Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * indxpath.c
4 : * Routines to determine which indexes are usable for scanning a
5 : * given relation, and create Paths accordingly.
6 : *
7 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : *
11 : * IDENTIFICATION
12 : * src/backend/optimizer/path/indxpath.c
13 : *
14 : *-------------------------------------------------------------------------
15 : */
16 : #include "postgres.h"
17 :
18 : #include <math.h>
19 :
20 : #include "access/stratnum.h"
21 : #include "access/sysattr.h"
22 : #include "access/transam.h"
23 : #include "catalog/pg_am.h"
24 : #include "catalog/pg_amop.h"
25 : #include "catalog/pg_operator.h"
26 : #include "catalog/pg_opfamily.h"
27 : #include "catalog/pg_type.h"
28 : #include "nodes/makefuncs.h"
29 : #include "nodes/nodeFuncs.h"
30 : #include "nodes/supportnodes.h"
31 : #include "optimizer/cost.h"
32 : #include "optimizer/optimizer.h"
33 : #include "optimizer/pathnode.h"
34 : #include "optimizer/paths.h"
35 : #include "optimizer/prep.h"
36 : #include "optimizer/restrictinfo.h"
37 : #include "utils/lsyscache.h"
38 : #include "utils/selfuncs.h"
39 :
40 :
41 : /* XXX see PartCollMatchesExprColl */
42 : #define IndexCollMatchesExprColl(idxcollation, exprcollation) \
43 : ((idxcollation) == InvalidOid || (idxcollation) == (exprcollation))
44 :
45 : /* Whether we are looking for plain indexscan, bitmap scan, or either */
46 : typedef enum
47 : {
48 : ST_INDEXSCAN, /* must support amgettuple */
49 : ST_BITMAPSCAN, /* must support amgetbitmap */
50 : ST_ANYSCAN, /* either is okay */
51 : } ScanTypeControl;
52 :
53 : /* Data structure for collecting qual clauses that match an index */
54 : typedef struct
55 : {
56 : bool nonempty; /* True if lists are not all empty */
57 : /* Lists of IndexClause nodes, one list per index column */
58 : List *indexclauses[INDEX_MAX_KEYS];
59 : } IndexClauseSet;
60 :
61 : /* Per-path data used within choose_bitmap_and() */
62 : typedef struct
63 : {
64 : Path *path; /* IndexPath, BitmapAndPath, or BitmapOrPath */
65 : List *quals; /* the WHERE clauses it uses */
66 : List *preds; /* predicates of its partial index(es) */
67 : Bitmapset *clauseids; /* quals+preds represented as a bitmapset */
68 : bool unclassifiable; /* has too many quals+preds to process? */
69 : } PathClauseUsage;
70 :
71 : /* Callback argument for ec_member_matches_indexcol */
72 : typedef struct
73 : {
74 : IndexOptInfo *index; /* index we're considering */
75 : int indexcol; /* index column we want to match to */
76 : } ec_member_matches_arg;
77 :
78 :
79 : static void consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
80 : IndexOptInfo *index,
81 : IndexClauseSet *rclauseset,
82 : IndexClauseSet *jclauseset,
83 : IndexClauseSet *eclauseset,
84 : List **bitindexpaths);
85 : static void consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
86 : IndexOptInfo *index,
87 : IndexClauseSet *rclauseset,
88 : IndexClauseSet *jclauseset,
89 : IndexClauseSet *eclauseset,
90 : List **bitindexpaths,
91 : List *indexjoinclauses,
92 : int considered_clauses,
93 : List **considered_relids);
94 : static void get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
95 : IndexOptInfo *index,
96 : IndexClauseSet *rclauseset,
97 : IndexClauseSet *jclauseset,
98 : IndexClauseSet *eclauseset,
99 : List **bitindexpaths,
100 : Relids relids,
101 : List **considered_relids);
102 : static bool eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
103 : List *indexjoinclauses);
104 : static void get_index_paths(PlannerInfo *root, RelOptInfo *rel,
105 : IndexOptInfo *index, IndexClauseSet *clauses,
106 : List **bitindexpaths);
107 : static List *build_index_paths(PlannerInfo *root, RelOptInfo *rel,
108 : IndexOptInfo *index, IndexClauseSet *clauses,
109 : bool useful_predicate,
110 : ScanTypeControl scantype,
111 : bool *skip_nonnative_saop);
112 : static List *build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
113 : List *clauses, List *other_clauses);
114 : static List *generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
115 : List *clauses, List *other_clauses);
116 : static Path *choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel,
117 : List *paths);
118 : static int path_usage_comparator(const void *a, const void *b);
119 : static Cost bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel,
120 : Path *ipath);
121 : static Cost bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel,
122 : List *paths);
123 : static PathClauseUsage *classify_index_clause_usage(Path *path,
124 : List **clauselist);
125 : static void find_indexpath_quals(Path *bitmapqual, List **quals, List **preds);
126 : static int find_list_position(Node *node, List **nodelist);
127 : static bool check_index_only(RelOptInfo *rel, IndexOptInfo *index);
128 : static double get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids);
129 : static double adjust_rowcount_for_semijoins(PlannerInfo *root,
130 : Index cur_relid,
131 : Index outer_relid,
132 : double rowcount);
133 : static double approximate_joinrel_size(PlannerInfo *root, Relids relids);
134 : static void match_restriction_clauses_to_index(PlannerInfo *root,
135 : IndexOptInfo *index,
136 : IndexClauseSet *clauseset);
137 : static void match_join_clauses_to_index(PlannerInfo *root,
138 : RelOptInfo *rel, IndexOptInfo *index,
139 : IndexClauseSet *clauseset,
140 : List **joinorclauses);
141 : static void match_eclass_clauses_to_index(PlannerInfo *root,
142 : IndexOptInfo *index,
143 : IndexClauseSet *clauseset);
144 : static void match_clauses_to_index(PlannerInfo *root,
145 : List *clauses,
146 : IndexOptInfo *index,
147 : IndexClauseSet *clauseset);
148 : static void match_clause_to_index(PlannerInfo *root,
149 : RestrictInfo *rinfo,
150 : IndexOptInfo *index,
151 : IndexClauseSet *clauseset);
152 : static IndexClause *match_clause_to_indexcol(PlannerInfo *root,
153 : RestrictInfo *rinfo,
154 : int indexcol,
155 : IndexOptInfo *index);
156 : static bool IsBooleanOpfamily(Oid opfamily);
157 : static IndexClause *match_boolean_index_clause(PlannerInfo *root,
158 : RestrictInfo *rinfo,
159 : int indexcol, IndexOptInfo *index);
160 : static IndexClause *match_opclause_to_indexcol(PlannerInfo *root,
161 : RestrictInfo *rinfo,
162 : int indexcol,
163 : IndexOptInfo *index);
164 : static IndexClause *match_funcclause_to_indexcol(PlannerInfo *root,
165 : RestrictInfo *rinfo,
166 : int indexcol,
167 : IndexOptInfo *index);
168 : static IndexClause *get_index_clause_from_support(PlannerInfo *root,
169 : RestrictInfo *rinfo,
170 : Oid funcid,
171 : int indexarg,
172 : int indexcol,
173 : IndexOptInfo *index);
174 : static IndexClause *match_saopclause_to_indexcol(PlannerInfo *root,
175 : RestrictInfo *rinfo,
176 : int indexcol,
177 : IndexOptInfo *index);
178 : static IndexClause *match_rowcompare_to_indexcol(PlannerInfo *root,
179 : RestrictInfo *rinfo,
180 : int indexcol,
181 : IndexOptInfo *index);
182 : static IndexClause *match_orclause_to_indexcol(PlannerInfo *root,
183 : RestrictInfo *rinfo,
184 : int indexcol,
185 : IndexOptInfo *index);
186 : static IndexClause *expand_indexqual_rowcompare(PlannerInfo *root,
187 : RestrictInfo *rinfo,
188 : int indexcol,
189 : IndexOptInfo *index,
190 : Oid expr_op,
191 : bool var_on_left);
192 : static void match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
193 : List **orderby_clauses_p,
194 : List **clause_columns_p);
195 : static Expr *match_clause_to_ordering_op(IndexOptInfo *index,
196 : int indexcol, Expr *clause, Oid pk_opfamily);
197 : static bool ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
198 : EquivalenceClass *ec, EquivalenceMember *em,
199 : void *arg);
200 :
201 :
202 : /*
203 : * create_index_paths()
204 : * Generate all interesting index paths for the given relation.
205 : * Candidate paths are added to the rel's pathlist (using add_path).
206 : *
207 : * To be considered for an index scan, an index must match one or more
208 : * restriction clauses or join clauses from the query's qual condition,
209 : * or match the query's ORDER BY condition, or have a predicate that
210 : * matches the query's qual condition.
211 : *
212 : * There are two basic kinds of index scans. A "plain" index scan uses
213 : * only restriction clauses (possibly none at all) in its indexqual,
214 : * so it can be applied in any context. A "parameterized" index scan uses
215 : * join clauses (plus restriction clauses, if available) in its indexqual.
216 : * When joining such a scan to one of the relations supplying the other
217 : * variables used in its indexqual, the parameterized scan must appear as
218 : * the inner relation of a nestloop join; it can't be used on the outer side,
219 : * nor in a merge or hash join. In that context, values for the other rels'
220 : * attributes are available and fixed during any one scan of the indexpath.
221 : *
222 : * An IndexPath is generated and submitted to add_path() for each plain or
223 : * parameterized index scan this routine deems potentially interesting for
224 : * the current query.
225 : *
226 : * 'rel' is the relation for which we want to generate index paths
227 : *
228 : * Note: check_index_predicates() must have been run previously for this rel.
229 : *
230 : * Note: in cases involving LATERAL references in the relation's tlist, it's
231 : * possible that rel->lateral_relids is nonempty. Currently, we include
232 : * lateral_relids into the parameterization reported for each path, but don't
233 : * take it into account otherwise. The fact that any such rels *must* be
234 : * available as parameter sources perhaps should influence our choices of
235 : * index quals ... but for now, it doesn't seem worth troubling over.
236 : * In particular, comments below about "unparameterized" paths should be read
237 : * as meaning "unparameterized so far as the indexquals are concerned".
238 : */
239 : void
240 403054 : create_index_paths(PlannerInfo *root, RelOptInfo *rel)
241 : {
242 : List *indexpaths;
243 : List *bitindexpaths;
244 : List *bitjoinpaths;
245 : List *joinorclauses;
246 : IndexClauseSet rclauseset;
247 : IndexClauseSet jclauseset;
248 : IndexClauseSet eclauseset;
249 : ListCell *lc;
250 :
251 : /* Skip the whole mess if no indexes */
252 403054 : if (rel->indexlist == NIL)
253 71124 : return;
254 :
255 : /* Bitmap paths are collected and then dealt with at the end */
256 331930 : bitindexpaths = bitjoinpaths = joinorclauses = NIL;
257 :
258 : /* Examine each index in turn */
259 1044518 : foreach(lc, rel->indexlist)
260 : {
261 712588 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
262 :
263 : /* Protect limited-size array in IndexClauseSets */
264 : Assert(index->nkeycolumns <= INDEX_MAX_KEYS);
265 :
266 : /*
267 : * Ignore partial indexes that do not match the query.
268 : * (generate_bitmap_or_paths() might be able to do something with
269 : * them, but that's of no concern here.)
270 : */
271 712588 : if (index->indpred != NIL && !index->predOK)
272 494 : continue;
273 :
274 : /*
275 : * Identify the restriction clauses that can match the index.
276 : */
277 24211196 : MemSet(&rclauseset, 0, sizeof(rclauseset));
278 712094 : match_restriction_clauses_to_index(root, index, &rclauseset);
279 :
280 : /*
281 : * Build index paths from the restriction clauses. These will be
282 : * non-parameterized paths. Plain paths go directly to add_path(),
283 : * bitmap paths are added to bitindexpaths to be handled below.
284 : */
285 712094 : get_index_paths(root, rel, index, &rclauseset,
286 : &bitindexpaths);
287 :
288 : /*
289 : * Identify the join clauses that can match the index. For the moment
290 : * we keep them separate from the restriction clauses. Note that this
291 : * step finds only "loose" join clauses that have not been merged into
292 : * EquivalenceClasses. Also, collect join OR clauses for later.
293 : */
294 24211196 : MemSet(&jclauseset, 0, sizeof(jclauseset));
295 712094 : match_join_clauses_to_index(root, rel, index,
296 : &jclauseset, &joinorclauses);
297 :
298 : /*
299 : * Look for EquivalenceClasses that can generate joinclauses matching
300 : * the index.
301 : */
302 24211196 : MemSet(&eclauseset, 0, sizeof(eclauseset));
303 712094 : match_eclass_clauses_to_index(root, index,
304 : &eclauseset);
305 :
306 : /*
307 : * If we found any plain or eclass join clauses, build parameterized
308 : * index paths using them.
309 : */
310 712094 : if (jclauseset.nonempty || eclauseset.nonempty)
311 138792 : consider_index_join_clauses(root, rel, index,
312 : &rclauseset,
313 : &jclauseset,
314 : &eclauseset,
315 : &bitjoinpaths);
316 : }
317 :
318 : /*
319 : * Generate BitmapOrPaths for any suitable OR-clauses present in the
320 : * restriction list. Add these to bitindexpaths.
321 : */
322 331930 : indexpaths = generate_bitmap_or_paths(root, rel,
323 : rel->baserestrictinfo, NIL);
324 331930 : bitindexpaths = list_concat(bitindexpaths, indexpaths);
325 :
326 : /*
327 : * Likewise, generate BitmapOrPaths for any suitable OR-clauses present in
328 : * the joinclause list. Add these to bitjoinpaths.
329 : */
330 331930 : indexpaths = generate_bitmap_or_paths(root, rel,
331 : joinorclauses, rel->baserestrictinfo);
332 331930 : bitjoinpaths = list_concat(bitjoinpaths, indexpaths);
333 :
334 : /*
335 : * If we found anything usable, generate a BitmapHeapPath for the most
336 : * promising combination of restriction bitmap index paths. Note there
337 : * will be only one such path no matter how many indexes exist. This
338 : * should be sufficient since there's basically only one figure of merit
339 : * (total cost) for such a path.
340 : */
341 331930 : if (bitindexpaths != NIL)
342 : {
343 : Path *bitmapqual;
344 : BitmapHeapPath *bpath;
345 :
346 199084 : bitmapqual = choose_bitmap_and(root, rel, bitindexpaths);
347 199084 : bpath = create_bitmap_heap_path(root, rel, bitmapqual,
348 : rel->lateral_relids, 1.0, 0);
349 199084 : add_path(rel, (Path *) bpath);
350 :
351 : /* create a partial bitmap heap path */
352 199084 : if (rel->consider_parallel && rel->lateral_relids == NULL)
353 143702 : create_partial_bitmap_paths(root, rel, bitmapqual);
354 : }
355 :
356 : /*
357 : * Likewise, if we found anything usable, generate BitmapHeapPaths for the
358 : * most promising combinations of join bitmap index paths. Our strategy
359 : * is to generate one such path for each distinct parameterization seen
360 : * among the available bitmap index paths. This may look pretty
361 : * expensive, but usually there won't be very many distinct
362 : * parameterizations. (This logic is quite similar to that in
363 : * consider_index_join_clauses, but we're working with whole paths not
364 : * individual clauses.)
365 : */
366 331930 : if (bitjoinpaths != NIL)
367 : {
368 : List *all_path_outers;
369 :
370 : /* Identify each distinct parameterization seen in bitjoinpaths */
371 126514 : all_path_outers = NIL;
372 279096 : foreach(lc, bitjoinpaths)
373 : {
374 152582 : Path *path = (Path *) lfirst(lc);
375 152582 : Relids required_outer = PATH_REQ_OUTER(path);
376 :
377 152582 : all_path_outers = list_append_unique(all_path_outers,
378 : required_outer);
379 : }
380 :
381 : /* Now, for each distinct parameterization set ... */
382 271910 : foreach(lc, all_path_outers)
383 : {
384 145396 : Relids max_outers = (Relids) lfirst(lc);
385 : List *this_path_set;
386 : Path *bitmapqual;
387 : Relids required_outer;
388 : double loop_count;
389 : BitmapHeapPath *bpath;
390 : ListCell *lcp;
391 :
392 : /* Identify all the bitmap join paths needing no more than that */
393 145396 : this_path_set = NIL;
394 348988 : foreach(lcp, bitjoinpaths)
395 : {
396 203592 : Path *path = (Path *) lfirst(lcp);
397 :
398 203592 : if (bms_is_subset(PATH_REQ_OUTER(path), max_outers))
399 159578 : this_path_set = lappend(this_path_set, path);
400 : }
401 :
402 : /*
403 : * Add in restriction bitmap paths, since they can be used
404 : * together with any join paths.
405 : */
406 145396 : this_path_set = list_concat(this_path_set, bitindexpaths);
407 :
408 : /* Select best AND combination for this parameterization */
409 145396 : bitmapqual = choose_bitmap_and(root, rel, this_path_set);
410 :
411 : /* And push that path into the mix */
412 145396 : required_outer = PATH_REQ_OUTER(bitmapqual);
413 145396 : loop_count = get_loop_count(root, rel->relid, required_outer);
414 145396 : bpath = create_bitmap_heap_path(root, rel, bitmapqual,
415 : required_outer, loop_count, 0);
416 145396 : add_path(rel, (Path *) bpath);
417 : }
418 : }
419 : }
420 :
421 : /*
422 : * consider_index_join_clauses
423 : * Given sets of join clauses for an index, decide which parameterized
424 : * index paths to build.
425 : *
426 : * Plain indexpaths are sent directly to add_path, while potential
427 : * bitmap indexpaths are added to *bitindexpaths for later processing.
428 : *
429 : * 'rel' is the index's heap relation
430 : * 'index' is the index for which we want to generate paths
431 : * 'rclauseset' is the collection of indexable restriction clauses
432 : * 'jclauseset' is the collection of indexable simple join clauses
433 : * 'eclauseset' is the collection of indexable clauses from EquivalenceClasses
434 : * '*bitindexpaths' is the list to add bitmap paths to
435 : */
436 : static void
437 138792 : consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
438 : IndexOptInfo *index,
439 : IndexClauseSet *rclauseset,
440 : IndexClauseSet *jclauseset,
441 : IndexClauseSet *eclauseset,
442 : List **bitindexpaths)
443 : {
444 138792 : int considered_clauses = 0;
445 138792 : List *considered_relids = NIL;
446 : int indexcol;
447 :
448 : /*
449 : * The strategy here is to identify every potentially useful set of outer
450 : * rels that can provide indexable join clauses. For each such set,
451 : * select all the join clauses available from those outer rels, add on all
452 : * the indexable restriction clauses, and generate plain and/or bitmap
453 : * index paths for that set of clauses. This is based on the assumption
454 : * that it's always better to apply a clause as an indexqual than as a
455 : * filter (qpqual); which is where an available clause would end up being
456 : * applied if we omit it from the indexquals.
457 : *
458 : * This looks expensive, but in most practical cases there won't be very
459 : * many distinct sets of outer rels to consider. As a safety valve when
460 : * that's not true, we use a heuristic: limit the number of outer rel sets
461 : * considered to a multiple of the number of clauses considered. (We'll
462 : * always consider using each individual join clause, though.)
463 : *
464 : * For simplicity in selecting relevant clauses, we represent each set of
465 : * outer rels as a maximum set of clause_relids --- that is, the indexed
466 : * relation itself is also included in the relids set. considered_relids
467 : * lists all relids sets we've already tried.
468 : */
469 353810 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
470 : {
471 : /* Consider each applicable simple join clause */
472 215018 : considered_clauses += list_length(jclauseset->indexclauses[indexcol]);
473 215018 : consider_index_join_outer_rels(root, rel, index,
474 : rclauseset, jclauseset, eclauseset,
475 : bitindexpaths,
476 : jclauseset->indexclauses[indexcol],
477 : considered_clauses,
478 : &considered_relids);
479 : /* Consider each applicable eclass join clause */
480 215018 : considered_clauses += list_length(eclauseset->indexclauses[indexcol]);
481 215018 : consider_index_join_outer_rels(root, rel, index,
482 : rclauseset, jclauseset, eclauseset,
483 : bitindexpaths,
484 : eclauseset->indexclauses[indexcol],
485 : considered_clauses,
486 : &considered_relids);
487 : }
488 138792 : }
489 :
490 : /*
491 : * consider_index_join_outer_rels
492 : * Generate parameterized paths based on clause relids in the clause list.
493 : *
494 : * Workhorse for consider_index_join_clauses; see notes therein for rationale.
495 : *
496 : * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset', and
497 : * 'bitindexpaths' as above
498 : * 'indexjoinclauses' is a list of IndexClauses for join clauses
499 : * 'considered_clauses' is the total number of clauses considered (so far)
500 : * '*considered_relids' is a list of all relids sets already considered
501 : */
502 : static void
503 430036 : consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
504 : IndexOptInfo *index,
505 : IndexClauseSet *rclauseset,
506 : IndexClauseSet *jclauseset,
507 : IndexClauseSet *eclauseset,
508 : List **bitindexpaths,
509 : List *indexjoinclauses,
510 : int considered_clauses,
511 : List **considered_relids)
512 : {
513 : ListCell *lc;
514 :
515 : /* Examine relids of each joinclause in the given list */
516 587214 : foreach(lc, indexjoinclauses)
517 : {
518 157178 : IndexClause *iclause = (IndexClause *) lfirst(lc);
519 157178 : Relids clause_relids = iclause->rinfo->clause_relids;
520 157178 : EquivalenceClass *parent_ec = iclause->rinfo->parent_ec;
521 : int num_considered_relids;
522 :
523 : /* If we already tried its relids set, no need to do so again */
524 157178 : if (list_member(*considered_relids, clause_relids))
525 7998 : continue;
526 :
527 : /*
528 : * Generate the union of this clause's relids set with each
529 : * previously-tried set. This ensures we try this clause along with
530 : * every interesting subset of previous clauses. However, to avoid
531 : * exponential growth of planning time when there are many clauses,
532 : * limit the number of relid sets accepted to 10 * considered_clauses.
533 : *
534 : * Note: get_join_index_paths appends entries to *considered_relids,
535 : * but we do not need to visit such newly-added entries within this
536 : * loop, so we don't use foreach() here. No real harm would be done
537 : * if we did visit them, since the subset check would reject them; but
538 : * it would waste some cycles.
539 : */
540 149180 : num_considered_relids = list_length(*considered_relids);
541 159934 : for (int pos = 0; pos < num_considered_relids; pos++)
542 : {
543 10754 : Relids oldrelids = (Relids) list_nth(*considered_relids, pos);
544 :
545 : /*
546 : * If either is a subset of the other, no new set is possible.
547 : * This isn't a complete test for redundancy, but it's easy and
548 : * cheap. get_join_index_paths will check more carefully if we
549 : * already generated the same relids set.
550 : */
551 10754 : if (bms_subset_compare(clause_relids, oldrelids) != BMS_DIFFERENT)
552 24 : continue;
553 :
554 : /*
555 : * If this clause was derived from an equivalence class, the
556 : * clause list may contain other clauses derived from the same
557 : * eclass. We should not consider that combining this clause with
558 : * one of those clauses generates a usefully different
559 : * parameterization; so skip if any clause derived from the same
560 : * eclass would already have been included when using oldrelids.
561 : */
562 21298 : if (parent_ec &&
563 10568 : eclass_already_used(parent_ec, oldrelids,
564 : indexjoinclauses))
565 7316 : continue;
566 :
567 : /*
568 : * If the number of relid sets considered exceeds our heuristic
569 : * limit, stop considering combinations of clauses. We'll still
570 : * consider the current clause alone, though (below this loop).
571 : */
572 3414 : if (list_length(*considered_relids) >= 10 * considered_clauses)
573 0 : break;
574 :
575 : /* OK, try the union set */
576 3414 : get_join_index_paths(root, rel, index,
577 : rclauseset, jclauseset, eclauseset,
578 : bitindexpaths,
579 : bms_union(clause_relids, oldrelids),
580 : considered_relids);
581 : }
582 :
583 : /* Also try this set of relids by itself */
584 149180 : get_join_index_paths(root, rel, index,
585 : rclauseset, jclauseset, eclauseset,
586 : bitindexpaths,
587 : clause_relids,
588 : considered_relids);
589 : }
590 430036 : }
591 :
592 : /*
593 : * get_join_index_paths
594 : * Generate index paths using clauses from the specified outer relations.
595 : * In addition to generating paths, relids is added to *considered_relids
596 : * if not already present.
597 : *
598 : * Workhorse for consider_index_join_clauses; see notes therein for rationale.
599 : *
600 : * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset',
601 : * 'bitindexpaths', 'considered_relids' as above
602 : * 'relids' is the current set of relids to consider (the target rel plus
603 : * one or more outer rels)
604 : */
605 : static void
606 152594 : get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
607 : IndexOptInfo *index,
608 : IndexClauseSet *rclauseset,
609 : IndexClauseSet *jclauseset,
610 : IndexClauseSet *eclauseset,
611 : List **bitindexpaths,
612 : Relids relids,
613 : List **considered_relids)
614 : {
615 : IndexClauseSet clauseset;
616 : int indexcol;
617 :
618 : /* If we already considered this relids set, don't repeat the work */
619 152594 : if (list_member(*considered_relids, relids))
620 0 : return;
621 :
622 : /* Identify indexclauses usable with this relids set */
623 5188196 : MemSet(&clauseset, 0, sizeof(clauseset));
624 :
625 394018 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
626 : {
627 : ListCell *lc;
628 :
629 : /* First find applicable simple join clauses */
630 276564 : foreach(lc, jclauseset->indexclauses[indexcol])
631 : {
632 35140 : IndexClause *iclause = (IndexClause *) lfirst(lc);
633 :
634 35140 : if (bms_is_subset(iclause->rinfo->clause_relids, relids))
635 34714 : clauseset.indexclauses[indexcol] =
636 34714 : lappend(clauseset.indexclauses[indexcol], iclause);
637 : }
638 :
639 : /*
640 : * Add applicable eclass join clauses. The clauses generated for each
641 : * column are redundant (cf generate_implied_equalities_for_column),
642 : * so we need at most one. This is the only exception to the general
643 : * rule of using all available index clauses.
644 : */
645 257660 : foreach(lc, eclauseset->indexclauses[indexcol])
646 : {
647 145456 : IndexClause *iclause = (IndexClause *) lfirst(lc);
648 :
649 145456 : if (bms_is_subset(iclause->rinfo->clause_relids, relids))
650 : {
651 129220 : clauseset.indexclauses[indexcol] =
652 129220 : lappend(clauseset.indexclauses[indexcol], iclause);
653 129220 : break;
654 : }
655 : }
656 :
657 : /* Add restriction clauses */
658 241424 : clauseset.indexclauses[indexcol] =
659 241424 : list_concat(clauseset.indexclauses[indexcol],
660 241424 : rclauseset->indexclauses[indexcol]);
661 :
662 241424 : if (clauseset.indexclauses[indexcol] != NIL)
663 192124 : clauseset.nonempty = true;
664 : }
665 :
666 : /* We should have found something, else caller passed silly relids */
667 : Assert(clauseset.nonempty);
668 :
669 : /* Build index path(s) using the collected set of clauses */
670 152594 : get_index_paths(root, rel, index, &clauseset, bitindexpaths);
671 :
672 : /*
673 : * Remember we considered paths for this set of relids.
674 : */
675 152594 : *considered_relids = lappend(*considered_relids, relids);
676 : }
677 :
678 : /*
679 : * eclass_already_used
680 : * True if any join clause usable with oldrelids was generated from
681 : * the specified equivalence class.
682 : */
683 : static bool
684 10568 : eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
685 : List *indexjoinclauses)
686 : {
687 : ListCell *lc;
688 :
689 14282 : foreach(lc, indexjoinclauses)
690 : {
691 11030 : IndexClause *iclause = (IndexClause *) lfirst(lc);
692 11030 : RestrictInfo *rinfo = iclause->rinfo;
693 :
694 22060 : if (rinfo->parent_ec == parent_ec &&
695 11030 : bms_is_subset(rinfo->clause_relids, oldrelids))
696 7316 : return true;
697 : }
698 3252 : return false;
699 : }
700 :
701 :
702 : /*
703 : * get_index_paths
704 : * Given an index and a set of index clauses for it, construct IndexPaths.
705 : *
706 : * Plain indexpaths are sent directly to add_path, while potential
707 : * bitmap indexpaths are added to *bitindexpaths for later processing.
708 : *
709 : * This is a fairly simple frontend to build_index_paths(). Its reason for
710 : * existence is mainly to handle ScalarArrayOpExpr quals properly. If the
711 : * index AM supports them natively, we should just include them in simple
712 : * index paths. If not, we should exclude them while building simple index
713 : * paths, and then make a separate attempt to include them in bitmap paths.
714 : */
715 : static void
716 864688 : get_index_paths(PlannerInfo *root, RelOptInfo *rel,
717 : IndexOptInfo *index, IndexClauseSet *clauses,
718 : List **bitindexpaths)
719 : {
720 : List *indexpaths;
721 864688 : bool skip_nonnative_saop = false;
722 : ListCell *lc;
723 :
724 : /*
725 : * Build simple index paths using the clauses. Allow ScalarArrayOpExpr
726 : * clauses only if the index AM supports them natively.
727 : */
728 864688 : indexpaths = build_index_paths(root, rel,
729 : index, clauses,
730 864688 : index->predOK,
731 : ST_ANYSCAN,
732 : &skip_nonnative_saop);
733 :
734 : /*
735 : * Submit all the ones that can form plain IndexScan plans to add_path. (A
736 : * plain IndexPath can represent either a plain IndexScan or an
737 : * IndexOnlyScan, but for our purposes here that distinction does not
738 : * matter. However, some of the indexes might support only bitmap scans,
739 : * and those we mustn't submit to add_path here.)
740 : *
741 : * Also, pick out the ones that are usable as bitmap scans. For that, we
742 : * must discard indexes that don't support bitmap scans, and we also are
743 : * only interested in paths that have some selectivity; we should discard
744 : * anything that was generated solely for ordering purposes.
745 : */
746 1382850 : foreach(lc, indexpaths)
747 : {
748 518162 : IndexPath *ipath = (IndexPath *) lfirst(lc);
749 :
750 518162 : if (index->amhasgettuple)
751 504366 : add_path(rel, (Path *) ipath);
752 :
753 518162 : if (index->amhasgetbitmap &&
754 518162 : (ipath->path.pathkeys == NIL ||
755 324638 : ipath->indexselectivity < 1.0))
756 377564 : *bitindexpaths = lappend(*bitindexpaths, ipath);
757 : }
758 :
759 : /*
760 : * If there were ScalarArrayOpExpr clauses that the index can't handle
761 : * natively, generate bitmap scan paths relying on executor-managed
762 : * ScalarArrayOpExpr.
763 : */
764 864688 : if (skip_nonnative_saop)
765 : {
766 32 : indexpaths = build_index_paths(root, rel,
767 : index, clauses,
768 : false,
769 : ST_BITMAPSCAN,
770 : NULL);
771 32 : *bitindexpaths = list_concat(*bitindexpaths, indexpaths);
772 : }
773 864688 : }
774 :
775 : /*
776 : * build_index_paths
777 : * Given an index and a set of index clauses for it, construct zero
778 : * or more IndexPaths. It also constructs zero or more partial IndexPaths.
779 : *
780 : * We return a list of paths because (1) this routine checks some cases
781 : * that should cause us to not generate any IndexPath, and (2) in some
782 : * cases we want to consider both a forward and a backward scan, so as
783 : * to obtain both sort orders. Note that the paths are just returned
784 : * to the caller and not immediately fed to add_path().
785 : *
786 : * At top level, useful_predicate should be exactly the index's predOK flag
787 : * (ie, true if it has a predicate that was proven from the restriction
788 : * clauses). When working on an arm of an OR clause, useful_predicate
789 : * should be true if the predicate required the current OR list to be proven.
790 : * Note that this routine should never be called at all if the index has an
791 : * unprovable predicate.
792 : *
793 : * scantype indicates whether we want to create plain indexscans, bitmap
794 : * indexscans, or both. When it's ST_BITMAPSCAN, we will not consider
795 : * index ordering while deciding if a Path is worth generating.
796 : *
797 : * If skip_nonnative_saop is non-NULL, we ignore ScalarArrayOpExpr clauses
798 : * unless the index AM supports them directly, and we set *skip_nonnative_saop
799 : * to true if we found any such clauses (caller must initialize the variable
800 : * to false). If it's NULL, we do not ignore ScalarArrayOpExpr clauses.
801 : *
802 : * 'rel' is the index's heap relation
803 : * 'index' is the index for which we want to generate paths
804 : * 'clauses' is the collection of indexable clauses (IndexClause nodes)
805 : * 'useful_predicate' indicates whether the index has a useful predicate
806 : * 'scantype' indicates whether we need plain or bitmap scan support
807 : * 'skip_nonnative_saop' indicates whether to accept SAOP if index AM doesn't
808 : */
809 : static List *
810 867822 : build_index_paths(PlannerInfo *root, RelOptInfo *rel,
811 : IndexOptInfo *index, IndexClauseSet *clauses,
812 : bool useful_predicate,
813 : ScanTypeControl scantype,
814 : bool *skip_nonnative_saop)
815 : {
816 867822 : List *result = NIL;
817 : IndexPath *ipath;
818 : List *index_clauses;
819 : Relids outer_relids;
820 : double loop_count;
821 : List *orderbyclauses;
822 : List *orderbyclausecols;
823 : List *index_pathkeys;
824 : List *useful_pathkeys;
825 : bool pathkeys_possibly_useful;
826 : bool index_is_ordered;
827 : bool index_only_scan;
828 : int indexcol;
829 :
830 : Assert(skip_nonnative_saop != NULL || scantype == ST_BITMAPSCAN);
831 :
832 : /*
833 : * Check that index supports the desired scan type(s)
834 : */
835 867822 : switch (scantype)
836 : {
837 0 : case ST_INDEXSCAN:
838 0 : if (!index->amhasgettuple)
839 0 : return NIL;
840 0 : break;
841 3134 : case ST_BITMAPSCAN:
842 3134 : if (!index->amhasgetbitmap)
843 0 : return NIL;
844 3134 : break;
845 864688 : case ST_ANYSCAN:
846 : /* either or both are OK */
847 864688 : break;
848 : }
849 :
850 : /*
851 : * 1. Combine the per-column IndexClause lists into an overall list.
852 : *
853 : * In the resulting list, clauses are ordered by index key, so that the
854 : * column numbers form a nondecreasing sequence. (This order is depended
855 : * on by btree and possibly other places.) The list can be empty, if the
856 : * index AM allows that.
857 : *
858 : * We also build a Relids set showing which outer rels are required by the
859 : * selected clauses. Any lateral_relids are included in that, but not
860 : * otherwise accounted for.
861 : */
862 867822 : index_clauses = NIL;
863 867822 : outer_relids = bms_copy(rel->lateral_relids);
864 2473450 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
865 : {
866 : ListCell *lc;
867 :
868 2050638 : foreach(lc, clauses->indexclauses[indexcol])
869 : {
870 444678 : IndexClause *iclause = (IndexClause *) lfirst(lc);
871 444678 : RestrictInfo *rinfo = iclause->rinfo;
872 :
873 444678 : if (skip_nonnative_saop && !index->amsearcharray &&
874 21842 : IsA(rinfo->clause, ScalarArrayOpExpr))
875 : {
876 : /*
877 : * Caller asked us to generate IndexPaths that omit any
878 : * ScalarArrayOpExpr clauses when the underlying index AM
879 : * lacks native support.
880 : *
881 : * We must omit this clause (and tell caller about it).
882 : */
883 32 : *skip_nonnative_saop = true;
884 32 : continue;
885 : }
886 :
887 : /* OK to include this clause */
888 444646 : index_clauses = lappend(index_clauses, iclause);
889 444646 : outer_relids = bms_add_members(outer_relids,
890 444646 : rinfo->clause_relids);
891 : }
892 :
893 : /*
894 : * If no clauses match the first index column, check for amoptionalkey
895 : * restriction. We can't generate a scan over an index with
896 : * amoptionalkey = false unless there's at least one index clause.
897 : * (When working on columns after the first, this test cannot fail. It
898 : * is always okay for columns after the first to not have any
899 : * clauses.)
900 : */
901 1605960 : if (index_clauses == NIL && !index->amoptionalkey)
902 332 : return NIL;
903 : }
904 :
905 : /* We do not want the index's rel itself listed in outer_relids */
906 867490 : outer_relids = bms_del_member(outer_relids, rel->relid);
907 :
908 : /* Compute loop_count for cost estimation purposes */
909 867490 : loop_count = get_loop_count(root, rel->relid, outer_relids);
910 :
911 : /*
912 : * 2. Compute pathkeys describing index's ordering, if any, then see how
913 : * many of them are actually useful for this query. This is not relevant
914 : * if we are only trying to build bitmap indexscans.
915 : */
916 1731846 : pathkeys_possibly_useful = (scantype != ST_BITMAPSCAN &&
917 864356 : has_useful_pathkeys(root, rel));
918 867490 : index_is_ordered = (index->sortopfamily != NULL);
919 867490 : if (index_is_ordered && pathkeys_possibly_useful)
920 : {
921 654072 : index_pathkeys = build_index_pathkeys(root, index,
922 : ForwardScanDirection);
923 654072 : useful_pathkeys = truncate_useless_pathkeys(root, rel,
924 : index_pathkeys);
925 654072 : orderbyclauses = NIL;
926 654072 : orderbyclausecols = NIL;
927 : }
928 213418 : else if (index->amcanorderbyop && pathkeys_possibly_useful)
929 : {
930 : /*
931 : * See if we can generate ordering operators for query_pathkeys or at
932 : * least some prefix thereof. Matching to just a prefix of the
933 : * query_pathkeys will allow an incremental sort to be considered on
934 : * the index's partially sorted results.
935 : */
936 1074 : match_pathkeys_to_index(index, root->query_pathkeys,
937 : &orderbyclauses,
938 : &orderbyclausecols);
939 1074 : if (list_length(root->query_pathkeys) == list_length(orderbyclauses))
940 468 : useful_pathkeys = root->query_pathkeys;
941 : else
942 606 : useful_pathkeys = list_copy_head(root->query_pathkeys,
943 : list_length(orderbyclauses));
944 : }
945 : else
946 : {
947 212344 : useful_pathkeys = NIL;
948 212344 : orderbyclauses = NIL;
949 212344 : orderbyclausecols = NIL;
950 : }
951 :
952 : /*
953 : * 3. Check if an index-only scan is possible. If we're not building
954 : * plain indexscans, this isn't relevant since bitmap scans don't support
955 : * index data retrieval anyway.
956 : */
957 1731846 : index_only_scan = (scantype != ST_BITMAPSCAN &&
958 864356 : check_index_only(rel, index));
959 :
960 : /*
961 : * 4. Generate an indexscan path if there are relevant restriction clauses
962 : * in the current clauses, OR the index ordering is potentially useful for
963 : * later merging or final output ordering, OR the index has a useful
964 : * predicate, OR an index-only scan is possible.
965 : */
966 867490 : if (index_clauses != NIL || useful_pathkeys != NIL || useful_predicate ||
967 : index_only_scan)
968 : {
969 520682 : ipath = create_index_path(root, index,
970 : index_clauses,
971 : orderbyclauses,
972 : orderbyclausecols,
973 : useful_pathkeys,
974 : ForwardScanDirection,
975 : index_only_scan,
976 : outer_relids,
977 : loop_count,
978 : false);
979 520682 : result = lappend(result, ipath);
980 :
981 : /*
982 : * If appropriate, consider parallel index scan. We don't allow
983 : * parallel index scan for bitmap index scans.
984 : */
985 520682 : if (index->amcanparallel &&
986 499796 : rel->consider_parallel && outer_relids == NULL &&
987 : scantype != ST_BITMAPSCAN)
988 : {
989 274218 : ipath = create_index_path(root, index,
990 : index_clauses,
991 : orderbyclauses,
992 : orderbyclausecols,
993 : useful_pathkeys,
994 : ForwardScanDirection,
995 : index_only_scan,
996 : outer_relids,
997 : loop_count,
998 : true);
999 :
1000 : /*
1001 : * if, after costing the path, we find that it's not worth using
1002 : * parallel workers, just free it.
1003 : */
1004 274218 : if (ipath->path.parallel_workers > 0)
1005 9956 : add_partial_path(rel, (Path *) ipath);
1006 : else
1007 264262 : pfree(ipath);
1008 : }
1009 : }
1010 :
1011 : /*
1012 : * 5. If the index is ordered, a backwards scan might be interesting.
1013 : */
1014 867490 : if (index_is_ordered && pathkeys_possibly_useful)
1015 : {
1016 654072 : index_pathkeys = build_index_pathkeys(root, index,
1017 : BackwardScanDirection);
1018 654072 : useful_pathkeys = truncate_useless_pathkeys(root, rel,
1019 : index_pathkeys);
1020 654072 : if (useful_pathkeys != NIL)
1021 : {
1022 614 : ipath = create_index_path(root, index,
1023 : index_clauses,
1024 : NIL,
1025 : NIL,
1026 : useful_pathkeys,
1027 : BackwardScanDirection,
1028 : index_only_scan,
1029 : outer_relids,
1030 : loop_count,
1031 : false);
1032 614 : result = lappend(result, ipath);
1033 :
1034 : /* If appropriate, consider parallel index scan */
1035 614 : if (index->amcanparallel &&
1036 614 : rel->consider_parallel && outer_relids == NULL &&
1037 : scantype != ST_BITMAPSCAN)
1038 : {
1039 512 : ipath = create_index_path(root, index,
1040 : index_clauses,
1041 : NIL,
1042 : NIL,
1043 : useful_pathkeys,
1044 : BackwardScanDirection,
1045 : index_only_scan,
1046 : outer_relids,
1047 : loop_count,
1048 : true);
1049 :
1050 : /*
1051 : * if, after costing the path, we find that it's not worth
1052 : * using parallel workers, just free it.
1053 : */
1054 512 : if (ipath->path.parallel_workers > 0)
1055 168 : add_partial_path(rel, (Path *) ipath);
1056 : else
1057 344 : pfree(ipath);
1058 : }
1059 : }
1060 : }
1061 :
1062 867490 : return result;
1063 : }
1064 :
1065 : /*
1066 : * build_paths_for_OR
1067 : * Given a list of restriction clauses from one arm of an OR clause,
1068 : * construct all matching IndexPaths for the relation.
1069 : *
1070 : * Here we must scan all indexes of the relation, since a bitmap OR tree
1071 : * can use multiple indexes.
1072 : *
1073 : * The caller actually supplies two lists of restriction clauses: some
1074 : * "current" ones and some "other" ones. Both lists can be used freely
1075 : * to match keys of the index, but an index must use at least one of the
1076 : * "current" clauses to be considered usable. The motivation for this is
1077 : * examples like
1078 : * WHERE (x = 42) AND (... OR (y = 52 AND z = 77) OR ....)
1079 : * While we are considering the y/z subclause of the OR, we can use "x = 42"
1080 : * as one of the available index conditions; but we shouldn't match the
1081 : * subclause to any index on x alone, because such a Path would already have
1082 : * been generated at the upper level. So we could use an index on x,y,z
1083 : * or an index on x,y for the OR subclause, but not an index on just x.
1084 : * When dealing with a partial index, a match of the index predicate to
1085 : * one of the "current" clauses also makes the index usable.
1086 : *
1087 : * 'rel' is the relation for which we want to generate index paths
1088 : * 'clauses' is the current list of clauses (RestrictInfo nodes)
1089 : * 'other_clauses' is the list of additional upper-level clauses
1090 : */
1091 : static List *
1092 10720 : build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
1093 : List *clauses, List *other_clauses)
1094 : {
1095 10720 : List *result = NIL;
1096 10720 : List *all_clauses = NIL; /* not computed till needed */
1097 : ListCell *lc;
1098 :
1099 37316 : foreach(lc, rel->indexlist)
1100 : {
1101 26596 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
1102 : IndexClauseSet clauseset;
1103 : List *indexpaths;
1104 : bool useful_predicate;
1105 :
1106 : /* Ignore index if it doesn't support bitmap scans */
1107 26596 : if (!index->amhasgetbitmap)
1108 23494 : continue;
1109 :
1110 : /*
1111 : * Ignore partial indexes that do not match the query. If a partial
1112 : * index is marked predOK then we know it's OK. Otherwise, we have to
1113 : * test whether the added clauses are sufficient to imply the
1114 : * predicate. If so, we can use the index in the current context.
1115 : *
1116 : * We set useful_predicate to true iff the predicate was proven using
1117 : * the current set of clauses. This is needed to prevent matching a
1118 : * predOK index to an arm of an OR, which would be a legal but
1119 : * pointlessly inefficient plan. (A better plan will be generated by
1120 : * just scanning the predOK index alone, no OR.)
1121 : */
1122 26596 : useful_predicate = false;
1123 26596 : if (index->indpred != NIL)
1124 : {
1125 168 : if (index->predOK)
1126 : {
1127 : /* Usable, but don't set useful_predicate */
1128 : }
1129 : else
1130 : {
1131 : /* Form all_clauses if not done already */
1132 144 : if (all_clauses == NIL)
1133 60 : all_clauses = list_concat_copy(clauses, other_clauses);
1134 :
1135 144 : if (!predicate_implied_by(index->indpred, all_clauses, false))
1136 96 : continue; /* can't use it at all */
1137 :
1138 48 : if (!predicate_implied_by(index->indpred, other_clauses, false))
1139 48 : useful_predicate = true;
1140 : }
1141 : }
1142 :
1143 : /*
1144 : * Identify the restriction clauses that can match the index.
1145 : */
1146 901000 : MemSet(&clauseset, 0, sizeof(clauseset));
1147 26500 : match_clauses_to_index(root, clauses, index, &clauseset);
1148 :
1149 : /*
1150 : * If no matches so far, and the index predicate isn't useful, we
1151 : * don't want it.
1152 : */
1153 26500 : if (!clauseset.nonempty && !useful_predicate)
1154 23398 : continue;
1155 :
1156 : /*
1157 : * Add "other" restriction clauses to the clauseset.
1158 : */
1159 3102 : match_clauses_to_index(root, other_clauses, index, &clauseset);
1160 :
1161 : /*
1162 : * Construct paths if possible.
1163 : */
1164 3102 : indexpaths = build_index_paths(root, rel,
1165 : index, &clauseset,
1166 : useful_predicate,
1167 : ST_BITMAPSCAN,
1168 : NULL);
1169 3102 : result = list_concat(result, indexpaths);
1170 : }
1171 :
1172 10720 : return result;
1173 : }
1174 :
1175 : /*
1176 : * Utility structure used to group similar OR-clause arguments in
1177 : * group_similar_or_args(). It represents information about the OR-clause
1178 : * argument and its matching index key.
1179 : */
1180 : typedef struct
1181 : {
1182 : int indexnum; /* index of the matching index, or -1 if no
1183 : * matching index */
1184 : int colnum; /* index of the matching column, or -1 if no
1185 : * matching index */
1186 : Oid opno; /* OID of the OpClause operator, or InvalidOid
1187 : * if not an OpExpr */
1188 : Oid inputcollid; /* OID of the OpClause input collation */
1189 : int argindex; /* index of the clause in the list of
1190 : * arguments */
1191 : int groupindex; /* value of argindex for the fist clause in
1192 : * the group of similar clauses */
1193 : } OrArgIndexMatch;
1194 :
1195 : /*
1196 : * Comparison function for OrArgIndexMatch which provides sort order placing
1197 : * similar OR-clause arguments together.
1198 : */
1199 : static int
1200 7066 : or_arg_index_match_cmp(const void *a, const void *b)
1201 : {
1202 7066 : const OrArgIndexMatch *match_a = (const OrArgIndexMatch *) a;
1203 7066 : const OrArgIndexMatch *match_b = (const OrArgIndexMatch *) b;
1204 :
1205 7066 : if (match_a->indexnum < match_b->indexnum)
1206 1342 : return -1;
1207 5724 : else if (match_a->indexnum > match_b->indexnum)
1208 3062 : return 1;
1209 :
1210 2662 : if (match_a->colnum < match_b->colnum)
1211 866 : return -1;
1212 1796 : else if (match_a->colnum > match_b->colnum)
1213 24 : return 1;
1214 :
1215 1772 : if (match_a->opno < match_b->opno)
1216 18 : return -1;
1217 1754 : else if (match_a->opno > match_b->opno)
1218 42 : return 1;
1219 :
1220 1712 : if (match_a->inputcollid < match_b->inputcollid)
1221 0 : return -1;
1222 1712 : else if (match_a->inputcollid > match_b->inputcollid)
1223 0 : return 1;
1224 :
1225 1712 : if (match_a->argindex < match_b->argindex)
1226 1634 : return -1;
1227 78 : else if (match_a->argindex > match_b->argindex)
1228 78 : return 1;
1229 :
1230 0 : return 0;
1231 : }
1232 :
1233 : /*
1234 : * Another comparison function for OrArgIndexMatch. It sorts groups together
1235 : * using groupindex. The group items are then sorted by argindex.
1236 : */
1237 : static int
1238 7156 : or_arg_index_match_cmp_group(const void *a, const void *b)
1239 : {
1240 7156 : const OrArgIndexMatch *match_a = (const OrArgIndexMatch *) a;
1241 7156 : const OrArgIndexMatch *match_b = (const OrArgIndexMatch *) b;
1242 :
1243 7156 : if (match_a->groupindex < match_b->groupindex)
1244 3506 : return -1;
1245 3650 : else if (match_a->groupindex > match_b->groupindex)
1246 3206 : return 1;
1247 :
1248 444 : if (match_a->argindex < match_b->argindex)
1249 444 : return -1;
1250 0 : else if (match_a->argindex > match_b->argindex)
1251 0 : return 1;
1252 :
1253 0 : return 0;
1254 : }
1255 :
1256 : /*
1257 : * group_similar_or_args
1258 : * Transform incoming OR-restrictinfo into a list of sub-restrictinfos,
1259 : * each of them containing a subset of similar OR-clause arguments from
1260 : * the source rinfo.
1261 : *
1262 : * Similar OR-clause arguments are of the form "indexkey op constant" having
1263 : * the same indexkey, operator, and collation. Constant may comprise either
1264 : * Const or Param. It may be employed later, during the
1265 : * match_clause_to_indexcol() to transform the whole OR-sub-rinfo to an SAOP
1266 : * clause.
1267 : *
1268 : * Returns the processed list of OR-clause arguments.
1269 : */
1270 : static List *
1271 8964 : group_similar_or_args(PlannerInfo *root, RelOptInfo *rel, RestrictInfo *rinfo)
1272 : {
1273 : int n;
1274 : int i;
1275 : int group_start;
1276 : OrArgIndexMatch *matches;
1277 8964 : bool matched = false;
1278 : ListCell *lc;
1279 : ListCell *lc2;
1280 : List *orargs;
1281 8964 : List *result = NIL;
1282 8964 : Index relid = rel->relid;
1283 :
1284 : Assert(IsA(rinfo->orclause, BoolExpr));
1285 8964 : orargs = ((BoolExpr *) rinfo->orclause)->args;
1286 8964 : n = list_length(orargs);
1287 :
1288 : /*
1289 : * To avoid N^2 behavior, take utility pass along the list of OR-clause
1290 : * arguments. For each argument, fill the OrArgIndexMatch structure,
1291 : * which will be used to sort these arguments at the next step.
1292 : */
1293 8964 : i = -1;
1294 8964 : matches = (OrArgIndexMatch *) palloc(sizeof(OrArgIndexMatch) * n);
1295 30218 : foreach(lc, orargs)
1296 : {
1297 21254 : Node *arg = lfirst(lc);
1298 : RestrictInfo *argrinfo;
1299 : OpExpr *clause;
1300 : Oid opno;
1301 : Node *leftop,
1302 : *rightop;
1303 : Node *nonConstExpr;
1304 : int indexnum;
1305 : int colnum;
1306 :
1307 21254 : i++;
1308 21254 : matches[i].argindex = i;
1309 21254 : matches[i].groupindex = i;
1310 21254 : matches[i].indexnum = -1;
1311 21254 : matches[i].colnum = -1;
1312 21254 : matches[i].opno = InvalidOid;
1313 21254 : matches[i].inputcollid = InvalidOid;
1314 :
1315 21254 : if (!IsA(arg, RestrictInfo))
1316 2364 : continue;
1317 :
1318 18890 : argrinfo = castNode(RestrictInfo, arg);
1319 :
1320 : /* Only operator clauses can match */
1321 18890 : if (!IsA(argrinfo->clause, OpExpr))
1322 7700 : continue;
1323 :
1324 11190 : clause = (OpExpr *) argrinfo->clause;
1325 11190 : opno = clause->opno;
1326 :
1327 : /* Only binary operators can match */
1328 11190 : if (list_length(clause->args) != 2)
1329 0 : continue;
1330 :
1331 : /*
1332 : * Ignore any RelabelType node above the operands. This is needed to
1333 : * be able to apply indexscanning in binary-compatible-operator cases.
1334 : * Note: we can assume there is at most one RelabelType node;
1335 : * eval_const_expressions() will have simplified if more than one.
1336 : */
1337 11190 : leftop = get_leftop(clause);
1338 11190 : if (IsA(leftop, RelabelType))
1339 204 : leftop = (Node *) ((RelabelType *) leftop)->arg;
1340 :
1341 11190 : rightop = get_rightop(clause);
1342 11190 : if (IsA(rightop, RelabelType))
1343 800 : rightop = (Node *) ((RelabelType *) rightop)->arg;
1344 :
1345 : /*
1346 : * Check for clauses of the form: (indexkey operator constant) or
1347 : * (constant operator indexkey). But we don't know a particular index
1348 : * yet. Therefore, we try to distinguish the potential index key and
1349 : * constant first, then search for a matching index key among all
1350 : * indexes.
1351 : */
1352 11190 : if (bms_is_member(relid, argrinfo->right_relids) &&
1353 1970 : !bms_is_member(relid, argrinfo->left_relids) &&
1354 1898 : !contain_volatile_functions(leftop))
1355 : {
1356 1898 : opno = get_commutator(opno);
1357 :
1358 1898 : if (!OidIsValid(opno))
1359 : {
1360 : /* commutator doesn't exist, we can't reverse the order */
1361 0 : continue;
1362 : }
1363 1898 : nonConstExpr = rightop;
1364 : }
1365 9292 : else if (bms_is_member(relid, argrinfo->left_relids) &&
1366 7376 : !bms_is_member(relid, argrinfo->right_relids) &&
1367 7304 : !contain_volatile_functions(rightop))
1368 : {
1369 7304 : nonConstExpr = leftop;
1370 : }
1371 : else
1372 : {
1373 1988 : continue;
1374 : }
1375 :
1376 : /*
1377 : * Match non-constant part to the index key. It's possible that a
1378 : * single non-constant part matches multiple index keys. It's OK, we
1379 : * just stop with first matching index key. Given that this choice is
1380 : * determined the same for every clause, we will group similar clauses
1381 : * together anyway.
1382 : */
1383 9202 : indexnum = 0;
1384 20024 : foreach(lc2, rel->indexlist)
1385 : {
1386 16354 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc2);
1387 :
1388 : /*
1389 : * Ignore index if it doesn't support bitmap scans or SAOP
1390 : * clauses.
1391 : */
1392 16354 : if (!index->amhasgetbitmap || !index->amsearcharray)
1393 54 : continue;
1394 :
1395 37006 : for (colnum = 0; colnum < index->nkeycolumns; colnum++)
1396 : {
1397 26238 : if (match_index_to_operand(nonConstExpr, colnum, index))
1398 : {
1399 5532 : matches[i].indexnum = indexnum;
1400 5532 : matches[i].colnum = colnum;
1401 5532 : matches[i].opno = opno;
1402 5532 : matches[i].inputcollid = clause->inputcollid;
1403 5532 : matched = true;
1404 5532 : break;
1405 : }
1406 : }
1407 :
1408 : /*
1409 : * Stop looping through the indexes, if we managed to match
1410 : * nonConstExpr to any index column.
1411 : */
1412 16300 : if (matches[i].indexnum >= 0)
1413 5532 : break;
1414 10768 : indexnum++;
1415 : }
1416 : }
1417 :
1418 : /*
1419 : * Fast-path check: if no clause is matching to the index column, we can
1420 : * just give up at this stage and return the clause list as-is.
1421 : */
1422 8964 : if (!matched)
1423 : {
1424 4992 : pfree(matches);
1425 4992 : return orargs;
1426 : }
1427 :
1428 : /*
1429 : * Sort clauses to make similar clauses go together. But at the same
1430 : * time, we would like to change the order of clauses as little as
1431 : * possible. To do so, we reorder each group of similar clauses so that
1432 : * the first item of the group stays in place, and all the other items are
1433 : * moved after it. So, if there are no similar clauses, the order of
1434 : * clauses stays the same. When there are some groups, required
1435 : * reordering happens while the rest of the clauses remain in their
1436 : * places. That is achieved by assigning a 'groupindex' to each clause:
1437 : * the number of the first item in the group in the original clause list.
1438 : */
1439 3972 : qsort(matches, n, sizeof(OrArgIndexMatch), or_arg_index_match_cmp);
1440 :
1441 : /* Assign groupindex to the sorted clauses */
1442 9506 : for (i = 1; i < n; i++)
1443 : {
1444 : /*
1445 : * When two clauses are similar and should belong to the same group,
1446 : * copy the 'groupindex' from the previous clause. Given we are
1447 : * considering clauses in direct order, all the clauses would have a
1448 : * 'groupindex' equal to the 'groupindex' of the first clause in the
1449 : * group.
1450 : */
1451 5534 : if (matches[i].indexnum == matches[i - 1].indexnum &&
1452 2542 : matches[i].colnum == matches[i - 1].colnum &&
1453 1664 : matches[i].opno == matches[i - 1].opno &&
1454 1616 : matches[i].inputcollid == matches[i - 1].inputcollid &&
1455 1616 : matches[i].indexnum != -1)
1456 444 : matches[i].groupindex = matches[i - 1].groupindex;
1457 : }
1458 :
1459 : /* Re-sort clauses first by groupindex then by argindex */
1460 3972 : qsort(matches, n, sizeof(OrArgIndexMatch), or_arg_index_match_cmp_group);
1461 :
1462 : /*
1463 : * Group similar clauses into single sub-restrictinfo. Side effect: the
1464 : * resulting list of restrictions will be sorted by indexnum and colnum.
1465 : */
1466 3972 : group_start = 0;
1467 13478 : for (i = 1; i <= n; i++)
1468 : {
1469 : /* Check if it's a group boundary */
1470 9506 : if (group_start >= 0 &&
1471 5534 : (i == n ||
1472 5534 : matches[i].indexnum != matches[group_start].indexnum ||
1473 2470 : matches[i].colnum != matches[group_start].colnum ||
1474 1610 : matches[i].opno != matches[group_start].opno ||
1475 1568 : matches[i].inputcollid != matches[group_start].inputcollid ||
1476 1568 : matches[i].indexnum == -1))
1477 : {
1478 : /*
1479 : * One clause in group: add it "as is" to the upper-level OR.
1480 : */
1481 9062 : if (i - group_start == 1)
1482 : {
1483 8750 : result = lappend(result,
1484 : list_nth(orargs,
1485 8750 : matches[group_start].argindex));
1486 : }
1487 : else
1488 : {
1489 : /*
1490 : * Two or more clauses in a group: create a nested OR.
1491 : */
1492 312 : List *args = NIL;
1493 312 : List *rargs = NIL;
1494 : RestrictInfo *subrinfo;
1495 : int j;
1496 :
1497 : Assert(i - group_start >= 2);
1498 :
1499 : /* Construct the list of nested OR arguments */
1500 1068 : for (j = group_start; j < i; j++)
1501 : {
1502 756 : Node *arg = list_nth(orargs, matches[j].argindex);
1503 :
1504 756 : rargs = lappend(rargs, arg);
1505 756 : if (IsA(arg, RestrictInfo))
1506 756 : args = lappend(args, ((RestrictInfo *) arg)->clause);
1507 : else
1508 0 : args = lappend(args, arg);
1509 : }
1510 :
1511 : /* Construct the nested OR and wrap it with RestrictInfo */
1512 312 : subrinfo = make_plain_restrictinfo(root,
1513 : make_orclause(args),
1514 : make_orclause(rargs),
1515 312 : rinfo->is_pushed_down,
1516 312 : rinfo->has_clone,
1517 312 : rinfo->is_clone,
1518 312 : rinfo->pseudoconstant,
1519 : rinfo->security_level,
1520 : rinfo->required_relids,
1521 : rinfo->incompatible_relids,
1522 : rinfo->outer_relids);
1523 312 : result = lappend(result, subrinfo);
1524 : }
1525 :
1526 9062 : group_start = i;
1527 : }
1528 : }
1529 3972 : pfree(matches);
1530 3972 : return result;
1531 : }
1532 :
1533 : /*
1534 : * make_bitmap_paths_for_or_group
1535 : * Generate bitmap paths for a group of similar OR-clause arguments
1536 : * produced by group_similar_or_args().
1537 : *
1538 : * This function considers two cases: (1) matching a group of clauses to
1539 : * the index as a whole, and (2) matching the individual clauses one-by-one.
1540 : * (1) typically comprises an optimal solution. If not, (2) typically
1541 : * comprises fair alternative.
1542 : *
1543 : * Ideally, we could consider all arbitrary splits of arguments into
1544 : * subgroups, but that could lead to unacceptable computational complexity.
1545 : * This is why we only consider two cases of above.
1546 : */
1547 : static List *
1548 306 : make_bitmap_paths_for_or_group(PlannerInfo *root, RelOptInfo *rel,
1549 : RestrictInfo *ri, List *other_clauses)
1550 : {
1551 306 : List *jointlist = NIL;
1552 306 : List *splitlist = NIL;
1553 : ListCell *lc;
1554 : List *orargs;
1555 306 : List *args = ((BoolExpr *) ri->orclause)->args;
1556 306 : Cost jointcost = 0.0,
1557 306 : splitcost = 0.0;
1558 : Path *bitmapqual;
1559 : List *indlist;
1560 :
1561 : /*
1562 : * First, try to match the whole group to the one index.
1563 : */
1564 306 : orargs = list_make1(ri);
1565 306 : indlist = build_paths_for_OR(root, rel,
1566 : orargs,
1567 : other_clauses);
1568 306 : if (indlist != NIL)
1569 : {
1570 300 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1571 300 : jointcost = bitmapqual->total_cost;
1572 300 : jointlist = list_make1(bitmapqual);
1573 : }
1574 :
1575 : /*
1576 : * If we manage to find a bitmap scan, which uses the group of OR-clause
1577 : * arguments as a whole, we can skip matching OR-clause arguments
1578 : * one-by-one as long as there are no other clauses, which can bring more
1579 : * efficiency to one-by-one case.
1580 : */
1581 306 : if (jointlist != NIL && other_clauses == NIL)
1582 84 : return jointlist;
1583 :
1584 : /*
1585 : * Also try to match all containing clauses one-by-one.
1586 : */
1587 768 : foreach(lc, args)
1588 : {
1589 552 : orargs = list_make1(lfirst(lc));
1590 :
1591 552 : indlist = build_paths_for_OR(root, rel,
1592 : orargs,
1593 : other_clauses);
1594 :
1595 552 : if (indlist == NIL)
1596 : {
1597 6 : splitlist = NIL;
1598 6 : break;
1599 : }
1600 :
1601 546 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1602 546 : splitcost += bitmapqual->total_cost;
1603 546 : splitlist = lappend(splitlist, bitmapqual);
1604 : }
1605 :
1606 : /*
1607 : * Pick the best option.
1608 : */
1609 222 : if (splitlist == NIL)
1610 6 : return jointlist;
1611 216 : else if (jointlist == NIL)
1612 0 : return splitlist;
1613 : else
1614 216 : return (jointcost < splitcost) ? jointlist : splitlist;
1615 : }
1616 :
1617 :
1618 : /*
1619 : * generate_bitmap_or_paths
1620 : * Look through the list of clauses to find OR clauses, and generate
1621 : * a BitmapOrPath for each one we can handle that way. Return a list
1622 : * of the generated BitmapOrPaths.
1623 : *
1624 : * other_clauses is a list of additional clauses that can be assumed true
1625 : * for the purpose of generating indexquals, but are not to be searched for
1626 : * ORs. (See build_paths_for_OR() for motivation.)
1627 : */
1628 : static List *
1629 665294 : generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
1630 : List *clauses, List *other_clauses)
1631 : {
1632 665294 : List *result = NIL;
1633 : List *all_clauses;
1634 : ListCell *lc;
1635 :
1636 : /*
1637 : * We can use both the current and other clauses as context for
1638 : * build_paths_for_OR; no need to remove ORs from the lists.
1639 : */
1640 665294 : all_clauses = list_concat_copy(clauses, other_clauses);
1641 :
1642 1033160 : foreach(lc, clauses)
1643 : {
1644 367866 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1645 : List *pathlist;
1646 : Path *bitmapqual;
1647 : ListCell *j;
1648 : List *groupedArgs;
1649 367866 : List *inner_other_clauses = NIL;
1650 :
1651 : /* Ignore RestrictInfos that aren't ORs */
1652 367866 : if (!restriction_is_or_clause(rinfo))
1653 358902 : continue;
1654 :
1655 : /*
1656 : * We must be able to match at least one index to each of the arms of
1657 : * the OR, else we can't use it.
1658 : */
1659 8964 : pathlist = NIL;
1660 :
1661 : /*
1662 : * Group the similar OR-clause arguments into dedicated RestrictInfos,
1663 : * because each of those RestrictInfos has a chance to match the index
1664 : * as a whole.
1665 : */
1666 8964 : groupedArgs = group_similar_or_args(root, rel, rinfo);
1667 :
1668 8964 : if (groupedArgs != ((BoolExpr *) rinfo->orclause)->args)
1669 : {
1670 : /*
1671 : * Some parts of the rinfo were probably grouped. In this case,
1672 : * we have a set of sub-rinfos that together are an exact
1673 : * duplicate of rinfo. Thus, we need to remove the rinfo from
1674 : * other clauses. match_clauses_to_index detects duplicated
1675 : * iclauses by comparing pointers to original rinfos that would be
1676 : * different. So, we must delete rinfo to avoid de-facto
1677 : * duplicated clauses in the index clauses list.
1678 : */
1679 3972 : inner_other_clauses = list_delete(list_copy(all_clauses), rinfo);
1680 : }
1681 :
1682 11184 : foreach(j, groupedArgs)
1683 : {
1684 10168 : Node *orarg = (Node *) lfirst(j);
1685 : List *indlist;
1686 :
1687 : /* OR arguments should be ANDs or sub-RestrictInfos */
1688 10168 : if (is_andclause(orarg))
1689 : {
1690 1434 : List *andargs = ((BoolExpr *) orarg)->args;
1691 :
1692 1434 : indlist = build_paths_for_OR(root, rel,
1693 : andargs,
1694 : all_clauses);
1695 :
1696 : /* Recurse in case there are sub-ORs */
1697 1434 : indlist = list_concat(indlist,
1698 1434 : generate_bitmap_or_paths(root, rel,
1699 : andargs,
1700 : all_clauses));
1701 : }
1702 8734 : else if (restriction_is_or_clause(castNode(RestrictInfo, orarg)))
1703 : {
1704 306 : RestrictInfo *ri = castNode(RestrictInfo, orarg);
1705 :
1706 : /*
1707 : * Generate bitmap paths for the group of similar OR-clause
1708 : * arguments.
1709 : */
1710 306 : indlist = make_bitmap_paths_for_or_group(root,
1711 : rel, ri,
1712 : inner_other_clauses);
1713 :
1714 306 : if (indlist == NIL)
1715 : {
1716 6 : pathlist = NIL;
1717 6 : break;
1718 : }
1719 : else
1720 : {
1721 300 : pathlist = list_concat(pathlist, indlist);
1722 300 : continue;
1723 : }
1724 : }
1725 : else
1726 : {
1727 8428 : RestrictInfo *ri = castNode(RestrictInfo, orarg);
1728 : List *orargs;
1729 :
1730 8428 : orargs = list_make1(ri);
1731 :
1732 8428 : indlist = build_paths_for_OR(root, rel,
1733 : orargs,
1734 : all_clauses);
1735 : }
1736 :
1737 : /*
1738 : * If nothing matched this arm, we can't do anything with this OR
1739 : * clause.
1740 : */
1741 9862 : if (indlist == NIL)
1742 : {
1743 7942 : pathlist = NIL;
1744 7942 : break;
1745 : }
1746 :
1747 : /*
1748 : * OK, pick the most promising AND combination, and add it to
1749 : * pathlist.
1750 : */
1751 1920 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1752 1920 : pathlist = lappend(pathlist, bitmapqual);
1753 : }
1754 :
1755 8964 : if (inner_other_clauses != NIL)
1756 2208 : list_free(inner_other_clauses);
1757 :
1758 : /*
1759 : * If we have a match for every arm, then turn them into a
1760 : * BitmapOrPath, and add to result list.
1761 : */
1762 8964 : if (pathlist != NIL)
1763 : {
1764 1016 : bitmapqual = (Path *) create_bitmap_or_path(root, rel, pathlist);
1765 1016 : result = lappend(result, bitmapqual);
1766 : }
1767 : }
1768 :
1769 665294 : return result;
1770 : }
1771 :
1772 :
1773 : /*
1774 : * choose_bitmap_and
1775 : * Given a nonempty list of bitmap paths, AND them into one path.
1776 : *
1777 : * This is a nontrivial decision since we can legally use any subset of the
1778 : * given path set. We want to choose a good tradeoff between selectivity
1779 : * and cost of computing the bitmap.
1780 : *
1781 : * The result is either a single one of the inputs, or a BitmapAndPath
1782 : * combining multiple inputs.
1783 : */
1784 : static Path *
1785 347246 : choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel, List *paths)
1786 : {
1787 347246 : int npaths = list_length(paths);
1788 : PathClauseUsage **pathinfoarray;
1789 : PathClauseUsage *pathinfo;
1790 : List *clauselist;
1791 347246 : List *bestpaths = NIL;
1792 347246 : Cost bestcost = 0;
1793 : int i,
1794 : j;
1795 : ListCell *l;
1796 :
1797 : Assert(npaths > 0); /* else caller error */
1798 347246 : if (npaths == 1)
1799 268572 : return (Path *) linitial(paths); /* easy case */
1800 :
1801 : /*
1802 : * In theory we should consider every nonempty subset of the given paths.
1803 : * In practice that seems like overkill, given the crude nature of the
1804 : * estimates, not to mention the possible effects of higher-level AND and
1805 : * OR clauses. Moreover, it's completely impractical if there are a large
1806 : * number of paths, since the work would grow as O(2^N).
1807 : *
1808 : * As a heuristic, we first check for paths using exactly the same sets of
1809 : * WHERE clauses + index predicate conditions, and reject all but the
1810 : * cheapest-to-scan in any such group. This primarily gets rid of indexes
1811 : * that include the interesting columns but also irrelevant columns. (In
1812 : * situations where the DBA has gone overboard on creating variant
1813 : * indexes, this can make for a very large reduction in the number of
1814 : * paths considered further.)
1815 : *
1816 : * We then sort the surviving paths with the cheapest-to-scan first, and
1817 : * for each path, consider using that path alone as the basis for a bitmap
1818 : * scan. Then we consider bitmap AND scans formed from that path plus
1819 : * each subsequent (higher-cost) path, adding on a subsequent path if it
1820 : * results in a reduction in the estimated total scan cost. This means we
1821 : * consider about O(N^2) rather than O(2^N) path combinations, which is
1822 : * quite tolerable, especially given than N is usually reasonably small
1823 : * because of the prefiltering step. The cheapest of these is returned.
1824 : *
1825 : * We will only consider AND combinations in which no two indexes use the
1826 : * same WHERE clause. This is a bit of a kluge: it's needed because
1827 : * costsize.c and clausesel.c aren't very smart about redundant clauses.
1828 : * They will usually double-count the redundant clauses, producing a
1829 : * too-small selectivity that makes a redundant AND step look like it
1830 : * reduces the total cost. Perhaps someday that code will be smarter and
1831 : * we can remove this limitation. (But note that this also defends
1832 : * against flat-out duplicate input paths, which can happen because
1833 : * match_join_clauses_to_index will find the same OR join clauses that
1834 : * extract_restriction_or_clauses has pulled OR restriction clauses out
1835 : * of.)
1836 : *
1837 : * For the same reason, we reject AND combinations in which an index
1838 : * predicate clause duplicates another clause. Here we find it necessary
1839 : * to be even stricter: we'll reject a partial index if any of its
1840 : * predicate clauses are implied by the set of WHERE clauses and predicate
1841 : * clauses used so far. This covers cases such as a condition "x = 42"
1842 : * used with a plain index, followed by a clauseless scan of a partial
1843 : * index "WHERE x >= 40 AND x < 50". The partial index has been accepted
1844 : * only because "x = 42" was present, and so allowing it would partially
1845 : * double-count selectivity. (We could use predicate_implied_by on
1846 : * regular qual clauses too, to have a more intelligent, but much more
1847 : * expensive, check for redundancy --- but in most cases simple equality
1848 : * seems to suffice.)
1849 : */
1850 :
1851 : /*
1852 : * Extract clause usage info and detect any paths that use exactly the
1853 : * same set of clauses; keep only the cheapest-to-scan of any such groups.
1854 : * The surviving paths are put into an array for qsort'ing.
1855 : */
1856 : pathinfoarray = (PathClauseUsage **)
1857 78674 : palloc(npaths * sizeof(PathClauseUsage *));
1858 78674 : clauselist = NIL;
1859 78674 : npaths = 0;
1860 258792 : foreach(l, paths)
1861 : {
1862 180118 : Path *ipath = (Path *) lfirst(l);
1863 :
1864 180118 : pathinfo = classify_index_clause_usage(ipath, &clauselist);
1865 :
1866 : /* If it's unclassifiable, treat it as distinct from all others */
1867 180118 : if (pathinfo->unclassifiable)
1868 : {
1869 0 : pathinfoarray[npaths++] = pathinfo;
1870 0 : continue;
1871 : }
1872 :
1873 281370 : for (i = 0; i < npaths; i++)
1874 : {
1875 248392 : if (!pathinfoarray[i]->unclassifiable &&
1876 124196 : bms_equal(pathinfo->clauseids, pathinfoarray[i]->clauseids))
1877 22944 : break;
1878 : }
1879 180118 : if (i < npaths)
1880 : {
1881 : /* duplicate clauseids, keep the cheaper one */
1882 : Cost ncost;
1883 : Cost ocost;
1884 : Selectivity nselec;
1885 : Selectivity oselec;
1886 :
1887 22944 : cost_bitmap_tree_node(pathinfo->path, &ncost, &nselec);
1888 22944 : cost_bitmap_tree_node(pathinfoarray[i]->path, &ocost, &oselec);
1889 22944 : if (ncost < ocost)
1890 5086 : pathinfoarray[i] = pathinfo;
1891 : }
1892 : else
1893 : {
1894 : /* not duplicate clauseids, add to array */
1895 157174 : pathinfoarray[npaths++] = pathinfo;
1896 : }
1897 : }
1898 :
1899 : /* If only one surviving path, we're done */
1900 78674 : if (npaths == 1)
1901 14416 : return pathinfoarray[0]->path;
1902 :
1903 : /* Sort the surviving paths by index access cost */
1904 64258 : qsort(pathinfoarray, npaths, sizeof(PathClauseUsage *),
1905 : path_usage_comparator);
1906 :
1907 : /*
1908 : * For each surviving index, consider it as an "AND group leader", and see
1909 : * whether adding on any of the later indexes results in an AND path with
1910 : * cheaper total cost than before. Then take the cheapest AND group.
1911 : *
1912 : * Note: paths that are either clauseless or unclassifiable will have
1913 : * empty clauseids, so that they will not be rejected by the clauseids
1914 : * filter here, nor will they cause later paths to be rejected by it.
1915 : */
1916 207016 : for (i = 0; i < npaths; i++)
1917 : {
1918 : Cost costsofar;
1919 : List *qualsofar;
1920 : Bitmapset *clauseidsofar;
1921 :
1922 142758 : pathinfo = pathinfoarray[i];
1923 142758 : paths = list_make1(pathinfo->path);
1924 142758 : costsofar = bitmap_scan_cost_est(root, rel, pathinfo->path);
1925 142758 : qualsofar = list_concat_copy(pathinfo->quals, pathinfo->preds);
1926 142758 : clauseidsofar = bms_copy(pathinfo->clauseids);
1927 :
1928 235938 : for (j = i + 1; j < npaths; j++)
1929 : {
1930 : Cost newcost;
1931 :
1932 93180 : pathinfo = pathinfoarray[j];
1933 : /* Check for redundancy */
1934 93180 : if (bms_overlap(pathinfo->clauseids, clauseidsofar))
1935 43798 : continue; /* consider it redundant */
1936 49382 : if (pathinfo->preds)
1937 : {
1938 24 : bool redundant = false;
1939 :
1940 : /* we check each predicate clause separately */
1941 24 : foreach(l, pathinfo->preds)
1942 : {
1943 24 : Node *np = (Node *) lfirst(l);
1944 :
1945 24 : if (predicate_implied_by(list_make1(np), qualsofar, false))
1946 : {
1947 24 : redundant = true;
1948 24 : break; /* out of inner foreach loop */
1949 : }
1950 : }
1951 24 : if (redundant)
1952 24 : continue;
1953 : }
1954 : /* tentatively add new path to paths, so we can estimate cost */
1955 49358 : paths = lappend(paths, pathinfo->path);
1956 49358 : newcost = bitmap_and_cost_est(root, rel, paths);
1957 49358 : if (newcost < costsofar)
1958 : {
1959 : /* keep new path in paths, update subsidiary variables */
1960 284 : costsofar = newcost;
1961 284 : qualsofar = list_concat(qualsofar, pathinfo->quals);
1962 284 : qualsofar = list_concat(qualsofar, pathinfo->preds);
1963 284 : clauseidsofar = bms_add_members(clauseidsofar,
1964 284 : pathinfo->clauseids);
1965 : }
1966 : else
1967 : {
1968 : /* reject new path, remove it from paths list */
1969 49074 : paths = list_truncate(paths, list_length(paths) - 1);
1970 : }
1971 : }
1972 :
1973 : /* Keep the cheapest AND-group (or singleton) */
1974 142758 : if (i == 0 || costsofar < bestcost)
1975 : {
1976 67786 : bestpaths = paths;
1977 67786 : bestcost = costsofar;
1978 : }
1979 :
1980 : /* some easy cleanup (we don't try real hard though) */
1981 142758 : list_free(qualsofar);
1982 : }
1983 :
1984 64258 : if (list_length(bestpaths) == 1)
1985 63998 : return (Path *) linitial(bestpaths); /* no need for AND */
1986 260 : return (Path *) create_bitmap_and_path(root, rel, bestpaths);
1987 : }
1988 :
1989 : /* qsort comparator to sort in increasing index access cost order */
1990 : static int
1991 85408 : path_usage_comparator(const void *a, const void *b)
1992 : {
1993 85408 : PathClauseUsage *pa = *(PathClauseUsage *const *) a;
1994 85408 : PathClauseUsage *pb = *(PathClauseUsage *const *) b;
1995 : Cost acost;
1996 : Cost bcost;
1997 : Selectivity aselec;
1998 : Selectivity bselec;
1999 :
2000 85408 : cost_bitmap_tree_node(pa->path, &acost, &aselec);
2001 85408 : cost_bitmap_tree_node(pb->path, &bcost, &bselec);
2002 :
2003 : /*
2004 : * If costs are the same, sort by selectivity.
2005 : */
2006 85408 : if (acost < bcost)
2007 56154 : return -1;
2008 29254 : if (acost > bcost)
2009 19130 : return 1;
2010 :
2011 10124 : if (aselec < bselec)
2012 4732 : return -1;
2013 5392 : if (aselec > bselec)
2014 1680 : return 1;
2015 :
2016 3712 : return 0;
2017 : }
2018 :
2019 : /*
2020 : * Estimate the cost of actually executing a bitmap scan with a single
2021 : * index path (which could be a BitmapAnd or BitmapOr node).
2022 : */
2023 : static Cost
2024 192116 : bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel, Path *ipath)
2025 : {
2026 : BitmapHeapPath bpath;
2027 :
2028 : /* Set up a dummy BitmapHeapPath */
2029 192116 : bpath.path.type = T_BitmapHeapPath;
2030 192116 : bpath.path.pathtype = T_BitmapHeapScan;
2031 192116 : bpath.path.parent = rel;
2032 192116 : bpath.path.pathtarget = rel->reltarget;
2033 192116 : bpath.path.param_info = ipath->param_info;
2034 192116 : bpath.path.pathkeys = NIL;
2035 192116 : bpath.bitmapqual = ipath;
2036 :
2037 : /*
2038 : * Check the cost of temporary path without considering parallelism.
2039 : * Parallel bitmap heap path will be considered at later stage.
2040 : */
2041 192116 : bpath.path.parallel_workers = 0;
2042 :
2043 : /* Now we can do cost_bitmap_heap_scan */
2044 192116 : cost_bitmap_heap_scan(&bpath.path, root, rel,
2045 : bpath.path.param_info,
2046 : ipath,
2047 : get_loop_count(root, rel->relid,
2048 192116 : PATH_REQ_OUTER(ipath)));
2049 :
2050 192116 : return bpath.path.total_cost;
2051 : }
2052 :
2053 : /*
2054 : * Estimate the cost of actually executing a BitmapAnd scan with the given
2055 : * inputs.
2056 : */
2057 : static Cost
2058 49358 : bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel, List *paths)
2059 : {
2060 : BitmapAndPath *apath;
2061 :
2062 : /*
2063 : * Might as well build a real BitmapAndPath here, as the work is slightly
2064 : * too complicated to be worth repeating just to save one palloc.
2065 : */
2066 49358 : apath = create_bitmap_and_path(root, rel, paths);
2067 :
2068 49358 : return bitmap_scan_cost_est(root, rel, (Path *) apath);
2069 : }
2070 :
2071 :
2072 : /*
2073 : * classify_index_clause_usage
2074 : * Construct a PathClauseUsage struct describing the WHERE clauses and
2075 : * index predicate clauses used by the given indexscan path.
2076 : * We consider two clauses the same if they are equal().
2077 : *
2078 : * At some point we might want to migrate this info into the Path data
2079 : * structure proper, but for the moment it's only needed within
2080 : * choose_bitmap_and().
2081 : *
2082 : * *clauselist is used and expanded as needed to identify all the distinct
2083 : * clauses seen across successive calls. Caller must initialize it to NIL
2084 : * before first call of a set.
2085 : */
2086 : static PathClauseUsage *
2087 180118 : classify_index_clause_usage(Path *path, List **clauselist)
2088 : {
2089 : PathClauseUsage *result;
2090 : Bitmapset *clauseids;
2091 : ListCell *lc;
2092 :
2093 180118 : result = (PathClauseUsage *) palloc(sizeof(PathClauseUsage));
2094 180118 : result->path = path;
2095 :
2096 : /* Recursively find the quals and preds used by the path */
2097 180118 : result->quals = NIL;
2098 180118 : result->preds = NIL;
2099 180118 : find_indexpath_quals(path, &result->quals, &result->preds);
2100 :
2101 : /*
2102 : * Some machine-generated queries have outlandish numbers of qual clauses.
2103 : * To avoid getting into O(N^2) behavior even in this preliminary
2104 : * classification step, we want to limit the number of entries we can
2105 : * accumulate in *clauselist. Treat any path with more than 100 quals +
2106 : * preds as unclassifiable, which will cause calling code to consider it
2107 : * distinct from all other paths.
2108 : */
2109 180118 : if (list_length(result->quals) + list_length(result->preds) > 100)
2110 : {
2111 0 : result->clauseids = NULL;
2112 0 : result->unclassifiable = true;
2113 0 : return result;
2114 : }
2115 :
2116 : /* Build up a bitmapset representing the quals and preds */
2117 180118 : clauseids = NULL;
2118 416776 : foreach(lc, result->quals)
2119 : {
2120 236658 : Node *node = (Node *) lfirst(lc);
2121 :
2122 236658 : clauseids = bms_add_member(clauseids,
2123 : find_list_position(node, clauselist));
2124 : }
2125 180412 : foreach(lc, result->preds)
2126 : {
2127 294 : Node *node = (Node *) lfirst(lc);
2128 :
2129 294 : clauseids = bms_add_member(clauseids,
2130 : find_list_position(node, clauselist));
2131 : }
2132 180118 : result->clauseids = clauseids;
2133 180118 : result->unclassifiable = false;
2134 :
2135 180118 : return result;
2136 : }
2137 :
2138 :
2139 : /*
2140 : * find_indexpath_quals
2141 : *
2142 : * Given the Path structure for a plain or bitmap indexscan, extract lists
2143 : * of all the index clauses and index predicate conditions used in the Path.
2144 : * These are appended to the initial contents of *quals and *preds (hence
2145 : * caller should initialize those to NIL).
2146 : *
2147 : * Note we are not trying to produce an accurate representation of the AND/OR
2148 : * semantics of the Path, but just find out all the base conditions used.
2149 : *
2150 : * The result lists contain pointers to the expressions used in the Path,
2151 : * but all the list cells are freshly built, so it's safe to destructively
2152 : * modify the lists (eg, by concat'ing with other lists).
2153 : */
2154 : static void
2155 182432 : find_indexpath_quals(Path *bitmapqual, List **quals, List **preds)
2156 : {
2157 182432 : if (IsA(bitmapqual, BitmapAndPath))
2158 : {
2159 0 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
2160 : ListCell *l;
2161 :
2162 0 : foreach(l, apath->bitmapquals)
2163 : {
2164 0 : find_indexpath_quals((Path *) lfirst(l), quals, preds);
2165 : }
2166 : }
2167 182432 : else if (IsA(bitmapqual, BitmapOrPath))
2168 : {
2169 1262 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
2170 : ListCell *l;
2171 :
2172 3576 : foreach(l, opath->bitmapquals)
2173 : {
2174 2314 : find_indexpath_quals((Path *) lfirst(l), quals, preds);
2175 : }
2176 : }
2177 181170 : else if (IsA(bitmapqual, IndexPath))
2178 : {
2179 181170 : IndexPath *ipath = (IndexPath *) bitmapqual;
2180 : ListCell *l;
2181 :
2182 417828 : foreach(l, ipath->indexclauses)
2183 : {
2184 236658 : IndexClause *iclause = (IndexClause *) lfirst(l);
2185 :
2186 236658 : *quals = lappend(*quals, iclause->rinfo->clause);
2187 : }
2188 181170 : *preds = list_concat(*preds, ipath->indexinfo->indpred);
2189 : }
2190 : else
2191 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
2192 182432 : }
2193 :
2194 :
2195 : /*
2196 : * find_list_position
2197 : * Return the given node's position (counting from 0) in the given
2198 : * list of nodes. If it's not equal() to any existing list member,
2199 : * add it at the end, and return that position.
2200 : */
2201 : static int
2202 236952 : find_list_position(Node *node, List **nodelist)
2203 : {
2204 : int i;
2205 : ListCell *lc;
2206 :
2207 236952 : i = 0;
2208 381522 : foreach(lc, *nodelist)
2209 : {
2210 214632 : Node *oldnode = (Node *) lfirst(lc);
2211 :
2212 214632 : if (equal(node, oldnode))
2213 70062 : return i;
2214 144570 : i++;
2215 : }
2216 :
2217 166890 : *nodelist = lappend(*nodelist, node);
2218 :
2219 166890 : return i;
2220 : }
2221 :
2222 :
2223 : /*
2224 : * check_index_only
2225 : * Determine whether an index-only scan is possible for this index.
2226 : */
2227 : static bool
2228 864356 : check_index_only(RelOptInfo *rel, IndexOptInfo *index)
2229 : {
2230 : bool result;
2231 864356 : Bitmapset *attrs_used = NULL;
2232 864356 : Bitmapset *index_canreturn_attrs = NULL;
2233 : ListCell *lc;
2234 : int i;
2235 :
2236 : /* Index-only scans must be enabled */
2237 864356 : if (!enable_indexonlyscan)
2238 3862 : return false;
2239 :
2240 : /*
2241 : * Check that all needed attributes of the relation are available from the
2242 : * index.
2243 : */
2244 :
2245 : /*
2246 : * First, identify all the attributes needed for joins or final output.
2247 : * Note: we must look at rel's targetlist, not the attr_needed data,
2248 : * because attr_needed isn't computed for inheritance child rels.
2249 : */
2250 860494 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
2251 :
2252 : /*
2253 : * Add all the attributes used by restriction clauses; but consider only
2254 : * those clauses not implied by the index predicate, since ones that are
2255 : * so implied don't need to be checked explicitly in the plan.
2256 : *
2257 : * Note: attributes used only in index quals would not be needed at
2258 : * runtime either, if we are certain that the index is not lossy. However
2259 : * it'd be complicated to account for that accurately, and it doesn't
2260 : * matter in most cases, since we'd conclude that such attributes are
2261 : * available from the index anyway.
2262 : */
2263 1776560 : foreach(lc, index->indrestrictinfo)
2264 : {
2265 916066 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2266 :
2267 916066 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
2268 : }
2269 :
2270 : /*
2271 : * Construct a bitmapset of columns that the index can return back in an
2272 : * index-only scan.
2273 : */
2274 2456748 : for (i = 0; i < index->ncolumns; i++)
2275 : {
2276 1596254 : int attno = index->indexkeys[i];
2277 :
2278 : /*
2279 : * For the moment, we just ignore index expressions. It might be nice
2280 : * to do something with them, later.
2281 : */
2282 1596254 : if (attno == 0)
2283 3286 : continue;
2284 :
2285 1592968 : if (index->canreturn[i])
2286 : index_canreturn_attrs =
2287 1317432 : bms_add_member(index_canreturn_attrs,
2288 : attno - FirstLowInvalidHeapAttributeNumber);
2289 : }
2290 :
2291 : /* Do we have all the necessary attributes? */
2292 860494 : result = bms_is_subset(attrs_used, index_canreturn_attrs);
2293 :
2294 860494 : bms_free(attrs_used);
2295 860494 : bms_free(index_canreturn_attrs);
2296 :
2297 860494 : return result;
2298 : }
2299 :
2300 : /*
2301 : * get_loop_count
2302 : * Choose the loop count estimate to use for costing a parameterized path
2303 : * with the given set of outer relids.
2304 : *
2305 : * Since we produce parameterized paths before we've begun to generate join
2306 : * relations, it's impossible to predict exactly how many times a parameterized
2307 : * path will be iterated; we don't know the size of the relation that will be
2308 : * on the outside of the nestloop. However, we should try to account for
2309 : * multiple iterations somehow in costing the path. The heuristic embodied
2310 : * here is to use the rowcount of the smallest other base relation needed in
2311 : * the join clauses used by the path. (We could alternatively consider the
2312 : * largest one, but that seems too optimistic.) This is of course the right
2313 : * answer for single-other-relation cases, and it seems like a reasonable
2314 : * zero-order approximation for multiway-join cases.
2315 : *
2316 : * In addition, we check to see if the other side of each join clause is on
2317 : * the inside of some semijoin that the current relation is on the outside of.
2318 : * If so, the only way that a parameterized path could be used is if the
2319 : * semijoin RHS has been unique-ified, so we should use the number of unique
2320 : * RHS rows rather than using the relation's raw rowcount.
2321 : *
2322 : * Note: for this to work, allpaths.c must establish all baserel size
2323 : * estimates before it begins to compute paths, or at least before it
2324 : * calls create_index_paths().
2325 : */
2326 : static double
2327 1205002 : get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids)
2328 : {
2329 : double result;
2330 : int outer_relid;
2331 :
2332 : /* For a non-parameterized path, just return 1.0 quickly */
2333 1205002 : if (outer_relids == NULL)
2334 821144 : return 1.0;
2335 :
2336 383858 : result = 0.0;
2337 383858 : outer_relid = -1;
2338 778642 : while ((outer_relid = bms_next_member(outer_relids, outer_relid)) >= 0)
2339 : {
2340 : RelOptInfo *outer_rel;
2341 : double rowcount;
2342 :
2343 : /* Paranoia: ignore bogus relid indexes */
2344 394784 : if (outer_relid >= root->simple_rel_array_size)
2345 0 : continue;
2346 394784 : outer_rel = root->simple_rel_array[outer_relid];
2347 394784 : if (outer_rel == NULL)
2348 254 : continue;
2349 : Assert(outer_rel->relid == outer_relid); /* sanity check on array */
2350 :
2351 : /* Other relation could be proven empty, if so ignore */
2352 394530 : if (IS_DUMMY_REL(outer_rel))
2353 24 : continue;
2354 :
2355 : /* Otherwise, rel's rows estimate should be valid by now */
2356 : Assert(outer_rel->rows > 0);
2357 :
2358 : /* Check to see if rel is on the inside of any semijoins */
2359 394506 : rowcount = adjust_rowcount_for_semijoins(root,
2360 : cur_relid,
2361 : outer_relid,
2362 : outer_rel->rows);
2363 :
2364 : /* Remember smallest row count estimate among the outer rels */
2365 394506 : if (result == 0.0 || result > rowcount)
2366 390828 : result = rowcount;
2367 : }
2368 : /* Return 1.0 if we found no valid relations (shouldn't happen) */
2369 383858 : return (result > 0.0) ? result : 1.0;
2370 : }
2371 :
2372 : /*
2373 : * Check to see if outer_relid is on the inside of any semijoin that cur_relid
2374 : * is on the outside of. If so, replace rowcount with the estimated number of
2375 : * unique rows from the semijoin RHS (assuming that's smaller, which it might
2376 : * not be). The estimate is crude but it's the best we can do at this stage
2377 : * of the proceedings.
2378 : */
2379 : static double
2380 394506 : adjust_rowcount_for_semijoins(PlannerInfo *root,
2381 : Index cur_relid,
2382 : Index outer_relid,
2383 : double rowcount)
2384 : {
2385 : ListCell *lc;
2386 :
2387 613982 : foreach(lc, root->join_info_list)
2388 : {
2389 219476 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
2390 :
2391 227728 : if (sjinfo->jointype == JOIN_SEMI &&
2392 11882 : bms_is_member(cur_relid, sjinfo->syn_lefthand) &&
2393 3630 : bms_is_member(outer_relid, sjinfo->syn_righthand))
2394 : {
2395 : /* Estimate number of unique-ified rows */
2396 : double nraw;
2397 : double nunique;
2398 :
2399 1336 : nraw = approximate_joinrel_size(root, sjinfo->syn_righthand);
2400 1336 : nunique = estimate_num_groups(root,
2401 : sjinfo->semi_rhs_exprs,
2402 : nraw,
2403 : NULL,
2404 : NULL);
2405 1336 : if (rowcount > nunique)
2406 462 : rowcount = nunique;
2407 : }
2408 : }
2409 394506 : return rowcount;
2410 : }
2411 :
2412 : /*
2413 : * Make an approximate estimate of the size of a joinrel.
2414 : *
2415 : * We don't have enough info at this point to get a good estimate, so we
2416 : * just multiply the base relation sizes together. Fortunately, this is
2417 : * the right answer anyway for the most common case with a single relation
2418 : * on the RHS of a semijoin. Also, estimate_num_groups() has only a weak
2419 : * dependency on its input_rows argument (it basically uses it as a clamp).
2420 : * So we might be able to get a fairly decent end result even with a severe
2421 : * overestimate of the RHS's raw size.
2422 : */
2423 : static double
2424 1336 : approximate_joinrel_size(PlannerInfo *root, Relids relids)
2425 : {
2426 1336 : double rowcount = 1.0;
2427 : int relid;
2428 :
2429 1336 : relid = -1;
2430 2876 : while ((relid = bms_next_member(relids, relid)) >= 0)
2431 : {
2432 : RelOptInfo *rel;
2433 :
2434 : /* Paranoia: ignore bogus relid indexes */
2435 1540 : if (relid >= root->simple_rel_array_size)
2436 0 : continue;
2437 1540 : rel = root->simple_rel_array[relid];
2438 1540 : if (rel == NULL)
2439 0 : continue;
2440 : Assert(rel->relid == relid); /* sanity check on array */
2441 :
2442 : /* Relation could be proven empty, if so ignore */
2443 1540 : if (IS_DUMMY_REL(rel))
2444 0 : continue;
2445 :
2446 : /* Otherwise, rel's rows estimate should be valid by now */
2447 : Assert(rel->rows > 0);
2448 :
2449 : /* Accumulate product */
2450 1540 : rowcount *= rel->rows;
2451 : }
2452 1336 : return rowcount;
2453 : }
2454 :
2455 :
2456 : /****************************************************************************
2457 : * ---- ROUTINES TO CHECK QUERY CLAUSES ----
2458 : ****************************************************************************/
2459 :
2460 : /*
2461 : * match_restriction_clauses_to_index
2462 : * Identify restriction clauses for the rel that match the index.
2463 : * Matching clauses are added to *clauseset.
2464 : */
2465 : static void
2466 712094 : match_restriction_clauses_to_index(PlannerInfo *root,
2467 : IndexOptInfo *index,
2468 : IndexClauseSet *clauseset)
2469 : {
2470 : /* We can ignore clauses that are implied by the index predicate */
2471 712094 : match_clauses_to_index(root, index->indrestrictinfo, index, clauseset);
2472 712094 : }
2473 :
2474 : /*
2475 : * match_join_clauses_to_index
2476 : * Identify join clauses for the rel that match the index.
2477 : * Matching clauses are added to *clauseset.
2478 : * Also, add any potentially usable join OR clauses to *joinorclauses.
2479 : * They also might be processed by match_clause_to_index() as a whole.
2480 : */
2481 : static void
2482 712094 : match_join_clauses_to_index(PlannerInfo *root,
2483 : RelOptInfo *rel, IndexOptInfo *index,
2484 : IndexClauseSet *clauseset,
2485 : List **joinorclauses)
2486 : {
2487 : ListCell *lc;
2488 :
2489 : /* Scan the rel's join clauses */
2490 960984 : foreach(lc, rel->joininfo)
2491 : {
2492 248890 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2493 :
2494 : /* Check if clause can be moved to this rel */
2495 248890 : if (!join_clause_is_movable_to(rinfo, rel))
2496 152066 : continue;
2497 :
2498 : /*
2499 : * Potentially usable, so see if it matches the index or is an OR. Use
2500 : * list_append_unique_ptr() here to avoid possible duplicates when
2501 : * processing the same clauses with different indexes.
2502 : */
2503 96824 : if (restriction_is_or_clause(rinfo))
2504 13220 : *joinorclauses = list_append_unique_ptr(*joinorclauses, rinfo);
2505 :
2506 96824 : match_clause_to_index(root, rinfo, index, clauseset);
2507 : }
2508 712094 : }
2509 :
2510 : /*
2511 : * match_eclass_clauses_to_index
2512 : * Identify EquivalenceClass join clauses for the rel that match the index.
2513 : * Matching clauses are added to *clauseset.
2514 : */
2515 : static void
2516 712094 : match_eclass_clauses_to_index(PlannerInfo *root, IndexOptInfo *index,
2517 : IndexClauseSet *clauseset)
2518 : {
2519 : int indexcol;
2520 :
2521 : /* No work if rel is not in any such ECs */
2522 712094 : if (!index->rel->has_eclass_joins)
2523 407956 : return;
2524 :
2525 799990 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2526 : {
2527 : ec_member_matches_arg arg;
2528 : List *clauses;
2529 :
2530 : /* Generate clauses, skipping any that join to lateral_referencers */
2531 495852 : arg.index = index;
2532 495852 : arg.indexcol = indexcol;
2533 495852 : clauses = generate_implied_equalities_for_column(root,
2534 : index->rel,
2535 : ec_member_matches_indexcol,
2536 : &arg,
2537 495852 : index->rel->lateral_referencers);
2538 :
2539 : /*
2540 : * We have to check whether the results actually do match the index,
2541 : * since for non-btree indexes the EC's equality operators might not
2542 : * be in the index opclass (cf ec_member_matches_indexcol).
2543 : */
2544 495852 : match_clauses_to_index(root, clauses, index, clauseset);
2545 : }
2546 : }
2547 :
2548 : /*
2549 : * match_clauses_to_index
2550 : * Perform match_clause_to_index() for each clause in a list.
2551 : * Matching clauses are added to *clauseset.
2552 : */
2553 : static void
2554 1237548 : match_clauses_to_index(PlannerInfo *root,
2555 : List *clauses,
2556 : IndexOptInfo *index,
2557 : IndexClauseSet *clauseset)
2558 : {
2559 : ListCell *lc;
2560 :
2561 2223328 : foreach(lc, clauses)
2562 : {
2563 985780 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
2564 :
2565 985780 : match_clause_to_index(root, rinfo, index, clauseset);
2566 : }
2567 1237548 : }
2568 :
2569 : /*
2570 : * match_clause_to_index
2571 : * Test whether a qual clause can be used with an index.
2572 : *
2573 : * If the clause is usable, add an IndexClause entry for it to the appropriate
2574 : * list in *clauseset. (*clauseset must be initialized to zeroes before first
2575 : * call.)
2576 : *
2577 : * Note: in some circumstances we may find the same RestrictInfos coming from
2578 : * multiple places. Defend against redundant outputs by refusing to add a
2579 : * clause twice (pointer equality should be a good enough check for this).
2580 : *
2581 : * Note: it's possible that a badly-defined index could have multiple matching
2582 : * columns. We always select the first match if so; this avoids scenarios
2583 : * wherein we get an inflated idea of the index's selectivity by using the
2584 : * same clause multiple times with different index columns.
2585 : */
2586 : static void
2587 1082604 : match_clause_to_index(PlannerInfo *root,
2588 : RestrictInfo *rinfo,
2589 : IndexOptInfo *index,
2590 : IndexClauseSet *clauseset)
2591 : {
2592 : int indexcol;
2593 :
2594 : /*
2595 : * Never match pseudoconstants to indexes. (Normally a match could not
2596 : * happen anyway, since a pseudoconstant clause couldn't contain a Var,
2597 : * but what if someone builds an expression index on a constant? It's not
2598 : * totally unreasonable to do so with a partial index, either.)
2599 : */
2600 1082604 : if (rinfo->pseudoconstant)
2601 13520 : return;
2602 :
2603 : /*
2604 : * If clause can't be used as an indexqual because it must wait till after
2605 : * some lower-security-level restriction clause, reject it.
2606 : */
2607 1069084 : if (!restriction_is_securely_promotable(rinfo, index->rel))
2608 474 : return;
2609 :
2610 : /* OK, check each index key column for a match */
2611 2357770 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2612 : {
2613 : IndexClause *iclause;
2614 : ListCell *lc;
2615 :
2616 : /* Ignore duplicates */
2617 1776880 : foreach(lc, clauseset->indexclauses[indexcol])
2618 : {
2619 78790 : iclause = (IndexClause *) lfirst(lc);
2620 :
2621 78790 : if (iclause->rinfo == rinfo)
2622 0 : return;
2623 : }
2624 :
2625 : /* OK, try to match the clause to the index column */
2626 1698090 : iclause = match_clause_to_indexcol(root,
2627 : rinfo,
2628 : indexcol,
2629 : index);
2630 1698090 : if (iclause)
2631 : {
2632 : /* Success, so record it */
2633 408930 : clauseset->indexclauses[indexcol] =
2634 408930 : lappend(clauseset->indexclauses[indexcol], iclause);
2635 408930 : clauseset->nonempty = true;
2636 408930 : return;
2637 : }
2638 : }
2639 : }
2640 :
2641 : /*
2642 : * match_clause_to_indexcol()
2643 : * Determine whether a restriction clause matches a column of an index,
2644 : * and if so, build an IndexClause node describing the details.
2645 : *
2646 : * To match an index normally, an operator clause:
2647 : *
2648 : * (1) must be in the form (indexkey op const) or (const op indexkey);
2649 : * and
2650 : * (2) must contain an operator which is in the index's operator family
2651 : * for this column; and
2652 : * (3) must match the collation of the index, if collation is relevant.
2653 : *
2654 : * Our definition of "const" is exceedingly liberal: we allow anything that
2655 : * doesn't involve a volatile function or a Var of the index's relation.
2656 : * In particular, Vars belonging to other relations of the query are
2657 : * accepted here, since a clause of that form can be used in a
2658 : * parameterized indexscan. It's the responsibility of higher code levels
2659 : * to manage restriction and join clauses appropriately.
2660 : *
2661 : * Note: we do need to check for Vars of the index's relation on the
2662 : * "const" side of the clause, since clauses like (a.f1 OP (b.f2 OP a.f3))
2663 : * are not processable by a parameterized indexscan on a.f1, whereas
2664 : * something like (a.f1 OP (b.f2 OP c.f3)) is.
2665 : *
2666 : * Presently, the executor can only deal with indexquals that have the
2667 : * indexkey on the left, so we can only use clauses that have the indexkey
2668 : * on the right if we can commute the clause to put the key on the left.
2669 : * We handle that by generating an IndexClause with the correctly-commuted
2670 : * opclause as a derived indexqual.
2671 : *
2672 : * If the index has a collation, the clause must have the same collation.
2673 : * For collation-less indexes, we assume it doesn't matter; this is
2674 : * necessary for cases like "hstore ? text", wherein hstore's operators
2675 : * don't care about collation but the clause will get marked with a
2676 : * collation anyway because of the text argument. (This logic is
2677 : * embodied in the macro IndexCollMatchesExprColl.)
2678 : *
2679 : * It is also possible to match RowCompareExpr clauses to indexes (but
2680 : * currently, only btree indexes handle this).
2681 : *
2682 : * It is also possible to match ScalarArrayOpExpr clauses to indexes, when
2683 : * the clause is of the form "indexkey op ANY (arrayconst)".
2684 : *
2685 : * It is also possible to match a list of OR clauses if it might be
2686 : * transformed into a single ScalarArrayOpExpr clause. On success,
2687 : * the returning index clause will contain a transformed clause.
2688 : *
2689 : * For boolean indexes, it is also possible to match the clause directly
2690 : * to the indexkey; or perhaps the clause is (NOT indexkey).
2691 : *
2692 : * And, last but not least, some operators and functions can be processed
2693 : * to derive (typically lossy) indexquals from a clause that isn't in
2694 : * itself indexable. If we see that any operand of an OpExpr or FuncExpr
2695 : * matches the index key, and the function has a planner support function
2696 : * attached to it, we'll invoke the support function to see if such an
2697 : * indexqual can be built.
2698 : *
2699 : * 'rinfo' is the clause to be tested (as a RestrictInfo node).
2700 : * 'indexcol' is a column number of 'index' (counting from 0).
2701 : * 'index' is the index of interest.
2702 : *
2703 : * Returns an IndexClause if the clause can be used with this index key,
2704 : * or NULL if not.
2705 : *
2706 : * NOTE: This routine always returns NULL if the clause is an AND clause.
2707 : * Higher-level routines deal with OR and AND clauses. OR clause can be
2708 : * matched as a whole by match_orclause_to_indexcol() though.
2709 : */
2710 : static IndexClause *
2711 1698090 : match_clause_to_indexcol(PlannerInfo *root,
2712 : RestrictInfo *rinfo,
2713 : int indexcol,
2714 : IndexOptInfo *index)
2715 : {
2716 : IndexClause *iclause;
2717 1698090 : Expr *clause = rinfo->clause;
2718 : Oid opfamily;
2719 :
2720 : Assert(indexcol < index->nkeycolumns);
2721 :
2722 : /*
2723 : * Historically this code has coped with NULL clauses. That's probably
2724 : * not possible anymore, but we might as well continue to cope.
2725 : */
2726 1698090 : if (clause == NULL)
2727 0 : return NULL;
2728 :
2729 : /* First check for boolean-index cases. */
2730 1698090 : opfamily = index->opfamily[indexcol];
2731 1698090 : if (IsBooleanOpfamily(opfamily))
2732 : {
2733 456 : iclause = match_boolean_index_clause(root, rinfo, indexcol, index);
2734 456 : if (iclause)
2735 298 : return iclause;
2736 : }
2737 :
2738 : /*
2739 : * Clause must be an opclause, funcclause, ScalarArrayOpExpr,
2740 : * RowCompareExpr, or OR-clause that could be converted to SAOP. Or, if
2741 : * the index supports it, we can handle IS NULL/NOT NULL clauses.
2742 : */
2743 1697792 : if (IsA(clause, OpExpr))
2744 : {
2745 1426380 : return match_opclause_to_indexcol(root, rinfo, indexcol, index);
2746 : }
2747 271412 : else if (IsA(clause, FuncExpr))
2748 : {
2749 29280 : return match_funcclause_to_indexcol(root, rinfo, indexcol, index);
2750 : }
2751 242132 : else if (IsA(clause, ScalarArrayOpExpr))
2752 : {
2753 77600 : return match_saopclause_to_indexcol(root, rinfo, indexcol, index);
2754 : }
2755 164532 : else if (IsA(clause, RowCompareExpr))
2756 : {
2757 504 : return match_rowcompare_to_indexcol(root, rinfo, indexcol, index);
2758 : }
2759 164028 : else if (restriction_is_or_clause(rinfo))
2760 : {
2761 43332 : return match_orclause_to_indexcol(root, rinfo, indexcol, index);
2762 : }
2763 120696 : else if (index->amsearchnulls && IsA(clause, NullTest))
2764 : {
2765 15496 : NullTest *nt = (NullTest *) clause;
2766 :
2767 30992 : if (!nt->argisrow &&
2768 15496 : match_index_to_operand((Node *) nt->arg, indexcol, index))
2769 : {
2770 1448 : iclause = makeNode(IndexClause);
2771 1448 : iclause->rinfo = rinfo;
2772 1448 : iclause->indexquals = list_make1(rinfo);
2773 1448 : iclause->lossy = false;
2774 1448 : iclause->indexcol = indexcol;
2775 1448 : iclause->indexcols = NIL;
2776 1448 : return iclause;
2777 : }
2778 : }
2779 :
2780 119248 : return NULL;
2781 : }
2782 :
2783 : /*
2784 : * IsBooleanOpfamily
2785 : * Detect whether an opfamily supports boolean equality as an operator.
2786 : *
2787 : * If the opfamily OID is in the range of built-in objects, we can rely
2788 : * on hard-wired knowledge of which built-in opfamilies support this.
2789 : * For extension opfamilies, there's no choice but to do a catcache lookup.
2790 : */
2791 : static bool
2792 2352506 : IsBooleanOpfamily(Oid opfamily)
2793 : {
2794 2352506 : if (opfamily < FirstNormalObjectId)
2795 2348868 : return IsBuiltinBooleanOpfamily(opfamily);
2796 : else
2797 3638 : return op_in_opfamily(BooleanEqualOperator, opfamily);
2798 : }
2799 :
2800 : /*
2801 : * match_boolean_index_clause
2802 : * Recognize restriction clauses that can be matched to a boolean index.
2803 : *
2804 : * The idea here is that, for an index on a boolean column that supports the
2805 : * BooleanEqualOperator, we can transform a plain reference to the indexkey
2806 : * into "indexkey = true", or "NOT indexkey" into "indexkey = false", etc,
2807 : * so as to make the expression indexable using the index's "=" operator.
2808 : * Since Postgres 8.1, we must do this because constant simplification does
2809 : * the reverse transformation; without this code there'd be no way to use
2810 : * such an index at all.
2811 : *
2812 : * This should be called only when IsBooleanOpfamily() recognizes the
2813 : * index's operator family. We check to see if the clause matches the
2814 : * index's key, and if so, build a suitable IndexClause.
2815 : */
2816 : static IndexClause *
2817 1732 : match_boolean_index_clause(PlannerInfo *root,
2818 : RestrictInfo *rinfo,
2819 : int indexcol,
2820 : IndexOptInfo *index)
2821 : {
2822 1732 : Node *clause = (Node *) rinfo->clause;
2823 1732 : Expr *op = NULL;
2824 :
2825 : /* Direct match? */
2826 1732 : if (match_index_to_operand(clause, indexcol, index))
2827 : {
2828 : /* convert to indexkey = TRUE */
2829 294 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2830 : (Expr *) clause,
2831 294 : (Expr *) makeBoolConst(true, false),
2832 : InvalidOid, InvalidOid);
2833 : }
2834 : /* NOT clause? */
2835 1438 : else if (is_notclause(clause))
2836 : {
2837 1208 : Node *arg = (Node *) get_notclausearg((Expr *) clause);
2838 :
2839 1208 : if (match_index_to_operand(arg, indexcol, index))
2840 : {
2841 : /* convert to indexkey = FALSE */
2842 1208 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2843 : (Expr *) arg,
2844 1208 : (Expr *) makeBoolConst(false, false),
2845 : InvalidOid, InvalidOid);
2846 : }
2847 : }
2848 :
2849 : /*
2850 : * Since we only consider clauses at top level of WHERE, we can convert
2851 : * indexkey IS TRUE and indexkey IS FALSE to index searches as well. The
2852 : * different meaning for NULL isn't important.
2853 : */
2854 230 : else if (clause && IsA(clause, BooleanTest))
2855 : {
2856 36 : BooleanTest *btest = (BooleanTest *) clause;
2857 36 : Node *arg = (Node *) btest->arg;
2858 :
2859 54 : if (btest->booltesttype == IS_TRUE &&
2860 18 : match_index_to_operand(arg, indexcol, index))
2861 : {
2862 : /* convert to indexkey = TRUE */
2863 18 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2864 : (Expr *) arg,
2865 18 : (Expr *) makeBoolConst(true, false),
2866 : InvalidOid, InvalidOid);
2867 : }
2868 36 : else if (btest->booltesttype == IS_FALSE &&
2869 18 : match_index_to_operand(arg, indexcol, index))
2870 : {
2871 : /* convert to indexkey = FALSE */
2872 18 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2873 : (Expr *) arg,
2874 18 : (Expr *) makeBoolConst(false, false),
2875 : InvalidOid, InvalidOid);
2876 : }
2877 : }
2878 :
2879 : /*
2880 : * If we successfully made an operator clause from the given qual, we must
2881 : * wrap it in an IndexClause. It's not lossy.
2882 : */
2883 1732 : if (op)
2884 : {
2885 1538 : IndexClause *iclause = makeNode(IndexClause);
2886 :
2887 1538 : iclause->rinfo = rinfo;
2888 1538 : iclause->indexquals = list_make1(make_simple_restrictinfo(root, op));
2889 1538 : iclause->lossy = false;
2890 1538 : iclause->indexcol = indexcol;
2891 1538 : iclause->indexcols = NIL;
2892 1538 : return iclause;
2893 : }
2894 :
2895 194 : return NULL;
2896 : }
2897 :
2898 : /*
2899 : * match_opclause_to_indexcol()
2900 : * Handles the OpExpr case for match_clause_to_indexcol(),
2901 : * which see for comments.
2902 : */
2903 : static IndexClause *
2904 1426380 : match_opclause_to_indexcol(PlannerInfo *root,
2905 : RestrictInfo *rinfo,
2906 : int indexcol,
2907 : IndexOptInfo *index)
2908 : {
2909 : IndexClause *iclause;
2910 1426380 : OpExpr *clause = (OpExpr *) rinfo->clause;
2911 : Node *leftop,
2912 : *rightop;
2913 : Oid expr_op;
2914 : Oid expr_coll;
2915 : Index index_relid;
2916 : Oid opfamily;
2917 : Oid idxcollation;
2918 :
2919 : /*
2920 : * Only binary operators need apply. (In theory, a planner support
2921 : * function could do something with a unary operator, but it seems
2922 : * unlikely to be worth the cycles to check.)
2923 : */
2924 1426380 : if (list_length(clause->args) != 2)
2925 0 : return NULL;
2926 :
2927 1426380 : leftop = (Node *) linitial(clause->args);
2928 1426380 : rightop = (Node *) lsecond(clause->args);
2929 1426380 : expr_op = clause->opno;
2930 1426380 : expr_coll = clause->inputcollid;
2931 :
2932 1426380 : index_relid = index->rel->relid;
2933 1426380 : opfamily = index->opfamily[indexcol];
2934 1426380 : idxcollation = index->indexcollations[indexcol];
2935 :
2936 : /*
2937 : * Check for clauses of the form: (indexkey operator constant) or
2938 : * (constant operator indexkey). See match_clause_to_indexcol's notes
2939 : * about const-ness.
2940 : *
2941 : * Note that we don't ask the support function about clauses that don't
2942 : * have one of these forms. Again, in principle it might be possible to
2943 : * do something, but it seems unlikely to be worth the cycles to check.
2944 : */
2945 1426380 : if (match_index_to_operand(leftop, indexcol, index) &&
2946 344978 : !bms_is_member(index_relid, rinfo->right_relids) &&
2947 344804 : !contain_volatile_functions(rightop))
2948 : {
2949 682982 : if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
2950 338178 : op_in_opfamily(expr_op, opfamily))
2951 : {
2952 330684 : iclause = makeNode(IndexClause);
2953 330684 : iclause->rinfo = rinfo;
2954 330684 : iclause->indexquals = list_make1(rinfo);
2955 330684 : iclause->lossy = false;
2956 330684 : iclause->indexcol = indexcol;
2957 330684 : iclause->indexcols = NIL;
2958 330684 : return iclause;
2959 : }
2960 :
2961 : /*
2962 : * If we didn't find a member of the index's opfamily, try the support
2963 : * function for the operator's underlying function.
2964 : */
2965 14120 : set_opfuncid(clause); /* make sure we have opfuncid */
2966 14120 : return get_index_clause_from_support(root,
2967 : rinfo,
2968 : clause->opfuncid,
2969 : 0, /* indexarg on left */
2970 : indexcol,
2971 : index);
2972 : }
2973 :
2974 1081576 : if (match_index_to_operand(rightop, indexcol, index) &&
2975 67430 : !bms_is_member(index_relid, rinfo->left_relids) &&
2976 67304 : !contain_volatile_functions(leftop))
2977 : {
2978 67304 : if (IndexCollMatchesExprColl(idxcollation, expr_coll))
2979 : {
2980 67292 : Oid comm_op = get_commutator(expr_op);
2981 :
2982 134584 : if (OidIsValid(comm_op) &&
2983 67292 : op_in_opfamily(comm_op, opfamily))
2984 : {
2985 : RestrictInfo *commrinfo;
2986 :
2987 : /* Build a commuted OpExpr and RestrictInfo */
2988 66816 : commrinfo = commute_restrictinfo(rinfo, comm_op);
2989 :
2990 : /* Make an IndexClause showing that as a derived qual */
2991 66816 : iclause = makeNode(IndexClause);
2992 66816 : iclause->rinfo = rinfo;
2993 66816 : iclause->indexquals = list_make1(commrinfo);
2994 66816 : iclause->lossy = false;
2995 66816 : iclause->indexcol = indexcol;
2996 66816 : iclause->indexcols = NIL;
2997 66816 : return iclause;
2998 : }
2999 : }
3000 :
3001 : /*
3002 : * If we didn't find a member of the index's opfamily, try the support
3003 : * function for the operator's underlying function.
3004 : */
3005 488 : set_opfuncid(clause); /* make sure we have opfuncid */
3006 488 : return get_index_clause_from_support(root,
3007 : rinfo,
3008 : clause->opfuncid,
3009 : 1, /* indexarg on right */
3010 : indexcol,
3011 : index);
3012 : }
3013 :
3014 1014272 : return NULL;
3015 : }
3016 :
3017 : /*
3018 : * match_funcclause_to_indexcol()
3019 : * Handles the FuncExpr case for match_clause_to_indexcol(),
3020 : * which see for comments.
3021 : */
3022 : static IndexClause *
3023 29280 : match_funcclause_to_indexcol(PlannerInfo *root,
3024 : RestrictInfo *rinfo,
3025 : int indexcol,
3026 : IndexOptInfo *index)
3027 : {
3028 29280 : FuncExpr *clause = (FuncExpr *) rinfo->clause;
3029 : int indexarg;
3030 : ListCell *lc;
3031 :
3032 : /*
3033 : * We have no built-in intelligence about function clauses, but if there's
3034 : * a planner support function, it might be able to do something. But, to
3035 : * cut down on wasted planning cycles, only call the support function if
3036 : * at least one argument matches the target index column.
3037 : *
3038 : * Note that we don't insist on the other arguments being pseudoconstants;
3039 : * the support function has to check that. This is to allow cases where
3040 : * only some of the other arguments need to be included in the indexqual.
3041 : */
3042 29280 : indexarg = 0;
3043 63012 : foreach(lc, clause->args)
3044 : {
3045 39444 : Node *op = (Node *) lfirst(lc);
3046 :
3047 39444 : if (match_index_to_operand(op, indexcol, index))
3048 : {
3049 5712 : return get_index_clause_from_support(root,
3050 : rinfo,
3051 : clause->funcid,
3052 : indexarg,
3053 : indexcol,
3054 : index);
3055 : }
3056 :
3057 33732 : indexarg++;
3058 : }
3059 :
3060 23568 : return NULL;
3061 : }
3062 :
3063 : /*
3064 : * get_index_clause_from_support()
3065 : * If the function has a planner support function, try to construct
3066 : * an IndexClause using indexquals created by the support function.
3067 : */
3068 : static IndexClause *
3069 20320 : get_index_clause_from_support(PlannerInfo *root,
3070 : RestrictInfo *rinfo,
3071 : Oid funcid,
3072 : int indexarg,
3073 : int indexcol,
3074 : IndexOptInfo *index)
3075 : {
3076 20320 : Oid prosupport = get_func_support(funcid);
3077 : SupportRequestIndexCondition req;
3078 : List *sresult;
3079 :
3080 20320 : if (!OidIsValid(prosupport))
3081 12224 : return NULL;
3082 :
3083 8096 : req.type = T_SupportRequestIndexCondition;
3084 8096 : req.root = root;
3085 8096 : req.funcid = funcid;
3086 8096 : req.node = (Node *) rinfo->clause;
3087 8096 : req.indexarg = indexarg;
3088 8096 : req.index = index;
3089 8096 : req.indexcol = indexcol;
3090 8096 : req.opfamily = index->opfamily[indexcol];
3091 8096 : req.indexcollation = index->indexcollations[indexcol];
3092 :
3093 8096 : req.lossy = true; /* default assumption */
3094 :
3095 : sresult = (List *)
3096 8096 : DatumGetPointer(OidFunctionCall1(prosupport,
3097 : PointerGetDatum(&req)));
3098 :
3099 8096 : if (sresult != NIL)
3100 : {
3101 1402 : IndexClause *iclause = makeNode(IndexClause);
3102 1402 : List *indexquals = NIL;
3103 : ListCell *lc;
3104 :
3105 : /*
3106 : * The support function API says it should just give back bare
3107 : * clauses, so here we must wrap each one in a RestrictInfo.
3108 : */
3109 4128 : foreach(lc, sresult)
3110 : {
3111 2726 : Expr *clause = (Expr *) lfirst(lc);
3112 :
3113 2726 : indexquals = lappend(indexquals,
3114 2726 : make_simple_restrictinfo(root, clause));
3115 : }
3116 :
3117 1402 : iclause->rinfo = rinfo;
3118 1402 : iclause->indexquals = indexquals;
3119 1402 : iclause->lossy = req.lossy;
3120 1402 : iclause->indexcol = indexcol;
3121 1402 : iclause->indexcols = NIL;
3122 :
3123 1402 : return iclause;
3124 : }
3125 :
3126 6694 : return NULL;
3127 : }
3128 :
3129 : /*
3130 : * match_saopclause_to_indexcol()
3131 : * Handles the ScalarArrayOpExpr case for match_clause_to_indexcol(),
3132 : * which see for comments.
3133 : */
3134 : static IndexClause *
3135 77600 : match_saopclause_to_indexcol(PlannerInfo *root,
3136 : RestrictInfo *rinfo,
3137 : int indexcol,
3138 : IndexOptInfo *index)
3139 : {
3140 77600 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
3141 : Node *leftop,
3142 : *rightop;
3143 : Relids right_relids;
3144 : Oid expr_op;
3145 : Oid expr_coll;
3146 : Index index_relid;
3147 : Oid opfamily;
3148 : Oid idxcollation;
3149 :
3150 : /* We only accept ANY clauses, not ALL */
3151 77600 : if (!saop->useOr)
3152 9094 : return NULL;
3153 68506 : leftop = (Node *) linitial(saop->args);
3154 68506 : rightop = (Node *) lsecond(saop->args);
3155 68506 : right_relids = pull_varnos(root, rightop);
3156 68506 : expr_op = saop->opno;
3157 68506 : expr_coll = saop->inputcollid;
3158 :
3159 68506 : index_relid = index->rel->relid;
3160 68506 : opfamily = index->opfamily[indexcol];
3161 68506 : idxcollation = index->indexcollations[indexcol];
3162 :
3163 : /*
3164 : * We must have indexkey on the left and a pseudo-constant array argument.
3165 : */
3166 68506 : if (match_index_to_operand(leftop, indexcol, index) &&
3167 7064 : !bms_is_member(index_relid, right_relids) &&
3168 7064 : !contain_volatile_functions(rightop))
3169 : {
3170 14122 : if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
3171 7058 : op_in_opfamily(expr_op, opfamily))
3172 : {
3173 7046 : IndexClause *iclause = makeNode(IndexClause);
3174 :
3175 7046 : iclause->rinfo = rinfo;
3176 7046 : iclause->indexquals = list_make1(rinfo);
3177 7046 : iclause->lossy = false;
3178 7046 : iclause->indexcol = indexcol;
3179 7046 : iclause->indexcols = NIL;
3180 7046 : return iclause;
3181 : }
3182 :
3183 : /*
3184 : * We do not currently ask support functions about ScalarArrayOpExprs,
3185 : * though in principle we could.
3186 : */
3187 : }
3188 :
3189 61460 : return NULL;
3190 : }
3191 :
3192 : /*
3193 : * match_rowcompare_to_indexcol()
3194 : * Handles the RowCompareExpr case for match_clause_to_indexcol(),
3195 : * which see for comments.
3196 : *
3197 : * In this routine we check whether the first column of the row comparison
3198 : * matches the target index column. This is sufficient to guarantee that some
3199 : * index condition can be constructed from the RowCompareExpr --- the rest
3200 : * is handled by expand_indexqual_rowcompare().
3201 : */
3202 : static IndexClause *
3203 504 : match_rowcompare_to_indexcol(PlannerInfo *root,
3204 : RestrictInfo *rinfo,
3205 : int indexcol,
3206 : IndexOptInfo *index)
3207 : {
3208 504 : RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3209 : Index index_relid;
3210 : Oid opfamily;
3211 : Oid idxcollation;
3212 : Node *leftop,
3213 : *rightop;
3214 : bool var_on_left;
3215 : Oid expr_op;
3216 : Oid expr_coll;
3217 :
3218 : /* Forget it if we're not dealing with a btree index */
3219 504 : if (index->relam != BTREE_AM_OID)
3220 0 : return NULL;
3221 :
3222 504 : index_relid = index->rel->relid;
3223 504 : opfamily = index->opfamily[indexcol];
3224 504 : idxcollation = index->indexcollations[indexcol];
3225 :
3226 : /*
3227 : * We could do the matching on the basis of insisting that the opfamily
3228 : * shown in the RowCompareExpr be the same as the index column's opfamily,
3229 : * but that could fail in the presence of reverse-sort opfamilies: it'd be
3230 : * a matter of chance whether RowCompareExpr had picked the forward or
3231 : * reverse-sort family. So look only at the operator, and match if it is
3232 : * a member of the index's opfamily (after commutation, if the indexkey is
3233 : * on the right). We'll worry later about whether any additional
3234 : * operators are matchable to the index.
3235 : */
3236 504 : leftop = (Node *) linitial(clause->largs);
3237 504 : rightop = (Node *) linitial(clause->rargs);
3238 504 : expr_op = linitial_oid(clause->opnos);
3239 504 : expr_coll = linitial_oid(clause->inputcollids);
3240 :
3241 : /* Collations must match, if relevant */
3242 504 : if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
3243 0 : return NULL;
3244 :
3245 : /*
3246 : * These syntactic tests are the same as in match_opclause_to_indexcol()
3247 : */
3248 504 : if (match_index_to_operand(leftop, indexcol, index) &&
3249 162 : !bms_is_member(index_relid, pull_varnos(root, rightop)) &&
3250 162 : !contain_volatile_functions(rightop))
3251 : {
3252 : /* OK, indexkey is on left */
3253 162 : var_on_left = true;
3254 : }
3255 342 : else if (match_index_to_operand(rightop, indexcol, index) &&
3256 24 : !bms_is_member(index_relid, pull_varnos(root, leftop)) &&
3257 24 : !contain_volatile_functions(leftop))
3258 : {
3259 : /* indexkey is on right, so commute the operator */
3260 24 : expr_op = get_commutator(expr_op);
3261 24 : if (expr_op == InvalidOid)
3262 0 : return NULL;
3263 24 : var_on_left = false;
3264 : }
3265 : else
3266 318 : return NULL;
3267 :
3268 : /* We're good if the operator is the right type of opfamily member */
3269 186 : switch (get_op_opfamily_strategy(expr_op, opfamily))
3270 : {
3271 186 : case BTLessStrategyNumber:
3272 : case BTLessEqualStrategyNumber:
3273 : case BTGreaterEqualStrategyNumber:
3274 : case BTGreaterStrategyNumber:
3275 186 : return expand_indexqual_rowcompare(root,
3276 : rinfo,
3277 : indexcol,
3278 : index,
3279 : expr_op,
3280 : var_on_left);
3281 : }
3282 :
3283 0 : return NULL;
3284 : }
3285 :
3286 : /*
3287 : * match_orclause_to_indexcol()
3288 : * Handles the OR-expr case for match_clause_to_indexcol() in the case
3289 : * when it could be transformed to ScalarArrayOpExpr.
3290 : *
3291 : * In this routine, we attempt to transform a list of OR-clause args into a
3292 : * single SAOP expression matching the target index column. On success,
3293 : * return an IndexClause, containing the transformed expression or NULL,
3294 : * if failed.
3295 : */
3296 : static IndexClause *
3297 43332 : match_orclause_to_indexcol(PlannerInfo *root,
3298 : RestrictInfo *rinfo,
3299 : int indexcol,
3300 : IndexOptInfo *index)
3301 : {
3302 : ListCell *lc;
3303 43332 : BoolExpr *orclause = (BoolExpr *) rinfo->orclause;
3304 43332 : Node *indexExpr = NULL;
3305 43332 : List *consts = NIL;
3306 43332 : ScalarArrayOpExpr *saopexpr = NULL;
3307 43332 : Oid matchOpno = InvalidOid;
3308 : IndexClause *iclause;
3309 43332 : Oid consttype = InvalidOid;
3310 43332 : Oid arraytype = InvalidOid;
3311 43332 : Oid inputcollid = InvalidOid;
3312 43332 : bool firstTime = true;
3313 43332 : bool haveNonConst = false;
3314 43332 : Index indexRelid = index->rel->relid;
3315 :
3316 : Assert(IsA(orclause, BoolExpr));
3317 : Assert(orclause->boolop == OR_EXPR);
3318 :
3319 : /* Ignore index if it doesn't support SAOP clauses */
3320 43332 : if (!index->amsearcharray)
3321 106 : return NULL;
3322 :
3323 : /*
3324 : * Try to convert a list of OR-clauses to a single SAOP expression. Each
3325 : * OR entry must be in the form: (indexkey operator constant) or (constant
3326 : * operator indexkey). Operators of all the entries must match. On
3327 : * discovery of anything unsupported, we give up by breaking out of the
3328 : * loop immediately and returning NULL.
3329 : */
3330 47762 : foreach(lc, orclause->args)
3331 : {
3332 : RestrictInfo *subRinfo;
3333 : OpExpr *subClause;
3334 : Oid opno;
3335 : Node *leftop,
3336 : *rightop;
3337 : Node *constExpr;
3338 :
3339 46712 : if (!IsA(lfirst(lc), RestrictInfo))
3340 5124 : break;
3341 :
3342 41588 : subRinfo = (RestrictInfo *) lfirst(lc);
3343 :
3344 : /* Only operator clauses can match */
3345 41588 : if (!IsA(subRinfo->clause, OpExpr))
3346 12734 : break;
3347 :
3348 28854 : subClause = (OpExpr *) subRinfo->clause;
3349 28854 : opno = subClause->opno;
3350 :
3351 : /* Only binary operators can match */
3352 28854 : if (list_length(subClause->args) != 2)
3353 0 : break;
3354 :
3355 : /*
3356 : * The parameters below must match between sub-rinfo and its parent as
3357 : * make_restrictinfo() fills them with the same values, and further
3358 : * modifications are also the same for the whole subtree. However,
3359 : * still make a sanity check.
3360 : */
3361 : Assert(subRinfo->is_pushed_down == rinfo->is_pushed_down);
3362 : Assert(subRinfo->is_clone == rinfo->is_clone);
3363 : Assert(subRinfo->security_level == rinfo->security_level);
3364 : Assert(bms_equal(subRinfo->incompatible_relids, rinfo->incompatible_relids));
3365 : Assert(bms_equal(subRinfo->outer_relids, rinfo->outer_relids));
3366 :
3367 : /*
3368 : * Also, check that required_relids in sub-rinfo is subset of parent's
3369 : * required_relids.
3370 : */
3371 : Assert(bms_is_subset(subRinfo->required_relids, rinfo->required_relids));
3372 :
3373 : /* Only the operator returning a boolean suit the transformation. */
3374 28854 : if (get_op_rettype(opno) != BOOLOID)
3375 0 : break;
3376 :
3377 : /*
3378 : * Check for clauses of the form: (indexkey operator constant) or
3379 : * (constant operator indexkey). See match_clause_to_indexcol's notes
3380 : * about const-ness.
3381 : */
3382 28854 : leftop = (Node *) linitial(subClause->args);
3383 28854 : rightop = (Node *) lsecond(subClause->args);
3384 28854 : if (match_index_to_operand(leftop, indexcol, index) &&
3385 6360 : !bms_is_member(indexRelid, subRinfo->right_relids) &&
3386 6330 : !contain_volatile_functions(rightop))
3387 : {
3388 6330 : indexExpr = leftop;
3389 6330 : constExpr = rightop;
3390 : }
3391 22524 : else if (match_index_to_operand(rightop, indexcol, index) &&
3392 194 : !bms_is_member(indexRelid, subRinfo->left_relids) &&
3393 188 : !contain_volatile_functions(leftop))
3394 : {
3395 188 : opno = get_commutator(opno);
3396 188 : if (!OidIsValid(opno))
3397 : {
3398 : /* commutator doesn't exist, we can't reverse the order */
3399 0 : break;
3400 : }
3401 188 : indexExpr = rightop;
3402 188 : constExpr = leftop;
3403 : }
3404 : else
3405 : {
3406 : break;
3407 : }
3408 :
3409 : /*
3410 : * Ignore any RelabelType node above the operands. This is needed to
3411 : * be able to apply indexscanning in binary-compatible-operator cases.
3412 : * Note: we can assume there is at most one RelabelType node;
3413 : * eval_const_expressions() will have simplified if more than one.
3414 : */
3415 6518 : if (IsA(constExpr, RelabelType))
3416 0 : constExpr = (Node *) ((RelabelType *) constExpr)->arg;
3417 6518 : if (IsA(indexExpr, RelabelType))
3418 12 : indexExpr = (Node *) ((RelabelType *) indexExpr)->arg;
3419 :
3420 : /* Forbid transformation for composite types, records. */
3421 13036 : if (type_is_rowtype(exprType(constExpr)) ||
3422 6518 : type_is_rowtype(exprType(indexExpr)))
3423 : break;
3424 :
3425 : /*
3426 : * Save information about the operator, type, and collation for the
3427 : * first matching qual. Then, check that subsequent quals match the
3428 : * first.
3429 : */
3430 6518 : if (firstTime)
3431 : {
3432 4760 : matchOpno = opno;
3433 4760 : consttype = exprType(constExpr);
3434 4760 : arraytype = get_array_type(consttype);
3435 4760 : inputcollid = subClause->inputcollid;
3436 :
3437 : /*
3438 : * Check that the operator is presented in the opfamily and that
3439 : * the expression collation matches the index collation. Also,
3440 : * there must be an array type to construct an array later.
3441 : */
3442 4760 : if (!IndexCollMatchesExprColl(index->indexcollations[indexcol], inputcollid) ||
3443 4634 : !op_in_opfamily(matchOpno, index->opfamily[indexcol]) ||
3444 : !OidIsValid(arraytype))
3445 : break;
3446 2922 : firstTime = false;
3447 : }
3448 : else
3449 : {
3450 1758 : if (opno != matchOpno ||
3451 3228 : inputcollid != subClause->inputcollid ||
3452 1614 : consttype != exprType(constExpr))
3453 : break;
3454 : }
3455 :
3456 : /*
3457 : * Check if our list of constants in match_clause_to_indexcol's
3458 : * understanding of const-ness have something other than Const.
3459 : */
3460 4536 : if (!IsA(constExpr, Const))
3461 368 : haveNonConst = true;
3462 4536 : consts = lappend(consts, constExpr);
3463 : }
3464 :
3465 : /*
3466 : * Handle failed conversion from breaking out of the loop because of an
3467 : * unsupported qual. Free the consts list and return NULL to indicate the
3468 : * conversion failed.
3469 : */
3470 43226 : if (lc != NULL)
3471 : {
3472 42176 : list_free(consts);
3473 42176 : return NULL;
3474 : }
3475 :
3476 1050 : saopexpr = make_SAOP_expr(matchOpno, indexExpr, consttype, inputcollid,
3477 : inputcollid, consts, haveNonConst);
3478 :
3479 : /*
3480 : * Finally, build an IndexClause based on the SAOP node. Use
3481 : * make_simple_restrictinfo() to get RestrictInfo with clean selectivity
3482 : * estimations, because they may differ from the estimation made for an OR
3483 : * clause. Although it is not a lossy expression, keep the original rinfo
3484 : * in iclause->rinfo as prescribed.
3485 : */
3486 1050 : iclause = makeNode(IndexClause);
3487 1050 : iclause->rinfo = rinfo;
3488 1050 : iclause->indexquals = list_make1(make_simple_restrictinfo(root,
3489 : &saopexpr->xpr));
3490 1050 : iclause->lossy = false;
3491 1050 : iclause->indexcol = indexcol;
3492 1050 : iclause->indexcols = NIL;
3493 1050 : return iclause;
3494 : }
3495 :
3496 : /*
3497 : * expand_indexqual_rowcompare --- expand a single indexqual condition
3498 : * that is a RowCompareExpr
3499 : *
3500 : * It's already known that the first column of the row comparison matches
3501 : * the specified column of the index. We can use additional columns of the
3502 : * row comparison as index qualifications, so long as they match the index
3503 : * in the "same direction", ie, the indexkeys are all on the same side of the
3504 : * clause and the operators are all the same-type members of the opfamilies.
3505 : *
3506 : * If all the columns of the RowCompareExpr match in this way, we just use it
3507 : * as-is, except for possibly commuting it to put the indexkeys on the left.
3508 : *
3509 : * Otherwise, we build a shortened RowCompareExpr (if more than one
3510 : * column matches) or a simple OpExpr (if the first-column match is all
3511 : * there is). In these cases the modified clause is always "<=" or ">="
3512 : * even when the original was "<" or ">" --- this is necessary to match all
3513 : * the rows that could match the original. (We are building a lossy version
3514 : * of the row comparison when we do this, so we set lossy = true.)
3515 : *
3516 : * Note: this is really just the last half of match_rowcompare_to_indexcol,
3517 : * but we split it out for comprehensibility.
3518 : */
3519 : static IndexClause *
3520 186 : expand_indexqual_rowcompare(PlannerInfo *root,
3521 : RestrictInfo *rinfo,
3522 : int indexcol,
3523 : IndexOptInfo *index,
3524 : Oid expr_op,
3525 : bool var_on_left)
3526 : {
3527 186 : IndexClause *iclause = makeNode(IndexClause);
3528 186 : RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3529 : int op_strategy;
3530 : Oid op_lefttype;
3531 : Oid op_righttype;
3532 : int matching_cols;
3533 : List *expr_ops;
3534 : List *opfamilies;
3535 : List *lefttypes;
3536 : List *righttypes;
3537 : List *new_ops;
3538 : List *var_args;
3539 : List *non_var_args;
3540 :
3541 186 : iclause->rinfo = rinfo;
3542 186 : iclause->indexcol = indexcol;
3543 :
3544 186 : if (var_on_left)
3545 : {
3546 162 : var_args = clause->largs;
3547 162 : non_var_args = clause->rargs;
3548 : }
3549 : else
3550 : {
3551 24 : var_args = clause->rargs;
3552 24 : non_var_args = clause->largs;
3553 : }
3554 :
3555 186 : get_op_opfamily_properties(expr_op, index->opfamily[indexcol], false,
3556 : &op_strategy,
3557 : &op_lefttype,
3558 : &op_righttype);
3559 :
3560 : /* Initialize returned list of which index columns are used */
3561 186 : iclause->indexcols = list_make1_int(indexcol);
3562 :
3563 : /* Build lists of ops, opfamilies and operator datatypes in case needed */
3564 186 : expr_ops = list_make1_oid(expr_op);
3565 186 : opfamilies = list_make1_oid(index->opfamily[indexcol]);
3566 186 : lefttypes = list_make1_oid(op_lefttype);
3567 186 : righttypes = list_make1_oid(op_righttype);
3568 :
3569 : /*
3570 : * See how many of the remaining columns match some index column in the
3571 : * same way. As in match_clause_to_indexcol(), the "other" side of any
3572 : * potential index condition is OK as long as it doesn't use Vars from the
3573 : * indexed relation.
3574 : */
3575 186 : matching_cols = 1;
3576 :
3577 354 : while (matching_cols < list_length(var_args))
3578 : {
3579 222 : Node *varop = (Node *) list_nth(var_args, matching_cols);
3580 222 : Node *constop = (Node *) list_nth(non_var_args, matching_cols);
3581 : int i;
3582 :
3583 222 : expr_op = list_nth_oid(clause->opnos, matching_cols);
3584 222 : if (!var_on_left)
3585 : {
3586 : /* indexkey is on right, so commute the operator */
3587 24 : expr_op = get_commutator(expr_op);
3588 24 : if (expr_op == InvalidOid)
3589 0 : break; /* operator is not usable */
3590 : }
3591 222 : if (bms_is_member(index->rel->relid, pull_varnos(root, constop)))
3592 0 : break; /* no good, Var on wrong side */
3593 222 : if (contain_volatile_functions(constop))
3594 0 : break; /* no good, volatile comparison value */
3595 :
3596 : /*
3597 : * The Var side can match any key column of the index.
3598 : */
3599 516 : for (i = 0; i < index->nkeycolumns; i++)
3600 : {
3601 462 : if (match_index_to_operand(varop, i, index) &&
3602 168 : get_op_opfamily_strategy(expr_op,
3603 168 : index->opfamily[i]) == op_strategy &&
3604 168 : IndexCollMatchesExprColl(index->indexcollations[i],
3605 : list_nth_oid(clause->inputcollids,
3606 : matching_cols)))
3607 : break;
3608 : }
3609 222 : if (i >= index->nkeycolumns)
3610 54 : break; /* no match found */
3611 :
3612 : /* Add column number to returned list */
3613 168 : iclause->indexcols = lappend_int(iclause->indexcols, i);
3614 :
3615 : /* Add operator info to lists */
3616 168 : get_op_opfamily_properties(expr_op, index->opfamily[i], false,
3617 : &op_strategy,
3618 : &op_lefttype,
3619 : &op_righttype);
3620 168 : expr_ops = lappend_oid(expr_ops, expr_op);
3621 168 : opfamilies = lappend_oid(opfamilies, index->opfamily[i]);
3622 168 : lefttypes = lappend_oid(lefttypes, op_lefttype);
3623 168 : righttypes = lappend_oid(righttypes, op_righttype);
3624 :
3625 : /* This column matches, keep scanning */
3626 168 : matching_cols++;
3627 : }
3628 :
3629 : /* Result is non-lossy if all columns are usable as index quals */
3630 186 : iclause->lossy = (matching_cols != list_length(clause->opnos));
3631 :
3632 : /*
3633 : * We can use rinfo->clause as-is if we have var on left and it's all
3634 : * usable as index quals.
3635 : */
3636 186 : if (var_on_left && !iclause->lossy)
3637 120 : iclause->indexquals = list_make1(rinfo);
3638 : else
3639 : {
3640 : /*
3641 : * We have to generate a modified rowcompare (possibly just one
3642 : * OpExpr). The painful part of this is changing < to <= or > to >=,
3643 : * so deal with that first.
3644 : */
3645 66 : if (!iclause->lossy)
3646 : {
3647 : /* very easy, just use the commuted operators */
3648 12 : new_ops = expr_ops;
3649 : }
3650 54 : else if (op_strategy == BTLessEqualStrategyNumber ||
3651 54 : op_strategy == BTGreaterEqualStrategyNumber)
3652 : {
3653 : /* easy, just use the same (possibly commuted) operators */
3654 0 : new_ops = list_truncate(expr_ops, matching_cols);
3655 : }
3656 : else
3657 : {
3658 : ListCell *opfamilies_cell;
3659 : ListCell *lefttypes_cell;
3660 : ListCell *righttypes_cell;
3661 :
3662 54 : if (op_strategy == BTLessStrategyNumber)
3663 30 : op_strategy = BTLessEqualStrategyNumber;
3664 24 : else if (op_strategy == BTGreaterStrategyNumber)
3665 24 : op_strategy = BTGreaterEqualStrategyNumber;
3666 : else
3667 0 : elog(ERROR, "unexpected strategy number %d", op_strategy);
3668 54 : new_ops = NIL;
3669 144 : forthree(opfamilies_cell, opfamilies,
3670 : lefttypes_cell, lefttypes,
3671 : righttypes_cell, righttypes)
3672 : {
3673 90 : Oid opfam = lfirst_oid(opfamilies_cell);
3674 90 : Oid lefttype = lfirst_oid(lefttypes_cell);
3675 90 : Oid righttype = lfirst_oid(righttypes_cell);
3676 :
3677 90 : expr_op = get_opfamily_member(opfam, lefttype, righttype,
3678 : op_strategy);
3679 90 : if (!OidIsValid(expr_op)) /* should not happen */
3680 0 : elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
3681 : op_strategy, lefttype, righttype, opfam);
3682 90 : new_ops = lappend_oid(new_ops, expr_op);
3683 : }
3684 : }
3685 :
3686 : /* If we have more than one matching col, create a subset rowcompare */
3687 66 : if (matching_cols > 1)
3688 : {
3689 48 : RowCompareExpr *rc = makeNode(RowCompareExpr);
3690 :
3691 48 : rc->cmptype = (CompareType) op_strategy;
3692 48 : rc->opnos = new_ops;
3693 48 : rc->opfamilies = list_copy_head(clause->opfamilies,
3694 : matching_cols);
3695 48 : rc->inputcollids = list_copy_head(clause->inputcollids,
3696 : matching_cols);
3697 48 : rc->largs = list_copy_head(var_args, matching_cols);
3698 48 : rc->rargs = list_copy_head(non_var_args, matching_cols);
3699 48 : iclause->indexquals = list_make1(make_simple_restrictinfo(root,
3700 : (Expr *) rc));
3701 : }
3702 : else
3703 : {
3704 : Expr *op;
3705 :
3706 : /* We don't report an index column list in this case */
3707 18 : iclause->indexcols = NIL;
3708 :
3709 18 : op = make_opclause(linitial_oid(new_ops), BOOLOID, false,
3710 18 : copyObject(linitial(var_args)),
3711 18 : copyObject(linitial(non_var_args)),
3712 : InvalidOid,
3713 18 : linitial_oid(clause->inputcollids));
3714 18 : iclause->indexquals = list_make1(make_simple_restrictinfo(root, op));
3715 : }
3716 : }
3717 :
3718 186 : return iclause;
3719 : }
3720 :
3721 :
3722 : /****************************************************************************
3723 : * ---- ROUTINES TO CHECK ORDERING OPERATORS ----
3724 : ****************************************************************************/
3725 :
3726 : /*
3727 : * match_pathkeys_to_index
3728 : * For the given 'index' and 'pathkeys', output a list of suitable ORDER
3729 : * BY expressions, each of the form "indexedcol operator pseudoconstant",
3730 : * along with an integer list of the index column numbers (zero based)
3731 : * that each clause would be used with.
3732 : *
3733 : * This attempts to find an ORDER BY and index column number for all items in
3734 : * the pathkey list, however, if we're unable to match any given pathkey to an
3735 : * index column, we return just the ones matched by the function so far. This
3736 : * allows callers who are interested in partial matches to get them. Callers
3737 : * can determine a partial match vs a full match by checking the outputted
3738 : * list lengths. A full match will have one item in the output lists for each
3739 : * item in the given 'pathkeys' list.
3740 : */
3741 : static void
3742 1074 : match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
3743 : List **orderby_clauses_p,
3744 : List **clause_columns_p)
3745 : {
3746 : ListCell *lc1;
3747 :
3748 1074 : *orderby_clauses_p = NIL; /* set default results */
3749 1074 : *clause_columns_p = NIL;
3750 :
3751 : /* Only indexes with the amcanorderbyop property are interesting here */
3752 1074 : if (!index->amcanorderbyop)
3753 0 : return;
3754 :
3755 1548 : foreach(lc1, pathkeys)
3756 : {
3757 1080 : PathKey *pathkey = (PathKey *) lfirst(lc1);
3758 1080 : bool found = false;
3759 : EquivalenceMemberIterator it;
3760 : EquivalenceMember *member;
3761 :
3762 :
3763 : /* Pathkey must request default sort order for the target opfamily */
3764 1080 : if (pathkey->pk_cmptype != COMPARE_LT || pathkey->pk_nulls_first)
3765 606 : return;
3766 :
3767 : /* If eclass is volatile, no hope of using an indexscan */
3768 1046 : if (pathkey->pk_eclass->ec_has_volatile)
3769 0 : return;
3770 :
3771 : /*
3772 : * Try to match eclass member expression(s) to index. Note that child
3773 : * EC members are considered, but only when they belong to the target
3774 : * relation. (Unlike regular members, the same expression could be a
3775 : * child member of more than one EC. Therefore, the same index could
3776 : * be considered to match more than one pathkey list, which is OK
3777 : * here. See also get_eclass_for_sort_expr.)
3778 : */
3779 1046 : setup_eclass_member_iterator(&it, pathkey->pk_eclass,
3780 1046 : index->rel->relids);
3781 1650 : while ((member = eclass_member_iterator_next(&it)) != NULL)
3782 : {
3783 : int indexcol;
3784 :
3785 : /* No possibility of match if it references other relations */
3786 1078 : if (!bms_equal(member->em_relids, index->rel->relids))
3787 32 : continue;
3788 :
3789 : /*
3790 : * We allow any column of the index to match each pathkey; they
3791 : * don't have to match left-to-right as you might expect. This is
3792 : * correct for GiST, and it doesn't matter for SP-GiST because
3793 : * that doesn't handle multiple columns anyway, and no other
3794 : * existing AMs support amcanorderbyop. We might need different
3795 : * logic in future for other implementations.
3796 : */
3797 1906 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
3798 : {
3799 : Expr *expr;
3800 :
3801 1334 : expr = match_clause_to_ordering_op(index,
3802 : indexcol,
3803 : member->em_expr,
3804 : pathkey->pk_opfamily);
3805 1334 : if (expr)
3806 : {
3807 474 : *orderby_clauses_p = lappend(*orderby_clauses_p, expr);
3808 474 : *clause_columns_p = lappend_int(*clause_columns_p, indexcol);
3809 474 : found = true;
3810 474 : break;
3811 : }
3812 : }
3813 :
3814 1046 : if (found) /* don't want to look at remaining members */
3815 474 : break;
3816 : }
3817 :
3818 : /*
3819 : * Return the matches found so far when this pathkey couldn't be
3820 : * matched to the index.
3821 : */
3822 1046 : if (!found)
3823 572 : return;
3824 : }
3825 : }
3826 :
3827 : /*
3828 : * match_clause_to_ordering_op
3829 : * Determines whether an ordering operator expression matches an
3830 : * index column.
3831 : *
3832 : * This is similar to, but simpler than, match_clause_to_indexcol.
3833 : * We only care about simple OpExpr cases. The input is a bare
3834 : * expression that is being ordered by, which must be of the form
3835 : * (indexkey op const) or (const op indexkey) where op is an ordering
3836 : * operator for the column's opfamily.
3837 : *
3838 : * 'index' is the index of interest.
3839 : * 'indexcol' is a column number of 'index' (counting from 0).
3840 : * 'clause' is the ordering expression to be tested.
3841 : * 'pk_opfamily' is the btree opfamily describing the required sort order.
3842 : *
3843 : * Note that we currently do not consider the collation of the ordering
3844 : * operator's result. In practical cases the result type will be numeric
3845 : * and thus have no collation, and it's not very clear what to match to
3846 : * if it did have a collation. The index's collation should match the
3847 : * ordering operator's input collation, not its result.
3848 : *
3849 : * If successful, return 'clause' as-is if the indexkey is on the left,
3850 : * otherwise a commuted copy of 'clause'. If no match, return NULL.
3851 : */
3852 : static Expr *
3853 1334 : match_clause_to_ordering_op(IndexOptInfo *index,
3854 : int indexcol,
3855 : Expr *clause,
3856 : Oid pk_opfamily)
3857 : {
3858 : Oid opfamily;
3859 : Oid idxcollation;
3860 : Node *leftop,
3861 : *rightop;
3862 : Oid expr_op;
3863 : Oid expr_coll;
3864 : Oid sortfamily;
3865 : bool commuted;
3866 :
3867 : Assert(indexcol < index->nkeycolumns);
3868 :
3869 1334 : opfamily = index->opfamily[indexcol];
3870 1334 : idxcollation = index->indexcollations[indexcol];
3871 :
3872 : /*
3873 : * Clause must be a binary opclause.
3874 : */
3875 1334 : if (!is_opclause(clause))
3876 860 : return NULL;
3877 474 : leftop = get_leftop(clause);
3878 474 : rightop = get_rightop(clause);
3879 474 : if (!leftop || !rightop)
3880 0 : return NULL;
3881 474 : expr_op = ((OpExpr *) clause)->opno;
3882 474 : expr_coll = ((OpExpr *) clause)->inputcollid;
3883 :
3884 : /*
3885 : * We can forget the whole thing right away if wrong collation.
3886 : */
3887 474 : if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
3888 0 : return NULL;
3889 :
3890 : /*
3891 : * Check for clauses of the form: (indexkey operator constant) or
3892 : * (constant operator indexkey).
3893 : */
3894 474 : if (match_index_to_operand(leftop, indexcol, index) &&
3895 450 : !contain_var_clause(rightop) &&
3896 450 : !contain_volatile_functions(rightop))
3897 : {
3898 450 : commuted = false;
3899 : }
3900 24 : else if (match_index_to_operand(rightop, indexcol, index) &&
3901 24 : !contain_var_clause(leftop) &&
3902 24 : !contain_volatile_functions(leftop))
3903 : {
3904 : /* Might match, but we need a commuted operator */
3905 24 : expr_op = get_commutator(expr_op);
3906 24 : if (expr_op == InvalidOid)
3907 0 : return NULL;
3908 24 : commuted = true;
3909 : }
3910 : else
3911 0 : return NULL;
3912 :
3913 : /*
3914 : * Is the (commuted) operator an ordering operator for the opfamily? And
3915 : * if so, does it yield the right sorting semantics?
3916 : */
3917 474 : sortfamily = get_op_opfamily_sortfamily(expr_op, opfamily);
3918 474 : if (sortfamily != pk_opfamily)
3919 0 : return NULL;
3920 :
3921 : /* We have a match. Return clause or a commuted version thereof. */
3922 474 : if (commuted)
3923 : {
3924 24 : OpExpr *newclause = makeNode(OpExpr);
3925 :
3926 : /* flat-copy all the fields of clause */
3927 24 : memcpy(newclause, clause, sizeof(OpExpr));
3928 :
3929 : /* commute it */
3930 24 : newclause->opno = expr_op;
3931 24 : newclause->opfuncid = InvalidOid;
3932 24 : newclause->args = list_make2(rightop, leftop);
3933 :
3934 24 : clause = (Expr *) newclause;
3935 : }
3936 :
3937 474 : return clause;
3938 : }
3939 :
3940 :
3941 : /****************************************************************************
3942 : * ---- ROUTINES TO DO PARTIAL INDEX PREDICATE TESTS ----
3943 : ****************************************************************************/
3944 :
3945 : /*
3946 : * check_index_predicates
3947 : * Set the predicate-derived IndexOptInfo fields for each index
3948 : * of the specified relation.
3949 : *
3950 : * predOK is set true if the index is partial and its predicate is satisfied
3951 : * for this query, ie the query's WHERE clauses imply the predicate.
3952 : *
3953 : * indrestrictinfo is set to the relation's baserestrictinfo list less any
3954 : * conditions that are implied by the index's predicate. (Obviously, for a
3955 : * non-partial index, this is the same as baserestrictinfo.) Such conditions
3956 : * can be dropped from the plan when using the index, in certain cases.
3957 : *
3958 : * At one time it was possible for this to get re-run after adding more
3959 : * restrictions to the rel, thus possibly letting us prove more indexes OK.
3960 : * That doesn't happen any more (at least not in the core code's usage),
3961 : * but this code still supports it in case extensions want to mess with the
3962 : * baserestrictinfo list. We assume that adding more restrictions can't make
3963 : * an index not predOK. We must recompute indrestrictinfo each time, though,
3964 : * to make sure any newly-added restrictions get into it if needed.
3965 : */
3966 : void
3967 403794 : check_index_predicates(PlannerInfo *root, RelOptInfo *rel)
3968 : {
3969 : List *clauselist;
3970 : bool have_partial;
3971 : bool is_target_rel;
3972 : Relids otherrels;
3973 : ListCell *lc;
3974 :
3975 : /* Indexes are available only on base or "other" member relations. */
3976 : Assert(IS_SIMPLE_REL(rel));
3977 :
3978 : /*
3979 : * Initialize the indrestrictinfo lists to be identical to
3980 : * baserestrictinfo, and check whether there are any partial indexes. If
3981 : * not, this is all we need to do.
3982 : */
3983 403794 : have_partial = false;
3984 1116610 : foreach(lc, rel->indexlist)
3985 : {
3986 712816 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
3987 :
3988 712816 : index->indrestrictinfo = rel->baserestrictinfo;
3989 712816 : if (index->indpred)
3990 980 : have_partial = true;
3991 : }
3992 403794 : if (!have_partial)
3993 403138 : return;
3994 :
3995 : /*
3996 : * Construct a list of clauses that we can assume true for the purpose of
3997 : * proving the index(es) usable. Restriction clauses for the rel are
3998 : * always usable, and so are any join clauses that are "movable to" this
3999 : * rel. Also, we can consider any EC-derivable join clauses (which must
4000 : * be "movable to" this rel, by definition).
4001 : */
4002 656 : clauselist = list_copy(rel->baserestrictinfo);
4003 :
4004 : /* Scan the rel's join clauses */
4005 656 : foreach(lc, rel->joininfo)
4006 : {
4007 0 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4008 :
4009 : /* Check if clause can be moved to this rel */
4010 0 : if (!join_clause_is_movable_to(rinfo, rel))
4011 0 : continue;
4012 :
4013 0 : clauselist = lappend(clauselist, rinfo);
4014 : }
4015 :
4016 : /*
4017 : * Add on any equivalence-derivable join clauses. Computing the correct
4018 : * relid sets for generate_join_implied_equalities is slightly tricky
4019 : * because the rel could be a child rel rather than a true baserel, and in
4020 : * that case we must subtract its parents' relid(s) from all_query_rels.
4021 : * Additionally, we mustn't consider clauses that are only computable
4022 : * after outer joins that can null the rel.
4023 : */
4024 656 : if (rel->reloptkind == RELOPT_OTHER_MEMBER_REL)
4025 72 : otherrels = bms_difference(root->all_query_rels,
4026 72 : find_childrel_parents(root, rel));
4027 : else
4028 584 : otherrels = bms_difference(root->all_query_rels, rel->relids);
4029 656 : otherrels = bms_del_members(otherrels, rel->nulling_relids);
4030 :
4031 656 : if (!bms_is_empty(otherrels))
4032 : clauselist =
4033 86 : list_concat(clauselist,
4034 86 : generate_join_implied_equalities(root,
4035 86 : bms_union(rel->relids,
4036 : otherrels),
4037 : otherrels,
4038 : rel,
4039 : NULL));
4040 :
4041 : /*
4042 : * Normally we remove quals that are implied by a partial index's
4043 : * predicate from indrestrictinfo, indicating that they need not be
4044 : * checked explicitly by an indexscan plan using this index. However, if
4045 : * the rel is a target relation of UPDATE/DELETE/MERGE/SELECT FOR UPDATE,
4046 : * we cannot remove such quals from the plan, because they need to be in
4047 : * the plan so that they will be properly rechecked by EvalPlanQual
4048 : * testing. Some day we might want to remove such quals from the main
4049 : * plan anyway and pass them through to EvalPlanQual via a side channel;
4050 : * but for now, we just don't remove implied quals at all for target
4051 : * relations.
4052 : */
4053 1204 : is_target_rel = (bms_is_member(rel->relid, root->all_result_relids) ||
4054 548 : get_plan_rowmark(root->rowMarks, rel->relid) != NULL);
4055 :
4056 : /*
4057 : * Now try to prove each index predicate true, and compute the
4058 : * indrestrictinfo lists for partial indexes. Note that we compute the
4059 : * indrestrictinfo list even for non-predOK indexes; this might seem
4060 : * wasteful, but we may be able to use such indexes in OR clauses, cf
4061 : * generate_bitmap_or_paths().
4062 : */
4063 2022 : foreach(lc, rel->indexlist)
4064 : {
4065 1366 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
4066 : ListCell *lcr;
4067 :
4068 1366 : if (index->indpred == NIL)
4069 386 : continue; /* ignore non-partial indexes here */
4070 :
4071 980 : if (!index->predOK) /* don't repeat work if already proven OK */
4072 980 : index->predOK = predicate_implied_by(index->indpred, clauselist,
4073 : false);
4074 :
4075 : /* If rel is an update target, leave indrestrictinfo as set above */
4076 980 : if (is_target_rel)
4077 168 : continue;
4078 :
4079 : /* Else compute indrestrictinfo as the non-implied quals */
4080 812 : index->indrestrictinfo = NIL;
4081 1914 : foreach(lcr, rel->baserestrictinfo)
4082 : {
4083 1102 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lcr);
4084 :
4085 : /* predicate_implied_by() assumes first arg is immutable */
4086 1102 : if (contain_mutable_functions((Node *) rinfo->clause) ||
4087 1102 : !predicate_implied_by(list_make1(rinfo->clause),
4088 : index->indpred, false))
4089 782 : index->indrestrictinfo = lappend(index->indrestrictinfo, rinfo);
4090 : }
4091 : }
4092 : }
4093 :
4094 : /****************************************************************************
4095 : * ---- ROUTINES TO CHECK EXTERNALLY-VISIBLE CONDITIONS ----
4096 : ****************************************************************************/
4097 :
4098 : /*
4099 : * ec_member_matches_indexcol
4100 : * Test whether an EquivalenceClass member matches an index column.
4101 : *
4102 : * This is a callback for use by generate_implied_equalities_for_column.
4103 : */
4104 : static bool
4105 470876 : ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
4106 : EquivalenceClass *ec, EquivalenceMember *em,
4107 : void *arg)
4108 : {
4109 470876 : IndexOptInfo *index = ((ec_member_matches_arg *) arg)->index;
4110 470876 : int indexcol = ((ec_member_matches_arg *) arg)->indexcol;
4111 : Oid curFamily;
4112 : Oid curCollation;
4113 :
4114 : Assert(indexcol < index->nkeycolumns);
4115 :
4116 470876 : curFamily = index->opfamily[indexcol];
4117 470876 : curCollation = index->indexcollations[indexcol];
4118 :
4119 : /*
4120 : * If it's a btree index, we can reject it if its opfamily isn't
4121 : * compatible with the EC, since no clause generated from the EC could be
4122 : * used with the index. For non-btree indexes, we can't easily tell
4123 : * whether clauses generated from the EC could be used with the index, so
4124 : * don't check the opfamily. This might mean we return "true" for a
4125 : * useless EC, so we have to recheck the results of
4126 : * generate_implied_equalities_for_column; see
4127 : * match_eclass_clauses_to_index.
4128 : */
4129 470876 : if (index->relam == BTREE_AM_OID &&
4130 470834 : !list_member_oid(ec->ec_opfamilies, curFamily))
4131 152568 : return false;
4132 :
4133 : /* We insist on collation match for all index types, though */
4134 318308 : if (!IndexCollMatchesExprColl(curCollation, ec->ec_collation))
4135 18 : return false;
4136 :
4137 318290 : return match_index_to_operand((Node *) em->em_expr, indexcol, index);
4138 : }
4139 :
4140 : /*
4141 : * relation_has_unique_index_for
4142 : * Determine whether the relation provably has at most one row satisfying
4143 : * a set of equality conditions, because the conditions constrain all
4144 : * columns of some unique index.
4145 : *
4146 : * The conditions are provided as a list of RestrictInfo nodes, where the
4147 : * caller has already determined that each condition is a mergejoinable
4148 : * equality with an expression in this relation on one side, and an
4149 : * expression not involving this relation on the other. The transient
4150 : * outer_is_left flag is used to identify which side we should look at:
4151 : * left side if outer_is_left is false, right side if it is true.
4152 : *
4153 : * The caller need only supply equality conditions arising from joins;
4154 : * this routine automatically adds in any usable baserestrictinfo clauses.
4155 : * (Note that the passed-in restrictlist will be destructively modified!)
4156 : *
4157 : * If extra_clauses isn't NULL, return baserestrictinfo clauses which were used
4158 : * to derive uniqueness.
4159 : */
4160 : bool
4161 216628 : relation_has_unique_index_for(PlannerInfo *root, RelOptInfo *rel,
4162 : List *restrictlist, List **extra_clauses)
4163 : {
4164 : ListCell *ic;
4165 :
4166 : /* Short-circuit if no indexes... */
4167 216628 : if (rel->indexlist == NIL)
4168 0 : return false;
4169 :
4170 : /*
4171 : * Examine the rel's restriction clauses for usable var = const clauses
4172 : * that we can add to the restrictlist.
4173 : */
4174 357368 : foreach(ic, rel->baserestrictinfo)
4175 : {
4176 140740 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(ic);
4177 :
4178 : /*
4179 : * Note: can_join won't be set for a restriction clause, but
4180 : * mergeopfamilies will be if it has a mergejoinable operator and
4181 : * doesn't contain volatile functions.
4182 : */
4183 140740 : if (restrictinfo->mergeopfamilies == NIL)
4184 56722 : continue; /* not mergejoinable */
4185 :
4186 : /*
4187 : * The clause certainly doesn't refer to anything but the given rel.
4188 : * If either side is pseudoconstant then we can use it.
4189 : */
4190 84018 : if (bms_is_empty(restrictinfo->left_relids))
4191 : {
4192 : /* righthand side is inner */
4193 60 : restrictinfo->outer_is_left = true;
4194 : }
4195 83958 : else if (bms_is_empty(restrictinfo->right_relids))
4196 : {
4197 : /* lefthand side is inner */
4198 83832 : restrictinfo->outer_is_left = false;
4199 : }
4200 : else
4201 126 : continue;
4202 :
4203 : /* OK, add to list */
4204 83892 : restrictlist = lappend(restrictlist, restrictinfo);
4205 : }
4206 :
4207 : /* Short-circuit the easy case */
4208 216628 : if (restrictlist == NIL)
4209 1114 : return false;
4210 :
4211 : /* Examine each index of the relation ... */
4212 548742 : foreach(ic, rel->indexlist)
4213 : {
4214 456342 : IndexOptInfo *ind = (IndexOptInfo *) lfirst(ic);
4215 : int c;
4216 456342 : List *exprs = NIL;
4217 :
4218 : /*
4219 : * If the index is not unique, or not immediately enforced, or if it's
4220 : * a partial index, it's useless here. We're unable to make use of
4221 : * predOK partial unique indexes due to the fact that
4222 : * check_index_predicates() also makes use of join predicates to
4223 : * determine if the partial index is usable. Here we need proofs that
4224 : * hold true before any joins are evaluated.
4225 : */
4226 456342 : if (!ind->unique || !ind->immediate || ind->indpred != NIL)
4227 124638 : continue;
4228 :
4229 : /*
4230 : * Try to find each index column in the list of conditions. This is
4231 : * O(N^2) or worse, but we expect all the lists to be short.
4232 : */
4233 555652 : for (c = 0; c < ind->nkeycolumns; c++)
4234 : {
4235 : ListCell *lc;
4236 :
4237 843478 : foreach(lc, restrictlist)
4238 : {
4239 634888 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4240 : Node *rexpr;
4241 :
4242 : /*
4243 : * The condition's equality operator must be a member of the
4244 : * index opfamily, else it is not asserting the right kind of
4245 : * equality behavior for this index. We check this first
4246 : * since it's probably cheaper than match_index_to_operand().
4247 : */
4248 634888 : if (!list_member_oid(rinfo->mergeopfamilies, ind->opfamily[c]))
4249 193990 : continue;
4250 :
4251 : /*
4252 : * XXX at some point we may need to check collations here too.
4253 : * For the moment we assume all collations reduce to the same
4254 : * notion of equality.
4255 : */
4256 :
4257 : /* OK, see if the condition operand matches the index key */
4258 440898 : if (rinfo->outer_is_left)
4259 177382 : rexpr = get_rightop(rinfo->clause);
4260 : else
4261 263516 : rexpr = get_leftop(rinfo->clause);
4262 :
4263 440898 : if (match_index_to_operand(rexpr, c, ind))
4264 : {
4265 223948 : if (bms_membership(rinfo->clause_relids) == BMS_SINGLETON)
4266 : {
4267 : MemoryContext oldMemCtx =
4268 52416 : MemoryContextSwitchTo(root->planner_cxt);
4269 :
4270 : /*
4271 : * Add filter clause into a list allowing caller to
4272 : * know if uniqueness have made not only by join
4273 : * clauses.
4274 : */
4275 : Assert(bms_is_empty(rinfo->left_relids) ||
4276 : bms_is_empty(rinfo->right_relids));
4277 52416 : if (extra_clauses)
4278 144 : exprs = lappend(exprs, rinfo);
4279 52416 : MemoryContextSwitchTo(oldMemCtx);
4280 : }
4281 :
4282 223948 : break; /* found a match; column is unique */
4283 : }
4284 : }
4285 :
4286 432538 : if (lc == NULL)
4287 208590 : break; /* no match; this index doesn't help us */
4288 : }
4289 :
4290 : /* Matched all key columns of this index? */
4291 331704 : if (c == ind->nkeycolumns)
4292 : {
4293 123114 : if (extra_clauses)
4294 654 : *extra_clauses = exprs;
4295 123114 : return true;
4296 : }
4297 : }
4298 :
4299 92400 : return false;
4300 : }
4301 :
4302 : /*
4303 : * indexcol_is_bool_constant_for_query
4304 : *
4305 : * If an index column is constrained to have a constant value by the query's
4306 : * WHERE conditions, then it's irrelevant for sort-order considerations.
4307 : * Usually that means we have a restriction clause WHERE indexcol = constant,
4308 : * which gets turned into an EquivalenceClass containing a constant, which
4309 : * is recognized as redundant by build_index_pathkeys(). But if the index
4310 : * column is a boolean variable (or expression), then we are not going to
4311 : * see WHERE indexcol = constant, because expression preprocessing will have
4312 : * simplified that to "WHERE indexcol" or "WHERE NOT indexcol". So we are not
4313 : * going to have a matching EquivalenceClass (unless the query also contains
4314 : * "ORDER BY indexcol"). To allow such cases to work the same as they would
4315 : * for non-boolean values, this function is provided to detect whether the
4316 : * specified index column matches a boolean restriction clause.
4317 : */
4318 : bool
4319 654416 : indexcol_is_bool_constant_for_query(PlannerInfo *root,
4320 : IndexOptInfo *index,
4321 : int indexcol)
4322 : {
4323 : ListCell *lc;
4324 :
4325 : /* If the index isn't boolean, we can't possibly get a match */
4326 654416 : if (!IsBooleanOpfamily(index->opfamily[indexcol]))
4327 651412 : return false;
4328 :
4329 : /* Check each restriction clause for the index's rel */
4330 3040 : foreach(lc, index->rel->baserestrictinfo)
4331 : {
4332 1276 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4333 :
4334 : /*
4335 : * As in match_clause_to_indexcol, never match pseudoconstants to
4336 : * indexes. (It might be semantically okay to do so here, but the
4337 : * odds of getting a match are negligible, so don't waste the cycles.)
4338 : */
4339 1276 : if (rinfo->pseudoconstant)
4340 0 : continue;
4341 :
4342 : /* See if we can match the clause's expression to the index column */
4343 1276 : if (match_boolean_index_clause(root, rinfo, indexcol, index))
4344 1240 : return true;
4345 : }
4346 :
4347 1764 : return false;
4348 : }
4349 :
4350 :
4351 : /****************************************************************************
4352 : * ---- ROUTINES TO CHECK OPERANDS ----
4353 : ****************************************************************************/
4354 :
4355 : /*
4356 : * match_index_to_operand()
4357 : * Generalized test for a match between an index's key
4358 : * and the operand on one side of a restriction or join clause.
4359 : *
4360 : * operand: the nodetree to be compared to the index
4361 : * indexcol: the column number of the index (counting from 0)
4362 : * index: the index of interest
4363 : *
4364 : * Note that we aren't interested in collations here; the caller must check
4365 : * for a collation match, if it's dealing with an operator where that matters.
4366 : *
4367 : * This is exported for use in selfuncs.c.
4368 : */
4369 : bool
4370 3725962 : match_index_to_operand(Node *operand,
4371 : int indexcol,
4372 : IndexOptInfo *index)
4373 : {
4374 : int indkey;
4375 :
4376 : /*
4377 : * Ignore any RelabelType node above the operand. This is needed to be
4378 : * able to apply indexscanning in binary-compatible-operator cases. Note:
4379 : * we can assume there is at most one RelabelType node;
4380 : * eval_const_expressions() will have simplified if more than one.
4381 : */
4382 3725962 : if (operand && IsA(operand, RelabelType))
4383 23396 : operand = (Node *) ((RelabelType *) operand)->arg;
4384 :
4385 3725962 : indkey = index->indexkeys[indexcol];
4386 3725962 : if (indkey != 0)
4387 : {
4388 : /*
4389 : * Simple index column; operand must be a matching Var.
4390 : */
4391 3719928 : if (operand && IsA(operand, Var) &&
4392 2768668 : index->rel->relid == ((Var *) operand)->varno &&
4393 2573102 : indkey == ((Var *) operand)->varattno &&
4394 907008 : ((Var *) operand)->varnullingrels == NULL)
4395 905390 : return true;
4396 : }
4397 : else
4398 : {
4399 : /*
4400 : * Index expression; find the correct expression. (This search could
4401 : * be avoided, at the cost of complicating all the callers of this
4402 : * routine; doesn't seem worth it.)
4403 : */
4404 : ListCell *indexpr_item;
4405 : int i;
4406 : Node *indexkey;
4407 :
4408 6034 : indexpr_item = list_head(index->indexprs);
4409 6034 : for (i = 0; i < indexcol; i++)
4410 : {
4411 0 : if (index->indexkeys[i] == 0)
4412 : {
4413 0 : if (indexpr_item == NULL)
4414 0 : elog(ERROR, "wrong number of index expressions");
4415 0 : indexpr_item = lnext(index->indexprs, indexpr_item);
4416 : }
4417 : }
4418 6034 : if (indexpr_item == NULL)
4419 0 : elog(ERROR, "wrong number of index expressions");
4420 6034 : indexkey = (Node *) lfirst(indexpr_item);
4421 :
4422 : /*
4423 : * Does it match the operand? Again, strip any relabeling.
4424 : */
4425 6034 : if (indexkey && IsA(indexkey, RelabelType))
4426 10 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
4427 :
4428 6034 : if (equal(indexkey, operand))
4429 2164 : return true;
4430 : }
4431 :
4432 2818408 : return false;
4433 : }
4434 :
4435 : /*
4436 : * is_pseudo_constant_for_index()
4437 : * Test whether the given expression can be used as an indexscan
4438 : * comparison value.
4439 : *
4440 : * An indexscan comparison value must not contain any volatile functions,
4441 : * and it can't contain any Vars of the index's own table. Vars of
4442 : * other tables are okay, though; in that case we'd be producing an
4443 : * indexqual usable in a parameterized indexscan. This is, therefore,
4444 : * a weaker condition than is_pseudo_constant_clause().
4445 : *
4446 : * This function is exported for use by planner support functions,
4447 : * which will have available the IndexOptInfo, but not any RestrictInfo
4448 : * infrastructure. It is making the same test made by functions above
4449 : * such as match_opclause_to_indexcol(), but those rely where possible
4450 : * on RestrictInfo information about variable membership.
4451 : *
4452 : * expr: the nodetree to be checked
4453 : * index: the index of interest
4454 : */
4455 : bool
4456 0 : is_pseudo_constant_for_index(PlannerInfo *root, Node *expr, IndexOptInfo *index)
4457 : {
4458 : /* pull_varnos is cheaper than volatility check, so do that first */
4459 0 : if (bms_is_member(index->rel->relid, pull_varnos(root, expr)))
4460 0 : return false; /* no good, contains Var of table */
4461 0 : if (contain_volatile_functions(expr))
4462 0 : return false; /* no good, volatile comparison value */
4463 0 : return true;
4464 : }
|
