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