Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * indxpath.c
4 : * Routines to determine which indexes are usable for scanning a
5 : * given relation, and create Paths accordingly.
6 : *
7 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : *
11 : * IDENTIFICATION
12 : * src/backend/optimizer/path/indxpath.c
13 : *
14 : *-------------------------------------------------------------------------
15 : */
16 : #include "postgres.h"
17 :
18 : #include <math.h>
19 :
20 : #include "access/stratnum.h"
21 : #include "access/sysattr.h"
22 : #include "catalog/pg_am.h"
23 : #include "catalog/pg_amop.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 "nodes/nodeFuncs.h"
29 : #include "nodes/supportnodes.h"
30 : #include "optimizer/cost.h"
31 : #include "optimizer/optimizer.h"
32 : #include "optimizer/pathnode.h"
33 : #include "optimizer/paths.h"
34 : #include "optimizer/prep.h"
35 : #include "optimizer/restrictinfo.h"
36 : #include "utils/array.h"
37 : #include "utils/lsyscache.h"
38 : #include "utils/selfuncs.h"
39 : #include "utils/syscache.h"
40 :
41 :
42 : /* XXX see PartCollMatchesExprColl */
43 : #define IndexCollMatchesExprColl(idxcollation, exprcollation) \
44 : ((idxcollation) == InvalidOid || (idxcollation) == (exprcollation))
45 :
46 : /* Whether we are looking for plain indexscan, bitmap scan, or either */
47 : typedef enum
48 : {
49 : ST_INDEXSCAN, /* must support amgettuple */
50 : ST_BITMAPSCAN, /* must support amgetbitmap */
51 : ST_ANYSCAN, /* either is okay */
52 : } ScanTypeControl;
53 :
54 : /* Data structure for collecting qual clauses that match an index */
55 : typedef struct
56 : {
57 : bool nonempty; /* True if lists are not all empty */
58 : /* Lists of IndexClause nodes, one list per index column */
59 : List *indexclauses[INDEX_MAX_KEYS];
60 : } IndexClauseSet;
61 :
62 : /* Per-path data used within choose_bitmap_and() */
63 : typedef struct
64 : {
65 : Path *path; /* IndexPath, BitmapAndPath, or BitmapOrPath */
66 : List *quals; /* the WHERE clauses it uses */
67 : List *preds; /* predicates of its partial index(es) */
68 : Bitmapset *clauseids; /* quals+preds represented as a bitmapset */
69 : bool unclassifiable; /* has too many quals+preds to process? */
70 : } PathClauseUsage;
71 :
72 : /* Callback argument for ec_member_matches_indexcol */
73 : typedef struct
74 : {
75 : IndexOptInfo *index; /* index we're considering */
76 : int indexcol; /* index column we want to match to */
77 : } ec_member_matches_arg;
78 :
79 :
80 : static void consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
81 : IndexOptInfo *index,
82 : IndexClauseSet *rclauseset,
83 : IndexClauseSet *jclauseset,
84 : IndexClauseSet *eclauseset,
85 : List **bitindexpaths);
86 : static void consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
87 : IndexOptInfo *index,
88 : IndexClauseSet *rclauseset,
89 : IndexClauseSet *jclauseset,
90 : IndexClauseSet *eclauseset,
91 : List **bitindexpaths,
92 : List *indexjoinclauses,
93 : int considered_clauses,
94 : List **considered_relids);
95 : static void get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
96 : IndexOptInfo *index,
97 : IndexClauseSet *rclauseset,
98 : IndexClauseSet *jclauseset,
99 : IndexClauseSet *eclauseset,
100 : List **bitindexpaths,
101 : Relids relids,
102 : List **considered_relids);
103 : static bool eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
104 : List *indexjoinclauses);
105 : static void get_index_paths(PlannerInfo *root, RelOptInfo *rel,
106 : IndexOptInfo *index, IndexClauseSet *clauses,
107 : List **bitindexpaths);
108 : static List *build_index_paths(PlannerInfo *root, RelOptInfo *rel,
109 : IndexOptInfo *index, IndexClauseSet *clauses,
110 : bool useful_predicate,
111 : ScanTypeControl scantype,
112 : bool *skip_nonnative_saop);
113 : static List *build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
114 : List *clauses, List *other_clauses);
115 : static List *generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
116 : List *clauses, List *other_clauses);
117 : static Path *choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel,
118 : List *paths);
119 : static int path_usage_comparator(const void *a, const void *b);
120 : static Cost bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel,
121 : Path *ipath);
122 : static Cost bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel,
123 : List *paths);
124 : static PathClauseUsage *classify_index_clause_usage(Path *path,
125 : List **clauselist);
126 : static void find_indexpath_quals(Path *bitmapqual, List **quals, List **preds);
127 : static int find_list_position(Node *node, List **nodelist);
128 : static bool check_index_only(RelOptInfo *rel, IndexOptInfo *index);
129 : static double get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids);
130 : static double adjust_rowcount_for_semijoins(PlannerInfo *root,
131 : Index cur_relid,
132 : Index outer_relid,
133 : double rowcount);
134 : static double approximate_joinrel_size(PlannerInfo *root, Relids relids);
135 : static void match_restriction_clauses_to_index(PlannerInfo *root,
136 : IndexOptInfo *index,
137 : IndexClauseSet *clauseset);
138 : static void match_join_clauses_to_index(PlannerInfo *root,
139 : RelOptInfo *rel, IndexOptInfo *index,
140 : IndexClauseSet *clauseset,
141 : List **joinorclauses);
142 : static void match_eclass_clauses_to_index(PlannerInfo *root,
143 : IndexOptInfo *index,
144 : IndexClauseSet *clauseset);
145 : static void match_clauses_to_index(PlannerInfo *root,
146 : List *clauses,
147 : IndexOptInfo *index,
148 : IndexClauseSet *clauseset);
149 : static void match_clause_to_index(PlannerInfo *root,
150 : RestrictInfo *rinfo,
151 : IndexOptInfo *index,
152 : IndexClauseSet *clauseset);
153 : static IndexClause *match_clause_to_indexcol(PlannerInfo *root,
154 : RestrictInfo *rinfo,
155 : int indexcol,
156 : IndexOptInfo *index);
157 : static bool IsBooleanOpfamily(Oid opfamily);
158 : static IndexClause *match_boolean_index_clause(PlannerInfo *root,
159 : RestrictInfo *rinfo,
160 : int indexcol, IndexOptInfo *index);
161 : static IndexClause *match_opclause_to_indexcol(PlannerInfo *root,
162 : RestrictInfo *rinfo,
163 : int indexcol,
164 : IndexOptInfo *index);
165 : static IndexClause *match_funcclause_to_indexcol(PlannerInfo *root,
166 : RestrictInfo *rinfo,
167 : int indexcol,
168 : IndexOptInfo *index);
169 : static IndexClause *get_index_clause_from_support(PlannerInfo *root,
170 : RestrictInfo *rinfo,
171 : Oid funcid,
172 : int indexarg,
173 : int indexcol,
174 : IndexOptInfo *index);
175 : static IndexClause *match_saopclause_to_indexcol(PlannerInfo *root,
176 : RestrictInfo *rinfo,
177 : int indexcol,
178 : IndexOptInfo *index);
179 : static IndexClause *match_rowcompare_to_indexcol(PlannerInfo *root,
180 : RestrictInfo *rinfo,
181 : int indexcol,
182 : IndexOptInfo *index);
183 : static IndexClause *match_orclause_to_indexcol(PlannerInfo *root,
184 : RestrictInfo *rinfo,
185 : int indexcol,
186 : IndexOptInfo *index);
187 : static IndexClause *expand_indexqual_rowcompare(PlannerInfo *root,
188 : RestrictInfo *rinfo,
189 : int indexcol,
190 : IndexOptInfo *index,
191 : Oid expr_op,
192 : bool var_on_left);
193 : static void match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
194 : List **orderby_clauses_p,
195 : List **clause_columns_p);
196 : static Expr *match_clause_to_ordering_op(IndexOptInfo *index,
197 : int indexcol, Expr *clause, Oid pk_opfamily);
198 : static bool ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
199 : EquivalenceClass *ec, EquivalenceMember *em,
200 : void *arg);
201 :
202 :
203 : /*
204 : * create_index_paths()
205 : * Generate all interesting index paths for the given relation.
206 : * Candidate paths are added to the rel's pathlist (using add_path).
207 : *
208 : * To be considered for an index scan, an index must match one or more
209 : * restriction clauses or join clauses from the query's qual condition,
210 : * or match the query's ORDER BY condition, or have a predicate that
211 : * matches the query's qual condition.
212 : *
213 : * There are two basic kinds of index scans. A "plain" index scan uses
214 : * only restriction clauses (possibly none at all) in its indexqual,
215 : * so it can be applied in any context. A "parameterized" index scan uses
216 : * join clauses (plus restriction clauses, if available) in its indexqual.
217 : * When joining such a scan to one of the relations supplying the other
218 : * variables used in its indexqual, the parameterized scan must appear as
219 : * the inner relation of a nestloop join; it can't be used on the outer side,
220 : * nor in a merge or hash join. In that context, values for the other rels'
221 : * attributes are available and fixed during any one scan of the indexpath.
222 : *
223 : * An IndexPath is generated and submitted to add_path() for each plain or
224 : * parameterized index scan this routine deems potentially interesting for
225 : * the current query.
226 : *
227 : * 'rel' is the relation for which we want to generate index paths
228 : *
229 : * Note: check_index_predicates() must have been run previously for this rel.
230 : *
231 : * Note: in cases involving LATERAL references in the relation's tlist, it's
232 : * possible that rel->lateral_relids is nonempty. Currently, we include
233 : * lateral_relids into the parameterization reported for each path, but don't
234 : * take it into account otherwise. The fact that any such rels *must* be
235 : * available as parameter sources perhaps should influence our choices of
236 : * index quals ... but for now, it doesn't seem worth troubling over.
237 : * In particular, comments below about "unparameterized" paths should be read
238 : * as meaning "unparameterized so far as the indexquals are concerned".
239 : */
240 : void
241 384248 : create_index_paths(PlannerInfo *root, RelOptInfo *rel)
242 : {
243 : List *indexpaths;
244 : List *bitindexpaths;
245 : List *bitjoinpaths;
246 : List *joinorclauses;
247 : IndexClauseSet rclauseset;
248 : IndexClauseSet jclauseset;
249 : IndexClauseSet eclauseset;
250 : ListCell *lc;
251 :
252 : /* Skip the whole mess if no indexes */
253 384248 : if (rel->indexlist == NIL)
254 68426 : return;
255 :
256 : /* Bitmap paths are collected and then dealt with at the end */
257 315822 : bitindexpaths = bitjoinpaths = joinorclauses = NIL;
258 :
259 : /* Examine each index in turn */
260 981702 : foreach(lc, rel->indexlist)
261 : {
262 665880 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
263 :
264 : /* Protect limited-size array in IndexClauseSets */
265 : Assert(index->nkeycolumns <= INDEX_MAX_KEYS);
266 :
267 : /*
268 : * Ignore partial indexes that do not match the query.
269 : * (generate_bitmap_or_paths() might be able to do something with
270 : * them, but that's of no concern here.)
271 : */
272 665880 : if (index->indpred != NIL && !index->predOK)
273 496 : continue;
274 :
275 : /*
276 : * Identify the restriction clauses that can match the index.
277 : */
278 22623056 : MemSet(&rclauseset, 0, sizeof(rclauseset));
279 665384 : match_restriction_clauses_to_index(root, index, &rclauseset);
280 :
281 : /*
282 : * Build index paths from the restriction clauses. These will be
283 : * non-parameterized paths. Plain paths go directly to add_path(),
284 : * bitmap paths are added to bitindexpaths to be handled below.
285 : */
286 665384 : get_index_paths(root, rel, index, &rclauseset,
287 : &bitindexpaths);
288 :
289 : /*
290 : * Identify the join clauses that can match the index. For the moment
291 : * we keep them separate from the restriction clauses. Note that this
292 : * step finds only "loose" join clauses that have not been merged into
293 : * EquivalenceClasses. Also, collect join OR clauses for later.
294 : */
295 22623056 : MemSet(&jclauseset, 0, sizeof(jclauseset));
296 665384 : match_join_clauses_to_index(root, rel, index,
297 : &jclauseset, &joinorclauses);
298 :
299 : /*
300 : * Look for EquivalenceClasses that can generate joinclauses matching
301 : * the index.
302 : */
303 22623056 : MemSet(&eclauseset, 0, sizeof(eclauseset));
304 665384 : match_eclass_clauses_to_index(root, index,
305 : &eclauseset);
306 :
307 : /*
308 : * If we found any plain or eclass join clauses, build parameterized
309 : * index paths using them.
310 : */
311 665384 : if (jclauseset.nonempty || eclauseset.nonempty)
312 126982 : consider_index_join_clauses(root, rel, index,
313 : &rclauseset,
314 : &jclauseset,
315 : &eclauseset,
316 : &bitjoinpaths);
317 : }
318 :
319 : /*
320 : * Generate BitmapOrPaths for any suitable OR-clauses present in the
321 : * restriction list. Add these to bitindexpaths.
322 : */
323 315822 : indexpaths = generate_bitmap_or_paths(root, rel,
324 : rel->baserestrictinfo, NIL);
325 315822 : bitindexpaths = list_concat(bitindexpaths, indexpaths);
326 :
327 : /*
328 : * Likewise, generate BitmapOrPaths for any suitable OR-clauses present in
329 : * the joinclause list. Add these to bitjoinpaths.
330 : */
331 315822 : indexpaths = generate_bitmap_or_paths(root, rel,
332 : joinorclauses, rel->baserestrictinfo);
333 315822 : bitjoinpaths = list_concat(bitjoinpaths, indexpaths);
334 :
335 : /*
336 : * If we found anything usable, generate a BitmapHeapPath for the most
337 : * promising combination of restriction bitmap index paths. Note there
338 : * will be only one such path no matter how many indexes exist. This
339 : * should be sufficient since there's basically only one figure of merit
340 : * (total cost) for such a path.
341 : */
342 315822 : if (bitindexpaths != NIL)
343 : {
344 : Path *bitmapqual;
345 : BitmapHeapPath *bpath;
346 :
347 190718 : bitmapqual = choose_bitmap_and(root, rel, bitindexpaths);
348 190718 : bpath = create_bitmap_heap_path(root, rel, bitmapqual,
349 : rel->lateral_relids, 1.0, 0);
350 190718 : add_path(rel, (Path *) bpath);
351 :
352 : /* create a partial bitmap heap path */
353 190718 : if (rel->consider_parallel && rel->lateral_relids == NULL)
354 136284 : create_partial_bitmap_paths(root, rel, bitmapqual);
355 : }
356 :
357 : /*
358 : * Likewise, if we found anything usable, generate BitmapHeapPaths for the
359 : * most promising combinations of join bitmap index paths. Our strategy
360 : * is to generate one such path for each distinct parameterization seen
361 : * among the available bitmap index paths. This may look pretty
362 : * expensive, but usually there won't be very many distinct
363 : * parameterizations. (This logic is quite similar to that in
364 : * consider_index_join_clauses, but we're working with whole paths not
365 : * individual clauses.)
366 : */
367 315822 : if (bitjoinpaths != NIL)
368 : {
369 : List *all_path_outers;
370 :
371 : /* Identify each distinct parameterization seen in bitjoinpaths */
372 115206 : all_path_outers = NIL;
373 257202 : foreach(lc, bitjoinpaths)
374 : {
375 141996 : Path *path = (Path *) lfirst(lc);
376 141996 : Relids required_outer = PATH_REQ_OUTER(path);
377 :
378 141996 : all_path_outers = list_append_unique(all_path_outers,
379 : required_outer);
380 : }
381 :
382 : /* Now, for each distinct parameterization set ... */
383 249420 : foreach(lc, all_path_outers)
384 : {
385 134214 : Relids max_outers = (Relids) lfirst(lc);
386 : List *this_path_set;
387 : Path *bitmapqual;
388 : Relids required_outer;
389 : double loop_count;
390 : BitmapHeapPath *bpath;
391 : ListCell *lcp;
392 :
393 : /* Identify all the bitmap join paths needing no more than that */
394 134214 : this_path_set = NIL;
395 329140 : foreach(lcp, bitjoinpaths)
396 : {
397 194926 : Path *path = (Path *) lfirst(lcp);
398 :
399 194926 : if (bms_is_subset(PATH_REQ_OUTER(path), max_outers))
400 148870 : this_path_set = lappend(this_path_set, path);
401 : }
402 :
403 : /*
404 : * Add in restriction bitmap paths, since they can be used
405 : * together with any join paths.
406 : */
407 134214 : this_path_set = list_concat(this_path_set, bitindexpaths);
408 :
409 : /* Select best AND combination for this parameterization */
410 134214 : bitmapqual = choose_bitmap_and(root, rel, this_path_set);
411 :
412 : /* And push that path into the mix */
413 134214 : required_outer = PATH_REQ_OUTER(bitmapqual);
414 134214 : loop_count = get_loop_count(root, rel->relid, required_outer);
415 134214 : bpath = create_bitmap_heap_path(root, rel, bitmapqual,
416 : required_outer, loop_count, 0);
417 134214 : add_path(rel, (Path *) bpath);
418 : }
419 : }
420 : }
421 :
422 : /*
423 : * consider_index_join_clauses
424 : * Given sets of join clauses for an index, decide which parameterized
425 : * index paths to build.
426 : *
427 : * Plain indexpaths are sent directly to add_path, while potential
428 : * bitmap indexpaths are added to *bitindexpaths for later processing.
429 : *
430 : * 'rel' is the index's heap relation
431 : * 'index' is the index for which we want to generate paths
432 : * 'rclauseset' is the collection of indexable restriction clauses
433 : * 'jclauseset' is the collection of indexable simple join clauses
434 : * 'eclauseset' is the collection of indexable clauses from EquivalenceClasses
435 : * '*bitindexpaths' is the list to add bitmap paths to
436 : */
437 : static void
438 126982 : consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
439 : IndexOptInfo *index,
440 : IndexClauseSet *rclauseset,
441 : IndexClauseSet *jclauseset,
442 : IndexClauseSet *eclauseset,
443 : List **bitindexpaths)
444 : {
445 126982 : int considered_clauses = 0;
446 126982 : List *considered_relids = NIL;
447 : int indexcol;
448 :
449 : /*
450 : * The strategy here is to identify every potentially useful set of outer
451 : * rels that can provide indexable join clauses. For each such set,
452 : * select all the join clauses available from those outer rels, add on all
453 : * the indexable restriction clauses, and generate plain and/or bitmap
454 : * index paths for that set of clauses. This is based on the assumption
455 : * that it's always better to apply a clause as an indexqual than as a
456 : * filter (qpqual); which is where an available clause would end up being
457 : * applied if we omit it from the indexquals.
458 : *
459 : * This looks expensive, but in most practical cases there won't be very
460 : * many distinct sets of outer rels to consider. As a safety valve when
461 : * that's not true, we use a heuristic: limit the number of outer rel sets
462 : * considered to a multiple of the number of clauses considered. (We'll
463 : * always consider using each individual join clause, though.)
464 : *
465 : * For simplicity in selecting relevant clauses, we represent each set of
466 : * outer rels as a maximum set of clause_relids --- that is, the indexed
467 : * relation itself is also included in the relids set. considered_relids
468 : * lists all relids sets we've already tried.
469 : */
470 315812 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
471 : {
472 : /* Consider each applicable simple join clause */
473 188830 : considered_clauses += list_length(jclauseset->indexclauses[indexcol]);
474 188830 : consider_index_join_outer_rels(root, rel, index,
475 : rclauseset, jclauseset, eclauseset,
476 : bitindexpaths,
477 : jclauseset->indexclauses[indexcol],
478 : considered_clauses,
479 : &considered_relids);
480 : /* Consider each applicable eclass join clause */
481 188830 : considered_clauses += list_length(eclauseset->indexclauses[indexcol]);
482 188830 : consider_index_join_outer_rels(root, rel, index,
483 : rclauseset, jclauseset, eclauseset,
484 : bitindexpaths,
485 : eclauseset->indexclauses[indexcol],
486 : considered_clauses,
487 : &considered_relids);
488 : }
489 126982 : }
490 :
491 : /*
492 : * consider_index_join_outer_rels
493 : * Generate parameterized paths based on clause relids in the clause list.
494 : *
495 : * Workhorse for consider_index_join_clauses; see notes therein for rationale.
496 : *
497 : * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset', and
498 : * 'bitindexpaths' as above
499 : * 'indexjoinclauses' is a list of IndexClauses for join clauses
500 : * 'considered_clauses' is the total number of clauses considered (so far)
501 : * '*considered_relids' is a list of all relids sets already considered
502 : */
503 : static void
504 377660 : consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
505 : IndexOptInfo *index,
506 : IndexClauseSet *rclauseset,
507 : IndexClauseSet *jclauseset,
508 : IndexClauseSet *eclauseset,
509 : List **bitindexpaths,
510 : List *indexjoinclauses,
511 : int considered_clauses,
512 : List **considered_relids)
513 : {
514 : ListCell *lc;
515 :
516 : /* Examine relids of each joinclause in the given list */
517 520118 : foreach(lc, indexjoinclauses)
518 : {
519 142458 : IndexClause *iclause = (IndexClause *) lfirst(lc);
520 142458 : Relids clause_relids = iclause->rinfo->clause_relids;
521 142458 : EquivalenceClass *parent_ec = iclause->rinfo->parent_ec;
522 : int num_considered_relids;
523 :
524 : /* If we already tried its relids set, no need to do so again */
525 142458 : if (list_member(*considered_relids, clause_relids))
526 3812 : continue;
527 :
528 : /*
529 : * Generate the union of this clause's relids set with each
530 : * previously-tried set. This ensures we try this clause along with
531 : * every interesting subset of previous clauses. However, to avoid
532 : * exponential growth of planning time when there are many clauses,
533 : * limit the number of relid sets accepted to 10 * considered_clauses.
534 : *
535 : * Note: get_join_index_paths appends entries to *considered_relids,
536 : * but we do not need to visit such newly-added entries within this
537 : * loop, so we don't use foreach() here. No real harm would be done
538 : * if we did visit them, since the subset check would reject them; but
539 : * it would waste some cycles.
540 : */
541 138646 : num_considered_relids = list_length(*considered_relids);
542 150676 : for (int pos = 0; pos < num_considered_relids; pos++)
543 : {
544 12030 : Relids oldrelids = (Relids) list_nth(*considered_relids, pos);
545 :
546 : /*
547 : * If either is a subset of the other, no new set is possible.
548 : * This isn't a complete test for redundancy, but it's easy and
549 : * cheap. get_join_index_paths will check more carefully if we
550 : * already generated the same relids set.
551 : */
552 12030 : if (bms_subset_compare(clause_relids, oldrelids) != BMS_DIFFERENT)
553 24 : continue;
554 :
555 : /*
556 : * If this clause was derived from an equivalence class, the
557 : * clause list may contain other clauses derived from the same
558 : * eclass. We should not consider that combining this clause with
559 : * one of those clauses generates a usefully different
560 : * parameterization; so skip if any clause derived from the same
561 : * eclass would already have been included when using oldrelids.
562 : */
563 23850 : if (parent_ec &&
564 11844 : eclass_already_used(parent_ec, oldrelids,
565 : indexjoinclauses))
566 8640 : continue;
567 :
568 : /*
569 : * If the number of relid sets considered exceeds our heuristic
570 : * limit, stop considering combinations of clauses. We'll still
571 : * consider the current clause alone, though (below this loop).
572 : */
573 3366 : if (list_length(*considered_relids) >= 10 * considered_clauses)
574 0 : break;
575 :
576 : /* OK, try the union set */
577 3366 : get_join_index_paths(root, rel, index,
578 : rclauseset, jclauseset, eclauseset,
579 : bitindexpaths,
580 : bms_union(clause_relids, oldrelids),
581 : considered_relids);
582 : }
583 :
584 : /* Also try this set of relids by itself */
585 138646 : get_join_index_paths(root, rel, index,
586 : rclauseset, jclauseset, eclauseset,
587 : bitindexpaths,
588 : clause_relids,
589 : considered_relids);
590 : }
591 377660 : }
592 :
593 : /*
594 : * get_join_index_paths
595 : * Generate index paths using clauses from the specified outer relations.
596 : * In addition to generating paths, relids is added to *considered_relids
597 : * if not already present.
598 : *
599 : * Workhorse for consider_index_join_clauses; see notes therein for rationale.
600 : *
601 : * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset',
602 : * 'bitindexpaths', 'considered_relids' as above
603 : * 'relids' is the current set of relids to consider (the target rel plus
604 : * one or more outer rels)
605 : */
606 : static void
607 142012 : get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
608 : IndexOptInfo *index,
609 : IndexClauseSet *rclauseset,
610 : IndexClauseSet *jclauseset,
611 : IndexClauseSet *eclauseset,
612 : List **bitindexpaths,
613 : Relids relids,
614 : List **considered_relids)
615 : {
616 : IndexClauseSet clauseset;
617 : int indexcol;
618 :
619 : /* If we already considered this relids set, don't repeat the work */
620 142012 : if (list_member(*considered_relids, relids))
621 0 : return;
622 :
623 : /* Identify indexclauses usable with this relids set */
624 4828408 : MemSet(&clauseset, 0, sizeof(clauseset));
625 :
626 359430 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
627 : {
628 : ListCell *lc;
629 :
630 : /* First find applicable simple join clauses */
631 250920 : foreach(lc, jclauseset->indexclauses[indexcol])
632 : {
633 33502 : IndexClause *iclause = (IndexClause *) lfirst(lc);
634 :
635 33502 : if (bms_is_subset(iclause->rinfo->clause_relids, relids))
636 33076 : clauseset.indexclauses[indexcol] =
637 33076 : lappend(clauseset.indexclauses[indexcol], iclause);
638 : }
639 :
640 : /*
641 : * Add applicable eclass join clauses. The clauses generated for each
642 : * column are redundant (cf generate_implied_equalities_for_column),
643 : * so we need at most one. This is the only exception to the general
644 : * rule of using all available index clauses.
645 : */
646 235240 : foreach(lc, eclauseset->indexclauses[indexcol])
647 : {
648 133864 : IndexClause *iclause = (IndexClause *) lfirst(lc);
649 :
650 133864 : if (bms_is_subset(iclause->rinfo->clause_relids, relids))
651 : {
652 116042 : clauseset.indexclauses[indexcol] =
653 116042 : lappend(clauseset.indexclauses[indexcol], iclause);
654 116042 : break;
655 : }
656 : }
657 :
658 : /* Add restriction clauses */
659 217418 : clauseset.indexclauses[indexcol] =
660 217418 : list_concat(clauseset.indexclauses[indexcol],
661 217418 : rclauseset->indexclauses[indexcol]);
662 :
663 217418 : if (clauseset.indexclauses[indexcol] != NIL)
664 174498 : clauseset.nonempty = true;
665 : }
666 :
667 : /* We should have found something, else caller passed silly relids */
668 : Assert(clauseset.nonempty);
669 :
670 : /* Build index path(s) using the collected set of clauses */
671 142012 : get_index_paths(root, rel, index, &clauseset, bitindexpaths);
672 :
673 : /*
674 : * Remember we considered paths for this set of relids.
675 : */
676 142012 : *considered_relids = lappend(*considered_relids, relids);
677 : }
678 :
679 : /*
680 : * eclass_already_used
681 : * True if any join clause usable with oldrelids was generated from
682 : * the specified equivalence class.
683 : */
684 : static bool
685 11844 : eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
686 : List *indexjoinclauses)
687 : {
688 : ListCell *lc;
689 :
690 15510 : foreach(lc, indexjoinclauses)
691 : {
692 12306 : IndexClause *iclause = (IndexClause *) lfirst(lc);
693 12306 : RestrictInfo *rinfo = iclause->rinfo;
694 :
695 24612 : if (rinfo->parent_ec == parent_ec &&
696 12306 : bms_is_subset(rinfo->clause_relids, oldrelids))
697 8640 : return true;
698 : }
699 3204 : return false;
700 : }
701 :
702 :
703 : /*
704 : * get_index_paths
705 : * Given an index and a set of index clauses for it, construct IndexPaths.
706 : *
707 : * Plain indexpaths are sent directly to add_path, while potential
708 : * bitmap indexpaths are added to *bitindexpaths for later processing.
709 : *
710 : * This is a fairly simple frontend to build_index_paths(). Its reason for
711 : * existence is mainly to handle ScalarArrayOpExpr quals properly. If the
712 : * index AM supports them natively, we should just include them in simple
713 : * index paths. If not, we should exclude them while building simple index
714 : * paths, and then make a separate attempt to include them in bitmap paths.
715 : */
716 : static void
717 807396 : get_index_paths(PlannerInfo *root, RelOptInfo *rel,
718 : IndexOptInfo *index, IndexClauseSet *clauses,
719 : List **bitindexpaths)
720 : {
721 : List *indexpaths;
722 807396 : bool skip_nonnative_saop = false;
723 : ListCell *lc;
724 :
725 : /*
726 : * Build simple index paths using the clauses. Allow ScalarArrayOpExpr
727 : * clauses only if the index AM supports them natively.
728 : */
729 807396 : indexpaths = build_index_paths(root, rel,
730 : index, clauses,
731 807396 : index->predOK,
732 : ST_ANYSCAN,
733 : &skip_nonnative_saop);
734 :
735 : /*
736 : * Submit all the ones that can form plain IndexScan plans to add_path. (A
737 : * plain IndexPath can represent either a plain IndexScan or an
738 : * IndexOnlyScan, but for our purposes here that distinction does not
739 : * matter. However, some of the indexes might support only bitmap scans,
740 : * and those we mustn't submit to add_path here.)
741 : *
742 : * Also, pick out the ones that are usable as bitmap scans. For that, we
743 : * must discard indexes that don't support bitmap scans, and we also are
744 : * only interested in paths that have some selectivity; we should discard
745 : * anything that was generated solely for ordering purposes.
746 : */
747 1290692 : foreach(lc, indexpaths)
748 : {
749 483296 : IndexPath *ipath = (IndexPath *) lfirst(lc);
750 :
751 483296 : if (index->amhasgettuple)
752 469896 : add_path(rel, (Path *) ipath);
753 :
754 483296 : if (index->amhasgetbitmap &&
755 483296 : (ipath->path.pathkeys == NIL ||
756 299250 : ipath->indexselectivity < 1.0))
757 356282 : *bitindexpaths = lappend(*bitindexpaths, ipath);
758 : }
759 :
760 : /*
761 : * If there were ScalarArrayOpExpr clauses that the index can't handle
762 : * natively, generate bitmap scan paths relying on executor-managed
763 : * ScalarArrayOpExpr.
764 : */
765 807396 : if (skip_nonnative_saop)
766 : {
767 32 : indexpaths = build_index_paths(root, rel,
768 : index, clauses,
769 : false,
770 : ST_BITMAPSCAN,
771 : NULL);
772 32 : *bitindexpaths = list_concat(*bitindexpaths, indexpaths);
773 : }
774 807396 : }
775 :
776 : /*
777 : * build_index_paths
778 : * Given an index and a set of index clauses for it, construct zero
779 : * or more IndexPaths. It also constructs zero or more partial IndexPaths.
780 : *
781 : * We return a list of paths because (1) this routine checks some cases
782 : * that should cause us to not generate any IndexPath, and (2) in some
783 : * cases we want to consider both a forward and a backward scan, so as
784 : * to obtain both sort orders. Note that the paths are just returned
785 : * to the caller and not immediately fed to add_path().
786 : *
787 : * At top level, useful_predicate should be exactly the index's predOK flag
788 : * (ie, true if it has a predicate that was proven from the restriction
789 : * clauses). When working on an arm of an OR clause, useful_predicate
790 : * should be true if the predicate required the current OR list to be proven.
791 : * Note that this routine should never be called at all if the index has an
792 : * unprovable predicate.
793 : *
794 : * scantype indicates whether we want to create plain indexscans, bitmap
795 : * indexscans, or both. When it's ST_BITMAPSCAN, we will not consider
796 : * index ordering while deciding if a Path is worth generating.
797 : *
798 : * If skip_nonnative_saop is non-NULL, we ignore ScalarArrayOpExpr clauses
799 : * unless the index AM supports them directly, and we set *skip_nonnative_saop
800 : * to true if we found any such clauses (caller must initialize the variable
801 : * to false). If it's NULL, we do not ignore ScalarArrayOpExpr clauses.
802 : *
803 : * 'rel' is the index's heap relation
804 : * 'index' is the index for which we want to generate paths
805 : * 'clauses' is the collection of indexable clauses (IndexClause nodes)
806 : * 'useful_predicate' indicates whether the index has a useful predicate
807 : * 'scantype' indicates whether we need plain or bitmap scan support
808 : * 'skip_nonnative_saop' indicates whether to accept SAOP if index AM doesn't
809 : */
810 : static List *
811 810416 : build_index_paths(PlannerInfo *root, RelOptInfo *rel,
812 : IndexOptInfo *index, IndexClauseSet *clauses,
813 : bool useful_predicate,
814 : ScanTypeControl scantype,
815 : bool *skip_nonnative_saop)
816 : {
817 810416 : List *result = NIL;
818 : IndexPath *ipath;
819 : List *index_clauses;
820 : Relids outer_relids;
821 : double loop_count;
822 : List *orderbyclauses;
823 : List *orderbyclausecols;
824 : List *index_pathkeys;
825 : List *useful_pathkeys;
826 : bool pathkeys_possibly_useful;
827 : bool index_is_ordered;
828 : bool index_only_scan;
829 : int indexcol;
830 :
831 : Assert(skip_nonnative_saop != NULL || scantype == ST_BITMAPSCAN);
832 :
833 : /*
834 : * Check that index supports the desired scan type(s)
835 : */
836 810416 : switch (scantype)
837 : {
838 0 : case ST_INDEXSCAN:
839 0 : if (!index->amhasgettuple)
840 0 : return NIL;
841 0 : break;
842 3020 : case ST_BITMAPSCAN:
843 3020 : if (!index->amhasgetbitmap)
844 0 : return NIL;
845 3020 : break;
846 807396 : case ST_ANYSCAN:
847 : /* either or both are OK */
848 807396 : break;
849 : }
850 :
851 : /*
852 : * 1. Combine the per-column IndexClause lists into an overall list.
853 : *
854 : * In the resulting list, clauses are ordered by index key, so that the
855 : * column numbers form a nondecreasing sequence. (This order is depended
856 : * on by btree and possibly other places.) The list can be empty, if the
857 : * index AM allows that.
858 : *
859 : * We also build a Relids set showing which outer rels are required by the
860 : * selected clauses. Any lateral_relids are included in that, but not
861 : * otherwise accounted for.
862 : */
863 810416 : index_clauses = NIL;
864 810416 : outer_relids = bms_copy(rel->lateral_relids);
865 2315746 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
866 : {
867 : ListCell *lc;
868 :
869 1922034 : foreach(lc, clauses->indexclauses[indexcol])
870 : {
871 416374 : IndexClause *iclause = (IndexClause *) lfirst(lc);
872 416374 : RestrictInfo *rinfo = iclause->rinfo;
873 :
874 416374 : if (skip_nonnative_saop && !index->amsearcharray &&
875 21266 : IsA(rinfo->clause, ScalarArrayOpExpr))
876 : {
877 : /*
878 : * Caller asked us to generate IndexPaths that omit any
879 : * ScalarArrayOpExpr clauses when the underlying index AM
880 : * lacks native support.
881 : *
882 : * We must omit this clause (and tell caller about it).
883 : */
884 32 : *skip_nonnative_saop = true;
885 32 : continue;
886 : }
887 :
888 : /* OK to include this clause */
889 416342 : index_clauses = lappend(index_clauses, iclause);
890 416342 : outer_relids = bms_add_members(outer_relids,
891 416342 : rinfo->clause_relids);
892 : }
893 :
894 : /*
895 : * If no clauses match the first index column, check for amoptionalkey
896 : * restriction. We can't generate a scan over an index with
897 : * amoptionalkey = false unless there's at least one index clause.
898 : * (When working on columns after the first, this test cannot fail. It
899 : * is always okay for columns after the first to not have any
900 : * clauses.)
901 : */
902 1505660 : if (index_clauses == NIL && !index->amoptionalkey)
903 330 : return NIL;
904 : }
905 :
906 : /* We do not want the index's rel itself listed in outer_relids */
907 810086 : outer_relids = bms_del_member(outer_relids, rel->relid);
908 :
909 : /* Compute loop_count for cost estimation purposes */
910 810086 : loop_count = get_loop_count(root, rel->relid, outer_relids);
911 :
912 : /*
913 : * 2. Compute pathkeys describing index's ordering, if any, then see how
914 : * many of them are actually useful for this query. This is not relevant
915 : * if we are only trying to build bitmap indexscans.
916 : */
917 1617152 : pathkeys_possibly_useful = (scantype != ST_BITMAPSCAN &&
918 807066 : has_useful_pathkeys(root, rel));
919 810086 : index_is_ordered = (index->sortopfamily != NULL);
920 810086 : if (index_is_ordered && pathkeys_possibly_useful)
921 : {
922 598500 : index_pathkeys = build_index_pathkeys(root, index,
923 : ForwardScanDirection);
924 598500 : useful_pathkeys = truncate_useless_pathkeys(root, rel,
925 : index_pathkeys);
926 598500 : orderbyclauses = NIL;
927 598500 : orderbyclausecols = NIL;
928 : }
929 211586 : else if (index->amcanorderbyop && pathkeys_possibly_useful)
930 : {
931 : /*
932 : * See if we can generate ordering operators for query_pathkeys or at
933 : * least some prefix thereof. Matching to just a prefix of the
934 : * query_pathkeys will allow an incremental sort to be considered on
935 : * the index's partially sorted results.
936 : */
937 1078 : match_pathkeys_to_index(index, root->query_pathkeys,
938 : &orderbyclauses,
939 : &orderbyclausecols);
940 1078 : if (list_length(root->query_pathkeys) == list_length(orderbyclauses))
941 472 : useful_pathkeys = root->query_pathkeys;
942 : else
943 606 : useful_pathkeys = list_copy_head(root->query_pathkeys,
944 : list_length(orderbyclauses));
945 : }
946 : else
947 : {
948 210508 : useful_pathkeys = NIL;
949 210508 : orderbyclauses = NIL;
950 210508 : orderbyclausecols = NIL;
951 : }
952 :
953 : /*
954 : * 3. Check if an index-only scan is possible. If we're not building
955 : * plain indexscans, this isn't relevant since bitmap scans don't support
956 : * index data retrieval anyway.
957 : */
958 1617152 : index_only_scan = (scantype != ST_BITMAPSCAN &&
959 807066 : check_index_only(rel, index));
960 :
961 : /*
962 : * 4. Generate an indexscan path if there are relevant restriction clauses
963 : * in the current clauses, OR the index ordering is potentially useful for
964 : * later merging or final output ordering, OR the index has a useful
965 : * predicate, OR an index-only scan is possible.
966 : */
967 810086 : if (index_clauses != NIL || useful_pathkeys != NIL || useful_predicate ||
968 : index_only_scan)
969 : {
970 485756 : ipath = create_index_path(root, index,
971 : index_clauses,
972 : orderbyclauses,
973 : orderbyclausecols,
974 : useful_pathkeys,
975 : ForwardScanDirection,
976 : index_only_scan,
977 : outer_relids,
978 : loop_count,
979 : false);
980 485756 : result = lappend(result, ipath);
981 :
982 : /*
983 : * If appropriate, consider parallel index scan. We don't allow
984 : * parallel index scan for bitmap index scans.
985 : */
986 485756 : if (index->amcanparallel &&
987 465374 : rel->consider_parallel && outer_relids == NULL &&
988 : scantype != ST_BITMAPSCAN)
989 : {
990 251816 : ipath = create_index_path(root, index,
991 : index_clauses,
992 : orderbyclauses,
993 : orderbyclausecols,
994 : useful_pathkeys,
995 : ForwardScanDirection,
996 : index_only_scan,
997 : outer_relids,
998 : loop_count,
999 : true);
1000 :
1001 : /*
1002 : * if, after costing the path, we find that it's not worth using
1003 : * parallel workers, just free it.
1004 : */
1005 251816 : if (ipath->path.parallel_workers > 0)
1006 9844 : add_partial_path(rel, (Path *) ipath);
1007 : else
1008 241972 : pfree(ipath);
1009 : }
1010 : }
1011 :
1012 : /*
1013 : * 5. If the index is ordered, a backwards scan might be interesting.
1014 : */
1015 810086 : if (index_is_ordered && pathkeys_possibly_useful)
1016 : {
1017 598500 : index_pathkeys = build_index_pathkeys(root, index,
1018 : BackwardScanDirection);
1019 598500 : useful_pathkeys = truncate_useless_pathkeys(root, rel,
1020 : index_pathkeys);
1021 598500 : if (useful_pathkeys != NIL)
1022 : {
1023 560 : ipath = create_index_path(root, index,
1024 : index_clauses,
1025 : NIL,
1026 : NIL,
1027 : useful_pathkeys,
1028 : BackwardScanDirection,
1029 : index_only_scan,
1030 : outer_relids,
1031 : loop_count,
1032 : false);
1033 560 : result = lappend(result, ipath);
1034 :
1035 : /* If appropriate, consider parallel index scan */
1036 560 : if (index->amcanparallel &&
1037 560 : rel->consider_parallel && outer_relids == NULL &&
1038 : scantype != ST_BITMAPSCAN)
1039 : {
1040 470 : ipath = create_index_path(root, index,
1041 : index_clauses,
1042 : NIL,
1043 : NIL,
1044 : useful_pathkeys,
1045 : BackwardScanDirection,
1046 : index_only_scan,
1047 : outer_relids,
1048 : loop_count,
1049 : true);
1050 :
1051 : /*
1052 : * if, after costing the path, we find that it's not worth
1053 : * using parallel workers, just free it.
1054 : */
1055 470 : if (ipath->path.parallel_workers > 0)
1056 168 : add_partial_path(rel, (Path *) ipath);
1057 : else
1058 302 : pfree(ipath);
1059 : }
1060 : }
1061 : }
1062 :
1063 810086 : return result;
1064 : }
1065 :
1066 : /*
1067 : * build_paths_for_OR
1068 : * Given a list of restriction clauses from one arm of an OR clause,
1069 : * construct all matching IndexPaths for the relation.
1070 : *
1071 : * Here we must scan all indexes of the relation, since a bitmap OR tree
1072 : * can use multiple indexes.
1073 : *
1074 : * The caller actually supplies two lists of restriction clauses: some
1075 : * "current" ones and some "other" ones. Both lists can be used freely
1076 : * to match keys of the index, but an index must use at least one of the
1077 : * "current" clauses to be considered usable. The motivation for this is
1078 : * examples like
1079 : * WHERE (x = 42) AND (... OR (y = 52 AND z = 77) OR ....)
1080 : * While we are considering the y/z subclause of the OR, we can use "x = 42"
1081 : * as one of the available index conditions; but we shouldn't match the
1082 : * subclause to any index on x alone, because such a Path would already have
1083 : * been generated at the upper level. So we could use an index on x,y,z
1084 : * or an index on x,y for the OR subclause, but not an index on just x.
1085 : * When dealing with a partial index, a match of the index predicate to
1086 : * one of the "current" clauses also makes the index usable.
1087 : *
1088 : * 'rel' is the relation for which we want to generate index paths
1089 : * 'clauses' is the current list of clauses (RestrictInfo nodes)
1090 : * 'other_clauses' is the list of additional upper-level clauses
1091 : */
1092 : static List *
1093 11106 : build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
1094 : List *clauses, List *other_clauses)
1095 : {
1096 11106 : List *result = NIL;
1097 11106 : List *all_clauses = NIL; /* not computed till needed */
1098 : ListCell *lc;
1099 :
1100 39058 : foreach(lc, rel->indexlist)
1101 : {
1102 27952 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
1103 : IndexClauseSet clauseset;
1104 : List *indexpaths;
1105 : bool useful_predicate;
1106 :
1107 : /* Ignore index if it doesn't support bitmap scans */
1108 27952 : if (!index->amhasgetbitmap)
1109 24964 : continue;
1110 :
1111 : /*
1112 : * Ignore partial indexes that do not match the query. If a partial
1113 : * index is marked predOK then we know it's OK. Otherwise, we have to
1114 : * test whether the added clauses are sufficient to imply the
1115 : * predicate. If so, we can use the index in the current context.
1116 : *
1117 : * We set useful_predicate to true iff the predicate was proven using
1118 : * the current set of clauses. This is needed to prevent matching a
1119 : * predOK index to an arm of an OR, which would be a legal but
1120 : * pointlessly inefficient plan. (A better plan will be generated by
1121 : * just scanning the predOK index alone, no OR.)
1122 : */
1123 27952 : useful_predicate = false;
1124 27952 : if (index->indpred != NIL)
1125 : {
1126 168 : if (index->predOK)
1127 : {
1128 : /* Usable, but don't set useful_predicate */
1129 : }
1130 : else
1131 : {
1132 : /* Form all_clauses if not done already */
1133 144 : if (all_clauses == NIL)
1134 60 : all_clauses = list_concat_copy(clauses, other_clauses);
1135 :
1136 144 : if (!predicate_implied_by(index->indpred, all_clauses, false))
1137 96 : continue; /* can't use it at all */
1138 :
1139 48 : if (!predicate_implied_by(index->indpred, other_clauses, false))
1140 48 : useful_predicate = true;
1141 : }
1142 : }
1143 :
1144 : /*
1145 : * Identify the restriction clauses that can match the index.
1146 : */
1147 947104 : MemSet(&clauseset, 0, sizeof(clauseset));
1148 27856 : match_clauses_to_index(root, clauses, index, &clauseset);
1149 :
1150 : /*
1151 : * If no matches so far, and the index predicate isn't useful, we
1152 : * don't want it.
1153 : */
1154 27856 : if (!clauseset.nonempty && !useful_predicate)
1155 24868 : continue;
1156 :
1157 : /*
1158 : * Add "other" restriction clauses to the clauseset.
1159 : */
1160 2988 : match_clauses_to_index(root, other_clauses, index, &clauseset);
1161 :
1162 : /*
1163 : * Construct paths if possible.
1164 : */
1165 2988 : indexpaths = build_index_paths(root, rel,
1166 : index, &clauseset,
1167 : useful_predicate,
1168 : ST_BITMAPSCAN,
1169 : NULL);
1170 2988 : result = list_concat(result, indexpaths);
1171 : }
1172 :
1173 11106 : return result;
1174 : }
1175 :
1176 : /*
1177 : * Utility structure used to group similar OR-clause arguments in
1178 : * group_similar_or_args(). It represents information about the OR-clause
1179 : * argument and its matching index key.
1180 : */
1181 : typedef struct
1182 : {
1183 : int indexnum; /* index of the matching index, or -1 if no
1184 : * matching index */
1185 : int colnum; /* index of the matching column, or -1 if no
1186 : * matching index */
1187 : Oid opno; /* OID of the OpClause operator, or InvalidOid
1188 : * if not an OpExpr */
1189 : Oid inputcollid; /* OID of the OpClause input collation */
1190 : int argindex; /* index of the clause in the list of
1191 : * arguments */
1192 : int groupindex; /* value of argindex for the fist clause in
1193 : * the group of similar clauses */
1194 : } OrArgIndexMatch;
1195 :
1196 : /*
1197 : * Comparison function for OrArgIndexMatch which provides sort order placing
1198 : * similar OR-clause arguments together.
1199 : */
1200 : static int
1201 6820 : or_arg_index_match_cmp(const void *a, const void *b)
1202 : {
1203 6820 : const OrArgIndexMatch *match_a = (const OrArgIndexMatch *) a;
1204 6820 : const OrArgIndexMatch *match_b = (const OrArgIndexMatch *) b;
1205 :
1206 6820 : if (match_a->indexnum < match_b->indexnum)
1207 1286 : return -1;
1208 5534 : else if (match_a->indexnum > match_b->indexnum)
1209 2916 : return 1;
1210 :
1211 2618 : if (match_a->colnum < match_b->colnum)
1212 830 : return -1;
1213 1788 : else if (match_a->colnum > match_b->colnum)
1214 24 : return 1;
1215 :
1216 1764 : if (match_a->opno < match_b->opno)
1217 18 : return -1;
1218 1746 : else if (match_a->opno > match_b->opno)
1219 42 : return 1;
1220 :
1221 1704 : if (match_a->inputcollid < match_b->inputcollid)
1222 0 : return -1;
1223 1704 : else if (match_a->inputcollid > match_b->inputcollid)
1224 0 : return 1;
1225 :
1226 1704 : if (match_a->argindex < match_b->argindex)
1227 1626 : return -1;
1228 78 : else if (match_a->argindex > match_b->argindex)
1229 78 : return 1;
1230 :
1231 0 : return 0;
1232 : }
1233 :
1234 : /*
1235 : * Another comparison function for OrArgIndexMatch. It sorts groups together
1236 : * using groupindex. The group items are then sorted by argindex.
1237 : */
1238 : static int
1239 6922 : or_arg_index_match_cmp_group(const void *a, const void *b)
1240 : {
1241 6922 : const OrArgIndexMatch *match_a = (const OrArgIndexMatch *) a;
1242 6922 : const OrArgIndexMatch *match_b = (const OrArgIndexMatch *) b;
1243 :
1244 6922 : if (match_a->groupindex < match_b->groupindex)
1245 3364 : return -1;
1246 3558 : else if (match_a->groupindex > match_b->groupindex)
1247 3114 : return 1;
1248 :
1249 444 : if (match_a->argindex < match_b->argindex)
1250 444 : return -1;
1251 0 : else if (match_a->argindex > match_b->argindex)
1252 0 : return 1;
1253 :
1254 0 : return 0;
1255 : }
1256 :
1257 : /*
1258 : * group_similar_or_args
1259 : * Transform incoming OR-restrictinfo into a list of sub-restrictinfos,
1260 : * each of them containing a subset of similar OR-clause arguments from
1261 : * the source rinfo.
1262 : *
1263 : * Similar OR-clause arguments are of the form "indexkey op constant" having
1264 : * the same indexkey, operator, and collation. Constant may comprise either
1265 : * Const or Param. It may be employed later, during the
1266 : * match_clause_to_indexcol() to transform the whole OR-sub-rinfo to an SAOP
1267 : * clause.
1268 : *
1269 : * Returns the processed list of OR-clause arguments.
1270 : */
1271 : static List *
1272 9414 : group_similar_or_args(PlannerInfo *root, RelOptInfo *rel, RestrictInfo *rinfo)
1273 : {
1274 : int n;
1275 : int i;
1276 : int group_start;
1277 : OrArgIndexMatch *matches;
1278 9414 : bool matched = false;
1279 : ListCell *lc;
1280 : ListCell *lc2;
1281 : List *orargs;
1282 9414 : List *result = NIL;
1283 9414 : Index relid = rel->relid;
1284 :
1285 : Assert(IsA(rinfo->orclause, BoolExpr));
1286 9414 : orargs = ((BoolExpr *) rinfo->orclause)->args;
1287 9414 : n = list_length(orargs);
1288 :
1289 : /*
1290 : * To avoid N^2 behavior, take utility pass along the list of OR-clause
1291 : * arguments. For each argument, fill the OrArgIndexMatch structure,
1292 : * which will be used to sort these arguments at the next step.
1293 : */
1294 9414 : i = -1;
1295 9414 : matches = (OrArgIndexMatch *) palloc(sizeof(OrArgIndexMatch) * n);
1296 31520 : foreach(lc, orargs)
1297 : {
1298 22106 : Node *arg = lfirst(lc);
1299 : RestrictInfo *argrinfo;
1300 : OpExpr *clause;
1301 : Oid opno;
1302 : Node *leftop,
1303 : *rightop;
1304 : Node *nonConstExpr;
1305 : int indexnum;
1306 : int colnum;
1307 :
1308 22106 : i++;
1309 22106 : matches[i].argindex = i;
1310 22106 : matches[i].groupindex = i;
1311 22106 : matches[i].indexnum = -1;
1312 22106 : matches[i].colnum = -1;
1313 22106 : matches[i].opno = InvalidOid;
1314 22106 : matches[i].inputcollid = InvalidOid;
1315 :
1316 22106 : if (!IsA(arg, RestrictInfo))
1317 2216 : continue;
1318 :
1319 19890 : argrinfo = castNode(RestrictInfo, arg);
1320 :
1321 : /* Only operator clauses can match */
1322 19890 : if (!IsA(argrinfo->clause, OpExpr))
1323 8888 : continue;
1324 :
1325 11002 : clause = (OpExpr *) argrinfo->clause;
1326 11002 : opno = clause->opno;
1327 :
1328 : /* Only binary operators can match */
1329 11002 : if (list_length(clause->args) != 2)
1330 0 : continue;
1331 :
1332 : /*
1333 : * Ignore any RelabelType node above the operands. This is needed to
1334 : * be able to apply indexscanning in binary-compatible-operator cases.
1335 : * Note: we can assume there is at most one RelabelType node;
1336 : * eval_const_expressions() will have simplified if more than one.
1337 : */
1338 11002 : leftop = get_leftop(clause);
1339 11002 : if (IsA(leftop, RelabelType))
1340 204 : leftop = (Node *) ((RelabelType *) leftop)->arg;
1341 :
1342 11002 : rightop = get_rightop(clause);
1343 11002 : if (IsA(rightop, RelabelType))
1344 764 : rightop = (Node *) ((RelabelType *) rightop)->arg;
1345 :
1346 : /*
1347 : * Check for clauses of the form: (indexkey operator constant) or
1348 : * (constant operator indexkey). But we don't know a particular index
1349 : * yet. Therefore, we try to distinguish the potential index key and
1350 : * constant first, then search for a matching index key among all
1351 : * indexes.
1352 : */
1353 11002 : if (bms_is_member(relid, argrinfo->right_relids) &&
1354 1880 : !bms_is_member(relid, argrinfo->left_relids) &&
1355 1808 : !contain_volatile_functions(leftop))
1356 : {
1357 1808 : opno = get_commutator(opno);
1358 :
1359 1808 : if (!OidIsValid(opno))
1360 : {
1361 : /* commutator doesn't exist, we can't reverse the order */
1362 0 : continue;
1363 : }
1364 1808 : nonConstExpr = rightop;
1365 : }
1366 9194 : else if (bms_is_member(relid, argrinfo->left_relids) &&
1367 7344 : !bms_is_member(relid, argrinfo->right_relids) &&
1368 7272 : !contain_volatile_functions(rightop))
1369 : {
1370 7272 : nonConstExpr = leftop;
1371 : }
1372 : else
1373 : {
1374 1922 : continue;
1375 : }
1376 :
1377 : /*
1378 : * Match non-constant part to the index key. It's possible that a
1379 : * single non-constant part matches multiple index keys. It's OK, we
1380 : * just stop with first matching index key. Given that this choice is
1381 : * determined the same for every clause, we will group similar clauses
1382 : * together anyway.
1383 : */
1384 9080 : indexnum = 0;
1385 19772 : foreach(lc2, rel->indexlist)
1386 : {
1387 16174 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc2);
1388 :
1389 : /*
1390 : * Ignore index if it doesn't support bitmap scans or SAOP
1391 : * clauses.
1392 : */
1393 16174 : if (!index->amhasgetbitmap || !index->amsearcharray)
1394 54 : continue;
1395 :
1396 36582 : for (colnum = 0; colnum < index->nkeycolumns; colnum++)
1397 : {
1398 25944 : if (match_index_to_operand(nonConstExpr, colnum, index))
1399 : {
1400 5482 : matches[i].indexnum = indexnum;
1401 5482 : matches[i].colnum = colnum;
1402 5482 : matches[i].opno = opno;
1403 5482 : matches[i].inputcollid = clause->inputcollid;
1404 5482 : matched = true;
1405 5482 : break;
1406 : }
1407 : }
1408 :
1409 : /*
1410 : * Stop looping through the indexes, if we managed to match
1411 : * nonConstExpr to any index column.
1412 : */
1413 16120 : if (matches[i].indexnum >= 0)
1414 5482 : break;
1415 10638 : indexnum++;
1416 : }
1417 : }
1418 :
1419 : /*
1420 : * Fast-path check: if no clause is matching to the index column, we can
1421 : * just give up at this stage and return the clause list as-is.
1422 : */
1423 9414 : if (!matched)
1424 : {
1425 5630 : pfree(matches);
1426 5630 : return orargs;
1427 : }
1428 :
1429 : /*
1430 : * Sort clauses to make similar clauses go together. But at the same
1431 : * time, we would like to change the order of clauses as little as
1432 : * possible. To do so, we reorder each group of similar clauses so that
1433 : * the first item of the group stays in place, and all the other items are
1434 : * moved after it. So, if there are no similar clauses, the order of
1435 : * clauses stays the same. When there are some groups, required
1436 : * reordering happens while the rest of the clauses remain in their
1437 : * places. That is achieved by assigning a 'groupindex' to each clause:
1438 : * the number of the first item in the group in the original clause list.
1439 : */
1440 3784 : qsort(matches, n, sizeof(OrArgIndexMatch), or_arg_index_match_cmp);
1441 :
1442 : /* Assign groupindex to the sorted clauses */
1443 9086 : for (i = 1; i < n; i++)
1444 : {
1445 : /*
1446 : * When two clauses are similar and should belong to the same group,
1447 : * copy the 'groupindex' from the previous clause. Given we are
1448 : * considering clauses in direct order, all the clauses would have a
1449 : * 'groupindex' equal to the 'groupindex' of the first clause in the
1450 : * group.
1451 : */
1452 5302 : if (matches[i].indexnum == matches[i - 1].indexnum &&
1453 2456 : matches[i].colnum == matches[i - 1].colnum &&
1454 1614 : matches[i].opno == matches[i - 1].opno &&
1455 1566 : matches[i].inputcollid == matches[i - 1].inputcollid &&
1456 1566 : matches[i].indexnum != -1)
1457 444 : matches[i].groupindex = matches[i - 1].groupindex;
1458 : }
1459 :
1460 : /* Re-sort clauses first by groupindex then by argindex */
1461 3784 : qsort(matches, n, sizeof(OrArgIndexMatch), or_arg_index_match_cmp_group);
1462 :
1463 : /*
1464 : * Group similar clauses into single sub-restrictinfo. Side effect: the
1465 : * resulting list of restrictions will be sorted by indexnum and colnum.
1466 : */
1467 3784 : group_start = 0;
1468 12870 : for (i = 1; i <= n; i++)
1469 : {
1470 : /* Check if it's a group boundary */
1471 9086 : if (group_start >= 0 &&
1472 5302 : (i == n ||
1473 5302 : matches[i].indexnum != matches[group_start].indexnum ||
1474 2384 : matches[i].colnum != matches[group_start].colnum ||
1475 1560 : matches[i].opno != matches[group_start].opno ||
1476 1518 : matches[i].inputcollid != matches[group_start].inputcollid ||
1477 1518 : matches[i].indexnum == -1))
1478 : {
1479 : /*
1480 : * One clause in group: add it "as is" to the upper-level OR.
1481 : */
1482 8642 : if (i - group_start == 1)
1483 : {
1484 8330 : result = lappend(result,
1485 : list_nth(orargs,
1486 8330 : matches[group_start].argindex));
1487 : }
1488 : else
1489 : {
1490 : /*
1491 : * Two or more clauses in a group: create a nested OR.
1492 : */
1493 312 : List *args = NIL;
1494 312 : List *rargs = NIL;
1495 : RestrictInfo *subrinfo;
1496 : int j;
1497 :
1498 : Assert(i - group_start >= 2);
1499 :
1500 : /* Construct the list of nested OR arguments */
1501 1068 : for (j = group_start; j < i; j++)
1502 : {
1503 756 : Node *arg = list_nth(orargs, matches[j].argindex);
1504 :
1505 756 : rargs = lappend(rargs, arg);
1506 756 : if (IsA(arg, RestrictInfo))
1507 756 : args = lappend(args, ((RestrictInfo *) arg)->clause);
1508 : else
1509 0 : args = lappend(args, arg);
1510 : }
1511 :
1512 : /* Construct the nested OR and wrap it with RestrictInfo */
1513 312 : subrinfo = make_plain_restrictinfo(root,
1514 : make_orclause(args),
1515 : make_orclause(rargs),
1516 312 : rinfo->is_pushed_down,
1517 312 : rinfo->has_clone,
1518 312 : rinfo->is_clone,
1519 312 : rinfo->pseudoconstant,
1520 : rinfo->security_level,
1521 : rinfo->required_relids,
1522 : rinfo->incompatible_relids,
1523 : rinfo->outer_relids);
1524 312 : result = lappend(result, subrinfo);
1525 : }
1526 :
1527 8642 : group_start = i;
1528 : }
1529 : }
1530 3784 : pfree(matches);
1531 3784 : return result;
1532 : }
1533 :
1534 : /*
1535 : * make_bitmap_paths_for_or_group
1536 : * Generate bitmap paths for a group of similar OR-clause arguments
1537 : * produced by group_similar_or_args().
1538 : *
1539 : * This function considers two cases: (1) matching a group of clauses to
1540 : * the index as a whole, and (2) matching the individual clauses one-by-one.
1541 : * (1) typically comprises an optimal solution. If not, (2) typically
1542 : * comprises fair alternative.
1543 : *
1544 : * Ideally, we could consider all arbitrary splits of arguments into
1545 : * subgroups, but that could lead to unacceptable computational complexity.
1546 : * This is why we only consider two cases of above.
1547 : */
1548 : static List *
1549 306 : make_bitmap_paths_for_or_group(PlannerInfo *root, RelOptInfo *rel,
1550 : RestrictInfo *ri, List *other_clauses)
1551 : {
1552 306 : List *jointlist = NIL;
1553 306 : List *splitlist = NIL;
1554 : ListCell *lc;
1555 : List *orargs;
1556 306 : List *args = ((BoolExpr *) ri->orclause)->args;
1557 306 : Cost jointcost = 0.0,
1558 306 : splitcost = 0.0;
1559 : Path *bitmapqual;
1560 : List *indlist;
1561 :
1562 : /*
1563 : * First, try to match the whole group to the one index.
1564 : */
1565 306 : orargs = list_make1(ri);
1566 306 : indlist = build_paths_for_OR(root, rel,
1567 : orargs,
1568 : other_clauses);
1569 306 : if (indlist != NIL)
1570 : {
1571 300 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1572 300 : jointcost = bitmapqual->total_cost;
1573 300 : jointlist = list_make1(bitmapqual);
1574 : }
1575 :
1576 : /*
1577 : * If we manage to find a bitmap scan, which uses the group of OR-clause
1578 : * arguments as a whole, we can skip matching OR-clause arguments
1579 : * one-by-one as long as there are no other clauses, which can bring more
1580 : * efficiency to one-by-one case.
1581 : */
1582 306 : if (jointlist != NIL && other_clauses == NIL)
1583 84 : return jointlist;
1584 :
1585 : /*
1586 : * Also try to match all containing clauses one-by-one.
1587 : */
1588 768 : foreach(lc, args)
1589 : {
1590 552 : orargs = list_make1(lfirst(lc));
1591 :
1592 552 : indlist = build_paths_for_OR(root, rel,
1593 : orargs,
1594 : other_clauses);
1595 :
1596 552 : if (indlist == NIL)
1597 : {
1598 6 : splitlist = NIL;
1599 6 : break;
1600 : }
1601 :
1602 546 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1603 546 : splitcost += bitmapqual->total_cost;
1604 546 : splitlist = lappend(splitlist, bitmapqual);
1605 : }
1606 :
1607 : /*
1608 : * Pick the best option.
1609 : */
1610 222 : if (splitlist == NIL)
1611 6 : return jointlist;
1612 216 : else if (jointlist == NIL)
1613 0 : return splitlist;
1614 : else
1615 216 : return (jointcost < splitcost) ? jointlist : splitlist;
1616 : }
1617 :
1618 :
1619 : /*
1620 : * generate_bitmap_or_paths
1621 : * Look through the list of clauses to find OR clauses, and generate
1622 : * a BitmapOrPath for each one we can handle that way. Return a list
1623 : * of the generated BitmapOrPaths.
1624 : *
1625 : * other_clauses is a list of additional clauses that can be assumed true
1626 : * for the purpose of generating indexquals, but are not to be searched for
1627 : * ORs. (See build_paths_for_OR() for motivation.)
1628 : */
1629 : static List *
1630 632960 : generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
1631 : List *clauses, List *other_clauses)
1632 : {
1633 632960 : List *result = NIL;
1634 : List *all_clauses;
1635 : ListCell *lc;
1636 :
1637 : /*
1638 : * We can use both the current and other clauses as context for
1639 : * build_paths_for_OR; no need to remove ORs from the lists.
1640 : */
1641 632960 : all_clauses = list_concat_copy(clauses, other_clauses);
1642 :
1643 987988 : foreach(lc, clauses)
1644 : {
1645 355028 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1646 : List *pathlist;
1647 : Path *bitmapqual;
1648 : ListCell *j;
1649 : List *groupedArgs;
1650 355028 : List *inner_other_clauses = NIL;
1651 :
1652 : /* Ignore RestrictInfos that aren't ORs */
1653 355028 : if (!restriction_is_or_clause(rinfo))
1654 345614 : continue;
1655 :
1656 : /*
1657 : * We must be able to match at least one index to each of the arms of
1658 : * the OR, else we can't use it.
1659 : */
1660 9414 : pathlist = NIL;
1661 :
1662 : /*
1663 : * Group the similar OR-clause arguments into dedicated RestrictInfos,
1664 : * because each of those RestrictInfos has a chance to match the index
1665 : * as a whole.
1666 : */
1667 9414 : groupedArgs = group_similar_or_args(root, rel, rinfo);
1668 :
1669 9414 : if (groupedArgs != ((BoolExpr *) rinfo->orclause)->args)
1670 : {
1671 : /*
1672 : * Some parts of the rinfo were probably grouped. In this case,
1673 : * we have a set of sub-rinfos that together are an exact
1674 : * duplicate of rinfo. Thus, we need to remove the rinfo from
1675 : * other clauses. match_clauses_to_index detects duplicated
1676 : * iclauses by comparing pointers to original rinfos that would be
1677 : * different. So, we must delete rinfo to avoid de-facto
1678 : * duplicated clauses in the index clauses list.
1679 : */
1680 3784 : inner_other_clauses = list_delete(list_copy(all_clauses), rinfo);
1681 : }
1682 :
1683 11526 : foreach(j, groupedArgs)
1684 : {
1685 10554 : Node *orarg = (Node *) lfirst(j);
1686 : List *indlist;
1687 :
1688 : /* OR arguments should be ANDs or sub-RestrictInfos */
1689 10554 : if (is_andclause(orarg))
1690 : {
1691 1316 : List *andargs = ((BoolExpr *) orarg)->args;
1692 :
1693 1316 : indlist = build_paths_for_OR(root, rel,
1694 : andargs,
1695 : all_clauses);
1696 :
1697 : /* Recurse in case there are sub-ORs */
1698 1316 : indlist = list_concat(indlist,
1699 1316 : generate_bitmap_or_paths(root, rel,
1700 : andargs,
1701 : all_clauses));
1702 : }
1703 9238 : else if (restriction_is_or_clause(castNode(RestrictInfo, orarg)))
1704 : {
1705 306 : RestrictInfo *ri = castNode(RestrictInfo, orarg);
1706 :
1707 : /*
1708 : * Generate bitmap paths for the group of similar OR-clause
1709 : * arguments.
1710 : */
1711 306 : indlist = make_bitmap_paths_for_or_group(root,
1712 : rel, ri,
1713 : inner_other_clauses);
1714 :
1715 306 : if (indlist == NIL)
1716 : {
1717 6 : pathlist = NIL;
1718 6 : break;
1719 : }
1720 : else
1721 : {
1722 300 : pathlist = list_concat(pathlist, indlist);
1723 300 : continue;
1724 : }
1725 : }
1726 : else
1727 : {
1728 8932 : RestrictInfo *ri = castNode(RestrictInfo, orarg);
1729 : List *orargs;
1730 :
1731 8932 : orargs = list_make1(ri);
1732 :
1733 8932 : indlist = build_paths_for_OR(root, rel,
1734 : orargs,
1735 : all_clauses);
1736 : }
1737 :
1738 : /*
1739 : * If nothing matched this arm, we can't do anything with this OR
1740 : * clause.
1741 : */
1742 10248 : if (indlist == NIL)
1743 : {
1744 8436 : pathlist = NIL;
1745 8436 : break;
1746 : }
1747 :
1748 : /*
1749 : * OK, pick the most promising AND combination, and add it to
1750 : * pathlist.
1751 : */
1752 1812 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1753 1812 : pathlist = lappend(pathlist, bitmapqual);
1754 : }
1755 :
1756 9414 : if (inner_other_clauses != NIL)
1757 2098 : list_free(inner_other_clauses);
1758 :
1759 : /*
1760 : * If we have a match for every arm, then turn them into a
1761 : * BitmapOrPath, and add to result list.
1762 : */
1763 9414 : if (pathlist != NIL)
1764 : {
1765 972 : bitmapqual = (Path *) create_bitmap_or_path(root, rel, pathlist);
1766 972 : result = lappend(result, bitmapqual);
1767 : }
1768 : }
1769 :
1770 632960 : return result;
1771 : }
1772 :
1773 :
1774 : /*
1775 : * choose_bitmap_and
1776 : * Given a nonempty list of bitmap paths, AND them into one path.
1777 : *
1778 : * This is a nontrivial decision since we can legally use any subset of the
1779 : * given path set. We want to choose a good tradeoff between selectivity
1780 : * and cost of computing the bitmap.
1781 : *
1782 : * The result is either a single one of the inputs, or a BitmapAndPath
1783 : * combining multiple inputs.
1784 : */
1785 : static Path *
1786 327590 : choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel, List *paths)
1787 : {
1788 327590 : int npaths = list_length(paths);
1789 : PathClauseUsage **pathinfoarray;
1790 : PathClauseUsage *pathinfo;
1791 : List *clauselist;
1792 327590 : List *bestpaths = NIL;
1793 327590 : Cost bestcost = 0;
1794 : int i,
1795 : j;
1796 : ListCell *l;
1797 :
1798 : Assert(npaths > 0); /* else caller error */
1799 327590 : if (npaths == 1)
1800 251326 : return (Path *) linitial(paths); /* easy case */
1801 :
1802 : /*
1803 : * In theory we should consider every nonempty subset of the given paths.
1804 : * In practice that seems like overkill, given the crude nature of the
1805 : * estimates, not to mention the possible effects of higher-level AND and
1806 : * OR clauses. Moreover, it's completely impractical if there are a large
1807 : * number of paths, since the work would grow as O(2^N).
1808 : *
1809 : * As a heuristic, we first check for paths using exactly the same sets of
1810 : * WHERE clauses + index predicate conditions, and reject all but the
1811 : * cheapest-to-scan in any such group. This primarily gets rid of indexes
1812 : * that include the interesting columns but also irrelevant columns. (In
1813 : * situations where the DBA has gone overboard on creating variant
1814 : * indexes, this can make for a very large reduction in the number of
1815 : * paths considered further.)
1816 : *
1817 : * We then sort the surviving paths with the cheapest-to-scan first, and
1818 : * for each path, consider using that path alone as the basis for a bitmap
1819 : * scan. Then we consider bitmap AND scans formed from that path plus
1820 : * each subsequent (higher-cost) path, adding on a subsequent path if it
1821 : * results in a reduction in the estimated total scan cost. This means we
1822 : * consider about O(N^2) rather than O(2^N) path combinations, which is
1823 : * quite tolerable, especially given than N is usually reasonably small
1824 : * because of the prefiltering step. The cheapest of these is returned.
1825 : *
1826 : * We will only consider AND combinations in which no two indexes use the
1827 : * same WHERE clause. This is a bit of a kluge: it's needed because
1828 : * costsize.c and clausesel.c aren't very smart about redundant clauses.
1829 : * They will usually double-count the redundant clauses, producing a
1830 : * too-small selectivity that makes a redundant AND step look like it
1831 : * reduces the total cost. Perhaps someday that code will be smarter and
1832 : * we can remove this limitation. (But note that this also defends
1833 : * against flat-out duplicate input paths, which can happen because
1834 : * match_join_clauses_to_index will find the same OR join clauses that
1835 : * extract_restriction_or_clauses has pulled OR restriction clauses out
1836 : * of.)
1837 : *
1838 : * For the same reason, we reject AND combinations in which an index
1839 : * predicate clause duplicates another clause. Here we find it necessary
1840 : * to be even stricter: we'll reject a partial index if any of its
1841 : * predicate clauses are implied by the set of WHERE clauses and predicate
1842 : * clauses used so far. This covers cases such as a condition "x = 42"
1843 : * used with a plain index, followed by a clauseless scan of a partial
1844 : * index "WHERE x >= 40 AND x < 50". The partial index has been accepted
1845 : * only because "x = 42" was present, and so allowing it would partially
1846 : * double-count selectivity. (We could use predicate_implied_by on
1847 : * regular qual clauses too, to have a more intelligent, but much more
1848 : * expensive, check for redundancy --- but in most cases simple equality
1849 : * seems to suffice.)
1850 : */
1851 :
1852 : /*
1853 : * Extract clause usage info and detect any paths that use exactly the
1854 : * same set of clauses; keep only the cheapest-to-scan of any such groups.
1855 : * The surviving paths are put into an array for qsort'ing.
1856 : */
1857 : pathinfoarray = (PathClauseUsage **)
1858 76264 : palloc(npaths * sizeof(PathClauseUsage *));
1859 76264 : clauselist = NIL;
1860 76264 : npaths = 0;
1861 251384 : foreach(l, paths)
1862 : {
1863 175120 : Path *ipath = (Path *) lfirst(l);
1864 :
1865 175120 : pathinfo = classify_index_clause_usage(ipath, &clauselist);
1866 :
1867 : /* If it's unclassifiable, treat it as distinct from all others */
1868 175120 : if (pathinfo->unclassifiable)
1869 : {
1870 0 : pathinfoarray[npaths++] = pathinfo;
1871 0 : continue;
1872 : }
1873 :
1874 273618 : for (i = 0; i < npaths; i++)
1875 : {
1876 242860 : if (!pathinfoarray[i]->unclassifiable &&
1877 121430 : bms_equal(pathinfo->clauseids, pathinfoarray[i]->clauseids))
1878 22932 : break;
1879 : }
1880 175120 : if (i < npaths)
1881 : {
1882 : /* duplicate clauseids, keep the cheaper one */
1883 : Cost ncost;
1884 : Cost ocost;
1885 : Selectivity nselec;
1886 : Selectivity oselec;
1887 :
1888 22932 : cost_bitmap_tree_node(pathinfo->path, &ncost, &nselec);
1889 22932 : cost_bitmap_tree_node(pathinfoarray[i]->path, &ocost, &oselec);
1890 22932 : if (ncost < ocost)
1891 5190 : pathinfoarray[i] = pathinfo;
1892 : }
1893 : else
1894 : {
1895 : /* not duplicate clauseids, add to array */
1896 152188 : pathinfoarray[npaths++] = pathinfo;
1897 : }
1898 : }
1899 :
1900 : /* If only one surviving path, we're done */
1901 76264 : if (npaths == 1)
1902 14492 : return pathinfoarray[0]->path;
1903 :
1904 : /* Sort the surviving paths by index access cost */
1905 61772 : qsort(pathinfoarray, npaths, sizeof(PathClauseUsage *),
1906 : path_usage_comparator);
1907 :
1908 : /*
1909 : * For each surviving index, consider it as an "AND group leader", and see
1910 : * whether adding on any of the later indexes results in an AND path with
1911 : * cheaper total cost than before. Then take the cheapest AND group.
1912 : *
1913 : * Note: paths that are either clauseless or unclassifiable will have
1914 : * empty clauseids, so that they will not be rejected by the clauseids
1915 : * filter here, nor will they cause later paths to be rejected by it.
1916 : */
1917 199468 : for (i = 0; i < npaths; i++)
1918 : {
1919 : Cost costsofar;
1920 : List *qualsofar;
1921 : Bitmapset *clauseidsofar;
1922 :
1923 137696 : pathinfo = pathinfoarray[i];
1924 137696 : paths = list_make1(pathinfo->path);
1925 137696 : costsofar = bitmap_scan_cost_est(root, rel, pathinfo->path);
1926 137696 : qualsofar = list_concat_copy(pathinfo->quals, pathinfo->preds);
1927 137696 : clauseidsofar = bms_copy(pathinfo->clauseids);
1928 :
1929 228210 : for (j = i + 1; j < npaths; j++)
1930 : {
1931 : Cost newcost;
1932 :
1933 90514 : pathinfo = pathinfoarray[j];
1934 : /* Check for redundancy */
1935 90514 : if (bms_overlap(pathinfo->clauseids, clauseidsofar))
1936 41508 : continue; /* consider it redundant */
1937 49006 : if (pathinfo->preds)
1938 : {
1939 24 : bool redundant = false;
1940 :
1941 : /* we check each predicate clause separately */
1942 24 : foreach(l, pathinfo->preds)
1943 : {
1944 24 : Node *np = (Node *) lfirst(l);
1945 :
1946 24 : if (predicate_implied_by(list_make1(np), qualsofar, false))
1947 : {
1948 24 : redundant = true;
1949 24 : break; /* out of inner foreach loop */
1950 : }
1951 : }
1952 24 : if (redundant)
1953 24 : continue;
1954 : }
1955 : /* tentatively add new path to paths, so we can estimate cost */
1956 48982 : paths = lappend(paths, pathinfo->path);
1957 48982 : newcost = bitmap_and_cost_est(root, rel, paths);
1958 48982 : if (newcost < costsofar)
1959 : {
1960 : /* keep new path in paths, update subsidiary variables */
1961 272 : costsofar = newcost;
1962 272 : qualsofar = list_concat(qualsofar, pathinfo->quals);
1963 272 : qualsofar = list_concat(qualsofar, pathinfo->preds);
1964 272 : clauseidsofar = bms_add_members(clauseidsofar,
1965 272 : pathinfo->clauseids);
1966 : }
1967 : else
1968 : {
1969 : /* reject new path, remove it from paths list */
1970 48710 : paths = list_truncate(paths, list_length(paths) - 1);
1971 : }
1972 : }
1973 :
1974 : /* Keep the cheapest AND-group (or singleton) */
1975 137696 : if (i == 0 || costsofar < bestcost)
1976 : {
1977 65520 : bestpaths = paths;
1978 65520 : bestcost = costsofar;
1979 : }
1980 :
1981 : /* some easy cleanup (we don't try real hard though) */
1982 137696 : list_free(qualsofar);
1983 : }
1984 :
1985 61772 : if (list_length(bestpaths) == 1)
1986 61524 : return (Path *) linitial(bestpaths); /* no need for AND */
1987 248 : return (Path *) create_bitmap_and_path(root, rel, bestpaths);
1988 : }
1989 :
1990 : /* qsort comparator to sort in increasing index access cost order */
1991 : static int
1992 85464 : path_usage_comparator(const void *a, const void *b)
1993 : {
1994 85464 : PathClauseUsage *pa = *(PathClauseUsage *const *) a;
1995 85464 : PathClauseUsage *pb = *(PathClauseUsage *const *) b;
1996 : Cost acost;
1997 : Cost bcost;
1998 : Selectivity aselec;
1999 : Selectivity bselec;
2000 :
2001 85464 : cost_bitmap_tree_node(pa->path, &acost, &aselec);
2002 85464 : cost_bitmap_tree_node(pb->path, &bcost, &bselec);
2003 :
2004 : /*
2005 : * If costs are the same, sort by selectivity.
2006 : */
2007 85464 : if (acost < bcost)
2008 48594 : return -1;
2009 36870 : if (acost > bcost)
2010 25482 : return 1;
2011 :
2012 11388 : if (aselec < bselec)
2013 3926 : return -1;
2014 7462 : if (aselec > bselec)
2015 1094 : return 1;
2016 :
2017 6368 : return 0;
2018 : }
2019 :
2020 : /*
2021 : * Estimate the cost of actually executing a bitmap scan with a single
2022 : * index path (which could be a BitmapAnd or BitmapOr node).
2023 : */
2024 : static Cost
2025 186678 : bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel, Path *ipath)
2026 : {
2027 : BitmapHeapPath bpath;
2028 :
2029 : /* Set up a dummy BitmapHeapPath */
2030 186678 : bpath.path.type = T_BitmapHeapPath;
2031 186678 : bpath.path.pathtype = T_BitmapHeapScan;
2032 186678 : bpath.path.parent = rel;
2033 186678 : bpath.path.pathtarget = rel->reltarget;
2034 186678 : bpath.path.param_info = ipath->param_info;
2035 186678 : bpath.path.pathkeys = NIL;
2036 186678 : bpath.bitmapqual = ipath;
2037 :
2038 : /*
2039 : * Check the cost of temporary path without considering parallelism.
2040 : * Parallel bitmap heap path will be considered at later stage.
2041 : */
2042 186678 : bpath.path.parallel_workers = 0;
2043 :
2044 : /* Now we can do cost_bitmap_heap_scan */
2045 186678 : cost_bitmap_heap_scan(&bpath.path, root, rel,
2046 : bpath.path.param_info,
2047 : ipath,
2048 : get_loop_count(root, rel->relid,
2049 186678 : PATH_REQ_OUTER(ipath)));
2050 :
2051 186678 : return bpath.path.total_cost;
2052 : }
2053 :
2054 : /*
2055 : * Estimate the cost of actually executing a BitmapAnd scan with the given
2056 : * inputs.
2057 : */
2058 : static Cost
2059 48982 : bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel, List *paths)
2060 : {
2061 : BitmapAndPath *apath;
2062 :
2063 : /*
2064 : * Might as well build a real BitmapAndPath here, as the work is slightly
2065 : * too complicated to be worth repeating just to save one palloc.
2066 : */
2067 48982 : apath = create_bitmap_and_path(root, rel, paths);
2068 :
2069 48982 : return bitmap_scan_cost_est(root, rel, (Path *) apath);
2070 : }
2071 :
2072 :
2073 : /*
2074 : * classify_index_clause_usage
2075 : * Construct a PathClauseUsage struct describing the WHERE clauses and
2076 : * index predicate clauses used by the given indexscan path.
2077 : * We consider two clauses the same if they are equal().
2078 : *
2079 : * At some point we might want to migrate this info into the Path data
2080 : * structure proper, but for the moment it's only needed within
2081 : * choose_bitmap_and().
2082 : *
2083 : * *clauselist is used and expanded as needed to identify all the distinct
2084 : * clauses seen across successive calls. Caller must initialize it to NIL
2085 : * before first call of a set.
2086 : */
2087 : static PathClauseUsage *
2088 175120 : classify_index_clause_usage(Path *path, List **clauselist)
2089 : {
2090 : PathClauseUsage *result;
2091 : Bitmapset *clauseids;
2092 : ListCell *lc;
2093 :
2094 175120 : result = (PathClauseUsage *) palloc(sizeof(PathClauseUsage));
2095 175120 : result->path = path;
2096 :
2097 : /* Recursively find the quals and preds used by the path */
2098 175120 : result->quals = NIL;
2099 175120 : result->preds = NIL;
2100 175120 : find_indexpath_quals(path, &result->quals, &result->preds);
2101 :
2102 : /*
2103 : * Some machine-generated queries have outlandish numbers of qual clauses.
2104 : * To avoid getting into O(N^2) behavior even in this preliminary
2105 : * classification step, we want to limit the number of entries we can
2106 : * accumulate in *clauselist. Treat any path with more than 100 quals +
2107 : * preds as unclassifiable, which will cause calling code to consider it
2108 : * distinct from all other paths.
2109 : */
2110 175120 : if (list_length(result->quals) + list_length(result->preds) > 100)
2111 : {
2112 0 : result->clauseids = NULL;
2113 0 : result->unclassifiable = true;
2114 0 : return result;
2115 : }
2116 :
2117 : /* Build up a bitmapset representing the quals and preds */
2118 175120 : clauseids = NULL;
2119 400982 : foreach(lc, result->quals)
2120 : {
2121 225862 : Node *node = (Node *) lfirst(lc);
2122 :
2123 225862 : clauseids = bms_add_member(clauseids,
2124 : find_list_position(node, clauselist));
2125 : }
2126 175414 : foreach(lc, result->preds)
2127 : {
2128 294 : Node *node = (Node *) lfirst(lc);
2129 :
2130 294 : clauseids = bms_add_member(clauseids,
2131 : find_list_position(node, clauselist));
2132 : }
2133 175120 : result->clauseids = clauseids;
2134 175120 : result->unclassifiable = false;
2135 :
2136 175120 : return result;
2137 : }
2138 :
2139 :
2140 : /*
2141 : * find_indexpath_quals
2142 : *
2143 : * Given the Path structure for a plain or bitmap indexscan, extract lists
2144 : * of all the index clauses and index predicate conditions used in the Path.
2145 : * These are appended to the initial contents of *quals and *preds (hence
2146 : * caller should initialize those to NIL).
2147 : *
2148 : * Note we are not trying to produce an accurate representation of the AND/OR
2149 : * semantics of the Path, but just find out all the base conditions used.
2150 : *
2151 : * The result lists contain pointers to the expressions used in the Path,
2152 : * but all the list cells are freshly built, so it's safe to destructively
2153 : * modify the lists (eg, by concat'ing with other lists).
2154 : */
2155 : static void
2156 177238 : find_indexpath_quals(Path *bitmapqual, List **quals, List **preds)
2157 : {
2158 177238 : if (IsA(bitmapqual, BitmapAndPath))
2159 : {
2160 0 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
2161 : ListCell *l;
2162 :
2163 0 : foreach(l, apath->bitmapquals)
2164 : {
2165 0 : find_indexpath_quals((Path *) lfirst(l), quals, preds);
2166 : }
2167 : }
2168 177238 : else if (IsA(bitmapqual, BitmapOrPath))
2169 : {
2170 1158 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
2171 : ListCell *l;
2172 :
2173 3276 : foreach(l, opath->bitmapquals)
2174 : {
2175 2118 : find_indexpath_quals((Path *) lfirst(l), quals, preds);
2176 : }
2177 : }
2178 176080 : else if (IsA(bitmapqual, IndexPath))
2179 : {
2180 176080 : IndexPath *ipath = (IndexPath *) bitmapqual;
2181 : ListCell *l;
2182 :
2183 401942 : foreach(l, ipath->indexclauses)
2184 : {
2185 225862 : IndexClause *iclause = (IndexClause *) lfirst(l);
2186 :
2187 225862 : *quals = lappend(*quals, iclause->rinfo->clause);
2188 : }
2189 176080 : *preds = list_concat(*preds, ipath->indexinfo->indpred);
2190 : }
2191 : else
2192 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
2193 177238 : }
2194 :
2195 :
2196 : /*
2197 : * find_list_position
2198 : * Return the given node's position (counting from 0) in the given
2199 : * list of nodes. If it's not equal() to any existing list member,
2200 : * add it at the end, and return that position.
2201 : */
2202 : static int
2203 226156 : find_list_position(Node *node, List **nodelist)
2204 : {
2205 : int i;
2206 : ListCell *lc;
2207 :
2208 226156 : i = 0;
2209 355448 : foreach(lc, *nodelist)
2210 : {
2211 196488 : Node *oldnode = (Node *) lfirst(lc);
2212 :
2213 196488 : if (equal(node, oldnode))
2214 67196 : return i;
2215 129292 : i++;
2216 : }
2217 :
2218 158960 : *nodelist = lappend(*nodelist, node);
2219 :
2220 158960 : return i;
2221 : }
2222 :
2223 :
2224 : /*
2225 : * check_index_only
2226 : * Determine whether an index-only scan is possible for this index.
2227 : */
2228 : static bool
2229 807066 : check_index_only(RelOptInfo *rel, IndexOptInfo *index)
2230 : {
2231 : bool result;
2232 807066 : Bitmapset *attrs_used = NULL;
2233 807066 : Bitmapset *index_canreturn_attrs = NULL;
2234 : ListCell *lc;
2235 : int i;
2236 :
2237 : /* Index-only scans must be enabled */
2238 807066 : if (!enable_indexonlyscan)
2239 3686 : return false;
2240 :
2241 : /*
2242 : * Check that all needed attributes of the relation are available from the
2243 : * index.
2244 : */
2245 :
2246 : /*
2247 : * First, identify all the attributes needed for joins or final output.
2248 : * Note: we must look at rel's targetlist, not the attr_needed data,
2249 : * because attr_needed isn't computed for inheritance child rels.
2250 : */
2251 803380 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
2252 :
2253 : /*
2254 : * Add all the attributes used by restriction clauses; but consider only
2255 : * those clauses not implied by the index predicate, since ones that are
2256 : * so implied don't need to be checked explicitly in the plan.
2257 : *
2258 : * Note: attributes used only in index quals would not be needed at
2259 : * runtime either, if we are certain that the index is not lossy. However
2260 : * it'd be complicated to account for that accurately, and it doesn't
2261 : * matter in most cases, since we'd conclude that such attributes are
2262 : * available from the index anyway.
2263 : */
2264 1661492 : foreach(lc, index->indrestrictinfo)
2265 : {
2266 858112 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2267 :
2268 858112 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
2269 : }
2270 :
2271 : /*
2272 : * Construct a bitmapset of columns that the index can return back in an
2273 : * index-only scan.
2274 : */
2275 2299750 : for (i = 0; i < index->ncolumns; i++)
2276 : {
2277 1496370 : int attno = index->indexkeys[i];
2278 :
2279 : /*
2280 : * For the moment, we just ignore index expressions. It might be nice
2281 : * to do something with them, later.
2282 : */
2283 1496370 : if (attno == 0)
2284 3274 : continue;
2285 :
2286 1493096 : if (index->canreturn[i])
2287 : index_canreturn_attrs =
2288 1217946 : bms_add_member(index_canreturn_attrs,
2289 : attno - FirstLowInvalidHeapAttributeNumber);
2290 : }
2291 :
2292 : /* Do we have all the necessary attributes? */
2293 803380 : result = bms_is_subset(attrs_used, index_canreturn_attrs);
2294 :
2295 803380 : bms_free(attrs_used);
2296 803380 : bms_free(index_canreturn_attrs);
2297 :
2298 803380 : return result;
2299 : }
2300 :
2301 : /*
2302 : * get_loop_count
2303 : * Choose the loop count estimate to use for costing a parameterized path
2304 : * with the given set of outer relids.
2305 : *
2306 : * Since we produce parameterized paths before we've begun to generate join
2307 : * relations, it's impossible to predict exactly how many times a parameterized
2308 : * path will be iterated; we don't know the size of the relation that will be
2309 : * on the outside of the nestloop. However, we should try to account for
2310 : * multiple iterations somehow in costing the path. The heuristic embodied
2311 : * here is to use the rowcount of the smallest other base relation needed in
2312 : * the join clauses used by the path. (We could alternatively consider the
2313 : * largest one, but that seems too optimistic.) This is of course the right
2314 : * answer for single-other-relation cases, and it seems like a reasonable
2315 : * zero-order approximation for multiway-join cases.
2316 : *
2317 : * In addition, we check to see if the other side of each join clause is on
2318 : * the inside of some semijoin that the current relation is on the outside of.
2319 : * If so, the only way that a parameterized path could be used is if the
2320 : * semijoin RHS has been unique-ified, so we should use the number of unique
2321 : * RHS rows rather than using the relation's raw rowcount.
2322 : *
2323 : * Note: for this to work, allpaths.c must establish all baserel size
2324 : * estimates before it begins to compute paths, or at least before it
2325 : * calls create_index_paths().
2326 : */
2327 : static double
2328 1130978 : get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids)
2329 : {
2330 : double result;
2331 : int outer_relid;
2332 :
2333 : /* For a non-parameterized path, just return 1.0 quickly */
2334 1130978 : if (outer_relids == NULL)
2335 768386 : return 1.0;
2336 :
2337 362592 : result = 0.0;
2338 362592 : outer_relid = -1;
2339 735950 : while ((outer_relid = bms_next_member(outer_relids, outer_relid)) >= 0)
2340 : {
2341 : RelOptInfo *outer_rel;
2342 : double rowcount;
2343 :
2344 : /* Paranoia: ignore bogus relid indexes */
2345 373358 : if (outer_relid >= root->simple_rel_array_size)
2346 0 : continue;
2347 373358 : outer_rel = root->simple_rel_array[outer_relid];
2348 373358 : if (outer_rel == NULL)
2349 254 : continue;
2350 : Assert(outer_rel->relid == outer_relid); /* sanity check on array */
2351 :
2352 : /* Other relation could be proven empty, if so ignore */
2353 373104 : if (IS_DUMMY_REL(outer_rel))
2354 24 : continue;
2355 :
2356 : /* Otherwise, rel's rows estimate should be valid by now */
2357 : Assert(outer_rel->rows > 0);
2358 :
2359 : /* Check to see if rel is on the inside of any semijoins */
2360 373080 : rowcount = adjust_rowcount_for_semijoins(root,
2361 : cur_relid,
2362 : outer_relid,
2363 : outer_rel->rows);
2364 :
2365 : /* Remember smallest row count estimate among the outer rels */
2366 373080 : if (result == 0.0 || result > rowcount)
2367 369516 : result = rowcount;
2368 : }
2369 : /* Return 1.0 if we found no valid relations (shouldn't happen) */
2370 362592 : return (result > 0.0) ? result : 1.0;
2371 : }
2372 :
2373 : /*
2374 : * Check to see if outer_relid is on the inside of any semijoin that cur_relid
2375 : * is on the outside of. If so, replace rowcount with the estimated number of
2376 : * unique rows from the semijoin RHS (assuming that's smaller, which it might
2377 : * not be). The estimate is crude but it's the best we can do at this stage
2378 : * of the proceedings.
2379 : */
2380 : static double
2381 373080 : adjust_rowcount_for_semijoins(PlannerInfo *root,
2382 : Index cur_relid,
2383 : Index outer_relid,
2384 : double rowcount)
2385 : {
2386 : ListCell *lc;
2387 :
2388 578594 : foreach(lc, root->join_info_list)
2389 : {
2390 205514 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
2391 :
2392 212670 : if (sjinfo->jointype == JOIN_SEMI &&
2393 10146 : bms_is_member(cur_relid, sjinfo->syn_lefthand) &&
2394 2990 : bms_is_member(outer_relid, sjinfo->syn_righthand))
2395 : {
2396 : /* Estimate number of unique-ified rows */
2397 : double nraw;
2398 : double nunique;
2399 :
2400 1020 : nraw = approximate_joinrel_size(root, sjinfo->syn_righthand);
2401 1020 : nunique = estimate_num_groups(root,
2402 : sjinfo->semi_rhs_exprs,
2403 : nraw,
2404 : NULL,
2405 : NULL);
2406 1020 : if (rowcount > nunique)
2407 346 : rowcount = nunique;
2408 : }
2409 : }
2410 373080 : return rowcount;
2411 : }
2412 :
2413 : /*
2414 : * Make an approximate estimate of the size of a joinrel.
2415 : *
2416 : * We don't have enough info at this point to get a good estimate, so we
2417 : * just multiply the base relation sizes together. Fortunately, this is
2418 : * the right answer anyway for the most common case with a single relation
2419 : * on the RHS of a semijoin. Also, estimate_num_groups() has only a weak
2420 : * dependency on its input_rows argument (it basically uses it as a clamp).
2421 : * So we might be able to get a fairly decent end result even with a severe
2422 : * overestimate of the RHS's raw size.
2423 : */
2424 : static double
2425 1020 : approximate_joinrel_size(PlannerInfo *root, Relids relids)
2426 : {
2427 1020 : double rowcount = 1.0;
2428 : int relid;
2429 :
2430 1020 : relid = -1;
2431 2196 : while ((relid = bms_next_member(relids, relid)) >= 0)
2432 : {
2433 : RelOptInfo *rel;
2434 :
2435 : /* Paranoia: ignore bogus relid indexes */
2436 1176 : if (relid >= root->simple_rel_array_size)
2437 0 : continue;
2438 1176 : rel = root->simple_rel_array[relid];
2439 1176 : if (rel == NULL)
2440 0 : continue;
2441 : Assert(rel->relid == relid); /* sanity check on array */
2442 :
2443 : /* Relation could be proven empty, if so ignore */
2444 1176 : if (IS_DUMMY_REL(rel))
2445 0 : continue;
2446 :
2447 : /* Otherwise, rel's rows estimate should be valid by now */
2448 : Assert(rel->rows > 0);
2449 :
2450 : /* Accumulate product */
2451 1176 : rowcount *= rel->rows;
2452 : }
2453 1020 : return rowcount;
2454 : }
2455 :
2456 :
2457 : /****************************************************************************
2458 : * ---- ROUTINES TO CHECK QUERY CLAUSES ----
2459 : ****************************************************************************/
2460 :
2461 : /*
2462 : * match_restriction_clauses_to_index
2463 : * Identify restriction clauses for the rel that match the index.
2464 : * Matching clauses are added to *clauseset.
2465 : */
2466 : static void
2467 665384 : match_restriction_clauses_to_index(PlannerInfo *root,
2468 : IndexOptInfo *index,
2469 : IndexClauseSet *clauseset)
2470 : {
2471 : /* We can ignore clauses that are implied by the index predicate */
2472 665384 : match_clauses_to_index(root, index->indrestrictinfo, index, clauseset);
2473 665384 : }
2474 :
2475 : /*
2476 : * match_join_clauses_to_index
2477 : * Identify join clauses for the rel that match the index.
2478 : * Matching clauses are added to *clauseset.
2479 : * Also, add any potentially usable join OR clauses to *joinorclauses.
2480 : * They also might be processed by match_clause_to_index() as a whole.
2481 : */
2482 : static void
2483 665384 : match_join_clauses_to_index(PlannerInfo *root,
2484 : RelOptInfo *rel, IndexOptInfo *index,
2485 : IndexClauseSet *clauseset,
2486 : List **joinorclauses)
2487 : {
2488 : ListCell *lc;
2489 :
2490 : /* Scan the rel's join clauses */
2491 904414 : foreach(lc, rel->joininfo)
2492 : {
2493 239030 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2494 :
2495 : /* Check if clause can be moved to this rel */
2496 239030 : if (!join_clause_is_movable_to(rinfo, rel))
2497 141798 : continue;
2498 :
2499 : /*
2500 : * Potentially usable, so see if it matches the index or is an OR. Use
2501 : * list_append_unique_ptr() here to avoid possible duplicates when
2502 : * processing the same clauses with different indexes.
2503 : */
2504 97232 : if (restriction_is_or_clause(rinfo))
2505 12696 : *joinorclauses = list_append_unique_ptr(*joinorclauses, rinfo);
2506 :
2507 97232 : match_clause_to_index(root, rinfo, index, clauseset);
2508 : }
2509 665384 : }
2510 :
2511 : /*
2512 : * match_eclass_clauses_to_index
2513 : * Identify EquivalenceClass join clauses for the rel that match the index.
2514 : * Matching clauses are added to *clauseset.
2515 : */
2516 : static void
2517 665384 : match_eclass_clauses_to_index(PlannerInfo *root, IndexOptInfo *index,
2518 : IndexClauseSet *clauseset)
2519 : {
2520 : int indexcol;
2521 :
2522 : /* No work if rel is not in any such ECs */
2523 665384 : if (!index->rel->has_eclass_joins)
2524 390712 : return;
2525 :
2526 718778 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2527 : {
2528 : ec_member_matches_arg arg;
2529 : List *clauses;
2530 :
2531 : /* Generate clauses, skipping any that join to lateral_referencers */
2532 444106 : arg.index = index;
2533 444106 : arg.indexcol = indexcol;
2534 444106 : clauses = generate_implied_equalities_for_column(root,
2535 : index->rel,
2536 : ec_member_matches_indexcol,
2537 : &arg,
2538 444106 : index->rel->lateral_referencers);
2539 :
2540 : /*
2541 : * We have to check whether the results actually do match the index,
2542 : * since for non-btree indexes the EC's equality operators might not
2543 : * be in the index opclass (cf ec_member_matches_indexcol).
2544 : */
2545 444106 : match_clauses_to_index(root, clauses, index, clauseset);
2546 : }
2547 : }
2548 :
2549 : /*
2550 : * match_clauses_to_index
2551 : * Perform match_clause_to_index() for each clause in a list.
2552 : * Matching clauses are added to *clauseset.
2553 : */
2554 : static void
2555 1140334 : match_clauses_to_index(PlannerInfo *root,
2556 : List *clauses,
2557 : IndexOptInfo *index,
2558 : IndexClauseSet *clauseset)
2559 : {
2560 : ListCell *lc;
2561 :
2562 2053010 : foreach(lc, clauses)
2563 : {
2564 912676 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
2565 :
2566 912676 : match_clause_to_index(root, rinfo, index, clauseset);
2567 : }
2568 1140334 : }
2569 :
2570 : /*
2571 : * match_clause_to_index
2572 : * Test whether a qual clause can be used with an index.
2573 : *
2574 : * If the clause is usable, add an IndexClause entry for it to the appropriate
2575 : * list in *clauseset. (*clauseset must be initialized to zeroes before first
2576 : * call.)
2577 : *
2578 : * Note: in some circumstances we may find the same RestrictInfos coming from
2579 : * multiple places. Defend against redundant outputs by refusing to add a
2580 : * clause twice (pointer equality should be a good enough check for this).
2581 : *
2582 : * Note: it's possible that a badly-defined index could have multiple matching
2583 : * columns. We always select the first match if so; this avoids scenarios
2584 : * wherein we get an inflated idea of the index's selectivity by using the
2585 : * same clause multiple times with different index columns.
2586 : */
2587 : static void
2588 1009908 : match_clause_to_index(PlannerInfo *root,
2589 : RestrictInfo *rinfo,
2590 : IndexOptInfo *index,
2591 : IndexClauseSet *clauseset)
2592 : {
2593 : int indexcol;
2594 :
2595 : /*
2596 : * Never match pseudoconstants to indexes. (Normally a match could not
2597 : * happen anyway, since a pseudoconstant clause couldn't contain a Var,
2598 : * but what if someone builds an expression index on a constant? It's not
2599 : * totally unreasonable to do so with a partial index, either.)
2600 : */
2601 1009908 : if (rinfo->pseudoconstant)
2602 12948 : return;
2603 :
2604 : /*
2605 : * If clause can't be used as an indexqual because it must wait till after
2606 : * some lower-security-level restriction clause, reject it.
2607 : */
2608 996960 : if (!restriction_is_securely_promotable(rinfo, index->rel))
2609 474 : return;
2610 :
2611 : /* OK, check each index key column for a match */
2612 2200258 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2613 : {
2614 : IndexClause *iclause;
2615 : ListCell *lc;
2616 :
2617 : /* Ignore duplicates */
2618 1660506 : foreach(lc, clauseset->indexclauses[indexcol])
2619 : {
2620 73122 : iclause = (IndexClause *) lfirst(lc);
2621 :
2622 73122 : if (iclause->rinfo == rinfo)
2623 0 : return;
2624 : }
2625 :
2626 : /* OK, try to match the clause to the index column */
2627 1587384 : iclause = match_clause_to_indexcol(root,
2628 : rinfo,
2629 : indexcol,
2630 : index);
2631 1587384 : if (iclause)
2632 : {
2633 : /* Success, so record it */
2634 383612 : clauseset->indexclauses[indexcol] =
2635 383612 : lappend(clauseset->indexclauses[indexcol], iclause);
2636 383612 : clauseset->nonempty = true;
2637 383612 : return;
2638 : }
2639 : }
2640 : }
2641 :
2642 : /*
2643 : * match_clause_to_indexcol()
2644 : * Determine whether a restriction clause matches a column of an index,
2645 : * and if so, build an IndexClause node describing the details.
2646 : *
2647 : * To match an index normally, an operator clause:
2648 : *
2649 : * (1) must be in the form (indexkey op const) or (const op indexkey);
2650 : * and
2651 : * (2) must contain an operator which is in the index's operator family
2652 : * for this column; and
2653 : * (3) must match the collation of the index, if collation is relevant.
2654 : *
2655 : * Our definition of "const" is exceedingly liberal: we allow anything that
2656 : * doesn't involve a volatile function or a Var of the index's relation.
2657 : * In particular, Vars belonging to other relations of the query are
2658 : * accepted here, since a clause of that form can be used in a
2659 : * parameterized indexscan. It's the responsibility of higher code levels
2660 : * to manage restriction and join clauses appropriately.
2661 : *
2662 : * Note: we do need to check for Vars of the index's relation on the
2663 : * "const" side of the clause, since clauses like (a.f1 OP (b.f2 OP a.f3))
2664 : * are not processable by a parameterized indexscan on a.f1, whereas
2665 : * something like (a.f1 OP (b.f2 OP c.f3)) is.
2666 : *
2667 : * Presently, the executor can only deal with indexquals that have the
2668 : * indexkey on the left, so we can only use clauses that have the indexkey
2669 : * on the right if we can commute the clause to put the key on the left.
2670 : * We handle that by generating an IndexClause with the correctly-commuted
2671 : * opclause as a derived indexqual.
2672 : *
2673 : * If the index has a collation, the clause must have the same collation.
2674 : * For collation-less indexes, we assume it doesn't matter; this is
2675 : * necessary for cases like "hstore ? text", wherein hstore's operators
2676 : * don't care about collation but the clause will get marked with a
2677 : * collation anyway because of the text argument. (This logic is
2678 : * embodied in the macro IndexCollMatchesExprColl.)
2679 : *
2680 : * It is also possible to match RowCompareExpr clauses to indexes (but
2681 : * currently, only btree indexes handle this).
2682 : *
2683 : * It is also possible to match ScalarArrayOpExpr clauses to indexes, when
2684 : * the clause is of the form "indexkey op ANY (arrayconst)".
2685 : *
2686 : * It is also possible to match a list of OR clauses if it might be
2687 : * transformed into a single ScalarArrayOpExpr clause. On success,
2688 : * the returning index clause will contain a transformed clause.
2689 : *
2690 : * For boolean indexes, it is also possible to match the clause directly
2691 : * to the indexkey; or perhaps the clause is (NOT indexkey).
2692 : *
2693 : * And, last but not least, some operators and functions can be processed
2694 : * to derive (typically lossy) indexquals from a clause that isn't in
2695 : * itself indexable. If we see that any operand of an OpExpr or FuncExpr
2696 : * matches the index key, and the function has a planner support function
2697 : * attached to it, we'll invoke the support function to see if such an
2698 : * indexqual can be built.
2699 : *
2700 : * 'rinfo' is the clause to be tested (as a RestrictInfo node).
2701 : * 'indexcol' is a column number of 'index' (counting from 0).
2702 : * 'index' is the index of interest.
2703 : *
2704 : * Returns an IndexClause if the clause can be used with this index key,
2705 : * or NULL if not.
2706 : *
2707 : * NOTE: This routine always returns NULL if the clause is an AND clause.
2708 : * Higher-level routines deal with OR and AND clauses. OR clause can be
2709 : * matched as a whole by match_orclause_to_indexcol() though.
2710 : */
2711 : static IndexClause *
2712 1587384 : match_clause_to_indexcol(PlannerInfo *root,
2713 : RestrictInfo *rinfo,
2714 : int indexcol,
2715 : IndexOptInfo *index)
2716 : {
2717 : IndexClause *iclause;
2718 1587384 : Expr *clause = rinfo->clause;
2719 : Oid opfamily;
2720 :
2721 : Assert(indexcol < index->nkeycolumns);
2722 :
2723 : /*
2724 : * Historically this code has coped with NULL clauses. That's probably
2725 : * not possible anymore, but we might as well continue to cope.
2726 : */
2727 1587384 : if (clause == NULL)
2728 0 : return NULL;
2729 :
2730 : /* First check for boolean-index cases. */
2731 1587384 : opfamily = index->opfamily[indexcol];
2732 1587384 : if (IsBooleanOpfamily(opfamily))
2733 : {
2734 446 : iclause = match_boolean_index_clause(root, rinfo, indexcol, index);
2735 446 : if (iclause)
2736 294 : return iclause;
2737 : }
2738 :
2739 : /*
2740 : * Clause must be an opclause, funcclause, ScalarArrayOpExpr,
2741 : * RowCompareExpr, or OR-clause that could be converted to SAOP. Or, if
2742 : * the index supports it, we can handle IS NULL/NOT NULL clauses.
2743 : */
2744 1587090 : if (IsA(clause, OpExpr))
2745 : {
2746 1303986 : return match_opclause_to_indexcol(root, rinfo, indexcol, index);
2747 : }
2748 283104 : else if (IsA(clause, FuncExpr))
2749 : {
2750 35724 : return match_funcclause_to_indexcol(root, rinfo, indexcol, index);
2751 : }
2752 247380 : else if (IsA(clause, ScalarArrayOpExpr))
2753 : {
2754 74142 : return match_saopclause_to_indexcol(root, rinfo, indexcol, index);
2755 : }
2756 173238 : else if (IsA(clause, RowCompareExpr))
2757 : {
2758 432 : return match_rowcompare_to_indexcol(root, rinfo, indexcol, index);
2759 : }
2760 172806 : else if (restriction_is_or_clause(rinfo))
2761 : {
2762 45584 : return match_orclause_to_indexcol(root, rinfo, indexcol, index);
2763 : }
2764 127222 : else if (index->amsearchnulls && IsA(clause, NullTest))
2765 : {
2766 14874 : NullTest *nt = (NullTest *) clause;
2767 :
2768 29748 : if (!nt->argisrow &&
2769 14874 : match_index_to_operand((Node *) nt->arg, indexcol, index))
2770 : {
2771 1442 : iclause = makeNode(IndexClause);
2772 1442 : iclause->rinfo = rinfo;
2773 1442 : iclause->indexquals = list_make1(rinfo);
2774 1442 : iclause->lossy = false;
2775 1442 : iclause->indexcol = indexcol;
2776 1442 : iclause->indexcols = NIL;
2777 1442 : return iclause;
2778 : }
2779 : }
2780 :
2781 125780 : return NULL;
2782 : }
2783 :
2784 : /*
2785 : * IsBooleanOpfamily
2786 : * Detect whether an opfamily supports boolean equality as an operator.
2787 : *
2788 : * If the opfamily OID is in the range of built-in objects, we can rely
2789 : * on hard-wired knowledge of which built-in opfamilies support this.
2790 : * For extension opfamilies, there's no choice but to do a catcache lookup.
2791 : */
2792 : static bool
2793 2160972 : IsBooleanOpfamily(Oid opfamily)
2794 : {
2795 2160972 : if (opfamily < FirstNormalObjectId)
2796 2157766 : return IsBuiltinBooleanOpfamily(opfamily);
2797 : else
2798 3206 : return op_in_opfamily(BooleanEqualOperator, opfamily);
2799 : }
2800 :
2801 : /*
2802 : * match_boolean_index_clause
2803 : * Recognize restriction clauses that can be matched to a boolean index.
2804 : *
2805 : * The idea here is that, for an index on a boolean column that supports the
2806 : * BooleanEqualOperator, we can transform a plain reference to the indexkey
2807 : * into "indexkey = true", or "NOT indexkey" into "indexkey = false", etc,
2808 : * so as to make the expression indexable using the index's "=" operator.
2809 : * Since Postgres 8.1, we must do this because constant simplification does
2810 : * the reverse transformation; without this code there'd be no way to use
2811 : * such an index at all.
2812 : *
2813 : * This should be called only when IsBooleanOpfamily() recognizes the
2814 : * index's operator family. We check to see if the clause matches the
2815 : * index's key, and if so, build a suitable IndexClause.
2816 : */
2817 : static IndexClause *
2818 1706 : match_boolean_index_clause(PlannerInfo *root,
2819 : RestrictInfo *rinfo,
2820 : int indexcol,
2821 : IndexOptInfo *index)
2822 : {
2823 1706 : Node *clause = (Node *) rinfo->clause;
2824 1706 : Expr *op = NULL;
2825 :
2826 : /* Direct match? */
2827 1706 : if (match_index_to_operand(clause, indexcol, index))
2828 : {
2829 : /* convert to indexkey = TRUE */
2830 274 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2831 : (Expr *) clause,
2832 274 : (Expr *) makeBoolConst(true, false),
2833 : InvalidOid, InvalidOid);
2834 : }
2835 : /* NOT clause? */
2836 1432 : else if (is_notclause(clause))
2837 : {
2838 1208 : Node *arg = (Node *) get_notclausearg((Expr *) clause);
2839 :
2840 1208 : if (match_index_to_operand(arg, indexcol, index))
2841 : {
2842 : /* convert to indexkey = FALSE */
2843 1208 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2844 : (Expr *) arg,
2845 1208 : (Expr *) makeBoolConst(false, false),
2846 : InvalidOid, InvalidOid);
2847 : }
2848 : }
2849 :
2850 : /*
2851 : * Since we only consider clauses at top level of WHERE, we can convert
2852 : * indexkey IS TRUE and indexkey IS FALSE to index searches as well. The
2853 : * different meaning for NULL isn't important.
2854 : */
2855 224 : else if (clause && IsA(clause, BooleanTest))
2856 : {
2857 36 : BooleanTest *btest = (BooleanTest *) clause;
2858 36 : Node *arg = (Node *) btest->arg;
2859 :
2860 54 : if (btest->booltesttype == IS_TRUE &&
2861 18 : match_index_to_operand(arg, indexcol, index))
2862 : {
2863 : /* convert to indexkey = TRUE */
2864 18 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2865 : (Expr *) arg,
2866 18 : (Expr *) makeBoolConst(true, false),
2867 : InvalidOid, InvalidOid);
2868 : }
2869 36 : else if (btest->booltesttype == IS_FALSE &&
2870 18 : match_index_to_operand(arg, indexcol, index))
2871 : {
2872 : /* convert to indexkey = FALSE */
2873 18 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2874 : (Expr *) arg,
2875 18 : (Expr *) makeBoolConst(false, false),
2876 : InvalidOid, InvalidOid);
2877 : }
2878 : }
2879 :
2880 : /*
2881 : * If we successfully made an operator clause from the given qual, we must
2882 : * wrap it in an IndexClause. It's not lossy.
2883 : */
2884 1706 : if (op)
2885 : {
2886 1518 : IndexClause *iclause = makeNode(IndexClause);
2887 :
2888 1518 : iclause->rinfo = rinfo;
2889 1518 : iclause->indexquals = list_make1(make_simple_restrictinfo(root, op));
2890 1518 : iclause->lossy = false;
2891 1518 : iclause->indexcol = indexcol;
2892 1518 : iclause->indexcols = NIL;
2893 1518 : return iclause;
2894 : }
2895 :
2896 188 : return NULL;
2897 : }
2898 :
2899 : /*
2900 : * match_opclause_to_indexcol()
2901 : * Handles the OpExpr case for match_clause_to_indexcol(),
2902 : * which see for comments.
2903 : */
2904 : static IndexClause *
2905 1303986 : match_opclause_to_indexcol(PlannerInfo *root,
2906 : RestrictInfo *rinfo,
2907 : int indexcol,
2908 : IndexOptInfo *index)
2909 : {
2910 : IndexClause *iclause;
2911 1303986 : OpExpr *clause = (OpExpr *) rinfo->clause;
2912 : Node *leftop,
2913 : *rightop;
2914 : Oid expr_op;
2915 : Oid expr_coll;
2916 : Index index_relid;
2917 : Oid opfamily;
2918 : Oid idxcollation;
2919 :
2920 : /*
2921 : * Only binary operators need apply. (In theory, a planner support
2922 : * function could do something with a unary operator, but it seems
2923 : * unlikely to be worth the cycles to check.)
2924 : */
2925 1303986 : if (list_length(clause->args) != 2)
2926 0 : return NULL;
2927 :
2928 1303986 : leftop = (Node *) linitial(clause->args);
2929 1303986 : rightop = (Node *) lsecond(clause->args);
2930 1303986 : expr_op = clause->opno;
2931 1303986 : expr_coll = clause->inputcollid;
2932 :
2933 1303986 : index_relid = index->rel->relid;
2934 1303986 : opfamily = index->opfamily[indexcol];
2935 1303986 : idxcollation = index->indexcollations[indexcol];
2936 :
2937 : /*
2938 : * Check for clauses of the form: (indexkey operator constant) or
2939 : * (constant operator indexkey). See match_clause_to_indexcol's notes
2940 : * about const-ness.
2941 : *
2942 : * Note that we don't ask the support function about clauses that don't
2943 : * have one of these forms. Again, in principle it might be possible to
2944 : * do something, but it seems unlikely to be worth the cycles to check.
2945 : */
2946 1303986 : if (match_index_to_operand(leftop, indexcol, index) &&
2947 323804 : !bms_is_member(index_relid, rinfo->right_relids) &&
2948 323630 : !contain_volatile_functions(rightop))
2949 : {
2950 640712 : if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
2951 317082 : op_in_opfamily(expr_op, opfamily))
2952 : {
2953 310218 : iclause = makeNode(IndexClause);
2954 310218 : iclause->rinfo = rinfo;
2955 310218 : iclause->indexquals = list_make1(rinfo);
2956 310218 : iclause->lossy = false;
2957 310218 : iclause->indexcol = indexcol;
2958 310218 : iclause->indexcols = NIL;
2959 310218 : return iclause;
2960 : }
2961 :
2962 : /*
2963 : * If we didn't find a member of the index's opfamily, try the support
2964 : * function for the operator's underlying function.
2965 : */
2966 13412 : set_opfuncid(clause); /* make sure we have opfuncid */
2967 13412 : return get_index_clause_from_support(root,
2968 : rinfo,
2969 : clause->opfuncid,
2970 : 0, /* indexarg on left */
2971 : indexcol,
2972 : index);
2973 : }
2974 :
2975 980356 : if (match_index_to_operand(rightop, indexcol, index) &&
2976 63442 : !bms_is_member(index_relid, rinfo->left_relids) &&
2977 63316 : !contain_volatile_functions(leftop))
2978 : {
2979 63316 : if (IndexCollMatchesExprColl(idxcollation, expr_coll))
2980 : {
2981 63304 : Oid comm_op = get_commutator(expr_op);
2982 :
2983 126608 : if (OidIsValid(comm_op) &&
2984 63304 : op_in_opfamily(comm_op, opfamily))
2985 : {
2986 : RestrictInfo *commrinfo;
2987 :
2988 : /* Build a commuted OpExpr and RestrictInfo */
2989 62838 : commrinfo = commute_restrictinfo(rinfo, comm_op);
2990 :
2991 : /* Make an IndexClause showing that as a derived qual */
2992 62838 : iclause = makeNode(IndexClause);
2993 62838 : iclause->rinfo = rinfo;
2994 62838 : iclause->indexquals = list_make1(commrinfo);
2995 62838 : iclause->lossy = false;
2996 62838 : iclause->indexcol = indexcol;
2997 62838 : iclause->indexcols = NIL;
2998 62838 : return iclause;
2999 : }
3000 : }
3001 :
3002 : /*
3003 : * If we didn't find a member of the index's opfamily, try the support
3004 : * function for the operator's underlying function.
3005 : */
3006 478 : set_opfuncid(clause); /* make sure we have opfuncid */
3007 478 : return get_index_clause_from_support(root,
3008 : rinfo,
3009 : clause->opfuncid,
3010 : 1, /* indexarg on right */
3011 : indexcol,
3012 : index);
3013 : }
3014 :
3015 917040 : return NULL;
3016 : }
3017 :
3018 : /*
3019 : * match_funcclause_to_indexcol()
3020 : * Handles the FuncExpr case for match_clause_to_indexcol(),
3021 : * which see for comments.
3022 : */
3023 : static IndexClause *
3024 35724 : match_funcclause_to_indexcol(PlannerInfo *root,
3025 : RestrictInfo *rinfo,
3026 : int indexcol,
3027 : IndexOptInfo *index)
3028 : {
3029 35724 : FuncExpr *clause = (FuncExpr *) rinfo->clause;
3030 : int indexarg;
3031 : ListCell *lc;
3032 :
3033 : /*
3034 : * We have no built-in intelligence about function clauses, but if there's
3035 : * a planner support function, it might be able to do something. But, to
3036 : * cut down on wasted planning cycles, only call the support function if
3037 : * at least one argument matches the target index column.
3038 : *
3039 : * Note that we don't insist on the other arguments being pseudoconstants;
3040 : * the support function has to check that. This is to allow cases where
3041 : * only some of the other arguments need to be included in the indexqual.
3042 : */
3043 35724 : indexarg = 0;
3044 85664 : foreach(lc, clause->args)
3045 : {
3046 57088 : Node *op = (Node *) lfirst(lc);
3047 :
3048 57088 : if (match_index_to_operand(op, indexcol, index))
3049 : {
3050 7148 : return get_index_clause_from_support(root,
3051 : rinfo,
3052 : clause->funcid,
3053 : indexarg,
3054 : indexcol,
3055 : index);
3056 : }
3057 :
3058 49940 : indexarg++;
3059 : }
3060 :
3061 28576 : return NULL;
3062 : }
3063 :
3064 : /*
3065 : * get_index_clause_from_support()
3066 : * If the function has a planner support function, try to construct
3067 : * an IndexClause using indexquals created by the support function.
3068 : */
3069 : static IndexClause *
3070 21038 : get_index_clause_from_support(PlannerInfo *root,
3071 : RestrictInfo *rinfo,
3072 : Oid funcid,
3073 : int indexarg,
3074 : int indexcol,
3075 : IndexOptInfo *index)
3076 : {
3077 21038 : Oid prosupport = get_func_support(funcid);
3078 : SupportRequestIndexCondition req;
3079 : List *sresult;
3080 :
3081 21038 : if (!OidIsValid(prosupport))
3082 13016 : return NULL;
3083 :
3084 8022 : req.type = T_SupportRequestIndexCondition;
3085 8022 : req.root = root;
3086 8022 : req.funcid = funcid;
3087 8022 : req.node = (Node *) rinfo->clause;
3088 8022 : req.indexarg = indexarg;
3089 8022 : req.index = index;
3090 8022 : req.indexcol = indexcol;
3091 8022 : req.opfamily = index->opfamily[indexcol];
3092 8022 : req.indexcollation = index->indexcollations[indexcol];
3093 :
3094 8022 : req.lossy = true; /* default assumption */
3095 :
3096 : sresult = (List *)
3097 8022 : DatumGetPointer(OidFunctionCall1(prosupport,
3098 : PointerGetDatum(&req)));
3099 :
3100 8022 : if (sresult != NIL)
3101 : {
3102 1390 : IndexClause *iclause = makeNode(IndexClause);
3103 1390 : List *indexquals = NIL;
3104 : ListCell *lc;
3105 :
3106 : /*
3107 : * The support function API says it should just give back bare
3108 : * clauses, so here we must wrap each one in a RestrictInfo.
3109 : */
3110 4092 : foreach(lc, sresult)
3111 : {
3112 2702 : Expr *clause = (Expr *) lfirst(lc);
3113 :
3114 2702 : indexquals = lappend(indexquals,
3115 2702 : make_simple_restrictinfo(root, clause));
3116 : }
3117 :
3118 1390 : iclause->rinfo = rinfo;
3119 1390 : iclause->indexquals = indexquals;
3120 1390 : iclause->lossy = req.lossy;
3121 1390 : iclause->indexcol = indexcol;
3122 1390 : iclause->indexcols = NIL;
3123 :
3124 1390 : return iclause;
3125 : }
3126 :
3127 6632 : return NULL;
3128 : }
3129 :
3130 : /*
3131 : * match_saopclause_to_indexcol()
3132 : * Handles the ScalarArrayOpExpr case for match_clause_to_indexcol(),
3133 : * which see for comments.
3134 : */
3135 : static IndexClause *
3136 74142 : match_saopclause_to_indexcol(PlannerInfo *root,
3137 : RestrictInfo *rinfo,
3138 : int indexcol,
3139 : IndexOptInfo *index)
3140 : {
3141 74142 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
3142 : Node *leftop,
3143 : *rightop;
3144 : Relids right_relids;
3145 : Oid expr_op;
3146 : Oid expr_coll;
3147 : Index index_relid;
3148 : Oid opfamily;
3149 : Oid idxcollation;
3150 :
3151 : /* We only accept ANY clauses, not ALL */
3152 74142 : if (!saop->useOr)
3153 13852 : return NULL;
3154 60290 : leftop = (Node *) linitial(saop->args);
3155 60290 : rightop = (Node *) lsecond(saop->args);
3156 60290 : right_relids = pull_varnos(root, rightop);
3157 60290 : expr_op = saop->opno;
3158 60290 : expr_coll = saop->inputcollid;
3159 :
3160 60290 : index_relid = index->rel->relid;
3161 60290 : opfamily = index->opfamily[indexcol];
3162 60290 : idxcollation = index->indexcollations[indexcol];
3163 :
3164 : /*
3165 : * We must have indexkey on the left and a pseudo-constant array argument.
3166 : */
3167 60290 : if (match_index_to_operand(leftop, indexcol, index) &&
3168 6248 : !bms_is_member(index_relid, right_relids) &&
3169 6248 : !contain_volatile_functions(rightop))
3170 : {
3171 12490 : if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
3172 6242 : op_in_opfamily(expr_op, opfamily))
3173 : {
3174 6230 : IndexClause *iclause = makeNode(IndexClause);
3175 :
3176 6230 : iclause->rinfo = rinfo;
3177 6230 : iclause->indexquals = list_make1(rinfo);
3178 6230 : iclause->lossy = false;
3179 6230 : iclause->indexcol = indexcol;
3180 6230 : iclause->indexcols = NIL;
3181 6230 : return iclause;
3182 : }
3183 :
3184 : /*
3185 : * We do not currently ask support functions about ScalarArrayOpExprs,
3186 : * though in principle we could.
3187 : */
3188 : }
3189 :
3190 54060 : return NULL;
3191 : }
3192 :
3193 : /*
3194 : * match_rowcompare_to_indexcol()
3195 : * Handles the RowCompareExpr case for match_clause_to_indexcol(),
3196 : * which see for comments.
3197 : *
3198 : * In this routine we check whether the first column of the row comparison
3199 : * matches the target index column. This is sufficient to guarantee that some
3200 : * index condition can be constructed from the RowCompareExpr --- the rest
3201 : * is handled by expand_indexqual_rowcompare().
3202 : */
3203 : static IndexClause *
3204 432 : match_rowcompare_to_indexcol(PlannerInfo *root,
3205 : RestrictInfo *rinfo,
3206 : int indexcol,
3207 : IndexOptInfo *index)
3208 : {
3209 432 : RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3210 : Index index_relid;
3211 : Oid opfamily;
3212 : Oid idxcollation;
3213 : Node *leftop,
3214 : *rightop;
3215 : bool var_on_left;
3216 : Oid expr_op;
3217 : Oid expr_coll;
3218 :
3219 : /* Forget it if we're not dealing with a btree index */
3220 432 : if (index->relam != BTREE_AM_OID)
3221 0 : return NULL;
3222 :
3223 432 : index_relid = index->rel->relid;
3224 432 : opfamily = index->opfamily[indexcol];
3225 432 : idxcollation = index->indexcollations[indexcol];
3226 :
3227 : /*
3228 : * We could do the matching on the basis of insisting that the opfamily
3229 : * shown in the RowCompareExpr be the same as the index column's opfamily,
3230 : * but that could fail in the presence of reverse-sort opfamilies: it'd be
3231 : * a matter of chance whether RowCompareExpr had picked the forward or
3232 : * reverse-sort family. So look only at the operator, and match if it is
3233 : * a member of the index's opfamily (after commutation, if the indexkey is
3234 : * on the right). We'll worry later about whether any additional
3235 : * operators are matchable to the index.
3236 : */
3237 432 : leftop = (Node *) linitial(clause->largs);
3238 432 : rightop = (Node *) linitial(clause->rargs);
3239 432 : expr_op = linitial_oid(clause->opnos);
3240 432 : expr_coll = linitial_oid(clause->inputcollids);
3241 :
3242 : /* Collations must match, if relevant */
3243 432 : if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
3244 0 : return NULL;
3245 :
3246 : /*
3247 : * These syntactic tests are the same as in match_opclause_to_indexcol()
3248 : */
3249 432 : if (match_index_to_operand(leftop, indexcol, index) &&
3250 126 : !bms_is_member(index_relid, pull_varnos(root, rightop)) &&
3251 126 : !contain_volatile_functions(rightop))
3252 : {
3253 : /* OK, indexkey is on left */
3254 126 : var_on_left = true;
3255 : }
3256 306 : else if (match_index_to_operand(rightop, indexcol, index) &&
3257 24 : !bms_is_member(index_relid, pull_varnos(root, leftop)) &&
3258 24 : !contain_volatile_functions(leftop))
3259 : {
3260 : /* indexkey is on right, so commute the operator */
3261 24 : expr_op = get_commutator(expr_op);
3262 24 : if (expr_op == InvalidOid)
3263 0 : return NULL;
3264 24 : var_on_left = false;
3265 : }
3266 : else
3267 282 : return NULL;
3268 :
3269 : /* We're good if the operator is the right type of opfamily member */
3270 150 : switch (get_op_opfamily_strategy(expr_op, opfamily))
3271 : {
3272 150 : case BTLessStrategyNumber:
3273 : case BTLessEqualStrategyNumber:
3274 : case BTGreaterEqualStrategyNumber:
3275 : case BTGreaterStrategyNumber:
3276 150 : return expand_indexqual_rowcompare(root,
3277 : rinfo,
3278 : indexcol,
3279 : index,
3280 : expr_op,
3281 : var_on_left);
3282 : }
3283 :
3284 0 : return NULL;
3285 : }
3286 :
3287 : /*
3288 : * match_orclause_to_indexcol()
3289 : * Handles the OR-expr case for match_clause_to_indexcol() in the case
3290 : * when it could be transformed to ScalarArrayOpExpr.
3291 : *
3292 : * In this routine, we attempt to transform a list of OR-clause args into a
3293 : * single SAOP expression matching the target index column. On success,
3294 : * return an IndexClause, containing the transformed expression or NULL,
3295 : * if failed.
3296 : */
3297 : static IndexClause *
3298 45584 : match_orclause_to_indexcol(PlannerInfo *root,
3299 : RestrictInfo *rinfo,
3300 : int indexcol,
3301 : IndexOptInfo *index)
3302 : {
3303 : ListCell *lc;
3304 45584 : BoolExpr *orclause = (BoolExpr *) rinfo->orclause;
3305 45584 : Node *indexExpr = NULL;
3306 45584 : List *consts = NIL;
3307 45584 : Node *arrayNode = NULL;
3308 45584 : ScalarArrayOpExpr *saopexpr = NULL;
3309 45584 : Oid matchOpno = InvalidOid;
3310 : IndexClause *iclause;
3311 45584 : Oid consttype = InvalidOid;
3312 45584 : Oid arraytype = InvalidOid;
3313 45584 : Oid inputcollid = InvalidOid;
3314 45584 : bool firstTime = true;
3315 45584 : bool haveNonConst = false;
3316 45584 : Index indexRelid = index->rel->relid;
3317 :
3318 : Assert(IsA(orclause, BoolExpr));
3319 : Assert(orclause->boolop == OR_EXPR);
3320 :
3321 : /* Ignore index if it doesn't support SAOP clauses */
3322 45584 : if (!index->amsearcharray)
3323 106 : return NULL;
3324 :
3325 : /*
3326 : * Try to convert a list of OR-clauses to a single SAOP expression. Each
3327 : * OR entry must be in the form: (indexkey operator constant) or (constant
3328 : * operator indexkey). Operators of all the entries must match. To be
3329 : * effective, give up on the first non-matching entry. Exit is
3330 : * implemented as a break from the loop, which is catched afterwards.
3331 : */
3332 49858 : foreach(lc, orclause->args)
3333 : {
3334 : RestrictInfo *subRinfo;
3335 : OpExpr *subClause;
3336 : Oid opno;
3337 : Node *leftop,
3338 : *rightop;
3339 : Node *constExpr;
3340 :
3341 48808 : if (!IsA(lfirst(lc), RestrictInfo))
3342 4742 : break;
3343 :
3344 44066 : subRinfo = (RestrictInfo *) lfirst(lc);
3345 :
3346 : /* Only operator clauses can match */
3347 44066 : if (!IsA(subRinfo->clause, OpExpr))
3348 15786 : break;
3349 :
3350 28280 : subClause = (OpExpr *) subRinfo->clause;
3351 28280 : opno = subClause->opno;
3352 :
3353 : /* Only binary operators can match */
3354 28280 : if (list_length(subClause->args) != 2)
3355 0 : break;
3356 :
3357 : /*
3358 : * The parameters below must match between sub-rinfo and its parent as
3359 : * make_restrictinfo() fills them with the same values, and further
3360 : * modifications are also the same for the whole subtree. However,
3361 : * still make a sanity check.
3362 : */
3363 : Assert(subRinfo->is_pushed_down == rinfo->is_pushed_down);
3364 : Assert(subRinfo->is_clone == rinfo->is_clone);
3365 : Assert(subRinfo->security_level == rinfo->security_level);
3366 : Assert(bms_equal(subRinfo->incompatible_relids, rinfo->incompatible_relids));
3367 : Assert(bms_equal(subRinfo->outer_relids, rinfo->outer_relids));
3368 :
3369 : /*
3370 : * Also, check that required_relids in sub-rinfo is subset of parent's
3371 : * required_relids.
3372 : */
3373 : Assert(bms_is_subset(subRinfo->required_relids, rinfo->required_relids));
3374 :
3375 : /* Only the operator returning a boolean suit the transformation. */
3376 28280 : if (get_op_rettype(opno) != BOOLOID)
3377 0 : break;
3378 :
3379 : /*
3380 : * Check for clauses of the form: (indexkey operator constant) or
3381 : * (constant operator indexkey). See match_clause_to_indexcol's notes
3382 : * about const-ness.
3383 : */
3384 28280 : leftop = (Node *) linitial(subClause->args);
3385 28280 : rightop = (Node *) lsecond(subClause->args);
3386 28280 : if (match_index_to_operand(leftop, indexcol, index) &&
3387 6140 : !bms_is_member(indexRelid, subRinfo->right_relids) &&
3388 6110 : !contain_volatile_functions(rightop))
3389 : {
3390 6110 : indexExpr = leftop;
3391 6110 : constExpr = rightop;
3392 : }
3393 22170 : else if (match_index_to_operand(rightop, indexcol, index) &&
3394 170 : !bms_is_member(indexRelid, subRinfo->left_relids) &&
3395 164 : !contain_volatile_functions(leftop))
3396 : {
3397 164 : opno = get_commutator(opno);
3398 164 : if (!OidIsValid(opno))
3399 : {
3400 : /* commutator doesn't exist, we can't reverse the order */
3401 0 : break;
3402 : }
3403 164 : indexExpr = rightop;
3404 164 : constExpr = leftop;
3405 : }
3406 : else
3407 : {
3408 : break;
3409 : }
3410 :
3411 : /*
3412 : * Ignore any RelabelType node above the operands. This is needed to
3413 : * be able to apply indexscanning in binary-compatible-operator cases.
3414 : * Note: we can assume there is at most one RelabelType node;
3415 : * eval_const_expressions() will have simplified if more than one.
3416 : */
3417 6274 : if (IsA(constExpr, RelabelType))
3418 0 : constExpr = (Node *) ((RelabelType *) constExpr)->arg;
3419 6274 : if (IsA(indexExpr, RelabelType))
3420 12 : indexExpr = (Node *) ((RelabelType *) indexExpr)->arg;
3421 :
3422 : /* Forbid transformation for composite types, records. */
3423 12548 : if (type_is_rowtype(exprType(constExpr)) ||
3424 6274 : type_is_rowtype(exprType(indexExpr)))
3425 : break;
3426 :
3427 : /*
3428 : * Save information about the operator, type, and collation for the
3429 : * first matching qual. Then, check that subsequent quals match the
3430 : * first.
3431 : */
3432 6274 : if (firstTime)
3433 : {
3434 4534 : matchOpno = opno;
3435 4534 : consttype = exprType(constExpr);
3436 4534 : arraytype = get_array_type(consttype);
3437 4534 : inputcollid = subClause->inputcollid;
3438 :
3439 : /*
3440 : * Check that the operator is presented in the opfamily and that
3441 : * the expression collation matches the index collation. Also,
3442 : * there must be an array type to construct an array later.
3443 : */
3444 4534 : if (!IndexCollMatchesExprColl(index->indexcollations[indexcol], inputcollid) ||
3445 4408 : !op_in_opfamily(matchOpno, index->opfamily[indexcol]) ||
3446 : !OidIsValid(arraytype))
3447 : break;
3448 2766 : firstTime = false;
3449 : }
3450 : else
3451 : {
3452 1740 : if (opno != matchOpno ||
3453 3228 : inputcollid != subClause->inputcollid ||
3454 1614 : consttype != exprType(constExpr))
3455 : break;
3456 : }
3457 :
3458 : /*
3459 : * Check if our list of constants in match_clause_to_indexcol's
3460 : * understanding of const-ness have something other than Const.
3461 : */
3462 4380 : if (!IsA(constExpr, Const))
3463 344 : haveNonConst = true;
3464 4380 : consts = lappend(consts, constExpr);
3465 : }
3466 :
3467 : /*
3468 : * Catch the break from the loop above. Normally, a foreach() loop ends
3469 : * up with a NULL list cell. A non-NULL list cell indicates a break from
3470 : * the foreach() loop. Free the consts list and return NULL then.
3471 : */
3472 45478 : if (lc != NULL)
3473 : {
3474 44428 : list_free(consts);
3475 44428 : return NULL;
3476 : }
3477 :
3478 : /*
3479 : * Assemble an array from the list of constants. It seems more profitable
3480 : * to build a const array. But in the presence of other nodes, we don't
3481 : * have a specific value here and must employ an ArrayExpr instead.
3482 : */
3483 1050 : if (haveNonConst)
3484 : {
3485 96 : ArrayExpr *arrayExpr = makeNode(ArrayExpr);
3486 :
3487 : /* array_collid will be set by parse_collate.c */
3488 96 : arrayExpr->element_typeid = consttype;
3489 96 : arrayExpr->array_typeid = arraytype;
3490 96 : arrayExpr->multidims = false;
3491 96 : arrayExpr->elements = consts;
3492 96 : arrayExpr->location = -1;
3493 :
3494 96 : arrayNode = (Node *) arrayExpr;
3495 : }
3496 : else
3497 : {
3498 : int16 typlen;
3499 : bool typbyval;
3500 : char typalign;
3501 : Datum *elems;
3502 954 : int i = 0;
3503 : ArrayType *arrayConst;
3504 :
3505 954 : get_typlenbyvalalign(consttype, &typlen, &typbyval, &typalign);
3506 :
3507 954 : elems = (Datum *) palloc(sizeof(Datum) * list_length(consts));
3508 4254 : foreach_node(Const, value, consts)
3509 : {
3510 : Assert(!value->constisnull);
3511 :
3512 2346 : elems[i++] = value->constvalue;
3513 : }
3514 :
3515 954 : arrayConst = construct_array(elems, i, consttype,
3516 : typlen, typbyval, typalign);
3517 954 : arrayNode = (Node *) makeConst(arraytype, -1, inputcollid,
3518 : -1, PointerGetDatum(arrayConst),
3519 : false, false);
3520 :
3521 954 : pfree(elems);
3522 954 : list_free(consts);
3523 : }
3524 :
3525 : /* Build the SAOP expression node */
3526 1050 : saopexpr = makeNode(ScalarArrayOpExpr);
3527 1050 : saopexpr->opno = matchOpno;
3528 1050 : saopexpr->opfuncid = get_opcode(matchOpno);
3529 1050 : saopexpr->hashfuncid = InvalidOid;
3530 1050 : saopexpr->negfuncid = InvalidOid;
3531 1050 : saopexpr->useOr = true;
3532 1050 : saopexpr->inputcollid = inputcollid;
3533 1050 : saopexpr->args = list_make2(indexExpr, arrayNode);
3534 1050 : saopexpr->location = -1;
3535 :
3536 : /*
3537 : * Finally, build an IndexClause based on the SAOP node. Use
3538 : * make_simple_restrictinfo() to get RestrictInfo with clean selectivity
3539 : * estimations, because they may differ from the estimation made for an OR
3540 : * clause. Although it is not a lossy expression, keep the original rinfo
3541 : * in iclause->rinfo as prescribed.
3542 : */
3543 1050 : iclause = makeNode(IndexClause);
3544 1050 : iclause->rinfo = rinfo;
3545 1050 : iclause->indexquals = list_make1(make_simple_restrictinfo(root,
3546 : &saopexpr->xpr));
3547 1050 : iclause->lossy = false;
3548 1050 : iclause->indexcol = indexcol;
3549 1050 : iclause->indexcols = NIL;
3550 1050 : return iclause;
3551 : }
3552 :
3553 : /*
3554 : * expand_indexqual_rowcompare --- expand a single indexqual condition
3555 : * that is a RowCompareExpr
3556 : *
3557 : * It's already known that the first column of the row comparison matches
3558 : * the specified column of the index. We can use additional columns of the
3559 : * row comparison as index qualifications, so long as they match the index
3560 : * in the "same direction", ie, the indexkeys are all on the same side of the
3561 : * clause and the operators are all the same-type members of the opfamilies.
3562 : *
3563 : * If all the columns of the RowCompareExpr match in this way, we just use it
3564 : * as-is, except for possibly commuting it to put the indexkeys on the left.
3565 : *
3566 : * Otherwise, we build a shortened RowCompareExpr (if more than one
3567 : * column matches) or a simple OpExpr (if the first-column match is all
3568 : * there is). In these cases the modified clause is always "<=" or ">="
3569 : * even when the original was "<" or ">" --- this is necessary to match all
3570 : * the rows that could match the original. (We are building a lossy version
3571 : * of the row comparison when we do this, so we set lossy = true.)
3572 : *
3573 : * Note: this is really just the last half of match_rowcompare_to_indexcol,
3574 : * but we split it out for comprehensibility.
3575 : */
3576 : static IndexClause *
3577 150 : expand_indexqual_rowcompare(PlannerInfo *root,
3578 : RestrictInfo *rinfo,
3579 : int indexcol,
3580 : IndexOptInfo *index,
3581 : Oid expr_op,
3582 : bool var_on_left)
3583 : {
3584 150 : IndexClause *iclause = makeNode(IndexClause);
3585 150 : RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3586 : int op_strategy;
3587 : Oid op_lefttype;
3588 : Oid op_righttype;
3589 : int matching_cols;
3590 : List *expr_ops;
3591 : List *opfamilies;
3592 : List *lefttypes;
3593 : List *righttypes;
3594 : List *new_ops;
3595 : List *var_args;
3596 : List *non_var_args;
3597 :
3598 150 : iclause->rinfo = rinfo;
3599 150 : iclause->indexcol = indexcol;
3600 :
3601 150 : if (var_on_left)
3602 : {
3603 126 : var_args = clause->largs;
3604 126 : non_var_args = clause->rargs;
3605 : }
3606 : else
3607 : {
3608 24 : var_args = clause->rargs;
3609 24 : non_var_args = clause->largs;
3610 : }
3611 :
3612 150 : get_op_opfamily_properties(expr_op, index->opfamily[indexcol], false,
3613 : &op_strategy,
3614 : &op_lefttype,
3615 : &op_righttype);
3616 :
3617 : /* Initialize returned list of which index columns are used */
3618 150 : iclause->indexcols = list_make1_int(indexcol);
3619 :
3620 : /* Build lists of ops, opfamilies and operator datatypes in case needed */
3621 150 : expr_ops = list_make1_oid(expr_op);
3622 150 : opfamilies = list_make1_oid(index->opfamily[indexcol]);
3623 150 : lefttypes = list_make1_oid(op_lefttype);
3624 150 : righttypes = list_make1_oid(op_righttype);
3625 :
3626 : /*
3627 : * See how many of the remaining columns match some index column in the
3628 : * same way. As in match_clause_to_indexcol(), the "other" side of any
3629 : * potential index condition is OK as long as it doesn't use Vars from the
3630 : * indexed relation.
3631 : */
3632 150 : matching_cols = 1;
3633 :
3634 282 : while (matching_cols < list_length(var_args))
3635 : {
3636 186 : Node *varop = (Node *) list_nth(var_args, matching_cols);
3637 186 : Node *constop = (Node *) list_nth(non_var_args, matching_cols);
3638 : int i;
3639 :
3640 186 : expr_op = list_nth_oid(clause->opnos, matching_cols);
3641 186 : if (!var_on_left)
3642 : {
3643 : /* indexkey is on right, so commute the operator */
3644 24 : expr_op = get_commutator(expr_op);
3645 24 : if (expr_op == InvalidOid)
3646 0 : break; /* operator is not usable */
3647 : }
3648 186 : if (bms_is_member(index->rel->relid, pull_varnos(root, constop)))
3649 0 : break; /* no good, Var on wrong side */
3650 186 : if (contain_volatile_functions(constop))
3651 0 : break; /* no good, volatile comparison value */
3652 :
3653 : /*
3654 : * The Var side can match any key column of the index.
3655 : */
3656 444 : for (i = 0; i < index->nkeycolumns; i++)
3657 : {
3658 390 : if (match_index_to_operand(varop, i, index) &&
3659 132 : get_op_opfamily_strategy(expr_op,
3660 132 : index->opfamily[i]) == op_strategy &&
3661 132 : IndexCollMatchesExprColl(index->indexcollations[i],
3662 : list_nth_oid(clause->inputcollids,
3663 : matching_cols)))
3664 : break;
3665 : }
3666 186 : if (i >= index->nkeycolumns)
3667 54 : break; /* no match found */
3668 :
3669 : /* Add column number to returned list */
3670 132 : iclause->indexcols = lappend_int(iclause->indexcols, i);
3671 :
3672 : /* Add operator info to lists */
3673 132 : get_op_opfamily_properties(expr_op, index->opfamily[i], false,
3674 : &op_strategy,
3675 : &op_lefttype,
3676 : &op_righttype);
3677 132 : expr_ops = lappend_oid(expr_ops, expr_op);
3678 132 : opfamilies = lappend_oid(opfamilies, index->opfamily[i]);
3679 132 : lefttypes = lappend_oid(lefttypes, op_lefttype);
3680 132 : righttypes = lappend_oid(righttypes, op_righttype);
3681 :
3682 : /* This column matches, keep scanning */
3683 132 : matching_cols++;
3684 : }
3685 :
3686 : /* Result is non-lossy if all columns are usable as index quals */
3687 150 : iclause->lossy = (matching_cols != list_length(clause->opnos));
3688 :
3689 : /*
3690 : * We can use rinfo->clause as-is if we have var on left and it's all
3691 : * usable as index quals.
3692 : */
3693 150 : if (var_on_left && !iclause->lossy)
3694 84 : iclause->indexquals = list_make1(rinfo);
3695 : else
3696 : {
3697 : /*
3698 : * We have to generate a modified rowcompare (possibly just one
3699 : * OpExpr). The painful part of this is changing < to <= or > to >=,
3700 : * so deal with that first.
3701 : */
3702 66 : if (!iclause->lossy)
3703 : {
3704 : /* very easy, just use the commuted operators */
3705 12 : new_ops = expr_ops;
3706 : }
3707 54 : else if (op_strategy == BTLessEqualStrategyNumber ||
3708 54 : op_strategy == BTGreaterEqualStrategyNumber)
3709 : {
3710 : /* easy, just use the same (possibly commuted) operators */
3711 0 : new_ops = list_truncate(expr_ops, matching_cols);
3712 : }
3713 : else
3714 : {
3715 : ListCell *opfamilies_cell;
3716 : ListCell *lefttypes_cell;
3717 : ListCell *righttypes_cell;
3718 :
3719 54 : if (op_strategy == BTLessStrategyNumber)
3720 30 : op_strategy = BTLessEqualStrategyNumber;
3721 24 : else if (op_strategy == BTGreaterStrategyNumber)
3722 24 : op_strategy = BTGreaterEqualStrategyNumber;
3723 : else
3724 0 : elog(ERROR, "unexpected strategy number %d", op_strategy);
3725 54 : new_ops = NIL;
3726 144 : forthree(opfamilies_cell, opfamilies,
3727 : lefttypes_cell, lefttypes,
3728 : righttypes_cell, righttypes)
3729 : {
3730 90 : Oid opfam = lfirst_oid(opfamilies_cell);
3731 90 : Oid lefttype = lfirst_oid(lefttypes_cell);
3732 90 : Oid righttype = lfirst_oid(righttypes_cell);
3733 :
3734 90 : expr_op = get_opfamily_member(opfam, lefttype, righttype,
3735 : op_strategy);
3736 90 : if (!OidIsValid(expr_op)) /* should not happen */
3737 0 : elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
3738 : op_strategy, lefttype, righttype, opfam);
3739 90 : new_ops = lappend_oid(new_ops, expr_op);
3740 : }
3741 : }
3742 :
3743 : /* If we have more than one matching col, create a subset rowcompare */
3744 66 : if (matching_cols > 1)
3745 : {
3746 48 : RowCompareExpr *rc = makeNode(RowCompareExpr);
3747 :
3748 48 : rc->cmptype = (CompareType) op_strategy;
3749 48 : rc->opnos = new_ops;
3750 48 : rc->opfamilies = list_copy_head(clause->opfamilies,
3751 : matching_cols);
3752 48 : rc->inputcollids = list_copy_head(clause->inputcollids,
3753 : matching_cols);
3754 48 : rc->largs = list_copy_head(var_args, matching_cols);
3755 48 : rc->rargs = list_copy_head(non_var_args, matching_cols);
3756 48 : iclause->indexquals = list_make1(make_simple_restrictinfo(root,
3757 : (Expr *) rc));
3758 : }
3759 : else
3760 : {
3761 : Expr *op;
3762 :
3763 : /* We don't report an index column list in this case */
3764 18 : iclause->indexcols = NIL;
3765 :
3766 18 : op = make_opclause(linitial_oid(new_ops), BOOLOID, false,
3767 18 : copyObject(linitial(var_args)),
3768 18 : copyObject(linitial(non_var_args)),
3769 : InvalidOid,
3770 18 : linitial_oid(clause->inputcollids));
3771 18 : iclause->indexquals = list_make1(make_simple_restrictinfo(root, op));
3772 : }
3773 : }
3774 :
3775 150 : return iclause;
3776 : }
3777 :
3778 :
3779 : /****************************************************************************
3780 : * ---- ROUTINES TO CHECK ORDERING OPERATORS ----
3781 : ****************************************************************************/
3782 :
3783 : /*
3784 : * match_pathkeys_to_index
3785 : * For the given 'index' and 'pathkeys', output a list of suitable ORDER
3786 : * BY expressions, each of the form "indexedcol operator pseudoconstant",
3787 : * along with an integer list of the index column numbers (zero based)
3788 : * that each clause would be used with.
3789 : *
3790 : * This attempts to find an ORDER BY and index column number for all items in
3791 : * the pathkey list, however, if we're unable to match any given pathkey to an
3792 : * index column, we return just the ones matched by the function so far. This
3793 : * allows callers who are interested in partial matches to get them. Callers
3794 : * can determine a partial match vs a full match by checking the outputted
3795 : * list lengths. A full match will have one item in the output lists for each
3796 : * item in the given 'pathkeys' list.
3797 : */
3798 : static void
3799 1078 : match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
3800 : List **orderby_clauses_p,
3801 : List **clause_columns_p)
3802 : {
3803 : ListCell *lc1;
3804 :
3805 1078 : *orderby_clauses_p = NIL; /* set default results */
3806 1078 : *clause_columns_p = NIL;
3807 :
3808 : /* Only indexes with the amcanorderbyop property are interesting here */
3809 1078 : if (!index->amcanorderbyop)
3810 0 : return;
3811 :
3812 1552 : foreach(lc1, pathkeys)
3813 : {
3814 1080 : PathKey *pathkey = (PathKey *) lfirst(lc1);
3815 1080 : bool found = false;
3816 : ListCell *lc2;
3817 :
3818 :
3819 : /* Pathkey must request default sort order for the target opfamily */
3820 1080 : if (pathkey->pk_strategy != BTLessStrategyNumber ||
3821 1046 : pathkey->pk_nulls_first)
3822 606 : return;
3823 :
3824 : /* If eclass is volatile, no hope of using an indexscan */
3825 1046 : if (pathkey->pk_eclass->ec_has_volatile)
3826 0 : return;
3827 :
3828 : /*
3829 : * Try to match eclass member expression(s) to index. Note that child
3830 : * EC members are considered, but only when they belong to the target
3831 : * relation. (Unlike regular members, the same expression could be a
3832 : * child member of more than one EC. Therefore, the same index could
3833 : * be considered to match more than one pathkey list, which is OK
3834 : * here. See also get_eclass_for_sort_expr.)
3835 : */
3836 1682 : foreach(lc2, pathkey->pk_eclass->ec_members)
3837 : {
3838 1110 : EquivalenceMember *member = (EquivalenceMember *) lfirst(lc2);
3839 : int indexcol;
3840 :
3841 : /* No possibility of match if it references other relations */
3842 1110 : if (!bms_equal(member->em_relids, index->rel->relids))
3843 64 : continue;
3844 :
3845 : /*
3846 : * We allow any column of the index to match each pathkey; they
3847 : * don't have to match left-to-right as you might expect. This is
3848 : * correct for GiST, and it doesn't matter for SP-GiST because
3849 : * that doesn't handle multiple columns anyway, and no other
3850 : * existing AMs support amcanorderbyop. We might need different
3851 : * logic in future for other implementations.
3852 : */
3853 1906 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
3854 : {
3855 : Expr *expr;
3856 :
3857 1334 : expr = match_clause_to_ordering_op(index,
3858 : indexcol,
3859 : member->em_expr,
3860 : pathkey->pk_opfamily);
3861 1334 : if (expr)
3862 : {
3863 474 : *orderby_clauses_p = lappend(*orderby_clauses_p, expr);
3864 474 : *clause_columns_p = lappend_int(*clause_columns_p, indexcol);
3865 474 : found = true;
3866 474 : break;
3867 : }
3868 : }
3869 :
3870 1046 : if (found) /* don't want to look at remaining members */
3871 474 : break;
3872 : }
3873 :
3874 : /*
3875 : * Return the matches found so far when this pathkey couldn't be
3876 : * matched to the index.
3877 : */
3878 1046 : if (!found)
3879 572 : return;
3880 : }
3881 : }
3882 :
3883 : /*
3884 : * match_clause_to_ordering_op
3885 : * Determines whether an ordering operator expression matches an
3886 : * index column.
3887 : *
3888 : * This is similar to, but simpler than, match_clause_to_indexcol.
3889 : * We only care about simple OpExpr cases. The input is a bare
3890 : * expression that is being ordered by, which must be of the form
3891 : * (indexkey op const) or (const op indexkey) where op is an ordering
3892 : * operator for the column's opfamily.
3893 : *
3894 : * 'index' is the index of interest.
3895 : * 'indexcol' is a column number of 'index' (counting from 0).
3896 : * 'clause' is the ordering expression to be tested.
3897 : * 'pk_opfamily' is the btree opfamily describing the required sort order.
3898 : *
3899 : * Note that we currently do not consider the collation of the ordering
3900 : * operator's result. In practical cases the result type will be numeric
3901 : * and thus have no collation, and it's not very clear what to match to
3902 : * if it did have a collation. The index's collation should match the
3903 : * ordering operator's input collation, not its result.
3904 : *
3905 : * If successful, return 'clause' as-is if the indexkey is on the left,
3906 : * otherwise a commuted copy of 'clause'. If no match, return NULL.
3907 : */
3908 : static Expr *
3909 1334 : match_clause_to_ordering_op(IndexOptInfo *index,
3910 : int indexcol,
3911 : Expr *clause,
3912 : Oid pk_opfamily)
3913 : {
3914 : Oid opfamily;
3915 : Oid idxcollation;
3916 : Node *leftop,
3917 : *rightop;
3918 : Oid expr_op;
3919 : Oid expr_coll;
3920 : Oid sortfamily;
3921 : bool commuted;
3922 :
3923 : Assert(indexcol < index->nkeycolumns);
3924 :
3925 1334 : opfamily = index->opfamily[indexcol];
3926 1334 : idxcollation = index->indexcollations[indexcol];
3927 :
3928 : /*
3929 : * Clause must be a binary opclause.
3930 : */
3931 1334 : if (!is_opclause(clause))
3932 860 : return NULL;
3933 474 : leftop = get_leftop(clause);
3934 474 : rightop = get_rightop(clause);
3935 474 : if (!leftop || !rightop)
3936 0 : return NULL;
3937 474 : expr_op = ((OpExpr *) clause)->opno;
3938 474 : expr_coll = ((OpExpr *) clause)->inputcollid;
3939 :
3940 : /*
3941 : * We can forget the whole thing right away if wrong collation.
3942 : */
3943 474 : if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
3944 0 : return NULL;
3945 :
3946 : /*
3947 : * Check for clauses of the form: (indexkey operator constant) or
3948 : * (constant operator indexkey).
3949 : */
3950 474 : if (match_index_to_operand(leftop, indexcol, index) &&
3951 450 : !contain_var_clause(rightop) &&
3952 450 : !contain_volatile_functions(rightop))
3953 : {
3954 450 : commuted = false;
3955 : }
3956 24 : else if (match_index_to_operand(rightop, indexcol, index) &&
3957 24 : !contain_var_clause(leftop) &&
3958 24 : !contain_volatile_functions(leftop))
3959 : {
3960 : /* Might match, but we need a commuted operator */
3961 24 : expr_op = get_commutator(expr_op);
3962 24 : if (expr_op == InvalidOid)
3963 0 : return NULL;
3964 24 : commuted = true;
3965 : }
3966 : else
3967 0 : return NULL;
3968 :
3969 : /*
3970 : * Is the (commuted) operator an ordering operator for the opfamily? And
3971 : * if so, does it yield the right sorting semantics?
3972 : */
3973 474 : sortfamily = get_op_opfamily_sortfamily(expr_op, opfamily);
3974 474 : if (sortfamily != pk_opfamily)
3975 0 : return NULL;
3976 :
3977 : /* We have a match. Return clause or a commuted version thereof. */
3978 474 : if (commuted)
3979 : {
3980 24 : OpExpr *newclause = makeNode(OpExpr);
3981 :
3982 : /* flat-copy all the fields of clause */
3983 24 : memcpy(newclause, clause, sizeof(OpExpr));
3984 :
3985 : /* commute it */
3986 24 : newclause->opno = expr_op;
3987 24 : newclause->opfuncid = InvalidOid;
3988 24 : newclause->args = list_make2(rightop, leftop);
3989 :
3990 24 : clause = (Expr *) newclause;
3991 : }
3992 :
3993 474 : return clause;
3994 : }
3995 :
3996 :
3997 : /****************************************************************************
3998 : * ---- ROUTINES TO DO PARTIAL INDEX PREDICATE TESTS ----
3999 : ****************************************************************************/
4000 :
4001 : /*
4002 : * check_index_predicates
4003 : * Set the predicate-derived IndexOptInfo fields for each index
4004 : * of the specified relation.
4005 : *
4006 : * predOK is set true if the index is partial and its predicate is satisfied
4007 : * for this query, ie the query's WHERE clauses imply the predicate.
4008 : *
4009 : * indrestrictinfo is set to the relation's baserestrictinfo list less any
4010 : * conditions that are implied by the index's predicate. (Obviously, for a
4011 : * non-partial index, this is the same as baserestrictinfo.) Such conditions
4012 : * can be dropped from the plan when using the index, in certain cases.
4013 : *
4014 : * At one time it was possible for this to get re-run after adding more
4015 : * restrictions to the rel, thus possibly letting us prove more indexes OK.
4016 : * That doesn't happen any more (at least not in the core code's usage),
4017 : * but this code still supports it in case extensions want to mess with the
4018 : * baserestrictinfo list. We assume that adding more restrictions can't make
4019 : * an index not predOK. We must recompute indrestrictinfo each time, though,
4020 : * to make sure any newly-added restrictions get into it if needed.
4021 : */
4022 : void
4023 384982 : check_index_predicates(PlannerInfo *root, RelOptInfo *rel)
4024 : {
4025 : List *clauselist;
4026 : bool have_partial;
4027 : bool is_target_rel;
4028 : Relids otherrels;
4029 : ListCell *lc;
4030 :
4031 : /* Indexes are available only on base or "other" member relations. */
4032 : Assert(IS_SIMPLE_REL(rel));
4033 :
4034 : /*
4035 : * Initialize the indrestrictinfo lists to be identical to
4036 : * baserestrictinfo, and check whether there are any partial indexes. If
4037 : * not, this is all we need to do.
4038 : */
4039 384982 : have_partial = false;
4040 1051072 : foreach(lc, rel->indexlist)
4041 : {
4042 666090 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
4043 :
4044 666090 : index->indrestrictinfo = rel->baserestrictinfo;
4045 666090 : if (index->indpred)
4046 984 : have_partial = true;
4047 : }
4048 384982 : if (!have_partial)
4049 384322 : return;
4050 :
4051 : /*
4052 : * Construct a list of clauses that we can assume true for the purpose of
4053 : * proving the index(es) usable. Restriction clauses for the rel are
4054 : * always usable, and so are any join clauses that are "movable to" this
4055 : * rel. Also, we can consider any EC-derivable join clauses (which must
4056 : * be "movable to" this rel, by definition).
4057 : */
4058 660 : clauselist = list_copy(rel->baserestrictinfo);
4059 :
4060 : /* Scan the rel's join clauses */
4061 660 : foreach(lc, rel->joininfo)
4062 : {
4063 0 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4064 :
4065 : /* Check if clause can be moved to this rel */
4066 0 : if (!join_clause_is_movable_to(rinfo, rel))
4067 0 : continue;
4068 :
4069 0 : clauselist = lappend(clauselist, rinfo);
4070 : }
4071 :
4072 : /*
4073 : * Add on any equivalence-derivable join clauses. Computing the correct
4074 : * relid sets for generate_join_implied_equalities is slightly tricky
4075 : * because the rel could be a child rel rather than a true baserel, and in
4076 : * that case we must subtract its parents' relid(s) from all_query_rels.
4077 : * Additionally, we mustn't consider clauses that are only computable
4078 : * after outer joins that can null the rel.
4079 : */
4080 660 : if (rel->reloptkind == RELOPT_OTHER_MEMBER_REL)
4081 72 : otherrels = bms_difference(root->all_query_rels,
4082 72 : find_childrel_parents(root, rel));
4083 : else
4084 588 : otherrels = bms_difference(root->all_query_rels, rel->relids);
4085 660 : otherrels = bms_del_members(otherrels, rel->nulling_relids);
4086 :
4087 660 : if (!bms_is_empty(otherrels))
4088 : clauselist =
4089 88 : list_concat(clauselist,
4090 88 : generate_join_implied_equalities(root,
4091 88 : bms_union(rel->relids,
4092 : otherrels),
4093 : otherrels,
4094 : rel,
4095 : NULL));
4096 :
4097 : /*
4098 : * Normally we remove quals that are implied by a partial index's
4099 : * predicate from indrestrictinfo, indicating that they need not be
4100 : * checked explicitly by an indexscan plan using this index. However, if
4101 : * the rel is a target relation of UPDATE/DELETE/MERGE/SELECT FOR UPDATE,
4102 : * we cannot remove such quals from the plan, because they need to be in
4103 : * the plan so that they will be properly rechecked by EvalPlanQual
4104 : * testing. Some day we might want to remove such quals from the main
4105 : * plan anyway and pass them through to EvalPlanQual via a side channel;
4106 : * but for now, we just don't remove implied quals at all for target
4107 : * relations.
4108 : */
4109 1208 : is_target_rel = (bms_is_member(rel->relid, root->all_result_relids) ||
4110 548 : get_plan_rowmark(root->rowMarks, rel->relid) != NULL);
4111 :
4112 : /*
4113 : * Now try to prove each index predicate true, and compute the
4114 : * indrestrictinfo lists for partial indexes. Note that we compute the
4115 : * indrestrictinfo list even for non-predOK indexes; this might seem
4116 : * wasteful, but we may be able to use such indexes in OR clauses, cf
4117 : * generate_bitmap_or_paths().
4118 : */
4119 2030 : foreach(lc, rel->indexlist)
4120 : {
4121 1370 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
4122 : ListCell *lcr;
4123 :
4124 1370 : if (index->indpred == NIL)
4125 386 : continue; /* ignore non-partial indexes here */
4126 :
4127 984 : if (!index->predOK) /* don't repeat work if already proven OK */
4128 984 : index->predOK = predicate_implied_by(index->indpred, clauselist,
4129 : false);
4130 :
4131 : /* If rel is an update target, leave indrestrictinfo as set above */
4132 984 : if (is_target_rel)
4133 172 : continue;
4134 :
4135 : /* Else compute indrestrictinfo as the non-implied quals */
4136 812 : index->indrestrictinfo = NIL;
4137 1914 : foreach(lcr, rel->baserestrictinfo)
4138 : {
4139 1102 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lcr);
4140 :
4141 : /* predicate_implied_by() assumes first arg is immutable */
4142 1102 : if (contain_mutable_functions((Node *) rinfo->clause) ||
4143 1102 : !predicate_implied_by(list_make1(rinfo->clause),
4144 : index->indpred, false))
4145 782 : index->indrestrictinfo = lappend(index->indrestrictinfo, rinfo);
4146 : }
4147 : }
4148 : }
4149 :
4150 : /****************************************************************************
4151 : * ---- ROUTINES TO CHECK EXTERNALLY-VISIBLE CONDITIONS ----
4152 : ****************************************************************************/
4153 :
4154 : /*
4155 : * ec_member_matches_indexcol
4156 : * Test whether an EquivalenceClass member matches an index column.
4157 : *
4158 : * This is a callback for use by generate_implied_equalities_for_column.
4159 : */
4160 : static bool
4161 392032 : ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
4162 : EquivalenceClass *ec, EquivalenceMember *em,
4163 : void *arg)
4164 : {
4165 392032 : IndexOptInfo *index = ((ec_member_matches_arg *) arg)->index;
4166 392032 : int indexcol = ((ec_member_matches_arg *) arg)->indexcol;
4167 : Oid curFamily;
4168 : Oid curCollation;
4169 :
4170 : Assert(indexcol < index->nkeycolumns);
4171 :
4172 392032 : curFamily = index->opfamily[indexcol];
4173 392032 : curCollation = index->indexcollations[indexcol];
4174 :
4175 : /*
4176 : * If it's a btree index, we can reject it if its opfamily isn't
4177 : * compatible with the EC, since no clause generated from the EC could be
4178 : * used with the index. For non-btree indexes, we can't easily tell
4179 : * whether clauses generated from the EC could be used with the index, so
4180 : * don't check the opfamily. This might mean we return "true" for a
4181 : * useless EC, so we have to recheck the results of
4182 : * generate_implied_equalities_for_column; see
4183 : * match_eclass_clauses_to_index.
4184 : */
4185 392032 : if (index->relam == BTREE_AM_OID &&
4186 391990 : !list_member_oid(ec->ec_opfamilies, curFamily))
4187 121648 : return false;
4188 :
4189 : /* We insist on collation match for all index types, though */
4190 270384 : if (!IndexCollMatchesExprColl(curCollation, ec->ec_collation))
4191 14 : return false;
4192 :
4193 270370 : return match_index_to_operand((Node *) em->em_expr, indexcol, index);
4194 : }
4195 :
4196 : /*
4197 : * relation_has_unique_index_for
4198 : * Determine whether the relation provably has at most one row satisfying
4199 : * a set of equality conditions, because the conditions constrain all
4200 : * columns of some unique index.
4201 : *
4202 : * The conditions can be represented in either or both of two ways:
4203 : * 1. A list of RestrictInfo nodes, where the caller has already determined
4204 : * that each condition is a mergejoinable equality with an expression in
4205 : * this relation on one side, and an expression not involving this relation
4206 : * on the other. The transient outer_is_left flag is used to identify which
4207 : * side we should look at: left side if outer_is_left is false, right side
4208 : * if it is true.
4209 : * 2. A list of expressions in this relation, and a corresponding list of
4210 : * equality operators. The caller must have already checked that the operators
4211 : * represent equality. (Note: the operators could be cross-type; the
4212 : * expressions should correspond to their RHS inputs.)
4213 : *
4214 : * The caller need only supply equality conditions arising from joins;
4215 : * this routine automatically adds in any usable baserestrictinfo clauses.
4216 : * (Note that the passed-in restrictlist will be destructively modified!)
4217 : */
4218 : bool
4219 936 : relation_has_unique_index_for(PlannerInfo *root, RelOptInfo *rel,
4220 : List *restrictlist,
4221 : List *exprlist, List *oprlist)
4222 : {
4223 936 : return relation_has_unique_index_ext(root, rel, restrictlist,
4224 : exprlist, oprlist, NULL);
4225 : }
4226 :
4227 : /*
4228 : * relation_has_unique_index_ext
4229 : * Same as relation_has_unique_index_for(), but supports extra_clauses
4230 : * parameter. If extra_clauses isn't NULL, return baserestrictinfo clauses
4231 : * which were used to derive uniqueness.
4232 : */
4233 : bool
4234 193306 : relation_has_unique_index_ext(PlannerInfo *root, RelOptInfo *rel,
4235 : List *restrictlist,
4236 : List *exprlist, List *oprlist,
4237 : List **extra_clauses)
4238 : {
4239 : ListCell *ic;
4240 :
4241 : Assert(list_length(exprlist) == list_length(oprlist));
4242 :
4243 : /* Short-circuit if no indexes... */
4244 193306 : if (rel->indexlist == NIL)
4245 466 : return false;
4246 :
4247 : /*
4248 : * Examine the rel's restriction clauses for usable var = const clauses
4249 : * that we can add to the restrictlist.
4250 : */
4251 319840 : foreach(ic, rel->baserestrictinfo)
4252 : {
4253 127000 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(ic);
4254 :
4255 : /*
4256 : * Note: can_join won't be set for a restriction clause, but
4257 : * mergeopfamilies will be if it has a mergejoinable operator and
4258 : * doesn't contain volatile functions.
4259 : */
4260 127000 : if (restrictinfo->mergeopfamilies == NIL)
4261 56612 : continue; /* not mergejoinable */
4262 :
4263 : /*
4264 : * The clause certainly doesn't refer to anything but the given rel.
4265 : * If either side is pseudoconstant then we can use it.
4266 : */
4267 70388 : if (bms_is_empty(restrictinfo->left_relids))
4268 : {
4269 : /* righthand side is inner */
4270 58 : restrictinfo->outer_is_left = true;
4271 : }
4272 70330 : else if (bms_is_empty(restrictinfo->right_relids))
4273 : {
4274 : /* lefthand side is inner */
4275 70204 : restrictinfo->outer_is_left = false;
4276 : }
4277 : else
4278 126 : continue;
4279 :
4280 : /* OK, add to list */
4281 70262 : restrictlist = lappend(restrictlist, restrictinfo);
4282 : }
4283 :
4284 : /* Short-circuit the easy case */
4285 192840 : if (restrictlist == NIL && exprlist == NIL)
4286 978 : return false;
4287 :
4288 : /* Examine each index of the relation ... */
4289 480330 : foreach(ic, rel->indexlist)
4290 : {
4291 402326 : IndexOptInfo *ind = (IndexOptInfo *) lfirst(ic);
4292 : int c;
4293 402326 : List *exprs = NIL;
4294 :
4295 : /*
4296 : * If the index is not unique, or not immediately enforced, or if it's
4297 : * a partial index, it's useless here. We're unable to make use of
4298 : * predOK partial unique indexes due to the fact that
4299 : * check_index_predicates() also makes use of join predicates to
4300 : * determine if the partial index is usable. Here we need proofs that
4301 : * hold true before any joins are evaluated.
4302 : */
4303 402326 : if (!ind->unique || !ind->immediate || ind->indpred != NIL)
4304 106004 : continue;
4305 :
4306 : /*
4307 : * Try to find each index column in the lists of conditions. This is
4308 : * O(N^2) or worse, but we expect all the lists to be short.
4309 : */
4310 494088 : for (c = 0; c < ind->nkeycolumns; c++)
4311 : {
4312 380230 : bool matched = false;
4313 : ListCell *lc;
4314 : ListCell *lc2;
4315 :
4316 707598 : foreach(lc, restrictlist)
4317 : {
4318 525134 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4319 : Node *rexpr;
4320 :
4321 : /*
4322 : * The condition's equality operator must be a member of the
4323 : * index opfamily, else it is not asserting the right kind of
4324 : * equality behavior for this index. We check this first
4325 : * since it's probably cheaper than match_index_to_operand().
4326 : */
4327 525134 : if (!list_member_oid(rinfo->mergeopfamilies, ind->opfamily[c]))
4328 150620 : continue;
4329 :
4330 : /*
4331 : * XXX at some point we may need to check collations here too.
4332 : * For the moment we assume all collations reduce to the same
4333 : * notion of equality.
4334 : */
4335 :
4336 : /* OK, see if the condition operand matches the index key */
4337 374514 : if (rinfo->outer_is_left)
4338 165412 : rexpr = get_rightop(rinfo->clause);
4339 : else
4340 209102 : rexpr = get_leftop(rinfo->clause);
4341 :
4342 374514 : if (match_index_to_operand(rexpr, c, ind))
4343 : {
4344 197766 : matched = true; /* column is unique */
4345 :
4346 197766 : if (bms_membership(rinfo->clause_relids) == BMS_SINGLETON)
4347 : {
4348 : MemoryContext oldMemCtx =
4349 49568 : MemoryContextSwitchTo(root->planner_cxt);
4350 :
4351 : /*
4352 : * Add filter clause into a list allowing caller to
4353 : * know if uniqueness have made not only by join
4354 : * clauses.
4355 : */
4356 : Assert(bms_is_empty(rinfo->left_relids) ||
4357 : bms_is_empty(rinfo->right_relids));
4358 49568 : if (extra_clauses)
4359 144 : exprs = lappend(exprs, rinfo);
4360 49568 : MemoryContextSwitchTo(oldMemCtx);
4361 : }
4362 :
4363 197766 : break;
4364 : }
4365 : }
4366 :
4367 380230 : if (matched)
4368 197766 : continue;
4369 :
4370 182622 : forboth(lc, exprlist, lc2, oprlist)
4371 : {
4372 158 : Node *expr = (Node *) lfirst(lc);
4373 158 : Oid opr = lfirst_oid(lc2);
4374 :
4375 : /* See if the expression matches the index key */
4376 158 : if (!match_index_to_operand(expr, c, ind))
4377 158 : continue;
4378 :
4379 : /*
4380 : * The equality operator must be a member of the index
4381 : * opfamily, else it is not asserting the right kind of
4382 : * equality behavior for this index. We assume the caller
4383 : * determined it is an equality operator, so we don't need to
4384 : * check any more tightly than this.
4385 : */
4386 0 : if (!op_in_opfamily(opr, ind->opfamily[c]))
4387 0 : continue;
4388 :
4389 : /*
4390 : * XXX at some point we may need to check collations here too.
4391 : * For the moment we assume all collations reduce to the same
4392 : * notion of equality.
4393 : */
4394 :
4395 0 : matched = true; /* column is unique */
4396 0 : break;
4397 : }
4398 :
4399 182464 : if (!matched)
4400 182464 : break; /* no match; this index doesn't help us */
4401 : }
4402 :
4403 : /* Matched all key columns of this index? */
4404 296322 : if (c == ind->nkeycolumns)
4405 : {
4406 113858 : if (extra_clauses)
4407 642 : *extra_clauses = exprs;
4408 113858 : return true;
4409 : }
4410 : }
4411 :
4412 78004 : return false;
4413 : }
4414 :
4415 : /*
4416 : * indexcol_is_bool_constant_for_query
4417 : *
4418 : * If an index column is constrained to have a constant value by the query's
4419 : * WHERE conditions, then it's irrelevant for sort-order considerations.
4420 : * Usually that means we have a restriction clause WHERE indexcol = constant,
4421 : * which gets turned into an EquivalenceClass containing a constant, which
4422 : * is recognized as redundant by build_index_pathkeys(). But if the index
4423 : * column is a boolean variable (or expression), then we are not going to
4424 : * see WHERE indexcol = constant, because expression preprocessing will have
4425 : * simplified that to "WHERE indexcol" or "WHERE NOT indexcol". So we are not
4426 : * going to have a matching EquivalenceClass (unless the query also contains
4427 : * "ORDER BY indexcol"). To allow such cases to work the same as they would
4428 : * for non-boolean values, this function is provided to detect whether the
4429 : * specified index column matches a boolean restriction clause.
4430 : */
4431 : bool
4432 573588 : indexcol_is_bool_constant_for_query(PlannerInfo *root,
4433 : IndexOptInfo *index,
4434 : int indexcol)
4435 : {
4436 : ListCell *lc;
4437 :
4438 : /* If the index isn't boolean, we can't possibly get a match */
4439 573588 : if (!IsBooleanOpfamily(index->opfamily[indexcol]))
4440 566124 : return false;
4441 :
4442 : /* Check each restriction clause for the index's rel */
4443 7500 : foreach(lc, index->rel->baserestrictinfo)
4444 : {
4445 1260 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4446 :
4447 : /*
4448 : * As in match_clause_to_indexcol, never match pseudoconstants to
4449 : * indexes. (It might be semantically okay to do so here, but the
4450 : * odds of getting a match are negligible, so don't waste the cycles.)
4451 : */
4452 1260 : if (rinfo->pseudoconstant)
4453 0 : continue;
4454 :
4455 : /* See if we can match the clause's expression to the index column */
4456 1260 : if (match_boolean_index_clause(root, rinfo, indexcol, index))
4457 1224 : return true;
4458 : }
4459 :
4460 6240 : return false;
4461 : }
4462 :
4463 :
4464 : /****************************************************************************
4465 : * ---- ROUTINES TO CHECK OPERANDS ----
4466 : ****************************************************************************/
4467 :
4468 : /*
4469 : * match_index_to_operand()
4470 : * Generalized test for a match between an index's key
4471 : * and the operand on one side of a restriction or join clause.
4472 : *
4473 : * operand: the nodetree to be compared to the index
4474 : * indexcol: the column number of the index (counting from 0)
4475 : * index: the index of interest
4476 : *
4477 : * Note that we aren't interested in collations here; the caller must check
4478 : * for a collation match, if it's dealing with an operator where that matters.
4479 : *
4480 : * This is exported for use in selfuncs.c.
4481 : */
4482 : bool
4483 3374544 : match_index_to_operand(Node *operand,
4484 : int indexcol,
4485 : IndexOptInfo *index)
4486 : {
4487 : int indkey;
4488 :
4489 : /*
4490 : * Ignore any RelabelType node above the operand. This is needed to be
4491 : * able to apply indexscanning in binary-compatible-operator cases. Note:
4492 : * we can assume there is at most one RelabelType node;
4493 : * eval_const_expressions() will have simplified if more than one.
4494 : */
4495 3374544 : if (operand && IsA(operand, RelabelType))
4496 23272 : operand = (Node *) ((RelabelType *) operand)->arg;
4497 :
4498 3374544 : indkey = index->indexkeys[indexcol];
4499 3374544 : if (indkey != 0)
4500 : {
4501 : /*
4502 : * Simple index column; operand must be a matching Var.
4503 : */
4504 3368558 : if (operand && IsA(operand, Var) &&
4505 2502608 : index->rel->relid == ((Var *) operand)->varno &&
4506 2313386 : indkey == ((Var *) operand)->varattno &&
4507 832336 : ((Var *) operand)->varnullingrels == NULL)
4508 831494 : return true;
4509 : }
4510 : else
4511 : {
4512 : /*
4513 : * Index expression; find the correct expression. (This search could
4514 : * be avoided, at the cost of complicating all the callers of this
4515 : * routine; doesn't seem worth it.)
4516 : */
4517 : ListCell *indexpr_item;
4518 : int i;
4519 : Node *indexkey;
4520 :
4521 5986 : indexpr_item = list_head(index->indexprs);
4522 5986 : for (i = 0; i < indexcol; i++)
4523 : {
4524 0 : if (index->indexkeys[i] == 0)
4525 : {
4526 0 : if (indexpr_item == NULL)
4527 0 : elog(ERROR, "wrong number of index expressions");
4528 0 : indexpr_item = lnext(index->indexprs, indexpr_item);
4529 : }
4530 : }
4531 5986 : if (indexpr_item == NULL)
4532 0 : elog(ERROR, "wrong number of index expressions");
4533 5986 : indexkey = (Node *) lfirst(indexpr_item);
4534 :
4535 : /*
4536 : * Does it match the operand? Again, strip any relabeling.
4537 : */
4538 5986 : if (indexkey && IsA(indexkey, RelabelType))
4539 10 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
4540 :
4541 5986 : if (equal(indexkey, operand))
4542 2164 : return true;
4543 : }
4544 :
4545 2540886 : return false;
4546 : }
4547 :
4548 : /*
4549 : * is_pseudo_constant_for_index()
4550 : * Test whether the given expression can be used as an indexscan
4551 : * comparison value.
4552 : *
4553 : * An indexscan comparison value must not contain any volatile functions,
4554 : * and it can't contain any Vars of the index's own table. Vars of
4555 : * other tables are okay, though; in that case we'd be producing an
4556 : * indexqual usable in a parameterized indexscan. This is, therefore,
4557 : * a weaker condition than is_pseudo_constant_clause().
4558 : *
4559 : * This function is exported for use by planner support functions,
4560 : * which will have available the IndexOptInfo, but not any RestrictInfo
4561 : * infrastructure. It is making the same test made by functions above
4562 : * such as match_opclause_to_indexcol(), but those rely where possible
4563 : * on RestrictInfo information about variable membership.
4564 : *
4565 : * expr: the nodetree to be checked
4566 : * index: the index of interest
4567 : */
4568 : bool
4569 0 : is_pseudo_constant_for_index(PlannerInfo *root, Node *expr, IndexOptInfo *index)
4570 : {
4571 : /* pull_varnos is cheaper than volatility check, so do that first */
4572 0 : if (bms_is_member(index->rel->relid, pull_varnos(root, expr)))
4573 0 : return false; /* no good, contains Var of table */
4574 0 : if (contain_volatile_functions(expr))
4575 0 : return false; /* no good, volatile comparison value */
4576 0 : return true;
4577 : }
|