1 /*-------------------------------------------------------------------------
4 * Routines for managing EquivalenceClasses
6 * See src/backend/optimizer/README for discussion of EquivalenceClasses.
9 * Portions Copyright (c) 1996-2012, PostgreSQL Global Development Group
10 * Portions Copyright (c) 1994, Regents of the University of California
13 * src/backend/optimizer/path/equivclass.c
15 *-------------------------------------------------------------------------
19 #include "access/skey.h"
20 #include "catalog/pg_type.h"
21 #include "nodes/makefuncs.h"
22 #include "nodes/nodeFuncs.h"
23 #include "optimizer/clauses.h"
24 #include "optimizer/pathnode.h"
25 #include "optimizer/paths.h"
26 #include "optimizer/planmain.h"
27 #include "optimizer/prep.h"
28 #include "optimizer/var.h"
29 #include "utils/lsyscache.h"
32 static EquivalenceMember *add_eq_member(EquivalenceClass *ec,
33 Expr *expr, Relids relids, Relids nullable_relids,
34 bool is_child, Oid datatype);
35 static void generate_base_implied_equalities_const(PlannerInfo *root,
36 EquivalenceClass *ec);
37 static void generate_base_implied_equalities_no_const(PlannerInfo *root,
38 EquivalenceClass *ec);
39 static void generate_base_implied_equalities_broken(PlannerInfo *root,
40 EquivalenceClass *ec);
41 static List *generate_join_implied_equalities_normal(PlannerInfo *root,
46 static List *generate_join_implied_equalities_broken(PlannerInfo *root,
48 Relids nominal_join_relids,
50 Relids nominal_inner_relids,
51 AppendRelInfo *inner_appinfo);
52 static Oid select_equality_operator(EquivalenceClass *ec,
53 Oid lefttype, Oid righttype);
54 static RestrictInfo *create_join_clause(PlannerInfo *root,
55 EquivalenceClass *ec, Oid opno,
56 EquivalenceMember *leftem,
57 EquivalenceMember *rightem,
58 EquivalenceClass *parent_ec);
59 static bool reconsider_outer_join_clause(PlannerInfo *root,
62 static bool reconsider_full_join_clause(PlannerInfo *root,
68 * The given clause has a mergejoinable operator and can be applied without
69 * any delay by an outer join, so its two sides can be considered equal
70 * anywhere they are both computable; moreover that equality can be
71 * extended transitively. Record this knowledge in the EquivalenceClass
72 * data structure. Returns TRUE if successful, FALSE if not (in which
73 * case caller should treat the clause as ordinary, not an equivalence).
75 * If below_outer_join is true, then the clause was found below the nullable
76 * side of an outer join, so its sides might validly be both NULL rather than
77 * strictly equal. We can still deduce equalities in such cases, but we take
78 * care to mark an EquivalenceClass if it came from any such clauses. Also,
79 * we have to check that both sides are either pseudo-constants or strict
80 * functions of Vars, else they might not both go to NULL above the outer
81 * join. (This is the reason why we need a failure return. It's more
82 * convenient to check this case here than at the call sites...)
84 * On success return, we have also initialized the clause's left_ec/right_ec
85 * fields to point to the EquivalenceClass representing it. This saves lookup
88 * Note: constructing merged EquivalenceClasses is a standard UNION-FIND
89 * problem, for which there exist better data structures than simple lists.
90 * If this code ever proves to be a bottleneck then it could be sped up ---
91 * but for now, simple is beautiful.
93 * Note: this is only called during planner startup, not during GEQO
94 * exploration, so we need not worry about whether we're in the right
98 process_equivalence(PlannerInfo *root, RestrictInfo *restrictinfo,
99 bool below_outer_join)
101 Expr *clause = restrictinfo->clause;
110 item1_nullable_relids,
111 item2_nullable_relids;
113 EquivalenceClass *ec1,
115 EquivalenceMember *em1,
119 /* Should not already be marked as having generated an eclass */
120 Assert(restrictinfo->left_ec == NULL);
121 Assert(restrictinfo->right_ec == NULL);
123 /* Extract info from given clause */
124 Assert(is_opclause(clause));
125 opno = ((OpExpr *) clause)->opno;
126 collation = ((OpExpr *) clause)->inputcollid;
127 item1 = (Expr *) get_leftop(clause);
128 item2 = (Expr *) get_rightop(clause);
129 item1_relids = restrictinfo->left_relids;
130 item2_relids = restrictinfo->right_relids;
133 * Ensure both input expressions expose the desired collation (their types
134 * should be OK already); see comments for canonicalize_ec_expression.
136 item1 = canonicalize_ec_expression(item1,
137 exprType((Node *) item1),
139 item2 = canonicalize_ec_expression(item2,
140 exprType((Node *) item2),
144 * Reject clauses of the form X=X. These are not as redundant as they
145 * might seem at first glance: assuming the operator is strict, this is
146 * really an expensive way to write X IS NOT NULL. So we must not risk
147 * just losing the clause, which would be possible if there is already a
148 * single-element EquivalenceClass containing X. The case is not common
149 * enough to be worth contorting the EC machinery for, so just reject the
150 * clause and let it be processed as a normal restriction clause.
152 if (equal(item1, item2))
153 return false; /* X=X is not a useful equivalence */
156 * If below outer join, check for strictness, else reject.
158 if (below_outer_join)
160 if (!bms_is_empty(item1_relids) &&
161 contain_nonstrict_functions((Node *) item1))
162 return false; /* LHS is non-strict but not constant */
163 if (!bms_is_empty(item2_relids) &&
164 contain_nonstrict_functions((Node *) item2))
165 return false; /* RHS is non-strict but not constant */
168 /* Calculate nullable-relid sets for each side of the clause */
169 item1_nullable_relids = bms_intersect(item1_relids,
170 restrictinfo->nullable_relids);
171 item2_nullable_relids = bms_intersect(item2_relids,
172 restrictinfo->nullable_relids);
175 * We use the declared input types of the operator, not exprType() of the
176 * inputs, as the nominal datatypes for opfamily lookup. This presumes
177 * that btree operators are always registered with amoplefttype and
178 * amoprighttype equal to their declared input types. We will need this
179 * info anyway to build EquivalenceMember nodes, and by extracting it now
180 * we can use type comparisons to short-circuit some equal() tests.
182 op_input_types(opno, &item1_type, &item2_type);
184 opfamilies = restrictinfo->mergeopfamilies;
187 * Sweep through the existing EquivalenceClasses looking for matches to
188 * item1 and item2. These are the possible outcomes:
190 * 1. We find both in the same EC. The equivalence is already known, so
191 * there's nothing to do.
193 * 2. We find both in different ECs. Merge the two ECs together.
195 * 3. We find just one. Add the other to its EC.
197 * 4. We find neither. Make a new, two-entry EC.
199 * Note: since all ECs are built through this process or the similar
200 * search in get_eclass_for_sort_expr(), it's impossible that we'd match
201 * an item in more than one existing nonvolatile EC. So it's okay to stop
202 * at the first match.
206 foreach(lc1, root->eq_classes)
208 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
211 /* Never match to a volatile EC */
212 if (cur_ec->ec_has_volatile)
216 * The collation has to match; check this first since it's cheaper
217 * than the opfamily comparison.
219 if (collation != cur_ec->ec_collation)
223 * A "match" requires matching sets of btree opfamilies. Use of
224 * equal() for this test has implications discussed in the comments
225 * for get_mergejoin_opfamilies().
227 if (!equal(opfamilies, cur_ec->ec_opfamilies))
230 foreach(lc2, cur_ec->ec_members)
232 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
234 Assert(!cur_em->em_is_child); /* no children yet */
237 * If below an outer join, don't match constants: they're not as
238 * constant as they look.
240 if ((below_outer_join || cur_ec->ec_below_outer_join) &&
245 item1_type == cur_em->em_datatype &&
246 equal(item1, cur_em->em_expr))
255 item2_type == cur_em->em_datatype &&
256 equal(item2, cur_em->em_expr))
269 /* Sweep finished, what did we find? */
273 /* If case 1, nothing to do, except add to sources */
276 ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
277 ec1->ec_below_outer_join |= below_outer_join;
278 /* mark the RI as associated with this eclass */
279 restrictinfo->left_ec = ec1;
280 restrictinfo->right_ec = ec1;
281 /* mark the RI as usable with this pair of EMs */
282 restrictinfo->left_em = em1;
283 restrictinfo->right_em = em2;
288 * Case 2: need to merge ec1 and ec2. We add ec2's items to ec1, then
289 * set ec2's ec_merged link to point to ec1 and remove ec2 from the
290 * eq_classes list. We cannot simply delete ec2 because that could
291 * leave dangling pointers in existing PathKeys. We leave it behind
292 * with a link so that the merged EC can be found.
294 ec1->ec_members = list_concat(ec1->ec_members, ec2->ec_members);
295 ec1->ec_sources = list_concat(ec1->ec_sources, ec2->ec_sources);
296 ec1->ec_derives = list_concat(ec1->ec_derives, ec2->ec_derives);
297 ec1->ec_relids = bms_join(ec1->ec_relids, ec2->ec_relids);
298 ec1->ec_has_const |= ec2->ec_has_const;
299 /* can't need to set has_volatile */
300 ec1->ec_below_outer_join |= ec2->ec_below_outer_join;
301 ec2->ec_merged = ec1;
302 root->eq_classes = list_delete_ptr(root->eq_classes, ec2);
303 /* just to avoid debugging confusion w/ dangling pointers: */
304 ec2->ec_members = NIL;
305 ec2->ec_sources = NIL;
306 ec2->ec_derives = NIL;
307 ec2->ec_relids = NULL;
308 ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
309 ec1->ec_below_outer_join |= below_outer_join;
310 /* mark the RI as associated with this eclass */
311 restrictinfo->left_ec = ec1;
312 restrictinfo->right_ec = ec1;
313 /* mark the RI as usable with this pair of EMs */
314 restrictinfo->left_em = em1;
315 restrictinfo->right_em = em2;
319 /* Case 3: add item2 to ec1 */
320 em2 = add_eq_member(ec1, item2, item2_relids, item2_nullable_relids,
322 ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
323 ec1->ec_below_outer_join |= below_outer_join;
324 /* mark the RI as associated with this eclass */
325 restrictinfo->left_ec = ec1;
326 restrictinfo->right_ec = ec1;
327 /* mark the RI as usable with this pair of EMs */
328 restrictinfo->left_em = em1;
329 restrictinfo->right_em = em2;
333 /* Case 3: add item1 to ec2 */
334 em1 = add_eq_member(ec2, item1, item1_relids, item1_nullable_relids,
336 ec2->ec_sources = lappend(ec2->ec_sources, restrictinfo);
337 ec2->ec_below_outer_join |= below_outer_join;
338 /* mark the RI as associated with this eclass */
339 restrictinfo->left_ec = ec2;
340 restrictinfo->right_ec = ec2;
341 /* mark the RI as usable with this pair of EMs */
342 restrictinfo->left_em = em1;
343 restrictinfo->right_em = em2;
347 /* Case 4: make a new, two-entry EC */
348 EquivalenceClass *ec = makeNode(EquivalenceClass);
350 ec->ec_opfamilies = opfamilies;
351 ec->ec_collation = collation;
352 ec->ec_members = NIL;
353 ec->ec_sources = list_make1(restrictinfo);
354 ec->ec_derives = NIL;
355 ec->ec_relids = NULL;
356 ec->ec_has_const = false;
357 ec->ec_has_volatile = false;
358 ec->ec_below_outer_join = below_outer_join;
359 ec->ec_broken = false;
361 ec->ec_merged = NULL;
362 em1 = add_eq_member(ec, item1, item1_relids, item1_nullable_relids,
364 em2 = add_eq_member(ec, item2, item2_relids, item2_nullable_relids,
367 root->eq_classes = lappend(root->eq_classes, ec);
369 /* mark the RI as associated with this eclass */
370 restrictinfo->left_ec = ec;
371 restrictinfo->right_ec = ec;
372 /* mark the RI as usable with this pair of EMs */
373 restrictinfo->left_em = em1;
374 restrictinfo->right_em = em2;
381 * canonicalize_ec_expression
383 * This function ensures that the expression exposes the expected type and
384 * collation, so that it will be equal() to other equivalence-class expressions
385 * that it ought to be equal() to.
387 * The rule for datatypes is that the exposed type should match what it would
388 * be for an input to an operator of the EC's opfamilies; which is usually
389 * the declared input type of the operator, but in the case of polymorphic
390 * operators no relabeling is wanted (compare the behavior of parse_coerce.c).
391 * Expressions coming in from quals will generally have the right type
392 * already, but expressions coming from indexkeys may not (because they are
393 * represented without any explicit relabel in pg_index), and the same problem
394 * occurs for sort expressions (because the parser is likewise cavalier about
395 * putting relabels on them). Such cases will be binary-compatible with the
396 * real operators, so adding a RelabelType is sufficient.
398 * Also, the expression's exposed collation must match the EC's collation.
399 * This is important because in comparisons like "foo < bar COLLATE baz",
400 * only one of the expressions has the correct exposed collation as we receive
401 * it from the parser. Forcing both of them to have it ensures that all
402 * variant spellings of such a construct behave the same. Again, we can
403 * stick on a RelabelType to force the right exposed collation. (It might
404 * work to not label the collation at all in EC members, but this is risky
405 * since some parts of the system expect exprCollation() to deliver the
406 * right answer for a sort key.)
408 * Note this code assumes that the expression has already been through
409 * eval_const_expressions, so there are no CollateExprs and no redundant
413 canonicalize_ec_expression(Expr *expr, Oid req_type, Oid req_collation)
415 Oid expr_type = exprType((Node *) expr);
418 * For a polymorphic-input-type opclass, just keep the same exposed type.
420 if (IsPolymorphicType(req_type))
421 req_type = expr_type;
424 * No work if the expression exposes the right type/collation already.
426 if (expr_type != req_type ||
427 exprCollation((Node *) expr) != req_collation)
430 * Strip any existing RelabelType, then add a new one if needed. This
431 * is to preserve the invariant of no redundant RelabelTypes.
433 * If we have to change the exposed type of the stripped expression,
434 * set typmod to -1 (since the new type may not have the same typmod
435 * interpretation). If we only have to change collation, preserve the
438 while (expr && IsA(expr, RelabelType))
439 expr = (Expr *) ((RelabelType *) expr)->arg;
441 if (exprType((Node *) expr) != req_type)
442 expr = (Expr *) makeRelabelType(expr,
446 COERCE_IMPLICIT_CAST);
447 else if (exprCollation((Node *) expr) != req_collation)
448 expr = (Expr *) makeRelabelType(expr,
450 exprTypmod((Node *) expr),
452 COERCE_IMPLICIT_CAST);
459 * add_eq_member - build a new EquivalenceMember and add it to an EC
461 static EquivalenceMember *
462 add_eq_member(EquivalenceClass *ec, Expr *expr, Relids relids,
463 Relids nullable_relids, bool is_child, Oid datatype)
465 EquivalenceMember *em = makeNode(EquivalenceMember);
468 em->em_relids = relids;
469 em->em_nullable_relids = nullable_relids;
470 em->em_is_const = false;
471 em->em_is_child = is_child;
472 em->em_datatype = datatype;
474 if (bms_is_empty(relids))
477 * No Vars, assume it's a pseudoconstant. This is correct for entries
478 * generated from process_equivalence(), because a WHERE clause can't
479 * contain aggregates or SRFs, and non-volatility was checked before
480 * process_equivalence() ever got called. But
481 * get_eclass_for_sort_expr() has to work harder. We put the tests
482 * there not here to save cycles in the equivalence case.
485 em->em_is_const = true;
486 ec->ec_has_const = true;
487 /* it can't affect ec_relids */
489 else if (!is_child) /* child members don't add to ec_relids */
491 ec->ec_relids = bms_add_members(ec->ec_relids, relids);
493 ec->ec_members = lappend(ec->ec_members, em);
500 * get_eclass_for_sort_expr
501 * Given an expression and opfamily/collation info, find an existing
502 * equivalence class it is a member of; if none, optionally build a new
503 * single-member EquivalenceClass for it.
505 * sortref is the SortGroupRef of the originating SortGroupClause, if any,
506 * or zero if not. (It should never be zero if the expression is volatile!)
508 * If rel is not NULL, it identifies a specific relation we're considering
509 * a path for, and indicates that child EC members for that relation can be
510 * considered. Otherwise child members are ignored. (Note: since child EC
511 * members aren't guaranteed unique, a non-NULL value means that there could
512 * be more than one EC that matches the expression; if so it's order-dependent
513 * which one you get. This is annoying but it only happens in corner cases,
514 * so for now we live with just reporting the first match. See also
515 * generate_implied_equalities_for_indexcol and match_pathkeys_to_index.)
517 * If create_it is TRUE, we'll build a new EquivalenceClass when there is no
518 * match. If create_it is FALSE, we just return NULL when no match.
520 * This can be used safely both before and after EquivalenceClass merging;
521 * since it never causes merging it does not invalidate any existing ECs
522 * or PathKeys. However, ECs added after path generation has begun are
523 * of limited usefulness, so usually it's best to create them beforehand.
525 * Note: opfamilies must be chosen consistently with the way
526 * process_equivalence() would do; that is, generated from a mergejoinable
527 * equality operator. Else we might fail to detect valid equivalences,
528 * generating poor (but not incorrect) plans.
531 get_eclass_for_sort_expr(PlannerInfo *root,
540 EquivalenceClass *newec;
541 EquivalenceMember *newem;
543 MemoryContext oldcontext;
546 * Ensure the expression exposes the correct type and collation.
548 expr = canonicalize_ec_expression(expr, opcintype, collation);
551 * Scan through the existing EquivalenceClasses for a match
553 foreach(lc1, root->eq_classes)
555 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
559 * Never match to a volatile EC, except when we are looking at another
560 * reference to the same volatile SortGroupClause.
562 if (cur_ec->ec_has_volatile &&
563 (sortref == 0 || sortref != cur_ec->ec_sortref))
566 if (collation != cur_ec->ec_collation)
568 if (!equal(opfamilies, cur_ec->ec_opfamilies))
571 foreach(lc2, cur_ec->ec_members)
573 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
576 * Ignore child members unless they match the request.
578 if (cur_em->em_is_child &&
579 !bms_equal(cur_em->em_relids, rel))
583 * If below an outer join, don't match constants: they're not as
584 * constant as they look.
586 if (cur_ec->ec_below_outer_join &&
590 if (opcintype == cur_em->em_datatype &&
591 equal(expr, cur_em->em_expr))
592 return cur_ec; /* Match! */
596 /* No match; does caller want a NULL result? */
601 * OK, build a new single-member EC
603 * Here, we must be sure that we construct the EC in the right context.
605 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
607 newec = makeNode(EquivalenceClass);
608 newec->ec_opfamilies = list_copy(opfamilies);
609 newec->ec_collation = collation;
610 newec->ec_members = NIL;
611 newec->ec_sources = NIL;
612 newec->ec_derives = NIL;
613 newec->ec_relids = NULL;
614 newec->ec_has_const = false;
615 newec->ec_has_volatile = contain_volatile_functions((Node *) expr);
616 newec->ec_below_outer_join = false;
617 newec->ec_broken = false;
618 newec->ec_sortref = sortref;
619 newec->ec_merged = NULL;
621 if (newec->ec_has_volatile && sortref == 0) /* should not happen */
622 elog(ERROR, "volatile EquivalenceClass has no sortref");
624 newem = add_eq_member(newec, copyObject(expr), pull_varnos((Node *) expr),
625 NULL, false, opcintype);
628 * add_eq_member doesn't check for volatile functions, set-returning
629 * functions, aggregates, or window functions, but such could appear in
630 * sort expressions; so we have to check whether its const-marking was
633 if (newec->ec_has_const)
635 if (newec->ec_has_volatile ||
636 expression_returns_set((Node *) expr) ||
637 contain_agg_clause((Node *) expr) ||
638 contain_window_function((Node *) expr))
640 newec->ec_has_const = false;
641 newem->em_is_const = false;
645 root->eq_classes = lappend(root->eq_classes, newec);
647 MemoryContextSwitchTo(oldcontext);
654 * generate_base_implied_equalities
655 * Generate any restriction clauses that we can deduce from equivalence
658 * When an EC contains pseudoconstants, our strategy is to generate
659 * "member = const1" clauses where const1 is the first constant member, for
660 * every other member (including other constants). If we are able to do this
661 * then we don't need any "var = var" comparisons because we've successfully
662 * constrained all the vars at their points of creation. If we fail to
663 * generate any of these clauses due to lack of cross-type operators, we fall
664 * back to the "ec_broken" strategy described below. (XXX if there are
665 * multiple constants of different types, it's possible that we might succeed
666 * in forming all the required clauses if we started from a different const
667 * member; but this seems a sufficiently hokey corner case to not be worth
668 * spending lots of cycles on.)
670 * For ECs that contain no pseudoconstants, we generate derived clauses
671 * "member1 = member2" for each pair of members belonging to the same base
672 * relation (actually, if there are more than two for the same base relation,
673 * we only need enough clauses to link each to each other). This provides
674 * the base case for the recursion: each row emitted by a base relation scan
675 * will constrain all computable members of the EC to be equal. As each
676 * join path is formed, we'll add additional derived clauses on-the-fly
677 * to maintain this invariant (see generate_join_implied_equalities).
679 * If the opfamilies used by the EC do not provide complete sets of cross-type
680 * equality operators, it is possible that we will fail to generate a clause
681 * that must be generated to maintain the invariant. (An example: given
682 * "WHERE a.x = b.y AND b.y = a.z", the scheme breaks down if we cannot
683 * generate "a.x = a.z" as a restriction clause for A.) In this case we mark
684 * the EC "ec_broken" and fall back to regurgitating its original source
685 * RestrictInfos at appropriate times. We do not try to retract any derived
686 * clauses already generated from the broken EC, so the resulting plan could
687 * be poor due to bad selectivity estimates caused by redundant clauses. But
688 * the correct solution to that is to fix the opfamilies ...
690 * Equality clauses derived by this function are passed off to
691 * process_implied_equality (in plan/initsplan.c) to be inserted into the
692 * restrictinfo datastructures. Note that this must be called after initial
693 * scanning of the quals and before Path construction begins.
695 * We make no attempt to avoid generating duplicate RestrictInfos here: we
696 * don't search ec_sources for matches, nor put the created RestrictInfos
697 * into ec_derives. Doing so would require some slightly ugly changes in
698 * initsplan.c's API, and there's no real advantage, because the clauses
699 * generated here can't duplicate anything we will generate for joins anyway.
702 generate_base_implied_equalities(PlannerInfo *root)
707 foreach(lc, root->eq_classes)
709 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
711 Assert(ec->ec_merged == NULL); /* else shouldn't be in list */
712 Assert(!ec->ec_broken); /* not yet anyway... */
714 /* Single-member ECs won't generate any deductions */
715 if (list_length(ec->ec_members) <= 1)
718 if (ec->ec_has_const)
719 generate_base_implied_equalities_const(root, ec);
721 generate_base_implied_equalities_no_const(root, ec);
723 /* Recover if we failed to generate required derived clauses */
725 generate_base_implied_equalities_broken(root, ec);
729 * This is also a handy place to mark base rels (which should all exist by
730 * now) with flags showing whether they have pending eclass joins.
732 for (rti = 1; rti < root->simple_rel_array_size; rti++)
734 RelOptInfo *brel = root->simple_rel_array[rti];
739 brel->has_eclass_joins = has_relevant_eclass_joinclause(root, brel);
744 * generate_base_implied_equalities when EC contains pseudoconstant(s)
747 generate_base_implied_equalities_const(PlannerInfo *root,
748 EquivalenceClass *ec)
750 EquivalenceMember *const_em = NULL;
754 * In the trivial case where we just had one "var = const" clause, push
755 * the original clause back into the main planner machinery. There is
756 * nothing to be gained by doing it differently, and we save the effort to
757 * re-build and re-analyze an equality clause that will be exactly
758 * equivalent to the old one.
760 if (list_length(ec->ec_members) == 2 &&
761 list_length(ec->ec_sources) == 1)
763 RestrictInfo *restrictinfo = (RestrictInfo *) linitial(ec->ec_sources);
765 if (bms_membership(restrictinfo->required_relids) != BMS_MULTIPLE)
767 distribute_restrictinfo_to_rels(root, restrictinfo);
772 /* Find the constant member to use */
773 foreach(lc, ec->ec_members)
775 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
777 if (cur_em->em_is_const)
783 Assert(const_em != NULL);
785 /* Generate a derived equality against each other member */
786 foreach(lc, ec->ec_members)
788 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
791 Assert(!cur_em->em_is_child); /* no children yet */
792 if (cur_em == const_em)
794 eq_op = select_equality_operator(ec,
796 const_em->em_datatype);
797 if (!OidIsValid(eq_op))
800 ec->ec_broken = true;
803 process_implied_equality(root, eq_op, ec->ec_collation,
804 cur_em->em_expr, const_em->em_expr,
805 bms_copy(ec->ec_relids),
806 bms_union(cur_em->em_nullable_relids,
807 const_em->em_nullable_relids),
808 ec->ec_below_outer_join,
809 cur_em->em_is_const);
814 * generate_base_implied_equalities when EC contains no pseudoconstants
817 generate_base_implied_equalities_no_const(PlannerInfo *root,
818 EquivalenceClass *ec)
820 EquivalenceMember **prev_ems;
824 * We scan the EC members once and track the last-seen member for each
825 * base relation. When we see another member of the same base relation,
826 * we generate "prev_mem = cur_mem". This results in the minimum number
827 * of derived clauses, but it's possible that it will fail when a
828 * different ordering would succeed. XXX FIXME: use a UNION-FIND
829 * algorithm similar to the way we build merged ECs. (Use a list-of-lists
832 prev_ems = (EquivalenceMember **)
833 palloc0(root->simple_rel_array_size * sizeof(EquivalenceMember *));
835 foreach(lc, ec->ec_members)
837 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
840 Assert(!cur_em->em_is_child); /* no children yet */
841 if (bms_membership(cur_em->em_relids) != BMS_SINGLETON)
843 relid = bms_singleton_member(cur_em->em_relids);
844 Assert(relid < root->simple_rel_array_size);
846 if (prev_ems[relid] != NULL)
848 EquivalenceMember *prev_em = prev_ems[relid];
851 eq_op = select_equality_operator(ec,
852 prev_em->em_datatype,
853 cur_em->em_datatype);
854 if (!OidIsValid(eq_op))
857 ec->ec_broken = true;
860 process_implied_equality(root, eq_op, ec->ec_collation,
861 prev_em->em_expr, cur_em->em_expr,
862 bms_copy(ec->ec_relids),
863 bms_union(prev_em->em_nullable_relids,
864 cur_em->em_nullable_relids),
865 ec->ec_below_outer_join,
868 prev_ems[relid] = cur_em;
874 * We also have to make sure that all the Vars used in the member clauses
875 * will be available at any join node we might try to reference them at.
876 * For the moment we force all the Vars to be available at all join nodes
877 * for this eclass. Perhaps this could be improved by doing some
878 * pre-analysis of which members we prefer to join, but it's no worse than
879 * what happened in the pre-8.3 code.
881 foreach(lc, ec->ec_members)
883 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
884 List *vars = pull_var_clause((Node *) cur_em->em_expr,
885 PVC_RECURSE_AGGREGATES,
886 PVC_INCLUDE_PLACEHOLDERS);
888 add_vars_to_targetlist(root, vars, ec->ec_relids, false);
894 * generate_base_implied_equalities cleanup after failure
896 * What we must do here is push any zero- or one-relation source RestrictInfos
897 * of the EC back into the main restrictinfo datastructures. Multi-relation
898 * clauses will be regurgitated later by generate_join_implied_equalities().
899 * (We do it this way to maintain continuity with the case that ec_broken
900 * becomes set only after we've gone up a join level or two.) However, for
901 * an EC that contains constants, we can adopt a simpler strategy and just
902 * throw back all the source RestrictInfos immediately; that works because
903 * we know that such an EC can't become broken later. (This rule justifies
904 * ignoring ec_has_const ECs in generate_join_implied_equalities, even when
908 generate_base_implied_equalities_broken(PlannerInfo *root,
909 EquivalenceClass *ec)
913 foreach(lc, ec->ec_sources)
915 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
917 if (ec->ec_has_const ||
918 bms_membership(restrictinfo->required_relids) != BMS_MULTIPLE)
919 distribute_restrictinfo_to_rels(root, restrictinfo);
925 * generate_join_implied_equalities
926 * Generate any join clauses that we can deduce from equivalence classes.
928 * At a join node, we must enforce restriction clauses sufficient to ensure
929 * that all equivalence-class members computable at that node are equal.
930 * Since the set of clauses to enforce can vary depending on which subset
931 * relations are the inputs, we have to compute this afresh for each join
932 * relation pair. Hence a fresh List of RestrictInfo nodes is built and
933 * passed back on each call.
935 * In addition to its use at join nodes, this can be applied to generate
936 * eclass-based join clauses for use in a parameterized scan of a base rel.
937 * The reason for the asymmetry of specifying the inner rel as a RelOptInfo
938 * and the outer rel by Relids is that this usage occurs before we have
939 * built any join RelOptInfos.
941 * An annoying special case for parameterized scans is that the inner rel can
942 * be an appendrel child (an "other rel"). In this case we must generate
943 * appropriate clauses using child EC members. add_child_rel_equivalences
944 * must already have been done for the child rel.
946 * The results are sufficient for use in merge, hash, and plain nestloop join
947 * methods. We do not worry here about selecting clauses that are optimal
948 * for use in a parameterized indexscan. indxpath.c makes its own selections
949 * of clauses to use, and if the ones we pick here are redundant with those,
950 * the extras will be eliminated at createplan time, using the parent_ec
951 * markers that we provide (see is_redundant_derived_clause()).
953 * Because the same join clauses are likely to be needed multiple times as
954 * we consider different join paths, we avoid generating multiple copies:
955 * whenever we select a particular pair of EquivalenceMembers to join,
956 * we check to see if the pair matches any original clause (in ec_sources)
957 * or previously-built clause (in ec_derives). This saves memory and allows
958 * re-use of information cached in RestrictInfos.
960 * join_relids should always equal bms_union(outer_relids, inner_rel->relids).
961 * We could simplify this function's API by computing it internally, but in
962 * all current uses, the caller has the value at hand anyway.
965 generate_join_implied_equalities(PlannerInfo *root,
968 RelOptInfo *inner_rel)
971 Relids inner_relids = inner_rel->relids;
972 Relids nominal_inner_relids;
973 Relids nominal_join_relids;
974 AppendRelInfo *inner_appinfo;
977 /* If inner rel is a child, extra setup work is needed */
978 if (inner_rel->reloptkind == RELOPT_OTHER_MEMBER_REL)
980 /* Lookup parent->child translation data */
981 inner_appinfo = find_childrel_appendrelinfo(root, inner_rel);
982 /* Construct relids for the parent rel */
983 nominal_inner_relids = bms_make_singleton(inner_appinfo->parent_relid);
984 /* ECs will be marked with the parent's relid, not the child's */
985 nominal_join_relids = bms_union(outer_relids, nominal_inner_relids);
989 inner_appinfo = NULL;
990 nominal_inner_relids = inner_relids;
991 nominal_join_relids = join_relids;
994 foreach(lc, root->eq_classes)
996 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
999 /* ECs containing consts do not need any further enforcement */
1000 if (ec->ec_has_const)
1003 /* Single-member ECs won't generate any deductions */
1004 if (list_length(ec->ec_members) <= 1)
1007 /* We can quickly ignore any that don't overlap the join, too */
1008 if (!bms_overlap(ec->ec_relids, nominal_join_relids))
1012 sublist = generate_join_implied_equalities_normal(root,
1018 /* Recover if we failed to generate required derived clauses */
1020 sublist = generate_join_implied_equalities_broken(root,
1022 nominal_join_relids,
1024 nominal_inner_relids,
1027 result = list_concat(result, sublist);
1034 * generate_join_implied_equalities for a still-valid EC
1037 generate_join_implied_equalities_normal(PlannerInfo *root,
1038 EquivalenceClass *ec,
1040 Relids outer_relids,
1041 Relids inner_relids)
1044 List *new_members = NIL;
1045 List *outer_members = NIL;
1046 List *inner_members = NIL;
1050 * First, scan the EC to identify member values that are computable at the
1051 * outer rel, at the inner rel, or at this relation but not in either
1052 * input rel. The outer-rel members should already be enforced equal,
1053 * likewise for the inner-rel members. We'll need to create clauses to
1054 * enforce that any newly computable members are all equal to each other
1055 * as well as to at least one input member, plus enforce at least one
1056 * outer-rel member equal to at least one inner-rel member.
1058 foreach(lc1, ec->ec_members)
1060 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc1);
1063 * We don't need to check explicitly for child EC members. This test
1064 * against join_relids will cause them to be ignored except when
1065 * considering a child inner rel, which is what we want.
1067 if (!bms_is_subset(cur_em->em_relids, join_relids))
1068 continue; /* not computable yet, or wrong child */
1070 if (bms_is_subset(cur_em->em_relids, outer_relids))
1071 outer_members = lappend(outer_members, cur_em);
1072 else if (bms_is_subset(cur_em->em_relids, inner_relids))
1073 inner_members = lappend(inner_members, cur_em);
1075 new_members = lappend(new_members, cur_em);
1079 * First, select the joinclause if needed. We can equate any one outer
1080 * member to any one inner member, but we have to find a datatype
1081 * combination for which an opfamily member operator exists. If we have
1082 * choices, we prefer simple Var members (possibly with RelabelType) since
1083 * these are (a) cheapest to compute at runtime and (b) most likely to
1084 * have useful statistics. Also, prefer operators that are also
1087 if (outer_members && inner_members)
1089 EquivalenceMember *best_outer_em = NULL;
1090 EquivalenceMember *best_inner_em = NULL;
1091 Oid best_eq_op = InvalidOid;
1092 int best_score = -1;
1093 RestrictInfo *rinfo;
1095 foreach(lc1, outer_members)
1097 EquivalenceMember *outer_em = (EquivalenceMember *) lfirst(lc1);
1100 foreach(lc2, inner_members)
1102 EquivalenceMember *inner_em = (EquivalenceMember *) lfirst(lc2);
1106 eq_op = select_equality_operator(ec,
1107 outer_em->em_datatype,
1108 inner_em->em_datatype);
1109 if (!OidIsValid(eq_op))
1112 if (IsA(outer_em->em_expr, Var) ||
1113 (IsA(outer_em->em_expr, RelabelType) &&
1114 IsA(((RelabelType *) outer_em->em_expr)->arg, Var)))
1116 if (IsA(inner_em->em_expr, Var) ||
1117 (IsA(inner_em->em_expr, RelabelType) &&
1118 IsA(((RelabelType *) inner_em->em_expr)->arg, Var)))
1120 if (op_hashjoinable(eq_op,
1121 exprType((Node *) outer_em->em_expr)))
1123 if (score > best_score)
1125 best_outer_em = outer_em;
1126 best_inner_em = inner_em;
1129 if (best_score == 3)
1130 break; /* no need to look further */
1133 if (best_score == 3)
1134 break; /* no need to look further */
1139 ec->ec_broken = true;
1144 * Create clause, setting parent_ec to mark it as redundant with other
1147 rinfo = create_join_clause(root, ec, best_eq_op,
1148 best_outer_em, best_inner_em,
1151 result = lappend(result, rinfo);
1155 * Now deal with building restrictions for any expressions that involve
1156 * Vars from both sides of the join. We have to equate all of these to
1157 * each other as well as to at least one old member (if any).
1159 * XXX as in generate_base_implied_equalities_no_const, we could be a lot
1160 * smarter here to avoid unnecessary failures in cross-type situations.
1161 * For now, use the same left-to-right method used there.
1165 List *old_members = list_concat(outer_members, inner_members);
1166 EquivalenceMember *prev_em = NULL;
1167 RestrictInfo *rinfo;
1169 /* For now, arbitrarily take the first old_member as the one to use */
1171 new_members = lappend(new_members, linitial(old_members));
1173 foreach(lc1, new_members)
1175 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc1);
1177 if (prev_em != NULL)
1181 eq_op = select_equality_operator(ec,
1182 prev_em->em_datatype,
1183 cur_em->em_datatype);
1184 if (!OidIsValid(eq_op))
1187 ec->ec_broken = true;
1190 /* do NOT set parent_ec, this qual is not redundant! */
1191 rinfo = create_join_clause(root, ec, eq_op,
1195 result = lappend(result, rinfo);
1205 * generate_join_implied_equalities cleanup after failure
1207 * Return any original RestrictInfos that are enforceable at this join.
1209 * In the case of a child inner relation, we have to translate the
1210 * original RestrictInfos from parent to child Vars.
1213 generate_join_implied_equalities_broken(PlannerInfo *root,
1214 EquivalenceClass *ec,
1215 Relids nominal_join_relids,
1216 Relids outer_relids,
1217 Relids nominal_inner_relids,
1218 AppendRelInfo *inner_appinfo)
1223 foreach(lc, ec->ec_sources)
1225 RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1226 Relids clause_relids = restrictinfo->required_relids;
1228 if (bms_is_subset(clause_relids, nominal_join_relids) &&
1229 !bms_is_subset(clause_relids, outer_relids) &&
1230 !bms_is_subset(clause_relids, nominal_inner_relids))
1231 result = lappend(result, restrictinfo);
1235 * If we have to translate, just brute-force apply adjust_appendrel_attrs
1236 * to all the RestrictInfos at once. This will result in returning
1237 * RestrictInfos that are not listed in ec_derives, but there shouldn't be
1238 * any duplication, and it's a sufficiently narrow corner case that we
1239 * shouldn't sweat too much over it anyway.
1242 result = (List *) adjust_appendrel_attrs(root, (Node *) result,
1250 * select_equality_operator
1251 * Select a suitable equality operator for comparing two EC members
1253 * Returns InvalidOid if no operator can be found for this datatype combination
1256 select_equality_operator(EquivalenceClass *ec, Oid lefttype, Oid righttype)
1260 foreach(lc, ec->ec_opfamilies)
1262 Oid opfamily = lfirst_oid(lc);
1265 opno = get_opfamily_member(opfamily, lefttype, righttype,
1266 BTEqualStrategyNumber);
1267 if (OidIsValid(opno))
1275 * create_join_clause
1276 * Find or make a RestrictInfo comparing the two given EC members
1277 * with the given operator.
1279 * parent_ec is either equal to ec (if the clause is a potentially-redundant
1280 * join clause) or NULL (if not). We have to treat this as part of the
1281 * match requirements --- it's possible that a clause comparing the same two
1282 * EMs is a join clause in one join path and a restriction clause in another.
1284 static RestrictInfo *
1285 create_join_clause(PlannerInfo *root,
1286 EquivalenceClass *ec, Oid opno,
1287 EquivalenceMember *leftem,
1288 EquivalenceMember *rightem,
1289 EquivalenceClass *parent_ec)
1291 RestrictInfo *rinfo;
1293 MemoryContext oldcontext;
1296 * Search to see if we already built a RestrictInfo for this pair of
1297 * EquivalenceMembers. We can use either original source clauses or
1298 * previously-derived clauses. The check on opno is probably redundant,
1301 foreach(lc, ec->ec_sources)
1303 rinfo = (RestrictInfo *) lfirst(lc);
1304 if (rinfo->left_em == leftem &&
1305 rinfo->right_em == rightem &&
1306 rinfo->parent_ec == parent_ec &&
1307 opno == ((OpExpr *) rinfo->clause)->opno)
1311 foreach(lc, ec->ec_derives)
1313 rinfo = (RestrictInfo *) lfirst(lc);
1314 if (rinfo->left_em == leftem &&
1315 rinfo->right_em == rightem &&
1316 rinfo->parent_ec == parent_ec &&
1317 opno == ((OpExpr *) rinfo->clause)->opno)
1322 * Not there, so build it, in planner context so we can re-use it. (Not
1323 * important in normal planning, but definitely so in GEQO.)
1325 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
1327 rinfo = build_implied_join_equality(opno,
1331 bms_union(leftem->em_relids,
1332 rightem->em_relids),
1333 bms_union(leftem->em_nullable_relids,
1334 rightem->em_nullable_relids));
1336 /* Mark the clause as redundant, or not */
1337 rinfo->parent_ec = parent_ec;
1340 * We know the correct values for left_ec/right_ec, ie this particular EC,
1341 * so we can just set them directly instead of forcing another lookup.
1343 rinfo->left_ec = ec;
1344 rinfo->right_ec = ec;
1346 /* Mark it as usable with these EMs */
1347 rinfo->left_em = leftem;
1348 rinfo->right_em = rightem;
1349 /* and save it for possible re-use */
1350 ec->ec_derives = lappend(ec->ec_derives, rinfo);
1352 MemoryContextSwitchTo(oldcontext);
1359 * reconsider_outer_join_clauses
1360 * Re-examine any outer-join clauses that were set aside by
1361 * distribute_qual_to_rels(), and see if we can derive any
1362 * EquivalenceClasses from them. Then, if they were not made
1363 * redundant, push them out into the regular join-clause lists.
1365 * When we have mergejoinable clauses A = B that are outer-join clauses,
1366 * we can't blindly combine them with other clauses A = C to deduce B = C,
1367 * since in fact the "equality" A = B won't necessarily hold above the
1368 * outer join (one of the variables might be NULL instead). Nonetheless
1369 * there are cases where we can add qual clauses using transitivity.
1371 * One case that we look for here is an outer-join clause OUTERVAR = INNERVAR
1372 * for which there is also an equivalence clause OUTERVAR = CONSTANT.
1373 * It is safe and useful to push a clause INNERVAR = CONSTANT into the
1374 * evaluation of the inner (nullable) relation, because any inner rows not
1375 * meeting this condition will not contribute to the outer-join result anyway.
1376 * (Any outer rows they could join to will be eliminated by the pushed-down
1377 * equivalence clause.)
1379 * Note that the above rule does not work for full outer joins; nor is it
1380 * very interesting to consider cases where the generated equivalence clause
1381 * would involve relations outside the outer join, since such clauses couldn't
1382 * be pushed into the inner side's scan anyway. So the restriction to
1383 * outervar = pseudoconstant is not really giving up anything.
1385 * For full-join cases, we can only do something useful if it's a FULL JOIN
1386 * USING and a merged column has an equivalence MERGEDVAR = CONSTANT.
1387 * By the time it gets here, the merged column will look like
1388 * COALESCE(LEFTVAR, RIGHTVAR)
1389 * and we will have a full-join clause LEFTVAR = RIGHTVAR that we can match
1390 * the COALESCE expression to. In this situation we can push LEFTVAR = CONSTANT
1391 * and RIGHTVAR = CONSTANT into the input relations, since any rows not
1392 * meeting these conditions cannot contribute to the join result.
1394 * Again, there isn't any traction to be gained by trying to deal with
1395 * clauses comparing a mergedvar to a non-pseudoconstant. So we can make
1396 * use of the EquivalenceClasses to search for matching variables that were
1397 * equivalenced to constants. The interesting outer-join clauses were
1398 * accumulated for us by distribute_qual_to_rels.
1400 * When we find one of these cases, we implement the changes we want by
1401 * generating a new equivalence clause INNERVAR = CONSTANT (or LEFTVAR, etc)
1402 * and pushing it into the EquivalenceClass structures. This is because we
1403 * may already know that INNERVAR is equivalenced to some other var(s), and
1404 * we'd like the constant to propagate to them too. Note that it would be
1405 * unsafe to merge any existing EC for INNERVAR with the OUTERVAR's EC ---
1406 * that could result in propagating constant restrictions from
1407 * INNERVAR to OUTERVAR, which would be very wrong.
1409 * It's possible that the INNERVAR is also an OUTERVAR for some other
1410 * outer-join clause, in which case the process can be repeated. So we repeat
1411 * looping over the lists of clauses until no further deductions can be made.
1412 * Whenever we do make a deduction, we remove the generating clause from the
1413 * lists, since we don't want to make the same deduction twice.
1415 * If we don't find any match for a set-aside outer join clause, we must
1416 * throw it back into the regular joinclause processing by passing it to
1417 * distribute_restrictinfo_to_rels(). If we do generate a derived clause,
1418 * however, the outer-join clause is redundant. We still throw it back,
1419 * because otherwise the join will be seen as a clauseless join and avoided
1420 * during join order searching; but we mark it as redundant to keep from
1421 * messing up the joinrel's size estimate. (This behavior means that the
1422 * API for this routine is uselessly complex: we could have just put all
1423 * the clauses into the regular processing initially. We keep it because
1424 * someday we might want to do something else, such as inserting "dummy"
1425 * joinclauses instead of real ones.)
1427 * Outer join clauses that are marked outerjoin_delayed are special: this
1428 * condition means that one or both VARs might go to null due to a lower
1429 * outer join. We can still push a constant through the clause, but only
1430 * if its operator is strict; and we *have to* throw the clause back into
1431 * regular joinclause processing. By keeping the strict join clause,
1432 * we ensure that any null-extended rows that are mistakenly generated due
1433 * to suppressing rows not matching the constant will be rejected at the
1434 * upper outer join. (This doesn't work for full-join clauses.)
1437 reconsider_outer_join_clauses(PlannerInfo *root)
1444 /* Outer loop repeats until we find no more deductions */
1449 /* Process the LEFT JOIN clauses */
1451 for (cell = list_head(root->left_join_clauses); cell; cell = next)
1453 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1456 if (reconsider_outer_join_clause(root, rinfo, true))
1459 /* remove it from the list */
1460 root->left_join_clauses =
1461 list_delete_cell(root->left_join_clauses, cell, prev);
1462 /* we throw it back anyway (see notes above) */
1463 /* but the thrown-back clause has no extra selectivity */
1464 rinfo->norm_selec = 2.0;
1465 rinfo->outer_selec = 1.0;
1466 distribute_restrictinfo_to_rels(root, rinfo);
1472 /* Process the RIGHT JOIN clauses */
1474 for (cell = list_head(root->right_join_clauses); cell; cell = next)
1476 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1479 if (reconsider_outer_join_clause(root, rinfo, false))
1482 /* remove it from the list */
1483 root->right_join_clauses =
1484 list_delete_cell(root->right_join_clauses, cell, prev);
1485 /* we throw it back anyway (see notes above) */
1486 /* but the thrown-back clause has no extra selectivity */
1487 rinfo->norm_selec = 2.0;
1488 rinfo->outer_selec = 1.0;
1489 distribute_restrictinfo_to_rels(root, rinfo);
1495 /* Process the FULL JOIN clauses */
1497 for (cell = list_head(root->full_join_clauses); cell; cell = next)
1499 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1502 if (reconsider_full_join_clause(root, rinfo))
1505 /* remove it from the list */
1506 root->full_join_clauses =
1507 list_delete_cell(root->full_join_clauses, cell, prev);
1508 /* we throw it back anyway (see notes above) */
1509 /* but the thrown-back clause has no extra selectivity */
1510 rinfo->norm_selec = 2.0;
1511 rinfo->outer_selec = 1.0;
1512 distribute_restrictinfo_to_rels(root, rinfo);
1519 /* Now, any remaining clauses have to be thrown back */
1520 foreach(cell, root->left_join_clauses)
1522 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1524 distribute_restrictinfo_to_rels(root, rinfo);
1526 foreach(cell, root->right_join_clauses)
1528 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1530 distribute_restrictinfo_to_rels(root, rinfo);
1532 foreach(cell, root->full_join_clauses)
1534 RestrictInfo *rinfo = (RestrictInfo *) lfirst(cell);
1536 distribute_restrictinfo_to_rels(root, rinfo);
1541 * reconsider_outer_join_clauses for a single LEFT/RIGHT JOIN clause
1543 * Returns TRUE if we were able to propagate a constant through the clause.
1546 reconsider_outer_join_clause(PlannerInfo *root, RestrictInfo *rinfo,
1556 Relids inner_relids,
1557 inner_nullable_relids;
1560 Assert(is_opclause(rinfo->clause));
1561 opno = ((OpExpr *) rinfo->clause)->opno;
1562 collation = ((OpExpr *) rinfo->clause)->inputcollid;
1564 /* If clause is outerjoin_delayed, operator must be strict */
1565 if (rinfo->outerjoin_delayed && !op_strict(opno))
1568 /* Extract needed info from the clause */
1569 op_input_types(opno, &left_type, &right_type);
1572 outervar = (Expr *) get_leftop(rinfo->clause);
1573 innervar = (Expr *) get_rightop(rinfo->clause);
1574 inner_datatype = right_type;
1575 inner_relids = rinfo->right_relids;
1579 outervar = (Expr *) get_rightop(rinfo->clause);
1580 innervar = (Expr *) get_leftop(rinfo->clause);
1581 inner_datatype = left_type;
1582 inner_relids = rinfo->left_relids;
1584 inner_nullable_relids = bms_intersect(inner_relids,
1585 rinfo->nullable_relids);
1587 /* Scan EquivalenceClasses for a match to outervar */
1588 foreach(lc1, root->eq_classes)
1590 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
1594 /* Ignore EC unless it contains pseudoconstants */
1595 if (!cur_ec->ec_has_const)
1597 /* Never match to a volatile EC */
1598 if (cur_ec->ec_has_volatile)
1600 /* It has to match the outer-join clause as to semantics, too */
1601 if (collation != cur_ec->ec_collation)
1603 if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
1605 /* Does it contain a match to outervar? */
1607 foreach(lc2, cur_ec->ec_members)
1609 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1611 Assert(!cur_em->em_is_child); /* no children yet */
1612 if (equal(outervar, cur_em->em_expr))
1619 continue; /* no match, so ignore this EC */
1622 * Yes it does! Try to generate a clause INNERVAR = CONSTANT for each
1623 * CONSTANT in the EC. Note that we must succeed with at least one
1624 * constant before we can decide to throw away the outer-join clause.
1627 foreach(lc2, cur_ec->ec_members)
1629 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1631 RestrictInfo *newrinfo;
1633 if (!cur_em->em_is_const)
1634 continue; /* ignore non-const members */
1635 eq_op = select_equality_operator(cur_ec,
1637 cur_em->em_datatype);
1638 if (!OidIsValid(eq_op))
1639 continue; /* can't generate equality */
1640 newrinfo = build_implied_join_equality(eq_op,
1641 cur_ec->ec_collation,
1644 bms_copy(inner_relids),
1645 bms_copy(inner_nullable_relids));
1646 if (process_equivalence(root, newrinfo, true))
1651 * If we were able to equate INNERVAR to any constant, report success.
1652 * Otherwise, fall out of the search loop, since we know the OUTERVAR
1653 * appears in at most one EC.
1661 return false; /* failed to make any deduction */
1665 * reconsider_outer_join_clauses for a single FULL JOIN clause
1667 * Returns TRUE if we were able to propagate a constant through the clause.
1670 reconsider_full_join_clause(PlannerInfo *root, RestrictInfo *rinfo)
1680 left_nullable_relids,
1681 right_nullable_relids;
1684 /* Can't use an outerjoin_delayed clause here */
1685 if (rinfo->outerjoin_delayed)
1688 /* Extract needed info from the clause */
1689 Assert(is_opclause(rinfo->clause));
1690 opno = ((OpExpr *) rinfo->clause)->opno;
1691 collation = ((OpExpr *) rinfo->clause)->inputcollid;
1692 op_input_types(opno, &left_type, &right_type);
1693 leftvar = (Expr *) get_leftop(rinfo->clause);
1694 rightvar = (Expr *) get_rightop(rinfo->clause);
1695 left_relids = rinfo->left_relids;
1696 right_relids = rinfo->right_relids;
1697 left_nullable_relids = bms_intersect(left_relids,
1698 rinfo->nullable_relids);
1699 right_nullable_relids = bms_intersect(right_relids,
1700 rinfo->nullable_relids);
1702 foreach(lc1, root->eq_classes)
1704 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
1705 EquivalenceMember *coal_em = NULL;
1711 /* Ignore EC unless it contains pseudoconstants */
1712 if (!cur_ec->ec_has_const)
1714 /* Never match to a volatile EC */
1715 if (cur_ec->ec_has_volatile)
1717 /* It has to match the outer-join clause as to semantics, too */
1718 if (collation != cur_ec->ec_collation)
1720 if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
1724 * Does it contain a COALESCE(leftvar, rightvar) construct?
1726 * We can assume the COALESCE() inputs are in the same order as the
1727 * join clause, since both were automatically generated in the cases
1730 * XXX currently this may fail to match in cross-type cases because
1731 * the COALESCE will contain typecast operations while the join clause
1732 * may not (if there is a cross-type mergejoin operator available for
1733 * the two column types). Is it OK to strip implicit coercions from
1734 * the COALESCE arguments?
1737 foreach(lc2, cur_ec->ec_members)
1739 coal_em = (EquivalenceMember *) lfirst(lc2);
1740 Assert(!coal_em->em_is_child); /* no children yet */
1741 if (IsA(coal_em->em_expr, CoalesceExpr))
1743 CoalesceExpr *cexpr = (CoalesceExpr *) coal_em->em_expr;
1747 if (list_length(cexpr->args) != 2)
1749 cfirst = (Node *) linitial(cexpr->args);
1750 csecond = (Node *) lsecond(cexpr->args);
1752 if (equal(leftvar, cfirst) && equal(rightvar, csecond))
1760 continue; /* no match, so ignore this EC */
1763 * Yes it does! Try to generate clauses LEFTVAR = CONSTANT and
1764 * RIGHTVAR = CONSTANT for each CONSTANT in the EC. Note that we must
1765 * succeed with at least one constant for each var before we can
1766 * decide to throw away the outer-join clause.
1768 matchleft = matchright = false;
1769 foreach(lc2, cur_ec->ec_members)
1771 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1773 RestrictInfo *newrinfo;
1775 if (!cur_em->em_is_const)
1776 continue; /* ignore non-const members */
1777 eq_op = select_equality_operator(cur_ec,
1779 cur_em->em_datatype);
1780 if (OidIsValid(eq_op))
1782 newrinfo = build_implied_join_equality(eq_op,
1783 cur_ec->ec_collation,
1786 bms_copy(left_relids),
1787 bms_copy(left_nullable_relids));
1788 if (process_equivalence(root, newrinfo, true))
1791 eq_op = select_equality_operator(cur_ec,
1793 cur_em->em_datatype);
1794 if (OidIsValid(eq_op))
1796 newrinfo = build_implied_join_equality(eq_op,
1797 cur_ec->ec_collation,
1800 bms_copy(right_relids),
1801 bms_copy(right_nullable_relids));
1802 if (process_equivalence(root, newrinfo, true))
1808 * If we were able to equate both vars to constants, we're done, and
1809 * we can throw away the full-join clause as redundant. Moreover, we
1810 * can remove the COALESCE entry from the EC, since the added
1811 * restrictions ensure it will always have the expected value. (We
1812 * don't bother trying to update ec_relids or ec_sources.)
1814 if (matchleft && matchright)
1816 cur_ec->ec_members = list_delete_ptr(cur_ec->ec_members, coal_em);
1821 * Otherwise, fall out of the search loop, since we know the COALESCE
1822 * appears in at most one EC (XXX might stop being true if we allow
1823 * stripping of coercions above?)
1828 return false; /* failed to make any deduction */
1834 * Detect whether two expressions are known equal due to equivalence
1837 * Actually, this only shows that the expressions are equal according
1838 * to some opfamily's notion of equality --- but we only use it for
1839 * selectivity estimation, so a fuzzy idea of equality is OK.
1841 * Note: does not bother to check for "equal(item1, item2)"; caller must
1842 * check that case if it's possible to pass identical items.
1845 exprs_known_equal(PlannerInfo *root, Node *item1, Node *item2)
1849 foreach(lc1, root->eq_classes)
1851 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
1852 bool item1member = false;
1853 bool item2member = false;
1856 /* Never match to a volatile EC */
1857 if (ec->ec_has_volatile)
1860 foreach(lc2, ec->ec_members)
1862 EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
1864 if (em->em_is_child)
1865 continue; /* ignore children here */
1866 if (equal(item1, em->em_expr))
1868 else if (equal(item2, em->em_expr))
1870 /* Exit as soon as equality is proven */
1871 if (item1member && item2member)
1880 * add_child_rel_equivalences
1881 * Search for EC members that reference the parent_rel, and
1882 * add transformed members referencing the child_rel.
1884 * Note that this function won't be called at all unless we have at least some
1885 * reason to believe that the EC members it generates will be useful.
1887 * parent_rel and child_rel could be derived from appinfo, but since the
1888 * caller has already computed them, we might as well just pass them in.
1891 add_child_rel_equivalences(PlannerInfo *root,
1892 AppendRelInfo *appinfo,
1893 RelOptInfo *parent_rel,
1894 RelOptInfo *child_rel)
1898 foreach(lc1, root->eq_classes)
1900 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
1904 * If this EC contains a volatile expression, then generating child
1905 * EMs would be downright dangerous, so skip it. We rely on a
1906 * volatile EC having only one EM.
1908 if (cur_ec->ec_has_volatile)
1911 /* No point in searching if parent rel not mentioned in eclass */
1912 if (!bms_is_subset(parent_rel->relids, cur_ec->ec_relids))
1915 foreach(lc2, cur_ec->ec_members)
1917 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1919 if (cur_em->em_is_const || cur_em->em_is_child)
1920 continue; /* ignore consts and children here */
1922 /* Does it reference parent_rel? */
1923 if (bms_overlap(cur_em->em_relids, parent_rel->relids))
1925 /* Yes, generate transformed child version */
1928 Relids new_nullable_relids;
1930 child_expr = (Expr *)
1931 adjust_appendrel_attrs(root,
1932 (Node *) cur_em->em_expr,
1936 * Transform em_relids to match. Note we do *not* do
1937 * pull_varnos(child_expr) here, as for example the
1938 * transformation might have substituted a constant, but we
1939 * don't want the child member to be marked as constant.
1941 new_relids = bms_difference(cur_em->em_relids,
1942 parent_rel->relids);
1943 new_relids = bms_add_members(new_relids, child_rel->relids);
1946 * And likewise for nullable_relids. Note this code assumes
1947 * parent and child relids are singletons.
1949 new_nullable_relids = cur_em->em_nullable_relids;
1950 if (bms_overlap(new_nullable_relids, parent_rel->relids))
1952 new_nullable_relids = bms_difference(new_nullable_relids,
1953 parent_rel->relids);
1954 new_nullable_relids = bms_add_members(new_nullable_relids,
1958 (void) add_eq_member(cur_ec, child_expr,
1959 new_relids, new_nullable_relids,
1960 true, cur_em->em_datatype);
1968 * mutate_eclass_expressions
1969 * Apply an expression tree mutator to all expressions stored in
1970 * equivalence classes.
1972 * This is a bit of a hack ... it's currently needed only by planagg.c,
1973 * which needs to do a global search-and-replace of MIN/MAX Aggrefs
1974 * after eclasses are already set up. Without changing the eclasses too,
1975 * subsequent matching of ORDER BY clauses would fail.
1977 * Note that we assume the mutation won't affect relation membership or any
1978 * other properties we keep track of (which is a bit bogus, but by the time
1979 * planagg.c runs, it no longer matters). Also we must be called in the
1980 * main planner memory context.
1983 mutate_eclass_expressions(PlannerInfo *root,
1984 Node *(*mutator) (),
1989 foreach(lc1, root->eq_classes)
1991 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
1994 foreach(lc2, cur_ec->ec_members)
1996 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
1998 cur_em->em_expr = (Expr *)
1999 mutator((Node *) cur_em->em_expr, context);
2006 * generate_implied_equalities_for_indexcol
2007 * Create EC-derived joinclauses usable with a specific index column.
2009 * We assume that any given index column could appear in only one EC.
2010 * (This should be true in all but the most pathological cases, and if it
2011 * isn't, we stop on the first match anyway.) Therefore, what we return
2012 * is a redundant list of clauses equating the index column to each of
2013 * the other-relation values it is known to be equal to. Any one of
2014 * these clauses can be used to create a parameterized indexscan, and there
2015 * is no value in using more than one. (But it *is* worthwhile to create
2016 * a separate parameterized path for each one, since that leads to different
2019 * The caller can pass a Relids set of rels we aren't interested in joining
2020 * to, so as to save the work of creating useless clauses.
2023 generate_implied_equalities_for_indexcol(PlannerInfo *root,
2024 IndexOptInfo *index,
2026 Relids prohibited_rels)
2029 RelOptInfo *rel = index->rel;
2030 bool is_child_rel = (rel->reloptkind == RELOPT_OTHER_MEMBER_REL);
2034 /* If it's a child rel, we'll need to know what its parent is */
2036 parent_relid = find_childrel_appendrelinfo(root, rel)->parent_relid;
2038 parent_relid = 0; /* not used, but keep compiler quiet */
2040 foreach(lc1, root->eq_classes)
2042 EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2043 EquivalenceMember *cur_em;
2047 * Won't generate joinclauses if const or single-member (the latter
2048 * test covers the volatile case too)
2050 if (cur_ec->ec_has_const || list_length(cur_ec->ec_members) <= 1)
2054 * No point in searching if rel not mentioned in eclass (but we can't
2055 * tell that for a child rel).
2057 if (!is_child_rel &&
2058 !bms_is_subset(rel->relids, cur_ec->ec_relids))
2062 * Scan members, looking for a match to the indexable column. Note
2063 * that child EC members are considered, but only when they belong to
2064 * the target relation. (Unlike regular members, the same expression
2065 * could be a child member of more than one EC. Therefore, it's
2066 * potentially order-dependent which EC a child relation's index
2067 * column gets matched to. This is annoying but it only happens in
2068 * corner cases, so for now we live with just reporting the first
2069 * match. See also get_eclass_for_sort_expr.)
2072 foreach(lc2, cur_ec->ec_members)
2074 cur_em = (EquivalenceMember *) lfirst(lc2);
2075 if (bms_equal(cur_em->em_relids, rel->relids) &&
2076 eclass_member_matches_indexcol(cur_ec, cur_em,
2086 * Found our match. Scan the other EC members and attempt to generate
2089 foreach(lc2, cur_ec->ec_members)
2091 EquivalenceMember *other_em = (EquivalenceMember *) lfirst(lc2);
2093 RestrictInfo *rinfo;
2095 if (other_em->em_is_child)
2096 continue; /* ignore children here */
2098 /* Make sure it'll be a join to a different rel */
2099 if (other_em == cur_em ||
2100 bms_overlap(other_em->em_relids, rel->relids))
2103 /* Forget it if caller doesn't want joins to this rel */
2104 if (bms_overlap(other_em->em_relids, prohibited_rels))
2108 * Also, if this is a child rel, avoid generating a useless join
2109 * to its parent rel.
2112 bms_is_member(parent_relid, other_em->em_relids))
2115 eq_op = select_equality_operator(cur_ec,
2116 cur_em->em_datatype,
2117 other_em->em_datatype);
2118 if (!OidIsValid(eq_op))
2121 /* set parent_ec to mark as redundant with other joinclauses */
2122 rinfo = create_join_clause(root, cur_ec, eq_op,
2126 result = lappend(result, rinfo);
2130 * If somehow we failed to create any join clauses, we might as well
2131 * keep scanning the ECs for another match. But if we did make any,
2132 * we're done, because we don't want to return non-redundant clauses.
2142 * have_relevant_eclass_joinclause
2143 * Detect whether there is an EquivalenceClass that could produce
2144 * a joinclause involving the two given relations.
2146 * This is essentially a very cut-down version of
2147 * generate_join_implied_equalities(). Note it's OK to occasionally say "yes"
2148 * incorrectly. Hence we don't bother with details like whether the lack of a
2149 * cross-type operator might prevent the clause from actually being generated.
2152 have_relevant_eclass_joinclause(PlannerInfo *root,
2153 RelOptInfo *rel1, RelOptInfo *rel2)
2157 foreach(lc1, root->eq_classes)
2159 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2162 * Won't generate joinclauses if single-member (this test covers the
2163 * volatile case too)
2165 if (list_length(ec->ec_members) <= 1)
2169 * We do not need to examine the individual members of the EC, because
2170 * all that we care about is whether each rel overlaps the relids of
2171 * at least one member, and a test on ec_relids is sufficient to prove
2172 * that. (As with have_relevant_joinclause(), it is not necessary
2173 * that the EC be able to form a joinclause relating exactly the two
2174 * given rels, only that it be able to form a joinclause mentioning
2175 * both, and this will surely be true if both of them overlap
2178 * Note we don't test ec_broken; if we did, we'd need a separate code
2179 * path to look through ec_sources. Checking the membership anyway is
2180 * OK as a possibly-overoptimistic heuristic.
2182 * We don't test ec_has_const either, even though a const eclass won't
2183 * generate real join clauses. This is because if we had "WHERE a.x =
2184 * b.y and a.x = 42", it is worth considering a join between a and b,
2185 * since the join result is likely to be small even though it'll end
2186 * up being an unqualified nestloop.
2188 if (bms_overlap(rel1->relids, ec->ec_relids) &&
2189 bms_overlap(rel2->relids, ec->ec_relids))
2198 * has_relevant_eclass_joinclause
2199 * Detect whether there is an EquivalenceClass that could produce
2200 * a joinclause involving the given relation and anything else.
2202 * This is the same as have_relevant_eclass_joinclause with the other rel
2203 * implicitly defined as "everything else in the query".
2206 has_relevant_eclass_joinclause(PlannerInfo *root, RelOptInfo *rel1)
2210 foreach(lc1, root->eq_classes)
2212 EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2215 * Won't generate joinclauses if single-member (this test covers the
2216 * volatile case too)
2218 if (list_length(ec->ec_members) <= 1)
2222 * Per the comment in have_relevant_eclass_joinclause, it's sufficient
2223 * to find an EC that mentions both this rel and some other rel.
2225 if (bms_overlap(rel1->relids, ec->ec_relids) &&
2226 !bms_is_subset(ec->ec_relids, rel1->relids))
2235 * eclass_useful_for_merging
2236 * Detect whether the EC could produce any mergejoinable join clauses
2237 * against the specified relation.
2239 * This is just a heuristic test and doesn't have to be exact; it's better
2240 * to say "yes" incorrectly than "no". Hence we don't bother with details
2241 * like whether the lack of a cross-type operator might prevent the clause
2242 * from actually being generated.
2245 eclass_useful_for_merging(EquivalenceClass *eclass,
2250 Assert(!eclass->ec_merged);
2253 * Won't generate joinclauses if const or single-member (the latter test
2254 * covers the volatile case too)
2256 if (eclass->ec_has_const || list_length(eclass->ec_members) <= 1)
2260 * Note we don't test ec_broken; if we did, we'd need a separate code path
2261 * to look through ec_sources. Checking the members anyway is OK as a
2262 * possibly-overoptimistic heuristic.
2265 /* If rel already includes all members of eclass, no point in searching */
2266 if (bms_is_subset(eclass->ec_relids, rel->relids))
2269 /* To join, we need a member not in the given rel */
2270 foreach(lc, eclass->ec_members)
2272 EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
2274 if (cur_em->em_is_child)
2275 continue; /* ignore children here */
2277 if (!bms_overlap(cur_em->em_relids, rel->relids))
2286 * is_redundant_derived_clause
2287 * Test whether rinfo is derived from same EC as any clause in clauselist;
2288 * if so, it can be presumed to represent a condition that's redundant
2289 * with that member of the list.
2292 is_redundant_derived_clause(RestrictInfo *rinfo, List *clauselist)
2294 EquivalenceClass *parent_ec = rinfo->parent_ec;
2297 /* Fail if it's not a potentially-redundant clause from some EC */
2298 if (parent_ec == NULL)
2301 foreach(lc, clauselist)
2303 RestrictInfo *otherrinfo = (RestrictInfo *) lfirst(lc);
2305 if (otherrinfo->parent_ec == parent_ec)