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