)
2287 : {
2288 87 : OpBtreeInterpretation *opinfo = (OpBtreeInterpretation *) lfirst(lc);
2289 87 : int strategy = opinfo->strategy;
366 drowley 2290 ECB :
366 drowley 2291 EUB : /* handle < / <= */
366 drowley 2292 GIC 87 : if (strategy == BTLessStrategyNumber ||
366 drowley 2293 ECB : strategy == BTLessEqualStrategyNumber)
2294 : {
2295 : /*
2296 : * < / <= is supported for monotonically increasing functions in
2297 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2298 : * for monotonically decreasing functions.
2299 : */
366 drowley 2300 CBC 66 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
366 drowley 2301 GIC 9 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2302 : {
366 drowley 2303 CBC 60 : *keep_original = false;
366 drowley 2304 GIC 60 : runopexpr = opexpr;
2305 60 : runoperator = opexpr->opno;
2306 : }
2307 66 : break;
2308 : }
2309 : /* handle > / >= */
2310 21 : else if (strategy == BTGreaterStrategyNumber ||
366 drowley 2311 ECB : strategy == BTGreaterEqualStrategyNumber)
2312 : {
2313 : /*
2314 : * > / >= is supported for monotonically decreasing functions in
2315 : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2316 : * for monotonically increasing functions.
2317 : */
366 drowley 2318 CBC 9 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
366 drowley 2319 GIC 6 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2320 : {
366 drowley 2321 CBC 9 : *keep_original = false;
366 drowley 2322 GIC 9 : runopexpr = opexpr;
2323 9 : runoperator = opexpr->opno;
2324 : }
2325 9 : break;
2326 : }
2327 : /* handle = */
2328 12 : else if (strategy == BTEqualStrategyNumber)
366 drowley 2329 ECB : {
2330 : int16 newstrategy;
2331 :
2332 : /*
2333 : * When both monotonically increasing and decreasing then the
2334 : * return value of the window function will be the same each time.
2335 : * We can simply use 'opexpr' as the run condition without
2336 : * modifying it.
2337 : */
366 drowley 2338 GIC 12 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
366 drowley 2339 ECB : {
366 drowley 2340 GIC 3 : *keep_original = false;
2341 3 : runopexpr = opexpr;
247 2342 3 : runoperator = opexpr->opno;
366 2343 3 : break;
2344 : }
2345 :
2346 : /*
2347 : * When monotonically increasing we make a qual with <wfunc> <=
2348 : * <value> or <value> >= <wfunc> in order to filter out values
366 drowley 2349 ECB : * which are above the value in the equality condition. For
2350 : * monotonically decreasing functions we want to filter values
2351 : * below the value in the equality condition.
2352 : */
366 drowley 2353 CBC 9 : if (res->monotonic & MONOTONICFUNC_INCREASING)
2354 9 : newstrategy = wfunc_left ? BTLessEqualStrategyNumber : BTGreaterEqualStrategyNumber;
2355 : else
366 drowley 2356 UIC 0 : newstrategy = wfunc_left ? BTGreaterEqualStrategyNumber : BTLessEqualStrategyNumber;
2357 :
2358 : /* We must keep the original equality qual */
366 drowley 2359 GIC 9 : *keep_original = true;
2360 9 : runopexpr = opexpr;
2361 :
2362 : /* determine the operator to use for the runCondition qual */
2363 9 : runoperator = get_opfamily_member(opinfo->opfamily_id,
366 drowley 2364 ECB : opinfo->oplefttype,
2365 : opinfo->oprighttype,
2366 : newstrategy);
366 drowley 2367 GBC 9 : break;
2368 : }
2369 : }
366 drowley 2370 ECB :
366 drowley 2371 CBC 87 : if (runopexpr != NULL)
2372 : {
2373 : Expr *newexpr;
366 drowley 2374 ECB :
2375 : /*
2376 : * Build the qual required for the run condition keeping the
2377 : * WindowFunc on the same side as it was originally.
2378 : */
366 drowley 2379 GIC 81 : if (wfunc_left)
2380 72 : newexpr = make_opclause(runoperator,
2381 : runopexpr->opresulttype,
366 drowley 2382 CBC 72 : runopexpr->opretset, (Expr *) wfunc,
2383 : otherexpr, runopexpr->opcollid,
2384 : runopexpr->inputcollid);
2385 : else
366 drowley 2386 GIC 9 : newexpr = make_opclause(runoperator,
2387 : runopexpr->opresulttype,
2388 9 : runopexpr->opretset,
2389 : otherexpr, (Expr *) wfunc,
366 drowley 2390 ECB : runopexpr->opcollid,
2391 : runopexpr->inputcollid);
2392 :
366 drowley 2393 CBC 81 : wclause->runCondition = lappend(wclause->runCondition, newexpr);
2394 :
2395 : /* record that this attno was used in a run condition */
317 drowley 2396 GIC 81 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
317 drowley 2397 ECB : attno - FirstLowInvalidHeapAttributeNumber);
366 drowley 2398 GIC 81 : return true;
366 drowley 2399 ECB : }
2400 :
2401 : /* unsupported OpExpr */
366 drowley 2402 GIC 6 : return false;
2403 : }
366 drowley 2404 ECB :
2405 : /*
2406 : * check_and_push_window_quals
2407 : * Check if 'clause' is a qual that can be pushed into a WindowFunc's
2408 : * WindowClause as a 'runCondition' qual. These, when present, allow
2409 : * some unnecessary work to be skipped during execution.
2410 : *
2411 : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2412 : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
317 2413 : * window quals for.
2414 : *
2415 : * Returns true if the caller still must keep the original qual or false if
2416 : * the caller can safely ignore the original qual because the WindowAgg node
2417 : * will use the runCondition to stop returning tuples.
2418 : */
2419 : static bool
366 drowley 2420 GIC 117 : check_and_push_window_quals(Query *subquery, RangeTblEntry *rte, Index rti,
2421 : Node *clause, Bitmapset **run_cond_attrs)
2422 : {
2423 117 : OpExpr *opexpr = (OpExpr *) clause;
2424 117 : bool keep_original = true;
2425 : Var *var1;
2426 : Var *var2;
2427 :
2428 : /* We're only able to use OpExprs with 2 operands */
2429 117 : if (!IsA(opexpr, OpExpr))
2430 9 : return true;
366 drowley 2431 ECB :
366 drowley 2432 GIC 108 : if (list_length(opexpr->args) != 2)
366 drowley 2433 UIC 0 : return true;
366 drowley 2434 ECB :
123 2435 : /*
2436 : * Currently, we restrict this optimization to strict OpExprs. The reason
2437 : * for this is that during execution, once the runcondition becomes false,
2438 : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2439 : * function result values, we set them to NULL. Having only strict
2440 : * OpExprs here ensures that we properly filter out the tuples with NULLs
2441 : * in the top-level WindowAgg.
2442 : */
123 drowley 2443 CBC 108 : set_opfuncid(opexpr);
123 drowley 2444 GBC 108 : if (!func_strict(opexpr->opfuncid))
123 drowley 2445 UIC 0 : return true;
2446 :
2447 : /*
2448 : * Check for plain Vars that reference window functions in the subquery.
2449 : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2450 : * be used as part of the run condition.
2451 : */
2452 :
2453 : /* Check the left side of the OpExpr */
366 drowley 2454 CBC 108 : var1 = linitial(opexpr->args);
2455 108 : if (IsA(var1, Var) && var1->varattno > 0)
366 drowley 2456 EUB : {
366 drowley 2457 GIC 90 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2458 90 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2459 :
2460 90 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
2461 : opexpr, true, &keep_original,
2462 : run_cond_attrs))
2463 72 : return keep_original;
2464 : }
366 drowley 2465 ECB :
2466 : /* and check the right side */
366 drowley 2467 GIC 36 : var2 = lsecond(opexpr->args);
366 drowley 2468 CBC 36 : if (IsA(var2, Var) && var2->varattno > 0)
366 drowley 2469 ECB : {
366 drowley 2470 GIC 21 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
366 drowley 2471 CBC 21 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2472 :
366 drowley 2473 GIC 21 : if (find_window_run_conditions(subquery, rte, rti, tle->resno, wfunc,
317 drowley 2474 ECB : opexpr, false, &keep_original,
2475 : run_cond_attrs))
366 drowley 2476 GIC 9 : return keep_original;
2477 : }
366 drowley 2478 ECB :
366 drowley 2479 CBC 27 : return true;
2480 : }
366 drowley 2481 ECB :
7937 tgl 2482 : /*
2483 : * set_subquery_pathlist
2589 2484 : * Generate SubqueryScan access paths for a subquery RTE
2485 : *
2486 : * We don't currently support generating parameterized paths for subqueries
3897 2487 : * by pushing join clauses down into them; it seems too expensive to re-plan
2488 : * the subquery multiple times to consider different alternatives.
2489 : * (XXX that could stand to be reconsidered, now that we use Paths.)
2589 2490 : * So the paths made here will be parameterized if the subquery contains
2491 : * LATERAL references, otherwise not. As long as that's true, there's no need
2492 : * for a separate set_subquery_size phase: just make the paths right away.
2493 : */
2494 : static void
6517 tgl 2495 GIC 3695 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2496 : Index rti, RangeTblEntry *rte)
2497 : {
6512 2498 3695 : Query *parse = root->parse;
7937 2499 3695 : Query *subquery = rte->subquery;
2500 : bool trivial_pathtarget;
2501 : Relids required_outer;
2502 : pushdown_safety_info safetyInfo;
2503 : double tuple_fraction;
2504 : RelOptInfo *sub_final_rel;
317 drowley 2505 3695 : Bitmapset *run_cond_attrs = NULL;
2506 : ListCell *lc;
7937 tgl 2507 ECB :
2508 : /*
2509 : * Must copy the Query so that planning doesn't mess up the RTE contents
5143 2510 : * (really really need to fix the planner to not scribble on its input,
3223 2511 : * someday ... but see remove_unused_subquery_outputs to start with).
2512 : */
5143 tgl 2513 GIC 3695 : subquery = copyObject(subquery);
2514 :
2515 : /*
2516 : * If it's a LATERAL subquery, it might contain some Vars of the current
3878 tgl 2517 ECB : * query level, requiring it to be treated as parameterized, even though
2518 : * we don't support pushing down join quals into subqueries.
2519 : */
3878 tgl 2520 GIC 3695 : required_outer = rel->lateral_relids;
2521 :
2522 : /*
2523 : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2524 : * flags as needed while recursing. In particular, we need a workspace
23 drowley 2525 ECB : * for keeping track of the reasons why columns are unsafe to reference.
2526 : * These reasons are stored in the bits inside unsafeFlags[i] when we
2527 : * discover reasons that column i of the subquery is unsafe to be used in
2528 : * a pushed-down qual.
2529 : */
3208 tgl 2530 GIC 3695 : memset(&safetyInfo, 0, sizeof(safetyInfo));
23 drowley 2531 3695 : safetyInfo.unsafeFlags = (unsigned char *)
23 drowley 2532 CBC 3695 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2533 :
2534 : /*
2535 : * If the subquery has the "security_barrier" flag, it means the subquery
2536 : * originated from a view that must enforce row-level security. Then we
2537 : * must not push down quals that contain leaky functions. (Ideally this
2538 : * would be checked inside subquery_is_pushdown_safe, but since we don't
2539 : * currently pass the RTE to that function, we must do it here.)
2540 : */
3208 tgl 2541 GIC 3695 : safetyInfo.unsafeLeaky = rte->security_barrier;
3208 tgl 2542 ECB :
7937 2543 : /*
2544 : * If there are any restriction clauses that have been attached to the
2545 : * subquery relation, consider pushing them down to become WHERE or HAVING
2546 : * quals of the subquery itself. This transformation is useful because it
2547 : * may allow us to generate a better plan for the subquery than evaluating
2548 : * all the subquery output rows and then filtering them.
2549 : *
2550 : * There are several cases where we cannot push down clauses. Restrictions
2551 : * involving the subquery are checked by subquery_is_pushdown_safe().
2552 : * Restrictions on individual clauses are checked by
6126 2553 : * qual_is_pushdown_safe(). Also, we don't want to push down
2554 : * pseudoconstant clauses; better to have the gating node above the
2555 : * subquery.
2556 : *
2557 : * Non-pushed-down clauses will get evaluated as qpquals of the
2558 : * SubqueryScan node.
2559 : *
2560 : * XXX Are there any cases where we want to make a policy decision not to
2561 : * push down a pushable qual, because it'd result in a worse plan?
2562 : */
7528 tgl 2563 GIC 4304 : if (rel->baserestrictinfo != NIL &&
3208 2564 609 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2565 : {
2566 : /* OK to consider pushing down individual quals */
7937 2567 557 : List *upperrestrictlist = NIL;
2568 : ListCell *l;
2569 :
6892 neilc 2570 1302 : foreach(l, rel->baserestrictinfo)
2571 : {
2572 745 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
366 drowley 2573 745 : Node *clause = (Node *) rinfo->clause;
2574 :
23 drowley 2575 CBC 745 : if (rinfo->pseudoconstant)
7937 tgl 2576 ECB : {
23 drowley 2577 GIC 2 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2578 2 : continue;
7937 tgl 2579 ECB : }
2580 :
23 drowley 2581 GIC 743 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
7937 tgl 2582 ECB : {
23 drowley 2583 GIC 420 : case PUSHDOWN_SAFE:
23 drowley 2584 ECB : /* Push it down */
23 drowley 2585 CBC 420 : subquery_push_qual(subquery, rte, rti, clause);
23 drowley 2586 GIC 420 : break;
23 drowley 2587 ECB :
23 drowley 2588 GIC 117 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
23 drowley 2589 ECB :
366 2590 : /*
2591 : * Since we can't push the qual down into the subquery,
2592 : * check if it happens to reference a window function. If
23 2593 : * so then it might be useful to use for the WindowAgg's
2594 : * runCondition.
366 2595 : */
23 drowley 2596 GIC 234 : if (!subquery->hasWindowFuncs ||
23 drowley 2597 CBC 117 : check_and_push_window_quals(subquery, rte, rti, clause,
23 drowley 2598 ECB : &run_cond_attrs))
2599 : {
2600 : /*
2601 : * subquery has no window funcs or the clause is not a
2602 : * suitable window run condition qual or it is, but
2603 : * the original must also be kept in the upper query.
2604 : */
23 drowley 2605 GIC 45 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2606 : }
2607 117 : break;
23 drowley 2608 ECB :
23 drowley 2609 CBC 206 : case PUSHDOWN_UNSAFE:
366 drowley 2610 GIC 206 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
23 2611 206 : break;
2612 : }
2613 : }
7937 tgl 2614 557 : rel->baserestrictinfo = upperrestrictlist;
2615 : /* We don't bother recomputing baserestrict_min_security */
2616 : }
7937 tgl 2617 ECB :
23 drowley 2618 GIC 3695 : pfree(safetyInfo.unsafeFlags);
7290 tgl 2619 ECB :
2620 : /*
3223 2621 : * The upper query might not use all the subquery's output columns; if
317 drowley 2622 : * not, we can simplify. Pass the attributes that were pushed down into
2623 : * WindowAgg run conditions to ensure we don't accidentally think those
2624 : * are unused.
2625 : */
317 drowley 2626 CBC 3695 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2627 :
2628 : /*
2629 : * We can safely pass the outer tuple_fraction down to the subquery if the
6385 bruce 2630 ECB : * outer level has no joining, aggregation, or sorting to do. Otherwise
2631 : * we'd better tell the subquery to plan for full retrieval. (XXX This
2632 : * could probably be made more intelligent ...)
2633 : */
6512 tgl 2634 GIC 3695 : if (parse->hasAggs ||
2635 3194 : parse->groupClause ||
2885 andres 2636 3191 : parse->groupingSets ||
173 tgl 2637 GNC 3191 : root->hasHavingQual ||
6512 tgl 2638 CBC 3191 : parse->distinctClause ||
6512 tgl 2639 GIC 5243 : parse->sortClause ||
2640 2245 : has_multiple_baserels(root))
2641 1947 : tuple_fraction = 0.0; /* default case */
2642 : else
2643 1748 : tuple_fraction = root->tuple_fraction;
2644 :
2645 : /* plan_params should not be in use in current query level */
3868 tgl 2646 CBC 3695 : Assert(root->plan_params == NIL);
3868 tgl 2647 ECB :
2589 2648 : /* Generate a subroot and Paths for the subquery */
2589 tgl 2649 CBC 3695 : rel->subroot = subquery_planner(root->glob, subquery,
5300 tgl 2650 ECB : root,
2589 2651 : false, tuple_fraction);
7937 2652 :
3868 2653 : /* Isolate the params needed by this specific subplan */
3868 tgl 2654 GIC 3695 : rel->subplan_params = root->plan_params;
3868 tgl 2655 CBC 3695 : root->plan_params = NIL;
2656 :
2657 : /*
3955 bruce 2658 ECB : * It's possible that constraint exclusion proved the subquery empty. If
2659 : * so, it's desirable to produce an unadorned dummy path so that we will
2660 : * recognize appropriate optimizations at this query level.
4215 tgl 2661 : */
2589 tgl 2662 GIC 3695 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2663 :
2664 3695 : if (IS_DUMMY_REL(sub_final_rel))
2665 : {
4215 tgl 2666 CBC 54 : set_dummy_rel_pathlist(rel);
2667 54 : return;
2668 : }
2669 :
2670 : /*
2671 : * Mark rel with estimated output rows, width, etc. Note that we have to
2672 : * do this before generating outer-query paths, else cost_subqueryscan is
2673 : * not happy.
2589 tgl 2674 ECB : */
4236 tgl 2675 GIC 3641 : set_subquery_size_estimates(root, rel);
7937 tgl 2676 ECB :
2677 : /*
2678 : * Also detect whether the reltarget is trivial, so that we can pass that
2679 : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2680 : * It's trivial if it fetches all the subplan output columns in order.
2681 : */
264 tgl 2682 GNC 3641 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2683 1099 : trivial_pathtarget = false;
2684 : else
2685 : {
2686 2542 : trivial_pathtarget = true;
2687 7096 : foreach(lc, rel->reltarget->exprs)
2688 : {
2689 4670 : Node *node = (Node *) lfirst(lc);
2690 : Var *var;
2691 :
2692 4670 : if (!IsA(node, Var))
2693 : {
264 tgl 2694 UNC 0 : trivial_pathtarget = false;
2695 0 : break;
2696 : }
264 tgl 2697 GNC 4670 : var = (Var *) node;
2698 4670 : if (var->varno != rti ||
2699 4670 : var->varattno != foreach_current_index(lc) + 1)
2700 : {
2701 116 : trivial_pathtarget = false;
2702 116 : break;
2703 : }
2704 : }
2705 : }
2706 :
2707 : /*
2589 tgl 2708 ECB : * For each Path that subquery_planner produced, make a SubqueryScanPath
2709 : * in the outer query.
2710 : */
2589 tgl 2711 GIC 7507 : foreach(lc, sub_final_rel->pathlist)
2712 : {
2713 3866 : Path *subpath = (Path *) lfirst(lc);
2714 : List *pathkeys;
2715 :
2716 : /* Convert subpath's pathkeys to outer representation */
2589 tgl 2717 CBC 3866 : pathkeys = convert_subquery_pathkeys(root,
2718 : rel,
2719 : subpath->pathkeys,
2720 : make_tlist_from_pathtarget(subpath->pathtarget));
2721 :
2722 : /* Generate outer path using this subpath */
2589 tgl 2723 GIC 3866 : add_path(rel, (Path *)
2589 tgl 2724 CBC 3866 : create_subqueryscan_path(root, rel, subpath,
2725 : trivial_pathtarget,
2589 tgl 2726 ECB : pathkeys, required_outer));
2727 : }
2728 :
1810 rhaas 2729 : /* If outer rel allows parallelism, do same for partial paths. */
1810 rhaas 2730 CBC 3641 : if (rel->consider_parallel && bms_is_empty(required_outer))
2731 : {
1810 rhaas 2732 ECB : /* If consider_parallel is false, there should be no partial paths. */
1810 rhaas 2733 GIC 1687 : Assert(sub_final_rel->consider_parallel ||
2734 : sub_final_rel->partial_pathlist == NIL);
1853 rhaas 2735 ECB :
2736 : /* Same for partial paths. */
1810 rhaas 2737 GBC 1702 : foreach(lc, sub_final_rel->partial_pathlist)
1810 rhaas 2738 EUB : {
1810 rhaas 2739 GIC 15 : Path *subpath = (Path *) lfirst(lc);
1810 rhaas 2740 ECB : List *pathkeys;
2741 :
2742 : /* Convert subpath's pathkeys to outer representation */
1810 rhaas 2743 GIC 15 : pathkeys = convert_subquery_pathkeys(root,
1810 rhaas 2744 ECB : rel,
2745 : subpath->pathkeys,
2746 : make_tlist_from_pathtarget(subpath->pathtarget));
2747 :
2748 : /* Generate outer path using this subpath */
1810 rhaas 2749 GIC 15 : add_partial_path(rel, (Path *)
2750 15 : create_subqueryscan_path(root, rel, subpath,
2751 : trivial_pathtarget,
2752 : pathkeys,
2753 : required_outer));
2754 : }
1853 rhaas 2755 ECB : }
2756 : }
8183 tgl 2757 :
2758 : /*
2759 : * set_function_pathlist
2760 : * Build the (single) access path for a function RTE
7637 2761 : */
2762 : static void
6517 tgl 2763 GIC 17699 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2764 : {
2765 : Relids required_outer;
3426 2766 17699 : List *pathkeys = NIL;
3897 tgl 2767 ECB :
2768 : /*
2769 : * We don't support pushing join clauses into the quals of a function
2770 : * scan, but it could still have required parameterization due to LATERAL
2771 : * refs in the function expression.
2772 : */
3878 tgl 2773 GIC 17699 : required_outer = rel->lateral_relids;
3897 tgl 2774 ECB :
2775 : /*
2776 : * The result is considered unordered unless ORDINALITY was used, in which
3426 2777 : * case it is ordered by the ordinal column (the last one). See if we
2778 : * care, by checking for uses of that Var in equivalence classes.
2779 : */
3426 tgl 2780 GIC 17699 : if (rte->funcordinality)
3426 tgl 2781 ECB : {
3426 tgl 2782 GIC 308 : AttrNumber ordattno = rel->max_attr;
3426 tgl 2783 CBC 308 : Var *var = NULL;
2784 : ListCell *lc;
2785 :
2786 : /*
2607 tgl 2787 ECB : * Is there a Var for it in rel's targetlist? If not, the query did
2788 : * not reference the ordinality column, or at least not in any way
2789 : * that would be interesting for sorting.
2790 : */
2582 tgl 2791 GIC 882 : foreach(lc, rel->reltarget->exprs)
2792 : {
3426 tgl 2793 CBC 879 : Var *node = (Var *) lfirst(lc);
3426 tgl 2794 ECB :
2795 : /* checking varno/varlevelsup is just paranoia */
3426 tgl 2796 GIC 879 : if (IsA(node, Var) &&
2797 879 : node->varattno == ordattno &&
2798 305 : node->varno == rel->relid &&
2799 305 : node->varlevelsup == 0)
2800 : {
2801 305 : var = node;
2802 305 : break;
2803 : }
2804 : }
2805 :
2806 : /*
3426 tgl 2807 ECB : * Try to build pathkeys for this Var with int8 sorting. We tell
2808 : * build_expression_pathkey not to build any new equivalence class; if
2809 : * the Var isn't already mentioned in some EC, it means that nothing
2810 : * cares about the ordering.
2811 : */
3426 tgl 2812 GIC 308 : if (var)
2813 305 : pathkeys = build_expression_pathkey(root,
2814 : (Expr *) var,
2815 : Int8LessOperator,
3426 tgl 2816 ECB : rel->relids,
2817 : false);
2818 : }
2819 :
2820 : /* Generate appropriate path */
3426 tgl 2821 GIC 17699 : add_path(rel, create_functionscan_path(root, rel,
2822 : pathkeys, required_outer));
7637 tgl 2823 CBC 17699 : }
2824 :
6094 mail 2825 ECB : /*
2826 : * set_values_pathlist
2827 : * Build the (single) access path for a VALUES RTE
2828 : */
2829 : static void
6094 mail 2830 GIC 3553 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2831 : {
2832 : Relids required_outer;
2833 :
3892 tgl 2834 ECB : /*
2835 : * We don't support pushing join clauses into the quals of a values scan,
3878 2836 : * but it could still have required parameterization due to LATERAL refs
2837 : * in the values expressions.
2838 : */
3878 tgl 2839 CBC 3553 : required_outer = rel->lateral_relids;
3892 tgl 2840 ECB :
6094 mail 2841 : /* Generate appropriate path */
3892 tgl 2842 CBC 3553 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
6094 mail 2843 GIC 3553 : }
6094 mail 2844 ECB :
2223 alvherre 2845 : /*
2846 : * set_tablefunc_pathlist
2847 : * Build the (single) access path for a table func RTE
2848 : */
2849 : static void
2223 alvherre 2850 GIC 108 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2851 : {
2852 : Relids required_outer;
2853 :
2854 : /*
2223 alvherre 2855 ECB : * We don't support pushing join clauses into the quals of a tablefunc
2856 : * scan, but it could still have required parameterization due to LATERAL
2857 : * refs in the function expression.
2858 : */
2223 alvherre 2859 GIC 108 : required_outer = rel->lateral_relids;
2860 :
2861 : /* Generate appropriate path */
2862 108 : add_path(rel, create_tablefuncscan_path(root, rel,
2863 : required_outer));
2223 alvherre 2864 CBC 108 : }
2865 :
5300 tgl 2866 ECB : /*
2867 : * set_cte_pathlist
2868 : * Build the (single) access path for a non-self-reference CTE RTE
2869 : *
2870 : * There's no need for a separate set_cte_size phase, since we don't
2871 : * support join-qual-parameterized paths for CTEs.
2872 : */
2873 : static void
5300 tgl 2874 GIC 1240 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2875 : {
2876 : Plan *cteplan;
2877 : PlannerInfo *cteroot;
2878 : Index levelsup;
2879 : int ndx;
2880 : ListCell *lc;
2881 : int plan_id;
3878 tgl 2882 ECB : Relids required_outer;
2883 :
2884 : /*
5300 2885 : * Find the referenced CTE, and locate the plan previously made for it.
2886 : */
5300 tgl 2887 GIC 1240 : levelsup = rte->ctelevelsup;
2888 1240 : cteroot = root;
2889 2194 : while (levelsup-- > 0)
2890 : {
2891 954 : cteroot = cteroot->parent_root;
2892 954 : if (!cteroot) /* shouldn't happen */
5300 tgl 2893 LBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
2894 : }
2895 :
2896 : /*
2897 : * Note: cte_plan_ids can be shorter than cteList, if we are still working
2898 : * on planning the CTEs (ie, this is a side-reference from another CTE).
2899 : * So we mustn't use forboth here.
2900 : */
5300 tgl 2901 GIC 1240 : ndx = 0;
5300 tgl 2902 CBC 1808 : foreach(lc, cteroot->parse->cteList)
2903 : {
5300 tgl 2904 GIC 1808 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
5300 tgl 2905 ECB :
5300 tgl 2906 GIC 1808 : if (strcmp(cte->ctename, rte->ctename) == 0)
5300 tgl 2907 CBC 1240 : break;
5300 tgl 2908 GIC 568 : ndx++;
2909 : }
2910 1240 : if (lc == NULL) /* shouldn't happen */
5300 tgl 2911 UIC 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
5300 tgl 2912 GIC 1240 : if (ndx >= list_length(cteroot->cte_plan_ids))
5300 tgl 2913 UIC 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
5300 tgl 2914 GIC 1240 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
353 2915 1240 : if (plan_id <= 0)
353 tgl 2916 UIC 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
5300 tgl 2917 CBC 1240 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2918 :
2919 : /* Mark rel with estimated output rows, width, etc */
2589 tgl 2920 GIC 1240 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
2921 :
2922 : /*
2923 : * We don't support pushing join clauses into the quals of a CTE scan, but
2924 : * it could still have required parameterization due to LATERAL refs in
2925 : * its tlist.
2926 : */
3878 2927 1240 : required_outer = rel->lateral_relids;
2928 :
2929 : /* Generate appropriate path */
3878 tgl 2930 CBC 1240 : add_path(rel, create_ctescan_path(root, rel, required_outer));
5300 2931 1240 : }
5300 tgl 2932 ECB :
2933 : /*
2200 kgrittn 2934 : * set_namedtuplestore_pathlist
2935 : * Build the (single) access path for a named tuplestore RTE
2200 kgrittn 2936 EUB : *
2937 : * There's no need for a separate set_namedtuplestore_size phase, since we
2938 : * don't support join-qual-parameterized paths for tuplestores.
2939 : */
2940 : static void
2200 kgrittn 2941 GIC 219 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
2942 : RangeTblEntry *rte)
2943 : {
2200 kgrittn 2944 ECB : Relids required_outer;
2945 :
2946 : /* Mark rel with estimated output rows, width, etc */
2200 kgrittn 2947 CBC 219 : set_namedtuplestore_size_estimates(root, rel);
2948 :
2200 kgrittn 2949 ECB : /*
2950 : * We don't support pushing join clauses into the quals of a tuplestore
2951 : * scan, but it could still have required parameterization due to LATERAL
2952 : * refs in its tlist.
2953 : */
2200 kgrittn 2954 GBC 219 : required_outer = rel->lateral_relids;
2200 kgrittn 2955 ECB :
2200 kgrittn 2956 EUB : /* Generate appropriate path */
2200 kgrittn 2957 CBC 219 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
2200 kgrittn 2958 ECB :
2200 kgrittn 2959 EUB : /* Select cheapest path (pretty easy in this case...) */
2200 kgrittn 2960 CBC 219 : set_cheapest(rel);
2200 kgrittn 2961 GIC 219 : }
2962 :
1532 tgl 2963 ECB : /*
2964 : * set_result_pathlist
2965 : * Build the (single) access path for an RTE_RESULT RTE
2966 : *
2967 : * There's no need for a separate set_result_size phase, since we
2968 : * don't support join-qual-parameterized paths for these RTEs.
2969 : */
2970 : static void
1532 tgl 2971 GIC 661 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
2972 : RangeTblEntry *rte)
1532 tgl 2973 ECB : {
2974 : Relids required_outer;
2975 :
2976 : /* Mark rel with estimated output rows, width, etc */
1532 tgl 2977 GIC 661 : set_result_size_estimates(root, rel);
2978 :
2979 : /*
2980 : * We don't support pushing join clauses into the quals of a Result scan,
2981 : * but it could still have required parameterization due to LATERAL refs
2982 : * in its tlist.
2983 : */
1532 tgl 2984 CBC 661 : required_outer = rel->lateral_relids;
2985 :
2986 : /* Generate appropriate path */
1532 tgl 2987 GIC 661 : add_path(rel, create_resultscan_path(root, rel, required_outer));
2988 :
2989 : /* Select cheapest path (pretty easy in this case...) */
1532 tgl 2990 CBC 661 : set_cheapest(rel);
1532 tgl 2991 GIC 661 : }
2992 :
2993 : /*
2994 : * set_worktable_pathlist
2995 : * Build the (single) access path for a self-reference CTE RTE
2996 : *
4090 tgl 2997 ECB : * There's no need for a separate set_worktable_size phase, since we don't
2998 : * support join-qual-parameterized paths for CTEs.
2999 : */
5300 3000 : static void
5300 tgl 3001 GIC 357 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3002 : {
2589 tgl 3003 ECB : Path *ctepath;
5300 3004 : PlannerInfo *cteroot;
3005 : Index levelsup;
3006 : Relids required_outer;
3007 :
3008 : /*
3009 : * We need to find the non-recursive term's path, which is in the plan
3010 : * level that's processing the recursive UNION, which is one level *below*
3011 : * where the CTE comes from.
3012 : */
5300 tgl 3013 GIC 357 : levelsup = rte->ctelevelsup;
5300 tgl 3014 CBC 357 : if (levelsup == 0) /* shouldn't happen */
5300 tgl 3015 UIC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
5300 tgl 3016 GIC 357 : levelsup--;
3017 357 : cteroot = root;
3018 792 : while (levelsup-- > 0)
3019 : {
5300 tgl 3020 CBC 435 : cteroot = cteroot->parent_root;
5300 tgl 3021 GIC 435 : if (!cteroot) /* shouldn't happen */
5300 tgl 3022 UIC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3023 : }
2589 tgl 3024 GIC 357 : ctepath = cteroot->non_recursive_path;
3025 357 : if (!ctepath) /* shouldn't happen */
2589 tgl 3026 UIC 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
5300 tgl 3027 ECB :
3028 : /* Mark rel with estimated output rows, width, etc */
2589 tgl 3029 GIC 357 : set_cte_size_estimates(root, rel, ctepath->rows);
5300 tgl 3030 ECB :
3031 : /*
3032 : * We don't support pushing join clauses into the quals of a worktable
3878 3033 : * scan, but it could still have required parameterization due to LATERAL
3522 3034 : * refs in its tlist. (I'm not sure this is actually possible given the
3035 : * restrictions on recursive references, but it's easy enough to support.)
3036 : */
3878 tgl 3037 GIC 357 : required_outer = rel->lateral_relids;
3038 :
3039 : /* Generate appropriate path */
3040 357 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
5300 3041 357 : }
3042 :
3043 : /*
2636 rhaas 3044 ECB : * generate_gather_paths
3045 : * Generate parallel access paths for a relation by pushing a Gather or
3046 : * Gather Merge on top of a partial path.
3047 : *
3048 : * This must not be called until after we're done creating all partial paths
3049 : * for the specified relation. (Otherwise, add_partial_path might delete a
3050 : * path that some GatherPath or GatherMergePath has a reference to.)
3051 : *
3052 : * If we're generating paths for a scan or join relation, override_rows will
3053 : * be false, and we'll just use the relation's size estimate. When we're
3054 : * being called for a partially-grouped path, though, we need to override
3055 : * the rowcount estimate. (It's not clear that the particular value we're
1868 3056 : * using here is actually best, but the underlying rel has no estimate so
3057 : * we must do something.)
2636 rhaas 3058 EUB : */
2636 rhaas 3059 ECB : void
1868 rhaas 3060 CBC 7235 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
2636 rhaas 3061 ECB : {
3062 : Path *cheapest_partial_path;
3063 : Path *simple_gather_path;
2222 3064 : ListCell *lc;
1868 rhaas 3065 EUB : double rows;
1868 rhaas 3066 GIC 7235 : double *rowsp = NULL;
2636 rhaas 3067 ECB :
3068 : /* If there are no partial paths, there's nothing to do here. */
2636 rhaas 3069 GBC 7235 : if (rel->partial_pathlist == NIL)
2636 rhaas 3070 UIC 0 : return;
3071 :
1868 rhaas 3072 ECB : /* Should we override the rel's rowcount estimate? */
1868 rhaas 3073 GIC 7235 : if (override_rows)
3074 826 : rowsp = &rows;
3075 :
3076 : /*
3077 : * The output of Gather is always unsorted, so there's only one partial
3078 : * path of interest: the cheapest one. That will be the one at the front
3079 : * of partial_pathlist because of the way add_partial_path works.
2636 rhaas 3080 ECB : */
2636 rhaas 3081 GIC 7235 : cheapest_partial_path = linitial(rel->partial_pathlist);
1868 3082 7235 : rows =
1868 rhaas 3083 CBC 7235 : cheapest_partial_path->rows * cheapest_partial_path->parallel_workers;
2636 rhaas 3084 ECB : simple_gather_path = (Path *)
2575 rhaas 3085 GIC 7235 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3086 : NULL, rowsp);
2636 3087 7235 : add_path(rel, simple_gather_path);
3088 :
3089 : /*
3090 : * For each useful ordering, we can consider an order-preserving Gather
3091 : * Merge.
3092 : */
2153 bruce 3093 15096 : foreach(lc, rel->partial_pathlist)
3094 : {
3095 7861 : Path *subpath = (Path *) lfirst(lc);
3096 : GatherMergePath *path;
3097 :
2222 rhaas 3098 7861 : if (subpath->pathkeys == NIL)
3099 7082 : continue;
3100 :
1868 3101 779 : rows = subpath->rows * subpath->parallel_workers;
2222 3102 779 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
1868 rhaas 3103 ECB : subpath->pathkeys, NULL, rowsp);
2222 rhaas 3104 GIC 779 : add_path(rel, &path->path);
3105 : }
3106 : }
3107 :
3108 : /*
1097 tomas.vondra 3109 ECB : * get_useful_pathkeys_for_relation
3110 : * Determine which orderings of a relation might be useful.
3111 : *
3112 : * Getting data in sorted order can be useful either because the requested
1097 tomas.vondra 3113 EUB : * order matches the final output ordering for the overall query we're
3114 : * planning, or because it enables an efficient merge join. Here, we try
3115 : * to figure out which pathkeys to consider.
1097 tomas.vondra 3116 ECB : *
3117 : * This allows us to do incremental sort on top of an index scan under a gather
3118 : * merge node, i.e. parallelized.
3119 : *
3120 : * If the require_parallel_safe is true, we also require the expressions to
3121 : * be parallel safe (which allows pushing the sort below Gather Merge).
3122 : *
3123 : * XXX At the moment this can only ever return a list with a single element,
3124 : * because it looks at query_pathkeys only. So we might return the pathkeys
3125 : * directly, but it seems plausible we'll want to consider other orderings
3126 : * in the future. For example, we might want to consider pathkeys useful for
3127 : * merge joins.
3128 : */
3129 : static List *
839 tomas.vondra 3130 CBC 7190 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3131 : bool require_parallel_safe)
3132 : {
1097 tomas.vondra 3133 GIC 7190 : List *useful_pathkeys_list = NIL;
3134 :
3135 : /*
1060 tgl 3136 ECB : * Considering query_pathkeys is always worth it, because it might allow
3137 : * us to avoid a total sort when we have a partially presorted path
887 tomas.vondra 3138 : * available or to push the total sort into the parallel portion of the
3139 : * query.
3140 : */
1097 tomas.vondra 3141 CBC 7190 : if (root->query_pathkeys)
1097 tomas.vondra 3142 ECB : {
3143 : ListCell *lc;
1060 tgl 3144 CBC 3070 : int npathkeys = 0; /* useful pathkeys */
1097 tomas.vondra 3145 ECB :
1097 tomas.vondra 3146 GIC 6784 : foreach(lc, root->query_pathkeys)
1097 tomas.vondra 3147 ECB : {
1097 tomas.vondra 3148 GIC 4455 : PathKey *pathkey = (PathKey *) lfirst(lc);
3149 4455 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3150 :
3151 : /*
3152 : * We can only build a sort for pathkeys that contain a
3153 : * safe-to-compute-early EC member computable from the current
3154 : * relation's reltarget, so ignore the remainder of the list as
3155 : * soon as we find a pathkey without such a member.
3156 : *
3157 : * It's still worthwhile to return any prefix of the pathkeys list
3158 : * that meets this requirement, as we may be able to do an
3159 : * incremental sort.
3160 : *
3161 : * If requested, ensure the sort expression is parallel-safe too.
3162 : */
719 tgl 3163 4455 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3164 : require_parallel_safe))
1097 tomas.vondra 3165 741 : break;
3166 :
3167 3714 : npathkeys++;
3168 : }
3169 :
3170 : /*
3171 : * The whole query_pathkeys list matches, so append it directly, to
3172 : * allow comparing pathkeys easily by comparing list pointer. If we
1060 tgl 3173 ECB : * have to truncate the pathkeys, we gotta do a copy though.
3174 : */
1097 tomas.vondra 3175 GIC 3070 : if (npathkeys == list_length(root->query_pathkeys))
1097 tomas.vondra 3176 CBC 2329 : useful_pathkeys_list = lappend(useful_pathkeys_list,
1097 tomas.vondra 3177 GIC 2329 : root->query_pathkeys);
3178 741 : else if (npathkeys > 0)
3179 204 : useful_pathkeys_list = lappend(useful_pathkeys_list,
270 drowley 3180 GNC 204 : list_copy_head(root->query_pathkeys,
3181 : npathkeys));
3182 : }
3183 :
1097 tomas.vondra 3184 CBC 7190 : return useful_pathkeys_list;
3185 : }
3186 :
1097 tomas.vondra 3187 ECB : /*
3188 : * generate_useful_gather_paths
3189 : * Generate parallel access paths for a relation by pushing a Gather or
3190 : * Gather Merge on top of a partial path.
3191 : *
3192 : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3193 : * paths (aiming to preserve the ordering), but also considers ordering that
3194 : * might be useful for nodes above the gather merge node, and tries to add
3195 : * a sort (regular or incremental) to provide that.
3196 : */
3197 : void
1097 tomas.vondra 3198 GIC 232369 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3199 : {
3200 : ListCell *lc;
3201 : double rows;
3202 232369 : double *rowsp = NULL;
3203 232369 : List *useful_pathkeys_list = NIL;
3204 232369 : Path *cheapest_partial_path = NULL;
3205 :
1097 tomas.vondra 3206 ECB : /* If there are no partial paths, there's nothing to do here. */
1097 tomas.vondra 3207 GIC 232369 : if (rel->partial_pathlist == NIL)
1097 tomas.vondra 3208 CBC 225179 : return;
3209 :
1097 tomas.vondra 3210 ECB : /* Should we override the rel's rowcount estimate? */
1097 tomas.vondra 3211 GIC 7190 : if (override_rows)
3212 781 : rowsp = &rows;
3213 :
3214 : /* generate the regular gather (merge) paths */
3215 7190 : generate_gather_paths(root, rel, override_rows);
3216 :
3217 : /* consider incremental sort for interesting orderings */
839 tomas.vondra 3218 CBC 7190 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
1097 tomas.vondra 3219 ECB :
3220 : /* used for explicit (full) sort paths */
1097 tomas.vondra 3221 CBC 7190 : cheapest_partial_path = linitial(rel->partial_pathlist);
1097 tomas.vondra 3222 ECB :
3223 : /*
3224 : * Consider sorted paths for each interesting ordering. We generate both
3225 : * incremental and full sort.
3226 : */
1097 tomas.vondra 3227 CBC 9723 : foreach(lc, useful_pathkeys_list)
3228 : {
1097 tomas.vondra 3229 GIC 2533 : List *useful_pathkeys = lfirst(lc);
3230 : ListCell *lc2;
3231 : bool is_sorted;
3232 : int presorted_keys;
3233 :
3234 5629 : foreach(lc2, rel->partial_pathlist)
3235 : {
3236 3096 : Path *subpath = (Path *) lfirst(lc2);
3237 : GatherMergePath *path;
3238 :
3239 3096 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3240 : subpath->pathkeys,
1060 tgl 3241 ECB : &presorted_keys);
3242 :
3243 : /*
3244 : * We don't need to consider the case where a subpath is already
1097 tomas.vondra 3245 : * fully sorted because generate_gather_paths already creates a
3246 : * gather merge path for every subpath that has pathkeys present.
3247 : *
3248 : * But since the subpath is already sorted, we know we don't need
3249 : * to consider adding a sort (full or incremental) on top of it,
642 drowley 3250 : * so we can continue here.
1097 tomas.vondra 3251 : */
1097 tomas.vondra 3252 GIC 3096 : if (is_sorted)
3253 540 : continue;
1097 tomas.vondra 3254 ECB :
3255 : /*
3256 : * Try at least sorting the cheapest path and also try
3257 : * incrementally sorting any path which is partially sorted
3258 : * already (no need to deal with paths which have presorted keys
3259 : * when incremental sort is disabled unless it's the cheapest
3260 : * input path).
3261 : */
114 drowley 3262 GNC 2556 : if (subpath != cheapest_partial_path &&
3263 93 : (presorted_keys == 0 || !enable_incremental_sort))
3264 27 : continue;
3265 :
3266 : /*
3267 : * Consider regular sort for any path that's not presorted or if
3268 : * incremental sort is disabled. We've no need to consider both
3269 : * sort and incremental sort on the same path. We assume that
3270 : * incremental sort is always faster when there are presorted
3271 : * keys.
3272 : *
3273 : * This is not redundant with the gather paths created in
1097 tomas.vondra 3274 ECB : * generate_gather_paths, because that doesn't generate ordered
3275 : * output. Here we add an explicit sort to match the useful
3276 : * ordering.
3277 : */
114 drowley 3278 GNC 2529 : if (presorted_keys == 0 || !enable_incremental_sort)
3279 : {
3280 2457 : subpath = (Path *) create_sort_path(root,
3281 : rel,
3282 : subpath,
3283 : useful_pathkeys,
3284 : -1.0);
3285 2457 : rows = subpath->rows * subpath->parallel_workers;
3286 : }
3287 : else
3288 72 : subpath = (Path *) create_incremental_sort_path(root,
3289 : rel,
3290 : subpath,
3291 : useful_pathkeys,
3292 : presorted_keys,
3293 : -1);
3294 2529 : path = create_gather_merge_path(root, rel,
3295 : subpath,
3296 2529 : rel->reltarget,
3297 : subpath->pathkeys,
3298 : NULL,
3299 : rowsp);
3300 :
3301 2529 : add_path(rel, &path->path);
3302 : }
1097 tomas.vondra 3303 ECB : }
3304 : }
3305 :
3306 : /*
3307 : * make_rel_from_joinlist
3308 : * Build access paths using a "joinlist" to guide the join path search.
3309 : *
6319 tgl 3310 : * See comments for deconstruct_jointree() for definition of the joinlist
3311 : * data structure.
3312 : */
3313 : static RelOptInfo *
6319 tgl 3314 GIC 129570 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3315 : {
3316 : int levels_needed;
3317 : List *initial_rels;
3318 : ListCell *jl;
8244 tgl 3319 ECB :
3320 : /*
6319 3321 : * Count the number of child joinlist nodes. This is the depth of the
3322 : * dynamic-programming algorithm we must employ to consider all ways of
3323 : * joining the child nodes.
3324 : */
6319 tgl 3325 GIC 129570 : levels_needed = list_length(joinlist);
8227 tgl 3326 ECB :
8227 tgl 3327 GIC 129570 : if (levels_needed <= 0)
8227 tgl 3328 UIC 0 : return NULL; /* nothing to do? */
3329 :
3330 : /*
3331 : * Construct a list of rels corresponding to the child joinlist nodes.
3332 : * This may contain both base rels and rels constructed according to
3333 : * sub-joinlists.
3334 : */
6319 tgl 3335 GIC 129570 : initial_rels = NIL;
3336 307129 : foreach(jl, joinlist)
3337 : {
3338 177559 : Node *jlnode = (Node *) lfirst(jl);
6319 tgl 3339 ECB : RelOptInfo *thisrel;
3340 :
6319 tgl 3341 GIC 177559 : if (IsA(jlnode, RangeTblRef))
3342 : {
3343 176119 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3344 :
3345 176119 : thisrel = find_base_rel(root, varno);
3346 : }
3347 1440 : else if (IsA(jlnode, List))
3348 : {
3349 : /* Recurse to handle subproblem */
6319 tgl 3350 CBC 1440 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3351 : }
6319 tgl 3352 ECB : else
6319 tgl 3353 EUB : {
6319 tgl 3354 UIC 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3355 : (int) nodeTag(jlnode));
3356 : thisrel = NULL; /* keep compiler quiet */
3357 : }
3358 :
6319 tgl 3359 GIC 177559 : initial_rels = lappend(initial_rels, thisrel);
8227 tgl 3360 ECB : }
3361 :
8227 tgl 3362 GIC 129570 : if (levels_needed == 1)
8227 tgl 3363 ECB : {
3364 : /*
3365 : * Single joinlist node, so we're done.
3366 : */
6892 neilc 3367 GIC 95056 : return (RelOptInfo *) linitial(initial_rels);
8227 tgl 3368 ECB : }
3369 : else
3370 : {
3371 : /*
3372 : * Consider the different orders in which we could join the rels,
3373 : * using a plugin, GEQO, or the regular join search code.
3374 : *
5567 3375 : * We put the initial_rels list into a PlannerInfo field because
3376 : * has_legal_joinclause() needs to look at it (ugly :-().
3377 : */
5567 tgl 3378 GIC 34514 : root->initial_rels = initial_rels;
5567 tgl 3379 EUB :
5674 tgl 3380 GIC 34514 : if (join_search_hook)
5674 tgl 3381 UIC 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
5674 tgl 3382 GIC 34514 : else if (enable_geqo && levels_needed >= geqo_threshold)
8227 3383 3 : return geqo(root, levels_needed, initial_rels);
8227 tgl 3384 ECB : else
5674 tgl 3385 GIC 34511 : return standard_join_search(root, levels_needed, initial_rels);
3386 : }
8244 tgl 3387 ECB : }
3388 :
3389 : /*
3390 : * standard_join_search
3391 : * Find possible joinpaths for a query by successively finding ways
8462 3392 : * to join component relations into join relations.
3393 : *
3394 : * 'levels_needed' is the number of iterations needed, ie, the number of
3395 : * independent jointree items in the query. This is > 1.
3396 : *
3397 : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3398 : * jointree item. These are the components to be joined together.
3399 : * Note that levels_needed == list_length(initial_rels).
3400 : *
3401 : * Returns the final level of join relations, i.e., the relation that is
3402 : * the result of joining all the original relations together.
5674 3403 : * At least one implementation path must be provided for this relation and
3404 : * all required sub-relations.
3405 : *
5674 tgl 3406 EUB : * To support loadable plugins that modify planner behavior by changing the
5674 tgl 3407 ECB : * join searching algorithm, we provide a hook variable that lets a plugin
3408 : * replace or supplement this function. Any such hook must return the same
3409 : * final join relation as the standard code would, but it might have a
3410 : * different set of implementation paths attached, and only the sub-joinrels
3411 : * needed for these paths need have been instantiated.
3412 : *
3413 : * Note to plugin authors: the functions invoked during standard_join_search()
3414 : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3415 : * than one join-order search, you'll probably need to save and restore the
3416 : * original states of those data structures. See geqo_eval() for an example.
3417 : */
3418 : RelOptInfo *
5674 tgl 3419 GIC 34511 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3420 : {
3421 : int lev;
3422 : RelOptInfo *rel;
3423 :
3424 : /*
3425 : * This function cannot be invoked recursively within any one planning
3426 : * problem, so join_rel_level[] can't be in use already.
3427 : */
4880 3428 34511 : Assert(root->join_rel_level == NULL);
3429 :
3430 : /*
3431 : * We employ a simple "dynamic programming" algorithm: we first find all
3432 : * ways to build joins of two jointree items, then all ways to build joins
3433 : * of three items (from two-item joins and single items), then four-item
3434 : * joins, and so on until we have considered all ways to join all the
3435 : * items into one rel.
3436 : *
3437 : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3438 : * set root->join_rel_level[1] to represent all the single-jointree-item
3439 : * relations.
3440 : */
3441 34511 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3442 :
3443 34511 : root->join_rel_level[1] = initial_rels;
9770 scrappy 3444 ECB :
8462 tgl 3445 GIC 82488 : for (lev = 2; lev <= levels_needed; lev++)
3446 : {
3447 : ListCell *lc;
3448 :
3449 : /*
3450 : * Determine all possible pairs of relations to be joined at this
3451 : * level, and build paths for making each one from every available
3452 : * pair of lower-level relations.
9240 bruce 3453 ECB : */
4880 tgl 3454 GIC 47977 : join_search_one_level(root, lev);
3455 :
3456 : /*
3457 : * Run generate_partitionwise_join_paths() and
3458 : * generate_useful_gather_paths() for each just-processed joinrel. We
3459 : * could not do this earlier because both regular and partial paths
3460 : * can get added to a particular joinrel at multiple times within
3461 : * join_search_one_level.
3462 : *
3463 : * After that, we're done creating paths for the joinrel, so run
3464 : * set_cheapest().
3465 : */
4880 tgl 3466 CBC 120992 : foreach(lc, root->join_rel_level[lev])
3467 : {
3468 73015 : rel = (RelOptInfo *) lfirst(lc);
3469 :
1878 peter_e 3470 ECB : /* Create paths for partitionwise joins. */
1878 peter_e 3471 GIC 73015 : generate_partitionwise_join_paths(root, rel);
3472 :
3473 : /*
3474 : * Except for the topmost scan/join rel, consider gathering
3475 : * partial paths. We'll do the same for the topmost scan/join rel
3476 : * once we know the final targetlist (see grouping_planner).
3477 : */
69 tgl 3478 GNC 73015 : if (!bms_equal(rel->relids, root->all_query_rels))
1097 tomas.vondra 3479 CBC 38723 : generate_useful_gather_paths(root, rel, false);
3480 :
3481 : /* Find and save the cheapest paths for this rel */
8454 tgl 3482 GIC 73015 : set_cheapest(rel);
3483 :
3484 : #ifdef OPTIMIZER_DEBUG
3485 : debug_print_rel(root, rel);
3486 : #endif
3487 : }
3488 : }
3489 :
3490 : /*
8227 tgl 3491 ECB : * We should have a single rel at the final level.
3492 : */
4880 tgl 3493 CBC 34511 : if (root->join_rel_level[levels_needed] == NIL)
7053 tgl 3494 UIC 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
4880 tgl 3495 GIC 34511 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
4880 tgl 3496 ECB :
4880 tgl 3497 GIC 34511 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3498 :
3499 34511 : root->join_rel_level = NULL;
3500 :
8462 3501 34511 : return rel;
3502 : }
9770 scrappy 3503 ECB :
3504 : /*****************************************************************************
3505 : * PUSHING QUALS DOWN INTO SUBQUERIES
3506 : *****************************************************************************/
7528 tgl 3507 :
3508 : /*
3509 : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3510 : *
3511 : * subquery is the particular component query being checked. topquery
3512 : * is the top component of a set-operations tree (the same Query if no
3513 : * set-op is involved).
3514 : *
3515 : * Conditions checked here:
3516 : *
3517 : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
7323 3518 : * since that could change the set of rows returned.
7528 tgl 3519 EUB : *
3208 tgl 3520 ECB : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3521 : * quals into it, because that could change the results.
7290 3522 : *
3523 : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
3208 3524 : * This is because upper-level quals should semantically be evaluated only
3525 : * once per distinct row, not once per original row, and if the qual is
3526 : * volatile then extra evaluations could change the results. (This issue
3527 : * does not apply to other forms of aggregation such as GROUP BY, because
3528 : * when those are present we push into HAVING not WHERE, so that the quals
3529 : * are still applied after aggregation.)
3530 : *
3531 : * 4. If the subquery contains window functions, we cannot push volatile quals
3532 : * into it. The issue here is a bit different from DISTINCT: a volatile qual
3533 : * might succeed for some rows of a window partition and fail for others,
3534 : * thereby changing the partition contents and thus the window functions'
3535 : * results for rows that remain.
3536 : *
3537 : * 5. If the subquery contains any set-returning functions in its targetlist,
3538 : * we cannot push volatile quals into it. That would push them below the SRFs
3539 : * and thereby change the number of times they are evaluated. Also, a
3540 : * volatile qual could succeed for some SRF output rows and fail for others,
3541 : * a behavior that cannot occur if it's evaluated before SRF expansion.
3542 : *
3543 : * 6. If the subquery has nonempty grouping sets, we cannot push down any
3544 : * quals. The concern here is that a qual referencing a "constant" grouping
3545 : * column could get constant-folded, which would be improper because the value
3546 : * is potentially nullable by grouping-set expansion. This restriction could
3547 : * be removed if we had a parsetree representation that shows that such
3548 : * grouping columns are not really constant. (There are other ideas that
3549 : * could be used to relax this restriction, but that's the approach most
3550 : * likely to get taken in the future. Note that there's not much to be gained
3551 : * so long as subquery_planner can't move HAVING clauses to WHERE within such
3552 : * a subquery.)
3553 : *
3554 : * In addition, we make several checks on the subquery's output columns to see
3555 : * if it is safe to reference them in pushed-down quals. If output column k
3556 : * is found to be unsafe to reference, we set the reason for that inside
3557 : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
3558 : * column k might not be referenced by some/all quals. The unsafeFlags[]
3559 : * array will be consulted later by qual_is_pushdown_safe(). It's better to
3560 : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
3561 : * because when the subquery involves set operations we have to check the
3562 : * output expressions in each arm of the set op.
3563 : *
3564 : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
3565 : * we're effectively assuming that the quals cannot distinguish values that
3566 : * the DISTINCT's equality operator sees as equal, yet there are many
3567 : * counterexamples to that assumption. However use of such a qual with a
3568 : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
3569 : * "equal" value will be chosen as the output value by the DISTINCT operation.
3570 : * So we don't worry too much about that. Another objection is that if the
3571 : * qual is expensive to evaluate, running it for each original row might cost
3572 : * more than we save by eliminating rows before the DISTINCT step. But it
3573 : * would be very hard to estimate that at this stage, and in practice pushdown
3574 : * seldom seems to make things worse, so we ignore that problem too.
3575 : *
3576 : * Note: likewise, pushing quals into a subquery with window functions is a
3577 : * bit dubious: the quals might remove some rows of a window partition while
3578 : * leaving others, causing changes in the window functions' results for the
3579 : * surviving rows. We insist that such a qual reference only partitioning
3580 : * columns, but again that only protects us if the qual does not distinguish
3581 : * values that the partitioning equality operator sees as equal. The risks
3582 : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
3583 : * occurs and thus there is no theoretical problem with such a qual. But
3584 : * we'll do this anyway because the potential performance benefits are very
3585 : * large, and we've seen no field complaints about the longstanding comparable
3586 : * behavior with DISTINCT.
3587 : */
3588 : static bool
7290 tgl 3589 GIC 683 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
3590 : pushdown_safety_info *safetyInfo)
3591 : {
3592 : SetOperationStmt *topop;
3593 :
3594 : /* Check point 1 */
7323 3595 683 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
7528 3596 46 : return false;
3597 :
3598 : /* Check point 6 */
960 3599 637 : if (subquery->groupClause && subquery->groupingSets)
3600 6 : return false;
3601 :
3602 : /* Check points 3, 4, and 5 */
2385 3603 631 : if (subquery->distinctClause ||
3604 595 : subquery->hasWindowFuncs ||
3605 469 : subquery->hasTargetSRFs)
3208 3606 251 : safetyInfo->unsafeVolatile = true;
3607 :
3608 : /*
3609 : * If we're at a leaf query, check for unsafe expressions in its target
3610 : * list, and mark any reasons why they're unsafe in unsafeFlags[].
3611 : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
3612 : * so no need to check them.)
3613 : */
3595 tgl 3614 CBC 631 : if (subquery->setOperations == NULL)
3208 tgl 3615 GIC 594 : check_output_expressions(subquery, safetyInfo);
3616 :
3617 : /* Are we at top level, or looking at a setop component? */
7528 3618 631 : if (subquery == topquery)
3619 : {
7528 tgl 3620 ECB : /* Top level, so check any component queries */
7528 tgl 3621 CBC 557 : if (subquery->setOperations != NULL)
7290 tgl 3622 GIC 37 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
3623 : safetyInfo))
7528 tgl 3624 LBC 0 : return false;
7528 tgl 3625 ECB : }
3626 : else
3627 : {
3628 : /* Setop component must not have more components (too weird) */
7528 tgl 3629 CBC 74 : if (subquery->setOperations != NULL)
7528 tgl 3630 LBC 0 : return false;
7290 tgl 3631 ECB : /* Check whether setop component output types match top level */
2238 peter_e 3632 GIC 74 : topop = castNode(SetOperationStmt, topquery->setOperations);
3633 74 : Assert(topop);
7290 tgl 3634 74 : compare_tlist_datatypes(subquery->targetList,
3635 : topop->colTypes,
3636 : safetyInfo);
3637 : }
7528 3638 631 : return true;
7528 tgl 3639 ECB : }
3640 :
3641 : /*
3642 : * Helper routine to recurse through setOperations tree
3643 : */
3644 : static bool
7290 tgl 3645 GIC 111 : recurse_pushdown_safe(Node *setOp, Query *topquery,
3208 tgl 3646 ECB : pushdown_safety_info *safetyInfo)
7528 3647 : {
7528 tgl 3648 GIC 111 : if (IsA(setOp, RangeTblRef))
7528 tgl 3649 EUB : {
7528 tgl 3650 GIC 74 : RangeTblRef *rtr = (RangeTblRef *) setOp;
3651 74 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
3652 74 : Query *subquery = rte->subquery;
3653 :
7528 tgl 3654 CBC 74 : Assert(subquery != NULL);
3208 tgl 3655 GBC 74 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
3656 : }
7528 tgl 3657 CBC 37 : else if (IsA(setOp, SetOperationStmt))
7528 tgl 3658 ECB : {
7528 tgl 3659 CBC 37 : SetOperationStmt *op = (SetOperationStmt *) setOp;
3660 :
3661 : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
7528 tgl 3662 GIC 37 : if (op->op == SETOP_EXCEPT)
7528 tgl 3663 LBC 0 : return false;
3664 : /* Else recurse */
3208 tgl 3665 GIC 37 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
7528 tgl 3666 UIC 0 : return false;
3208 tgl 3667 GIC 37 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
7528 tgl 3668 UIC 0 : return false;
3669 : }
7528 tgl 3670 ECB : else
3671 : {
7198 tgl 3672 UIC 0 : elog(ERROR, "unrecognized node type: %d",
7528 tgl 3673 ECB : (int) nodeTag(setOp));
3674 : }
7528 tgl 3675 CBC 37 : return true;
7528 tgl 3676 ECB : }
3677 :
3678 : /*
3595 3679 : * check_output_expressions - check subquery's output expressions for safety
3680 : *
3681 : * There are several cases in which it's unsafe to push down an upper-level
3260 bruce 3682 : * qual if it references a particular output column of a subquery. We check
3683 : * each output column of the subquery and set flags in unsafeFlags[k] when we
23 drowley 3684 : * see that column is unsafe for a pushed-down qual to reference. The
3685 : * conditions checked here are:
3686 : *
3595 tgl 3687 : * 1. We must not push down any quals that refer to subselect outputs that
3595 tgl 3688 EUB : * return sets, else we'd introduce functions-returning-sets into the
3689 : * subquery's WHERE/HAVING quals.
3595 tgl 3690 ECB : *
3595 tgl 3691 EUB : * 2. We must not push down any quals that refer to subselect outputs that
3595 tgl 3692 ECB : * contain volatile functions, for fear of introducing strange results due
3595 tgl 3693 EUB : * to multiple evaluation of a volatile function.
3694 : *
3695 : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
3696 : * refer to non-DISTINCT output columns, because that could change the set
3697 : * of rows returned. (This condition is vacuous for DISTINCT, because then
3698 : * there are no non-DISTINCT output columns, so we needn't check. Note that
3699 : * subquery_is_pushdown_safe already reported that we can't use volatile
3208 tgl 3700 ECB : * quals if there's DISTINCT or DISTINCT ON.)
3701 : *
3702 : * 4. If the subquery has any window functions, we must not push down quals
3703 : * that reference any output columns that are not listed in all the subquery's
3704 : * window PARTITION BY clauses. We can push down quals that use only
3705 : * partitioning columns because they should succeed or fail identically for
3706 : * every row of any one window partition, and totally excluding some
3707 : * partitions will not change a window function's results for remaining
3708 : * partitions. (Again, this also requires nonvolatile quals, but
3709 : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
3710 : * unsafe for this reason can still have WindowClause run conditions pushed
3711 : * down.
3712 : */
3713 : static void
3208 tgl 3714 GIC 594 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
3715 : {
3716 : ListCell *lc;
3717 :
3595 3718 4016 : foreach(lc, subquery->targetList)
3719 : {
3720 3422 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
3721 :
3722 3422 : if (tle->resjunk)
3723 69 : continue; /* ignore resjunk columns */
3724 :
3725 : /* Functions returning sets are unsafe (point 1) */
2399 3726 3353 : if (subquery->hasTargetSRFs &&
23 drowley 3727 295 : (safetyInfo->unsafeFlags[tle->resno] &
3728 295 : UNSAFE_HAS_SET_FUNC) == 0 &&
2399 tgl 3729 295 : expression_returns_set((Node *) tle->expr))
3730 : {
23 drowley 3731 164 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
3595 tgl 3732 164 : continue;
3733 : }
3734 :
3735 : /* Volatile functions are unsafe (point 2) */
23 drowley 3736 3189 : if ((safetyInfo->unsafeFlags[tle->resno] &
3737 3183 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
3738 3183 : contain_volatile_functions((Node *) tle->expr))
3595 tgl 3739 ECB : {
23 drowley 3740 GIC 39 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
3595 tgl 3741 39 : continue;
3742 : }
3595 tgl 3743 ECB :
3744 : /* If subquery uses DISTINCT ON, check point 3 */
3595 tgl 3745 CBC 3150 : if (subquery->hasDistinctOn &&
23 drowley 3746 UIC 0 : (safetyInfo->unsafeFlags[tle->resno] &
23 drowley 3747 LBC 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3595 tgl 3748 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
3749 : {
3750 : /* non-DISTINCT column, so mark it unsafe */
23 drowley 3751 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
3595 tgl 3752 0 : continue;
3595 tgl 3753 ECB : }
3208 3754 :
3755 : /* If subquery uses window functions, check point 4 */
3208 tgl 3756 CBC 3150 : if (subquery->hasWindowFuncs &&
23 drowley 3757 567 : (safetyInfo->unsafeFlags[tle->resno] &
23 drowley 3758 GIC 1086 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
3208 tgl 3759 567 : !targetIsInAllPartitionLists(tle, subquery))
3760 : {
3208 tgl 3761 ECB : /* not present in all PARTITION BY clauses, so mark it unsafe */
23 drowley 3762 CBC 519 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
3208 tgl 3763 519 : continue;
3764 : }
3595 tgl 3765 ECB : }
3595 tgl 3766 CBC 594 : }
3767 :
3768 : /*
3769 : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
3595 tgl 3770 ECB : * push quals into each component query, but the quals can only reference
3595 tgl 3771 EUB : * subquery columns that suffer no type coercions in the set operation.
3772 : * Otherwise there are possible semantic gotchas. So, we check the
3773 : * component queries to see if any of them have output types different from
3774 : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
3775 : * unsafeFlags[k] if column k has different type in any component.
6086 3776 : *
3777 : * We don't have to care about typmods here: the only allowed difference
3778 : * between set-op input and output typmods is input is a specific typmod
3779 : * and output is -1, and that does not require a coercion.
3780 : *
3595 tgl 3781 ECB : * tlist is a subquery tlist.
3782 : * colTypes is an OID list of the top-level setop's output column types.
23 drowley 3783 : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
7290 tgl 3784 : */
3785 : static void
7290 tgl 3786 GIC 74 : compare_tlist_datatypes(List *tlist, List *colTypes,
3208 tgl 3787 ECB : pushdown_safety_info *safetyInfo)
7290 3788 : {
3789 : ListCell *l;
6892 neilc 3790 GIC 74 : ListCell *colType = list_head(colTypes);
7290 tgl 3791 ECB :
6892 neilc 3792 GIC 240 : foreach(l, tlist)
3793 : {
3794 166 : TargetEntry *tle = (TargetEntry *) lfirst(l);
3795 :
6577 tgl 3796 166 : if (tle->resjunk)
7290 tgl 3797 UIC 0 : continue; /* ignore resjunk columns */
6892 neilc 3798 GIC 166 : if (colType == NULL)
7290 tgl 3799 UIC 0 : elog(ERROR, "wrong number of tlist entries");
6577 tgl 3800 GIC 166 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
23 drowley 3801 14 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
1364 tgl 3802 166 : colType = lnext(colTypes, colType);
3803 : }
6892 neilc 3804 74 : if (colType != NULL)
7290 tgl 3805 UIC 0 : elog(ERROR, "wrong number of tlist entries");
7290 tgl 3806 GIC 74 : }
3807 :
3808 : /*
3809 : * targetIsInAllPartitionLists
3810 : * True if the TargetEntry is listed in the PARTITION BY clause
3208 tgl 3811 ECB : * of every window defined in the query.
3812 : *
3813 : * It would be safe to ignore windows not actually used by any window
3814 : * function, but it's not easy to get that info at this stage; and it's
3815 : * unlikely to be useful to spend any extra cycles getting it, since
3816 : * unreferenced window definitions are probably infrequent in practice.
3817 : */
3818 : static bool
3208 tgl 3819 CBC 567 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
3820 : {
3208 tgl 3821 ECB : ListCell *lc;
3208 tgl 3822 EUB :
3208 tgl 3823 CBC 627 : foreach(lc, query->windowClause)
3208 tgl 3824 EUB : {
3208 tgl 3825 CBC 579 : WindowClause *wc = (WindowClause *) lfirst(lc);
3208 tgl 3826 ECB :
3208 tgl 3827 CBC 579 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
3208 tgl 3828 GIC 519 : return false;
3208 tgl 3829 ECB : }
3208 tgl 3830 GBC 48 : return true;
3208 tgl 3831 ECB : }
3832 :
3833 : /*
3834 : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
3835 : *
3836 : * rinfo is a restriction clause applying to the given subquery (whose RTE
3837 : * has index rti in the parent query).
3838 : *
3839 : * Conditions checked here:
3840 : *
3841 : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
3842 : * unclear that it will work correctly: SubLinks will already have been
3843 : * transformed into SubPlans in the qual, but not in the subquery). Note that
741 drowley 3844 : * SubLinks that transform to initplans are safe, and will be accepted here
3845 : * because what we'll see in the qual is just a Param referencing the initplan
3846 : * output.
3847 : *
3848 : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
3849 : * functions.
3208 tgl 3850 : *
3851 : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
741 drowley 3852 : * functions that are passed Var nodes, and therefore might reveal values from
3853 : * the subquery as side effects.
3854 : *
3855 : * 4. rinfo's clause must not refer to the whole-row output of the subquery
3856 : * (since there is no easy way to name that within the subquery itself).
3857 : *
3858 : * 5. rinfo's clause must not refer to any subquery output columns that were
3859 : * found to be unsafe to reference by subquery_is_pushdown_safe().
3860 : */
3861 : static pushdown_safe_type
741 drowley 3862 GIC 743 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
3863 : pushdown_safety_info *safetyInfo)
3864 : {
23 3865 743 : pushdown_safe_type safe = PUSHDOWN_SAFE;
741 3866 743 : Node *qual = (Node *) rinfo->clause;
3867 : List *vars;
3868 : ListCell *vl;
3869 :
3870 : /* Refuse subselects (point 1) */
7323 tgl 3871 743 : if (contain_subplans(qual))
23 drowley 3872 33 : return PUSHDOWN_UNSAFE;
3873 :
3874 : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
3208 tgl 3875 1011 : if (safetyInfo->unsafeVolatile &&
741 drowley 3876 301 : contain_volatile_functions((Node *) rinfo))
23 3877 9 : return PUSHDOWN_UNSAFE;
3878 :
3879 : /* Refuse leaky quals if told to (point 3) */
3208 tgl 3880 839 : if (safetyInfo->unsafeLeaky &&
2904 sfrost 3881 138 : contain_leaked_vars(qual))
23 drowley 3882 69 : return PUSHDOWN_UNSAFE;
3883 :
7323 tgl 3884 ECB : /*
3885 : * Examine all Vars used in clause. Since it's a restriction clause, all
3886 : * such Vars must refer to subselect output columns ... unless this is
3887 : * part of a LATERAL subquery, in which case there could be lateral
3888 : * references.
3889 : *
3890 : * By omitting the relevant flags, this also gives us a cheap sanity check
3891 : * that no aggregates or window functions appear in the qual. Those would
3892 : * be unsafe to push down, but at least for the moment we could never see
3893 : * any in a qual anyhow.
3894 : */
2586 tgl 3895 CBC 632 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
7323 tgl 3896 GIC 1226 : foreach(vl, vars)
3897 : {
7188 bruce 3898 CBC 689 : Var *var = (Var *) lfirst(vl);
7323 tgl 3899 ECB :
5283 3900 : /*
3901 : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
3902 : * It's not clear whether a PHV could safely be pushed down, and even
3903 : * less clear whether such a situation could arise in any cases of
3904 : * practical interest anyway. So for the moment, just refuse to push
3905 : * down.
3906 : */
5283 tgl 3907 GIC 689 : if (!IsA(var, Var))
3908 : {
23 drowley 3909 UIC 0 : safe = PUSHDOWN_UNSAFE;
5283 tgl 3910 0 : break;
3911 : }
3912 :
3913 : /*
3914 : * Punt if we find any lateral references. It would be safe to push
3915 : * these down, but we'd have to convert them into outer references,
3916 : * which subquery_push_qual lacks the infrastructure to do. The case
3917 : * arises so seldom that it doesn't seem worth working hard on.
1000 tgl 3918 ECB : */
1000 tgl 3919 CBC 689 : if (var->varno != rti)
3920 : {
23 drowley 3921 6 : safe = PUSHDOWN_UNSAFE;
1000 tgl 3922 GIC 6 : break;
3923 : }
3924 :
3925 : /* Subqueries have no system columns */
3595 3926 683 : Assert(var->varattno >= 0);
3927 :
3928 : /* Check point 4 */
6264 3929 683 : if (var->varattno == 0)
6264 tgl 3930 ECB : {
23 drowley 3931 UIC 0 : safe = PUSHDOWN_UNSAFE;
6264 tgl 3932 UBC 0 : break;
6264 tgl 3933 EUB : }
3934 :
3935 : /* Check point 5 */
23 drowley 3936 GIC 683 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
3937 : {
3938 245 : if (safetyInfo->unsafeFlags[var->varattno] &
3939 : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
3940 : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
3941 : {
23 drowley 3942 CBC 89 : safe = PUSHDOWN_UNSAFE;
23 drowley 3943 GIC 89 : break;
23 drowley 3944 ECB : }
3945 : else
3946 : {
3947 : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
23 drowley 3948 GIC 156 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
23 drowley 3949 ECB : /* don't break, we might find another Var that's unsafe */
3950 : }
3951 : }
7323 tgl 3952 : }
3953 :
6888 neilc 3954 GBC 632 : list_free(vars);
7323 tgl 3955 EUB :
7323 tgl 3956 GIC 632 : return safe;
3957 : }
3958 :
7528 tgl 3959 ECB : /*
3960 : * subquery_push_qual - push down a qual that we have determined is safe
3961 : */
3962 : static void
6518 tgl 3963 GIC 470 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
3964 : {
7528 tgl 3965 CBC 470 : if (subquery->setOperations != NULL)
7528 tgl 3966 ECB : {
3967 : /* Recurse to push it separately to each component query */
6807 tgl 3968 GIC 25 : recurse_push_qual(subquery->setOperations, subquery,
3969 : rte, rti, qual);
3970 : }
7528 tgl 3971 ECB : else
3972 : {
3973 : /*
3974 : * We need to replace Vars in the qual (which must refer to outputs of
3975 : * the subquery) with copies of the subquery's targetlist expressions.
3976 : * Note that at this point, any uplevel Vars in the qual should have
6385 bruce 3977 : * been replaced with Params, so they need no work.
3978 : *
7522 3979 : * This step also ensures that when we are pushing into a setop tree,
3980 : * each component query gets its own copy of the qual.
3981 : */
3804 tgl 3982 GIC 445 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
3983 : subquery->targetList,
3984 : REPLACEVARS_REPORT_ERROR, 0,
3985 : &subquery->hasSubLinks);
6604 tgl 3986 ECB :
3987 : /*
6385 bruce 3988 : * Now attach the qual to the proper place: normally WHERE, but if the
3989 : * subquery uses grouping or aggregation, put it in HAVING (since the
3990 : * qual really refers to the group-result rows).
6604 tgl 3991 : */
2885 andres 3992 GIC 445 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
6604 tgl 3993 72 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
3994 : else
3995 373 : subquery->jointree->quals =
3996 373 : make_and_qual(subquery->jointree->quals, qual);
3997 :
3998 : /*
3999 : * We need not change the subquery's hasAggs or hasSubLinks flags,
4000 : * since we can't be pushing down any aggregates that weren't there
4001 : * before, and we don't push down subselects at all.
4002 : */
4003 : }
7528 4004 470 : }
7528 tgl 4005 ECB :
4006 : /*
4007 : * Helper routine to recurse through setOperations tree
4008 : */
4009 : static void
7528 tgl 4010 GIC 75 : recurse_push_qual(Node *setOp, Query *topquery,
4011 : RangeTblEntry *rte, Index rti, Node *qual)
4012 : {
4013 75 : if (IsA(setOp, RangeTblRef))
4014 : {
7528 tgl 4015 CBC 50 : RangeTblRef *rtr = (RangeTblRef *) setOp;
6908 4016 50 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
6908 tgl 4017 GIC 50 : Query *subquery = subrte->subquery;
7528 tgl 4018 ECB :
7528 tgl 4019 CBC 50 : Assert(subquery != NULL);
6518 tgl 4020 GIC 50 : subquery_push_qual(subquery, rte, rti, qual);
4021 : }
7528 4022 25 : else if (IsA(setOp, SetOperationStmt))
4023 : {
4024 25 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4025 :
6518 4026 25 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
6518 tgl 4027 CBC 25 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4028 : }
4029 : else
4030 : {
7198 tgl 4031 UIC 0 : elog(ERROR, "unrecognized node type: %d",
4032 : (int) nodeTag(setOp));
7528 tgl 4033 ECB : }
7528 tgl 4034 GIC 75 : }
4035 :
3223 tgl 4036 ECB : /*****************************************************************************
4037 : * SIMPLIFYING SUBQUERY TARGETLISTS
4038 : *****************************************************************************/
4039 :
4040 : /*
4041 : * remove_unused_subquery_outputs
4042 : * Remove subquery targetlist items we don't need
4043 : *
4044 : * It's possible, even likely, that the upper query does not read all the
4045 : * output columns of the subquery. We can remove any such outputs that are
4046 : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4047 : * affect semantics otherwise (e.g., volatile functions can't be removed).
4048 : * This is useful not only because we might be able to remove expensive-to-
4049 : * compute expressions, but because deletion of output columns might allow
4050 : * optimizations such as join removal to occur within the subquery.
4051 : *
4052 : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4053 : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
317 drowley 4054 EUB : * parameter is modifed by the function, so callers must make a copy if they
4055 : * need to use the passed in Bitmapset after calling this function.
4056 : *
3223 tgl 4057 ECB : * To avoid affecting column numbering in the targetlist, we don't physically
4058 : * remove unused tlist entries, but rather replace their expressions with NULL
4059 : * constants. This is implemented by modifying subquery->targetList.
4060 : */
4061 : static void
317 drowley 4062 GIC 3695 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4063 : Bitmapset *extra_used_attrs)
4064 : {
4065 : Bitmapset *attrs_used;
4066 : ListCell *lc;
4067 :
4068 : /*
4069 : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4070 : * the current callers use the Bitmapset after calling this function.
4071 : */
4072 3695 : attrs_used = extra_used_attrs;
4073 :
4074 : /*
4075 : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4076 : * could update all the child SELECTs' tlists, but it seems not worth the
4077 : * trouble presently.
4078 : */
3223 tgl 4079 3695 : if (subquery->setOperations)
4080 712 : return;
4081 :
4082 : /*
4083 : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4084 : * time: all its output columns must be used in the distinctClause.
3223 tgl 4085 ECB : */
3223 tgl 4086 GIC 3191 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4087 107 : return;
4088 :
4089 : /*
4090 : * Collect a bitmap of all the output column numbers used by the upper
4091 : * query.
4092 : *
4093 : * Add all the attributes needed for joins or final output. Note: we must
4094 : * look at rel's targetlist, not the attr_needed data, because attr_needed
3223 tgl 4095 ECB : * isn't computed for inheritance child rels, cf set_append_rel_size().
4096 : * (XXX might be worth changing that sometime.)
4097 : */
2582 tgl 4098 GIC 3084 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4099 :
4100 : /* Add all the attributes used by un-pushed-down restriction clauses. */
3223 4101 3401 : foreach(lc, rel->baserestrictinfo)
3223 tgl 4102 ECB : {
3223 tgl 4103 CBC 317 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4104 :
3223 tgl 4105 GIC 317 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4106 : }
4107 :
4108 : /*
3223 tgl 4109 ECB : * If there's a whole-row reference to the subquery, we can't remove
4110 : * anything.
4111 : */
3223 tgl 4112 GIC 3084 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4113 101 : return;
4114 :
4115 : /*
4116 : * Run through the tlist and zap entries we don't need. It's okay to
4117 : * modify the tlist items in-place because set_subquery_pathlist made a
4118 : * copy of the subquery.
4119 : */
4120 14955 : foreach(lc, subquery->targetList)
3223 tgl 4121 ECB : {
3223 tgl 4122 GIC 11972 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4123 11972 : Node *texpr = (Node *) tle->expr;
3223 tgl 4124 ECB :
4125 : /*
4126 : * If it has a sortgroupref number, it's used in some sort/group
4127 : * clause so we'd better not remove it. Also, don't remove any
4128 : * resjunk columns, since their reason for being has nothing to do
4129 : * with anybody reading the subquery's output. (It's likely that
4130 : * resjunk columns in a sub-SELECT would always have ressortgroupref
4131 : * set, but even if they don't, it seems imprudent to remove them.)
4132 : */
3223 tgl 4133 GIC 11972 : if (tle->ressortgroupref || tle->resjunk)
4134 1595 : continue;
3223 tgl 4135 ECB :
4136 : /*
4137 : * If it's used by the upper query, we can't remove it.
4138 : */
3223 tgl 4139 GIC 10377 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4140 : attrs_used))
4141 5900 : continue;
4142 :
3223 tgl 4143 ECB : /*
4144 : * If it contains a set-returning function, we can't remove it since
4145 : * that could change the number of rows returned by the subquery.
4146 : */
2399 tgl 4147 GIC 4901 : if (subquery->hasTargetSRFs &&
4148 424 : expression_returns_set(texpr))
3223 4149 300 : continue;
4150 :
4151 : /*
4152 : * If it contains volatile functions, we daren't remove it for fear
4153 : * that the user is expecting their side-effects to happen.
4154 : */
4155 4177 : if (contain_volatile_functions(texpr))
3223 tgl 4156 CBC 13 : continue;
3223 tgl 4157 ECB :
4158 : /*
4159 : * OK, we don't need it. Replace the expression with a NULL constant.
4160 : * Preserve the exposed type of the expression, in case something
4161 : * looks at the rowtype of the subquery's result.
4162 : */
3223 tgl 4163 GIC 4164 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
3223 tgl 4164 ECB : exprTypmod(texpr),
4165 : exprCollation(texpr));
4166 : }
4167 : }
4168 :
4169 : /*
2223 rhaas 4170 : * create_partial_bitmap_paths
4171 : * Build partial bitmap heap path for the relation
4172 : */
4173 : void
2223 rhaas 4174 GIC 51609 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4175 : Path *bitmapqual)
4176 : {
4177 : int parallel_workers;
2223 rhaas 4178 ECB : double pages_fetched;
4179 :
4180 : /* Compute heap pages for bitmap heap scan */
2223 rhaas 4181 GIC 51609 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4182 : NULL, NULL);
4183 :
1892 4184 51609 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4185 : max_parallel_workers_per_gather);
2223 rhaas 4186 ECB :
2223 rhaas 4187 GIC 51609 : if (parallel_workers <= 0)
4188 49585 : return;
4189 :
4190 2024 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4191 : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4192 : }
4193 :
4194 : /*
4195 : * Compute the number of parallel workers that should be used to scan a
4196 : * relation. We compute the parallel workers based on the size of the heap to
2244 rhaas 4197 ECB : * be scanned and the size of the index to be scanned, then choose a minimum
4198 : * of those.
4199 : *
4200 : * "heap_pages" is the number of pages from the table that we expect to scan, or
4201 : * -1 if we don't expect to scan any.
4202 : *
4203 : * "index_pages" is the number of pages from the index that we expect to scan, or
2217 4204 : * -1 if we don't expect to scan any.
4205 : *
4206 : * "max_workers" is caller's limit on the number of workers. This typically
1892 4207 : * comes from a GUC.
4208 : */
4209 : int
1892 rhaas 4210 CBC 259853 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
1892 rhaas 4211 ECB : int max_workers)
4212 : {
2244 rhaas 4213 CBC 259853 : int parallel_workers = 0;
4214 :
4215 : /*
4216 : * If the user has set the parallel_workers reloption, use that; otherwise
4217 : * select a default number of workers.
4218 : */
2272 rhaas 4219 GIC 259853 : if (rel->rel_parallel_workers != -1)
4220 912 : parallel_workers = rel->rel_parallel_workers;
4221 : else
4222 : {
4223 : /*
4224 : * If the number of pages being scanned is insufficient to justify a
4225 : * parallel scan, just return zero ... unless it's an inheritance
4226 : * child. In that case, we want to generate a parallel path here
4227 : * anyway. It might not be worthwhile just for this relation, but
4228 : * when combined with all of its inheritance siblings it may well pay
4229 : * off.
4230 : */
2217 4231 258941 : if (rel->reloptkind == RELOPT_BASEREL &&
4232 241139 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
2118 tgl 4233 CBC 7733 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
2272 rhaas 4234 GIC 240828 : return 0;
4235 :
2217 rhaas 4236 CBC 18113 : if (heap_pages >= 0)
4237 : {
4238 : int heap_parallel_threshold;
2217 rhaas 4239 GIC 17183 : int heap_parallel_workers = 1;
4240 :
4241 : /*
2244 rhaas 4242 ECB : * Select the number of workers based on the log of the size of
4243 : * the relation. This probably needs to be a good deal more
4244 : * sophisticated, but we need something here for now. Note that
4245 : * the upper limit of the min_parallel_table_scan_size GUC is
4246 : * chosen to prevent overflow here.
4247 : */
2244 rhaas 4248 GIC 17183 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4249 19459 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4250 : {
4251 2276 : heap_parallel_workers++;
4252 2276 : heap_parallel_threshold *= 3;
4253 2276 : if (heap_parallel_threshold > INT_MAX / 3)
2244 rhaas 4254 LBC 0 : break; /* avoid overflow */
2244 rhaas 4255 ECB : }
4256 :
2244 rhaas 4257 CBC 17183 : parallel_workers = heap_parallel_workers;
4258 : }
2244 rhaas 4259 ECB :
2217 rhaas 4260 GIC 18113 : if (index_pages >= 0)
4261 : {
2217 rhaas 4262 CBC 4608 : int index_parallel_workers = 1;
4263 : int index_parallel_threshold;
4264 :
4265 : /* same calculation as for heap_pages above */
2244 rhaas 4266 GIC 4608 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4267 4734 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4268 : {
4269 126 : index_parallel_workers++;
4270 126 : index_parallel_threshold *= 3;
2244 rhaas 4271 CBC 126 : if (index_parallel_threshold > INT_MAX / 3)
2244 rhaas 4272 LBC 0 : break; /* avoid overflow */
4273 : }
2244 rhaas 4274 ECB :
2244 rhaas 4275 CBC 4608 : if (parallel_workers > 0)
4276 3678 : parallel_workers = Min(parallel_workers, index_parallel_workers);
2244 rhaas 4277 EUB : else
2244 rhaas 4278 GIC 930 : parallel_workers = index_parallel_workers;
4279 : }
2272 rhaas 4280 ECB : }
4281 :
4282 : /* In no case use more than caller supplied maximum number of workers */
1892 rhaas 4283 CBC 19025 : parallel_workers = Min(parallel_workers, max_workers);
4284 :
2272 4285 19025 : return parallel_workers;
4286 : }
4287 :
4288 : /*
1878 peter_e 4289 ECB : * generate_partitionwise_join_paths
4290 : * Create paths representing partitionwise join for given partitioned
4291 : * join relation.
2011 rhaas 4292 : *
4293 : * This must not be called until after we are done adding paths for all
4294 : * child-joins. Otherwise, add_path might delete a path to which some path
2011 rhaas 4295 EUB : * generated here has a reference.
4296 : */
4297 : void
1878 peter_e 4298 CBC 76650 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
2011 rhaas 4299 ECB : {
2011 rhaas 4300 GIC 76650 : List *live_children = NIL;
2011 rhaas 4301 ECB : int cnt_parts;
4302 : int num_parts;
4303 : RelOptInfo **part_rels;
4304 :
4305 : /* Handle only join relations here. */
2011 rhaas 4306 CBC 76650 : if (!IS_JOIN_REL(rel))
2011 rhaas 4307 UIC 0 : return;
2011 rhaas 4308 ECB :
4309 : /* We've nothing to do if the relation is not partitioned. */
1889 rhaas 4310 GIC 76650 : if (!IS_PARTITIONED_REL(rel))
2011 4311 75882 : return;
4312 :
4313 : /* The relation should have consider_partitionwise_join set. */
1682 efujita 4314 768 : Assert(rel->consider_partitionwise_join);
4315 :
4316 : /* Guard against stack overflow due to overly deep partition hierarchy. */
2011 rhaas 4317 768 : check_stack_depth();
4318 :
4319 768 : num_parts = rel->nparts;
4320 768 : part_rels = rel->part_rels;
2011 rhaas 4321 ECB :
4322 : /* Collect non-dummy child-joins. */
2011 rhaas 4323 CBC 2881 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4324 : {
2011 rhaas 4325 GIC 2113 : RelOptInfo *child_rel = part_rels[cnt_parts];
4326 :
4327 : /* If it's been pruned entirely, it's certainly dummy. */
1471 tgl 4328 2113 : if (child_rel == NULL)
1471 tgl 4329 CBC 26 : continue;
1889 rhaas 4330 EUB :
4331 : /* Make partitionwise join paths for this partitioned child-join. */
1878 peter_e 4332 GIC 2087 : generate_partitionwise_join_paths(root, child_rel);
2011 rhaas 4333 ECB :
126 tgl 4334 : /* If we failed to make any path for this child, we must give up. */
126 tgl 4335 GIC 2087 : if (child_rel->pathlist == NIL)
4336 : {
126 tgl 4337 ECB : /*
4338 : * Mark the parent joinrel as unpartitioned so that later
4339 : * functions treat it correctly.
4340 : */
126 tgl 4341 UIC 0 : rel->nparts = 0;
126 tgl 4342 LBC 0 : return;
126 tgl 4343 ECB : }
4344 :
4345 : /* Else, identify the cheapest path for it. */
1494 tgl 4346 CBC 2087 : set_cheapest(child_rel);
4347 :
126 tgl 4348 ECB : /* Dummy children need not be scanned, so ignore those. */
2011 rhaas 4349 GIC 2087 : if (IS_DUMMY_REL(child_rel))
2011 rhaas 4350 UIC 0 : continue;
2011 rhaas 4351 ECB :
4352 : #ifdef OPTIMIZER_DEBUG
4353 : debug_print_rel(root, child_rel);
4354 : #endif
4355 :
2011 rhaas 4356 GIC 2087 : live_children = lappend(live_children, child_rel);
4357 : }
2011 rhaas 4358 ECB :
4359 : /* If all child-joins are dummy, parent join is also dummy. */
2011 rhaas 4360 GIC 768 : if (!live_children)
4361 : {
2011 rhaas 4362 UIC 0 : mark_dummy_rel(rel);
4363 0 : return;
2011 rhaas 4364 EUB : }
4365 :
4366 : /* Build additional paths for this rel from child-join paths. */
2011 rhaas 4367 GIC 768 : add_paths_to_append_rel(root, rel, live_children);
4368 768 : list_free(live_children);
2011 rhaas 4369 ECB : }
4370 :
4371 :
7528 tgl 4372 : /*****************************************************************************
7528 tgl 4373 EUB : * DEBUG SUPPORT
4374 : *****************************************************************************/
4375 :
4376 : #ifdef OPTIMIZER_DEBUG
4377 :
4378 : static void
2535 tgl 4379 ECB : print_relids(PlannerInfo *root, Relids relids)
4380 : {
4381 : int x;
4382 : bool first = true;
7843 4383 :
4384 : x = -1;
3054 tgl 4385 EUB : while ((x = bms_next_member(relids, x)) >= 0)
7843 4386 : {
4387 : if (!first)
4388 : printf(" ");
4389 : if (x < root->simple_rel_array_size &&
2535 tgl 4390 ECB : root->simple_rte_array[x])
4391 : printf("%s", root->simple_rte_array[x]->eref->aliasname);
4392 : else
4393 : printf("%d", x);
4394 : first = false;
4395 : }
4396 : }
4397 :
4398 : static void
4399 : print_restrictclauses(PlannerInfo *root, List *clauses)
4400 : {
4401 : ListCell *l;
4402 :
4403 : foreach(l, clauses)
4404 : {
4405 : RestrictInfo *c = lfirst(l);
4406 :
4407 : print_expr((Node *) c->clause, root->parse->rtable);
4408 : if (lnext(clauses, l))
4409 : printf(", ");
4410 : }
4411 : }
4412 :
4413 : static void
4414 : print_path(PlannerInfo *root, Path *path, int indent)
4415 : {
4416 : const char *ptype;
4417 : bool join = false;
4418 : Path *subpath = NULL;
4419 : int i;
4420 :
4421 : switch (nodeTag(path))
4422 : {
4423 : case T_Path:
4424 : switch (path->pathtype)
4425 : {
4426 : case T_SeqScan:
4427 : ptype = "SeqScan";
4428 : break;
4429 : case T_SampleScan:
4430 : ptype = "SampleScan";
4431 : break;
4432 : case T_FunctionScan:
4433 : ptype = "FunctionScan";
4434 : break;
4435 : case T_TableFuncScan:
4436 : ptype = "TableFuncScan";
4437 : break;
4438 : case T_ValuesScan:
4439 : ptype = "ValuesScan";
4440 : break;
4441 : case T_CteScan:
4442 : ptype = "CteScan";
4443 : break;
4444 : case T_NamedTuplestoreScan:
4445 : ptype = "NamedTuplestoreScan";
4446 : break;
4447 : case T_Result:
4448 : ptype = "Result";
4449 : break;
4450 : case T_WorkTableScan:
4451 : ptype = "WorkTableScan";
4452 : break;
4453 : default:
4454 : ptype = "???Path";
4455 : break;
4456 : }
4457 : break;
4458 : case T_IndexPath:
4459 : ptype = "IdxScan";
4460 : break;
4461 : case T_BitmapHeapPath:
4462 : ptype = "BitmapHeapScan";
4463 : break;
4464 : case T_BitmapAndPath:
4465 : ptype = "BitmapAndPath";
4466 : break;
4467 : case T_BitmapOrPath:
4468 : ptype = "BitmapOrPath";
4469 : break;
4470 : case T_TidPath:
4471 : ptype = "TidScan";
4472 : break;
4473 : case T_SubqueryScanPath:
4474 : ptype = "SubqueryScan";
4475 : break;
4476 : case T_ForeignPath:
4477 : ptype = "ForeignScan";
4478 : break;
4479 : case T_CustomPath:
4480 : ptype = "CustomScan";
4481 : break;
4482 : case T_NestPath:
4483 : ptype = "NestLoop";
4484 : join = true;
4485 : break;
4486 : case T_MergePath:
4487 : ptype = "MergeJoin";
4488 : join = true;
4489 : break;
4490 : case T_HashPath:
4491 : ptype = "HashJoin";
4492 : join = true;
4493 : break;
4494 : case T_AppendPath:
4495 : ptype = "Append";
4496 : break;
4497 : case T_MergeAppendPath:
4498 : ptype = "MergeAppend";
4499 : break;
4500 : case T_GroupResultPath:
4501 : ptype = "GroupResult";
4502 : break;
4503 : case T_MaterialPath:
4504 : ptype = "Material";
4505 : subpath = ((MaterialPath *) path)->subpath;
4506 : break;
4507 : case T_MemoizePath:
4508 : ptype = "Memoize";
4509 : subpath = ((MemoizePath *) path)->subpath;
4510 : break;
4511 : case T_UniquePath:
4512 : ptype = "Unique";
4513 : subpath = ((UniquePath *) path)->subpath;
4514 : break;
4515 : case T_GatherPath:
4516 : ptype = "Gather";
4517 : subpath = ((GatherPath *) path)->subpath;
4518 : break;
4519 : case T_GatherMergePath:
4520 : ptype = "GatherMerge";
4521 : subpath = ((GatherMergePath *) path)->subpath;
4522 : break;
4523 : case T_ProjectionPath:
4524 : ptype = "Projection";
4525 : subpath = ((ProjectionPath *) path)->subpath;
4526 : break;
4527 : case T_ProjectSetPath:
4528 : ptype = "ProjectSet";
4529 : subpath = ((ProjectSetPath *) path)->subpath;
4530 : break;
4531 : case T_SortPath:
4532 : ptype = "Sort";
4533 : subpath = ((SortPath *) path)->subpath;
4534 : break;
4535 : case T_IncrementalSortPath:
4536 : ptype = "IncrementalSort";
4537 : subpath = ((SortPath *) path)->subpath;
4538 : break;
4539 : case T_GroupPath:
4540 : ptype = "Group";
4541 : subpath = ((GroupPath *) path)->subpath;
4542 : break;
4543 : case T_UpperUniquePath:
4544 : ptype = "UpperUnique";
4545 : subpath = ((UpperUniquePath *) path)->subpath;
4546 : break;
4547 : case T_AggPath:
4548 : ptype = "Agg";
4549 : subpath = ((AggPath *) path)->subpath;
4550 : break;
4551 : case T_GroupingSetsPath:
4552 : ptype = "GroupingSets";
4553 : subpath = ((GroupingSetsPath *) path)->subpath;
4554 : break;
4555 : case T_MinMaxAggPath:
4556 : ptype = "MinMaxAgg";
4557 : break;
4558 : case T_WindowAggPath:
4559 : ptype = "WindowAgg";
4560 : subpath = ((WindowAggPath *) path)->subpath;
4561 : break;
4562 : case T_SetOpPath:
4563 : ptype = "SetOp";
4564 : subpath = ((SetOpPath *) path)->subpath;
4565 : break;
4566 : case T_RecursiveUnionPath:
4567 : ptype = "RecursiveUnion";
4568 : break;
4569 : case T_LockRowsPath:
4570 : ptype = "LockRows";
4571 : subpath = ((LockRowsPath *) path)->subpath;
4572 : break;
4573 : case T_ModifyTablePath:
4574 : ptype = "ModifyTable";
4575 : break;
4576 : case T_LimitPath:
4577 : ptype = "Limit";
4578 : subpath = ((LimitPath *) path)->subpath;
4579 : break;
4580 : default:
4581 : ptype = "???Path";
4582 : break;
4583 : }
4584 :
4585 : for (i = 0; i < indent; i++)
4586 : printf("\t");
4587 : printf("%s", ptype);
4588 :
4589 : if (path->parent)
4590 : {
4591 : printf("(");
4592 : print_relids(root, path->parent->relids);
4593 : printf(")");
4594 : }
4595 : if (path->param_info)
4596 : {
4597 : printf(" required_outer (");
4598 : print_relids(root, path->param_info->ppi_req_outer);
4599 : printf(")");
4600 : }
4601 : printf(" rows=%.0f cost=%.2f..%.2f\n",
4602 : path->rows, path->startup_cost, path->total_cost);
4603 :
4604 : if (path->pathkeys)
4605 : {
4606 : for (i = 0; i < indent; i++)
4607 : printf("\t");
4608 : printf(" pathkeys: ");
4609 : print_pathkeys(path->pathkeys, root->parse->rtable);
4610 : }
4611 :
4612 : if (join)
4613 : {
4614 : JoinPath *jp = (JoinPath *) path;
4615 :
4616 : for (i = 0; i < indent; i++)
4617 : printf("\t");
4618 : printf(" clauses: ");
4619 : print_restrictclauses(root, jp->joinrestrictinfo);
4620 : printf("\n");
4621 :
4622 : if (IsA(path, MergePath))
4623 : {
4624 : MergePath *mp = (MergePath *) path;
4625 :
4626 : for (i = 0; i < indent; i++)
4627 : printf("\t");
4628 : printf(" sortouter=%d sortinner=%d materializeinner=%d\n",
4629 : ((mp->outersortkeys) ? 1 : 0),
4630 : ((mp->innersortkeys) ? 1 : 0),
4631 : ((mp->materialize_inner) ? 1 : 0));
4632 : }
4633 :
4634 : print_path(root, jp->outerjoinpath, indent + 1);
4635 : print_path(root, jp->innerjoinpath, indent + 1);
4636 : }
4637 :
4638 : if (subpath)
4639 : print_path(root, subpath, indent + 1);
4640 : }
4641 :
4642 : void
4643 : debug_print_rel(PlannerInfo *root, RelOptInfo *rel)
4644 : {
4645 : ListCell *l;
4646 :
4647 : printf("RELOPTINFO (");
4648 : print_relids(root, rel->relids);
4649 : printf("): rows=%.0f width=%d\n", rel->rows, rel->reltarget->width);
4650 :
4651 : if (rel->baserestrictinfo)
4652 : {
4653 : printf("\tbaserestrictinfo: ");
4654 : print_restrictclauses(root, rel->baserestrictinfo);
4655 : printf("\n");
4656 : }
4657 :
4658 : if (rel->joininfo)
4659 : {
4660 : printf("\tjoininfo: ");
4661 : print_restrictclauses(root, rel->joininfo);
4662 : printf("\n");
4663 : }
4664 :
4665 : printf("\tpath list:\n");
4666 : foreach(l, rel->pathlist)
4667 : print_path(root, lfirst(l), 1);
4668 : if (rel->cheapest_parameterized_paths)
4669 : {
4670 : printf("\n\tcheapest parameterized paths:\n");
4671 : foreach(l, rel->cheapest_parameterized_paths)
4672 : print_path(root, lfirst(l), 1);
4673 : }
4674 : if (rel->cheapest_startup_path)
4675 : {
4676 : printf("\n\tcheapest startup path:\n");
4677 : print_path(root, rel->cheapest_startup_path, 1);
4678 : }
4679 : if (rel->cheapest_total_path)
4680 : {
4681 : printf("\n\tcheapest total path:\n");
4682 : print_path(root, rel->cheapest_total_path, 1);
4683 : }
4684 : printf("\n");
4685 : fflush(stdout);
4686 : }
4687 :
4688 : #endif /* OPTIMIZER_DEBUG */
|
