1 /*-------------------------------------------------------------------------
2 *
3 * equivclass.c
4 * Routines for managing EquivalenceClasses
5 *
6 * See src/backend/optimizer/README for discussion of EquivalenceClasses.
7 *
8 *
9 * Portions Copyright (c) 1996-2018, PostgreSQL Global Development Group
10 * Portions Copyright (c) 1994, Regents of the University of California
11 *
12 * IDENTIFICATION
13 * src/backend/optimizer/path/equivclass.c
14 *
15 *-------------------------------------------------------------------------
16 */
17 #include "postgres.h"
18
19 #include <limits.h>
20
21 #include "access/stratnum.h"
22 #include "catalog/pg_type.h"
23 #include "nodes/makefuncs.h"
24 #include "nodes/nodeFuncs.h"
25 #include "optimizer/clauses.h"
26 #include "optimizer/pathnode.h"
27 #include "optimizer/paths.h"
28 #include "optimizer/planmain.h"
29 #include "optimizer/prep.h"
30 #include "optimizer/restrictinfo.h"
31 #include "optimizer/var.h"
32 #include "utils/lsyscache.h"
33
34
35 static EquivalenceMember *add_eq_member(EquivalenceClass *ec,
36 Expr *expr, Relids relids, Relids nullable_relids,
37 bool is_child, Oid datatype);
38 static void generate_base_implied_equalities_const(PlannerInfo *root,
39 EquivalenceClass *ec);
40 static void generate_base_implied_equalities_no_const(PlannerInfo *root,
41 EquivalenceClass *ec);
42 static void generate_base_implied_equalities_broken(PlannerInfo *root,
43 EquivalenceClass *ec);
44 static List *generate_join_implied_equalities_normal(PlannerInfo *root,
45 EquivalenceClass *ec,
46 Relids join_relids,
47 Relids outer_relids,
48 Relids inner_relids);
49 static List *generate_join_implied_equalities_broken(PlannerInfo *root,
50 EquivalenceClass *ec,
51 Relids nominal_join_relids,
52 Relids outer_relids,
53 Relids nominal_inner_relids,
54 RelOptInfo *inner_rel);
55 static Oid select_equality_operator(EquivalenceClass *ec,
56 Oid lefttype, Oid righttype);
57 static RestrictInfo *create_join_clause(PlannerInfo *root,
58 EquivalenceClass *ec, Oid opno,
59 EquivalenceMember *leftem,
60 EquivalenceMember *rightem,
61 EquivalenceClass *parent_ec);
62 static bool reconsider_outer_join_clause(PlannerInfo *root,
63 RestrictInfo *rinfo,
64 bool outer_on_left);
65 static bool reconsider_full_join_clause(PlannerInfo *root,
66 RestrictInfo *rinfo);
67
68
69 /*
70 * process_equivalence
71 * The given clause has a mergejoinable operator and can be applied without
72 * any delay by an outer join, so its two sides can be considered equal
73 * anywhere they are both computable; moreover that equality can be
74 * extended transitively. Record this knowledge in the EquivalenceClass
75 * data structure, if applicable. Returns true if successful, false if not
76 * (in which case caller should treat the clause as ordinary, not an
77 * equivalence).
78 *
79 * In some cases, although we cannot convert a clause into EquivalenceClass
80 * knowledge, we can still modify it to a more useful form than the original.
81 * Then, *p_restrictinfo will be replaced by a new RestrictInfo, which is what
82 * the caller should use for further processing.
83 *
84 * If below_outer_join is true, then the clause was found below the nullable
85 * side of an outer join, so its sides might validly be both NULL rather than
86 * strictly equal. We can still deduce equalities in such cases, but we take
87 * care to mark an EquivalenceClass if it came from any such clauses. Also,
88 * we have to check that both sides are either pseudo-constants or strict
89 * functions of Vars, else they might not both go to NULL above the outer
90 * join. (This is the main reason why we need a failure return. It's more
91 * convenient to check this case here than at the call sites...)
92 *
93 * We also reject proposed equivalence clauses if they contain leaky functions
94 * and have security_level above zero. The EC evaluation rules require us to
95 * apply certain tests at certain joining levels, and we can't tolerate
96 * delaying any test on security_level grounds. By rejecting candidate clauses
97 * that might require security delays, we ensure it's safe to apply an EC
98 * clause as soon as it's supposed to be applied.
99 *
100 * On success return, we have also initialized the clause's left_ec/right_ec
101 * fields to point to the EquivalenceClass representing it. This saves lookup
102 * effort later.
103 *
104 * Note: constructing merged EquivalenceClasses is a standard UNION-FIND
105 * problem, for which there exist better data structures than simple lists.
106 * If this code ever proves to be a bottleneck then it could be sped up ---
107 * but for now, simple is beautiful.
108 *
109 * Note: this is only called during planner startup, not during GEQO
110 * exploration, so we need not worry about whether we're in the right
111 * memory context.
112 */
113 bool
process_equivalence(PlannerInfo * root,RestrictInfo ** p_restrictinfo,bool below_outer_join)114 process_equivalence(PlannerInfo *root,
115 RestrictInfo **p_restrictinfo,
116 bool below_outer_join)
117 {
118 RestrictInfo *restrictinfo = *p_restrictinfo;
119 Expr *clause = restrictinfo->clause;
120 Oid opno,
121 collation,
122 item1_type,
123 item2_type;
124 Expr *item1;
125 Expr *item2;
126 Relids item1_relids,
127 item2_relids,
128 item1_nullable_relids,
129 item2_nullable_relids;
130 List *opfamilies;
131 EquivalenceClass *ec1,
132 *ec2;
133 EquivalenceMember *em1,
134 *em2;
135 ListCell *lc1;
136
137 /* Should not already be marked as having generated an eclass */
138 Assert(restrictinfo->left_ec == NULL);
139 Assert(restrictinfo->right_ec == NULL);
140
141 /* Reject if it is potentially postponable by security considerations */
142 if (restrictinfo->security_level > 0 && !restrictinfo->leakproof)
143 return false;
144
145 /* Extract info from given clause */
146 Assert(is_opclause(clause));
147 opno = ((OpExpr *) clause)->opno;
148 collation = ((OpExpr *) clause)->inputcollid;
149 item1 = (Expr *) get_leftop(clause);
150 item2 = (Expr *) get_rightop(clause);
151 item1_relids = restrictinfo->left_relids;
152 item2_relids = restrictinfo->right_relids;
153
154 /*
155 * Ensure both input expressions expose the desired collation (their types
156 * should be OK already); see comments for canonicalize_ec_expression.
157 */
158 item1 = canonicalize_ec_expression(item1,
159 exprType((Node *) item1),
160 collation);
161 item2 = canonicalize_ec_expression(item2,
162 exprType((Node *) item2),
163 collation);
164
165 /*
166 * Clauses of the form X=X cannot be translated into EquivalenceClasses.
167 * We'd either end up with a single-entry EC, losing the knowledge that
168 * the clause was present at all, or else make an EC with duplicate
169 * entries, causing other issues.
170 */
171 if (equal(item1, item2))
172 {
173 /*
174 * If the operator is strict, then the clause can be treated as just
175 * "X IS NOT NULL". (Since we know we are considering a top-level
176 * qual, we can ignore the difference between FALSE and NULL results.)
177 * It's worth making the conversion because we'll typically get a much
178 * better selectivity estimate than we would for X=X.
179 *
180 * If the operator is not strict, we can't be sure what it will do
181 * with NULLs, so don't attempt to optimize it.
182 */
183 set_opfuncid((OpExpr *) clause);
184 if (func_strict(((OpExpr *) clause)->opfuncid))
185 {
186 NullTest *ntest = makeNode(NullTest);
187
188 ntest->arg = item1;
189 ntest->nulltesttype = IS_NOT_NULL;
190 ntest->argisrow = false; /* correct even if composite arg */
191 ntest->location = -1;
192
193 *p_restrictinfo =
194 make_restrictinfo((Expr *) ntest,
195 restrictinfo->is_pushed_down,
196 restrictinfo->outerjoin_delayed,
197 restrictinfo->pseudoconstant,
198 restrictinfo->security_level,
199 NULL,
200 restrictinfo->outer_relids,
201 restrictinfo->nullable_relids);
202 }
203 return false;
204 }
205
206 /*
207 * If below outer join, check for strictness, else reject.
208 */
209 if (below_outer_join)
210 {
211 if (!bms_is_empty(item1_relids) &&
212 contain_nonstrict_functions((Node *) item1))
213 return false; /* LHS is non-strict but not constant */
214 if (!bms_is_empty(item2_relids) &&
215 contain_nonstrict_functions((Node *) item2))
216 return false; /* RHS is non-strict but not constant */
217 }
218
219 /* Calculate nullable-relid sets for each side of the clause */
220 item1_nullable_relids = bms_intersect(item1_relids,
221 restrictinfo->nullable_relids);
222 item2_nullable_relids = bms_intersect(item2_relids,
223 restrictinfo->nullable_relids);
224
225 /*
226 * We use the declared input types of the operator, not exprType() of the
227 * inputs, as the nominal datatypes for opfamily lookup. This presumes
228 * that btree operators are always registered with amoplefttype and
229 * amoprighttype equal to their declared input types. We will need this
230 * info anyway to build EquivalenceMember nodes, and by extracting it now
231 * we can use type comparisons to short-circuit some equal() tests.
232 */
233 op_input_types(opno, &item1_type, &item2_type);
234
235 opfamilies = restrictinfo->mergeopfamilies;
236
237 /*
238 * Sweep through the existing EquivalenceClasses looking for matches to
239 * item1 and item2. These are the possible outcomes:
240 *
241 * 1. We find both in the same EC. The equivalence is already known, so
242 * there's nothing to do.
243 *
244 * 2. We find both in different ECs. Merge the two ECs together.
245 *
246 * 3. We find just one. Add the other to its EC.
247 *
248 * 4. We find neither. Make a new, two-entry EC.
249 *
250 * Note: since all ECs are built through this process or the similar
251 * search in get_eclass_for_sort_expr(), it's impossible that we'd match
252 * an item in more than one existing nonvolatile EC. So it's okay to stop
253 * at the first match.
254 */
255 ec1 = ec2 = NULL;
256 em1 = em2 = NULL;
257 foreach(lc1, root->eq_classes)
258 {
259 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
260 ListCell *lc2;
261
262 /* Never match to a volatile EC */
263 if (cur_ec->ec_has_volatile)
264 continue;
265
266 /*
267 * The collation has to match; check this first since it's cheaper
268 * than the opfamily comparison.
269 */
270 if (collation != cur_ec->ec_collation)
271 continue;
272
273 /*
274 * A "match" requires matching sets of btree opfamilies. Use of
275 * equal() for this test has implications discussed in the comments
276 * for get_mergejoin_opfamilies().
277 */
278 if (!equal(opfamilies, cur_ec->ec_opfamilies))
279 continue;
280
281 foreach(lc2, cur_ec->ec_members)
282 {
283 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
284
285 Assert(!cur_em->em_is_child); /* no children yet */
286
287 /*
288 * If below an outer join, don't match constants: they're not as
289 * constant as they look.
290 */
291 if ((below_outer_join || cur_ec->ec_below_outer_join) &&
292 cur_em->em_is_const)
293 continue;
294
295 if (!ec1 &&
296 item1_type == cur_em->em_datatype &&
297 equal(item1, cur_em->em_expr))
298 {
299 ec1 = cur_ec;
300 em1 = cur_em;
301 if (ec2)
302 break;
303 }
304
305 if (!ec2 &&
306 item2_type == cur_em->em_datatype &&
307 equal(item2, cur_em->em_expr))
308 {
309 ec2 = cur_ec;
310 em2 = cur_em;
311 if (ec1)
312 break;
313 }
314 }
315
316 if (ec1 && ec2)
317 break;
318 }
319
320 /* Sweep finished, what did we find? */
321
322 if (ec1 && ec2)
323 {
324 /* If case 1, nothing to do, except add to sources */
325 if (ec1 == ec2)
326 {
327 ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
328 ec1->ec_below_outer_join |= below_outer_join;
329 ec1->ec_min_security = Min(ec1->ec_min_security,
330 restrictinfo->security_level);
331 ec1->ec_max_security = Max(ec1->ec_max_security,
332 restrictinfo->security_level);
333 /* mark the RI as associated with this eclass */
334 restrictinfo->left_ec = ec1;
335 restrictinfo->right_ec = ec1;
336 /* mark the RI as usable with this pair of EMs */
337 restrictinfo->left_em = em1;
338 restrictinfo->right_em = em2;
339 return true;
340 }
341
342 /*
343 * Case 2: need to merge ec1 and ec2. This should never happen after
344 * we've built any canonical pathkeys; if it did, those pathkeys might
345 * be rendered non-canonical by the merge.
346 */
347 if (root->canon_pathkeys != NIL)
348 elog(ERROR, "too late to merge equivalence classes");
349
350 /*
351 * We add ec2's items to ec1, then set ec2's ec_merged link to point
352 * to ec1 and remove ec2 from the eq_classes list. We cannot simply
353 * delete ec2 because that could leave dangling pointers in existing
354 * PathKeys. We leave it behind with a link so that the merged EC can
355 * be found.
356 */
357 ec1->ec_members = list_concat(ec1->ec_members, ec2->ec_members);
358 ec1->ec_sources = list_concat(ec1->ec_sources, ec2->ec_sources);
359 ec1->ec_derives = list_concat(ec1->ec_derives, ec2->ec_derives);
360 ec1->ec_relids = bms_join(ec1->ec_relids, ec2->ec_relids);
361 ec1->ec_has_const |= ec2->ec_has_const;
362 /* can't need to set has_volatile */
363 ec1->ec_below_outer_join |= ec2->ec_below_outer_join;
364 ec1->ec_min_security = Min(ec1->ec_min_security,
365 ec2->ec_min_security);
366 ec1->ec_max_security = Max(ec1->ec_max_security,
367 ec2->ec_max_security);
368 ec2->ec_merged = ec1;
369 root->eq_classes = list_delete_ptr(root->eq_classes, ec2);
370 /* just to avoid debugging confusion w/ dangling pointers: */
371 ec2->ec_members = NIL;
372 ec2->ec_sources = NIL;
373 ec2->ec_derives = NIL;
374 ec2->ec_relids = NULL;
375 ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
376 ec1->ec_below_outer_join |= below_outer_join;
377 ec1->ec_min_security = Min(ec1->ec_min_security,
378 restrictinfo->security_level);
379 ec1->ec_max_security = Max(ec1->ec_max_security,
380 restrictinfo->security_level);
381 /* mark the RI as associated with this eclass */
382 restrictinfo->left_ec = ec1;
383 restrictinfo->right_ec = ec1;
384 /* mark the RI as usable with this pair of EMs */
385 restrictinfo->left_em = em1;
386 restrictinfo->right_em = em2;
387 }
388 else if (ec1)
389 {
390 /* Case 3: add item2 to ec1 */
391 em2 = add_eq_member(ec1, item2, item2_relids, item2_nullable_relids,
392 false, item2_type);
393 ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
394 ec1->ec_below_outer_join |= below_outer_join;
395 ec1->ec_min_security = Min(ec1->ec_min_security,
396 restrictinfo->security_level);
397 ec1->ec_max_security = Max(ec1->ec_max_security,
398 restrictinfo->security_level);
399 /* mark the RI as associated with this eclass */
400 restrictinfo->left_ec = ec1;
401 restrictinfo->right_ec = ec1;
402 /* mark the RI as usable with this pair of EMs */
403 restrictinfo->left_em = em1;
404 restrictinfo->right_em = em2;
405 }
406 else if (ec2)
407 {
408 /* Case 3: add item1 to ec2 */
409 em1 = add_eq_member(ec2, item1, item1_relids, item1_nullable_relids,
410 false, item1_type);
411 ec2->ec_sources = lappend(ec2->ec_sources, restrictinfo);
412 ec2->ec_below_outer_join |= below_outer_join;
413 ec2->ec_min_security = Min(ec2->ec_min_security,
414 restrictinfo->security_level);
415 ec2->ec_max_security = Max(ec2->ec_max_security,
416 restrictinfo->security_level);
417 /* mark the RI as associated with this eclass */
418 restrictinfo->left_ec = ec2;
419 restrictinfo->right_ec = ec2;
420 /* mark the RI as usable with this pair of EMs */
421 restrictinfo->left_em = em1;
422 restrictinfo->right_em = em2;
423 }
424 else
425 {
426 /* Case 4: make a new, two-entry EC */
427 EquivalenceClass *ec = makeNode(EquivalenceClass);
428
429 ec->ec_opfamilies = opfamilies;
430 ec->ec_collation = collation;
431 ec->ec_members = NIL;
432 ec->ec_sources = list_make1(restrictinfo);
433 ec->ec_derives = NIL;
434 ec->ec_relids = NULL;
435 ec->ec_has_const = false;
436 ec->ec_has_volatile = false;
437 ec->ec_below_outer_join = below_outer_join;
438 ec->ec_broken = false;
439 ec->ec_sortref = 0;
440 ec->ec_min_security = restrictinfo->security_level;
441 ec->ec_max_security = restrictinfo->security_level;
442 ec->ec_merged = NULL;
443 em1 = add_eq_member(ec, item1, item1_relids, item1_nullable_relids,
444 false, item1_type);
445 em2 = add_eq_member(ec, item2, item2_relids, item2_nullable_relids,
446 false, item2_type);
447
448 root->eq_classes = lappend(root->eq_classes, ec);
449
450 /* mark the RI as associated with this eclass */
451 restrictinfo->left_ec = ec;
452 restrictinfo->right_ec = ec;
453 /* mark the RI as usable with this pair of EMs */
454 restrictinfo->left_em = em1;
455 restrictinfo->right_em = em2;
456 }
457
458 return true;
459 }
460
461 /*
462 * canonicalize_ec_expression
463 *
464 * This function ensures that the expression exposes the expected type and
465 * collation, so that it will be equal() to other equivalence-class expressions
466 * that it ought to be equal() to.
467 *
468 * The rule for datatypes is that the exposed type should match what it would
469 * be for an input to an operator of the EC's opfamilies; which is usually
470 * the declared input type of the operator, but in the case of polymorphic
471 * operators no relabeling is wanted (compare the behavior of parse_coerce.c).
472 * Expressions coming in from quals will generally have the right type
473 * already, but expressions coming from indexkeys may not (because they are
474 * represented without any explicit relabel in pg_index), and the same problem
475 * occurs for sort expressions (because the parser is likewise cavalier about
476 * putting relabels on them). Such cases will be binary-compatible with the
477 * real operators, so adding a RelabelType is sufficient.
478 *
479 * Also, the expression's exposed collation must match the EC's collation.
480 * This is important because in comparisons like "foo < bar COLLATE baz",
481 * only one of the expressions has the correct exposed collation as we receive
482 * it from the parser. Forcing both of them to have it ensures that all
483 * variant spellings of such a construct behave the same. Again, we can
484 * stick on a RelabelType to force the right exposed collation. (It might
485 * work to not label the collation at all in EC members, but this is risky
486 * since some parts of the system expect exprCollation() to deliver the
487 * right answer for a sort key.)
488 *
489 * Note this code assumes that the expression has already been through
490 * eval_const_expressions, so there are no CollateExprs and no redundant
491 * RelabelTypes.
492 */
493 Expr *
canonicalize_ec_expression(Expr * expr,Oid req_type,Oid req_collation)494 canonicalize_ec_expression(Expr *expr, Oid req_type, Oid req_collation)
495 {
496 Oid expr_type = exprType((Node *) expr);
497
498 /*
499 * For a polymorphic-input-type opclass, just keep the same exposed type.
500 * RECORD opclasses work like polymorphic-type ones for this purpose.
501 */
502 if (IsPolymorphicType(req_type) || req_type == RECORDOID)
503 req_type = expr_type;
504
505 /*
506 * No work if the expression exposes the right type/collation already.
507 */
508 if (expr_type != req_type ||
509 exprCollation((Node *) expr) != req_collation)
510 {
511 /*
512 * Strip any existing RelabelType, then add a new one if needed. This
513 * is to preserve the invariant of no redundant RelabelTypes.
514 *
515 * If we have to change the exposed type of the stripped expression,
516 * set typmod to -1 (since the new type may not have the same typmod
517 * interpretation). If we only have to change collation, preserve the
518 * exposed typmod.
519 */
520 while (expr && IsA(expr, RelabelType))
521 expr = (Expr *) ((RelabelType *) expr)->arg;
522
523 if (exprType((Node *) expr) != req_type)
524 expr = (Expr *) makeRelabelType(expr,
525 req_type,
526 -1,
527 req_collation,
528 COERCE_IMPLICIT_CAST);
529 else if (exprCollation((Node *) expr) != req_collation)
530 expr = (Expr *) makeRelabelType(expr,
531 req_type,
532 exprTypmod((Node *) expr),
533 req_collation,
534 COERCE_IMPLICIT_CAST);
535 }
536
537 return expr;
538 }
539
540 /*
541 * add_eq_member - build a new EquivalenceMember and add it to an EC
542 */
543 static EquivalenceMember *
add_eq_member(EquivalenceClass * ec,Expr * expr,Relids relids,Relids nullable_relids,bool is_child,Oid datatype)544 add_eq_member(EquivalenceClass *ec, Expr *expr, Relids relids,
545 Relids nullable_relids, bool is_child, Oid datatype)
546 {
547 EquivalenceMember *em = makeNode(EquivalenceMember);
548
549 em->em_expr = expr;
550 em->em_relids = relids;
551 em->em_nullable_relids = nullable_relids;
552 em->em_is_const = false;
553 em->em_is_child = is_child;
554 em->em_datatype = datatype;
555
556 if (bms_is_empty(relids))
557 {
558 /*
559 * No Vars, assume it's a pseudoconstant. This is correct for entries
560 * generated from process_equivalence(), because a WHERE clause can't
561 * contain aggregates or SRFs, and non-volatility was checked before
562 * process_equivalence() ever got called. But
563 * get_eclass_for_sort_expr() has to work harder. We put the tests
564 * there not here to save cycles in the equivalence case.
565 */
566 Assert(!is_child);
567 em->em_is_const = true;
568 ec->ec_has_const = true;
569 /* it can't affect ec_relids */
570 }
571 else if (!is_child) /* child members don't add to ec_relids */
572 {
573 ec->ec_relids = bms_add_members(ec->ec_relids, relids);
574 }
575 ec->ec_members = lappend(ec->ec_members, em);
576
577 return em;
578 }
579
580
581 /*
582 * get_eclass_for_sort_expr
583 * Given an expression and opfamily/collation info, find an existing
584 * equivalence class it is a member of; if none, optionally build a new
585 * single-member EquivalenceClass for it.
586 *
587 * expr is the expression, and nullable_relids is the set of base relids
588 * that are potentially nullable below it. We actually only care about
589 * the set of such relids that are used in the expression; but for caller
590 * convenience, we perform that intersection step here. The caller need
591 * only be sure that nullable_relids doesn't omit any nullable rels that
592 * might appear in the expr.
593 *
594 * sortref is the SortGroupRef of the originating SortGroupClause, if any,
595 * or zero if not. (It should never be zero if the expression is volatile!)
596 *
597 * If rel is not NULL, it identifies a specific relation we're considering
598 * a path for, and indicates that child EC members for that relation can be
599 * considered. Otherwise child members are ignored. (Note: since child EC
600 * members aren't guaranteed unique, a non-NULL value means that there could
601 * be more than one EC that matches the expression; if so it's order-dependent
602 * which one you get. This is annoying but it only happens in corner cases,
603 * so for now we live with just reporting the first match. See also
604 * generate_implied_equalities_for_column and match_pathkeys_to_index.)
605 *
606 * If create_it is true, we'll build a new EquivalenceClass when there is no
607 * match. If create_it is false, we just return NULL when no match.
608 *
609 * This can be used safely both before and after EquivalenceClass merging;
610 * since it never causes merging it does not invalidate any existing ECs
611 * or PathKeys. However, ECs added after path generation has begun are
612 * of limited usefulness, so usually it's best to create them beforehand.
613 *
614 * Note: opfamilies must be chosen consistently with the way
615 * process_equivalence() would do; that is, generated from a mergejoinable
616 * equality operator. Else we might fail to detect valid equivalences,
617 * generating poor (but not incorrect) plans.
618 */
619 EquivalenceClass *
get_eclass_for_sort_expr(PlannerInfo * root,Expr * expr,Relids nullable_relids,List * opfamilies,Oid opcintype,Oid collation,Index sortref,Relids rel,bool create_it)620 get_eclass_for_sort_expr(PlannerInfo *root,
621 Expr *expr,
622 Relids nullable_relids,
623 List *opfamilies,
624 Oid opcintype,
625 Oid collation,
626 Index sortref,
627 Relids rel,
628 bool create_it)
629 {
630 Relids expr_relids;
631 EquivalenceClass *newec;
632 EquivalenceMember *newem;
633 ListCell *lc1;
634 MemoryContext oldcontext;
635
636 /*
637 * Ensure the expression exposes the correct type and collation.
638 */
639 expr = canonicalize_ec_expression(expr, opcintype, collation);
640
641 /*
642 * Scan through the existing EquivalenceClasses for a match
643 */
644 foreach(lc1, root->eq_classes)
645 {
646 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
647 ListCell *lc2;
648
649 /*
650 * Never match to a volatile EC, except when we are looking at another
651 * reference to the same volatile SortGroupClause.
652 */
653 if (cur_ec->ec_has_volatile &&
654 (sortref == 0 || sortref != cur_ec->ec_sortref))
655 continue;
656
657 if (collation != cur_ec->ec_collation)
658 continue;
659 if (!equal(opfamilies, cur_ec->ec_opfamilies))
660 continue;
661
662 foreach(lc2, cur_ec->ec_members)
663 {
664 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
665
666 /*
667 * Ignore child members unless they match the request.
668 */
669 if (cur_em->em_is_child &&
670 !bms_equal(cur_em->em_relids, rel))
671 continue;
672
673 /*
674 * If below an outer join, don't match constants: they're not as
675 * constant as they look.
676 */
677 if (cur_ec->ec_below_outer_join &&
678 cur_em->em_is_const)
679 continue;
680
681 if (opcintype == cur_em->em_datatype &&
682 equal(expr, cur_em->em_expr))
683 return cur_ec; /* Match! */
684 }
685 }
686
687 /* No match; does caller want a NULL result? */
688 if (!create_it)
689 return NULL;
690
691 /*
692 * OK, build a new single-member EC
693 *
694 * Here, we must be sure that we construct the EC in the right context.
695 */
696 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
697
698 newec = makeNode(EquivalenceClass);
699 newec->ec_opfamilies = list_copy(opfamilies);
700 newec->ec_collation = collation;
701 newec->ec_members = NIL;
702 newec->ec_sources = NIL;
703 newec->ec_derives = NIL;
704 newec->ec_relids = NULL;
705 newec->ec_has_const = false;
706 newec->ec_has_volatile = contain_volatile_functions((Node *) expr);
707 newec->ec_below_outer_join = false;
708 newec->ec_broken = false;
709 newec->ec_sortref = sortref;
710 newec->ec_min_security = UINT_MAX;
711 newec->ec_max_security = 0;
712 newec->ec_merged = NULL;
713
714 if (newec->ec_has_volatile && sortref == 0) /* should not happen */
715 elog(ERROR, "volatile EquivalenceClass has no sortref");
716
717 /*
718 * Get the precise set of nullable relids appearing in the expression.
719 */
720 expr_relids = pull_varnos((Node *) expr);
721 nullable_relids = bms_intersect(nullable_relids, expr_relids);
722
723 newem = add_eq_member(newec, copyObject(expr), expr_relids,
724 nullable_relids, false, opcintype);
725
726 /*
727 * add_eq_member doesn't check for volatile functions, set-returning
728 * functions, aggregates, or window functions, but such could appear in
729 * sort expressions; so we have to check whether its const-marking was
730 * correct.
731 */
732 if (newec->ec_has_const)
733 {
734 if (newec->ec_has_volatile ||
735 expression_returns_set((Node *) expr) ||
736 contain_agg_clause((Node *) expr) ||
737 contain_window_function((Node *) expr))
738 {
739 newec->ec_has_const = false;
740 newem->em_is_const = false;
741 }
742 }
743
744 root->eq_classes = lappend(root->eq_classes, newec);
745
746 MemoryContextSwitchTo(oldcontext);
747
748 return newec;
749 }
750
751
752 /*
753 * generate_base_implied_equalities
754 * Generate any restriction clauses that we can deduce from equivalence
755 * classes.
756 *
757 * When an EC contains pseudoconstants, our strategy is to generate
758 * "member = const1" clauses where const1 is the first constant member, for
759 * every other member (including other constants). If we are able to do this
760 * then we don't need any "var = var" comparisons because we've successfully
761 * constrained all the vars at their points of creation. If we fail to
762 * generate any of these clauses due to lack of cross-type operators, we fall
763 * back to the "ec_broken" strategy described below. (XXX if there are
764 * multiple constants of different types, it's possible that we might succeed
765 * in forming all the required clauses if we started from a different const
766 * member; but this seems a sufficiently hokey corner case to not be worth
767 * spending lots of cycles on.)
768 *
769 * For ECs that contain no pseudoconstants, we generate derived clauses
770 * "member1 = member2" for each pair of members belonging to the same base
771 * relation (actually, if there are more than two for the same base relation,
772 * we only need enough clauses to link each to each other). This provides
773 * the base case for the recursion: each row emitted by a base relation scan
774 * will constrain all computable members of the EC to be equal. As each
775 * join path is formed, we'll add additional derived clauses on-the-fly
776 * to maintain this invariant (see generate_join_implied_equalities).
777 *
778 * If the opfamilies used by the EC do not provide complete sets of cross-type
779 * equality operators, it is possible that we will fail to generate a clause
780 * that must be generated to maintain the invariant. (An example: given
781 * "WHERE a.x = b.y AND b.y = a.z", the scheme breaks down if we cannot
782 * generate "a.x = a.z" as a restriction clause for A.) In this case we mark
783 * the EC "ec_broken" and fall back to regurgitating its original source
784 * RestrictInfos at appropriate times. We do not try to retract any derived
785 * clauses already generated from the broken EC, so the resulting plan could
786 * be poor due to bad selectivity estimates caused by redundant clauses. But
787 * the correct solution to that is to fix the opfamilies ...
788 *
789 * Equality clauses derived by this function are passed off to
790 * process_implied_equality (in plan/initsplan.c) to be inserted into the
791 * restrictinfo datastructures. Note that this must be called after initial
792 * scanning of the quals and before Path construction begins.
793 *
794 * We make no attempt to avoid generating duplicate RestrictInfos here: we
795 * don't search ec_sources for matches, nor put the created RestrictInfos
796 * into ec_derives. Doing so would require some slightly ugly changes in
797 * initsplan.c's API, and there's no real advantage, because the clauses
798 * generated here can't duplicate anything we will generate for joins anyway.
799 */
800 void
generate_base_implied_equalities(PlannerInfo * root)801 generate_base_implied_equalities(PlannerInfo *root)
802 {
803 ListCell *lc;
804 Index rti;
805
806 foreach(lc, root->eq_classes)
807 {
808 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
809
810 Assert(ec->ec_merged == NULL); /* else shouldn't be in list */
811 Assert(!ec->ec_broken); /* not yet anyway... */
812
813 /* Single-member ECs won't generate any deductions */
814 if (list_length(ec->ec_members) <= 1)
815 continue;
816
817 if (ec->ec_has_const)
818 generate_base_implied_equalities_const(root, ec);
819 else
820 generate_base_implied_equalities_no_const(root, ec);
821
822 /* Recover if we failed to generate required derived clauses */
823 if (ec->ec_broken)
824 generate_base_implied_equalities_broken(root, ec);
825 }
826
827 /*
828 * This is also a handy place to mark base rels (which should all exist by
829 * now) with flags showing whether they have pending eclass joins.
830 */
831 for (rti = 1; rti < root->simple_rel_array_size; rti++)
832 {
833 RelOptInfo *brel = root->simple_rel_array[rti];
834
835 if (brel == NULL)
836 continue;
837
838 brel->has_eclass_joins = has_relevant_eclass_joinclause(root, brel);
839 }
840 }
841
842 /*
843 * generate_base_implied_equalities when EC contains pseudoconstant(s)
844 */
845 static void
generate_base_implied_equalities_const(PlannerInfo * root,EquivalenceClass * ec)846 generate_base_implied_equalities_const(PlannerInfo *root,
847 EquivalenceClass *ec)
848 {
849 EquivalenceMember *const_em = NULL;
850 ListCell *lc;
851
852 /*
853 * In the trivial case where we just had one "var = const" clause, push
854 * the original clause back into the main planner machinery. There is
855 * nothing to be gained by doing it differently, and we save the effort to
856 * re-build and re-analyze an equality clause that will be exactly
857 * equivalent to the old one.
858 */
859 if (list_length(ec->ec_members) == 2 &&
860 list_length(ec->ec_sources) == 1)
861 {
862 RestrictInfo *restrictinfo = (RestrictInfo *) linitial(ec->ec_sources);
863
864 if (bms_membership(restrictinfo->required_relids) != BMS_MULTIPLE)
865 {
866 distribute_restrictinfo_to_rels(root, restrictinfo);
867 return;
868 }
869 }
870
871 /*
872 * Find the constant member to use. We prefer an actual constant to
873 * pseudo-constants (such as Params), because the constraint exclusion
874 * machinery might be able to exclude relations on the basis of generated
875 * "var = const" equalities, but "var = param" won't work for that.
876 */
877 foreach(lc, ec->ec_members)
878 {
879 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
880
881 if (cur_em->em_is_const)
882 {
883 const_em = cur_em;
884 if (IsA(cur_em->em_expr, Const))
885 break;
886 }
887 }
888 Assert(const_em != NULL);
889
890 /* Generate a derived equality against each other member */
891 foreach(lc, ec->ec_members)
892 {
893 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
894 Oid eq_op;
895
896 Assert(!cur_em->em_is_child); /* no children yet */
897 if (cur_em == const_em)
898 continue;
899 eq_op = select_equality_operator(ec,
900 cur_em->em_datatype,
901 const_em->em_datatype);
902 if (!OidIsValid(eq_op))
903 {
904 /* failed... */
905 ec->ec_broken = true;
906 break;
907 }
908 process_implied_equality(root, eq_op, ec->ec_collation,
909 cur_em->em_expr, const_em->em_expr,
910 bms_copy(ec->ec_relids),
911 bms_union(cur_em->em_nullable_relids,
912 const_em->em_nullable_relids),
913 ec->ec_min_security,
914 ec->ec_below_outer_join,
915 cur_em->em_is_const);
916 }
917 }
918
919 /*
920 * generate_base_implied_equalities when EC contains no pseudoconstants
921 */
922 static void
generate_base_implied_equalities_no_const(PlannerInfo * root,EquivalenceClass * ec)923 generate_base_implied_equalities_no_const(PlannerInfo *root,
924 EquivalenceClass *ec)
925 {
926 EquivalenceMember **prev_ems;
927 ListCell *lc;
928
929 /*
930 * We scan the EC members once and track the last-seen member for each
931 * base relation. When we see another member of the same base relation,
932 * we generate "prev_mem = cur_mem". This results in the minimum number
933 * of derived clauses, but it's possible that it will fail when a
934 * different ordering would succeed. XXX FIXME: use a UNION-FIND
935 * algorithm similar to the way we build merged ECs. (Use a list-of-lists
936 * for each rel.)
937 */
938 prev_ems = (EquivalenceMember **)
939 palloc0(root->simple_rel_array_size * sizeof(EquivalenceMember *));
940
941 foreach(lc, ec->ec_members)
942 {
943 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
944 int relid;
945
946 Assert(!cur_em->em_is_child); /* no children yet */
947 if (!bms_get_singleton_member(cur_em->em_relids, &relid))
948 continue;
949 Assert(relid < root->simple_rel_array_size);
950
951 if (prev_ems[relid] != NULL)
952 {
953 EquivalenceMember *prev_em = prev_ems[relid];
954 Oid eq_op;
955
956 eq_op = select_equality_operator(ec,
957 prev_em->em_datatype,
958 cur_em->em_datatype);
959 if (!OidIsValid(eq_op))
960 {
961 /* failed... */
962 ec->ec_broken = true;
963 break;
964 }
965 process_implied_equality(root, eq_op, ec->ec_collation,
966 prev_em->em_expr, cur_em->em_expr,
967 bms_copy(ec->ec_relids),
968 bms_union(prev_em->em_nullable_relids,
969 cur_em->em_nullable_relids),
970 ec->ec_min_security,
971 ec->ec_below_outer_join,
972 false);
973 }
974 prev_ems[relid] = cur_em;
975 }
976
977 pfree(prev_ems);
978
979 /*
980 * We also have to make sure that all the Vars used in the member clauses
981 * will be available at any join node we might try to reference them at.
982 * For the moment we force all the Vars to be available at all join nodes
983 * for this eclass. Perhaps this could be improved by doing some
984 * pre-analysis of which members we prefer to join, but it's no worse than
985 * what happened in the pre-8.3 code.
986 */
987 foreach(lc, ec->ec_members)
988 {
989 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
990 List *vars = pull_var_clause((Node *) cur_em->em_expr,
991 PVC_RECURSE_AGGREGATES |
992 PVC_RECURSE_WINDOWFUNCS |
993 PVC_INCLUDE_PLACEHOLDERS);
994
995 add_vars_to_targetlist(root, vars, ec->ec_relids, false);
996 list_free(vars);
997 }
998 }
999
1000 /*
1001 * generate_base_implied_equalities cleanup after failure
1002 *
1003 * What we must do here is push any zero- or one-relation source RestrictInfos
1004 * of the EC back into the main restrictinfo datastructures. Multi-relation
1005 * clauses will be regurgitated later by generate_join_implied_equalities().
1006 * (We do it this way to maintain continuity with the case that ec_broken
1007 * becomes set only after we've gone up a join level or two.) However, for
1008 * an EC that contains constants, we can adopt a simpler strategy and just
1009 * throw back all the source RestrictInfos immediately; that works because
1010 * we know that such an EC can't become broken later. (This rule justifies
1011 * ignoring ec_has_const ECs in generate_join_implied_equalities, even when
1012 * they are broken.)
1013 */
1014 static void
generate_base_implied_equalities_broken(PlannerInfo * root,EquivalenceClass * ec)1015 generate_base_implied_equalities_broken(PlannerInfo *root,
1016 EquivalenceClass *ec)
1017 {
1018 ListCell *lc;
1019
1020 foreach(lc, ec->ec_sources)
1021 {
1022 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1023
1024 if (ec->ec_has_const ||
1025 bms_membership(restrictinfo->required_relids) != BMS_MULTIPLE)
1026 distribute_restrictinfo_to_rels(root, restrictinfo);
1027 }
1028 }
1029
1030
1031 /*
1032 * generate_join_implied_equalities
1033 * Generate any join clauses that we can deduce from equivalence classes.
1034 *
1035 * At a join node, we must enforce restriction clauses sufficient to ensure
1036 * that all equivalence-class members computable at that node are equal.
1037 * Since the set of clauses to enforce can vary depending on which subset
1038 * relations are the inputs, we have to compute this afresh for each join
1039 * relation pair. Hence a fresh List of RestrictInfo nodes is built and
1040 * passed back on each call.
1041 *
1042 * In addition to its use at join nodes, this can be applied to generate
1043 * eclass-based join clauses for use in a parameterized scan of a base rel.
1044 * The reason for the asymmetry of specifying the inner rel as a RelOptInfo
1045 * and the outer rel by Relids is that this usage occurs before we have
1046 * built any join RelOptInfos.
1047 *
1048 * An annoying special case for parameterized scans is that the inner rel can
1049 * be an appendrel child (an "other rel"). In this case we must generate
1050 * appropriate clauses using child EC members. add_child_rel_equivalences
1051 * must already have been done for the child rel.
1052 *
1053 * The results are sufficient for use in merge, hash, and plain nestloop join
1054 * methods. We do not worry here about selecting clauses that are optimal
1055 * for use in a parameterized indexscan. indxpath.c makes its own selections
1056 * of clauses to use, and if the ones we pick here are redundant with those,
1057 * the extras will be eliminated at createplan time, using the parent_ec
1058 * markers that we provide (see is_redundant_derived_clause()).
1059 *
1060 * Because the same join clauses are likely to be needed multiple times as
1061 * we consider different join paths, we avoid generating multiple copies:
1062 * whenever we select a particular pair of EquivalenceMembers to join,
1063 * we check to see if the pair matches any original clause (in ec_sources)
1064 * or previously-built clause (in ec_derives). This saves memory and allows
1065 * re-use of information cached in RestrictInfos.
1066 *
1067 * join_relids should always equal bms_union(outer_relids, inner_rel->relids).
1068 * We could simplify this function's API by computing it internally, but in
1069 * most current uses, the caller has the value at hand anyway.
1070 */
1071 List *
generate_join_implied_equalities(PlannerInfo * root,Relids join_relids,Relids outer_relids,RelOptInfo * inner_rel)1072 generate_join_implied_equalities(PlannerInfo *root,
1073 Relids join_relids,
1074 Relids outer_relids,
1075 RelOptInfo *inner_rel)
1076 {
1077 return generate_join_implied_equalities_for_ecs(root,
1078 root->eq_classes,
1079 join_relids,
1080 outer_relids,
1081 inner_rel);
1082 }
1083
1084 /*
1085 * generate_join_implied_equalities_for_ecs
1086 * As above, but consider only the listed ECs.
1087 */
1088 List *
generate_join_implied_equalities_for_ecs(PlannerInfo * root,List * eclasses,Relids join_relids,Relids outer_relids,RelOptInfo * inner_rel)1089 generate_join_implied_equalities_for_ecs(PlannerInfo *root,
1090 List *eclasses,
1091 Relids join_relids,
1092 Relids outer_relids,
1093 RelOptInfo *inner_rel)
1094 {
1095 List *result = NIL;
1096 Relids inner_relids = inner_rel->relids;
1097 Relids nominal_inner_relids;
1098 Relids nominal_join_relids;
1099 ListCell *lc;
1100
1101 /* If inner rel is a child, extra setup work is needed */
1102 if (IS_OTHER_REL(inner_rel))
1103 {
1104 Assert(!bms_is_empty(inner_rel->top_parent_relids));
1105
1106 /* Fetch relid set for the topmost parent rel */
1107 nominal_inner_relids = inner_rel->top_parent_relids;
1108 /* ECs will be marked with the parent's relid, not the child's */
1109 nominal_join_relids = bms_union(outer_relids, nominal_inner_relids);
1110 }
1111 else
1112 {
1113 nominal_inner_relids = inner_relids;
1114 nominal_join_relids = join_relids;
1115 }
1116
1117 foreach(lc, eclasses)
1118 {
1119 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
1120 List *sublist = NIL;
1121
1122 /* ECs containing consts do not need any further enforcement */
1123 if (ec->ec_has_const)
1124 continue;
1125
1126 /* Single-member ECs won't generate any deductions */
1127 if (list_length(ec->ec_members) <= 1)
1128 continue;
1129
1130 /* We can quickly ignore any that don't overlap the join, too */
1131 if (!bms_overlap(ec->ec_relids, nominal_join_relids))
1132 continue;
1133
1134 if (!ec->ec_broken)
1135 sublist = generate_join_implied_equalities_normal(root,
1136 ec,
1137 join_relids,
1138 outer_relids,
1139 inner_relids);
1140
1141 /* Recover if we failed to generate required derived clauses */
1142 if (ec->ec_broken)
1143 sublist = generate_join_implied_equalities_broken(root,
1144 ec,
1145 nominal_join_relids,
1146 outer_relids,
1147 nominal_inner_relids,
1148 inner_rel);
1149
1150 result = list_concat(result, sublist);
1151 }
1152
1153 return result;
1154 }
1155
1156 /*
1157 * generate_join_implied_equalities for a still-valid EC
1158 */
1159 static List *
generate_join_implied_equalities_normal(PlannerInfo * root,EquivalenceClass * ec,Relids join_relids,Relids outer_relids,Relids inner_relids)1160 generate_join_implied_equalities_normal(PlannerInfo *root,
1161 EquivalenceClass *ec,
1162 Relids join_relids,
1163 Relids outer_relids,
1164 Relids inner_relids)
1165 {
1166 List *result = NIL;
1167 List *new_members = NIL;
1168 List *outer_members = NIL;
1169 List *inner_members = NIL;
1170 ListCell *lc1;
1171
1172 /*
1173 * First, scan the EC to identify member values that are computable at the
1174 * outer rel, at the inner rel, or at this relation but not in either
1175 * input rel. The outer-rel members should already be enforced equal,
1176 * likewise for the inner-rel members. We'll need to create clauses to
1177 * enforce that any newly computable members are all equal to each other
1178 * as well as to at least one input member, plus enforce at least one
1179 * outer-rel member equal to at least one inner-rel member.
1180 */
1181 foreach(lc1, ec->ec_members)
1182 {
1183 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc1);
1184
1185 /*
1186 * We don't need to check explicitly for child EC members. This test
1187 * against join_relids will cause them to be ignored except when
1188 * considering a child inner rel, which is what we want.
1189 */
1190 if (!bms_is_subset(cur_em->em_relids, join_relids))
1191 continue; /* not computable yet, or wrong child */
1192
1193 if (bms_is_subset(cur_em->em_relids, outer_relids))
1194 outer_members = lappend(outer_members, cur_em);
1195 else if (bms_is_subset(cur_em->em_relids, inner_relids))
1196 inner_members = lappend(inner_members, cur_em);
1197 else
1198 new_members = lappend(new_members, cur_em);
1199 }
1200
1201 /*
1202 * First, select the joinclause if needed. We can equate any one outer
1203 * member to any one inner member, but we have to find a datatype
1204 * combination for which an opfamily member operator exists. If we have
1205 * choices, we prefer simple Var members (possibly with RelabelType) since
1206 * these are (a) cheapest to compute at runtime and (b) most likely to
1207 * have useful statistics. Also, prefer operators that are also
1208 * hashjoinable.
1209 */
1210 if (outer_members && inner_members)
1211 {
1212 EquivalenceMember *best_outer_em = NULL;
1213 EquivalenceMember *best_inner_em = NULL;
1214 Oid best_eq_op = InvalidOid;
1215 int best_score = -1;
1216 RestrictInfo *rinfo;
1217
1218 foreach(lc1, outer_members)
1219 {
1220 EquivalenceMember *outer_em = (EquivalenceMember *) lfirst(lc1);
1221 ListCell *lc2;
1222
1223 foreach(lc2, inner_members)
1224 {
1225 EquivalenceMember *inner_em = (EquivalenceMember *) lfirst(lc2);
1226 Oid eq_op;
1227 int score;
1228
1229 eq_op = select_equality_operator(ec,
1230 outer_em->em_datatype,
1231 inner_em->em_datatype);
1232 if (!OidIsValid(eq_op))
1233 continue;
1234 score = 0;
1235 if (IsA(outer_em->em_expr, Var) ||
1236 (IsA(outer_em->em_expr, RelabelType) &&
1237 IsA(((RelabelType *) outer_em->em_expr)->arg, Var)))
1238 score++;
1239 if (IsA(inner_em->em_expr, Var) ||
1240 (IsA(inner_em->em_expr, RelabelType) &&
1241 IsA(((RelabelType *) inner_em->em_expr)->arg, Var)))
1242 score++;
1243 if (op_hashjoinable(eq_op,
1244 exprType((Node *) outer_em->em_expr)))
1245 score++;
1246 if (score > best_score)
1247 {
1248 best_outer_em = outer_em;
1249 best_inner_em = inner_em;
1250 best_eq_op = eq_op;
1251 best_score = score;
1252 if (best_score == 3)
1253 break; /* no need to look further */
1254 }
1255 }
1256 if (best_score == 3)
1257 break; /* no need to look further */
1258 }
1259 if (best_score < 0)
1260 {
1261 /* failed... */
1262 ec->ec_broken = true;
1263 return NIL;
1264 }
1265
1266 /*
1267 * Create clause, setting parent_ec to mark it as redundant with other
1268 * joinclauses
1269 */
1270 rinfo = create_join_clause(root, ec, best_eq_op,
1271 best_outer_em, best_inner_em,
1272 ec);
1273
1274 result = lappend(result, rinfo);
1275 }
1276
1277 /*
1278 * Now deal with building restrictions for any expressions that involve
1279 * Vars from both sides of the join. We have to equate all of these to
1280 * each other as well as to at least one old member (if any).
1281 *
1282 * XXX as in generate_base_implied_equalities_no_const, we could be a lot
1283 * smarter here to avoid unnecessary failures in cross-type situations.
1284 * For now, use the same left-to-right method used there.
1285 */
1286 if (new_members)
1287 {
1288 List *old_members = list_concat(outer_members, inner_members);
1289 EquivalenceMember *prev_em = NULL;
1290 RestrictInfo *rinfo;
1291
1292 /* For now, arbitrarily take the first old_member as the one to use */
1293 if (old_members)
1294 new_members = lappend(new_members, linitial(old_members));
1295
1296 foreach(lc1, new_members)
1297 {
1298 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc1);
1299
1300 if (prev_em != NULL)
1301 {
1302 Oid eq_op;
1303
1304 eq_op = select_equality_operator(ec,
1305 prev_em->em_datatype,
1306 cur_em->em_datatype);
1307 if (!OidIsValid(eq_op))
1308 {
1309 /* failed... */
1310 ec->ec_broken = true;
1311 return NIL;
1312 }
1313 /* do NOT set parent_ec, this qual is not redundant! */
1314 rinfo = create_join_clause(root, ec, eq_op,
1315 prev_em, cur_em,
1316 NULL);
1317
1318 result = lappend(result, rinfo);
1319 }
1320 prev_em = cur_em;
1321 }
1322 }
1323
1324 return result;
1325 }
1326
1327 /*
1328 * generate_join_implied_equalities cleanup after failure
1329 *
1330 * Return any original RestrictInfos that are enforceable at this join.
1331 *
1332 * In the case of a child inner relation, we have to translate the
1333 * original RestrictInfos from parent to child Vars.
1334 */
1335 static List *
generate_join_implied_equalities_broken(PlannerInfo * root,EquivalenceClass * ec,Relids nominal_join_relids,Relids outer_relids,Relids nominal_inner_relids,RelOptInfo * inner_rel)1336 generate_join_implied_equalities_broken(PlannerInfo *root,
1337 EquivalenceClass *ec,
1338 Relids nominal_join_relids,
1339 Relids outer_relids,
1340 Relids nominal_inner_relids,
1341 RelOptInfo *inner_rel)
1342 {
1343 List *result = NIL;
1344 ListCell *lc;
1345
1346 foreach(lc, ec->ec_sources)
1347 {
1348 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1349 Relids clause_relids = restrictinfo->required_relids;
1350
1351 if (bms_is_subset(clause_relids, nominal_join_relids) &&
1352 !bms_is_subset(clause_relids, outer_relids) &&
1353 !bms_is_subset(clause_relids, nominal_inner_relids))
1354 result = lappend(result, restrictinfo);
1355 }
1356
1357 /*
1358 * If we have to translate, just brute-force apply adjust_appendrel_attrs
1359 * to all the RestrictInfos at once. This will result in returning
1360 * RestrictInfos that are not listed in ec_derives, but there shouldn't be
1361 * any duplication, and it's a sufficiently narrow corner case that we
1362 * shouldn't sweat too much over it anyway.
1363 *
1364 * Since inner_rel might be an indirect descendant of the baserel
1365 * mentioned in the ec_sources clauses, we have to be prepared to apply
1366 * multiple levels of Var translation.
1367 */
1368 if (IS_OTHER_REL(inner_rel) && result != NIL)
1369 result = (List *) adjust_appendrel_attrs_multilevel(root,
1370 (Node *) result,
1371 inner_rel->relids,
1372 inner_rel->top_parent_relids);
1373
1374 return result;
1375 }
1376
1377
1378 /*
1379 * select_equality_operator
1380 * Select a suitable equality operator for comparing two EC members
1381 *
1382 * Returns InvalidOid if no operator can be found for this datatype combination
1383 */
1384 static Oid
select_equality_operator(EquivalenceClass * ec,Oid lefttype,Oid righttype)1385 select_equality_operator(EquivalenceClass *ec, Oid lefttype, Oid righttype)
1386 {
1387 ListCell *lc;
1388
1389 foreach(lc, ec->ec_opfamilies)
1390 {
1391 Oid opfamily = lfirst_oid(lc);
1392 Oid opno;
1393
1394 opno = get_opfamily_member(opfamily, lefttype, righttype,
1395 BTEqualStrategyNumber);
1396 if (!OidIsValid(opno))
1397 continue;
1398 /* If no barrier quals in query, don't worry about leaky operators */
1399 if (ec->ec_max_security == 0)
1400 return opno;
1401 /* Otherwise, insist that selected operators be leakproof */
1402 if (get_func_leakproof(get_opcode(opno)))
1403 return opno;
1404 }
1405 return InvalidOid;
1406 }
1407
1408
1409 /*
1410 * create_join_clause
1411 * Find or make a RestrictInfo comparing the two given EC members
1412 * with the given operator.
1413 *
1414 * parent_ec is either equal to ec (if the clause is a potentially-redundant
1415 * join clause) or NULL (if not). We have to treat this as part of the
1416 * match requirements --- it's possible that a clause comparing the same two
1417 * EMs is a join clause in one join path and a restriction clause in another.
1418 */
1419 static RestrictInfo *
create_join_clause(PlannerInfo * root,EquivalenceClass * ec,Oid opno,EquivalenceMember * leftem,EquivalenceMember * rightem,EquivalenceClass * parent_ec)1420 create_join_clause(PlannerInfo *root,
1421 EquivalenceClass *ec, Oid opno,
1422 EquivalenceMember *leftem,
1423 EquivalenceMember *rightem,
1424 EquivalenceClass *parent_ec)
1425 {
1426 RestrictInfo *rinfo;
1427 ListCell *lc;
1428 MemoryContext oldcontext;
1429
1430 /*
1431 * Search to see if we already built a RestrictInfo for this pair of
1432 * EquivalenceMembers. We can use either original source clauses or
1433 * previously-derived clauses. The check on opno is probably redundant,
1434 * but be safe ...
1435 */
1436 foreach(lc, ec->ec_sources)
1437 {
1438 rinfo = (RestrictInfo *) lfirst(lc);
1439 if (rinfo->left_em == leftem &&
1440 rinfo->right_em == rightem &&
1441 rinfo->parent_ec == parent_ec &&
1442 opno == ((OpExpr *) rinfo->clause)->opno)
1443 return rinfo;
1444 }
1445
1446 foreach(lc, ec->ec_derives)
1447 {
1448 rinfo = (RestrictInfo *) lfirst(lc);
1449 if (rinfo->left_em == leftem &&
1450 rinfo->right_em == rightem &&
1451 rinfo->parent_ec == parent_ec &&
1452 opno == ((OpExpr *) rinfo->clause)->opno)
1453 return rinfo;
1454 }
1455
1456 /*
1457 * Not there, so build it, in planner context so we can re-use it. (Not
1458 * important in normal planning, but definitely so in GEQO.)
1459 */
1460 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
1461
1462 rinfo = build_implied_join_equality(opno,
1463 ec->ec_collation,
1464 leftem->em_expr,
1465 rightem->em_expr,
1466 bms_union(leftem->em_relids,
1467 rightem->em_relids),
1468 bms_union(leftem->em_nullable_relids,
1469 rightem->em_nullable_relids),
1470 ec->ec_min_security);
1471
1472 /* Mark the clause as redundant, or not */
1473 rinfo->parent_ec = parent_ec;
1474
1475 /*
1476 * We know the correct values for left_ec/right_ec, ie this particular EC,
1477 * so we can just set them directly instead of forcing another lookup.
1478 */
1479 rinfo->left_ec = ec;
1480 rinfo->right_ec = ec;
1481
1482 /* Mark it as usable with these EMs */
1483 rinfo->left_em = leftem;
1484 rinfo->right_em = rightem;
1485 /* and save it for possible re-use */
1486 ec->ec_derives = lappend(ec->ec_derives, rinfo);
1487
1488 MemoryContextSwitchTo(oldcontext);
1489
1490 return rinfo;
1491 }
1492
1493
1494 /*
1495 * reconsider_outer_join_clauses
1496 * Re-examine any outer-join clauses that were set aside by
1497 * distribute_qual_to_rels(), and see if we can derive any
1498 * EquivalenceClasses from them. Then, if they were not made
1499 * redundant, push them out into the regular join-clause lists.
1500 *
1501 * When we have mergejoinable clauses A = B that are outer-join clauses,
1502 * we can't blindly combine them with other clauses A = C to deduce B = C,
1503 * since in fact the "equality" A = B won't necessarily hold above the
1504 * outer join (one of the variables might be NULL instead). Nonetheless
1505 * there are cases where we can add qual clauses using transitivity.
1506 *
1507 * One case that we look for here is an outer-join clause OUTERVAR = INNERVAR
1508 * for which there is also an equivalence clause OUTERVAR = CONSTANT.
1509 * It is safe and useful to push a clause INNERVAR = CONSTANT into the
1510 * evaluation of the inner (nullable) relation, because any inner rows not
1511 * meeting this condition will not contribute to the outer-join result anyway.
1512 * (Any outer rows they could join to will be eliminated by the pushed-down
1513 * equivalence clause.)
1514 *
1515 * Note that the above rule does not work for full outer joins; nor is it
1516 * very interesting to consider cases where the generated equivalence clause
1517 * would involve relations outside the outer join, since such clauses couldn't
1518 * be pushed into the inner side's scan anyway. So the restriction to
1519 * outervar = pseudoconstant is not really giving up anything.
1520 *
1521 * For full-join cases, we can only do something useful if it's a FULL JOIN
1522 * USING and a merged column has an equivalence MERGEDVAR = CONSTANT.
1523 * By the time it gets here, the merged column will look like
1524 * COALESCE(LEFTVAR, RIGHTVAR)
1525 * and we will have a full-join clause LEFTVAR = RIGHTVAR that we can match
1526 * the COALESCE expression to. In this situation we can push LEFTVAR = CONSTANT
1527 * and RIGHTVAR = CONSTANT into the input relations, since any rows not
1528 * meeting these conditions cannot contribute to the join result.
1529 *
1530 * Again, there isn't any traction to be gained by trying to deal with
1531 * clauses comparing a mergedvar to a non-pseudoconstant. So we can make
1532 * use of the EquivalenceClasses to search for matching variables that were
1533 * equivalenced to constants. The interesting outer-join clauses were
1534 * accumulated for us by distribute_qual_to_rels.
1535 *
1536 * When we find one of these cases, we implement the changes we want by
1537 * generating a new equivalence clause INNERVAR = CONSTANT (or LEFTVAR, etc)
1538 * and pushing it into the EquivalenceClass structures. This is because we
1539 * may already know that INNERVAR is equivalenced to some other var(s), and
1540 * we'd like the constant to propagate to them too. Note that it would be
1541 * unsafe to merge any existing EC for INNERVAR with the OUTERVAR's EC ---
1542 * that could result in propagating constant restrictions from
1543 * INNERVAR to OUTERVAR, which would be very wrong.
1544 *
1545 * It's possible that the INNERVAR is also an OUTERVAR for some other
1546 * outer-join clause, in which case the process can be repeated. So we repeat
1547 * looping over the lists of clauses until no further deductions can be made.
1548 * Whenever we do make a deduction, we remove the generating clause from the
1549 * lists, since we don't want to make the same deduction twice.
1550 *
1551 * If we don't find any match for a set-aside outer join clause, we must
1552 * throw it back into the regular joinclause processing by passing it to
1553 * distribute_restrictinfo_to_rels(). If we do generate a derived clause,
1554 * however, the outer-join clause is redundant. We still throw it back,
1555 * because otherwise the join will be seen as a clauseless join and avoided
1556 * during join order searching; but we mark it as redundant to keep from
1557 * messing up the joinrel's size estimate. (This behavior means that the
1558 * API for this routine is uselessly complex: we could have just put all
1559 * the clauses into the regular processing initially. We keep it because
1560 * someday we might want to do something else, such as inserting "dummy"
1561 * joinclauses instead of real ones.)
1562 *
1563 * Outer join clauses that are marked outerjoin_delayed are special: this
1564 * condition means that one or both VARs might go to null due to a lower
1565 * outer join. We can still push a constant through the clause, but only
1566 * if its operator is strict; and we *have to* throw the clause back into
1567 * regular joinclause processing. By keeping the strict join clause,
1568 * we ensure that any null-extended rows that are mistakenly generated due
1569 * to suppressing rows not matching the constant will be rejected at the
1570 * upper outer join. (This doesn't work for full-join clauses.)
1571 */
1572 void
reconsider_outer_join_clauses(PlannerInfo * root)1573 reconsider_outer_join_clauses(PlannerInfo *root)
1574 {
1575 bool found;
1576 ListCell *cell;
1577 ListCell *prev;
1578 ListCell *next;
1579
1580 /* Outer loop repeats until we find no more deductions */
1581 do
1582 {
1583 found = false;
1584
1585 /* Process the LEFT JOIN clauses */
1586 prev = NULL;
1587 for (cell = list_head(root->left_join_clauses); cell; cell = next)
1588 {
1589 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1590
1591 next = lnext(cell);
1592 if (reconsider_outer_join_clause(root, rinfo, true))
1593 {
1594 found = true;
1595 /* remove it from the list */
1596 root->left_join_clauses =
1597 list_delete_cell(root->left_join_clauses, cell, prev);
1598 /* we throw it back anyway (see notes above) */
1599 /* but the thrown-back clause has no extra selectivity */
1600 rinfo->norm_selec = 2.0;
1601 rinfo->outer_selec = 1.0;
1602 distribute_restrictinfo_to_rels(root, rinfo);
1603 }
1604 else
1605 prev = cell;
1606 }
1607
1608 /* Process the RIGHT JOIN clauses */
1609 prev = NULL;
1610 for (cell = list_head(root->right_join_clauses); cell; cell = next)
1611 {
1612 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1613
1614 next = lnext(cell);
1615 if (reconsider_outer_join_clause(root, rinfo, false))
1616 {
1617 found = true;
1618 /* remove it from the list */
1619 root->right_join_clauses =
1620 list_delete_cell(root->right_join_clauses, cell, prev);
1621 /* we throw it back anyway (see notes above) */
1622 /* but the thrown-back clause has no extra selectivity */
1623 rinfo->norm_selec = 2.0;
1624 rinfo->outer_selec = 1.0;
1625 distribute_restrictinfo_to_rels(root, rinfo);
1626 }
1627 else
1628 prev = cell;
1629 }
1630
1631 /* Process the FULL JOIN clauses */
1632 prev = NULL;
1633 for (cell = list_head(root->full_join_clauses); cell; cell = next)
1634 {
1635 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1636
1637 next = lnext(cell);
1638 if (reconsider_full_join_clause(root, rinfo))
1639 {
1640 found = true;
1641 /* remove it from the list */
1642 root->full_join_clauses =
1643 list_delete_cell(root->full_join_clauses, cell, prev);
1644 /* we throw it back anyway (see notes above) */
1645 /* but the thrown-back clause has no extra selectivity */
1646 rinfo->norm_selec = 2.0;
1647 rinfo->outer_selec = 1.0;
1648 distribute_restrictinfo_to_rels(root, rinfo);
1649 }
1650 else
1651 prev = cell;
1652 }
1653 } while (found);
1654
1655 /* Now, any remaining clauses have to be thrown back */
1656 foreach(cell, root->left_join_clauses)
1657 {
1658 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1659
1660 distribute_restrictinfo_to_rels(root, rinfo);
1661 }
1662 foreach(cell, root->right_join_clauses)
1663 {
1664 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1665
1666 distribute_restrictinfo_to_rels(root, rinfo);
1667 }
1668 foreach(cell, root->full_join_clauses)
1669 {
1670 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1671
1672 distribute_restrictinfo_to_rels(root, rinfo);
1673 }
1674 }
1675
1676 /*
1677 * reconsider_outer_join_clauses for a single LEFT/RIGHT JOIN clause
1678 *
1679 * Returns true if we were able to propagate a constant through the clause.
1680 */
1681 static bool
reconsider_outer_join_clause(PlannerInfo * root,RestrictInfo * rinfo,bool outer_on_left)1682 reconsider_outer_join_clause(PlannerInfo *root, RestrictInfo *rinfo,
1683 bool outer_on_left)
1684 {
1685 Expr *outervar,
1686 *innervar;
1687 Oid opno,
1688 collation,
1689 left_type,
1690 right_type,
1691 inner_datatype;
1692 Relids inner_relids,
1693 inner_nullable_relids;
1694 ListCell *lc1;
1695
1696 Assert(is_opclause(rinfo->clause));
1697 opno = ((OpExpr *) rinfo->clause)->opno;
1698 collation = ((OpExpr *) rinfo->clause)->inputcollid;
1699
1700 /* If clause is outerjoin_delayed, operator must be strict */
1701 if (rinfo->outerjoin_delayed && !op_strict(opno))
1702 return false;
1703
1704 /* Extract needed info from the clause */
1705 op_input_types(opno, &left_type, &right_type);
1706 if (outer_on_left)
1707 {
1708 outervar = (Expr *) get_leftop(rinfo->clause);
1709 innervar = (Expr *) get_rightop(rinfo->clause);
1710 inner_datatype = right_type;
1711 inner_relids = rinfo->right_relids;
1712 }
1713 else
1714 {
1715 outervar = (Expr *) get_rightop(rinfo->clause);
1716 innervar = (Expr *) get_leftop(rinfo->clause);
1717 inner_datatype = left_type;
1718 inner_relids = rinfo->left_relids;
1719 }
1720 inner_nullable_relids = bms_intersect(inner_relids,
1721 rinfo->nullable_relids);
1722
1723 /* Scan EquivalenceClasses for a match to outervar */
1724 foreach(lc1, root->eq_classes)
1725 {
1726 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
1727 bool match;
1728 ListCell *lc2;
1729
1730 /* Ignore EC unless it contains pseudoconstants */
1731 if (!cur_ec->ec_has_const)
1732 continue;
1733 /* Never match to a volatile EC */
1734 if (cur_ec->ec_has_volatile)
1735 continue;
1736 /* It has to match the outer-join clause as to semantics, too */
1737 if (collation != cur_ec->ec_collation)
1738 continue;
1739 if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
1740 continue;
1741 /* Does it contain a match to outervar? */
1742 match = false;
1743 foreach(lc2, cur_ec->ec_members)
1744 {
1745 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1746
1747 Assert(!cur_em->em_is_child); /* no children yet */
1748 if (equal(outervar, cur_em->em_expr))
1749 {
1750 match = true;
1751 break;
1752 }
1753 }
1754 if (!match)
1755 continue; /* no match, so ignore this EC */
1756
1757 /*
1758 * Yes it does! Try to generate a clause INNERVAR = CONSTANT for each
1759 * CONSTANT in the EC. Note that we must succeed with at least one
1760 * constant before we can decide to throw away the outer-join clause.
1761 */
1762 match = false;
1763 foreach(lc2, cur_ec->ec_members)
1764 {
1765 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1766 Oid eq_op;
1767 RestrictInfo *newrinfo;
1768
1769 if (!cur_em->em_is_const)
1770 continue; /* ignore non-const members */
1771 eq_op = select_equality_operator(cur_ec,
1772 inner_datatype,
1773 cur_em->em_datatype);
1774 if (!OidIsValid(eq_op))
1775 continue; /* can't generate equality */
1776 newrinfo = build_implied_join_equality(eq_op,
1777 cur_ec->ec_collation,
1778 innervar,
1779 cur_em->em_expr,
1780 bms_copy(inner_relids),
1781 bms_copy(inner_nullable_relids),
1782 cur_ec->ec_min_security);
1783 if (process_equivalence(root, &newrinfo, true))
1784 match = true;
1785 }
1786
1787 /*
1788 * If we were able to equate INNERVAR to any constant, report success.
1789 * Otherwise, fall out of the search loop, since we know the OUTERVAR
1790 * appears in at most one EC.
1791 */
1792 if (match)
1793 return true;
1794 else
1795 break;
1796 }
1797
1798 return false; /* failed to make any deduction */
1799 }
1800
1801 /*
1802 * reconsider_outer_join_clauses for a single FULL JOIN clause
1803 *
1804 * Returns true if we were able to propagate a constant through the clause.
1805 */
1806 static bool
reconsider_full_join_clause(PlannerInfo * root,RestrictInfo * rinfo)1807 reconsider_full_join_clause(PlannerInfo *root, RestrictInfo *rinfo)
1808 {
1809 Expr *leftvar;
1810 Expr *rightvar;
1811 Oid opno,
1812 collation,
1813 left_type,
1814 right_type;
1815 Relids left_relids,
1816 right_relids,
1817 left_nullable_relids,
1818 right_nullable_relids;
1819 ListCell *lc1;
1820
1821 /* Can't use an outerjoin_delayed clause here */
1822 if (rinfo->outerjoin_delayed)
1823 return false;
1824
1825 /* Extract needed info from the clause */
1826 Assert(is_opclause(rinfo->clause));
1827 opno = ((OpExpr *) rinfo->clause)->opno;
1828 collation = ((OpExpr *) rinfo->clause)->inputcollid;
1829 op_input_types(opno, &left_type, &right_type);
1830 leftvar = (Expr *) get_leftop(rinfo->clause);
1831 rightvar = (Expr *) get_rightop(rinfo->clause);
1832 left_relids = rinfo->left_relids;
1833 right_relids = rinfo->right_relids;
1834 left_nullable_relids = bms_intersect(left_relids,
1835 rinfo->nullable_relids);
1836 right_nullable_relids = bms_intersect(right_relids,
1837 rinfo->nullable_relids);
1838
1839 foreach(lc1, root->eq_classes)
1840 {
1841 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
1842 EquivalenceMember *coal_em = NULL;
1843 bool match;
1844 bool matchleft;
1845 bool matchright;
1846 ListCell *lc2;
1847
1848 /* Ignore EC unless it contains pseudoconstants */
1849 if (!cur_ec->ec_has_const)
1850 continue;
1851 /* Never match to a volatile EC */
1852 if (cur_ec->ec_has_volatile)
1853 continue;
1854 /* It has to match the outer-join clause as to semantics, too */
1855 if (collation != cur_ec->ec_collation)
1856 continue;
1857 if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
1858 continue;
1859
1860 /*
1861 * Does it contain a COALESCE(leftvar, rightvar) construct?
1862 *
1863 * We can assume the COALESCE() inputs are in the same order as the
1864 * join clause, since both were automatically generated in the cases
1865 * we care about.
1866 *
1867 * XXX currently this may fail to match in cross-type cases because
1868 * the COALESCE will contain typecast operations while the join clause
1869 * may not (if there is a cross-type mergejoin operator available for
1870 * the two column types). Is it OK to strip implicit coercions from
1871 * the COALESCE arguments?
1872 */
1873 match = false;
1874 foreach(lc2, cur_ec->ec_members)
1875 {
1876 coal_em = (EquivalenceMember *) lfirst(lc2);
1877 Assert(!coal_em->em_is_child); /* no children yet */
1878 if (IsA(coal_em->em_expr, CoalesceExpr))
1879 {
1880 CoalesceExpr *cexpr = (CoalesceExpr *) coal_em->em_expr;
1881 Node *cfirst;
1882 Node *csecond;
1883
1884 if (list_length(cexpr->args) != 2)
1885 continue;
1886 cfirst = (Node *) linitial(cexpr->args);
1887 csecond = (Node *) lsecond(cexpr->args);
1888
1889 if (equal(leftvar, cfirst) && equal(rightvar, csecond))
1890 {
1891 match = true;
1892 break;
1893 }
1894 }
1895 }
1896 if (!match)
1897 continue; /* no match, so ignore this EC */
1898
1899 /*
1900 * Yes it does! Try to generate clauses LEFTVAR = CONSTANT and
1901 * RIGHTVAR = CONSTANT for each CONSTANT in the EC. Note that we must
1902 * succeed with at least one constant for each var before we can
1903 * decide to throw away the outer-join clause.
1904 */
1905 matchleft = matchright = false;
1906 foreach(lc2, cur_ec->ec_members)
1907 {
1908 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1909 Oid eq_op;
1910 RestrictInfo *newrinfo;
1911
1912 if (!cur_em->em_is_const)
1913 continue; /* ignore non-const members */
1914 eq_op = select_equality_operator(cur_ec,
1915 left_type,
1916 cur_em->em_datatype);
1917 if (OidIsValid(eq_op))
1918 {
1919 newrinfo = build_implied_join_equality(eq_op,
1920 cur_ec->ec_collation,
1921 leftvar,
1922 cur_em->em_expr,
1923 bms_copy(left_relids),
1924 bms_copy(left_nullable_relids),
1925 cur_ec->ec_min_security);
1926 if (process_equivalence(root, &newrinfo, true))
1927 matchleft = true;
1928 }
1929 eq_op = select_equality_operator(cur_ec,
1930 right_type,
1931 cur_em->em_datatype);
1932 if (OidIsValid(eq_op))
1933 {
1934 newrinfo = build_implied_join_equality(eq_op,
1935 cur_ec->ec_collation,
1936 rightvar,
1937 cur_em->em_expr,
1938 bms_copy(right_relids),
1939 bms_copy(right_nullable_relids),
1940 cur_ec->ec_min_security);
1941 if (process_equivalence(root, &newrinfo, true))
1942 matchright = true;
1943 }
1944 }
1945
1946 /*
1947 * If we were able to equate both vars to constants, we're done, and
1948 * we can throw away the full-join clause as redundant. Moreover, we
1949 * can remove the COALESCE entry from the EC, since the added
1950 * restrictions ensure it will always have the expected value. (We
1951 * don't bother trying to update ec_relids or ec_sources.)
1952 */
1953 if (matchleft && matchright)
1954 {
1955 cur_ec->ec_members = list_delete_ptr(cur_ec->ec_members, coal_em);
1956 return true;
1957 }
1958
1959 /*
1960 * Otherwise, fall out of the search loop, since we know the COALESCE
1961 * appears in at most one EC (XXX might stop being true if we allow
1962 * stripping of coercions above?)
1963 */
1964 break;
1965 }
1966
1967 return false; /* failed to make any deduction */
1968 }
1969
1970
1971 /*
1972 * exprs_known_equal
1973 * Detect whether two expressions are known equal due to equivalence
1974 * relationships.
1975 *
1976 * Actually, this only shows that the expressions are equal according
1977 * to some opfamily's notion of equality --- but we only use it for
1978 * selectivity estimation, so a fuzzy idea of equality is OK.
1979 *
1980 * Note: does not bother to check for "equal(item1, item2)"; caller must
1981 * check that case if it's possible to pass identical items.
1982 */
1983 bool
exprs_known_equal(PlannerInfo * root,Node * item1,Node * item2)1984 exprs_known_equal(PlannerInfo *root, Node *item1, Node *item2)
1985 {
1986 ListCell *lc1;
1987
1988 foreach(lc1, root->eq_classes)
1989 {
1990 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
1991 bool item1member = false;
1992 bool item2member = false;
1993 ListCell *lc2;
1994
1995 /* Never match to a volatile EC */
1996 if (ec->ec_has_volatile)
1997 continue;
1998
1999 foreach(lc2, ec->ec_members)
2000 {
2001 EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
2002
2003 if (em->em_is_child)
2004 continue; /* ignore children here */
2005 if (equal(item1, em->em_expr))
2006 item1member = true;
2007 else if (equal(item2, em->em_expr))
2008 item2member = true;
2009 /* Exit as soon as equality is proven */
2010 if (item1member && item2member)
2011 return true;
2012 }
2013 }
2014 return false;
2015 }
2016
2017
2018 /*
2019 * match_eclasses_to_foreign_key_col
2020 * See whether a foreign key column match is proven by any eclass.
2021 *
2022 * If the referenced and referencing Vars of the fkey's colno'th column are
2023 * known equal due to any eclass, return that eclass; otherwise return NULL.
2024 * (In principle there might be more than one matching eclass if multiple
2025 * collations are involved, but since collation doesn't matter for equality,
2026 * we ignore that fine point here.) This is much like exprs_known_equal,
2027 * except that we insist on the comparison operator matching the eclass, so
2028 * that the result is definite not approximate.
2029 */
2030 EquivalenceClass *
match_eclasses_to_foreign_key_col(PlannerInfo * root,ForeignKeyOptInfo * fkinfo,int colno)2031 match_eclasses_to_foreign_key_col(PlannerInfo *root,
2032 ForeignKeyOptInfo *fkinfo,
2033 int colno)
2034 {
2035 Index var1varno = fkinfo->con_relid;
2036 AttrNumber var1attno = fkinfo->conkey[colno];
2037 Index var2varno = fkinfo->ref_relid;
2038 AttrNumber var2attno = fkinfo->confkey[colno];
2039 Oid eqop = fkinfo->conpfeqop[colno];
2040 List *opfamilies = NIL; /* compute only if needed */
2041 ListCell *lc1;
2042
2043 foreach(lc1, root->eq_classes)
2044 {
2045 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2046 bool item1member = false;
2047 bool item2member = false;
2048 ListCell *lc2;
2049
2050 /* Never match to a volatile EC */
2051 if (ec->ec_has_volatile)
2052 continue;
2053 /* Note: it seems okay to match to "broken" eclasses here */
2054
2055 /*
2056 * If eclass visibly doesn't have members for both rels, there's no
2057 * need to grovel through the members.
2058 */
2059 if (!bms_is_member(var1varno, ec->ec_relids) ||
2060 !bms_is_member(var2varno, ec->ec_relids))
2061 continue;
2062
2063 foreach(lc2, ec->ec_members)
2064 {
2065 EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
2066 Var *var;
2067
2068 if (em->em_is_child)
2069 continue; /* ignore children here */
2070
2071 /* EM must be a Var, possibly with RelabelType */
2072 var = (Var *) em->em_expr;
2073 while (var && IsA(var, RelabelType))
2074 var = (Var *) ((RelabelType *) var)->arg;
2075 if (!(var && IsA(var, Var)))
2076 continue;
2077
2078 /* Match? */
2079 if (var->varno == var1varno && var->varattno == var1attno)
2080 item1member = true;
2081 else if (var->varno == var2varno && var->varattno == var2attno)
2082 item2member = true;
2083
2084 /* Have we found both PK and FK column in this EC? */
2085 if (item1member && item2member)
2086 {
2087 /*
2088 * Succeed if eqop matches EC's opfamilies. We could test
2089 * this before scanning the members, but it's probably cheaper
2090 * to test for member matches first.
2091 */
2092 if (opfamilies == NIL) /* compute if we didn't already */
2093 opfamilies = get_mergejoin_opfamilies(eqop);
2094 if (equal(opfamilies, ec->ec_opfamilies))
2095 return ec;
2096 /* Otherwise, done with this EC, move on to the next */
2097 break;
2098 }
2099 }
2100 }
2101 return NULL;
2102 }
2103
2104
2105 /*
2106 * add_child_rel_equivalences
2107 * Search for EC members that reference the parent_rel, and
2108 * add transformed members referencing the child_rel.
2109 *
2110 * Note that this function won't be called at all unless we have at least some
2111 * reason to believe that the EC members it generates will be useful.
2112 *
2113 * parent_rel and child_rel could be derived from appinfo, but since the
2114 * caller has already computed them, we might as well just pass them in.
2115 */
2116 void
add_child_rel_equivalences(PlannerInfo * root,AppendRelInfo * appinfo,RelOptInfo * parent_rel,RelOptInfo * child_rel)2117 add_child_rel_equivalences(PlannerInfo *root,
2118 AppendRelInfo *appinfo,
2119 RelOptInfo *parent_rel,
2120 RelOptInfo *child_rel)
2121 {
2122 ListCell *lc1;
2123
2124 foreach(lc1, root->eq_classes)
2125 {
2126 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2127 ListCell *lc2;
2128
2129 /*
2130 * If this EC contains a volatile expression, then generating child
2131 * EMs would be downright dangerous, so skip it. We rely on a
2132 * volatile EC having only one EM.
2133 */
2134 if (cur_ec->ec_has_volatile)
2135 continue;
2136
2137 /*
2138 * No point in searching if parent rel not mentioned in eclass; but we
2139 * can't tell that for sure if parent rel is itself a child.
2140 */
2141 if (parent_rel->reloptkind == RELOPT_BASEREL &&
2142 !bms_is_subset(parent_rel->relids, cur_ec->ec_relids))
2143 continue;
2144
2145 foreach(lc2, cur_ec->ec_members)
2146 {
2147 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2148
2149 if (cur_em->em_is_const)
2150 continue; /* ignore consts here */
2151
2152 /* Does it reference parent_rel? */
2153 if (bms_overlap(cur_em->em_relids, parent_rel->relids))
2154 {
2155 /* Yes, generate transformed child version */
2156 Expr *child_expr;
2157 Relids new_relids;
2158 Relids new_nullable_relids;
2159
2160 child_expr = (Expr *)
2161 adjust_appendrel_attrs(root,
2162 (Node *) cur_em->em_expr,
2163 1, &appinfo);
2164
2165 /*
2166 * Transform em_relids to match. Note we do *not* do
2167 * pull_varnos(child_expr) here, as for example the
2168 * transformation might have substituted a constant, but we
2169 * don't want the child member to be marked as constant.
2170 */
2171 new_relids = bms_difference(cur_em->em_relids,
2172 parent_rel->relids);
2173 new_relids = bms_add_members(new_relids, child_rel->relids);
2174
2175 /*
2176 * And likewise for nullable_relids. Note this code assumes
2177 * parent and child relids are singletons.
2178 */
2179 new_nullable_relids = cur_em->em_nullable_relids;
2180 if (bms_overlap(new_nullable_relids, parent_rel->relids))
2181 {
2182 new_nullable_relids = bms_difference(new_nullable_relids,
2183 parent_rel->relids);
2184 new_nullable_relids = bms_add_members(new_nullable_relids,
2185 child_rel->relids);
2186 }
2187
2188 (void) add_eq_member(cur_ec, child_expr,
2189 new_relids, new_nullable_relids,
2190 true, cur_em->em_datatype);
2191 }
2192 }
2193 }
2194 }
2195
2196
2197 /*
2198 * generate_implied_equalities_for_column
2199 * Create EC-derived joinclauses usable with a specific column.
2200 *
2201 * This is used by indxpath.c to extract potentially indexable joinclauses
2202 * from ECs, and can be used by foreign data wrappers for similar purposes.
2203 * We assume that only expressions in Vars of a single table are of interest,
2204 * but the caller provides a callback function to identify exactly which
2205 * such expressions it would like to know about.
2206 *
2207 * We assume that any given table/index column could appear in only one EC.
2208 * (This should be true in all but the most pathological cases, and if it
2209 * isn't, we stop on the first match anyway.) Therefore, what we return
2210 * is a redundant list of clauses equating the table/index column to each of
2211 * the other-relation values it is known to be equal to. Any one of
2212 * these clauses can be used to create a parameterized path, and there
2213 * is no value in using more than one. (But it *is* worthwhile to create
2214 * a separate parameterized path for each one, since that leads to different
2215 * join orders.)
2216 *
2217 * The caller can pass a Relids set of rels we aren't interested in joining
2218 * to, so as to save the work of creating useless clauses.
2219 */
2220 List *
generate_implied_equalities_for_column(PlannerInfo * root,RelOptInfo * rel,ec_matches_callback_type callback,void * callback_arg,Relids prohibited_rels)2221 generate_implied_equalities_for_column(PlannerInfo *root,
2222 RelOptInfo *rel,
2223 ec_matches_callback_type callback,
2224 void *callback_arg,
2225 Relids prohibited_rels)
2226 {
2227 List *result = NIL;
2228 bool is_child_rel = (rel->reloptkind == RELOPT_OTHER_MEMBER_REL);
2229 Relids parent_relids;
2230 ListCell *lc1;
2231
2232 /* Indexes are available only on base or "other" member relations. */
2233 Assert(IS_SIMPLE_REL(rel));
2234
2235 /* If it's a child rel, we'll need to know what its parent(s) are */
2236 if (is_child_rel)
2237 parent_relids = find_childrel_parents(root, rel);
2238 else
2239 parent_relids = NULL; /* not used, but keep compiler quiet */
2240
2241 foreach(lc1, root->eq_classes)
2242 {
2243 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2244 EquivalenceMember *cur_em;
2245 ListCell *lc2;
2246
2247 /*
2248 * Won't generate joinclauses if const or single-member (the latter
2249 * test covers the volatile case too)
2250 */
2251 if (cur_ec->ec_has_const || list_length(cur_ec->ec_members) <= 1)
2252 continue;
2253
2254 /*
2255 * No point in searching if rel not mentioned in eclass (but we can't
2256 * tell that for a child rel).
2257 */
2258 if (!is_child_rel &&
2259 !bms_is_subset(rel->relids, cur_ec->ec_relids))
2260 continue;
2261
2262 /*
2263 * Scan members, looking for a match to the target column. Note that
2264 * child EC members are considered, but only when they belong to the
2265 * target relation. (Unlike regular members, the same expression
2266 * could be a child member of more than one EC. Therefore, it's
2267 * potentially order-dependent which EC a child relation's target
2268 * column gets matched to. This is annoying but it only happens in
2269 * corner cases, so for now we live with just reporting the first
2270 * match. See also get_eclass_for_sort_expr.)
2271 */
2272 cur_em = NULL;
2273 foreach(lc2, cur_ec->ec_members)
2274 {
2275 cur_em = (EquivalenceMember *) lfirst(lc2);
2276 if (bms_equal(cur_em->em_relids, rel->relids) &&
2277 callback(root, rel, cur_ec, cur_em, callback_arg))
2278 break;
2279 cur_em = NULL;
2280 }
2281
2282 if (!cur_em)
2283 continue;
2284
2285 /*
2286 * Found our match. Scan the other EC members and attempt to generate
2287 * joinclauses.
2288 */
2289 foreach(lc2, cur_ec->ec_members)
2290 {
2291 EquivalenceMember *other_em = (EquivalenceMember *) lfirst(lc2);
2292 Oid eq_op;
2293 RestrictInfo *rinfo;
2294
2295 if (other_em->em_is_child)
2296 continue; /* ignore children here */
2297
2298 /* Make sure it'll be a join to a different rel */
2299 if (other_em == cur_em ||
2300 bms_overlap(other_em->em_relids, rel->relids))
2301 continue;
2302
2303 /* Forget it if caller doesn't want joins to this rel */
2304 if (bms_overlap(other_em->em_relids, prohibited_rels))
2305 continue;
2306
2307 /*
2308 * Also, if this is a child rel, avoid generating a useless join
2309 * to its parent rel(s).
2310 */
2311 if (is_child_rel &&
2312 bms_overlap(parent_relids, other_em->em_relids))
2313 continue;
2314
2315 eq_op = select_equality_operator(cur_ec,
2316 cur_em->em_datatype,
2317 other_em->em_datatype);
2318 if (!OidIsValid(eq_op))
2319 continue;
2320
2321 /* set parent_ec to mark as redundant with other joinclauses */
2322 rinfo = create_join_clause(root, cur_ec, eq_op,
2323 cur_em, other_em,
2324 cur_ec);
2325
2326 result = lappend(result, rinfo);
2327 }
2328
2329 /*
2330 * If somehow we failed to create any join clauses, we might as well
2331 * keep scanning the ECs for another match. But if we did make any,
2332 * we're done, because we don't want to return non-redundant clauses.
2333 */
2334 if (result)
2335 break;
2336 }
2337
2338 return result;
2339 }
2340
2341 /*
2342 * have_relevant_eclass_joinclause
2343 * Detect whether there is an EquivalenceClass that could produce
2344 * a joinclause involving the two given relations.
2345 *
2346 * This is essentially a very cut-down version of
2347 * generate_join_implied_equalities(). Note it's OK to occasionally say "yes"
2348 * incorrectly. Hence we don't bother with details like whether the lack of a
2349 * cross-type operator might prevent the clause from actually being generated.
2350 */
2351 bool
have_relevant_eclass_joinclause(PlannerInfo * root,RelOptInfo * rel1,RelOptInfo * rel2)2352 have_relevant_eclass_joinclause(PlannerInfo *root,
2353 RelOptInfo *rel1, RelOptInfo *rel2)
2354 {
2355 ListCell *lc1;
2356
2357 foreach(lc1, root->eq_classes)
2358 {
2359 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2360
2361 /*
2362 * Won't generate joinclauses if single-member (this test covers the
2363 * volatile case too)
2364 */
2365 if (list_length(ec->ec_members) <= 1)
2366 continue;
2367
2368 /*
2369 * We do not need to examine the individual members of the EC, because
2370 * all that we care about is whether each rel overlaps the relids of
2371 * at least one member, and a test on ec_relids is sufficient to prove
2372 * that. (As with have_relevant_joinclause(), it is not necessary
2373 * that the EC be able to form a joinclause relating exactly the two
2374 * given rels, only that it be able to form a joinclause mentioning
2375 * both, and this will surely be true if both of them overlap
2376 * ec_relids.)
2377 *
2378 * Note we don't test ec_broken; if we did, we'd need a separate code
2379 * path to look through ec_sources. Checking the membership anyway is
2380 * OK as a possibly-overoptimistic heuristic.
2381 *
2382 * We don't test ec_has_const either, even though a const eclass won't
2383 * generate real join clauses. This is because if we had "WHERE a.x =
2384 * b.y and a.x = 42", it is worth considering a join between a and b,
2385 * since the join result is likely to be small even though it'll end
2386 * up being an unqualified nestloop.
2387 */
2388 if (bms_overlap(rel1->relids, ec->ec_relids) &&
2389 bms_overlap(rel2->relids, ec->ec_relids))
2390 return true;
2391 }
2392
2393 return false;
2394 }
2395
2396
2397 /*
2398 * has_relevant_eclass_joinclause
2399 * Detect whether there is an EquivalenceClass that could produce
2400 * a joinclause involving the given relation and anything else.
2401 *
2402 * This is the same as have_relevant_eclass_joinclause with the other rel
2403 * implicitly defined as "everything else in the query".
2404 */
2405 bool
has_relevant_eclass_joinclause(PlannerInfo * root,RelOptInfo * rel1)2406 has_relevant_eclass_joinclause(PlannerInfo *root, RelOptInfo *rel1)
2407 {
2408 ListCell *lc1;
2409
2410 foreach(lc1, root->eq_classes)
2411 {
2412 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2413
2414 /*
2415 * Won't generate joinclauses if single-member (this test covers the
2416 * volatile case too)
2417 */
2418 if (list_length(ec->ec_members) <= 1)
2419 continue;
2420
2421 /*
2422 * Per the comment in have_relevant_eclass_joinclause, it's sufficient
2423 * to find an EC that mentions both this rel and some other rel.
2424 */
2425 if (bms_overlap(rel1->relids, ec->ec_relids) &&
2426 !bms_is_subset(ec->ec_relids, rel1->relids))
2427 return true;
2428 }
2429
2430 return false;
2431 }
2432
2433
2434 /*
2435 * eclass_useful_for_merging
2436 * Detect whether the EC could produce any mergejoinable join clauses
2437 * against the specified relation.
2438 *
2439 * This is just a heuristic test and doesn't have to be exact; it's better
2440 * to say "yes" incorrectly than "no". Hence we don't bother with details
2441 * like whether the lack of a cross-type operator might prevent the clause
2442 * from actually being generated.
2443 */
2444 bool
eclass_useful_for_merging(PlannerInfo * root,EquivalenceClass * eclass,RelOptInfo * rel)2445 eclass_useful_for_merging(PlannerInfo *root,
2446 EquivalenceClass *eclass,
2447 RelOptInfo *rel)
2448 {
2449 Relids relids;
2450 ListCell *lc;
2451
2452 Assert(!eclass->ec_merged);
2453
2454 /*
2455 * Won't generate joinclauses if const or single-member (the latter test
2456 * covers the volatile case too)
2457 */
2458 if (eclass->ec_has_const || list_length(eclass->ec_members) <= 1)
2459 return false;
2460
2461 /*
2462 * Note we don't test ec_broken; if we did, we'd need a separate code path
2463 * to look through ec_sources. Checking the members anyway is OK as a
2464 * possibly-overoptimistic heuristic.
2465 */
2466
2467 /* If specified rel is a child, we must consider the topmost parent rel */
2468 if (IS_OTHER_REL(rel))
2469 {
2470 Assert(!bms_is_empty(rel->top_parent_relids));
2471 relids = rel->top_parent_relids;
2472 }
2473 else
2474 relids = rel->relids;
2475
2476 /* If rel already includes all members of eclass, no point in searching */
2477 if (bms_is_subset(eclass->ec_relids, relids))
2478 return false;
2479
2480 /* To join, we need a member not in the given rel */
2481 foreach(lc, eclass->ec_members)
2482 {
2483 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
2484
2485 if (cur_em->em_is_child)
2486 continue; /* ignore children here */
2487
2488 if (!bms_overlap(cur_em->em_relids, relids))
2489 return true;
2490 }
2491
2492 return false;
2493 }
2494
2495
2496 /*
2497 * is_redundant_derived_clause
2498 * Test whether rinfo is derived from same EC as any clause in clauselist;
2499 * if so, it can be presumed to represent a condition that's redundant
2500 * with that member of the list.
2501 */
2502 bool
is_redundant_derived_clause(RestrictInfo * rinfo,List * clauselist)2503 is_redundant_derived_clause(RestrictInfo *rinfo, List *clauselist)
2504 {
2505 EquivalenceClass *parent_ec = rinfo->parent_ec;
2506 ListCell *lc;
2507
2508 /* Fail if it's not a potentially-redundant clause from some EC */
2509 if (parent_ec == NULL)
2510 return false;
2511
2512 foreach(lc, clauselist)
2513 {
2514 RestrictInfo *otherrinfo = (RestrictInfo *) lfirst(lc);
2515
2516 if (otherrinfo->parent_ec == parent_ec)
2517 return true;
2518 }
2519
2520 return false;
2521 }
2522