Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * indxpath.c
4 : * Routines to determine which indexes are usable for scanning a
5 : * given relation, and create Paths accordingly.
6 : *
7 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : *
11 : * IDENTIFICATION
12 : * src/backend/optimizer/path/indxpath.c
13 : *
14 : *-------------------------------------------------------------------------
15 : */
16 : #include "postgres.h"
17 :
18 : #include <math.h>
19 :
20 : #include "access/stratnum.h"
21 : #include "access/sysattr.h"
22 : #include "access/transam.h"
23 : #include "catalog/pg_am.h"
24 : #include "catalog/pg_amop.h"
25 : #include "catalog/pg_operator.h"
26 : #include "catalog/pg_opfamily.h"
27 : #include "catalog/pg_type.h"
28 : #include "nodes/makefuncs.h"
29 : #include "nodes/nodeFuncs.h"
30 : #include "nodes/supportnodes.h"
31 : #include "optimizer/cost.h"
32 : #include "optimizer/optimizer.h"
33 : #include "optimizer/pathnode.h"
34 : #include "optimizer/paths.h"
35 : #include "optimizer/prep.h"
36 : #include "optimizer/restrictinfo.h"
37 : #include "utils/lsyscache.h"
38 : #include "utils/selfuncs.h"
39 :
40 :
41 : /* XXX see PartCollMatchesExprColl */
42 : #define IndexCollMatchesExprColl(idxcollation, exprcollation) \
43 : ((idxcollation) == InvalidOid || (idxcollation) == (exprcollation))
44 :
45 : /* Whether we are looking for plain indexscan, bitmap scan, or either */
46 : typedef enum
47 : {
48 : ST_INDEXSCAN, /* must support amgettuple */
49 : ST_BITMAPSCAN, /* must support amgetbitmap */
50 : ST_ANYSCAN, /* either is okay */
51 : } ScanTypeControl;
52 :
53 : /* Data structure for collecting qual clauses that match an index */
54 : typedef struct
55 : {
56 : bool nonempty; /* True if lists are not all empty */
57 : /* Lists of IndexClause nodes, one list per index column */
58 : List *indexclauses[INDEX_MAX_KEYS];
59 : } IndexClauseSet;
60 :
61 : /* Per-path data used within choose_bitmap_and() */
62 : typedef struct
63 : {
64 : Path *path; /* IndexPath, BitmapAndPath, or BitmapOrPath */
65 : List *quals; /* the WHERE clauses it uses */
66 : List *preds; /* predicates of its partial index(es) */
67 : Bitmapset *clauseids; /* quals+preds represented as a bitmapset */
68 : bool unclassifiable; /* has too many quals+preds to process? */
69 : } PathClauseUsage;
70 :
71 : /* Callback argument for ec_member_matches_indexcol */
72 : typedef struct
73 : {
74 : IndexOptInfo *index; /* index we're considering */
75 : int indexcol; /* index column we want to match to */
76 : } ec_member_matches_arg;
77 :
78 :
79 : static void consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
80 : IndexOptInfo *index,
81 : IndexClauseSet *rclauseset,
82 : IndexClauseSet *jclauseset,
83 : IndexClauseSet *eclauseset,
84 : List **bitindexpaths);
85 : static void consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
86 : IndexOptInfo *index,
87 : IndexClauseSet *rclauseset,
88 : IndexClauseSet *jclauseset,
89 : IndexClauseSet *eclauseset,
90 : List **bitindexpaths,
91 : List *indexjoinclauses,
92 : int considered_clauses,
93 : List **considered_relids);
94 : static void get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
95 : IndexOptInfo *index,
96 : IndexClauseSet *rclauseset,
97 : IndexClauseSet *jclauseset,
98 : IndexClauseSet *eclauseset,
99 : List **bitindexpaths,
100 : Relids relids,
101 : List **considered_relids);
102 : static bool eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
103 : List *indexjoinclauses);
104 : static void get_index_paths(PlannerInfo *root, RelOptInfo *rel,
105 : IndexOptInfo *index, IndexClauseSet *clauses,
106 : List **bitindexpaths);
107 : static List *build_index_paths(PlannerInfo *root, RelOptInfo *rel,
108 : IndexOptInfo *index, IndexClauseSet *clauses,
109 : bool useful_predicate,
110 : ScanTypeControl scantype,
111 : bool *skip_nonnative_saop);
112 : static List *build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
113 : List *clauses, List *other_clauses);
114 : static List *generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
115 : List *clauses, List *other_clauses);
116 : static Path *choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel,
117 : List *paths);
118 : static int path_usage_comparator(const void *a, const void *b);
119 : static Cost bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel,
120 : Path *ipath);
121 : static Cost bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel,
122 : List *paths);
123 : static PathClauseUsage *classify_index_clause_usage(Path *path,
124 : List **clauselist);
125 : static void find_indexpath_quals(Path *bitmapqual, List **quals, List **preds);
126 : static int find_list_position(Node *node, List **nodelist);
127 : static bool check_index_only(RelOptInfo *rel, IndexOptInfo *index);
128 : static double get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids);
129 : static double adjust_rowcount_for_semijoins(PlannerInfo *root,
130 : Index cur_relid,
131 : Index outer_relid,
132 : double rowcount);
133 : static double approximate_joinrel_size(PlannerInfo *root, Relids relids);
134 : static void match_restriction_clauses_to_index(PlannerInfo *root,
135 : IndexOptInfo *index,
136 : IndexClauseSet *clauseset);
137 : static void match_join_clauses_to_index(PlannerInfo *root,
138 : RelOptInfo *rel, IndexOptInfo *index,
139 : IndexClauseSet *clauseset,
140 : List **joinorclauses);
141 : static void match_eclass_clauses_to_index(PlannerInfo *root,
142 : IndexOptInfo *index,
143 : IndexClauseSet *clauseset);
144 : static void match_clauses_to_index(PlannerInfo *root,
145 : List *clauses,
146 : IndexOptInfo *index,
147 : IndexClauseSet *clauseset);
148 : static void match_clause_to_index(PlannerInfo *root,
149 : RestrictInfo *rinfo,
150 : IndexOptInfo *index,
151 : IndexClauseSet *clauseset);
152 : static IndexClause *match_clause_to_indexcol(PlannerInfo *root,
153 : RestrictInfo *rinfo,
154 : int indexcol,
155 : IndexOptInfo *index);
156 : static bool IsBooleanOpfamily(Oid opfamily);
157 : static IndexClause *match_boolean_index_clause(PlannerInfo *root,
158 : RestrictInfo *rinfo,
159 : int indexcol, IndexOptInfo *index);
160 : static IndexClause *match_opclause_to_indexcol(PlannerInfo *root,
161 : RestrictInfo *rinfo,
162 : int indexcol,
163 : IndexOptInfo *index);
164 : static IndexClause *match_funcclause_to_indexcol(PlannerInfo *root,
165 : RestrictInfo *rinfo,
166 : int indexcol,
167 : IndexOptInfo *index);
168 : static IndexClause *get_index_clause_from_support(PlannerInfo *root,
169 : RestrictInfo *rinfo,
170 : Oid funcid,
171 : int indexarg,
172 : int indexcol,
173 : IndexOptInfo *index);
174 : static IndexClause *match_saopclause_to_indexcol(PlannerInfo *root,
175 : RestrictInfo *rinfo,
176 : int indexcol,
177 : IndexOptInfo *index);
178 : static IndexClause *match_rowcompare_to_indexcol(PlannerInfo *root,
179 : RestrictInfo *rinfo,
180 : int indexcol,
181 : IndexOptInfo *index);
182 : static IndexClause *match_orclause_to_indexcol(PlannerInfo *root,
183 : RestrictInfo *rinfo,
184 : int indexcol,
185 : IndexOptInfo *index);
186 : static IndexClause *expand_indexqual_rowcompare(PlannerInfo *root,
187 : RestrictInfo *rinfo,
188 : int indexcol,
189 : IndexOptInfo *index,
190 : Oid expr_op,
191 : bool var_on_left);
192 : static void match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
193 : List **orderby_clauses_p,
194 : List **clause_columns_p);
195 : static Expr *match_clause_to_ordering_op(IndexOptInfo *index,
196 : int indexcol, Expr *clause, Oid pk_opfamily);
197 : static bool ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
198 : EquivalenceClass *ec, EquivalenceMember *em,
199 : void *arg);
200 :
201 :
202 : /*
203 : * create_index_paths()
204 : * Generate all interesting index paths for the given relation.
205 : * Candidate paths are added to the rel's pathlist (using add_path).
206 : *
207 : * To be considered for an index scan, an index must match one or more
208 : * restriction clauses or join clauses from the query's qual condition,
209 : * or match the query's ORDER BY condition, or have a predicate that
210 : * matches the query's qual condition.
211 : *
212 : * There are two basic kinds of index scans. A "plain" index scan uses
213 : * only restriction clauses (possibly none at all) in its indexqual,
214 : * so it can be applied in any context. A "parameterized" index scan uses
215 : * join clauses (plus restriction clauses, if available) in its indexqual.
216 : * When joining such a scan to one of the relations supplying the other
217 : * variables used in its indexqual, the parameterized scan must appear as
218 : * the inner relation of a nestloop join; it can't be used on the outer side,
219 : * nor in a merge or hash join. In that context, values for the other rels'
220 : * attributes are available and fixed during any one scan of the indexpath.
221 : *
222 : * An IndexPath is generated and submitted to add_path() for each plain or
223 : * parameterized index scan this routine deems potentially interesting for
224 : * the current query.
225 : *
226 : * 'rel' is the relation for which we want to generate index paths
227 : *
228 : * Note: check_index_predicates() must have been run previously for this rel.
229 : *
230 : * Note: in cases involving LATERAL references in the relation's tlist, it's
231 : * possible that rel->lateral_relids is nonempty. Currently, we include
232 : * lateral_relids into the parameterization reported for each path, but don't
233 : * take it into account otherwise. The fact that any such rels *must* be
234 : * available as parameter sources perhaps should influence our choices of
235 : * index quals ... but for now, it doesn't seem worth troubling over.
236 : * In particular, comments below about "unparameterized" paths should be read
237 : * as meaning "unparameterized so far as the indexquals are concerned".
238 : */
239 : void
240 416526 : create_index_paths(PlannerInfo *root, RelOptInfo *rel)
241 : {
242 : List *indexpaths;
243 : List *bitindexpaths;
244 : List *bitjoinpaths;
245 : List *joinorclauses;
246 : IndexClauseSet rclauseset;
247 : IndexClauseSet jclauseset;
248 : IndexClauseSet eclauseset;
249 : ListCell *lc;
250 :
251 : /* Skip the whole mess if no indexes */
252 416526 : if (rel->indexlist == NIL)
253 72174 : return;
254 :
255 : /* Bitmap paths are collected and then dealt with at the end */
256 344352 : bitindexpaths = bitjoinpaths = joinorclauses = NIL;
257 :
258 : /* Examine each index in turn */
259 1077616 : foreach(lc, rel->indexlist)
260 : {
261 733264 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
262 :
263 : /* Protect limited-size array in IndexClauseSets */
264 : Assert(index->nkeycolumns <= INDEX_MAX_KEYS);
265 :
266 : /*
267 : * Ignore partial indexes that do not match the query.
268 : * (generate_bitmap_or_paths() might be able to do something with
269 : * them, but that's of no concern here.)
270 : */
271 733264 : if (index->indpred != NIL && !index->predOK)
272 538 : continue;
273 :
274 : /*
275 : * Identify the restriction clauses that can match the index.
276 : */
277 24912684 : MemSet(&rclauseset, 0, sizeof(rclauseset));
278 732726 : match_restriction_clauses_to_index(root, index, &rclauseset);
279 :
280 : /*
281 : * Build index paths from the restriction clauses. These will be
282 : * non-parameterized paths. Plain paths go directly to add_path(),
283 : * bitmap paths are added to bitindexpaths to be handled below.
284 : */
285 732726 : get_index_paths(root, rel, index, &rclauseset,
286 : &bitindexpaths);
287 :
288 : /*
289 : * Identify the join clauses that can match the index. For the moment
290 : * we keep them separate from the restriction clauses. Note that this
291 : * step finds only "loose" join clauses that have not been merged into
292 : * EquivalenceClasses. Also, collect join OR clauses for later.
293 : */
294 24912684 : MemSet(&jclauseset, 0, sizeof(jclauseset));
295 732726 : match_join_clauses_to_index(root, rel, index,
296 : &jclauseset, &joinorclauses);
297 :
298 : /*
299 : * Look for EquivalenceClasses that can generate joinclauses matching
300 : * the index.
301 : */
302 24912684 : MemSet(&eclauseset, 0, sizeof(eclauseset));
303 732726 : match_eclass_clauses_to_index(root, index,
304 : &eclauseset);
305 :
306 : /*
307 : * If we found any plain or eclass join clauses, build parameterized
308 : * index paths using them.
309 : */
310 732726 : if (jclauseset.nonempty || eclauseset.nonempty)
311 142526 : consider_index_join_clauses(root, rel, index,
312 : &rclauseset,
313 : &jclauseset,
314 : &eclauseset,
315 : &bitjoinpaths);
316 : }
317 :
318 : /*
319 : * Generate BitmapOrPaths for any suitable OR-clauses present in the
320 : * restriction list. Add these to bitindexpaths.
321 : */
322 344352 : indexpaths = generate_bitmap_or_paths(root, rel,
323 : rel->baserestrictinfo, NIL);
324 344352 : bitindexpaths = list_concat(bitindexpaths, indexpaths);
325 :
326 : /*
327 : * Likewise, generate BitmapOrPaths for any suitable OR-clauses present in
328 : * the joinclause list. Add these to bitjoinpaths.
329 : */
330 344352 : indexpaths = generate_bitmap_or_paths(root, rel,
331 : joinorclauses, rel->baserestrictinfo);
332 344352 : bitjoinpaths = list_concat(bitjoinpaths, indexpaths);
333 :
334 : /*
335 : * If we found anything usable, generate a BitmapHeapPath for the most
336 : * promising combination of restriction bitmap index paths. Note there
337 : * will be only one such path no matter how many indexes exist. This
338 : * should be sufficient since there's basically only one figure of merit
339 : * (total cost) for such a path.
340 : */
341 344352 : if (bitindexpaths != NIL)
342 : {
343 : Path *bitmapqual;
344 : BitmapHeapPath *bpath;
345 :
346 208308 : bitmapqual = choose_bitmap_and(root, rel, bitindexpaths);
347 208308 : bpath = create_bitmap_heap_path(root, rel, bitmapqual,
348 : rel->lateral_relids, 1.0, 0);
349 208308 : add_path(rel, (Path *) bpath);
350 :
351 : /* create a partial bitmap heap path */
352 208308 : if (rel->consider_parallel && rel->lateral_relids == NULL)
353 146184 : create_partial_bitmap_paths(root, rel, bitmapqual);
354 : }
355 :
356 : /*
357 : * Likewise, if we found anything usable, generate BitmapHeapPaths for the
358 : * most promising combinations of join bitmap index paths. Our strategy
359 : * is to generate one such path for each distinct parameterization seen
360 : * among the available bitmap index paths. This may look pretty
361 : * expensive, but usually there won't be very many distinct
362 : * parameterizations. (This logic is quite similar to that in
363 : * consider_index_join_clauses, but we're working with whole paths not
364 : * individual clauses.)
365 : */
366 344352 : if (bitjoinpaths != NIL)
367 : {
368 : List *all_path_outers;
369 :
370 : /* Identify each distinct parameterization seen in bitjoinpaths */
371 129676 : all_path_outers = NIL;
372 286240 : foreach(lc, bitjoinpaths)
373 : {
374 156564 : Path *path = (Path *) lfirst(lc);
375 156564 : Relids required_outer = PATH_REQ_OUTER(path);
376 :
377 156564 : all_path_outers = list_append_unique(all_path_outers,
378 : required_outer);
379 : }
380 :
381 : /* Now, for each distinct parameterization set ... */
382 278942 : foreach(lc, all_path_outers)
383 : {
384 149266 : Relids max_outers = (Relids) lfirst(lc);
385 : List *this_path_set;
386 : Path *bitmapqual;
387 : Relids required_outer;
388 : double loop_count;
389 : BitmapHeapPath *bpath;
390 : ListCell *lcp;
391 :
392 : /* Identify all the bitmap join paths needing no more than that */
393 149266 : this_path_set = NIL;
394 358516 : foreach(lcp, bitjoinpaths)
395 : {
396 209250 : Path *path = (Path *) lfirst(lcp);
397 :
398 209250 : if (bms_is_subset(PATH_REQ_OUTER(path), max_outers))
399 163740 : this_path_set = lappend(this_path_set, path);
400 : }
401 :
402 : /*
403 : * Add in restriction bitmap paths, since they can be used
404 : * together with any join paths.
405 : */
406 149266 : this_path_set = list_concat(this_path_set, bitindexpaths);
407 :
408 : /* Select best AND combination for this parameterization */
409 149266 : bitmapqual = choose_bitmap_and(root, rel, this_path_set);
410 :
411 : /* And push that path into the mix */
412 149266 : required_outer = PATH_REQ_OUTER(bitmapqual);
413 149266 : loop_count = get_loop_count(root, rel->relid, required_outer);
414 149266 : bpath = create_bitmap_heap_path(root, rel, bitmapqual,
415 : required_outer, loop_count, 0);
416 149266 : add_path(rel, (Path *) bpath);
417 : }
418 : }
419 : }
420 :
421 : /*
422 : * consider_index_join_clauses
423 : * Given sets of join clauses for an index, decide which parameterized
424 : * index paths to build.
425 : *
426 : * Plain indexpaths are sent directly to add_path, while potential
427 : * bitmap indexpaths are added to *bitindexpaths for later processing.
428 : *
429 : * 'rel' is the index's heap relation
430 : * 'index' is the index for which we want to generate paths
431 : * 'rclauseset' is the collection of indexable restriction clauses
432 : * 'jclauseset' is the collection of indexable simple join clauses
433 : * 'eclauseset' is the collection of indexable clauses from EquivalenceClasses
434 : * '*bitindexpaths' is the list to add bitmap paths to
435 : */
436 : static void
437 142526 : consider_index_join_clauses(PlannerInfo *root, RelOptInfo *rel,
438 : IndexOptInfo *index,
439 : IndexClauseSet *rclauseset,
440 : IndexClauseSet *jclauseset,
441 : IndexClauseSet *eclauseset,
442 : List **bitindexpaths)
443 : {
444 142526 : int considered_clauses = 0;
445 142526 : List *considered_relids = NIL;
446 : int indexcol;
447 :
448 : /*
449 : * The strategy here is to identify every potentially useful set of outer
450 : * rels that can provide indexable join clauses. For each such set,
451 : * select all the join clauses available from those outer rels, add on all
452 : * the indexable restriction clauses, and generate plain and/or bitmap
453 : * index paths for that set of clauses. This is based on the assumption
454 : * that it's always better to apply a clause as an indexqual than as a
455 : * filter (qpqual); which is where an available clause would end up being
456 : * applied if we omit it from the indexquals.
457 : *
458 : * This looks expensive, but in most practical cases there won't be very
459 : * many distinct sets of outer rels to consider. As a safety valve when
460 : * that's not true, we use a heuristic: limit the number of outer rel sets
461 : * considered to a multiple of the number of clauses considered. (We'll
462 : * always consider using each individual join clause, though.)
463 : *
464 : * For simplicity in selecting relevant clauses, we represent each set of
465 : * outer rels as a maximum set of clause_relids --- that is, the indexed
466 : * relation itself is also included in the relids set. considered_relids
467 : * lists all relids sets we've already tried.
468 : */
469 362840 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
470 : {
471 : /* Consider each applicable simple join clause */
472 220314 : considered_clauses += list_length(jclauseset->indexclauses[indexcol]);
473 220314 : consider_index_join_outer_rels(root, rel, index,
474 : rclauseset, jclauseset, eclauseset,
475 : bitindexpaths,
476 : jclauseset->indexclauses[indexcol],
477 : considered_clauses,
478 : &considered_relids);
479 : /* Consider each applicable eclass join clause */
480 220314 : considered_clauses += list_length(eclauseset->indexclauses[indexcol]);
481 220314 : consider_index_join_outer_rels(root, rel, index,
482 : rclauseset, jclauseset, eclauseset,
483 : bitindexpaths,
484 : eclauseset->indexclauses[indexcol],
485 : considered_clauses,
486 : &considered_relids);
487 : }
488 142526 : }
489 :
490 : /*
491 : * consider_index_join_outer_rels
492 : * Generate parameterized paths based on clause relids in the clause list.
493 : *
494 : * Workhorse for consider_index_join_clauses; see notes therein for rationale.
495 : *
496 : * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset', and
497 : * 'bitindexpaths' as above
498 : * 'indexjoinclauses' is a list of IndexClauses for join clauses
499 : * 'considered_clauses' is the total number of clauses considered (so far)
500 : * '*considered_relids' is a list of all relids sets already considered
501 : */
502 : static void
503 440628 : consider_index_join_outer_rels(PlannerInfo *root, RelOptInfo *rel,
504 : IndexOptInfo *index,
505 : IndexClauseSet *rclauseset,
506 : IndexClauseSet *jclauseset,
507 : IndexClauseSet *eclauseset,
508 : List **bitindexpaths,
509 : List *indexjoinclauses,
510 : int considered_clauses,
511 : List **considered_relids)
512 : {
513 : ListCell *lc;
514 :
515 : /* Examine relids of each joinclause in the given list */
516 601762 : foreach(lc, indexjoinclauses)
517 : {
518 161134 : IndexClause *iclause = (IndexClause *) lfirst(lc);
519 161134 : Relids clause_relids = iclause->rinfo->clause_relids;
520 161134 : EquivalenceClass *parent_ec = iclause->rinfo->parent_ec;
521 : int num_considered_relids;
522 :
523 : /* If we already tried its relids set, no need to do so again */
524 161134 : if (list_member(*considered_relids, clause_relids))
525 8064 : continue;
526 :
527 : /*
528 : * Generate the union of this clause's relids set with each
529 : * previously-tried set. This ensures we try this clause along with
530 : * every interesting subset of previous clauses. However, to avoid
531 : * exponential growth of planning time when there are many clauses,
532 : * limit the number of relid sets accepted to 10 * considered_clauses.
533 : *
534 : * Note: get_join_index_paths appends entries to *considered_relids,
535 : * but we do not need to visit such newly-added entries within this
536 : * loop, so we don't use foreach() here. No real harm would be done
537 : * if we did visit them, since the subset check would reject them; but
538 : * it would waste some cycles.
539 : */
540 153070 : num_considered_relids = list_length(*considered_relids);
541 163980 : for (int pos = 0; pos < num_considered_relids; pos++)
542 : {
543 10910 : Relids oldrelids = (Relids) list_nth(*considered_relids, pos);
544 :
545 : /*
546 : * If either is a subset of the other, no new set is possible.
547 : * This isn't a complete test for redundancy, but it's easy and
548 : * cheap. get_join_index_paths will check more carefully if we
549 : * already generated the same relids set.
550 : */
551 10910 : if (bms_subset_compare(clause_relids, oldrelids) != BMS_DIFFERENT)
552 24 : continue;
553 :
554 : /*
555 : * If this clause was derived from an equivalence class, the
556 : * clause list may contain other clauses derived from the same
557 : * eclass. We should not consider that combining this clause with
558 : * one of those clauses generates a usefully different
559 : * parameterization; so skip if any clause derived from the same
560 : * eclass would already have been included when using oldrelids.
561 : */
562 21610 : if (parent_ec &&
563 10724 : eclass_already_used(parent_ec, oldrelids,
564 : indexjoinclauses))
565 7388 : continue;
566 :
567 : /*
568 : * If the number of relid sets considered exceeds our heuristic
569 : * limit, stop considering combinations of clauses. We'll still
570 : * consider the current clause alone, though (below this loop).
571 : */
572 3498 : if (list_length(*considered_relids) >= 10 * considered_clauses)
573 0 : break;
574 :
575 : /* OK, try the union set */
576 3498 : get_join_index_paths(root, rel, index,
577 : rclauseset, jclauseset, eclauseset,
578 : bitindexpaths,
579 : bms_union(clause_relids, oldrelids),
580 : considered_relids);
581 : }
582 :
583 : /* Also try this set of relids by itself */
584 153070 : get_join_index_paths(root, rel, index,
585 : rclauseset, jclauseset, eclauseset,
586 : bitindexpaths,
587 : clause_relids,
588 : considered_relids);
589 : }
590 440628 : }
591 :
592 : /*
593 : * get_join_index_paths
594 : * Generate index paths using clauses from the specified outer relations.
595 : * In addition to generating paths, relids is added to *considered_relids
596 : * if not already present.
597 : *
598 : * Workhorse for consider_index_join_clauses; see notes therein for rationale.
599 : *
600 : * 'rel', 'index', 'rclauseset', 'jclauseset', 'eclauseset',
601 : * 'bitindexpaths', 'considered_relids' as above
602 : * 'relids' is the current set of relids to consider (the target rel plus
603 : * one or more outer rels)
604 : */
605 : static void
606 156568 : get_join_index_paths(PlannerInfo *root, RelOptInfo *rel,
607 : IndexOptInfo *index,
608 : IndexClauseSet *rclauseset,
609 : IndexClauseSet *jclauseset,
610 : IndexClauseSet *eclauseset,
611 : List **bitindexpaths,
612 : Relids relids,
613 : List **considered_relids)
614 : {
615 : IndexClauseSet clauseset;
616 : int indexcol;
617 :
618 : /* If we already considered this relids set, don't repeat the work */
619 156568 : if (list_member(*considered_relids, relids))
620 0 : return;
621 :
622 : /* Identify indexclauses usable with this relids set */
623 5323312 : MemSet(&clauseset, 0, sizeof(clauseset));
624 :
625 403744 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
626 : {
627 : ListCell *lc;
628 :
629 : /* First find applicable simple join clauses */
630 282988 : foreach(lc, jclauseset->indexclauses[indexcol])
631 : {
632 35812 : IndexClause *iclause = (IndexClause *) lfirst(lc);
633 :
634 35812 : if (bms_is_subset(iclause->rinfo->clause_relids, relids))
635 35386 : clauseset.indexclauses[indexcol] =
636 35386 : lappend(clauseset.indexclauses[indexcol], iclause);
637 : }
638 :
639 : /*
640 : * Add applicable eclass join clauses. The clauses generated for each
641 : * column are redundant (cf generate_implied_equalities_for_column),
642 : * so we need at most one. This is the only exception to the general
643 : * rule of using all available index clauses.
644 : */
645 263668 : foreach(lc, eclauseset->indexclauses[indexcol])
646 : {
647 149164 : IndexClause *iclause = (IndexClause *) lfirst(lc);
648 :
649 149164 : if (bms_is_subset(iclause->rinfo->clause_relids, relids))
650 : {
651 132672 : clauseset.indexclauses[indexcol] =
652 132672 : lappend(clauseset.indexclauses[indexcol], iclause);
653 132672 : break;
654 : }
655 : }
656 :
657 : /* Add restriction clauses */
658 247176 : clauseset.indexclauses[indexcol] =
659 247176 : list_concat(clauseset.indexclauses[indexcol],
660 247176 : rclauseset->indexclauses[indexcol]);
661 :
662 247176 : if (clauseset.indexclauses[indexcol] != NIL)
663 196676 : clauseset.nonempty = true;
664 : }
665 :
666 : /* We should have found something, else caller passed silly relids */
667 : Assert(clauseset.nonempty);
668 :
669 : /* Build index path(s) using the collected set of clauses */
670 156568 : get_index_paths(root, rel, index, &clauseset, bitindexpaths);
671 :
672 : /*
673 : * Remember we considered paths for this set of relids.
674 : */
675 156568 : *considered_relids = lappend(*considered_relids, relids);
676 : }
677 :
678 : /*
679 : * eclass_already_used
680 : * True if any join clause usable with oldrelids was generated from
681 : * the specified equivalence class.
682 : */
683 : static bool
684 10724 : eclass_already_used(EquivalenceClass *parent_ec, Relids oldrelids,
685 : List *indexjoinclauses)
686 : {
687 : ListCell *lc;
688 :
689 14522 : foreach(lc, indexjoinclauses)
690 : {
691 11186 : IndexClause *iclause = (IndexClause *) lfirst(lc);
692 11186 : RestrictInfo *rinfo = iclause->rinfo;
693 :
694 22372 : if (rinfo->parent_ec == parent_ec &&
695 11186 : bms_is_subset(rinfo->clause_relids, oldrelids))
696 7388 : return true;
697 : }
698 3336 : return false;
699 : }
700 :
701 :
702 : /*
703 : * get_index_paths
704 : * Given an index and a set of index clauses for it, construct IndexPaths.
705 : *
706 : * Plain indexpaths are sent directly to add_path, while potential
707 : * bitmap indexpaths are added to *bitindexpaths for later processing.
708 : *
709 : * This is a fairly simple frontend to build_index_paths(). Its reason for
710 : * existence is mainly to handle ScalarArrayOpExpr quals properly. If the
711 : * index AM supports them natively, we should just include them in simple
712 : * index paths. If not, we should exclude them while building simple index
713 : * paths, and then make a separate attempt to include them in bitmap paths.
714 : */
715 : static void
716 889294 : get_index_paths(PlannerInfo *root, RelOptInfo *rel,
717 : IndexOptInfo *index, IndexClauseSet *clauses,
718 : List **bitindexpaths)
719 : {
720 : List *indexpaths;
721 889294 : bool skip_nonnative_saop = false;
722 : ListCell *lc;
723 :
724 : /*
725 : * Build simple index paths using the clauses. Allow ScalarArrayOpExpr
726 : * clauses only if the index AM supports them natively.
727 : */
728 889294 : indexpaths = build_index_paths(root, rel,
729 : index, clauses,
730 889294 : index->predOK,
731 : ST_ANYSCAN,
732 : &skip_nonnative_saop);
733 :
734 : /*
735 : * Submit all the ones that can form plain IndexScan plans to add_path. (A
736 : * plain IndexPath can represent either a plain IndexScan or an
737 : * IndexOnlyScan, but for our purposes here that distinction does not
738 : * matter. However, some of the indexes might support only bitmap scans,
739 : * and those we mustn't submit to add_path here.)
740 : *
741 : * Also, pick out the ones that are usable as bitmap scans. For that, we
742 : * must discard indexes that don't support bitmap scans, and we also are
743 : * only interested in paths that have some selectivity; we should discard
744 : * anything that was generated solely for ordering purposes.
745 : */
746 1425228 : foreach(lc, indexpaths)
747 : {
748 535934 : IndexPath *ipath = (IndexPath *) lfirst(lc);
749 :
750 535934 : if (index->amhasgettuple)
751 522138 : add_path(rel, (Path *) ipath);
752 :
753 535934 : if (index->amhasgetbitmap &&
754 535934 : (ipath->path.pathkeys == NIL ||
755 333470 : ipath->indexselectivity < 1.0))
756 391106 : *bitindexpaths = lappend(*bitindexpaths, ipath);
757 : }
758 :
759 : /*
760 : * If there were ScalarArrayOpExpr clauses that the index can't handle
761 : * natively, generate bitmap scan paths relying on executor-managed
762 : * ScalarArrayOpExpr.
763 : */
764 889294 : if (skip_nonnative_saop)
765 : {
766 32 : indexpaths = build_index_paths(root, rel,
767 : index, clauses,
768 : false,
769 : ST_BITMAPSCAN,
770 : NULL);
771 32 : *bitindexpaths = list_concat(*bitindexpaths, indexpaths);
772 : }
773 889294 : }
774 :
775 : /*
776 : * build_index_paths
777 : * Given an index and a set of index clauses for it, construct zero
778 : * or more IndexPaths. It also constructs zero or more partial IndexPaths.
779 : *
780 : * We return a list of paths because (1) this routine checks some cases
781 : * that should cause us to not generate any IndexPath, and (2) in some
782 : * cases we want to consider both a forward and a backward scan, so as
783 : * to obtain both sort orders. Note that the paths are just returned
784 : * to the caller and not immediately fed to add_path().
785 : *
786 : * At top level, useful_predicate should be exactly the index's predOK flag
787 : * (ie, true if it has a predicate that was proven from the restriction
788 : * clauses). When working on an arm of an OR clause, useful_predicate
789 : * should be true if the predicate required the current OR list to be proven.
790 : * Note that this routine should never be called at all if the index has an
791 : * unprovable predicate.
792 : *
793 : * scantype indicates whether we want to create plain indexscans, bitmap
794 : * indexscans, or both. When it's ST_BITMAPSCAN, we will not consider
795 : * index ordering while deciding if a Path is worth generating.
796 : *
797 : * If skip_nonnative_saop is non-NULL, we ignore ScalarArrayOpExpr clauses
798 : * unless the index AM supports them directly, and we set *skip_nonnative_saop
799 : * to true if we found any such clauses (caller must initialize the variable
800 : * to false). If it's NULL, we do not ignore ScalarArrayOpExpr clauses.
801 : *
802 : * 'rel' is the index's heap relation
803 : * 'index' is the index for which we want to generate paths
804 : * 'clauses' is the collection of indexable clauses (IndexClause nodes)
805 : * 'useful_predicate' indicates whether the index has a useful predicate
806 : * 'scantype' indicates whether we need plain or bitmap scan support
807 : * 'skip_nonnative_saop' indicates whether to accept SAOP if index AM doesn't
808 : */
809 : static List *
810 892480 : build_index_paths(PlannerInfo *root, RelOptInfo *rel,
811 : IndexOptInfo *index, IndexClauseSet *clauses,
812 : bool useful_predicate,
813 : ScanTypeControl scantype,
814 : bool *skip_nonnative_saop)
815 : {
816 892480 : List *result = NIL;
817 : IndexPath *ipath;
818 : List *index_clauses;
819 : Relids outer_relids;
820 : double loop_count;
821 : List *orderbyclauses;
822 : List *orderbyclausecols;
823 : List *index_pathkeys;
824 : List *useful_pathkeys;
825 : bool pathkeys_possibly_useful;
826 : bool index_is_ordered;
827 : bool index_only_scan;
828 : int indexcol;
829 :
830 : Assert(skip_nonnative_saop != NULL || scantype == ST_BITMAPSCAN);
831 :
832 : /*
833 : * Check that index supports the desired scan type(s)
834 : */
835 892480 : switch (scantype)
836 : {
837 0 : case ST_INDEXSCAN:
838 0 : if (!index->amhasgettuple)
839 0 : return NIL;
840 0 : break;
841 3186 : case ST_BITMAPSCAN:
842 3186 : if (!index->amhasgetbitmap)
843 0 : return NIL;
844 3186 : break;
845 889294 : case ST_ANYSCAN:
846 : /* either or both are OK */
847 889294 : break;
848 : }
849 :
850 : /*
851 : * 1. Combine the per-column IndexClause lists into an overall list.
852 : *
853 : * In the resulting list, clauses are ordered by index key, so that the
854 : * column numbers form a nondecreasing sequence. (This order is depended
855 : * on by btree and possibly other places.) The list can be empty, if the
856 : * index AM allows that.
857 : *
858 : * We also build a Relids set showing which outer rels are required by the
859 : * selected clauses. Any lateral_relids are included in that, but not
860 : * otherwise accounted for.
861 : */
862 892480 : index_clauses = NIL;
863 892480 : outer_relids = bms_copy(rel->lateral_relids);
864 2531034 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
865 : {
866 : ListCell *lc;
867 :
868 2098006 : foreach(lc, clauses->indexclauses[indexcol])
869 : {
870 459108 : IndexClause *iclause = (IndexClause *) lfirst(lc);
871 459108 : RestrictInfo *rinfo = iclause->rinfo;
872 :
873 459108 : if (skip_nonnative_saop && !index->amsearcharray &&
874 21854 : IsA(rinfo->clause, ScalarArrayOpExpr))
875 : {
876 : /*
877 : * Caller asked us to generate IndexPaths that omit any
878 : * ScalarArrayOpExpr clauses when the underlying index AM
879 : * lacks native support.
880 : *
881 : * We must omit this clause (and tell caller about it).
882 : */
883 32 : *skip_nonnative_saop = true;
884 32 : continue;
885 : }
886 :
887 : /* OK to include this clause */
888 459076 : index_clauses = lappend(index_clauses, iclause);
889 459076 : outer_relids = bms_add_members(outer_relids,
890 459076 : rinfo->clause_relids);
891 : }
892 :
893 : /*
894 : * If no clauses match the first index column, check for amoptionalkey
895 : * restriction. We can't generate a scan over an index with
896 : * amoptionalkey = false unless there's at least one index clause.
897 : * (When working on columns after the first, this test cannot fail. It
898 : * is always okay for columns after the first to not have any
899 : * clauses.)
900 : */
901 1638898 : if (index_clauses == NIL && !index->amoptionalkey)
902 344 : return NIL;
903 : }
904 :
905 : /* We do not want the index's rel itself listed in outer_relids */
906 892136 : outer_relids = bms_del_member(outer_relids, rel->relid);
907 :
908 : /* Compute loop_count for cost estimation purposes */
909 892136 : loop_count = get_loop_count(root, rel->relid, outer_relids);
910 :
911 : /*
912 : * 2. Compute pathkeys describing index's ordering, if any, then see how
913 : * many of them are actually useful for this query. This is not relevant
914 : * if we are only trying to build bitmap indexscans.
915 : */
916 1781086 : pathkeys_possibly_useful = (scantype != ST_BITMAPSCAN &&
917 888950 : has_useful_pathkeys(root, rel));
918 892136 : index_is_ordered = (index->sortopfamily != NULL);
919 892136 : if (index_is_ordered && pathkeys_possibly_useful)
920 : {
921 670450 : index_pathkeys = build_index_pathkeys(root, index,
922 : ForwardScanDirection);
923 670450 : useful_pathkeys = truncate_useless_pathkeys(root, rel,
924 : index_pathkeys);
925 670450 : orderbyclauses = NIL;
926 670450 : orderbyclausecols = NIL;
927 : }
928 221686 : else if (index->amcanorderbyop && pathkeys_possibly_useful)
929 : {
930 : /*
931 : * See if we can generate ordering operators for query_pathkeys or at
932 : * least some prefix thereof. Matching to just a prefix of the
933 : * query_pathkeys will allow an incremental sort to be considered on
934 : * the index's partially sorted results.
935 : */
936 1074 : match_pathkeys_to_index(index, root->query_pathkeys,
937 : &orderbyclauses,
938 : &orderbyclausecols);
939 1074 : if (list_length(root->query_pathkeys) == list_length(orderbyclauses))
940 468 : useful_pathkeys = root->query_pathkeys;
941 : else
942 606 : useful_pathkeys = list_copy_head(root->query_pathkeys,
943 : list_length(orderbyclauses));
944 : }
945 : else
946 : {
947 220612 : useful_pathkeys = NIL;
948 220612 : orderbyclauses = NIL;
949 220612 : orderbyclausecols = NIL;
950 : }
951 :
952 : /*
953 : * 3. Check if an index-only scan is possible. If we're not building
954 : * plain indexscans, this isn't relevant since bitmap scans don't support
955 : * index data retrieval anyway.
956 : */
957 1781086 : index_only_scan = (scantype != ST_BITMAPSCAN &&
958 888950 : check_index_only(rel, index));
959 :
960 : /*
961 : * 4. Generate an indexscan path if there are relevant restriction clauses
962 : * in the current clauses, OR the index ordering is potentially useful for
963 : * later merging or final output ordering, OR the index has a useful
964 : * predicate, OR an index-only scan is possible.
965 : */
966 892136 : if (index_clauses != NIL || useful_pathkeys != NIL || useful_predicate ||
967 : index_only_scan)
968 : {
969 538506 : ipath = create_index_path(root, index,
970 : index_clauses,
971 : orderbyclauses,
972 : orderbyclausecols,
973 : useful_pathkeys,
974 : ForwardScanDirection,
975 : index_only_scan,
976 : outer_relids,
977 : loop_count,
978 : false);
979 538506 : result = lappend(result, ipath);
980 :
981 : /*
982 : * If appropriate, consider parallel index scan. We don't allow
983 : * parallel index scan for bitmap index scans.
984 : */
985 538506 : if (index->amcanparallel &&
986 517608 : rel->consider_parallel && outer_relids == NULL &&
987 : scantype != ST_BITMAPSCAN)
988 : {
989 280880 : ipath = create_index_path(root, index,
990 : index_clauses,
991 : orderbyclauses,
992 : orderbyclausecols,
993 : useful_pathkeys,
994 : ForwardScanDirection,
995 : index_only_scan,
996 : outer_relids,
997 : loop_count,
998 : true);
999 :
1000 : /*
1001 : * if, after costing the path, we find that it's not worth using
1002 : * parallel workers, just free it.
1003 : */
1004 280880 : if (ipath->path.parallel_workers > 0)
1005 10070 : add_partial_path(rel, (Path *) ipath);
1006 : else
1007 270810 : pfree(ipath);
1008 : }
1009 : }
1010 :
1011 : /*
1012 : * 5. If the index is ordered, a backwards scan might be interesting.
1013 : */
1014 892136 : if (index_is_ordered && pathkeys_possibly_useful)
1015 : {
1016 670450 : index_pathkeys = build_index_pathkeys(root, index,
1017 : BackwardScanDirection);
1018 670450 : useful_pathkeys = truncate_useless_pathkeys(root, rel,
1019 : index_pathkeys);
1020 670450 : if (useful_pathkeys != NIL)
1021 : {
1022 614 : ipath = create_index_path(root, index,
1023 : index_clauses,
1024 : NIL,
1025 : NIL,
1026 : useful_pathkeys,
1027 : BackwardScanDirection,
1028 : index_only_scan,
1029 : outer_relids,
1030 : loop_count,
1031 : false);
1032 614 : result = lappend(result, ipath);
1033 :
1034 : /* If appropriate, consider parallel index scan */
1035 614 : if (index->amcanparallel &&
1036 614 : rel->consider_parallel && outer_relids == NULL &&
1037 : scantype != ST_BITMAPSCAN)
1038 : {
1039 512 : ipath = create_index_path(root, index,
1040 : index_clauses,
1041 : NIL,
1042 : NIL,
1043 : useful_pathkeys,
1044 : BackwardScanDirection,
1045 : index_only_scan,
1046 : outer_relids,
1047 : loop_count,
1048 : true);
1049 :
1050 : /*
1051 : * if, after costing the path, we find that it's not worth
1052 : * using parallel workers, just free it.
1053 : */
1054 512 : if (ipath->path.parallel_workers > 0)
1055 168 : add_partial_path(rel, (Path *) ipath);
1056 : else
1057 344 : pfree(ipath);
1058 : }
1059 : }
1060 : }
1061 :
1062 892136 : return result;
1063 : }
1064 :
1065 : /*
1066 : * build_paths_for_OR
1067 : * Given a list of restriction clauses from one arm of an OR clause,
1068 : * construct all matching IndexPaths for the relation.
1069 : *
1070 : * Here we must scan all indexes of the relation, since a bitmap OR tree
1071 : * can use multiple indexes.
1072 : *
1073 : * The caller actually supplies two lists of restriction clauses: some
1074 : * "current" ones and some "other" ones. Both lists can be used freely
1075 : * to match keys of the index, but an index must use at least one of the
1076 : * "current" clauses to be considered usable. The motivation for this is
1077 : * examples like
1078 : * WHERE (x = 42) AND (... OR (y = 52 AND z = 77) OR ....)
1079 : * While we are considering the y/z subclause of the OR, we can use "x = 42"
1080 : * as one of the available index conditions; but we shouldn't match the
1081 : * subclause to any index on x alone, because such a Path would already have
1082 : * been generated at the upper level. So we could use an index on x,y,z
1083 : * or an index on x,y for the OR subclause, but not an index on just x.
1084 : * When dealing with a partial index, a match of the index predicate to
1085 : * one of the "current" clauses also makes the index usable.
1086 : *
1087 : * 'rel' is the relation for which we want to generate index paths
1088 : * 'clauses' is the current list of clauses (RestrictInfo nodes)
1089 : * 'other_clauses' is the list of additional upper-level clauses
1090 : */
1091 : static List *
1092 10876 : build_paths_for_OR(PlannerInfo *root, RelOptInfo *rel,
1093 : List *clauses, List *other_clauses)
1094 : {
1095 10876 : List *result = NIL;
1096 10876 : List *all_clauses = NIL; /* not computed till needed */
1097 : ListCell *lc;
1098 :
1099 37876 : foreach(lc, rel->indexlist)
1100 : {
1101 27000 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
1102 : IndexClauseSet clauseset;
1103 : List *indexpaths;
1104 : bool useful_predicate;
1105 :
1106 : /* Ignore index if it doesn't support bitmap scans */
1107 27000 : if (!index->amhasgetbitmap)
1108 23846 : continue;
1109 :
1110 : /*
1111 : * Ignore partial indexes that do not match the query. If a partial
1112 : * index is marked predOK then we know it's OK. Otherwise, we have to
1113 : * test whether the added clauses are sufficient to imply the
1114 : * predicate. If so, we can use the index in the current context.
1115 : *
1116 : * We set useful_predicate to true iff the predicate was proven using
1117 : * the current set of clauses. This is needed to prevent matching a
1118 : * predOK index to an arm of an OR, which would be a legal but
1119 : * pointlessly inefficient plan. (A better plan will be generated by
1120 : * just scanning the predOK index alone, no OR.)
1121 : */
1122 27000 : useful_predicate = false;
1123 27000 : if (index->indpred != NIL)
1124 : {
1125 168 : if (index->predOK)
1126 : {
1127 : /* Usable, but don't set useful_predicate */
1128 : }
1129 : else
1130 : {
1131 : /* Form all_clauses if not done already */
1132 144 : if (all_clauses == NIL)
1133 60 : all_clauses = list_concat_copy(clauses, other_clauses);
1134 :
1135 144 : if (!predicate_implied_by(index->indpred, all_clauses, false))
1136 96 : continue; /* can't use it at all */
1137 :
1138 48 : if (!predicate_implied_by(index->indpred, other_clauses, false))
1139 48 : useful_predicate = true;
1140 : }
1141 : }
1142 :
1143 : /*
1144 : * Identify the restriction clauses that can match the index.
1145 : */
1146 914736 : MemSet(&clauseset, 0, sizeof(clauseset));
1147 26904 : match_clauses_to_index(root, clauses, index, &clauseset);
1148 :
1149 : /*
1150 : * If no matches so far, and the index predicate isn't useful, we
1151 : * don't want it.
1152 : */
1153 26904 : if (!clauseset.nonempty && !useful_predicate)
1154 23750 : continue;
1155 :
1156 : /*
1157 : * Add "other" restriction clauses to the clauseset.
1158 : */
1159 3154 : match_clauses_to_index(root, other_clauses, index, &clauseset);
1160 :
1161 : /*
1162 : * Construct paths if possible.
1163 : */
1164 3154 : indexpaths = build_index_paths(root, rel,
1165 : index, &clauseset,
1166 : useful_predicate,
1167 : ST_BITMAPSCAN,
1168 : NULL);
1169 3154 : result = list_concat(result, indexpaths);
1170 : }
1171 :
1172 10876 : return result;
1173 : }
1174 :
1175 : /*
1176 : * Utility structure used to group similar OR-clause arguments in
1177 : * group_similar_or_args(). It represents information about the OR-clause
1178 : * argument and its matching index key.
1179 : */
1180 : typedef struct
1181 : {
1182 : int indexnum; /* index of the matching index, or -1 if no
1183 : * matching index */
1184 : int colnum; /* index of the matching column, or -1 if no
1185 : * matching index */
1186 : Oid opno; /* OID of the OpClause operator, or InvalidOid
1187 : * if not an OpExpr */
1188 : Oid inputcollid; /* OID of the OpClause input collation */
1189 : int argindex; /* index of the clause in the list of
1190 : * arguments */
1191 : int groupindex; /* value of argindex for the fist clause in
1192 : * the group of similar clauses */
1193 : } OrArgIndexMatch;
1194 :
1195 : /*
1196 : * Comparison function for OrArgIndexMatch which provides sort order placing
1197 : * similar OR-clause arguments together.
1198 : */
1199 : static int
1200 7150 : or_arg_index_match_cmp(const void *a, const void *b)
1201 : {
1202 7150 : const OrArgIndexMatch *match_a = (const OrArgIndexMatch *) a;
1203 7150 : const OrArgIndexMatch *match_b = (const OrArgIndexMatch *) b;
1204 :
1205 7150 : if (match_a->indexnum < match_b->indexnum)
1206 1362 : return -1;
1207 5788 : else if (match_a->indexnum > match_b->indexnum)
1208 3102 : return 1;
1209 :
1210 2686 : if (match_a->colnum < match_b->colnum)
1211 874 : return -1;
1212 1812 : else if (match_a->colnum > match_b->colnum)
1213 24 : return 1;
1214 :
1215 1788 : if (match_a->opno < match_b->opno)
1216 18 : return -1;
1217 1770 : else if (match_a->opno > match_b->opno)
1218 42 : return 1;
1219 :
1220 1728 : if (match_a->inputcollid < match_b->inputcollid)
1221 0 : return -1;
1222 1728 : else if (match_a->inputcollid > match_b->inputcollid)
1223 0 : return 1;
1224 :
1225 1728 : if (match_a->argindex < match_b->argindex)
1226 1650 : return -1;
1227 78 : else if (match_a->argindex > match_b->argindex)
1228 78 : return 1;
1229 :
1230 0 : return 0;
1231 : }
1232 :
1233 : /*
1234 : * Another comparison function for OrArgIndexMatch. It sorts groups together
1235 : * using groupindex. The group items are then sorted by argindex.
1236 : */
1237 : static int
1238 7240 : or_arg_index_match_cmp_group(const void *a, const void *b)
1239 : {
1240 7240 : const OrArgIndexMatch *match_a = (const OrArgIndexMatch *) a;
1241 7240 : const OrArgIndexMatch *match_b = (const OrArgIndexMatch *) b;
1242 :
1243 7240 : if (match_a->groupindex < match_b->groupindex)
1244 3550 : return -1;
1245 3690 : else if (match_a->groupindex > match_b->groupindex)
1246 3246 : return 1;
1247 :
1248 444 : if (match_a->argindex < match_b->argindex)
1249 444 : return -1;
1250 0 : else if (match_a->argindex > match_b->argindex)
1251 0 : return 1;
1252 :
1253 0 : return 0;
1254 : }
1255 :
1256 : /*
1257 : * group_similar_or_args
1258 : * Transform incoming OR-restrictinfo into a list of sub-restrictinfos,
1259 : * each of them containing a subset of similar OR-clause arguments from
1260 : * the source rinfo.
1261 : *
1262 : * Similar OR-clause arguments are of the form "indexkey op constant" having
1263 : * the same indexkey, operator, and collation. Constant may comprise either
1264 : * Const or Param. It may be employed later, during the
1265 : * match_clause_to_indexcol() to transform the whole OR-sub-rinfo to an SAOP
1266 : * clause.
1267 : *
1268 : * Returns the processed list of OR-clause arguments.
1269 : */
1270 : static List *
1271 9092 : group_similar_or_args(PlannerInfo *root, RelOptInfo *rel, RestrictInfo *rinfo)
1272 : {
1273 : int n;
1274 : int i;
1275 : int group_start;
1276 : OrArgIndexMatch *matches;
1277 9092 : bool matched = false;
1278 : ListCell *lc;
1279 : ListCell *lc2;
1280 : List *orargs;
1281 9092 : List *result = NIL;
1282 9092 : Index relid = rel->relid;
1283 :
1284 : Assert(IsA(rinfo->orclause, BoolExpr));
1285 9092 : orargs = ((BoolExpr *) rinfo->orclause)->args;
1286 9092 : n = list_length(orargs);
1287 :
1288 : /*
1289 : * To avoid N^2 behavior, take utility pass along the list of OR-clause
1290 : * arguments. For each argument, fill the OrArgIndexMatch structure,
1291 : * which will be used to sort these arguments at the next step.
1292 : */
1293 9092 : i = -1;
1294 9092 : matches = palloc_array(OrArgIndexMatch, n);
1295 30622 : foreach(lc, orargs)
1296 : {
1297 21530 : Node *arg = lfirst(lc);
1298 : RestrictInfo *argrinfo;
1299 : OpExpr *clause;
1300 : Oid opno;
1301 : Node *leftop,
1302 : *rightop;
1303 : Node *nonConstExpr;
1304 : int indexnum;
1305 : int colnum;
1306 :
1307 21530 : i++;
1308 21530 : matches[i].argindex = i;
1309 21530 : matches[i].groupindex = i;
1310 21530 : matches[i].indexnum = -1;
1311 21530 : matches[i].colnum = -1;
1312 21530 : matches[i].opno = InvalidOid;
1313 21530 : matches[i].inputcollid = InvalidOid;
1314 :
1315 21530 : if (!IsA(arg, RestrictInfo))
1316 2458 : continue;
1317 :
1318 19072 : argrinfo = castNode(RestrictInfo, arg);
1319 :
1320 : /* Only operator clauses can match */
1321 19072 : if (!IsA(argrinfo->clause, OpExpr))
1322 7790 : continue;
1323 :
1324 11282 : clause = (OpExpr *) argrinfo->clause;
1325 11282 : opno = clause->opno;
1326 :
1327 : /* Only binary operators can match */
1328 11282 : if (list_length(clause->args) != 2)
1329 0 : continue;
1330 :
1331 : /*
1332 : * Ignore any RelabelType node above the operands. This is needed to
1333 : * be able to apply indexscanning in binary-compatible-operator cases.
1334 : * Note: we can assume there is at most one RelabelType node;
1335 : * eval_const_expressions() will have simplified if more than one.
1336 : */
1337 11282 : leftop = get_leftop(clause);
1338 11282 : if (IsA(leftop, RelabelType))
1339 204 : leftop = (Node *) ((RelabelType *) leftop)->arg;
1340 :
1341 11282 : rightop = get_rightop(clause);
1342 11282 : if (IsA(rightop, RelabelType))
1343 808 : rightop = (Node *) ((RelabelType *) rightop)->arg;
1344 :
1345 : /*
1346 : * Check for clauses of the form: (indexkey operator constant) or
1347 : * (constant operator indexkey). But we don't know a particular index
1348 : * yet. Therefore, we try to distinguish the potential index key and
1349 : * constant first, then search for a matching index key among all
1350 : * indexes.
1351 : */
1352 11282 : if (bms_is_member(relid, argrinfo->right_relids) &&
1353 1986 : !bms_is_member(relid, argrinfo->left_relids) &&
1354 1914 : !contain_volatile_functions(leftop))
1355 : {
1356 1914 : opno = get_commutator(opno);
1357 :
1358 1914 : if (!OidIsValid(opno))
1359 : {
1360 : /* commutator doesn't exist, we can't reverse the order */
1361 0 : continue;
1362 : }
1363 1914 : nonConstExpr = rightop;
1364 : }
1365 9368 : else if (bms_is_member(relid, argrinfo->left_relids) &&
1366 7436 : !bms_is_member(relid, argrinfo->right_relids) &&
1367 7364 : !contain_volatile_functions(rightop))
1368 : {
1369 7364 : nonConstExpr = leftop;
1370 : }
1371 : else
1372 : {
1373 2004 : continue;
1374 : }
1375 :
1376 : /*
1377 : * Match non-constant part to the index key. It's possible that a
1378 : * single non-constant part matches multiple index keys. It's OK, we
1379 : * just stop with first matching index key. Given that this choice is
1380 : * determined the same for every clause, we will group similar clauses
1381 : * together anyway.
1382 : */
1383 9278 : indexnum = 0;
1384 20188 : foreach(lc2, rel->indexlist)
1385 : {
1386 16502 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc2);
1387 :
1388 : /*
1389 : * Ignore index if it doesn't support bitmap scans or SAOP
1390 : * clauses.
1391 : */
1392 16502 : if (!index->amhasgetbitmap || !index->amsearcharray)
1393 54 : continue;
1394 :
1395 37322 : for (colnum = 0; colnum < index->nkeycolumns; colnum++)
1396 : {
1397 26466 : if (match_index_to_operand(nonConstExpr, colnum, index))
1398 : {
1399 5592 : matches[i].indexnum = indexnum;
1400 5592 : matches[i].colnum = colnum;
1401 5592 : matches[i].opno = opno;
1402 5592 : matches[i].inputcollid = clause->inputcollid;
1403 5592 : matched = true;
1404 5592 : break;
1405 : }
1406 : }
1407 :
1408 : /*
1409 : * Stop looping through the indexes, if we managed to match
1410 : * nonConstExpr to any index column.
1411 : */
1412 16448 : if (matches[i].indexnum >= 0)
1413 5592 : break;
1414 10856 : indexnum++;
1415 : }
1416 : }
1417 :
1418 : /*
1419 : * Fast-path check: if no clause is matching to the index column, we can
1420 : * just give up at this stage and return the clause list as-is.
1421 : */
1422 9092 : if (!matched)
1423 : {
1424 5068 : pfree(matches);
1425 5068 : return orargs;
1426 : }
1427 :
1428 : /*
1429 : * Sort clauses to make similar clauses go together. But at the same
1430 : * time, we would like to change the order of clauses as little as
1431 : * possible. To do so, we reorder each group of similar clauses so that
1432 : * the first item of the group stays in place, and all the other items are
1433 : * moved after it. So, if there are no similar clauses, the order of
1434 : * clauses stays the same. When there are some groups, required
1435 : * reordering happens while the rest of the clauses remain in their
1436 : * places. That is achieved by assigning a 'groupindex' to each clause:
1437 : * the number of the first item in the group in the original clause list.
1438 : */
1439 4024 : qsort(matches, n, sizeof(OrArgIndexMatch), or_arg_index_match_cmp);
1440 :
1441 : /* Assign groupindex to the sorted clauses */
1442 9626 : for (i = 1; i < n; i++)
1443 : {
1444 : /*
1445 : * When two clauses are similar and should belong to the same group,
1446 : * copy the 'groupindex' from the previous clause. Given we are
1447 : * considering clauses in direct order, all the clauses would have a
1448 : * 'groupindex' equal to the 'groupindex' of the first clause in the
1449 : * group.
1450 : */
1451 5602 : if (matches[i].indexnum == matches[i - 1].indexnum &&
1452 2566 : matches[i].colnum == matches[i - 1].colnum &&
1453 1680 : matches[i].opno == matches[i - 1].opno &&
1454 1632 : matches[i].inputcollid == matches[i - 1].inputcollid &&
1455 1632 : matches[i].indexnum != -1)
1456 444 : matches[i].groupindex = matches[i - 1].groupindex;
1457 : }
1458 :
1459 : /* Re-sort clauses first by groupindex then by argindex */
1460 4024 : qsort(matches, n, sizeof(OrArgIndexMatch), or_arg_index_match_cmp_group);
1461 :
1462 : /*
1463 : * Group similar clauses into single sub-restrictinfo. Side effect: the
1464 : * resulting list of restrictions will be sorted by indexnum and colnum.
1465 : */
1466 4024 : group_start = 0;
1467 13650 : for (i = 1; i <= n; i++)
1468 : {
1469 : /* Check if it's a group boundary */
1470 9626 : if (group_start >= 0 &&
1471 5602 : (i == n ||
1472 5602 : matches[i].indexnum != matches[group_start].indexnum ||
1473 2494 : matches[i].colnum != matches[group_start].colnum ||
1474 1626 : matches[i].opno != matches[group_start].opno ||
1475 1584 : matches[i].inputcollid != matches[group_start].inputcollid ||
1476 1584 : matches[i].indexnum == -1))
1477 : {
1478 : /*
1479 : * One clause in group: add it "as is" to the upper-level OR.
1480 : */
1481 9182 : if (i - group_start == 1)
1482 : {
1483 8870 : result = lappend(result,
1484 : list_nth(orargs,
1485 8870 : matches[group_start].argindex));
1486 : }
1487 : else
1488 : {
1489 : /*
1490 : * Two or more clauses in a group: create a nested OR.
1491 : */
1492 312 : List *args = NIL;
1493 312 : List *rargs = NIL;
1494 : RestrictInfo *subrinfo;
1495 : int j;
1496 :
1497 : Assert(i - group_start >= 2);
1498 :
1499 : /* Construct the list of nested OR arguments */
1500 1068 : for (j = group_start; j < i; j++)
1501 : {
1502 756 : Node *arg = list_nth(orargs, matches[j].argindex);
1503 :
1504 756 : rargs = lappend(rargs, arg);
1505 756 : if (IsA(arg, RestrictInfo))
1506 756 : args = lappend(args, ((RestrictInfo *) arg)->clause);
1507 : else
1508 0 : args = lappend(args, arg);
1509 : }
1510 :
1511 : /* Construct the nested OR and wrap it with RestrictInfo */
1512 312 : subrinfo = make_plain_restrictinfo(root,
1513 : make_orclause(args),
1514 : make_orclause(rargs),
1515 312 : rinfo->is_pushed_down,
1516 312 : rinfo->has_clone,
1517 312 : rinfo->is_clone,
1518 312 : rinfo->pseudoconstant,
1519 : rinfo->security_level,
1520 : rinfo->required_relids,
1521 : rinfo->incompatible_relids,
1522 : rinfo->outer_relids);
1523 312 : result = lappend(result, subrinfo);
1524 : }
1525 :
1526 9182 : group_start = i;
1527 : }
1528 : }
1529 4024 : pfree(matches);
1530 4024 : return result;
1531 : }
1532 :
1533 : /*
1534 : * make_bitmap_paths_for_or_group
1535 : * Generate bitmap paths for a group of similar OR-clause arguments
1536 : * produced by group_similar_or_args().
1537 : *
1538 : * This function considers two cases: (1) matching a group of clauses to
1539 : * the index as a whole, and (2) matching the individual clauses one-by-one.
1540 : * (1) typically comprises an optimal solution. If not, (2) typically
1541 : * comprises fair alternative.
1542 : *
1543 : * Ideally, we could consider all arbitrary splits of arguments into
1544 : * subgroups, but that could lead to unacceptable computational complexity.
1545 : * This is why we only consider two cases of above.
1546 : */
1547 : static List *
1548 306 : make_bitmap_paths_for_or_group(PlannerInfo *root, RelOptInfo *rel,
1549 : RestrictInfo *ri, List *other_clauses)
1550 : {
1551 306 : List *jointlist = NIL;
1552 306 : List *splitlist = NIL;
1553 : ListCell *lc;
1554 : List *orargs;
1555 306 : List *args = ((BoolExpr *) ri->orclause)->args;
1556 306 : Cost jointcost = 0.0,
1557 306 : splitcost = 0.0;
1558 : Path *bitmapqual;
1559 : List *indlist;
1560 :
1561 : /*
1562 : * First, try to match the whole group to the one index.
1563 : */
1564 306 : orargs = list_make1(ri);
1565 306 : indlist = build_paths_for_OR(root, rel,
1566 : orargs,
1567 : other_clauses);
1568 306 : if (indlist != NIL)
1569 : {
1570 300 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1571 300 : jointcost = bitmapqual->total_cost;
1572 300 : jointlist = list_make1(bitmapqual);
1573 : }
1574 :
1575 : /*
1576 : * If we manage to find a bitmap scan, which uses the group of OR-clause
1577 : * arguments as a whole, we can skip matching OR-clause arguments
1578 : * one-by-one as long as there are no other clauses, which can bring more
1579 : * efficiency to one-by-one case.
1580 : */
1581 306 : if (jointlist != NIL && other_clauses == NIL)
1582 84 : return jointlist;
1583 :
1584 : /*
1585 : * Also try to match all containing clauses one-by-one.
1586 : */
1587 768 : foreach(lc, args)
1588 : {
1589 552 : orargs = list_make1(lfirst(lc));
1590 :
1591 552 : indlist = build_paths_for_OR(root, rel,
1592 : orargs,
1593 : other_clauses);
1594 :
1595 552 : if (indlist == NIL)
1596 : {
1597 6 : splitlist = NIL;
1598 6 : break;
1599 : }
1600 :
1601 546 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1602 546 : splitcost += bitmapqual->total_cost;
1603 546 : splitlist = lappend(splitlist, bitmapqual);
1604 : }
1605 :
1606 : /*
1607 : * Pick the best option.
1608 : */
1609 222 : if (splitlist == NIL)
1610 6 : return jointlist;
1611 216 : else if (jointlist == NIL)
1612 0 : return splitlist;
1613 : else
1614 216 : return (jointcost < splitcost) ? jointlist : splitlist;
1615 : }
1616 :
1617 :
1618 : /*
1619 : * generate_bitmap_or_paths
1620 : * Look through the list of clauses to find OR clauses, and generate
1621 : * a BitmapOrPath for each one we can handle that way. Return a list
1622 : * of the generated BitmapOrPaths.
1623 : *
1624 : * other_clauses is a list of additional clauses that can be assumed true
1625 : * for the purpose of generating indexquals, but are not to be searched for
1626 : * ORs. (See build_paths_for_OR() for motivation.)
1627 : */
1628 : static List *
1629 690190 : generate_bitmap_or_paths(PlannerInfo *root, RelOptInfo *rel,
1630 : List *clauses, List *other_clauses)
1631 : {
1632 690190 : List *result = NIL;
1633 : List *all_clauses;
1634 : ListCell *lc;
1635 :
1636 : /*
1637 : * We can use both the current and other clauses as context for
1638 : * build_paths_for_OR; no need to remove ORs from the lists.
1639 : */
1640 690190 : all_clauses = list_concat_copy(clauses, other_clauses);
1641 :
1642 1069816 : foreach(lc, clauses)
1643 : {
1644 379626 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1645 : List *pathlist;
1646 : Path *bitmapqual;
1647 : ListCell *j;
1648 : List *groupedArgs;
1649 379626 : List *inner_other_clauses = NIL;
1650 :
1651 : /* Ignore RestrictInfos that aren't ORs */
1652 379626 : if (!restriction_is_or_clause(rinfo))
1653 370534 : continue;
1654 :
1655 : /*
1656 : * We must be able to match at least one index to each of the arms of
1657 : * the OR, else we can't use it.
1658 : */
1659 9092 : pathlist = NIL;
1660 :
1661 : /*
1662 : * Group the similar OR-clause arguments into dedicated RestrictInfos,
1663 : * because each of those RestrictInfos has a chance to match the index
1664 : * as a whole.
1665 : */
1666 9092 : groupedArgs = group_similar_or_args(root, rel, rinfo);
1667 :
1668 9092 : if (groupedArgs != ((BoolExpr *) rinfo->orclause)->args)
1669 : {
1670 : /*
1671 : * Some parts of the rinfo were probably grouped. In this case,
1672 : * we have a set of sub-rinfos that together are an exact
1673 : * duplicate of rinfo. Thus, we need to remove the rinfo from
1674 : * other clauses. match_clauses_to_index detects duplicated
1675 : * iclauses by comparing pointers to original rinfos that would be
1676 : * different. So, we must delete rinfo to avoid de-facto
1677 : * duplicated clauses in the index clauses list.
1678 : */
1679 4024 : inner_other_clauses = list_delete(list_copy(all_clauses), rinfo);
1680 : }
1681 :
1682 11364 : foreach(j, groupedArgs)
1683 : {
1684 10324 : Node *orarg = (Node *) lfirst(j);
1685 : List *indlist;
1686 :
1687 : /* OR arguments should be ANDs or sub-RestrictInfos */
1688 10324 : if (is_andclause(orarg))
1689 : {
1690 1486 : List *andargs = ((BoolExpr *) orarg)->args;
1691 :
1692 1486 : indlist = build_paths_for_OR(root, rel,
1693 : andargs,
1694 : all_clauses);
1695 :
1696 : /* Recurse in case there are sub-ORs */
1697 1486 : indlist = list_concat(indlist,
1698 1486 : generate_bitmap_or_paths(root, rel,
1699 : andargs,
1700 : all_clauses));
1701 : }
1702 8838 : else if (restriction_is_or_clause(castNode(RestrictInfo, orarg)))
1703 : {
1704 306 : RestrictInfo *ri = castNode(RestrictInfo, orarg);
1705 :
1706 : /*
1707 : * Generate bitmap paths for the group of similar OR-clause
1708 : * arguments.
1709 : */
1710 306 : indlist = make_bitmap_paths_for_or_group(root,
1711 : rel, ri,
1712 : inner_other_clauses);
1713 :
1714 306 : if (indlist == NIL)
1715 : {
1716 6 : pathlist = NIL;
1717 6 : break;
1718 : }
1719 : else
1720 : {
1721 300 : pathlist = list_concat(pathlist, indlist);
1722 300 : continue;
1723 : }
1724 : }
1725 : else
1726 : {
1727 8532 : RestrictInfo *ri = castNode(RestrictInfo, orarg);
1728 : List *orargs;
1729 :
1730 8532 : orargs = list_make1(ri);
1731 :
1732 8532 : indlist = build_paths_for_OR(root, rel,
1733 : orargs,
1734 : all_clauses);
1735 : }
1736 :
1737 : /*
1738 : * If nothing matched this arm, we can't do anything with this OR
1739 : * clause.
1740 : */
1741 10018 : if (indlist == NIL)
1742 : {
1743 8046 : pathlist = NIL;
1744 8046 : break;
1745 : }
1746 :
1747 : /*
1748 : * OK, pick the most promising AND combination, and add it to
1749 : * pathlist.
1750 : */
1751 1972 : bitmapqual = choose_bitmap_and(root, rel, indlist);
1752 1972 : pathlist = lappend(pathlist, bitmapqual);
1753 : }
1754 :
1755 9092 : if (inner_other_clauses != NIL)
1756 2240 : list_free(inner_other_clauses);
1757 :
1758 : /*
1759 : * If we have a match for every arm, then turn them into a
1760 : * BitmapOrPath, and add to result list.
1761 : */
1762 9092 : if (pathlist != NIL)
1763 : {
1764 1040 : bitmapqual = (Path *) create_bitmap_or_path(root, rel, pathlist);
1765 1040 : result = lappend(result, bitmapqual);
1766 : }
1767 : }
1768 :
1769 690190 : return result;
1770 : }
1771 :
1772 :
1773 : /*
1774 : * choose_bitmap_and
1775 : * Given a nonempty list of bitmap paths, AND them into one path.
1776 : *
1777 : * This is a nontrivial decision since we can legally use any subset of the
1778 : * given path set. We want to choose a good tradeoff between selectivity
1779 : * and cost of computing the bitmap.
1780 : *
1781 : * The result is either a single one of the inputs, or a BitmapAndPath
1782 : * combining multiple inputs.
1783 : */
1784 : static Path *
1785 360392 : choose_bitmap_and(PlannerInfo *root, RelOptInfo *rel, List *paths)
1786 : {
1787 360392 : int npaths = list_length(paths);
1788 : PathClauseUsage **pathinfoarray;
1789 : PathClauseUsage *pathinfo;
1790 : List *clauselist;
1791 360392 : List *bestpaths = NIL;
1792 360392 : Cost bestcost = 0;
1793 : int i,
1794 : j;
1795 : ListCell *l;
1796 :
1797 : Assert(npaths > 0); /* else caller error */
1798 360392 : if (npaths == 1)
1799 280342 : return (Path *) linitial(paths); /* easy case */
1800 :
1801 : /*
1802 : * In theory we should consider every nonempty subset of the given paths.
1803 : * In practice that seems like overkill, given the crude nature of the
1804 : * estimates, not to mention the possible effects of higher-level AND and
1805 : * OR clauses. Moreover, it's completely impractical if there are a large
1806 : * number of paths, since the work would grow as O(2^N).
1807 : *
1808 : * As a heuristic, we first check for paths using exactly the same sets of
1809 : * WHERE clauses + index predicate conditions, and reject all but the
1810 : * cheapest-to-scan in any such group. This primarily gets rid of indexes
1811 : * that include the interesting columns but also irrelevant columns. (In
1812 : * situations where the DBA has gone overboard on creating variant
1813 : * indexes, this can make for a very large reduction in the number of
1814 : * paths considered further.)
1815 : *
1816 : * We then sort the surviving paths with the cheapest-to-scan first, and
1817 : * for each path, consider using that path alone as the basis for a bitmap
1818 : * scan. Then we consider bitmap AND scans formed from that path plus
1819 : * each subsequent (higher-cost) path, adding on a subsequent path if it
1820 : * results in a reduction in the estimated total scan cost. This means we
1821 : * consider about O(N^2) rather than O(2^N) path combinations, which is
1822 : * quite tolerable, especially given than N is usually reasonably small
1823 : * because of the prefiltering step. The cheapest of these is returned.
1824 : *
1825 : * We will only consider AND combinations in which no two indexes use the
1826 : * same WHERE clause. This is a bit of a kluge: it's needed because
1827 : * costsize.c and clausesel.c aren't very smart about redundant clauses.
1828 : * They will usually double-count the redundant clauses, producing a
1829 : * too-small selectivity that makes a redundant AND step look like it
1830 : * reduces the total cost. Perhaps someday that code will be smarter and
1831 : * we can remove this limitation. (But note that this also defends
1832 : * against flat-out duplicate input paths, which can happen because
1833 : * match_join_clauses_to_index will find the same OR join clauses that
1834 : * extract_restriction_or_clauses has pulled OR restriction clauses out
1835 : * of.)
1836 : *
1837 : * For the same reason, we reject AND combinations in which an index
1838 : * predicate clause duplicates another clause. Here we find it necessary
1839 : * to be even stricter: we'll reject a partial index if any of its
1840 : * predicate clauses are implied by the set of WHERE clauses and predicate
1841 : * clauses used so far. This covers cases such as a condition "x = 42"
1842 : * used with a plain index, followed by a clauseless scan of a partial
1843 : * index "WHERE x >= 40 AND x < 50". The partial index has been accepted
1844 : * only because "x = 42" was present, and so allowing it would partially
1845 : * double-count selectivity. (We could use predicate_implied_by on
1846 : * regular qual clauses too, to have a more intelligent, but much more
1847 : * expensive, check for redundancy --- but in most cases simple equality
1848 : * seems to suffice.)
1849 : */
1850 :
1851 : /*
1852 : * Extract clause usage info and detect any paths that use exactly the
1853 : * same set of clauses; keep only the cheapest-to-scan of any such groups.
1854 : * The surviving paths are put into an array for qsort'ing.
1855 : */
1856 80050 : pathinfoarray = palloc_array(PathClauseUsage *, npaths);
1857 80050 : clauselist = NIL;
1858 80050 : npaths = 0;
1859 263298 : foreach(l, paths)
1860 : {
1861 183248 : Path *ipath = (Path *) lfirst(l);
1862 :
1863 183248 : pathinfo = classify_index_clause_usage(ipath, &clauselist);
1864 :
1865 : /* If it's unclassifiable, treat it as distinct from all others */
1866 183248 : if (pathinfo->unclassifiable)
1867 : {
1868 0 : pathinfoarray[npaths++] = pathinfo;
1869 0 : continue;
1870 : }
1871 :
1872 286152 : for (i = 0; i < npaths; i++)
1873 : {
1874 252560 : if (!pathinfoarray[i]->unclassifiable &&
1875 126280 : bms_equal(pathinfo->clauseids, pathinfoarray[i]->clauseids))
1876 23376 : break;
1877 : }
1878 183248 : if (i < npaths)
1879 : {
1880 : /* duplicate clauseids, keep the cheaper one */
1881 : Cost ncost;
1882 : Cost ocost;
1883 : Selectivity nselec;
1884 : Selectivity oselec;
1885 :
1886 23376 : cost_bitmap_tree_node(pathinfo->path, &ncost, &nselec);
1887 23376 : cost_bitmap_tree_node(pathinfoarray[i]->path, &ocost, &oselec);
1888 23376 : if (ncost < ocost)
1889 5110 : pathinfoarray[i] = pathinfo;
1890 : }
1891 : else
1892 : {
1893 : /* not duplicate clauseids, add to array */
1894 159872 : pathinfoarray[npaths++] = pathinfo;
1895 : }
1896 : }
1897 :
1898 : /* If only one surviving path, we're done */
1899 80050 : if (npaths == 1)
1900 14640 : return pathinfoarray[0]->path;
1901 :
1902 : /* Sort the surviving paths by index access cost */
1903 65410 : qsort(pathinfoarray, npaths, sizeof(PathClauseUsage *),
1904 : path_usage_comparator);
1905 :
1906 : /*
1907 : * For each surviving index, consider it as an "AND group leader", and see
1908 : * whether adding on any of the later indexes results in an AND path with
1909 : * cheaper total cost than before. Then take the cheapest AND group.
1910 : *
1911 : * Note: paths that are either clauseless or unclassifiable will have
1912 : * empty clauseids, so that they will not be rejected by the clauseids
1913 : * filter here, nor will they cause later paths to be rejected by it.
1914 : */
1915 210642 : for (i = 0; i < npaths; i++)
1916 : {
1917 : Cost costsofar;
1918 : List *qualsofar;
1919 : Bitmapset *clauseidsofar;
1920 :
1921 145232 : pathinfo = pathinfoarray[i];
1922 145232 : paths = list_make1(pathinfo->path);
1923 145232 : costsofar = bitmap_scan_cost_est(root, rel, pathinfo->path);
1924 145232 : qualsofar = list_concat_copy(pathinfo->quals, pathinfo->preds);
1925 145232 : clauseidsofar = bms_copy(pathinfo->clauseids);
1926 :
1927 239904 : for (j = i + 1; j < npaths; j++)
1928 : {
1929 : Cost newcost;
1930 :
1931 94672 : pathinfo = pathinfoarray[j];
1932 : /* Check for redundancy */
1933 94672 : if (bms_overlap(pathinfo->clauseids, clauseidsofar))
1934 44554 : continue; /* consider it redundant */
1935 50118 : if (pathinfo->preds)
1936 : {
1937 24 : bool redundant = false;
1938 :
1939 : /* we check each predicate clause separately */
1940 24 : foreach(l, pathinfo->preds)
1941 : {
1942 24 : Node *np = (Node *) lfirst(l);
1943 :
1944 24 : if (predicate_implied_by(list_make1(np), qualsofar, false))
1945 : {
1946 24 : redundant = true;
1947 24 : break; /* out of inner foreach loop */
1948 : }
1949 : }
1950 24 : if (redundant)
1951 24 : continue;
1952 : }
1953 : /* tentatively add new path to paths, so we can estimate cost */
1954 50094 : paths = lappend(paths, pathinfo->path);
1955 50094 : newcost = bitmap_and_cost_est(root, rel, paths);
1956 50094 : if (newcost < costsofar)
1957 : {
1958 : /* keep new path in paths, update subsidiary variables */
1959 292 : costsofar = newcost;
1960 292 : qualsofar = list_concat(qualsofar, pathinfo->quals);
1961 292 : qualsofar = list_concat(qualsofar, pathinfo->preds);
1962 292 : clauseidsofar = bms_add_members(clauseidsofar,
1963 292 : pathinfo->clauseids);
1964 : }
1965 : else
1966 : {
1967 : /* reject new path, remove it from paths list */
1968 49802 : paths = list_truncate(paths, list_length(paths) - 1);
1969 : }
1970 : }
1971 :
1972 : /* Keep the cheapest AND-group (or singleton) */
1973 145232 : if (i == 0 || costsofar < bestcost)
1974 : {
1975 69676 : bestpaths = paths;
1976 69676 : bestcost = costsofar;
1977 : }
1978 :
1979 : /* some easy cleanup (we don't try real hard though) */
1980 145232 : list_free(qualsofar);
1981 : }
1982 :
1983 65410 : if (list_length(bestpaths) == 1)
1984 65142 : return (Path *) linitial(bestpaths); /* no need for AND */
1985 268 : return (Path *) create_bitmap_and_path(root, rel, bestpaths);
1986 : }
1987 :
1988 : /* qsort comparator to sort in increasing index access cost order */
1989 : static int
1990 86708 : path_usage_comparator(const void *a, const void *b)
1991 : {
1992 86708 : PathClauseUsage *pa = *(PathClauseUsage *const *) a;
1993 86708 : PathClauseUsage *pb = *(PathClauseUsage *const *) b;
1994 : Cost acost;
1995 : Cost bcost;
1996 : Selectivity aselec;
1997 : Selectivity bselec;
1998 :
1999 86708 : cost_bitmap_tree_node(pa->path, &acost, &aselec);
2000 86708 : cost_bitmap_tree_node(pb->path, &bcost, &bselec);
2001 :
2002 : /*
2003 : * If costs are the same, sort by selectivity.
2004 : */
2005 86708 : if (acost < bcost)
2006 57148 : return -1;
2007 29560 : if (acost > bcost)
2008 19476 : return 1;
2009 :
2010 10084 : if (aselec < bselec)
2011 5208 : return -1;
2012 4876 : if (aselec > bselec)
2013 1102 : return 1;
2014 :
2015 3774 : return 0;
2016 : }
2017 :
2018 : /*
2019 : * Estimate the cost of actually executing a bitmap scan with a single
2020 : * index path (which could be a BitmapAnd or BitmapOr node).
2021 : */
2022 : static Cost
2023 195326 : bitmap_scan_cost_est(PlannerInfo *root, RelOptInfo *rel, Path *ipath)
2024 : {
2025 : BitmapHeapPath bpath;
2026 :
2027 : /* Set up a dummy BitmapHeapPath */
2028 195326 : bpath.path.type = T_BitmapHeapPath;
2029 195326 : bpath.path.pathtype = T_BitmapHeapScan;
2030 195326 : bpath.path.parent = rel;
2031 195326 : bpath.path.pathtarget = rel->reltarget;
2032 195326 : bpath.path.param_info = ipath->param_info;
2033 195326 : bpath.path.pathkeys = NIL;
2034 195326 : bpath.bitmapqual = ipath;
2035 :
2036 : /*
2037 : * Check the cost of temporary path without considering parallelism.
2038 : * Parallel bitmap heap path will be considered at later stage.
2039 : */
2040 195326 : bpath.path.parallel_workers = 0;
2041 :
2042 : /* Now we can do cost_bitmap_heap_scan */
2043 195326 : cost_bitmap_heap_scan(&bpath.path, root, rel,
2044 : bpath.path.param_info,
2045 : ipath,
2046 : get_loop_count(root, rel->relid,
2047 195326 : PATH_REQ_OUTER(ipath)));
2048 :
2049 195326 : return bpath.path.total_cost;
2050 : }
2051 :
2052 : /*
2053 : * Estimate the cost of actually executing a BitmapAnd scan with the given
2054 : * inputs.
2055 : */
2056 : static Cost
2057 50094 : bitmap_and_cost_est(PlannerInfo *root, RelOptInfo *rel, List *paths)
2058 : {
2059 : BitmapAndPath *apath;
2060 :
2061 : /*
2062 : * Might as well build a real BitmapAndPath here, as the work is slightly
2063 : * too complicated to be worth repeating just to save one palloc.
2064 : */
2065 50094 : apath = create_bitmap_and_path(root, rel, paths);
2066 :
2067 50094 : return bitmap_scan_cost_est(root, rel, (Path *) apath);
2068 : }
2069 :
2070 :
2071 : /*
2072 : * classify_index_clause_usage
2073 : * Construct a PathClauseUsage struct describing the WHERE clauses and
2074 : * index predicate clauses used by the given indexscan path.
2075 : * We consider two clauses the same if they are equal().
2076 : *
2077 : * At some point we might want to migrate this info into the Path data
2078 : * structure proper, but for the moment it's only needed within
2079 : * choose_bitmap_and().
2080 : *
2081 : * *clauselist is used and expanded as needed to identify all the distinct
2082 : * clauses seen across successive calls. Caller must initialize it to NIL
2083 : * before first call of a set.
2084 : */
2085 : static PathClauseUsage *
2086 183248 : classify_index_clause_usage(Path *path, List **clauselist)
2087 : {
2088 : PathClauseUsage *result;
2089 : Bitmapset *clauseids;
2090 : ListCell *lc;
2091 :
2092 183248 : result = palloc_object(PathClauseUsage);
2093 183248 : result->path = path;
2094 :
2095 : /* Recursively find the quals and preds used by the path */
2096 183248 : result->quals = NIL;
2097 183248 : result->preds = NIL;
2098 183248 : find_indexpath_quals(path, &result->quals, &result->preds);
2099 :
2100 : /*
2101 : * Some machine-generated queries have outlandish numbers of qual clauses.
2102 : * To avoid getting into O(N^2) behavior even in this preliminary
2103 : * classification step, we want to limit the number of entries we can
2104 : * accumulate in *clauselist. Treat any path with more than 100 quals +
2105 : * preds as unclassifiable, which will cause calling code to consider it
2106 : * distinct from all other paths.
2107 : */
2108 183248 : if (list_length(result->quals) + list_length(result->preds) > 100)
2109 : {
2110 0 : result->clauseids = NULL;
2111 0 : result->unclassifiable = true;
2112 0 : return result;
2113 : }
2114 :
2115 : /* Build up a bitmapset representing the quals and preds */
2116 183248 : clauseids = NULL;
2117 423672 : foreach(lc, result->quals)
2118 : {
2119 240424 : Node *node = (Node *) lfirst(lc);
2120 :
2121 240424 : clauseids = bms_add_member(clauseids,
2122 : find_list_position(node, clauselist));
2123 : }
2124 183536 : foreach(lc, result->preds)
2125 : {
2126 288 : Node *node = (Node *) lfirst(lc);
2127 :
2128 288 : clauseids = bms_add_member(clauseids,
2129 : find_list_position(node, clauselist));
2130 : }
2131 183248 : result->clauseids = clauseids;
2132 183248 : result->unclassifiable = false;
2133 :
2134 183248 : return result;
2135 : }
2136 :
2137 :
2138 : /*
2139 : * find_indexpath_quals
2140 : *
2141 : * Given the Path structure for a plain or bitmap indexscan, extract lists
2142 : * of all the index clauses and index predicate conditions used in the Path.
2143 : * These are appended to the initial contents of *quals and *preds (hence
2144 : * caller should initialize those to NIL).
2145 : *
2146 : * Note we are not trying to produce an accurate representation of the AND/OR
2147 : * semantics of the Path, but just find out all the base conditions used.
2148 : *
2149 : * The result lists contain pointers to the expressions used in the Path,
2150 : * but all the list cells are freshly built, so it's safe to destructively
2151 : * modify the lists (eg, by concat'ing with other lists).
2152 : */
2153 : static void
2154 185658 : find_indexpath_quals(Path *bitmapqual, List **quals, List **preds)
2155 : {
2156 185658 : if (IsA(bitmapqual, BitmapAndPath))
2157 : {
2158 0 : BitmapAndPath *apath = (BitmapAndPath *) bitmapqual;
2159 : ListCell *l;
2160 :
2161 0 : foreach(l, apath->bitmapquals)
2162 : {
2163 0 : find_indexpath_quals((Path *) lfirst(l), quals, preds);
2164 : }
2165 : }
2166 185658 : else if (IsA(bitmapqual, BitmapOrPath))
2167 : {
2168 1310 : BitmapOrPath *opath = (BitmapOrPath *) bitmapqual;
2169 : ListCell *l;
2170 :
2171 3720 : foreach(l, opath->bitmapquals)
2172 : {
2173 2410 : find_indexpath_quals((Path *) lfirst(l), quals, preds);
2174 : }
2175 : }
2176 184348 : else if (IsA(bitmapqual, IndexPath))
2177 : {
2178 184348 : IndexPath *ipath = (IndexPath *) bitmapqual;
2179 : ListCell *l;
2180 :
2181 424772 : foreach(l, ipath->indexclauses)
2182 : {
2183 240424 : IndexClause *iclause = (IndexClause *) lfirst(l);
2184 :
2185 240424 : *quals = lappend(*quals, iclause->rinfo->clause);
2186 : }
2187 184348 : *preds = list_concat(*preds, ipath->indexinfo->indpred);
2188 : }
2189 : else
2190 0 : elog(ERROR, "unrecognized node type: %d", nodeTag(bitmapqual));
2191 185658 : }
2192 :
2193 :
2194 : /*
2195 : * find_list_position
2196 : * Return the given node's position (counting from 0) in the given
2197 : * list of nodes. If it's not equal() to any existing list member,
2198 : * add it at the end, and return that position.
2199 : */
2200 : static int
2201 240712 : find_list_position(Node *node, List **nodelist)
2202 : {
2203 : int i;
2204 : ListCell *lc;
2205 :
2206 240712 : i = 0;
2207 386990 : foreach(lc, *nodelist)
2208 : {
2209 217504 : Node *oldnode = (Node *) lfirst(lc);
2210 :
2211 217504 : if (equal(node, oldnode))
2212 71226 : return i;
2213 146278 : i++;
2214 : }
2215 :
2216 169486 : *nodelist = lappend(*nodelist, node);
2217 :
2218 169486 : return i;
2219 : }
2220 :
2221 :
2222 : /*
2223 : * check_index_only
2224 : * Determine whether an index-only scan is possible for this index.
2225 : */
2226 : static bool
2227 888950 : check_index_only(RelOptInfo *rel, IndexOptInfo *index)
2228 : {
2229 : bool result;
2230 888950 : Bitmapset *attrs_used = NULL;
2231 888950 : Bitmapset *index_canreturn_attrs = NULL;
2232 : ListCell *lc;
2233 : int i;
2234 :
2235 : /* Index-only scans must be enabled */
2236 888950 : if (!enable_indexonlyscan)
2237 3860 : return false;
2238 :
2239 : /*
2240 : * Check that all needed attributes of the relation are available from the
2241 : * index.
2242 : */
2243 :
2244 : /*
2245 : * First, identify all the attributes needed for joins or final output.
2246 : * Note: we must look at rel's targetlist, not the attr_needed data,
2247 : * because attr_needed isn't computed for inheritance child rels.
2248 : */
2249 885090 : pull_varattnos((Node *) rel->reltarget->exprs, rel->relid, &attrs_used);
2250 :
2251 : /*
2252 : * Add all the attributes used by restriction clauses; but consider only
2253 : * those clauses not implied by the index predicate, since ones that are
2254 : * so implied don't need to be checked explicitly in the plan.
2255 : *
2256 : * Note: attributes used only in index quals would not be needed at
2257 : * runtime either, if we are certain that the index is not lossy. However
2258 : * it'd be complicated to account for that accurately, and it doesn't
2259 : * matter in most cases, since we'd conclude that such attributes are
2260 : * available from the index anyway.
2261 : */
2262 1822368 : foreach(lc, index->indrestrictinfo)
2263 : {
2264 937278 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2265 :
2266 937278 : pull_varattnos((Node *) rinfo->clause, rel->relid, &attrs_used);
2267 : }
2268 :
2269 : /*
2270 : * Construct a bitmapset of columns that the index can return back in an
2271 : * index-only scan.
2272 : */
2273 2514196 : for (i = 0; i < index->ncolumns; i++)
2274 : {
2275 1629106 : int attno = index->indexkeys[i];
2276 :
2277 : /*
2278 : * For the moment, we just ignore index expressions. It might be nice
2279 : * to do something with them, later.
2280 : */
2281 1629106 : if (attno == 0)
2282 3310 : continue;
2283 :
2284 1625796 : if (index->canreturn[i])
2285 : index_canreturn_attrs =
2286 1350248 : bms_add_member(index_canreturn_attrs,
2287 : attno - FirstLowInvalidHeapAttributeNumber);
2288 : }
2289 :
2290 : /* Do we have all the necessary attributes? */
2291 885090 : result = bms_is_subset(attrs_used, index_canreturn_attrs);
2292 :
2293 885090 : bms_free(attrs_used);
2294 885090 : bms_free(index_canreturn_attrs);
2295 :
2296 885090 : return result;
2297 : }
2298 :
2299 : /*
2300 : * get_loop_count
2301 : * Choose the loop count estimate to use for costing a parameterized path
2302 : * with the given set of outer relids.
2303 : *
2304 : * Since we produce parameterized paths before we've begun to generate join
2305 : * relations, it's impossible to predict exactly how many times a parameterized
2306 : * path will be iterated; we don't know the size of the relation that will be
2307 : * on the outside of the nestloop. However, we should try to account for
2308 : * multiple iterations somehow in costing the path. The heuristic embodied
2309 : * here is to use the rowcount of the smallest other base relation needed in
2310 : * the join clauses used by the path. (We could alternatively consider the
2311 : * largest one, but that seems too optimistic.) This is of course the right
2312 : * answer for single-other-relation cases, and it seems like a reasonable
2313 : * zero-order approximation for multiway-join cases.
2314 : *
2315 : * In addition, we check to see if the other side of each join clause is on
2316 : * the inside of some semijoin that the current relation is on the outside of.
2317 : * If so, the only way that a parameterized path could be used is if the
2318 : * semijoin RHS has been unique-ified, so we should use the number of unique
2319 : * RHS rows rather than using the relation's raw rowcount.
2320 : *
2321 : * Note: for this to work, allpaths.c must establish all baserel size
2322 : * estimates before it begins to compute paths, or at least before it
2323 : * calls create_index_paths().
2324 : */
2325 : static double
2326 1236728 : get_loop_count(PlannerInfo *root, Index cur_relid, Relids outer_relids)
2327 : {
2328 : double result;
2329 : int outer_relid;
2330 :
2331 : /* For a non-parameterized path, just return 1.0 quickly */
2332 1236728 : if (outer_relids == NULL)
2333 843266 : return 1.0;
2334 :
2335 393462 : result = 0.0;
2336 393462 : outer_relid = -1;
2337 798102 : while ((outer_relid = bms_next_member(outer_relids, outer_relid)) >= 0)
2338 : {
2339 : RelOptInfo *outer_rel;
2340 : double rowcount;
2341 :
2342 : /* Paranoia: ignore bogus relid indexes */
2343 404640 : if (outer_relid >= root->simple_rel_array_size)
2344 0 : continue;
2345 404640 : outer_rel = root->simple_rel_array[outer_relid];
2346 404640 : if (outer_rel == NULL)
2347 254 : continue;
2348 : Assert(outer_rel->relid == outer_relid); /* sanity check on array */
2349 :
2350 : /* Other relation could be proven empty, if so ignore */
2351 404386 : if (IS_DUMMY_REL(outer_rel))
2352 24 : continue;
2353 :
2354 : /* Otherwise, rel's rows estimate should be valid by now */
2355 : Assert(outer_rel->rows > 0);
2356 :
2357 : /* Check to see if rel is on the inside of any semijoins */
2358 404362 : rowcount = adjust_rowcount_for_semijoins(root,
2359 : cur_relid,
2360 : outer_relid,
2361 : outer_rel->rows);
2362 :
2363 : /* Remember smallest row count estimate among the outer rels */
2364 404362 : if (result == 0.0 || result > rowcount)
2365 400678 : result = rowcount;
2366 : }
2367 : /* Return 1.0 if we found no valid relations (shouldn't happen) */
2368 393462 : return (result > 0.0) ? result : 1.0;
2369 : }
2370 :
2371 : /*
2372 : * Check to see if outer_relid is on the inside of any semijoin that cur_relid
2373 : * is on the outside of. If so, replace rowcount with the estimated number of
2374 : * unique rows from the semijoin RHS (assuming that's smaller, which it might
2375 : * not be). The estimate is crude but it's the best we can do at this stage
2376 : * of the proceedings.
2377 : */
2378 : static double
2379 404362 : adjust_rowcount_for_semijoins(PlannerInfo *root,
2380 : Index cur_relid,
2381 : Index outer_relid,
2382 : double rowcount)
2383 : {
2384 : ListCell *lc;
2385 :
2386 627266 : foreach(lc, root->join_info_list)
2387 : {
2388 222904 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
2389 :
2390 231252 : if (sjinfo->jointype == JOIN_SEMI &&
2391 12074 : bms_is_member(cur_relid, sjinfo->syn_lefthand) &&
2392 3726 : bms_is_member(outer_relid, sjinfo->syn_righthand))
2393 : {
2394 : /* Estimate number of unique-ified rows */
2395 : double nraw;
2396 : double nunique;
2397 :
2398 1360 : nraw = approximate_joinrel_size(root, sjinfo->syn_righthand);
2399 1360 : nunique = estimate_num_groups(root,
2400 : sjinfo->semi_rhs_exprs,
2401 : nraw,
2402 : NULL,
2403 : NULL);
2404 1360 : if (rowcount > nunique)
2405 462 : rowcount = nunique;
2406 : }
2407 : }
2408 404362 : return rowcount;
2409 : }
2410 :
2411 : /*
2412 : * Make an approximate estimate of the size of a joinrel.
2413 : *
2414 : * We don't have enough info at this point to get a good estimate, so we
2415 : * just multiply the base relation sizes together. Fortunately, this is
2416 : * the right answer anyway for the most common case with a single relation
2417 : * on the RHS of a semijoin. Also, estimate_num_groups() has only a weak
2418 : * dependency on its input_rows argument (it basically uses it as a clamp).
2419 : * So we might be able to get a fairly decent end result even with a severe
2420 : * overestimate of the RHS's raw size.
2421 : */
2422 : static double
2423 1360 : approximate_joinrel_size(PlannerInfo *root, Relids relids)
2424 : {
2425 1360 : double rowcount = 1.0;
2426 : int relid;
2427 :
2428 1360 : relid = -1;
2429 2924 : while ((relid = bms_next_member(relids, relid)) >= 0)
2430 : {
2431 : RelOptInfo *rel;
2432 :
2433 : /* Paranoia: ignore bogus relid indexes */
2434 1564 : if (relid >= root->simple_rel_array_size)
2435 0 : continue;
2436 1564 : rel = root->simple_rel_array[relid];
2437 1564 : if (rel == NULL)
2438 0 : continue;
2439 : Assert(rel->relid == relid); /* sanity check on array */
2440 :
2441 : /* Relation could be proven empty, if so ignore */
2442 1564 : if (IS_DUMMY_REL(rel))
2443 0 : continue;
2444 :
2445 : /* Otherwise, rel's rows estimate should be valid by now */
2446 : Assert(rel->rows > 0);
2447 :
2448 : /* Accumulate product */
2449 1564 : rowcount *= rel->rows;
2450 : }
2451 1360 : return rowcount;
2452 : }
2453 :
2454 :
2455 : /****************************************************************************
2456 : * ---- ROUTINES TO CHECK QUERY CLAUSES ----
2457 : ****************************************************************************/
2458 :
2459 : /*
2460 : * match_restriction_clauses_to_index
2461 : * Identify restriction clauses for the rel that match the index.
2462 : * Matching clauses are added to *clauseset.
2463 : */
2464 : static void
2465 732726 : match_restriction_clauses_to_index(PlannerInfo *root,
2466 : IndexOptInfo *index,
2467 : IndexClauseSet *clauseset)
2468 : {
2469 : /* We can ignore clauses that are implied by the index predicate */
2470 732726 : match_clauses_to_index(root, index->indrestrictinfo, index, clauseset);
2471 732726 : }
2472 :
2473 : /*
2474 : * match_join_clauses_to_index
2475 : * Identify join clauses for the rel that match the index.
2476 : * Matching clauses are added to *clauseset.
2477 : * Also, add any potentially usable join OR clauses to *joinorclauses.
2478 : * They also might be processed by match_clause_to_index() as a whole.
2479 : */
2480 : static void
2481 732726 : match_join_clauses_to_index(PlannerInfo *root,
2482 : RelOptInfo *rel, IndexOptInfo *index,
2483 : IndexClauseSet *clauseset,
2484 : List **joinorclauses)
2485 : {
2486 : ListCell *lc;
2487 :
2488 : /* Scan the rel's join clauses */
2489 986190 : foreach(lc, rel->joininfo)
2490 : {
2491 253464 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2492 :
2493 : /* Check if clause can be moved to this rel */
2494 253464 : if (!join_clause_is_movable_to(rinfo, rel))
2495 154902 : continue;
2496 :
2497 : /*
2498 : * Potentially usable, so see if it matches the index or is an OR. Use
2499 : * list_append_unique_ptr() here to avoid possible duplicates when
2500 : * processing the same clauses with different indexes.
2501 : */
2502 98562 : if (restriction_is_or_clause(rinfo))
2503 13380 : *joinorclauses = list_append_unique_ptr(*joinorclauses, rinfo);
2504 :
2505 98562 : match_clause_to_index(root, rinfo, index, clauseset);
2506 : }
2507 732726 : }
2508 :
2509 : /*
2510 : * match_eclass_clauses_to_index
2511 : * Identify EquivalenceClass join clauses for the rel that match the index.
2512 : * Matching clauses are added to *clauseset.
2513 : */
2514 : static void
2515 732726 : match_eclass_clauses_to_index(PlannerInfo *root, IndexOptInfo *index,
2516 : IndexClauseSet *clauseset)
2517 : {
2518 : int indexcol;
2519 :
2520 : /* No work if rel is not in any such ECs */
2521 732726 : if (!index->rel->has_eclass_joins)
2522 421060 : return;
2523 :
2524 818944 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2525 : {
2526 : ec_member_matches_arg arg;
2527 : List *clauses;
2528 :
2529 : /* Generate clauses, skipping any that join to lateral_referencers */
2530 507278 : arg.index = index;
2531 507278 : arg.indexcol = indexcol;
2532 507278 : clauses = generate_implied_equalities_for_column(root,
2533 : index->rel,
2534 : ec_member_matches_indexcol,
2535 : &arg,
2536 507278 : index->rel->lateral_referencers);
2537 :
2538 : /*
2539 : * We have to check whether the results actually do match the index,
2540 : * since for non-btree indexes the EC's equality operators might not
2541 : * be in the index opclass (cf ec_member_matches_indexcol).
2542 : */
2543 507278 : match_clauses_to_index(root, clauses, index, clauseset);
2544 : }
2545 : }
2546 :
2547 : /*
2548 : * match_clauses_to_index
2549 : * Perform match_clause_to_index() for each clause in a list.
2550 : * Matching clauses are added to *clauseset.
2551 : */
2552 : static void
2553 1270062 : match_clauses_to_index(PlannerInfo *root,
2554 : List *clauses,
2555 : IndexOptInfo *index,
2556 : IndexClauseSet *clauseset)
2557 : {
2558 : ListCell *lc;
2559 :
2560 2279482 : foreach(lc, clauses)
2561 : {
2562 1009420 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
2563 :
2564 1009420 : match_clause_to_index(root, rinfo, index, clauseset);
2565 : }
2566 1270062 : }
2567 :
2568 : /*
2569 : * match_clause_to_index
2570 : * Test whether a qual clause can be used with an index.
2571 : *
2572 : * If the clause is usable, add an IndexClause entry for it to the appropriate
2573 : * list in *clauseset. (*clauseset must be initialized to zeroes before first
2574 : * call.)
2575 : *
2576 : * Note: in some circumstances we may find the same RestrictInfos coming from
2577 : * multiple places. Defend against redundant outputs by refusing to add a
2578 : * clause twice (pointer equality should be a good enough check for this).
2579 : *
2580 : * Note: it's possible that a badly-defined index could have multiple matching
2581 : * columns. We always select the first match if so; this avoids scenarios
2582 : * wherein we get an inflated idea of the index's selectivity by using the
2583 : * same clause multiple times with different index columns.
2584 : */
2585 : static void
2586 1107982 : match_clause_to_index(PlannerInfo *root,
2587 : RestrictInfo *rinfo,
2588 : IndexOptInfo *index,
2589 : IndexClauseSet *clauseset)
2590 : {
2591 : int indexcol;
2592 :
2593 : /*
2594 : * Never match pseudoconstants to indexes. (Normally a match could not
2595 : * happen anyway, since a pseudoconstant clause couldn't contain a Var,
2596 : * but what if someone builds an expression index on a constant? It's not
2597 : * totally unreasonable to do so with a partial index, either.)
2598 : */
2599 1107982 : if (rinfo->pseudoconstant)
2600 13832 : return;
2601 :
2602 : /*
2603 : * If clause can't be used as an indexqual because it must wait till after
2604 : * some lower-security-level restriction clause, reject it.
2605 : */
2606 1094150 : if (!restriction_is_securely_promotable(rinfo, index->rel))
2607 522 : return;
2608 :
2609 : /* OK, check each index key column for a match */
2610 2401606 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
2611 : {
2612 : IndexClause *iclause;
2613 : ListCell *lc;
2614 :
2615 : /* Ignore duplicates */
2616 1810560 : foreach(lc, clauseset->indexclauses[indexcol])
2617 : {
2618 79818 : iclause = (IndexClause *) lfirst(lc);
2619 :
2620 79818 : if (iclause->rinfo == rinfo)
2621 0 : return;
2622 : }
2623 :
2624 : /* OK, try to match the clause to the index column */
2625 1730742 : iclause = match_clause_to_indexcol(root,
2626 : rinfo,
2627 : indexcol,
2628 : index);
2629 1730742 : if (iclause)
2630 : {
2631 : /* Success, so record it */
2632 422764 : clauseset->indexclauses[indexcol] =
2633 422764 : lappend(clauseset->indexclauses[indexcol], iclause);
2634 422764 : clauseset->nonempty = true;
2635 422764 : return;
2636 : }
2637 : }
2638 : }
2639 :
2640 : /*
2641 : * match_clause_to_indexcol()
2642 : * Determine whether a restriction clause matches a column of an index,
2643 : * and if so, build an IndexClause node describing the details.
2644 : *
2645 : * To match an index normally, an operator clause:
2646 : *
2647 : * (1) must be in the form (indexkey op const) or (const op indexkey);
2648 : * and
2649 : * (2) must contain an operator which is in the index's operator family
2650 : * for this column; and
2651 : * (3) must match the collation of the index, if collation is relevant.
2652 : *
2653 : * Our definition of "const" is exceedingly liberal: we allow anything that
2654 : * doesn't involve a volatile function or a Var of the index's relation.
2655 : * In particular, Vars belonging to other relations of the query are
2656 : * accepted here, since a clause of that form can be used in a
2657 : * parameterized indexscan. It's the responsibility of higher code levels
2658 : * to manage restriction and join clauses appropriately.
2659 : *
2660 : * Note: we do need to check for Vars of the index's relation on the
2661 : * "const" side of the clause, since clauses like (a.f1 OP (b.f2 OP a.f3))
2662 : * are not processable by a parameterized indexscan on a.f1, whereas
2663 : * something like (a.f1 OP (b.f2 OP c.f3)) is.
2664 : *
2665 : * Presently, the executor can only deal with indexquals that have the
2666 : * indexkey on the left, so we can only use clauses that have the indexkey
2667 : * on the right if we can commute the clause to put the key on the left.
2668 : * We handle that by generating an IndexClause with the correctly-commuted
2669 : * opclause as a derived indexqual.
2670 : *
2671 : * If the index has a collation, the clause must have the same collation.
2672 : * For collation-less indexes, we assume it doesn't matter; this is
2673 : * necessary for cases like "hstore ? text", wherein hstore's operators
2674 : * don't care about collation but the clause will get marked with a
2675 : * collation anyway because of the text argument. (This logic is
2676 : * embodied in the macro IndexCollMatchesExprColl.)
2677 : *
2678 : * It is also possible to match RowCompareExpr clauses to indexes (but
2679 : * currently, only btree indexes handle this).
2680 : *
2681 : * It is also possible to match ScalarArrayOpExpr clauses to indexes, when
2682 : * the clause is of the form "indexkey op ANY (arrayconst)".
2683 : *
2684 : * It is also possible to match a list of OR clauses if it might be
2685 : * transformed into a single ScalarArrayOpExpr clause. On success,
2686 : * the returning index clause will contain a transformed clause.
2687 : *
2688 : * For boolean indexes, it is also possible to match the clause directly
2689 : * to the indexkey; or perhaps the clause is (NOT indexkey).
2690 : *
2691 : * And, last but not least, some operators and functions can be processed
2692 : * to derive (typically lossy) indexquals from a clause that isn't in
2693 : * itself indexable. If we see that any operand of an OpExpr or FuncExpr
2694 : * matches the index key, and the function has a planner support function
2695 : * attached to it, we'll invoke the support function to see if such an
2696 : * indexqual can be built.
2697 : *
2698 : * 'rinfo' is the clause to be tested (as a RestrictInfo node).
2699 : * 'indexcol' is a column number of 'index' (counting from 0).
2700 : * 'index' is the index of interest.
2701 : *
2702 : * Returns an IndexClause if the clause can be used with this index key,
2703 : * or NULL if not.
2704 : *
2705 : * NOTE: This routine always returns NULL if the clause is an AND clause.
2706 : * Higher-level routines deal with OR and AND clauses. OR clause can be
2707 : * matched as a whole by match_orclause_to_indexcol() though.
2708 : */
2709 : static IndexClause *
2710 1730742 : match_clause_to_indexcol(PlannerInfo *root,
2711 : RestrictInfo *rinfo,
2712 : int indexcol,
2713 : IndexOptInfo *index)
2714 : {
2715 : IndexClause *iclause;
2716 1730742 : Expr *clause = rinfo->clause;
2717 : Oid opfamily;
2718 :
2719 : Assert(indexcol < index->nkeycolumns);
2720 :
2721 : /*
2722 : * Historically this code has coped with NULL clauses. That's probably
2723 : * not possible anymore, but we might as well continue to cope.
2724 : */
2725 1730742 : if (clause == NULL)
2726 0 : return NULL;
2727 :
2728 : /* First check for boolean-index cases. */
2729 1730742 : opfamily = index->opfamily[indexcol];
2730 1730742 : if (IsBooleanOpfamily(opfamily))
2731 : {
2732 490 : iclause = match_boolean_index_clause(root, rinfo, indexcol, index);
2733 490 : if (iclause)
2734 314 : return iclause;
2735 : }
2736 :
2737 : /*
2738 : * Clause must be an opclause, funcclause, ScalarArrayOpExpr,
2739 : * RowCompareExpr, or OR-clause that could be converted to SAOP. Or, if
2740 : * the index supports it, we can handle IS NULL/NOT NULL clauses.
2741 : */
2742 1730428 : if (IsA(clause, OpExpr))
2743 : {
2744 1453992 : return match_opclause_to_indexcol(root, rinfo, indexcol, index);
2745 : }
2746 276436 : else if (IsA(clause, FuncExpr))
2747 : {
2748 29990 : return match_funcclause_to_indexcol(root, rinfo, indexcol, index);
2749 : }
2750 246446 : else if (IsA(clause, ScalarArrayOpExpr))
2751 : {
2752 80210 : return match_saopclause_to_indexcol(root, rinfo, indexcol, index);
2753 : }
2754 166236 : else if (IsA(clause, RowCompareExpr))
2755 : {
2756 504 : return match_rowcompare_to_indexcol(root, rinfo, indexcol, index);
2757 : }
2758 165732 : else if (restriction_is_or_clause(rinfo))
2759 : {
2760 43958 : return match_orclause_to_indexcol(root, rinfo, indexcol, index);
2761 : }
2762 121774 : else if (index->amsearchnulls && IsA(clause, NullTest))
2763 : {
2764 15640 : NullTest *nt = (NullTest *) clause;
2765 :
2766 31280 : if (!nt->argisrow &&
2767 15640 : match_index_to_operand((Node *) nt->arg, indexcol, index))
2768 : {
2769 1448 : iclause = makeNode(IndexClause);
2770 1448 : iclause->rinfo = rinfo;
2771 1448 : iclause->indexquals = list_make1(rinfo);
2772 1448 : iclause->lossy = false;
2773 1448 : iclause->indexcol = indexcol;
2774 1448 : iclause->indexcols = NIL;
2775 1448 : return iclause;
2776 : }
2777 : }
2778 :
2779 120326 : return NULL;
2780 : }
2781 :
2782 : /*
2783 : * IsBooleanOpfamily
2784 : * Detect whether an opfamily supports boolean equality as an operator.
2785 : *
2786 : * If the opfamily OID is in the range of built-in objects, we can rely
2787 : * on hard-wired knowledge of which built-in opfamilies support this.
2788 : * For extension opfamilies, there's no choice but to do a catcache lookup.
2789 : */
2790 : static bool
2791 2399766 : IsBooleanOpfamily(Oid opfamily)
2792 : {
2793 2399766 : if (opfamily < FirstNormalObjectId)
2794 2396128 : return IsBuiltinBooleanOpfamily(opfamily);
2795 : else
2796 3638 : return op_in_opfamily(BooleanEqualOperator, opfamily);
2797 : }
2798 :
2799 : /*
2800 : * match_boolean_index_clause
2801 : * Recognize restriction clauses that can be matched to a boolean index.
2802 : *
2803 : * The idea here is that, for an index on a boolean column that supports the
2804 : * BooleanEqualOperator, we can transform a plain reference to the indexkey
2805 : * into "indexkey = true", or "NOT indexkey" into "indexkey = false", etc,
2806 : * so as to make the expression indexable using the index's "=" operator.
2807 : * Since Postgres 8.1, we must do this because constant simplification does
2808 : * the reverse transformation; without this code there'd be no way to use
2809 : * such an index at all.
2810 : *
2811 : * This should be called only when IsBooleanOpfamily() recognizes the
2812 : * index's operator family. We check to see if the clause matches the
2813 : * index's key, and if so, build a suitable IndexClause.
2814 : */
2815 : static IndexClause *
2816 1782 : match_boolean_index_clause(PlannerInfo *root,
2817 : RestrictInfo *rinfo,
2818 : int indexcol,
2819 : IndexOptInfo *index)
2820 : {
2821 1782 : Node *clause = (Node *) rinfo->clause;
2822 1782 : Expr *op = NULL;
2823 :
2824 : /* Direct match? */
2825 1782 : if (match_index_to_operand(clause, indexcol, index))
2826 : {
2827 : /* convert to indexkey = TRUE */
2828 314 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2829 : (Expr *) clause,
2830 314 : (Expr *) makeBoolConst(true, false),
2831 : InvalidOid, InvalidOid);
2832 : }
2833 : /* NOT clause? */
2834 1468 : else if (is_notclause(clause))
2835 : {
2836 1208 : Node *arg = (Node *) get_notclausearg((Expr *) clause);
2837 :
2838 1208 : if (match_index_to_operand(arg, indexcol, index))
2839 : {
2840 : /* convert to indexkey = FALSE */
2841 1208 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2842 : (Expr *) arg,
2843 1208 : (Expr *) makeBoolConst(false, false),
2844 : InvalidOid, InvalidOid);
2845 : }
2846 : }
2847 :
2848 : /*
2849 : * Since we only consider clauses at top level of WHERE, we can convert
2850 : * indexkey IS TRUE and indexkey IS FALSE to index searches as well. The
2851 : * different meaning for NULL isn't important.
2852 : */
2853 260 : else if (clause && IsA(clause, BooleanTest))
2854 : {
2855 48 : BooleanTest *btest = (BooleanTest *) clause;
2856 48 : Node *arg = (Node *) btest->arg;
2857 :
2858 78 : if (btest->booltesttype == IS_TRUE &&
2859 30 : match_index_to_operand(arg, indexcol, index))
2860 : {
2861 : /* convert to indexkey = TRUE */
2862 30 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2863 : (Expr *) arg,
2864 30 : (Expr *) makeBoolConst(true, false),
2865 : InvalidOid, InvalidOid);
2866 : }
2867 36 : else if (btest->booltesttype == IS_FALSE &&
2868 18 : match_index_to_operand(arg, indexcol, index))
2869 : {
2870 : /* convert to indexkey = FALSE */
2871 18 : op = make_opclause(BooleanEqualOperator, BOOLOID, false,
2872 : (Expr *) arg,
2873 18 : (Expr *) makeBoolConst(false, false),
2874 : InvalidOid, InvalidOid);
2875 : }
2876 : }
2877 :
2878 : /*
2879 : * If we successfully made an operator clause from the given qual, we must
2880 : * wrap it in an IndexClause. It's not lossy.
2881 : */
2882 1782 : if (op)
2883 : {
2884 1570 : IndexClause *iclause = makeNode(IndexClause);
2885 :
2886 1570 : iclause->rinfo = rinfo;
2887 1570 : iclause->indexquals = list_make1(make_simple_restrictinfo(root, op));
2888 1570 : iclause->lossy = false;
2889 1570 : iclause->indexcol = indexcol;
2890 1570 : iclause->indexcols = NIL;
2891 1570 : return iclause;
2892 : }
2893 :
2894 212 : return NULL;
2895 : }
2896 :
2897 : /*
2898 : * match_opclause_to_indexcol()
2899 : * Handles the OpExpr case for match_clause_to_indexcol(),
2900 : * which see for comments.
2901 : */
2902 : static IndexClause *
2903 1453992 : match_opclause_to_indexcol(PlannerInfo *root,
2904 : RestrictInfo *rinfo,
2905 : int indexcol,
2906 : IndexOptInfo *index)
2907 : {
2908 : IndexClause *iclause;
2909 1453992 : OpExpr *clause = (OpExpr *) rinfo->clause;
2910 : Node *leftop,
2911 : *rightop;
2912 : Oid expr_op;
2913 : Oid expr_coll;
2914 : Index index_relid;
2915 : Oid opfamily;
2916 : Oid idxcollation;
2917 :
2918 : /*
2919 : * Only binary operators need apply. (In theory, a planner support
2920 : * function could do something with a unary operator, but it seems
2921 : * unlikely to be worth the cycles to check.)
2922 : */
2923 1453992 : if (list_length(clause->args) != 2)
2924 0 : return NULL;
2925 :
2926 1453992 : leftop = (Node *) linitial(clause->args);
2927 1453992 : rightop = (Node *) lsecond(clause->args);
2928 1453992 : expr_op = clause->opno;
2929 1453992 : expr_coll = clause->inputcollid;
2930 :
2931 1453992 : index_relid = index->rel->relid;
2932 1453992 : opfamily = index->opfamily[indexcol];
2933 1453992 : idxcollation = index->indexcollations[indexcol];
2934 :
2935 : /*
2936 : * Check for clauses of the form: (indexkey operator constant) or
2937 : * (constant operator indexkey). See match_clause_to_indexcol's notes
2938 : * about const-ness.
2939 : *
2940 : * Note that we don't ask the support function about clauses that don't
2941 : * have one of these forms. Again, in principle it might be possible to
2942 : * do something, but it seems unlikely to be worth the cycles to check.
2943 : */
2944 1453992 : if (match_index_to_operand(leftop, indexcol, index) &&
2945 357036 : !bms_is_member(index_relid, rinfo->right_relids) &&
2946 356862 : !contain_volatile_functions(rightop))
2947 : {
2948 706990 : if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
2949 350128 : op_in_opfamily(expr_op, opfamily))
2950 : {
2951 342536 : iclause = makeNode(IndexClause);
2952 342536 : iclause->rinfo = rinfo;
2953 342536 : iclause->indexquals = list_make1(rinfo);
2954 342536 : iclause->lossy = false;
2955 342536 : iclause->indexcol = indexcol;
2956 342536 : iclause->indexcols = NIL;
2957 342536 : return iclause;
2958 : }
2959 :
2960 : /*
2961 : * If we didn't find a member of the index's opfamily, try the support
2962 : * function for the operator's underlying function.
2963 : */
2964 14326 : set_opfuncid(clause); /* make sure we have opfuncid */
2965 14326 : return get_index_clause_from_support(root,
2966 : rinfo,
2967 : clause->opfuncid,
2968 : 0, /* indexarg on left */
2969 : indexcol,
2970 : index);
2971 : }
2972 :
2973 1097130 : if (match_index_to_operand(rightop, indexcol, index) &&
2974 69084 : !bms_is_member(index_relid, rinfo->left_relids) &&
2975 68958 : !contain_volatile_functions(leftop))
2976 : {
2977 68958 : if (IndexCollMatchesExprColl(idxcollation, expr_coll))
2978 : {
2979 68946 : Oid comm_op = get_commutator(expr_op);
2980 :
2981 137892 : if (OidIsValid(comm_op) &&
2982 68946 : op_in_opfamily(comm_op, opfamily))
2983 : {
2984 : RestrictInfo *commrinfo;
2985 :
2986 : /* Build a commuted OpExpr and RestrictInfo */
2987 68470 : commrinfo = commute_restrictinfo(rinfo, comm_op);
2988 :
2989 : /* Make an IndexClause showing that as a derived qual */
2990 68470 : iclause = makeNode(IndexClause);
2991 68470 : iclause->rinfo = rinfo;
2992 68470 : iclause->indexquals = list_make1(commrinfo);
2993 68470 : iclause->lossy = false;
2994 68470 : iclause->indexcol = indexcol;
2995 68470 : iclause->indexcols = NIL;
2996 68470 : return iclause;
2997 : }
2998 : }
2999 :
3000 : /*
3001 : * If we didn't find a member of the index's opfamily, try the support
3002 : * function for the operator's underlying function.
3003 : */
3004 488 : set_opfuncid(clause); /* make sure we have opfuncid */
3005 488 : return get_index_clause_from_support(root,
3006 : rinfo,
3007 : clause->opfuncid,
3008 : 1, /* indexarg on right */
3009 : indexcol,
3010 : index);
3011 : }
3012 :
3013 1028172 : return NULL;
3014 : }
3015 :
3016 : /*
3017 : * match_funcclause_to_indexcol()
3018 : * Handles the FuncExpr case for match_clause_to_indexcol(),
3019 : * which see for comments.
3020 : */
3021 : static IndexClause *
3022 29990 : match_funcclause_to_indexcol(PlannerInfo *root,
3023 : RestrictInfo *rinfo,
3024 : int indexcol,
3025 : IndexOptInfo *index)
3026 : {
3027 29990 : FuncExpr *clause = (FuncExpr *) rinfo->clause;
3028 : int indexarg;
3029 : ListCell *lc;
3030 :
3031 : /*
3032 : * We have no built-in intelligence about function clauses, but if there's
3033 : * a planner support function, it might be able to do something. But, to
3034 : * cut down on wasted planning cycles, only call the support function if
3035 : * at least one argument matches the target index column.
3036 : *
3037 : * Note that we don't insist on the other arguments being pseudoconstants;
3038 : * the support function has to check that. This is to allow cases where
3039 : * only some of the other arguments need to be included in the indexqual.
3040 : */
3041 29990 : indexarg = 0;
3042 64502 : foreach(lc, clause->args)
3043 : {
3044 40336 : Node *op = (Node *) lfirst(lc);
3045 :
3046 40336 : if (match_index_to_operand(op, indexcol, index))
3047 : {
3048 5824 : return get_index_clause_from_support(root,
3049 : rinfo,
3050 : clause->funcid,
3051 : indexarg,
3052 : indexcol,
3053 : index);
3054 : }
3055 :
3056 34512 : indexarg++;
3057 : }
3058 :
3059 24166 : return NULL;
3060 : }
3061 :
3062 : /*
3063 : * get_index_clause_from_support()
3064 : * If the function has a planner support function, try to construct
3065 : * an IndexClause using indexquals created by the support function.
3066 : */
3067 : static IndexClause *
3068 20638 : get_index_clause_from_support(PlannerInfo *root,
3069 : RestrictInfo *rinfo,
3070 : Oid funcid,
3071 : int indexarg,
3072 : int indexcol,
3073 : IndexOptInfo *index)
3074 : {
3075 20638 : Oid prosupport = get_func_support(funcid);
3076 : SupportRequestIndexCondition req;
3077 : List *sresult;
3078 :
3079 20638 : if (!OidIsValid(prosupport))
3080 12436 : return NULL;
3081 :
3082 8202 : req.type = T_SupportRequestIndexCondition;
3083 8202 : req.root = root;
3084 8202 : req.funcid = funcid;
3085 8202 : req.node = (Node *) rinfo->clause;
3086 8202 : req.indexarg = indexarg;
3087 8202 : req.index = index;
3088 8202 : req.indexcol = indexcol;
3089 8202 : req.opfamily = index->opfamily[indexcol];
3090 8202 : req.indexcollation = index->indexcollations[indexcol];
3091 :
3092 8202 : req.lossy = true; /* default assumption */
3093 :
3094 : sresult = (List *)
3095 8202 : DatumGetPointer(OidFunctionCall1(prosupport,
3096 : PointerGetDatum(&req)));
3097 :
3098 8202 : if (sresult != NIL)
3099 : {
3100 1402 : IndexClause *iclause = makeNode(IndexClause);
3101 1402 : List *indexquals = NIL;
3102 : ListCell *lc;
3103 :
3104 : /*
3105 : * The support function API says it should just give back bare
3106 : * clauses, so here we must wrap each one in a RestrictInfo.
3107 : */
3108 4128 : foreach(lc, sresult)
3109 : {
3110 2726 : Expr *clause = (Expr *) lfirst(lc);
3111 :
3112 2726 : indexquals = lappend(indexquals,
3113 2726 : make_simple_restrictinfo(root, clause));
3114 : }
3115 :
3116 1402 : iclause->rinfo = rinfo;
3117 1402 : iclause->indexquals = indexquals;
3118 1402 : iclause->lossy = req.lossy;
3119 1402 : iclause->indexcol = indexcol;
3120 1402 : iclause->indexcols = NIL;
3121 :
3122 1402 : return iclause;
3123 : }
3124 :
3125 6800 : return NULL;
3126 : }
3127 :
3128 : /*
3129 : * match_saopclause_to_indexcol()
3130 : * Handles the ScalarArrayOpExpr case for match_clause_to_indexcol(),
3131 : * which see for comments.
3132 : */
3133 : static IndexClause *
3134 80210 : match_saopclause_to_indexcol(PlannerInfo *root,
3135 : RestrictInfo *rinfo,
3136 : int indexcol,
3137 : IndexOptInfo *index)
3138 : {
3139 80210 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
3140 : Node *leftop,
3141 : *rightop;
3142 : Relids right_relids;
3143 : Oid expr_op;
3144 : Oid expr_coll;
3145 : Index index_relid;
3146 : Oid opfamily;
3147 : Oid idxcollation;
3148 :
3149 : /* We only accept ANY clauses, not ALL */
3150 80210 : if (!saop->useOr)
3151 9322 : return NULL;
3152 70888 : leftop = (Node *) linitial(saop->args);
3153 70888 : rightop = (Node *) lsecond(saop->args);
3154 70888 : right_relids = pull_varnos(root, rightop);
3155 70888 : expr_op = saop->opno;
3156 70888 : expr_coll = saop->inputcollid;
3157 :
3158 70888 : index_relid = index->rel->relid;
3159 70888 : opfamily = index->opfamily[indexcol];
3160 70888 : idxcollation = index->indexcollations[indexcol];
3161 :
3162 : /*
3163 : * We must have indexkey on the left and a pseudo-constant array argument.
3164 : */
3165 70888 : if (match_index_to_operand(leftop, indexcol, index) &&
3166 7376 : !bms_is_member(index_relid, right_relids) &&
3167 7376 : !contain_volatile_functions(rightop))
3168 : {
3169 14746 : if (IndexCollMatchesExprColl(idxcollation, expr_coll) &&
3170 7370 : op_in_opfamily(expr_op, opfamily))
3171 : {
3172 7358 : IndexClause *iclause = makeNode(IndexClause);
3173 :
3174 7358 : iclause->rinfo = rinfo;
3175 7358 : iclause->indexquals = list_make1(rinfo);
3176 7358 : iclause->lossy = false;
3177 7358 : iclause->indexcol = indexcol;
3178 7358 : iclause->indexcols = NIL;
3179 7358 : return iclause;
3180 : }
3181 :
3182 : /*
3183 : * We do not currently ask support functions about ScalarArrayOpExprs,
3184 : * though in principle we could.
3185 : */
3186 : }
3187 :
3188 63530 : return NULL;
3189 : }
3190 :
3191 : /*
3192 : * match_rowcompare_to_indexcol()
3193 : * Handles the RowCompareExpr case for match_clause_to_indexcol(),
3194 : * which see for comments.
3195 : *
3196 : * In this routine we check whether the first column of the row comparison
3197 : * matches the target index column. This is sufficient to guarantee that some
3198 : * index condition can be constructed from the RowCompareExpr --- the rest
3199 : * is handled by expand_indexqual_rowcompare().
3200 : */
3201 : static IndexClause *
3202 504 : match_rowcompare_to_indexcol(PlannerInfo *root,
3203 : RestrictInfo *rinfo,
3204 : int indexcol,
3205 : IndexOptInfo *index)
3206 : {
3207 504 : RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3208 : Index index_relid;
3209 : Oid opfamily;
3210 : Oid idxcollation;
3211 : Node *leftop,
3212 : *rightop;
3213 : bool var_on_left;
3214 : Oid expr_op;
3215 : Oid expr_coll;
3216 :
3217 : /* Forget it if we're not dealing with a btree index */
3218 504 : if (index->relam != BTREE_AM_OID)
3219 0 : return NULL;
3220 :
3221 504 : index_relid = index->rel->relid;
3222 504 : opfamily = index->opfamily[indexcol];
3223 504 : idxcollation = index->indexcollations[indexcol];
3224 :
3225 : /*
3226 : * We could do the matching on the basis of insisting that the opfamily
3227 : * shown in the RowCompareExpr be the same as the index column's opfamily,
3228 : * but that could fail in the presence of reverse-sort opfamilies: it'd be
3229 : * a matter of chance whether RowCompareExpr had picked the forward or
3230 : * reverse-sort family. So look only at the operator, and match if it is
3231 : * a member of the index's opfamily (after commutation, if the indexkey is
3232 : * on the right). We'll worry later about whether any additional
3233 : * operators are matchable to the index.
3234 : */
3235 504 : leftop = (Node *) linitial(clause->largs);
3236 504 : rightop = (Node *) linitial(clause->rargs);
3237 504 : expr_op = linitial_oid(clause->opnos);
3238 504 : expr_coll = linitial_oid(clause->inputcollids);
3239 :
3240 : /* Collations must match, if relevant */
3241 504 : if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
3242 0 : return NULL;
3243 :
3244 : /*
3245 : * These syntactic tests are the same as in match_opclause_to_indexcol()
3246 : */
3247 504 : if (match_index_to_operand(leftop, indexcol, index) &&
3248 162 : !bms_is_member(index_relid, pull_varnos(root, rightop)) &&
3249 162 : !contain_volatile_functions(rightop))
3250 : {
3251 : /* OK, indexkey is on left */
3252 162 : var_on_left = true;
3253 : }
3254 342 : else if (match_index_to_operand(rightop, indexcol, index) &&
3255 24 : !bms_is_member(index_relid, pull_varnos(root, leftop)) &&
3256 24 : !contain_volatile_functions(leftop))
3257 : {
3258 : /* indexkey is on right, so commute the operator */
3259 24 : expr_op = get_commutator(expr_op);
3260 24 : if (expr_op == InvalidOid)
3261 0 : return NULL;
3262 24 : var_on_left = false;
3263 : }
3264 : else
3265 318 : return NULL;
3266 :
3267 : /* We're good if the operator is the right type of opfamily member */
3268 186 : switch (get_op_opfamily_strategy(expr_op, opfamily))
3269 : {
3270 186 : case BTLessStrategyNumber:
3271 : case BTLessEqualStrategyNumber:
3272 : case BTGreaterEqualStrategyNumber:
3273 : case BTGreaterStrategyNumber:
3274 186 : return expand_indexqual_rowcompare(root,
3275 : rinfo,
3276 : indexcol,
3277 : index,
3278 : expr_op,
3279 : var_on_left);
3280 : }
3281 :
3282 0 : return NULL;
3283 : }
3284 :
3285 : /*
3286 : * match_orclause_to_indexcol()
3287 : * Handles the OR-expr case for match_clause_to_indexcol() in the case
3288 : * when it could be transformed to ScalarArrayOpExpr.
3289 : *
3290 : * In this routine, we attempt to transform a list of OR-clause args into a
3291 : * single SAOP expression matching the target index column. On success,
3292 : * return an IndexClause containing the transformed expression.
3293 : * Return NULL if the transformation fails.
3294 : */
3295 : static IndexClause *
3296 43958 : match_orclause_to_indexcol(PlannerInfo *root,
3297 : RestrictInfo *rinfo,
3298 : int indexcol,
3299 : IndexOptInfo *index)
3300 : {
3301 43958 : BoolExpr *orclause = (BoolExpr *) rinfo->orclause;
3302 43958 : List *consts = NIL;
3303 43958 : Node *indexExpr = NULL;
3304 43958 : Oid matchOpno = InvalidOid;
3305 43958 : Oid consttype = InvalidOid;
3306 43958 : Oid arraytype = InvalidOid;
3307 43958 : Oid inputcollid = InvalidOid;
3308 43958 : bool firstTime = true;
3309 43958 : bool haveNonConst = false;
3310 43958 : Index indexRelid = index->rel->relid;
3311 : ScalarArrayOpExpr *saopexpr;
3312 : IndexClause *iclause;
3313 : ListCell *lc;
3314 :
3315 : /* Forget it if index doesn't support SAOP clauses */
3316 43958 : if (!index->amsearcharray)
3317 106 : return NULL;
3318 :
3319 : /*
3320 : * Try to convert a list of OR-clauses to a single SAOP expression. Each
3321 : * OR entry must be in the form: (indexkey operator constant) or (constant
3322 : * operator indexkey). Operators of all the entries must match. On
3323 : * discovery of anything unsupported, we give up by breaking out of the
3324 : * loop immediately and returning NULL.
3325 : */
3326 48432 : foreach(lc, orclause->args)
3327 : {
3328 47382 : RestrictInfo *subRinfo = (RestrictInfo *) lfirst(lc);
3329 : OpExpr *subClause;
3330 : Oid opno;
3331 : Node *leftop,
3332 : *rightop;
3333 : Node *constExpr;
3334 :
3335 : /* If it's not a RestrictInfo (i.e. it's a sub-AND), we can't use it */
3336 47382 : if (!IsA(subRinfo, RestrictInfo))
3337 5284 : break;
3338 :
3339 : /* Only operator clauses can match */
3340 42098 : if (!IsA(subRinfo->clause, OpExpr))
3341 12944 : break;
3342 :
3343 29154 : subClause = (OpExpr *) subRinfo->clause;
3344 29154 : opno = subClause->opno;
3345 :
3346 : /* Only binary operators can match */
3347 29154 : if (list_length(subClause->args) != 2)
3348 0 : break;
3349 :
3350 : /*
3351 : * Check for clauses of the form: (indexkey operator constant) or
3352 : * (constant operator indexkey). These tests should agree with
3353 : * match_opclause_to_indexcol.
3354 : */
3355 29154 : leftop = (Node *) linitial(subClause->args);
3356 29154 : rightop = (Node *) lsecond(subClause->args);
3357 29154 : if (match_index_to_operand(leftop, indexcol, index) &&
3358 6420 : !bms_is_member(indexRelid, subRinfo->right_relids) &&
3359 6390 : !contain_volatile_functions(rightop))
3360 : {
3361 6390 : indexExpr = leftop;
3362 6390 : constExpr = rightop;
3363 : }
3364 22764 : else if (match_index_to_operand(rightop, indexcol, index) &&
3365 194 : !bms_is_member(indexRelid, subRinfo->left_relids) &&
3366 188 : !contain_volatile_functions(leftop))
3367 : {
3368 188 : opno = get_commutator(opno);
3369 188 : if (!OidIsValid(opno))
3370 : {
3371 : /* commutator doesn't exist, we can't reverse the order */
3372 0 : break;
3373 : }
3374 188 : indexExpr = rightop;
3375 188 : constExpr = leftop;
3376 : }
3377 : else
3378 : {
3379 : break;
3380 : }
3381 :
3382 : /*
3383 : * Save information about the operator, type, and collation for the
3384 : * first matching qual. Then, check that subsequent quals match the
3385 : * first.
3386 : */
3387 6578 : if (firstTime)
3388 : {
3389 4820 : matchOpno = opno;
3390 4820 : consttype = exprType(constExpr);
3391 4820 : arraytype = get_array_type(consttype);
3392 4820 : inputcollid = subClause->inputcollid;
3393 :
3394 : /*
3395 : * Check that the operator is presented in the opfamily and that
3396 : * the expression collation matches the index collation. Also,
3397 : * there must be an array type to construct an array later.
3398 : */
3399 4820 : if (!IndexCollMatchesExprColl(index->indexcollations[indexcol],
3400 4694 : inputcollid) ||
3401 4694 : !op_in_opfamily(matchOpno, index->opfamily[indexcol]) ||
3402 : !OidIsValid(arraytype))
3403 : break;
3404 :
3405 : /*
3406 : * Disallow if either type is RECORD, mainly because we can't be
3407 : * positive that all the RHS expressions are the same record type.
3408 : */
3409 2966 : if (consttype == RECORDOID || exprType(indexExpr) == RECORDOID)
3410 : break;
3411 :
3412 2966 : firstTime = false;
3413 : }
3414 : else
3415 : {
3416 1758 : if (matchOpno != opno ||
3417 3228 : inputcollid != subClause->inputcollid ||
3418 1614 : consttype != exprType(constExpr))
3419 : break;
3420 : }
3421 :
3422 : /*
3423 : * The righthand inputs don't necessarily have to be plain Consts, but
3424 : * make_SAOP_expr needs to know if any are not.
3425 : */
3426 4580 : if (!IsA(constExpr, Const))
3427 368 : haveNonConst = true;
3428 :
3429 4580 : consts = lappend(consts, constExpr);
3430 : }
3431 :
3432 : /*
3433 : * Handle failed conversion from breaking out of the loop because of an
3434 : * unsupported qual. Also check that we have an indexExpr, just in case
3435 : * the OR list was somehow empty (it shouldn't be). Return NULL to
3436 : * indicate the conversion failed.
3437 : */
3438 43852 : if (lc != NULL || indexExpr == NULL)
3439 : {
3440 42802 : list_free(consts); /* might as well */
3441 42802 : return NULL;
3442 : }
3443 :
3444 : /*
3445 : * Build the new SAOP node. We use the indexExpr from the last OR arm;
3446 : * since all the arms passed match_index_to_operand, it shouldn't matter
3447 : * which one we use. But using "inputcollid" twice is a bit of a cheat:
3448 : * we might end up with an array Const node that is labeled with a
3449 : * collation despite its elements being of a noncollatable type. But
3450 : * nothing is likely to complain about that, so we don't bother being more
3451 : * accurate.
3452 : */
3453 1050 : saopexpr = make_SAOP_expr(matchOpno, indexExpr, consttype, inputcollid,
3454 : inputcollid, consts, haveNonConst);
3455 : Assert(saopexpr != NULL);
3456 :
3457 : /*
3458 : * Finally, build an IndexClause based on the SAOP node. It's not lossy.
3459 : */
3460 1050 : iclause = makeNode(IndexClause);
3461 1050 : iclause->rinfo = rinfo;
3462 1050 : iclause->indexquals = list_make1(make_simple_restrictinfo(root,
3463 : (Expr *) saopexpr));
3464 1050 : iclause->lossy = false;
3465 1050 : iclause->indexcol = indexcol;
3466 1050 : iclause->indexcols = NIL;
3467 1050 : return iclause;
3468 : }
3469 :
3470 : /*
3471 : * expand_indexqual_rowcompare --- expand a single indexqual condition
3472 : * that is a RowCompareExpr
3473 : *
3474 : * It's already known that the first column of the row comparison matches
3475 : * the specified column of the index. We can use additional columns of the
3476 : * row comparison as index qualifications, so long as they match the index
3477 : * in the "same direction", ie, the indexkeys are all on the same side of the
3478 : * clause and the operators are all the same-type members of the opfamilies.
3479 : *
3480 : * If all the columns of the RowCompareExpr match in this way, we just use it
3481 : * as-is, except for possibly commuting it to put the indexkeys on the left.
3482 : *
3483 : * Otherwise, we build a shortened RowCompareExpr (if more than one
3484 : * column matches) or a simple OpExpr (if the first-column match is all
3485 : * there is). In these cases the modified clause is always "<=" or ">="
3486 : * even when the original was "<" or ">" --- this is necessary to match all
3487 : * the rows that could match the original. (We are building a lossy version
3488 : * of the row comparison when we do this, so we set lossy = true.)
3489 : *
3490 : * Note: this is really just the last half of match_rowcompare_to_indexcol,
3491 : * but we split it out for comprehensibility.
3492 : */
3493 : static IndexClause *
3494 186 : expand_indexqual_rowcompare(PlannerInfo *root,
3495 : RestrictInfo *rinfo,
3496 : int indexcol,
3497 : IndexOptInfo *index,
3498 : Oid expr_op,
3499 : bool var_on_left)
3500 : {
3501 186 : IndexClause *iclause = makeNode(IndexClause);
3502 186 : RowCompareExpr *clause = (RowCompareExpr *) rinfo->clause;
3503 : int op_strategy;
3504 : Oid op_lefttype;
3505 : Oid op_righttype;
3506 : int matching_cols;
3507 : List *expr_ops;
3508 : List *opfamilies;
3509 : List *lefttypes;
3510 : List *righttypes;
3511 : List *new_ops;
3512 : List *var_args;
3513 : List *non_var_args;
3514 :
3515 186 : iclause->rinfo = rinfo;
3516 186 : iclause->indexcol = indexcol;
3517 :
3518 186 : if (var_on_left)
3519 : {
3520 162 : var_args = clause->largs;
3521 162 : non_var_args = clause->rargs;
3522 : }
3523 : else
3524 : {
3525 24 : var_args = clause->rargs;
3526 24 : non_var_args = clause->largs;
3527 : }
3528 :
3529 186 : get_op_opfamily_properties(expr_op, index->opfamily[indexcol], false,
3530 : &op_strategy,
3531 : &op_lefttype,
3532 : &op_righttype);
3533 :
3534 : /* Initialize returned list of which index columns are used */
3535 186 : iclause->indexcols = list_make1_int(indexcol);
3536 :
3537 : /* Build lists of ops, opfamilies and operator datatypes in case needed */
3538 186 : expr_ops = list_make1_oid(expr_op);
3539 186 : opfamilies = list_make1_oid(index->opfamily[indexcol]);
3540 186 : lefttypes = list_make1_oid(op_lefttype);
3541 186 : righttypes = list_make1_oid(op_righttype);
3542 :
3543 : /*
3544 : * See how many of the remaining columns match some index column in the
3545 : * same way. As in match_clause_to_indexcol(), the "other" side of any
3546 : * potential index condition is OK as long as it doesn't use Vars from the
3547 : * indexed relation.
3548 : */
3549 186 : matching_cols = 1;
3550 :
3551 354 : while (matching_cols < list_length(var_args))
3552 : {
3553 222 : Node *varop = (Node *) list_nth(var_args, matching_cols);
3554 222 : Node *constop = (Node *) list_nth(non_var_args, matching_cols);
3555 : int i;
3556 :
3557 222 : expr_op = list_nth_oid(clause->opnos, matching_cols);
3558 222 : if (!var_on_left)
3559 : {
3560 : /* indexkey is on right, so commute the operator */
3561 24 : expr_op = get_commutator(expr_op);
3562 24 : if (expr_op == InvalidOid)
3563 0 : break; /* operator is not usable */
3564 : }
3565 222 : if (bms_is_member(index->rel->relid, pull_varnos(root, constop)))
3566 0 : break; /* no good, Var on wrong side */
3567 222 : if (contain_volatile_functions(constop))
3568 0 : break; /* no good, volatile comparison value */
3569 :
3570 : /*
3571 : * The Var side can match any key column of the index.
3572 : */
3573 516 : for (i = 0; i < index->nkeycolumns; i++)
3574 : {
3575 462 : if (match_index_to_operand(varop, i, index) &&
3576 168 : get_op_opfamily_strategy(expr_op,
3577 168 : index->opfamily[i]) == op_strategy &&
3578 168 : IndexCollMatchesExprColl(index->indexcollations[i],
3579 : list_nth_oid(clause->inputcollids,
3580 : matching_cols)))
3581 : break;
3582 : }
3583 222 : if (i >= index->nkeycolumns)
3584 54 : break; /* no match found */
3585 :
3586 : /* Add column number to returned list */
3587 168 : iclause->indexcols = lappend_int(iclause->indexcols, i);
3588 :
3589 : /* Add operator info to lists */
3590 168 : get_op_opfamily_properties(expr_op, index->opfamily[i], false,
3591 : &op_strategy,
3592 : &op_lefttype,
3593 : &op_righttype);
3594 168 : expr_ops = lappend_oid(expr_ops, expr_op);
3595 168 : opfamilies = lappend_oid(opfamilies, index->opfamily[i]);
3596 168 : lefttypes = lappend_oid(lefttypes, op_lefttype);
3597 168 : righttypes = lappend_oid(righttypes, op_righttype);
3598 :
3599 : /* This column matches, keep scanning */
3600 168 : matching_cols++;
3601 : }
3602 :
3603 : /* Result is non-lossy if all columns are usable as index quals */
3604 186 : iclause->lossy = (matching_cols != list_length(clause->opnos));
3605 :
3606 : /*
3607 : * We can use rinfo->clause as-is if we have var on left and it's all
3608 : * usable as index quals.
3609 : */
3610 186 : if (var_on_left && !iclause->lossy)
3611 120 : iclause->indexquals = list_make1(rinfo);
3612 : else
3613 : {
3614 : /*
3615 : * We have to generate a modified rowcompare (possibly just one
3616 : * OpExpr). The painful part of this is changing < to <= or > to >=,
3617 : * so deal with that first.
3618 : */
3619 66 : if (!iclause->lossy)
3620 : {
3621 : /* very easy, just use the commuted operators */
3622 12 : new_ops = expr_ops;
3623 : }
3624 54 : else if (op_strategy == BTLessEqualStrategyNumber ||
3625 54 : op_strategy == BTGreaterEqualStrategyNumber)
3626 : {
3627 : /* easy, just use the same (possibly commuted) operators */
3628 0 : new_ops = list_truncate(expr_ops, matching_cols);
3629 : }
3630 : else
3631 : {
3632 : ListCell *opfamilies_cell;
3633 : ListCell *lefttypes_cell;
3634 : ListCell *righttypes_cell;
3635 :
3636 54 : if (op_strategy == BTLessStrategyNumber)
3637 30 : op_strategy = BTLessEqualStrategyNumber;
3638 24 : else if (op_strategy == BTGreaterStrategyNumber)
3639 24 : op_strategy = BTGreaterEqualStrategyNumber;
3640 : else
3641 0 : elog(ERROR, "unexpected strategy number %d", op_strategy);
3642 54 : new_ops = NIL;
3643 144 : forthree(opfamilies_cell, opfamilies,
3644 : lefttypes_cell, lefttypes,
3645 : righttypes_cell, righttypes)
3646 : {
3647 90 : Oid opfam = lfirst_oid(opfamilies_cell);
3648 90 : Oid lefttype = lfirst_oid(lefttypes_cell);
3649 90 : Oid righttype = lfirst_oid(righttypes_cell);
3650 :
3651 90 : expr_op = get_opfamily_member(opfam, lefttype, righttype,
3652 : op_strategy);
3653 90 : if (!OidIsValid(expr_op)) /* should not happen */
3654 0 : elog(ERROR, "missing operator %d(%u,%u) in opfamily %u",
3655 : op_strategy, lefttype, righttype, opfam);
3656 90 : new_ops = lappend_oid(new_ops, expr_op);
3657 : }
3658 : }
3659 :
3660 : /* If we have more than one matching col, create a subset rowcompare */
3661 66 : if (matching_cols > 1)
3662 : {
3663 48 : RowCompareExpr *rc = makeNode(RowCompareExpr);
3664 :
3665 48 : rc->cmptype = (CompareType) op_strategy;
3666 48 : rc->opnos = new_ops;
3667 48 : rc->opfamilies = list_copy_head(clause->opfamilies,
3668 : matching_cols);
3669 48 : rc->inputcollids = list_copy_head(clause->inputcollids,
3670 : matching_cols);
3671 48 : rc->largs = list_copy_head(var_args, matching_cols);
3672 48 : rc->rargs = list_copy_head(non_var_args, matching_cols);
3673 48 : iclause->indexquals = list_make1(make_simple_restrictinfo(root,
3674 : (Expr *) rc));
3675 : }
3676 : else
3677 : {
3678 : Expr *op;
3679 :
3680 : /* We don't report an index column list in this case */
3681 18 : iclause->indexcols = NIL;
3682 :
3683 18 : op = make_opclause(linitial_oid(new_ops), BOOLOID, false,
3684 18 : copyObject(linitial(var_args)),
3685 18 : copyObject(linitial(non_var_args)),
3686 : InvalidOid,
3687 18 : linitial_oid(clause->inputcollids));
3688 18 : iclause->indexquals = list_make1(make_simple_restrictinfo(root, op));
3689 : }
3690 : }
3691 :
3692 186 : return iclause;
3693 : }
3694 :
3695 :
3696 : /****************************************************************************
3697 : * ---- ROUTINES TO CHECK ORDERING OPERATORS ----
3698 : ****************************************************************************/
3699 :
3700 : /*
3701 : * match_pathkeys_to_index
3702 : * For the given 'index' and 'pathkeys', output a list of suitable ORDER
3703 : * BY expressions, each of the form "indexedcol operator pseudoconstant",
3704 : * along with an integer list of the index column numbers (zero based)
3705 : * that each clause would be used with.
3706 : *
3707 : * This attempts to find an ORDER BY and index column number for all items in
3708 : * the pathkey list, however, if we're unable to match any given pathkey to an
3709 : * index column, we return just the ones matched by the function so far. This
3710 : * allows callers who are interested in partial matches to get them. Callers
3711 : * can determine a partial match vs a full match by checking the outputted
3712 : * list lengths. A full match will have one item in the output lists for each
3713 : * item in the given 'pathkeys' list.
3714 : */
3715 : static void
3716 1074 : match_pathkeys_to_index(IndexOptInfo *index, List *pathkeys,
3717 : List **orderby_clauses_p,
3718 : List **clause_columns_p)
3719 : {
3720 : ListCell *lc1;
3721 :
3722 1074 : *orderby_clauses_p = NIL; /* set default results */
3723 1074 : *clause_columns_p = NIL;
3724 :
3725 : /* Only indexes with the amcanorderbyop property are interesting here */
3726 1074 : if (!index->amcanorderbyop)
3727 0 : return;
3728 :
3729 1548 : foreach(lc1, pathkeys)
3730 : {
3731 1080 : PathKey *pathkey = (PathKey *) lfirst(lc1);
3732 1080 : bool found = false;
3733 : EquivalenceMemberIterator it;
3734 : EquivalenceMember *member;
3735 :
3736 :
3737 : /* Pathkey must request default sort order for the target opfamily */
3738 1080 : if (pathkey->pk_cmptype != COMPARE_LT || pathkey->pk_nulls_first)
3739 606 : return;
3740 :
3741 : /* If eclass is volatile, no hope of using an indexscan */
3742 1046 : if (pathkey->pk_eclass->ec_has_volatile)
3743 0 : return;
3744 :
3745 : /*
3746 : * Try to match eclass member expression(s) to index. Note that child
3747 : * EC members are considered, but only when they belong to the target
3748 : * relation. (Unlike regular members, the same expression could be a
3749 : * child member of more than one EC. Therefore, the same index could
3750 : * be considered to match more than one pathkey list, which is OK
3751 : * here. See also get_eclass_for_sort_expr.)
3752 : */
3753 1046 : setup_eclass_member_iterator(&it, pathkey->pk_eclass,
3754 1046 : index->rel->relids);
3755 1650 : while ((member = eclass_member_iterator_next(&it)) != NULL)
3756 : {
3757 : int indexcol;
3758 :
3759 : /* No possibility of match if it references other relations */
3760 1078 : if (!bms_equal(member->em_relids, index->rel->relids))
3761 32 : continue;
3762 :
3763 : /*
3764 : * We allow any column of the index to match each pathkey; they
3765 : * don't have to match left-to-right as you might expect. This is
3766 : * correct for GiST, and it doesn't matter for SP-GiST because
3767 : * that doesn't handle multiple columns anyway, and no other
3768 : * existing AMs support amcanorderbyop. We might need different
3769 : * logic in future for other implementations.
3770 : */
3771 1906 : for (indexcol = 0; indexcol < index->nkeycolumns; indexcol++)
3772 : {
3773 : Expr *expr;
3774 :
3775 1334 : expr = match_clause_to_ordering_op(index,
3776 : indexcol,
3777 : member->em_expr,
3778 : pathkey->pk_opfamily);
3779 1334 : if (expr)
3780 : {
3781 474 : *orderby_clauses_p = lappend(*orderby_clauses_p, expr);
3782 474 : *clause_columns_p = lappend_int(*clause_columns_p, indexcol);
3783 474 : found = true;
3784 474 : break;
3785 : }
3786 : }
3787 :
3788 1046 : if (found) /* don't want to look at remaining members */
3789 474 : break;
3790 : }
3791 :
3792 : /*
3793 : * Return the matches found so far when this pathkey couldn't be
3794 : * matched to the index.
3795 : */
3796 1046 : if (!found)
3797 572 : return;
3798 : }
3799 : }
3800 :
3801 : /*
3802 : * match_clause_to_ordering_op
3803 : * Determines whether an ordering operator expression matches an
3804 : * index column.
3805 : *
3806 : * This is similar to, but simpler than, match_clause_to_indexcol.
3807 : * We only care about simple OpExpr cases. The input is a bare
3808 : * expression that is being ordered by, which must be of the form
3809 : * (indexkey op const) or (const op indexkey) where op is an ordering
3810 : * operator for the column's opfamily.
3811 : *
3812 : * 'index' is the index of interest.
3813 : * 'indexcol' is a column number of 'index' (counting from 0).
3814 : * 'clause' is the ordering expression to be tested.
3815 : * 'pk_opfamily' is the btree opfamily describing the required sort order.
3816 : *
3817 : * Note that we currently do not consider the collation of the ordering
3818 : * operator's result. In practical cases the result type will be numeric
3819 : * and thus have no collation, and it's not very clear what to match to
3820 : * if it did have a collation. The index's collation should match the
3821 : * ordering operator's input collation, not its result.
3822 : *
3823 : * If successful, return 'clause' as-is if the indexkey is on the left,
3824 : * otherwise a commuted copy of 'clause'. If no match, return NULL.
3825 : */
3826 : static Expr *
3827 1334 : match_clause_to_ordering_op(IndexOptInfo *index,
3828 : int indexcol,
3829 : Expr *clause,
3830 : Oid pk_opfamily)
3831 : {
3832 : Oid opfamily;
3833 : Oid idxcollation;
3834 : Node *leftop,
3835 : *rightop;
3836 : Oid expr_op;
3837 : Oid expr_coll;
3838 : Oid sortfamily;
3839 : bool commuted;
3840 :
3841 : Assert(indexcol < index->nkeycolumns);
3842 :
3843 1334 : opfamily = index->opfamily[indexcol];
3844 1334 : idxcollation = index->indexcollations[indexcol];
3845 :
3846 : /*
3847 : * Clause must be a binary opclause.
3848 : */
3849 1334 : if (!is_opclause(clause))
3850 860 : return NULL;
3851 474 : leftop = get_leftop(clause);
3852 474 : rightop = get_rightop(clause);
3853 474 : if (!leftop || !rightop)
3854 0 : return NULL;
3855 474 : expr_op = ((OpExpr *) clause)->opno;
3856 474 : expr_coll = ((OpExpr *) clause)->inputcollid;
3857 :
3858 : /*
3859 : * We can forget the whole thing right away if wrong collation.
3860 : */
3861 474 : if (!IndexCollMatchesExprColl(idxcollation, expr_coll))
3862 0 : return NULL;
3863 :
3864 : /*
3865 : * Check for clauses of the form: (indexkey operator constant) or
3866 : * (constant operator indexkey).
3867 : */
3868 474 : if (match_index_to_operand(leftop, indexcol, index) &&
3869 450 : !contain_var_clause(rightop) &&
3870 450 : !contain_volatile_functions(rightop))
3871 : {
3872 450 : commuted = false;
3873 : }
3874 24 : else if (match_index_to_operand(rightop, indexcol, index) &&
3875 24 : !contain_var_clause(leftop) &&
3876 24 : !contain_volatile_functions(leftop))
3877 : {
3878 : /* Might match, but we need a commuted operator */
3879 24 : expr_op = get_commutator(expr_op);
3880 24 : if (expr_op == InvalidOid)
3881 0 : return NULL;
3882 24 : commuted = true;
3883 : }
3884 : else
3885 0 : return NULL;
3886 :
3887 : /*
3888 : * Is the (commuted) operator an ordering operator for the opfamily? And
3889 : * if so, does it yield the right sorting semantics?
3890 : */
3891 474 : sortfamily = get_op_opfamily_sortfamily(expr_op, opfamily);
3892 474 : if (sortfamily != pk_opfamily)
3893 0 : return NULL;
3894 :
3895 : /* We have a match. Return clause or a commuted version thereof. */
3896 474 : if (commuted)
3897 : {
3898 24 : OpExpr *newclause = makeNode(OpExpr);
3899 :
3900 : /* flat-copy all the fields of clause */
3901 24 : memcpy(newclause, clause, sizeof(OpExpr));
3902 :
3903 : /* commute it */
3904 24 : newclause->opno = expr_op;
3905 24 : newclause->opfuncid = InvalidOid;
3906 24 : newclause->args = list_make2(rightop, leftop);
3907 :
3908 24 : clause = (Expr *) newclause;
3909 : }
3910 :
3911 474 : return clause;
3912 : }
3913 :
3914 :
3915 : /****************************************************************************
3916 : * ---- ROUTINES TO DO PARTIAL INDEX PREDICATE TESTS ----
3917 : ****************************************************************************/
3918 :
3919 : /*
3920 : * check_index_predicates
3921 : * Set the predicate-derived IndexOptInfo fields for each index
3922 : * of the specified relation.
3923 : *
3924 : * predOK is set true if the index is partial and its predicate is satisfied
3925 : * for this query, ie the query's WHERE clauses imply the predicate.
3926 : *
3927 : * indrestrictinfo is set to the relation's baserestrictinfo list less any
3928 : * conditions that are implied by the index's predicate. (Obviously, for a
3929 : * non-partial index, this is the same as baserestrictinfo.) Such conditions
3930 : * can be dropped from the plan when using the index, in certain cases.
3931 : *
3932 : * At one time it was possible for this to get re-run after adding more
3933 : * restrictions to the rel, thus possibly letting us prove more indexes OK.
3934 : * That doesn't happen any more (at least not in the core code's usage),
3935 : * but this code still supports it in case extensions want to mess with the
3936 : * baserestrictinfo list. We assume that adding more restrictions can't make
3937 : * an index not predOK. We must recompute indrestrictinfo each time, though,
3938 : * to make sure any newly-added restrictions get into it if needed.
3939 : */
3940 : void
3941 417270 : check_index_predicates(PlannerInfo *root, RelOptInfo *rel)
3942 : {
3943 : List *clauselist;
3944 : bool have_partial;
3945 : bool is_target_rel;
3946 : Relids otherrels;
3947 : ListCell *lc;
3948 :
3949 : /* Indexes are available only on base or "other" member relations. */
3950 : Assert(IS_SIMPLE_REL(rel));
3951 :
3952 : /*
3953 : * Initialize the indrestrictinfo lists to be identical to
3954 : * baserestrictinfo, and check whether there are any partial indexes. If
3955 : * not, this is all we need to do.
3956 : */
3957 417270 : have_partial = false;
3958 1150770 : foreach(lc, rel->indexlist)
3959 : {
3960 733500 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
3961 :
3962 733500 : index->indrestrictinfo = rel->baserestrictinfo;
3963 733500 : if (index->indpred)
3964 1032 : have_partial = true;
3965 : }
3966 417270 : if (!have_partial)
3967 416562 : return;
3968 :
3969 : /*
3970 : * Construct a list of clauses that we can assume true for the purpose of
3971 : * proving the index(es) usable. Restriction clauses for the rel are
3972 : * always usable, and so are any join clauses that are "movable to" this
3973 : * rel. Also, we can consider any EC-derivable join clauses (which must
3974 : * be "movable to" this rel, by definition).
3975 : */
3976 708 : clauselist = list_copy(rel->baserestrictinfo);
3977 :
3978 : /* Scan the rel's join clauses */
3979 708 : foreach(lc, rel->joininfo)
3980 : {
3981 0 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
3982 :
3983 : /* Check if clause can be moved to this rel */
3984 0 : if (!join_clause_is_movable_to(rinfo, rel))
3985 0 : continue;
3986 :
3987 0 : clauselist = lappend(clauselist, rinfo);
3988 : }
3989 :
3990 : /*
3991 : * Add on any equivalence-derivable join clauses. Computing the correct
3992 : * relid sets for generate_join_implied_equalities is slightly tricky
3993 : * because the rel could be a child rel rather than a true baserel, and in
3994 : * that case we must subtract its parents' relid(s) from all_query_rels.
3995 : * Additionally, we mustn't consider clauses that are only computable
3996 : * after outer joins that can null the rel.
3997 : */
3998 708 : if (rel->reloptkind == RELOPT_OTHER_MEMBER_REL)
3999 72 : otherrels = bms_difference(root->all_query_rels,
4000 72 : find_childrel_parents(root, rel));
4001 : else
4002 636 : otherrels = bms_difference(root->all_query_rels, rel->relids);
4003 708 : otherrels = bms_del_members(otherrels, rel->nulling_relids);
4004 :
4005 708 : if (!bms_is_empty(otherrels))
4006 : clauselist =
4007 88 : list_concat(clauselist,
4008 88 : generate_join_implied_equalities(root,
4009 88 : bms_union(rel->relids,
4010 : otherrels),
4011 : otherrels,
4012 : rel,
4013 : NULL));
4014 :
4015 : /*
4016 : * Normally we remove quals that are implied by a partial index's
4017 : * predicate from indrestrictinfo, indicating that they need not be
4018 : * checked explicitly by an indexscan plan using this index. However, if
4019 : * the rel is a target relation of UPDATE/DELETE/MERGE/SELECT FOR UPDATE,
4020 : * we cannot remove such quals from the plan, because they need to be in
4021 : * the plan so that they will be properly rechecked by EvalPlanQual
4022 : * testing. Some day we might want to remove such quals from the main
4023 : * plan anyway and pass them through to EvalPlanQual via a side channel;
4024 : * but for now, we just don't remove implied quals at all for target
4025 : * relations.
4026 : */
4027 1292 : is_target_rel = (bms_is_member(rel->relid, root->all_result_relids) ||
4028 584 : get_plan_rowmark(root->rowMarks, rel->relid) != NULL);
4029 :
4030 : /*
4031 : * Now try to prove each index predicate true, and compute the
4032 : * indrestrictinfo lists for partial indexes. Note that we compute the
4033 : * indrestrictinfo list even for non-predOK indexes; this might seem
4034 : * wasteful, but we may be able to use such indexes in OR clauses, cf
4035 : * generate_bitmap_or_paths().
4036 : */
4037 2174 : foreach(lc, rel->indexlist)
4038 : {
4039 1466 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
4040 : ListCell *lcr;
4041 :
4042 1466 : if (index->indpred == NIL)
4043 434 : continue; /* ignore non-partial indexes here */
4044 :
4045 1032 : if (!index->predOK) /* don't repeat work if already proven OK */
4046 1032 : index->predOK = predicate_implied_by(index->indpred, clauselist,
4047 : false);
4048 :
4049 : /* If rel is an update target, leave indrestrictinfo as set above */
4050 1032 : if (is_target_rel)
4051 184 : continue;
4052 :
4053 : /*
4054 : * If index is !amoptionalkey, also leave indrestrictinfo as set
4055 : * above. Otherwise we risk removing all quals for the first index
4056 : * key and then not being able to generate an indexscan at all. It
4057 : * would be better to be more selective, but we've not yet identified
4058 : * which if any of the quals match the first index key.
4059 : */
4060 848 : if (!index->amoptionalkey)
4061 36 : continue;
4062 :
4063 : /* Else compute indrestrictinfo as the non-implied quals */
4064 812 : index->indrestrictinfo = NIL;
4065 1914 : foreach(lcr, rel->baserestrictinfo)
4066 : {
4067 1102 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lcr);
4068 :
4069 : /* predicate_implied_by() assumes first arg is immutable */
4070 1102 : if (contain_mutable_functions((Node *) rinfo->clause) ||
4071 1102 : !predicate_implied_by(list_make1(rinfo->clause),
4072 : index->indpred, false))
4073 782 : index->indrestrictinfo = lappend(index->indrestrictinfo, rinfo);
4074 : }
4075 : }
4076 : }
4077 :
4078 : /****************************************************************************
4079 : * ---- ROUTINES TO CHECK EXTERNALLY-VISIBLE CONDITIONS ----
4080 : ****************************************************************************/
4081 :
4082 : /*
4083 : * ec_member_matches_indexcol
4084 : * Test whether an EquivalenceClass member matches an index column.
4085 : *
4086 : * This is a callback for use by generate_implied_equalities_for_column.
4087 : */
4088 : static bool
4089 483566 : ec_member_matches_indexcol(PlannerInfo *root, RelOptInfo *rel,
4090 : EquivalenceClass *ec, EquivalenceMember *em,
4091 : void *arg)
4092 : {
4093 483566 : IndexOptInfo *index = ((ec_member_matches_arg *) arg)->index;
4094 483566 : int indexcol = ((ec_member_matches_arg *) arg)->indexcol;
4095 : Oid curFamily;
4096 : Oid curCollation;
4097 :
4098 : Assert(indexcol < index->nkeycolumns);
4099 :
4100 483566 : curFamily = index->opfamily[indexcol];
4101 483566 : curCollation = index->indexcollations[indexcol];
4102 :
4103 : /*
4104 : * If it's a btree index, we can reject it if its opfamily isn't
4105 : * compatible with the EC, since no clause generated from the EC could be
4106 : * used with the index. For non-btree indexes, we can't easily tell
4107 : * whether clauses generated from the EC could be used with the index, so
4108 : * don't check the opfamily. This might mean we return "true" for a
4109 : * useless EC, so we have to recheck the results of
4110 : * generate_implied_equalities_for_column; see
4111 : * match_eclass_clauses_to_index.
4112 : */
4113 483566 : if (index->relam == BTREE_AM_OID &&
4114 483524 : !list_member_oid(ec->ec_opfamilies, curFamily))
4115 156556 : return false;
4116 :
4117 : /* We insist on collation match for all index types, though */
4118 327010 : if (!IndexCollMatchesExprColl(curCollation, ec->ec_collation))
4119 18 : return false;
4120 :
4121 326992 : return match_index_to_operand((Node *) em->em_expr, indexcol, index);
4122 : }
4123 :
4124 : /*
4125 : * relation_has_unique_index_for
4126 : * Determine whether the relation provably has at most one row satisfying
4127 : * a set of equality conditions, because the conditions constrain all
4128 : * columns of some unique index.
4129 : *
4130 : * The conditions are provided as a list of RestrictInfo nodes, where the
4131 : * caller has already determined that each condition is a mergejoinable
4132 : * equality with an expression in this relation on one side, and an
4133 : * expression not involving this relation on the other. The transient
4134 : * outer_is_left flag is used to identify which side we should look at:
4135 : * left side if outer_is_left is false, right side if it is true.
4136 : *
4137 : * The caller need only supply equality conditions arising from joins;
4138 : * this routine automatically adds in any usable baserestrictinfo clauses.
4139 : * (Note that the passed-in restrictlist will be destructively modified!)
4140 : *
4141 : * If extra_clauses isn't NULL, return baserestrictinfo clauses which were used
4142 : * to derive uniqueness.
4143 : */
4144 : bool
4145 221940 : relation_has_unique_index_for(PlannerInfo *root, RelOptInfo *rel,
4146 : List *restrictlist, List **extra_clauses)
4147 : {
4148 : ListCell *ic;
4149 :
4150 : /* Short-circuit if no indexes... */
4151 221940 : if (rel->indexlist == NIL)
4152 0 : return false;
4153 :
4154 : /*
4155 : * Examine the rel's restriction clauses for usable var = const clauses
4156 : * that we can add to the restrictlist.
4157 : */
4158 366176 : foreach(ic, rel->baserestrictinfo)
4159 : {
4160 144236 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(ic);
4161 :
4162 : /*
4163 : * Note: can_join won't be set for a restriction clause, but
4164 : * mergeopfamilies will be if it has a mergejoinable operator and
4165 : * doesn't contain volatile functions.
4166 : */
4167 144236 : if (restrictinfo->mergeopfamilies == NIL)
4168 58218 : continue; /* not mergejoinable */
4169 :
4170 : /*
4171 : * The clause certainly doesn't refer to anything but the given rel.
4172 : * If either side is pseudoconstant then we can use it.
4173 : */
4174 86018 : if (bms_is_empty(restrictinfo->left_relids))
4175 : {
4176 : /* righthand side is inner */
4177 62 : restrictinfo->outer_is_left = true;
4178 : }
4179 85956 : else if (bms_is_empty(restrictinfo->right_relids))
4180 : {
4181 : /* lefthand side is inner */
4182 85830 : restrictinfo->outer_is_left = false;
4183 : }
4184 : else
4185 126 : continue;
4186 :
4187 : /* OK, add to list */
4188 85892 : restrictlist = lappend(restrictlist, restrictinfo);
4189 : }
4190 :
4191 : /* Short-circuit the easy case */
4192 221940 : if (restrictlist == NIL)
4193 1112 : return false;
4194 :
4195 : /* Examine each index of the relation ... */
4196 563600 : foreach(ic, rel->indexlist)
4197 : {
4198 468770 : IndexOptInfo *ind = (IndexOptInfo *) lfirst(ic);
4199 : int c;
4200 468770 : List *exprs = NIL;
4201 :
4202 : /*
4203 : * If the index is not unique, or not immediately enforced, or if it's
4204 : * a partial index, it's useless here. We're unable to make use of
4205 : * predOK partial unique indexes due to the fact that
4206 : * check_index_predicates() also makes use of join predicates to
4207 : * determine if the partial index is usable. Here we need proofs that
4208 : * hold true before any joins are evaluated.
4209 : */
4210 468770 : if (!ind->unique || !ind->immediate || ind->indpred != NIL)
4211 128794 : continue;
4212 :
4213 : /*
4214 : * Try to find each index column in the list of conditions. This is
4215 : * O(N^2) or worse, but we expect all the lists to be short.
4216 : */
4217 568834 : for (c = 0; c < ind->nkeycolumns; c++)
4218 : {
4219 : ListCell *lc;
4220 :
4221 863080 : foreach(lc, restrictlist)
4222 : {
4223 649102 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4224 : Node *rexpr;
4225 :
4226 : /*
4227 : * The condition's equality operator must be a member of the
4228 : * index opfamily, else it is not asserting the right kind of
4229 : * equality behavior for this index. We check this first
4230 : * since it's probably cheaper than match_index_to_operand().
4231 : */
4232 649102 : if (!list_member_oid(rinfo->mergeopfamilies, ind->opfamily[c]))
4233 198916 : continue;
4234 :
4235 : /*
4236 : * XXX at some point we may need to check collations here too.
4237 : * For the moment we assume all collations reduce to the same
4238 : * notion of equality.
4239 : */
4240 :
4241 : /* OK, see if the condition operand matches the index key */
4242 450186 : if (rinfo->outer_is_left)
4243 180286 : rexpr = get_rightop(rinfo->clause);
4244 : else
4245 269900 : rexpr = get_leftop(rinfo->clause);
4246 :
4247 450186 : if (match_index_to_operand(rexpr, c, ind))
4248 : {
4249 228858 : if (bms_membership(rinfo->clause_relids) == BMS_SINGLETON)
4250 : {
4251 : MemoryContext oldMemCtx =
4252 53578 : MemoryContextSwitchTo(root->planner_cxt);
4253 :
4254 : /*
4255 : * Add filter clause into a list allowing caller to
4256 : * know if uniqueness have made not only by join
4257 : * clauses.
4258 : */
4259 : Assert(bms_is_empty(rinfo->left_relids) ||
4260 : bms_is_empty(rinfo->right_relids));
4261 53578 : if (extra_clauses)
4262 144 : exprs = lappend(exprs, rinfo);
4263 53578 : MemoryContextSwitchTo(oldMemCtx);
4264 : }
4265 :
4266 228858 : break; /* found a match; column is unique */
4267 : }
4268 : }
4269 :
4270 442836 : if (lc == NULL)
4271 213978 : break; /* no match; this index doesn't help us */
4272 : }
4273 :
4274 : /* Matched all key columns of this index? */
4275 339976 : if (c == ind->nkeycolumns)
4276 : {
4277 125998 : if (extra_clauses)
4278 654 : *extra_clauses = exprs;
4279 125998 : return true;
4280 : }
4281 : }
4282 :
4283 94830 : return false;
4284 : }
4285 :
4286 : /*
4287 : * indexcol_is_bool_constant_for_query
4288 : *
4289 : * If an index column is constrained to have a constant value by the query's
4290 : * WHERE conditions, then it's irrelevant for sort-order considerations.
4291 : * Usually that means we have a restriction clause WHERE indexcol = constant,
4292 : * which gets turned into an EquivalenceClass containing a constant, which
4293 : * is recognized as redundant by build_index_pathkeys(). But if the index
4294 : * column is a boolean variable (or expression), then we are not going to
4295 : * see WHERE indexcol = constant, because expression preprocessing will have
4296 : * simplified that to "WHERE indexcol" or "WHERE NOT indexcol". So we are not
4297 : * going to have a matching EquivalenceClass (unless the query also contains
4298 : * "ORDER BY indexcol"). To allow such cases to work the same as they would
4299 : * for non-boolean values, this function is provided to detect whether the
4300 : * specified index column matches a boolean restriction clause.
4301 : */
4302 : bool
4303 669024 : indexcol_is_bool_constant_for_query(PlannerInfo *root,
4304 : IndexOptInfo *index,
4305 : int indexcol)
4306 : {
4307 : ListCell *lc;
4308 :
4309 : /* If the index isn't boolean, we can't possibly get a match */
4310 669024 : if (!IsBooleanOpfamily(index->opfamily[indexcol]))
4311 665992 : return false;
4312 :
4313 : /* Check each restriction clause for the index's rel */
4314 3068 : foreach(lc, index->rel->baserestrictinfo)
4315 : {
4316 1292 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
4317 :
4318 : /*
4319 : * As in match_clause_to_indexcol, never match pseudoconstants to
4320 : * indexes. (It might be semantically okay to do so here, but the
4321 : * odds of getting a match are negligible, so don't waste the cycles.)
4322 : */
4323 1292 : if (rinfo->pseudoconstant)
4324 0 : continue;
4325 :
4326 : /* See if we can match the clause's expression to the index column */
4327 1292 : if (match_boolean_index_clause(root, rinfo, indexcol, index))
4328 1256 : return true;
4329 : }
4330 :
4331 1776 : return false;
4332 : }
4333 :
4334 :
4335 : /****************************************************************************
4336 : * ---- ROUTINES TO CHECK OPERANDS ----
4337 : ****************************************************************************/
4338 :
4339 : /*
4340 : * match_index_to_operand()
4341 : * Generalized test for a match between an index's key
4342 : * and the operand on one side of a restriction or join clause.
4343 : *
4344 : * operand: the nodetree to be compared to the index
4345 : * indexcol: the column number of the index (counting from 0)
4346 : * index: the index of interest
4347 : *
4348 : * Note that we aren't interested in collations here; the caller must check
4349 : * for a collation match, if it's dealing with an operator where that matters.
4350 : *
4351 : * This is exported for use in selfuncs.c.
4352 : */
4353 : bool
4354 3795222 : match_index_to_operand(Node *operand,
4355 : int indexcol,
4356 : IndexOptInfo *index)
4357 : {
4358 : int indkey;
4359 :
4360 : /*
4361 : * Ignore any RelabelType node above the operand. This is needed to be
4362 : * able to apply indexscanning in binary-compatible-operator cases. Note:
4363 : * we can assume there is at most one RelabelType node;
4364 : * eval_const_expressions() will have simplified if more than one.
4365 : */
4366 3795222 : if (operand && IsA(operand, RelabelType))
4367 23726 : operand = (Node *) ((RelabelType *) operand)->arg;
4368 :
4369 3795222 : indkey = index->indexkeys[indexcol];
4370 3795222 : if (indkey != 0)
4371 : {
4372 : /*
4373 : * Simple index column; operand must be a matching Var.
4374 : */
4375 3789128 : if (operand && IsA(operand, Var) &&
4376 2825350 : index->rel->relid == ((Var *) operand)->varno &&
4377 2625858 : indkey == ((Var *) operand)->varattno &&
4378 931692 : ((Var *) operand)->varnullingrels == NULL)
4379 930062 : return true;
4380 : }
4381 : else
4382 : {
4383 : /*
4384 : * Index expression; find the correct expression. (This search could
4385 : * be avoided, at the cost of complicating all the callers of this
4386 : * routine; doesn't seem worth it.)
4387 : */
4388 : ListCell *indexpr_item;
4389 : int i;
4390 : Node *indexkey;
4391 :
4392 6094 : indexpr_item = list_head(index->indexprs);
4393 6094 : for (i = 0; i < indexcol; i++)
4394 : {
4395 0 : if (index->indexkeys[i] == 0)
4396 : {
4397 0 : if (indexpr_item == NULL)
4398 0 : elog(ERROR, "wrong number of index expressions");
4399 0 : indexpr_item = lnext(index->indexprs, indexpr_item);
4400 : }
4401 : }
4402 6094 : if (indexpr_item == NULL)
4403 0 : elog(ERROR, "wrong number of index expressions");
4404 6094 : indexkey = (Node *) lfirst(indexpr_item);
4405 :
4406 : /*
4407 : * Does it match the operand? Again, strip any relabeling.
4408 : */
4409 6094 : if (indexkey && IsA(indexkey, RelabelType))
4410 10 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
4411 :
4412 6094 : if (equal(indexkey, operand))
4413 2176 : return true;
4414 : }
4415 :
4416 2862984 : return false;
4417 : }
4418 :
4419 : /*
4420 : * is_pseudo_constant_for_index()
4421 : * Test whether the given expression can be used as an indexscan
4422 : * comparison value.
4423 : *
4424 : * An indexscan comparison value must not contain any volatile functions,
4425 : * and it can't contain any Vars of the index's own table. Vars of
4426 : * other tables are okay, though; in that case we'd be producing an
4427 : * indexqual usable in a parameterized indexscan. This is, therefore,
4428 : * a weaker condition than is_pseudo_constant_clause().
4429 : *
4430 : * This function is exported for use by planner support functions,
4431 : * which will have available the IndexOptInfo, but not any RestrictInfo
4432 : * infrastructure. It is making the same test made by functions above
4433 : * such as match_opclause_to_indexcol(), but those rely where possible
4434 : * on RestrictInfo information about variable membership.
4435 : *
4436 : * expr: the nodetree to be checked
4437 : * index: the index of interest
4438 : */
4439 : bool
4440 0 : is_pseudo_constant_for_index(PlannerInfo *root, Node *expr, IndexOptInfo *index)
4441 : {
4442 : /* pull_varnos is cheaper than volatility check, so do that first */
4443 0 : if (bms_is_member(index->rel->relid, pull_varnos(root, expr)))
4444 0 : return false; /* no good, contains Var of table */
4445 0 : if (contain_volatile_functions(expr))
4446 0 : return false; /* no good, volatile comparison value */
4447 0 : return true;
4448 : }
|
