?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Disallow empty passwords in LDAP authentication, the same way
[PostgreSQL.git] / src / backend / optimizer / path / pathkeys.c
blob93491589ec900391f8a1208630e17e8f8dfbee04
1 /*-------------------------------------------------------------------------
2 *
3 * pathkeys.c
4 * Utilities for matching and building path keys
5 *
6 * See src/backend/optimizer/README for a great deal of information about
7 * the nature and use of path keys.
8 *
9 *
10 * Portions Copyright (c) 1996-2009, PostgreSQL Global Development Group
11 * Portions Copyright (c) 1994, Regents of the University of California
12 *
13 * IDENTIFICATION
14 * $PostgreSQL$
15 *
16 *-------------------------------------------------------------------------
17 */
18 #include "postgres.h"
19
20 #include "access/skey.h"
21 #include "catalog/pg_type.h"
22 #include "nodes/makefuncs.h"
23 #include "nodes/nodeFuncs.h"
24 #include "nodes/plannodes.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/pathnode.h"
27 #include "optimizer/paths.h"
28 #include "optimizer/tlist.h"
29 #include "parser/parsetree.h"
30 #include "utils/lsyscache.h"
31
32
33 static PathKey *makePathKey(EquivalenceClass *eclass, Oid opfamily,
34 int strategy, bool nulls_first);
35 static PathKey *make_canonical_pathkey(PlannerInfo *root,
36 EquivalenceClass *eclass, Oid opfamily,
37 int strategy, bool nulls_first);
38 static bool pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys);
39 static PathKey *make_pathkey_from_sortinfo(PlannerInfo *root,
40 Expr *expr, Oid ordering_op,
41 bool nulls_first,
42 Index sortref,
43 bool canonicalize);
44 static Var *find_indexkey_var(PlannerInfo *root, RelOptInfo *rel,
45 AttrNumber varattno);
46 static bool right_merge_direction(PlannerInfo *root, PathKey *pathkey);
47
48
49 /****************************************************************************
50 * PATHKEY CONSTRUCTION AND REDUNDANCY TESTING
51 ****************************************************************************/
52
53 /*
54 * makePathKey
55 * create a PathKey node
56 *
57 * This does not promise to create a canonical PathKey, it's merely a
58 * convenience routine to build the specified node.
59 */
60 static PathKey *
61 makePathKey(EquivalenceClass *eclass, Oid opfamily,
62 int strategy, bool nulls_first)
63 {
64 PathKey *pk = makeNode(PathKey);
65
66 pk->pk_eclass = eclass;
67 pk->pk_opfamily = opfamily;
68 pk->pk_strategy = strategy;
69 pk->pk_nulls_first = nulls_first;
70
71 return pk;
72 }
73
74 /*
75 * make_canonical_pathkey
76 * Given the parameters for a PathKey, find any pre-existing matching
77 * pathkey in the query's list of "canonical" pathkeys. Make a new
78 * entry if there's not one already.
79 *
80 * Note that this function must not be used until after we have completed
81 * merging EquivalenceClasses.
82 */
83 static PathKey *
84 make_canonical_pathkey(PlannerInfo *root,
85 EquivalenceClass *eclass, Oid opfamily,
86 int strategy, bool nulls_first)
87 {
88 PathKey *pk;
89 ListCell *lc;
90 MemoryContext oldcontext;
91
92 /* The passed eclass might be non-canonical, so chase up to the top */
93 while (eclass->ec_merged)
94 eclass = eclass->ec_merged;
95
96 foreach(lc, root->canon_pathkeys)
97 {
98 pk = (PathKey *) lfirst(lc);
99 if (eclass == pk->pk_eclass &&
100 opfamily == pk->pk_opfamily &&
101 strategy == pk->pk_strategy &&
102 nulls_first == pk->pk_nulls_first)
103 return pk;
104 }
105
106 /*
107 * Be sure canonical pathkeys are allocated in the main planning context.
108 * Not an issue in normal planning, but it is for GEQO.
109 */
110 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
111
112 pk = makePathKey(eclass, opfamily, strategy, nulls_first);
113 root->canon_pathkeys = lappend(root->canon_pathkeys, pk);
114
115 MemoryContextSwitchTo(oldcontext);
116
117 return pk;
118 }
119
120 /*
121 * pathkey_is_redundant
122 * Is a pathkey redundant with one already in the given list?
123 *
124 * Both the given pathkey and the list members must be canonical for this
125 * to work properly. We detect two cases:
126 *
127 * 1. If the new pathkey's equivalence class contains a constant, and isn't
128 * below an outer join, then we can disregard it as a sort key. An example:
129 * SELECT ... WHERE x = 42 ORDER BY x, y;
130 * We may as well just sort by y. Note that because of opfamily matching,
131 * this is semantically correct: we know that the equality constraint is one
132 * that actually binds the variable to a single value in the terms of any
133 * ordering operator that might go with the eclass. This rule not only lets
134 * us simplify (or even skip) explicit sorts, but also allows matching index
135 * sort orders to a query when there are don't-care index columns.
136 *
137 * 2. If the new pathkey's equivalence class is the same as that of any
138 * existing member of the pathkey list, then it is redundant. Some examples:
139 * SELECT ... ORDER BY x, x;
140 * SELECT ... ORDER BY x, x DESC;
141 * SELECT ... WHERE x = y ORDER BY x, y;
142 * In all these cases the second sort key cannot distinguish values that are
143 * considered equal by the first, and so there's no point in using it.
144 * Note in particular that we need not compare opfamily (all the opfamilies
145 * of the EC have the same notion of equality) nor sort direction.
146 *
147 * Because the equivclass.c machinery forms only one copy of any EC per query,
148 * pointer comparison is enough to decide whether canonical ECs are the same.
149 */
150 static bool
151 pathkey_is_redundant(PathKey *new_pathkey, List *pathkeys)
152 {
153 EquivalenceClass *new_ec = new_pathkey->pk_eclass;
154 ListCell *lc;
155
156 /* Assert we've been given canonical pathkeys */
157 Assert(!new_ec->ec_merged);
158
159 /* Check for EC containing a constant --- unconditionally redundant */
160 if (EC_MUST_BE_REDUNDANT(new_ec))
161 return true;
162
163 /* If same EC already used in list, then redundant */
164 foreach(lc, pathkeys)
165 {
166 PathKey *old_pathkey = (PathKey *) lfirst(lc);
167
168 /* Assert we've been given canonical pathkeys */
169 Assert(!old_pathkey->pk_eclass->ec_merged);
170
171 if (new_ec == old_pathkey->pk_eclass)
172 return true;
173 }
174
175 return false;
176 }
177
178 /*
179 * canonicalize_pathkeys
180 * Convert a not-necessarily-canonical pathkeys list to canonical form.
181 *
182 * Note that this function must not be used until after we have completed
183 * merging EquivalenceClasses.
184 */
185 List *
186 canonicalize_pathkeys(PlannerInfo *root, List *pathkeys)
187 {
188 List *new_pathkeys = NIL;
189 ListCell *l;
190
191 foreach(l, pathkeys)
192 {
193 PathKey *pathkey = (PathKey *) lfirst(l);
194 EquivalenceClass *eclass;
195 PathKey *cpathkey;
196
197 /* Find the canonical (merged) EquivalenceClass */
198 eclass = pathkey->pk_eclass;
199 while (eclass->ec_merged)
200 eclass = eclass->ec_merged;
201
202 /*
203 * If we can tell it's redundant just from the EC, skip.
204 * pathkey_is_redundant would notice that, but we needn't even bother
205 * constructing the node...
206 */
207 if (EC_MUST_BE_REDUNDANT(eclass))
208 continue;
209
210 /* OK, build a canonicalized PathKey struct */
211 cpathkey = make_canonical_pathkey(root,
212 eclass,
213 pathkey->pk_opfamily,
214 pathkey->pk_strategy,
215 pathkey->pk_nulls_first);
216
217 /* Add to list unless redundant */
218 if (!pathkey_is_redundant(cpathkey, new_pathkeys))
219 new_pathkeys = lappend(new_pathkeys, cpathkey);
220 }
221 return new_pathkeys;
222 }
223
224 /*
225 * make_pathkey_from_sortinfo
226 * Given an expression, a sortop, and a nulls-first flag, create
227 * a PathKey. If canonicalize = true, the result is a "canonical"
228 * PathKey, otherwise not. (But note it might be redundant anyway.)
229 *
230 * If the PathKey is being generated from a SortGroupClause, sortref should be
231 * the SortGroupClause's SortGroupRef; otherwise zero.
232 *
233 * canonicalize should always be TRUE after EquivalenceClass merging has
234 * been performed, but FALSE if we haven't done EquivalenceClass merging yet.
235 */
236 static PathKey *
237 make_pathkey_from_sortinfo(PlannerInfo *root,
238 Expr *expr, Oid ordering_op,
239 bool nulls_first,
240 Index sortref,
241 bool canonicalize)
242 {
243 Oid opfamily,
244 opcintype;
245 int16 strategy;
246 Oid equality_op;
247 List *opfamilies;
248 EquivalenceClass *eclass;
249
250 /*
251 * An ordering operator fully determines the behavior of its opfamily, so
252 * could only meaningfully appear in one family --- or perhaps two if one
253 * builds a reverse-sort opfamily, but there's not much point in that
254 * anymore. But EquivalenceClasses need to contain opfamily lists based
255 * on the family membership of equality operators, which could easily be
256 * bigger. So, look up the equality operator that goes with the ordering
257 * operator (this should be unique) and get its membership.
258 */
259
260 /* Find the operator in pg_amop --- failure shouldn't happen */
261 if (!get_ordering_op_properties(ordering_op,
262 &opfamily, &opcintype, &strategy))
263 elog(ERROR, "operator %u is not a valid ordering operator",
264 ordering_op);
265 /* Get matching equality operator */
266 equality_op = get_opfamily_member(opfamily,
267 opcintype,
268 opcintype,
269 BTEqualStrategyNumber);
270 if (!OidIsValid(equality_op)) /* shouldn't happen */
271 elog(ERROR, "could not find equality operator for ordering operator %u",
272 ordering_op);
273 opfamilies = get_mergejoin_opfamilies(equality_op);
274 if (!opfamilies) /* certainly should find some */
275 elog(ERROR, "could not find opfamilies for ordering operator %u",
276 ordering_op);
277
278 /*
279 * When dealing with binary-compatible opclasses, we have to ensure that
280 * the exposed type of the expression tree matches the declared input type
281 * of the opclass, except when that is a polymorphic type (compare the
282 * behavior of parse_coerce.c). This ensures that we can correctly match
283 * the indexkey or sortclause expression to other expressions we find in
284 * the query, because arguments of ordinary operator expressions will be
285 * cast that way. (We have to do this for indexkeys because they are
286 * represented without any explicit relabel in pg_index, and for sort
287 * clauses because the parser is likewise cavalier about putting relabels
288 * on them.)
289 */
290 if (exprType((Node *) expr) != opcintype &&
291 !IsPolymorphicType(opcintype))
292 {
293 /* Strip any existing RelabelType, and add a new one if needed */
294 while (expr && IsA(expr, RelabelType))
295 expr = (Expr *) ((RelabelType *) expr)->arg;
296 if (exprType((Node *) expr) != opcintype)
297 expr = (Expr *) makeRelabelType(expr,
298 opcintype,
299 -1,
300 COERCE_DONTCARE);
301 }
302
303 /* Now find or create a matching EquivalenceClass */
304 eclass = get_eclass_for_sort_expr(root, expr, opcintype, opfamilies,
305 sortref);
306
307 /* And finally we can find or create a PathKey node */
308 if (canonicalize)
309 return make_canonical_pathkey(root, eclass, opfamily,
310 strategy, nulls_first);
311 else
312 return makePathKey(eclass, opfamily, strategy, nulls_first);
313 }
314
315
316 /****************************************************************************
317 * PATHKEY COMPARISONS
318 ****************************************************************************/
319
320 /*
321 * compare_pathkeys
322 * Compare two pathkeys to see if they are equivalent, and if not whether
323 * one is "better" than the other.
324 *
325 * This function may only be applied to canonicalized pathkey lists.
326 * In the canonical representation, pathkeys can be checked for equality
327 * by simple pointer comparison.
328 */
329 PathKeysComparison
330 compare_pathkeys(List *keys1, List *keys2)
331 {
332 ListCell *key1,
333 *key2;
334
335 /*
336 * Fall out quickly if we are passed two identical lists. This mostly
337 * catches the case where both are NIL, but that's common enough to
338 * warrant the test.
339 */
340 if (keys1 == keys2)
341 return PATHKEYS_EQUAL;
342
343 forboth(key1, keys1, key2, keys2)
344 {
345 PathKey *pathkey1 = (PathKey *) lfirst(key1);
346 PathKey *pathkey2 = (PathKey *) lfirst(key2);
347
348 /*
349 * XXX would like to check that we've been given canonicalized input,
350 * but PlannerInfo not accessible here...
351 */
352 #ifdef NOT_USED
353 Assert(list_member_ptr(root->canon_pathkeys, pathkey1));
354 Assert(list_member_ptr(root->canon_pathkeys, pathkey2));
355 #endif
356
357 if (pathkey1 != pathkey2)
358 return PATHKEYS_DIFFERENT; /* no need to keep looking */
359 }
360
361 /*
362 * If we reached the end of only one list, the other is longer and
363 * therefore not a subset.
364 */
365 if (key1 != NULL)
366 return PATHKEYS_BETTER1; /* key1 is longer */
367 if (key2 != NULL)
368 return PATHKEYS_BETTER2; /* key2 is longer */
369 return PATHKEYS_EQUAL;
370 }
371
372 /*
373 * pathkeys_contained_in
374 * Common special case of compare_pathkeys: we just want to know
375 * if keys2 are at least as well sorted as keys1.
376 */
377 bool
378 pathkeys_contained_in(List *keys1, List *keys2)
379 {
380 switch (compare_pathkeys(keys1, keys2))
381 {
382 case PATHKEYS_EQUAL:
383 case PATHKEYS_BETTER2:
384 return true;
385 default:
386 break;
387 }
388 return false;
389 }
390
391 /*
392 * get_cheapest_path_for_pathkeys
393 * Find the cheapest path (according to the specified criterion) that
394 * satisfies the given pathkeys. Return NULL if no such path.
395 *
396 * 'paths' is a list of possible paths that all generate the same relation
397 * 'pathkeys' represents a required ordering (already canonicalized!)
398 * 'cost_criterion' is STARTUP_COST or TOTAL_COST
399 */
400 Path *
401 get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
402 CostSelector cost_criterion)
403 {
404 Path *matched_path = NULL;
405 ListCell *l;
406
407 foreach(l, paths)
408 {
409 Path *path = (Path *) lfirst(l);
410
411 /*
412 * Since cost comparison is a lot cheaper than pathkey comparison, do
413 * that first. (XXX is that still true?)
414 */
415 if (matched_path != NULL &&
416 compare_path_costs(matched_path, path, cost_criterion) <= 0)
417 continue;
418
419 if (pathkeys_contained_in(pathkeys, path->pathkeys))
420 matched_path = path;
421 }
422 return matched_path;
423 }
424
425 /*
426 * get_cheapest_fractional_path_for_pathkeys
427 * Find the cheapest path (for retrieving a specified fraction of all
428 * the tuples) that satisfies the given pathkeys.
429 * Return NULL if no such path.
430 *
431 * See compare_fractional_path_costs() for the interpretation of the fraction
432 * parameter.
433 *
434 * 'paths' is a list of possible paths that all generate the same relation
435 * 'pathkeys' represents a required ordering (already canonicalized!)
436 * 'fraction' is the fraction of the total tuples expected to be retrieved
437 */
438 Path *
439 get_cheapest_fractional_path_for_pathkeys(List *paths,
440 List *pathkeys,
441 double fraction)
442 {
443 Path *matched_path = NULL;
444 ListCell *l;
445
446 foreach(l, paths)
447 {
448 Path *path = (Path *) lfirst(l);
449
450 /*
451 * Since cost comparison is a lot cheaper than pathkey comparison, do
452 * that first.
453 */
454 if (matched_path != NULL &&
455 compare_fractional_path_costs(matched_path, path, fraction) <= 0)
456 continue;
457
458 if (pathkeys_contained_in(pathkeys, path->pathkeys))
459 matched_path = path;
460 }
461 return matched_path;
462 }
463
464 /****************************************************************************
465 * NEW PATHKEY FORMATION
466 ****************************************************************************/
467
468 /*
469 * build_index_pathkeys
470 * Build a pathkeys list that describes the ordering induced by an index
471 * scan using the given index. (Note that an unordered index doesn't
472 * induce any ordering; such an index will have no sortop OIDS in
473 * its sortops arrays, and we will return NIL.)
474 *
475 * If 'scandir' is BackwardScanDirection, attempt to build pathkeys
476 * representing a backwards scan of the index. Return NIL if can't do it.
477 *
478 * The result is canonical, meaning that redundant pathkeys are removed;
479 * it may therefore have fewer entries than there are index columns.
480 *
481 * We generate the full pathkeys list whether or not all are useful for the
482 * current query. Caller should do truncate_useless_pathkeys().
483 */
484 List *
485 build_index_pathkeys(PlannerInfo *root,
486 IndexOptInfo *index,
487 ScanDirection scandir)
488 {
489 List *retval = NIL;
490 ListCell *indexprs_item = list_head(index->indexprs);
491 int i;
492
493 for (i = 0; i < index->ncolumns; i++)
494 {
495 Oid sortop;
496 bool nulls_first;
497 int ikey;
498 Expr *indexkey;
499 PathKey *cpathkey;
500
501 if (ScanDirectionIsBackward(scandir))
502 {
503 sortop = index->revsortop[i];
504 nulls_first = !index->nulls_first[i];
505 }
506 else
507 {
508 sortop = index->fwdsortop[i];
509 nulls_first = index->nulls_first[i];
510 }
511
512 if (!OidIsValid(sortop))
513 break; /* no more orderable columns */
514
515 ikey = index->indexkeys[i];
516 if (ikey != 0)
517 {
518 /* simple index column */
519 indexkey = (Expr *) find_indexkey_var(root, index->rel, ikey);
520 }
521 else
522 {
523 /* expression --- assume we need not copy it */
524 if (indexprs_item == NULL)
525 elog(ERROR, "wrong number of index expressions");
526 indexkey = (Expr *) lfirst(indexprs_item);
527 indexprs_item = lnext(indexprs_item);
528 }
529
530 /* OK, make a canonical pathkey for this sort key */
531 cpathkey = make_pathkey_from_sortinfo(root,
532 indexkey,
533 sortop,
534 nulls_first,
535 0,
536 true);
537
538 /* Add to list unless redundant */
539 if (!pathkey_is_redundant(cpathkey, retval))
540 retval = lappend(retval, cpathkey);
541 }
542
543 return retval;
544 }
545
546 /*
547 * Find or make a Var node for the specified attribute of the rel.
548 *
549 * We first look for the var in the rel's target list, because that's
550 * easy and fast. But the var might not be there (this should normally
551 * only happen for vars that are used in WHERE restriction clauses,
552 * but not in join clauses or in the SELECT target list). In that case,
553 * gin up a Var node the hard way.
554 */
555 static Var *
556 find_indexkey_var(PlannerInfo *root, RelOptInfo *rel, AttrNumber varattno)
557 {
558 ListCell *temp;
559 Index relid;
560 Oid reloid,
561 vartypeid;
562 int32 type_mod;
563
564 foreach(temp, rel->reltargetlist)
565 {
566 Var *var = (Var *) lfirst(temp);
567
568 if (IsA(var, Var) &&
569 var->varattno == varattno)
570 return var;
571 }
572
573 relid = rel->relid;
574 reloid = getrelid(relid, root->parse->rtable);
575 get_atttypetypmod(reloid, varattno, &vartypeid, &type_mod);
576
577 return makeVar(relid, varattno, vartypeid, type_mod, 0);
578 }
579
580 /*
581 * convert_subquery_pathkeys
582 * Build a pathkeys list that describes the ordering of a subquery's
583 * result, in the terms of the outer query. This is essentially a
584 * task of conversion.
585 *
586 * 'rel': outer query's RelOptInfo for the subquery relation.
587 * 'subquery_pathkeys': the subquery's output pathkeys, in its terms.
588 *
589 * It is not necessary for caller to do truncate_useless_pathkeys(),
590 * because we select keys in a way that takes usefulness of the keys into
591 * account.
592 */
593 List *
594 convert_subquery_pathkeys(PlannerInfo *root, RelOptInfo *rel,
595 List *subquery_pathkeys)
596 {
597 List *retval = NIL;
598 int retvallen = 0;
599 int outer_query_keys = list_length(root->query_pathkeys);
600 List *sub_tlist = rel->subplan->targetlist;
601 ListCell *i;
602
603 foreach(i, subquery_pathkeys)
604 {
605 PathKey *sub_pathkey = (PathKey *) lfirst(i);
606 EquivalenceClass *sub_eclass = sub_pathkey->pk_eclass;
607 PathKey *best_pathkey = NULL;
608
609 if (sub_eclass->ec_has_volatile)
610 {
611 /*
612 * If the sub_pathkey's EquivalenceClass is volatile, then it must
613 * have come from an ORDER BY clause, and we have to match it to
614 * that same targetlist entry.
615 */
616 TargetEntry *tle;
617
618 if (sub_eclass->ec_sortref == 0) /* can't happen */
619 elog(ERROR, "volatile EquivalenceClass has no sortref");
620 tle = get_sortgroupref_tle(sub_eclass->ec_sortref, sub_tlist);
621 Assert(tle);
622 /* resjunk items aren't visible to outer query */
623 if (!tle->resjunk)
624 {
625 /* We can represent this sub_pathkey */
626 EquivalenceMember *sub_member;
627 Expr *outer_expr;
628 EquivalenceClass *outer_ec;
629
630 Assert(list_length(sub_eclass->ec_members) == 1);
631 sub_member = (EquivalenceMember *) linitial(sub_eclass->ec_members);
632 outer_expr = (Expr *)
633 makeVar(rel->relid,
634 tle->resno,
635 exprType((Node *) tle->expr),
636 exprTypmod((Node *) tle->expr),
637 0);
638 outer_ec =
639 get_eclass_for_sort_expr(root,
640 outer_expr,
641 sub_member->em_datatype,
642 sub_eclass->ec_opfamilies,
643 0);
644 best_pathkey =
645 make_canonical_pathkey(root,
646 outer_ec,
647 sub_pathkey->pk_opfamily,
648 sub_pathkey->pk_strategy,
649 sub_pathkey->pk_nulls_first);
650 }
651 }
652 else
653 {
654 /*
655 * Otherwise, the sub_pathkey's EquivalenceClass could contain
656 * multiple elements (representing knowledge that multiple items
657 * are effectively equal). Each element might match none, one, or
658 * more of the output columns that are visible to the outer query.
659 * This means we may have multiple possible representations of the
660 * sub_pathkey in the context of the outer query. Ideally we
661 * would generate them all and put them all into an EC of the
662 * outer query, thereby propagating equality knowledge up to the
663 * outer query. Right now we cannot do so, because the outer
664 * query's EquivalenceClasses are already frozen when this is
665 * called. Instead we prefer the one that has the highest "score"
666 * (number of EC peers, plus one if it matches the outer
667 * query_pathkeys). This is the most likely to be useful in the
668 * outer query.
669 */
670 int best_score = -1;
671 ListCell *j;
672
673 foreach(j, sub_eclass->ec_members)
674 {
675 EquivalenceMember *sub_member = (EquivalenceMember *) lfirst(j);
676 Expr *sub_expr = sub_member->em_expr;
677 Expr *sub_stripped;
678 ListCell *k;
679
680 /*
681 * We handle two cases: the sub_pathkey key can be either an
682 * exact match for a targetlist entry, or it could match after
683 * stripping RelabelType nodes. (We need that case since
684 * make_pathkey_from_sortinfo could add or remove
685 * RelabelType.)
686 */
687 sub_stripped = sub_expr;
688 while (sub_stripped && IsA(sub_stripped, RelabelType))
689 sub_stripped = ((RelabelType *) sub_stripped)->arg;
690
691 foreach(k, sub_tlist)
692 {
693 TargetEntry *tle = (TargetEntry *) lfirst(k);
694 Expr *outer_expr;
695 EquivalenceClass *outer_ec;
696 PathKey *outer_pk;
697 int score;
698
699 /* resjunk items aren't visible to outer query */
700 if (tle->resjunk)
701 continue;
702
703 if (equal(tle->expr, sub_expr))
704 {
705 /* Exact match */
706 outer_expr = (Expr *)
707 makeVar(rel->relid,
708 tle->resno,
709 exprType((Node *) tle->expr),
710 exprTypmod((Node *) tle->expr),
711 0);
712 }
713 else
714 {
715 Expr *tle_stripped;
716
717 tle_stripped = tle->expr;
718 while (tle_stripped && IsA(tle_stripped, RelabelType))
719 tle_stripped = ((RelabelType *) tle_stripped)->arg;
720
721 if (equal(tle_stripped, sub_stripped))
722 {
723 /* Match after discarding RelabelType */
724 outer_expr = (Expr *)
725 makeVar(rel->relid,
726 tle->resno,
727 exprType((Node *) tle->expr),
728 exprTypmod((Node *) tle->expr),
729 0);
730 if (exprType((Node *) outer_expr) !=
731 exprType((Node *) sub_expr))
732 outer_expr = (Expr *)
733 makeRelabelType(outer_expr,
734 exprType((Node *) sub_expr),
735 -1,
736 COERCE_DONTCARE);
737 }
738 else
739 continue;
740 }
741
742 /* Found a representation for this sub_pathkey */
743 outer_ec = get_eclass_for_sort_expr(root,
744 outer_expr,
745 sub_member->em_datatype,
746 sub_eclass->ec_opfamilies,
747 0);
748 outer_pk = make_canonical_pathkey(root,
749 outer_ec,
750 sub_pathkey->pk_opfamily,
751 sub_pathkey->pk_strategy,
752 sub_pathkey->pk_nulls_first);
753 /* score = # of equivalence peers */
754 score = list_length(outer_ec->ec_members) - 1;
755 /* +1 if it matches the proper query_pathkeys item */
756 if (retvallen < outer_query_keys &&
757 list_nth(root->query_pathkeys, retvallen) == outer_pk)
758 score++;
759 if (score > best_score)
760 {
761 best_pathkey = outer_pk;
762 best_score = score;
763 }
764 }
765 }
766 }
767
768 /*
769 * If we couldn't find a representation of this sub_pathkey, we're
770 * done (we can't use the ones to its right, either).
771 */
772 if (!best_pathkey)
773 break;
774
775 /*
776 * Eliminate redundant ordering info; could happen if outer query
777 * equivalences subquery keys...
778 */
779 if (!pathkey_is_redundant(best_pathkey, retval))
780 {
781 retval = lappend(retval, best_pathkey);
782 retvallen++;
783 }
784 }
785
786 return retval;
787 }
788
789 /*
790 * build_join_pathkeys
791 * Build the path keys for a join relation constructed by mergejoin or
792 * nestloop join. This is normally the same as the outer path's keys.
793 *
794 * EXCEPTION: in a FULL or RIGHT join, we cannot treat the result as
795 * having the outer path's path keys, because null lefthand rows may be
796 * inserted at random points. It must be treated as unsorted.
797 *
798 * We truncate away any pathkeys that are uninteresting for higher joins.
799 *
800 * 'joinrel' is the join relation that paths are being formed for
801 * 'jointype' is the join type (inner, left, full, etc)
802 * 'outer_pathkeys' is the list of the current outer path's path keys
803 *
804 * Returns the list of new path keys.
805 */
806 List *
807 build_join_pathkeys(PlannerInfo *root,
808 RelOptInfo *joinrel,
809 JoinType jointype,
810 List *outer_pathkeys)
811 {
812 if (jointype == JOIN_FULL || jointype == JOIN_RIGHT)
813 return NIL;
814
815 /*
816 * This used to be quite a complex bit of code, but now that all pathkey
817 * sublists start out life canonicalized, we don't have to do a darn thing
818 * here!
819 *
820 * We do, however, need to truncate the pathkeys list, since it may
821 * contain pathkeys that were useful for forming this joinrel but are
822 * uninteresting to higher levels.
823 */
824 return truncate_useless_pathkeys(root, joinrel, outer_pathkeys);
825 }
826
827 /****************************************************************************
828 * PATHKEYS AND SORT CLAUSES
829 ****************************************************************************/
830
831 /*
832 * make_pathkeys_for_sortclauses
833 * Generate a pathkeys list that represents the sort order specified
834 * by a list of SortGroupClauses
835 *
836 * If canonicalize is TRUE, the resulting PathKeys are all in canonical form;
837 * otherwise not. canonicalize should always be TRUE after EquivalenceClass
838 * merging has been performed, but FALSE if we haven't done EquivalenceClass
839 * merging yet. (We provide this option because grouping_planner() needs to
840 * be able to represent requested pathkeys before the equivalence classes have
841 * been created for the query.)
842 *
843 * 'sortclauses' is a list of SortGroupClause nodes
844 * 'tlist' is the targetlist to find the referenced tlist entries in
845 */
846 List *
847 make_pathkeys_for_sortclauses(PlannerInfo *root,
848 List *sortclauses,
849 List *tlist,
850 bool canonicalize)
851 {
852 List *pathkeys = NIL;
853 ListCell *l;
854
855 foreach(l, sortclauses)
856 {
857 SortGroupClause *sortcl = (SortGroupClause *) lfirst(l);
858 Expr *sortkey;
859 PathKey *pathkey;
860
861 sortkey = (Expr *) get_sortgroupclause_expr(sortcl, tlist);
862 Assert(OidIsValid(sortcl->sortop));
863 pathkey = make_pathkey_from_sortinfo(root,
864 sortkey,
865 sortcl->sortop,
866 sortcl->nulls_first,
867 sortcl->tleSortGroupRef,
868 canonicalize);
869
870 /* Canonical form eliminates redundant ordering keys */
871 if (canonicalize)
872 {
873 if (!pathkey_is_redundant(pathkey, pathkeys))
874 pathkeys = lappend(pathkeys, pathkey);
875 }
876 else
877 pathkeys = lappend(pathkeys, pathkey);
878 }
879 return pathkeys;
880 }
881
882 /****************************************************************************
883 * PATHKEYS AND MERGECLAUSES
884 ****************************************************************************/
885
886 /*
887 * cache_mergeclause_eclasses
888 * Make the cached EquivalenceClass links valid in a mergeclause
889 * restrictinfo.
890 *
891 * RestrictInfo contains fields in which we may cache pointers to
892 * EquivalenceClasses for the left and right inputs of the mergeclause.
893 * (If the mergeclause is a true equivalence clause these will be the
894 * same EquivalenceClass, otherwise not.)
895 */
896 void
897 cache_mergeclause_eclasses(PlannerInfo *root, RestrictInfo *restrictinfo)
898 {
899 Assert(restrictinfo->mergeopfamilies != NIL);
900
901 /* the cached values should be either both set or both not */
902 if (restrictinfo->left_ec == NULL)
903 {
904 Expr *clause = restrictinfo->clause;
905 Oid lefttype,
906 righttype;
907
908 /* Need the declared input types of the operator */
909 op_input_types(((OpExpr *) clause)->opno, &lefttype, &righttype);
910
911 /* Find or create a matching EquivalenceClass for each side */
912 restrictinfo->left_ec =
913 get_eclass_for_sort_expr(root,
914 (Expr *) get_leftop(clause),
915 lefttype,
916 restrictinfo->mergeopfamilies,
917 0);
918 restrictinfo->right_ec =
919 get_eclass_for_sort_expr(root,
920 (Expr *) get_rightop(clause),
921 righttype,
922 restrictinfo->mergeopfamilies,
923 0);
924 }
925 else
926 Assert(restrictinfo->right_ec != NULL);
927 }
928
929 /*
930 * find_mergeclauses_for_pathkeys
931 * This routine attempts to find a set of mergeclauses that can be
932 * used with a specified ordering for one of the input relations.
933 * If successful, it returns a list of mergeclauses.
934 *
935 * 'pathkeys' is a pathkeys list showing the ordering of an input path.
936 * 'outer_keys' is TRUE if these keys are for the outer input path,
937 * FALSE if for inner.
938 * 'restrictinfos' is a list of mergejoinable restriction clauses for the
939 * join relation being formed.
940 *
941 * The restrictinfos must be marked (via outer_is_left) to show which side
942 * of each clause is associated with the current outer path. (See
943 * select_mergejoin_clauses())
944 *
945 * The result is NIL if no merge can be done, else a maximal list of
946 * usable mergeclauses (represented as a list of their restrictinfo nodes).
947 */
948 List *
949 find_mergeclauses_for_pathkeys(PlannerInfo *root,
950 List *pathkeys,
951 bool outer_keys,
952 List *restrictinfos)
953 {
954 List *mergeclauses = NIL;
955 ListCell *i;
956
957 /* make sure we have eclasses cached in the clauses */
958 foreach(i, restrictinfos)
959 {
960 RestrictInfo *rinfo = (RestrictInfo *) lfirst(i);
961
962 cache_mergeclause_eclasses(root, rinfo);
963 }
964
965 foreach(i, pathkeys)
966 {
967 PathKey *pathkey = (PathKey *) lfirst(i);
968 EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
969 List *matched_restrictinfos = NIL;
970 ListCell *j;
971
972 /*----------
973 * A mergejoin clause matches a pathkey if it has the same EC.
974 * If there are multiple matching clauses, take them all. In plain
975 * inner-join scenarios we expect only one match, because
976 * equivalence-class processing will have removed any redundant
977 * mergeclauses. However, in outer-join scenarios there might be
978 * multiple matches. An example is
979 *
980 * select * from a full join b
981 * on a.v1 = b.v1 and a.v2 = b.v2 and a.v1 = b.v2;
982 *
983 * Given the pathkeys ({a.v1}, {a.v2}) it is okay to return all three
984 * clauses (in the order a.v1=b.v1, a.v1=b.v2, a.v2=b.v2) and indeed
985 * we *must* do so or we will be unable to form a valid plan.
986 *
987 * We expect that the given pathkeys list is canonical, which means
988 * no two members have the same EC, so it's not possible for this
989 * code to enter the same mergeclause into the result list twice.
990 *
991 * XXX it's possible that multiple matching clauses might have
992 * different ECs on the other side, in which case the order we put
993 * them into our result makes a difference in the pathkeys required
994 * for the other input path. However this routine hasn't got any info
995 * about which order would be best, so for now we disregard that case
996 * (which is probably a corner case anyway).
997 *----------
998 */
999 foreach(j, restrictinfos)
1000 {
1001 RestrictInfo *rinfo = (RestrictInfo *) lfirst(j);
1002 EquivalenceClass *clause_ec;
1003
1004 if (outer_keys)
1005 clause_ec = rinfo->outer_is_left ?
1006 rinfo->left_ec : rinfo->right_ec;
1007 else
1008 clause_ec = rinfo->outer_is_left ?
1009 rinfo->right_ec : rinfo->left_ec;
1010 if (clause_ec == pathkey_ec)
1011 matched_restrictinfos = lappend(matched_restrictinfos, rinfo);
1012 }
1013
1014 /*
1015 * If we didn't find a mergeclause, we're done --- any additional
1016 * sort-key positions in the pathkeys are useless. (But we can still
1017 * mergejoin if we found at least one mergeclause.)
1018 */
1019 if (matched_restrictinfos == NIL)
1020 break;
1021
1022 /*
1023 * If we did find usable mergeclause(s) for this sort-key position,
1024 * add them to result list.
1025 */
1026 mergeclauses = list_concat(mergeclauses, matched_restrictinfos);
1027 }
1028
1029 return mergeclauses;
1030 }
1031
1032 /*
1033 * select_outer_pathkeys_for_merge
1034 * Builds a pathkey list representing a possible sort ordering
1035 * that can be used with the given mergeclauses.
1036 *
1037 * 'mergeclauses' is a list of RestrictInfos for mergejoin clauses
1038 * that will be used in a merge join.
1039 * 'joinrel' is the join relation we are trying to construct.
1040 *
1041 * The restrictinfos must be marked (via outer_is_left) to show which side
1042 * of each clause is associated with the current outer path. (See
1043 * select_mergejoin_clauses())
1044 *
1045 * Returns a pathkeys list that can be applied to the outer relation.
1046 *
1047 * Since we assume here that a sort is required, there is no particular use
1048 * in matching any available ordering of the outerrel. (joinpath.c has an
1049 * entirely separate code path for considering sort-free mergejoins.) Rather,
1050 * it's interesting to try to match the requested query_pathkeys so that a
1051 * second output sort may be avoided; and failing that, we try to list "more
1052 * popular" keys (those with the most unmatched EquivalenceClass peers)
1053 * earlier, in hopes of making the resulting ordering useful for as many
1054 * higher-level mergejoins as possible.
1055 */
1056 List *
1057 select_outer_pathkeys_for_merge(PlannerInfo *root,
1058 List *mergeclauses,
1059 RelOptInfo *joinrel)
1060 {
1061 List *pathkeys = NIL;
1062 int nClauses = list_length(mergeclauses);
1063 EquivalenceClass **ecs;
1064 int *scores;
1065 int necs;
1066 ListCell *lc;
1067 int j;
1068
1069 /* Might have no mergeclauses */
1070 if (nClauses == 0)
1071 return NIL;
1072
1073 /*
1074 * Make arrays of the ECs used by the mergeclauses (dropping any
1075 * duplicates) and their "popularity" scores.
1076 */
1077 ecs = (EquivalenceClass **) palloc(nClauses * sizeof(EquivalenceClass *));
1078 scores = (int *) palloc(nClauses * sizeof(int));
1079 necs = 0;
1080
1081 foreach(lc, mergeclauses)
1082 {
1083 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1084 EquivalenceClass *oeclass;
1085 int score;
1086 ListCell *lc2;
1087
1088 /* get the outer eclass */
1089 cache_mergeclause_eclasses(root, rinfo);
1090
1091 if (rinfo->outer_is_left)
1092 oeclass = rinfo->left_ec;
1093 else
1094 oeclass = rinfo->right_ec;
1095
1096 /* reject duplicates */
1097 for (j = 0; j < necs; j++)
1098 {
1099 if (ecs[j] == oeclass)
1100 break;
1101 }
1102 if (j < necs)
1103 continue;
1104
1105 /* compute score */
1106 score = 0;
1107 foreach(lc2, oeclass->ec_members)
1108 {
1109 EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
1110
1111 /* Potential future join partner? */
1112 if (!em->em_is_const && !em->em_is_child &&
1113 !bms_overlap(em->em_relids, joinrel->relids))
1114 score++;
1115 }
1116
1117 ecs[necs] = oeclass;
1118 scores[necs] = score;
1119 necs++;
1120 }
1121
1122 /*
1123 * Find out if we have all the ECs mentioned in query_pathkeys; if so we
1124 * can generate a sort order that's also useful for final output. There is
1125 * no percentage in a partial match, though, so we have to have 'em all.
1126 */
1127 if (root->query_pathkeys)
1128 {
1129 foreach(lc, root->query_pathkeys)
1130 {
1131 PathKey *query_pathkey = (PathKey *) lfirst(lc);
1132 EquivalenceClass *query_ec = query_pathkey->pk_eclass;
1133
1134 for (j = 0; j < necs; j++)
1135 {
1136 if (ecs[j] == query_ec)
1137 break; /* found match */
1138 }
1139 if (j >= necs)
1140 break; /* didn't find match */
1141 }
1142 /* if we got to the end of the list, we have them all */
1143 if (lc == NULL)
1144 {
1145 /* copy query_pathkeys as starting point for our output */
1146 pathkeys = list_copy(root->query_pathkeys);
1147 /* mark their ECs as already-emitted */
1148 foreach(lc, root->query_pathkeys)
1149 {
1150 PathKey *query_pathkey = (PathKey *) lfirst(lc);
1151 EquivalenceClass *query_ec = query_pathkey->pk_eclass;
1152
1153 for (j = 0; j < necs; j++)
1154 {
1155 if (ecs[j] == query_ec)
1156 {
1157 scores[j] = -1;
1158 break;
1159 }
1160 }
1161 }
1162 }
1163 }
1164
1165 /*
1166 * Add remaining ECs to the list in popularity order, using a default sort
1167 * ordering. (We could use qsort() here, but the list length is usually
1168 * so small it's not worth it.)
1169 */
1170 for (;;)
1171 {
1172 int best_j;
1173 int best_score;
1174 EquivalenceClass *ec;
1175 PathKey *pathkey;
1176
1177 best_j = 0;
1178 best_score = scores[0];
1179 for (j = 1; j < necs; j++)
1180 {
1181 if (scores[j] > best_score)
1182 {
1183 best_j = j;
1184 best_score = scores[j];
1185 }
1186 }
1187 if (best_score < 0)
1188 break; /* all done */
1189 ec = ecs[best_j];
1190 scores[best_j] = -1;
1191 pathkey = make_canonical_pathkey(root,
1192 ec,
1193 linitial_oid(ec->ec_opfamilies),
1194 BTLessStrategyNumber,
1195 false);
1196 /* can't be redundant because no duplicate ECs */
1197 Assert(!pathkey_is_redundant(pathkey, pathkeys));
1198 pathkeys = lappend(pathkeys, pathkey);
1199 }
1200
1201 pfree(ecs);
1202 pfree(scores);
1203
1204 return pathkeys;
1205 }
1206
1207 /*
1208 * make_inner_pathkeys_for_merge
1209 * Builds a pathkey list representing the explicit sort order that
1210 * must be applied to an inner path to make it usable with the
1211 * given mergeclauses.
1212 *
1213 * 'mergeclauses' is a list of RestrictInfos for mergejoin clauses
1214 * that will be used in a merge join.
1215 * 'outer_pathkeys' are the already-known canonical pathkeys for the outer
1216 * side of the join.
1217 *
1218 * The restrictinfos must be marked (via outer_is_left) to show which side
1219 * of each clause is associated with the current outer path. (See
1220 * select_mergejoin_clauses())
1221 *
1222 * Returns a pathkeys list that can be applied to the inner relation.
1223 *
1224 * Note that it is not this routine's job to decide whether sorting is
1225 * actually needed for a particular input path. Assume a sort is necessary;
1226 * just make the keys, eh?
1227 */
1228 List *
1229 make_inner_pathkeys_for_merge(PlannerInfo *root,
1230 List *mergeclauses,
1231 List *outer_pathkeys)
1232 {
1233 List *pathkeys = NIL;
1234 EquivalenceClass *lastoeclass;
1235 PathKey *opathkey;
1236 ListCell *lc;
1237 ListCell *lop;
1238
1239 lastoeclass = NULL;
1240 opathkey = NULL;
1241 lop = list_head(outer_pathkeys);
1242
1243 foreach(lc, mergeclauses)
1244 {
1245 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1246 EquivalenceClass *oeclass;
1247 EquivalenceClass *ieclass;
1248 PathKey *pathkey;
1249
1250 cache_mergeclause_eclasses(root, rinfo);
1251
1252 if (rinfo->outer_is_left)
1253 {
1254 oeclass = rinfo->left_ec;
1255 ieclass = rinfo->right_ec;
1256 }
1257 else
1258 {
1259 oeclass = rinfo->right_ec;
1260 ieclass = rinfo->left_ec;
1261 }
1262
1263 /* outer eclass should match current or next pathkeys */
1264 /* we check this carefully for debugging reasons */
1265 if (oeclass != lastoeclass)
1266 {
1267 if (!lop)
1268 elog(ERROR, "too few pathkeys for mergeclauses");
1269 opathkey = (PathKey *) lfirst(lop);
1270 lop = lnext(lop);
1271 lastoeclass = opathkey->pk_eclass;
1272 if (oeclass != lastoeclass)
1273 elog(ERROR, "outer pathkeys do not match mergeclause");
1274 }
1275
1276 /*
1277 * Often, we'll have same EC on both sides, in which case the outer
1278 * pathkey is also canonical for the inner side, and we can skip a
1279 * useless search.
1280 */
1281 if (ieclass == oeclass)
1282 pathkey = opathkey;
1283 else
1284 pathkey = make_canonical_pathkey(root,
1285 ieclass,
1286 opathkey->pk_opfamily,
1287 opathkey->pk_strategy,
1288 opathkey->pk_nulls_first);
1289
1290 /*
1291 * Don't generate redundant pathkeys (can happen if multiple
1292 * mergeclauses refer to same EC).
1293 */
1294 if (!pathkey_is_redundant(pathkey, pathkeys))
1295 pathkeys = lappend(pathkeys, pathkey);
1296 }
1297
1298 return pathkeys;
1299 }
1300
1301 /****************************************************************************
1302 * PATHKEY USEFULNESS CHECKS
1303 *
1304 * We only want to remember as many of the pathkeys of a path as have some
1305 * potential use, either for subsequent mergejoins or for meeting the query's
1306 * requested output ordering. This ensures that add_path() won't consider
1307 * a path to have a usefully different ordering unless it really is useful.
1308 * These routines check for usefulness of given pathkeys.
1309 ****************************************************************************/
1310
1311 /*
1312 * pathkeys_useful_for_merging
1313 * Count the number of pathkeys that may be useful for mergejoins
1314 * above the given relation.
1315 *
1316 * We consider a pathkey potentially useful if it corresponds to the merge
1317 * ordering of either side of any joinclause for the rel. This might be
1318 * overoptimistic, since joinclauses that require different other relations
1319 * might never be usable at the same time, but trying to be exact is likely
1320 * to be more trouble than it's worth.
1321 *
1322 * To avoid doubling the number of mergejoin paths considered, we would like
1323 * to consider only one of the two scan directions (ASC or DESC) as useful
1324 * for merging for any given target column. The choice is arbitrary unless
1325 * one of the directions happens to match an ORDER BY key, in which case
1326 * that direction should be preferred, in hopes of avoiding a final sort step.
1327 * right_merge_direction() implements this heuristic.
1328 */
1329 int
1330 pathkeys_useful_for_merging(PlannerInfo *root, RelOptInfo *rel, List *pathkeys)
1331 {
1332 int useful = 0;
1333 ListCell *i;
1334
1335 foreach(i, pathkeys)
1336 {
1337 PathKey *pathkey = (PathKey *) lfirst(i);
1338 bool matched = false;
1339 ListCell *j;
1340
1341 /* If "wrong" direction, not useful for merging */
1342 if (!right_merge_direction(root, pathkey))
1343 break;
1344
1345 /*
1346 * First look into the EquivalenceClass of the pathkey, to see if
1347 * there are any members not yet joined to the rel. If so, it's
1348 * surely possible to generate a mergejoin clause using them.
1349 */
1350 if (rel->has_eclass_joins &&
1351 eclass_useful_for_merging(pathkey->pk_eclass, rel))
1352 matched = true;
1353 else
1354 {
1355 /*
1356 * Otherwise search the rel's joininfo list, which contains
1357 * non-EquivalenceClass-derivable join clauses that might
1358 * nonetheless be mergejoinable.
1359 */
1360 foreach(j, rel->joininfo)
1361 {
1362 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(j);
1363
1364 if (restrictinfo->mergeopfamilies == NIL)
1365 continue;
1366 cache_mergeclause_eclasses(root, restrictinfo);
1367
1368 if (pathkey->pk_eclass == restrictinfo->left_ec ||
1369 pathkey->pk_eclass == restrictinfo->right_ec)
1370 {
1371 matched = true;
1372 break;
1373 }
1374 }
1375 }
1376
1377 /*
1378 * If we didn't find a mergeclause, we're done --- any additional
1379 * sort-key positions in the pathkeys are useless. (But we can still
1380 * mergejoin if we found at least one mergeclause.)
1381 */
1382 if (matched)
1383 useful++;
1384 else
1385 break;
1386 }
1387
1388 return useful;
1389 }
1390
1391 /*
1392 * right_merge_direction
1393 * Check whether the pathkey embodies the preferred sort direction
1394 * for merging its target column.
1395 */
1396 static bool
1397 right_merge_direction(PlannerInfo *root, PathKey *pathkey)
1398 {
1399 ListCell *l;
1400
1401 foreach(l, root->query_pathkeys)
1402 {
1403 PathKey *query_pathkey = (PathKey *) lfirst(l);
1404
1405 if (pathkey->pk_eclass == query_pathkey->pk_eclass &&
1406 pathkey->pk_opfamily == query_pathkey->pk_opfamily)
1407 {
1408 /*
1409 * Found a matching query sort column. Prefer this pathkey's
1410 * direction iff it matches. Note that we ignore pk_nulls_first,
1411 * which means that a sort might be needed anyway ... but we still
1412 * want to prefer only one of the two possible directions, and we
1413 * might as well use this one.
1414 */
1415 return (pathkey->pk_strategy == query_pathkey->pk_strategy);
1416 }
1417 }
1418
1419 /* If no matching ORDER BY request, prefer the ASC direction */
1420 return (pathkey->pk_strategy == BTLessStrategyNumber);
1421 }
1422
1423 /*
1424 * pathkeys_useful_for_ordering
1425 * Count the number of pathkeys that are useful for meeting the
1426 * query's requested output ordering.
1427 *
1428 * Unlike merge pathkeys, this is an all-or-nothing affair: it does us
1429 * no good to order by just the first key(s) of the requested ordering.
1430 * So the result is always either 0 or list_length(root->query_pathkeys).
1431 */
1432 int
1433 pathkeys_useful_for_ordering(PlannerInfo *root, List *pathkeys)
1434 {
1435 if (root->query_pathkeys == NIL)
1436 return 0; /* no special ordering requested */
1437
1438 if (pathkeys == NIL)
1439 return 0; /* unordered path */
1440
1441 if (pathkeys_contained_in(root->query_pathkeys, pathkeys))
1442 {
1443 /* It's useful ... or at least the first N keys are */
1444 return list_length(root->query_pathkeys);
1445 }
1446
1447 return 0; /* path ordering not useful */
1448 }
1449
1450 /*
1451 * truncate_useless_pathkeys
1452 * Shorten the given pathkey list to just the useful pathkeys.
1453 */
1454 List *
1455 truncate_useless_pathkeys(PlannerInfo *root,
1456 RelOptInfo *rel,
1457 List *pathkeys)
1458 {
1459 int nuseful;
1460 int nuseful2;
1461
1462 nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
1463 nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
1464 if (nuseful2 > nuseful)
1465 nuseful = nuseful2;
1466
1467 /*
1468 * Note: not safe to modify input list destructively, but we can avoid
1469 * copying the list if we're not actually going to change it
1470 */
1471 if (nuseful == 0)
1472 return NIL;
1473 else if (nuseful == list_length(pathkeys))
1474 return pathkeys;
1475 else
1476 return list_truncate(list_copy(pathkeys), nuseful);
1477 }
1478
1479 /*
1480 * has_useful_pathkeys
1481 * Detect whether the specified rel could have any pathkeys that are
1482 * useful according to truncate_useless_pathkeys().
1483 *
1484 * This is a cheap test that lets us skip building pathkeys at all in very
1485 * simple queries. It's OK to err in the direction of returning "true" when
1486 * there really aren't any usable pathkeys, but erring in the other direction
1487 * is bad --- so keep this in sync with the routines above!
1488 *
1489 * We could make the test more complex, for example checking to see if any of
1490 * the joinclauses are really mergejoinable, but that likely wouldn't win
1491 * often enough to repay the extra cycles. Queries with neither a join nor
1492 * a sort are reasonably common, though, so this much work seems worthwhile.
1493 */
1494 bool
1495 has_useful_pathkeys(PlannerInfo *root, RelOptInfo *rel)
1496 {
1497 if (rel->joininfo != NIL || rel->has_eclass_joins)
1498 return true; /* might be able to use pathkeys for merging */
1499 if (root->query_pathkeys != NIL)
1500 return true; /* might be able to use them for ordering */
1501 return false; /* definitely useless */
1502 }
PostgreSQL from CVS