Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * plancat.c
4 : * routines for accessing the system catalogs
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : *
11 : * IDENTIFICATION
12 : * src/backend/optimizer/util/plancat.c
13 : *
14 : *-------------------------------------------------------------------------
15 : */
16 : #include "postgres.h"
17 :
18 : #include <math.h>
19 :
20 : #include "access/genam.h"
21 : #include "access/htup_details.h"
22 : #include "access/nbtree.h"
23 : #include "access/sysattr.h"
24 : #include "access/table.h"
25 : #include "access/tableam.h"
26 : #include "access/transam.h"
27 : #include "access/xlog.h"
28 : #include "catalog/catalog.h"
29 : #include "catalog/dependency.h"
30 : #include "catalog/heap.h"
31 : #include "catalog/pg_am.h"
32 : #include "catalog/pg_proc.h"
33 : #include "catalog/pg_statistic_ext.h"
34 : #include "foreign/fdwapi.h"
35 : #include "miscadmin.h"
36 : #include "nodes/makefuncs.h"
37 : #include "nodes/supportnodes.h"
38 : #include "optimizer/clauses.h"
39 : #include "optimizer/cost.h"
40 : #include "optimizer/optimizer.h"
41 : #include "optimizer/plancat.h"
42 : #include "optimizer/prep.h"
43 : #include "parser/parse_relation.h"
44 : #include "parser/parsetree.h"
45 : #include "partitioning/partdesc.h"
46 : #include "rewrite/rewriteManip.h"
47 : #include "statistics/statistics.h"
48 : #include "storage/bufmgr.h"
49 : #include "utils/builtins.h"
50 : #include "utils/lsyscache.h"
51 : #include "utils/partcache.h"
52 : #include "utils/rel.h"
53 : #include "utils/snapmgr.h"
54 : #include "utils/syscache.h"
55 :
56 : /* GUC parameter */
57 : int constraint_exclusion = CONSTRAINT_EXCLUSION_PARTITION;
58 :
59 : /* Hook for plugins to get control in get_relation_info() */
60 : get_relation_info_hook_type get_relation_info_hook = NULL;
61 :
62 :
63 : static void get_relation_foreign_keys(PlannerInfo *root, RelOptInfo *rel,
64 : Relation relation, bool inhparent);
65 : static bool infer_collation_opclass_match(InferenceElem *elem, Relation idxRel,
66 : List *idxExprs);
67 : static List *get_relation_constraints(PlannerInfo *root,
68 : Oid relationObjectId, RelOptInfo *rel,
69 : bool include_noinherit,
70 : bool include_notnull,
71 : bool include_partition);
72 : static List *build_index_tlist(PlannerInfo *root, IndexOptInfo *index,
73 : Relation heapRelation);
74 : static List *get_relation_statistics(RelOptInfo *rel, Relation relation);
75 : static void set_relation_partition_info(PlannerInfo *root, RelOptInfo *rel,
76 : Relation relation);
77 : static PartitionScheme find_partition_scheme(PlannerInfo *root, Relation rel);
78 : static void set_baserel_partition_key_exprs(Relation relation,
79 : RelOptInfo *rel);
80 : static void set_baserel_partition_constraint(Relation relation,
81 : RelOptInfo *rel);
82 :
83 :
84 : /*
85 : * get_relation_info -
86 : * Retrieves catalog information for a given relation.
87 : *
88 : * Given the Oid of the relation, return the following info into fields
89 : * of the RelOptInfo struct:
90 : *
91 : * min_attr lowest valid AttrNumber
92 : * max_attr highest valid AttrNumber
93 : * indexlist list of IndexOptInfos for relation's indexes
94 : * statlist list of StatisticExtInfo for relation's statistic objects
95 : * serverid if it's a foreign table, the server OID
96 : * fdwroutine if it's a foreign table, the FDW function pointers
97 : * pages number of pages
98 : * tuples number of tuples
99 : * rel_parallel_workers user-defined number of parallel workers
100 : *
101 : * Also, add information about the relation's foreign keys to root->fkey_list.
102 : *
103 : * Also, initialize the attr_needed[] and attr_widths[] arrays. In most
104 : * cases these are left as zeroes, but sometimes we need to compute attr
105 : * widths here, and we may as well cache the results for costsize.c.
106 : *
107 : * If inhparent is true, all we need to do is set up the attr arrays:
108 : * the RelOptInfo actually represents the appendrel formed by an inheritance
109 : * tree, and so the parent rel's physical size and index information isn't
110 : * important for it.
111 : */
112 : void
113 282604 : get_relation_info(PlannerInfo *root, Oid relationObjectId, bool inhparent,
114 : RelOptInfo *rel)
115 : {
116 282604 : Index varno = rel->relid;
117 : Relation relation;
118 : bool hasindex;
119 282604 : List *indexinfos = NIL;
120 :
121 : /*
122 : * We need not lock the relation since it was already locked, either by
123 : * the rewriter or when expand_inherited_rtentry() added it to the query's
124 : * rangetable.
125 : */
126 282604 : relation = table_open(relationObjectId, NoLock);
127 :
128 : /* Temporary and unlogged relations are inaccessible during recovery. */
129 282604 : if (!RelationNeedsWAL(relation) && RecoveryInProgress())
130 0 : ereport(ERROR,
131 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
132 : errmsg("cannot access temporary or unlogged relations during recovery")));
133 :
134 282604 : rel->min_attr = FirstLowInvalidHeapAttributeNumber + 1;
135 282604 : rel->max_attr = RelationGetNumberOfAttributes(relation);
136 282604 : rel->reltablespace = RelationGetForm(relation)->reltablespace;
137 :
138 : Assert(rel->max_attr >= rel->min_attr);
139 282604 : rel->attr_needed = (Relids *)
140 282604 : palloc0((rel->max_attr - rel->min_attr + 1) * sizeof(Relids));
141 282604 : rel->attr_widths = (int32 *)
142 282604 : palloc0((rel->max_attr - rel->min_attr + 1) * sizeof(int32));
143 :
144 : /*
145 : * Estimate relation size --- unless it's an inheritance parent, in which
146 : * case the size we want is not the rel's own size but the size of its
147 : * inheritance tree. That will be computed in set_append_rel_size().
148 : */
149 282604 : if (!inhparent)
150 248276 : estimate_rel_size(relation, rel->attr_widths - rel->min_attr,
151 : &rel->pages, &rel->tuples, &rel->allvisfrac);
152 :
153 : /* Retrieve the parallel_workers reloption, or -1 if not set. */
154 282604 : rel->rel_parallel_workers = RelationGetParallelWorkers(relation, -1);
155 :
156 : /*
157 : * Make list of indexes. Ignore indexes on system catalogs if told to.
158 : * Don't bother with indexes for an inheritance parent, either.
159 : */
160 282604 : if (inhparent ||
161 248276 : (IgnoreSystemIndexes && IsSystemRelation(relation)))
162 34328 : hasindex = false;
163 : else
164 248276 : hasindex = relation->rd_rel->relhasindex;
165 :
166 282604 : if (hasindex)
167 : {
168 : List *indexoidlist;
169 : LOCKMODE lmode;
170 : ListCell *l;
171 :
172 206790 : indexoidlist = RelationGetIndexList(relation);
173 :
174 : /*
175 : * For each index, we get the same type of lock that the executor will
176 : * need, and do not release it. This saves a couple of trips to the
177 : * shared lock manager while not creating any real loss of
178 : * concurrency, because no schema changes could be happening on the
179 : * index while we hold lock on the parent rel, and no lock type used
180 : * for queries blocks any other kind of index operation.
181 : */
182 206790 : lmode = root->simple_rte_array[varno]->rellockmode;
183 :
184 623986 : foreach(l, indexoidlist)
185 : {
186 417196 : Oid indexoid = lfirst_oid(l);
187 : Relation indexRelation;
188 : Form_pg_index index;
189 : IndexAmRoutine *amroutine;
190 : IndexOptInfo *info;
191 : int ncolumns,
192 : nkeycolumns;
193 : int i;
194 :
195 : /*
196 : * Extract info from the relation descriptor for the index.
197 : */
198 417196 : indexRelation = index_open(indexoid, lmode);
199 417196 : index = indexRelation->rd_index;
200 :
201 : /*
202 : * Ignore invalid indexes, since they can't safely be used for
203 : * queries. Note that this is OK because the data structure we
204 : * are constructing is only used by the planner --- the executor
205 : * still needs to insert into "invalid" indexes, if they're marked
206 : * indisready.
207 : */
208 417196 : if (!index->indisvalid)
209 : {
210 18 : index_close(indexRelation, NoLock);
211 18 : continue;
212 : }
213 :
214 : /*
215 : * Ignore partitioned indexes, since they are not usable for
216 : * queries.
217 : */
218 417178 : if (indexRelation->rd_rel->relkind == RELKIND_PARTITIONED_INDEX)
219 : {
220 4 : index_close(indexRelation, NoLock);
221 4 : continue;
222 : }
223 :
224 : /*
225 : * If the index is valid, but cannot yet be used, ignore it; but
226 : * mark the plan we are generating as transient. See
227 : * src/backend/access/heap/README.HOT for discussion.
228 : */
229 417174 : if (index->indcheckxmin &&
230 80 : !TransactionIdPrecedes(HeapTupleHeaderGetXmin(indexRelation->rd_indextuple->t_data),
231 : TransactionXmin))
232 : {
233 2 : root->glob->transientPlan = true;
234 2 : index_close(indexRelation, NoLock);
235 2 : continue;
236 : }
237 :
238 417172 : info = makeNode(IndexOptInfo);
239 :
240 417172 : info->indexoid = index->indexrelid;
241 417172 : info->reltablespace =
242 417172 : RelationGetForm(indexRelation)->reltablespace;
243 417172 : info->rel = rel;
244 417172 : info->ncolumns = ncolumns = index->indnatts;
245 417172 : info->nkeycolumns = nkeycolumns = index->indnkeyatts;
246 :
247 417172 : info->indexkeys = (int *) palloc(sizeof(int) * ncolumns);
248 417172 : info->indexcollations = (Oid *) palloc(sizeof(Oid) * nkeycolumns);
249 417172 : info->opfamily = (Oid *) palloc(sizeof(Oid) * nkeycolumns);
250 417172 : info->opcintype = (Oid *) palloc(sizeof(Oid) * nkeycolumns);
251 417172 : info->canreturn = (bool *) palloc(sizeof(bool) * ncolumns);
252 :
253 1167690 : for (i = 0; i < ncolumns; i++)
254 : {
255 750518 : info->indexkeys[i] = index->indkey.values[i];
256 750518 : info->canreturn[i] = index_can_return(indexRelation, i + 1);
257 : }
258 :
259 1167390 : for (i = 0; i < nkeycolumns; i++)
260 : {
261 750218 : info->opfamily[i] = indexRelation->rd_opfamily[i];
262 750218 : info->opcintype[i] = indexRelation->rd_opcintype[i];
263 750218 : info->indexcollations[i] = indexRelation->rd_indcollation[i];
264 : }
265 :
266 417172 : info->relam = indexRelation->rd_rel->relam;
267 :
268 : /* We copy just the fields we need, not all of rd_indam */
269 417172 : amroutine = indexRelation->rd_indam;
270 417172 : info->amcanorderbyop = amroutine->amcanorderbyop;
271 417172 : info->amoptionalkey = amroutine->amoptionalkey;
272 417172 : info->amsearcharray = amroutine->amsearcharray;
273 417172 : info->amsearchnulls = amroutine->amsearchnulls;
274 417172 : info->amcanparallel = amroutine->amcanparallel;
275 417172 : info->amhasgettuple = (amroutine->amgettuple != NULL);
276 834344 : info->amhasgetbitmap = amroutine->amgetbitmap != NULL &&
277 417172 : relation->rd_tableam->scan_bitmap_next_block != NULL;
278 417172 : info->amcostestimate = amroutine->amcostestimate;
279 : Assert(info->amcostestimate != NULL);
280 :
281 : /* Fetch index opclass options */
282 417172 : info->opclassoptions = RelationGetIndexAttOptions(indexRelation, true);
283 :
284 : /*
285 : * Fetch the ordering information for the index, if any.
286 : */
287 417172 : if (info->relam == BTREE_AM_OID)
288 : {
289 : /*
290 : * If it's a btree index, we can use its opfamily OIDs
291 : * directly as the sort ordering opfamily OIDs.
292 : */
293 : Assert(amroutine->amcanorder);
294 :
295 407924 : info->sortopfamily = info->opfamily;
296 407924 : info->reverse_sort = (bool *) palloc(sizeof(bool) * nkeycolumns);
297 407924 : info->nulls_first = (bool *) palloc(sizeof(bool) * nkeycolumns);
298 :
299 1033352 : for (i = 0; i < nkeycolumns; i++)
300 : {
301 625428 : int16 opt = indexRelation->rd_indoption[i];
302 :
303 625428 : info->reverse_sort[i] = (opt & INDOPTION_DESC) != 0;
304 625428 : info->nulls_first[i] = (opt & INDOPTION_NULLS_FIRST) != 0;
305 : }
306 : }
307 9248 : else if (amroutine->amcanorder)
308 : {
309 : /*
310 : * Otherwise, identify the corresponding btree opfamilies by
311 : * trying to map this index's "<" operators into btree. Since
312 : * "<" uniquely defines the behavior of a sort order, this is
313 : * a sufficient test.
314 : *
315 : * XXX This method is rather slow and also requires the
316 : * undesirable assumption that the other index AM numbers its
317 : * strategies the same as btree. It'd be better to have a way
318 : * to explicitly declare the corresponding btree opfamily for
319 : * each opfamily of the other index type. But given the lack
320 : * of current or foreseeable amcanorder index types, it's not
321 : * worth expending more effort on now.
322 : */
323 0 : info->sortopfamily = (Oid *) palloc(sizeof(Oid) * nkeycolumns);
324 0 : info->reverse_sort = (bool *) palloc(sizeof(bool) * nkeycolumns);
325 0 : info->nulls_first = (bool *) palloc(sizeof(bool) * nkeycolumns);
326 :
327 0 : for (i = 0; i < nkeycolumns; i++)
328 : {
329 0 : int16 opt = indexRelation->rd_indoption[i];
330 : Oid ltopr;
331 : Oid btopfamily;
332 : Oid btopcintype;
333 : int16 btstrategy;
334 :
335 0 : info->reverse_sort[i] = (opt & INDOPTION_DESC) != 0;
336 0 : info->nulls_first[i] = (opt & INDOPTION_NULLS_FIRST) != 0;
337 :
338 0 : ltopr = get_opfamily_member(info->opfamily[i],
339 0 : info->opcintype[i],
340 0 : info->opcintype[i],
341 : BTLessStrategyNumber);
342 0 : if (OidIsValid(ltopr) &&
343 0 : get_ordering_op_properties(ltopr,
344 : &btopfamily,
345 : &btopcintype,
346 0 : &btstrategy) &&
347 0 : btopcintype == info->opcintype[i] &&
348 0 : btstrategy == BTLessStrategyNumber)
349 : {
350 : /* Successful mapping */
351 0 : info->sortopfamily[i] = btopfamily;
352 : }
353 : else
354 : {
355 : /* Fail ... quietly treat index as unordered */
356 0 : info->sortopfamily = NULL;
357 0 : info->reverse_sort = NULL;
358 0 : info->nulls_first = NULL;
359 0 : break;
360 : }
361 : }
362 : }
363 : else
364 : {
365 9248 : info->sortopfamily = NULL;
366 9248 : info->reverse_sort = NULL;
367 9248 : info->nulls_first = NULL;
368 : }
369 :
370 : /*
371 : * Fetch the index expressions and predicate, if any. We must
372 : * modify the copies we obtain from the relcache to have the
373 : * correct varno for the parent relation, so that they match up
374 : * correctly against qual clauses.
375 : */
376 417172 : info->indexprs = RelationGetIndexExpressions(indexRelation);
377 417172 : info->indpred = RelationGetIndexPredicate(indexRelation);
378 417172 : if (info->indexprs && varno != 1)
379 1056 : ChangeVarNodes((Node *) info->indexprs, 1, varno, 0);
380 417172 : if (info->indpred && varno != 1)
381 84 : ChangeVarNodes((Node *) info->indpred, 1, varno, 0);
382 :
383 : /* Build targetlist using the completed indexprs data */
384 417172 : info->indextlist = build_index_tlist(root, info, relation);
385 :
386 417172 : info->indrestrictinfo = NIL; /* set later, in indxpath.c */
387 417172 : info->predOK = false; /* set later, in indxpath.c */
388 417172 : info->unique = index->indisunique;
389 417172 : info->immediate = index->indimmediate;
390 417172 : info->hypothetical = false;
391 :
392 : /*
393 : * Estimate the index size. If it's not a partial index, we lock
394 : * the number-of-tuples estimate to equal the parent table; if it
395 : * is partial then we have to use the same methods as we would for
396 : * a table, except we can be sure that the index is not larger
397 : * than the table.
398 : */
399 417172 : if (info->indpred == NIL)
400 : {
401 416492 : info->pages = RelationGetNumberOfBlocks(indexRelation);
402 416492 : info->tuples = rel->tuples;
403 : }
404 : else
405 : {
406 : double allvisfrac; /* dummy */
407 :
408 680 : estimate_rel_size(indexRelation, NULL,
409 : &info->pages, &info->tuples, &allvisfrac);
410 680 : if (info->tuples > rel->tuples)
411 0 : info->tuples = rel->tuples;
412 : }
413 :
414 417172 : if (info->relam == BTREE_AM_OID)
415 : {
416 : /* For btrees, get tree height while we have the index open */
417 407924 : info->tree_height = _bt_getrootheight(indexRelation);
418 : }
419 : else
420 : {
421 : /* For other index types, just set it to "unknown" for now */
422 9248 : info->tree_height = -1;
423 : }
424 :
425 417172 : index_close(indexRelation, NoLock);
426 :
427 : /*
428 : * We've historically used lcons() here. It'd make more sense to
429 : * use lappend(), but that causes the planner to change behavior
430 : * in cases where two indexes seem equally attractive. For now,
431 : * stick with lcons() --- few tables should have so many indexes
432 : * that the O(N^2) behavior of lcons() is really a problem.
433 : */
434 417172 : indexinfos = lcons(info, indexinfos);
435 : }
436 :
437 206790 : list_free(indexoidlist);
438 : }
439 :
440 282604 : rel->indexlist = indexinfos;
441 :
442 282604 : rel->statlist = get_relation_statistics(rel, relation);
443 :
444 : /* Grab foreign-table info using the relcache, while we have it */
445 282604 : if (relation->rd_rel->relkind == RELKIND_FOREIGN_TABLE)
446 : {
447 1846 : rel->serverid = GetForeignServerIdByRelId(RelationGetRelid(relation));
448 1846 : rel->fdwroutine = GetFdwRoutineForRelation(relation, true);
449 : }
450 : else
451 : {
452 280758 : rel->serverid = InvalidOid;
453 280758 : rel->fdwroutine = NULL;
454 : }
455 :
456 : /* Collect info about relation's foreign keys, if relevant */
457 282594 : get_relation_foreign_keys(root, rel, relation, inhparent);
458 :
459 : /*
460 : * Collect info about relation's partitioning scheme, if any. Only
461 : * inheritance parents may be partitioned.
462 : */
463 282594 : if (inhparent && relation->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
464 9760 : set_relation_partition_info(root, rel, relation);
465 :
466 282594 : table_close(relation, NoLock);
467 :
468 : /*
469 : * Allow a plugin to editorialize on the info we obtained from the
470 : * catalogs. Actions might include altering the assumed relation size,
471 : * removing an index, or adding a hypothetical index to the indexlist.
472 : */
473 282594 : if (get_relation_info_hook)
474 0 : (*get_relation_info_hook) (root, relationObjectId, inhparent, rel);
475 282594 : }
476 :
477 : /*
478 : * get_relation_foreign_keys -
479 : * Retrieves foreign key information for a given relation.
480 : *
481 : * ForeignKeyOptInfos for relevant foreign keys are created and added to
482 : * root->fkey_list. We do this now while we have the relcache entry open.
483 : * We could sometimes avoid making useless ForeignKeyOptInfos if we waited
484 : * until all RelOptInfos have been built, but the cost of re-opening the
485 : * relcache entries would probably exceed any savings.
486 : */
487 : static void
488 282594 : get_relation_foreign_keys(PlannerInfo *root, RelOptInfo *rel,
489 : Relation relation, bool inhparent)
490 : {
491 282594 : List *rtable = root->parse->rtable;
492 : List *cachedfkeys;
493 : ListCell *lc;
494 :
495 : /*
496 : * If it's not a baserel, we don't care about its FKs. Also, if the query
497 : * references only a single relation, we can skip the lookup since no FKs
498 : * could satisfy the requirements below.
499 : */
500 540888 : if (rel->reloptkind != RELOPT_BASEREL ||
501 258294 : list_length(rtable) < 2)
502 151826 : return;
503 :
504 : /*
505 : * If it's the parent of an inheritance tree, ignore its FKs. We could
506 : * make useful FK-based deductions if we found that all members of the
507 : * inheritance tree have equivalent FK constraints, but detecting that
508 : * would require code that hasn't been written.
509 : */
510 130768 : if (inhparent)
511 2866 : return;
512 :
513 : /*
514 : * Extract data about relation's FKs from the relcache. Note that this
515 : * list belongs to the relcache and might disappear in a cache flush, so
516 : * we must not do any further catalog access within this function.
517 : */
518 127902 : cachedfkeys = RelationGetFKeyList(relation);
519 :
520 : /*
521 : * Figure out which FKs are of interest for this query, and create
522 : * ForeignKeyOptInfos for them. We want only FKs that reference some
523 : * other RTE of the current query. In queries containing self-joins,
524 : * there might be more than one other RTE for a referenced table, and we
525 : * should make a ForeignKeyOptInfo for each occurrence.
526 : *
527 : * Ideally, we would ignore RTEs that correspond to non-baserels, but it's
528 : * too hard to identify those here, so we might end up making some useless
529 : * ForeignKeyOptInfos. If so, match_foreign_keys_to_quals() will remove
530 : * them again.
531 : */
532 129564 : foreach(lc, cachedfkeys)
533 : {
534 1662 : ForeignKeyCacheInfo *cachedfk = (ForeignKeyCacheInfo *) lfirst(lc);
535 : Index rti;
536 : ListCell *lc2;
537 :
538 : /* conrelid should always be that of the table we're considering */
539 : Assert(cachedfk->conrelid == RelationGetRelid(relation));
540 :
541 : /* Scan to find other RTEs matching confrelid */
542 1662 : rti = 0;
543 7520 : foreach(lc2, rtable)
544 : {
545 5858 : RangeTblEntry *rte = (RangeTblEntry *) lfirst(lc2);
546 : ForeignKeyOptInfo *info;
547 :
548 5858 : rti++;
549 : /* Ignore if not the correct table */
550 5858 : if (rte->rtekind != RTE_RELATION ||
551 4152 : rte->relid != cachedfk->confrelid)
552 4612 : continue;
553 : /* Ignore if it's an inheritance parent; doesn't really match */
554 1246 : if (rte->inh)
555 102 : continue;
556 : /* Ignore self-referential FKs; we only care about joins */
557 1144 : if (rti == rel->relid)
558 40 : continue;
559 :
560 : /* OK, let's make an entry */
561 1104 : info = makeNode(ForeignKeyOptInfo);
562 1104 : info->con_relid = rel->relid;
563 1104 : info->ref_relid = rti;
564 1104 : info->nkeys = cachedfk->nkeys;
565 1104 : memcpy(info->conkey, cachedfk->conkey, sizeof(info->conkey));
566 1104 : memcpy(info->confkey, cachedfk->confkey, sizeof(info->confkey));
567 1104 : memcpy(info->conpfeqop, cachedfk->conpfeqop, sizeof(info->conpfeqop));
568 : /* zero out fields to be filled by match_foreign_keys_to_quals */
569 1104 : info->nmatched_ec = 0;
570 1104 : info->nmatched_rcols = 0;
571 1104 : info->nmatched_ri = 0;
572 1104 : memset(info->eclass, 0, sizeof(info->eclass));
573 1104 : memset(info->rinfos, 0, sizeof(info->rinfos));
574 :
575 1104 : root->fkey_list = lappend(root->fkey_list, info);
576 : }
577 : }
578 : }
579 :
580 : /*
581 : * infer_arbiter_indexes -
582 : * Determine the unique indexes used to arbitrate speculative insertion.
583 : *
584 : * Uses user-supplied inference clause expressions and predicate to match a
585 : * unique index from those defined and ready on the heap relation (target).
586 : * An exact match is required on columns/expressions (although they can appear
587 : * in any order). However, the predicate given by the user need only restrict
588 : * insertion to a subset of some part of the table covered by some particular
589 : * unique index (in particular, a partial unique index) in order to be
590 : * inferred.
591 : *
592 : * The implementation does not consider which B-Tree operator class any
593 : * particular available unique index attribute uses, unless one was specified
594 : * in the inference specification. The same is true of collations. In
595 : * particular, there is no system dependency on the default operator class for
596 : * the purposes of inference. If no opclass (or collation) is specified, then
597 : * all matching indexes (that may or may not match the default in terms of
598 : * each attribute opclass/collation) are used for inference.
599 : */
600 : List *
601 984 : infer_arbiter_indexes(PlannerInfo *root)
602 : {
603 984 : OnConflictExpr *onconflict = root->parse->onConflict;
604 :
605 : /* Iteration state */
606 : RangeTblEntry *rte;
607 : Relation relation;
608 984 : Oid indexOidFromConstraint = InvalidOid;
609 : List *indexList;
610 : ListCell *l;
611 :
612 : /* Normalized inference attributes and inference expressions: */
613 984 : Bitmapset *inferAttrs = NULL;
614 984 : List *inferElems = NIL;
615 :
616 : /* Results */
617 984 : List *results = NIL;
618 :
619 : /*
620 : * Quickly return NIL for ON CONFLICT DO NOTHING without an inference
621 : * specification or named constraint. ON CONFLICT DO UPDATE statements
622 : * must always provide one or the other (but parser ought to have caught
623 : * that already).
624 : */
625 984 : if (onconflict->arbiterElems == NIL &&
626 156 : onconflict->constraint == InvalidOid)
627 124 : return NIL;
628 :
629 : /*
630 : * We need not lock the relation since it was already locked, either by
631 : * the rewriter or when expand_inherited_rtentry() added it to the query's
632 : * rangetable.
633 : */
634 860 : rte = rt_fetch(root->parse->resultRelation, root->parse->rtable);
635 :
636 860 : relation = table_open(rte->relid, NoLock);
637 :
638 : /*
639 : * Build normalized/BMS representation of plain indexed attributes, as
640 : * well as a separate list of expression items. This simplifies matching
641 : * the cataloged definition of indexes.
642 : */
643 1884 : foreach(l, onconflict->arbiterElems)
644 : {
645 1024 : InferenceElem *elem = (InferenceElem *) lfirst(l);
646 : Var *var;
647 : int attno;
648 :
649 1024 : if (!IsA(elem->expr, Var))
650 : {
651 : /* If not a plain Var, just shove it in inferElems for now */
652 116 : inferElems = lappend(inferElems, elem->expr);
653 116 : continue;
654 : }
655 :
656 908 : var = (Var *) elem->expr;
657 908 : attno = var->varattno;
658 :
659 908 : if (attno == 0)
660 0 : ereport(ERROR,
661 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
662 : errmsg("whole row unique index inference specifications are not supported")));
663 :
664 908 : inferAttrs = bms_add_member(inferAttrs,
665 : attno - FirstLowInvalidHeapAttributeNumber);
666 : }
667 :
668 : /*
669 : * Lookup named constraint's index. This is not immediately returned
670 : * because some additional sanity checks are required.
671 : */
672 860 : if (onconflict->constraint != InvalidOid)
673 : {
674 32 : indexOidFromConstraint = get_constraint_index(onconflict->constraint);
675 :
676 32 : if (indexOidFromConstraint == InvalidOid)
677 0 : ereport(ERROR,
678 : (errcode(ERRCODE_WRONG_OBJECT_TYPE),
679 : errmsg("constraint in ON CONFLICT clause has no associated index")));
680 : }
681 :
682 : /*
683 : * Using that representation, iterate through the list of indexes on the
684 : * target relation to try and find a match
685 : */
686 860 : indexList = RelationGetIndexList(relation);
687 :
688 1960 : foreach(l, indexList)
689 : {
690 1132 : Oid indexoid = lfirst_oid(l);
691 : Relation idxRel;
692 : Form_pg_index idxForm;
693 : Bitmapset *indexedAttrs;
694 : List *idxExprs;
695 : List *predExprs;
696 : AttrNumber natt;
697 : ListCell *el;
698 :
699 : /*
700 : * Extract info from the relation descriptor for the index. Obtain
701 : * the same lock type that the executor will ultimately use.
702 : *
703 : * Let executor complain about !indimmediate case directly, because
704 : * enforcement needs to occur there anyway when an inference clause is
705 : * omitted.
706 : */
707 1132 : idxRel = index_open(indexoid, rte->rellockmode);
708 1132 : idxForm = idxRel->rd_index;
709 :
710 1132 : if (!idxForm->indisvalid)
711 4 : goto next;
712 :
713 : /*
714 : * Note that we do not perform a check against indcheckxmin (like e.g.
715 : * get_relation_info()) here to eliminate candidates, because
716 : * uniqueness checking only cares about the most recently committed
717 : * tuple versions.
718 : */
719 :
720 : /*
721 : * Look for match on "ON constraint_name" variant, which may not be
722 : * unique constraint. This can only be a constraint name.
723 : */
724 1128 : if (indexOidFromConstraint == idxForm->indexrelid)
725 : {
726 32 : if (!idxForm->indisunique && onconflict->action == ONCONFLICT_UPDATE)
727 4 : ereport(ERROR,
728 : (errcode(ERRCODE_WRONG_OBJECT_TYPE),
729 : errmsg("ON CONFLICT DO UPDATE not supported with exclusion constraints")));
730 :
731 28 : results = lappend_oid(results, idxForm->indexrelid);
732 28 : list_free(indexList);
733 28 : index_close(idxRel, NoLock);
734 28 : table_close(relation, NoLock);
735 28 : return results;
736 : }
737 1096 : else if (indexOidFromConstraint != InvalidOid)
738 : {
739 : /* No point in further work for index in named constraint case */
740 12 : goto next;
741 : }
742 :
743 : /*
744 : * Only considering conventional inference at this point (not named
745 : * constraints), so index under consideration can be immediately
746 : * skipped if it's not unique
747 : */
748 1084 : if (!idxForm->indisunique)
749 4 : goto next;
750 :
751 : /* Build BMS representation of plain (non expression) index attrs */
752 1080 : indexedAttrs = NULL;
753 2520 : for (natt = 0; natt < idxForm->indnkeyatts; natt++)
754 : {
755 1440 : int attno = idxRel->rd_index->indkey.values[natt];
756 :
757 1440 : if (attno != 0)
758 1228 : indexedAttrs = bms_add_member(indexedAttrs,
759 : attno - FirstLowInvalidHeapAttributeNumber);
760 : }
761 :
762 : /* Non-expression attributes (if any) must match */
763 1080 : if (!bms_equal(indexedAttrs, inferAttrs))
764 240 : goto next;
765 :
766 : /* Expression attributes (if any) must match */
767 840 : idxExprs = RelationGetIndexExpressions(idxRel);
768 1908 : foreach(el, onconflict->arbiterElems)
769 : {
770 1100 : InferenceElem *elem = (InferenceElem *) lfirst(el);
771 :
772 : /*
773 : * Ensure that collation/opclass aspects of inference expression
774 : * element match. Even though this loop is primarily concerned
775 : * with matching expressions, it is a convenient point to check
776 : * this for both expressions and ordinary (non-expression)
777 : * attributes appearing as inference elements.
778 : */
779 1100 : if (!infer_collation_opclass_match(elem, idxRel, idxExprs))
780 32 : goto next;
781 :
782 : /*
783 : * Plain Vars don't factor into count of expression elements, and
784 : * the question of whether or not they satisfy the index
785 : * definition has already been considered (they must).
786 : */
787 1076 : if (IsA(elem->expr, Var))
788 960 : continue;
789 :
790 : /*
791 : * Might as well avoid redundant check in the rare cases where
792 : * infer_collation_opclass_match() is required to do real work.
793 : * Otherwise, check that element expression appears in cataloged
794 : * index definition.
795 : */
796 116 : if (elem->infercollid != InvalidOid ||
797 204 : elem->inferopclass != InvalidOid ||
798 100 : list_member(idxExprs, elem->expr))
799 108 : continue;
800 :
801 8 : goto next;
802 : }
803 :
804 : /*
805 : * Now that all inference elements were matched, ensure that the
806 : * expression elements from inference clause are not missing any
807 : * cataloged expressions. This does the right thing when unique
808 : * indexes redundantly repeat the same attribute, or if attributes
809 : * redundantly appear multiple times within an inference clause.
810 : */
811 808 : if (list_difference(idxExprs, inferElems) != NIL)
812 36 : goto next;
813 :
814 : /*
815 : * If it's a partial index, its predicate must be implied by the ON
816 : * CONFLICT's WHERE clause.
817 : */
818 772 : predExprs = RelationGetIndexPredicate(idxRel);
819 :
820 772 : if (!predicate_implied_by(predExprs, (List *) onconflict->arbiterWhere, false))
821 24 : goto next;
822 :
823 748 : results = lappend_oid(results, idxForm->indexrelid);
824 1100 : next:
825 1100 : index_close(idxRel, NoLock);
826 : }
827 :
828 828 : list_free(indexList);
829 828 : table_close(relation, NoLock);
830 :
831 828 : if (results == NIL)
832 116 : ereport(ERROR,
833 : (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
834 : errmsg("there is no unique or exclusion constraint matching the ON CONFLICT specification")));
835 :
836 712 : return results;
837 : }
838 :
839 : /*
840 : * infer_collation_opclass_match - ensure infer element opclass/collation match
841 : *
842 : * Given unique index inference element from inference specification, if
843 : * collation was specified, or if opclass was specified, verify that there is
844 : * at least one matching indexed attribute (occasionally, there may be more).
845 : * Skip this in the common case where inference specification does not include
846 : * collation or opclass (instead matching everything, regardless of cataloged
847 : * collation/opclass of indexed attribute).
848 : *
849 : * At least historically, Postgres has not offered collations or opclasses
850 : * with alternative-to-default notions of equality, so these additional
851 : * criteria should only be required infrequently.
852 : *
853 : * Don't give up immediately when an inference element matches some attribute
854 : * cataloged as indexed but not matching additional opclass/collation
855 : * criteria. This is done so that the implementation is as forgiving as
856 : * possible of redundancy within cataloged index attributes (or, less
857 : * usefully, within inference specification elements). If collations actually
858 : * differ between apparently redundantly indexed attributes (redundant within
859 : * or across indexes), then there really is no redundancy as such.
860 : *
861 : * Note that if an inference element specifies an opclass and a collation at
862 : * once, both must match in at least one particular attribute within index
863 : * catalog definition in order for that inference element to be considered
864 : * inferred/satisfied.
865 : */
866 : static bool
867 1100 : infer_collation_opclass_match(InferenceElem *elem, Relation idxRel,
868 : List *idxExprs)
869 : {
870 : AttrNumber natt;
871 1100 : Oid inferopfamily = InvalidOid; /* OID of opclass opfamily */
872 1100 : Oid inferopcinputtype = InvalidOid; /* OID of opclass input type */
873 1100 : int nplain = 0; /* # plain attrs observed */
874 :
875 : /*
876 : * If inference specification element lacks collation/opclass, then no
877 : * need to check for exact match.
878 : */
879 1100 : if (elem->infercollid == InvalidOid && elem->inferopclass == InvalidOid)
880 1024 : return true;
881 :
882 : /*
883 : * Lookup opfamily and input type, for matching indexes
884 : */
885 76 : if (elem->inferopclass)
886 : {
887 56 : inferopfamily = get_opclass_family(elem->inferopclass);
888 56 : inferopcinputtype = get_opclass_input_type(elem->inferopclass);
889 : }
890 :
891 164 : for (natt = 1; natt <= idxRel->rd_att->natts; natt++)
892 : {
893 140 : Oid opfamily = idxRel->rd_opfamily[natt - 1];
894 140 : Oid opcinputtype = idxRel->rd_opcintype[natt - 1];
895 140 : Oid collation = idxRel->rd_indcollation[natt - 1];
896 140 : int attno = idxRel->rd_index->indkey.values[natt - 1];
897 :
898 140 : if (attno != 0)
899 112 : nplain++;
900 :
901 140 : if (elem->inferopclass != InvalidOid &&
902 44 : (inferopfamily != opfamily || inferopcinputtype != opcinputtype))
903 : {
904 : /* Attribute needed to match opclass, but didn't */
905 60 : continue;
906 : }
907 :
908 80 : if (elem->infercollid != InvalidOid &&
909 56 : elem->infercollid != collation)
910 : {
911 : /* Attribute needed to match collation, but didn't */
912 24 : continue;
913 : }
914 :
915 : /* If one matching index att found, good enough -- return true */
916 56 : if (IsA(elem->expr, Var))
917 : {
918 36 : if (((Var *) elem->expr)->varattno == attno)
919 36 : return true;
920 : }
921 20 : else if (attno == 0)
922 : {
923 20 : Node *nattExpr = list_nth(idxExprs, (natt - 1) - nplain);
924 :
925 : /*
926 : * Note that unlike routines like match_index_to_operand() we
927 : * don't need to care about RelabelType. Neither the index
928 : * definition nor the inference clause should contain them.
929 : */
930 20 : if (equal(elem->expr, nattExpr))
931 16 : return true;
932 : }
933 : }
934 :
935 24 : return false;
936 : }
937 :
938 : /*
939 : * estimate_rel_size - estimate # pages and # tuples in a table or index
940 : *
941 : * We also estimate the fraction of the pages that are marked all-visible in
942 : * the visibility map, for use in estimation of index-only scans.
943 : *
944 : * If attr_widths isn't NULL, it points to the zero-index entry of the
945 : * relation's attr_widths[] cache; we fill this in if we have need to compute
946 : * the attribute widths for estimation purposes.
947 : */
948 : void
949 270820 : estimate_rel_size(Relation rel, int32 *attr_widths,
950 : BlockNumber *pages, double *tuples, double *allvisfrac)
951 : {
952 : BlockNumber curpages;
953 : BlockNumber relpages;
954 : double reltuples;
955 : BlockNumber relallvisible;
956 : double density;
957 :
958 270820 : switch (rel->rd_rel->relkind)
959 : {
960 267848 : case RELKIND_RELATION:
961 : case RELKIND_MATVIEW:
962 : case RELKIND_TOASTVALUE:
963 267848 : table_relation_estimate_size(rel, attr_widths, pages, tuples,
964 : allvisfrac);
965 267848 : break;
966 :
967 680 : case RELKIND_INDEX:
968 :
969 : /*
970 : * XXX: It'd probably be good to move this into a callback,
971 : * individual index types e.g. know if they have a metapage.
972 : */
973 :
974 : /* it has storage, ok to call the smgr */
975 680 : curpages = RelationGetNumberOfBlocks(rel);
976 :
977 : /* coerce values in pg_class to more desirable types */
978 680 : relpages = (BlockNumber) rel->rd_rel->relpages;
979 680 : reltuples = (double) rel->rd_rel->reltuples;
980 680 : relallvisible = (BlockNumber) rel->rd_rel->relallvisible;
981 :
982 : /* report estimated # pages */
983 680 : *pages = curpages;
984 : /* quick exit if rel is clearly empty */
985 680 : if (curpages == 0)
986 : {
987 0 : *tuples = 0;
988 0 : *allvisfrac = 0;
989 0 : break;
990 : }
991 : /* coerce values in pg_class to more desirable types */
992 680 : relpages = (BlockNumber) rel->rd_rel->relpages;
993 680 : reltuples = (double) rel->rd_rel->reltuples;
994 680 : relallvisible = (BlockNumber) rel->rd_rel->relallvisible;
995 :
996 : /*
997 : * Discount the metapage while estimating the number of tuples.
998 : * This is a kluge because it assumes more than it ought to about
999 : * index structure. Currently it's OK for btree, hash, and GIN
1000 : * indexes but suspect for GiST indexes.
1001 : */
1002 680 : if (relpages > 0)
1003 : {
1004 680 : curpages--;
1005 680 : relpages--;
1006 : }
1007 :
1008 : /* estimate number of tuples from previous tuple density */
1009 680 : if (relpages > 0)
1010 412 : density = reltuples / (double) relpages;
1011 : else
1012 : {
1013 : /*
1014 : * When we have no data because the relation was truncated,
1015 : * estimate tuple width from attribute datatypes. We assume
1016 : * here that the pages are completely full, which is OK for
1017 : * tables (since they've presumably not been VACUUMed yet) but
1018 : * is probably an overestimate for indexes. Fortunately
1019 : * get_relation_info() can clamp the overestimate to the
1020 : * parent table's size.
1021 : *
1022 : * Note: this code intentionally disregards alignment
1023 : * considerations, because (a) that would be gilding the lily
1024 : * considering how crude the estimate is, and (b) it creates
1025 : * platform dependencies in the default plans which are kind
1026 : * of a headache for regression testing.
1027 : *
1028 : * XXX: Should this logic be more index specific?
1029 : */
1030 : int32 tuple_width;
1031 :
1032 268 : tuple_width = get_rel_data_width(rel, attr_widths);
1033 268 : tuple_width += MAXALIGN(SizeofHeapTupleHeader);
1034 268 : tuple_width += sizeof(ItemIdData);
1035 : /* note: integer division is intentional here */
1036 268 : density = (BLCKSZ - SizeOfPageHeaderData) / tuple_width;
1037 : }
1038 680 : *tuples = rint(density * (double) curpages);
1039 :
1040 : /*
1041 : * We use relallvisible as-is, rather than scaling it up like we
1042 : * do for the pages and tuples counts, on the theory that any
1043 : * pages added since the last VACUUM are most likely not marked
1044 : * all-visible. But costsize.c wants it converted to a fraction.
1045 : */
1046 680 : if (relallvisible == 0 || curpages <= 0)
1047 680 : *allvisfrac = 0;
1048 0 : else if ((double) relallvisible >= curpages)
1049 0 : *allvisfrac = 1;
1050 : else
1051 0 : *allvisfrac = (double) relallvisible / curpages;
1052 680 : break;
1053 :
1054 416 : case RELKIND_SEQUENCE:
1055 : /* Sequences always have a known size */
1056 416 : *pages = 1;
1057 416 : *tuples = 1;
1058 416 : *allvisfrac = 0;
1059 416 : break;
1060 1846 : case RELKIND_FOREIGN_TABLE:
1061 : /* Just use whatever's in pg_class */
1062 1846 : *pages = rel->rd_rel->relpages;
1063 1846 : *tuples = rel->rd_rel->reltuples;
1064 1846 : *allvisfrac = 0;
1065 1846 : break;
1066 30 : default:
1067 : /* else it has no disk storage; probably shouldn't get here? */
1068 30 : *pages = 0;
1069 30 : *tuples = 0;
1070 30 : *allvisfrac = 0;
1071 30 : break;
1072 : }
1073 270820 : }
1074 :
1075 :
1076 : /*
1077 : * get_rel_data_width
1078 : *
1079 : * Estimate the average width of (the data part of) the relation's tuples.
1080 : *
1081 : * If attr_widths isn't NULL, it points to the zero-index entry of the
1082 : * relation's attr_widths[] cache; use and update that cache as appropriate.
1083 : *
1084 : * Currently we ignore dropped columns. Ideally those should be included
1085 : * in the result, but we haven't got any way to get info about them; and
1086 : * since they might be mostly NULLs, treating them as zero-width is not
1087 : * necessarily the wrong thing anyway.
1088 : */
1089 : int32
1090 103042 : get_rel_data_width(Relation rel, int32 *attr_widths)
1091 : {
1092 103042 : int32 tuple_width = 0;
1093 : int i;
1094 :
1095 491230 : for (i = 1; i <= RelationGetNumberOfAttributes(rel); i++)
1096 : {
1097 388188 : Form_pg_attribute att = TupleDescAttr(rel->rd_att, i - 1);
1098 : int32 item_width;
1099 :
1100 388188 : if (att->attisdropped)
1101 1648 : continue;
1102 :
1103 : /* use previously cached data, if any */
1104 386540 : if (attr_widths != NULL && attr_widths[i] > 0)
1105 : {
1106 4152 : tuple_width += attr_widths[i];
1107 4152 : continue;
1108 : }
1109 :
1110 : /* This should match set_rel_width() in costsize.c */
1111 382388 : item_width = get_attavgwidth(RelationGetRelid(rel), i);
1112 382388 : if (item_width <= 0)
1113 : {
1114 381518 : item_width = get_typavgwidth(att->atttypid, att->atttypmod);
1115 : Assert(item_width > 0);
1116 : }
1117 382388 : if (attr_widths != NULL)
1118 325976 : attr_widths[i] = item_width;
1119 382388 : tuple_width += item_width;
1120 : }
1121 :
1122 103042 : return tuple_width;
1123 : }
1124 :
1125 : /*
1126 : * get_relation_data_width
1127 : *
1128 : * External API for get_rel_data_width: same behavior except we have to
1129 : * open the relcache entry.
1130 : */
1131 : int32
1132 1382 : get_relation_data_width(Oid relid, int32 *attr_widths)
1133 : {
1134 : int32 result;
1135 : Relation relation;
1136 :
1137 : /* As above, assume relation is already locked */
1138 1382 : relation = table_open(relid, NoLock);
1139 :
1140 1382 : result = get_rel_data_width(relation, attr_widths);
1141 :
1142 1382 : table_close(relation, NoLock);
1143 :
1144 1382 : return result;
1145 : }
1146 :
1147 :
1148 : /*
1149 : * get_relation_constraints
1150 : *
1151 : * Retrieve the applicable constraint expressions of the given relation.
1152 : *
1153 : * Returns a List (possibly empty) of constraint expressions. Each one
1154 : * has been canonicalized, and its Vars are changed to have the varno
1155 : * indicated by rel->relid. This allows the expressions to be easily
1156 : * compared to expressions taken from WHERE.
1157 : *
1158 : * If include_noinherit is true, it's okay to include constraints that
1159 : * are marked NO INHERIT.
1160 : *
1161 : * If include_notnull is true, "col IS NOT NULL" expressions are generated
1162 : * and added to the result for each column that's marked attnotnull.
1163 : *
1164 : * If include_partition is true, and the relation is a partition,
1165 : * also include the partitioning constraints.
1166 : *
1167 : * Note: at present this is invoked at most once per relation per planner
1168 : * run, and in many cases it won't be invoked at all, so there seems no
1169 : * point in caching the data in RelOptInfo.
1170 : */
1171 : static List *
1172 14876 : get_relation_constraints(PlannerInfo *root,
1173 : Oid relationObjectId, RelOptInfo *rel,
1174 : bool include_noinherit,
1175 : bool include_notnull,
1176 : bool include_partition)
1177 : {
1178 14876 : List *result = NIL;
1179 14876 : Index varno = rel->relid;
1180 : Relation relation;
1181 : TupleConstr *constr;
1182 :
1183 : /*
1184 : * We assume the relation has already been safely locked.
1185 : */
1186 14876 : relation = table_open(relationObjectId, NoLock);
1187 :
1188 14876 : constr = relation->rd_att->constr;
1189 14876 : if (constr != NULL)
1190 : {
1191 6340 : int num_check = constr->num_check;
1192 : int i;
1193 :
1194 6796 : for (i = 0; i < num_check; i++)
1195 : {
1196 : Node *cexpr;
1197 :
1198 : /*
1199 : * If this constraint hasn't been fully validated yet, we must
1200 : * ignore it here. Also ignore if NO INHERIT and we weren't told
1201 : * that that's safe.
1202 : */
1203 456 : if (!constr->check[i].ccvalid)
1204 28 : continue;
1205 428 : if (constr->check[i].ccnoinherit && !include_noinherit)
1206 0 : continue;
1207 :
1208 428 : cexpr = stringToNode(constr->check[i].ccbin);
1209 :
1210 : /*
1211 : * Run each expression through const-simplification and
1212 : * canonicalization. This is not just an optimization, but is
1213 : * necessary, because we will be comparing it to
1214 : * similarly-processed qual clauses, and may fail to detect valid
1215 : * matches without this. This must match the processing done to
1216 : * qual clauses in preprocess_expression()! (We can skip the
1217 : * stuff involving subqueries, however, since we don't allow any
1218 : * in check constraints.)
1219 : */
1220 428 : cexpr = eval_const_expressions(root, cexpr);
1221 :
1222 428 : cexpr = (Node *) canonicalize_qual((Expr *) cexpr, true);
1223 :
1224 : /* Fix Vars to have the desired varno */
1225 428 : if (varno != 1)
1226 416 : ChangeVarNodes(cexpr, 1, varno, 0);
1227 :
1228 : /*
1229 : * Finally, convert to implicit-AND format (that is, a List) and
1230 : * append the resulting item(s) to our output list.
1231 : */
1232 428 : result = list_concat(result,
1233 428 : make_ands_implicit((Expr *) cexpr));
1234 : }
1235 :
1236 : /* Add NOT NULL constraints in expression form, if requested */
1237 6340 : if (include_notnull && constr->has_not_null)
1238 : {
1239 6024 : int natts = relation->rd_att->natts;
1240 :
1241 25734 : for (i = 1; i <= natts; i++)
1242 : {
1243 19710 : Form_pg_attribute att = TupleDescAttr(relation->rd_att, i - 1);
1244 :
1245 19710 : if (att->attnotnull && !att->attisdropped)
1246 : {
1247 7020 : NullTest *ntest = makeNode(NullTest);
1248 :
1249 7020 : ntest->arg = (Expr *) makeVar(varno,
1250 : i,
1251 : att->atttypid,
1252 : att->atttypmod,
1253 : att->attcollation,
1254 : 0);
1255 7020 : ntest->nulltesttype = IS_NOT_NULL;
1256 :
1257 : /*
1258 : * argisrow=false is correct even for a composite column,
1259 : * because attnotnull does not represent a SQL-spec IS NOT
1260 : * NULL test in such a case, just IS DISTINCT FROM NULL.
1261 : */
1262 7020 : ntest->argisrow = false;
1263 7020 : ntest->location = -1;
1264 7020 : result = lappend(result, ntest);
1265 : }
1266 : }
1267 : }
1268 : }
1269 :
1270 : /*
1271 : * Add partitioning constraints, if requested.
1272 : */
1273 14876 : if (include_partition && relation->rd_rel->relispartition)
1274 : {
1275 : /* make sure rel->partition_qual is set */
1276 8 : set_baserel_partition_constraint(relation, rel);
1277 8 : result = list_concat(result, rel->partition_qual);
1278 : }
1279 :
1280 14876 : table_close(relation, NoLock);
1281 :
1282 14876 : return result;
1283 : }
1284 :
1285 : /*
1286 : * get_relation_statistics
1287 : * Retrieve extended statistics defined on the table.
1288 : *
1289 : * Returns a List (possibly empty) of StatisticExtInfo objects describing
1290 : * the statistics. Note that this doesn't load the actual statistics data,
1291 : * just the identifying metadata. Only stats actually built are considered.
1292 : */
1293 : static List *
1294 282604 : get_relation_statistics(RelOptInfo *rel, Relation relation)
1295 : {
1296 : List *statoidlist;
1297 282604 : List *stainfos = NIL;
1298 : ListCell *l;
1299 :
1300 282604 : statoidlist = RelationGetStatExtList(relation);
1301 :
1302 283040 : foreach(l, statoidlist)
1303 : {
1304 436 : Oid statOid = lfirst_oid(l);
1305 : Form_pg_statistic_ext staForm;
1306 : HeapTuple htup;
1307 : HeapTuple dtup;
1308 436 : Bitmapset *keys = NULL;
1309 : int i;
1310 :
1311 436 : htup = SearchSysCache1(STATEXTOID, ObjectIdGetDatum(statOid));
1312 436 : if (!HeapTupleIsValid(htup))
1313 0 : elog(ERROR, "cache lookup failed for statistics object %u", statOid);
1314 436 : staForm = (Form_pg_statistic_ext) GETSTRUCT(htup);
1315 :
1316 436 : dtup = SearchSysCache1(STATEXTDATASTXOID, ObjectIdGetDatum(statOid));
1317 436 : if (!HeapTupleIsValid(dtup))
1318 0 : elog(ERROR, "cache lookup failed for statistics object %u", statOid);
1319 :
1320 : /*
1321 : * First, build the array of columns covered. This is ultimately
1322 : * wasted if no stats within the object have actually been built, but
1323 : * it doesn't seem worth troubling over that case.
1324 : */
1325 1648 : for (i = 0; i < staForm->stxkeys.dim1; i++)
1326 1212 : keys = bms_add_member(keys, staForm->stxkeys.values[i]);
1327 :
1328 : /* add one StatisticExtInfo for each kind built */
1329 436 : if (statext_is_kind_built(dtup, STATS_EXT_NDISTINCT))
1330 : {
1331 44 : StatisticExtInfo *info = makeNode(StatisticExtInfo);
1332 :
1333 44 : info->statOid = statOid;
1334 44 : info->rel = rel;
1335 44 : info->kind = STATS_EXT_NDISTINCT;
1336 44 : info->keys = bms_copy(keys);
1337 :
1338 44 : stainfos = lappend(stainfos, info);
1339 : }
1340 :
1341 436 : if (statext_is_kind_built(dtup, STATS_EXT_DEPENDENCIES))
1342 : {
1343 200 : StatisticExtInfo *info = makeNode(StatisticExtInfo);
1344 :
1345 200 : info->statOid = statOid;
1346 200 : info->rel = rel;
1347 200 : info->kind = STATS_EXT_DEPENDENCIES;
1348 200 : info->keys = bms_copy(keys);
1349 :
1350 200 : stainfos = lappend(stainfos, info);
1351 : }
1352 :
1353 436 : if (statext_is_kind_built(dtup, STATS_EXT_MCV))
1354 : {
1355 256 : StatisticExtInfo *info = makeNode(StatisticExtInfo);
1356 :
1357 256 : info->statOid = statOid;
1358 256 : info->rel = rel;
1359 256 : info->kind = STATS_EXT_MCV;
1360 256 : info->keys = bms_copy(keys);
1361 :
1362 256 : stainfos = lappend(stainfos, info);
1363 : }
1364 :
1365 436 : ReleaseSysCache(htup);
1366 436 : ReleaseSysCache(dtup);
1367 436 : bms_free(keys);
1368 : }
1369 :
1370 282604 : list_free(statoidlist);
1371 :
1372 282604 : return stainfos;
1373 : }
1374 :
1375 : /*
1376 : * relation_excluded_by_constraints
1377 : *
1378 : * Detect whether the relation need not be scanned because it has either
1379 : * self-inconsistent restrictions, or restrictions inconsistent with the
1380 : * relation's applicable constraints.
1381 : *
1382 : * Note: this examines only rel->relid, rel->reloptkind, and
1383 : * rel->baserestrictinfo; therefore it can be called before filling in
1384 : * other fields of the RelOptInfo.
1385 : */
1386 : bool
1387 292522 : relation_excluded_by_constraints(PlannerInfo *root,
1388 : RelOptInfo *rel, RangeTblEntry *rte)
1389 : {
1390 : bool include_noinherit;
1391 : bool include_notnull;
1392 292522 : bool include_partition = false;
1393 : List *safe_restrictions;
1394 : List *constraint_pred;
1395 : List *safe_constraints;
1396 : ListCell *lc;
1397 :
1398 : /* As of now, constraint exclusion works only with simple relations. */
1399 : Assert(IS_SIMPLE_REL(rel));
1400 :
1401 : /*
1402 : * If there are no base restriction clauses, we have no hope of proving
1403 : * anything below, so fall out quickly.
1404 : */
1405 292522 : if (rel->baserestrictinfo == NIL)
1406 125922 : return false;
1407 :
1408 : /*
1409 : * Regardless of the setting of constraint_exclusion, detect
1410 : * constant-FALSE-or-NULL restriction clauses. Because const-folding will
1411 : * reduce "anything AND FALSE" to just "FALSE", any such case should
1412 : * result in exactly one baserestrictinfo entry. This doesn't fire very
1413 : * often, but it seems cheap enough to be worth doing anyway. (Without
1414 : * this, we'd miss some optimizations that 9.5 and earlier found via much
1415 : * more roundabout methods.)
1416 : */
1417 166600 : if (list_length(rel->baserestrictinfo) == 1)
1418 : {
1419 120412 : RestrictInfo *rinfo = (RestrictInfo *) linitial(rel->baserestrictinfo);
1420 120412 : Expr *clause = rinfo->clause;
1421 :
1422 120412 : if (clause && IsA(clause, Const) &&
1423 198 : (((Const *) clause)->constisnull ||
1424 198 : !DatumGetBool(((Const *) clause)->constvalue)))
1425 198 : return true;
1426 : }
1427 :
1428 : /*
1429 : * Skip further tests, depending on constraint_exclusion.
1430 : */
1431 166402 : switch (constraint_exclusion)
1432 : {
1433 52 : case CONSTRAINT_EXCLUSION_OFF:
1434 : /* In 'off' mode, never make any further tests */
1435 52 : return false;
1436 :
1437 166254 : case CONSTRAINT_EXCLUSION_PARTITION:
1438 :
1439 : /*
1440 : * When constraint_exclusion is set to 'partition' we only handle
1441 : * appendrel members. Normally, they are RELOPT_OTHER_MEMBER_REL
1442 : * relations, but we also consider inherited target relations as
1443 : * appendrel members for the purposes of constraint exclusion
1444 : * (since, indeed, they were appendrel members earlier in
1445 : * inheritance_planner).
1446 : *
1447 : * In both cases, partition pruning was already applied, so there
1448 : * is no need to consider the rel's partition constraints here.
1449 : */
1450 166254 : if (rel->reloptkind == RELOPT_OTHER_MEMBER_REL ||
1451 153072 : (rel->relid == root->parse->resultRelation &&
1452 10236 : root->inhTargetKind != INHKIND_NONE))
1453 : break; /* appendrel member, so process it */
1454 151242 : return false;
1455 :
1456 96 : case CONSTRAINT_EXCLUSION_ON:
1457 :
1458 : /*
1459 : * In 'on' mode, always apply constraint exclusion. If we are
1460 : * considering a baserel that is a partition (i.e., it was
1461 : * directly named rather than expanded from a parent table), then
1462 : * its partition constraints haven't been considered yet, so
1463 : * include them in the processing here.
1464 : */
1465 96 : if (rel->reloptkind == RELOPT_BASEREL &&
1466 76 : !(rel->relid == root->parse->resultRelation &&
1467 24 : root->inhTargetKind != INHKIND_NONE))
1468 64 : include_partition = true;
1469 96 : break; /* always try to exclude */
1470 : }
1471 :
1472 : /*
1473 : * Check for self-contradictory restriction clauses. We dare not make
1474 : * deductions with non-immutable functions, but any immutable clauses that
1475 : * are self-contradictory allow us to conclude the scan is unnecessary.
1476 : *
1477 : * Note: strip off RestrictInfo because predicate_refuted_by() isn't
1478 : * expecting to see any in its predicate argument.
1479 : */
1480 15012 : safe_restrictions = NIL;
1481 35112 : foreach(lc, rel->baserestrictinfo)
1482 : {
1483 20004 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1484 :
1485 20004 : if (!contain_mutable_functions((Node *) rinfo->clause))
1486 19068 : safe_restrictions = lappend(safe_restrictions, rinfo->clause);
1487 : }
1488 :
1489 : /*
1490 : * We can use weak refutation here, since we're comparing restriction
1491 : * clauses with restriction clauses.
1492 : */
1493 15108 : if (predicate_refuted_by(safe_restrictions, safe_restrictions, true))
1494 60 : return true;
1495 :
1496 : /*
1497 : * Only plain relations have constraints, so stop here for other rtekinds.
1498 : */
1499 15048 : if (rte->rtekind != RTE_RELATION)
1500 172 : return false;
1501 :
1502 : /*
1503 : * If we are scanning just this table, we can use NO INHERIT constraints,
1504 : * but not if we're scanning its children too. (Note that partitioned
1505 : * tables should never have NO INHERIT constraints; but it's not necessary
1506 : * for us to assume that here.)
1507 : */
1508 14876 : include_noinherit = !rte->inh;
1509 :
1510 : /*
1511 : * Currently, attnotnull constraints must be treated as NO INHERIT unless
1512 : * this is a partitioned table. In future we might track their
1513 : * inheritance status more accurately, allowing this to be refined.
1514 : */
1515 14876 : include_notnull = (!rte->inh || rte->relkind == RELKIND_PARTITIONED_TABLE);
1516 :
1517 : /*
1518 : * Fetch the appropriate set of constraint expressions.
1519 : */
1520 14876 : constraint_pred = get_relation_constraints(root, rte->relid, rel,
1521 : include_noinherit,
1522 : include_notnull,
1523 : include_partition);
1524 :
1525 : /*
1526 : * We do not currently enforce that CHECK constraints contain only
1527 : * immutable functions, so it's necessary to check here. We daren't draw
1528 : * conclusions from plan-time evaluation of non-immutable functions. Since
1529 : * they're ANDed, we can just ignore any mutable constraints in the list,
1530 : * and reason about the rest.
1531 : */
1532 14876 : safe_constraints = NIL;
1533 22400 : foreach(lc, constraint_pred)
1534 : {
1535 7524 : Node *pred = (Node *) lfirst(lc);
1536 :
1537 7524 : if (!contain_mutable_functions(pred))
1538 7524 : safe_constraints = lappend(safe_constraints, pred);
1539 : }
1540 :
1541 : /*
1542 : * The constraints are effectively ANDed together, so we can just try to
1543 : * refute the entire collection at once. This may allow us to make proofs
1544 : * that would fail if we took them individually.
1545 : *
1546 : * Note: we use rel->baserestrictinfo, not safe_restrictions as might seem
1547 : * an obvious optimization. Some of the clauses might be OR clauses that
1548 : * have volatile and nonvolatile subclauses, and it's OK to make
1549 : * deductions with the nonvolatile parts.
1550 : *
1551 : * We need strong refutation because we have to prove that the constraints
1552 : * would yield false, not just NULL.
1553 : */
1554 14876 : if (predicate_refuted_by(safe_constraints, rel->baserestrictinfo, false))
1555 56 : return true;
1556 :
1557 14820 : return false;
1558 : }
1559 :
1560 :
1561 : /*
1562 : * build_physical_tlist
1563 : *
1564 : * Build a targetlist consisting of exactly the relation's user attributes,
1565 : * in order. The executor can special-case such tlists to avoid a projection
1566 : * step at runtime, so we use such tlists preferentially for scan nodes.
1567 : *
1568 : * Exception: if there are any dropped or missing columns, we punt and return
1569 : * NIL. Ideally we would like to handle these cases too. However this
1570 : * creates problems for ExecTypeFromTL, which may be asked to build a tupdesc
1571 : * for a tlist that includes vars of no-longer-existent types. In theory we
1572 : * could dig out the required info from the pg_attribute entries of the
1573 : * relation, but that data is not readily available to ExecTypeFromTL.
1574 : * For now, we don't apply the physical-tlist optimization when there are
1575 : * dropped cols.
1576 : *
1577 : * We also support building a "physical" tlist for subqueries, functions,
1578 : * values lists, table expressions, and CTEs, since the same optimization can
1579 : * occur in SubqueryScan, FunctionScan, ValuesScan, CteScan, TableFunc,
1580 : * NamedTuplestoreScan, and WorkTableScan nodes.
1581 : */
1582 : List *
1583 78244 : build_physical_tlist(PlannerInfo *root, RelOptInfo *rel)
1584 : {
1585 78244 : List *tlist = NIL;
1586 78244 : Index varno = rel->relid;
1587 78244 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
1588 : Relation relation;
1589 : Query *subquery;
1590 : Var *var;
1591 : ListCell *l;
1592 : int attrno,
1593 : numattrs;
1594 : List *colvars;
1595 :
1596 78244 : switch (rte->rtekind)
1597 : {
1598 68356 : case RTE_RELATION:
1599 : /* Assume we already have adequate lock */
1600 68356 : relation = table_open(rte->relid, NoLock);
1601 :
1602 68356 : numattrs = RelationGetNumberOfAttributes(relation);
1603 1265492 : for (attrno = 1; attrno <= numattrs; attrno++)
1604 : {
1605 1197272 : Form_pg_attribute att_tup = TupleDescAttr(relation->rd_att,
1606 : attrno - 1);
1607 :
1608 1197272 : if (att_tup->attisdropped || att_tup->atthasmissing)
1609 : {
1610 : /* found a dropped or missing col, so punt */
1611 136 : tlist = NIL;
1612 136 : break;
1613 : }
1614 :
1615 1197136 : var = makeVar(varno,
1616 : attrno,
1617 : att_tup->atttypid,
1618 : att_tup->atttypmod,
1619 : att_tup->attcollation,
1620 : 0);
1621 :
1622 1197136 : tlist = lappend(tlist,
1623 1197136 : makeTargetEntry((Expr *) var,
1624 : attrno,
1625 : NULL,
1626 : false));
1627 : }
1628 :
1629 68356 : table_close(relation, NoLock);
1630 68356 : break;
1631 :
1632 1004 : case RTE_SUBQUERY:
1633 1004 : subquery = rte->subquery;
1634 3764 : foreach(l, subquery->targetList)
1635 : {
1636 2760 : TargetEntry *tle = (TargetEntry *) lfirst(l);
1637 :
1638 : /*
1639 : * A resjunk column of the subquery can be reflected as
1640 : * resjunk in the physical tlist; we need not punt.
1641 : */
1642 2760 : var = makeVarFromTargetEntry(varno, tle);
1643 :
1644 2760 : tlist = lappend(tlist,
1645 5520 : makeTargetEntry((Expr *) var,
1646 2760 : tle->resno,
1647 : NULL,
1648 2760 : tle->resjunk));
1649 : }
1650 1004 : break;
1651 :
1652 8884 : case RTE_FUNCTION:
1653 : case RTE_TABLEFUNC:
1654 : case RTE_VALUES:
1655 : case RTE_CTE:
1656 : case RTE_NAMEDTUPLESTORE:
1657 : case RTE_RESULT:
1658 : /* Not all of these can have dropped cols, but share code anyway */
1659 8884 : expandRTE(rte, varno, 0, -1, true /* include dropped */ ,
1660 : NULL, &colvars);
1661 32526 : foreach(l, colvars)
1662 : {
1663 23642 : var = (Var *) lfirst(l);
1664 :
1665 : /*
1666 : * A non-Var in expandRTE's output means a dropped column;
1667 : * must punt.
1668 : */
1669 23642 : if (!IsA(var, Var))
1670 : {
1671 0 : tlist = NIL;
1672 0 : break;
1673 : }
1674 :
1675 23642 : tlist = lappend(tlist,
1676 23642 : makeTargetEntry((Expr *) var,
1677 23642 : var->varattno,
1678 : NULL,
1679 : false));
1680 : }
1681 8884 : break;
1682 :
1683 0 : default:
1684 : /* caller error */
1685 0 : elog(ERROR, "unsupported RTE kind %d in build_physical_tlist",
1686 : (int) rte->rtekind);
1687 : break;
1688 : }
1689 :
1690 78244 : return tlist;
1691 : }
1692 :
1693 : /*
1694 : * build_index_tlist
1695 : *
1696 : * Build a targetlist representing the columns of the specified index.
1697 : * Each column is represented by a Var for the corresponding base-relation
1698 : * column, or an expression in base-relation Vars, as appropriate.
1699 : *
1700 : * There are never any dropped columns in indexes, so unlike
1701 : * build_physical_tlist, we need no failure case.
1702 : */
1703 : static List *
1704 417172 : build_index_tlist(PlannerInfo *root, IndexOptInfo *index,
1705 : Relation heapRelation)
1706 : {
1707 417172 : List *tlist = NIL;
1708 417172 : Index varno = index->rel->relid;
1709 : ListCell *indexpr_item;
1710 : int i;
1711 :
1712 417172 : indexpr_item = list_head(index->indexprs);
1713 1167690 : for (i = 0; i < index->ncolumns; i++)
1714 : {
1715 750518 : int indexkey = index->indexkeys[i];
1716 : Expr *indexvar;
1717 :
1718 750518 : if (indexkey != 0)
1719 : {
1720 : /* simple column */
1721 : const FormData_pg_attribute *att_tup;
1722 :
1723 749018 : if (indexkey < 0)
1724 0 : att_tup = SystemAttributeDefinition(indexkey);
1725 : else
1726 749018 : att_tup = TupleDescAttr(heapRelation->rd_att, indexkey - 1);
1727 :
1728 749018 : indexvar = (Expr *) makeVar(varno,
1729 : indexkey,
1730 : att_tup->atttypid,
1731 : att_tup->atttypmod,
1732 : att_tup->attcollation,
1733 : 0);
1734 : }
1735 : else
1736 : {
1737 : /* expression column */
1738 1500 : if (indexpr_item == NULL)
1739 0 : elog(ERROR, "wrong number of index expressions");
1740 1500 : indexvar = (Expr *) lfirst(indexpr_item);
1741 1500 : indexpr_item = lnext(index->indexprs, indexpr_item);
1742 : }
1743 :
1744 750518 : tlist = lappend(tlist,
1745 750518 : makeTargetEntry(indexvar,
1746 750518 : i + 1,
1747 : NULL,
1748 : false));
1749 : }
1750 417172 : if (indexpr_item != NULL)
1751 0 : elog(ERROR, "wrong number of index expressions");
1752 :
1753 417172 : return tlist;
1754 : }
1755 :
1756 : /*
1757 : * restriction_selectivity
1758 : *
1759 : * Returns the selectivity of a specified restriction operator clause.
1760 : * This code executes registered procedures stored in the
1761 : * operator relation, by calling the function manager.
1762 : *
1763 : * See clause_selectivity() for the meaning of the additional parameters.
1764 : */
1765 : Selectivity
1766 340982 : restriction_selectivity(PlannerInfo *root,
1767 : Oid operatorid,
1768 : List *args,
1769 : Oid inputcollid,
1770 : int varRelid)
1771 : {
1772 340982 : RegProcedure oprrest = get_oprrest(operatorid);
1773 : float8 result;
1774 :
1775 : /*
1776 : * if the oprrest procedure is missing for whatever reason, use a
1777 : * selectivity of 0.5
1778 : */
1779 340982 : if (!oprrest)
1780 56 : return (Selectivity) 0.5;
1781 :
1782 340926 : result = DatumGetFloat8(OidFunctionCall4Coll(oprrest,
1783 : inputcollid,
1784 : PointerGetDatum(root),
1785 : ObjectIdGetDatum(operatorid),
1786 : PointerGetDatum(args),
1787 : Int32GetDatum(varRelid)));
1788 :
1789 340926 : if (result < 0.0 || result > 1.0)
1790 0 : elog(ERROR, "invalid restriction selectivity: %f", result);
1791 :
1792 340926 : return (Selectivity) result;
1793 : }
1794 :
1795 : /*
1796 : * join_selectivity
1797 : *
1798 : * Returns the selectivity of a specified join operator clause.
1799 : * This code executes registered procedures stored in the
1800 : * operator relation, by calling the function manager.
1801 : *
1802 : * See clause_selectivity() for the meaning of the additional parameters.
1803 : */
1804 : Selectivity
1805 124500 : join_selectivity(PlannerInfo *root,
1806 : Oid operatorid,
1807 : List *args,
1808 : Oid inputcollid,
1809 : JoinType jointype,
1810 : SpecialJoinInfo *sjinfo)
1811 : {
1812 124500 : RegProcedure oprjoin = get_oprjoin(operatorid);
1813 : float8 result;
1814 :
1815 : /*
1816 : * if the oprjoin procedure is missing for whatever reason, use a
1817 : * selectivity of 0.5
1818 : */
1819 124500 : if (!oprjoin)
1820 100 : return (Selectivity) 0.5;
1821 :
1822 124400 : result = DatumGetFloat8(OidFunctionCall5Coll(oprjoin,
1823 : inputcollid,
1824 : PointerGetDatum(root),
1825 : ObjectIdGetDatum(operatorid),
1826 : PointerGetDatum(args),
1827 : Int16GetDatum(jointype),
1828 : PointerGetDatum(sjinfo)));
1829 :
1830 124400 : if (result < 0.0 || result > 1.0)
1831 0 : elog(ERROR, "invalid join selectivity: %f", result);
1832 :
1833 124400 : return (Selectivity) result;
1834 : }
1835 :
1836 : /*
1837 : * function_selectivity
1838 : *
1839 : * Returns the selectivity of a specified boolean function clause.
1840 : * This code executes registered procedures stored in the
1841 : * pg_proc relation, by calling the function manager.
1842 : *
1843 : * See clause_selectivity() for the meaning of the additional parameters.
1844 : */
1845 : Selectivity
1846 5804 : function_selectivity(PlannerInfo *root,
1847 : Oid funcid,
1848 : List *args,
1849 : Oid inputcollid,
1850 : bool is_join,
1851 : int varRelid,
1852 : JoinType jointype,
1853 : SpecialJoinInfo *sjinfo)
1854 : {
1855 5804 : RegProcedure prosupport = get_func_support(funcid);
1856 : SupportRequestSelectivity req;
1857 : SupportRequestSelectivity *sresult;
1858 :
1859 : /*
1860 : * If no support function is provided, use our historical default
1861 : * estimate, 0.3333333. This seems a pretty unprincipled choice, but
1862 : * Postgres has been using that estimate for function calls since 1992.
1863 : * The hoariness of this behavior suggests that we should not be in too
1864 : * much hurry to use another value.
1865 : */
1866 5804 : if (!prosupport)
1867 5800 : return (Selectivity) 0.3333333;
1868 :
1869 4 : req.type = T_SupportRequestSelectivity;
1870 4 : req.root = root;
1871 4 : req.funcid = funcid;
1872 4 : req.args = args;
1873 4 : req.inputcollid = inputcollid;
1874 4 : req.is_join = is_join;
1875 4 : req.varRelid = varRelid;
1876 4 : req.jointype = jointype;
1877 4 : req.sjinfo = sjinfo;
1878 4 : req.selectivity = -1; /* to catch failure to set the value */
1879 :
1880 4 : sresult = (SupportRequestSelectivity *)
1881 4 : DatumGetPointer(OidFunctionCall1(prosupport,
1882 : PointerGetDatum(&req)));
1883 :
1884 : /* If support function fails, use default */
1885 4 : if (sresult != &req)
1886 0 : return (Selectivity) 0.3333333;
1887 :
1888 4 : if (req.selectivity < 0.0 || req.selectivity > 1.0)
1889 0 : elog(ERROR, "invalid function selectivity: %f", req.selectivity);
1890 :
1891 4 : return (Selectivity) req.selectivity;
1892 : }
1893 :
1894 : /*
1895 : * add_function_cost
1896 : *
1897 : * Get an estimate of the execution cost of a function, and *add* it to
1898 : * the contents of *cost. The estimate may include both one-time and
1899 : * per-tuple components, since QualCost does.
1900 : *
1901 : * The funcid must always be supplied. If it is being called as the
1902 : * implementation of a specific parsetree node (FuncExpr, OpExpr,
1903 : * WindowFunc, etc), pass that as "node", else pass NULL.
1904 : *
1905 : * In some usages root might be NULL, too.
1906 : */
1907 : void
1908 718342 : add_function_cost(PlannerInfo *root, Oid funcid, Node *node,
1909 : QualCost *cost)
1910 : {
1911 : HeapTuple proctup;
1912 : Form_pg_proc procform;
1913 :
1914 718342 : proctup = SearchSysCache1(PROCOID, ObjectIdGetDatum(funcid));
1915 718342 : if (!HeapTupleIsValid(proctup))
1916 0 : elog(ERROR, "cache lookup failed for function %u", funcid);
1917 718342 : procform = (Form_pg_proc) GETSTRUCT(proctup);
1918 :
1919 718342 : if (OidIsValid(procform->prosupport))
1920 : {
1921 : SupportRequestCost req;
1922 : SupportRequestCost *sresult;
1923 :
1924 26004 : req.type = T_SupportRequestCost;
1925 26004 : req.root = root;
1926 26004 : req.funcid = funcid;
1927 26004 : req.node = node;
1928 :
1929 : /* Initialize cost fields so that support function doesn't have to */
1930 26004 : req.startup = 0;
1931 26004 : req.per_tuple = 0;
1932 :
1933 26004 : sresult = (SupportRequestCost *)
1934 26004 : DatumGetPointer(OidFunctionCall1(procform->prosupport,
1935 : PointerGetDatum(&req)));
1936 :
1937 26004 : if (sresult == &req)
1938 : {
1939 : /* Success, so accumulate support function's estimate into *cost */
1940 12 : cost->startup += req.startup;
1941 12 : cost->per_tuple += req.per_tuple;
1942 12 : ReleaseSysCache(proctup);
1943 12 : return;
1944 : }
1945 : }
1946 :
1947 : /* No support function, or it failed, so rely on procost */
1948 718330 : cost->per_tuple += procform->procost * cpu_operator_cost;
1949 :
1950 718330 : ReleaseSysCache(proctup);
1951 : }
1952 :
1953 : /*
1954 : * get_function_rows
1955 : *
1956 : * Get an estimate of the number of rows returned by a set-returning function.
1957 : *
1958 : * The funcid must always be supplied. In current usage, the calling node
1959 : * will always be supplied, and will be either a FuncExpr or OpExpr.
1960 : * But it's a good idea to not fail if it's NULL.
1961 : *
1962 : * In some usages root might be NULL, too.
1963 : *
1964 : * Note: this returns the unfiltered result of the support function, if any.
1965 : * It's usually a good idea to apply clamp_row_est() to the result, but we
1966 : * leave it to the caller to do so.
1967 : */
1968 : double
1969 32612 : get_function_rows(PlannerInfo *root, Oid funcid, Node *node)
1970 : {
1971 : HeapTuple proctup;
1972 : Form_pg_proc procform;
1973 : double result;
1974 :
1975 32612 : proctup = SearchSysCache1(PROCOID, ObjectIdGetDatum(funcid));
1976 32612 : if (!HeapTupleIsValid(proctup))
1977 0 : elog(ERROR, "cache lookup failed for function %u", funcid);
1978 32612 : procform = (Form_pg_proc) GETSTRUCT(proctup);
1979 :
1980 : Assert(procform->proretset); /* else caller error */
1981 :
1982 32612 : if (OidIsValid(procform->prosupport))
1983 : {
1984 : SupportRequestRows req;
1985 : SupportRequestRows *sresult;
1986 :
1987 19570 : req.type = T_SupportRequestRows;
1988 19570 : req.root = root;
1989 19570 : req.funcid = funcid;
1990 19570 : req.node = node;
1991 :
1992 19570 : req.rows = 0; /* just for sanity */
1993 :
1994 19570 : sresult = (SupportRequestRows *)
1995 19570 : DatumGetPointer(OidFunctionCall1(procform->prosupport,
1996 : PointerGetDatum(&req)));
1997 :
1998 19570 : if (sresult == &req)
1999 : {
2000 : /* Success */
2001 19382 : ReleaseSysCache(proctup);
2002 19382 : return req.rows;
2003 : }
2004 : }
2005 :
2006 : /* No support function, or it failed, so rely on prorows */
2007 13230 : result = procform->prorows;
2008 :
2009 13230 : ReleaseSysCache(proctup);
2010 :
2011 13230 : return result;
2012 : }
2013 :
2014 : /*
2015 : * has_unique_index
2016 : *
2017 : * Detect whether there is a unique index on the specified attribute
2018 : * of the specified relation, thus allowing us to conclude that all
2019 : * the (non-null) values of the attribute are distinct.
2020 : *
2021 : * This function does not check the index's indimmediate property, which
2022 : * means that uniqueness may transiently fail to hold intra-transaction.
2023 : * That's appropriate when we are making statistical estimates, but beware
2024 : * of using this for any correctness proofs.
2025 : */
2026 : bool
2027 844310 : has_unique_index(RelOptInfo *rel, AttrNumber attno)
2028 : {
2029 : ListCell *ilist;
2030 :
2031 2034110 : foreach(ilist, rel->indexlist)
2032 : {
2033 1471092 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
2034 :
2035 : /*
2036 : * Note: ignore partial indexes, since they don't allow us to conclude
2037 : * that all attr values are distinct, *unless* they are marked predOK
2038 : * which means we know the index's predicate is satisfied by the
2039 : * query. We don't take any interest in expressional indexes either.
2040 : * Also, a multicolumn unique index doesn't allow us to conclude that
2041 : * just the specified attr is unique.
2042 : */
2043 1471092 : if (index->unique &&
2044 1067874 : index->nkeycolumns == 1 &&
2045 559402 : index->indexkeys[0] == attno &&
2046 281320 : (index->indpred == NIL || index->predOK))
2047 281292 : return true;
2048 : }
2049 563018 : return false;
2050 : }
2051 :
2052 :
2053 : /*
2054 : * has_row_triggers
2055 : *
2056 : * Detect whether the specified relation has any row-level triggers for event.
2057 : */
2058 : bool
2059 344 : has_row_triggers(PlannerInfo *root, Index rti, CmdType event)
2060 : {
2061 344 : RangeTblEntry *rte = planner_rt_fetch(rti, root);
2062 : Relation relation;
2063 : TriggerDesc *trigDesc;
2064 344 : bool result = false;
2065 :
2066 : /* Assume we already have adequate lock */
2067 344 : relation = table_open(rte->relid, NoLock);
2068 :
2069 344 : trigDesc = relation->trigdesc;
2070 344 : switch (event)
2071 : {
2072 88 : case CMD_INSERT:
2073 88 : if (trigDesc &&
2074 22 : (trigDesc->trig_insert_after_row ||
2075 10 : trigDesc->trig_insert_before_row))
2076 22 : result = true;
2077 88 : break;
2078 152 : case CMD_UPDATE:
2079 152 : if (trigDesc &&
2080 48 : (trigDesc->trig_update_after_row ||
2081 28 : trigDesc->trig_update_before_row))
2082 36 : result = true;
2083 152 : break;
2084 104 : case CMD_DELETE:
2085 104 : if (trigDesc &&
2086 30 : (trigDesc->trig_delete_after_row ||
2087 18 : trigDesc->trig_delete_before_row))
2088 16 : result = true;
2089 104 : break;
2090 0 : default:
2091 0 : elog(ERROR, "unrecognized CmdType: %d", (int) event);
2092 : break;
2093 : }
2094 :
2095 344 : table_close(relation, NoLock);
2096 344 : return result;
2097 : }
2098 :
2099 : bool
2100 270 : has_stored_generated_columns(PlannerInfo *root, Index rti)
2101 : {
2102 270 : RangeTblEntry *rte = planner_rt_fetch(rti, root);
2103 : Relation relation;
2104 : TupleDesc tupdesc;
2105 270 : bool result = false;
2106 :
2107 : /* Assume we already have adequate lock */
2108 270 : relation = table_open(rte->relid, NoLock);
2109 :
2110 270 : tupdesc = RelationGetDescr(relation);
2111 270 : result = tupdesc->constr && tupdesc->constr->has_generated_stored;
2112 :
2113 270 : table_close(relation, NoLock);
2114 :
2115 270 : return result;
2116 : }
2117 :
2118 : /*
2119 : * set_relation_partition_info
2120 : *
2121 : * Set partitioning scheme and related information for a partitioned table.
2122 : */
2123 : static void
2124 9760 : set_relation_partition_info(PlannerInfo *root, RelOptInfo *rel,
2125 : Relation relation)
2126 : {
2127 : PartitionDesc partdesc;
2128 :
2129 : /* Create the PartitionDirectory infrastructure if we didn't already */
2130 9760 : if (root->glob->partition_directory == NULL)
2131 6662 : root->glob->partition_directory =
2132 6662 : CreatePartitionDirectory(CurrentMemoryContext);
2133 :
2134 9760 : partdesc = PartitionDirectoryLookup(root->glob->partition_directory,
2135 : relation);
2136 9760 : rel->part_scheme = find_partition_scheme(root, relation);
2137 : Assert(partdesc != NULL && rel->part_scheme != NULL);
2138 9760 : rel->boundinfo = partdesc->boundinfo;
2139 9760 : rel->nparts = partdesc->nparts;
2140 9760 : set_baserel_partition_key_exprs(relation, rel);
2141 9760 : set_baserel_partition_constraint(relation, rel);
2142 9760 : }
2143 :
2144 : /*
2145 : * find_partition_scheme
2146 : *
2147 : * Find or create a PartitionScheme for this Relation.
2148 : */
2149 : static PartitionScheme
2150 9760 : find_partition_scheme(PlannerInfo *root, Relation relation)
2151 : {
2152 9760 : PartitionKey partkey = RelationGetPartitionKey(relation);
2153 : ListCell *lc;
2154 : int partnatts,
2155 : i;
2156 : PartitionScheme part_scheme;
2157 :
2158 : /* A partitioned table should have a partition key. */
2159 : Assert(partkey != NULL);
2160 :
2161 9760 : partnatts = partkey->partnatts;
2162 :
2163 : /* Search for a matching partition scheme and return if found one. */
2164 10916 : foreach(lc, root->part_schemes)
2165 : {
2166 3454 : part_scheme = lfirst(lc);
2167 :
2168 : /* Match partitioning strategy and number of keys. */
2169 3454 : if (partkey->strategy != part_scheme->strategy ||
2170 2870 : partnatts != part_scheme->partnatts)
2171 856 : continue;
2172 :
2173 : /* Match partition key type properties. */
2174 2598 : if (memcmp(partkey->partopfamily, part_scheme->partopfamily,
2175 2298 : sizeof(Oid) * partnatts) != 0 ||
2176 2298 : memcmp(partkey->partopcintype, part_scheme->partopcintype,
2177 2298 : sizeof(Oid) * partnatts) != 0 ||
2178 2298 : memcmp(partkey->partcollation, part_scheme->partcollation,
2179 : sizeof(Oid) * partnatts) != 0)
2180 300 : continue;
2181 :
2182 : /*
2183 : * Length and byval information should match when partopcintype
2184 : * matches.
2185 : */
2186 : Assert(memcmp(partkey->parttyplen, part_scheme->parttyplen,
2187 : sizeof(int16) * partnatts) == 0);
2188 : Assert(memcmp(partkey->parttypbyval, part_scheme->parttypbyval,
2189 : sizeof(bool) * partnatts) == 0);
2190 :
2191 : /*
2192 : * If partopfamily and partopcintype matched, must have the same
2193 : * partition comparison functions. Note that we cannot reliably
2194 : * Assert the equality of function structs themselves for they might
2195 : * be different across PartitionKey's, so just Assert for the function
2196 : * OIDs.
2197 : */
2198 : #ifdef USE_ASSERT_CHECKING
2199 : for (i = 0; i < partkey->partnatts; i++)
2200 : Assert(partkey->partsupfunc[i].fn_oid ==
2201 : part_scheme->partsupfunc[i].fn_oid);
2202 : #endif
2203 :
2204 : /* Found matching partition scheme. */
2205 2298 : return part_scheme;
2206 : }
2207 :
2208 : /*
2209 : * Did not find matching partition scheme. Create one copying relevant
2210 : * information from the relcache. We need to copy the contents of the
2211 : * array since the relcache entry may not survive after we have closed the
2212 : * relation.
2213 : */
2214 7462 : part_scheme = (PartitionScheme) palloc0(sizeof(PartitionSchemeData));
2215 7462 : part_scheme->strategy = partkey->strategy;
2216 7462 : part_scheme->partnatts = partkey->partnatts;
2217 :
2218 7462 : part_scheme->partopfamily = (Oid *) palloc(sizeof(Oid) * partnatts);
2219 7462 : memcpy(part_scheme->partopfamily, partkey->partopfamily,
2220 : sizeof(Oid) * partnatts);
2221 :
2222 7462 : part_scheme->partopcintype = (Oid *) palloc(sizeof(Oid) * partnatts);
2223 7462 : memcpy(part_scheme->partopcintype, partkey->partopcintype,
2224 : sizeof(Oid) * partnatts);
2225 :
2226 7462 : part_scheme->partcollation = (Oid *) palloc(sizeof(Oid) * partnatts);
2227 7462 : memcpy(part_scheme->partcollation, partkey->partcollation,
2228 : sizeof(Oid) * partnatts);
2229 :
2230 7462 : part_scheme->parttyplen = (int16 *) palloc(sizeof(int16) * partnatts);
2231 7462 : memcpy(part_scheme->parttyplen, partkey->parttyplen,
2232 : sizeof(int16) * partnatts);
2233 :
2234 7462 : part_scheme->parttypbyval = (bool *) palloc(sizeof(bool) * partnatts);
2235 7462 : memcpy(part_scheme->parttypbyval, partkey->parttypbyval,
2236 : sizeof(bool) * partnatts);
2237 :
2238 7462 : part_scheme->partsupfunc = (FmgrInfo *)
2239 7462 : palloc(sizeof(FmgrInfo) * partnatts);
2240 15704 : for (i = 0; i < partnatts; i++)
2241 8242 : fmgr_info_copy(&part_scheme->partsupfunc[i], &partkey->partsupfunc[i],
2242 : CurrentMemoryContext);
2243 :
2244 : /* Add the partitioning scheme to PlannerInfo. */
2245 7462 : root->part_schemes = lappend(root->part_schemes, part_scheme);
2246 :
2247 7462 : return part_scheme;
2248 : }
2249 :
2250 : /*
2251 : * set_baserel_partition_key_exprs
2252 : *
2253 : * Builds partition key expressions for the given base relation and fills
2254 : * rel->partexprs.
2255 : */
2256 : static void
2257 9760 : set_baserel_partition_key_exprs(Relation relation,
2258 : RelOptInfo *rel)
2259 : {
2260 9760 : PartitionKey partkey = RelationGetPartitionKey(relation);
2261 : int partnatts;
2262 : int cnt;
2263 : List **partexprs;
2264 : ListCell *lc;
2265 9760 : Index varno = rel->relid;
2266 :
2267 : Assert(IS_SIMPLE_REL(rel) && rel->relid > 0);
2268 :
2269 : /* A partitioned table should have a partition key. */
2270 : Assert(partkey != NULL);
2271 :
2272 9760 : partnatts = partkey->partnatts;
2273 9760 : partexprs = (List **) palloc(sizeof(List *) * partnatts);
2274 9760 : lc = list_head(partkey->partexprs);
2275 :
2276 20320 : for (cnt = 0; cnt < partnatts; cnt++)
2277 : {
2278 : Expr *partexpr;
2279 10560 : AttrNumber attno = partkey->partattrs[cnt];
2280 :
2281 10560 : if (attno != InvalidAttrNumber)
2282 : {
2283 : /* Single column partition key is stored as a Var node. */
2284 : Assert(attno > 0);
2285 :
2286 30072 : partexpr = (Expr *) makeVar(varno, attno,
2287 10024 : partkey->parttypid[cnt],
2288 10024 : partkey->parttypmod[cnt],
2289 10024 : partkey->parttypcoll[cnt], 0);
2290 : }
2291 : else
2292 : {
2293 536 : if (lc == NULL)
2294 0 : elog(ERROR, "wrong number of partition key expressions");
2295 :
2296 : /* Re-stamp the expression with given varno. */
2297 536 : partexpr = (Expr *) copyObject(lfirst(lc));
2298 536 : ChangeVarNodes((Node *) partexpr, 1, varno, 0);
2299 536 : lc = lnext(partkey->partexprs, lc);
2300 : }
2301 :
2302 : /* Base relations have a single expression per key. */
2303 10560 : partexprs[cnt] = list_make1(partexpr);
2304 : }
2305 :
2306 9760 : rel->partexprs = partexprs;
2307 :
2308 : /*
2309 : * A base relation does not have nullable partition key expressions, since
2310 : * no outer join is involved. We still allocate an array of empty
2311 : * expression lists to keep partition key expression handling code simple.
2312 : * See build_joinrel_partition_info() and match_expr_to_partition_keys().
2313 : */
2314 9760 : rel->nullable_partexprs = (List **) palloc0(sizeof(List *) * partnatts);
2315 9760 : }
2316 :
2317 : /*
2318 : * set_baserel_partition_constraint
2319 : *
2320 : * Builds the partition constraint for the given base relation and sets it
2321 : * in the given RelOptInfo. All Var nodes are restamped with the relid of the
2322 : * given relation.
2323 : */
2324 : static void
2325 9768 : set_baserel_partition_constraint(Relation relation, RelOptInfo *rel)
2326 : {
2327 : List *partconstr;
2328 :
2329 9768 : if (rel->partition_qual) /* already done */
2330 0 : return;
2331 :
2332 : /*
2333 : * Run the partition quals through const-simplification similar to check
2334 : * constraints. We skip canonicalize_qual, though, because partition
2335 : * quals should be in canonical form already; also, since the qual is in
2336 : * implicit-AND format, we'd have to explicitly convert it to explicit-AND
2337 : * format and back again.
2338 : */
2339 9768 : partconstr = RelationGetPartitionQual(relation);
2340 9768 : if (partconstr)
2341 : {
2342 1942 : partconstr = (List *) expression_planner((Expr *) partconstr);
2343 1942 : if (rel->relid != 1)
2344 1898 : ChangeVarNodes((Node *) partconstr, 1, rel->relid, 0);
2345 1942 : rel->partition_qual = partconstr;
2346 : }
2347 : }
|
