1 /*-------------------------------------------------------------------------
2 *
3 * analyzejoins.c
4 * Routines for simplifying joins after initial query analysis
5 *
6 * While we do a great deal of join simplification in prep/prepjointree.c,
7 * certain optimizations cannot be performed at that stage for lack of
8 * detailed information about the query. The routines here are invoked
9 * after initsplan.c has done its work, and can do additional join removal
10 * and simplification steps based on the information extracted. The penalty
11 * is that we have to work harder to clean up after ourselves when we modify
12 * the query, since the derived data structures have to be updated too.
13 *
14 * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
15 * Portions Copyright (c) 1994, Regents of the University of California
16 *
17 *
18 * IDENTIFICATION
19 * src/backend/optimizer/plan/analyzejoins.c
20 *
21 *-------------------------------------------------------------------------
22 */
23 #include "postgres.h"
24
25 #include "nodes/nodeFuncs.h"
26 #include "optimizer/clauses.h"
27 #include "optimizer/joininfo.h"
28 #include "optimizer/pathnode.h"
29 #include "optimizer/paths.h"
30 #include "optimizer/planmain.h"
31 #include "optimizer/tlist.h"
32 #include "optimizer/var.h"
33 #include "utils/lsyscache.h"
34
35 /* local functions */
36 static bool join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo);
37 static void remove_rel_from_query(PlannerInfo *root, int relid,
38 Relids joinrelids);
39 static List *remove_rel_from_joinlist(List *joinlist, int relid, int *nremoved);
40 static bool rel_supports_distinctness(PlannerInfo *root, RelOptInfo *rel);
41 static bool rel_is_distinct_for(PlannerInfo *root, RelOptInfo *rel,
42 List *clause_list);
43 static Oid distinct_col_search(int colno, List *colnos, List *opids);
44 static bool is_innerrel_unique_for(PlannerInfo *root,
45 Relids outerrelids,
46 RelOptInfo *innerrel,
47 JoinType jointype,
48 List *restrictlist);
49
50
51 /*
52 * remove_useless_joins
53 * Check for relations that don't actually need to be joined at all,
54 * and remove them from the query.
55 *
56 * We are passed the current joinlist and return the updated list. Other
57 * data structures that have to be updated are accessible via "root".
58 */
59 List *
remove_useless_joins(PlannerInfo * root,List * joinlist)60 remove_useless_joins(PlannerInfo *root, List *joinlist)
61 {
62 ListCell *lc;
63
64 /*
65 * We are only interested in relations that are left-joined to, so we can
66 * scan the join_info_list to find them easily.
67 */
68 restart:
69 foreach(lc, root->join_info_list)
70 {
71 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
72 int innerrelid;
73 int nremoved;
74
75 /* Skip if not removable */
76 if (!join_is_removable(root, sjinfo))
77 continue;
78
79 /*
80 * Currently, join_is_removable can only succeed when the sjinfo's
81 * righthand is a single baserel. Remove that rel from the query and
82 * joinlist.
83 */
84 innerrelid = bms_singleton_member(sjinfo->min_righthand);
85
86 remove_rel_from_query(root, innerrelid,
87 bms_union(sjinfo->min_lefthand,
88 sjinfo->min_righthand));
89
90 /* We verify that exactly one reference gets removed from joinlist */
91 nremoved = 0;
92 joinlist = remove_rel_from_joinlist(joinlist, innerrelid, &nremoved);
93 if (nremoved != 1)
94 elog(ERROR, "failed to find relation %d in joinlist", innerrelid);
95
96 /*
97 * We can delete this SpecialJoinInfo from the list too, since it's no
98 * longer of interest.
99 */
100 root->join_info_list = list_delete_ptr(root->join_info_list, sjinfo);
101
102 /*
103 * Restart the scan. This is necessary to ensure we find all
104 * removable joins independently of ordering of the join_info_list
105 * (note that removal of attr_needed bits may make a join appear
106 * removable that did not before). Also, since we just deleted the
107 * current list cell, we'd have to have some kluge to continue the
108 * list scan anyway.
109 */
110 goto restart;
111 }
112
113 return joinlist;
114 }
115
116 /*
117 * clause_sides_match_join
118 * Determine whether a join clause is of the right form to use in this join.
119 *
120 * We already know that the clause is a binary opclause referencing only the
121 * rels in the current join. The point here is to check whether it has the
122 * form "outerrel_expr op innerrel_expr" or "innerrel_expr op outerrel_expr",
123 * rather than mixing outer and inner vars on either side. If it matches,
124 * we set the transient flag outer_is_left to identify which side is which.
125 */
126 static inline bool
clause_sides_match_join(RestrictInfo * rinfo,Relids outerrelids,Relids innerrelids)127 clause_sides_match_join(RestrictInfo *rinfo, Relids outerrelids,
128 Relids innerrelids)
129 {
130 if (bms_is_subset(rinfo->left_relids, outerrelids) &&
131 bms_is_subset(rinfo->right_relids, innerrelids))
132 {
133 /* lefthand side is outer */
134 rinfo->outer_is_left = true;
135 return true;
136 }
137 else if (bms_is_subset(rinfo->left_relids, innerrelids) &&
138 bms_is_subset(rinfo->right_relids, outerrelids))
139 {
140 /* righthand side is outer */
141 rinfo->outer_is_left = false;
142 return true;
143 }
144 return false; /* no good for these input relations */
145 }
146
147 /*
148 * join_is_removable
149 * Check whether we need not perform this special join at all, because
150 * it will just duplicate its left input.
151 *
152 * This is true for a left join for which the join condition cannot match
153 * more than one inner-side row. (There are other possibly interesting
154 * cases, but we don't have the infrastructure to prove them.) We also
155 * have to check that the inner side doesn't generate any variables needed
156 * above the join.
157 */
158 static bool
join_is_removable(PlannerInfo * root,SpecialJoinInfo * sjinfo)159 join_is_removable(PlannerInfo *root, SpecialJoinInfo *sjinfo)
160 {
161 int innerrelid;
162 RelOptInfo *innerrel;
163 Relids joinrelids;
164 List *clause_list = NIL;
165 ListCell *l;
166 int attroff;
167
168 /*
169 * Must be a non-delaying left join to a single baserel, else we aren't
170 * going to be able to do anything with it.
171 */
172 if (sjinfo->jointype != JOIN_LEFT ||
173 sjinfo->delay_upper_joins)
174 return false;
175
176 if (!bms_get_singleton_member(sjinfo->min_righthand, &innerrelid))
177 return false;
178
179 innerrel = find_base_rel(root, innerrelid);
180
181 /*
182 * Before we go to the effort of checking whether any innerrel variables
183 * are needed above the join, make a quick check to eliminate cases in
184 * which we will surely be unable to prove uniqueness of the innerrel.
185 */
186 if (!rel_supports_distinctness(root, innerrel))
187 return false;
188
189 /* Compute the relid set for the join we are considering */
190 joinrelids = bms_union(sjinfo->min_lefthand, sjinfo->min_righthand);
191
192 /*
193 * We can't remove the join if any inner-rel attributes are used above the
194 * join.
195 *
196 * Note that this test only detects use of inner-rel attributes in higher
197 * join conditions and the target list. There might be such attributes in
198 * pushed-down conditions at this join, too. We check that case below.
199 *
200 * As a micro-optimization, it seems better to start with max_attr and
201 * count down rather than starting with min_attr and counting up, on the
202 * theory that the system attributes are somewhat less likely to be wanted
203 * and should be tested last.
204 */
205 for (attroff = innerrel->max_attr - innerrel->min_attr;
206 attroff >= 0;
207 attroff--)
208 {
209 if (!bms_is_subset(innerrel->attr_needed[attroff], joinrelids))
210 return false;
211 }
212
213 /*
214 * Similarly check that the inner rel isn't needed by any PlaceHolderVars
215 * that will be used above the join. We only need to fail if such a PHV
216 * actually references some inner-rel attributes; but the correct check
217 * for that is relatively expensive, so we first check against ph_eval_at,
218 * which must mention the inner rel if the PHV uses any inner-rel attrs as
219 * non-lateral references. Note that if the PHV's syntactic scope is just
220 * the inner rel, we can't drop the rel even if the PHV is variable-free.
221 */
222 foreach(l, root->placeholder_list)
223 {
224 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
225
226 if (bms_overlap(phinfo->ph_lateral, innerrel->relids))
227 return false; /* it references innerrel laterally */
228 if (bms_is_subset(phinfo->ph_needed, joinrelids))
229 continue; /* PHV is not used above the join */
230 if (!bms_overlap(phinfo->ph_eval_at, innerrel->relids))
231 continue; /* it definitely doesn't reference innerrel */
232 if (bms_is_subset(phinfo->ph_eval_at, innerrel->relids))
233 return false; /* there isn't any other place to eval PHV */
234 if (bms_overlap(pull_varnos((Node *) phinfo->ph_var->phexpr),
235 innerrel->relids))
236 return false; /* it does reference innerrel */
237 }
238
239 /*
240 * Search for mergejoinable clauses that constrain the inner rel against
241 * either the outer rel or a pseudoconstant. If an operator is
242 * mergejoinable then it behaves like equality for some btree opclass, so
243 * it's what we want. The mergejoinability test also eliminates clauses
244 * containing volatile functions, which we couldn't depend on.
245 */
246 foreach(l, innerrel->joininfo)
247 {
248 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(l);
249
250 /*
251 * If it's not a join clause for this outer join, we can't use it.
252 * Note that if the clause is pushed-down, then it is logically from
253 * above the outer join, even if it references no other rels (it might
254 * be from WHERE, for example).
255 */
256 if (RINFO_IS_PUSHED_DOWN(restrictinfo, joinrelids))
257 {
258 /*
259 * If such a clause actually references the inner rel then join
260 * removal has to be disallowed. We have to check this despite
261 * the previous attr_needed checks because of the possibility of
262 * pushed-down clauses referencing the rel.
263 */
264 if (bms_is_member(innerrelid, restrictinfo->clause_relids))
265 return false;
266 continue; /* else, ignore; not useful here */
267 }
268
269 /* Ignore if it's not a mergejoinable clause */
270 if (!restrictinfo->can_join ||
271 restrictinfo->mergeopfamilies == NIL)
272 continue; /* not mergejoinable */
273
274 /*
275 * Check if clause has the form "outer op inner" or "inner op outer",
276 * and if so mark which side is inner.
277 */
278 if (!clause_sides_match_join(restrictinfo, sjinfo->min_lefthand,
279 innerrel->relids))
280 continue; /* no good for these input relations */
281
282 /* OK, add to list */
283 clause_list = lappend(clause_list, restrictinfo);
284 }
285
286 /*
287 * Now that we have the relevant equality join clauses, try to prove the
288 * innerrel distinct.
289 */
290 if (rel_is_distinct_for(root, innerrel, clause_list))
291 return true;
292
293 /*
294 * Some day it would be nice to check for other methods of establishing
295 * distinctness.
296 */
297 return false;
298 }
299
300
301 /*
302 * Remove the target relid from the planner's data structures, having
303 * determined that there is no need to include it in the query.
304 *
305 * We are not terribly thorough here. We must make sure that the rel is
306 * no longer treated as a baserel, and that attributes of other baserels
307 * are no longer marked as being needed at joins involving this rel.
308 * Also, join quals involving the rel have to be removed from the joininfo
309 * lists, but only if they belong to the outer join identified by joinrelids.
310 */
311 static void
remove_rel_from_query(PlannerInfo * root,int relid,Relids joinrelids)312 remove_rel_from_query(PlannerInfo *root, int relid, Relids joinrelids)
313 {
314 RelOptInfo *rel = find_base_rel(root, relid);
315 List *joininfos;
316 Index rti;
317 ListCell *l;
318 ListCell *nextl;
319
320 /*
321 * Mark the rel as "dead" to show it is no longer part of the join tree.
322 * (Removing it from the baserel array altogether seems too risky.)
323 */
324 rel->reloptkind = RELOPT_DEADREL;
325
326 /*
327 * Remove references to the rel from other baserels' attr_needed arrays.
328 */
329 for (rti = 1; rti < root->simple_rel_array_size; rti++)
330 {
331 RelOptInfo *otherrel = root->simple_rel_array[rti];
332 int attroff;
333
334 /* there may be empty slots corresponding to non-baserel RTEs */
335 if (otherrel == NULL)
336 continue;
337
338 Assert(otherrel->relid == rti); /* sanity check on array */
339
340 /* no point in processing target rel itself */
341 if (otherrel == rel)
342 continue;
343
344 for (attroff = otherrel->max_attr - otherrel->min_attr;
345 attroff >= 0;
346 attroff--)
347 {
348 otherrel->attr_needed[attroff] =
349 bms_del_member(otherrel->attr_needed[attroff], relid);
350 }
351 }
352
353 /*
354 * Likewise remove references from SpecialJoinInfo data structures.
355 *
356 * This is relevant in case the outer join we're deleting is nested inside
357 * other outer joins: the upper joins' relid sets have to be adjusted. The
358 * RHS of the target outer join will be made empty here, but that's OK
359 * since caller will delete that SpecialJoinInfo entirely.
360 */
361 foreach(l, root->join_info_list)
362 {
363 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
364
365 sjinfo->min_lefthand = bms_del_member(sjinfo->min_lefthand, relid);
366 sjinfo->min_righthand = bms_del_member(sjinfo->min_righthand, relid);
367 sjinfo->syn_lefthand = bms_del_member(sjinfo->syn_lefthand, relid);
368 sjinfo->syn_righthand = bms_del_member(sjinfo->syn_righthand, relid);
369 }
370
371 /*
372 * Likewise remove references from PlaceHolderVar data structures,
373 * removing any no-longer-needed placeholders entirely.
374 *
375 * Removal is a bit tricker than it might seem: we can remove PHVs that
376 * are used at the target rel and/or in the join qual, but not those that
377 * are used at join partner rels or above the join. It's not that easy to
378 * distinguish PHVs used at partner rels from those used in the join qual,
379 * since they will both have ph_needed sets that are subsets of
380 * joinrelids. However, a PHV used at a partner rel could not have the
381 * target rel in ph_eval_at, so we check that while deciding whether to
382 * remove or just update the PHV. There is no corresponding test in
383 * join_is_removable because it doesn't need to distinguish those cases.
384 */
385 for (l = list_head(root->placeholder_list); l != NULL; l = nextl)
386 {
387 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
388
389 nextl = lnext(l);
390 Assert(!bms_is_member(relid, phinfo->ph_lateral));
391 if (bms_is_subset(phinfo->ph_needed, joinrelids) &&
392 bms_is_member(relid, phinfo->ph_eval_at))
393 root->placeholder_list = list_delete_ptr(root->placeholder_list,
394 phinfo);
395 else
396 {
397 phinfo->ph_eval_at = bms_del_member(phinfo->ph_eval_at, relid);
398 Assert(!bms_is_empty(phinfo->ph_eval_at));
399 phinfo->ph_needed = bms_del_member(phinfo->ph_needed, relid);
400 }
401 }
402
403 /*
404 * Remove any joinquals referencing the rel from the joininfo lists.
405 *
406 * In some cases, a joinqual has to be put back after deleting its
407 * reference to the target rel. This can occur for pseudoconstant and
408 * outerjoin-delayed quals, which can get marked as requiring the rel in
409 * order to force them to be evaluated at or above the join. We can't
410 * just discard them, though. Only quals that logically belonged to the
411 * outer join being discarded should be removed from the query.
412 *
413 * We must make a copy of the rel's old joininfo list before starting the
414 * loop, because otherwise remove_join_clause_from_rels would destroy the
415 * list while we're scanning it.
416 */
417 joininfos = list_copy(rel->joininfo);
418 foreach(l, joininfos)
419 {
420 RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
421
422 remove_join_clause_from_rels(root, rinfo, rinfo->required_relids);
423
424 if (RINFO_IS_PUSHED_DOWN(rinfo, joinrelids))
425 {
426 /* Recheck that qual doesn't actually reference the target rel */
427 Assert(!bms_is_member(relid, rinfo->clause_relids));
428
429 /*
430 * The required_relids probably aren't shared with anything else,
431 * but let's copy them just to be sure.
432 */
433 rinfo->required_relids = bms_copy(rinfo->required_relids);
434 rinfo->required_relids = bms_del_member(rinfo->required_relids,
435 relid);
436 distribute_restrictinfo_to_rels(root, rinfo);
437 }
438 }
439
440 /*
441 * There may be references to the rel in root->fkey_list, but if so,
442 * match_foreign_keys_to_quals() will get rid of them.
443 */
444 }
445
446 /*
447 * Remove any occurrences of the target relid from a joinlist structure.
448 *
449 * It's easiest to build a whole new list structure, so we handle it that
450 * way. Efficiency is not a big deal here.
451 *
452 * *nremoved is incremented by the number of occurrences removed (there
453 * should be exactly one, but the caller checks that).
454 */
455 static List *
remove_rel_from_joinlist(List * joinlist,int relid,int * nremoved)456 remove_rel_from_joinlist(List *joinlist, int relid, int *nremoved)
457 {
458 List *result = NIL;
459 ListCell *jl;
460
461 foreach(jl, joinlist)
462 {
463 Node *jlnode = (Node *) lfirst(jl);
464
465 if (IsA(jlnode, RangeTblRef))
466 {
467 int varno = ((RangeTblRef *) jlnode)->rtindex;
468
469 if (varno == relid)
470 (*nremoved)++;
471 else
472 result = lappend(result, jlnode);
473 }
474 else if (IsA(jlnode, List))
475 {
476 /* Recurse to handle subproblem */
477 List *sublist;
478
479 sublist = remove_rel_from_joinlist((List *) jlnode,
480 relid, nremoved);
481 /* Avoid including empty sub-lists in the result */
482 if (sublist)
483 result = lappend(result, sublist);
484 }
485 else
486 {
487 elog(ERROR, "unrecognized joinlist node type: %d",
488 (int) nodeTag(jlnode));
489 }
490 }
491
492 return result;
493 }
494
495
496 /*
497 * reduce_unique_semijoins
498 * Check for semijoins that can be simplified to plain inner joins
499 * because the inner relation is provably unique for the join clauses.
500 *
501 * Ideally this would happen during reduce_outer_joins, but we don't have
502 * enough information at that point.
503 *
504 * To perform the strength reduction when applicable, we need only delete
505 * the semijoin's SpecialJoinInfo from root->join_info_list. (We don't
506 * bother fixing the join type attributed to it in the query jointree,
507 * since that won't be consulted again.)
508 */
509 void
reduce_unique_semijoins(PlannerInfo * root)510 reduce_unique_semijoins(PlannerInfo *root)
511 {
512 ListCell *lc;
513 ListCell *next;
514
515 /*
516 * Scan the join_info_list to find semijoins. We can't use foreach
517 * because we may delete the current cell.
518 */
519 for (lc = list_head(root->join_info_list); lc != NULL; lc = next)
520 {
521 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(lc);
522 int innerrelid;
523 RelOptInfo *innerrel;
524 Relids joinrelids;
525 List *restrictlist;
526
527 next = lnext(lc);
528
529 /*
530 * Must be a non-delaying semijoin to a single baserel, else we aren't
531 * going to be able to do anything with it. (It's probably not
532 * possible for delay_upper_joins to be set on a semijoin, but we
533 * might as well check.)
534 */
535 if (sjinfo->jointype != JOIN_SEMI ||
536 sjinfo->delay_upper_joins)
537 continue;
538
539 if (!bms_get_singleton_member(sjinfo->min_righthand, &innerrelid))
540 continue;
541
542 innerrel = find_base_rel(root, innerrelid);
543
544 /*
545 * Before we trouble to run generate_join_implied_equalities, make a
546 * quick check to eliminate cases in which we will surely be unable to
547 * prove uniqueness of the innerrel.
548 */
549 if (!rel_supports_distinctness(root, innerrel))
550 continue;
551
552 /* Compute the relid set for the join we are considering */
553 joinrelids = bms_union(sjinfo->min_lefthand, sjinfo->min_righthand);
554
555 /*
556 * Since we're only considering a single-rel RHS, any join clauses it
557 * has must be clauses linking it to the semijoin's min_lefthand. We
558 * can also consider EC-derived join clauses.
559 */
560 restrictlist =
561 list_concat(generate_join_implied_equalities(root,
562 joinrelids,
563 sjinfo->min_lefthand,
564 innerrel),
565 innerrel->joininfo);
566
567 /* Test whether the innerrel is unique for those clauses. */
568 if (!innerrel_is_unique(root, sjinfo->min_lefthand, innerrel,
569 JOIN_SEMI, restrictlist, true))
570 continue;
571
572 /* OK, remove the SpecialJoinInfo from the list. */
573 root->join_info_list = list_delete_ptr(root->join_info_list, sjinfo);
574 }
575 }
576
577
578 /*
579 * rel_supports_distinctness
580 * Could the relation possibly be proven distinct on some set of columns?
581 *
582 * This is effectively a pre-checking function for rel_is_distinct_for().
583 * It must return TRUE if rel_is_distinct_for() could possibly return TRUE
584 * with this rel, but it should not expend a lot of cycles. The idea is
585 * that callers can avoid doing possibly-expensive processing to compute
586 * rel_is_distinct_for()'s argument lists if the call could not possibly
587 * succeed.
588 */
589 static bool
rel_supports_distinctness(PlannerInfo * root,RelOptInfo * rel)590 rel_supports_distinctness(PlannerInfo *root, RelOptInfo *rel)
591 {
592 /* We only know about baserels ... */
593 if (rel->reloptkind != RELOPT_BASEREL)
594 return false;
595 if (rel->rtekind == RTE_RELATION)
596 {
597 /*
598 * For a plain relation, we only know how to prove uniqueness by
599 * reference to unique indexes. Make sure there's at least one
600 * suitable unique index. It must be immediately enforced, and if
601 * it's a partial index, it must match the query. (Keep these
602 * conditions in sync with relation_has_unique_index_for!)
603 */
604 ListCell *lc;
605
606 foreach(lc, rel->indexlist)
607 {
608 IndexOptInfo *ind = (IndexOptInfo *) lfirst(lc);
609
610 if (ind->unique && ind->immediate &&
611 (ind->indpred == NIL || ind->predOK))
612 return true;
613 }
614 }
615 else if (rel->rtekind == RTE_SUBQUERY)
616 {
617 Query *subquery = root->simple_rte_array[rel->relid]->subquery;
618
619 /* Check if the subquery has any qualities that support distinctness */
620 if (query_supports_distinctness(subquery))
621 return true;
622 }
623 /* We have no proof rules for any other rtekinds. */
624 return false;
625 }
626
627 /*
628 * rel_is_distinct_for
629 * Does the relation return only distinct rows according to clause_list?
630 *
631 * clause_list is a list of join restriction clauses involving this rel and
632 * some other one. Return true if no two rows emitted by this rel could
633 * possibly join to the same row of the other rel.
634 *
635 * The caller must have already determined that each condition is a
636 * mergejoinable equality with an expression in this relation on one side, and
637 * an expression not involving this relation on the other. The transient
638 * outer_is_left flag is used to identify which side references this relation:
639 * left side if outer_is_left is false, right side if it is true.
640 *
641 * Note that the passed-in clause_list may be destructively modified! This
642 * is OK for current uses, because the clause_list is built by the caller for
643 * the sole purpose of passing to this function.
644 */
645 static bool
rel_is_distinct_for(PlannerInfo * root,RelOptInfo * rel,List * clause_list)646 rel_is_distinct_for(PlannerInfo *root, RelOptInfo *rel, List *clause_list)
647 {
648 /*
649 * We could skip a couple of tests here if we assume all callers checked
650 * rel_supports_distinctness first, but it doesn't seem worth taking any
651 * risk for.
652 */
653 if (rel->reloptkind != RELOPT_BASEREL)
654 return false;
655 if (rel->rtekind == RTE_RELATION)
656 {
657 /*
658 * Examine the indexes to see if we have a matching unique index.
659 * relation_has_unique_index_for automatically adds any usable
660 * restriction clauses for the rel, so we needn't do that here.
661 */
662 if (relation_has_unique_index_for(root, rel, clause_list, NIL, NIL))
663 return true;
664 }
665 else if (rel->rtekind == RTE_SUBQUERY)
666 {
667 Index relid = rel->relid;
668 Query *subquery = root->simple_rte_array[relid]->subquery;
669 List *colnos = NIL;
670 List *opids = NIL;
671 ListCell *l;
672
673 /*
674 * Build the argument lists for query_is_distinct_for: a list of
675 * output column numbers that the query needs to be distinct over, and
676 * a list of equality operators that the output columns need to be
677 * distinct according to.
678 *
679 * (XXX we are not considering restriction clauses attached to the
680 * subquery; is that worth doing?)
681 */
682 foreach(l, clause_list)
683 {
684 RestrictInfo *rinfo = lfirst_node(RestrictInfo, l);
685 Oid op;
686 Var *var;
687
688 /*
689 * Get the equality operator we need uniqueness according to.
690 * (This might be a cross-type operator and thus not exactly the
691 * same operator the subquery would consider; that's all right
692 * since query_is_distinct_for can resolve such cases.) The
693 * caller's mergejoinability test should have selected only
694 * OpExprs.
695 */
696 op = castNode(OpExpr, rinfo->clause)->opno;
697
698 /* caller identified the inner side for us */
699 if (rinfo->outer_is_left)
700 var = (Var *) get_rightop(rinfo->clause);
701 else
702 var = (Var *) get_leftop(rinfo->clause);
703
704 /*
705 * We may ignore any RelabelType node above the operand. (There
706 * won't be more than one, since eval_const_expressions() has been
707 * applied already.)
708 */
709 if (var && IsA(var, RelabelType))
710 var = (Var *) ((RelabelType *) var)->arg;
711
712 /*
713 * If inner side isn't a Var referencing a subquery output column,
714 * this clause doesn't help us.
715 */
716 if (!var || !IsA(var, Var) ||
717 var->varno != relid || var->varlevelsup != 0)
718 continue;
719
720 colnos = lappend_int(colnos, var->varattno);
721 opids = lappend_oid(opids, op);
722 }
723
724 if (query_is_distinct_for(subquery, colnos, opids))
725 return true;
726 }
727 return false;
728 }
729
730
731 /*
732 * query_supports_distinctness - could the query possibly be proven distinct
733 * on some set of output columns?
734 *
735 * This is effectively a pre-checking function for query_is_distinct_for().
736 * It must return TRUE if query_is_distinct_for() could possibly return TRUE
737 * with this query, but it should not expend a lot of cycles. The idea is
738 * that callers can avoid doing possibly-expensive processing to compute
739 * query_is_distinct_for()'s argument lists if the call could not possibly
740 * succeed.
741 */
742 bool
query_supports_distinctness(Query * query)743 query_supports_distinctness(Query *query)
744 {
745 /* SRFs break distinctness except with DISTINCT, see below */
746 if (query->hasTargetSRFs && query->distinctClause == NIL)
747 return false;
748
749 /* check for features we can prove distinctness with */
750 if (query->distinctClause != NIL ||
751 query->groupClause != NIL ||
752 query->groupingSets != NIL ||
753 query->hasAggs ||
754 query->havingQual ||
755 query->setOperations)
756 return true;
757
758 return false;
759 }
760
761 /*
762 * query_is_distinct_for - does query never return duplicates of the
763 * specified columns?
764 *
765 * query is a not-yet-planned subquery (in current usage, it's always from
766 * a subquery RTE, which the planner avoids scribbling on).
767 *
768 * colnos is an integer list of output column numbers (resno's). We are
769 * interested in whether rows consisting of just these columns are certain
770 * to be distinct. "Distinctness" is defined according to whether the
771 * corresponding upper-level equality operators listed in opids would think
772 * the values are distinct. (Note: the opids entries could be cross-type
773 * operators, and thus not exactly the equality operators that the subquery
774 * would use itself. We use equality_ops_are_compatible() to check
775 * compatibility. That looks at btree or hash opfamily membership, and so
776 * should give trustworthy answers for all operators that we might need
777 * to deal with here.)
778 */
779 bool
query_is_distinct_for(Query * query,List * colnos,List * opids)780 query_is_distinct_for(Query *query, List *colnos, List *opids)
781 {
782 ListCell *l;
783 Oid opid;
784
785 Assert(list_length(colnos) == list_length(opids));
786
787 /*
788 * DISTINCT (including DISTINCT ON) guarantees uniqueness if all the
789 * columns in the DISTINCT clause appear in colnos and operator semantics
790 * match. This is true even if there are SRFs in the DISTINCT columns or
791 * elsewhere in the tlist.
792 */
793 if (query->distinctClause)
794 {
795 foreach(l, query->distinctClause)
796 {
797 SortGroupClause *sgc = (SortGroupClause *) lfirst(l);
798 TargetEntry *tle = get_sortgroupclause_tle(sgc,
799 query->targetList);
800
801 opid = distinct_col_search(tle->resno, colnos, opids);
802 if (!OidIsValid(opid) ||
803 !equality_ops_are_compatible(opid, sgc->eqop))
804 break; /* exit early if no match */
805 }
806 if (l == NULL) /* had matches for all? */
807 return true;
808 }
809
810 /*
811 * Otherwise, a set-returning function in the query's targetlist can
812 * result in returning duplicate rows, despite any grouping that might
813 * occur before tlist evaluation. (If all tlist SRFs are within GROUP BY
814 * columns, it would be safe because they'd be expanded before grouping.
815 * But it doesn't currently seem worth the effort to check for that.)
816 */
817 if (query->hasTargetSRFs)
818 return false;
819
820 /*
821 * Similarly, GROUP BY without GROUPING SETS guarantees uniqueness if all
822 * the grouped columns appear in colnos and operator semantics match.
823 */
824 if (query->groupClause && !query->groupingSets)
825 {
826 foreach(l, query->groupClause)
827 {
828 SortGroupClause *sgc = (SortGroupClause *) lfirst(l);
829 TargetEntry *tle = get_sortgroupclause_tle(sgc,
830 query->targetList);
831
832 opid = distinct_col_search(tle->resno, colnos, opids);
833 if (!OidIsValid(opid) ||
834 !equality_ops_are_compatible(opid, sgc->eqop))
835 break; /* exit early if no match */
836 }
837 if (l == NULL) /* had matches for all? */
838 return true;
839 }
840 else if (query->groupingSets)
841 {
842 /*
843 * If we have grouping sets with expressions, we probably don't have
844 * uniqueness and analysis would be hard. Punt.
845 */
846 if (query->groupClause)
847 return false;
848
849 /*
850 * If we have no groupClause (therefore no grouping expressions), we
851 * might have one or many empty grouping sets. If there's just one,
852 * then we're returning only one row and are certainly unique. But
853 * otherwise, we know we're certainly not unique.
854 */
855 if (list_length(query->groupingSets) == 1 &&
856 ((GroupingSet *) linitial(query->groupingSets))->kind == GROUPING_SET_EMPTY)
857 return true;
858 else
859 return false;
860 }
861 else
862 {
863 /*
864 * If we have no GROUP BY, but do have aggregates or HAVING, then the
865 * result is at most one row so it's surely unique, for any operators.
866 */
867 if (query->hasAggs || query->havingQual)
868 return true;
869 }
870
871 /*
872 * UNION, INTERSECT, EXCEPT guarantee uniqueness of the whole output row,
873 * except with ALL.
874 */
875 if (query->setOperations)
876 {
877 SetOperationStmt *topop = castNode(SetOperationStmt, query->setOperations);
878
879 Assert(topop->op != SETOP_NONE);
880
881 if (!topop->all)
882 {
883 ListCell *lg;
884
885 /* We're good if all the nonjunk output columns are in colnos */
886 lg = list_head(topop->groupClauses);
887 foreach(l, query->targetList)
888 {
889 TargetEntry *tle = (TargetEntry *) lfirst(l);
890 SortGroupClause *sgc;
891
892 if (tle->resjunk)
893 continue; /* ignore resjunk columns */
894
895 /* non-resjunk columns should have grouping clauses */
896 Assert(lg != NULL);
897 sgc = (SortGroupClause *) lfirst(lg);
898 lg = lnext(lg);
899
900 opid = distinct_col_search(tle->resno, colnos, opids);
901 if (!OidIsValid(opid) ||
902 !equality_ops_are_compatible(opid, sgc->eqop))
903 break; /* exit early if no match */
904 }
905 if (l == NULL) /* had matches for all? */
906 return true;
907 }
908 }
909
910 /*
911 * XXX Are there any other cases in which we can easily see the result
912 * must be distinct?
913 *
914 * If you do add more smarts to this function, be sure to update
915 * query_supports_distinctness() to match.
916 */
917
918 return false;
919 }
920
921 /*
922 * distinct_col_search - subroutine for query_is_distinct_for
923 *
924 * If colno is in colnos, return the corresponding element of opids,
925 * else return InvalidOid. (Ordinarily colnos would not contain duplicates,
926 * but if it does, we arbitrarily select the first match.)
927 */
928 static Oid
distinct_col_search(int colno,List * colnos,List * opids)929 distinct_col_search(int colno, List *colnos, List *opids)
930 {
931 ListCell *lc1,
932 *lc2;
933
934 forboth(lc1, colnos, lc2, opids)
935 {
936 if (colno == lfirst_int(lc1))
937 return lfirst_oid(lc2);
938 }
939 return InvalidOid;
940 }
941
942
943 /*
944 * innerrel_is_unique
945 * Check if the innerrel provably contains at most one tuple matching any
946 * tuple from the outerrel, based on join clauses in the 'restrictlist'.
947 *
948 * We need an actual RelOptInfo for the innerrel, but it's sufficient to
949 * identify the outerrel by its Relids. This asymmetry supports use of this
950 * function before joinrels have been built.
951 *
952 * The proof must be made based only on clauses that will be "joinquals"
953 * rather than "otherquals" at execution. For an inner join there's no
954 * difference; but if the join is outer, we must ignore pushed-down quals,
955 * as those will become "otherquals". Note that this means the answer might
956 * vary depending on whether IS_OUTER_JOIN(jointype); since we cache the
957 * answer without regard to that, callers must take care not to call this
958 * with jointypes that would be classified differently by IS_OUTER_JOIN().
959 *
960 * The actual proof is undertaken by is_innerrel_unique_for(); this function
961 * is a frontend that is mainly concerned with caching the answers.
962 * In particular, the force_cache argument allows overriding the internal
963 * heuristic about whether to cache negative answers; it should be "true"
964 * if making an inquiry that is not part of the normal bottom-up join search
965 * sequence.
966 */
967 bool
innerrel_is_unique(PlannerInfo * root,Relids outerrelids,RelOptInfo * innerrel,JoinType jointype,List * restrictlist,bool force_cache)968 innerrel_is_unique(PlannerInfo *root,
969 Relids outerrelids,
970 RelOptInfo *innerrel,
971 JoinType jointype,
972 List *restrictlist,
973 bool force_cache)
974 {
975 MemoryContext old_context;
976 ListCell *lc;
977
978 /* Certainly can't prove uniqueness when there are no joinclauses */
979 if (restrictlist == NIL)
980 return false;
981
982 /*
983 * Make a quick check to eliminate cases in which we will surely be unable
984 * to prove uniqueness of the innerrel.
985 */
986 if (!rel_supports_distinctness(root, innerrel))
987 return false;
988
989 /*
990 * Query the cache to see if we've managed to prove that innerrel is
991 * unique for any subset of this outerrel. We don't need an exact match,
992 * as extra outerrels can't make the innerrel any less unique (or more
993 * formally, the restrictlist for a join to a superset outerrel must be a
994 * superset of the conditions we successfully used before).
995 */
996 foreach(lc, innerrel->unique_for_rels)
997 {
998 Relids unique_for_rels = (Relids) lfirst(lc);
999
1000 if (bms_is_subset(unique_for_rels, outerrelids))
1001 return true; /* Success! */
1002 }
1003
1004 /*
1005 * Conversely, we may have already determined that this outerrel, or some
1006 * superset thereof, cannot prove this innerrel to be unique.
1007 */
1008 foreach(lc, innerrel->non_unique_for_rels)
1009 {
1010 Relids unique_for_rels = (Relids) lfirst(lc);
1011
1012 if (bms_is_subset(outerrelids, unique_for_rels))
1013 return false;
1014 }
1015
1016 /* No cached information, so try to make the proof. */
1017 if (is_innerrel_unique_for(root, outerrelids, innerrel,
1018 jointype, restrictlist))
1019 {
1020 /*
1021 * Cache the positive result for future probes, being sure to keep it
1022 * in the planner_cxt even if we are working in GEQO.
1023 *
1024 * Note: one might consider trying to isolate the minimal subset of
1025 * the outerrels that proved the innerrel unique. But it's not worth
1026 * the trouble, because the planner builds up joinrels incrementally
1027 * and so we'll see the minimally sufficient outerrels before any
1028 * supersets of them anyway.
1029 */
1030 old_context = MemoryContextSwitchTo(root->planner_cxt);
1031 innerrel->unique_for_rels = lappend(innerrel->unique_for_rels,
1032 bms_copy(outerrelids));
1033 MemoryContextSwitchTo(old_context);
1034
1035 return true; /* Success! */
1036 }
1037 else
1038 {
1039 /*
1040 * None of the join conditions for outerrel proved innerrel unique, so
1041 * we can safely reject this outerrel or any subset of it in future
1042 * checks.
1043 *
1044 * However, in normal planning mode, caching this knowledge is totally
1045 * pointless; it won't be queried again, because we build up joinrels
1046 * from smaller to larger. It is useful in GEQO mode, where the
1047 * knowledge can be carried across successive planning attempts; and
1048 * it's likely to be useful when using join-search plugins, too. Hence
1049 * cache when join_search_private is non-NULL. (Yeah, that's a hack,
1050 * but it seems reasonable.)
1051 *
1052 * Also, allow callers to override that heuristic and force caching;
1053 * that's useful for reduce_unique_semijoins, which calls here before
1054 * the normal join search starts.
1055 */
1056 if (force_cache || root->join_search_private)
1057 {
1058 old_context = MemoryContextSwitchTo(root->planner_cxt);
1059 innerrel->non_unique_for_rels =
1060 lappend(innerrel->non_unique_for_rels,
1061 bms_copy(outerrelids));
1062 MemoryContextSwitchTo(old_context);
1063 }
1064
1065 return false;
1066 }
1067 }
1068
1069 /*
1070 * is_innerrel_unique_for
1071 * Check if the innerrel provably contains at most one tuple matching any
1072 * tuple from the outerrel, based on join clauses in the 'restrictlist'.
1073 */
1074 static bool
is_innerrel_unique_for(PlannerInfo * root,Relids outerrelids,RelOptInfo * innerrel,JoinType jointype,List * restrictlist)1075 is_innerrel_unique_for(PlannerInfo *root,
1076 Relids outerrelids,
1077 RelOptInfo *innerrel,
1078 JoinType jointype,
1079 List *restrictlist)
1080 {
1081 Relids joinrelids = bms_union(outerrelids, innerrel->relids);
1082 List *clause_list = NIL;
1083 ListCell *lc;
1084
1085 /*
1086 * Search for mergejoinable clauses that constrain the inner rel against
1087 * the outer rel. If an operator is mergejoinable then it behaves like
1088 * equality for some btree opclass, so it's what we want. The
1089 * mergejoinability test also eliminates clauses containing volatile
1090 * functions, which we couldn't depend on.
1091 */
1092 foreach(lc, restrictlist)
1093 {
1094 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1095
1096 /*
1097 * As noted above, if it's a pushed-down clause and we're at an outer
1098 * join, we can't use it.
1099 */
1100 if (IS_OUTER_JOIN(jointype) &&
1101 RINFO_IS_PUSHED_DOWN(restrictinfo, joinrelids))
1102 continue;
1103
1104 /* Ignore if it's not a mergejoinable clause */
1105 if (!restrictinfo->can_join ||
1106 restrictinfo->mergeopfamilies == NIL)
1107 continue; /* not mergejoinable */
1108
1109 /*
1110 * Check if clause has the form "outer op inner" or "inner op outer",
1111 * and if so mark which side is inner.
1112 */
1113 if (!clause_sides_match_join(restrictinfo, outerrelids,
1114 innerrel->relids))
1115 continue; /* no good for these input relations */
1116
1117 /* OK, add to list */
1118 clause_list = lappend(clause_list, restrictinfo);
1119 }
1120
1121 /* Let rel_is_distinct_for() do the hard work */
1122 return rel_is_distinct_for(root, innerrel, clause_list);
1123 }
1124