Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * allpaths.c
4 : : * Routines to find possible search paths for processing a query
5 : : *
6 : : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : *
10 : : * IDENTIFICATION
11 : : * src/backend/optimizer/path/allpaths.c
12 : : *
13 : : *-------------------------------------------------------------------------
14 : : */
15 : :
16 : : #include "postgres.h"
17 : :
18 : : #include <limits.h>
19 : : #include <math.h>
20 : :
21 : : #include "access/sysattr.h"
22 : : #include "access/tsmapi.h"
23 : : #include "catalog/pg_class.h"
24 : : #include "catalog/pg_operator.h"
25 : : #include "catalog/pg_proc.h"
26 : : #include "foreign/fdwapi.h"
27 : : #include "miscadmin.h"
28 : : #include "nodes/makefuncs.h"
29 : : #include "nodes/nodeFuncs.h"
30 : : #include "nodes/supportnodes.h"
31 : : #ifdef OPTIMIZER_DEBUG
32 : : #include "nodes/print.h"
33 : : #endif
34 : : #include "optimizer/appendinfo.h"
35 : : #include "optimizer/clauses.h"
36 : : #include "optimizer/cost.h"
37 : : #include "optimizer/geqo.h"
38 : : #include "optimizer/optimizer.h"
39 : : #include "optimizer/pathnode.h"
40 : : #include "optimizer/paths.h"
41 : : #include "optimizer/plancat.h"
42 : : #include "optimizer/planner.h"
43 : : #include "optimizer/prep.h"
44 : : #include "optimizer/tlist.h"
45 : : #include "parser/parse_clause.h"
46 : : #include "parser/parsetree.h"
47 : : #include "partitioning/partbounds.h"
48 : : #include "port/pg_bitutils.h"
49 : : #include "rewrite/rewriteManip.h"
50 : : #include "utils/lsyscache.h"
51 : : #include "utils/selfuncs.h"
52 : :
53 : :
54 : : /* Bitmask flags for pushdown_safety_info.unsafeFlags */
55 : : #define UNSAFE_HAS_VOLATILE_FUNC (1 << 0)
56 : : #define UNSAFE_HAS_SET_FUNC (1 << 1)
57 : : #define UNSAFE_NOTIN_DISTINCTON_CLAUSE (1 << 2)
58 : : #define UNSAFE_NOTIN_PARTITIONBY_CLAUSE (1 << 3)
59 : : #define UNSAFE_TYPE_MISMATCH (1 << 4)
60 : :
61 : : /* results of subquery_is_pushdown_safe */
62 : : typedef struct pushdown_safety_info
63 : : {
64 : : unsigned char *unsafeFlags; /* bitmask of reasons why this target list
65 : : * column is unsafe for qual pushdown, or 0 if
66 : : * no reason. */
67 : : bool unsafeVolatile; /* don't push down volatile quals */
68 : : bool unsafeLeaky; /* don't push down leaky quals */
69 : : } pushdown_safety_info;
70 : :
71 : : /* Return type for qual_is_pushdown_safe */
72 : : typedef enum pushdown_safe_type
73 : : {
74 : : PUSHDOWN_UNSAFE, /* unsafe to push qual into subquery */
75 : : PUSHDOWN_SAFE, /* safe to push qual into subquery */
76 : : PUSHDOWN_WINDOWCLAUSE_RUNCOND, /* unsafe, but may work as WindowClause
77 : : * run condition */
78 : : } pushdown_safe_type;
79 : :
80 : : /* These parameters are set by GUC */
81 : : bool enable_geqo = false; /* just in case GUC doesn't set it */
82 : : bool enable_eager_aggregate = true;
83 : : int geqo_threshold;
84 : : double min_eager_agg_group_size;
85 : : int min_parallel_table_scan_size;
86 : : int min_parallel_index_scan_size;
87 : :
88 : : /* Hook for plugins to get control in set_rel_pathlist() */
89 : : set_rel_pathlist_hook_type set_rel_pathlist_hook = NULL;
90 : :
91 : : /* Hook for plugins to replace standard_join_search() */
92 : : join_search_hook_type join_search_hook = NULL;
93 : :
94 : :
95 : : static void set_base_rel_consider_startup(PlannerInfo *root);
96 : : static void set_base_rel_sizes(PlannerInfo *root);
97 : : static void setup_simple_grouped_rels(PlannerInfo *root);
98 : : static void set_base_rel_pathlists(PlannerInfo *root);
99 : : static void set_rel_size(PlannerInfo *root, RelOptInfo *rel,
100 : : Index rti, RangeTblEntry *rte);
101 : : static void set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
102 : : Index rti, RangeTblEntry *rte);
103 : : static void set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel,
104 : : RangeTblEntry *rte);
105 : : static void create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel);
106 : : static void set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
107 : : RangeTblEntry *rte);
108 : : static void set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
109 : : RangeTblEntry *rte);
110 : : static void set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel,
111 : : RangeTblEntry *rte);
112 : : static void set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
113 : : RangeTblEntry *rte);
114 : : static void set_foreign_size(PlannerInfo *root, RelOptInfo *rel,
115 : : RangeTblEntry *rte);
116 : : static void set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel,
117 : : RangeTblEntry *rte);
118 : : static void set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
119 : : Index rti, RangeTblEntry *rte);
120 : : static void set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
121 : : Index rti, RangeTblEntry *rte);
122 : : static void set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel);
123 : : static void generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
124 : : List *live_childrels,
125 : : List *all_child_pathkeys);
126 : : static Path *get_cheapest_parameterized_child_path(PlannerInfo *root,
127 : : RelOptInfo *rel,
128 : : Relids required_outer);
129 : : static void accumulate_append_subpath(Path *path,
130 : : List **subpaths,
131 : : List **special_subpaths);
132 : : static Path *get_singleton_append_subpath(Path *path);
133 : : static void set_dummy_rel_pathlist(RelOptInfo *rel);
134 : : static void set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
135 : : Index rti, RangeTblEntry *rte);
136 : : static void set_function_pathlist(PlannerInfo *root, RelOptInfo *rel,
137 : : RangeTblEntry *rte);
138 : : static void set_values_pathlist(PlannerInfo *root, RelOptInfo *rel,
139 : : RangeTblEntry *rte);
140 : : static void set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel,
141 : : RangeTblEntry *rte);
142 : : static void set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel,
143 : : RangeTblEntry *rte);
144 : : static void set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
145 : : RangeTblEntry *rte);
146 : : static void set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
147 : : RangeTblEntry *rte);
148 : : static void set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel,
149 : : RangeTblEntry *rte);
150 : : static RelOptInfo *make_rel_from_joinlist(PlannerInfo *root, List *joinlist);
151 : : static bool subquery_is_pushdown_safe(Query *subquery, Query *topquery,
152 : : pushdown_safety_info *safetyInfo);
153 : : static bool recurse_pushdown_safe(Node *setOp, Query *topquery,
154 : : pushdown_safety_info *safetyInfo);
155 : : static void check_output_expressions(Query *subquery,
156 : : pushdown_safety_info *safetyInfo);
157 : : static void compare_tlist_datatypes(List *tlist, List *colTypes,
158 : : pushdown_safety_info *safetyInfo);
159 : : static bool targetIsInAllPartitionLists(TargetEntry *tle, Query *query);
160 : : static pushdown_safe_type qual_is_pushdown_safe(Query *subquery, Index rti,
161 : : RestrictInfo *rinfo,
162 : : pushdown_safety_info *safetyInfo);
163 : : static void subquery_push_qual(Query *subquery,
164 : : RangeTblEntry *rte, Index rti, Node *qual);
165 : : static void recurse_push_qual(Node *setOp, Query *topquery,
166 : : RangeTblEntry *rte, Index rti, Node *qual);
167 : : static void remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
168 : : Bitmapset *extra_used_attrs);
169 : :
170 : :
171 : : /*
172 : : * make_one_rel
173 : : * Finds all possible access paths for executing a query, returning a
174 : : * single rel that represents the join of all base rels in the query.
175 : : */
176 : : RelOptInfo *
7251 tgl@sss.pgh.pa.us 177 :CBC 162949 : make_one_rel(PlannerInfo *root, List *joinlist)
178 : : {
179 : : RelOptInfo *rel;
180 : : Index rti;
181 : : double total_pages;
182 : :
183 : : /* Mark base rels as to whether we care about fast-start plans */
3799 184 : 162949 : set_base_rel_consider_startup(root);
185 : :
186 : : /*
187 : : * Compute size estimates and consider_parallel flags for each base rel.
188 : : */
5022 189 : 162949 : set_base_rel_sizes(root);
190 : :
191 : : /*
192 : : * Build grouped relations for simple rels (i.e., base or "other" member
193 : : * relations) where possible.
194 : : */
19 rguo@postgresql.org 195 :GNC 162932 : setup_simple_grouped_rels(root);
196 : :
197 : : /*
198 : : * We should now have size estimates for every actual table involved in
199 : : * the query, and we also know which if any have been deleted from the
200 : : * query by join removal, pruned by partition pruning, or eliminated by
201 : : * constraint exclusion. So we can now compute total_table_pages.
202 : : *
203 : : * Note that appendrels are not double-counted here, even though we don't
204 : : * bother to distinguish RelOptInfos for appendrel parents, because the
205 : : * parents will have pages = 0.
206 : : *
207 : : * XXX if a table is self-joined, we will count it once per appearance,
208 : : * which perhaps is the wrong thing ... but that's not completely clear,
209 : : * and detecting self-joins here is difficult, so ignore it for now.
210 : : */
2546 tgl@sss.pgh.pa.us 211 :CBC 162932 : total_pages = 0;
212 [ + + ]: 504333 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
213 : : {
214 : 341401 : RelOptInfo *brel = root->simple_rel_array[rti];
215 : :
216 : : /* there may be empty slots corresponding to non-baserel RTEs */
217 [ + + ]: 341401 : if (brel == NULL)
218 : 80263 : continue;
219 : :
220 [ - + ]: 261138 : Assert(brel->relid == rti); /* sanity check on array */
221 : :
222 [ + + ]: 261138 : if (IS_DUMMY_REL(brel))
223 : 698 : continue;
224 : :
225 [ + + + - ]: 260440 : if (IS_SIMPLE_REL(brel))
226 : 260440 : total_pages += (double) brel->pages;
227 : : }
228 : 162932 : root->total_table_pages = total_pages;
229 : :
230 : : /*
231 : : * Generate access paths for each base rel.
232 : : */
9115 233 : 162932 : set_base_rel_pathlists(root);
234 : :
235 : : /*
236 : : * Generate access paths for the entire join tree.
237 : : */
7251 238 : 162932 : rel = make_rel_from_joinlist(root, joinlist);
239 : :
240 : : /*
241 : : * The result should join all and only the query's base + outer-join rels.
242 : : */
1001 243 [ - + ]: 162932 : Assert(bms_equal(rel->relids, root->all_query_rels));
244 : :
5022 245 : 162932 : return rel;
246 : : }
247 : :
248 : : /*
249 : : * set_base_rel_consider_startup
250 : : * Set the consider_[param_]startup flags for each base-relation entry.
251 : : *
252 : : * For the moment, we only deal with consider_param_startup here; because the
253 : : * logic for consider_startup is pretty trivial and is the same for every base
254 : : * relation, we just let build_simple_rel() initialize that flag correctly to
255 : : * start with. If that logic ever gets more complicated it would probably
256 : : * be better to move it here.
257 : : */
258 : : static void
3799 259 : 162949 : set_base_rel_consider_startup(PlannerInfo *root)
260 : : {
261 : : /*
262 : : * Since parameterized paths can only be used on the inside of a nestloop
263 : : * join plan, there is usually little value in considering fast-start
264 : : * plans for them. However, for relations that are on the RHS of a SEMI
265 : : * or ANTI join, a fast-start plan can be useful because we're only going
266 : : * to care about fetching one tuple anyway.
267 : : *
268 : : * To minimize growth of planning time, we currently restrict this to
269 : : * cases where the RHS is a single base relation, not a join; there is no
270 : : * provision for consider_param_startup to get set at all on joinrels.
271 : : * Also we don't worry about appendrels. costsize.c's costing rules for
272 : : * nestloop semi/antijoins don't consider such cases either.
273 : : */
274 : : ListCell *lc;
275 : :
276 [ + + + + : 184068 : foreach(lc, root->join_info_list)
+ + ]
277 : : {
278 : 21119 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
279 : : int varno;
280 : :
281 [ + + + + : 25384 : if ((sjinfo->jointype == JOIN_SEMI || sjinfo->jointype == JOIN_ANTI) &&
+ + ]
282 : 4265 : bms_get_singleton_member(sjinfo->syn_righthand, &varno))
283 : : {
284 : 4165 : RelOptInfo *rel = find_base_rel(root, varno);
285 : :
286 : 4165 : rel->consider_param_startup = true;
287 : : }
288 : : }
289 : 162949 : }
290 : :
291 : : /*
292 : : * set_base_rel_sizes
293 : : * Set the size estimates (rows and widths) for each base-relation entry.
294 : : * Also determine whether to consider parallel paths for base relations.
295 : : *
296 : : * We do this in a separate pass over the base rels so that rowcount
297 : : * estimates are available for parameterized path generation, and also so
298 : : * that each rel's consider_parallel flag is set correctly before we begin to
299 : : * generate paths.
300 : : */
301 : : static void
5022 302 : 162949 : set_base_rel_sizes(PlannerInfo *root)
303 : : {
304 : : Index rti;
305 : :
306 [ + + ]: 504351 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
307 : : {
308 : 341419 : RelOptInfo *rel = root->simple_rel_array[rti];
309 : : RangeTblEntry *rte;
310 : :
311 : : /* there may be empty slots corresponding to non-baserel RTEs */
312 [ + + ]: 341419 : if (rel == NULL)
313 : 80264 : continue;
314 : :
3050 315 [ - + ]: 261155 : Assert(rel->relid == rti); /* sanity check on array */
316 : :
317 : : /* ignore RTEs that are "other rels" */
5022 318 [ + + ]: 261155 : if (rel->reloptkind != RELOPT_BASEREL)
319 : 29081 : continue;
320 : :
3638 rhaas@postgresql.org 321 : 232074 : rte = root->simple_rte_array[rti];
322 : :
323 : : /*
324 : : * If parallelism is allowable for this query in general, see whether
325 : : * it's allowable for this rel in particular. We have to do this
326 : : * before set_rel_size(), because (a) if this rel is an inheritance
327 : : * parent, set_append_rel_size() will use and perhaps change the rel's
328 : : * consider_parallel flag, and (b) for some RTE types, set_rel_size()
329 : : * goes ahead and makes paths immediately.
330 : : */
331 [ + + ]: 232074 : if (root->glob->parallelModeOK)
332 : 185141 : set_rel_consider_parallel(root, rel, rte);
333 : :
334 : 232074 : set_rel_size(root, rel, rti, rte);
335 : : }
10702 scrappy@hub.org 336 : 162932 : }
337 : :
338 : : /*
339 : : * setup_simple_grouped_rels
340 : : * For each simple relation, build a grouped simple relation if eager
341 : : * aggregation is possible and if this relation can produce grouped paths.
342 : : */
343 : : static void
19 rguo@postgresql.org 344 :GNC 162932 : setup_simple_grouped_rels(PlannerInfo *root)
345 : : {
346 : : Index rti;
347 : :
348 : : /*
349 : : * If there are no aggregate expressions or grouping expressions, eager
350 : : * aggregation is not possible.
351 : : */
352 [ + + ]: 162932 : if (root->agg_clause_list == NIL ||
353 [ + + ]: 330 : root->group_expr_list == NIL)
354 : 162635 : return;
355 : :
356 [ + + ]: 2545 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
357 : : {
358 : 2248 : RelOptInfo *rel = root->simple_rel_array[rti];
359 : :
360 : : /* there may be empty slots corresponding to non-baserel RTEs */
361 [ + + ]: 2248 : if (rel == NULL)
362 : 712 : continue;
363 : :
364 [ - + ]: 1536 : Assert(rel->relid == rti); /* sanity check on array */
365 [ + + - + ]: 1536 : Assert(IS_SIMPLE_REL(rel)); /* sanity check on rel */
366 : :
367 : 1536 : (void) build_simple_grouped_rel(root, rel);
368 : : }
369 : : }
370 : :
371 : : /*
372 : : * set_base_rel_pathlists
373 : : * Finds all paths available for scanning each base-relation entry.
374 : : * Sequential scan and any available indices are considered.
375 : : * Each useful path is attached to its relation's 'pathlist' field.
376 : : */
377 : : static void
7449 tgl@sss.pgh.pa.us 378 :CBC 162932 : set_base_rel_pathlists(PlannerInfo *root)
379 : : {
380 : : Index rti;
381 : :
7209 382 [ + + ]: 504333 : for (rti = 1; rti < root->simple_rel_array_size; rti++)
383 : : {
384 : 341401 : RelOptInfo *rel = root->simple_rel_array[rti];
385 : :
386 : : /* there may be empty slots corresponding to non-baserel RTEs */
7448 387 [ + + ]: 341401 : if (rel == NULL)
388 : 80263 : continue;
389 : :
3050 390 [ - + ]: 261138 : Assert(rel->relid == rti); /* sanity check on array */
391 : :
392 : : /* ignore RTEs that are "other rels" */
7448 393 [ + + ]: 261138 : if (rel->reloptkind != RELOPT_BASEREL)
394 : 29081 : continue;
395 : :
6764 396 : 232057 : set_rel_pathlist(root, rel, rti, root->simple_rte_array[rti]);
397 : : }
7206 398 : 162932 : }
399 : :
400 : : /*
401 : : * set_rel_size
402 : : * Set size estimates for a base relation
403 : : */
404 : : static void
5022 405 : 261000 : set_rel_size(PlannerInfo *root, RelOptInfo *rel,
406 : : Index rti, RangeTblEntry *rte)
407 : : {
5147 408 [ + + + + ]: 493074 : if (rel->reloptkind == RELOPT_BASEREL &&
409 : 232074 : relation_excluded_by_constraints(root, rel, rte))
410 : : {
411 : : /*
412 : : * We proved we don't need to scan the rel via constraint exclusion,
413 : : * so set up a single dummy path for it. Here we only check this for
414 : : * regular baserels; if it's an otherrel, CE was already checked in
415 : : * set_append_rel_size().
416 : : *
417 : : * In this case, we go ahead and set up the relation's path right away
418 : : * instead of leaving it for set_rel_pathlist to do. This is because
419 : : * we don't have a convention for marking a rel as dummy except by
420 : : * assigning a dummy path to it.
421 : : */
422 : 330 : set_dummy_rel_pathlist(rel);
423 : : }
424 [ + + ]: 260670 : else if (rte->inh)
425 : : {
426 : : /* It's an "append relation", process accordingly */
5022 427 : 12851 : set_append_rel_size(root, rel, rti, rte);
428 : : }
429 : : else
430 : : {
5361 431 [ + + + + : 247819 : switch (rel->rtekind)
+ + + +
- ]
432 : : {
433 : 205093 : case RTE_RELATION:
434 [ + + ]: 205093 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
435 : : {
436 : : /* Foreign table */
5022 437 : 1229 : set_foreign_size(root, rel, rte);
438 : : }
3142 rhaas@postgresql.org 439 [ + + ]: 203864 : else if (rte->relkind == RELKIND_PARTITIONED_TABLE)
440 : : {
441 : : /*
442 : : * We could get here if asked to scan a partitioned table
443 : : * with ONLY. In that case we shouldn't scan any of the
444 : : * partitions, so mark it as a dummy rel.
445 : : */
446 : 20 : set_dummy_rel_pathlist(rel);
447 : : }
3818 simon@2ndQuadrant.co 448 [ + + ]: 203844 : else if (rte->tablesample != NULL)
449 : : {
450 : : /* Sampled relation */
451 : 153 : set_tablesample_rel_size(root, rel, rte);
452 : : }
453 : : else
454 : : {
455 : : /* Plain relation */
5022 tgl@sss.pgh.pa.us 456 : 203691 : set_plain_rel_size(root, rel, rte);
457 : : }
5361 458 : 205076 : break;
459 : 8757 : case RTE_SUBQUERY:
460 : :
461 : : /*
462 : : * Subqueries don't support making a choice between
463 : : * parameterized and unparameterized paths, so just go ahead
464 : : * and build their paths immediately.
465 : : */
466 : 8757 : set_subquery_pathlist(root, rel, rti, rte);
467 : 8757 : break;
468 : 24598 : case RTE_FUNCTION:
5022 469 : 24598 : set_function_size_estimates(root, rel);
5361 470 : 24598 : break;
3155 alvherre@alvh.no-ip. 471 : 311 : case RTE_TABLEFUNC:
472 : 311 : set_tablefunc_size_estimates(root, rel);
473 : 311 : break;
5361 tgl@sss.pgh.pa.us 474 : 4125 : case RTE_VALUES:
5022 475 : 4125 : set_values_size_estimates(root, rel);
5361 476 : 4125 : break;
477 : 2594 : case RTE_CTE:
478 : :
479 : : /*
480 : : * CTEs don't support making a choice between parameterized
481 : : * and unparameterized paths, so just go ahead and build their
482 : : * paths immediately.
483 : : */
484 [ + + ]: 2594 : if (rte->self_reference)
485 : 470 : set_worktable_pathlist(root, rel, rte);
486 : : else
487 : 2124 : set_cte_pathlist(root, rel, rte);
488 : 2594 : break;
3132 kgrittn@postgresql.o 489 : 243 : case RTE_NAMEDTUPLESTORE:
490 : : /* Might as well just build the path immediately */
491 : 243 : set_namedtuplestore_pathlist(root, rel, rte);
492 : 243 : break;
2464 tgl@sss.pgh.pa.us 493 : 2098 : case RTE_RESULT:
494 : : /* Might as well just build the path immediately */
495 : 2098 : set_result_pathlist(root, rel, rte);
496 : 2098 : break;
5361 tgl@sss.pgh.pa.us 497 :UBC 0 : default:
498 [ # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
499 : : break;
500 : : }
501 : : }
502 : :
503 : : /*
504 : : * We insist that all non-dummy rels have a nonzero rowcount estimate.
505 : : */
3746 tgl@sss.pgh.pa.us 506 [ + + - + ]:CBC 260982 : Assert(rel->rows > 0 || IS_DUMMY_REL(rel));
5022 507 : 260982 : }
508 : :
509 : : /*
510 : : * set_rel_pathlist
511 : : * Build access paths for a base relation
512 : : */
513 : : static void
514 : 261051 : set_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
515 : : Index rti, RangeTblEntry *rte)
516 : : {
517 [ + + ]: 261051 : if (IS_DUMMY_REL(rel))
518 : : {
519 : : /* We already proved the relation empty, so nothing more to do */
520 : : }
521 [ + + ]: 260422 : else if (rte->inh)
522 : : {
523 : : /* It's an "append relation", process accordingly */
524 : 12703 : set_append_rel_pathlist(root, rel, rti, rte);
525 : : }
526 : : else
527 : : {
528 [ + + + + : 247719 : switch (rel->rtekind)
+ + + +
- ]
529 : : {
530 : 205056 : case RTE_RELATION:
531 [ + + ]: 205056 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
532 : : {
533 : : /* Foreign table */
534 : 1227 : set_foreign_pathlist(root, rel, rte);
535 : : }
3818 simon@2ndQuadrant.co 536 [ + + ]: 203829 : else if (rte->tablesample != NULL)
537 : : {
538 : : /* Sampled relation */
539 : 153 : set_tablesample_rel_pathlist(root, rel, rte);
540 : : }
541 : : else
542 : : {
543 : : /* Plain relation */
5022 tgl@sss.pgh.pa.us 544 : 203676 : set_plain_rel_pathlist(root, rel, rte);
545 : : }
546 : 205056 : break;
547 : 8694 : case RTE_SUBQUERY:
548 : : /* Subquery --- fully handled during set_rel_size */
549 : 8694 : break;
550 : 24598 : case RTE_FUNCTION:
551 : : /* RangeFunction */
552 : 24598 : set_function_pathlist(root, rel, rte);
553 : 24598 : break;
3155 alvherre@alvh.no-ip. 554 : 311 : case RTE_TABLEFUNC:
555 : : /* Table Function */
556 : 311 : set_tablefunc_pathlist(root, rel, rte);
557 : 311 : break;
5022 tgl@sss.pgh.pa.us 558 : 4125 : case RTE_VALUES:
559 : : /* Values list */
560 : 4125 : set_values_pathlist(root, rel, rte);
561 : 4125 : break;
562 : 2594 : case RTE_CTE:
563 : : /* CTE reference --- fully handled during set_rel_size */
564 : 2594 : break;
3132 kgrittn@postgresql.o 565 : 243 : case RTE_NAMEDTUPLESTORE:
566 : : /* tuplestore reference --- fully handled during set_rel_size */
567 : 243 : break;
2464 tgl@sss.pgh.pa.us 568 : 2098 : case RTE_RESULT:
569 : : /* simple Result --- fully handled during set_rel_size */
570 : 2098 : break;
5022 tgl@sss.pgh.pa.us 571 :UBC 0 : default:
572 [ # # ]: 0 : elog(ERROR, "unexpected rtekind: %d", (int) rel->rtekind);
573 : : break;
574 : : }
575 : : }
576 : :
577 : : /*
578 : : * Allow a plugin to editorialize on the set of Paths for this base
579 : : * relation. It could add new paths (such as CustomPaths) by calling
580 : : * add_path(), or add_partial_path() if parallel aware. It could also
581 : : * delete or modify paths added by the core code.
582 : : */
2452 tgl@sss.pgh.pa.us 583 [ - + ]:CBC 261051 : if (set_rel_pathlist_hook)
2452 tgl@sss.pgh.pa.us 584 :UBC 0 : (*set_rel_pathlist_hook) (root, rel, rti, rte);
585 : :
586 : : /*
587 : : * If this is a baserel, we should normally consider gathering any partial
588 : : * paths we may have created for it. We have to do this after calling the
589 : : * set_rel_pathlist_hook, else it cannot add partial paths to be included
590 : : * here.
591 : : *
592 : : * However, if this is an inheritance child, skip it. Otherwise, we could
593 : : * end up with a very large number of gather nodes, each trying to grab
594 : : * its own pool of workers. Instead, we'll consider gathering partial
595 : : * paths for the parent appendrel.
596 : : *
597 : : * Also, if this is the topmost scan/join rel, we postpone gathering until
598 : : * the final scan/join targetlist is available (see grouping_planner).
599 : : */
2786 rhaas@postgresql.org 600 [ + + ]:CBC 261051 : if (rel->reloptkind == RELOPT_BASEREL &&
1001 tgl@sss.pgh.pa.us 601 [ + + ]: 232057 : !bms_equal(rel->relids, root->all_query_rels))
2029 tomas.vondra@postgre 602 : 119234 : generate_useful_gather_paths(root, rel, false);
603 : :
604 : : /* Now find the cheapest of the paths for this rel */
3993 tgl@sss.pgh.pa.us 605 : 261051 : set_cheapest(rel);
606 : :
607 : : /*
608 : : * If a grouped relation for this rel exists, build partial aggregation
609 : : * paths for it.
610 : : *
611 : : * Note that this can only happen after we've called set_cheapest() for
612 : : * this base rel, because we need its cheapest paths.
613 : : */
19 rguo@postgresql.org 614 :GNC 261051 : set_grouped_rel_pathlist(root, rel);
615 : :
616 : : #ifdef OPTIMIZER_DEBUG
617 : : pprint(rel);
618 : : #endif
9115 tgl@sss.pgh.pa.us 619 :CBC 261051 : }
620 : :
621 : : /*
622 : : * set_plain_rel_size
623 : : * Set size estimates for a plain relation (no subquery, no inheritance)
624 : : */
625 : : static void
5022 626 : 203691 : set_plain_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
627 : : {
628 : : /*
629 : : * Test any partial indexes of rel for applicability. We must do this
630 : : * first since partial unique indexes can affect size estimates.
631 : : */
3497 632 : 203691 : check_index_predicates(root, rel);
633 : :
634 : : /* Mark rel with estimated output rows, width, etc */
6764 635 : 203691 : set_baserel_size_estimates(root, rel);
5022 636 : 203676 : }
637 : :
638 : : /*
639 : : * If this relation could possibly be scanned from within a worker, then set
640 : : * its consider_parallel flag.
641 : : */
642 : : static void
3638 rhaas@postgresql.org 643 : 206780 : set_rel_consider_parallel(PlannerInfo *root, RelOptInfo *rel,
644 : : RangeTblEntry *rte)
645 : : {
646 : : /*
647 : : * The flag has previously been initialized to false, so we can just
648 : : * return if it becomes clear that we can't safely set it.
649 : : */
3403 tgl@sss.pgh.pa.us 650 [ - + ]: 206780 : Assert(!rel->consider_parallel);
651 : :
652 : : /* Don't call this if parallelism is disallowed for the entire query. */
3638 rhaas@postgresql.org 653 [ - + ]: 206780 : Assert(root->glob->parallelModeOK);
654 : :
655 : : /* This should only be called for baserels and appendrel children. */
3129 656 [ + + - + ]: 206780 : Assert(IS_SIMPLE_REL(rel));
657 : :
658 : : /* Assorted checks based on rtekind. */
3638 659 [ + + - + : 206780 : switch (rte->rtekind)
+ + + + +
- - ]
660 : : {
661 : 178679 : case RTE_RELATION:
662 : :
663 : : /*
664 : : * Currently, parallel workers can't access the leader's temporary
665 : : * tables. We could possibly relax this if we wrote all of its
666 : : * local buffers at the start of the query and made no changes
667 : : * thereafter (maybe we could allow hint bit changes), and if we
668 : : * taught the workers to read them. Writing a large number of
669 : : * temporary buffers could be expensive, though, and we don't have
670 : : * the rest of the necessary infrastructure right now anyway. So
671 : : * for now, bail out if we see a temporary table.
672 : : */
673 [ + + ]: 178679 : if (get_rel_persistence(rte->relid) == RELPERSISTENCE_TEMP)
674 : 4263 : return;
675 : :
676 : : /*
677 : : * Table sampling can be pushed down to workers if the sample
678 : : * function and its arguments are safe.
679 : : */
680 [ + + ]: 174416 : if (rte->tablesample != NULL)
681 : : {
3254 tgl@sss.pgh.pa.us 682 : 165 : char proparallel = func_parallel(rte->tablesample->tsmhandler);
683 : :
3638 rhaas@postgresql.org 684 [ + + ]: 165 : if (proparallel != PROPARALLEL_SAFE)
685 : 18 : return;
3356 tgl@sss.pgh.pa.us 686 [ + + ]: 147 : if (!is_parallel_safe(root, (Node *) rte->tablesample->args))
3638 rhaas@postgresql.org 687 : 6 : return;
688 : : }
689 : :
690 : : /*
691 : : * Ask FDWs whether they can support performing a ForeignScan
692 : : * within a worker. Most often, the answer will be no. For
693 : : * example, if the nature of the FDW is such that it opens a TCP
694 : : * connection with a remote server, each parallel worker would end
695 : : * up with a separate connection, and these connections might not
696 : : * be appropriately coordinated between workers and the leader.
697 : : */
3531 698 [ + + ]: 174392 : if (rte->relkind == RELKIND_FOREIGN_TABLE)
699 : : {
700 [ - + ]: 782 : Assert(rel->fdwroutine);
701 [ + + ]: 782 : if (!rel->fdwroutine->IsForeignScanParallelSafe)
702 : 746 : return;
703 [ - + ]: 36 : if (!rel->fdwroutine->IsForeignScanParallelSafe(root, rel, rte))
3531 rhaas@postgresql.org 704 :UBC 0 : return;
705 : : }
706 : :
707 : : /*
708 : : * There are additional considerations for appendrels, which we'll
709 : : * deal with in set_append_rel_size and set_append_rel_pathlist.
710 : : * For now, just set consider_parallel based on the rel's own
711 : : * quals and targetlist.
712 : : */
3638 rhaas@postgresql.org 713 :CBC 173646 : break;
714 : :
715 : 9620 : case RTE_SUBQUERY:
716 : :
717 : : /*
718 : : * There's no intrinsic problem with scanning a subquery-in-FROM
719 : : * (as distinct from a SubPlan or InitPlan) in a parallel worker.
720 : : * If the subquery doesn't happen to have any parallel-safe paths,
721 : : * then flagging it as consider_parallel won't change anything,
722 : : * but that's true for plain tables, too. We must set
723 : : * consider_parallel based on the rel's own quals and targetlist,
724 : : * so that if a subquery path is parallel-safe but the quals and
725 : : * projection we're sticking onto it are not, we correctly mark
726 : : * the SubqueryScanPath as not parallel-safe. (Note that
727 : : * set_subquery_pathlist() might push some of these quals down
728 : : * into the subquery itself, but that doesn't change anything.)
729 : : *
730 : : * We can't push sub-select containing LIMIT/OFFSET to workers as
731 : : * there is no guarantee that the row order will be fully
732 : : * deterministic, and applying LIMIT/OFFSET will lead to
733 : : * inconsistent results at the top-level. (In some cases, where
734 : : * the result is ordered, we could relax this restriction. But it
735 : : * doesn't currently seem worth expending extra effort to do so.)
736 : : */
737 : : {
2600 akapila@postgresql.o 738 : 9620 : Query *subquery = castNode(Query, rte->subquery);
739 : :
740 [ + + ]: 9620 : if (limit_needed(subquery))
741 : 254 : return;
742 : : }
3403 tgl@sss.pgh.pa.us 743 : 9366 : break;
744 : :
3638 rhaas@postgresql.org 745 :UBC 0 : case RTE_JOIN:
746 : : /* Shouldn't happen; we're only considering baserels here. */
747 : 0 : Assert(false);
748 : : return;
749 : :
3638 rhaas@postgresql.org 750 :CBC 12529 : case RTE_FUNCTION:
751 : : /* Check for parallel-restricted functions. */
3356 tgl@sss.pgh.pa.us 752 [ + + ]: 12529 : if (!is_parallel_safe(root, (Node *) rte->functions))
3638 rhaas@postgresql.org 753 : 6208 : return;
754 : 6321 : break;
755 : :
3155 alvherre@alvh.no-ip. 756 : 311 : case RTE_TABLEFUNC:
757 : : /* not parallel safe */
758 : 311 : return;
759 : :
3638 rhaas@postgresql.org 760 : 1416 : case RTE_VALUES:
761 : : /* Check for parallel-restricted functions. */
3356 tgl@sss.pgh.pa.us 762 [ + + ]: 1416 : if (!is_parallel_safe(root, (Node *) rte->values_lists))
763 : 6 : return;
3638 rhaas@postgresql.org 764 : 1410 : break;
765 : :
766 : 2113 : case RTE_CTE:
767 : :
768 : : /*
769 : : * CTE tuplestores aren't shared among parallel workers, so we
770 : : * force all CTE scans to happen in the leader. Also, populating
771 : : * the CTE would require executing a subplan that's not available
772 : : * in the worker, might be parallel-restricted, and must get
773 : : * executed only once.
774 : : */
775 : 2113 : return;
776 : :
3132 kgrittn@postgresql.o 777 : 229 : case RTE_NAMEDTUPLESTORE:
778 : :
779 : : /*
780 : : * tuplestore cannot be shared, at least without more
781 : : * infrastructure to support that.
782 : : */
783 : 229 : return;
784 : :
2464 tgl@sss.pgh.pa.us 785 : 1883 : case RTE_RESULT:
786 : : /* RESULT RTEs, in themselves, are no problem. */
787 : 1883 : break;
412 rguo@postgresql.org 788 :UBC 0 : case RTE_GROUP:
789 : : /* Shouldn't happen; we're only considering baserels here. */
790 : 0 : Assert(false);
791 : : return;
792 : : }
793 : :
794 : : /*
795 : : * If there's anything in baserestrictinfo that's parallel-restricted, we
796 : : * give up on parallelizing access to this relation. We could consider
797 : : * instead postponing application of the restricted quals until we're
798 : : * above all the parallelism in the plan tree, but it's not clear that
799 : : * that would be a win in very many cases, and it might be tricky to make
800 : : * outer join clauses work correctly. It would likely break equivalence
801 : : * classes, too.
802 : : */
3356 tgl@sss.pgh.pa.us 803 [ + + ]:CBC 192626 : if (!is_parallel_safe(root, (Node *) rel->baserestrictinfo))
3638 rhaas@postgresql.org 804 : 13596 : return;
805 : :
806 : : /*
807 : : * Likewise, if the relation's outputs are not parallel-safe, give up.
808 : : * (Usually, they're just Vars, but sometimes they're not.)
809 : : */
3356 tgl@sss.pgh.pa.us 810 [ + + ]: 179030 : if (!is_parallel_safe(root, (Node *) rel->reltarget->exprs))
3427 rhaas@postgresql.org 811 : 9 : return;
812 : :
813 : : /* We have a winner. */
3638 814 : 179021 : rel->consider_parallel = true;
815 : : }
816 : :
817 : : /*
818 : : * set_plain_rel_pathlist
819 : : * Build access paths for a plain relation (no subquery, no inheritance)
820 : : */
821 : : static void
5022 tgl@sss.pgh.pa.us 822 : 203676 : set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
823 : : {
824 : : Relids required_outer;
825 : :
826 : : /*
827 : : * We don't support pushing join clauses into the quals of a seqscan, but
828 : : * it could still have required parameterization due to LATERAL refs in
829 : : * its tlist.
830 : : */
4810 831 : 203676 : required_outer = rel->lateral_relids;
832 : :
833 : : /*
834 : : * Consider TID scans.
835 : : *
836 : : * If create_tidscan_paths returns true, then a TID scan path is forced.
837 : : * This happens when rel->baserestrictinfo contains CurrentOfExpr, because
838 : : * the executor can't handle any other type of path for such queries.
839 : : * Hence, we return without adding any other paths.
840 : : */
462 rhaas@postgresql.org 841 [ + + ]: 203676 : if (create_tidscan_paths(root, rel))
842 : 202 : return;
843 : :
844 : : /* Consider sequential scan */
3638 845 : 203474 : add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
846 : :
847 : : /* If appropriate, consider parallel sequential scan */
3568 848 [ + + + + ]: 203474 : if (rel->consider_parallel && required_outer == NULL)
3467 tgl@sss.pgh.pa.us 849 : 153198 : create_plain_partial_paths(root, rel);
850 : :
851 : : /* Consider index scans */
7490 852 : 203474 : create_index_paths(root, rel);
853 : : }
854 : :
855 : : /*
856 : : * create_plain_partial_paths
857 : : * Build partial access paths for parallel scan of a plain relation
858 : : */
859 : : static void
3467 860 : 153198 : create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
861 : : {
862 : : int parallel_workers;
863 : :
2824 rhaas@postgresql.org 864 : 153198 : parallel_workers = compute_parallel_worker(rel, rel->pages, -1,
865 : : max_parallel_workers_per_gather);
866 : :
867 : : /* If any limit was set to zero, the user doesn't want a parallel scan. */
3427 tgl@sss.pgh.pa.us 868 [ + + ]: 153198 : if (parallel_workers <= 0)
869 : 139487 : return;
870 : :
871 : : /* Add an unordered partial path based on a parallel sequential scan. */
rhaas@postgresql.org 872 : 13711 : add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
873 : : }
874 : :
875 : : /*
876 : : * set_tablesample_rel_size
877 : : * Set size estimates for a sampled relation
878 : : */
879 : : static void
3818 simon@2ndQuadrant.co 880 : 153 : set_tablesample_rel_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
881 : : {
3747 tgl@sss.pgh.pa.us 882 : 153 : TableSampleClause *tsc = rte->tablesample;
883 : : TsmRoutine *tsm;
884 : : BlockNumber pages;
885 : : double tuples;
886 : :
887 : : /*
888 : : * Test any partial indexes of rel for applicability. We must do this
889 : : * first since partial unique indexes can affect size estimates.
890 : : */
3497 891 : 153 : check_index_predicates(root, rel);
892 : :
893 : : /*
894 : : * Call the sampling method's estimation function to estimate the number
895 : : * of pages it will read and the number of tuples it will return. (Note:
896 : : * we assume the function returns sane values.)
897 : : */
3747 898 : 153 : tsm = GetTsmRoutine(tsc->tsmhandler);
899 : 153 : tsm->SampleScanGetSampleSize(root, rel, tsc->args,
900 : : &pages, &tuples);
901 : :
902 : : /*
903 : : * For the moment, because we will only consider a SampleScan path for the
904 : : * rel, it's okay to just overwrite the pages and tuples estimates for the
905 : : * whole relation. If we ever consider multiple path types for sampled
906 : : * rels, we'll need more complication.
907 : : */
908 : 153 : rel->pages = pages;
909 : 153 : rel->tuples = tuples;
910 : :
911 : : /* Mark rel with estimated output rows, width, etc */
3818 simon@2ndQuadrant.co 912 : 153 : set_baserel_size_estimates(root, rel);
913 : 153 : }
914 : :
915 : : /*
916 : : * set_tablesample_rel_pathlist
917 : : * Build access paths for a sampled relation
918 : : */
919 : : static void
920 : 153 : set_tablesample_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
921 : : {
922 : : Relids required_outer;
923 : : Path *path;
924 : :
925 : : /*
926 : : * We don't support pushing join clauses into the quals of a samplescan,
927 : : * but it could still have required parameterization due to LATERAL refs
928 : : * in its tlist or TABLESAMPLE arguments.
929 : : */
930 : 153 : required_outer = rel->lateral_relids;
931 : :
932 : : /* Consider sampled scan */
933 : 153 : path = create_samplescan_path(root, rel, required_outer);
934 : :
935 : : /*
936 : : * If the sampling method does not support repeatable scans, we must avoid
937 : : * plans that would scan the rel multiple times. Ideally, we'd simply
938 : : * avoid putting the rel on the inside of a nestloop join; but adding such
939 : : * a consideration to the planner seems like a great deal of complication
940 : : * to support an uncommon usage of second-rate sampling methods. Instead,
941 : : * if there is a risk that the query might perform an unsafe join, just
942 : : * wrap the SampleScan in a Materialize node. We can check for joins by
943 : : * counting the membership of all_query_rels (note that this correctly
944 : : * counts inheritance trees as single rels). If we're inside a subquery,
945 : : * we can't easily check whether a join might occur in the outer query, so
946 : : * just assume one is possible.
947 : : *
948 : : * GetTsmRoutine is relatively expensive compared to the other tests here,
949 : : * so check repeatable_across_scans last, even though that's a bit odd.
950 : : */
3747 tgl@sss.pgh.pa.us 951 [ + + + + ]: 293 : if ((root->query_level > 1 ||
1001 952 : 140 : bms_membership(root->all_query_rels) != BMS_SINGLETON) &&
3050 953 [ + + ]: 49 : !(GetTsmRoutine(rte->tablesample->tsmhandler)->repeatable_across_scans))
954 : : {
3747 955 : 4 : path = (Path *) create_material_path(rel, path);
956 : : }
957 : :
958 : 153 : add_path(rel, path);
959 : :
960 : : /* For the moment, at least, there are no other paths to consider */
3818 simon@2ndQuadrant.co 961 : 153 : }
962 : :
963 : : /*
964 : : * set_foreign_size
965 : : * Set size estimates for a foreign table RTE
966 : : */
967 : : static void
5022 tgl@sss.pgh.pa.us 968 : 1229 : set_foreign_size(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
969 : : {
970 : : /* Mark rel with estimated output rows, width, etc */
971 : 1229 : set_foreign_size_estimates(root, rel);
972 : :
973 : : /* Let FDW adjust the size estimates, if it can */
4980 974 : 1229 : rel->fdwroutine->GetForeignRelSize(root, rel, rte->relid);
975 : :
976 : : /* ... but do not let it set the rows estimate to zero */
3746 977 : 1227 : rel->rows = clamp_row_est(rel->rows);
978 : :
979 : : /*
980 : : * Also, make sure rel->tuples is not insane relative to rel->rows.
981 : : * Notably, this ensures sanity if pg_class.reltuples contains -1 and the
982 : : * FDW doesn't do anything to replace that.
983 : : */
1942 984 [ + + ]: 1227 : rel->tuples = Max(rel->tuples, rel->rows);
5022 985 : 1227 : }
986 : :
987 : : /*
988 : : * set_foreign_pathlist
989 : : * Build access paths for a foreign table RTE
990 : : */
991 : : static void
992 : 1227 : set_foreign_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
993 : : {
994 : : /* Call the FDW's GetForeignPaths function to generate path(s) */
4980 995 : 1227 : rel->fdwroutine->GetForeignPaths(root, rel, rte->relid);
5022 996 : 1227 : }
997 : :
998 : : /*
999 : : * set_append_rel_size
1000 : : * Set size estimates for a simple "append relation"
1001 : : *
1002 : : * The passed-in rel and RTE represent the entire append relation. The
1003 : : * relation's contents are computed by appending together the output of the
1004 : : * individual member relations. Note that in the non-partitioned inheritance
1005 : : * case, the first member relation is actually the same table as is mentioned
1006 : : * in the parent RTE ... but it has a different RTE and RelOptInfo. This is
1007 : : * a good thing because their outputs are not the same size.
1008 : : */
1009 : : static void
1010 : 12851 : set_append_rel_size(PlannerInfo *root, RelOptInfo *rel,
1011 : : Index rti, RangeTblEntry *rte)
1012 : : {
8926 1013 : 12851 : int parentRTindex = rti;
1014 : : bool has_live_children;
1015 : : double parent_tuples;
1016 : : double parent_rows;
1017 : : double parent_size;
1018 : : double *parent_attrsizes;
1019 : : int nattrs;
1020 : : ListCell *l;
1021 : :
1022 : : /* Guard against stack overflow due to overly deep inheritance tree. */
2965 rhaas@postgresql.org 1023 : 12851 : check_stack_depth();
1024 : :
3129 1025 [ + + - + ]: 12851 : Assert(IS_SIMPLE_REL(rel));
1026 : :
1027 : : /*
1028 : : * If this is a partitioned baserel, set the consider_partitionwise_join
1029 : : * flag; currently, we only consider partitionwise joins with the baserel
1030 : : * if its targetlist doesn't contain a whole-row Var.
1031 : : */
2614 efujita@postgresql.o 1032 [ + + ]: 12851 : if (enable_partitionwise_join &&
1033 [ + + ]: 2491 : rel->reloptkind == RELOPT_BASEREL &&
1034 [ + - ]: 1981 : rte->relkind == RELKIND_PARTITIONED_TABLE &&
1001 tgl@sss.pgh.pa.us 1035 [ + + ]: 1981 : bms_is_empty(rel->attr_needed[InvalidAttrNumber - rel->min_attr]))
2614 efujita@postgresql.o 1036 : 1943 : rel->consider_partitionwise_join = true;
1037 : :
1038 : : /*
1039 : : * Initialize to compute size estimates for whole append relation.
1040 : : *
1041 : : * We handle tuples estimates by setting "tuples" to the total number of
1042 : : * tuples accumulated from each live child, rather than using "rows".
1043 : : * Although an appendrel itself doesn't directly enforce any quals, its
1044 : : * child relations may. Therefore, setting "tuples" equal to "rows" for
1045 : : * an appendrel isn't always appropriate, and can lead to inaccurate cost
1046 : : * estimates. For example, when estimating the number of distinct values
1047 : : * from an appendrel, we would be unable to adjust the estimate based on
1048 : : * the restriction selectivity (see estimate_num_groups).
1049 : : *
1050 : : * We handle width estimates by weighting the widths of different child
1051 : : * rels proportionally to their number of rows. This is sensible because
1052 : : * the use of width estimates is mainly to compute the total relation
1053 : : * "footprint" if we have to sort or hash it. To do this, we sum the
1054 : : * total equivalent size (in "double" arithmetic) and then divide by the
1055 : : * total rowcount estimate. This is done separately for the total rel
1056 : : * width and each attribute.
1057 : : *
1058 : : * Note: if you consider changing this logic, beware that child rels could
1059 : : * have zero rows and/or width, if they were excluded by constraints.
1060 : : */
3746 tgl@sss.pgh.pa.us 1061 : 12851 : has_live_children = false;
252 rguo@postgresql.org 1062 : 12851 : parent_tuples = 0;
6331 tgl@sss.pgh.pa.us 1063 : 12851 : parent_rows = 0;
1064 : 12851 : parent_size = 0;
1065 : 12851 : nattrs = rel->max_attr - rel->min_attr + 1;
1066 : 12851 : parent_attrsizes = (double *) palloc0(nattrs * sizeof(double));
1067 : :
7209 1068 [ + + + + : 67638 : foreach(l, root->append_rel_list)
+ + ]
1069 : : {
1070 : 54788 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1071 : : int childRTindex;
1072 : : RangeTblEntry *childRTE;
1073 : : RelOptInfo *childrel;
1074 : : List *childrinfos;
1075 : : ListCell *parentvars;
1076 : : ListCell *childvars;
1077 : : ListCell *lc;
1078 : :
1079 : : /* append_rel_list contains all append rels; ignore others */
1080 [ + + ]: 54788 : if (appinfo->parent_relid != parentRTindex)
1081 : 25931 : continue;
1082 : :
1083 : 29028 : childRTindex = appinfo->child_relid;
6764 1084 : 29028 : childRTE = root->simple_rte_array[childRTindex];
1085 : :
1086 : : /*
1087 : : * The child rel's RelOptInfo was already created during
1088 : : * add_other_rels_to_query.
1089 : : */
6978 1090 : 29028 : childrel = find_base_rel(root, childRTindex);
1091 [ - + ]: 29028 : Assert(childrel->reloptkind == RELOPT_OTHER_MEMBER_REL);
1092 : :
1093 : : /* We may have already proven the child to be dummy. */
2403 1094 [ + + ]: 29028 : if (IS_DUMMY_REL(childrel))
2460 alvherre@alvh.no-ip. 1095 : 9 : continue;
1096 : :
1097 : : /*
1098 : : * We have to copy the parent's targetlist and quals to the child,
1099 : : * with appropriate substitution of variables. However, the
1100 : : * baserestrictinfo quals were already copied/substituted when the
1101 : : * child RelOptInfo was built. So we don't need any additional setup
1102 : : * before applying constraint exclusion.
1103 : : */
6418 tgl@sss.pgh.pa.us 1104 [ + + ]: 29019 : if (relation_excluded_by_constraints(root, childrel, childRTE))
1105 : : {
1106 : : /*
1107 : : * This child need not be scanned, so we can omit it from the
1108 : : * appendrel.
1109 : : */
6729 1110 : 93 : set_dummy_rel_pathlist(childrel);
6764 1111 : 93 : continue;
1112 : : }
1113 : :
1114 : : /*
1115 : : * Constraint exclusion failed, so copy the parent's join quals and
1116 : : * targetlist to the child, with appropriate variable substitutions.
1117 : : *
1118 : : * We skip join quals that came from above outer joins that can null
1119 : : * this rel, since they would be of no value while generating paths
1120 : : * for the child. This saves some effort while processing the child
1121 : : * rel, and it also avoids an implementation restriction in
1122 : : * adjust_appendrel_attrs (it can't apply nullingrels to a non-Var).
1123 : : */
960 1124 : 28926 : childrinfos = NIL;
1125 [ + + + + : 35457 : foreach(lc, rel->joininfo)
+ + ]
1126 : : {
1127 : 6531 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1128 : :
1129 [ + + ]: 6531 : if (!bms_overlap(rinfo->clause_relids, rel->nulling_relids))
1130 : 5370 : childrinfos = lappend(childrinfos,
1131 : 5370 : adjust_appendrel_attrs(root,
1132 : : (Node *) rinfo,
1133 : : 1, &appinfo));
1134 : : }
1135 : 28926 : childrel->joininfo = childrinfos;
1136 : :
1137 : : /*
1138 : : * Now for the child's targetlist.
1139 : : *
1140 : : * NB: the resulting childrel->reltarget->exprs may contain arbitrary
1141 : : * expressions, which otherwise would not occur in a rel's targetlist.
1142 : : * Code that might be looking at an appendrel child must cope with
1143 : : * such. (Normally, a rel's targetlist would only include Vars and
1144 : : * PlaceHolderVars.) XXX we do not bother to update the cost or width
1145 : : * fields of childrel->reltarget; not clear if that would be useful.
1146 : : */
2471 efujita@postgresql.o 1147 : 57852 : childrel->reltarget->exprs = (List *)
1148 : 28926 : adjust_appendrel_attrs(root,
1149 : 28926 : (Node *) rel->reltarget->exprs,
1150 : : 1, &appinfo);
1151 : :
1152 : : /*
1153 : : * We have to make child entries in the EquivalenceClass data
1154 : : * structures as well. This is needed either if the parent
1155 : : * participates in some eclass joins (because we will want to consider
1156 : : * inner-indexscan joins on the individual children) or if the parent
1157 : : * has useful pathkeys (because we should try to build MergeAppend
1158 : : * paths that produce those sort orderings).
1159 : : */
1160 [ + + + + ]: 28926 : if (rel->has_eclass_joins || has_useful_pathkeys(root, rel))
1161 : 17703 : add_child_rel_equivalences(root, appinfo, rel, childrel);
1162 : 28926 : childrel->has_eclass_joins = rel->has_eclass_joins;
1163 : :
1164 : : /*
1165 : : * Note: we could compute appropriate attr_needed data for the child's
1166 : : * variables, by transforming the parent's attr_needed through the
1167 : : * translated_vars mapping. However, currently there's no need
1168 : : * because attr_needed is only examined for base relations not
1169 : : * otherrels. So we just leave the child's attr_needed empty.
1170 : : */
1171 : :
1172 : : /*
1173 : : * If we consider partitionwise joins with the parent rel, do the same
1174 : : * for partitioned child rels.
1175 : : *
1176 : : * Note: here we abuse the consider_partitionwise_join flag by setting
1177 : : * it for child rels that are not themselves partitioned. We do so to
1178 : : * tell try_partitionwise_join() that the child rel is sufficiently
1179 : : * valid to be used as a per-partition input, even if it later gets
1180 : : * proven to be dummy. (It's not usable until we've set up the
1181 : : * reltarget and EC entries, which we just did.)
1182 : : */
1183 [ + + ]: 28926 : if (rel->consider_partitionwise_join)
2614 1184 : 6574 : childrel->consider_partitionwise_join = true;
1185 : :
1186 : : /*
1187 : : * If parallelism is allowable for this query in general, see whether
1188 : : * it's allowable for this childrel in particular. But if we've
1189 : : * already decided the appendrel is not parallel-safe as a whole,
1190 : : * there's no point in considering parallelism for this child. For
1191 : : * consistency, do this before calling set_rel_size() for the child.
1192 : : */
3403 tgl@sss.pgh.pa.us 1193 [ + + + + ]: 28926 : if (root->glob->parallelModeOK && rel->consider_parallel)
1194 : 21639 : set_rel_consider_parallel(root, childrel, childRTE);
1195 : :
1196 : : /*
1197 : : * Compute the child's size.
1198 : : */
5022 1199 : 28926 : set_rel_size(root, childrel, childRTindex, childRTE);
1200 : :
1201 : : /*
1202 : : * It is possible that constraint exclusion detected a contradiction
1203 : : * within a child subquery, even though we didn't prove one above. If
1204 : : * so, we can skip this child.
1205 : : */
1206 [ + + ]: 28925 : if (IS_DUMMY_REL(childrel))
5147 1207 : 69 : continue;
1208 : :
1209 : : /* We have at least one live child. */
3746 1210 : 28856 : has_live_children = true;
1211 : :
1212 : : /*
1213 : : * If any live child is not parallel-safe, treat the whole appendrel
1214 : : * as not parallel-safe. In future we might be able to generate plans
1215 : : * in which some children are farmed out to workers while others are
1216 : : * not; but we don't have that today, so it's a waste to consider
1217 : : * partial paths anywhere in the appendrel unless it's all safe.
1218 : : * (Child rels visited before this one will be unmarked in
1219 : : * set_append_rel_pathlist().)
1220 : : */
3403 1221 [ + + ]: 28856 : if (!childrel->consider_parallel)
1222 : 7633 : rel->consider_parallel = false;
1223 : :
1224 : : /*
1225 : : * Accumulate size information from each live child.
1226 : : */
3746 1227 [ - + ]: 28856 : Assert(childrel->rows > 0);
1228 : :
252 rguo@postgresql.org 1229 : 28856 : parent_tuples += childrel->tuples;
3746 tgl@sss.pgh.pa.us 1230 : 28856 : parent_rows += childrel->rows;
3514 1231 : 28856 : parent_size += childrel->reltarget->width * childrel->rows;
1232 : :
1233 : : /*
1234 : : * Accumulate per-column estimates too. We need not do anything for
1235 : : * PlaceHolderVars in the parent list. If child expression isn't a
1236 : : * Var, or we didn't record a width estimate for it, we have to fall
1237 : : * back on a datatype-based estimate.
1238 : : *
1239 : : * By construction, child's targetlist is 1-to-1 with parent's.
1240 : : */
1241 [ + + + + : 94953 : forboth(parentvars, rel->reltarget->exprs,
+ + + + +
+ + - +
+ ]
1242 : : childvars, childrel->reltarget->exprs)
1243 : : {
3746 1244 : 66097 : Var *parentvar = (Var *) lfirst(parentvars);
1245 : 66097 : Node *childvar = (Node *) lfirst(childvars);
1246 : :
1671 1247 [ + + + + ]: 66097 : if (IsA(parentvar, Var) && parentvar->varno == parentRTindex)
1248 : : {
3746 1249 : 59641 : int pndx = parentvar->varattno - rel->min_attr;
1250 : 59641 : int32 child_width = 0;
1251 : :
1252 [ + + ]: 59641 : if (IsA(childvar, Var) &&
1253 [ + + ]: 57251 : ((Var *) childvar)->varno == childrel->relid)
1254 : : {
1255 : 57218 : int cndx = ((Var *) childvar)->varattno - childrel->min_attr;
1256 : :
1257 : 57218 : child_width = childrel->attr_widths[cndx];
1258 : : }
1259 [ + + ]: 59641 : if (child_width <= 0)
1260 : 2423 : child_width = get_typavgwidth(exprType(childvar),
1261 : : exprTypmod(childvar));
1262 [ - + ]: 59641 : Assert(child_width > 0);
1263 : 59641 : parent_attrsizes[pndx] += child_width * childrel->rows;
1264 : : }
1265 : : }
1266 : : }
1267 : :
1268 [ + + ]: 12850 : if (has_live_children)
1269 : : {
1270 : : /*
1271 : : * Save the finished size estimates.
1272 : : */
1273 : : int i;
1274 : :
1275 [ - + ]: 12703 : Assert(parent_rows > 0);
252 rguo@postgresql.org 1276 : 12703 : rel->tuples = parent_tuples;
3746 tgl@sss.pgh.pa.us 1277 : 12703 : rel->rows = parent_rows;
3514 1278 : 12703 : rel->reltarget->width = rint(parent_size / parent_rows);
6331 1279 [ + + ]: 118245 : for (i = 0; i < nattrs; i++)
1280 : 105542 : rel->attr_widths[i] = rint(parent_attrsizes[i] / parent_rows);
1281 : :
1282 : : /*
1283 : : * Note that we leave rel->pages as zero; this is important to avoid
1284 : : * double-counting the appendrel tree in total_table_pages.
1285 : : */
1286 : : }
1287 : : else
1288 : : {
1289 : : /*
1290 : : * All children were excluded by constraints, so mark the whole
1291 : : * appendrel dummy. We must do this in this phase so that the rel's
1292 : : * dummy-ness is visible when we generate paths for other rels.
1293 : : */
3746 1294 : 147 : set_dummy_rel_pathlist(rel);
1295 : : }
1296 : :
6331 1297 : 12850 : pfree(parent_attrsizes);
5022 1298 : 12850 : }
1299 : :
1300 : : /*
1301 : : * set_append_rel_pathlist
1302 : : * Build access paths for an "append relation"
1303 : : */
1304 : : static void
1305 : 12703 : set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
1306 : : Index rti, RangeTblEntry *rte)
1307 : : {
1308 : 12703 : int parentRTindex = rti;
1309 : 12703 : List *live_childrels = NIL;
1310 : : ListCell *l;
1311 : :
1312 : : /*
1313 : : * Generate access paths for each member relation, and remember the
1314 : : * non-dummy children.
1315 : : */
1316 [ + - + + : 67283 : foreach(l, root->append_rel_list)
+ + ]
1317 : : {
1318 : 54580 : AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
1319 : : int childRTindex;
1320 : : RangeTblEntry *childRTE;
1321 : : RelOptInfo *childrel;
1322 : :
1323 : : /* append_rel_list contains all append rels; ignore others */
1324 [ + + ]: 54580 : if (appinfo->parent_relid != parentRTindex)
1325 : 25586 : continue;
1326 : :
1327 : : /* Re-locate the child RTE and RelOptInfo */
1328 : 28994 : childRTindex = appinfo->child_relid;
1329 : 28994 : childRTE = root->simple_rte_array[childRTindex];
1330 : 28994 : childrel = root->simple_rel_array[childRTindex];
1331 : :
1332 : : /*
1333 : : * If set_append_rel_size() decided the parent appendrel was
1334 : : * parallel-unsafe at some point after visiting this child rel, we
1335 : : * need to propagate the unsafety marking down to the child, so that
1336 : : * we don't generate useless partial paths for it.
1337 : : */
3403 1338 [ + + ]: 28994 : if (!rel->consider_parallel)
1339 : 7731 : childrel->consider_parallel = false;
1340 : :
1341 : : /*
1342 : : * Compute the child's access paths.
1343 : : */
5022 1344 : 28994 : set_rel_pathlist(root, childrel, childRTindex, childRTE);
1345 : :
1346 : : /*
1347 : : * If child is dummy, ignore it.
1348 : : */
1349 [ + + ]: 28994 : if (IS_DUMMY_REL(childrel))
1350 : 138 : continue;
1351 : :
1352 : : /*
1353 : : * Child is live, so add it to the live_childrels list for use below.
1354 : : */
4825 1355 : 28856 : live_childrels = lappend(live_childrels, childrel);
1356 : : }
1357 : :
1358 : : /* Add paths to the append relation. */
3149 rhaas@postgresql.org 1359 : 12703 : add_paths_to_append_rel(root, rel, live_childrels);
1360 : 12703 : }
1361 : :
1362 : : /*
1363 : : * set_grouped_rel_pathlist
1364 : : * If a grouped relation for the given 'rel' exists, build partial
1365 : : * aggregation paths for it.
1366 : : */
1367 : : static void
19 rguo@postgresql.org 1368 :GNC 261051 : set_grouped_rel_pathlist(PlannerInfo *root, RelOptInfo *rel)
1369 : : {
1370 : : RelOptInfo *grouped_rel;
1371 : :
1372 : : /*
1373 : : * If there are no aggregate expressions or grouping expressions, eager
1374 : : * aggregation is not possible.
1375 : : */
1376 [ + + ]: 261051 : if (root->agg_clause_list == NIL ||
1377 [ + + ]: 1668 : root->group_expr_list == NIL)
1378 : 259515 : return;
1379 : :
1380 : : /* Add paths to the grouped base relation if one exists. */
1381 : 1536 : grouped_rel = rel->grouped_rel;
1382 [ + + ]: 1536 : if (grouped_rel)
1383 : : {
1384 [ - + ]: 293 : Assert(IS_GROUPED_REL(grouped_rel));
1385 : :
1386 : 293 : generate_grouped_paths(root, grouped_rel, rel);
1387 : 293 : set_cheapest(grouped_rel);
1388 : : }
1389 : : }
1390 : :
1391 : :
1392 : : /*
1393 : : * add_paths_to_append_rel
1394 : : * Generate paths for the given append relation given the set of non-dummy
1395 : : * child rels.
1396 : : *
1397 : : * The function collects all parameterizations and orderings supported by the
1398 : : * non-dummy children. For every such parameterization or ordering, it creates
1399 : : * an append path collecting one path from each non-dummy child with given
1400 : : * parameterization or ordering. Similarly it collects partial paths from
1401 : : * non-dummy children to create partial append paths.
1402 : : */
1403 : : void
3149 rhaas@postgresql.org 1404 :CBC 23265 : add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
1405 : : List *live_childrels)
1406 : : {
1407 : 23265 : List *subpaths = NIL;
1408 : 23265 : bool subpaths_valid = true;
753 drowley@postgresql.o 1409 : 23265 : List *startup_subpaths = NIL;
1410 : 23265 : bool startup_subpaths_valid = true;
3149 rhaas@postgresql.org 1411 : 23265 : List *partial_subpaths = NIL;
2883 1412 : 23265 : List *pa_partial_subpaths = NIL;
1413 : 23265 : List *pa_nonpartial_subpaths = NIL;
3149 1414 : 23265 : bool partial_subpaths_valid = true;
1415 : : bool pa_subpaths_valid;
1416 : 23265 : List *all_child_pathkeys = NIL;
1417 : 23265 : List *all_child_outers = NIL;
1418 : : ListCell *l;
2883 1419 : 23265 : double partial_rows = -1;
1420 : :
1421 : : /* If appropriate, consider parallel append */
2686 akapila@postgresql.o 1422 [ + + + + ]: 23265 : pa_subpaths_valid = enable_parallel_append && rel->consider_parallel;
1423 : :
1424 : : /*
1425 : : * For every non-dummy child, remember the cheapest path. Also, identify
1426 : : * all pathkeys (orderings) and parameterizations (required_outer sets)
1427 : : * available for the non-dummy member relations.
1428 : : */
3149 rhaas@postgresql.org 1429 [ + - + + : 74477 : foreach(l, live_childrels)
+ + ]
1430 : : {
1431 : 51212 : RelOptInfo *childrel = lfirst(l);
1432 : : ListCell *lcp;
2883 1433 : 51212 : Path *cheapest_partial_path = NULL;
1434 : :
1435 : : /*
1436 : : * If child has an unparameterized cheapest-total path, add that to
1437 : : * the unparameterized Append path we are constructing for the parent.
1438 : : * If not, there's no workable unparameterized path.
1439 : : *
1440 : : * With partitionwise aggregates, the child rel's pathlist may be
1441 : : * empty, so don't assume that a path exists here.
1442 : : */
2776 1443 [ + - ]: 51212 : if (childrel->pathlist != NIL &&
1444 [ + + ]: 51212 : childrel->cheapest_total_path->param_info == NULL)
2883 1445 : 50846 : accumulate_append_subpath(childrel->cheapest_total_path,
1446 : : &subpaths, NULL);
1447 : : else
4825 tgl@sss.pgh.pa.us 1448 : 366 : subpaths_valid = false;
1449 : :
1450 : : /*
1451 : : * When the planner is considering cheap startup plans, we'll also
1452 : : * collect all the cheapest_startup_paths (if set) and build an
1453 : : * AppendPath containing those as subpaths.
1454 : : */
748 drowley@postgresql.o 1455 [ + + + + ]: 51212 : if (rel->consider_startup && childrel->cheapest_startup_path != NULL)
1456 : 811 : {
1457 : : Path *cheapest_path;
1458 : :
1459 : : /*
1460 : : * With an indication of how many tuples the query should provide,
1461 : : * the optimizer tries to choose the path optimal for that
1462 : : * specific number of tuples.
1463 : : */
231 akorotkov@postgresql 1464 [ + - ]: 811 : if (root->tuple_fraction > 0.0)
1465 : : cheapest_path =
1466 : 811 : get_cheapest_fractional_path(childrel,
1467 : : root->tuple_fraction);
1468 : : else
231 akorotkov@postgresql 1469 :UBC 0 : cheapest_path = childrel->cheapest_startup_path;
1470 : :
1471 : : /* cheapest_startup_path must not be a parameterized path. */
231 akorotkov@postgresql 1472 [ - + ]:CBC 811 : Assert(cheapest_path->param_info == NULL);
1473 : 811 : accumulate_append_subpath(cheapest_path,
1474 : : &startup_subpaths,
1475 : : NULL);
1476 : : }
1477 : : else
753 drowley@postgresql.o 1478 : 50401 : startup_subpaths_valid = false;
1479 : :
1480 : :
1481 : : /* Same idea, but for a partial plan. */
3568 rhaas@postgresql.org 1482 [ + + ]: 51212 : if (childrel->partial_pathlist != NIL)
1483 : : {
2883 1484 : 31697 : cheapest_partial_path = linitial(childrel->partial_pathlist);
1485 : 31697 : accumulate_append_subpath(cheapest_partial_path,
1486 : : &partial_subpaths, NULL);
1487 : : }
1488 : : else
3568 1489 : 19515 : partial_subpaths_valid = false;
1490 : :
1491 : : /*
1492 : : * Same idea, but for a parallel append mixing partial and non-partial
1493 : : * paths.
1494 : : */
2883 1495 [ + + ]: 51212 : if (pa_subpaths_valid)
1496 : : {
1497 : 38642 : Path *nppath = NULL;
1498 : :
1499 : : nppath =
1500 : 38642 : get_cheapest_parallel_safe_total_inner(childrel->pathlist);
1501 : :
1502 [ + + + + ]: 38642 : if (cheapest_partial_path == NULL && nppath == NULL)
1503 : : {
1504 : : /* Neither a partial nor a parallel-safe path? Forget it. */
1505 : 273 : pa_subpaths_valid = false;
1506 : : }
1507 [ + + + + ]: 38369 : else if (nppath == NULL ||
1508 : 31472 : (cheapest_partial_path != NULL &&
1509 [ + + ]: 31472 : cheapest_partial_path->total_cost < nppath->total_cost))
1510 : : {
1511 : : /* Partial path is cheaper or the only option. */
1512 [ - + ]: 31402 : Assert(cheapest_partial_path != NULL);
1513 : 31402 : accumulate_append_subpath(cheapest_partial_path,
1514 : : &pa_partial_subpaths,
1515 : : &pa_nonpartial_subpaths);
1516 : : }
1517 : : else
1518 : : {
1519 : : /*
1520 : : * Either we've got only a non-partial path, or we think that
1521 : : * a single backend can execute the best non-partial path
1522 : : * faster than all the parallel backends working together can
1523 : : * execute the best partial path.
1524 : : *
1525 : : * It might make sense to be more aggressive here. Even if
1526 : : * the best non-partial path is more expensive than the best
1527 : : * partial path, it could still be better to choose the
1528 : : * non-partial path if there are several such paths that can
1529 : : * be given to different workers. For now, we don't try to
1530 : : * figure that out.
1531 : : */
1532 : 6967 : accumulate_append_subpath(nppath,
1533 : : &pa_nonpartial_subpaths,
1534 : : NULL);
1535 : : }
1536 : : }
1537 : :
1538 : : /*
1539 : : * Collect lists of all the available path orderings and
1540 : : * parameterizations for all the children. We use these as a
1541 : : * heuristic to indicate which sort orderings and parameterizations we
1542 : : * should build Append and MergeAppend paths for.
1543 : : */
5022 tgl@sss.pgh.pa.us 1544 [ + - + + : 120958 : foreach(lcp, childrel->pathlist)
+ + ]
1545 : : {
1546 : 69746 : Path *childpath = (Path *) lfirst(lcp);
1547 : 69746 : List *childkeys = childpath->pathkeys;
4939 1548 [ + + ]: 69746 : Relids childouter = PATH_REQ_OUTER(childpath);
1549 : :
1550 : : /* Unsorted paths don't contribute to pathkey list */
5022 1551 [ + + ]: 69746 : if (childkeys != NIL)
1552 : : {
1553 : : ListCell *lpk;
1554 : 18397 : bool found = false;
1555 : :
1556 : : /* Have we already seen this ordering? */
1557 [ + + + + : 18500 : foreach(lpk, all_child_pathkeys)
+ + ]
1558 : : {
1559 : 12428 : List *existing_pathkeys = (List *) lfirst(lpk);
1560 : :
1561 [ + + ]: 12428 : if (compare_pathkeys(existing_pathkeys,
1562 : : childkeys) == PATHKEYS_EQUAL)
1563 : : {
1564 : 12325 : found = true;
1565 : 12325 : break;
1566 : : }
1567 : : }
1568 [ + + ]: 18397 : if (!found)
1569 : : {
1570 : : /* No, so add it to all_child_pathkeys */
1571 : 6072 : all_child_pathkeys = lappend(all_child_pathkeys,
1572 : : childkeys);
1573 : : }
1574 : : }
1575 : :
1576 : : /* Unparameterized paths don't contribute to param-set list */
1577 [ + + ]: 69746 : if (childouter)
1578 : : {
1579 : : ListCell *lco;
1580 : 3298 : bool found = false;
1581 : :
1582 : : /* Have we already seen this param set? */
1583 [ + + + + : 3658 : foreach(lco, all_child_outers)
+ + ]
1584 : : {
4887 bruce@momjian.us 1585 : 2405 : Relids existing_outers = (Relids) lfirst(lco);
1586 : :
5022 tgl@sss.pgh.pa.us 1587 [ + + ]: 2405 : if (bms_equal(existing_outers, childouter))
1588 : : {
1589 : 2045 : found = true;
1590 : 2045 : break;
1591 : : }
1592 : : }
1593 [ + + ]: 3298 : if (!found)
1594 : : {
1595 : : /* No, so add it to all_child_outers */
1596 : 1253 : all_child_outers = lappend(all_child_outers,
1597 : : childouter);
1598 : : }
1599 : : }
1600 : : }
1601 : : }
1602 : :
1603 : : /*
1604 : : * If we found unparameterized paths for all children, build an unordered,
1605 : : * unparameterized Append path for the rel. (Note: this is correct even
1606 : : * if we have zero or one live subpath due to constraint exclusion.)
1607 : : */
4825 1608 [ + + ]: 23265 : if (subpaths_valid)
2760 alvherre@alvh.no-ip. 1609 : 23109 : add_path(rel, (Path *) create_append_path(root, rel, subpaths, NIL,
1610 : : NIL, NULL, 0, false,
1611 : : -1));
1612 : :
1613 : : /* build an AppendPath for the cheap startup paths, if valid */
753 drowley@postgresql.o 1614 [ + + ]: 23265 : if (startup_subpaths_valid)
1615 : 332 : add_path(rel, (Path *) create_append_path(root, rel, startup_subpaths,
1616 : : NIL, NIL, NULL, 0, false, -1));
1617 : :
1618 : : /*
1619 : : * Consider an append of unordered, unparameterized partial paths. Make
1620 : : * it parallel-aware if possible.
1621 : : */
2426 tgl@sss.pgh.pa.us 1622 [ + + + - ]: 23265 : if (partial_subpaths_valid && partial_subpaths != NIL)
1623 : : {
1624 : : AppendPath *appendpath;
1625 : : ListCell *lc;
3427 rhaas@postgresql.org 1626 : 13893 : int parallel_workers = 0;
1627 : :
1628 : : /* Find the highest number of workers requested for any subpath. */
3568 1629 [ + - + + : 47827 : foreach(lc, partial_subpaths)
+ + ]
1630 : : {
1631 : 33934 : Path *path = lfirst(lc);
1632 : :
3427 1633 : 33934 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1634 : : }
1635 [ - + ]: 13893 : Assert(parallel_workers > 0);
1636 : :
1637 : : /*
1638 : : * If the use of parallel append is permitted, always request at least
1639 : : * log2(# of children) workers. We assume it can be useful to have
1640 : : * extra workers in this case because they will be spread out across
1641 : : * the children. The precise formula is just a guess, but we don't
1642 : : * want to end up with a radically different answer for a table with N
1643 : : * partitions vs. an unpartitioned table with the same data, so the
1644 : : * use of some kind of log-scaling here seems to make some sense.
1645 : : */
2883 1646 [ + + ]: 13893 : if (enable_parallel_append)
1647 : : {
1648 [ + + ]: 13869 : parallel_workers = Max(parallel_workers,
1649 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1650 : 13869 : parallel_workers = Min(parallel_workers,
1651 : : max_parallel_workers_per_gather);
1652 : : }
1653 [ - + ]: 13893 : Assert(parallel_workers > 0);
1654 : :
1655 : : /* Generate a partial append path. */
2760 alvherre@alvh.no-ip. 1656 : 13893 : appendpath = create_append_path(root, rel, NIL, partial_subpaths,
1657 : : NIL, NULL, parallel_workers,
1658 : : enable_parallel_append,
1659 : : -1);
1660 : :
1661 : : /*
1662 : : * Make sure any subsequent partial paths use the same row count
1663 : : * estimate.
1664 : : */
2883 rhaas@postgresql.org 1665 : 13893 : partial_rows = appendpath->path.rows;
1666 : :
1667 : : /* Add the path. */
1668 : 13893 : add_partial_path(rel, (Path *) appendpath);
1669 : : }
1670 : :
1671 : : /*
1672 : : * Consider a parallel-aware append using a mix of partial and non-partial
1673 : : * paths. (This only makes sense if there's at least one child which has
1674 : : * a non-partial path that is substantially cheaper than any partial path;
1675 : : * otherwise, we should use the append path added in the previous step.)
1676 : : */
1677 [ + + + + ]: 23265 : if (pa_subpaths_valid && pa_nonpartial_subpaths != NIL)
1678 : : {
1679 : : AppendPath *appendpath;
1680 : : ListCell *lc;
1681 : 2482 : int parallel_workers = 0;
1682 : :
1683 : : /*
1684 : : * Find the highest number of workers requested for any partial
1685 : : * subpath.
1686 : : */
1687 [ + + + + : 2933 : foreach(lc, pa_partial_subpaths)
+ + ]
1688 : : {
1689 : 451 : Path *path = lfirst(lc);
1690 : :
1691 : 451 : parallel_workers = Max(parallel_workers, path->parallel_workers);
1692 : : }
1693 : :
1694 : : /*
1695 : : * Same formula here as above. It's even more important in this
1696 : : * instance because the non-partial paths won't contribute anything to
1697 : : * the planned number of parallel workers.
1698 : : */
1699 [ + - ]: 2482 : parallel_workers = Max(parallel_workers,
1700 : : pg_leftmost_one_pos32(list_length(live_childrels)) + 1);
1701 : 2482 : parallel_workers = Min(parallel_workers,
1702 : : max_parallel_workers_per_gather);
1703 [ - + ]: 2482 : Assert(parallel_workers > 0);
1704 : :
2760 alvherre@alvh.no-ip. 1705 : 2482 : appendpath = create_append_path(root, rel, pa_nonpartial_subpaths,
1706 : : pa_partial_subpaths,
1707 : : NIL, NULL, parallel_workers, true,
1708 : : partial_rows);
3568 rhaas@postgresql.org 1709 : 2482 : add_partial_path(rel, (Path *) appendpath);
1710 : : }
1711 : :
1712 : : /*
1713 : : * Also build unparameterized ordered append paths based on the collected
1714 : : * list of child pathkeys.
1715 : : */
4825 tgl@sss.pgh.pa.us 1716 [ + + ]: 23265 : if (subpaths_valid)
2397 1717 : 23109 : generate_orderedappend_paths(root, rel, live_childrels,
1718 : : all_child_pathkeys);
1719 : :
1720 : : /*
1721 : : * Build Append paths for each parameterization seen among the child rels.
1722 : : * (This may look pretty expensive, but in most cases of practical
1723 : : * interest, the child rels will expose mostly the same parameterizations,
1724 : : * so that not that many cases actually get considered here.)
1725 : : *
1726 : : * The Append node itself cannot enforce quals, so all qual checking must
1727 : : * be done in the child paths. This means that to have a parameterized
1728 : : * Append path, we must have the exact same parameterization for each
1729 : : * child path; otherwise some children might be failing to check the
1730 : : * moved-down quals. To make them match up, we can try to increase the
1731 : : * parameterization of lesser-parameterized paths.
1732 : : */
5022 1733 [ + + + + : 24518 : foreach(l, all_child_outers)
+ + ]
1734 : : {
4887 bruce@momjian.us 1735 : 1253 : Relids required_outer = (Relids) lfirst(l);
1736 : : ListCell *lcr;
1737 : :
1738 : : /* Select the child paths for an Append with this parameterization */
5022 tgl@sss.pgh.pa.us 1739 : 1253 : subpaths = NIL;
4825 1740 : 1253 : subpaths_valid = true;
5022 1741 [ + - + + : 4599 : foreach(lcr, live_childrels)
+ + ]
1742 : : {
1743 : 3352 : RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
1744 : : Path *subpath;
1745 : :
2776 rhaas@postgresql.org 1746 [ - + ]: 3352 : if (childrel->pathlist == NIL)
1747 : : {
1748 : : /* failed to make a suitable path for this child */
2776 rhaas@postgresql.org 1749 :UBC 0 : subpaths_valid = false;
1750 : 0 : break;
1751 : : }
1752 : :
4495 tgl@sss.pgh.pa.us 1753 :CBC 3352 : subpath = get_cheapest_parameterized_child_path(root,
1754 : : childrel,
1755 : : required_outer);
1756 [ + + ]: 3352 : if (subpath == NULL)
1757 : : {
1758 : : /* failed to make a suitable path for this child */
1759 : 6 : subpaths_valid = false;
1760 : 6 : break;
1761 : : }
1729 1762 : 3346 : accumulate_append_subpath(subpath, &subpaths, NULL);
1763 : : }
1764 : :
4825 1765 [ + + ]: 1253 : if (subpaths_valid)
4939 1766 : 1247 : add_path(rel, (Path *)
2760 alvherre@alvh.no-ip. 1767 : 1247 : create_append_path(root, rel, subpaths, NIL,
1768 : : NIL, required_outer, 0, false,
1769 : : -1));
1770 : : }
1771 : :
1772 : : /*
1773 : : * When there is only a single child relation, the Append path can inherit
1774 : : * any ordering available for the child rel's path, so that it's useful to
1775 : : * consider ordered partial paths. Above we only considered the cheapest
1776 : : * partial path for each child, but let's also make paths using any
1777 : : * partial paths that have pathkeys.
1778 : : */
2408 tgl@sss.pgh.pa.us 1779 [ + + ]: 23265 : if (list_length(live_childrels) == 1)
1780 : : {
1781 : 7189 : RelOptInfo *childrel = (RelOptInfo *) linitial(live_childrels);
1782 : :
1783 : : /* skip the cheapest partial path, since we already used that above */
1820 drowley@postgresql.o 1784 [ + + + + : 7291 : for_each_from(l, childrel->partial_pathlist, 1)
+ + ]
1785 : : {
2408 tgl@sss.pgh.pa.us 1786 : 102 : Path *path = (Path *) lfirst(l);
1787 : : AppendPath *appendpath;
1788 : :
1789 : : /* skip paths with no pathkeys. */
1820 drowley@postgresql.o 1790 [ - + ]: 102 : if (path->pathkeys == NIL)
2408 tgl@sss.pgh.pa.us 1791 :UBC 0 : continue;
1792 : :
2408 tgl@sss.pgh.pa.us 1793 :CBC 102 : appendpath = create_append_path(root, rel, NIL, list_make1(path),
1794 : : NIL, NULL,
1795 : : path->parallel_workers, true,
1796 : : partial_rows);
1797 : 102 : add_partial_path(rel, (Path *) appendpath);
1798 : : }
1799 : : }
5022 1800 : 23265 : }
1801 : :
1802 : : /*
1803 : : * generate_orderedappend_paths
1804 : : * Generate ordered append paths for an append relation
1805 : : *
1806 : : * Usually we generate MergeAppend paths here, but there are some special
1807 : : * cases where we can generate simple Append paths, because the subpaths
1808 : : * can provide tuples in the required order already.
1809 : : *
1810 : : * We generate a path for each ordering (pathkey list) appearing in
1811 : : * all_child_pathkeys.
1812 : : *
1813 : : * We consider both cheapest-startup and cheapest-total cases, ie, for each
1814 : : * interesting ordering, collect all the cheapest startup subpaths and all the
1815 : : * cheapest total paths, and build a suitable path for each case.
1816 : : *
1817 : : * We don't currently generate any parameterized ordered paths here. While
1818 : : * it would not take much more code here to do so, it's very unclear that it
1819 : : * is worth the planning cycles to investigate such paths: there's little
1820 : : * use for an ordered path on the inside of a nestloop. In fact, it's likely
1821 : : * that the current coding of add_path would reject such paths out of hand,
1822 : : * because add_path gives no credit for sort ordering of parameterized paths,
1823 : : * and a parameterized MergeAppend is going to be more expensive than the
1824 : : * corresponding parameterized Append path. If we ever try harder to support
1825 : : * parameterized mergejoin plans, it might be worth adding support for
1826 : : * parameterized paths here to feed such joins. (See notes in
1827 : : * optimizer/README for why that might not ever happen, though.)
1828 : : */
1829 : : static void
2397 1830 : 23109 : generate_orderedappend_paths(PlannerInfo *root, RelOptInfo *rel,
1831 : : List *live_childrels,
1832 : : List *all_child_pathkeys)
1833 : : {
1834 : : ListCell *lcp;
1835 : 23109 : List *partition_pathkeys = NIL;
1836 : 23109 : List *partition_pathkeys_desc = NIL;
1837 : 23109 : bool partition_pathkeys_partial = true;
1838 : 23109 : bool partition_pathkeys_desc_partial = true;
1839 : :
1840 : : /*
1841 : : * Some partitioned table setups may allow us to use an Append node
1842 : : * instead of a MergeAppend. This is possible in cases such as RANGE
1843 : : * partitioned tables where it's guaranteed that an earlier partition must
1844 : : * contain rows which come earlier in the sort order. To detect whether
1845 : : * this is relevant, build pathkey descriptions of the partition ordering,
1846 : : * for both forward and reverse scans.
1847 : : */
1848 [ + + + + : 37312 : if (rel->part_scheme != NULL && IS_SIMPLE_REL(rel) &&
+ + + + ]
1546 drowley@postgresql.o 1849 : 14203 : partitions_are_ordered(rel->boundinfo, rel->live_parts))
1850 : : {
2397 tgl@sss.pgh.pa.us 1851 : 11915 : partition_pathkeys = build_partition_pathkeys(root, rel,
1852 : : ForwardScanDirection,
1853 : : &partition_pathkeys_partial);
1854 : :
1855 : 11915 : partition_pathkeys_desc = build_partition_pathkeys(root, rel,
1856 : : BackwardScanDirection,
1857 : : &partition_pathkeys_desc_partial);
1858 : :
1859 : : /*
1860 : : * You might think we should truncate_useless_pathkeys here, but
1861 : : * allowing partition keys which are a subset of the query's pathkeys
1862 : : * can often be useful. For example, consider a table partitioned by
1863 : : * RANGE (a, b), and a query with ORDER BY a, b, c. If we have child
1864 : : * paths that can produce the a, b, c ordering (perhaps via indexes on
1865 : : * (a, b, c)) then it works to consider the appendrel output as
1866 : : * ordered by a, b, c.
1867 : : */
1868 : : }
1869 : :
1870 : : /* Now consider each interesting sort ordering */
5022 1871 [ + + + + : 29151 : foreach(lcp, all_child_pathkeys)
+ + ]
1872 : : {
1873 : 6042 : List *pathkeys = (List *) lfirst(lcp);
5314 bruce@momjian.us 1874 : 6042 : List *startup_subpaths = NIL;
1875 : 6042 : List *total_subpaths = NIL;
1384 tomas.vondra@postgre 1876 : 6042 : List *fractional_subpaths = NIL;
5314 bruce@momjian.us 1877 : 6042 : bool startup_neq_total = false;
1878 : : bool match_partition_order;
1879 : : bool match_partition_order_desc;
1880 : : int end_index;
1881 : : int first_index;
1882 : : int direction;
1883 : :
1884 : : /*
1885 : : * Determine if this sort ordering matches any partition pathkeys we
1886 : : * have, for both ascending and descending partition order. If the
1887 : : * partition pathkeys happen to be contained in pathkeys then it still
1888 : : * works, as described above, providing that the partition pathkeys
1889 : : * are complete and not just a prefix of the partition keys. (In such
1890 : : * cases we'll be relying on the child paths to have sorted the
1891 : : * lower-order columns of the required pathkeys.)
1892 : : */
2397 tgl@sss.pgh.pa.us 1893 : 6042 : match_partition_order =
1894 [ + + ]: 10950 : pathkeys_contained_in(pathkeys, partition_pathkeys) ||
1895 [ + + + + ]: 5012 : (!partition_pathkeys_partial &&
1896 : 104 : pathkeys_contained_in(partition_pathkeys, pathkeys));
1897 : :
1898 [ + + + + ]: 15732 : match_partition_order_desc = !match_partition_order &&
1899 : 4854 : (pathkeys_contained_in(pathkeys, partition_pathkeys_desc) ||
1900 [ + + + + ]: 4868 : (!partition_pathkeys_desc_partial &&
1901 : 32 : pathkeys_contained_in(partition_pathkeys_desc, pathkeys)));
1902 : :
1903 : : /*
1904 : : * When the required pathkeys match the reverse of the partition
1905 : : * order, we must build the list of paths in reverse starting with the
1906 : : * last matching partition first. We can get away without making any
1907 : : * special cases for this in the loop below by just looping backward
1908 : : * over the child relations in this case.
1909 : : */
980 drowley@postgresql.o 1910 [ + + ]: 6042 : if (match_partition_order_desc)
1911 : : {
1912 : : /* loop backward */
1913 : 24 : first_index = list_length(live_childrels) - 1;
1914 : 24 : end_index = -1;
1915 : 24 : direction = -1;
1916 : :
1917 : : /*
1918 : : * Set this to true to save us having to check for
1919 : : * match_partition_order_desc in the loop below.
1920 : : */
1921 : 24 : match_partition_order = true;
1922 : : }
1923 : : else
1924 : : {
1925 : : /* for all other case, loop forward */
1926 : 6018 : first_index = 0;
1927 : 6018 : end_index = list_length(live_childrels);
1928 : 6018 : direction = 1;
1929 : : }
1930 : :
1931 : : /* Select the child paths for this ordering... */
1932 [ + + ]: 21699 : for (int i = first_index; i != end_index; i += direction)
1933 : : {
1934 : 15657 : RelOptInfo *childrel = list_nth_node(RelOptInfo, live_childrels, i);
1935 : : Path *cheapest_startup,
1936 : : *cheapest_total,
1384 tomas.vondra@postgre 1937 : 15657 : *cheapest_fractional = NULL;
1938 : :
1939 : : /* Locate the right paths, if they are available. */
1940 : : cheapest_startup =
5492 tgl@sss.pgh.pa.us 1941 : 15657 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1942 : : pathkeys,
1943 : : NULL,
1944 : : STARTUP_COST,
1945 : : false);
1946 : : cheapest_total =
1947 : 15657 : get_cheapest_path_for_pathkeys(childrel->pathlist,
1948 : : pathkeys,
1949 : : NULL,
1950 : : TOTAL_COST,
1951 : : false);
1952 : :
1953 : : /*
1954 : : * If we can't find any paths with the right order just use the
1955 : : * cheapest-total path; we'll have to sort it later.
1956 : : */
5022 1957 [ + + - + ]: 15657 : if (cheapest_startup == NULL || cheapest_total == NULL)
1958 : : {
4939 1959 : 170 : cheapest_startup = cheapest_total =
1960 : : childrel->cheapest_total_path;
1961 : : /* Assert we do have an unparameterized path for this child */
4825 1962 [ - + ]: 170 : Assert(cheapest_total->param_info == NULL);
1963 : : }
1964 : :
1965 : : /*
1966 : : * When building a fractional path, determine a cheapest
1967 : : * fractional path for each child relation too. Looking at startup
1968 : : * and total costs is not enough, because the cheapest fractional
1969 : : * path may be dominated by two separate paths (one for startup,
1970 : : * one for total).
1971 : : *
1972 : : * When needed (building fractional path), determine the cheapest
1973 : : * fractional path too.
1974 : : */
1384 tomas.vondra@postgre 1975 [ + + ]: 15657 : if (root->tuple_fraction > 0)
1976 : : {
162 akorotkov@postgresql 1977 : 430 : double path_fraction = root->tuple_fraction;
1978 : :
1979 : : /*
1980 : : * Merge Append considers only live children relations. Dummy
1981 : : * relations must be filtered out before.
1982 : : */
1983 [ - + ]: 430 : Assert(childrel->rows > 0);
1984 : :
1985 : : /* Convert absolute limit to a path fraction */
1986 [ + - ]: 430 : if (path_fraction >= 1.0)
1987 : 430 : path_fraction /= childrel->rows;
1988 : :
1989 : : cheapest_fractional =
1384 tomas.vondra@postgre 1990 : 430 : get_cheapest_fractional_path_for_pathkeys(childrel->pathlist,
1991 : : pathkeys,
1992 : : NULL,
1993 : : path_fraction);
1994 : :
1995 : : /*
1996 : : * If we found no path with matching pathkeys, use the
1997 : : * cheapest total path instead.
1998 : : *
1999 : : * XXX We might consider partially sorted paths too (with an
2000 : : * incremental sort on top). But we'd have to build all the
2001 : : * incremental paths, do the costing etc.
2002 : : */
2003 [ + + ]: 430 : if (!cheapest_fractional)
2004 : 22 : cheapest_fractional = cheapest_total;
2005 : : }
2006 : :
2007 : : /*
2008 : : * Notice whether we actually have different paths for the
2009 : : * "cheapest" and "total" cases; frequently there will be no point
2010 : : * in two create_merge_append_path() calls.
2011 : : */
5492 tgl@sss.pgh.pa.us 2012 [ + + ]: 15657 : if (cheapest_startup != cheapest_total)
2013 : 48 : startup_neq_total = true;
2014 : :
2015 : : /*
2016 : : * Collect the appropriate child paths. The required logic varies
2017 : : * for the Append and MergeAppend cases.
2018 : : */
2397 2019 [ + + ]: 15657 : if (match_partition_order)
2020 : : {
2021 : : /*
2022 : : * We're going to make a plain Append path. We don't need
2023 : : * most of what accumulate_append_subpath would do, but we do
2024 : : * want to cut out child Appends or MergeAppends if they have
2025 : : * just a single subpath (and hence aren't doing anything
2026 : : * useful).
2027 : : */
2028 : 3238 : cheapest_startup = get_singleton_append_subpath(cheapest_startup);
2029 : 3238 : cheapest_total = get_singleton_append_subpath(cheapest_total);
2030 : :
2031 : 3238 : startup_subpaths = lappend(startup_subpaths, cheapest_startup);
2032 : 3238 : total_subpaths = lappend(total_subpaths, cheapest_total);
2033 : :
1384 tomas.vondra@postgre 2034 [ + + ]: 3238 : if (cheapest_fractional)
2035 : : {
2036 : 72 : cheapest_fractional = get_singleton_append_subpath(cheapest_fractional);
2037 : 72 : fractional_subpaths = lappend(fractional_subpaths, cheapest_fractional);
2038 : : }
2039 : : }
2040 : : else
2041 : : {
2042 : : /*
2043 : : * Otherwise, rely on accumulate_append_subpath to collect the
2044 : : * child paths for the MergeAppend.
2045 : : */
2397 tgl@sss.pgh.pa.us 2046 : 12419 : accumulate_append_subpath(cheapest_startup,
2047 : : &startup_subpaths, NULL);
2048 : 12419 : accumulate_append_subpath(cheapest_total,
2049 : : &total_subpaths, NULL);
2050 : :
1384 tomas.vondra@postgre 2051 [ + + ]: 12419 : if (cheapest_fractional)
2052 : 358 : accumulate_append_subpath(cheapest_fractional,
2053 : : &fractional_subpaths, NULL);
2054 : : }
2055 : : }
2056 : :
2057 : : /* ... and build the Append or MergeAppend paths */
980 drowley@postgresql.o 2058 [ + + ]: 6042 : if (match_partition_order)
2059 : : {
2060 : : /* We only need Append */
2397 tgl@sss.pgh.pa.us 2061 : 1212 : add_path(rel, (Path *) create_append_path(root,
2062 : : rel,
2063 : : startup_subpaths,
2064 : : NIL,
2065 : : pathkeys,
2066 : : NULL,
2067 : : 0,
2068 : : false,
2069 : : -1));
2070 [ - + ]: 1212 : if (startup_neq_total)
2397 tgl@sss.pgh.pa.us 2071 :UBC 0 : add_path(rel, (Path *) create_append_path(root,
2072 : : rel,
2073 : : total_subpaths,
2074 : : NIL,
2075 : : pathkeys,
2076 : : NULL,
2077 : : 0,
2078 : : false,
2079 : : -1));
2080 : :
1384 tomas.vondra@postgre 2081 [ + + ]:CBC 1212 : if (fractional_subpaths)
2082 : 36 : add_path(rel, (Path *) create_append_path(root,
2083 : : rel,
2084 : : fractional_subpaths,
2085 : : NIL,
2086 : : pathkeys,
2087 : : NULL,
2088 : : 0,
2089 : : false,
2090 : : -1));
2091 : : }
2092 : : else
2093 : : {
2094 : : /* We need MergeAppend */
5492 tgl@sss.pgh.pa.us 2095 : 4830 : add_path(rel, (Path *) create_merge_append_path(root,
2096 : : rel,
2097 : : startup_subpaths,
2098 : : pathkeys,
2099 : : NULL));
2397 2100 [ + + ]: 4830 : if (startup_neq_total)
2101 : 30 : add_path(rel, (Path *) create_merge_append_path(root,
2102 : : rel,
2103 : : total_subpaths,
2104 : : pathkeys,
2105 : : NULL));
2106 : :
1384 tomas.vondra@postgre 2107 [ + + ]: 4830 : if (fractional_subpaths)
2108 : 128 : add_path(rel, (Path *) create_merge_append_path(root,
2109 : : rel,
2110 : : fractional_subpaths,
2111 : : pathkeys,
2112 : : NULL));
2113 : : }
2114 : : }
10702 scrappy@hub.org 2115 : 23109 : }
2116 : :
2117 : : /*
2118 : : * get_cheapest_parameterized_child_path
2119 : : * Get cheapest path for this relation that has exactly the requested
2120 : : * parameterization.
2121 : : *
2122 : : * Returns NULL if unable to create such a path.
2123 : : */
2124 : : static Path *
4495 tgl@sss.pgh.pa.us 2125 : 3352 : get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
2126 : : Relids required_outer)
2127 : : {
2128 : : Path *cheapest;
2129 : : ListCell *lc;
2130 : :
2131 : : /*
2132 : : * Look up the cheapest existing path with no more than the needed
2133 : : * parameterization. If it has exactly the needed parameterization, we're
2134 : : * done.
2135 : : */
2136 : 3352 : cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
2137 : : NIL,
2138 : : required_outer,
2139 : : TOTAL_COST,
2140 : : false);
2141 [ - + ]: 3352 : Assert(cheapest != NULL);
2142 [ + + + + ]: 3352 : if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
2143 : 3182 : return cheapest;
2144 : :
2145 : : /*
2146 : : * Otherwise, we can "reparameterize" an existing path to match the given
2147 : : * parameterization, which effectively means pushing down additional
2148 : : * joinquals to be checked within the path's scan. However, some existing
2149 : : * paths might check the available joinquals already while others don't;
2150 : : * therefore, it's not clear which existing path will be cheapest after
2151 : : * reparameterization. We have to go through them all and find out.
2152 : : */
2153 : 170 : cheapest = NULL;
2154 [ + - + + : 590 : foreach(lc, rel->pathlist)
+ + ]
2155 : : {
2156 : 420 : Path *path = (Path *) lfirst(lc);
2157 : :
2158 : : /* Can't use it if it needs more than requested parameterization */
2159 [ + + + + ]: 420 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
2160 : 12 : continue;
2161 : :
2162 : : /*
2163 : : * Reparameterization can only increase the path's cost, so if it's
2164 : : * already more expensive than the current cheapest, forget it.
2165 : : */
2166 [ + + + + ]: 636 : if (cheapest != NULL &&
2167 : 228 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2168 : 192 : continue;
2169 : :
2170 : : /* Reparameterize if needed, then recheck cost */
2171 [ + + + + ]: 216 : if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
2172 : : {
2173 : 178 : path = reparameterize_path(root, path, required_outer, 1.0);
2174 [ + + ]: 178 : if (path == NULL)
2175 : 16 : continue; /* failed to reparameterize this one */
2176 [ + - - + ]: 162 : Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
2177 : :
2178 [ - + - - ]: 162 : if (cheapest != NULL &&
4495 tgl@sss.pgh.pa.us 2179 :UBC 0 : compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
2180 : 0 : continue;
2181 : : }
2182 : :
2183 : : /* We have a new best path */
4495 tgl@sss.pgh.pa.us 2184 :CBC 200 : cheapest = path;
2185 : : }
2186 : :
2187 : : /* Return the best path, or NULL if we found no suitable candidate */
2188 : 170 : return cheapest;
2189 : : }
2190 : :
2191 : : /*
2192 : : * accumulate_append_subpath
2193 : : * Add a subpath to the list being built for an Append or MergeAppend.
2194 : : *
2195 : : * It's possible that the child is itself an Append or MergeAppend path, in
2196 : : * which case we can "cut out the middleman" and just add its child paths to
2197 : : * our own list. (We don't try to do this earlier because we need to apply
2198 : : * both levels of transformation to the quals.)
2199 : : *
2200 : : * Note that if we omit a child MergeAppend in this way, we are effectively
2201 : : * omitting a sort step, which seems fine: if the parent is to be an Append,
2202 : : * its result would be unsorted anyway, while if the parent is to be a
2203 : : * MergeAppend, there's no point in a separate sort on a child.
2204 : : *
2205 : : * Normally, either path is a partial path and subpaths is a list of partial
2206 : : * paths, or else path is a non-partial plan and subpaths is a list of those.
2207 : : * However, if path is a parallel-aware Append, then we add its partial path
2208 : : * children to subpaths and the rest to special_subpaths. If the latter is
2209 : : * NULL, we don't flatten the path at all (unless it contains only partial
2210 : : * paths).
2211 : : */
2212 : : static void
1729 2213 : 150265 : accumulate_append_subpath(Path *path, List **subpaths, List **special_subpaths)
2214 : : {
5492 2215 [ + + ]: 150265 : if (IsA(path, AppendPath))
2216 : : {
5314 bruce@momjian.us 2217 : 7936 : AppendPath *apath = (AppendPath *) path;
2218 : :
2883 rhaas@postgresql.org 2219 [ + + + + ]: 7936 : if (!apath->path.parallel_aware || apath->first_partial_path == 0)
2220 : : {
2268 tgl@sss.pgh.pa.us 2221 : 7768 : *subpaths = list_concat(*subpaths, apath->subpaths);
2883 rhaas@postgresql.org 2222 : 7768 : return;
2223 : : }
2224 [ + + ]: 168 : else if (special_subpaths != NULL)
2225 : : {
2226 : : List *new_special_subpaths;
2227 : :
2228 : : /* Split Parallel Append into partial and non-partial subpaths */
2229 : 84 : *subpaths = list_concat(*subpaths,
2230 : 84 : list_copy_tail(apath->subpaths,
2231 : : apath->first_partial_path));
1202 drowley@postgresql.o 2232 : 84 : new_special_subpaths = list_copy_head(apath->subpaths,
2233 : : apath->first_partial_path);
2883 rhaas@postgresql.org 2234 : 84 : *special_subpaths = list_concat(*special_subpaths,
2235 : : new_special_subpaths);
2847 2236 : 84 : return;
2237 : : }
2238 : : }
4231 tgl@sss.pgh.pa.us 2239 [ + + ]: 142329 : else if (IsA(path, MergeAppendPath))
2240 : : {
2241 : 610 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2242 : :
2268 2243 : 610 : *subpaths = list_concat(*subpaths, mpath->subpaths);
2883 rhaas@postgresql.org 2244 : 610 : return;
2245 : : }
2246 : :
2247 : 141803 : *subpaths = lappend(*subpaths, path);
2248 : : }
2249 : :
2250 : : /*
2251 : : * get_singleton_append_subpath
2252 : : * Returns the single subpath of an Append/MergeAppend, or just
2253 : : * return 'path' if it's not a single sub-path Append/MergeAppend.
2254 : : *
2255 : : * Note: 'path' must not be a parallel-aware path.
2256 : : */
2257 : : static Path *
2397 tgl@sss.pgh.pa.us 2258 : 6548 : get_singleton_append_subpath(Path *path)
2259 : : {
2260 [ - + ]: 6548 : Assert(!path->parallel_aware);
2261 : :
2262 [ + + ]: 6548 : if (IsA(path, AppendPath))
2263 : : {
2264 : 194 : AppendPath *apath = (AppendPath *) path;
2265 : :
2266 [ + + ]: 194 : if (list_length(apath->subpaths) == 1)
2267 : 96 : return (Path *) linitial(apath->subpaths);
2268 : : }
2269 [ + + ]: 6354 : else if (IsA(path, MergeAppendPath))
2270 : : {
2271 : 174 : MergeAppendPath *mpath = (MergeAppendPath *) path;
2272 : :
2273 [ - + ]: 174 : if (list_length(mpath->subpaths) == 1)
2397 tgl@sss.pgh.pa.us 2274 :UBC 0 : return (Path *) linitial(mpath->subpaths);
2275 : : }
2276 : :
2397 tgl@sss.pgh.pa.us 2277 :CBC 6452 : return path;
2278 : : }
2279 : :
2280 : : /*
2281 : : * set_dummy_rel_pathlist
2282 : : * Build a dummy path for a relation that's been excluded by constraints
2283 : : *
2284 : : * Rather than inventing a special "dummy" path type, we represent this as an
2285 : : * AppendPath with no members (see also IS_DUMMY_APPEND/IS_DUMMY_REL macros).
2286 : : *
2287 : : * (See also mark_dummy_rel, which does basically the same thing, but is
2288 : : * typically used to change a rel into dummy state after we already made
2289 : : * paths for it.)
2290 : : */
2291 : : static void
6729 2292 : 653 : set_dummy_rel_pathlist(RelOptInfo *rel)
2293 : : {
2294 : : /* Set dummy size estimates --- we leave attr_widths[] as zeroes */
2295 : 653 : rel->rows = 0;
3514 2296 : 653 : rel->reltarget->width = 0;
2297 : :
2298 : : /* Discard any pre-existing paths; no further need for them */
5022 2299 : 653 : rel->pathlist = NIL;
3568 rhaas@postgresql.org 2300 : 653 : rel->partial_pathlist = NIL;
2301 : :
2302 : : /* Set up the dummy path */
2419 tgl@sss.pgh.pa.us 2303 : 653 : add_path(rel, (Path *) create_append_path(NULL, rel, NIL, NIL,
2304 : : NIL, rel->lateral_relids,
2305 : : 0, false, -1));
2306 : :
2307 : : /*
2308 : : * We set the cheapest-path fields immediately, just in case they were
2309 : : * pointing at some discarded path. This is redundant in current usage
2310 : : * because set_rel_pathlist will do it later, but it's cheap so we keep it
2311 : : * for safety and consistency with mark_dummy_rel.
2312 : : */
6729 2313 : 653 : set_cheapest(rel);
2314 : 653 : }
2315 : :
2316 : : /*
2317 : : * find_window_run_conditions
2318 : : * Determine if 'wfunc' is really a WindowFunc and call its prosupport
2319 : : * function to determine the function's monotonic properties. We then
2320 : : * see if 'opexpr' can be used to short-circuit execution.
2321 : : *
2322 : : * For example row_number() over (order by ...) always produces a value one
2323 : : * higher than the previous. If someone has a window function in a subquery
2324 : : * and has a WHERE clause in the outer query to filter rows <= 10, then we may
2325 : : * as well stop processing the windowagg once the row number reaches 11. Here
2326 : : * we check if 'opexpr' might help us to stop doing needless extra processing
2327 : : * in WindowAgg nodes.
2328 : : *
2329 : : * '*keep_original' is set to true if the caller should also use 'opexpr' for
2330 : : * its original purpose. This is set to false if the caller can assume that
2331 : : * the run condition will handle all of the required filtering.
2332 : : *
2333 : : * Returns true if 'opexpr' was found to be useful and was added to the
2334 : : * WindowFunc's runCondition. We also set *keep_original accordingly and add
2335 : : * 'attno' to *run_cond_attrs offset by FirstLowInvalidHeapAttributeNumber.
2336 : : * If the 'opexpr' cannot be used then we set *keep_original to true and
2337 : : * return false.
2338 : : */
2339 : : static bool
27 drowley@postgresql.o 2340 :GNC 120 : find_window_run_conditions(Query *subquery, AttrNumber attno,
2341 : : WindowFunc *wfunc, OpExpr *opexpr, bool wfunc_left,
2342 : : bool *keep_original, Bitmapset **run_cond_attrs)
2343 : : {
2344 : : Oid prosupport;
2345 : : Expr *otherexpr;
2346 : : SupportRequestWFuncMonotonic req;
2347 : : SupportRequestWFuncMonotonic *res;
2348 : : WindowClause *wclause;
2349 : : List *opinfos;
2350 : : OpExpr *runopexpr;
2351 : : Oid runoperator;
2352 : : ListCell *lc;
2353 : :
1298 drowley@postgresql.o 2354 :CBC 120 : *keep_original = true;
2355 : :
2356 [ - + ]: 120 : while (IsA(wfunc, RelabelType))
1298 drowley@postgresql.o 2357 :UBC 0 : wfunc = (WindowFunc *) ((RelabelType *) wfunc)->arg;
2358 : :
2359 : : /* we can only work with window functions */
1298 drowley@postgresql.o 2360 [ + + ]:CBC 120 : if (!IsA(wfunc, WindowFunc))
2361 : 12 : return false;
2362 : :
2363 : : /* can't use it if there are subplans in the WindowFunc */
955 2364 [ + + ]: 108 : if (contain_subplans((Node *) wfunc))
2365 : 3 : return false;
2366 : :
1298 2367 : 105 : prosupport = get_func_support(wfunc->winfnoid);
2368 : :
2369 : : /* Check if there's a support function for 'wfunc' */
2370 [ + + ]: 105 : if (!OidIsValid(prosupport))
2371 : 9 : return false;
2372 : :
2373 : : /* get the Expr from the other side of the OpExpr */
2374 [ + + ]: 96 : if (wfunc_left)
2375 : 84 : otherexpr = lsecond(opexpr->args);
2376 : : else
2377 : 12 : otherexpr = linitial(opexpr->args);
2378 : :
2379 : : /*
2380 : : * The value being compared must not change during the evaluation of the
2381 : : * window partition.
2382 : : */
2383 [ - + ]: 96 : if (!is_pseudo_constant_clause((Node *) otherexpr))
1298 drowley@postgresql.o 2384 :UBC 0 : return false;
2385 : :
2386 : : /* find the window clause belonging to the window function */
1298 drowley@postgresql.o 2387 :CBC 96 : wclause = (WindowClause *) list_nth(subquery->windowClause,
2388 : 96 : wfunc->winref - 1);
2389 : :
2390 : 96 : req.type = T_SupportRequestWFuncMonotonic;
2391 : 96 : req.window_func = wfunc;
2392 : 96 : req.window_clause = wclause;
2393 : :
2394 : : /* call the support function */
2395 : : res = (SupportRequestWFuncMonotonic *)
2396 : 96 : DatumGetPointer(OidFunctionCall1(prosupport,
2397 : : PointerGetDatum(&req)));
2398 : :
2399 : : /*
2400 : : * Nothing to do if the function is neither monotonically increasing nor
2401 : : * monotonically decreasing.
2402 : : */
2403 [ + - - + ]: 96 : if (res == NULL || res->monotonic == MONOTONICFUNC_NONE)
1298 drowley@postgresql.o 2404 :UBC 0 : return false;
2405 : :
1298 drowley@postgresql.o 2406 :CBC 96 : runopexpr = NULL;
2407 : 96 : runoperator = InvalidOid;
204 peter@eisentraut.org 2408 : 96 : opinfos = get_op_index_interpretation(opexpr->opno);
2409 : :
1298 drowley@postgresql.o 2410 [ + - + - : 96 : foreach(lc, opinfos)
+ - ]
2411 : : {
204 peter@eisentraut.org 2412 : 96 : OpIndexInterpretation *opinfo = (OpIndexInterpretation *) lfirst(lc);
2413 : 96 : CompareType cmptype = opinfo->cmptype;
2414 : :
2415 : : /* handle < / <= */
2416 [ + + + + ]: 96 : if (cmptype == COMPARE_LT || cmptype == COMPARE_LE)
2417 : : {
2418 : : /*
2419 : : * < / <= is supported for monotonically increasing functions in
2420 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2421 : : * for monotonically decreasing functions.
2422 : : */
1298 drowley@postgresql.o 2423 [ + + + + ]: 69 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)) ||
2424 [ + + + + ]: 9 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)))
2425 : : {
2426 : 63 : *keep_original = false;
2427 : 63 : runopexpr = opexpr;
2428 : 63 : runoperator = opexpr->opno;
2429 : : }
2430 : 69 : break;
2431 : : }
2432 : : /* handle > / >= */
204 peter@eisentraut.org 2433 [ + + + + ]: 27 : else if (cmptype == COMPARE_GT || cmptype == COMPARE_GE)
2434 : : {
2435 : : /*
2436 : : * > / >= is supported for monotonically decreasing functions in
2437 : : * the form <wfunc> op <pseudoconst> and <pseudoconst> op <wfunc>
2438 : : * for monotonically increasing functions.
2439 : : */
1298 drowley@postgresql.o 2440 [ + + - + ]: 9 : if ((wfunc_left && (res->monotonic & MONOTONICFUNC_DECREASING)) ||
2441 [ + - + - ]: 6 : (!wfunc_left && (res->monotonic & MONOTONICFUNC_INCREASING)))
2442 : : {
2443 : 9 : *keep_original = false;
2444 : 9 : runopexpr = opexpr;
2445 : 9 : runoperator = opexpr->opno;
2446 : : }
2447 : 9 : break;
2448 : : }
2449 : : /* handle = */
204 peter@eisentraut.org 2450 [ + - ]: 18 : else if (cmptype == COMPARE_EQ)
2451 : : {
2452 : : CompareType newcmptype;
2453 : :
2454 : : /*
2455 : : * When both monotonically increasing and decreasing then the
2456 : : * return value of the window function will be the same each time.
2457 : : * We can simply use 'opexpr' as the run condition without
2458 : : * modifying it.
2459 : : */
1298 drowley@postgresql.o 2460 [ + + ]: 18 : if ((res->monotonic & MONOTONICFUNC_BOTH) == MONOTONICFUNC_BOTH)
2461 : : {
2462 : 3 : *keep_original = false;
2463 : 3 : runopexpr = opexpr;
1179 2464 : 3 : runoperator = opexpr->opno;
1298 2465 : 3 : break;
2466 : : }
2467 : :
2468 : : /*
2469 : : * When monotonically increasing we make a qual with <wfunc> <=
2470 : : * <value> or <value> >= <wfunc> in order to filter out values
2471 : : * which are above the value in the equality condition. For
2472 : : * monotonically decreasing functions we want to filter values
2473 : : * below the value in the equality condition.
2474 : : */
2475 [ + - ]: 15 : if (res->monotonic & MONOTONICFUNC_INCREASING)
204 peter@eisentraut.org 2476 [ + - ]: 15 : newcmptype = wfunc_left ? COMPARE_LE : COMPARE_GE;
2477 : : else
204 peter@eisentraut.org 2478 [ # # ]:UBC 0 : newcmptype = wfunc_left ? COMPARE_GE : COMPARE_LE;
2479 : :
2480 : : /* We must keep the original equality qual */
1298 drowley@postgresql.o 2481 :CBC 15 : *keep_original = true;
2482 : 15 : runopexpr = opexpr;
2483 : :
2484 : : /* determine the operator to use for the WindowFuncRunCondition */
204 peter@eisentraut.org 2485 : 15 : runoperator = get_opfamily_member_for_cmptype(opinfo->opfamily_id,
2486 : : opinfo->oplefttype,
2487 : : opinfo->oprighttype,
2488 : : newcmptype);
1298 drowley@postgresql.o 2489 : 15 : break;
2490 : : }
2491 : : }
2492 : :
2493 [ + + ]: 96 : if (runopexpr != NULL)
2494 : : {
2495 : : WindowFuncRunCondition *wfuncrc;
2496 : :
540 2497 : 90 : wfuncrc = makeNode(WindowFuncRunCondition);
2498 : 90 : wfuncrc->opno = runoperator;
2499 : 90 : wfuncrc->inputcollid = runopexpr->inputcollid;
2500 : 90 : wfuncrc->wfunc_left = wfunc_left;
2501 : 90 : wfuncrc->arg = copyObject(otherexpr);
2502 : :
2503 : 90 : wfunc->runCondition = lappend(wfunc->runCondition, wfuncrc);
2504 : :
2505 : : /* record that this attno was used in a run condition */
1249 2506 : 90 : *run_cond_attrs = bms_add_member(*run_cond_attrs,
2507 : : attno - FirstLowInvalidHeapAttributeNumber);
1298 2508 : 90 : return true;
2509 : : }
2510 : :
2511 : : /* unsupported OpExpr */
2512 : 6 : return false;
2513 : : }
2514 : :
2515 : : /*
2516 : : * check_and_push_window_quals
2517 : : * Check if 'clause' is a qual that can be pushed into a WindowFunc
2518 : : * as a 'runCondition' qual. These, when present, allow some unnecessary
2519 : : * work to be skipped during execution.
2520 : : *
2521 : : * 'run_cond_attrs' will be populated with all targetlist resnos of subquery
2522 : : * targets (offset by FirstLowInvalidHeapAttributeNumber) that we pushed
2523 : : * window quals for.
2524 : : *
2525 : : * Returns true if the caller still must keep the original qual or false if
2526 : : * the caller can safely ignore the original qual because the WindowAgg node
2527 : : * will use the runCondition to stop returning tuples.
2528 : : */
2529 : : static bool
27 drowley@postgresql.o 2530 :GNC 126 : check_and_push_window_quals(Query *subquery, Node *clause,
2531 : : Bitmapset **run_cond_attrs)
2532 : : {
1298 drowley@postgresql.o 2533 :CBC 126 : OpExpr *opexpr = (OpExpr *) clause;
2534 : 126 : bool keep_original = true;
2535 : : Var *var1;
2536 : : Var *var2;
2537 : :
2538 : : /* We're only able to use OpExprs with 2 operands */
2539 [ + + ]: 126 : if (!IsA(opexpr, OpExpr))
2540 : 9 : return true;
2541 : :
2542 [ - + ]: 117 : if (list_length(opexpr->args) != 2)
1298 drowley@postgresql.o 2543 :UBC 0 : return true;
2544 : :
2545 : : /*
2546 : : * Currently, we restrict this optimization to strict OpExprs. The reason
2547 : : * for this is that during execution, once the runcondition becomes false,
2548 : : * we stop evaluating WindowFuncs. To avoid leaving around stale window
2549 : : * function result values, we set them to NULL. Having only strict
2550 : : * OpExprs here ensures that we properly filter out the tuples with NULLs
2551 : : * in the top-level WindowAgg.
2552 : : */
1055 drowley@postgresql.o 2553 :CBC 117 : set_opfuncid(opexpr);
2554 [ - + ]: 117 : if (!func_strict(opexpr->opfuncid))
1055 drowley@postgresql.o 2555 :UBC 0 : return true;
2556 : :
2557 : : /*
2558 : : * Check for plain Vars that reference window functions in the subquery.
2559 : : * If we find any, we'll ask find_window_run_conditions() if 'opexpr' can
2560 : : * be used as part of the run condition.
2561 : : */
2562 : :
2563 : : /* Check the left side of the OpExpr */
1298 drowley@postgresql.o 2564 :CBC 117 : var1 = linitial(opexpr->args);
2565 [ + + + - ]: 117 : if (IsA(var1, Var) && var1->varattno > 0)
2566 : : {
2567 : 99 : TargetEntry *tle = list_nth(subquery->targetList, var1->varattno - 1);
2568 : 99 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2569 : :
27 drowley@postgresql.o 2570 [ + + ]:GNC 99 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2571 : : true, &keep_original, run_cond_attrs))
1298 drowley@postgresql.o 2572 :CBC 81 : return keep_original;
2573 : : }
2574 : :
2575 : : /* and check the right side */
2576 : 36 : var2 = lsecond(opexpr->args);
2577 [ + + + - ]: 36 : if (IsA(var2, Var) && var2->varattno > 0)
2578 : : {
2579 : 21 : TargetEntry *tle = list_nth(subquery->targetList, var2->varattno - 1);
2580 : 21 : WindowFunc *wfunc = (WindowFunc *) tle->expr;
2581 : :
27 drowley@postgresql.o 2582 [ + + ]:GNC 21 : if (find_window_run_conditions(subquery, tle->resno, wfunc, opexpr,
2583 : : false, &keep_original, run_cond_attrs))
1298 drowley@postgresql.o 2584 :CBC 9 : return keep_original;
2585 : : }
2586 : :
2587 : 27 : return true;
2588 : : }
2589 : :
2590 : : /*
2591 : : * set_subquery_pathlist
2592 : : * Generate SubqueryScan access paths for a subquery RTE
2593 : : *
2594 : : * We don't currently support generating parameterized paths for subqueries
2595 : : * by pushing join clauses down into them; it seems too expensive to re-plan
2596 : : * the subquery multiple times to consider different alternatives.
2597 : : * (XXX that could stand to be reconsidered, now that we use Paths.)
2598 : : * So the paths made here will be parameterized if the subquery contains
2599 : : * LATERAL references, otherwise not. As long as that's true, there's no need
2600 : : * for a separate set_subquery_size phase: just make the paths right away.
2601 : : */
2602 : : static void
7449 tgl@sss.pgh.pa.us 2603 : 8757 : set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
2604 : : Index rti, RangeTblEntry *rte)
2605 : : {
7444 2606 : 8757 : Query *parse = root->parse;
8869 2607 : 8757 : Query *subquery = rte->subquery;
2608 : : bool trivial_pathtarget;
2609 : : Relids required_outer;
2610 : : pushdown_safety_info safetyInfo;
2611 : : double tuple_fraction;
2612 : : RelOptInfo *sub_final_rel;
1249 drowley@postgresql.o 2613 : 8757 : Bitmapset *run_cond_attrs = NULL;
2614 : : ListCell *lc;
2615 : : char *plan_name;
2616 : :
2617 : : /*
2618 : : * Must copy the Query so that planning doesn't mess up the RTE contents
2619 : : * (really really need to fix the planner to not scribble on its input,
2620 : : * someday ... but see remove_unused_subquery_outputs to start with).
2621 : : */
6075 tgl@sss.pgh.pa.us 2622 : 8757 : subquery = copyObject(subquery);
2623 : :
2624 : : /*
2625 : : * If it's a LATERAL subquery, it might contain some Vars of the current
2626 : : * query level, requiring it to be treated as parameterized, even though
2627 : : * we don't support pushing down join quals into subqueries.
2628 : : */
4810 2629 : 8757 : required_outer = rel->lateral_relids;
2630 : :
2631 : : /*
2632 : : * Zero out result area for subquery_is_pushdown_safe, so that it can set
2633 : : * flags as needed while recursing. In particular, we need a workspace
2634 : : * for keeping track of the reasons why columns are unsafe to reference.
2635 : : * These reasons are stored in the bits inside unsafeFlags[i] when we
2636 : : * discover reasons that column i of the subquery is unsafe to be used in
2637 : : * a pushed-down qual.
2638 : : */
4140 2639 : 8757 : memset(&safetyInfo, 0, sizeof(safetyInfo));
955 drowley@postgresql.o 2640 : 8757 : safetyInfo.unsafeFlags = (unsigned char *)
2641 : 8757 : palloc0((list_length(subquery->targetList) + 1) * sizeof(unsigned char));
2642 : :
2643 : : /*
2644 : : * If the subquery has the "security_barrier" flag, it means the subquery
2645 : : * originated from a view that must enforce row-level security. Then we
2646 : : * must not push down quals that contain leaky functions. (Ideally this
2647 : : * would be checked inside subquery_is_pushdown_safe, but since we don't
2648 : : * currently pass the RTE to that function, we must do it here.)
2649 : : */
4140 tgl@sss.pgh.pa.us 2650 : 8757 : safetyInfo.unsafeLeaky = rte->security_barrier;
2651 : :
2652 : : /*
2653 : : * If there are any restriction clauses that have been attached to the
2654 : : * subquery relation, consider pushing them down to become WHERE or HAVING
2655 : : * quals of the subquery itself. This transformation is useful because it
2656 : : * may allow us to generate a better plan for the subquery than evaluating
2657 : : * all the subquery output rows and then filtering them.
2658 : : *
2659 : : * There are several cases where we cannot push down clauses. Restrictions
2660 : : * involving the subquery are checked by subquery_is_pushdown_safe().
2661 : : * Restrictions on individual clauses are checked by
2662 : : * qual_is_pushdown_safe(). Also, we don't want to push down
2663 : : * pseudoconstant clauses; better to have the gating node above the
2664 : : * subquery.
2665 : : *
2666 : : * Non-pushed-down clauses will get evaluated as qpquals of the
2667 : : * SubqueryScan node.
2668 : : *
2669 : : * XXX Are there any cases where we want to make a policy decision not to
2670 : : * push down a pushable qual, because it'd result in a worse plan?
2671 : : */
8460 2672 [ + + + + ]: 9705 : if (rel->baserestrictinfo != NIL &&
4140 2673 : 948 : subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
2674 : : {
2675 : : /* OK to consider pushing down individual quals */
8869 2676 : 875 : List *upperrestrictlist = NIL;
2677 : : ListCell *l;
2678 : :
7824 neilc@samurai.com 2679 [ + - + + : 2226 : foreach(l, rel->baserestrictinfo)
+ + ]
2680 : : {
2681 : 1351 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
1298 drowley@postgresql.o 2682 : 1351 : Node *clause = (Node *) rinfo->clause;
2683 : :
955 2684 [ + + ]: 1351 : if (rinfo->pseudoconstant)
2685 : : {
2686 : 2 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2687 : 2 : continue;
2688 : : }
2689 : :
2690 [ + + + - ]: 1349 : switch (qual_is_pushdown_safe(subquery, rti, rinfo, &safetyInfo))
2691 : : {
2692 : 996 : case PUSHDOWN_SAFE:
2693 : : /* Push it down */
2694 : 996 : subquery_push_qual(subquery, rte, rti, clause);
2695 : 996 : break;
2696 : :
2697 : 126 : case PUSHDOWN_WINDOWCLAUSE_RUNCOND:
2698 : :
2699 : : /*
2700 : : * Since we can't push the qual down into the subquery,
2701 : : * check if it happens to reference a window function. If
2702 : : * so then it might be useful to use for the WindowAgg's
2703 : : * runCondition.
2704 : : */
2705 [ + - + + ]: 252 : if (!subquery->hasWindowFuncs ||
27 drowley@postgresql.o 2706 :GNC 126 : check_and_push_window_quals(subquery, clause,
2707 : : &run_cond_attrs))
2708 : : {
2709 : : /*
2710 : : * subquery has no window funcs or the clause is not a
2711 : : * suitable window run condition qual or it is, but
2712 : : * the original must also be kept in the upper query.
2713 : : */
955 drowley@postgresql.o 2714 :CBC 51 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
2715 : : }
2716 : 126 : break;
2717 : :
2718 : 227 : case PUSHDOWN_UNSAFE:
1298 2719 : 227 : upperrestrictlist = lappend(upperrestrictlist, rinfo);
955 2720 : 227 : break;
2721 : : }
2722 : : }
8869 tgl@sss.pgh.pa.us 2723 : 875 : rel->baserestrictinfo = upperrestrictlist;
2724 : : /* We don't bother recomputing baserestrict_min_security */
2725 : : }
2726 : :
955 drowley@postgresql.o 2727 : 8757 : pfree(safetyInfo.unsafeFlags);
2728 : :
2729 : : /*
2730 : : * The upper query might not use all the subquery's output columns; if
2731 : : * not, we can simplify. Pass the attributes that were pushed down into
2732 : : * WindowAgg run conditions to ensure we don't accidentally think those
2733 : : * are unused.
2734 : : */
1249 2735 : 8757 : remove_unused_subquery_outputs(subquery, rel, run_cond_attrs);
2736 : :
2737 : : /*
2738 : : * We can safely pass the outer tuple_fraction down to the subquery if the
2739 : : * outer level has no joining, aggregation, or sorting to do. Otherwise
2740 : : * we'd better tell the subquery to plan for full retrieval. (XXX This
2741 : : * could probably be made more intelligent ...)
2742 : : */
7444 tgl@sss.pgh.pa.us 2743 [ + + ]: 8757 : if (parse->hasAggs ||
2744 [ + + ]: 8135 : parse->groupClause ||
3817 andres@anarazel.de 2745 [ + - ]: 8126 : parse->groupingSets ||
1105 tgl@sss.pgh.pa.us 2746 [ + - ]: 8126 : root->hasHavingQual ||
7444 2747 [ + + ]: 8126 : parse->distinctClause ||
2748 [ + + + + ]: 10980 : parse->sortClause ||
748 2749 : 3108 : bms_membership(root->all_baserels) == BMS_MULTIPLE)
7444 2750 : 6310 : tuple_fraction = 0.0; /* default case */
2751 : : else
2752 : 2447 : tuple_fraction = root->tuple_fraction;
2753 : :
2754 : : /* plan_params should not be in use in current query level */
4800 2755 [ - + ]: 8757 : Assert(root->plan_params == NIL);
2756 : :
2757 : : /* Generate a subroot and Paths for the subquery */
20 rhaas@postgresql.org 2758 :GNC 8757 : plan_name = choose_plan_name(root->glob, rte->eref->aliasname, false);
2759 : 8757 : rel->subroot = subquery_planner(root->glob, subquery, plan_name,
2760 : : root, false, tuple_fraction, NULL);
2761 : :
2762 : : /* Isolate the params needed by this specific subplan */
4800 tgl@sss.pgh.pa.us 2763 :CBC 8757 : rel->subplan_params = root->plan_params;
2764 : 8757 : root->plan_params = NIL;
2765 : :
2766 : : /*
2767 : : * It's possible that constraint exclusion proved the subquery empty. If
2768 : : * so, it's desirable to produce an unadorned dummy path so that we will
2769 : : * recognize appropriate optimizations at this query level.
2770 : : */
3521 2771 : 8757 : sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
2772 : :
2773 [ + + ]: 8757 : if (IS_DUMMY_REL(sub_final_rel))
2774 : : {
5147 2775 : 63 : set_dummy_rel_pathlist(rel);
2776 : 63 : return;
2777 : : }
2778 : :
2779 : : /*
2780 : : * Mark rel with estimated output rows, width, etc. Note that we have to
2781 : : * do this before generating outer-query paths, else cost_subqueryscan is
2782 : : * not happy.
2783 : : */
5168 2784 : 8694 : set_subquery_size_estimates(root, rel);
2785 : :
2786 : : /*
2787 : : * Also detect whether the reltarget is trivial, so that we can pass that
2788 : : * info to cost_subqueryscan (rather than re-deriving it multiple times).
2789 : : * It's trivial if it fetches all the subplan output columns in order.
2790 : : */
1196 2791 [ + + ]: 8694 : if (list_length(rel->reltarget->exprs) != list_length(subquery->targetList))
2792 : 3855 : trivial_pathtarget = false;
2793 : : else
2794 : : {
2795 : 4839 : trivial_pathtarget = true;
2796 [ + + + + : 15834 : foreach(lc, rel->reltarget->exprs)
+ + ]
2797 : : {
2798 : 11144 : Node *node = (Node *) lfirst(lc);
2799 : : Var *var;
2800 : :
2801 [ - + ]: 11144 : if (!IsA(node, Var))
2802 : : {
1196 tgl@sss.pgh.pa.us 2803 :UBC 0 : trivial_pathtarget = false;
2804 : 0 : break;
2805 : : }
1196 tgl@sss.pgh.pa.us 2806 :CBC 11144 : var = (Var *) node;
2807 [ + - ]: 11144 : if (var->varno != rti ||
2808 [ + + ]: 11144 : var->varattno != foreach_current_index(lc) + 1)
2809 : : {
2810 : 149 : trivial_pathtarget = false;
2811 : 149 : break;
2812 : : }
2813 : : }
2814 : : }
2815 : :
2816 : : /*
2817 : : * For each Path that subquery_planner produced, make a SubqueryScanPath
2818 : : * in the outer query.
2819 : : */
3521 2820 [ + - + + : 18383 : foreach(lc, sub_final_rel->pathlist)
+ + ]
2821 : : {
2822 : 9689 : Path *subpath = (Path *) lfirst(lc);
2823 : : List *pathkeys;
2824 : :
2825 : : /* Convert subpath's pathkeys to outer representation */
2826 : 9689 : pathkeys = convert_subquery_pathkeys(root,
2827 : : rel,
2828 : : subpath->pathkeys,
2829 : : make_tlist_from_pathtarget(subpath->pathtarget));
2830 : :
2831 : : /* Generate outer path using this subpath */
2832 : 9689 : add_path(rel, (Path *)
2833 : 9689 : create_subqueryscan_path(root, rel, subpath,
2834 : : trivial_pathtarget,
2835 : : pathkeys, required_outer));
2836 : : }
2837 : :
2838 : : /* If outer rel allows parallelism, do same for partial paths. */
2742 rhaas@postgresql.org 2839 [ + + + + ]: 8694 : if (rel->consider_parallel && bms_is_empty(required_outer))
2840 : : {
2841 : : /* If consider_parallel is false, there should be no partial paths. */
2842 [ + + - + ]: 6570 : Assert(sub_final_rel->consider_parallel ||
2843 : : sub_final_rel->partial_pathlist == NIL);
2844 : :
2845 : : /* Same for partial paths. */
2846 [ + + + + : 6591 : foreach(lc, sub_final_rel->partial_pathlist)
+ + ]
2847 : : {
2848 : 21 : Path *subpath = (Path *) lfirst(lc);
2849 : : List *pathkeys;
2850 : :
2851 : : /* Convert subpath's pathkeys to outer representation */
2852 : 21 : pathkeys = convert_subquery_pathkeys(root,
2853 : : rel,
2854 : : subpath->pathkeys,
2855 : : make_tlist_from_pathtarget(subpath->pathtarget));
2856 : :
2857 : : /* Generate outer path using this subpath */
2858 : 21 : add_partial_path(rel, (Path *)
2859 : 21 : create_subqueryscan_path(root, rel, subpath,
2860 : : trivial_pathtarget,
2861 : : pathkeys,
2862 : : required_outer));
2863 : : }
2864 : : }
2865 : : }
2866 : :
2867 : : /*
2868 : : * set_function_pathlist
2869 : : * Build the (single) access path for a function RTE
2870 : : */
2871 : : static void
7449 tgl@sss.pgh.pa.us 2872 : 24598 : set_function_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2873 : : {
2874 : : Relids required_outer;
4358 2875 : 24598 : List *pathkeys = NIL;
2876 : :
2877 : : /*
2878 : : * We don't support pushing join clauses into the quals of a function
2879 : : * scan, but it could still have required parameterization due to LATERAL
2880 : : * refs in the function expression.
2881 : : */
4810 2882 : 24598 : required_outer = rel->lateral_relids;
2883 : :
2884 : : /*
2885 : : * The result is considered unordered unless ORDINALITY was used, in which
2886 : : * case it is ordered by the ordinal column (the last one). See if we
2887 : : * care, by checking for uses of that Var in equivalence classes.
2888 : : */
4358 2889 [ + + ]: 24598 : if (rte->funcordinality)
2890 : : {
2891 : 463 : AttrNumber ordattno = rel->max_attr;
2892 : 463 : Var *var = NULL;
2893 : : ListCell *lc;
2894 : :
2895 : : /*
2896 : : * Is there a Var for it in rel's targetlist? If not, the query did
2897 : : * not reference the ordinality column, or at least not in any way
2898 : : * that would be interesting for sorting.
2899 : : */
3514 2900 [ + - + + : 1055 : foreach(lc, rel->reltarget->exprs)
+ + ]
2901 : : {
4358 2902 : 1052 : Var *node = (Var *) lfirst(lc);
2903 : :
2904 : : /* checking varno/varlevelsup is just paranoia */
2905 [ + - ]: 1052 : if (IsA(node, Var) &&
2906 [ + + ]: 1052 : node->varattno == ordattno &&
2907 [ + - ]: 460 : node->varno == rel->relid &&
2908 [ + - ]: 460 : node->varlevelsup == 0)
2909 : : {
2910 : 460 : var = node;
2911 : 460 : break;
2912 : : }
2913 : : }
2914 : :
2915 : : /*
2916 : : * Try to build pathkeys for this Var with int8 sorting. We tell
2917 : : * build_expression_pathkey not to build any new equivalence class; if
2918 : : * the Var isn't already mentioned in some EC, it means that nothing
2919 : : * cares about the ordering.
2920 : : */
2921 [ + + ]: 463 : if (var)
2922 : 460 : pathkeys = build_expression_pathkey(root,
2923 : : (Expr *) var,
2924 : : Int8LessOperator,
2925 : : rel->relids,
2926 : : false);
2927 : : }
2928 : :
2929 : : /* Generate appropriate path */
2930 : 24598 : add_path(rel, create_functionscan_path(root, rel,
2931 : : pathkeys, required_outer));
8569 2932 : 24598 : }
2933 : :
2934 : : /*
2935 : : * set_values_pathlist
2936 : : * Build the (single) access path for a VALUES RTE
2937 : : */
2938 : : static void
7026 mail@joeconway.com 2939 : 4125 : set_values_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2940 : : {
2941 : : Relids required_outer;
2942 : :
2943 : : /*
2944 : : * We don't support pushing join clauses into the quals of a values scan,
2945 : : * but it could still have required parameterization due to LATERAL refs
2946 : : * in the values expressions.
2947 : : */
4810 tgl@sss.pgh.pa.us 2948 : 4125 : required_outer = rel->lateral_relids;
2949 : :
2950 : : /* Generate appropriate path */
4824 2951 : 4125 : add_path(rel, create_valuesscan_path(root, rel, required_outer));
7026 mail@joeconway.com 2952 : 4125 : }
2953 : :
2954 : : /*
2955 : : * set_tablefunc_pathlist
2956 : : * Build the (single) access path for a table func RTE
2957 : : */
2958 : : static void
3155 alvherre@alvh.no-ip. 2959 : 311 : set_tablefunc_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2960 : : {
2961 : : Relids required_outer;
2962 : :
2963 : : /*
2964 : : * We don't support pushing join clauses into the quals of a tablefunc
2965 : : * scan, but it could still have required parameterization due to LATERAL
2966 : : * refs in the function expression.
2967 : : */
2968 : 311 : required_outer = rel->lateral_relids;
2969 : :
2970 : : /* Generate appropriate path */
2971 : 311 : add_path(rel, create_tablefuncscan_path(root, rel,
2972 : : required_outer));
2973 : 311 : }
2974 : :
2975 : : /*
2976 : : * set_cte_pathlist
2977 : : * Build the (single) access path for a non-self-reference CTE RTE
2978 : : *
2979 : : * There's no need for a separate set_cte_size phase, since we don't
2980 : : * support join-qual-parameterized paths for CTEs.
2981 : : */
2982 : : static void
6232 tgl@sss.pgh.pa.us 2983 : 2124 : set_cte_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
2984 : : {
2985 : : Path *ctepath;
2986 : : Plan *cteplan;
2987 : : PlannerInfo *cteroot;
2988 : : Index levelsup;
2989 : : List *pathkeys;
2990 : : int ndx;
2991 : : ListCell *lc;
2992 : : int plan_id;
2993 : : Relids required_outer;
2994 : :
2995 : : /*
2996 : : * Find the referenced CTE, and locate the path and plan previously made
2997 : : * for it.
2998 : : */
2999 : 2124 : levelsup = rte->ctelevelsup;
3000 : 2124 : cteroot = root;
3001 [ + + ]: 3706 : while (levelsup-- > 0)
3002 : : {
3003 : 1582 : cteroot = cteroot->parent_root;
3004 [ - + ]: 1582 : if (!cteroot) /* shouldn't happen */
6232 tgl@sss.pgh.pa.us 3005 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3006 : : }
3007 : :
3008 : : /*
3009 : : * Note: cte_plan_ids can be shorter than cteList, if we are still working
3010 : : * on planning the CTEs (ie, this is a side-reference from another CTE).
3011 : : * So we mustn't use forboth here.
3012 : : */
6232 tgl@sss.pgh.pa.us 3013 :CBC 2124 : ndx = 0;
3014 [ + - + - : 2915 : foreach(lc, cteroot->parse->cteList)
+ - ]
3015 : : {
3016 : 2915 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
3017 : :
3018 [ + + ]: 2915 : if (strcmp(cte->ctename, rte->ctename) == 0)
3019 : 2124 : break;
3020 : 791 : ndx++;
3021 : : }
3022 [ - + ]: 2124 : if (lc == NULL) /* shouldn't happen */
6232 tgl@sss.pgh.pa.us 3023 [ # # ]:UBC 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6232 tgl@sss.pgh.pa.us 3024 [ - + ]:CBC 2124 : if (ndx >= list_length(cteroot->cte_plan_ids))
6232 tgl@sss.pgh.pa.us 3025 [ # # ]:UBC 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6232 tgl@sss.pgh.pa.us 3026 :CBC 2124 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
1285 3027 [ - + ]: 2124 : if (plan_id <= 0)
1285 tgl@sss.pgh.pa.us 3028 [ # # ]:UBC 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
3029 : :
580 tgl@sss.pgh.pa.us 3030 [ - + ]:CBC 2124 : Assert(list_length(root->glob->subpaths) == list_length(root->glob->subplans));
3031 : 2124 : ctepath = (Path *) list_nth(root->glob->subpaths, plan_id - 1);
6232 3032 : 2124 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
3033 : :
3034 : : /* Mark rel with estimated output rows, width, etc */
3521 3035 : 2124 : set_cte_size_estimates(root, rel, cteplan->plan_rows);
3036 : :
3037 : : /* Convert the ctepath's pathkeys to outer query's representation */
580 3038 : 2124 : pathkeys = convert_subquery_pathkeys(root,
3039 : : rel,
3040 : : ctepath->pathkeys,
3041 : : cteplan->targetlist);
3042 : :
3043 : : /*
3044 : : * We don't support pushing join clauses into the quals of a CTE scan, but
3045 : : * it could still have required parameterization due to LATERAL refs in
3046 : : * its tlist.
3047 : : */
4810 3048 : 2124 : required_outer = rel->lateral_relids;
3049 : :
3050 : : /* Generate appropriate path */
580 3051 : 2124 : add_path(rel, create_ctescan_path(root, rel, pathkeys, required_outer));
6232 3052 : 2124 : }
3053 : :
3054 : : /*
3055 : : * set_namedtuplestore_pathlist
3056 : : * Build the (single) access path for a named tuplestore RTE
3057 : : *
3058 : : * There's no need for a separate set_namedtuplestore_size phase, since we
3059 : : * don't support join-qual-parameterized paths for tuplestores.
3060 : : */
3061 : : static void
3132 kgrittn@postgresql.o 3062 : 243 : set_namedtuplestore_pathlist(PlannerInfo *root, RelOptInfo *rel,
3063 : : RangeTblEntry *rte)
3064 : : {
3065 : : Relids required_outer;
3066 : :
3067 : : /* Mark rel with estimated output rows, width, etc */
3068 : 243 : set_namedtuplestore_size_estimates(root, rel);
3069 : :
3070 : : /*
3071 : : * We don't support pushing join clauses into the quals of a tuplestore
3072 : : * scan, but it could still have required parameterization due to LATERAL
3073 : : * refs in its tlist.
3074 : : */
3075 : 243 : required_outer = rel->lateral_relids;
3076 : :
3077 : : /* Generate appropriate path */
3078 : 243 : add_path(rel, create_namedtuplestorescan_path(root, rel, required_outer));
3079 : 243 : }
3080 : :
3081 : : /*
3082 : : * set_result_pathlist
3083 : : * Build the (single) access path for an RTE_RESULT RTE
3084 : : *
3085 : : * There's no need for a separate set_result_size phase, since we
3086 : : * don't support join-qual-parameterized paths for these RTEs.
3087 : : */
3088 : : static void
2464 tgl@sss.pgh.pa.us 3089 : 2098 : set_result_pathlist(PlannerInfo *root, RelOptInfo *rel,
3090 : : RangeTblEntry *rte)
3091 : : {
3092 : : Relids required_outer;
3093 : :
3094 : : /* Mark rel with estimated output rows, width, etc */
3095 : 2098 : set_result_size_estimates(root, rel);
3096 : :
3097 : : /*
3098 : : * We don't support pushing join clauses into the quals of a Result scan,
3099 : : * but it could still have required parameterization due to LATERAL refs
3100 : : * in its tlist.
3101 : : */
3102 : 2098 : required_outer = rel->lateral_relids;
3103 : :
3104 : : /* Generate appropriate path */
3105 : 2098 : add_path(rel, create_resultscan_path(root, rel, required_outer));
3106 : 2098 : }
3107 : :
3108 : : /*
3109 : : * set_worktable_pathlist
3110 : : * Build the (single) access path for a self-reference CTE RTE
3111 : : *
3112 : : * There's no need for a separate set_worktable_size phase, since we don't
3113 : : * support join-qual-parameterized paths for CTEs.
3114 : : */
3115 : : static void
6232 3116 : 470 : set_worktable_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
3117 : : {
3118 : : Path *ctepath;
3119 : : PlannerInfo *cteroot;
3120 : : Index levelsup;
3121 : : Relids required_outer;
3122 : :
3123 : : /*
3124 : : * We need to find the non-recursive term's path, which is in the plan
3125 : : * level that's processing the recursive UNION, which is one level *below*
3126 : : * where the CTE comes from.
3127 : : */
3128 : 470 : levelsup = rte->ctelevelsup;
3129 [ - + ]: 470 : if (levelsup == 0) /* shouldn't happen */
6232 tgl@sss.pgh.pa.us 3130 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6232 tgl@sss.pgh.pa.us 3131 :CBC 470 : levelsup--;
3132 : 470 : cteroot = root;
3133 [ + + ]: 1146 : while (levelsup-- > 0)
3134 : : {
3135 : 676 : cteroot = cteroot->parent_root;
3136 [ - + ]: 676 : if (!cteroot) /* shouldn't happen */
6232 tgl@sss.pgh.pa.us 3137 [ # # ]:UBC 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
3138 : : }
3521 tgl@sss.pgh.pa.us 3139 :CBC 470 : ctepath = cteroot->non_recursive_path;
3140 [ - + ]: 470 : if (!ctepath) /* shouldn't happen */
3521 tgl@sss.pgh.pa.us 3141 [ # # ]:UBC 0 : elog(ERROR, "could not find path for CTE \"%s\"", rte->ctename);
3142 : :
3143 : : /* Mark rel with estimated output rows, width, etc */
3521 tgl@sss.pgh.pa.us 3144 :CBC 470 : set_cte_size_estimates(root, rel, ctepath->rows);
3145 : :
3146 : : /*
3147 : : * We don't support pushing join clauses into the quals of a worktable
3148 : : * scan, but it could still have required parameterization due to LATERAL
3149 : : * refs in its tlist. (I'm not sure this is actually possible given the
3150 : : * restrictions on recursive references, but it's easy enough to support.)
3151 : : */
4810 3152 : 470 : required_outer = rel->lateral_relids;
3153 : :
3154 : : /* Generate appropriate path */
3155 : 470 : add_path(rel, create_worktablescan_path(root, rel, required_outer));
6232 3156 : 470 : }
3157 : :
3158 : : /*
3159 : : * generate_gather_paths
3160 : : * Generate parallel access paths for a relation by pushing a Gather or
3161 : : * Gather Merge on top of a partial path.
3162 : : *
3163 : : * This must not be called until after we're done creating all partial paths
3164 : : * for the specified relation. (Otherwise, add_partial_path might delete a
3165 : : * path that some GatherPath or GatherMergePath has a reference to.)
3166 : : *
3167 : : * If we're generating paths for a scan or join relation, override_rows will
3168 : : * be false, and we'll just use the relation's size estimate. When we're
3169 : : * being called for a partially-grouped or partially-distinct path, though, we
3170 : : * need to override the rowcount estimate. (It's not clear that the
3171 : : * particular value we're using here is actually best, but the underlying rel
3172 : : * has no estimate so we must do something.)
3173 : : */
3174 : : void
2800 rhaas@postgresql.org 3175 : 12372 : generate_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3176 : : {
3177 : : Path *cheapest_partial_path;
3178 : : Path *simple_gather_path;
3179 : : ListCell *lc;
3180 : : double rows;
3181 : 12372 : double *rowsp = NULL;
3182 : :
3183 : : /* If there are no partial paths, there's nothing to do here. */
3568 3184 [ - + ]: 12372 : if (rel->partial_pathlist == NIL)
3568 rhaas@postgresql.org 3185 :UBC 0 : return;
3186 : :
3187 : : /* Should we override the rel's rowcount estimate? */
2800 rhaas@postgresql.org 3188 [ + + ]:CBC 12372 : if (override_rows)
3189 : 3037 : rowsp = &rows;
3190 : :
3191 : : /*
3192 : : * The output of Gather is always unsorted, so there's only one partial
3193 : : * path of interest: the cheapest one. That will be the one at the front
3194 : : * of partial_pathlist because of the way add_partial_path works.
3195 : : */
3568 3196 : 12372 : cheapest_partial_path = linitial(rel->partial_pathlist);
461 rguo@postgresql.org 3197 : 12372 : rows = compute_gather_rows(cheapest_partial_path);
3198 : : simple_gather_path = (Path *)
3507 rhaas@postgresql.org 3199 : 12372 : create_gather_path(root, rel, cheapest_partial_path, rel->reltarget,
3200 : : NULL, rowsp);
3568 3201 : 12372 : add_path(rel, simple_gather_path);
3202 : :
3203 : : /*
3204 : : * For each useful ordering, we can consider an order-preserving Gather
3205 : : * Merge.
3206 : : */
3085 bruce@momjian.us 3207 [ + - + + : 27405 : foreach(lc, rel->partial_pathlist)
+ + ]
3208 : : {
3209 : 15033 : Path *subpath = (Path *) lfirst(lc);
3210 : : GatherMergePath *path;
3211 : :
3154 rhaas@postgresql.org 3212 [ + + ]: 15033 : if (subpath->pathkeys == NIL)
3213 : 12018 : continue;
3214 : :
461 rguo@postgresql.org 3215 : 3015 : rows = compute_gather_rows(subpath);
3154 rhaas@postgresql.org 3216 : 3015 : path = create_gather_merge_path(root, rel, subpath, rel->reltarget,
3217 : : subpath->pathkeys, NULL, rowsp);
3218 : 3015 : add_path(rel, &path->path);
3219 : : }
3220 : : }
3221 : :
3222 : : /*
3223 : : * get_useful_pathkeys_for_relation
3224 : : * Determine which orderings of a relation might be useful.
3225 : : *
3226 : : * Getting data in sorted order can be useful either because the requested
3227 : : * order matches the final output ordering for the overall query we're
3228 : : * planning, or because it enables an efficient merge join. Here, we try
3229 : : * to figure out which pathkeys to consider.
3230 : : *
3231 : : * This allows us to do incremental sort on top of an index scan under a gather
3232 : : * merge node, i.e. parallelized.
3233 : : *
3234 : : * If the require_parallel_safe is true, we also require the expressions to
3235 : : * be parallel safe (which allows pushing the sort below Gather Merge).
3236 : : *
3237 : : * XXX At the moment this can only ever return a list with a single element,
3238 : : * because it looks at query_pathkeys only. So we might return the pathkeys
3239 : : * directly, but it seems plausible we'll want to consider other orderings
3240 : : * in the future. For example, we might want to consider pathkeys useful for
3241 : : * merge joins.
3242 : : */
3243 : : static List *
1771 tomas.vondra@postgre 3244 : 12372 : get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel,
3245 : : bool require_parallel_safe)
3246 : : {
2029 3247 : 12372 : List *useful_pathkeys_list = NIL;
3248 : :
3249 : : /*
3250 : : * Considering query_pathkeys is always worth it, because it might allow
3251 : : * us to avoid a total sort when we have a partially presorted path
3252 : : * available or to push the total sort into the parallel portion of the
3253 : : * query.
3254 : : */
3255 [ + + ]: 12372 : if (root->query_pathkeys)
3256 : : {
3257 : : ListCell *lc;
1992 tgl@sss.pgh.pa.us 3258 : 7462 : int npathkeys = 0; /* useful pathkeys */
3259 : :
2029 tomas.vondra@postgre 3260 [ + - + + : 13043 : foreach(lc, root->query_pathkeys)
+ + ]
3261 : : {
3262 : 9434 : PathKey *pathkey = (PathKey *) lfirst(lc);
3263 : 9434 : EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
3264 : :
3265 : : /*
3266 : : * We can only build a sort for pathkeys that contain a
3267 : : * safe-to-compute-early EC member computable from the current
3268 : : * relation's reltarget, so ignore the remainder of the list as
3269 : : * soon as we find a pathkey without such a member.
3270 : : *
3271 : : * It's still worthwhile to return any prefix of the pathkeys list
3272 : : * that meets this requirement, as we may be able to do an
3273 : : * incremental sort.
3274 : : *
3275 : : * If requested, ensure the sort expression is parallel-safe too.
3276 : : */
1651 tgl@sss.pgh.pa.us 3277 [ + + ]: 9434 : if (!relation_can_be_sorted_early(root, rel, pathkey_ec,
3278 : : require_parallel_safe))
2029 tomas.vondra@postgre 3279 : 3853 : break;
3280 : :
3281 : 5581 : npathkeys++;
3282 : : }
3283 : :
3284 : : /*
3285 : : * The whole query_pathkeys list matches, so append it directly, to
3286 : : * allow comparing pathkeys easily by comparing list pointer. If we
3287 : : * have to truncate the pathkeys, we gotta do a copy though.
3288 : : */
3289 [ + + ]: 7462 : if (npathkeys == list_length(root->query_pathkeys))
3290 : 3609 : useful_pathkeys_list = lappend(useful_pathkeys_list,
3291 : 3609 : root->query_pathkeys);
3292 [ + + ]: 3853 : else if (npathkeys > 0)
3293 : 237 : useful_pathkeys_list = lappend(useful_pathkeys_list,
1202 drowley@postgresql.o 3294 : 237 : list_copy_head(root->query_pathkeys,
3295 : : npathkeys));
3296 : : }
3297 : :
2029 tomas.vondra@postgre 3298 : 12372 : return useful_pathkeys_list;
3299 : : }
3300 : :
3301 : : /*
3302 : : * generate_useful_gather_paths
3303 : : * Generate parallel access paths for a relation by pushing a Gather or
3304 : : * Gather Merge on top of a partial path.
3305 : : *
3306 : : * Unlike plain generate_gather_paths, this looks both at pathkeys of input
3307 : : * paths (aiming to preserve the ordering), but also considers ordering that
3308 : : * might be useful for nodes above the gather merge node, and tries to add
3309 : : * a sort (regular or incremental) to provide that.
3310 : : */
3311 : : void
3312 : 307088 : generate_useful_gather_paths(PlannerInfo *root, RelOptInfo *rel, bool override_rows)
3313 : : {
3314 : : ListCell *lc;
3315 : : double rows;
3316 : 307088 : double *rowsp = NULL;
3317 : 307088 : List *useful_pathkeys_list = NIL;
3318 : 307088 : Path *cheapest_partial_path = NULL;
3319 : :
3320 : : /* If there are no partial paths, there's nothing to do here. */
3321 [ + + ]: 307088 : if (rel->partial_pathlist == NIL)
3322 : 294716 : return;
3323 : :
3324 : : /* Should we override the rel's rowcount estimate? */
3325 [ + + ]: 12372 : if (override_rows)
3326 : 3037 : rowsp = &rows;
3327 : :
3328 : : /* generate the regular gather (merge) paths */
3329 : 12372 : generate_gather_paths(root, rel, override_rows);
3330 : :
3331 : : /* consider incremental sort for interesting orderings */
1771 3332 : 12372 : useful_pathkeys_list = get_useful_pathkeys_for_relation(root, rel, true);
3333 : :
3334 : : /* used for explicit (full) sort paths */
2029 3335 : 12372 : cheapest_partial_path = linitial(rel->partial_pathlist);
3336 : :
3337 : : /*
3338 : : * Consider sorted paths for each interesting ordering. We generate both
3339 : : * incremental and full sort.
3340 : : */
3341 [ + + + + : 16218 : foreach(lc, useful_pathkeys_list)
+ + ]
3342 : : {
3343 : 3846 : List *useful_pathkeys = lfirst(lc);
3344 : : ListCell *lc2;
3345 : : bool is_sorted;
3346 : : int presorted_keys;
3347 : :
3348 [ + - + + : 8758 : foreach(lc2, rel->partial_pathlist)
+ + ]
3349 : : {
3350 : 4912 : Path *subpath = (Path *) lfirst(lc2);
3351 : : GatherMergePath *path;
3352 : :
3353 : 4912 : is_sorted = pathkeys_count_contained_in(useful_pathkeys,
3354 : : subpath->pathkeys,
3355 : : &presorted_keys);
3356 : :
3357 : : /*
3358 : : * We don't need to consider the case where a subpath is already
3359 : : * fully sorted because generate_gather_paths already creates a
3360 : : * gather merge path for every subpath that has pathkeys present.
3361 : : *
3362 : : * But since the subpath is already sorted, we know we don't need
3363 : : * to consider adding a sort (full or incremental) on top of it,
3364 : : * so we can continue here.
3365 : : */
3366 [ + + ]: 4912 : if (is_sorted)
3367 : 1199 : continue;
3368 : :
3369 : : /*
3370 : : * Try at least sorting the cheapest path and also try
3371 : : * incrementally sorting any path which is partially sorted
3372 : : * already (no need to deal with paths which have presorted keys
3373 : : * when incremental sort is disabled unless it's the cheapest
3374 : : * input path).
3375 : : */
1046 drowley@postgresql.o 3376 [ + + ]: 3713 : if (subpath != cheapest_partial_path &&
3377 [ + + + + ]: 123 : (presorted_keys == 0 || !enable_incremental_sort))
3378 : 51 : continue;
3379 : :
3380 : : /*
3381 : : * Consider regular sort for any path that's not presorted or if
3382 : : * incremental sort is disabled. We've no need to consider both
3383 : : * sort and incremental sort on the same path. We assume that
3384 : : * incremental sort is always faster when there are presorted
3385 : : * keys.
3386 : : *
3387 : : * This is not redundant with the gather paths created in
3388 : : * generate_gather_paths, because that doesn't generate ordered
3389 : : * output. Here we add an explicit sort to match the useful
3390 : : * ordering.
3391 : : */
3392 [ + + + + ]: 3662 : if (presorted_keys == 0 || !enable_incremental_sort)
3393 : : {
3394 : 3584 : subpath = (Path *) create_sort_path(root,
3395 : : rel,
3396 : : subpath,
3397 : : useful_pathkeys,
3398 : : -1.0);
3399 : : }
3400 : : else
3401 : 78 : subpath = (Path *) create_incremental_sort_path(root,
3402 : : rel,
3403 : : subpath,
3404 : : useful_pathkeys,
3405 : : presorted_keys,
3406 : : -1);
461 rguo@postgresql.org 3407 : 3662 : rows = compute_gather_rows(subpath);
1046 drowley@postgresql.o 3408 : 3662 : path = create_gather_merge_path(root, rel,
3409 : : subpath,
3410 : 3662 : rel->reltarget,
3411 : : subpath->pathkeys,
3412 : : NULL,
3413 : : rowsp);
3414 : :
3415 : 3662 : add_path(rel, &path->path);
3416 : : }
3417 : : }
3418 : : }
3419 : :
3420 : : /*
3421 : : * generate_grouped_paths
3422 : : * Generate paths for a grouped relation by adding sorted and hashed
3423 : : * partial aggregation paths on top of paths of the ungrouped relation.
3424 : : *
3425 : : * The information needed is provided by the RelAggInfo structure stored in
3426 : : * "grouped_rel".
3427 : : */
3428 : : void
19 rguo@postgresql.org 3429 :GNC 449 : generate_grouped_paths(PlannerInfo *root, RelOptInfo *grouped_rel,
3430 : : RelOptInfo *rel)
3431 : : {
3432 : 449 : RelAggInfo *agg_info = grouped_rel->agg_info;
3433 : : AggClauseCosts agg_costs;
3434 : : bool can_hash;
3435 : : bool can_sort;
3436 : 449 : Path *cheapest_total_path = NULL;
3437 : 449 : Path *cheapest_partial_path = NULL;
3438 : 449 : double dNumGroups = 0;
3439 : 449 : double dNumPartialGroups = 0;
3440 : 449 : List *group_pathkeys = NIL;
3441 : :
3442 [ - + ]: 449 : if (IS_DUMMY_REL(rel))
3443 : : {
19 rguo@postgresql.org 3444 :UNC 0 : mark_dummy_rel(grouped_rel);
3445 : 0 : return;
3446 : : }
3447 : :
3448 : : /*
3449 : : * We push partial aggregation only to the lowest possible level in the
3450 : : * join tree that is deemed useful.
3451 : : */
13 rguo@postgresql.org 3452 [ + - ]:GNC 449 : if (!bms_equal(agg_info->apply_agg_at, rel->relids) ||
19 3453 [ - + ]: 449 : !agg_info->agg_useful)
19 rguo@postgresql.org 3454 :UNC 0 : return;
3455 : :
19 rguo@postgresql.org 3456 [ + - + - :GNC 2694 : MemSet(&agg_costs, 0, sizeof(AggClauseCosts));
+ - + - +
+ ]
3457 : 449 : get_agg_clause_costs(root, AGGSPLIT_INITIAL_SERIAL, &agg_costs);
3458 : :
3459 : : /*
3460 : : * Determine whether it's possible to perform sort-based implementations
3461 : : * of grouping, and generate the pathkeys that represent the grouping
3462 : : * requirements in that case.
3463 : : */
3464 : 449 : can_sort = grouping_is_sortable(agg_info->group_clauses);
3465 [ + - ]: 449 : if (can_sort)
3466 : : {
3467 : : RelOptInfo *top_grouped_rel;
3468 : : List *top_group_tlist;
3469 : :
3470 [ + + - + ]: 251 : top_grouped_rel = IS_OTHER_REL(rel) ?
3471 [ + + ]: 700 : rel->top_parent->grouped_rel : grouped_rel;
3472 : : top_group_tlist =
3473 : 449 : make_tlist_from_pathtarget(top_grouped_rel->agg_info->target);
3474 : :
3475 : : group_pathkeys =
3476 : 449 : make_pathkeys_for_sortclauses(root, agg_info->group_clauses,
3477 : : top_group_tlist);
3478 : : }
3479 : :
3480 : : /*
3481 : : * Determine whether we should consider hash-based implementations of
3482 : : * grouping.
3483 : : */
3484 [ - + ]: 449 : Assert(root->numOrderedAggs == 0);
3485 [ + - + - ]: 898 : can_hash = (agg_info->group_clauses != NIL &&
3486 : 449 : grouping_is_hashable(agg_info->group_clauses));
3487 : :
3488 : : /*
3489 : : * Consider whether we should generate partially aggregated non-partial
3490 : : * paths. We can only do this if we have a non-partial path.
3491 : : */
3492 [ + - ]: 449 : if (rel->pathlist != NIL)
3493 : : {
3494 : 449 : cheapest_total_path = rel->cheapest_total_path;
3495 [ - + ]: 449 : Assert(cheapest_total_path != NULL);
3496 : : }
3497 : :
3498 : : /*
3499 : : * If parallelism is possible for grouped_rel, then we should consider
3500 : : * generating partially-grouped partial paths. However, if the ungrouped
3501 : : * rel has no partial paths, then we can't.
3502 : : */
3503 [ + + + + ]: 449 : if (grouped_rel->consider_parallel && rel->partial_pathlist != NIL)
3504 : : {
3505 : 366 : cheapest_partial_path = linitial(rel->partial_pathlist);
3506 [ - + ]: 366 : Assert(cheapest_partial_path != NULL);
3507 : : }
3508 : :
3509 : : /* Estimate number of partial groups. */
3510 [ + - ]: 449 : if (cheapest_total_path != NULL)
3511 : 449 : dNumGroups = estimate_num_groups(root,
3512 : : agg_info->group_exprs,
3513 : : cheapest_total_path->rows,
3514 : : NULL, NULL);
3515 [ + + ]: 449 : if (cheapest_partial_path != NULL)
3516 : 366 : dNumPartialGroups = estimate_num_groups(root,
3517 : : agg_info->group_exprs,
3518 : : cheapest_partial_path->rows,
3519 : : NULL, NULL);
3520 : :
3521 [ + - + - ]: 449 : if (can_sort && cheapest_total_path != NULL)
3522 : : {
3523 : : ListCell *lc;
3524 : :
3525 : : /*
3526 : : * Use any available suitably-sorted path as input, and also consider
3527 : : * sorting the cheapest-total path and incremental sort on any paths
3528 : : * with presorted keys.
3529 : : *
3530 : : * To save planning time, we ignore parameterized input paths unless
3531 : : * they are the cheapest-total path.
3532 : : */
3533 [ + - + + : 1087 : foreach(lc, rel->pathlist)
+ + ]
3534 : : {
3535 : 638 : Path *input_path = (Path *) lfirst(lc);
3536 : : Path *path;
3537 : : bool is_sorted;
3538 : : int presorted_keys;
3539 : :
3540 : : /*
3541 : : * Ignore parameterized paths that are not the cheapest-total
3542 : : * path.
3543 : : */
3544 [ + + + - ]: 638 : if (input_path->param_info &&
3545 : : input_path != cheapest_total_path)
3546 : 15 : continue;
3547 : :
3548 : 635 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3549 : : input_path->pathkeys,
3550 : : &presorted_keys);
3551 : :
3552 : : /*
3553 : : * Ignore paths that are not suitably or partially sorted, unless
3554 : : * they are the cheapest total path (no need to deal with paths
3555 : : * which have presorted keys when incremental sort is disabled).
3556 : : */
3557 [ + + + + ]: 635 : if (!is_sorted && input_path != cheapest_total_path &&
3558 [ + + - + ]: 84 : (presorted_keys == 0 || !enable_incremental_sort))
3559 : 12 : continue;
3560 : :
3561 : : /*
3562 : : * Since the path originates from a non-grouped relation that is
3563 : : * not aware of eager aggregation, we must ensure that it provides
3564 : : * the correct input for partial aggregation.
3565 : : */
3566 : 623 : path = (Path *) create_projection_path(root,
3567 : : grouped_rel,
3568 : : input_path,
3569 : 623 : agg_info->agg_input);
3570 : :
3571 [ + + ]: 623 : if (!is_sorted)
3572 : : {
3573 : : /*
3574 : : * We've no need to consider both a sort and incremental sort.
3575 : : * We'll just do a sort if there are no presorted keys and an
3576 : : * incremental sort when there are presorted keys.
3577 : : */
3578 [ + + - + ]: 518 : if (presorted_keys == 0 || !enable_incremental_sort)
3579 : 446 : path = (Path *) create_sort_path(root,
3580 : : grouped_rel,
3581 : : path,
3582 : : group_pathkeys,
3583 : : -1.0);
3584 : : else
3585 : 72 : path = (Path *) create_incremental_sort_path(root,
3586 : : grouped_rel,
3587 : : path,
3588 : : group_pathkeys,
3589 : : presorted_keys,
3590 : : -1.0);
3591 : : }
3592 : :
3593 : : /*
3594 : : * qual is NIL because the HAVING clause cannot be evaluated until
3595 : : * the final value of the aggregate is known.
3596 : : */
3597 : 623 : path = (Path *) create_agg_path(root,
3598 : : grouped_rel,
3599 : : path,
3600 : 623 : agg_info->target,
3601 : : AGG_SORTED,
3602 : : AGGSPLIT_INITIAL_SERIAL,
3603 : : agg_info->group_clauses,
3604 : : NIL,
3605 : : &agg_costs,
3606 : : dNumGroups);
3607 : :
3608 : 623 : add_path(grouped_rel, path);
3609 : : }
3610 : : }
3611 : :
3612 [ + - + + ]: 449 : if (can_sort && cheapest_partial_path != NULL)
3613 : : {
3614 : : ListCell *lc;
3615 : :
3616 : : /* Similar to above logic, but for partial paths. */
3617 [ + - + + : 852 : foreach(lc, rel->partial_pathlist)
+ + ]
3618 : : {
3619 : 486 : Path *input_path = (Path *) lfirst(lc);
3620 : : Path *path;
3621 : : bool is_sorted;
3622 : : int presorted_keys;
3623 : :
3624 : 486 : is_sorted = pathkeys_count_contained_in(group_pathkeys,
3625 : : input_path->pathkeys,
3626 : : &presorted_keys);
3627 : :
3628 : : /*
3629 : : * Ignore paths that are not suitably or partially sorted, unless
3630 : : * they are the cheapest partial path (no need to deal with paths
3631 : : * which have presorted keys when incremental sort is disabled).
3632 : : */
3633 [ + + + + ]: 486 : if (!is_sorted && input_path != cheapest_partial_path &&
3634 [ + - - + ]: 48 : (presorted_keys == 0 || !enable_incremental_sort))
19 rguo@postgresql.org 3635 :UNC 0 : continue;
3636 : :
3637 : : /*
3638 : : * Since the path originates from a non-grouped relation that is
3639 : : * not aware of eager aggregation, we must ensure that it provides
3640 : : * the correct input for partial aggregation.
3641 : : */
19 rguo@postgresql.org 3642 :GNC 486 : path = (Path *) create_projection_path(root,
3643 : : grouped_rel,
3644 : : input_path,
3645 : 486 : agg_info->agg_input);
3646 : :
3647 [ + + ]: 486 : if (!is_sorted)
3648 : : {
3649 : : /*
3650 : : * We've no need to consider both a sort and incremental sort.
3651 : : * We'll just do a sort if there are no presorted keys and an
3652 : : * incremental sort when there are presorted keys.
3653 : : */
3654 [ + + - + ]: 414 : if (presorted_keys == 0 || !enable_incremental_sort)
3655 : 366 : path = (Path *) create_sort_path(root,
3656 : : grouped_rel,
3657 : : path,
3658 : : group_pathkeys,
3659 : : -1.0);
3660 : : else
3661 : 48 : path = (Path *) create_incremental_sort_path(root,
3662 : : grouped_rel,
3663 : : path,
3664 : : group_pathkeys,
3665 : : presorted_keys,
3666 : : -1.0);
3667 : : }
3668 : :
3669 : : /*
3670 : : * qual is NIL because the HAVING clause cannot be evaluated until
3671 : : * the final value of the aggregate is known.
3672 : : */
3673 : 486 : path = (Path *) create_agg_path(root,
3674 : : grouped_rel,
3675 : : path,
3676 : 486 : agg_info->target,
3677 : : AGG_SORTED,
3678 : : AGGSPLIT_INITIAL_SERIAL,
3679 : : agg_info->group_clauses,
3680 : : NIL,
3681 : : &agg_costs,
3682 : : dNumPartialGroups);
3683 : :
3684 : 486 : add_partial_path(grouped_rel, path);
3685 : : }
3686 : : }
3687 : :
3688 : : /*
3689 : : * Add a partially-grouped HashAgg Path where possible
3690 : : */
3691 [ + - + - ]: 449 : if (can_hash && cheapest_total_path != NULL)
3692 : : {
3693 : : Path *path;
3694 : :
3695 : : /*
3696 : : * Since the path originates from a non-grouped relation that is not
3697 : : * aware of eager aggregation, we must ensure that it provides the
3698 : : * correct input for partial aggregation.
3699 : : */
3700 : 449 : path = (Path *) create_projection_path(root,
3701 : : grouped_rel,
3702 : : cheapest_total_path,
3703 : 449 : agg_info->agg_input);
3704 : :
3705 : : /*
3706 : : * qual is NIL because the HAVING clause cannot be evaluated until the
3707 : : * final value of the aggregate is known.
3708 : : */
3709 : 449 : path = (Path *) create_agg_path(root,
3710 : : grouped_rel,
3711 : : path,
3712 : 449 : agg_info->target,
3713 : : AGG_HASHED,
3714 : : AGGSPLIT_INITIAL_SERIAL,
3715 : : agg_info->group_clauses,
3716 : : NIL,
3717 : : &agg_costs,
3718 : : dNumGroups);
3719 : :
3720 : 449 : add_path(grouped_rel, path);
3721 : : }
3722 : :
3723 : : /*
3724 : : * Now add a partially-grouped HashAgg partial Path where possible
3725 : : */
3726 [ + - + + ]: 449 : if (can_hash && cheapest_partial_path != NULL)
3727 : : {
3728 : : Path *path;
3729 : :
3730 : : /*
3731 : : * Since the path originates from a non-grouped relation that is not
3732 : : * aware of eager aggregation, we must ensure that it provides the
3733 : : * correct input for partial aggregation.
3734 : : */
3735 : 366 : path = (Path *) create_projection_path(root,
3736 : : grouped_rel,
3737 : : cheapest_partial_path,
3738 : 366 : agg_info->agg_input);
3739 : :
3740 : : /*
3741 : : * qual is NIL because the HAVING clause cannot be evaluated until the
3742 : : * final value of the aggregate is known.
3743 : : */
3744 : 366 : path = (Path *) create_agg_path(root,
3745 : : grouped_rel,
3746 : : path,
3747 : 366 : agg_info->target,
3748 : : AGG_HASHED,
3749 : : AGGSPLIT_INITIAL_SERIAL,
3750 : : agg_info->group_clauses,
3751 : : NIL,
3752 : : &agg_costs,
3753 : : dNumPartialGroups);
3754 : :
3755 : 366 : add_partial_path(grouped_rel, path);
3756 : : }
3757 : : }
3758 : :
3759 : : /*
3760 : : * make_rel_from_joinlist
3761 : : * Build access paths using a "joinlist" to guide the join path search.
3762 : : *
3763 : : * See comments for deconstruct_jointree() for definition of the joinlist
3764 : : * data structure.
3765 : : */
3766 : : static RelOptInfo *
7251 tgl@sss.pgh.pa.us 3767 :CBC 164621 : make_rel_from_joinlist(PlannerInfo *root, List *joinlist)
3768 : : {
3769 : : int levels_needed;
3770 : : List *initial_rels;
3771 : : ListCell *jl;
3772 : :
3773 : : /*
3774 : : * Count the number of child joinlist nodes. This is the depth of the
3775 : : * dynamic-programming algorithm we must employ to consider all ways of
3776 : : * joining the child nodes.
3777 : : */
3778 : 164621 : levels_needed = list_length(joinlist);
3779 : :
9159 3780 [ - + ]: 164621 : if (levels_needed <= 0)
9159 tgl@sss.pgh.pa.us 3781 :UBC 0 : return NULL; /* nothing to do? */
3782 : :
3783 : : /*
3784 : : * Construct a list of rels corresponding to the child joinlist nodes.
3785 : : * This may contain both base rels and rels constructed according to
3786 : : * sub-joinlists.
3787 : : */
7251 tgl@sss.pgh.pa.us 3788 :CBC 164621 : initial_rels = NIL;
3789 [ + - + + : 398367 : foreach(jl, joinlist)
+ + ]
3790 : : {
3791 : 233746 : Node *jlnode = (Node *) lfirst(jl);
3792 : : RelOptInfo *thisrel;
3793 : :
3794 [ + + ]: 233746 : if (IsA(jlnode, RangeTblRef))
3795 : : {
3796 : 232057 : int varno = ((RangeTblRef *) jlnode)->rtindex;
3797 : :
3798 : 232057 : thisrel = find_base_rel(root, varno);
3799 : : }
3800 [ + - ]: 1689 : else if (IsA(jlnode, List))
3801 : : {
3802 : : /* Recurse to handle subproblem */
3803 : 1689 : thisrel = make_rel_from_joinlist(root, (List *) jlnode);
3804 : : }
3805 : : else
3806 : : {
7251 tgl@sss.pgh.pa.us 3807 [ # # ]:UBC 0 : elog(ERROR, "unrecognized joinlist node type: %d",
3808 : : (int) nodeTag(jlnode));
3809 : : thisrel = NULL; /* keep compiler quiet */
3810 : : }
3811 : :
7251 tgl@sss.pgh.pa.us 3812 :CBC 233746 : initial_rels = lappend(initial_rels, thisrel);
3813 : : }
3814 : :
9159 3815 [ + + ]: 164621 : if (levels_needed == 1)
3816 : : {
3817 : : /*
3818 : : * Single joinlist node, so we're done.
3819 : : */
7824 neilc@samurai.com 3820 : 114263 : return (RelOptInfo *) linitial(initial_rels);
3821 : : }
3822 : : else
3823 : : {
3824 : : /*
3825 : : * Consider the different orders in which we could join the rels,
3826 : : * using a plugin, GEQO, or the regular join search code.
3827 : : *
3828 : : * We put the initial_rels list into a PlannerInfo field because
3829 : : * has_legal_joinclause() needs to look at it (ugly :-().
3830 : : */
6499 tgl@sss.pgh.pa.us 3831 : 50358 : root->initial_rels = initial_rels;
3832 : :
6606 3833 [ - + ]: 50358 : if (join_search_hook)
6606 tgl@sss.pgh.pa.us 3834 :UBC 0 : return (*join_search_hook) (root, levels_needed, initial_rels);
6606 tgl@sss.pgh.pa.us 3835 [ + - + + ]:CBC 50358 : else if (enable_geqo && levels_needed >= geqo_threshold)
9159 3836 : 21 : return geqo(root, levels_needed, initial_rels);
3837 : : else
6606 3838 : 50337 : return standard_join_search(root, levels_needed, initial_rels);
3839 : : }
3840 : : }
3841 : :
3842 : : /*
3843 : : * standard_join_search
3844 : : * Find possible joinpaths for a query by successively finding ways
3845 : : * to join component relations into join relations.
3846 : : *
3847 : : * 'levels_needed' is the number of iterations needed, ie, the number of
3848 : : * independent jointree items in the query. This is > 1.
3849 : : *
3850 : : * 'initial_rels' is a list of RelOptInfo nodes for each independent
3851 : : * jointree item. These are the components to be joined together.
3852 : : * Note that levels_needed == list_length(initial_rels).
3853 : : *
3854 : : * Returns the final level of join relations, i.e., the relation that is
3855 : : * the result of joining all the original relations together.
3856 : : * At least one implementation path must be provided for this relation and
3857 : : * all required sub-relations.
3858 : : *
3859 : : * To support loadable plugins that modify planner behavior by changing the
3860 : : * join searching algorithm, we provide a hook variable that lets a plugin
3861 : : * replace or supplement this function. Any such hook must return the same
3862 : : * final join relation as the standard code would, but it might have a
3863 : : * different set of implementation paths attached, and only the sub-joinrels
3864 : : * needed for these paths need have been instantiated.
3865 : : *
3866 : : * Note to plugin authors: the functions invoked during standard_join_search()
3867 : : * modify root->join_rel_list and root->join_rel_hash. If you want to do more
3868 : : * than one join-order search, you'll probably need to save and restore the
3869 : : * original states of those data structures. See geqo_eval() for an example.
3870 : : */
3871 : : RelOptInfo *
3872 : 50337 : standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
3873 : : {
3874 : : int lev;
3875 : : RelOptInfo *rel;
3876 : :
3877 : : /*
3878 : : * This function cannot be invoked recursively within any one planning
3879 : : * problem, so join_rel_level[] can't be in use already.
3880 : : */
5812 3881 [ - + ]: 50337 : Assert(root->join_rel_level == NULL);
3882 : :
3883 : : /*
3884 : : * We employ a simple "dynamic programming" algorithm: we first find all
3885 : : * ways to build joins of two jointree items, then all ways to build joins
3886 : : * of three items (from two-item joins and single items), then four-item
3887 : : * joins, and so on until we have considered all ways to join all the
3888 : : * items into one rel.
3889 : : *
3890 : : * root->join_rel_level[j] is a list of all the j-item rels. Initially we
3891 : : * set root->join_rel_level[1] to represent all the single-jointree-item
3892 : : * relations.
3893 : : */
3894 : 50337 : root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
3895 : :
3896 : 50337 : root->join_rel_level[1] = initial_rels;
3897 : :
9394 3898 [ + + ]: 119432 : for (lev = 2; lev <= levels_needed; lev++)
3899 : : {
3900 : : ListCell *lc;
3901 : :
3902 : : /*
3903 : : * Determine all possible pairs of relations to be joined at this
3904 : : * level, and build paths for making each one from every available
3905 : : * pair of lower-level relations.
3906 : : */
5812 3907 : 69095 : join_search_one_level(root, lev);
3908 : :
3909 : : /*
3910 : : * Run generate_partitionwise_join_paths() and
3911 : : * generate_useful_gather_paths() for each just-processed joinrel. We
3912 : : * could not do this earlier because both regular and partial paths
3913 : : * can get added to a particular joinrel at multiple times within
3914 : : * join_search_one_level.
3915 : : *
3916 : : * After that, we're done creating paths for the joinrel, so run
3917 : : * set_cheapest().
3918 : : *
3919 : : * In addition, we also run generate_grouped_paths() for the grouped
3920 : : * relation of each just-processed joinrel, and run set_cheapest() for
3921 : : * the grouped relation afterwards.
3922 : : */
3923 [ + + + + : 175074 : foreach(lc, root->join_rel_level[lev])
+ + ]
3924 : : {
3925 : : bool is_top_rel;
3926 : :
3927 : 105979 : rel = (RelOptInfo *) lfirst(lc);
3928 : :
19 rguo@postgresql.org 3929 :GNC 105979 : is_top_rel = bms_equal(rel->relids, root->all_query_rels);
3930 : :
3931 : : /* Create paths for partitionwise joins. */
2810 peter_e@gmx.net 3932 :CBC 105979 : generate_partitionwise_join_paths(root, rel);
3933 : :
3934 : : /*
3935 : : * Except for the topmost scan/join rel, consider gathering
3936 : : * partial paths. We'll do the same for the topmost scan/join rel
3937 : : * once we know the final targetlist (see grouping_planner's and
3938 : : * its call to apply_scanjoin_target_to_paths).
3939 : : */
19 rguo@postgresql.org 3940 [ + + ]:GNC 105979 : if (!is_top_rel)
2029 tomas.vondra@postgre 3941 :CBC 55891 : generate_useful_gather_paths(root, rel, false);
3942 : :
3943 : : /* Find and save the cheapest paths for this rel */
9386 tgl@sss.pgh.pa.us 3944 : 105979 : set_cheapest(rel);
3945 : :
3946 : : /*
3947 : : * Except for the topmost scan/join rel, consider generating
3948 : : * partial aggregation paths for the grouped relation on top of
3949 : : * the paths of this rel. After that, we're done creating paths
3950 : : * for the grouped relation, so run set_cheapest().
3951 : : */
19 rguo@postgresql.org 3952 [ + + + + ]:GNC 105979 : if (rel->grouped_rel != NULL && !is_top_rel)
3953 : : {
3954 : 36 : RelOptInfo *grouped_rel = rel->grouped_rel;
3955 : :
3956 [ - + ]: 36 : Assert(IS_GROUPED_REL(grouped_rel));
3957 : :
3958 : 36 : generate_grouped_paths(root, grouped_rel, rel);
3959 : 36 : set_cheapest(grouped_rel);
3960 : : }
3961 : :
3962 : : #ifdef OPTIMIZER_DEBUG
3963 : : pprint(rel);
3964 : : #endif
3965 : : }
3966 : : }
3967 : :
3968 : : /*
3969 : : * We should have a single rel at the final level.
3970 : : */
5812 tgl@sss.pgh.pa.us 3971 [ - + ]:CBC 50337 : if (root->join_rel_level[levels_needed] == NIL)
7985 tgl@sss.pgh.pa.us 3972 [ # # ]:UBC 0 : elog(ERROR, "failed to build any %d-way joins", levels_needed);
5812 tgl@sss.pgh.pa.us 3973 [ - + ]:CBC 50337 : Assert(list_length(root->join_rel_level[levels_needed]) == 1);
3974 : :
3975 : 50337 : rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
3976 : :
3977 : 50337 : root->join_rel_level = NULL;
3978 : :
9394 3979 : 50337 : return rel;
3980 : : }
3981 : :
3982 : : /*****************************************************************************
3983 : : * PUSHING QUALS DOWN INTO SUBQUERIES
3984 : : *****************************************************************************/
3985 : :
3986 : : /*
3987 : : * subquery_is_pushdown_safe - is a subquery safe for pushing down quals?
3988 : : *
3989 : : * subquery is the particular component query being checked. topquery
3990 : : * is the top component of a set-operations tree (the same Query if no
3991 : : * set-op is involved).
3992 : : *
3993 : : * Conditions checked here:
3994 : : *
3995 : : * 1. If the subquery has a LIMIT clause, we must not push down any quals,
3996 : : * since that could change the set of rows returned.
3997 : : *
3998 : : * 2. If the subquery contains EXCEPT or EXCEPT ALL set ops we cannot push
3999 : : * quals into it, because that could change the results.
4000 : : *
4001 : : * 3. If the subquery uses DISTINCT, we cannot push volatile quals into it.
4002 : : * This is because upper-level quals should semantically be evaluated only
4003 : : * once per distinct row, not once per original row, and if the qual is
4004 : : * volatile then extra evaluations could change the results. (This issue
4005 : : * does not apply to other forms of aggregation such as GROUP BY, because
4006 : : * when those are present we push into HAVING not WHERE, so that the quals
4007 : : * are still applied after aggregation.)
4008 : : *
4009 : : * 4. If the subquery contains window functions, we cannot push volatile quals
4010 : : * into it. The issue here is a bit different from DISTINCT: a volatile qual
4011 : : * might succeed for some rows of a window partition and fail for others,
4012 : : * thereby changing the partition contents and thus the window functions'
4013 : : * results for rows that remain.
4014 : : *
4015 : : * 5. If the subquery contains any set-returning functions in its targetlist,
4016 : : * we cannot push volatile quals into it. That would push them below the SRFs
4017 : : * and thereby change the number of times they are evaluated. Also, a
4018 : : * volatile qual could succeed for some SRF output rows and fail for others,
4019 : : * a behavior that cannot occur if it's evaluated before SRF expansion.
4020 : : *
4021 : : * 6. If the subquery has nonempty grouping sets, we cannot push down any
4022 : : * quals. The concern here is that a qual referencing a "constant" grouping
4023 : : * column could get constant-folded, which would be improper because the value
4024 : : * is potentially nullable by grouping-set expansion. This restriction could
4025 : : * be removed if we had a parsetree representation that shows that such
4026 : : * grouping columns are not really constant. (There are other ideas that
4027 : : * could be used to relax this restriction, but that's the approach most
4028 : : * likely to get taken in the future. Note that there's not much to be gained
4029 : : * so long as subquery_planner can't move HAVING clauses to WHERE within such
4030 : : * a subquery.)
4031 : : *
4032 : : * In addition, we make several checks on the subquery's output columns to see
4033 : : * if it is safe to reference them in pushed-down quals. If output column k
4034 : : * is found to be unsafe to reference, we set the reason for that inside
4035 : : * safetyInfo->unsafeFlags[k], but we don't reject the subquery overall since
4036 : : * column k might not be referenced by some/all quals. The unsafeFlags[]
4037 : : * array will be consulted later by qual_is_pushdown_safe(). It's better to
4038 : : * do it this way than to make the checks directly in qual_is_pushdown_safe(),
4039 : : * because when the subquery involves set operations we have to check the
4040 : : * output expressions in each arm of the set op.
4041 : : *
4042 : : * Note: pushing quals into a DISTINCT subquery is theoretically dubious:
4043 : : * we're effectively assuming that the quals cannot distinguish values that
4044 : : * the DISTINCT's equality operator sees as equal, yet there are many
4045 : : * counterexamples to that assumption. However use of such a qual with a
4046 : : * DISTINCT subquery would be unsafe anyway, since there's no guarantee which
4047 : : * "equal" value will be chosen as the output value by the DISTINCT operation.
4048 : : * So we don't worry too much about that. Another objection is that if the
4049 : : * qual is expensive to evaluate, running it for each original row might cost
4050 : : * more than we save by eliminating rows before the DISTINCT step. But it
4051 : : * would be very hard to estimate that at this stage, and in practice pushdown
4052 : : * seldom seems to make things worse, so we ignore that problem too.
4053 : : *
4054 : : * Note: likewise, pushing quals into a subquery with window functions is a
4055 : : * bit dubious: the quals might remove some rows of a window partition while
4056 : : * leaving others, causing changes in the window functions' results for the
4057 : : * surviving rows. We insist that such a qual reference only partitioning
4058 : : * columns, but again that only protects us if the qual does not distinguish
4059 : : * values that the partitioning equality operator sees as equal. The risks
4060 : : * here are perhaps larger than for DISTINCT, since no de-duplication of rows
4061 : : * occurs and thus there is no theoretical problem with such a qual. But
4062 : : * we'll do this anyway because the potential performance benefits are very
4063 : : * large, and we've seen no field complaints about the longstanding comparable
4064 : : * behavior with DISTINCT.
4065 : : */
4066 : : static bool
8222 4067 : 1136 : subquery_is_pushdown_safe(Query *subquery, Query *topquery,
4068 : : pushdown_safety_info *safetyInfo)
4069 : : {
4070 : : SetOperationStmt *topop;
4071 : :
4072 : : /* Check point 1 */
8255 4073 [ + + + + ]: 1136 : if (subquery->limitOffset != NULL || subquery->limitCount != NULL)
8460 4074 : 67 : return false;
4075 : :
4076 : : /* Check point 6 */
1892 4077 [ + + + + ]: 1069 : if (subquery->groupClause && subquery->groupingSets)
4078 : 6 : return false;
4079 : :
4080 : : /* Check points 3, 4, and 5 */
3317 4081 [ + + ]: 1063 : if (subquery->distinctClause ||
4082 [ + + ]: 1021 : subquery->hasWindowFuncs ||
4083 [ + + ]: 888 : subquery->hasTargetSRFs)
4140 4084 : 273 : safetyInfo->unsafeVolatile = true;
4085 : :
4086 : : /*
4087 : : * If we're at a leaf query, check for unsafe expressions in its target
4088 : : * list, and mark any reasons why they're unsafe in unsafeFlags[].
4089 : : * (Non-leaf nodes in setop trees have only simple Vars in their tlists,
4090 : : * so no need to check them.)
4091 : : */
4527 4092 [ + + ]: 1063 : if (subquery->setOperations == NULL)
4140 4093 : 969 : check_output_expressions(subquery, safetyInfo);
4094 : :
4095 : : /* Are we at top level, or looking at a setop component? */
8460 4096 [ + + ]: 1063 : if (subquery == topquery)
4097 : : {
4098 : : /* Top level, so check any component queries */
4099 [ + + ]: 875 : if (subquery->setOperations != NULL)
8222 4100 [ - + ]: 94 : if (!recurse_pushdown_safe(subquery->setOperations, topquery,
4101 : : safetyInfo))
8460 tgl@sss.pgh.pa.us 4102 :UBC 0 : return false;
4103 : : }
4104 : : else
4105 : : {
4106 : : /* Setop component must not have more components (too weird) */
8460 tgl@sss.pgh.pa.us 4107 [ - + ]:CBC 188 : if (subquery->setOperations != NULL)
8460 tgl@sss.pgh.pa.us 4108 :UBC 0 : return false;
4109 : : /* Check whether setop component output types match top level */
3170 peter_e@gmx.net 4110 :CBC 188 : topop = castNode(SetOperationStmt, topquery->setOperations);
4111 [ - + ]: 188 : Assert(topop);
8222 tgl@sss.pgh.pa.us 4112 : 188 : compare_tlist_datatypes(subquery->targetList,
4113 : : topop->colTypes,
4114 : : safetyInfo);
4115 : : }
8460 4116 : 1063 : return true;
4117 : : }
4118 : :
4119 : : /*
4120 : : * Helper routine to recurse through setOperations tree
4121 : : */
4122 : : static bool
8222 4123 : 282 : recurse_pushdown_safe(Node *setOp, Query *topquery,
4124 : : pushdown_safety_info *safetyInfo)
4125 : : {
8460 4126 [ + + ]: 282 : if (IsA(setOp, RangeTblRef))
4127 : : {
4128 : 188 : RangeTblRef *rtr = (RangeTblRef *) setOp;
4129 : 188 : RangeTblEntry *rte = rt_fetch(rtr->rtindex, topquery->rtable);
4130 : 188 : Query *subquery = rte->subquery;
4131 : :
4132 [ - + ]: 188 : Assert(subquery != NULL);
4140 4133 : 188 : return subquery_is_pushdown_safe(subquery, topquery, safetyInfo);
4134 : : }
8460 4135 [ + - ]: 94 : else if (IsA(setOp, SetOperationStmt))
4136 : : {
4137 : 94 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4138 : :
4139 : : /* EXCEPT is no good (point 2 for subquery_is_pushdown_safe) */
4140 [ - + ]: 94 : if (op->op == SETOP_EXCEPT)
8460 tgl@sss.pgh.pa.us 4141 :UBC 0 : return false;
4142 : : /* Else recurse */
4140 tgl@sss.pgh.pa.us 4143 [ - + ]:CBC 94 : if (!recurse_pushdown_safe(op->larg, topquery, safetyInfo))
8460 tgl@sss.pgh.pa.us 4144 :UBC 0 : return false;
4140 tgl@sss.pgh.pa.us 4145 [ - + ]:CBC 94 : if (!recurse_pushdown_safe(op->rarg, topquery, safetyInfo))
8460 tgl@sss.pgh.pa.us 4146 :UBC 0 : return false;
4147 : : }
4148 : : else
4149 : : {
8130 4150 [ # # ]: 0 : elog(ERROR, "unrecognized node type: %d",
4151 : : (int) nodeTag(setOp));
4152 : : }
8460 tgl@sss.pgh.pa.us 4153 :CBC 94 : return true;
4154 : : }
4155 : :
4156 : : /*
4157 : : * check_output_expressions - check subquery's output expressions for safety
4158 : : *
4159 : : * There are several cases in which it's unsafe to push down an upper-level
4160 : : * qual if it references a particular output column of a subquery. We check
4161 : : * each output column of the subquery and set flags in unsafeFlags[k] when we
4162 : : * see that column is unsafe for a pushed-down qual to reference. The
4163 : : * conditions checked here are:
4164 : : *
4165 : : * 1. We must not push down any quals that refer to subselect outputs that
4166 : : * return sets, else we'd introduce functions-returning-sets into the
4167 : : * subquery's WHERE/HAVING quals.
4168 : : *
4169 : : * 2. We must not push down any quals that refer to subselect outputs that
4170 : : * contain volatile functions, for fear of introducing strange results due
4171 : : * to multiple evaluation of a volatile function.
4172 : : *
4173 : : * 3. If the subquery uses DISTINCT ON, we must not push down any quals that
4174 : : * refer to non-DISTINCT output columns, because that could change the set
4175 : : * of rows returned. (This condition is vacuous for DISTINCT, because then
4176 : : * there are no non-DISTINCT output columns, so we needn't check. Note that
4177 : : * subquery_is_pushdown_safe already reported that we can't use volatile
4178 : : * quals if there's DISTINCT or DISTINCT ON.)
4179 : : *
4180 : : * 4. If the subquery has any window functions, we must not push down quals
4181 : : * that reference any output columns that are not listed in all the subquery's
4182 : : * window PARTITION BY clauses. We can push down quals that use only
4183 : : * partitioning columns because they should succeed or fail identically for
4184 : : * every row of any one window partition, and totally excluding some
4185 : : * partitions will not change a window function's results for remaining
4186 : : * partitions. (Again, this also requires nonvolatile quals, but
4187 : : * subquery_is_pushdown_safe handles that.). Subquery columns marked as
4188 : : * unsafe for this reason can still have WindowClause run conditions pushed
4189 : : * down.
4190 : : */
4191 : : static void
4140 4192 : 969 : check_output_expressions(Query *subquery, pushdown_safety_info *safetyInfo)
4193 : : {
4194 : : ListCell *lc;
4195 : :
4527 4196 [ + - + + : 9071 : foreach(lc, subquery->targetList)
+ + ]
4197 : : {
4198 : 8102 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4199 : :
4200 [ + + ]: 8102 : if (tle->resjunk)
4201 : 67 : continue; /* ignore resjunk columns */
4202 : :
4203 : : /* Functions returning sets are unsafe (point 1) */
3331 4204 [ + + ]: 8035 : if (subquery->hasTargetSRFs &&
955 drowley@postgresql.o 4205 [ + - ]: 334 : (safetyInfo->unsafeFlags[tle->resno] &
4206 [ + + ]: 334 : UNSAFE_HAS_SET_FUNC) == 0 &&
3331 tgl@sss.pgh.pa.us 4207 : 334 : expression_returns_set((Node *) tle->expr))
4208 : : {
955 drowley@postgresql.o 4209 : 188 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_SET_FUNC;
4527 tgl@sss.pgh.pa.us 4210 : 188 : continue;
4211 : : }
4212 : :
4213 : : /* Volatile functions are unsafe (point 2) */
955 drowley@postgresql.o 4214 [ + + ]: 7847 : if ((safetyInfo->unsafeFlags[tle->resno] &
4215 [ + + ]: 7841 : UNSAFE_HAS_VOLATILE_FUNC) == 0 &&
4216 : 7841 : contain_volatile_functions((Node *) tle->expr))
4217 : : {
4218 : 39 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_HAS_VOLATILE_FUNC;
4527 tgl@sss.pgh.pa.us 4219 : 39 : continue;
4220 : : }
4221 : :
4222 : : /* If subquery uses DISTINCT ON, check point 3 */
4223 [ - + ]: 7808 : if (subquery->hasDistinctOn &&
955 drowley@postgresql.o 4224 [ # # ]:UBC 0 : (safetyInfo->unsafeFlags[tle->resno] &
4225 : 0 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4527 tgl@sss.pgh.pa.us 4226 [ # # ]: 0 : !targetIsInSortList(tle, InvalidOid, subquery->distinctClause))
4227 : : {
4228 : : /* non-DISTINCT column, so mark it unsafe */
955 drowley@postgresql.o 4229 : 0 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_DISTINCTON_CLAUSE;
4527 tgl@sss.pgh.pa.us 4230 : 0 : continue;
4231 : : }
4232 : :
4233 : : /* If subquery uses window functions, check point 4 */
4140 tgl@sss.pgh.pa.us 4234 [ + + ]:CBC 7808 : if (subquery->hasWindowFuncs &&
955 drowley@postgresql.o 4235 [ + - ]: 545 : (safetyInfo->unsafeFlags[tle->resno] &
4236 : 1046 : UNSAFE_NOTIN_DISTINCTON_CLAUSE) == 0 &&
4140 tgl@sss.pgh.pa.us 4237 [ + + ]: 545 : !targetIsInAllPartitionLists(tle, subquery))
4238 : : {
4239 : : /* not present in all PARTITION BY clauses, so mark it unsafe */
955 drowley@postgresql.o 4240 : 501 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_NOTIN_PARTITIONBY_CLAUSE;
4140 tgl@sss.pgh.pa.us 4241 : 501 : continue;
4242 : : }
4243 : : }
4527 4244 : 969 : }
4245 : :
4246 : : /*
4247 : : * For subqueries using UNION/UNION ALL/INTERSECT/INTERSECT ALL, we can
4248 : : * push quals into each component query, but the quals can only reference
4249 : : * subquery columns that suffer no type coercions in the set operation.
4250 : : * Otherwise there are possible semantic gotchas. So, we check the
4251 : : * component queries to see if any of them have output types different from
4252 : : * the top-level setop outputs. We set the UNSAFE_TYPE_MISMATCH bit in
4253 : : * unsafeFlags[k] if column k has different type in any component.
4254 : : *
4255 : : * We don't have to care about typmods here: the only allowed difference
4256 : : * between set-op input and output typmods is input is a specific typmod
4257 : : * and output is -1, and that does not require a coercion.
4258 : : *
4259 : : * tlist is a subquery tlist.
4260 : : * colTypes is an OID list of the top-level setop's output column types.
4261 : : * safetyInfo is the pushdown_safety_info to set unsafeFlags[] for.
4262 : : */
4263 : : static void
8222 4264 : 188 : compare_tlist_datatypes(List *tlist, List *colTypes,
4265 : : pushdown_safety_info *safetyInfo)
4266 : : {
4267 : : ListCell *l;
7824 neilc@samurai.com 4268 : 188 : ListCell *colType = list_head(colTypes);
4269 : :
4270 [ + - + + : 588 : foreach(l, tlist)
+ + ]
4271 : : {
4272 : 400 : TargetEntry *tle = (TargetEntry *) lfirst(l);
4273 : :
7509 tgl@sss.pgh.pa.us 4274 [ - + ]: 400 : if (tle->resjunk)
8222 tgl@sss.pgh.pa.us 4275 :UBC 0 : continue; /* ignore resjunk columns */
7824 neilc@samurai.com 4276 [ - + ]:CBC 400 : if (colType == NULL)
8222 tgl@sss.pgh.pa.us 4277 [ # # ]:UBC 0 : elog(ERROR, "wrong number of tlist entries");
7509 tgl@sss.pgh.pa.us 4278 [ + + ]:CBC 400 : if (exprType((Node *) tle->expr) != lfirst_oid(colType))
955 drowley@postgresql.o 4279 : 68 : safetyInfo->unsafeFlags[tle->resno] |= UNSAFE_TYPE_MISMATCH;
2296 tgl@sss.pgh.pa.us 4280 : 400 : colType = lnext(colTypes, colType);
4281 : : }
7824 neilc@samurai.com 4282 [ - + ]: 188 : if (colType != NULL)
8222 tgl@sss.pgh.pa.us 4283 [ # # ]:UBC 0 : elog(ERROR, "wrong number of tlist entries");
8222 tgl@sss.pgh.pa.us 4284 :CBC 188 : }
4285 : :
4286 : : /*
4287 : : * targetIsInAllPartitionLists
4288 : : * True if the TargetEntry is listed in the PARTITION BY clause
4289 : : * of every window defined in the query.
4290 : : *
4291 : : * It would be safe to ignore windows not actually used by any window
4292 : : * function, but it's not easy to get that info at this stage; and it's
4293 : : * unlikely to be useful to spend any extra cycles getting it, since
4294 : : * unreferenced window definitions are probably infrequent in practice.
4295 : : */
4296 : : static bool
4140 4297 : 545 : targetIsInAllPartitionLists(TargetEntry *tle, Query *query)
4298 : : {
4299 : : ListCell *lc;
4300 : :
4301 [ + - + + : 601 : foreach(lc, query->windowClause)
+ + ]
4302 : : {
4303 : 557 : WindowClause *wc = (WindowClause *) lfirst(lc);
4304 : :
4305 [ + + ]: 557 : if (!targetIsInSortList(tle, InvalidOid, wc->partitionClause))
4306 : 501 : return false;
4307 : : }
4308 : 44 : return true;
4309 : : }
4310 : :
4311 : : /*
4312 : : * qual_is_pushdown_safe - is a particular rinfo safe to push down?
4313 : : *
4314 : : * rinfo is a restriction clause applying to the given subquery (whose RTE
4315 : : * has index rti in the parent query).
4316 : : *
4317 : : * Conditions checked here:
4318 : : *
4319 : : * 1. rinfo's clause must not contain any SubPlans (mainly because it's
4320 : : * unclear that it will work correctly: SubLinks will already have been
4321 : : * transformed into SubPlans in the qual, but not in the subquery). Note that
4322 : : * SubLinks that transform to initplans are safe, and will be accepted here
4323 : : * because what we'll see in the qual is just a Param referencing the initplan
4324 : : * output.
4325 : : *
4326 : : * 2. If unsafeVolatile is set, rinfo's clause must not contain any volatile
4327 : : * functions.
4328 : : *
4329 : : * 3. If unsafeLeaky is set, rinfo's clause must not contain any leaky
4330 : : * functions that are passed Var nodes, and therefore might reveal values from
4331 : : * the subquery as side effects.
4332 : : *
4333 : : * 4. rinfo's clause must not refer to the whole-row output of the subquery
4334 : : * (since there is no easy way to name that within the subquery itself).
4335 : : *
4336 : : * 5. rinfo's clause must not refer to any subquery output columns that were
4337 : : * found to be unsafe to reference by subquery_is_pushdown_safe().
4338 : : */
4339 : : static pushdown_safe_type
1673 drowley@postgresql.o 4340 : 1349 : qual_is_pushdown_safe(Query *subquery, Index rti, RestrictInfo *rinfo,
4341 : : pushdown_safety_info *safetyInfo)
4342 : : {
955 4343 : 1349 : pushdown_safe_type safe = PUSHDOWN_SAFE;
1673 4344 : 1349 : Node *qual = (Node *) rinfo->clause;
4345 : : List *vars;
4346 : : ListCell *vl;
4347 : :
4348 : : /* Refuse subselects (point 1) */
8255 tgl@sss.pgh.pa.us 4349 [ + + ]: 1349 : if (contain_subplans(qual))
955 drowley@postgresql.o 4350 : 33 : return PUSHDOWN_UNSAFE;
4351 : :
4352 : : /* Refuse volatile quals if we found they'd be unsafe (point 2) */
4140 tgl@sss.pgh.pa.us 4353 [ + + + + ]: 1642 : if (safetyInfo->unsafeVolatile &&
1673 drowley@postgresql.o 4354 : 326 : contain_volatile_functions((Node *) rinfo))
955 4355 : 9 : return PUSHDOWN_UNSAFE;
4356 : :
4357 : : /* Refuse leaky quals if told to (point 3) */
4140 tgl@sss.pgh.pa.us 4358 [ + + + + ]: 1896 : if (safetyInfo->unsafeLeaky &&
3836 sfrost@snowman.net 4359 : 589 : contain_leaked_vars(qual))
955 drowley@postgresql.o 4360 : 81 : return PUSHDOWN_UNSAFE;
4361 : :
4362 : : /*
4363 : : * Examine all Vars used in clause. Since it's a restriction clause, all
4364 : : * such Vars must refer to subselect output columns ... unless this is
4365 : : * part of a LATERAL subquery, in which case there could be lateral
4366 : : * references.
4367 : : *
4368 : : * By omitting the relevant flags, this also gives us a cheap sanity check
4369 : : * that no aggregates or window functions appear in the qual. Those would
4370 : : * be unsafe to push down, but at least for the moment we could never see
4371 : : * any in a qual anyhow.
4372 : : */
3518 tgl@sss.pgh.pa.us 4373 : 1226 : vars = pull_var_clause(qual, PVC_INCLUDE_PLACEHOLDERS);
8255 4374 [ + + + + : 2402 : foreach(vl, vars)
+ + ]
4375 : : {
8120 bruce@momjian.us 4376 : 1280 : Var *var = (Var *) lfirst(vl);
4377 : :
4378 : : /*
4379 : : * XXX Punt if we find any PlaceHolderVars in the restriction clause.
4380 : : * It's not clear whether a PHV could safely be pushed down, and even
4381 : : * less clear whether such a situation could arise in any cases of
4382 : : * practical interest anyway. So for the moment, just refuse to push
4383 : : * down.
4384 : : */
6215 tgl@sss.pgh.pa.us 4385 [ - + ]: 1280 : if (!IsA(var, Var))
4386 : : {
955 drowley@postgresql.o 4387 :UBC 0 : safe = PUSHDOWN_UNSAFE;
6215 tgl@sss.pgh.pa.us 4388 : 0 : break;
4389 : : }
4390 : :
4391 : : /*
4392 : : * Punt if we find any lateral references. It would be safe to push
4393 : : * these down, but we'd have to convert them into outer references,
4394 : : * which subquery_push_qual lacks the infrastructure to do. The case
4395 : : * arises so seldom that it doesn't seem worth working hard on.
4396 : : */
1932 tgl@sss.pgh.pa.us 4397 [ + + ]:CBC 1280 : if (var->varno != rti)
4398 : : {
955 drowley@postgresql.o 4399 : 6 : safe = PUSHDOWN_UNSAFE;
1932 tgl@sss.pgh.pa.us 4400 : 6 : break;
4401 : : }
4402 : :
4403 : : /* Subqueries have no system columns */
4527 4404 [ - + ]: 1274 : Assert(var->varattno >= 0);
4405 : :
4406 : : /* Check point 4 */
7196 4407 [ - + ]: 1274 : if (var->varattno == 0)
4408 : : {
955 drowley@postgresql.o 4409 :UBC 0 : safe = PUSHDOWN_UNSAFE;
7196 tgl@sss.pgh.pa.us 4410 : 0 : break;
4411 : : }
4412 : :
4413 : : /* Check point 5 */
955 drowley@postgresql.o 4414 [ + + ]:CBC 1274 : if (safetyInfo->unsafeFlags[var->varattno] != 0)
4415 : : {
4416 [ + + ]: 263 : if (safetyInfo->unsafeFlags[var->varattno] &
4417 : : (UNSAFE_HAS_VOLATILE_FUNC | UNSAFE_HAS_SET_FUNC |
4418 : : UNSAFE_NOTIN_DISTINCTON_CLAUSE | UNSAFE_TYPE_MISMATCH))
4419 : : {
4420 : 98 : safe = PUSHDOWN_UNSAFE;
4421 : 98 : break;
4422 : : }
4423 : : else
4424 : : {
4425 : : /* UNSAFE_NOTIN_PARTITIONBY_CLAUSE is ok for run conditions */
4426 : 165 : safe = PUSHDOWN_WINDOWCLAUSE_RUNCOND;
4427 : : /* don't break, we might find another Var that's unsafe */
4428 : : }
4429 : : }
4430 : : }
4431 : :
7820 neilc@samurai.com 4432 : 1226 : list_free(vars);
4433 : :
8255 tgl@sss.pgh.pa.us 4434 : 1226 : return safe;
4435 : : }
4436 : :
4437 : : /*
4438 : : * subquery_push_qual - push down a qual that we have determined is safe
4439 : : */
4440 : : static void
7450 4441 : 1160 : subquery_push_qual(Query *subquery, RangeTblEntry *rte, Index rti, Node *qual)
4442 : : {
8460 4443 [ + + ]: 1160 : if (subquery->setOperations != NULL)
4444 : : {
4445 : : /* Recurse to push it separately to each component query */
7739 4446 : 82 : recurse_push_qual(subquery->setOperations, subquery,
4447 : : rte, rti, qual);
4448 : : }
4449 : : else
4450 : : {
4451 : : /*
4452 : : * We need to replace Vars in the qual (which must refer to outputs of
4453 : : * the subquery) with copies of the subquery's targetlist expressions.
4454 : : * Note that at this point, any uplevel Vars in the qual should have
4455 : : * been replaced with Params, so they need no work.
4456 : : *
4457 : : * This step also ensures that when we are pushing into a setop tree,
4458 : : * each component query gets its own copy of the qual.
4459 : : */
4736 4460 : 1078 : qual = ReplaceVarsFromTargetList(qual, rti, 0, rte,
4461 : : subquery->targetList,
4462 : : subquery->resultRelation,
4463 : : REPLACEVARS_REPORT_ERROR, 0,
4464 : : &subquery->hasSubLinks);
4465 : :
4466 : : /*
4467 : : * Now attach the qual to the proper place: normally WHERE, but if the
4468 : : * subquery uses grouping or aggregation, put it in HAVING (since the
4469 : : * qual really refers to the group-result rows).
4470 : : */
3817 andres@anarazel.de 4471 [ + + + - : 1078 : if (subquery->hasAggs || subquery->groupClause || subquery->groupingSets || subquery->havingQual)
+ - - + ]
7536 tgl@sss.pgh.pa.us 4472 : 139 : subquery->havingQual = make_and_qual(subquery->havingQual, qual);
4473 : : else
4474 : 939 : subquery->jointree->quals =
4475 : 939 : make_and_qual(subquery->jointree->quals, qual);
4476 : :
4477 : : /*
4478 : : * We need not change the subquery's hasAggs or hasSubLinks flags,
4479 : : * since we can't be pushing down any aggregates that weren't there
4480 : : * before, and we don't push down subselects at all.
4481 : : */
4482 : : }
8460 4483 : 1160 : }
4484 : :
4485 : : /*
4486 : : * Helper routine to recurse through setOperations tree
4487 : : */
4488 : : static void
4489 : 246 : recurse_push_qual(Node *setOp, Query *topquery,
4490 : : RangeTblEntry *rte, Index rti, Node *qual)
4491 : : {
4492 [ + + ]: 246 : if (IsA(setOp, RangeTblRef))
4493 : : {
4494 : 164 : RangeTblRef *rtr = (RangeTblRef *) setOp;
7840 4495 : 164 : RangeTblEntry *subrte = rt_fetch(rtr->rtindex, topquery->rtable);
4496 : 164 : Query *subquery = subrte->subquery;
4497 : :
8460 4498 [ - + ]: 164 : Assert(subquery != NULL);
7450 4499 : 164 : subquery_push_qual(subquery, rte, rti, qual);
4500 : : }
8460 4501 [ + - ]: 82 : else if (IsA(setOp, SetOperationStmt))
4502 : : {
4503 : 82 : SetOperationStmt *op = (SetOperationStmt *) setOp;
4504 : :
7450 4505 : 82 : recurse_push_qual(op->larg, topquery, rte, rti, qual);
4506 : 82 : recurse_push_qual(op->rarg, topquery, rte, rti, qual);
4507 : : }
4508 : : else
4509 : : {
8130 tgl@sss.pgh.pa.us 4510 [ # # ]:UBC 0 : elog(ERROR, "unrecognized node type: %d",
4511 : : (int) nodeTag(setOp));
4512 : : }
8460 tgl@sss.pgh.pa.us 4513 :CBC 246 : }
4514 : :
4515 : : /*****************************************************************************
4516 : : * SIMPLIFYING SUBQUERY TARGETLISTS
4517 : : *****************************************************************************/
4518 : :
4519 : : /*
4520 : : * remove_unused_subquery_outputs
4521 : : * Remove subquery targetlist items we don't need
4522 : : *
4523 : : * It's possible, even likely, that the upper query does not read all the
4524 : : * output columns of the subquery. We can remove any such outputs that are
4525 : : * not needed by the subquery itself (e.g., as sort/group columns) and do not
4526 : : * affect semantics otherwise (e.g., volatile functions can't be removed).
4527 : : * This is useful not only because we might be able to remove expensive-to-
4528 : : * compute expressions, but because deletion of output columns might allow
4529 : : * optimizations such as join removal to occur within the subquery.
4530 : : *
4531 : : * extra_used_attrs can be passed as non-NULL to mark any columns (offset by
4532 : : * FirstLowInvalidHeapAttributeNumber) that we should not remove. This
4533 : : * parameter is modified by the function, so callers must make a copy if they
4534 : : * need to use the passed in Bitmapset after calling this function.
4535 : : *
4536 : : * To avoid affecting column numbering in the targetlist, we don't physically
4537 : : * remove unused tlist entries, but rather replace their expressions with NULL
4538 : : * constants. This is implemented by modifying subquery->targetList.
4539 : : */
4540 : : static void
1249 drowley@postgresql.o 4541 : 8757 : remove_unused_subquery_outputs(Query *subquery, RelOptInfo *rel,
4542 : : Bitmapset *extra_used_attrs)
4543 : : {
4544 : : Bitmapset *attrs_used;
4545 : : ListCell *lc;
4546 : :
4547 : : /*
4548 : : * Just point directly to extra_used_attrs. No need to bms_copy as none of
4549 : : * the current callers use the Bitmapset after calling this function.
4550 : : */
4551 : 8757 : attrs_used = extra_used_attrs;
4552 : :
4553 : : /*
4554 : : * Do nothing if subquery has UNION/INTERSECT/EXCEPT: in principle we
4555 : : * could update all the child SELECTs' tlists, but it seems not worth the
4556 : : * trouble presently.
4557 : : */
4155 tgl@sss.pgh.pa.us 4558 [ + + ]: 8757 : if (subquery->setOperations)
4559 : 1011 : return;
4560 : :
4561 : : /*
4562 : : * If subquery has regular DISTINCT (not DISTINCT ON), we're wasting our
4563 : : * time: all its output columns must be used in the distinctClause.
4564 : : */
4565 [ + + + + ]: 8324 : if (subquery->distinctClause && !subquery->hasDistinctOn)
4566 : 426 : return;
4567 : :
4568 : : /*
4569 : : * Collect a bitmap of all the output column numbers used by the upper
4570 : : * query.
4571 : : *
4572 : : * Add all the attributes needed for joins or final output. Note: we must
4573 : : * look at rel's targetlist, not the attr_needed data, because attr_needed
4574 : : * isn't computed for inheritance child rels, cf set_append_rel_size().
4575 : : * (XXX might be worth changing that sometime.)
4576 : : */
3514 4577 : 7898 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
4578 : :
4579 : : /* Add all the attributes used by un-pushed-down restriction clauses. */
4155 4580 [ + + + + : 8266 : foreach(lc, rel->baserestrictinfo)
+ + ]
4581 : : {
4582 : 368 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4583 : :
4584 : 368 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
4585 : : }
4586 : :
4587 : : /*
4588 : : * If there's a whole-row reference to the subquery, we can't remove
4589 : : * anything.
4590 : : */
4591 [ + + ]: 7898 : if (bms_is_member(0 - FirstLowInvalidHeapAttributeNumber, attrs_used))
4592 : 152 : return;
4593 : :
4594 : : /*
4595 : : * Run through the tlist and zap entries we don't need. It's okay to
4596 : : * modify the tlist items in-place because set_subquery_pathlist made a
4597 : : * copy of the subquery.
4598 : : */
4599 [ + + + + : 44281 : foreach(lc, subquery->targetList)
+ + ]
4600 : : {
4601 : 36535 : TargetEntry *tle = (TargetEntry *) lfirst(lc);
4602 : 36535 : Node *texpr = (Node *) tle->expr;
4603 : :
4604 : : /*
4605 : : * If it has a sortgroupref number, it's used in some sort/group
4606 : : * clause so we'd better not remove it. Also, don't remove any
4607 : : * resjunk columns, since their reason for being has nothing to do
4608 : : * with anybody reading the subquery's output. (It's likely that
4609 : : * resjunk columns in a sub-SELECT would always have ressortgroupref
4610 : : * set, but even if they don't, it seems imprudent to remove them.)
4611 : : */
4612 [ + + - + ]: 36535 : if (tle->ressortgroupref || tle->resjunk)
4613 : 1311 : continue;
4614 : :
4615 : : /*
4616 : : * If it's used by the upper query, we can't remove it.
4617 : : */
4618 [ + + ]: 35224 : if (bms_is_member(tle->resno - FirstLowInvalidHeapAttributeNumber,
4619 : : attrs_used))
4620 : 23814 : continue;
4621 : :
4622 : : /*
4623 : : * If it contains a set-returning function, we can't remove it since
4624 : : * that could change the number of rows returned by the subquery.
4625 : : */
3331 4626 [ + + + + ]: 11930 : if (subquery->hasTargetSRFs &&
4627 : 520 : expression_returns_set(texpr))
4155 4628 : 386 : continue;
4629 : :
4630 : : /*
4631 : : * If it contains volatile functions, we daren't remove it for fear
4632 : : * that the user is expecting their side-effects to happen.
4633 : : */
4634 [ + + ]: 11024 : if (contain_volatile_functions(texpr))
4635 : 15 : continue;
4636 : :
4637 : : /*
4638 : : * OK, we don't need it. Replace the expression with a NULL constant.
4639 : : * Preserve the exposed type of the expression, in case something
4640 : : * looks at the rowtype of the subquery's result.
4641 : : */
4642 : 11009 : tle->expr = (Expr *) makeNullConst(exprType(texpr),
4643 : : exprTypmod(texpr),
4644 : : exprCollation(texpr));
4645 : : }
4646 : : }
4647 : :
4648 : : /*
4649 : : * create_partial_bitmap_paths
4650 : : * Build partial bitmap heap path for the relation
4651 : : */
4652 : : void
3155 rhaas@postgresql.org 4653 : 73271 : create_partial_bitmap_paths(PlannerInfo *root, RelOptInfo *rel,
4654 : : Path *bitmapqual)
4655 : : {
4656 : : int parallel_workers;
4657 : : double pages_fetched;
4658 : :
4659 : : /* Compute heap pages for bitmap heap scan */
4660 : 73271 : pages_fetched = compute_bitmap_pages(root, rel, bitmapqual, 1.0,
4661 : : NULL, NULL);
4662 : :
2824 4663 : 73271 : parallel_workers = compute_parallel_worker(rel, pages_fetched, -1,
4664 : : max_parallel_workers_per_gather);
4665 : :
3155 4666 [ + + ]: 73271 : if (parallel_workers <= 0)
4667 : 71188 : return;
4668 : :
4669 : 2083 : add_partial_path(rel, (Path *) create_bitmap_heap_path(root, rel,
4670 : : bitmapqual, rel->lateral_relids, 1.0, parallel_workers));
4671 : : }
4672 : :
4673 : : /*
4674 : : * Compute the number of parallel workers that should be used to scan a
4675 : : * relation. We compute the parallel workers based on the size of the heap to
4676 : : * be scanned and the size of the index to be scanned, then choose a minimum
4677 : : * of those.
4678 : : *
4679 : : * "heap_pages" is the number of pages from the table that we expect to scan, or
4680 : : * -1 if we don't expect to scan any.
4681 : : *
4682 : : * "index_pages" is the number of pages from the index that we expect to scan, or
4683 : : * -1 if we don't expect to scan any.
4684 : : *
4685 : : * "max_workers" is caller's limit on the number of workers. This typically
4686 : : * comes from a GUC.
4687 : : */
4688 : : int
2824 4689 : 382253 : compute_parallel_worker(RelOptInfo *rel, double heap_pages, double index_pages,
4690 : : int max_workers)
4691 : : {
3176 4692 : 382253 : int parallel_workers = 0;
4693 : :
4694 : : /*
4695 : : * If the user has set the parallel_workers reloption, use that; otherwise
4696 : : * select a default number of workers.
4697 : : */
3204 4698 [ + + ]: 382253 : if (rel->rel_parallel_workers != -1)
4699 : 957 : parallel_workers = rel->rel_parallel_workers;
4700 : : else
4701 : : {
4702 : : /*
4703 : : * If the number of pages being scanned is insufficient to justify a
4704 : : * parallel scan, just return zero ... unless it's an inheritance
4705 : : * child. In that case, we want to generate a parallel path here
4706 : : * anyway. It might not be worthwhile just for this relation, but
4707 : : * when combined with all of its inheritance siblings it may well pay
4708 : : * off.
4709 : : */
3149 4710 [ + + + + ]: 381296 : if (rel->reloptkind == RELOPT_BASEREL &&
4711 [ + + + + ]: 361545 : ((heap_pages >= 0 && heap_pages < min_parallel_table_scan_size) ||
3050 tgl@sss.pgh.pa.us 4712 [ + + ]: 11103 : (index_pages >= 0 && index_pages < min_parallel_index_scan_size)))
3204 rhaas@postgresql.org 4713 : 361152 : return 0;
4714 : :
3149 4715 [ + + ]: 20144 : if (heap_pages >= 0)
4716 : : {
4717 : : int heap_parallel_threshold;
4718 : 19077 : int heap_parallel_workers = 1;
4719 : :
4720 : : /*
4721 : : * Select the number of workers based on the log of the size of
4722 : : * the relation. This probably needs to be a good deal more
4723 : : * sophisticated, but we need something here for now. Note that
4724 : : * the upper limit of the min_parallel_table_scan_size GUC is
4725 : : * chosen to prevent overflow here.
4726 : : */
3176 4727 : 19077 : heap_parallel_threshold = Max(min_parallel_table_scan_size, 1);
4728 [ + + ]: 21511 : while (heap_pages >= (BlockNumber) (heap_parallel_threshold * 3))
4729 : : {
4730 : 2434 : heap_parallel_workers++;
4731 : 2434 : heap_parallel_threshold *= 3;
4732 [ - + ]: 2434 : if (heap_parallel_threshold > INT_MAX / 3)
3176 rhaas@postgresql.org 4733 :UBC 0 : break; /* avoid overflow */
4734 : : }
4735 : :
3176 rhaas@postgresql.org 4736 :CBC 19077 : parallel_workers = heap_parallel_workers;
4737 : : }
4738 : :
3149 4739 [ + + ]: 20144 : if (index_pages >= 0)
4740 : : {
4741 : 4901 : int index_parallel_workers = 1;
4742 : : int index_parallel_threshold;
4743 : :
4744 : : /* same calculation as for heap_pages above */
3176 4745 : 4901 : index_parallel_threshold = Max(min_parallel_index_scan_size, 1);
4746 [ + + ]: 5039 : while (index_pages >= (BlockNumber) (index_parallel_threshold * 3))
4747 : : {
4748 : 138 : index_parallel_workers++;
4749 : 138 : index_parallel_threshold *= 3;
4750 [ - + ]: 138 : if (index_parallel_threshold > INT_MAX / 3)
3176 rhaas@postgresql.org 4751 :UBC 0 : break; /* avoid overflow */
4752 : : }
4753 : :
3176 rhaas@postgresql.org 4754 [ + + ]:CBC 4901 : if (parallel_workers > 0)
4755 : 3834 : parallel_workers = Min(parallel_workers, index_parallel_workers);
4756 : : else
4757 : 1067 : parallel_workers = index_parallel_workers;
4758 : : }
4759 : : }
4760 : :
4761 : : /* In no case use more than caller supplied maximum number of workers */
2824 4762 : 21101 : parallel_workers = Min(parallel_workers, max_workers);
4763 : :
3204 4764 : 21101 : return parallel_workers;
4765 : : }
4766 : :
4767 : : /*
4768 : : * generate_partitionwise_join_paths
4769 : : * Create paths representing partitionwise join for given partitioned
4770 : : * join relation.
4771 : : *
4772 : : * This must not be called until after we are done adding paths for all
4773 : : * child-joins. Otherwise, add_path might delete a path to which some path
4774 : : * generated here has a reference.
4775 : : */
4776 : : void
2810 peter_e@gmx.net 4777 : 118464 : generate_partitionwise_join_paths(PlannerInfo *root, RelOptInfo *rel)
4778 : : {
2943 rhaas@postgresql.org 4779 : 118464 : List *live_children = NIL;
4780 : : int cnt_parts;
4781 : : int num_parts;
4782 : : RelOptInfo **part_rels;
4783 : :
4784 : : /* Handle only join relations here. */
4785 [ + + - + ]: 118464 : if (!IS_JOIN_REL(rel))
2943 rhaas@postgresql.org 4786 :UBC 0 : return;
4787 : :
4788 : : /* We've nothing to do if the relation is not partitioned. */
2821 rhaas@postgresql.org 4789 [ + + + + :CBC 118464 : if (!IS_PARTITIONED_REL(rel))
+ + + - +
+ ]
2943 4790 : 114891 : return;
4791 : :
4792 : : /* The relation should have consider_partitionwise_join set. */
2614 efujita@postgresql.o 4793 [ - + ]: 3573 : Assert(rel->consider_partitionwise_join);
4794 : :
4795 : : /* Guard against stack overflow due to overly deep partition hierarchy. */
2943 rhaas@postgresql.org 4796 : 3573 : check_stack_depth();
4797 : :
4798 : 3573 : num_parts = rel->nparts;
4799 : 3573 : part_rels = rel->part_rels;
4800 : :
4801 : : /* Collect non-dummy child-joins. */
4802 [ + + ]: 12724 : for (cnt_parts = 0; cnt_parts < num_parts; cnt_parts++)
4803 : : {
4804 : 9151 : RelOptInfo *child_rel = part_rels[cnt_parts];
4805 : :
4806 : : /* If it's been pruned entirely, it's certainly dummy. */
2403 tgl@sss.pgh.pa.us 4807 [ + + ]: 9151 : if (child_rel == NULL)
4808 : 32 : continue;
4809 : :
4810 : : /* Make partitionwise join paths for this partitioned child-join. */
2810 peter_e@gmx.net 4811 : 9119 : generate_partitionwise_join_paths(root, child_rel);
4812 : :
4813 : : /* If we failed to make any path for this child, we must give up. */
1058 tgl@sss.pgh.pa.us 4814 [ - + ]: 9119 : if (child_rel->pathlist == NIL)
4815 : : {
4816 : : /*
4817 : : * Mark the parent joinrel as unpartitioned so that later
4818 : : * functions treat it correctly.
4819 : : */
1058 tgl@sss.pgh.pa.us 4820 :UBC 0 : rel->nparts = 0;
4821 : 0 : return;
4822 : : }
4823 : :
4824 : : /* Else, identify the cheapest path for it. */
2426 tgl@sss.pgh.pa.us 4825 :CBC 9119 : set_cheapest(child_rel);
4826 : :
4827 : : /* Dummy children need not be scanned, so ignore those. */
2943 rhaas@postgresql.org 4828 [ - + ]: 9119 : if (IS_DUMMY_REL(child_rel))
2943 rhaas@postgresql.org 4829 :UBC 0 : continue;
4830 : :
4831 : : /*
4832 : : * Except for the topmost scan/join rel, consider generating partial
4833 : : * aggregation paths for the grouped relation on top of the paths of
4834 : : * this partitioned child-join. After that, we're done creating paths
4835 : : * for the grouped relation, so run set_cheapest().
4836 : : */
19 rguo@postgresql.org 4837 [ + + ]:GNC 9119 : if (child_rel->grouped_rel != NULL &&
4838 [ + + - + : 6438 : !bms_equal(IS_OTHER_REL(rel) ?
+ + ]
4839 : : rel->top_parent_relids : rel->relids,
4840 [ + - ]: 6438 : root->all_query_rels))
4841 : : {
4842 : 120 : RelOptInfo *grouped_rel = child_rel->grouped_rel;
4843 : :
4844 [ - + ]: 120 : Assert(IS_GROUPED_REL(grouped_rel));
4845 : :
4846 : 120 : generate_grouped_paths(root, grouped_rel, child_rel);
4847 : 120 : set_cheapest(grouped_rel);
4848 : : }
4849 : :
4850 : : #ifdef OPTIMIZER_DEBUG
4851 : : pprint(child_rel);
4852 : : #endif
4853 : :
2943 rhaas@postgresql.org 4854 :CBC 9119 : live_children = lappend(live_children, child_rel);
4855 : : }
4856 : :
4857 : : /* If all child-joins are dummy, parent join is also dummy. */
4858 [ - + ]: 3573 : if (!live_children)
4859 : : {
2943 rhaas@postgresql.org 4860 :UBC 0 : mark_dummy_rel(rel);
4861 : 0 : return;
4862 : : }
4863 : :
4864 : : /* Build additional paths for this rel from child-join paths. */
2943 rhaas@postgresql.org 4865 :CBC 3573 : add_paths_to_append_rel(root, rel, live_children);
4866 : 3573 : list_free(live_children);
4867 : : }
|
