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) 1996-2018, 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_collation.h"
24 #include "catalog/pg_operator.h"
25 #include "catalog/pg_opfamily.h"
26 #include "catalog/pg_type.h"
27 #include "nodes/makefuncs.h"
28 #include "optimizer/clauses.h"
29 #include "optimizer/cost.h"
30 #include "optimizer/pathnode.h"
31 #include "optimizer/paths.h"
32 #include "optimizer/predtest.h"
33 #include "optimizer/prep.h"
34 #include "optimizer/restrictinfo.h"
35 #include "optimizer/var.h"
36 #include "utils/builtins.h"
37 #include "utils/bytea.h"
38 #include "utils/lsyscache.h"
39 #include "utils/pg_locale.h"
40 #include "utils/selfuncs.h"
41
42
43 /* XXX see PartCollMatchesExprColl */
44 #define IndexCollMatchesExprColl(idxcollation, exprcollation) \
45 ((idxcollation) == InvalidOid || (idxcollation) == (exprcollation))
46
47 /* Whether we are looking for plain indexscan, bitmap scan, or either */
48 typedef enum
49 {
50 ST_INDEXSCAN, /* must support amgettuple */
51 ST_BITMAPSCAN, /* must support amgetbitmap */
52 ST_ANYSCAN /* either is okay */
53 } ScanTypeControl;
54
55 /* Data structure for collecting qual clauses that match an index */
56 typedef struct
57 {
58 bool nonempty; /* True if lists are not all empty */
59 /* Lists of RestrictInfos, one per index column */
60 List *indexclauses[INDEX_MAX_KEYS];
61 } IndexClauseSet;
62
63 /* Per-path data used within choose_bitmap_and() */
64 typedef struct
65 {
66 Path *path; /* IndexPath, BitmapAndPath, or BitmapOrPath */
67 List *quals; /* the WHERE clauses it uses */
68 List *preds; /* predicates of its partial index(es) */
69 Bitmapset *clauseids; /* quals+preds represented as a bitmapset */
70 bool unclassifiable; /* has too many quals+preds to process? */
71 } PathClauseUsage;
72
73 /* Callback argument for ec_member_matches_indexcol */
74 typedef struct
75 {
76 IndexOptInfo *index; /* index we're considering */
77 int indexcol; /* index column we want to match to */
78 } ec_member_matches_arg;
79
80
81 static void consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
82 IndexOptInfo *index,
83 IndexClauseSet *rclauseset,
84 IndexClauseSet *jclauseset,
85 IndexClauseSet *eclauseset,
86 List **bitindexpaths);
87 static void consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
88 IndexOptInfo *index,
89 IndexClauseSet *rclauseset,
90 IndexClauseSet *jclauseset,
91 IndexClauseSet *eclauseset,
92 List **bitindexpaths,
93 List *indexjoinclauses,
94 int considered_clauses,
95 List **considered_relids);
96 static void get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
97 IndexOptInfo *index,
98 IndexClauseSet *rclauseset,
99 IndexClauseSet *jclauseset,
100 IndexClauseSet *eclauseset,
101 List **bitindexpaths,
102 Relids relids,
103 List **considered_relids);
104 static bool eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
105 List *indexjoinclauses);
106 static bool bms_equal_any(Relids relids, List *relids_list);
107 static void get_index_paths(PlannerInfo *root, RelOptInfo *rel,
108 IndexOptInfo *index, IndexClauseSet *clauses,
109 List **bitindexpaths);
110 static List *build_index_paths(PlannerInfo *root, RelOptInfo *rel,
111 IndexOptInfo *index, IndexClauseSet *clauses,
112 bool useful_predicate,
113 ScanTypeControl scantype,
114 bool *skip_nonnative_saop,
115 bool *skip_lower_saop);
116 static List *build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
117 List *clauses, List *other_clauses);
118 static List *generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
119 List *clauses, List *other_clauses);
120 static Path *choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel,
121 List *paths);
122 static int path_usage_comparator(const void *a, const void *b);
123 static Cost bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel,
124 Path *ipath);
125 static Cost bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel,
126 List *paths);
127 static PathClauseUsage *classify_index_clause_usage(Path *path,
128 List **clauselist);
129 static void find_indexpath_quals(Path *bitmapqual, List **quals, List **preds);
130 static int find_list_position(Node *node, List **nodelist);
131 static bool check_index_only(RelOptInfo *rel, IndexOptInfo *index);
132 static double get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids);
133 static double adjust_rowcount_for_semijoins(PlannerInfo *root,
134 Index cur_relid,
135 Index outer_relid,
136 double rowcount);
137 static double approximate_joinrel_size(PlannerInfo *root, Relids relids);
138 static void match_restriction_clauses_to_index(RelOptInfo *rel,
139 IndexOptInfo *index,
140 IndexClauseSet *clauseset);
141 static void match_join_clauses_to_index(PlannerInfo *root,
142 RelOptInfo *rel, IndexOptInfo *index,
143 IndexClauseSet *clauseset,
144 List **joinorclauses);
145 static void match_eclass_clauses_to_index(PlannerInfo *root,
146 IndexOptInfo *index,
147 IndexClauseSet *clauseset);
148 static void match_clauses_to_index(IndexOptInfo *index,
149 List *clauses,
150 IndexClauseSet *clauseset);
151 static void match_clause_to_index(IndexOptInfo *index,
152 RestrictInfo *rinfo,
153 IndexClauseSet *clauseset);
154 static bool match_clause_to_indexcol(IndexOptInfo *index,
155 int indexcol,
156 RestrictInfo *rinfo);
157 static bool is_indexable_operator(Oid expr_op, Oid opfamily,
158 bool indexkey_on_left);
159 static bool match_rowcompare_to_indexcol(IndexOptInfo *index,
160 int indexcol,
161 Oid opfamily,
162 Oid idxcollation,
163 RowCompareExpr *clause);
164 static void match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
165 List **orderby_clauses_p,
166 List **clause_columns_p);
167 static Expr *match_clause_to_ordering_op(IndexOptInfo *index,
168 int indexcol, Expr *clause, Oid pk_opfamily);
169 static bool ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
170 EquivalenceClass *ec, EquivalenceMember *em,
171 void *arg);
172 static bool match_boolean_index_clause(Node *clause, int indexcol,
173 IndexOptInfo *index);
174 static bool match_special_index_operator(Expr *clause,
175 Oid opfamily, Oid idxcollation,
176 bool indexkey_on_left);
177 static Expr *expand_boolean_index_clause(Node *clause, int indexcol,
178 IndexOptInfo *index);
179 static List *expand_indexqual_opclause(RestrictInfo *rinfo,
180 Oid opfamily, Oid idxcollation);
181 static RestrictInfo *expand_indexqual_rowcompare(RestrictInfo *rinfo,
182 IndexOptInfo *index,
183 int indexcol);
184 static List *prefix_quals(Node *leftop, Oid opfamily, Oid collation,
185 Const *prefix, Pattern_Prefix_Status pstatus);
186 static List *network_prefix_quals(Node *leftop, Oid expr_op, Oid opfamily,
187 Datum rightop);
188 static Datum string_to_datum(const char *str, Oid datatype);
189 static Const *string_to_const(const char *str, Oid datatype);
190
191
192 /*
193 * create_index_paths()
194 * Generate all interesting index paths for the given relation.
195 * Candidate paths are added to the rel's pathlist (using add_path).
196 *
197 * To be considered for an index scan, an index must match one or more
198 * restriction clauses or join clauses from the query's qual condition,
199 * or match the query's ORDER BY condition, or have a predicate that
200 * matches the query's qual condition.
201 *
202 * There are two basic kinds of index scans. A "plain" index scan uses
203 * only restriction clauses (possibly none at all) in its indexqual,
204 * so it can be applied in any context. A "parameterized" index scan uses
205 * join clauses (plus restriction clauses, if available) in its indexqual.
206 * When joining such a scan to one of the relations supplying the other
207 * variables used in its indexqual, the parameterized scan must appear as
208 * the inner relation of a nestloop join; it can't be used on the outer side,
209 * nor in a merge or hash join. In that context, values for the other rels'
210 * attributes are available and fixed during any one scan of the indexpath.
211 *
212 * An IndexPath is generated and submitted to add_path() for each plain or
213 * parameterized index scan this routine deems potentially interesting for
214 * the current query.
215 *
216 * 'rel' is the relation for which we want to generate index paths
217 *
218 * Note: check_index_predicates() must have been run previously for this rel.
219 *
220 * Note: in cases involving LATERAL references in the relation's tlist, it's
221 * possible that rel->lateral_relids is nonempty. Currently, we include
222 * lateral_relids into the parameterization reported for each path, but don't
223 * take it into account otherwise. The fact that any such rels *must* be
224 * available as parameter sources perhaps should influence our choices of
225 * index quals ... but for now, it doesn't seem worth troubling over.
226 * In particular, comments below about "unparameterized" paths should be read
227 * as meaning "unparameterized so far as the indexquals are concerned".
228 */
229 void
create_index_paths(PlannerInfo * root,RelOptInfo * rel)230 create_index_paths(PlannerInfo *root, RelOptInfo *rel)
231 {
232 List *indexpaths;
233 List *bitindexpaths;
234 List *bitjoinpaths;
235 List *joinorclauses;
236 IndexClauseSet rclauseset;
237 IndexClauseSet jclauseset;
238 IndexClauseSet eclauseset;
239 ListCell *lc;
240
241 /* Skip the whole mess if no indexes */
242 if (rel->indexlist == NIL)
243 return;
244
245 /* Bitmap paths are collected and then dealt with at the end */
246 bitindexpaths = bitjoinpaths = joinorclauses = NIL;
247
248 /* Examine each index in turn */
249 foreach(lc, rel->indexlist)
250 {
251 IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
252
253 /* Protect limited-size array in IndexClauseSets */
254 Assert(index->nkeycolumns <= INDEX_MAX_KEYS);
255
256 /*
257 * Ignore partial indexes that do not match the query.
258 * (generate_bitmap_or_paths() might be able to do something with
259 * them, but that's of no concern here.)
260 */
261 if (index->indpred != NIL && !index->predOK)
262 continue;
263
264 /*
265 * Identify the restriction clauses that can match the index.
266 */
267 MemSet(&rclauseset, 0, sizeof(rclauseset));
268 match_restriction_clauses_to_index(rel, index, &rclauseset);
269
270 /*
271 * Build index paths from the restriction clauses. These will be
272 * non-parameterized paths. Plain paths go directly to add_path(),
273 * bitmap paths are added to bitindexpaths to be handled below.
274 */
275 get_index_paths(root, rel, index, &rclauseset,
276 &bitindexpaths);
277
278 /*
279 * Identify the join clauses that can match the index. For the moment
280 * we keep them separate from the restriction clauses. Note that this
281 * step finds only "loose" join clauses that have not been merged into
282 * EquivalenceClasses. Also, collect join OR clauses for later.
283 */
284 MemSet(&jclauseset, 0, sizeof(jclauseset));
285 match_join_clauses_to_index(root, rel, index,
286 &jclauseset, &joinorclauses);
287
288 /*
289 * Look for EquivalenceClasses that can generate joinclauses matching
290 * the index.
291 */
292 MemSet(&eclauseset, 0, sizeof(eclauseset));
293 match_eclass_clauses_to_index(root, index,
294 &eclauseset);
295
296 /*
297 * If we found any plain or eclass join clauses, build parameterized
298 * index paths using them.
299 */
300 if (jclauseset.nonempty || eclauseset.nonempty)
301 consider_index_join_clauses(root, rel, index,
302 &rclauseset,
303 &jclauseset,
304 &eclauseset,
305 &bitjoinpaths);
306 }
307
308 /*
309 * Generate BitmapOrPaths for any suitable OR-clauses present in the
310 * restriction list. Add these to bitindexpaths.
311 */
312 indexpaths = generate_bitmap_or_paths(root, rel,
313 rel->baserestrictinfo, NIL);
314 bitindexpaths = list_concat(bitindexpaths, indexpaths);
315
316 /*
317 * Likewise, generate BitmapOrPaths for any suitable OR-clauses present in
318 * the joinclause list. Add these to bitjoinpaths.
319 */
320 indexpaths = generate_bitmap_or_paths(root, rel,
321 joinorclauses, rel->baserestrictinfo);
322 bitjoinpaths = list_concat(bitjoinpaths, indexpaths);
323
324 /*
325 * If we found anything usable, generate a BitmapHeapPath for the most
326 * promising combination of restriction bitmap index paths. Note there
327 * will be only one such path no matter how many indexes exist. This
328 * should be sufficient since there's basically only one figure of merit
329 * (total cost) for such a path.
330 */
331 if (bitindexpaths != NIL)
332 {
333 Path *bitmapqual;
334 BitmapHeapPath *bpath;
335
336 bitmapqual = choose_bitmap_and(root, rel, bitindexpaths);
337 bpath = create_bitmap_heap_path(root, rel, bitmapqual,
338 rel->lateral_relids, 1.0, 0);
339 add_path(rel, (Path *) bpath);
340
341 /* create a partial bitmap heap path */
342 if (rel->consider_parallel && rel->lateral_relids == NULL)
343 create_partial_bitmap_paths(root, rel, bitmapqual);
344 }
345
346 /*
347 * Likewise, if we found anything usable, generate BitmapHeapPaths for the
348 * most promising combinations of join bitmap index paths. Our strategy
349 * is to generate one such path for each distinct parameterization seen
350 * among the available bitmap index paths. This may look pretty
351 * expensive, but usually there won't be very many distinct
352 * parameterizations. (This logic is quite similar to that in
353 * consider_index_join_clauses, but we're working with whole paths not
354 * individual clauses.)
355 */
356 if (bitjoinpaths != NIL)
357 {
358 List *all_path_outers;
359 ListCell *lc;
360
361 /* Identify each distinct parameterization seen in bitjoinpaths */
362 all_path_outers = NIL;
363 foreach(lc, bitjoinpaths)
364 {
365 Path *path = (Path *) lfirst(lc);
366 Relids required_outer = PATH_REQ_OUTER(path);
367
368 if (!bms_equal_any(required_outer, all_path_outers))
369 all_path_outers = lappend(all_path_outers, required_outer);
370 }
371
372 /* Now, for each distinct parameterization set ... */
373 foreach(lc, all_path_outers)
374 {
375 Relids max_outers = (Relids) lfirst(lc);
376 List *this_path_set;
377 Path *bitmapqual;
378 Relids required_outer;
379 double loop_count;
380 BitmapHeapPath *bpath;
381 ListCell *lcp;
382
383 /* Identify all the bitmap join paths needing no more than that */
384 this_path_set = NIL;
385 foreach(lcp, bitjoinpaths)
386 {
387 Path *path = (Path *) lfirst(lcp);
388
389 if (bms_is_subset(PATH_REQ_OUTER(path), max_outers))
390 this_path_set = lappend(this_path_set, path);
391 }
392
393 /*
394 * Add in restriction bitmap paths, since they can be used
395 * together with any join paths.
396 */
397 this_path_set = list_concat(this_path_set, bitindexpaths);
398
399 /* Select best AND combination for this parameterization */
400 bitmapqual = choose_bitmap_and(root, rel, this_path_set);
401
402 /* And push that path into the mix */
403 required_outer = PATH_REQ_OUTER(bitmapqual);
404 loop_count = get_loop_count(root, rel->relid, required_outer);
405 bpath = create_bitmap_heap_path(root, rel, bitmapqual,
406 required_outer, loop_count, 0);
407 add_path(rel, (Path *) bpath);
408 }
409 }
410 }
411
412 /*
413 * consider_index_join_clauses
414 * Given sets of join clauses for an index, decide which parameterized
415 * index paths to build.
416 *
417 * Plain indexpaths are sent directly to add_path, while potential
418 * bitmap indexpaths are added to *bitindexpaths for later processing.
419 *
420 * 'rel' is the index's heap relation
421 * 'index' is the index for which we want to generate paths
422 * 'rclauseset' is the collection of indexable restriction clauses
423 * 'jclauseset' is the collection of indexable simple join clauses
424 * 'eclauseset' is the collection of indexable clauses from EquivalenceClasses
425 * '*bitindexpaths' is the list to add bitmap paths to
426 */
427 static void
consider_index_join_clauses(PlannerInfo * root,RelOptInfo * rel,IndexOptInfo * index,IndexClauseSet * rclauseset,IndexClauseSet * jclauseset,IndexClauseSet * eclauseset,List ** bitindexpaths)428 consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
429 IndexOptInfo *index,
430 IndexClauseSet *rclauseset,
431 IndexClauseSet *jclauseset,
432 IndexClauseSet *eclauseset,
433 List **bitindexpaths)
434 {
435 int considered_clauses = 0;
436 List *considered_relids = NIL;
437 int indexcol;
438
439 /*
440 * The strategy here is to identify every potentially useful set of outer
441 * rels that can provide indexable join clauses. For each such set,
442 * select all the join clauses available from those outer rels, add on all
443 * the indexable restriction clauses, and generate plain and/or bitmap
444 * index paths for that set of clauses. This is based on the assumption
445 * that it's always better to apply a clause as an indexqual than as a
446 * filter (qpqual); which is where an available clause would end up being
447 * applied if we omit it from the indexquals.
448 *
449 * This looks expensive, but in most practical cases there won't be very
450 * many distinct sets of outer rels to consider. As a safety valve when
451 * that's not true, we use a heuristic: limit the number of outer rel sets
452 * considered to a multiple of the number of clauses considered. (We'll
453 * always consider using each individual join clause, though.)
454 *
455 * For simplicity in selecting relevant clauses, we represent each set of
456 * outer rels as a maximum set of clause_relids --- that is, the indexed
457 * relation itself is also included in the relids set. considered_relids
458 * lists all relids sets we've already tried.
459 */
460 for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
461 {
462 /* Consider each applicable simple join clause */
463 considered_clauses += list_length(jclauseset->indexclauses[indexcol]);
464 consider_index_join_outer_rels(root, rel, index,
465 rclauseset, jclauseset, eclauseset,
466 bitindexpaths,
467 jclauseset->indexclauses[indexcol],
468 considered_clauses,
469 &considered_relids);
470 /* Consider each applicable eclass join clause */
471 considered_clauses += list_length(eclauseset->indexclauses[indexcol]);
472 consider_index_join_outer_rels(root, rel, index,
473 rclauseset, jclauseset, eclauseset,
474 bitindexpaths,
475 eclauseset->indexclauses[indexcol],
476 considered_clauses,
477 &considered_relids);
478 }
479 }
480
481 /*
482 * consider_index_join_outer_rels
483 * Generate parameterized paths based on clause relids in the clause list.
484 *
485 * Workhorse for consider_index_join_clauses; see notes therein for rationale.
486 *
487 * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset', and
488 * 'bitindexpaths' as above
489 * 'indexjoinclauses' is a list of RestrictInfos for join clauses
490 * 'considered_clauses' is the total number of clauses considered (so far)
491 * '*considered_relids' is a list of all relids sets already considered
492 */
493 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)494 consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
495 IndexOptInfo *index,
496 IndexClauseSet *rclauseset,
497 IndexClauseSet *jclauseset,
498 IndexClauseSet *eclauseset,
499 List **bitindexpaths,
500 List *indexjoinclauses,
501 int considered_clauses,
502 List **considered_relids)
503 {
504 ListCell *lc;
505
506 /* Examine relids of each joinclause in the given list */
507 foreach(lc, indexjoinclauses)
508 {
509 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
510 Relids clause_relids = rinfo->clause_relids;
511 ListCell *lc2;
512
513 /* If we already tried its relids set, no need to do so again */
514 if (bms_equal_any(clause_relids, *considered_relids))
515 continue;
516
517 /*
518 * Generate the union of this clause's relids set with each
519 * previously-tried set. This ensures we try this clause along with
520 * every interesting subset of previous clauses. However, to avoid
521 * exponential growth of planning time when there are many clauses,
522 * limit the number of relid sets accepted to 10 * considered_clauses.
523 *
524 * Note: get_join_index_paths adds entries to *considered_relids, but
525 * it prepends them to the list, so that we won't visit new entries
526 * during the inner foreach loop. No real harm would be done if we
527 * did, since the subset check would reject them; but it would waste
528 * some cycles.
529 */
530 foreach(lc2, *considered_relids)
531 {
532 Relids oldrelids = (Relids) lfirst(lc2);
533
534 /*
535 * If either is a subset of the other, no new set is possible.
536 * This isn't a complete test for redundancy, but it's easy and
537 * cheap. get_join_index_paths will check more carefully if we
538 * already generated the same relids set.
539 */
540 if (bms_subset_compare(clause_relids, oldrelids) != BMS_DIFFERENT)
541 continue;
542
543 /*
544 * If this clause was derived from an equivalence class, the
545 * clause list may contain other clauses derived from the same
546 * eclass. We should not consider that combining this clause with
547 * one of those clauses generates a usefully different
548 * parameterization; so skip if any clause derived from the same
549 * eclass would already have been included when using oldrelids.
550 */
551 if (rinfo->parent_ec &&
552 eclass_already_used(rinfo->parent_ec, oldrelids,
553 indexjoinclauses))
554 continue;
555
556 /*
557 * If the number of relid sets considered exceeds our heuristic
558 * limit, stop considering combinations of clauses. We'll still
559 * consider the current clause alone, though (below this loop).
560 */
561 if (list_length(*considered_relids) >= 10 * considered_clauses)
562 break;
563
564 /* OK, try the union set */
565 get_join_index_paths(root, rel, index,
566 rclauseset, jclauseset, eclauseset,
567 bitindexpaths,
568 bms_union(clause_relids, oldrelids),
569 considered_relids);
570 }
571
572 /* Also try this set of relids by itself */
573 get_join_index_paths(root, rel, index,
574 rclauseset, jclauseset, eclauseset,
575 bitindexpaths,
576 clause_relids,
577 considered_relids);
578 }
579 }
580
581 /*
582 * get_join_index_paths
583 * Generate index paths using clauses from the specified outer relations.
584 * In addition to generating paths, relids is added to *considered_relids
585 * if not already present.
586 *
587 * Workhorse for consider_index_join_clauses; see notes therein for rationale.
588 *
589 * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset',
590 * 'bitindexpaths', 'considered_relids' as above
591 * 'relids' is the current set of relids to consider (the target rel plus
592 * one or more outer rels)
593 */
594 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)595 get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
596 IndexOptInfo *index,
597 IndexClauseSet *rclauseset,
598 IndexClauseSet *jclauseset,
599 IndexClauseSet *eclauseset,
600 List **bitindexpaths,
601 Relids relids,
602 List **considered_relids)
603 {
604 IndexClauseSet clauseset;
605 int indexcol;
606
607 /* If we already considered this relids set, don't repeat the work */
608 if (bms_equal_any(relids, *considered_relids))
609 return;
610
611 /* Identify indexclauses usable with this relids set */
612 MemSet(&clauseset, 0, sizeof(clauseset));
613
614 for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
615 {
616 ListCell *lc;
617
618 /* First find applicable simple join clauses */
619 foreach(lc, jclauseset->indexclauses[indexcol])
620 {
621 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
622
623 if (bms_is_subset(rinfo->clause_relids, relids))
624 clauseset.indexclauses[indexcol] =
625 lappend(clauseset.indexclauses[indexcol], rinfo);
626 }
627
628 /*
629 * Add applicable eclass join clauses. The clauses generated for each
630 * column are redundant (cf generate_implied_equalities_for_column),
631 * so we need at most one. This is the only exception to the general
632 * rule of using all available index clauses.
633 */
634 foreach(lc, eclauseset->indexclauses[indexcol])
635 {
636 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
637
638 if (bms_is_subset(rinfo->clause_relids, relids))
639 {
640 clauseset.indexclauses[indexcol] =
641 lappend(clauseset.indexclauses[indexcol], rinfo);
642 break;
643 }
644 }
645
646 /* Add restriction clauses (this is nondestructive to rclauseset) */
647 clauseset.indexclauses[indexcol] =
648 list_concat(clauseset.indexclauses[indexcol],
649 rclauseset->indexclauses[indexcol]);
650
651 if (clauseset.indexclauses[indexcol] != NIL)
652 clauseset.nonempty = true;
653 }
654
655 /* We should have found something, else caller passed silly relids */
656 Assert(clauseset.nonempty);
657
658 /* Build index path(s) using the collected set of clauses */
659 get_index_paths(root, rel, index, &clauseset, bitindexpaths);
660
661 /*
662 * Remember we considered paths for this set of relids. We use lcons not
663 * lappend to avoid confusing the loop in consider_index_join_outer_rels.
664 */
665 *considered_relids = lcons(relids, *considered_relids);
666 }
667
668 /*
669 * eclass_already_used
670 * True if any join clause usable with oldrelids was generated from
671 * the specified equivalence class.
672 */
673 static bool
eclass_already_used(EquivalenceClass * parent_ec,Relids oldrelids,List * indexjoinclauses)674 eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
675 List *indexjoinclauses)
676 {
677 ListCell *lc;
678
679 foreach(lc, indexjoinclauses)
680 {
681 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
682
683 if (rinfo->parent_ec == parent_ec &&
684 bms_is_subset(rinfo->clause_relids, oldrelids))
685 return true;
686 }
687 return false;
688 }
689
690 /*
691 * bms_equal_any
692 * True if relids is bms_equal to any member of relids_list
693 *
694 * Perhaps this should be in bitmapset.c someday.
695 */
696 static bool
bms_equal_any(Relids relids,List * relids_list)697 bms_equal_any(Relids relids, List *relids_list)
698 {
699 ListCell *lc;
700
701 foreach(lc, relids_list)
702 {
703 if (bms_equal(relids, (Relids) lfirst(lc)))
704 return true;
705 }
706 return false;
707 }
708
709
710 /*
711 * get_index_paths
712 * Given an index and a set of index clauses for it, construct IndexPaths.
713 *
714 * Plain indexpaths are sent directly to add_path, while potential
715 * bitmap indexpaths are added to *bitindexpaths for later processing.
716 *
717 * This is a fairly simple frontend to build_index_paths(). Its reason for
718 * existence is mainly to handle ScalarArrayOpExpr quals properly. If the
719 * index AM supports them natively, we should just include them in simple
720 * index paths. If not, we should exclude them while building simple index
721 * paths, and then make a separate attempt to include them in bitmap paths.
722 * Furthermore, we should consider excluding lower-order ScalarArrayOpExpr
723 * quals so as to create ordered paths.
724 */
725 static void
get_index_paths(PlannerInfo * root,RelOptInfo * rel,IndexOptInfo * index,IndexClauseSet * clauses,List ** bitindexpaths)726 get_index_paths(PlannerInfo *root, RelOptInfo *rel,
727 IndexOptInfo *index, IndexClauseSet *clauses,
728 List **bitindexpaths)
729 {
730 List *indexpaths;
731 bool skip_nonnative_saop = false;
732 bool skip_lower_saop = false;
733 ListCell *lc;
734
735 /*
736 * Build simple index paths using the clauses. Allow ScalarArrayOpExpr
737 * clauses only if the index AM supports them natively, and skip any such
738 * clauses for index columns after the first (so that we produce ordered
739 * paths if possible).
740 */
741 indexpaths = build_index_paths(root, rel,
742 index, clauses,
743 index->predOK,
744 ST_ANYSCAN,
745 &skip_nonnative_saop,
746 &skip_lower_saop);
747
748 /*
749 * If we skipped any lower-order ScalarArrayOpExprs on an index with an AM
750 * that supports them, then try again including those clauses. This will
751 * produce paths with more selectivity but no ordering.
752 */
753 if (skip_lower_saop)
754 {
755 indexpaths = list_concat(indexpaths,
756 build_index_paths(root, rel,
757 index, clauses,
758 index->predOK,
759 ST_ANYSCAN,
760 &skip_nonnative_saop,
761 NULL));
762 }
763
764 /*
765 * Submit all the ones that can form plain IndexScan plans to add_path. (A
766 * plain IndexPath can represent either a plain IndexScan or an
767 * IndexOnlyScan, but for our purposes here that distinction does not
768 * matter. However, some of the indexes might support only bitmap scans,
769 * and those we mustn't submit to add_path here.)
770 *
771 * Also, pick out the ones that are usable as bitmap scans. For that, we
772 * must discard indexes that don't support bitmap scans, and we also are
773 * only interested in paths that have some selectivity; we should discard
774 * anything that was generated solely for ordering purposes.
775 */
776 foreach(lc, indexpaths)
777 {
778 IndexPath *ipath = (IndexPath *) lfirst(lc);
779
780 if (index->amhasgettuple)
781 add_path(rel, (Path *) ipath);
782
783 if (index->amhasgetbitmap &&
784 (ipath->path.pathkeys == NIL ||
785 ipath->indexselectivity < 1.0))
786 *bitindexpaths = lappend(*bitindexpaths, ipath);
787 }
788
789 /*
790 * If there were ScalarArrayOpExpr clauses that the index can't handle
791 * natively, generate bitmap scan paths relying on executor-managed
792 * ScalarArrayOpExpr.
793 */
794 if (skip_nonnative_saop)
795 {
796 indexpaths = build_index_paths(root, rel,
797 index, clauses,
798 false,
799 ST_BITMAPSCAN,
800 NULL,
801 NULL);
802 *bitindexpaths = list_concat(*bitindexpaths, indexpaths);
803 }
804 }
805
806 /*
807 * build_index_paths
808 * Given an index and a set of index clauses for it, construct zero
809 * or more IndexPaths. It also constructs zero or more partial IndexPaths.
810 *
811 * We return a list of paths because (1) this routine checks some cases
812 * that should cause us to not generate any IndexPath, and (2) in some
813 * cases we want to consider both a forward and a backward scan, so as
814 * to obtain both sort orders. Note that the paths are just returned
815 * to the caller and not immediately fed to add_path().
816 *
817 * At top level, useful_predicate should be exactly the index's predOK flag
818 * (ie, true if it has a predicate that was proven from the restriction
819 * clauses). When working on an arm of an OR clause, useful_predicate
820 * should be true if the predicate required the current OR list to be proven.
821 * Note that this routine should never be called at all if the index has an
822 * unprovable predicate.
823 *
824 * scantype indicates whether we want to create plain indexscans, bitmap
825 * indexscans, or both. When it's ST_BITMAPSCAN, we will not consider
826 * index ordering while deciding if a Path is worth generating.
827 *
828 * If skip_nonnative_saop is non-NULL, we ignore ScalarArrayOpExpr clauses
829 * unless the index AM supports them directly, and we set *skip_nonnative_saop
830 * to true if we found any such clauses (caller must initialize the variable
831 * to false). If it's NULL, we do not ignore ScalarArrayOpExpr clauses.
832 *
833 * If skip_lower_saop is non-NULL, we ignore ScalarArrayOpExpr clauses for
834 * non-first index columns, and we set *skip_lower_saop to true if we found
835 * any such clauses (caller must initialize the variable to false). If it's
836 * NULL, we do not ignore non-first ScalarArrayOpExpr clauses, but they will
837 * result in considering the scan's output to be unordered.
838 *
839 * 'rel' is the index's heap relation
840 * 'index' is the index for which we want to generate paths
841 * 'clauses' is the collection of indexable clauses (RestrictInfo nodes)
842 * 'useful_predicate' indicates whether the index has a useful predicate
843 * 'scantype' indicates whether we need plain or bitmap scan support
844 * 'skip_nonnative_saop' indicates whether to accept SAOP if index AM doesn't
845 * 'skip_lower_saop' indicates whether to accept non-first-column SAOP
846 */
847 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)848 build_index_paths(PlannerInfo *root, RelOptInfo *rel,
849 IndexOptInfo *index, IndexClauseSet *clauses,
850 bool useful_predicate,
851 ScanTypeControl scantype,
852 bool *skip_nonnative_saop,
853 bool *skip_lower_saop)
854 {
855 List *result = NIL;
856 IndexPath *ipath;
857 List *index_clauses;
858 List *clause_columns;
859 Relids outer_relids;
860 double loop_count;
861 List *orderbyclauses;
862 List *orderbyclausecols;
863 List *index_pathkeys;
864 List *useful_pathkeys;
865 bool found_lower_saop_clause;
866 bool pathkeys_possibly_useful;
867 bool index_is_ordered;
868 bool index_only_scan;
869 int indexcol;
870
871 /*
872 * Check that index supports the desired scan type(s)
873 */
874 switch (scantype)
875 {
876 case ST_INDEXSCAN:
877 if (!index->amhasgettuple)
878 return NIL;
879 break;
880 case ST_BITMAPSCAN:
881 if (!index->amhasgetbitmap)
882 return NIL;
883 break;
884 case ST_ANYSCAN:
885 /* either or both are OK */
886 break;
887 }
888
889 /*
890 * 1. Collect the index clauses into a single list.
891 *
892 * We build a list of RestrictInfo nodes for clauses to be used with this
893 * index, along with an integer list of the index column numbers (zero
894 * based) that each clause should be used with. The clauses are ordered
895 * by index key, so that the column numbers form a nondecreasing sequence.
896 * (This order is depended on by btree and possibly other places.) The
897 * lists can be empty, if the index AM allows that.
898 *
899 * found_lower_saop_clause is set true if we accept a ScalarArrayOpExpr
900 * index clause for a non-first index column. This prevents us from
901 * assuming that the scan result is ordered. (Actually, the result is
902 * still ordered if there are equality constraints for all earlier
903 * columns, but it seems too expensive and non-modular for this code to be
904 * aware of that refinement.)
905 *
906 * We also build a Relids set showing which outer rels are required by the
907 * selected clauses. Any lateral_relids are included in that, but not
908 * otherwise accounted for.
909 */
910 index_clauses = NIL;
911 clause_columns = NIL;
912 found_lower_saop_clause = false;
913 outer_relids = bms_copy(rel->lateral_relids);
914 for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
915 {
916 ListCell *lc;
917
918 foreach(lc, clauses->indexclauses[indexcol])
919 {
920 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
921
922 if (IsA(rinfo->clause, ScalarArrayOpExpr))
923 {
924 if (!index->amsearcharray)
925 {
926 if (skip_nonnative_saop)
927 {
928 /* Ignore because not supported by index */
929 *skip_nonnative_saop = true;
930 continue;
931 }
932 /* Caller had better intend this only for bitmap scan */
933 Assert(scantype == ST_BITMAPSCAN);
934 }
935 if (indexcol > 0)
936 {
937 if (skip_lower_saop)
938 {
939 /* Caller doesn't want to lose index ordering */
940 *skip_lower_saop = true;
941 continue;
942 }
943 found_lower_saop_clause = true;
944 }
945 }
946 index_clauses = lappend(index_clauses, rinfo);
947 clause_columns = lappend_int(clause_columns, indexcol);
948 outer_relids = bms_add_members(outer_relids,
949 rinfo->clause_relids);
950 }
951
952 /*
953 * If no clauses match the first index column, check for amoptionalkey
954 * restriction. We can't generate a scan over an index with
955 * amoptionalkey = false unless there's at least one index clause.
956 * (When working on columns after the first, this test cannot fail. It
957 * is always okay for columns after the first to not have any
958 * clauses.)
959 */
960 if (index_clauses == NIL && !index->amoptionalkey)
961 return NIL;
962 }
963
964 /* We do not want the index's rel itself listed in outer_relids */
965 outer_relids = bms_del_member(outer_relids, rel->relid);
966 /* Enforce convention that outer_relids is exactly NULL if empty */
967 if (bms_is_empty(outer_relids))
968 outer_relids = NULL;
969
970 /* Compute loop_count for cost estimation purposes */
971 loop_count = get_loop_count(root, rel->relid, outer_relids);
972
973 /*
974 * 2. Compute pathkeys describing index's ordering, if any, then see how
975 * many of them are actually useful for this query. This is not relevant
976 * if we are only trying to build bitmap indexscans, nor if we have to
977 * assume the scan is unordered.
978 */
979 pathkeys_possibly_useful = (scantype != ST_BITMAPSCAN &&
980 !found_lower_saop_clause &&
981 has_useful_pathkeys(root, rel));
982 index_is_ordered = (index->sortopfamily != NULL);
983 if (index_is_ordered && pathkeys_possibly_useful)
984 {
985 index_pathkeys = build_index_pathkeys(root, index,
986 ForwardScanDirection);
987 useful_pathkeys = truncate_useless_pathkeys(root, rel,
988 index_pathkeys);
989 orderbyclauses = NIL;
990 orderbyclausecols = NIL;
991 }
992 else if (index->amcanorderbyop && pathkeys_possibly_useful)
993 {
994 /* see if we can generate ordering operators for query_pathkeys */
995 match_pathkeys_to_index(index, root->query_pathkeys,
996 &orderbyclauses,
997 &orderbyclausecols);
998 if (orderbyclauses)
999 useful_pathkeys = root->query_pathkeys;
1000 else
1001 useful_pathkeys = NIL;
1002 }
1003 else
1004 {
1005 useful_pathkeys = NIL;
1006 orderbyclauses = NIL;
1007 orderbyclausecols = NIL;
1008 }
1009
1010 /*
1011 * 3. Check if an index-only scan is possible. If we're not building
1012 * plain indexscans, this isn't relevant since bitmap scans don't support
1013 * index data retrieval anyway.
1014 */
1015 index_only_scan = (scantype != ST_BITMAPSCAN &&
1016 check_index_only(rel, index));
1017
1018 /*
1019 * 4. Generate an indexscan path if there are relevant restriction clauses
1020 * in the current clauses, OR the index ordering is potentially useful for
1021 * later merging or final output ordering, OR the index has a useful
1022 * predicate, OR an index-only scan is possible.
1023 */
1024 if (index_clauses != NIL || useful_pathkeys != NIL || useful_predicate ||
1025 index_only_scan)
1026 {
1027 ipath = create_index_path(root, index,
1028 index_clauses,
1029 clause_columns,
1030 orderbyclauses,
1031 orderbyclausecols,
1032 useful_pathkeys,
1033 index_is_ordered ?
1034 ForwardScanDirection :
1035 NoMovementScanDirection,
1036 index_only_scan,
1037 outer_relids,
1038 loop_count,
1039 false);
1040 result = lappend(result, ipath);
1041
1042 /*
1043 * If appropriate, consider parallel index scan. We don't allow
1044 * parallel index scan for bitmap index scans.
1045 */
1046 if (index->amcanparallel &&
1047 rel->consider_parallel && outer_relids == NULL &&
1048 scantype != ST_BITMAPSCAN)
1049 {
1050 ipath = create_index_path(root, index,
1051 index_clauses,
1052 clause_columns,
1053 orderbyclauses,
1054 orderbyclausecols,
1055 useful_pathkeys,
1056 index_is_ordered ?
1057 ForwardScanDirection :
1058 NoMovementScanDirection,
1059 index_only_scan,
1060 outer_relids,
1061 loop_count,
1062 true);
1063
1064 /*
1065 * if, after costing the path, we find that it's not worth using
1066 * parallel workers, just free it.
1067 */
1068 if (ipath->path.parallel_workers > 0)
1069 add_partial_path(rel, (Path *) ipath);
1070 else
1071 pfree(ipath);
1072 }
1073 }
1074
1075 /*
1076 * 5. If the index is ordered, a backwards scan might be interesting.
1077 */
1078 if (index_is_ordered && pathkeys_possibly_useful)
1079 {
1080 index_pathkeys = build_index_pathkeys(root, index,
1081 BackwardScanDirection);
1082 useful_pathkeys = truncate_useless_pathkeys(root, rel,
1083 index_pathkeys);
1084 if (useful_pathkeys != NIL)
1085 {
1086 ipath = create_index_path(root, index,
1087 index_clauses,
1088 clause_columns,
1089 NIL,
1090 NIL,
1091 useful_pathkeys,
1092 BackwardScanDirection,
1093 index_only_scan,
1094 outer_relids,
1095 loop_count,
1096 false);
1097 result = lappend(result, ipath);
1098
1099 /* If appropriate, consider parallel index scan */
1100 if (index->amcanparallel &&
1101 rel->consider_parallel && outer_relids == NULL &&
1102 scantype != ST_BITMAPSCAN)
1103 {
1104 ipath = create_index_path(root, index,
1105 index_clauses,
1106 clause_columns,
1107 NIL,
1108 NIL,
1109 useful_pathkeys,
1110 BackwardScanDirection,
1111 index_only_scan,
1112 outer_relids,
1113 loop_count,
1114 true);
1115
1116 /*
1117 * if, after costing the path, we find that it's not worth
1118 * using parallel workers, just free it.
1119 */
1120 if (ipath->path.parallel_workers > 0)
1121 add_partial_path(rel, (Path *) ipath);
1122 else
1123 pfree(ipath);
1124 }
1125 }
1126 }
1127
1128 return result;
1129 }
1130
1131 /*
1132 * build_paths_for_OR
1133 * Given a list of restriction clauses from one arm of an OR clause,
1134 * construct all matching IndexPaths for the relation.
1135 *
1136 * Here we must scan all indexes of the relation, since a bitmap OR tree
1137 * can use multiple indexes.
1138 *
1139 * The caller actually supplies two lists of restriction clauses: some
1140 * "current" ones and some "other" ones. Both lists can be used freely
1141 * to match keys of the index, but an index must use at least one of the
1142 * "current" clauses to be considered usable. The motivation for this is
1143 * examples like
1144 * WHERE (x = 42) AND (... OR (y = 52 AND z = 77) OR ....)
1145 * While we are considering the y/z subclause of the OR, we can use "x = 42"
1146 * as one of the available index conditions; but we shouldn't match the
1147 * subclause to any index on x alone, because such a Path would already have
1148 * been generated at the upper level. So we could use an index on x,y,z
1149 * or an index on x,y for the OR subclause, but not an index on just x.
1150 * When dealing with a partial index, a match of the index predicate to
1151 * one of the "current" clauses also makes the index usable.
1152 *
1153 * 'rel' is the relation for which we want to generate index paths
1154 * 'clauses' is the current list of clauses (RestrictInfo nodes)
1155 * 'other_clauses' is the list of additional upper-level clauses
1156 */
1157 static List *
build_paths_for_OR(PlannerInfo * root,RelOptInfo * rel,List * clauses,List * other_clauses)1158 build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
1159 List *clauses, List *other_clauses)
1160 {
1161 List *result = NIL;
1162 List *all_clauses = NIL; /* not computed till needed */
1163 ListCell *lc;
1164
1165 foreach(lc, rel->indexlist)
1166 {
1167 IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
1168 IndexClauseSet clauseset;
1169 List *indexpaths;
1170 bool useful_predicate;
1171
1172 /* Ignore index if it doesn't support bitmap scans */
1173 if (!index->amhasgetbitmap)
1174 continue;
1175
1176 /*
1177 * Ignore partial indexes that do not match the query. If a partial
1178 * index is marked predOK then we know it's OK. Otherwise, we have to
1179 * test whether the added clauses are sufficient to imply the
1180 * predicate. If so, we can use the index in the current context.
1181 *
1182 * We set useful_predicate to true iff the predicate was proven using
1183 * the current set of clauses. This is needed to prevent matching a
1184 * predOK index to an arm of an OR, which would be a legal but
1185 * pointlessly inefficient plan. (A better plan will be generated by
1186 * just scanning the predOK index alone, no OR.)
1187 */
1188 useful_predicate = false;
1189 if (index->indpred != NIL)
1190 {
1191 if (index->predOK)
1192 {
1193 /* Usable, but don't set useful_predicate */
1194 }
1195 else
1196 {
1197 /* Form all_clauses if not done already */
1198 if (all_clauses == NIL)
1199 all_clauses = list_concat(list_copy(clauses),
1200 other_clauses);
1201
1202 if (!predicate_implied_by(index->indpred, all_clauses, false))
1203 continue; /* can't use it at all */
1204
1205 if (!predicate_implied_by(index->indpred, other_clauses, false))
1206 useful_predicate = true;
1207 }
1208 }
1209
1210 /*
1211 * Identify the restriction clauses that can match the index.
1212 */
1213 MemSet(&clauseset, 0, sizeof(clauseset));
1214 match_clauses_to_index(index, clauses, &clauseset);
1215
1216 /*
1217 * If no matches so far, and the index predicate isn't useful, we
1218 * don't want it.
1219 */
1220 if (!clauseset.nonempty && !useful_predicate)
1221 continue;
1222
1223 /*
1224 * Add "other" restriction clauses to the clauseset.
1225 */
1226 match_clauses_to_index(index, other_clauses, &clauseset);
1227
1228 /*
1229 * Construct paths if possible.
1230 */
1231 indexpaths = build_index_paths(root, rel,
1232 index, &clauseset,
1233 useful_predicate,
1234 ST_BITMAPSCAN,
1235 NULL,
1236 NULL);
1237 result = list_concat(result, indexpaths);
1238 }
1239
1240 return result;
1241 }
1242
1243 /*
1244 * generate_bitmap_or_paths
1245 * Look through the list of clauses to find OR clauses, and generate
1246 * a BitmapOrPath for each one we can handle that way. Return a list
1247 * of the generated BitmapOrPaths.
1248 *
1249 * other_clauses is a list of additional clauses that can be assumed true
1250 * for the purpose of generating indexquals, but are not to be searched for
1251 * ORs. (See build_paths_for_OR() for motivation.)
1252 */
1253 static List *
generate_bitmap_or_paths(PlannerInfo * root,RelOptInfo * rel,List * clauses,List * other_clauses)1254 generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
1255 List *clauses, List *other_clauses)
1256 {
1257 List *result = NIL;
1258 List *all_clauses;
1259 ListCell *lc;
1260
1261 /*
1262 * We can use both the current and other clauses as context for
1263 * build_paths_for_OR; no need to remove ORs from the lists.
1264 */
1265 all_clauses = list_concat(list_copy(clauses), other_clauses);
1266
1267 foreach(lc, clauses)
1268 {
1269 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1270 List *pathlist;
1271 Path *bitmapqual;
1272 ListCell *j;
1273
1274 /* Ignore RestrictInfos that aren't ORs */
1275 if (!restriction_is_or_clause(rinfo))
1276 continue;
1277
1278 /*
1279 * We must be able to match at least one index to each of the arms of
1280 * the OR, else we can't use it.
1281 */
1282 pathlist = NIL;
1283 foreach(j, ((BoolExpr *) rinfo->orclause)->args)
1284 {
1285 Node *orarg = (Node *) lfirst(j);
1286 List *indlist;
1287
1288 /* OR arguments should be ANDs or sub-RestrictInfos */
1289 if (and_clause(orarg))
1290 {
1291 List *andargs = ((BoolExpr *) orarg)->args;
1292
1293 indlist = build_paths_for_OR(root, rel,
1294 andargs,
1295 all_clauses);
1296
1297 /* Recurse in case there are sub-ORs */
1298 indlist = list_concat(indlist,
1299 generate_bitmap_or_paths(root, rel,
1300 andargs,
1301 all_clauses));
1302 }
1303 else
1304 {
1305 RestrictInfo *rinfo = castNode(RestrictInfo, orarg);
1306 List *orargs;
1307
1308 Assert(!restriction_is_or_clause(rinfo));
1309 orargs = list_make1(rinfo);
1310
1311 indlist = build_paths_for_OR(root, rel,
1312 orargs,
1313 all_clauses);
1314 }
1315
1316 /*
1317 * If nothing matched this arm, we can't do anything with this OR
1318 * clause.
1319 */
1320 if (indlist == NIL)
1321 {
1322 pathlist = NIL;
1323 break;
1324 }
1325
1326 /*
1327 * OK, pick the most promising AND combination, and add it to
1328 * pathlist.
1329 */
1330 bitmapqual = choose_bitmap_and(root, rel, indlist);
1331 pathlist = lappend(pathlist, bitmapqual);
1332 }
1333
1334 /*
1335 * If we have a match for every arm, then turn them into a
1336 * BitmapOrPath, and add to result list.
1337 */
1338 if (pathlist != NIL)
1339 {
1340 bitmapqual = (Path *) create_bitmap_or_path(root, rel, pathlist);
1341 result = lappend(result, bitmapqual);
1342 }
1343 }
1344
1345 return result;
1346 }
1347
1348
1349 /*
1350 * choose_bitmap_and
1351 * Given a nonempty list of bitmap paths, AND them into one path.
1352 *
1353 * This is a nontrivial decision since we can legally use any subset of the
1354 * given path set. We want to choose a good tradeoff between selectivity
1355 * and cost of computing the bitmap.
1356 *
1357 * The result is either a single one of the inputs, or a BitmapAndPath
1358 * combining multiple inputs.
1359 */
1360 static Path *
choose_bitmap_and(PlannerInfo * root,RelOptInfo * rel,List * paths)1361 choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel, List *paths)
1362 {
1363 int npaths = list_length(paths);
1364 PathClauseUsage **pathinfoarray;
1365 PathClauseUsage *pathinfo;
1366 List *clauselist;
1367 List *bestpaths = NIL;
1368 Cost bestcost = 0;
1369 int i,
1370 j;
1371 ListCell *l;
1372
1373 Assert(npaths > 0); /* else caller error */
1374 if (npaths == 1)
1375 return (Path *) linitial(paths); /* easy case */
1376
1377 /*
1378 * In theory we should consider every nonempty subset of the given paths.
1379 * In practice that seems like overkill, given the crude nature of the
1380 * estimates, not to mention the possible effects of higher-level AND and
1381 * OR clauses. Moreover, it's completely impractical if there are a large
1382 * number of paths, since the work would grow as O(2^N).
1383 *
1384 * As a heuristic, we first check for paths using exactly the same sets of
1385 * WHERE clauses + index predicate conditions, and reject all but the
1386 * cheapest-to-scan in any such group. This primarily gets rid of indexes
1387 * that include the interesting columns but also irrelevant columns. (In
1388 * situations where the DBA has gone overboard on creating variant
1389 * indexes, this can make for a very large reduction in the number of
1390 * paths considered further.)
1391 *
1392 * We then sort the surviving paths with the cheapest-to-scan first, and
1393 * for each path, consider using that path alone as the basis for a bitmap
1394 * scan. Then we consider bitmap AND scans formed from that path plus
1395 * each subsequent (higher-cost) path, adding on a subsequent path if it
1396 * results in a reduction in the estimated total scan cost. This means we
1397 * consider about O(N^2) rather than O(2^N) path combinations, which is
1398 * quite tolerable, especially given than N is usually reasonably small
1399 * because of the prefiltering step. The cheapest of these is returned.
1400 *
1401 * We will only consider AND combinations in which no two indexes use the
1402 * same WHERE clause. This is a bit of a kluge: it's needed because
1403 * costsize.c and clausesel.c aren't very smart about redundant clauses.
1404 * They will usually double-count the redundant clauses, producing a
1405 * too-small selectivity that makes a redundant AND step look like it
1406 * reduces the total cost. Perhaps someday that code will be smarter and
1407 * we can remove this limitation. (But note that this also defends
1408 * against flat-out duplicate input paths, which can happen because
1409 * match_join_clauses_to_index will find the same OR join clauses that
1410 * extract_restriction_or_clauses has pulled OR restriction clauses out
1411 * of.)
1412 *
1413 * For the same reason, we reject AND combinations in which an index
1414 * predicate clause duplicates another clause. Here we find it necessary
1415 * to be even stricter: we'll reject a partial index if any of its
1416 * predicate clauses are implied by the set of WHERE clauses and predicate
1417 * clauses used so far. This covers cases such as a condition "x = 42"
1418 * used with a plain index, followed by a clauseless scan of a partial
1419 * index "WHERE x >= 40 AND x < 50". The partial index has been accepted
1420 * only because "x = 42" was present, and so allowing it would partially
1421 * double-count selectivity. (We could use predicate_implied_by on
1422 * regular qual clauses too, to have a more intelligent, but much more
1423 * expensive, check for redundancy --- but in most cases simple equality
1424 * seems to suffice.)
1425 */
1426
1427 /*
1428 * Extract clause usage info and detect any paths that use exactly the
1429 * same set of clauses; keep only the cheapest-to-scan of any such groups.
1430 * The surviving paths are put into an array for qsort'ing.
1431 */
1432 pathinfoarray = (PathClauseUsage **)
1433 palloc(npaths * sizeof(PathClauseUsage *));
1434 clauselist = NIL;
1435 npaths = 0;
1436 foreach(l, paths)
1437 {
1438 Path *ipath = (Path *) lfirst(l);
1439
1440 pathinfo = classify_index_clause_usage(ipath, &clauselist);
1441
1442 /* If it's unclassifiable, treat it as distinct from all others */
1443 if (pathinfo->unclassifiable)
1444 {
1445 pathinfoarray[npaths++] = pathinfo;
1446 continue;
1447 }
1448
1449 for (i = 0; i < npaths; i++)
1450 {
1451 if (!pathinfoarray[i]->unclassifiable &&
1452 bms_equal(pathinfo->clauseids, pathinfoarray[i]->clauseids))
1453 break;
1454 }
1455 if (i < npaths)
1456 {
1457 /* duplicate clauseids, keep the cheaper one */
1458 Cost ncost;
1459 Cost ocost;
1460 Selectivity nselec;
1461 Selectivity oselec;
1462
1463 cost_bitmap_tree_node(pathinfo->path, &ncost, &nselec);
1464 cost_bitmap_tree_node(pathinfoarray[i]->path, &ocost, &oselec);
1465 if (ncost < ocost)
1466 pathinfoarray[i] = pathinfo;
1467 }
1468 else
1469 {
1470 /* not duplicate clauseids, add to array */
1471 pathinfoarray[npaths++] = pathinfo;
1472 }
1473 }
1474
1475 /* If only one surviving path, we're done */
1476 if (npaths == 1)
1477 return pathinfoarray[0]->path;
1478
1479 /* Sort the surviving paths by index access cost */
1480 qsort(pathinfoarray, npaths, sizeof(PathClauseUsage *),
1481 path_usage_comparator);
1482
1483 /*
1484 * For each surviving index, consider it as an "AND group leader", and see
1485 * whether adding on any of the later indexes results in an AND path with
1486 * cheaper total cost than before. Then take the cheapest AND group.
1487 *
1488 * Note: paths that are either clauseless or unclassifiable will have
1489 * empty clauseids, so that they will not be rejected by the clauseids
1490 * filter here, nor will they cause later paths to be rejected by it.
1491 */
1492 for (i = 0; i < npaths; i++)
1493 {
1494 Cost costsofar;
1495 List *qualsofar;
1496 Bitmapset *clauseidsofar;
1497 ListCell *lastcell;
1498
1499 pathinfo = pathinfoarray[i];
1500 paths = list_make1(pathinfo->path);
1501 costsofar = bitmap_scan_cost_est(root, rel, pathinfo->path);
1502 qualsofar = list_concat(list_copy(pathinfo->quals),
1503 list_copy(pathinfo->preds));
1504 clauseidsofar = bms_copy(pathinfo->clauseids);
1505 lastcell = list_head(paths); /* for quick deletions */
1506
1507 for (j = i + 1; j < npaths; j++)
1508 {
1509 Cost newcost;
1510
1511 pathinfo = pathinfoarray[j];
1512 /* Check for redundancy */
1513 if (bms_overlap(pathinfo->clauseids, clauseidsofar))
1514 continue; /* consider it redundant */
1515 if (pathinfo->preds)
1516 {
1517 bool redundant = false;
1518
1519 /* we check each predicate clause separately */
1520 foreach(l, pathinfo->preds)
1521 {
1522 Node *np = (Node *) lfirst(l);
1523
1524 if (predicate_implied_by(list_make1(np), qualsofar, false))
1525 {
1526 redundant = true;
1527 break; /* out of inner foreach loop */
1528 }
1529 }
1530 if (redundant)
1531 continue;
1532 }
1533 /* tentatively add new path to paths, so we can estimate cost */
1534 paths = lappend(paths, pathinfo->path);
1535 newcost = bitmap_and_cost_est(root, rel, paths);
1536 if (newcost < costsofar)
1537 {
1538 /* keep new path in paths, update subsidiary variables */
1539 costsofar = newcost;
1540 qualsofar = list_concat(qualsofar,
1541 list_copy(pathinfo->quals));
1542 qualsofar = list_concat(qualsofar,
1543 list_copy(pathinfo->preds));
1544 clauseidsofar = bms_add_members(clauseidsofar,
1545 pathinfo->clauseids);
1546 lastcell = lnext(lastcell);
1547 }
1548 else
1549 {
1550 /* reject new path, remove it from paths list */
1551 paths = list_delete_cell(paths, lnext(lastcell), lastcell);
1552 }
1553 Assert(lnext(lastcell) == NULL);
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 indexquals 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
1764 *quals = list_concat(*quals, get_actual_clauses(ipath->indexclauses));
1765 *preds = list_concat(*preds, list_copy(ipath->indexinfo->indpred));
1766 }
1767 else
1768 elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
1769 }
1770
1771
1772 /*
1773 * find_list_position
1774 * Return the given node's position (counting from 0) in the given
1775 * list of nodes. If it's not equal() to any existing list member,
1776 * add it at the end, and return that position.
1777 */
1778 static int
find_list_position(Node * node,List ** nodelist)1779 find_list_position(Node *node, List **nodelist)
1780 {
1781 int i;
1782 ListCell *lc;
1783
1784 i = 0;
1785 foreach(lc, *nodelist)
1786 {
1787 Node *oldnode = (Node *) lfirst(lc);
1788
1789 if (equal(node, oldnode))
1790 return i;
1791 i++;
1792 }
1793
1794 *nodelist = lappend(*nodelist, node);
1795
1796 return i;
1797 }
1798
1799
1800 /*
1801 * check_index_only
1802 * Determine whether an index-only scan is possible for this index.
1803 */
1804 static bool
check_index_only(RelOptInfo * rel,IndexOptInfo * index)1805 check_index_only(RelOptInfo *rel, IndexOptInfo *index)
1806 {
1807 bool result;
1808 Bitmapset *attrs_used = NULL;
1809 Bitmapset *index_canreturn_attrs = NULL;
1810 Bitmapset *index_cannotreturn_attrs = NULL;
1811 ListCell *lc;
1812 int i;
1813
1814 /* Index-only scans must be enabled */
1815 if (!enable_indexonlyscan)
1816 return false;
1817
1818 /*
1819 * Check that all needed attributes of the relation are available from the
1820 * index.
1821 */
1822
1823 /*
1824 * First, identify all the attributes needed for joins or final output.
1825 * Note: we must look at rel's targetlist, not the attr_needed data,
1826 * because attr_needed isn't computed for inheritance child rels.
1827 */
1828 pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
1829
1830 /*
1831 * Add all the attributes used by restriction clauses; but consider only
1832 * those clauses not implied by the index predicate, since ones that are
1833 * so implied don't need to be checked explicitly in the plan.
1834 *
1835 * Note: attributes used only in index quals would not be needed at
1836 * runtime either, if we are certain that the index is not lossy. However
1837 * it'd be complicated to account for that accurately, and it doesn't
1838 * matter in most cases, since we'd conclude that such attributes are
1839 * available from the index anyway.
1840 */
1841 foreach(lc, index->indrestrictinfo)
1842 {
1843 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1844
1845 pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
1846 }
1847
1848 /*
1849 * Construct a bitmapset of columns that the index can return back in an
1850 * index-only scan. If there are multiple index columns containing the
1851 * same attribute, all of them must be capable of returning the value,
1852 * since we might recheck operators on any of them. (Potentially we could
1853 * be smarter about that, but it's such a weird situation that it doesn't
1854 * seem worth spending a lot of sweat on.)
1855 */
1856 for (i = 0; i < index->ncolumns; i++)
1857 {
1858 int attno = index->indexkeys[i];
1859
1860 /*
1861 * For the moment, we just ignore index expressions. It might be nice
1862 * to do something with them, later.
1863 */
1864 if (attno == 0)
1865 continue;
1866
1867 if (index->canreturn[i])
1868 index_canreturn_attrs =
1869 bms_add_member(index_canreturn_attrs,
1870 attno - FirstLowInvalidHeapAttributeNumber);
1871 else
1872 index_cannotreturn_attrs =
1873 bms_add_member(index_cannotreturn_attrs,
1874 attno - FirstLowInvalidHeapAttributeNumber);
1875 }
1876
1877 index_canreturn_attrs = bms_del_members(index_canreturn_attrs,
1878 index_cannotreturn_attrs);
1879
1880 /* Do we have all the necessary attributes? */
1881 result = bms_is_subset(attrs_used, index_canreturn_attrs);
1882
1883 bms_free(attrs_used);
1884 bms_free(index_canreturn_attrs);
1885 bms_free(index_cannotreturn_attrs);
1886
1887 return result;
1888 }
1889
1890 /*
1891 * get_loop_count
1892 * Choose the loop count estimate to use for costing a parameterized path
1893 * with the given set of outer relids.
1894 *
1895 * Since we produce parameterized paths before we've begun to generate join
1896 * relations, it's impossible to predict exactly how many times a parameterized
1897 * path will be iterated; we don't know the size of the relation that will be
1898 * on the outside of the nestloop. However, we should try to account for
1899 * multiple iterations somehow in costing the path. The heuristic embodied
1900 * here is to use the rowcount of the smallest other base relation needed in
1901 * the join clauses used by the path. (We could alternatively consider the
1902 * largest one, but that seems too optimistic.) This is of course the right
1903 * answer for single-other-relation cases, and it seems like a reasonable
1904 * zero-order approximation for multiway-join cases.
1905 *
1906 * In addition, we check to see if the other side of each join clause is on
1907 * the inside of some semijoin that the current relation is on the outside of.
1908 * If so, the only way that a parameterized path could be used is if the
1909 * semijoin RHS has been unique-ified, so we should use the number of unique
1910 * RHS rows rather than using the relation's raw rowcount.
1911 *
1912 * Note: for this to work, allpaths.c must establish all baserel size
1913 * estimates before it begins to compute paths, or at least before it
1914 * calls create_index_paths().
1915 */
1916 static double
get_loop_count(PlannerInfo * root,Index cur_relid,Relids outer_relids)1917 get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids)
1918 {
1919 double result;
1920 int outer_relid;
1921
1922 /* For a non-parameterized path, just return 1.0 quickly */
1923 if (outer_relids == NULL)
1924 return 1.0;
1925
1926 result = 0.0;
1927 outer_relid = -1;
1928 while ((outer_relid = bms_next_member(outer_relids, outer_relid)) >= 0)
1929 {
1930 RelOptInfo *outer_rel;
1931 double rowcount;
1932
1933 /* Paranoia: ignore bogus relid indexes */
1934 if (outer_relid >= root->simple_rel_array_size)
1935 continue;
1936 outer_rel = root->simple_rel_array[outer_relid];
1937 if (outer_rel == NULL)
1938 continue;
1939 Assert(outer_rel->relid == outer_relid); /* sanity check on array */
1940
1941 /* Other relation could be proven empty, if so ignore */
1942 if (IS_DUMMY_REL(outer_rel))
1943 continue;
1944
1945 /* Otherwise, rel's rows estimate should be valid by now */
1946 Assert(outer_rel->rows > 0);
1947
1948 /* Check to see if rel is on the inside of any semijoins */
1949 rowcount = adjust_rowcount_for_semijoins(root,
1950 cur_relid,
1951 outer_relid,
1952 outer_rel->rows);
1953
1954 /* Remember smallest row count estimate among the outer rels */
1955 if (result == 0.0 || result > rowcount)
1956 result = rowcount;
1957 }
1958 /* Return 1.0 if we found no valid relations (shouldn't happen) */
1959 return (result > 0.0) ? result : 1.0;
1960 }
1961
1962 /*
1963 * Check to see if outer_relid is on the inside of any semijoin that cur_relid
1964 * is on the outside of. If so, replace rowcount with the estimated number of
1965 * unique rows from the semijoin RHS (assuming that's smaller, which it might
1966 * not be). The estimate is crude but it's the best we can do at this stage
1967 * of the proceedings.
1968 */
1969 static double
adjust_rowcount_for_semijoins(PlannerInfo * root,Index cur_relid,Index outer_relid,double rowcount)1970 adjust_rowcount_for_semijoins(PlannerInfo *root,
1971 Index cur_relid,
1972 Index outer_relid,
1973 double rowcount)
1974 {
1975 ListCell *lc;
1976
1977 foreach(lc, root->join_info_list)
1978 {
1979 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
1980
1981 if (sjinfo->jointype == JOIN_SEMI &&
1982 bms_is_member(cur_relid, sjinfo->syn_lefthand) &&
1983 bms_is_member(outer_relid, sjinfo->syn_righthand))
1984 {
1985 /* Estimate number of unique-ified rows */
1986 double nraw;
1987 double nunique;
1988
1989 nraw = approximate_joinrel_size(root, sjinfo->syn_righthand);
1990 nunique = estimate_num_groups(root,
1991 sjinfo->semi_rhs_exprs,
1992 nraw,
1993 NULL);
1994 if (rowcount > nunique)
1995 rowcount = nunique;
1996 }
1997 }
1998 return rowcount;
1999 }
2000
2001 /*
2002 * Make an approximate estimate of the size of a joinrel.
2003 *
2004 * We don't have enough info at this point to get a good estimate, so we
2005 * just multiply the base relation sizes together. Fortunately, this is
2006 * the right answer anyway for the most common case with a single relation
2007 * on the RHS of a semijoin. Also, estimate_num_groups() has only a weak
2008 * dependency on its input_rows argument (it basically uses it as a clamp).
2009 * So we might be able to get a fairly decent end result even with a severe
2010 * overestimate of the RHS's raw size.
2011 */
2012 static double
approximate_joinrel_size(PlannerInfo * root,Relids relids)2013 approximate_joinrel_size(PlannerInfo *root, Relids relids)
2014 {
2015 double rowcount = 1.0;
2016 int relid;
2017
2018 relid = -1;
2019 while ((relid = bms_next_member(relids, relid)) >= 0)
2020 {
2021 RelOptInfo *rel;
2022
2023 /* Paranoia: ignore bogus relid indexes */
2024 if (relid >= root->simple_rel_array_size)
2025 continue;
2026 rel = root->simple_rel_array[relid];
2027 if (rel == NULL)
2028 continue;
2029 Assert(rel->relid == relid); /* sanity check on array */
2030
2031 /* Relation could be proven empty, if so ignore */
2032 if (IS_DUMMY_REL(rel))
2033 continue;
2034
2035 /* Otherwise, rel's rows estimate should be valid by now */
2036 Assert(rel->rows > 0);
2037
2038 /* Accumulate product */
2039 rowcount *= rel->rows;
2040 }
2041 return rowcount;
2042 }
2043
2044
2045 /****************************************************************************
2046 * ---- ROUTINES TO CHECK QUERY CLAUSES ----
2047 ****************************************************************************/
2048
2049 /*
2050 * match_restriction_clauses_to_index
2051 * Identify restriction clauses for the rel that match the index.
2052 * Matching clauses are added to *clauseset.
2053 */
2054 static void
match_restriction_clauses_to_index(RelOptInfo * rel,IndexOptInfo * index,IndexClauseSet * clauseset)2055 match_restriction_clauses_to_index(RelOptInfo *rel, IndexOptInfo *index,
2056 IndexClauseSet *clauseset)
2057 {
2058 /* We can ignore clauses that are implied by the index predicate */
2059 match_clauses_to_index(index, index->indrestrictinfo, clauseset);
2060 }
2061
2062 /*
2063 * match_join_clauses_to_index
2064 * Identify join clauses for the rel that match the index.
2065 * Matching clauses are added to *clauseset.
2066 * Also, add any potentially usable join OR clauses to *joinorclauses.
2067 */
2068 static void
match_join_clauses_to_index(PlannerInfo * root,RelOptInfo * rel,IndexOptInfo * index,IndexClauseSet * clauseset,List ** joinorclauses)2069 match_join_clauses_to_index(PlannerInfo *root,
2070 RelOptInfo *rel, IndexOptInfo *index,
2071 IndexClauseSet *clauseset,
2072 List **joinorclauses)
2073 {
2074 ListCell *lc;
2075
2076 /* Scan the rel's join clauses */
2077 foreach(lc, rel->joininfo)
2078 {
2079 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2080
2081 /* Check if clause can be moved to this rel */
2082 if (!join_clause_is_movable_to(rinfo, rel))
2083 continue;
2084
2085 /* Potentially usable, so see if it matches the index or is an OR */
2086 if (restriction_is_or_clause(rinfo))
2087 *joinorclauses = lappend(*joinorclauses, rinfo);
2088 else
2089 match_clause_to_index(index, rinfo, clauseset);
2090 }
2091 }
2092
2093 /*
2094 * match_eclass_clauses_to_index
2095 * Identify EquivalenceClass join clauses for the rel that match the index.
2096 * Matching clauses are added to *clauseset.
2097 */
2098 static void
match_eclass_clauses_to_index(PlannerInfo * root,IndexOptInfo * index,IndexClauseSet * clauseset)2099 match_eclass_clauses_to_index(PlannerInfo *root, IndexOptInfo *index,
2100 IndexClauseSet *clauseset)
2101 {
2102 int indexcol;
2103
2104 /* No work if rel is not in any such ECs */
2105 if (!index->rel->has_eclass_joins)
2106 return;
2107
2108 for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2109 {
2110 ec_member_matches_arg arg;
2111 List *clauses;
2112
2113 /* Generate clauses, skipping any that join to lateral_referencers */
2114 arg.index = index;
2115 arg.indexcol = indexcol;
2116 clauses = generate_implied_equalities_for_column(root,
2117 index->rel,
2118 ec_member_matches_indexcol,
2119 (void *) &arg,
2120 index->rel->lateral_referencers);
2121
2122 /*
2123 * We have to check whether the results actually do match the index,
2124 * since for non-btree indexes the EC's equality operators might not
2125 * be in the index opclass (cf ec_member_matches_indexcol).
2126 */
2127 match_clauses_to_index(index, clauses, clauseset);
2128 }
2129 }
2130
2131 /*
2132 * match_clauses_to_index
2133 * Perform match_clause_to_index() for each clause in a list.
2134 * Matching clauses are added to *clauseset.
2135 */
2136 static void
match_clauses_to_index(IndexOptInfo * index,List * clauses,IndexClauseSet * clauseset)2137 match_clauses_to_index(IndexOptInfo *index,
2138 List *clauses,
2139 IndexClauseSet *clauseset)
2140 {
2141 ListCell *lc;
2142
2143 foreach(lc, clauses)
2144 {
2145 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
2146
2147 match_clause_to_index(index, rinfo, clauseset);
2148 }
2149 }
2150
2151 /*
2152 * match_clause_to_index
2153 * Test whether a qual clause can be used with an index.
2154 *
2155 * If the clause is usable, add it to the appropriate list in *clauseset.
2156 * *clauseset must be initialized to zeroes before first call.
2157 *
2158 * Note: in some circumstances we may find the same RestrictInfos coming from
2159 * multiple places. Defend against redundant outputs by refusing to add a
2160 * clause twice (pointer equality should be a good enough check for this).
2161 *
2162 * Note: it's possible that a badly-defined index could have multiple matching
2163 * columns. We always select the first match if so; this avoids scenarios
2164 * wherein we get an inflated idea of the index's selectivity by using the
2165 * same clause multiple times with different index columns.
2166 */
2167 static void
match_clause_to_index(IndexOptInfo * index,RestrictInfo * rinfo,IndexClauseSet * clauseset)2168 match_clause_to_index(IndexOptInfo *index,
2169 RestrictInfo *rinfo,
2170 IndexClauseSet *clauseset)
2171 {
2172 int indexcol;
2173
2174 /*
2175 * Never match pseudoconstants to indexes. (Normally a match could not
2176 * happen anyway, since a pseudoconstant clause couldn't contain a Var,
2177 * but what if someone builds an expression index on a constant? It's not
2178 * totally unreasonable to do so with a partial index, either.)
2179 */
2180 if (rinfo->pseudoconstant)
2181 return;
2182
2183 /*
2184 * If clause can't be used as an indexqual because it must wait till after
2185 * some lower-security-level restriction clause, reject it.
2186 */
2187 if (!restriction_is_securely_promotable(rinfo, index->rel))
2188 return;
2189
2190 /* OK, check each index key column for a match */
2191 for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2192 {
2193 if (match_clause_to_indexcol(index,
2194 indexcol,
2195 rinfo))
2196 {
2197 clauseset->indexclauses[indexcol] =
2198 list_append_unique_ptr(clauseset->indexclauses[indexcol],
2199 rinfo);
2200 clauseset->nonempty = true;
2201 return;
2202 }
2203 }
2204 }
2205
2206 /*
2207 * match_clause_to_indexcol()
2208 * Determines whether a restriction clause matches a column of an index.
2209 *
2210 * To match an index normally, the clause:
2211 *
2212 * (1) must be in the form (indexkey op const) or (const op indexkey);
2213 * and
2214 * (2) must contain an operator which is in the same family as the index
2215 * operator for this column, or is a "special" operator as recognized
2216 * by match_special_index_operator();
2217 * and
2218 * (3) must match the collation of the index, if collation is relevant.
2219 *
2220 * Our definition of "const" is exceedingly liberal: we allow anything that
2221 * doesn't involve a volatile function or a Var of the index's relation.
2222 * In particular, Vars belonging to other relations of the query are
2223 * accepted here, since a clause of that form can be used in a
2224 * parameterized indexscan. It's the responsibility of higher code levels
2225 * to manage restriction and join clauses appropriately.
2226 *
2227 * Note: we do need to check for Vars of the index's relation on the
2228 * "const" side of the clause, since clauses like (a.f1 OP (b.f2 OP a.f3))
2229 * are not processable by a parameterized indexscan on a.f1, whereas
2230 * something like (a.f1 OP (b.f2 OP c.f3)) is.
2231 *
2232 * Presently, the executor can only deal with indexquals that have the
2233 * indexkey on the left, so we can only use clauses that have the indexkey
2234 * on the right if we can commute the clause to put the key on the left.
2235 * We do not actually do the commuting here, but we check whether a
2236 * suitable commutator operator is available.
2237 *
2238 * If the index has a collation, the clause must have the same collation.
2239 * For collation-less indexes, we assume it doesn't matter; this is
2240 * necessary for cases like "hstore ? text", wherein hstore's operators
2241 * don't care about collation but the clause will get marked with a
2242 * collation anyway because of the text argument. (This logic is
2243 * embodied in the macro IndexCollMatchesExprColl.)
2244 *
2245 * It is also possible to match RowCompareExpr clauses to indexes (but
2246 * currently, only btree indexes handle this). In this routine we will
2247 * report a match if the first column of the row comparison matches the
2248 * target index column. This is sufficient to guarantee that some index
2249 * condition can be constructed from the RowCompareExpr --- whether the
2250 * remaining columns match the index too is considered in
2251 * adjust_rowcompare_for_index().
2252 *
2253 * It is also possible to match ScalarArrayOpExpr clauses to indexes, when
2254 * the clause is of the form "indexkey op ANY (arrayconst)".
2255 *
2256 * For boolean indexes, it is also possible to match the clause directly
2257 * to the indexkey; or perhaps the clause is (NOT indexkey).
2258 *
2259 * 'index' is the index of interest.
2260 * 'indexcol' is a column number of 'index' (counting from 0).
2261 * 'rinfo' is the clause to be tested (as a RestrictInfo node).
2262 *
2263 * Returns true if the clause can be used with this index key.
2264 *
2265 * NOTE: returns false if clause is an OR or AND clause; it is the
2266 * responsibility of higher-level routines to cope with those.
2267 */
2268 static bool
match_clause_to_indexcol(IndexOptInfo * index,int indexcol,RestrictInfo * rinfo)2269 match_clause_to_indexcol(IndexOptInfo *index,
2270 int indexcol,
2271 RestrictInfo *rinfo)
2272 {
2273 Expr *clause = rinfo->clause;
2274 Index index_relid = index->rel->relid;
2275 Oid opfamily;
2276 Oid idxcollation;
2277 Node *leftop,
2278 *rightop;
2279 Relids left_relids;
2280 Relids right_relids;
2281 Oid expr_op;
2282 Oid expr_coll;
2283 bool plain_op;
2284
2285 Assert(indexcol < index->nkeycolumns);
2286
2287 opfamily = index->opfamily[indexcol];
2288 idxcollation = index->indexcollations[indexcol];
2289
2290 /* First check for boolean-index cases. */
2291 if (IsBooleanOpfamily(opfamily))
2292 {
2293 if (match_boolean_index_clause((Node *) clause, indexcol, index))
2294 return true;
2295 }
2296
2297 /*
2298 * Clause must be a binary opclause, or possibly a ScalarArrayOpExpr
2299 * (which is always binary, by definition). Or it could be a
2300 * RowCompareExpr, which we pass off to match_rowcompare_to_indexcol().
2301 * Or, if the index supports it, we can handle IS NULL/NOT NULL clauses.
2302 */
2303 if (is_opclause(clause))
2304 {
2305 leftop = get_leftop(clause);
2306 rightop = get_rightop(clause);
2307 if (!leftop || !rightop)
2308 return false;
2309 left_relids = rinfo->left_relids;
2310 right_relids = rinfo->right_relids;
2311 expr_op = ((OpExpr *) clause)->opno;
2312 expr_coll = ((OpExpr *) clause)->inputcollid;
2313 plain_op = true;
2314 }
2315 else if (clause && IsA(clause, ScalarArrayOpExpr))
2316 {
2317 ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
2318
2319 /* We only accept ANY clauses, not ALL */
2320 if (!saop->useOr)
2321 return false;
2322 leftop = (Node *) linitial(saop->args);
2323 rightop = (Node *) lsecond(saop->args);
2324 left_relids = NULL; /* not actually needed */
2325 right_relids = pull_varnos(rightop);
2326 expr_op = saop->opno;
2327 expr_coll = saop->inputcollid;
2328 plain_op = false;
2329 }
2330 else if (clause && IsA(clause, RowCompareExpr))
2331 {
2332 return match_rowcompare_to_indexcol(index, indexcol,
2333 opfamily, idxcollation,
2334 (RowCompareExpr *) clause);
2335 }
2336 else if (index->amsearchnulls && IsA(clause, NullTest))
2337 {
2338 NullTest *nt = (NullTest *) clause;
2339
2340 if (!nt->argisrow &&
2341 match_index_to_operand((Node *) nt->arg, indexcol, index))
2342 return true;
2343 return false;
2344 }
2345 else
2346 return false;
2347
2348 /*
2349 * Check for clauses of the form: (indexkey operator constant) or
2350 * (constant operator indexkey). See above notes about const-ness.
2351 */
2352 if (match_index_to_operand(leftop, indexcol, index) &&
2353 !bms_is_member(index_relid, right_relids) &&
2354 !contain_volatile_functions(rightop))
2355 {
2356 if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
2357 is_indexable_operator(expr_op, opfamily, true))
2358 return true;
2359
2360 /*
2361 * If we didn't find a member of the index's opfamily, see whether it
2362 * is a "special" indexable operator.
2363 */
2364 if (plain_op &&
2365 match_special_index_operator(clause, opfamily,
2366 idxcollation, true))
2367 return true;
2368 return false;
2369 }
2370
2371 if (plain_op &&
2372 match_index_to_operand(rightop, indexcol, index) &&
2373 !bms_is_member(index_relid, left_relids) &&
2374 !contain_volatile_functions(leftop))
2375 {
2376 if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
2377 is_indexable_operator(expr_op, opfamily, false))
2378 return true;
2379
2380 /*
2381 * If we didn't find a member of the index's opfamily, see whether it
2382 * is a "special" indexable operator.
2383 */
2384 if (match_special_index_operator(clause, opfamily,
2385 idxcollation, false))
2386 return true;
2387 return false;
2388 }
2389
2390 return false;
2391 }
2392
2393 /*
2394 * is_indexable_operator
2395 * Does the operator match the specified index opfamily?
2396 *
2397 * If the indexkey is on the right, what we actually want to know
2398 * is whether the operator has a commutator operator that matches
2399 * the opfamily.
2400 */
2401 static bool
is_indexable_operator(Oid expr_op,Oid opfamily,bool indexkey_on_left)2402 is_indexable_operator(Oid expr_op, Oid opfamily, bool indexkey_on_left)
2403 {
2404 /* Get the commuted operator if necessary */
2405 if (!indexkey_on_left)
2406 {
2407 expr_op = get_commutator(expr_op);
2408 if (expr_op == InvalidOid)
2409 return false;
2410 }
2411
2412 /* OK if the (commuted) operator is a member of the index's opfamily */
2413 return op_in_opfamily(expr_op, opfamily);
2414 }
2415
2416 /*
2417 * match_rowcompare_to_indexcol()
2418 * Handles the RowCompareExpr case for match_clause_to_indexcol(),
2419 * which see for comments.
2420 */
2421 static bool
match_rowcompare_to_indexcol(IndexOptInfo * index,int indexcol,Oid opfamily,Oid idxcollation,RowCompareExpr * clause)2422 match_rowcompare_to_indexcol(IndexOptInfo *index,
2423 int indexcol,
2424 Oid opfamily,
2425 Oid idxcollation,
2426 RowCompareExpr *clause)
2427 {
2428 Index index_relid = index->rel->relid;
2429 Node *leftop,
2430 *rightop;
2431 Oid expr_op;
2432 Oid expr_coll;
2433
2434 /* Forget it if we're not dealing with a btree index */
2435 if (index->relam != BTREE_AM_OID)
2436 return false;
2437
2438 /*
2439 * We could do the matching on the basis of insisting that the opfamily
2440 * shown in the RowCompareExpr be the same as the index column's opfamily,
2441 * but that could fail in the presence of reverse-sort opfamilies: it'd be
2442 * a matter of chance whether RowCompareExpr had picked the forward or
2443 * reverse-sort family. So look only at the operator, and match if it is
2444 * a member of the index's opfamily (after commutation, if the indexkey is
2445 * on the right). We'll worry later about whether any additional
2446 * operators are matchable to the index.
2447 */
2448 leftop = (Node *) linitial(clause->largs);
2449 rightop = (Node *) linitial(clause->rargs);
2450 expr_op = linitial_oid(clause->opnos);
2451 expr_coll = linitial_oid(clause->inputcollids);
2452
2453 /* Collations must match, if relevant */
2454 if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
2455 return false;
2456
2457 /*
2458 * These syntactic tests are the same as in match_clause_to_indexcol()
2459 */
2460 if (match_index_to_operand(leftop, indexcol, index) &&
2461 !bms_is_member(index_relid, pull_varnos(rightop)) &&
2462 !contain_volatile_functions(rightop))
2463 {
2464 /* OK, indexkey is on left */
2465 }
2466 else if (match_index_to_operand(rightop, indexcol, index) &&
2467 !bms_is_member(index_relid, pull_varnos(leftop)) &&
2468 !contain_volatile_functions(leftop))
2469 {
2470 /* indexkey is on right, so commute the operator */
2471 expr_op = get_commutator(expr_op);
2472 if (expr_op == InvalidOid)
2473 return false;
2474 }
2475 else
2476 return false;
2477
2478 /* We're good if the operator is the right type of opfamily member */
2479 switch (get_op_opfamily_strategy(expr_op, opfamily))
2480 {
2481 case BTLessStrategyNumber:
2482 case BTLessEqualStrategyNumber:
2483 case BTGreaterEqualStrategyNumber:
2484 case BTGreaterStrategyNumber:
2485 return true;
2486 }
2487
2488 return false;
2489 }
2490
2491
2492 /****************************************************************************
2493 * ---- ROUTINES TO CHECK ORDERING OPERATORS ----
2494 ****************************************************************************/
2495
2496 /*
2497 * match_pathkeys_to_index
2498 * Test whether an index can produce output ordered according to the
2499 * given pathkeys using "ordering operators".
2500 *
2501 * If it can, return a list of suitable ORDER BY expressions, each of the form
2502 * "indexedcol operator pseudoconstant", along with an integer list of the
2503 * index column numbers (zero based) that each clause would be used with.
2504 * NIL lists are returned if the ordering is not achievable this way.
2505 *
2506 * On success, the result list is ordered by pathkeys, and in fact is
2507 * one-to-one with the requested pathkeys.
2508 */
2509 static void
match_pathkeys_to_index(IndexOptInfo * index,List * pathkeys,List ** orderby_clauses_p,List ** clause_columns_p)2510 match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
2511 List **orderby_clauses_p,
2512 List **clause_columns_p)
2513 {
2514 List *orderby_clauses = NIL;
2515 List *clause_columns = NIL;
2516 ListCell *lc1;
2517
2518 *orderby_clauses_p = NIL; /* set default results */
2519 *clause_columns_p = NIL;
2520
2521 /* Only indexes with the amcanorderbyop property are interesting here */
2522 if (!index->amcanorderbyop)
2523 return;
2524
2525 foreach(lc1, pathkeys)
2526 {
2527 PathKey *pathkey = (PathKey *) lfirst(lc1);
2528 bool found = false;
2529 ListCell *lc2;
2530
2531 /*
2532 * Note: for any failure to match, we just return NIL immediately.
2533 * There is no value in matching just some of the pathkeys.
2534 */
2535
2536 /* Pathkey must request default sort order for the target opfamily */
2537 if (pathkey->pk_strategy != BTLessStrategyNumber ||
2538 pathkey->pk_nulls_first)
2539 return;
2540
2541 /* If eclass is volatile, no hope of using an indexscan */
2542 if (pathkey->pk_eclass->ec_has_volatile)
2543 return;
2544
2545 /*
2546 * Try to match eclass member expression(s) to index. Note that child
2547 * EC members are considered, but only when they belong to the target
2548 * relation. (Unlike regular members, the same expression could be a
2549 * child member of more than one EC. Therefore, the same index could
2550 * be considered to match more than one pathkey list, which is OK
2551 * here. See also get_eclass_for_sort_expr.)
2552 */
2553 foreach(lc2, pathkey->pk_eclass->ec_members)
2554 {
2555 EquivalenceMember *member = (EquivalenceMember *) lfirst(lc2);
2556 int indexcol;
2557
2558 /* No possibility of match if it references other relations */
2559 if (!bms_equal(member->em_relids, index->rel->relids))
2560 continue;
2561
2562 /*
2563 * We allow any column of the index to match each pathkey; they
2564 * don't have to match left-to-right as you might expect. This is
2565 * correct for GiST, and it doesn't matter for SP-GiST because
2566 * that doesn't handle multiple columns anyway, and no other
2567 * existing AMs support amcanorderbyop. We might need different
2568 * logic in future for other implementations.
2569 */
2570 for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2571 {
2572 Expr *expr;
2573
2574 expr = match_clause_to_ordering_op(index,
2575 indexcol,
2576 member->em_expr,
2577 pathkey->pk_opfamily);
2578 if (expr)
2579 {
2580 orderby_clauses = lappend(orderby_clauses, expr);
2581 clause_columns = lappend_int(clause_columns, indexcol);
2582 found = true;
2583 break;
2584 }
2585 }
2586
2587 if (found) /* don't want to look at remaining members */
2588 break;
2589 }
2590
2591 if (!found) /* fail if no match for this pathkey */
2592 return;
2593 }
2594
2595 *orderby_clauses_p = orderby_clauses; /* success! */
2596 *clause_columns_p = clause_columns;
2597 }
2598
2599 /*
2600 * match_clause_to_ordering_op
2601 * Determines whether an ordering operator expression matches an
2602 * index column.
2603 *
2604 * This is similar to, but simpler than, match_clause_to_indexcol.
2605 * We only care about simple OpExpr cases. The input is a bare
2606 * expression that is being ordered by, which must be of the form
2607 * (indexkey op const) or (const op indexkey) where op is an ordering
2608 * operator for the column's opfamily.
2609 *
2610 * 'index' is the index of interest.
2611 * 'indexcol' is a column number of 'index' (counting from 0).
2612 * 'clause' is the ordering expression to be tested.
2613 * 'pk_opfamily' is the btree opfamily describing the required sort order.
2614 *
2615 * Note that we currently do not consider the collation of the ordering
2616 * operator's result. In practical cases the result type will be numeric
2617 * and thus have no collation, and it's not very clear what to match to
2618 * if it did have a collation. The index's collation should match the
2619 * ordering operator's input collation, not its result.
2620 *
2621 * If successful, return 'clause' as-is if the indexkey is on the left,
2622 * otherwise a commuted copy of 'clause'. If no match, return NULL.
2623 */
2624 static Expr *
match_clause_to_ordering_op(IndexOptInfo * index,int indexcol,Expr * clause,Oid pk_opfamily)2625 match_clause_to_ordering_op(IndexOptInfo *index,
2626 int indexcol,
2627 Expr *clause,
2628 Oid pk_opfamily)
2629 {
2630 Oid opfamily;
2631 Oid idxcollation;
2632 Node *leftop,
2633 *rightop;
2634 Oid expr_op;
2635 Oid expr_coll;
2636 Oid sortfamily;
2637 bool commuted;
2638
2639 Assert(indexcol < index->nkeycolumns);
2640
2641 opfamily = index->opfamily[indexcol];
2642 idxcollation = index->indexcollations[indexcol];
2643
2644 /*
2645 * Clause must be a binary opclause.
2646 */
2647 if (!is_opclause(clause))
2648 return NULL;
2649 leftop = get_leftop(clause);
2650 rightop = get_rightop(clause);
2651 if (!leftop || !rightop)
2652 return NULL;
2653 expr_op = ((OpExpr *) clause)->opno;
2654 expr_coll = ((OpExpr *) clause)->inputcollid;
2655
2656 /*
2657 * We can forget the whole thing right away if wrong collation.
2658 */
2659 if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
2660 return NULL;
2661
2662 /*
2663 * Check for clauses of the form: (indexkey operator constant) or
2664 * (constant operator indexkey).
2665 */
2666 if (match_index_to_operand(leftop, indexcol, index) &&
2667 !contain_var_clause(rightop) &&
2668 !contain_volatile_functions(rightop))
2669 {
2670 commuted = false;
2671 }
2672 else if (match_index_to_operand(rightop, indexcol, index) &&
2673 !contain_var_clause(leftop) &&
2674 !contain_volatile_functions(leftop))
2675 {
2676 /* Might match, but we need a commuted operator */
2677 expr_op = get_commutator(expr_op);
2678 if (expr_op == InvalidOid)
2679 return NULL;
2680 commuted = true;
2681 }
2682 else
2683 return NULL;
2684
2685 /*
2686 * Is the (commuted) operator an ordering operator for the opfamily? And
2687 * if so, does it yield the right sorting semantics?
2688 */
2689 sortfamily = get_op_opfamily_sortfamily(expr_op, opfamily);
2690 if (sortfamily != pk_opfamily)
2691 return NULL;
2692
2693 /* We have a match. Return clause or a commuted version thereof. */
2694 if (commuted)
2695 {
2696 OpExpr *newclause = makeNode(OpExpr);
2697
2698 /* flat-copy all the fields of clause */
2699 memcpy(newclause, clause, sizeof(OpExpr));
2700
2701 /* commute it */
2702 newclause->opno = expr_op;
2703 newclause->opfuncid = InvalidOid;
2704 newclause->args = list_make2(rightop, leftop);
2705
2706 clause = (Expr *) newclause;
2707 }
2708
2709 return clause;
2710 }
2711
2712
2713 /****************************************************************************
2714 * ---- ROUTINES TO DO PARTIAL INDEX PREDICATE TESTS ----
2715 ****************************************************************************/
2716
2717 /*
2718 * check_index_predicates
2719 * Set the predicate-derived IndexOptInfo fields for each index
2720 * of the specified relation.
2721 *
2722 * predOK is set true if the index is partial and its predicate is satisfied
2723 * for this query, ie the query's WHERE clauses imply the predicate.
2724 *
2725 * indrestrictinfo is set to the relation's baserestrictinfo list less any
2726 * conditions that are implied by the index's predicate. (Obviously, for a
2727 * non-partial index, this is the same as baserestrictinfo.) Such conditions
2728 * can be dropped from the plan when using the index, in certain cases.
2729 *
2730 * At one time it was possible for this to get re-run after adding more
2731 * restrictions to the rel, thus possibly letting us prove more indexes OK.
2732 * That doesn't happen any more (at least not in the core code's usage),
2733 * but this code still supports it in case extensions want to mess with the
2734 * baserestrictinfo list. We assume that adding more restrictions can't make
2735 * an index not predOK. We must recompute indrestrictinfo each time, though,
2736 * to make sure any newly-added restrictions get into it if needed.
2737 */
2738 void
check_index_predicates(PlannerInfo * root,RelOptInfo * rel)2739 check_index_predicates(PlannerInfo *root, RelOptInfo *rel)
2740 {
2741 List *clauselist;
2742 bool have_partial;
2743 bool is_target_rel;
2744 Relids otherrels;
2745 ListCell *lc;
2746
2747 /* Indexes are available only on base or "other" member relations. */
2748 Assert(IS_SIMPLE_REL(rel));
2749
2750 /*
2751 * Initialize the indrestrictinfo lists to be identical to
2752 * baserestrictinfo, and check whether there are any partial indexes. If
2753 * not, this is all we need to do.
2754 */
2755 have_partial = false;
2756 foreach(lc, rel->indexlist)
2757 {
2758 IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
2759
2760 index->indrestrictinfo = rel->baserestrictinfo;
2761 if (index->indpred)
2762 have_partial = true;
2763 }
2764 if (!have_partial)
2765 return;
2766
2767 /*
2768 * Construct a list of clauses that we can assume true for the purpose of
2769 * proving the index(es) usable. Restriction clauses for the rel are
2770 * always usable, and so are any join clauses that are "movable to" this
2771 * rel. Also, we can consider any EC-derivable join clauses (which must
2772 * be "movable to" this rel, by definition).
2773 */
2774 clauselist = list_copy(rel->baserestrictinfo);
2775
2776 /* Scan the rel's join clauses */
2777 foreach(lc, rel->joininfo)
2778 {
2779 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2780
2781 /* Check if clause can be moved to this rel */
2782 if (!join_clause_is_movable_to(rinfo, rel))
2783 continue;
2784
2785 clauselist = lappend(clauselist, rinfo);
2786 }
2787
2788 /*
2789 * Add on any equivalence-derivable join clauses. Computing the correct
2790 * relid sets for generate_join_implied_equalities is slightly tricky
2791 * because the rel could be a child rel rather than a true baserel, and in
2792 * that case we must remove its parents' relid(s) from all_baserels.
2793 */
2794 if (rel->reloptkind == RELOPT_OTHER_MEMBER_REL)
2795 otherrels = bms_difference(root->all_baserels,
2796 find_childrel_parents(root, rel));
2797 else
2798 otherrels = bms_difference(root->all_baserels, rel->relids);
2799
2800 if (!bms_is_empty(otherrels))
2801 clauselist =
2802 list_concat(clauselist,
2803 generate_join_implied_equalities(root,
2804 bms_union(rel->relids,
2805 otherrels),
2806 otherrels,
2807 rel));
2808
2809 /*
2810 * Normally we remove quals that are implied by a partial index's
2811 * predicate from indrestrictinfo, indicating that they need not be
2812 * checked explicitly by an indexscan plan using this index. However, if
2813 * the rel is a target relation of UPDATE/DELETE/SELECT FOR UPDATE, we
2814 * cannot remove such quals from the plan, because they need to be in the
2815 * plan so that they will be properly rechecked by EvalPlanQual testing.
2816 * Some day we might want to remove such quals from the main plan anyway
2817 * and pass them through to EvalPlanQual via a side channel; but for now,
2818 * we just don't remove implied quals at all for target relations.
2819 */
2820 is_target_rel = (rel->relid == root->parse->resultRelation ||
2821 get_plan_rowmark(root->rowMarks, rel->relid) != NULL);
2822
2823 /*
2824 * Now try to prove each index predicate true, and compute the
2825 * indrestrictinfo lists for partial indexes. Note that we compute the
2826 * indrestrictinfo list even for non-predOK indexes; this might seem
2827 * wasteful, but we may be able to use such indexes in OR clauses, cf
2828 * generate_bitmap_or_paths().
2829 */
2830 foreach(lc, rel->indexlist)
2831 {
2832 IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
2833 ListCell *lcr;
2834
2835 if (index->indpred == NIL)
2836 continue; /* ignore non-partial indexes here */
2837
2838 if (!index->predOK) /* don't repeat work if already proven OK */
2839 index->predOK = predicate_implied_by(index->indpred, clauselist,
2840 false);
2841
2842 /* If rel is an update target, leave indrestrictinfo as set above */
2843 if (is_target_rel)
2844 continue;
2845
2846 /* Else compute indrestrictinfo as the non-implied quals */
2847 index->indrestrictinfo = NIL;
2848 foreach(lcr, rel->baserestrictinfo)
2849 {
2850 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lcr);
2851
2852 /* predicate_implied_by() assumes first arg is immutable */
2853 if (contain_mutable_functions((Node *) rinfo->clause) ||
2854 !predicate_implied_by(list_make1(rinfo->clause),
2855 index->indpred, false))
2856 index->indrestrictinfo = lappend(index->indrestrictinfo, rinfo);
2857 }
2858 }
2859 }
2860
2861 /****************************************************************************
2862 * ---- ROUTINES TO CHECK EXTERNALLY-VISIBLE CONDITIONS ----
2863 ****************************************************************************/
2864
2865 /*
2866 * ec_member_matches_indexcol
2867 * Test whether an EquivalenceClass member matches an index column.
2868 *
2869 * This is a callback for use by generate_implied_equalities_for_column.
2870 */
2871 static bool
ec_member_matches_indexcol(PlannerInfo * root,RelOptInfo * rel,EquivalenceClass * ec,EquivalenceMember * em,void * arg)2872 ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
2873 EquivalenceClass *ec, EquivalenceMember *em,
2874 void *arg)
2875 {
2876 IndexOptInfo *index = ((ec_member_matches_arg *) arg)->index;
2877 int indexcol = ((ec_member_matches_arg *) arg)->indexcol;
2878 Oid curFamily;
2879 Oid curCollation;
2880
2881 Assert(indexcol < index->nkeycolumns);
2882
2883 curFamily = index->opfamily[indexcol];
2884 curCollation = index->indexcollations[indexcol];
2885
2886 /*
2887 * If it's a btree index, we can reject it if its opfamily isn't
2888 * compatible with the EC, since no clause generated from the EC could be
2889 * used with the index. For non-btree indexes, we can't easily tell
2890 * whether clauses generated from the EC could be used with the index, so
2891 * don't check the opfamily. This might mean we return "true" for a
2892 * useless EC, so we have to recheck the results of
2893 * generate_implied_equalities_for_column; see
2894 * match_eclass_clauses_to_index.
2895 */
2896 if (index->relam == BTREE_AM_OID &&
2897 !list_member_oid(ec->ec_opfamilies, curFamily))
2898 return false;
2899
2900 /* We insist on collation match for all index types, though */
2901 if (!IndexCollMatchesExprColl(curCollation, ec->ec_collation))
2902 return false;
2903
2904 return match_index_to_operand((Node *) em->em_expr, indexcol, index);
2905 }
2906
2907 /*
2908 * relation_has_unique_index_for
2909 * Determine whether the relation provably has at most one row satisfying
2910 * a set of equality conditions, because the conditions constrain all
2911 * columns of some unique index.
2912 *
2913 * The conditions can be represented in either or both of two ways:
2914 * 1. A list of RestrictInfo nodes, where the caller has already determined
2915 * that each condition is a mergejoinable equality with an expression in
2916 * this relation on one side, and an expression not involving this relation
2917 * on the other. The transient outer_is_left flag is used to identify which
2918 * side we should look at: left side if outer_is_left is false, right side
2919 * if it is true.
2920 * 2. A list of expressions in this relation, and a corresponding list of
2921 * equality operators. The caller must have already checked that the operators
2922 * represent equality. (Note: the operators could be cross-type; the
2923 * expressions should correspond to their RHS inputs.)
2924 *
2925 * The caller need only supply equality conditions arising from joins;
2926 * this routine automatically adds in any usable baserestrictinfo clauses.
2927 * (Note that the passed-in restrictlist will be destructively modified!)
2928 */
2929 bool
relation_has_unique_index_for(PlannerInfo * root,RelOptInfo * rel,List * restrictlist,List * exprlist,List * oprlist)2930 relation_has_unique_index_for(PlannerInfo *root, RelOptInfo *rel,
2931 List *restrictlist,
2932 List *exprlist, List *oprlist)
2933 {
2934 ListCell *ic;
2935
2936 Assert(list_length(exprlist) == list_length(oprlist));
2937
2938 /* Short-circuit if no indexes... */
2939 if (rel->indexlist == NIL)
2940 return false;
2941
2942 /*
2943 * Examine the rel's restriction clauses for usable var = const clauses
2944 * that we can add to the restrictlist.
2945 */
2946 foreach(ic, rel->baserestrictinfo)
2947 {
2948 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(ic);
2949
2950 /*
2951 * Note: can_join won't be set for a restriction clause, but
2952 * mergeopfamilies will be if it has a mergejoinable operator and
2953 * doesn't contain volatile functions.
2954 */
2955 if (restrictinfo->mergeopfamilies == NIL)
2956 continue; /* not mergejoinable */
2957
2958 /*
2959 * The clause certainly doesn't refer to anything but the given rel.
2960 * If either side is pseudoconstant then we can use it.
2961 */
2962 if (bms_is_empty(restrictinfo->left_relids))
2963 {
2964 /* righthand side is inner */
2965 restrictinfo->outer_is_left = true;
2966 }
2967 else if (bms_is_empty(restrictinfo->right_relids))
2968 {
2969 /* lefthand side is inner */
2970 restrictinfo->outer_is_left = false;
2971 }
2972 else
2973 continue;
2974
2975 /* OK, add to list */
2976 restrictlist = lappend(restrictlist, restrictinfo);
2977 }
2978
2979 /* Short-circuit the easy case */
2980 if (restrictlist == NIL && exprlist == NIL)
2981 return false;
2982
2983 /* Examine each index of the relation ... */
2984 foreach(ic, rel->indexlist)
2985 {
2986 IndexOptInfo *ind = (IndexOptInfo *) lfirst(ic);
2987 int c;
2988
2989 /*
2990 * If the index is not unique, or not immediately enforced, or if it's
2991 * a partial index that doesn't match the query, it's useless here.
2992 */
2993 if (!ind->unique || !ind->immediate ||
2994 (ind->indpred != NIL && !ind->predOK))
2995 continue;
2996
2997 /*
2998 * Try to find each index column in the lists of conditions. This is
2999 * O(N^2) or worse, but we expect all the lists to be short.
3000 */
3001 for (c = 0; c < ind->nkeycolumns; c++)
3002 {
3003 bool matched = false;
3004 ListCell *lc;
3005 ListCell *lc2;
3006
3007 foreach(lc, restrictlist)
3008 {
3009 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
3010 Node *rexpr;
3011
3012 /*
3013 * The condition's equality operator must be a member of the
3014 * index opfamily, else it is not asserting the right kind of
3015 * equality behavior for this index. We check this first
3016 * since it's probably cheaper than match_index_to_operand().
3017 */
3018 if (!list_member_oid(rinfo->mergeopfamilies, ind->opfamily[c]))
3019 continue;
3020
3021 /*
3022 * XXX at some point we may need to check collations here too.
3023 * For the moment we assume all collations reduce to the same
3024 * notion of equality.
3025 */
3026
3027 /* OK, see if the condition operand matches the index key */
3028 if (rinfo->outer_is_left)
3029 rexpr = get_rightop(rinfo->clause);
3030 else
3031 rexpr = get_leftop(rinfo->clause);
3032
3033 if (match_index_to_operand(rexpr, c, ind))
3034 {
3035 matched = true; /* column is unique */
3036 break;
3037 }
3038 }
3039
3040 if (matched)
3041 continue;
3042
3043 forboth(lc, exprlist, lc2, oprlist)
3044 {
3045 Node *expr = (Node *) lfirst(lc);
3046 Oid opr = lfirst_oid(lc2);
3047
3048 /* See if the expression matches the index key */
3049 if (!match_index_to_operand(expr, c, ind))
3050 continue;
3051
3052 /*
3053 * The equality operator must be a member of the index
3054 * opfamily, else it is not asserting the right kind of
3055 * equality behavior for this index. We assume the caller
3056 * determined it is an equality operator, so we don't need to
3057 * check any more tightly than this.
3058 */
3059 if (!op_in_opfamily(opr, ind->opfamily[c]))
3060 continue;
3061
3062 /*
3063 * XXX at some point we may need to check collations here too.
3064 * For the moment we assume all collations reduce to the same
3065 * notion of equality.
3066 */
3067
3068 matched = true; /* column is unique */
3069 break;
3070 }
3071
3072 if (!matched)
3073 break; /* no match; this index doesn't help us */
3074 }
3075
3076 /* Matched all key columns of this index? */
3077 if (c == ind->nkeycolumns)
3078 return true;
3079 }
3080
3081 return false;
3082 }
3083
3084 /*
3085 * indexcol_is_bool_constant_for_query
3086 *
3087 * If an index column is constrained to have a constant value by the query's
3088 * WHERE conditions, then it's irrelevant for sort-order considerations.
3089 * Usually that means we have a restriction clause WHERE indexcol = constant,
3090 * which gets turned into an EquivalenceClass containing a constant, which
3091 * is recognized as redundant by build_index_pathkeys(). But if the index
3092 * column is a boolean variable (or expression), then we are not going to
3093 * see WHERE indexcol = constant, because expression preprocessing will have
3094 * simplified that to "WHERE indexcol" or "WHERE NOT indexcol". So we are not
3095 * going to have a matching EquivalenceClass (unless the query also contains
3096 * "ORDER BY indexcol"). To allow such cases to work the same as they would
3097 * for non-boolean values, this function is provided to detect whether the
3098 * specified index column matches a boolean restriction clause.
3099 */
3100 bool
indexcol_is_bool_constant_for_query(IndexOptInfo * index,int indexcol)3101 indexcol_is_bool_constant_for_query(IndexOptInfo *index, int indexcol)
3102 {
3103 ListCell *lc;
3104
3105 /* If the index isn't boolean, we can't possibly get a match */
3106 if (!IsBooleanOpfamily(index->opfamily[indexcol]))
3107 return false;
3108
3109 /* Check each restriction clause for the index's rel */
3110 foreach(lc, index->rel->baserestrictinfo)
3111 {
3112 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
3113
3114 /*
3115 * As in match_clause_to_indexcol, never match pseudoconstants to
3116 * indexes. (It might be semantically okay to do so here, but the
3117 * odds of getting a match are negligible, so don't waste the cycles.)
3118 */
3119 if (rinfo->pseudoconstant)
3120 continue;
3121
3122 /* See if we can match the clause's expression to the index column */
3123 if (match_boolean_index_clause((Node *) rinfo->clause, indexcol, index))
3124 return true;
3125 }
3126
3127 return false;
3128 }
3129
3130
3131 /****************************************************************************
3132 * ---- ROUTINES TO CHECK OPERANDS ----
3133 ****************************************************************************/
3134
3135 /*
3136 * match_index_to_operand()
3137 * Generalized test for a match between an index's key
3138 * and the operand on one side of a restriction or join clause.
3139 *
3140 * operand: the nodetree to be compared to the index
3141 * indexcol: the column number of the index (counting from 0)
3142 * index: the index of interest
3143 *
3144 * Note that we aren't interested in collations here; the caller must check
3145 * for a collation match, if it's dealing with an operator where that matters.
3146 *
3147 * This is exported for use in selfuncs.c.
3148 */
3149 bool
match_index_to_operand(Node * operand,int indexcol,IndexOptInfo * index)3150 match_index_to_operand(Node *operand,
3151 int indexcol,
3152 IndexOptInfo *index)
3153 {
3154 int indkey;
3155
3156 /*
3157 * Ignore any RelabelType node above the operand. This is needed to be
3158 * able to apply indexscanning in binary-compatible-operator cases. Note:
3159 * we can assume there is at most one RelabelType node;
3160 * eval_const_expressions() will have simplified if more than one.
3161 */
3162 if (operand && IsA(operand, RelabelType))
3163 operand = (Node *) ((RelabelType *) operand)->arg;
3164
3165 indkey = index->indexkeys[indexcol];
3166 if (indkey != 0)
3167 {
3168 /*
3169 * Simple index column; operand must be a matching Var.
3170 */
3171 if (operand && IsA(operand, Var) &&
3172 index->rel->relid == ((Var *) operand)->varno &&
3173 indkey == ((Var *) operand)->varattno)
3174 return true;
3175 }
3176 else
3177 {
3178 /*
3179 * Index expression; find the correct expression. (This search could
3180 * be avoided, at the cost of complicating all the callers of this
3181 * routine; doesn't seem worth it.)
3182 */
3183 ListCell *indexpr_item;
3184 int i;
3185 Node *indexkey;
3186
3187 indexpr_item = list_head(index->indexprs);
3188 for (i = 0; i < indexcol; i++)
3189 {
3190 if (index->indexkeys[i] == 0)
3191 {
3192 if (indexpr_item == NULL)
3193 elog(ERROR, "wrong number of index expressions");
3194 indexpr_item = lnext(indexpr_item);
3195 }
3196 }
3197 if (indexpr_item == NULL)
3198 elog(ERROR, "wrong number of index expressions");
3199 indexkey = (Node *) lfirst(indexpr_item);
3200
3201 /*
3202 * Does it match the operand? Again, strip any relabeling.
3203 */
3204 if (indexkey && IsA(indexkey, RelabelType))
3205 indexkey = (Node *) ((RelabelType *) indexkey)->arg;
3206
3207 if (equal(indexkey, operand))
3208 return true;
3209 }
3210
3211 return false;
3212 }
3213
3214 /****************************************************************************
3215 * ---- ROUTINES FOR "SPECIAL" INDEXABLE OPERATORS ----
3216 ****************************************************************************/
3217
3218 /*
3219 * These routines handle special optimization of operators that can be
3220 * used with index scans even though they are not known to the executor's
3221 * indexscan machinery. The key idea is that these operators allow us
3222 * to derive approximate indexscan qual clauses, such that any tuples
3223 * that pass the operator clause itself must also satisfy the simpler
3224 * indexscan condition(s). Then we can use the indexscan machinery
3225 * to avoid scanning as much of the table as we'd otherwise have to,
3226 * while applying the original operator as a qpqual condition to ensure
3227 * we deliver only the tuples we want. (In essence, we're using a regular
3228 * index as if it were a lossy index.)
3229 *
3230 * An example of what we're doing is
3231 * textfield LIKE 'abc%'
3232 * from which we can generate the indexscanable conditions
3233 * textfield >= 'abc' AND textfield < 'abd'
3234 * which allow efficient scanning of an index on textfield.
3235 * (In reality, character set and collation issues make the transformation
3236 * from LIKE to indexscan limits rather harder than one might think ...
3237 * but that's the basic idea.)
3238 *
3239 * Another thing that we do with this machinery is to provide special
3240 * smarts for "boolean" indexes (that is, indexes on boolean columns
3241 * that support boolean equality). We can transform a plain reference
3242 * to the indexkey into "indexkey = true", or "NOT indexkey" into
3243 * "indexkey = false", so as to make the expression indexable using the
3244 * regular index operators. (As of Postgres 8.1, we must do this here
3245 * because constant simplification does the reverse transformation;
3246 * without this code there'd be no way to use such an index at all.)
3247 *
3248 * Three routines are provided here:
3249 *
3250 * match_special_index_operator() is just an auxiliary function for
3251 * match_clause_to_indexcol(); after the latter fails to recognize a
3252 * restriction opclause's operator as a member of an index's opfamily,
3253 * it asks match_special_index_operator() whether the clause should be
3254 * considered an indexqual anyway.
3255 *
3256 * match_boolean_index_clause() similarly detects clauses that can be
3257 * converted into boolean equality operators.
3258 *
3259 * expand_indexqual_conditions() converts a list of RestrictInfo nodes
3260 * (with implicit AND semantics across list elements) into a list of clauses
3261 * that the executor can actually handle. For operators that are members of
3262 * the index's opfamily this transformation is a no-op, but clauses recognized
3263 * by match_special_index_operator() or match_boolean_index_clause() must be
3264 * converted into one or more "regular" indexqual conditions.
3265 */
3266
3267 /*
3268 * match_boolean_index_clause
3269 * Recognize restriction clauses that can be matched to a boolean index.
3270 *
3271 * This should be called only when IsBooleanOpfamily() recognizes the
3272 * index's operator family. We check to see if the clause matches the
3273 * index's key.
3274 */
3275 static bool
match_boolean_index_clause(Node * clause,int indexcol,IndexOptInfo * index)3276 match_boolean_index_clause(Node *clause,
3277 int indexcol,
3278 IndexOptInfo *index)
3279 {
3280 /* Direct match? */
3281 if (match_index_to_operand(clause, indexcol, index))
3282 return true;
3283 /* NOT clause? */
3284 if (not_clause(clause))
3285 {
3286 if (match_index_to_operand((Node *) get_notclausearg((Expr *) clause),
3287 indexcol, index))
3288 return true;
3289 }
3290
3291 /*
3292 * Since we only consider clauses at top level of WHERE, we can convert
3293 * indexkey IS TRUE and indexkey IS FALSE to index searches as well. The
3294 * different meaning for NULL isn't important.
3295 */
3296 else if (clause && IsA(clause, BooleanTest))
3297 {
3298 BooleanTest *btest = (BooleanTest *) clause;
3299
3300 if (btest->booltesttype == IS_TRUE ||
3301 btest->booltesttype == IS_FALSE)
3302 if (match_index_to_operand((Node *) btest->arg,
3303 indexcol, index))
3304 return true;
3305 }
3306 return false;
3307 }
3308
3309 /*
3310 * match_special_index_operator
3311 * Recognize restriction clauses that can be used to generate
3312 * additional indexscanable qualifications.
3313 *
3314 * The given clause is already known to be a binary opclause having
3315 * the form (indexkey OP pseudoconst) or (pseudoconst OP indexkey),
3316 * but the OP proved not to be one of the index's opfamily operators.
3317 * Return 'true' if we can do something with it anyway.
3318 */
3319 static bool
match_special_index_operator(Expr * clause,Oid opfamily,Oid idxcollation,bool indexkey_on_left)3320 match_special_index_operator(Expr *clause, Oid opfamily, Oid idxcollation,
3321 bool indexkey_on_left)
3322 {
3323 bool isIndexable = false;
3324 Node *rightop;
3325 Oid expr_op;
3326 Oid expr_coll;
3327 Const *patt;
3328 Const *prefix = NULL;
3329 Pattern_Prefix_Status pstatus = Pattern_Prefix_None;
3330
3331 /*
3332 * Currently, all known special operators require the indexkey on the
3333 * left, but this test could be pushed into the switch statement if some
3334 * are added that do not...
3335 */
3336 if (!indexkey_on_left)
3337 return false;
3338
3339 /* we know these will succeed */
3340 rightop = get_rightop(clause);
3341 expr_op = ((OpExpr *) clause)->opno;
3342 expr_coll = ((OpExpr *) clause)->inputcollid;
3343
3344 /* again, required for all current special ops: */
3345 if (!IsA(rightop, Const) ||
3346 ((Const *) rightop)->constisnull)
3347 return false;
3348 patt = (Const *) rightop;
3349
3350 switch (expr_op)
3351 {
3352 case OID_TEXT_LIKE_OP:
3353 case OID_BPCHAR_LIKE_OP:
3354 case OID_NAME_LIKE_OP:
3355 /* the right-hand const is type text for all of these */
3356 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Like, expr_coll,
3357 &prefix, NULL);
3358 isIndexable = (pstatus != Pattern_Prefix_None);
3359 break;
3360
3361 case OID_BYTEA_LIKE_OP:
3362 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Like, expr_coll,
3363 &prefix, NULL);
3364 isIndexable = (pstatus != Pattern_Prefix_None);
3365 break;
3366
3367 case OID_TEXT_ICLIKE_OP:
3368 case OID_BPCHAR_ICLIKE_OP:
3369 case OID_NAME_ICLIKE_OP:
3370 /* the right-hand const is type text for all of these */
3371 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Like_IC, expr_coll,
3372 &prefix, NULL);
3373 isIndexable = (pstatus != Pattern_Prefix_None);
3374 break;
3375
3376 case OID_TEXT_REGEXEQ_OP:
3377 case OID_BPCHAR_REGEXEQ_OP:
3378 case OID_NAME_REGEXEQ_OP:
3379 /* the right-hand const is type text for all of these */
3380 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Regex, expr_coll,
3381 &prefix, NULL);
3382 isIndexable = (pstatus != Pattern_Prefix_None);
3383 break;
3384
3385 case OID_TEXT_ICREGEXEQ_OP:
3386 case OID_BPCHAR_ICREGEXEQ_OP:
3387 case OID_NAME_ICREGEXEQ_OP:
3388 /* the right-hand const is type text for all of these */
3389 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Regex_IC, expr_coll,
3390 &prefix, NULL);
3391 isIndexable = (pstatus != Pattern_Prefix_None);
3392 break;
3393
3394 case OID_INET_SUB_OP:
3395 case OID_INET_SUBEQ_OP:
3396 isIndexable = true;
3397 break;
3398 }
3399
3400 if (prefix)
3401 {
3402 pfree(DatumGetPointer(prefix->constvalue));
3403 pfree(prefix);
3404 }
3405
3406 /* done if the expression doesn't look indexable */
3407 if (!isIndexable)
3408 return false;
3409
3410 /*
3411 * Must also check that index's opfamily supports the operators we will
3412 * want to apply. (A hash index, for example, will not support ">=".)
3413 * Currently, only btree and spgist support the operators we need.
3414 *
3415 * Note: actually, in the Pattern_Prefix_Exact case, we only need "=" so a
3416 * hash index would work. Currently it doesn't seem worth checking for
3417 * that, however.
3418 *
3419 * We insist on the opfamily being the specific one we expect, else we'd
3420 * do the wrong thing if someone were to make a reverse-sort opfamily with
3421 * the same operators.
3422 *
3423 * The non-pattern opclasses will not sort the way we need in most non-C
3424 * locales. We can use such an index anyway for an exact match (simple
3425 * equality), but not for prefix-match cases. Note that here we are
3426 * looking at the index's collation, not the expression's collation --
3427 * this test is *not* dependent on the LIKE/regex operator's collation.
3428 */
3429 switch (expr_op)
3430 {
3431 case OID_TEXT_LIKE_OP:
3432 case OID_TEXT_ICLIKE_OP:
3433 case OID_TEXT_REGEXEQ_OP:
3434 case OID_TEXT_ICREGEXEQ_OP:
3435 isIndexable =
3436 (opfamily == TEXT_PATTERN_BTREE_FAM_OID) ||
3437 (opfamily == TEXT_SPGIST_FAM_OID) ||
3438 (opfamily == TEXT_BTREE_FAM_OID &&
3439 (pstatus == Pattern_Prefix_Exact ||
3440 lc_collate_is_c(idxcollation)));
3441 break;
3442
3443 case OID_BPCHAR_LIKE_OP:
3444 case OID_BPCHAR_ICLIKE_OP:
3445 case OID_BPCHAR_REGEXEQ_OP:
3446 case OID_BPCHAR_ICREGEXEQ_OP:
3447 isIndexable =
3448 (opfamily == BPCHAR_PATTERN_BTREE_FAM_OID) ||
3449 (opfamily == BPCHAR_BTREE_FAM_OID &&
3450 (pstatus == Pattern_Prefix_Exact ||
3451 lc_collate_is_c(idxcollation)));
3452 break;
3453
3454 case OID_NAME_LIKE_OP:
3455 case OID_NAME_ICLIKE_OP:
3456 case OID_NAME_REGEXEQ_OP:
3457 case OID_NAME_ICREGEXEQ_OP:
3458 /* name uses locale-insensitive sorting */
3459 isIndexable = (opfamily == NAME_BTREE_FAM_OID);
3460 break;
3461
3462 case OID_BYTEA_LIKE_OP:
3463 isIndexable = (opfamily == BYTEA_BTREE_FAM_OID);
3464 break;
3465
3466 case OID_INET_SUB_OP:
3467 case OID_INET_SUBEQ_OP:
3468 isIndexable = (opfamily == NETWORK_BTREE_FAM_OID);
3469 break;
3470 }
3471
3472 return isIndexable;
3473 }
3474
3475 /*
3476 * expand_indexqual_conditions
3477 * Given a list of RestrictInfo nodes, produce a list of directly usable
3478 * index qual clauses.
3479 *
3480 * Standard qual clauses (those in the index's opfamily) are passed through
3481 * unchanged. Boolean clauses and "special" index operators are expanded
3482 * into clauses that the indexscan machinery will know what to do with.
3483 * RowCompare clauses are simplified if necessary to create a clause that is
3484 * fully checkable by the index.
3485 *
3486 * In addition to the expressions themselves, there are auxiliary lists
3487 * of the index column numbers that the clauses are meant to be used with;
3488 * we generate an updated column number list for the result. (This is not
3489 * the identical list because one input clause sometimes produces more than
3490 * one output clause.)
3491 *
3492 * The input clauses are sorted by column number, and so the output is too.
3493 * (This is depended on in various places in both planner and executor.)
3494 */
3495 void
expand_indexqual_conditions(IndexOptInfo * index,List * indexclauses,List * indexclausecols,List ** indexquals_p,List ** indexqualcols_p)3496 expand_indexqual_conditions(IndexOptInfo *index,
3497 List *indexclauses, List *indexclausecols,
3498 List **indexquals_p, List **indexqualcols_p)
3499 {
3500 List *indexquals = NIL;
3501 List *indexqualcols = NIL;
3502 ListCell *lcc,
3503 *lci;
3504
3505 forboth(lcc, indexclauses, lci, indexclausecols)
3506 {
3507 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lcc);
3508 int indexcol = lfirst_int(lci);
3509 Expr *clause = rinfo->clause;
3510 Oid curFamily;
3511 Oid curCollation;
3512
3513 Assert(indexcol < index->nkeycolumns);
3514
3515 curFamily = index->opfamily[indexcol];
3516 curCollation = index->indexcollations[indexcol];
3517
3518 /* First check for boolean cases */
3519 if (IsBooleanOpfamily(curFamily))
3520 {
3521 Expr *boolqual;
3522
3523 boolqual = expand_boolean_index_clause((Node *) clause,
3524 indexcol,
3525 index);
3526 if (boolqual)
3527 {
3528 indexquals = lappend(indexquals,
3529 make_simple_restrictinfo(boolqual));
3530 indexqualcols = lappend_int(indexqualcols, indexcol);
3531 continue;
3532 }
3533 }
3534
3535 /*
3536 * Else it must be an opclause (usual case), ScalarArrayOp,
3537 * RowCompare, or NullTest
3538 */
3539 if (is_opclause(clause))
3540 {
3541 indexquals = list_concat(indexquals,
3542 expand_indexqual_opclause(rinfo,
3543 curFamily,
3544 curCollation));
3545 /* expand_indexqual_opclause can produce multiple clauses */
3546 while (list_length(indexqualcols) < list_length(indexquals))
3547 indexqualcols = lappend_int(indexqualcols, indexcol);
3548 }
3549 else if (IsA(clause, ScalarArrayOpExpr))
3550 {
3551 /* no extra work at this time */
3552 indexquals = lappend(indexquals, rinfo);
3553 indexqualcols = lappend_int(indexqualcols, indexcol);
3554 }
3555 else if (IsA(clause, RowCompareExpr))
3556 {
3557 indexquals = lappend(indexquals,
3558 expand_indexqual_rowcompare(rinfo,
3559 index,
3560 indexcol));
3561 indexqualcols = lappend_int(indexqualcols, indexcol);
3562 }
3563 else if (IsA(clause, NullTest))
3564 {
3565 Assert(index->amsearchnulls);
3566 indexquals = lappend(indexquals, rinfo);
3567 indexqualcols = lappend_int(indexqualcols, indexcol);
3568 }
3569 else
3570 elog(ERROR, "unsupported indexqual type: %d",
3571 (int) nodeTag(clause));
3572 }
3573
3574 *indexquals_p = indexquals;
3575 *indexqualcols_p = indexqualcols;
3576 }
3577
3578 /*
3579 * expand_boolean_index_clause
3580 * Convert a clause recognized by match_boolean_index_clause into
3581 * a boolean equality operator clause.
3582 *
3583 * Returns NULL if the clause isn't a boolean index qual.
3584 */
3585 static Expr *
expand_boolean_index_clause(Node * clause,int indexcol,IndexOptInfo * index)3586 expand_boolean_index_clause(Node *clause,
3587 int indexcol,
3588 IndexOptInfo *index)
3589 {
3590 /* Direct match? */
3591 if (match_index_to_operand(clause, indexcol, index))
3592 {
3593 /* convert to indexkey = TRUE */
3594 return make_opclause(BooleanEqualOperator, BOOLOID, false,
3595 (Expr *) clause,
3596 (Expr *) makeBoolConst(true, false),
3597 InvalidOid, InvalidOid);
3598 }
3599 /* NOT clause? */
3600 if (not_clause(clause))
3601 {
3602 Node *arg = (Node *) get_notclausearg((Expr *) clause);
3603
3604 /* It must have matched the indexkey */
3605 Assert(match_index_to_operand(arg, indexcol, index));
3606 /* convert to indexkey = FALSE */
3607 return make_opclause(BooleanEqualOperator, BOOLOID, false,
3608 (Expr *) arg,
3609 (Expr *) makeBoolConst(false, false),
3610 InvalidOid, InvalidOid);
3611 }
3612 if (clause && IsA(clause, BooleanTest))
3613 {
3614 BooleanTest *btest = (BooleanTest *) clause;
3615 Node *arg = (Node *) btest->arg;
3616
3617 /* It must have matched the indexkey */
3618 Assert(match_index_to_operand(arg, indexcol, index));
3619 if (btest->booltesttype == IS_TRUE)
3620 {
3621 /* convert to indexkey = TRUE */
3622 return make_opclause(BooleanEqualOperator, BOOLOID, false,
3623 (Expr *) arg,
3624 (Expr *) makeBoolConst(true, false),
3625 InvalidOid, InvalidOid);
3626 }
3627 if (btest->booltesttype == IS_FALSE)
3628 {
3629 /* convert to indexkey = FALSE */
3630 return make_opclause(BooleanEqualOperator, BOOLOID, false,
3631 (Expr *) arg,
3632 (Expr *) makeBoolConst(false, false),
3633 InvalidOid, InvalidOid);
3634 }
3635 /* Oops */
3636 Assert(false);
3637 }
3638
3639 return NULL;
3640 }
3641
3642 /*
3643 * expand_indexqual_opclause --- expand a single indexqual condition
3644 * that is an operator clause
3645 *
3646 * The input is a single RestrictInfo, the output a list of RestrictInfos.
3647 *
3648 * In the base case this is just list_make1(), but we have to be prepared to
3649 * expand special cases that were accepted by match_special_index_operator().
3650 */
3651 static List *
expand_indexqual_opclause(RestrictInfo * rinfo,Oid opfamily,Oid idxcollation)3652 expand_indexqual_opclause(RestrictInfo *rinfo, Oid opfamily, Oid idxcollation)
3653 {
3654 Expr *clause = rinfo->clause;
3655
3656 /* we know these will succeed */
3657 Node *leftop = get_leftop(clause);
3658 Node *rightop = get_rightop(clause);
3659 Oid expr_op = ((OpExpr *) clause)->opno;
3660 Oid expr_coll = ((OpExpr *) clause)->inputcollid;
3661 Const *patt = (Const *) rightop;
3662 Const *prefix = NULL;
3663 Pattern_Prefix_Status pstatus;
3664
3665 /*
3666 * LIKE and regex operators are not members of any btree index opfamily,
3667 * but they can be members of opfamilies for more exotic index types such
3668 * as GIN. Therefore, we should only do expansion if the operator is
3669 * actually not in the opfamily. But checking that requires a syscache
3670 * lookup, so it's best to first see if the operator is one we are
3671 * interested in.
3672 */
3673 switch (expr_op)
3674 {
3675 case OID_TEXT_LIKE_OP:
3676 case OID_BPCHAR_LIKE_OP:
3677 case OID_NAME_LIKE_OP:
3678 case OID_BYTEA_LIKE_OP:
3679 if (!op_in_opfamily(expr_op, opfamily))
3680 {
3681 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Like, expr_coll,
3682 &prefix, NULL);
3683 return prefix_quals(leftop, opfamily, idxcollation, prefix, pstatus);
3684 }
3685 break;
3686
3687 case OID_TEXT_ICLIKE_OP:
3688 case OID_BPCHAR_ICLIKE_OP:
3689 case OID_NAME_ICLIKE_OP:
3690 if (!op_in_opfamily(expr_op, opfamily))
3691 {
3692 /* the right-hand const is type text for all of these */
3693 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Like_IC, expr_coll,
3694 &prefix, NULL);
3695 return prefix_quals(leftop, opfamily, idxcollation, prefix, pstatus);
3696 }
3697 break;
3698
3699 case OID_TEXT_REGEXEQ_OP:
3700 case OID_BPCHAR_REGEXEQ_OP:
3701 case OID_NAME_REGEXEQ_OP:
3702 if (!op_in_opfamily(expr_op, opfamily))
3703 {
3704 /* the right-hand const is type text for all of these */
3705 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Regex, expr_coll,
3706 &prefix, NULL);
3707 return prefix_quals(leftop, opfamily, idxcollation, prefix, pstatus);
3708 }
3709 break;
3710
3711 case OID_TEXT_ICREGEXEQ_OP:
3712 case OID_BPCHAR_ICREGEXEQ_OP:
3713 case OID_NAME_ICREGEXEQ_OP:
3714 if (!op_in_opfamily(expr_op, opfamily))
3715 {
3716 /* the right-hand const is type text for all of these */
3717 pstatus = pattern_fixed_prefix(patt, Pattern_Type_Regex_IC, expr_coll,
3718 &prefix, NULL);
3719 return prefix_quals(leftop, opfamily, idxcollation, prefix, pstatus);
3720 }
3721 break;
3722
3723 case OID_INET_SUB_OP:
3724 case OID_INET_SUBEQ_OP:
3725 if (!op_in_opfamily(expr_op, opfamily))
3726 {
3727 return network_prefix_quals(leftop, expr_op, opfamily,
3728 patt->constvalue);
3729 }
3730 break;
3731 }
3732
3733 /* Default case: just make a list of the unmodified indexqual */
3734 return list_make1(rinfo);
3735 }
3736
3737 /*
3738 * expand_indexqual_rowcompare --- expand a single indexqual condition
3739 * that is a RowCompareExpr
3740 *
3741 * This is a thin wrapper around adjust_rowcompare_for_index; we export the
3742 * latter so that createplan.c can use it to re-discover which columns of the
3743 * index are used by a row comparison indexqual.
3744 */
3745 static RestrictInfo *
expand_indexqual_rowcompare(RestrictInfo * rinfo,IndexOptInfo * index,int indexcol)3746 expand_indexqual_rowcompare(RestrictInfo *rinfo,
3747 IndexOptInfo *index,
3748 int indexcol)
3749 {
3750 RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3751 Expr *newclause;
3752 List *indexcolnos;
3753 bool var_on_left;
3754
3755 newclause = adjust_rowcompare_for_index(clause,
3756 index,
3757 indexcol,
3758 &indexcolnos,
3759 &var_on_left);
3760
3761 /*
3762 * If we didn't have to change the RowCompareExpr, return the original
3763 * RestrictInfo.
3764 */
3765 if (newclause == (Expr *) clause)
3766 return rinfo;
3767
3768 /* Else we need a new RestrictInfo */
3769 return make_simple_restrictinfo(newclause);
3770 }
3771
3772 /*
3773 * adjust_rowcompare_for_index --- expand a single indexqual condition
3774 * that is a RowCompareExpr
3775 *
3776 * It's already known that the first column of the row comparison matches
3777 * the specified column of the index. We can use additional columns of the
3778 * row comparison as index qualifications, so long as they match the index
3779 * in the "same direction", ie, the indexkeys are all on the same side of the
3780 * clause and the operators are all the same-type members of the opfamilies.
3781 * If all the columns of the RowCompareExpr match in this way, we just use it
3782 * as-is. Otherwise, we build a shortened RowCompareExpr (if more than one
3783 * column matches) or a simple OpExpr (if the first-column match is all
3784 * there is). In these cases the modified clause is always "<=" or ">="
3785 * even when the original was "<" or ">" --- this is necessary to match all
3786 * the rows that could match the original. (We are essentially building a
3787 * lossy version of the row comparison when we do this.)
3788 *
3789 * *indexcolnos receives an integer list of the index column numbers (zero
3790 * based) used in the resulting expression. The reason we need to return
3791 * that is that if the index is selected for use, createplan.c will need to
3792 * call this again to extract that list. (This is a bit grotty, but row
3793 * comparison indexquals aren't used enough to justify finding someplace to
3794 * keep the information in the Path representation.) Since createplan.c
3795 * also needs to know which side of the RowCompareExpr is the index side,
3796 * we also return *var_on_left_p rather than re-deducing that there.
3797 */
3798 Expr *
adjust_rowcompare_for_index(RowCompareExpr * clause,IndexOptInfo * index,int indexcol,List ** indexcolnos,bool * var_on_left_p)3799 adjust_rowcompare_for_index(RowCompareExpr *clause,
3800 IndexOptInfo *index,
3801 int indexcol,
3802 List **indexcolnos,
3803 bool *var_on_left_p)
3804 {
3805 bool var_on_left;
3806 int op_strategy;
3807 Oid op_lefttype;
3808 Oid op_righttype;
3809 int matching_cols;
3810 Oid expr_op;
3811 List *opfamilies;
3812 List *lefttypes;
3813 List *righttypes;
3814 List *new_ops;
3815 ListCell *largs_cell;
3816 ListCell *rargs_cell;
3817 ListCell *opnos_cell;
3818 ListCell *collids_cell;
3819
3820 /* We have to figure out (again) how the first col matches */
3821 var_on_left = match_index_to_operand((Node *) linitial(clause->largs),
3822 indexcol, index);
3823 Assert(var_on_left ||
3824 match_index_to_operand((Node *) linitial(clause->rargs),
3825 indexcol, index));
3826 *var_on_left_p = var_on_left;
3827
3828 expr_op = linitial_oid(clause->opnos);
3829 if (!var_on_left)
3830 expr_op = get_commutator(expr_op);
3831 get_op_opfamily_properties(expr_op, index->opfamily[indexcol], false,
3832 &op_strategy,
3833 &op_lefttype,
3834 &op_righttype);
3835
3836 /* Initialize returned list of which index columns are used */
3837 *indexcolnos = list_make1_int(indexcol);
3838
3839 /* Build lists of the opfamilies and operator datatypes in case needed */
3840 opfamilies = list_make1_oid(index->opfamily[indexcol]);
3841 lefttypes = list_make1_oid(op_lefttype);
3842 righttypes = list_make1_oid(op_righttype);
3843
3844 /*
3845 * See how many of the remaining columns match some index column in the
3846 * same way. As in match_clause_to_indexcol(), the "other" side of any
3847 * potential index condition is OK as long as it doesn't use Vars from the
3848 * indexed relation.
3849 */
3850 matching_cols = 1;
3851 largs_cell = lnext(list_head(clause->largs));
3852 rargs_cell = lnext(list_head(clause->rargs));
3853 opnos_cell = lnext(list_head(clause->opnos));
3854 collids_cell = lnext(list_head(clause->inputcollids));
3855
3856 while (largs_cell != NULL)
3857 {
3858 Node *varop;
3859 Node *constop;
3860 int i;
3861
3862 expr_op = lfirst_oid(opnos_cell);
3863 if (var_on_left)
3864 {
3865 varop = (Node *) lfirst(largs_cell);
3866 constop = (Node *) lfirst(rargs_cell);
3867 }
3868 else
3869 {
3870 varop = (Node *) lfirst(rargs_cell);
3871 constop = (Node *) lfirst(largs_cell);
3872 /* indexkey is on right, so commute the operator */
3873 expr_op = get_commutator(expr_op);
3874 if (expr_op == InvalidOid)
3875 break; /* operator is not usable */
3876 }
3877 if (bms_is_member(index->rel->relid, pull_varnos(constop)))
3878 break; /* no good, Var on wrong side */
3879 if (contain_volatile_functions(constop))
3880 break; /* no good, volatile comparison value */
3881
3882 /*
3883 * The Var side can match any key column of the index.
3884 */
3885 for (i = 0; i < index->nkeycolumns; i++)
3886 {
3887 if (match_index_to_operand(varop, i, index) &&
3888 get_op_opfamily_strategy(expr_op,
3889 index->opfamily[i]) == op_strategy &&
3890 IndexCollMatchesExprColl(index->indexcollations[i],
3891 lfirst_oid(collids_cell)))
3892
3893 break;
3894 }
3895 if (i >= index->nkeycolumns)
3896 break; /* no match found */
3897
3898 /* Add column number to returned list */
3899 *indexcolnos = lappend_int(*indexcolnos, i);
3900
3901 /* Add opfamily and datatypes to lists */
3902 get_op_opfamily_properties(expr_op, index->opfamily[i], false,
3903 &op_strategy,
3904 &op_lefttype,
3905 &op_righttype);
3906 opfamilies = lappend_oid(opfamilies, index->opfamily[i]);
3907 lefttypes = lappend_oid(lefttypes, op_lefttype);
3908 righttypes = lappend_oid(righttypes, op_righttype);
3909
3910 /* This column matches, keep scanning */
3911 matching_cols++;
3912 largs_cell = lnext(largs_cell);
3913 rargs_cell = lnext(rargs_cell);
3914 opnos_cell = lnext(opnos_cell);
3915 collids_cell = lnext(collids_cell);
3916 }
3917
3918 /* Return clause as-is if it's all usable as index quals */
3919 if (matching_cols == list_length(clause->opnos))
3920 return (Expr *) clause;
3921
3922 /*
3923 * We have to generate a subset rowcompare (possibly just one OpExpr). The
3924 * painful part of this is changing < to <= or > to >=, so deal with that
3925 * first.
3926 */
3927 if (op_strategy == BTLessEqualStrategyNumber ||
3928 op_strategy == BTGreaterEqualStrategyNumber)
3929 {
3930 /* easy, just use the same operators */
3931 new_ops = list_truncate(list_copy(clause->opnos), matching_cols);
3932 }
3933 else
3934 {
3935 ListCell *opfamilies_cell;
3936 ListCell *lefttypes_cell;
3937 ListCell *righttypes_cell;
3938
3939 if (op_strategy == BTLessStrategyNumber)
3940 op_strategy = BTLessEqualStrategyNumber;
3941 else if (op_strategy == BTGreaterStrategyNumber)
3942 op_strategy = BTGreaterEqualStrategyNumber;
3943 else
3944 elog(ERROR, "unexpected strategy number %d", op_strategy);
3945 new_ops = NIL;
3946 lefttypes_cell = list_head(lefttypes);
3947 righttypes_cell = list_head(righttypes);
3948 foreach(opfamilies_cell, opfamilies)
3949 {
3950 Oid opfam = lfirst_oid(opfamilies_cell);
3951 Oid lefttype = lfirst_oid(lefttypes_cell);
3952 Oid righttype = lfirst_oid(righttypes_cell);
3953
3954 expr_op = get_opfamily_member(opfam, lefttype, righttype,
3955 op_strategy);
3956 if (!OidIsValid(expr_op)) /* should not happen */
3957 elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
3958 op_strategy, lefttype, righttype, opfam);
3959 if (!var_on_left)
3960 {
3961 expr_op = get_commutator(expr_op);
3962 if (!OidIsValid(expr_op)) /* should not happen */
3963 elog(ERROR, "could not find commutator of operator %d(%u,%u) of opfamily %u",
3964 op_strategy, lefttype, righttype, opfam);
3965 }
3966 new_ops = lappend_oid(new_ops, expr_op);
3967 lefttypes_cell = lnext(lefttypes_cell);
3968 righttypes_cell = lnext(righttypes_cell);
3969 }
3970 }
3971
3972 /* If we have more than one matching col, create a subset rowcompare */
3973 if (matching_cols > 1)
3974 {
3975 RowCompareExpr *rc = makeNode(RowCompareExpr);
3976
3977 if (var_on_left)
3978 rc->rctype = (RowCompareType) op_strategy;
3979 else
3980 rc->rctype = (op_strategy == BTLessEqualStrategyNumber) ?
3981 ROWCOMPARE_GE : ROWCOMPARE_LE;
3982 rc->opnos = new_ops;
3983 rc->opfamilies = list_truncate(list_copy(clause->opfamilies),
3984 matching_cols);
3985 rc->inputcollids = list_truncate(list_copy(clause->inputcollids),
3986 matching_cols);
3987 rc->largs = list_truncate(copyObject(clause->largs),
3988 matching_cols);
3989 rc->rargs = list_truncate(copyObject(clause->rargs),
3990 matching_cols);
3991 return (Expr *) rc;
3992 }
3993 else
3994 {
3995 return make_opclause(linitial_oid(new_ops), BOOLOID, false,
3996 copyObject(linitial(clause->largs)),
3997 copyObject(linitial(clause->rargs)),
3998 InvalidOid,
3999 linitial_oid(clause->inputcollids));
4000 }
4001 }
4002
4003 /*
4004 * Given a fixed prefix that all the "leftop" values must have,
4005 * generate suitable indexqual condition(s). opfamily is the index
4006 * operator family; we use it to deduce the appropriate comparison
4007 * operators and operand datatypes. collation is the input collation to use.
4008 */
4009 static List *
prefix_quals(Node * leftop,Oid opfamily,Oid collation,Const * prefix_const,Pattern_Prefix_Status pstatus)4010 prefix_quals(Node *leftop, Oid opfamily, Oid collation,
4011 Const *prefix_const, Pattern_Prefix_Status pstatus)
4012 {
4013 List *result;
4014 Oid datatype;
4015 Oid oproid;
4016 Expr *expr;
4017 FmgrInfo ltproc;
4018 Const *greaterstr;
4019
4020 Assert(pstatus != Pattern_Prefix_None);
4021
4022 switch (opfamily)
4023 {
4024 case TEXT_BTREE_FAM_OID:
4025 case TEXT_PATTERN_BTREE_FAM_OID:
4026 case TEXT_SPGIST_FAM_OID:
4027 datatype = TEXTOID;
4028 break;
4029
4030 case BPCHAR_BTREE_FAM_OID:
4031 case BPCHAR_PATTERN_BTREE_FAM_OID:
4032 datatype = BPCHAROID;
4033 break;
4034
4035 case NAME_BTREE_FAM_OID:
4036 datatype = NAMEOID;
4037 break;
4038
4039 case BYTEA_BTREE_FAM_OID:
4040 datatype = BYTEAOID;
4041 break;
4042
4043 default:
4044 /* shouldn't get here */
4045 elog(ERROR, "unexpected opfamily: %u", opfamily);
4046 return NIL;
4047 }
4048
4049 /*
4050 * If necessary, coerce the prefix constant to the right type. The given
4051 * prefix constant is either text or bytea type.
4052 */
4053 if (prefix_const->consttype != datatype)
4054 {
4055 char *prefix;
4056
4057 switch (prefix_const->consttype)
4058 {
4059 case TEXTOID:
4060 prefix = TextDatumGetCString(prefix_const->constvalue);
4061 break;
4062 case BYTEAOID:
4063 prefix = DatumGetCString(DirectFunctionCall1(byteaout,
4064 prefix_const->constvalue));
4065 break;
4066 default:
4067 elog(ERROR, "unexpected const type: %u",
4068 prefix_const->consttype);
4069 return NIL;
4070 }
4071 prefix_const = string_to_const(prefix, datatype);
4072 pfree(prefix);
4073 }
4074
4075 /*
4076 * If we found an exact-match pattern, generate an "=" indexqual.
4077 */
4078 if (pstatus == Pattern_Prefix_Exact)
4079 {
4080 oproid = get_opfamily_member(opfamily, datatype, datatype,
4081 BTEqualStrategyNumber);
4082 if (oproid == InvalidOid)
4083 elog(ERROR, "no = operator for opfamily %u", opfamily);
4084 expr = make_opclause(oproid, BOOLOID, false,
4085 (Expr *) leftop, (Expr *) prefix_const,
4086 InvalidOid, collation);
4087 result = list_make1(make_simple_restrictinfo(expr));
4088 return result;
4089 }
4090
4091 /*
4092 * Otherwise, we have a nonempty required prefix of the values.
4093 *
4094 * We can always say "x >= prefix".
4095 */
4096 oproid = get_opfamily_member(opfamily, datatype, datatype,
4097 BTGreaterEqualStrategyNumber);
4098 if (oproid == InvalidOid)
4099 elog(ERROR, "no >= operator for opfamily %u", opfamily);
4100 expr = make_opclause(oproid, BOOLOID, false,
4101 (Expr *) leftop, (Expr *) prefix_const,
4102 InvalidOid, collation);
4103 result = list_make1(make_simple_restrictinfo(expr));
4104
4105 /*-------
4106 * If we can create a string larger than the prefix, we can say
4107 * "x < greaterstr". NB: we rely on make_greater_string() to generate
4108 * a guaranteed-greater string, not just a probably-greater string.
4109 * In general this is only guaranteed in C locale, so we'd better be
4110 * using a C-locale index collation.
4111 *-------
4112 */
4113 oproid = get_opfamily_member(opfamily, datatype, datatype,
4114 BTLessStrategyNumber);
4115 if (oproid == InvalidOid)
4116 elog(ERROR, "no < operator for opfamily %u", opfamily);
4117 fmgr_info(get_opcode(oproid), <proc);
4118 greaterstr = make_greater_string(prefix_const, <proc, collation);
4119 if (greaterstr)
4120 {
4121 expr = make_opclause(oproid, BOOLOID, false,
4122 (Expr *) leftop, (Expr *) greaterstr,
4123 InvalidOid, collation);
4124 result = lappend(result, make_simple_restrictinfo(expr));
4125 }
4126
4127 return result;
4128 }
4129
4130 /*
4131 * Given a leftop and a rightop, and an inet-family sup/sub operator,
4132 * generate suitable indexqual condition(s). expr_op is the original
4133 * operator, and opfamily is the index opfamily.
4134 */
4135 static List *
network_prefix_quals(Node * leftop,Oid expr_op,Oid opfamily,Datum rightop)4136 network_prefix_quals(Node *leftop, Oid expr_op, Oid opfamily, Datum rightop)
4137 {
4138 bool is_eq;
4139 Oid datatype;
4140 Oid opr1oid;
4141 Oid opr2oid;
4142 Datum opr1right;
4143 Datum opr2right;
4144 List *result;
4145 Expr *expr;
4146
4147 switch (expr_op)
4148 {
4149 case OID_INET_SUB_OP:
4150 datatype = INETOID;
4151 is_eq = false;
4152 break;
4153 case OID_INET_SUBEQ_OP:
4154 datatype = INETOID;
4155 is_eq = true;
4156 break;
4157 default:
4158 elog(ERROR, "unexpected operator: %u", expr_op);
4159 return NIL;
4160 }
4161
4162 /*
4163 * create clause "key >= network_scan_first( rightop )", or ">" if the
4164 * operator disallows equality.
4165 */
4166 if (is_eq)
4167 {
4168 opr1oid = get_opfamily_member(opfamily, datatype, datatype,
4169 BTGreaterEqualStrategyNumber);
4170 if (opr1oid == InvalidOid)
4171 elog(ERROR, "no >= operator for opfamily %u", opfamily);
4172 }
4173 else
4174 {
4175 opr1oid = get_opfamily_member(opfamily, datatype, datatype,
4176 BTGreaterStrategyNumber);
4177 if (opr1oid == InvalidOid)
4178 elog(ERROR, "no > operator for opfamily %u", opfamily);
4179 }
4180
4181 opr1right = network_scan_first(rightop);
4182
4183 expr = make_opclause(opr1oid, BOOLOID, false,
4184 (Expr *) leftop,
4185 (Expr *) makeConst(datatype, -1,
4186 InvalidOid, /* not collatable */
4187 -1, opr1right,
4188 false, false),
4189 InvalidOid, InvalidOid);
4190 result = list_make1(make_simple_restrictinfo(expr));
4191
4192 /* create clause "key <= network_scan_last( rightop )" */
4193
4194 opr2oid = get_opfamily_member(opfamily, datatype, datatype,
4195 BTLessEqualStrategyNumber);
4196 if (opr2oid == InvalidOid)
4197 elog(ERROR, "no <= operator for opfamily %u", opfamily);
4198
4199 opr2right = network_scan_last(rightop);
4200
4201 expr = make_opclause(opr2oid, BOOLOID, false,
4202 (Expr *) leftop,
4203 (Expr *) makeConst(datatype, -1,
4204 InvalidOid, /* not collatable */
4205 -1, opr2right,
4206 false, false),
4207 InvalidOid, InvalidOid);
4208 result = lappend(result, make_simple_restrictinfo(expr));
4209
4210 return result;
4211 }
4212
4213 /*
4214 * Handy subroutines for match_special_index_operator() and friends.
4215 */
4216
4217 /*
4218 * Generate a Datum of the appropriate type from a C string.
4219 * Note that all of the supported types are pass-by-ref, so the
4220 * returned value should be pfree'd if no longer needed.
4221 */
4222 static Datum
string_to_datum(const char * str,Oid datatype)4223 string_to_datum(const char *str, Oid datatype)
4224 {
4225 /*
4226 * We cheat a little by assuming that CStringGetTextDatum() will do for
4227 * bpchar and varchar constants too...
4228 */
4229 if (datatype == NAMEOID)
4230 return DirectFunctionCall1(namein, CStringGetDatum(str));
4231 else if (datatype == BYTEAOID)
4232 return DirectFunctionCall1(byteain, CStringGetDatum(str));
4233 else
4234 return CStringGetTextDatum(str);
4235 }
4236
4237 /*
4238 * Generate a Const node of the appropriate type from a C string.
4239 */
4240 static Const *
string_to_const(const char * str,Oid datatype)4241 string_to_const(const char *str, Oid datatype)
4242 {
4243 Datum conval = string_to_datum(str, datatype);
4244 Oid collation;
4245 int constlen;
4246
4247 /*
4248 * We only need to support a few datatypes here, so hard-wire properties
4249 * instead of incurring the expense of catalog lookups.
4250 */
4251 switch (datatype)
4252 {
4253 case TEXTOID:
4254 case VARCHAROID:
4255 case BPCHAROID:
4256 collation = DEFAULT_COLLATION_OID;
4257 constlen = -1;
4258 break;
4259
4260 case NAMEOID:
4261 collation = InvalidOid;
4262 constlen = NAMEDATALEN;
4263 break;
4264
4265 case BYTEAOID:
4266 collation = InvalidOid;
4267 constlen = -1;
4268 break;
4269
4270 default:
4271 elog(ERROR, "unexpected datatype in string_to_const: %u",
4272 datatype);
4273 return NULL;
4274 }
4275
4276 return makeConst(datatype, -1, collation, constlen,
4277 conval, false, false);
4278 }
4279