Line data Source code
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) 1996-2025, 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 "common/hashfn.h"
24 : #include "nodes/makefuncs.h"
25 : #include "nodes/nodeFuncs.h"
26 : #include "optimizer/appendinfo.h"
27 : #include "optimizer/clauses.h"
28 : #include "optimizer/optimizer.h"
29 : #include "optimizer/pathnode.h"
30 : #include "optimizer/paths.h"
31 : #include "optimizer/planmain.h"
32 : #include "optimizer/restrictinfo.h"
33 : #include "rewrite/rewriteManip.h"
34 : #include "utils/lsyscache.h"
35 :
36 :
37 : static EquivalenceMember *make_eq_member(EquivalenceClass *ec,
38 : Expr *expr, Relids relids,
39 : JoinDomain *jdomain,
40 : EquivalenceMember *parent,
41 : Oid datatype);
42 : static EquivalenceMember *add_eq_member(EquivalenceClass *ec,
43 : Expr *expr, Relids relids,
44 : JoinDomain *jdomain,
45 : Oid datatype);
46 : static EquivalenceMember *add_child_eq_member(PlannerInfo *root,
47 : EquivalenceClass *ec,
48 : int ec_index, Expr *expr,
49 : Relids relids,
50 : JoinDomain *jdomain,
51 : EquivalenceMember *parent_em,
52 : Oid datatype,
53 : Index child_relid);
54 : static void generate_base_implied_equalities_const(PlannerInfo *root,
55 : EquivalenceClass *ec);
56 : static void generate_base_implied_equalities_no_const(PlannerInfo *root,
57 : EquivalenceClass *ec);
58 : static void generate_base_implied_equalities_broken(PlannerInfo *root,
59 : EquivalenceClass *ec);
60 : static List *generate_join_implied_equalities_normal(PlannerInfo *root,
61 : EquivalenceClass *ec,
62 : Relids join_relids,
63 : Relids outer_relids,
64 : Relids inner_relids);
65 : static List *generate_join_implied_equalities_broken(PlannerInfo *root,
66 : EquivalenceClass *ec,
67 : Relids nominal_join_relids,
68 : Relids outer_relids,
69 : Relids nominal_inner_relids,
70 : RelOptInfo *inner_rel);
71 : static Oid select_equality_operator(EquivalenceClass *ec,
72 : Oid lefttype, Oid righttype);
73 : static RestrictInfo *create_join_clause(PlannerInfo *root,
74 : EquivalenceClass *ec, Oid opno,
75 : EquivalenceMember *leftem,
76 : EquivalenceMember *rightem,
77 : EquivalenceClass *parent_ec);
78 : static bool reconsider_outer_join_clause(PlannerInfo *root,
79 : OuterJoinClauseInfo *ojcinfo,
80 : bool outer_on_left);
81 : static bool reconsider_full_join_clause(PlannerInfo *root,
82 : OuterJoinClauseInfo *ojcinfo);
83 : static JoinDomain *find_join_domain(PlannerInfo *root, Relids relids);
84 : static Bitmapset *get_eclass_indexes_for_relids(PlannerInfo *root,
85 : Relids relids);
86 : static Bitmapset *get_common_eclass_indexes(PlannerInfo *root, Relids relids1,
87 : Relids relids2);
88 : static void ec_build_derives_hash(PlannerInfo *root, EquivalenceClass *ec);
89 : static void ec_add_derived_clauses(EquivalenceClass *ec, List *clauses);
90 : static void ec_add_derived_clause(EquivalenceClass *ec, RestrictInfo *clause);
91 : static void ec_add_clause_to_derives_hash(EquivalenceClass *ec, RestrictInfo *rinfo);
92 : static RestrictInfo *ec_search_clause_for_ems(PlannerInfo *root, EquivalenceClass *ec,
93 : EquivalenceMember *leftem,
94 : EquivalenceMember *rightem,
95 : EquivalenceClass *parent_ec);
96 : static RestrictInfo *ec_search_derived_clause_for_ems(PlannerInfo *root,
97 : EquivalenceClass *ec,
98 : EquivalenceMember *leftem,
99 : EquivalenceMember *rightem,
100 : EquivalenceClass *parent_ec);
101 :
102 : /*
103 : * Hash key identifying a derived clause.
104 : *
105 : * This structure should not be filled manually. Use fill_ec_derives_key() to
106 : * set it up in canonical form.
107 : */
108 : typedef struct
109 : {
110 : EquivalenceMember *em1;
111 : EquivalenceMember *em2;
112 : EquivalenceClass *parent_ec;
113 : } ECDerivesKey;
114 :
115 : /* Hash table entry in ec_derives_hash. */
116 : typedef struct
117 : {
118 : uint32 status;
119 : ECDerivesKey key;
120 : RestrictInfo *rinfo;
121 : } ECDerivesEntry;
122 :
123 : /* Threshold for switching from list to hash table */
124 : #define EC_DERIVES_HASH_THRESHOLD 32
125 :
126 : #define SH_PREFIX derives
127 : #define SH_ELEMENT_TYPE ECDerivesEntry
128 : #define SH_KEY_TYPE ECDerivesKey
129 : #define SH_KEY key
130 : #define SH_HASH_KEY(tb, key) \
131 : hash_bytes((const unsigned char *) &(key), sizeof(ECDerivesKey))
132 : #define SH_EQUAL(tb, a, b) \
133 : ((a).em1 == (b).em1 && (a).em2 == (b).em2 && (a).parent_ec == (b).parent_ec)
134 : #define SH_SCOPE static inline
135 : #define SH_DECLARE
136 : #define SH_DEFINE
137 : #include "lib/simplehash.h"
138 :
139 : /*
140 : * process_equivalence
141 : * The given clause has a mergejoinable operator and is not an outer-join
142 : * qualification, so its two sides can be considered equal
143 : * anywhere they are both computable; moreover that equality can be
144 : * extended transitively. Record this knowledge in the EquivalenceClass
145 : * data structure, if applicable. Returns true if successful, false if not
146 : * (in which case caller should treat the clause as ordinary, not an
147 : * equivalence).
148 : *
149 : * In some cases, although we cannot convert a clause into EquivalenceClass
150 : * knowledge, we can still modify it to a more useful form than the original.
151 : * Then, *p_restrictinfo will be replaced by a new RestrictInfo, which is what
152 : * the caller should use for further processing.
153 : *
154 : * jdomain is the join domain within which the given clause was found.
155 : * This limits the applicability of deductions from the EquivalenceClass,
156 : * as described in optimizer/README.
157 : *
158 : * We reject proposed equivalence clauses if they contain leaky functions
159 : * and have security_level above zero. The EC evaluation rules require us to
160 : * apply certain tests at certain joining levels, and we can't tolerate
161 : * delaying any test on security_level grounds. By rejecting candidate clauses
162 : * that might require security delays, we ensure it's safe to apply an EC
163 : * clause as soon as it's supposed to be applied.
164 : *
165 : * On success return, we have also initialized the clause's left_ec/right_ec
166 : * fields to point to the EquivalenceClass representing it. This saves lookup
167 : * effort later.
168 : *
169 : * Note: constructing merged EquivalenceClasses is a standard UNION-FIND
170 : * problem, for which there exist better data structures than simple lists.
171 : * If this code ever proves to be a bottleneck then it could be sped up ---
172 : * but for now, simple is beautiful.
173 : *
174 : * Note: this is only called during planner startup, not during GEQO
175 : * exploration, so we need not worry about whether we're in the right
176 : * memory context.
177 : */
178 : bool
179 313090 : process_equivalence(PlannerInfo *root,
180 : RestrictInfo **p_restrictinfo,
181 : JoinDomain *jdomain)
182 : {
183 313090 : RestrictInfo *restrictinfo = *p_restrictinfo;
184 313090 : Expr *clause = restrictinfo->clause;
185 : Oid opno,
186 : collation,
187 : item1_type,
188 : item2_type;
189 : Expr *item1;
190 : Expr *item2;
191 : Relids item1_relids,
192 : item2_relids;
193 : List *opfamilies;
194 : EquivalenceClass *ec1,
195 : *ec2;
196 : EquivalenceMember *em1,
197 : *em2;
198 : ListCell *lc1;
199 : int ec2_idx;
200 :
201 : /* Should not already be marked as having generated an eclass */
202 : Assert(restrictinfo->left_ec == NULL);
203 : Assert(restrictinfo->right_ec == NULL);
204 :
205 : /* Reject if it is potentially postponable by security considerations */
206 313090 : if (restrictinfo->security_level > 0 && !restrictinfo->leakproof)
207 208 : return false;
208 :
209 : /* Extract info from given clause */
210 : Assert(is_opclause(clause));
211 312882 : opno = ((OpExpr *) clause)->opno;
212 312882 : collation = ((OpExpr *) clause)->inputcollid;
213 312882 : item1 = (Expr *) get_leftop(clause);
214 312882 : item2 = (Expr *) get_rightop(clause);
215 312882 : item1_relids = restrictinfo->left_relids;
216 312882 : item2_relids = restrictinfo->right_relids;
217 :
218 : /*
219 : * Ensure both input expressions expose the desired collation (their types
220 : * should be OK already); see comments for canonicalize_ec_expression.
221 : */
222 312882 : item1 = canonicalize_ec_expression(item1,
223 : exprType((Node *) item1),
224 : collation);
225 312882 : item2 = canonicalize_ec_expression(item2,
226 : exprType((Node *) item2),
227 : collation);
228 :
229 : /*
230 : * Clauses of the form X=X cannot be translated into EquivalenceClasses.
231 : * We'd either end up with a single-entry EC, losing the knowledge that
232 : * the clause was present at all, or else make an EC with duplicate
233 : * entries, causing other issues.
234 : */
235 312882 : if (equal(item1, item2))
236 : {
237 : /*
238 : * If the operator is strict, then the clause can be treated as just
239 : * "X IS NOT NULL". (Since we know we are considering a top-level
240 : * qual, we can ignore the difference between FALSE and NULL results.)
241 : * It's worth making the conversion because we'll typically get a much
242 : * better selectivity estimate than we would for X=X.
243 : *
244 : * If the operator is not strict, we can't be sure what it will do
245 : * with NULLs, so don't attempt to optimize it.
246 : */
247 54 : set_opfuncid((OpExpr *) clause);
248 54 : if (func_strict(((OpExpr *) clause)->opfuncid))
249 : {
250 54 : NullTest *ntest = makeNode(NullTest);
251 :
252 54 : ntest->arg = item1;
253 54 : ntest->nulltesttype = IS_NOT_NULL;
254 54 : ntest->argisrow = false; /* correct even if composite arg */
255 54 : ntest->location = -1;
256 :
257 54 : *p_restrictinfo =
258 54 : make_restrictinfo(root,
259 : (Expr *) ntest,
260 54 : restrictinfo->is_pushed_down,
261 54 : restrictinfo->has_clone,
262 54 : restrictinfo->is_clone,
263 54 : restrictinfo->pseudoconstant,
264 : restrictinfo->security_level,
265 : NULL,
266 : restrictinfo->incompatible_relids,
267 : restrictinfo->outer_relids);
268 : }
269 54 : return false;
270 : }
271 :
272 : /*
273 : * We use the declared input types of the operator, not exprType() of the
274 : * inputs, as the nominal datatypes for opfamily lookup. This presumes
275 : * that btree operators are always registered with amoplefttype and
276 : * amoprighttype equal to their declared input types. We will need this
277 : * info anyway to build EquivalenceMember nodes, and by extracting it now
278 : * we can use type comparisons to short-circuit some equal() tests.
279 : */
280 312828 : op_input_types(opno, &item1_type, &item2_type);
281 :
282 312828 : opfamilies = restrictinfo->mergeopfamilies;
283 :
284 : /*
285 : * Sweep through the existing EquivalenceClasses looking for matches to
286 : * item1 and item2. These are the possible outcomes:
287 : *
288 : * 1. We find both in the same EC. The equivalence is already known, so
289 : * there's nothing to do.
290 : *
291 : * 2. We find both in different ECs. Merge the two ECs together.
292 : *
293 : * 3. We find just one. Add the other to its EC.
294 : *
295 : * 4. We find neither. Make a new, two-entry EC.
296 : *
297 : * Note: since all ECs are built through this process or the similar
298 : * search in get_eclass_for_sort_expr(), it's impossible that we'd match
299 : * an item in more than one existing nonvolatile EC. So it's okay to stop
300 : * at the first match.
301 : */
302 312828 : ec1 = ec2 = NULL;
303 312828 : em1 = em2 = NULL;
304 312828 : ec2_idx = -1;
305 520376 : foreach(lc1, root->eq_classes)
306 : {
307 207602 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
308 : ListCell *lc2;
309 :
310 : /* Never match to a volatile EC */
311 207602 : if (cur_ec->ec_has_volatile)
312 0 : continue;
313 :
314 : /*
315 : * The collation has to match; check this first since it's cheaper
316 : * than the opfamily comparison.
317 : */
318 207602 : if (collation != cur_ec->ec_collation)
319 13354 : continue;
320 :
321 : /*
322 : * A "match" requires matching sets of btree opfamilies. Use of
323 : * equal() for this test has implications discussed in the comments
324 : * for get_mergejoin_opfamilies().
325 : */
326 194248 : if (!equal(opfamilies, cur_ec->ec_opfamilies))
327 51162 : continue;
328 :
329 : /* We don't expect any children yet */
330 : Assert(cur_ec->ec_childmembers == NULL);
331 :
332 431622 : foreach(lc2, cur_ec->ec_members)
333 : {
334 288590 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
335 :
336 : /* Child members should not exist in ec_members */
337 : Assert(!cur_em->em_is_child);
338 :
339 : /*
340 : * Match constants only within the same JoinDomain (see
341 : * optimizer/README).
342 : */
343 288590 : if (cur_em->em_is_const && cur_em->em_jdomain != jdomain)
344 4630 : continue;
345 :
346 283960 : if (!ec1 &&
347 542016 : item1_type == cur_em->em_datatype &&
348 270876 : equal(item1, cur_em->em_expr))
349 : {
350 18862 : ec1 = cur_ec;
351 18862 : em1 = cur_em;
352 18862 : if (ec2)
353 30 : break;
354 : }
355 :
356 283930 : if (!ec2 &&
357 561836 : item2_type == cur_em->em_datatype &&
358 280726 : equal(item2, cur_em->em_expr))
359 : {
360 4764 : ec2 = cur_ec;
361 4764 : ec2_idx = foreach_current_index(lc1);
362 4764 : em2 = cur_em;
363 4764 : if (ec1)
364 24 : break;
365 : }
366 : }
367 :
368 143086 : if (ec1 && ec2)
369 54 : break;
370 : }
371 :
372 : /* Sweep finished, what did we find? */
373 :
374 312828 : if (ec1 && ec2)
375 : {
376 : /* If case 1, nothing to do, except add to sources */
377 54 : if (ec1 == ec2)
378 : {
379 12 : ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
380 12 : ec1->ec_min_security = Min(ec1->ec_min_security,
381 : restrictinfo->security_level);
382 12 : ec1->ec_max_security = Max(ec1->ec_max_security,
383 : restrictinfo->security_level);
384 : /* mark the RI as associated with this eclass */
385 12 : restrictinfo->left_ec = ec1;
386 12 : restrictinfo->right_ec = ec1;
387 : /* mark the RI as usable with this pair of EMs */
388 12 : restrictinfo->left_em = em1;
389 12 : restrictinfo->right_em = em2;
390 12 : return true;
391 : }
392 :
393 : /*
394 : * Case 2: need to merge ec1 and ec2. This should never happen after
395 : * the ECs have reached canonical state; otherwise, pathkeys could be
396 : * rendered non-canonical by the merge, and relation eclass indexes
397 : * would get broken by removal of an eq_classes list entry.
398 : */
399 42 : if (root->ec_merging_done)
400 0 : elog(ERROR, "too late to merge equivalence classes");
401 :
402 : /*
403 : * We add ec2's items to ec1, then set ec2's ec_merged link to point
404 : * to ec1 and remove ec2 from the eq_classes list. We cannot simply
405 : * delete ec2 because that could leave dangling pointers in existing
406 : * PathKeys. We leave it behind with a link so that the merged EC can
407 : * be found.
408 : */
409 42 : ec1->ec_members = list_concat(ec1->ec_members, ec2->ec_members);
410 42 : ec1->ec_sources = list_concat(ec1->ec_sources, ec2->ec_sources);
411 :
412 : /*
413 : * Appends ec2's derived clauses to ec1->ec_derives_list and adds them
414 : * to ec1->ec_derives_hash if present.
415 : */
416 42 : ec_add_derived_clauses(ec1, ec2->ec_derives_list);
417 42 : ec1->ec_relids = bms_join(ec1->ec_relids, ec2->ec_relids);
418 42 : ec1->ec_has_const |= ec2->ec_has_const;
419 : /* can't need to set has_volatile */
420 42 : ec1->ec_min_security = Min(ec1->ec_min_security,
421 : ec2->ec_min_security);
422 42 : ec1->ec_max_security = Max(ec1->ec_max_security,
423 : ec2->ec_max_security);
424 42 : ec2->ec_merged = ec1;
425 42 : root->eq_classes = list_delete_nth_cell(root->eq_classes, ec2_idx);
426 : /* just to avoid debugging confusion w/ dangling pointers: */
427 42 : ec2->ec_members = NIL;
428 42 : ec2->ec_sources = NIL;
429 42 : ec_clear_derived_clauses(ec2);
430 42 : ec2->ec_relids = NULL;
431 42 : ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
432 42 : ec1->ec_min_security = Min(ec1->ec_min_security,
433 : restrictinfo->security_level);
434 42 : ec1->ec_max_security = Max(ec1->ec_max_security,
435 : restrictinfo->security_level);
436 : /* mark the RI as associated with this eclass */
437 42 : restrictinfo->left_ec = ec1;
438 42 : restrictinfo->right_ec = ec1;
439 : /* mark the RI as usable with this pair of EMs */
440 42 : restrictinfo->left_em = em1;
441 42 : restrictinfo->right_em = em2;
442 : }
443 312774 : else if (ec1)
444 : {
445 : /* Case 3: add item2 to ec1 */
446 18808 : em2 = add_eq_member(ec1, item2, item2_relids,
447 : jdomain, item2_type);
448 18808 : ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
449 18808 : ec1->ec_min_security = Min(ec1->ec_min_security,
450 : restrictinfo->security_level);
451 18808 : ec1->ec_max_security = Max(ec1->ec_max_security,
452 : restrictinfo->security_level);
453 : /* mark the RI as associated with this eclass */
454 18808 : restrictinfo->left_ec = ec1;
455 18808 : restrictinfo->right_ec = ec1;
456 : /* mark the RI as usable with this pair of EMs */
457 18808 : restrictinfo->left_em = em1;
458 18808 : restrictinfo->right_em = em2;
459 : }
460 293966 : else if (ec2)
461 : {
462 : /* Case 3: add item1 to ec2 */
463 4710 : em1 = add_eq_member(ec2, item1, item1_relids,
464 : jdomain, item1_type);
465 4710 : ec2->ec_sources = lappend(ec2->ec_sources, restrictinfo);
466 4710 : ec2->ec_min_security = Min(ec2->ec_min_security,
467 : restrictinfo->security_level);
468 4710 : ec2->ec_max_security = Max(ec2->ec_max_security,
469 : restrictinfo->security_level);
470 : /* mark the RI as associated with this eclass */
471 4710 : restrictinfo->left_ec = ec2;
472 4710 : restrictinfo->right_ec = ec2;
473 : /* mark the RI as usable with this pair of EMs */
474 4710 : restrictinfo->left_em = em1;
475 4710 : restrictinfo->right_em = em2;
476 : }
477 : else
478 : {
479 : /* Case 4: make a new, two-entry EC */
480 289256 : EquivalenceClass *ec = makeNode(EquivalenceClass);
481 :
482 289256 : ec->ec_opfamilies = opfamilies;
483 289256 : ec->ec_collation = collation;
484 289256 : ec->ec_childmembers_size = 0;
485 289256 : ec->ec_members = NIL;
486 289256 : ec->ec_childmembers = NULL;
487 289256 : ec->ec_sources = list_make1(restrictinfo);
488 289256 : ec->ec_derives_list = NIL;
489 289256 : ec->ec_derives_hash = NULL;
490 289256 : ec->ec_relids = NULL;
491 289256 : ec->ec_has_const = false;
492 289256 : ec->ec_has_volatile = false;
493 289256 : ec->ec_broken = false;
494 289256 : ec->ec_sortref = 0;
495 289256 : ec->ec_min_security = restrictinfo->security_level;
496 289256 : ec->ec_max_security = restrictinfo->security_level;
497 289256 : ec->ec_merged = NULL;
498 289256 : em1 = add_eq_member(ec, item1, item1_relids,
499 : jdomain, item1_type);
500 289256 : em2 = add_eq_member(ec, item2, item2_relids,
501 : jdomain, item2_type);
502 :
503 289256 : root->eq_classes = lappend(root->eq_classes, ec);
504 :
505 : /* mark the RI as associated with this eclass */
506 289256 : restrictinfo->left_ec = ec;
507 289256 : restrictinfo->right_ec = ec;
508 : /* mark the RI as usable with this pair of EMs */
509 289256 : restrictinfo->left_em = em1;
510 289256 : restrictinfo->right_em = em2;
511 : }
512 :
513 312816 : return true;
514 : }
515 :
516 : /*
517 : * canonicalize_ec_expression
518 : *
519 : * This function ensures that the expression exposes the expected type and
520 : * collation, so that it will be equal() to other equivalence-class expressions
521 : * that it ought to be equal() to.
522 : *
523 : * The rule for datatypes is that the exposed type should match what it would
524 : * be for an input to an operator of the EC's opfamilies; which is usually
525 : * the declared input type of the operator, but in the case of polymorphic
526 : * operators no relabeling is wanted (compare the behavior of parse_coerce.c).
527 : * Expressions coming in from quals will generally have the right type
528 : * already, but expressions coming from indexkeys may not (because they are
529 : * represented without any explicit relabel in pg_index), and the same problem
530 : * occurs for sort expressions (because the parser is likewise cavalier about
531 : * putting relabels on them). Such cases will be binary-compatible with the
532 : * real operators, so adding a RelabelType is sufficient.
533 : *
534 : * Also, the expression's exposed collation must match the EC's collation.
535 : * This is important because in comparisons like "foo < bar COLLATE baz",
536 : * only one of the expressions has the correct exposed collation as we receive
537 : * it from the parser. Forcing both of them to have it ensures that all
538 : * variant spellings of such a construct behave the same. Again, we can
539 : * stick on a RelabelType to force the right exposed collation. (It might
540 : * work to not label the collation at all in EC members, but this is risky
541 : * since some parts of the system expect exprCollation() to deliver the
542 : * right answer for a sort key.)
543 : */
544 : Expr *
545 2846346 : canonicalize_ec_expression(Expr *expr, Oid req_type, Oid req_collation)
546 : {
547 2846346 : Oid expr_type = exprType((Node *) expr);
548 :
549 : /*
550 : * For a polymorphic-input-type opclass, just keep the same exposed type.
551 : * RECORD opclasses work like polymorphic-type ones for this purpose.
552 : */
553 2846346 : if (IsPolymorphicType(req_type) || req_type == RECORDOID)
554 7866 : req_type = expr_type;
555 :
556 : /*
557 : * No work if the expression exposes the right type/collation already.
558 : */
559 5621850 : if (expr_type != req_type ||
560 2775504 : exprCollation((Node *) expr) != req_collation)
561 : {
562 : /*
563 : * If we have to change the type of the expression, set typmod to -1,
564 : * since the new type may not have the same typmod interpretation.
565 : * When we only have to change collation, preserve the exposed typmod.
566 : */
567 : int32 req_typmod;
568 :
569 72450 : if (expr_type != req_type)
570 70842 : req_typmod = -1;
571 : else
572 1608 : req_typmod = exprTypmod((Node *) expr);
573 :
574 : /*
575 : * Use applyRelabelType so that we preserve const-flatness. This is
576 : * important since eval_const_expressions has already been applied.
577 : */
578 72450 : expr = (Expr *) applyRelabelType((Node *) expr,
579 : req_type, req_typmod, req_collation,
580 : COERCE_IMPLICIT_CAST, -1, false);
581 : }
582 :
583 2846346 : return expr;
584 : }
585 :
586 : /*
587 : * make_eq_member
588 : * Build a new EquivalenceMember without adding it to an EC. If 'parent'
589 : * is NULL, the result will be a parent member, otherwise a child member.
590 : */
591 : static EquivalenceMember *
592 945732 : make_eq_member(EquivalenceClass *ec, Expr *expr, Relids relids,
593 : JoinDomain *jdomain, EquivalenceMember *parent, Oid datatype)
594 : {
595 945732 : EquivalenceMember *em = makeNode(EquivalenceMember);
596 :
597 945732 : em->em_expr = expr;
598 945732 : em->em_relids = relids;
599 945732 : em->em_is_const = false;
600 945732 : em->em_is_child = (parent != NULL);
601 945732 : em->em_datatype = datatype;
602 945732 : em->em_jdomain = jdomain;
603 945732 : em->em_parent = parent;
604 :
605 945732 : if (bms_is_empty(relids))
606 : {
607 : /*
608 : * No Vars, assume it's a pseudoconstant. This is correct for entries
609 : * generated from process_equivalence(), because a WHERE clause can't
610 : * contain aggregates or SRFs, and non-volatility was checked before
611 : * process_equivalence() ever got called. But
612 : * get_eclass_for_sort_expr() has to work harder. We put the tests
613 : * there not here to save cycles in the equivalence case.
614 : */
615 : Assert(!parent);
616 233752 : em->em_is_const = true;
617 233752 : ec->ec_has_const = true;
618 : /* it can't affect ec_relids */
619 : }
620 :
621 945732 : return em;
622 : }
623 :
624 : /*
625 : * add_eq_member - build a new non-child EquivalenceMember and add it to 'ec'.
626 : */
627 : static EquivalenceMember *
628 858670 : add_eq_member(EquivalenceClass *ec, Expr *expr, Relids relids,
629 : JoinDomain *jdomain, Oid datatype)
630 : {
631 858670 : EquivalenceMember *em = make_eq_member(ec, expr, relids, jdomain,
632 : NULL, datatype);
633 :
634 : /* add to the members list */
635 858670 : ec->ec_members = lappend(ec->ec_members, em);
636 :
637 : /* record the relids for parent members */
638 858670 : ec->ec_relids = bms_add_members(ec->ec_relids, relids);
639 :
640 858670 : return em;
641 : }
642 :
643 : /*
644 : * add_child_eq_member
645 : * Create an em_is_child=true EquivalenceMember and add it to 'ec'.
646 : *
647 : * 'root' is the PlannerInfo that 'ec' belongs to.
648 : * 'ec' is the EquivalenceClass to add the child member to.
649 : * 'ec_index' the index of 'ec' within root->eq_classes, or -1 if maintaining
650 : * the RelOptInfo.eclass_indexes isn't needed.
651 : * 'expr' is the em_expr for the new member.
652 : * 'relids' is the 'em_relids' for the new member.
653 : * 'jdomain' is the 'em_jdomain' for the new member.
654 : * 'parent_em' is the parent member of the child to create.
655 : * 'datatype' is the em_datatype of the new member.
656 : * 'child_relid' defines which element of ec_childmembers to add this member
657 : * to. This is generally a RELOPT_OTHER_MEMBER_REL, but for set operations
658 : * can be a RELOPT_BASEREL representing the set-op children.
659 : */
660 : static EquivalenceMember *
661 87062 : add_child_eq_member(PlannerInfo *root, EquivalenceClass *ec, int ec_index,
662 : Expr *expr, Relids relids, JoinDomain *jdomain,
663 : EquivalenceMember *parent_em, Oid datatype,
664 : Index child_relid)
665 : {
666 : EquivalenceMember *em;
667 :
668 : Assert(parent_em != NULL);
669 :
670 : /*
671 : * Allocate the array to store child members; an array of Lists indexed by
672 : * relid, or expand the existing one, if necessary.
673 : */
674 87062 : if (unlikely(ec->ec_childmembers_size < root->simple_rel_array_size))
675 : {
676 26228 : if (ec->ec_childmembers == NULL)
677 26228 : ec->ec_childmembers = palloc0_array(List *, root->simple_rel_array_size);
678 : else
679 0 : ec->ec_childmembers = repalloc0_array(ec->ec_childmembers, List *,
680 : ec->ec_childmembers_size,
681 : root->simple_rel_array_size);
682 :
683 26228 : ec->ec_childmembers_size = root->simple_rel_array_size;
684 : }
685 :
686 87062 : em = make_eq_member(ec, expr, relids, jdomain, parent_em, datatype);
687 :
688 : /* add member to the ec_childmembers List for the given child_relid */
689 87062 : ec->ec_childmembers[child_relid] = lappend(ec->ec_childmembers[child_relid], em);
690 :
691 : /* Record this EC index for the child rel */
692 87062 : if (ec_index >= 0)
693 : {
694 48030 : RelOptInfo *child_rel = root->simple_rel_array[child_relid];
695 :
696 48030 : child_rel->eclass_indexes =
697 48030 : bms_add_member(child_rel->eclass_indexes, ec_index);
698 : }
699 :
700 87062 : return em;
701 : }
702 :
703 :
704 : /*
705 : * get_eclass_for_sort_expr
706 : * Given an expression and opfamily/collation info, find an existing
707 : * equivalence class it is a member of; if none, optionally build a new
708 : * single-member EquivalenceClass for it.
709 : *
710 : * sortref is the SortGroupRef of the originating SortGroupClause, if any,
711 : * or zero if not. (It should never be zero if the expression is volatile!)
712 : *
713 : * If rel is not NULL, it identifies a specific relation we're considering
714 : * a path for, and indicates that child EC members for that relation can be
715 : * considered. Otherwise child members are ignored. (Note: since child EC
716 : * members aren't guaranteed unique, a non-NULL value means that there could
717 : * be more than one EC that matches the expression; if so it's order-dependent
718 : * which one you get. This is annoying but it only happens in corner cases,
719 : * so for now we live with just reporting the first match. See also
720 : * generate_implied_equalities_for_column and match_pathkeys_to_index.)
721 : *
722 : * If create_it is true, we'll build a new EquivalenceClass when there is no
723 : * match. If create_it is false, we just return NULL when no match.
724 : *
725 : * This can be used safely both before and after EquivalenceClass merging;
726 : * since it never causes merging it does not invalidate any existing ECs
727 : * or PathKeys. However, ECs added after path generation has begun are
728 : * of limited usefulness, so usually it's best to create them beforehand.
729 : *
730 : * Note: opfamilies must be chosen consistently with the way
731 : * process_equivalence() would do; that is, generated from a mergejoinable
732 : * equality operator. Else we might fail to detect valid equivalences,
733 : * generating poor (but not incorrect) plans.
734 : */
735 : EquivalenceClass *
736 2074712 : get_eclass_for_sort_expr(PlannerInfo *root,
737 : Expr *expr,
738 : List *opfamilies,
739 : Oid opcintype,
740 : Oid collation,
741 : Index sortref,
742 : Relids rel,
743 : bool create_it)
744 : {
745 : JoinDomain *jdomain;
746 : Relids expr_relids;
747 : EquivalenceClass *newec;
748 : EquivalenceMember *newem;
749 : ListCell *lc1;
750 : MemoryContext oldcontext;
751 :
752 : /*
753 : * Ensure the expression exposes the correct type and collation.
754 : */
755 2074712 : expr = canonicalize_ec_expression(expr, opcintype, collation);
756 :
757 : /*
758 : * Since SortGroupClause nodes are top-level expressions (GROUP BY, ORDER
759 : * BY, etc), they can be presumed to belong to the top JoinDomain.
760 : */
761 2074712 : jdomain = linitial_node(JoinDomain, root->join_domains);
762 :
763 : /*
764 : * Scan through the existing EquivalenceClasses for a match
765 : */
766 7010478 : foreach(lc1, root->eq_classes)
767 : {
768 6090362 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
769 : EquivalenceMemberIterator it;
770 : EquivalenceMember *cur_em;
771 :
772 : /*
773 : * Never match to a volatile EC, except when we are looking at another
774 : * reference to the same volatile SortGroupClause.
775 : */
776 6090362 : if (cur_ec->ec_has_volatile &&
777 36 : (sortref == 0 || sortref != cur_ec->ec_sortref))
778 2631654 : continue;
779 :
780 6089838 : if (collation != cur_ec->ec_collation)
781 1590626 : continue;
782 4499212 : if (!equal(opfamilies, cur_ec->ec_opfamilies))
783 1040504 : continue;
784 :
785 3458708 : setup_eclass_member_iterator(&it, cur_ec, rel);
786 7724762 : while ((cur_em = eclass_member_iterator_next(&it)) != NULL)
787 : {
788 : /*
789 : * Ignore child members unless they match the request.
790 : */
791 5420650 : if (cur_em->em_is_child &&
792 98364 : !bms_equal(cur_em->em_relids, rel))
793 0 : continue;
794 :
795 : /*
796 : * Match constants only within the same JoinDomain (see
797 : * optimizer/README).
798 : */
799 5420650 : if (cur_em->em_is_const && cur_em->em_jdomain != jdomain)
800 97162 : continue;
801 :
802 10595770 : if (opcintype == cur_em->em_datatype &&
803 5272282 : equal(expr, cur_em->em_expr))
804 1154596 : return cur_ec; /* Match! */
805 : }
806 : }
807 :
808 : /* No match; does caller want a NULL result? */
809 920116 : if (!create_it)
810 663476 : return NULL;
811 :
812 : /*
813 : * OK, build a new single-member EC
814 : *
815 : * Here, we must be sure that we construct the EC in the right context.
816 : */
817 256640 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
818 :
819 256640 : newec = makeNode(EquivalenceClass);
820 256640 : newec->ec_opfamilies = list_copy(opfamilies);
821 256640 : newec->ec_collation = collation;
822 256640 : newec->ec_childmembers_size = 0;
823 256640 : newec->ec_members = NIL;
824 256640 : newec->ec_childmembers = NULL;
825 256640 : newec->ec_sources = NIL;
826 256640 : newec->ec_derives_list = NIL;
827 256640 : newec->ec_derives_hash = NULL;
828 256640 : newec->ec_relids = NULL;
829 256640 : newec->ec_has_const = false;
830 256640 : newec->ec_has_volatile = contain_volatile_functions((Node *) expr);
831 256640 : newec->ec_broken = false;
832 256640 : newec->ec_sortref = sortref;
833 256640 : newec->ec_min_security = UINT_MAX;
834 256640 : newec->ec_max_security = 0;
835 256640 : newec->ec_merged = NULL;
836 :
837 256640 : if (newec->ec_has_volatile && sortref == 0) /* should not happen */
838 0 : elog(ERROR, "volatile EquivalenceClass has no sortref");
839 :
840 : /*
841 : * Get the precise set of relids appearing in the expression.
842 : */
843 256640 : expr_relids = pull_varnos(root, (Node *) expr);
844 :
845 256640 : newem = add_eq_member(newec, copyObject(expr), expr_relids,
846 : jdomain, opcintype);
847 :
848 : /*
849 : * add_eq_member doesn't check for volatile functions, set-returning
850 : * functions, aggregates, or window functions, but such could appear in
851 : * sort expressions; so we have to check whether its const-marking was
852 : * correct.
853 : */
854 256640 : if (newec->ec_has_const)
855 : {
856 18152 : if (newec->ec_has_volatile ||
857 17870 : expression_returns_set((Node *) expr) ||
858 17544 : contain_agg_clause((Node *) expr) ||
859 8684 : contain_window_function((Node *) expr))
860 : {
861 464 : newec->ec_has_const = false;
862 464 : newem->em_is_const = false;
863 : }
864 : }
865 :
866 256640 : root->eq_classes = lappend(root->eq_classes, newec);
867 :
868 : /*
869 : * If EC merging is already complete, we have to mop up by adding the new
870 : * EC to the eclass_indexes of the relation(s) mentioned in it.
871 : */
872 256640 : if (root->ec_merging_done)
873 : {
874 152142 : int ec_index = list_length(root->eq_classes) - 1;
875 152142 : int i = -1;
876 :
877 293414 : while ((i = bms_next_member(newec->ec_relids, i)) > 0)
878 : {
879 141272 : RelOptInfo *rel = root->simple_rel_array[i];
880 :
881 : /* ignore the RTE_GROUP RTE */
882 141272 : if (i == root->group_rtindex)
883 604 : continue;
884 :
885 140668 : if (rel == NULL) /* must be an outer join */
886 : {
887 : Assert(bms_is_member(i, root->outer_join_rels));
888 6486 : continue;
889 : }
890 :
891 : Assert(rel->reloptkind == RELOPT_BASEREL);
892 :
893 134182 : rel->eclass_indexes = bms_add_member(rel->eclass_indexes,
894 : ec_index);
895 : }
896 : }
897 :
898 256640 : MemoryContextSwitchTo(oldcontext);
899 :
900 256640 : return newec;
901 : }
902 :
903 : /*
904 : * find_ec_member_matching_expr
905 : * Locate an EquivalenceClass member matching the given expr, if any;
906 : * return NULL if no match.
907 : *
908 : * "Matching" is defined as "equal after stripping RelabelTypes".
909 : * This is used for identifying sort expressions, and we need to allow
910 : * binary-compatible relabeling for some cases involving binary-compatible
911 : * sort operators.
912 : *
913 : * Child EC members are ignored unless they belong to given 'relids'.
914 : */
915 : EquivalenceMember *
916 363520 : find_ec_member_matching_expr(EquivalenceClass *ec,
917 : Expr *expr,
918 : Relids relids)
919 : {
920 : EquivalenceMemberIterator it;
921 : EquivalenceMember *em;
922 :
923 : /* We ignore binary-compatible relabeling on both ends */
924 386092 : while (expr && IsA(expr, RelabelType))
925 22572 : expr = ((RelabelType *) expr)->arg;
926 :
927 363520 : setup_eclass_member_iterator(&it, ec, relids);
928 612620 : while ((em = eclass_member_iterator_next(&it)) != NULL)
929 : {
930 : Expr *emexpr;
931 :
932 : /*
933 : * We shouldn't be trying to sort by an equivalence class that
934 : * contains a constant, so no need to consider such cases any further.
935 : */
936 396090 : if (em->em_is_const)
937 0 : continue;
938 :
939 : /*
940 : * Ignore child members unless they belong to the requested rel.
941 : */
942 396090 : if (em->em_is_child &&
943 10126 : !bms_is_subset(em->em_relids, relids))
944 4080 : continue;
945 :
946 : /*
947 : * Match if same expression (after stripping relabel).
948 : */
949 392010 : emexpr = em->em_expr;
950 397784 : while (emexpr && IsA(emexpr, RelabelType))
951 5774 : emexpr = ((RelabelType *) emexpr)->arg;
952 :
953 392010 : if (equal(emexpr, expr))
954 146990 : return em;
955 : }
956 :
957 216530 : return NULL;
958 : }
959 :
960 : /*
961 : * find_computable_ec_member
962 : * Locate an EquivalenceClass member that can be computed from the
963 : * expressions appearing in "exprs"; return NULL if no match.
964 : *
965 : * "exprs" can be either a list of bare expression trees, or a list of
966 : * TargetEntry nodes. Typically it will contain Vars and possibly Aggrefs
967 : * and WindowFuncs; however, when considering an appendrel member the list
968 : * could contain arbitrary expressions. We consider an EC member to be
969 : * computable if all the Vars, PlaceHolderVars, Aggrefs, and WindowFuncs
970 : * it needs are present in "exprs".
971 : *
972 : * There is some subtlety in that definition: for example, if an EC member is
973 : * Var_A + 1 while what is in "exprs" is Var_A + 2, it's still computable.
974 : * This works because in the final plan tree, the EC member's expression will
975 : * be computed as part of the same plan node targetlist that is currently
976 : * represented by "exprs". So if we have Var_A available for the existing
977 : * tlist member, it must be OK to use it in the EC expression too.
978 : *
979 : * Unlike find_ec_member_matching_expr, there's no special provision here
980 : * for binary-compatible relabeling. This is intentional: if we have to
981 : * compute an expression in this way, setrefs.c is going to insist on exact
982 : * matches of Vars to the source tlist.
983 : *
984 : * Child EC members are ignored unless they belong to given 'relids'.
985 : * Also, non-parallel-safe expressions are ignored if 'require_parallel_safe'.
986 : *
987 : * Note: some callers pass root == NULL for notational reasons. This is OK
988 : * when require_parallel_safe is false.
989 : */
990 : EquivalenceMember *
991 4992 : find_computable_ec_member(PlannerInfo *root,
992 : EquivalenceClass *ec,
993 : List *exprs,
994 : Relids relids,
995 : bool require_parallel_safe)
996 : {
997 : List *exprvars;
998 : EquivalenceMemberIterator it;
999 : EquivalenceMember *em;
1000 :
1001 : /*
1002 : * Pull out the Vars and quasi-Vars present in "exprs". In the typical
1003 : * non-appendrel case, this is just another representation of the same
1004 : * list. However, it does remove the distinction between the case of a
1005 : * list of plain expressions and a list of TargetEntrys.
1006 : */
1007 4992 : exprvars = pull_var_clause((Node *) exprs,
1008 : PVC_INCLUDE_AGGREGATES |
1009 : PVC_INCLUDE_WINDOWFUNCS |
1010 : PVC_INCLUDE_PLACEHOLDERS |
1011 : PVC_INCLUDE_CONVERTROWTYPES);
1012 :
1013 4992 : setup_eclass_member_iterator(&it, ec, relids);
1014 10060 : while ((em = eclass_member_iterator_next(&it)) != NULL)
1015 : {
1016 : List *emvars;
1017 : ListCell *lc2;
1018 :
1019 : /*
1020 : * We shouldn't be trying to sort by an equivalence class that
1021 : * contains a constant, so no need to consider such cases any further.
1022 : */
1023 5538 : if (em->em_is_const)
1024 0 : continue;
1025 :
1026 : /*
1027 : * Ignore child members unless they belong to the requested rel.
1028 : */
1029 5538 : if (em->em_is_child &&
1030 324 : !bms_is_subset(em->em_relids, relids))
1031 132 : continue;
1032 :
1033 : /*
1034 : * Match if all Vars and quasi-Vars are present in "exprs".
1035 : */
1036 5406 : emvars = pull_var_clause((Node *) em->em_expr,
1037 : PVC_INCLUDE_AGGREGATES |
1038 : PVC_INCLUDE_WINDOWFUNCS |
1039 : PVC_INCLUDE_PLACEHOLDERS);
1040 6106 : foreach(lc2, emvars)
1041 : {
1042 5606 : if (!list_member(exprvars, lfirst(lc2)))
1043 4906 : break;
1044 : }
1045 5406 : list_free(emvars);
1046 5406 : if (lc2)
1047 4906 : continue; /* we hit a non-available Var */
1048 :
1049 : /*
1050 : * If requested, reject expressions that are not parallel-safe. We
1051 : * check this last because it's a rather expensive test.
1052 : */
1053 500 : if (require_parallel_safe &&
1054 128 : !is_parallel_safe(root, (Node *) em->em_expr))
1055 30 : continue;
1056 :
1057 470 : return em; /* found usable expression */
1058 : }
1059 :
1060 4522 : return NULL;
1061 : }
1062 :
1063 : /*
1064 : * relation_can_be_sorted_early
1065 : * Can this relation be sorted on this EC before the final output step?
1066 : *
1067 : * To succeed, we must find an EC member that prepare_sort_from_pathkeys knows
1068 : * how to sort on, given the rel's reltarget as input. There are also a few
1069 : * additional constraints based on the fact that the desired sort will be done
1070 : * "early", within the scan/join part of the plan. Also, non-parallel-safe
1071 : * expressions are ignored if 'require_parallel_safe'.
1072 : *
1073 : * At some point we might want to return the identified EquivalenceMember,
1074 : * but for now, callers only want to know if there is one.
1075 : */
1076 : bool
1077 12810 : relation_can_be_sorted_early(PlannerInfo *root, RelOptInfo *rel,
1078 : EquivalenceClass *ec, bool require_parallel_safe)
1079 : {
1080 12810 : PathTarget *target = rel->reltarget;
1081 : EquivalenceMember *em;
1082 : ListCell *lc;
1083 :
1084 : /*
1085 : * Reject volatile ECs immediately; such sorts must always be postponed.
1086 : */
1087 12810 : if (ec->ec_has_volatile)
1088 72 : return false;
1089 :
1090 : /*
1091 : * Try to find an EM directly matching some reltarget member.
1092 : */
1093 25936 : foreach(lc, target->exprs)
1094 : {
1095 21316 : Expr *targetexpr = (Expr *) lfirst(lc);
1096 :
1097 21316 : em = find_ec_member_matching_expr(ec, targetexpr, rel->relids);
1098 21316 : if (!em)
1099 13198 : continue;
1100 :
1101 : /*
1102 : * Reject expressions involving set-returning functions, as those
1103 : * can't be computed early either. (Note: this test and the following
1104 : * one are effectively checking properties of targetexpr, so there's
1105 : * no point in asking whether some other EC member would be better.)
1106 : */
1107 8118 : if (expression_returns_set((Node *) em->em_expr))
1108 0 : continue;
1109 :
1110 : /*
1111 : * If requested, reject expressions that are not parallel-safe. We
1112 : * check this last because it's a rather expensive test.
1113 : */
1114 8118 : if (require_parallel_safe &&
1115 8118 : !is_parallel_safe(root, (Node *) em->em_expr))
1116 0 : continue;
1117 :
1118 8118 : return true;
1119 : }
1120 :
1121 : /*
1122 : * Try to find an expression computable from the reltarget.
1123 : */
1124 4620 : em = find_computable_ec_member(root, ec, target->exprs, rel->relids,
1125 : require_parallel_safe);
1126 4620 : if (!em)
1127 4522 : return false;
1128 :
1129 : /*
1130 : * Reject expressions involving set-returning functions, as those can't be
1131 : * computed early either. (There's no point in looking for another EC
1132 : * member in this case; since SRFs can't appear in WHERE, they cannot
1133 : * belong to multi-member ECs.)
1134 : */
1135 98 : if (expression_returns_set((Node *) em->em_expr))
1136 12 : return false;
1137 :
1138 86 : return true;
1139 : }
1140 :
1141 : /*
1142 : * generate_base_implied_equalities
1143 : * Generate any restriction clauses that we can deduce from equivalence
1144 : * classes.
1145 : *
1146 : * When an EC contains pseudoconstants, our strategy is to generate
1147 : * "member = const1" clauses where const1 is the first constant member, for
1148 : * every other member (including other constants). If we are able to do this
1149 : * then we don't need any "var = var" comparisons because we've successfully
1150 : * constrained all the vars at their points of creation. If we fail to
1151 : * generate any of these clauses due to lack of cross-type operators, we fall
1152 : * back to the "ec_broken" strategy described below. (XXX if there are
1153 : * multiple constants of different types, it's possible that we might succeed
1154 : * in forming all the required clauses if we started from a different const
1155 : * member; but this seems a sufficiently hokey corner case to not be worth
1156 : * spending lots of cycles on.)
1157 : *
1158 : * For ECs that contain no pseudoconstants, we generate derived clauses
1159 : * "member1 = member2" for each pair of members belonging to the same base
1160 : * relation (actually, if there are more than two for the same base relation,
1161 : * we only need enough clauses to link each to each other). This provides
1162 : * the base case for the recursion: each row emitted by a base relation scan
1163 : * will constrain all computable members of the EC to be equal. As each
1164 : * join path is formed, we'll add additional derived clauses on-the-fly
1165 : * to maintain this invariant (see generate_join_implied_equalities).
1166 : *
1167 : * If the opfamilies used by the EC do not provide complete sets of cross-type
1168 : * equality operators, it is possible that we will fail to generate a clause
1169 : * that must be generated to maintain the invariant. (An example: given
1170 : * "WHERE a.x = b.y AND b.y = a.z", the scheme breaks down if we cannot
1171 : * generate "a.x = a.z" as a restriction clause for A.) In this case we mark
1172 : * the EC "ec_broken" and fall back to regurgitating its original source
1173 : * RestrictInfos at appropriate times. We do not try to retract any derived
1174 : * clauses already generated from the broken EC, so the resulting plan could
1175 : * be poor due to bad selectivity estimates caused by redundant clauses. But
1176 : * the correct solution to that is to fix the opfamilies ...
1177 : *
1178 : * Equality clauses derived by this function are passed off to
1179 : * process_implied_equality (in plan/initsplan.c) to be inserted into the
1180 : * restrictinfo datastructures. Note that this must be called after initial
1181 : * scanning of the quals and before Path construction begins.
1182 : *
1183 : * We make no attempt to avoid generating duplicate RestrictInfos here: we
1184 : * don't search existing source or derived clauses in the EC for matches. It
1185 : * doesn't really seem worth the trouble to do so.
1186 : */
1187 : void
1188 332410 : generate_base_implied_equalities(PlannerInfo *root)
1189 : {
1190 : int ec_index;
1191 : ListCell *lc;
1192 :
1193 : /*
1194 : * At this point, we're done absorbing knowledge of equivalences in the
1195 : * query, so no further EC merging should happen, and ECs remaining in the
1196 : * eq_classes list can be considered canonical. (But note that it's still
1197 : * possible for new single-member ECs to be added through
1198 : * get_eclass_for_sort_expr().)
1199 : */
1200 332410 : root->ec_merging_done = true;
1201 :
1202 332410 : ec_index = 0;
1203 726122 : foreach(lc, root->eq_classes)
1204 : {
1205 393712 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
1206 393712 : bool can_generate_joinclause = false;
1207 : int i;
1208 :
1209 : Assert(ec->ec_merged == NULL); /* else shouldn't be in list */
1210 : Assert(!ec->ec_broken); /* not yet anyway... */
1211 :
1212 : /*
1213 : * Generate implied equalities that are restriction clauses.
1214 : * Single-member ECs won't generate any deductions, either here or at
1215 : * the join level.
1216 : */
1217 393712 : if (list_length(ec->ec_members) > 1)
1218 : {
1219 291200 : if (ec->ec_has_const)
1220 224472 : generate_base_implied_equalities_const(root, ec);
1221 : else
1222 66728 : generate_base_implied_equalities_no_const(root, ec);
1223 :
1224 : /* Recover if we failed to generate required derived clauses */
1225 291200 : if (ec->ec_broken)
1226 30 : generate_base_implied_equalities_broken(root, ec);
1227 :
1228 : /* Detect whether this EC might generate join clauses */
1229 291200 : can_generate_joinclause =
1230 291200 : (bms_membership(ec->ec_relids) == BMS_MULTIPLE);
1231 : }
1232 :
1233 : /*
1234 : * Mark the base rels cited in each eclass (which should all exist by
1235 : * now) with the eq_classes indexes of all eclasses mentioning them.
1236 : * This will let us avoid searching in subsequent lookups. While
1237 : * we're at it, we can mark base rels that have pending eclass joins;
1238 : * this is a cheap version of has_relevant_eclass_joinclause().
1239 : */
1240 393712 : i = -1;
1241 877990 : while ((i = bms_next_member(ec->ec_relids, i)) > 0)
1242 : {
1243 484278 : RelOptInfo *rel = root->simple_rel_array[i];
1244 :
1245 : /* ignore the RTE_GROUP RTE */
1246 484278 : if (i == root->group_rtindex)
1247 0 : continue;
1248 :
1249 484278 : if (rel == NULL) /* must be an outer join */
1250 : {
1251 : Assert(bms_is_member(i, root->outer_join_rels));
1252 5114 : continue;
1253 : }
1254 :
1255 : Assert(rel->reloptkind == RELOPT_BASEREL);
1256 :
1257 479164 : rel->eclass_indexes = bms_add_member(rel->eclass_indexes,
1258 : ec_index);
1259 :
1260 479164 : if (can_generate_joinclause)
1261 170140 : rel->has_eclass_joins = true;
1262 : }
1263 :
1264 393712 : ec_index++;
1265 : }
1266 332410 : }
1267 :
1268 : /*
1269 : * generate_base_implied_equalities when EC contains pseudoconstant(s)
1270 : */
1271 : static void
1272 224472 : generate_base_implied_equalities_const(PlannerInfo *root,
1273 : EquivalenceClass *ec)
1274 : {
1275 224472 : EquivalenceMember *const_em = NULL;
1276 : ListCell *lc;
1277 :
1278 : /*
1279 : * In the trivial case where we just had one "var = const" clause, push
1280 : * the original clause back into the main planner machinery. There is
1281 : * nothing to be gained by doing it differently, and we save the effort to
1282 : * re-build and re-analyze an equality clause that will be exactly
1283 : * equivalent to the old one.
1284 : */
1285 429992 : if (list_length(ec->ec_members) == 2 &&
1286 205520 : list_length(ec->ec_sources) == 1)
1287 : {
1288 205520 : RestrictInfo *restrictinfo = (RestrictInfo *) linitial(ec->ec_sources);
1289 :
1290 205520 : distribute_restrictinfo_to_rels(root, restrictinfo);
1291 205520 : return;
1292 : }
1293 :
1294 : /* We don't expect any children yet */
1295 : Assert(ec->ec_childmembers == NULL);
1296 :
1297 : /*
1298 : * Find the constant member to use. We prefer an actual constant to
1299 : * pseudo-constants (such as Params), because the constraint exclusion
1300 : * machinery might be able to exclude relations on the basis of generated
1301 : * "var = const" equalities, but "var = param" won't work for that.
1302 : */
1303 43948 : foreach(lc, ec->ec_members)
1304 : {
1305 43868 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1306 :
1307 43868 : if (cur_em->em_is_const)
1308 : {
1309 18958 : const_em = cur_em;
1310 18958 : if (IsA(cur_em->em_expr, Const))
1311 18872 : break;
1312 : }
1313 : }
1314 : Assert(const_em != NULL);
1315 :
1316 : /* Generate a derived equality against each other member */
1317 75928 : foreach(lc, ec->ec_members)
1318 : {
1319 57006 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1320 : Oid eq_op;
1321 : RestrictInfo *rinfo;
1322 :
1323 : /* Child members should not exist in ec_members */
1324 : Assert(!cur_em->em_is_child);
1325 57006 : if (cur_em == const_em)
1326 18928 : continue;
1327 38078 : eq_op = select_equality_operator(ec,
1328 : cur_em->em_datatype,
1329 : const_em->em_datatype);
1330 38078 : if (!OidIsValid(eq_op))
1331 : {
1332 : /* failed... */
1333 30 : ec->ec_broken = true;
1334 30 : break;
1335 : }
1336 :
1337 : /*
1338 : * We use the constant's em_jdomain as qualscope, so that if the
1339 : * generated clause is variable-free (i.e, both EMs are consts) it
1340 : * will be enforced at the join domain level.
1341 : */
1342 38048 : rinfo = process_implied_equality(root, eq_op, ec->ec_collation,
1343 : cur_em->em_expr, const_em->em_expr,
1344 38048 : const_em->em_jdomain->jd_relids,
1345 : ec->ec_min_security,
1346 38048 : cur_em->em_is_const);
1347 :
1348 : /*
1349 : * If the clause didn't degenerate to a constant, fill in the correct
1350 : * markings for a mergejoinable clause, and save it as a derived
1351 : * clause. (We will not re-use such clauses directly, but selectivity
1352 : * estimation may consult those later. Note that this use of derived
1353 : * clauses does not overlap with its use for join clauses, since we
1354 : * never generate join clauses from an ec_has_const eclass.)
1355 : */
1356 38048 : if (rinfo && rinfo->mergeopfamilies)
1357 : {
1358 : /* it's not redundant, so don't set parent_ec */
1359 37916 : rinfo->left_ec = rinfo->right_ec = ec;
1360 37916 : rinfo->left_em = cur_em;
1361 37916 : rinfo->right_em = const_em;
1362 37916 : ec_add_derived_clause(ec, rinfo);
1363 : }
1364 : }
1365 : }
1366 :
1367 : /*
1368 : * generate_base_implied_equalities when EC contains no pseudoconstants
1369 : */
1370 : static void
1371 66728 : generate_base_implied_equalities_no_const(PlannerInfo *root,
1372 : EquivalenceClass *ec)
1373 : {
1374 : EquivalenceMember **prev_ems;
1375 : ListCell *lc;
1376 :
1377 : /*
1378 : * We scan the EC members once and track the last-seen member for each
1379 : * base relation. When we see another member of the same base relation,
1380 : * we generate "prev_em = cur_em". This results in the minimum number of
1381 : * derived clauses, but it's possible that it will fail when a different
1382 : * ordering would succeed. XXX FIXME: use a UNION-FIND algorithm similar
1383 : * to the way we build merged ECs. (Use a list-of-lists for each rel.)
1384 : */
1385 : prev_ems = (EquivalenceMember **)
1386 66728 : palloc0(root->simple_rel_array_size * sizeof(EquivalenceMember *));
1387 :
1388 : /* We don't expect any children yet */
1389 : Assert(ec->ec_childmembers == NULL);
1390 :
1391 202656 : foreach(lc, ec->ec_members)
1392 : {
1393 135928 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1394 : int relid;
1395 :
1396 : /* Child members should not exist in ec_members */
1397 : Assert(!cur_em->em_is_child);
1398 :
1399 135928 : if (!bms_get_singleton_member(cur_em->em_relids, &relid))
1400 192 : continue;
1401 : Assert(relid < root->simple_rel_array_size);
1402 :
1403 135736 : if (prev_ems[relid] != NULL)
1404 : {
1405 442 : EquivalenceMember *prev_em = prev_ems[relid];
1406 : Oid eq_op;
1407 : RestrictInfo *rinfo;
1408 :
1409 442 : eq_op = select_equality_operator(ec,
1410 : prev_em->em_datatype,
1411 : cur_em->em_datatype);
1412 442 : if (!OidIsValid(eq_op))
1413 : {
1414 : /* failed... */
1415 0 : ec->ec_broken = true;
1416 0 : break;
1417 : }
1418 :
1419 : /*
1420 : * The expressions aren't constants, so the passed qualscope will
1421 : * never be used to place the generated clause. We just need to
1422 : * be sure it covers both expressions, which em_relids should do.
1423 : */
1424 442 : rinfo = process_implied_equality(root, eq_op, ec->ec_collation,
1425 : prev_em->em_expr, cur_em->em_expr,
1426 : cur_em->em_relids,
1427 : ec->ec_min_security,
1428 : false);
1429 :
1430 : /*
1431 : * If the clause didn't degenerate to a constant, fill in the
1432 : * correct markings for a mergejoinable clause. We don't record
1433 : * it as a derived clause, since we don't currently need to
1434 : * re-find such clauses, and don't want to clutter the
1435 : * derived-clause set with non-join clauses.
1436 : */
1437 442 : if (rinfo && rinfo->mergeopfamilies)
1438 : {
1439 : /* it's not redundant, so don't set parent_ec */
1440 442 : rinfo->left_ec = rinfo->right_ec = ec;
1441 442 : rinfo->left_em = prev_em;
1442 442 : rinfo->right_em = cur_em;
1443 : }
1444 : }
1445 135736 : prev_ems[relid] = cur_em;
1446 : }
1447 :
1448 66728 : pfree(prev_ems);
1449 :
1450 : /*
1451 : * We also have to make sure that all the Vars used in the member clauses
1452 : * will be available at any join node we might try to reference them at.
1453 : * For the moment we force all the Vars to be available at all join nodes
1454 : * for this eclass. Perhaps this could be improved by doing some
1455 : * pre-analysis of which members we prefer to join, but it's no worse than
1456 : * what happened in the pre-8.3 code. (Note: rebuild_eclass_attr_needed
1457 : * needs to match this code.)
1458 : */
1459 202656 : foreach(lc, ec->ec_members)
1460 : {
1461 135928 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1462 135928 : List *vars = pull_var_clause((Node *) cur_em->em_expr,
1463 : PVC_RECURSE_AGGREGATES |
1464 : PVC_RECURSE_WINDOWFUNCS |
1465 : PVC_INCLUDE_PLACEHOLDERS);
1466 :
1467 135928 : add_vars_to_targetlist(root, vars, ec->ec_relids);
1468 135928 : list_free(vars);
1469 : }
1470 66728 : }
1471 :
1472 : /*
1473 : * generate_base_implied_equalities cleanup after failure
1474 : *
1475 : * What we must do here is push any zero- or one-relation source RestrictInfos
1476 : * of the EC back into the main restrictinfo datastructures. Multi-relation
1477 : * clauses will be regurgitated later by generate_join_implied_equalities().
1478 : * (We do it this way to maintain continuity with the case that ec_broken
1479 : * becomes set only after we've gone up a join level or two.) However, for
1480 : * an EC that contains constants, we can adopt a simpler strategy and just
1481 : * throw back all the source RestrictInfos immediately; that works because
1482 : * we know that such an EC can't become broken later. (This rule justifies
1483 : * ignoring ec_has_const ECs in generate_join_implied_equalities, even when
1484 : * they are broken.)
1485 : */
1486 : static void
1487 30 : generate_base_implied_equalities_broken(PlannerInfo *root,
1488 : EquivalenceClass *ec)
1489 : {
1490 : ListCell *lc;
1491 :
1492 96 : foreach(lc, ec->ec_sources)
1493 : {
1494 66 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1495 :
1496 66 : if (ec->ec_has_const ||
1497 0 : bms_membership(restrictinfo->required_relids) != BMS_MULTIPLE)
1498 66 : distribute_restrictinfo_to_rels(root, restrictinfo);
1499 : }
1500 30 : }
1501 :
1502 :
1503 : /*
1504 : * generate_join_implied_equalities
1505 : * Generate any join clauses that we can deduce from equivalence classes.
1506 : *
1507 : * At a join node, we must enforce restriction clauses sufficient to ensure
1508 : * that all equivalence-class members computable at that node are equal.
1509 : * Since the set of clauses to enforce can vary depending on which subset
1510 : * relations are the inputs, we have to compute this afresh for each join
1511 : * relation pair. Hence a fresh List of RestrictInfo nodes is built and
1512 : * passed back on each call.
1513 : *
1514 : * In addition to its use at join nodes, this can be applied to generate
1515 : * eclass-based join clauses for use in a parameterized scan of a base rel.
1516 : * The reason for the asymmetry of specifying the inner rel as a RelOptInfo
1517 : * and the outer rel by Relids is that this usage occurs before we have
1518 : * built any join RelOptInfos.
1519 : *
1520 : * An annoying special case for parameterized scans is that the inner rel can
1521 : * be an appendrel child (an "other rel"). In this case we must generate
1522 : * appropriate clauses using child EC members. add_child_rel_equivalences
1523 : * must already have been done for the child rel.
1524 : *
1525 : * The results are sufficient for use in merge, hash, and plain nestloop join
1526 : * methods. We do not worry here about selecting clauses that are optimal
1527 : * for use in a parameterized indexscan. indxpath.c makes its own selections
1528 : * of clauses to use, and if the ones we pick here are redundant with those,
1529 : * the extras will be eliminated at createplan time, using the parent_ec
1530 : * markers that we provide (see is_redundant_derived_clause()).
1531 : *
1532 : * Because the same join clauses are likely to be needed multiple times as
1533 : * we consider different join paths, we avoid generating multiple copies:
1534 : * whenever we select a particular pair of EquivalenceMembers to join,
1535 : * we check to see if the pair matches any original clause (in ec_sources)
1536 : * or previously-built derived clause. This saves memory and allows
1537 : * re-use of information cached in RestrictInfos. We also avoid generating
1538 : * commutative duplicates, i.e. if the algorithm selects "a.x = b.y" but
1539 : * we already have "b.y = a.x", we return the existing clause.
1540 : *
1541 : * If we are considering an outer join, sjinfo is the associated OJ info,
1542 : * otherwise it can be NULL.
1543 : *
1544 : * join_relids should always equal bms_union(outer_relids, inner_rel->relids)
1545 : * plus whatever add_outer_joins_to_relids() would add. We could simplify
1546 : * this function's API by computing it internally, but most callers have the
1547 : * value at hand anyway.
1548 : */
1549 : List *
1550 521660 : generate_join_implied_equalities(PlannerInfo *root,
1551 : Relids join_relids,
1552 : Relids outer_relids,
1553 : RelOptInfo *inner_rel,
1554 : SpecialJoinInfo *sjinfo)
1555 : {
1556 521660 : List *result = NIL;
1557 521660 : Relids inner_relids = inner_rel->relids;
1558 : Relids nominal_inner_relids;
1559 : Relids nominal_join_relids;
1560 : Bitmapset *matching_ecs;
1561 : int i;
1562 :
1563 : /* If inner rel is a child, extra setup work is needed */
1564 521660 : if (IS_OTHER_REL(inner_rel))
1565 : {
1566 : Assert(!bms_is_empty(inner_rel->top_parent_relids));
1567 :
1568 : /* Fetch relid set for the topmost parent rel */
1569 7334 : nominal_inner_relids = inner_rel->top_parent_relids;
1570 : /* ECs will be marked with the parent's relid, not the child's */
1571 7334 : nominal_join_relids = bms_union(outer_relids, nominal_inner_relids);
1572 7334 : nominal_join_relids = add_outer_joins_to_relids(root,
1573 : nominal_join_relids,
1574 : sjinfo,
1575 : NULL);
1576 : }
1577 : else
1578 : {
1579 514326 : nominal_inner_relids = inner_relids;
1580 514326 : nominal_join_relids = join_relids;
1581 : }
1582 :
1583 : /*
1584 : * Examine all potentially-relevant eclasses.
1585 : *
1586 : * If we are considering an outer join, we must include "join" clauses
1587 : * that mention either input rel plus the outer join's relid; these
1588 : * represent post-join filter clauses that have to be applied at this
1589 : * join. We don't have infrastructure that would let us identify such
1590 : * eclasses cheaply, so just fall back to considering all eclasses
1591 : * mentioning anything in nominal_join_relids.
1592 : *
1593 : * At inner joins, we can be smarter: only consider eclasses mentioning
1594 : * both input rels.
1595 : */
1596 521660 : if (sjinfo && sjinfo->ojrelid != 0)
1597 100654 : matching_ecs = get_eclass_indexes_for_relids(root, nominal_join_relids);
1598 : else
1599 421006 : matching_ecs = get_common_eclass_indexes(root, nominal_inner_relids,
1600 : outer_relids);
1601 :
1602 521660 : i = -1;
1603 1511614 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
1604 : {
1605 989954 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
1606 989954 : List *sublist = NIL;
1607 :
1608 : /* ECs containing consts do not need any further enforcement */
1609 989954 : if (ec->ec_has_const)
1610 134798 : continue;
1611 :
1612 : /* Single-member ECs won't generate any deductions */
1613 855156 : if (list_length(ec->ec_members) <= 1)
1614 468226 : continue;
1615 :
1616 : /* Sanity check that this eclass overlaps the join */
1617 : Assert(bms_overlap(ec->ec_relids, nominal_join_relids));
1618 :
1619 386930 : if (!ec->ec_broken)
1620 386606 : sublist = generate_join_implied_equalities_normal(root,
1621 : ec,
1622 : join_relids,
1623 : outer_relids,
1624 : inner_relids);
1625 :
1626 : /* Recover if we failed to generate required derived clauses */
1627 386930 : if (ec->ec_broken)
1628 360 : sublist = generate_join_implied_equalities_broken(root,
1629 : ec,
1630 : nominal_join_relids,
1631 : outer_relids,
1632 : nominal_inner_relids,
1633 : inner_rel);
1634 :
1635 386930 : result = list_concat(result, sublist);
1636 : }
1637 :
1638 521660 : return result;
1639 : }
1640 :
1641 : /*
1642 : * generate_join_implied_equalities_for_ecs
1643 : * As above, but consider only the listed ECs.
1644 : *
1645 : * For the sole current caller, we can assume sjinfo == NULL, that is we are
1646 : * not interested in outer-join filter clauses. This might need to change
1647 : * in future.
1648 : */
1649 : List *
1650 962 : generate_join_implied_equalities_for_ecs(PlannerInfo *root,
1651 : List *eclasses,
1652 : Relids join_relids,
1653 : Relids outer_relids,
1654 : RelOptInfo *inner_rel)
1655 : {
1656 962 : List *result = NIL;
1657 962 : Relids inner_relids = inner_rel->relids;
1658 : Relids nominal_inner_relids;
1659 : Relids nominal_join_relids;
1660 : ListCell *lc;
1661 :
1662 : /* If inner rel is a child, extra setup work is needed */
1663 962 : if (IS_OTHER_REL(inner_rel))
1664 : {
1665 : Assert(!bms_is_empty(inner_rel->top_parent_relids));
1666 :
1667 : /* Fetch relid set for the topmost parent rel */
1668 0 : nominal_inner_relids = inner_rel->top_parent_relids;
1669 : /* ECs will be marked with the parent's relid, not the child's */
1670 0 : nominal_join_relids = bms_union(outer_relids, nominal_inner_relids);
1671 : }
1672 : else
1673 : {
1674 962 : nominal_inner_relids = inner_relids;
1675 962 : nominal_join_relids = join_relids;
1676 : }
1677 :
1678 1972 : foreach(lc, eclasses)
1679 : {
1680 1010 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
1681 1010 : List *sublist = NIL;
1682 :
1683 : /* ECs containing consts do not need any further enforcement */
1684 1010 : if (ec->ec_has_const)
1685 0 : continue;
1686 :
1687 : /* Single-member ECs won't generate any deductions */
1688 1010 : if (list_length(ec->ec_members) <= 1)
1689 0 : continue;
1690 :
1691 : /* We can quickly ignore any that don't overlap the join, too */
1692 1010 : if (!bms_overlap(ec->ec_relids, nominal_join_relids))
1693 0 : continue;
1694 :
1695 1010 : if (!ec->ec_broken)
1696 1010 : sublist = generate_join_implied_equalities_normal(root,
1697 : ec,
1698 : join_relids,
1699 : outer_relids,
1700 : inner_relids);
1701 :
1702 : /* Recover if we failed to generate required derived clauses */
1703 1010 : if (ec->ec_broken)
1704 0 : sublist = generate_join_implied_equalities_broken(root,
1705 : ec,
1706 : nominal_join_relids,
1707 : outer_relids,
1708 : nominal_inner_relids,
1709 : inner_rel);
1710 :
1711 1010 : result = list_concat(result, sublist);
1712 : }
1713 :
1714 962 : return result;
1715 : }
1716 :
1717 : /*
1718 : * generate_join_implied_equalities for a still-valid EC
1719 : */
1720 : static List *
1721 387616 : generate_join_implied_equalities_normal(PlannerInfo *root,
1722 : EquivalenceClass *ec,
1723 : Relids join_relids,
1724 : Relids outer_relids,
1725 : Relids inner_relids)
1726 : {
1727 387616 : List *result = NIL;
1728 387616 : List *new_members = NIL;
1729 387616 : List *outer_members = NIL;
1730 387616 : List *inner_members = NIL;
1731 : EquivalenceMemberIterator it;
1732 : EquivalenceMember *cur_em;
1733 :
1734 : /*
1735 : * First, scan the EC to identify member values that are computable at the
1736 : * outer rel, at the inner rel, or at this relation but not in either
1737 : * input rel. The outer-rel members should already be enforced equal,
1738 : * likewise for the inner-rel members. We'll need to create clauses to
1739 : * enforce that any newly computable members are all equal to each other
1740 : * as well as to at least one input member, plus enforce at least one
1741 : * outer-rel member equal to at least one inner-rel member.
1742 : */
1743 387616 : setup_eclass_member_iterator(&it, ec, join_relids);
1744 1256510 : while ((cur_em = eclass_member_iterator_next(&it)) != NULL)
1745 : {
1746 : /*
1747 : * We don't need to check explicitly for child EC members. This test
1748 : * against join_relids will cause them to be ignored except when
1749 : * considering a child inner rel, which is what we want.
1750 : */
1751 868894 : if (!bms_is_subset(cur_em->em_relids, join_relids))
1752 87816 : continue; /* not computable yet, or wrong child */
1753 :
1754 781078 : if (bms_is_subset(cur_em->em_relids, outer_relids))
1755 458816 : outer_members = lappend(outer_members, cur_em);
1756 322262 : else if (bms_is_subset(cur_em->em_relids, inner_relids))
1757 320888 : inner_members = lappend(inner_members, cur_em);
1758 : else
1759 1374 : new_members = lappend(new_members, cur_em);
1760 : }
1761 :
1762 : /*
1763 : * First, select the joinclause if needed. We can equate any one outer
1764 : * member to any one inner member, but we have to find a datatype
1765 : * combination for which an opfamily member operator exists. If we have
1766 : * choices, we prefer simple Var members (possibly with RelabelType) since
1767 : * these are (a) cheapest to compute at runtime and (b) most likely to
1768 : * have useful statistics. Also, prefer operators that are also
1769 : * hashjoinable.
1770 : */
1771 387616 : if (outer_members && inner_members)
1772 : {
1773 309016 : EquivalenceMember *best_outer_em = NULL;
1774 309016 : EquivalenceMember *best_inner_em = NULL;
1775 309016 : Oid best_eq_op = InvalidOid;
1776 309016 : int best_score = -1;
1777 : RestrictInfo *rinfo;
1778 : ListCell *lc1;
1779 :
1780 324318 : foreach(lc1, outer_members)
1781 : {
1782 309088 : EquivalenceMember *outer_em = (EquivalenceMember *) lfirst(lc1);
1783 : ListCell *lc2;
1784 :
1785 324402 : foreach(lc2, inner_members)
1786 : {
1787 309100 : EquivalenceMember *inner_em = (EquivalenceMember *) lfirst(lc2);
1788 : Oid eq_op;
1789 : int score;
1790 :
1791 309100 : eq_op = select_equality_operator(ec,
1792 : outer_em->em_datatype,
1793 : inner_em->em_datatype);
1794 309100 : if (!OidIsValid(eq_op))
1795 36 : continue;
1796 309064 : score = 0;
1797 309064 : if (IsA(outer_em->em_expr, Var) ||
1798 17012 : (IsA(outer_em->em_expr, RelabelType) &&
1799 3938 : IsA(((RelabelType *) outer_em->em_expr)->arg, Var)))
1800 295942 : score++;
1801 309064 : if (IsA(inner_em->em_expr, Var) ||
1802 11192 : (IsA(inner_em->em_expr, RelabelType) &&
1803 8652 : IsA(((RelabelType *) inner_em->em_expr)->arg, Var)))
1804 306506 : score++;
1805 309064 : if (op_hashjoinable(eq_op,
1806 309064 : exprType((Node *) outer_em->em_expr)))
1807 308986 : score++;
1808 309064 : if (score > best_score)
1809 : {
1810 308980 : best_outer_em = outer_em;
1811 308980 : best_inner_em = inner_em;
1812 308980 : best_eq_op = eq_op;
1813 308980 : best_score = score;
1814 308980 : if (best_score == 3)
1815 293786 : break; /* no need to look further */
1816 : }
1817 : }
1818 309088 : if (best_score == 3)
1819 293786 : break; /* no need to look further */
1820 : }
1821 309016 : if (best_score < 0)
1822 : {
1823 : /* failed... */
1824 36 : ec->ec_broken = true;
1825 36 : return NIL;
1826 : }
1827 :
1828 : /*
1829 : * Create clause, setting parent_ec to mark it as redundant with other
1830 : * joinclauses
1831 : */
1832 308980 : rinfo = create_join_clause(root, ec, best_eq_op,
1833 : best_outer_em, best_inner_em,
1834 : ec);
1835 :
1836 308980 : result = lappend(result, rinfo);
1837 : }
1838 :
1839 : /*
1840 : * Now deal with building restrictions for any expressions that involve
1841 : * Vars from both sides of the join. We have to equate all of these to
1842 : * each other as well as to at least one old member (if any).
1843 : *
1844 : * XXX as in generate_base_implied_equalities_no_const, we could be a lot
1845 : * smarter here to avoid unnecessary failures in cross-type situations.
1846 : * For now, use the same left-to-right method used there.
1847 : */
1848 387580 : if (new_members)
1849 : {
1850 1338 : List *old_members = list_concat(outer_members, inner_members);
1851 1338 : EquivalenceMember *prev_em = NULL;
1852 : RestrictInfo *rinfo;
1853 : ListCell *lc1;
1854 :
1855 : /* For now, arbitrarily take the first old_member as the one to use */
1856 1338 : if (old_members)
1857 1104 : new_members = lappend(new_members, linitial(old_members));
1858 :
1859 3816 : foreach(lc1, new_members)
1860 : {
1861 2478 : cur_em = (EquivalenceMember *) lfirst(lc1);
1862 :
1863 2478 : if (prev_em != NULL)
1864 : {
1865 : Oid eq_op;
1866 :
1867 1140 : eq_op = select_equality_operator(ec,
1868 : prev_em->em_datatype,
1869 : cur_em->em_datatype);
1870 1140 : if (!OidIsValid(eq_op))
1871 : {
1872 : /* failed... */
1873 0 : ec->ec_broken = true;
1874 0 : return NIL;
1875 : }
1876 : /* do NOT set parent_ec, this qual is not redundant! */
1877 1140 : rinfo = create_join_clause(root, ec, eq_op,
1878 : prev_em, cur_em,
1879 : NULL);
1880 :
1881 1140 : result = lappend(result, rinfo);
1882 : }
1883 2478 : prev_em = cur_em;
1884 : }
1885 : }
1886 :
1887 387580 : return result;
1888 : }
1889 :
1890 : /*
1891 : * generate_join_implied_equalities cleanup after failure
1892 : *
1893 : * Return any original RestrictInfos that are enforceable at this join.
1894 : *
1895 : * In the case of a child inner relation, we have to translate the
1896 : * original RestrictInfos from parent to child Vars.
1897 : */
1898 : static List *
1899 360 : generate_join_implied_equalities_broken(PlannerInfo *root,
1900 : EquivalenceClass *ec,
1901 : Relids nominal_join_relids,
1902 : Relids outer_relids,
1903 : Relids nominal_inner_relids,
1904 : RelOptInfo *inner_rel)
1905 : {
1906 360 : List *result = NIL;
1907 : ListCell *lc;
1908 :
1909 984 : foreach(lc, ec->ec_sources)
1910 : {
1911 624 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1912 624 : Relids clause_relids = restrictinfo->required_relids;
1913 :
1914 624 : if (bms_is_subset(clause_relids, nominal_join_relids) &&
1915 336 : !bms_is_subset(clause_relids, outer_relids) &&
1916 312 : !bms_is_subset(clause_relids, nominal_inner_relids))
1917 312 : result = lappend(result, restrictinfo);
1918 : }
1919 :
1920 : /*
1921 : * If we have to translate, just brute-force apply adjust_appendrel_attrs
1922 : * to all the RestrictInfos at once. This will result in returning
1923 : * RestrictInfos that are not included in EC's derived clauses, but there
1924 : * shouldn't be any duplication, and it's a sufficiently narrow corner
1925 : * case that we shouldn't sweat too much over it anyway.
1926 : *
1927 : * Since inner_rel might be an indirect descendant of the baserel
1928 : * mentioned in the ec_sources clauses, we have to be prepared to apply
1929 : * multiple levels of Var translation.
1930 : */
1931 360 : if (IS_OTHER_REL(inner_rel) && result != NIL)
1932 162 : result = (List *) adjust_appendrel_attrs_multilevel(root,
1933 : (Node *) result,
1934 : inner_rel,
1935 162 : inner_rel->top_parent);
1936 :
1937 360 : return result;
1938 : }
1939 :
1940 :
1941 : /*
1942 : * select_equality_operator
1943 : * Select a suitable equality operator for comparing two EC members
1944 : *
1945 : * Returns InvalidOid if no operator can be found for this datatype combination
1946 : */
1947 : static Oid
1948 460414 : select_equality_operator(EquivalenceClass *ec, Oid lefttype, Oid righttype)
1949 : {
1950 : ListCell *lc;
1951 :
1952 460480 : foreach(lc, ec->ec_opfamilies)
1953 : {
1954 460414 : Oid opfamily = lfirst_oid(lc);
1955 : Oid opno;
1956 :
1957 460414 : opno = get_opfamily_member_for_cmptype(opfamily, lefttype, righttype, COMPARE_EQ);
1958 460414 : if (!OidIsValid(opno))
1959 66 : continue;
1960 : /* If no barrier quals in query, don't worry about leaky operators */
1961 460348 : if (ec->ec_max_security == 0)
1962 460348 : return opno;
1963 : /* Otherwise, insist that selected operators be leakproof */
1964 428 : if (get_func_leakproof(get_opcode(opno)))
1965 428 : return opno;
1966 : }
1967 66 : return InvalidOid;
1968 : }
1969 :
1970 :
1971 : /*
1972 : * create_join_clause
1973 : * Find or make a RestrictInfo comparing the two given EC members
1974 : * with the given operator (or, possibly, its commutator, because
1975 : * the ordering of the operands in the result is not guaranteed).
1976 : *
1977 : * parent_ec is either equal to ec (if the clause is a potentially-redundant
1978 : * join clause) or NULL (if not). We have to treat this as part of the
1979 : * match requirements --- it's possible that a clause comparing the same two
1980 : * EMs is a join clause in one join path and a restriction clause in another.
1981 : */
1982 : static RestrictInfo *
1983 423874 : create_join_clause(PlannerInfo *root,
1984 : EquivalenceClass *ec, Oid opno,
1985 : EquivalenceMember *leftem,
1986 : EquivalenceMember *rightem,
1987 : EquivalenceClass *parent_ec)
1988 : {
1989 : RestrictInfo *rinfo;
1990 423874 : RestrictInfo *parent_rinfo = NULL;
1991 : MemoryContext oldcontext;
1992 :
1993 423874 : rinfo = ec_search_clause_for_ems(root, ec, leftem, rightem, parent_ec);
1994 423874 : if (rinfo)
1995 348566 : return rinfo;
1996 :
1997 : /*
1998 : * Not there, so build it, in planner context so we can re-use it. (Not
1999 : * important in normal planning, but definitely so in GEQO.)
2000 : */
2001 75308 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
2002 :
2003 : /*
2004 : * If either EM is a child, recursively create the corresponding
2005 : * parent-to-parent clause, so that we can duplicate its rinfo_serial.
2006 : */
2007 75308 : if (leftem->em_is_child || rightem->em_is_child)
2008 : {
2009 4122 : EquivalenceMember *leftp = leftem->em_parent ? leftem->em_parent : leftem;
2010 4122 : EquivalenceMember *rightp = rightem->em_parent ? rightem->em_parent : rightem;
2011 :
2012 4122 : parent_rinfo = create_join_clause(root, ec, opno,
2013 : leftp, rightp,
2014 : parent_ec);
2015 : }
2016 :
2017 75308 : rinfo = build_implied_join_equality(root,
2018 : opno,
2019 : ec->ec_collation,
2020 : leftem->em_expr,
2021 : rightem->em_expr,
2022 75308 : bms_union(leftem->em_relids,
2023 75308 : rightem->em_relids),
2024 : ec->ec_min_security);
2025 :
2026 : /*
2027 : * If either EM is a child, force the clause's clause_relids to include
2028 : * the relid(s) of the child rel. In normal cases it would already, but
2029 : * not if we are considering appendrel child relations with pseudoconstant
2030 : * translated variables (i.e., UNION ALL sub-selects with constant output
2031 : * items). We must do this so that join_clause_is_movable_into() will
2032 : * think that the clause should be evaluated at the correct place.
2033 : */
2034 75308 : if (leftem->em_is_child)
2035 3576 : rinfo->clause_relids = bms_add_members(rinfo->clause_relids,
2036 3576 : leftem->em_relids);
2037 75308 : if (rightem->em_is_child)
2038 546 : rinfo->clause_relids = bms_add_members(rinfo->clause_relids,
2039 546 : rightem->em_relids);
2040 :
2041 : /* If it's a child clause, copy the parent's rinfo_serial */
2042 75308 : if (parent_rinfo)
2043 4122 : rinfo->rinfo_serial = parent_rinfo->rinfo_serial;
2044 :
2045 : /* Mark the clause as redundant, or not */
2046 75308 : rinfo->parent_ec = parent_ec;
2047 :
2048 : /*
2049 : * We know the correct values for left_ec/right_ec, ie this particular EC,
2050 : * so we can just set them directly instead of forcing another lookup.
2051 : */
2052 75308 : rinfo->left_ec = ec;
2053 75308 : rinfo->right_ec = ec;
2054 :
2055 : /* Mark it as usable with these EMs */
2056 75308 : rinfo->left_em = leftem;
2057 75308 : rinfo->right_em = rightem;
2058 : /* and save it for possible re-use */
2059 75308 : ec_add_derived_clause(ec, rinfo);
2060 :
2061 75308 : MemoryContextSwitchTo(oldcontext);
2062 :
2063 75308 : return rinfo;
2064 : }
2065 :
2066 :
2067 : /*
2068 : * reconsider_outer_join_clauses
2069 : * Re-examine any outer-join clauses that were set aside by
2070 : * distribute_qual_to_rels(), and see if we can derive any
2071 : * EquivalenceClasses from them. Then, if they were not made
2072 : * redundant, push them out into the regular join-clause lists.
2073 : *
2074 : * When we have mergejoinable clauses A = B that are outer-join clauses,
2075 : * we can't blindly combine them with other clauses A = C to deduce B = C,
2076 : * since in fact the "equality" A = B won't necessarily hold above the
2077 : * outer join (one of the variables might be NULL instead). Nonetheless
2078 : * there are cases where we can add qual clauses using transitivity.
2079 : *
2080 : * One case that we look for here is an outer-join clause OUTERVAR = INNERVAR
2081 : * for which there is also an equivalence clause OUTERVAR = CONSTANT.
2082 : * It is safe and useful to push a clause INNERVAR = CONSTANT into the
2083 : * evaluation of the inner (nullable) relation, because any inner rows not
2084 : * meeting this condition will not contribute to the outer-join result anyway.
2085 : * (Any outer rows they could join to will be eliminated by the pushed-down
2086 : * equivalence clause.)
2087 : *
2088 : * Note that the above rule does not work for full outer joins; nor is it
2089 : * very interesting to consider cases where the generated equivalence clause
2090 : * would involve relations outside the outer join, since such clauses couldn't
2091 : * be pushed into the inner side's scan anyway. So the restriction to
2092 : * outervar = pseudoconstant is not really giving up anything.
2093 : *
2094 : * For full-join cases, we can only do something useful if it's a FULL JOIN
2095 : * USING and a merged column has an equivalence MERGEDVAR = CONSTANT.
2096 : * By the time it gets here, the merged column will look like
2097 : * COALESCE(LEFTVAR, RIGHTVAR)
2098 : * and we will have a full-join clause LEFTVAR = RIGHTVAR that we can match
2099 : * the COALESCE expression to. In this situation we can push LEFTVAR = CONSTANT
2100 : * and RIGHTVAR = CONSTANT into the input relations, since any rows not
2101 : * meeting these conditions cannot contribute to the join result.
2102 : *
2103 : * Again, there isn't any traction to be gained by trying to deal with
2104 : * clauses comparing a mergedvar to a non-pseudoconstant. So we can make
2105 : * use of the EquivalenceClasses to search for matching variables that were
2106 : * equivalenced to constants. The interesting outer-join clauses were
2107 : * accumulated for us by distribute_qual_to_rels.
2108 : *
2109 : * When we find one of these cases, we implement the changes we want by
2110 : * generating a new equivalence clause INNERVAR = CONSTANT (or LEFTVAR, etc)
2111 : * and pushing it into the EquivalenceClass structures. This is because we
2112 : * may already know that INNERVAR is equivalenced to some other var(s), and
2113 : * we'd like the constant to propagate to them too. Note that it would be
2114 : * unsafe to merge any existing EC for INNERVAR with the OUTERVAR's EC ---
2115 : * that could result in propagating constant restrictions from
2116 : * INNERVAR to OUTERVAR, which would be very wrong.
2117 : *
2118 : * It's possible that the INNERVAR is also an OUTERVAR for some other
2119 : * outer-join clause, in which case the process can be repeated. So we repeat
2120 : * looping over the lists of clauses until no further deductions can be made.
2121 : * Whenever we do make a deduction, we remove the generating clause from the
2122 : * lists, since we don't want to make the same deduction twice.
2123 : *
2124 : * If we don't find any match for a set-aside outer join clause, we must
2125 : * throw it back into the regular joinclause processing by passing it to
2126 : * distribute_restrictinfo_to_rels(). If we do generate a derived clause,
2127 : * however, the outer-join clause is redundant. We must still put some
2128 : * clause into the regular processing, because otherwise the join will be
2129 : * seen as a clauseless join and avoided during join order searching.
2130 : * We handle this by generating a constant-TRUE clause that is marked with
2131 : * the same required_relids etc as the removed outer-join clause, thus
2132 : * making it a join clause between the correct relations.
2133 : */
2134 : void
2135 334372 : reconsider_outer_join_clauses(PlannerInfo *root)
2136 : {
2137 : bool found;
2138 : ListCell *cell;
2139 :
2140 : /* Outer loop repeats until we find no more deductions */
2141 : do
2142 : {
2143 334372 : found = false;
2144 :
2145 : /* Process the LEFT JOIN clauses */
2146 362716 : foreach(cell, root->left_join_clauses)
2147 : {
2148 28344 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2149 :
2150 28344 : if (reconsider_outer_join_clause(root, ojcinfo, true))
2151 : {
2152 920 : RestrictInfo *rinfo = ojcinfo->rinfo;
2153 :
2154 920 : found = true;
2155 : /* remove it from the list */
2156 920 : root->left_join_clauses =
2157 920 : foreach_delete_current(root->left_join_clauses, cell);
2158 : /* throw back a dummy replacement clause (see notes above) */
2159 920 : rinfo = make_restrictinfo(root,
2160 920 : (Expr *) makeBoolConst(true, false),
2161 920 : rinfo->is_pushed_down,
2162 920 : rinfo->has_clone,
2163 920 : rinfo->is_clone,
2164 : false, /* pseudoconstant */
2165 : 0, /* security_level */
2166 : rinfo->required_relids,
2167 : rinfo->incompatible_relids,
2168 : rinfo->outer_relids);
2169 920 : distribute_restrictinfo_to_rels(root, rinfo);
2170 : }
2171 : }
2172 :
2173 : /* Process the RIGHT JOIN clauses */
2174 363528 : foreach(cell, root->right_join_clauses)
2175 : {
2176 29156 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2177 :
2178 29156 : if (reconsider_outer_join_clause(root, ojcinfo, false))
2179 : {
2180 1048 : RestrictInfo *rinfo = ojcinfo->rinfo;
2181 :
2182 1048 : found = true;
2183 : /* remove it from the list */
2184 1048 : root->right_join_clauses =
2185 1048 : foreach_delete_current(root->right_join_clauses, cell);
2186 : /* throw back a dummy replacement clause (see notes above) */
2187 1048 : rinfo = make_restrictinfo(root,
2188 1048 : (Expr *) makeBoolConst(true, false),
2189 1048 : rinfo->is_pushed_down,
2190 1048 : rinfo->has_clone,
2191 1048 : rinfo->is_clone,
2192 : false, /* pseudoconstant */
2193 : 0, /* security_level */
2194 : rinfo->required_relids,
2195 : rinfo->incompatible_relids,
2196 : rinfo->outer_relids);
2197 1048 : distribute_restrictinfo_to_rels(root, rinfo);
2198 : }
2199 : }
2200 :
2201 : /* Process the FULL JOIN clauses */
2202 335618 : foreach(cell, root->full_join_clauses)
2203 : {
2204 1246 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2205 :
2206 1246 : if (reconsider_full_join_clause(root, ojcinfo))
2207 : {
2208 6 : RestrictInfo *rinfo = ojcinfo->rinfo;
2209 :
2210 6 : found = true;
2211 : /* remove it from the list */
2212 6 : root->full_join_clauses =
2213 6 : foreach_delete_current(root->full_join_clauses, cell);
2214 : /* throw back a dummy replacement clause (see notes above) */
2215 6 : rinfo = make_restrictinfo(root,
2216 6 : (Expr *) makeBoolConst(true, false),
2217 6 : rinfo->is_pushed_down,
2218 6 : rinfo->has_clone,
2219 6 : rinfo->is_clone,
2220 : false, /* pseudoconstant */
2221 : 0, /* security_level */
2222 : rinfo->required_relids,
2223 : rinfo->incompatible_relids,
2224 : rinfo->outer_relids);
2225 6 : distribute_restrictinfo_to_rels(root, rinfo);
2226 : }
2227 : }
2228 334372 : } while (found);
2229 :
2230 : /* Now, any remaining clauses have to be thrown back */
2231 359112 : foreach(cell, root->left_join_clauses)
2232 : {
2233 26702 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2234 :
2235 26702 : distribute_restrictinfo_to_rels(root, ojcinfo->rinfo);
2236 : }
2237 359428 : foreach(cell, root->right_join_clauses)
2238 : {
2239 27018 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2240 :
2241 27018 : distribute_restrictinfo_to_rels(root, ojcinfo->rinfo);
2242 : }
2243 333650 : foreach(cell, root->full_join_clauses)
2244 : {
2245 1240 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2246 :
2247 1240 : distribute_restrictinfo_to_rels(root, ojcinfo->rinfo);
2248 : }
2249 332410 : }
2250 :
2251 : /*
2252 : * reconsider_outer_join_clauses for a single LEFT/RIGHT JOIN clause
2253 : *
2254 : * Returns true if we were able to propagate a constant through the clause.
2255 : */
2256 : static bool
2257 57500 : reconsider_outer_join_clause(PlannerInfo *root, OuterJoinClauseInfo *ojcinfo,
2258 : bool outer_on_left)
2259 : {
2260 57500 : RestrictInfo *rinfo = ojcinfo->rinfo;
2261 57500 : SpecialJoinInfo *sjinfo = ojcinfo->sjinfo;
2262 : Expr *outervar,
2263 : *innervar;
2264 : Oid opno,
2265 : collation,
2266 : left_type,
2267 : right_type,
2268 : inner_datatype;
2269 : Relids inner_relids;
2270 : ListCell *lc1;
2271 :
2272 : Assert(is_opclause(rinfo->clause));
2273 57500 : opno = ((OpExpr *) rinfo->clause)->opno;
2274 57500 : collation = ((OpExpr *) rinfo->clause)->inputcollid;
2275 :
2276 : /* Extract needed info from the clause */
2277 57500 : op_input_types(opno, &left_type, &right_type);
2278 57500 : if (outer_on_left)
2279 : {
2280 28344 : outervar = (Expr *) get_leftop(rinfo->clause);
2281 28344 : innervar = (Expr *) get_rightop(rinfo->clause);
2282 28344 : inner_datatype = right_type;
2283 28344 : inner_relids = rinfo->right_relids;
2284 : }
2285 : else
2286 : {
2287 29156 : outervar = (Expr *) get_rightop(rinfo->clause);
2288 29156 : innervar = (Expr *) get_leftop(rinfo->clause);
2289 29156 : inner_datatype = left_type;
2290 29156 : inner_relids = rinfo->left_relids;
2291 : }
2292 :
2293 : /* Scan EquivalenceClasses for a match to outervar */
2294 366500 : foreach(lc1, root->eq_classes)
2295 : {
2296 310968 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2297 : bool match;
2298 : ListCell *lc2;
2299 :
2300 : /* We don't expect any children yet */
2301 : Assert(cur_ec->ec_childmembers == NULL);
2302 :
2303 : /* Ignore EC unless it contains pseudoconstants */
2304 310968 : if (!cur_ec->ec_has_const)
2305 234132 : continue;
2306 : /* Never match to a volatile EC */
2307 76836 : if (cur_ec->ec_has_volatile)
2308 0 : continue;
2309 : /* It has to match the outer-join clause as to semantics, too */
2310 76836 : if (collation != cur_ec->ec_collation)
2311 2750 : continue;
2312 74086 : if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
2313 17216 : continue;
2314 : /* Does it contain a match to outervar? */
2315 56870 : match = false;
2316 176480 : foreach(lc2, cur_ec->ec_members)
2317 : {
2318 121578 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2319 :
2320 : /* Child members should not exist in ec_members */
2321 : Assert(!cur_em->em_is_child);
2322 121578 : if (equal(outervar, cur_em->em_expr))
2323 : {
2324 1968 : match = true;
2325 1968 : break;
2326 : }
2327 : }
2328 56870 : if (!match)
2329 54902 : continue; /* no match, so ignore this EC */
2330 :
2331 : /*
2332 : * Yes it does! Try to generate a clause INNERVAR = CONSTANT for each
2333 : * CONSTANT in the EC. Note that we must succeed with at least one
2334 : * constant before we can decide to throw away the outer-join clause.
2335 : */
2336 1968 : match = false;
2337 6958 : foreach(lc2, cur_ec->ec_members)
2338 : {
2339 4990 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2340 : Oid eq_op;
2341 : RestrictInfo *newrinfo;
2342 : JoinDomain *jdomain;
2343 :
2344 4990 : if (!cur_em->em_is_const)
2345 2980 : continue; /* ignore non-const members */
2346 2010 : eq_op = select_equality_operator(cur_ec,
2347 : inner_datatype,
2348 : cur_em->em_datatype);
2349 2010 : if (!OidIsValid(eq_op))
2350 0 : continue; /* can't generate equality */
2351 2010 : newrinfo = build_implied_join_equality(root,
2352 : eq_op,
2353 : cur_ec->ec_collation,
2354 : innervar,
2355 : cur_em->em_expr,
2356 : bms_copy(inner_relids),
2357 : cur_ec->ec_min_security);
2358 : /* This equality holds within the OJ's child JoinDomain */
2359 2010 : jdomain = find_join_domain(root, sjinfo->syn_righthand);
2360 2010 : if (process_equivalence(root, &newrinfo, jdomain))
2361 2010 : match = true;
2362 : }
2363 :
2364 : /*
2365 : * If we were able to equate INNERVAR to any constant, report success.
2366 : * Otherwise, fall out of the search loop, since we know the OUTERVAR
2367 : * appears in at most one EC.
2368 : */
2369 1968 : if (match)
2370 1968 : return true;
2371 : else
2372 0 : break;
2373 : }
2374 :
2375 55532 : return false; /* failed to make any deduction */
2376 : }
2377 :
2378 : /*
2379 : * reconsider_outer_join_clauses for a single FULL JOIN clause
2380 : *
2381 : * Returns true if we were able to propagate a constant through the clause.
2382 : */
2383 : static bool
2384 1246 : reconsider_full_join_clause(PlannerInfo *root, OuterJoinClauseInfo *ojcinfo)
2385 : {
2386 1246 : RestrictInfo *rinfo = ojcinfo->rinfo;
2387 1246 : SpecialJoinInfo *sjinfo = ojcinfo->sjinfo;
2388 1246 : Relids fjrelids = bms_make_singleton(sjinfo->ojrelid);
2389 : Expr *leftvar;
2390 : Expr *rightvar;
2391 : Oid opno,
2392 : collation,
2393 : left_type,
2394 : right_type;
2395 : Relids left_relids,
2396 : right_relids;
2397 : ListCell *lc1;
2398 :
2399 : /* Extract needed info from the clause */
2400 : Assert(is_opclause(rinfo->clause));
2401 1246 : opno = ((OpExpr *) rinfo->clause)->opno;
2402 1246 : collation = ((OpExpr *) rinfo->clause)->inputcollid;
2403 1246 : op_input_types(opno, &left_type, &right_type);
2404 1246 : leftvar = (Expr *) get_leftop(rinfo->clause);
2405 1246 : rightvar = (Expr *) get_rightop(rinfo->clause);
2406 1246 : left_relids = rinfo->left_relids;
2407 1246 : right_relids = rinfo->right_relids;
2408 :
2409 6318 : foreach(lc1, root->eq_classes)
2410 : {
2411 5078 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2412 5078 : EquivalenceMember *coal_em = NULL;
2413 : bool match;
2414 : bool matchleft;
2415 : bool matchright;
2416 : ListCell *lc2;
2417 5078 : int coal_idx = -1;
2418 :
2419 : /* We don't expect any children yet */
2420 : Assert(cur_ec->ec_childmembers == NULL);
2421 :
2422 : /* Ignore EC unless it contains pseudoconstants */
2423 5078 : if (!cur_ec->ec_has_const)
2424 4782 : continue;
2425 : /* Never match to a volatile EC */
2426 296 : if (cur_ec->ec_has_volatile)
2427 0 : continue;
2428 : /* It has to match the outer-join clause as to semantics, too */
2429 296 : if (collation != cur_ec->ec_collation)
2430 36 : continue;
2431 260 : if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
2432 0 : continue;
2433 :
2434 : /*
2435 : * Does it contain a COALESCE(leftvar, rightvar) construct?
2436 : *
2437 : * We can assume the COALESCE() inputs are in the same order as the
2438 : * join clause, since both were automatically generated in the cases
2439 : * we care about.
2440 : *
2441 : * XXX currently this may fail to match in cross-type cases because
2442 : * the COALESCE will contain typecast operations while the join clause
2443 : * may not (if there is a cross-type mergejoin operator available for
2444 : * the two column types). Is it OK to strip implicit coercions from
2445 : * the COALESCE arguments?
2446 : */
2447 260 : match = false;
2448 762 : foreach(lc2, cur_ec->ec_members)
2449 : {
2450 508 : coal_em = (EquivalenceMember *) lfirst(lc2);
2451 :
2452 : /* Child members should not exist in ec_members */
2453 : Assert(!coal_em->em_is_child);
2454 508 : if (IsA(coal_em->em_expr, CoalesceExpr))
2455 : {
2456 18 : CoalesceExpr *cexpr = (CoalesceExpr *) coal_em->em_expr;
2457 : Node *cfirst;
2458 : Node *csecond;
2459 :
2460 18 : if (list_length(cexpr->args) != 2)
2461 12 : continue;
2462 6 : cfirst = (Node *) linitial(cexpr->args);
2463 6 : csecond = (Node *) lsecond(cexpr->args);
2464 :
2465 : /*
2466 : * The COALESCE arguments will be marked as possibly nulled by
2467 : * the full join, while we wish to generate clauses that apply
2468 : * to the join's inputs. So we must strip the join from the
2469 : * nullingrels fields of cfirst/csecond before comparing them
2470 : * to leftvar/rightvar. (Perhaps with a less hokey
2471 : * representation for FULL JOIN USING output columns, this
2472 : * wouldn't be needed?)
2473 : */
2474 6 : cfirst = remove_nulling_relids(cfirst, fjrelids, NULL);
2475 6 : csecond = remove_nulling_relids(csecond, fjrelids, NULL);
2476 :
2477 6 : if (equal(leftvar, cfirst) && equal(rightvar, csecond))
2478 : {
2479 6 : coal_idx = foreach_current_index(lc2);
2480 6 : match = true;
2481 6 : break;
2482 : }
2483 : }
2484 : }
2485 260 : if (!match)
2486 254 : continue; /* no match, so ignore this EC */
2487 :
2488 : /*
2489 : * Yes it does! Try to generate clauses LEFTVAR = CONSTANT and
2490 : * RIGHTVAR = CONSTANT for each CONSTANT in the EC. Note that we must
2491 : * succeed with at least one constant for each var before we can
2492 : * decide to throw away the outer-join clause.
2493 : */
2494 6 : matchleft = matchright = false;
2495 18 : foreach(lc2, cur_ec->ec_members)
2496 : {
2497 12 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2498 : Oid eq_op;
2499 : RestrictInfo *newrinfo;
2500 : JoinDomain *jdomain;
2501 :
2502 12 : if (!cur_em->em_is_const)
2503 6 : continue; /* ignore non-const members */
2504 6 : eq_op = select_equality_operator(cur_ec,
2505 : left_type,
2506 : cur_em->em_datatype);
2507 6 : if (OidIsValid(eq_op))
2508 : {
2509 6 : newrinfo = build_implied_join_equality(root,
2510 : eq_op,
2511 : cur_ec->ec_collation,
2512 : leftvar,
2513 : cur_em->em_expr,
2514 : bms_copy(left_relids),
2515 : cur_ec->ec_min_security);
2516 : /* This equality holds within the lefthand child JoinDomain */
2517 6 : jdomain = find_join_domain(root, sjinfo->syn_lefthand);
2518 6 : if (process_equivalence(root, &newrinfo, jdomain))
2519 6 : matchleft = true;
2520 : }
2521 6 : eq_op = select_equality_operator(cur_ec,
2522 : right_type,
2523 : cur_em->em_datatype);
2524 6 : if (OidIsValid(eq_op))
2525 : {
2526 6 : newrinfo = build_implied_join_equality(root,
2527 : eq_op,
2528 : cur_ec->ec_collation,
2529 : rightvar,
2530 : cur_em->em_expr,
2531 : bms_copy(right_relids),
2532 : cur_ec->ec_min_security);
2533 : /* This equality holds within the righthand child JoinDomain */
2534 6 : jdomain = find_join_domain(root, sjinfo->syn_righthand);
2535 6 : if (process_equivalence(root, &newrinfo, jdomain))
2536 6 : matchright = true;
2537 : }
2538 : }
2539 :
2540 : /*
2541 : * If we were able to equate both vars to constants, we're done, and
2542 : * we can throw away the full-join clause as redundant. Moreover, we
2543 : * can remove the COALESCE entry from the EC, since the added
2544 : * restrictions ensure it will always have the expected value. (We
2545 : * don't bother trying to update ec_relids or ec_sources.)
2546 : */
2547 6 : if (matchleft && matchright)
2548 : {
2549 6 : cur_ec->ec_members = list_delete_nth_cell(cur_ec->ec_members, coal_idx);
2550 6 : return true;
2551 : }
2552 :
2553 : /*
2554 : * Otherwise, fall out of the search loop, since we know the COALESCE
2555 : * appears in at most one EC (XXX might stop being true if we allow
2556 : * stripping of coercions above?)
2557 : */
2558 0 : break;
2559 : }
2560 :
2561 1240 : return false; /* failed to make any deduction */
2562 : }
2563 :
2564 : /*
2565 : * rebuild_eclass_attr_needed
2566 : * Put back attr_needed bits for Vars/PHVs needed for join eclasses.
2567 : *
2568 : * This is used to rebuild attr_needed/ph_needed sets after removal of a
2569 : * useless outer join. It should match what
2570 : * generate_base_implied_equalities_no_const did, except that we call
2571 : * add_vars_to_attr_needed not add_vars_to_targetlist.
2572 : */
2573 : void
2574 11296 : rebuild_eclass_attr_needed(PlannerInfo *root)
2575 : {
2576 : ListCell *lc;
2577 :
2578 59722 : foreach(lc, root->eq_classes)
2579 : {
2580 48426 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
2581 :
2582 : /*
2583 : * We don't expect any EC child members to exist at this point. Ensure
2584 : * that's the case, otherwise, we might be getting asked to do
2585 : * something this function hasn't been coded for.
2586 : */
2587 : Assert(ec->ec_childmembers == NULL);
2588 :
2589 : /* Need do anything only for a multi-member, no-const EC. */
2590 48426 : if (list_length(ec->ec_members) > 1 && !ec->ec_has_const)
2591 : {
2592 : ListCell *lc2;
2593 :
2594 12060 : foreach(lc2, ec->ec_members)
2595 : {
2596 8054 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2597 8054 : List *vars = pull_var_clause((Node *) cur_em->em_expr,
2598 : PVC_RECURSE_AGGREGATES |
2599 : PVC_RECURSE_WINDOWFUNCS |
2600 : PVC_INCLUDE_PLACEHOLDERS);
2601 :
2602 8054 : add_vars_to_attr_needed(root, vars, ec->ec_relids);
2603 8054 : list_free(vars);
2604 : }
2605 : }
2606 : }
2607 11296 : }
2608 :
2609 : /*
2610 : * find_join_domain
2611 : * Find the highest JoinDomain enclosed within the given relid set.
2612 : *
2613 : * (We could avoid this search at the cost of complicating APIs elsewhere,
2614 : * which doesn't seem worth it.)
2615 : */
2616 : static JoinDomain *
2617 2022 : find_join_domain(PlannerInfo *root, Relids relids)
2618 : {
2619 : ListCell *lc;
2620 :
2621 4134 : foreach(lc, root->join_domains)
2622 : {
2623 4134 : JoinDomain *jdomain = (JoinDomain *) lfirst(lc);
2624 :
2625 4134 : if (bms_is_subset(jdomain->jd_relids, relids))
2626 2022 : return jdomain;
2627 : }
2628 0 : elog(ERROR, "failed to find appropriate JoinDomain");
2629 : return NULL; /* keep compiler quiet */
2630 : }
2631 :
2632 :
2633 : /*
2634 : * exprs_known_equal
2635 : * Detect whether two expressions are known equal due to equivalence
2636 : * relationships.
2637 : *
2638 : * If opfamily is given, the expressions must be known equal per the semantics
2639 : * of that opfamily (note it has to be a btree opfamily, since those are the
2640 : * only opfamilies equivclass.c deals with). If opfamily is InvalidOid, we'll
2641 : * return true if they're equal according to any opfamily, which is fuzzy but
2642 : * OK for estimation purposes.
2643 : *
2644 : * Note: does not bother to check for "equal(item1, item2)"; caller must
2645 : * check that case if it's possible to pass identical items.
2646 : */
2647 : bool
2648 4772 : exprs_known_equal(PlannerInfo *root, Node *item1, Node *item2, Oid opfamily)
2649 : {
2650 : ListCell *lc1;
2651 :
2652 65592 : foreach(lc1, root->eq_classes)
2653 : {
2654 60994 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2655 60994 : bool item1member = false;
2656 60994 : bool item2member = false;
2657 : ListCell *lc2;
2658 :
2659 : /* Never match to a volatile EC */
2660 60994 : if (ec->ec_has_volatile)
2661 0 : continue;
2662 :
2663 : /*
2664 : * It's okay to consider ec_broken ECs here. Brokenness just means we
2665 : * couldn't derive all the implied clauses we'd have liked to; it does
2666 : * not invalidate our knowledge that the members are equal.
2667 : */
2668 :
2669 : /* Ignore if this EC doesn't use specified opfamily */
2670 60994 : if (OidIsValid(opfamily) &&
2671 648 : !list_member_oid(ec->ec_opfamilies, opfamily))
2672 228 : continue;
2673 :
2674 : /* Ignore children here */
2675 133192 : foreach(lc2, ec->ec_members)
2676 : {
2677 72600 : EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
2678 :
2679 : /* Child members should not exist in ec_members */
2680 : Assert(!em->em_is_child);
2681 72600 : if (equal(item1, em->em_expr))
2682 3016 : item1member = true;
2683 69584 : else if (equal(item2, em->em_expr))
2684 3440 : item2member = true;
2685 : /* Exit as soon as equality is proven */
2686 72600 : if (item1member && item2member)
2687 174 : return true;
2688 : }
2689 : }
2690 4598 : return false;
2691 : }
2692 :
2693 :
2694 : /*
2695 : * match_eclasses_to_foreign_key_col
2696 : * See whether a foreign key column match is proven by any eclass.
2697 : *
2698 : * If the referenced and referencing Vars of the fkey's colno'th column are
2699 : * known equal due to any eclass, return that eclass; otherwise return NULL.
2700 : * (In principle there might be more than one matching eclass if multiple
2701 : * collations are involved, but since collation doesn't matter for equality,
2702 : * we ignore that fine point here.) This is much like exprs_known_equal,
2703 : * except for the format of the input.
2704 : *
2705 : * On success, we also set fkinfo->eclass[colno] to the matching eclass,
2706 : * and set fkinfo->fk_eclass_member[colno] to the eclass member for the
2707 : * referencing Var.
2708 : */
2709 : EquivalenceClass *
2710 2402 : match_eclasses_to_foreign_key_col(PlannerInfo *root,
2711 : ForeignKeyOptInfo *fkinfo,
2712 : int colno)
2713 : {
2714 2402 : Index var1varno = fkinfo->con_relid;
2715 2402 : AttrNumber var1attno = fkinfo->conkey[colno];
2716 2402 : Index var2varno = fkinfo->ref_relid;
2717 2402 : AttrNumber var2attno = fkinfo->confkey[colno];
2718 2402 : Oid eqop = fkinfo->conpfeqop[colno];
2719 2402 : RelOptInfo *rel1 = root->simple_rel_array[var1varno];
2720 2402 : RelOptInfo *rel2 = root->simple_rel_array[var2varno];
2721 2402 : List *opfamilies = NIL; /* compute only if needed */
2722 : Bitmapset *matching_ecs;
2723 : int i;
2724 :
2725 : /* Consider only eclasses mentioning both relations */
2726 : Assert(root->ec_merging_done);
2727 : Assert(IS_SIMPLE_REL(rel1));
2728 : Assert(IS_SIMPLE_REL(rel2));
2729 2402 : matching_ecs = bms_intersect(rel1->eclass_indexes,
2730 2402 : rel2->eclass_indexes);
2731 :
2732 2402 : i = -1;
2733 2498 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
2734 : {
2735 438 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes,
2736 : i);
2737 438 : EquivalenceMember *item1_em = NULL;
2738 438 : EquivalenceMember *item2_em = NULL;
2739 : ListCell *lc2;
2740 :
2741 : /* Never match to a volatile EC */
2742 438 : if (ec->ec_has_volatile)
2743 0 : continue;
2744 :
2745 : /*
2746 : * It's okay to consider "broken" ECs here, see exprs_known_equal.
2747 : * Ignore children here.
2748 : */
2749 1074 : foreach(lc2, ec->ec_members)
2750 : {
2751 978 : EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
2752 : Var *var;
2753 :
2754 : /* Child members should not exist in ec_members */
2755 : Assert(!em->em_is_child);
2756 :
2757 : /* EM must be a Var, possibly with RelabelType */
2758 978 : var = (Var *) em->em_expr;
2759 978 : while (var && IsA(var, RelabelType))
2760 0 : var = (Var *) ((RelabelType *) var)->arg;
2761 978 : if (!(var && IsA(var, Var)))
2762 6 : continue;
2763 :
2764 : /* Match? */
2765 972 : if (var->varno == var1varno && var->varattno == var1attno)
2766 342 : item1_em = em;
2767 630 : else if (var->varno == var2varno && var->varattno == var2attno)
2768 342 : item2_em = em;
2769 :
2770 : /* Have we found both PK and FK column in this EC? */
2771 972 : if (item1_em && item2_em)
2772 : {
2773 : /*
2774 : * Succeed if eqop matches EC's opfamilies. We could test
2775 : * this before scanning the members, but it's probably cheaper
2776 : * to test for member matches first.
2777 : */
2778 342 : if (opfamilies == NIL) /* compute if we didn't already */
2779 342 : opfamilies = get_mergejoin_opfamilies(eqop);
2780 342 : if (equal(opfamilies, ec->ec_opfamilies))
2781 : {
2782 342 : fkinfo->eclass[colno] = ec;
2783 342 : fkinfo->fk_eclass_member[colno] = item2_em;
2784 342 : return ec;
2785 : }
2786 : /* Otherwise, done with this EC, move on to the next */
2787 0 : break;
2788 : }
2789 : }
2790 : }
2791 2060 : return NULL;
2792 : }
2793 :
2794 : /*
2795 : * find_derived_clause_for_ec_member
2796 : * Search for a previously-derived clause mentioning the given EM.
2797 : *
2798 : * The eclass should be an ec_has_const EC, of which the EM is a non-const
2799 : * member. This should ensure there is just one derived clause mentioning
2800 : * the EM (and equating it to a constant).
2801 : * Returns NULL if no such clause can be found.
2802 : */
2803 : RestrictInfo *
2804 6 : find_derived_clause_for_ec_member(PlannerInfo *root,
2805 : EquivalenceClass *ec,
2806 : EquivalenceMember *em)
2807 : {
2808 : Assert(ec->ec_has_const);
2809 : Assert(!em->em_is_const);
2810 :
2811 6 : return ec_search_derived_clause_for_ems(root, ec, em, NULL, NULL);
2812 : }
2813 :
2814 :
2815 : /*
2816 : * add_child_rel_equivalences
2817 : * Search for EC members that reference the root parent of child_rel, and
2818 : * add transformed members referencing the child_rel.
2819 : *
2820 : * Note that this function won't be called at all unless we have at least some
2821 : * reason to believe that the EC members it generates will be useful.
2822 : *
2823 : * parent_rel and child_rel could be derived from appinfo, but since the
2824 : * caller has already computed them, we might as well just pass them in.
2825 : *
2826 : * The passed-in AppendRelInfo is not used when the parent_rel is not a
2827 : * top-level baserel, since it shows the mapping from the parent_rel but
2828 : * we need to translate EC expressions that refer to the top-level parent.
2829 : * Using it is faster than using adjust_appendrel_attrs_multilevel(), though,
2830 : * so we prefer it when we can.
2831 : */
2832 : void
2833 28810 : add_child_rel_equivalences(PlannerInfo *root,
2834 : AppendRelInfo *appinfo,
2835 : RelOptInfo *parent_rel,
2836 : RelOptInfo *child_rel)
2837 : {
2838 28810 : Relids top_parent_relids = child_rel->top_parent_relids;
2839 28810 : Relids child_relids = child_rel->relids;
2840 : int i;
2841 :
2842 : /*
2843 : * EC merging should be complete already, so we can use the parent rel's
2844 : * eclass_indexes to avoid searching all of root->eq_classes.
2845 : */
2846 : Assert(root->ec_merging_done);
2847 : Assert(IS_SIMPLE_REL(parent_rel));
2848 :
2849 28810 : i = -1;
2850 79114 : while ((i = bms_next_member(parent_rel->eclass_indexes, i)) >= 0)
2851 : {
2852 50304 : EquivalenceClass *cur_ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
2853 :
2854 : /*
2855 : * If this EC contains a volatile expression, then generating child
2856 : * EMs would be downright dangerous, so skip it. We rely on a
2857 : * volatile EC having only one EM.
2858 : */
2859 50304 : if (cur_ec->ec_has_volatile)
2860 0 : continue;
2861 :
2862 : /* Sanity check eclass_indexes only contain ECs for parent_rel */
2863 : Assert(bms_is_subset(top_parent_relids, cur_ec->ec_relids));
2864 :
2865 174952 : foreach_node(EquivalenceMember, cur_em, cur_ec->ec_members)
2866 : {
2867 74344 : if (cur_em->em_is_const)
2868 3276 : continue; /* ignore consts here */
2869 :
2870 : /* Child members should not exist in ec_members */
2871 : Assert(!cur_em->em_is_child);
2872 :
2873 : /*
2874 : * Consider only members that reference and can be computed at
2875 : * child's topmost parent rel. In particular we want to exclude
2876 : * parent-rel Vars that have nonempty varnullingrels. Translating
2877 : * those might fail, if the transformed expression wouldn't be a
2878 : * simple Var; and in any case it wouldn't produce a member that
2879 : * has any use in creating plans for the child rel.
2880 : */
2881 71068 : if (bms_is_subset(cur_em->em_relids, top_parent_relids) &&
2882 48030 : !bms_is_empty(cur_em->em_relids))
2883 : {
2884 : /* OK, generate transformed child version */
2885 : Expr *child_expr;
2886 : Relids new_relids;
2887 :
2888 48030 : if (parent_rel->reloptkind == RELOPT_BASEREL)
2889 : {
2890 : /* Simple single-level transformation */
2891 : child_expr = (Expr *)
2892 40020 : adjust_appendrel_attrs(root,
2893 40020 : (Node *) cur_em->em_expr,
2894 : 1, &appinfo);
2895 : }
2896 : else
2897 : {
2898 : /* Must do multi-level transformation */
2899 : child_expr = (Expr *)
2900 8010 : adjust_appendrel_attrs_multilevel(root,
2901 8010 : (Node *) cur_em->em_expr,
2902 : child_rel,
2903 8010 : child_rel->top_parent);
2904 : }
2905 :
2906 : /*
2907 : * Transform em_relids to match. Note we do *not* do
2908 : * pull_varnos(child_expr) here, as for example the
2909 : * transformation might have substituted a constant, but we
2910 : * don't want the child member to be marked as constant.
2911 : */
2912 48030 : new_relids = bms_difference(cur_em->em_relids,
2913 : top_parent_relids);
2914 48030 : new_relids = bms_add_members(new_relids, child_relids);
2915 :
2916 48030 : add_child_eq_member(root,
2917 : cur_ec,
2918 : i,
2919 : child_expr,
2920 : new_relids,
2921 : cur_em->em_jdomain,
2922 : cur_em,
2923 : cur_em->em_datatype,
2924 : child_rel->relid);
2925 : }
2926 : }
2927 : }
2928 28810 : }
2929 :
2930 : /*
2931 : * add_child_join_rel_equivalences
2932 : * Like add_child_rel_equivalences(), but for joinrels
2933 : *
2934 : * Here we find the ECs relevant to the top parent joinrel and add transformed
2935 : * member expressions that refer to this child joinrel.
2936 : *
2937 : * Note that this function won't be called at all unless we have at least some
2938 : * reason to believe that the EC members it generates will be useful.
2939 : */
2940 : void
2941 4378 : add_child_join_rel_equivalences(PlannerInfo *root,
2942 : int nappinfos, AppendRelInfo **appinfos,
2943 : RelOptInfo *parent_joinrel,
2944 : RelOptInfo *child_joinrel)
2945 : {
2946 4378 : Relids top_parent_relids = child_joinrel->top_parent_relids;
2947 4378 : Relids child_relids = child_joinrel->relids;
2948 : Bitmapset *matching_ecs;
2949 : MemoryContext oldcontext;
2950 : int i;
2951 :
2952 : Assert(IS_JOIN_REL(child_joinrel) && IS_JOIN_REL(parent_joinrel));
2953 :
2954 : /* We need consider only ECs that mention the parent joinrel */
2955 4378 : matching_ecs = get_eclass_indexes_for_relids(root, top_parent_relids);
2956 :
2957 : /*
2958 : * If we're being called during GEQO join planning, we still have to
2959 : * create any new EC members in the main planner context, to avoid having
2960 : * a corrupt EC data structure after the GEQO context is reset. This is
2961 : * problematic since we'll leak memory across repeated GEQO cycles. For
2962 : * now, though, bloat is better than crash. If it becomes a real issue
2963 : * we'll have to do something to avoid generating duplicate EC members.
2964 : */
2965 4378 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
2966 :
2967 4378 : i = -1;
2968 20166 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
2969 : {
2970 15788 : EquivalenceClass *cur_ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
2971 :
2972 : /*
2973 : * If this EC contains a volatile expression, then generating child
2974 : * EMs would be downright dangerous, so skip it. We rely on a
2975 : * volatile EC having only one EM.
2976 : */
2977 15788 : if (cur_ec->ec_has_volatile)
2978 0 : continue;
2979 :
2980 : /* Sanity check on get_eclass_indexes_for_relids result */
2981 : Assert(bms_overlap(top_parent_relids, cur_ec->ec_relids));
2982 :
2983 54152 : foreach_node(EquivalenceMember, cur_em, cur_ec->ec_members)
2984 : {
2985 22576 : if (cur_em->em_is_const)
2986 2232 : continue; /* ignore consts here */
2987 :
2988 : /* Child members should not exist in ec_members */
2989 : Assert(!cur_em->em_is_child);
2990 :
2991 : /*
2992 : * We may ignore expressions that reference a single baserel,
2993 : * because add_child_rel_equivalences should have handled them.
2994 : */
2995 20344 : if (bms_membership(cur_em->em_relids) != BMS_MULTIPLE)
2996 17722 : continue;
2997 :
2998 : /* Does this member reference child's topmost parent rel? */
2999 2622 : if (bms_overlap(cur_em->em_relids, top_parent_relids))
3000 : {
3001 : /* Yes, generate transformed child version */
3002 : Expr *child_expr;
3003 : Relids new_relids;
3004 :
3005 2622 : if (parent_joinrel->reloptkind == RELOPT_JOINREL)
3006 : {
3007 : /* Simple single-level transformation */
3008 : child_expr = (Expr *)
3009 2526 : adjust_appendrel_attrs(root,
3010 2526 : (Node *) cur_em->em_expr,
3011 : nappinfos, appinfos);
3012 : }
3013 : else
3014 : {
3015 : /* Must do multi-level transformation */
3016 : Assert(parent_joinrel->reloptkind == RELOPT_OTHER_JOINREL);
3017 : child_expr = (Expr *)
3018 96 : adjust_appendrel_attrs_multilevel(root,
3019 96 : (Node *) cur_em->em_expr,
3020 : child_joinrel,
3021 96 : child_joinrel->top_parent);
3022 : }
3023 :
3024 : /*
3025 : * Transform em_relids to match. Note we do *not* do
3026 : * pull_varnos(child_expr) here, as for example the
3027 : * transformation might have substituted a constant, but we
3028 : * don't want the child member to be marked as constant.
3029 : */
3030 2622 : new_relids = bms_difference(cur_em->em_relids,
3031 : top_parent_relids);
3032 2622 : new_relids = bms_add_members(new_relids, child_relids);
3033 :
3034 : /*
3035 : * Add new child member to the EquivalenceClass. Because this
3036 : * is a RELOPT_OTHER_JOINREL which has multiple component
3037 : * relids, there is no ideal place to store these members in
3038 : * the class. Ordinarily, child members are stored in the
3039 : * ec_childmembers[] array element corresponding to their
3040 : * relid, however, here we have multiple component relids, so
3041 : * there's no single ec_childmembers[] array element to store
3042 : * this member. So that we still correctly find this member
3043 : * in loops iterating over an EquivalenceMemberIterator, we
3044 : * opt to store the member in the ec_childmembers array in
3045 : * only the first component relid slot of the array. This
3046 : * allows the member to be found, providing callers of
3047 : * setup_eclass_member_iterator() specify all the component
3048 : * relids for the RELOPT_OTHER_JOINREL, which they do. If we
3049 : * opted to store the member in each ec_childmembers[] element
3050 : * for all the component relids, then that would just result
3051 : * in eclass_member_iterator_next() finding the member
3052 : * multiple times, which is a waste of effort.
3053 : */
3054 2622 : add_child_eq_member(root,
3055 : cur_ec,
3056 : -1,
3057 : child_expr,
3058 : new_relids,
3059 : cur_em->em_jdomain,
3060 : cur_em,
3061 : cur_em->em_datatype,
3062 2622 : bms_next_member(child_joinrel->relids, -1));
3063 : }
3064 : }
3065 : }
3066 :
3067 4378 : MemoryContextSwitchTo(oldcontext);
3068 4378 : }
3069 :
3070 : /*
3071 : * add_setop_child_rel_equivalences
3072 : * Add equivalence members for each non-resjunk target in 'child_tlist'
3073 : * to the EquivalenceClass in the corresponding setop_pathkey's pk_eclass.
3074 : *
3075 : * 'root' is the PlannerInfo belonging to the top-level set operation.
3076 : * 'child_rel' is the RelOptInfo of the child relation we're adding
3077 : * EquivalenceMembers for.
3078 : * 'child_tlist' is the target list for the setop child relation. The target
3079 : * list expressions are what we add as EquivalenceMembers.
3080 : * 'setop_pathkeys' is a list of PathKeys which must contain an entry for each
3081 : * non-resjunk target in 'child_tlist'.
3082 : */
3083 : void
3084 11902 : add_setop_child_rel_equivalences(PlannerInfo *root, RelOptInfo *child_rel,
3085 : List *child_tlist, List *setop_pathkeys)
3086 : {
3087 : ListCell *lc;
3088 11902 : ListCell *lc2 = list_head(setop_pathkeys);
3089 :
3090 48312 : foreach(lc, child_tlist)
3091 : {
3092 36410 : TargetEntry *tle = lfirst_node(TargetEntry, lc);
3093 : EquivalenceMember *parent_em;
3094 : PathKey *pk;
3095 :
3096 36410 : if (tle->resjunk)
3097 0 : continue;
3098 :
3099 36410 : if (lc2 == NULL)
3100 0 : elog(ERROR, "too few pathkeys for set operation");
3101 :
3102 36410 : pk = lfirst_node(PathKey, lc2);
3103 36410 : parent_em = linitial(pk->pk_eclass->ec_members);
3104 :
3105 : /*
3106 : * We can safely pass the parent member as the first member in the
3107 : * ec_members list as this is added first in generate_union_paths,
3108 : * likewise, the JoinDomain can be that of the initial member of the
3109 : * Pathkey's EquivalenceClass. We pass -1 for ec_index since we
3110 : * maintain the eclass_indexes for the child_rel after the loop.
3111 : */
3112 36410 : add_child_eq_member(root,
3113 : pk->pk_eclass,
3114 : -1,
3115 : tle->expr,
3116 : child_rel->relids,
3117 : parent_em->em_jdomain,
3118 : parent_em,
3119 36410 : exprType((Node *) tle->expr),
3120 : child_rel->relid);
3121 :
3122 36410 : lc2 = lnext(setop_pathkeys, lc2);
3123 : }
3124 :
3125 : /*
3126 : * transformSetOperationStmt() ensures that the targetlist never contains
3127 : * any resjunk columns, so all eclasses that exist in 'root' must have
3128 : * received a new member in the loop above. Add them to the child_rel's
3129 : * eclass_indexes.
3130 : */
3131 11902 : child_rel->eclass_indexes = bms_add_range(child_rel->eclass_indexes, 0,
3132 11902 : list_length(root->eq_classes) - 1);
3133 11902 : }
3134 :
3135 : /*
3136 : * setup_eclass_member_iterator
3137 : * Setup an EquivalenceMemberIterator 'it' to iterate over all parent
3138 : * EquivalenceMembers and child members belonging to the given 'ec'.
3139 : *
3140 : * This iterator returns:
3141 : * - All parent members stored directly in ec_members for 'ec', and;
3142 : * - Any child member added to the given ec by add_child_eq_member() where
3143 : * the child_relid specified in the add_child_eq_member() call is a member
3144 : * of the 'child_relids' parameter.
3145 : *
3146 : * Note:
3147 : * The given 'child_relids' must remain allocated and not be changed for the
3148 : * lifetime of the iterator.
3149 : *
3150 : * Parameters:
3151 : * 'it' is a pointer to the iterator to set up. Normally stack allocated.
3152 : * 'ec' is the EquivalenceClass from which to iterate members for.
3153 : * 'child_relids' is the relids to return child members for.
3154 : */
3155 : void
3156 4765654 : setup_eclass_member_iterator(EquivalenceMemberIterator *it,
3157 : EquivalenceClass *ec, Relids child_relids)
3158 : {
3159 4765654 : it->ec = ec;
3160 : /* no need to set this if the class has no child members array set */
3161 4765654 : it->child_relids = ec->ec_childmembers != NULL ? child_relids : NULL;
3162 4765654 : it->current_relid = -1;
3163 4765654 : it->current_list = ec->ec_members;
3164 4765654 : it->current_cell = list_head(it->current_list);
3165 4765654 : }
3166 :
3167 : /*
3168 : * eclass_member_iterator_next
3169 : * Get the next EquivalenceMember from the EquivalenceMemberIterator 'it',
3170 : * as setup by setup_eclass_member_iterator(). NULL is returned if there
3171 : * are no members left, after which callers must not call
3172 : * eclass_member_iterator_next() again for the given iterator.
3173 : */
3174 : EquivalenceMember *
3175 11139580 : eclass_member_iterator_next(EquivalenceMemberIterator *it)
3176 : {
3177 11139580 : while (it->current_list != NULL)
3178 : {
3179 11118264 : while (it->current_cell != NULL)
3180 : {
3181 : EquivalenceMember *em;
3182 :
3183 7657636 : nextcell:
3184 7782540 : em = lfirst_node(EquivalenceMember, it->current_cell);
3185 7782540 : it->current_cell = lnext(it->current_list, it->current_cell);
3186 7782540 : return em;
3187 : }
3188 :
3189 : /* Search for the next list to return members from */
3190 3576970 : while ((it->current_relid = bms_next_member(it->child_relids, it->current_relid)) > 0)
3191 : {
3192 : /*
3193 : * Be paranoid in case we're given relids above what we've sized
3194 : * the ec_childmembers array to.
3195 : */
3196 241246 : if (it->current_relid >= it->ec->ec_childmembers_size)
3197 0 : return NULL;
3198 :
3199 241246 : it->current_list = it->ec->ec_childmembers[it->current_relid];
3200 :
3201 : /* If there are members in this list, use it. */
3202 241246 : if (it->current_list != NIL)
3203 : {
3204 : /* point current_cell to the head of this list */
3205 124904 : it->current_cell = list_head(it->current_list);
3206 124904 : goto nextcell;
3207 : }
3208 : }
3209 3335724 : return NULL;
3210 : }
3211 :
3212 21316 : return NULL;
3213 : }
3214 :
3215 : /*
3216 : * generate_implied_equalities_for_column
3217 : * Create EC-derived joinclauses usable with a specific column.
3218 : *
3219 : * This is used by indxpath.c to extract potentially indexable joinclauses
3220 : * from ECs, and can be used by foreign data wrappers for similar purposes.
3221 : * We assume that only expressions in Vars of a single table are of interest,
3222 : * but the caller provides a callback function to identify exactly which
3223 : * such expressions it would like to know about.
3224 : *
3225 : * We assume that any given table/index column could appear in only one EC.
3226 : * (This should be true in all but the most pathological cases, and if it
3227 : * isn't, we stop on the first match anyway.) Therefore, what we return
3228 : * is a redundant list of clauses equating the table/index column to each of
3229 : * the other-relation values it is known to be equal to. Any one of
3230 : * these clauses can be used to create a parameterized path, and there
3231 : * is no value in using more than one. (But it *is* worthwhile to create
3232 : * a separate parameterized path for each one, since that leads to different
3233 : * join orders.)
3234 : *
3235 : * The caller can pass a Relids set of rels we aren't interested in joining
3236 : * to, so as to save the work of creating useless clauses.
3237 : */
3238 : List *
3239 613444 : generate_implied_equalities_for_column(PlannerInfo *root,
3240 : RelOptInfo *rel,
3241 : ec_matches_callback_type callback,
3242 : void *callback_arg,
3243 : Relids prohibited_rels)
3244 : {
3245 613444 : List *result = NIL;
3246 613444 : bool is_child_rel = (rel->reloptkind == RELOPT_OTHER_MEMBER_REL);
3247 : Relids parent_relids;
3248 : int i;
3249 :
3250 : /* Should be OK to rely on eclass_indexes */
3251 : Assert(root->ec_merging_done);
3252 :
3253 : /* Indexes are available only on base or "other" member relations. */
3254 : Assert(IS_SIMPLE_REL(rel));
3255 :
3256 : /* If it's a child rel, we'll need to know what its parent(s) are */
3257 613444 : if (is_child_rel)
3258 11974 : parent_relids = find_childrel_parents(root, rel);
3259 : else
3260 601470 : parent_relids = NULL; /* not used, but keep compiler quiet */
3261 :
3262 613444 : i = -1;
3263 1788674 : while ((i = bms_next_member(rel->eclass_indexes, i)) >= 0)
3264 : {
3265 1277828 : EquivalenceClass *cur_ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
3266 : EquivalenceMemberIterator it;
3267 : EquivalenceMember *cur_em;
3268 : ListCell *lc2;
3269 :
3270 : /* Sanity check eclass_indexes only contain ECs for rel */
3271 : Assert(is_child_rel || bms_is_subset(rel->relids, cur_ec->ec_relids));
3272 :
3273 : /*
3274 : * Won't generate joinclauses if const or single-member (the latter
3275 : * test covers the volatile case too)
3276 : */
3277 1277828 : if (cur_ec->ec_has_const || list_length(cur_ec->ec_members) <= 1)
3278 1174784 : continue;
3279 :
3280 : /*
3281 : * Scan members, looking for a match to the target column. Note that
3282 : * child EC members are considered, but only when they belong to the
3283 : * target relation. (Unlike regular members, the same expression
3284 : * could be a child member of more than one EC. Therefore, it's
3285 : * potentially order-dependent which EC a child relation's target
3286 : * column gets matched to. This is annoying but it only happens in
3287 : * corner cases, so for now we live with just reporting the first
3288 : * match. See also get_eclass_for_sort_expr.)
3289 : */
3290 546172 : setup_eclass_member_iterator(&it, cur_ec, rel->relids);
3291 1527968 : while ((cur_em = eclass_member_iterator_next(&it)) != NULL)
3292 : {
3293 1632172 : if (bms_equal(cur_em->em_relids, rel->relids) &&
3294 547332 : callback(root, rel, cur_ec, cur_em, callback_arg))
3295 103044 : break;
3296 : }
3297 :
3298 546172 : if (!cur_em)
3299 443128 : continue;
3300 :
3301 : /*
3302 : * Found our match. Scan the other EC members and attempt to generate
3303 : * joinclauses. Ignore children here.
3304 : */
3305 316316 : foreach(lc2, cur_ec->ec_members)
3306 : {
3307 213272 : EquivalenceMember *other_em = (EquivalenceMember *) lfirst(lc2);
3308 : Oid eq_op;
3309 : RestrictInfo *rinfo;
3310 :
3311 : /* Child members should not exist in ec_members */
3312 : Assert(!other_em->em_is_child);
3313 :
3314 : /* Make sure it'll be a join to a different rel */
3315 326616 : if (other_em == cur_em ||
3316 113344 : bms_overlap(other_em->em_relids, rel->relids))
3317 100284 : continue;
3318 :
3319 : /* Forget it if caller doesn't want joins to this rel */
3320 112988 : if (bms_overlap(other_em->em_relids, prohibited_rels))
3321 156 : continue;
3322 :
3323 : /*
3324 : * Also, if this is a child rel, avoid generating a useless join
3325 : * to its parent rel(s).
3326 : */
3327 119692 : if (is_child_rel &&
3328 6860 : bms_overlap(parent_relids, other_em->em_relids))
3329 3200 : continue;
3330 :
3331 109632 : eq_op = select_equality_operator(cur_ec,
3332 : cur_em->em_datatype,
3333 : other_em->em_datatype);
3334 109632 : if (!OidIsValid(eq_op))
3335 0 : continue;
3336 :
3337 : /* set parent_ec to mark as redundant with other joinclauses */
3338 109632 : rinfo = create_join_clause(root, cur_ec, eq_op,
3339 : cur_em, other_em,
3340 : cur_ec);
3341 :
3342 109632 : result = lappend(result, rinfo);
3343 : }
3344 :
3345 : /*
3346 : * If somehow we failed to create any join clauses, we might as well
3347 : * keep scanning the ECs for another match. But if we did make any,
3348 : * we're done, because we don't want to return non-redundant clauses.
3349 : */
3350 103044 : if (result)
3351 102598 : break;
3352 : }
3353 :
3354 613444 : return result;
3355 : }
3356 :
3357 : /*
3358 : * have_relevant_eclass_joinclause
3359 : * Detect whether there is an EquivalenceClass that could produce
3360 : * a joinclause involving the two given relations.
3361 : *
3362 : * This is essentially a very cut-down version of
3363 : * generate_join_implied_equalities(). Note it's OK to occasionally say "yes"
3364 : * incorrectly. Hence we don't bother with details like whether the lack of a
3365 : * cross-type operator might prevent the clause from actually being generated.
3366 : * False negatives are not always fatal either: they will discourage, but not
3367 : * completely prevent, investigation of particular join pathways.
3368 : */
3369 : bool
3370 182694 : have_relevant_eclass_joinclause(PlannerInfo *root,
3371 : RelOptInfo *rel1, RelOptInfo *rel2)
3372 : {
3373 : Bitmapset *matching_ecs;
3374 : int i;
3375 :
3376 : /*
3377 : * Examine only eclasses mentioning both rel1 and rel2.
3378 : *
3379 : * Note that we do not consider the possibility of an eclass generating
3380 : * "join" clauses that mention just one of the rels plus an outer join
3381 : * that could be formed from them. Although such clauses must be
3382 : * correctly enforced when we form the outer join, they don't seem like
3383 : * sufficient reason to prioritize this join over other ones. The join
3384 : * ordering rules will force the join to be made when necessary.
3385 : */
3386 182694 : matching_ecs = get_common_eclass_indexes(root, rel1->relids,
3387 : rel2->relids);
3388 :
3389 182694 : i = -1;
3390 182712 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
3391 : {
3392 154778 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes,
3393 : i);
3394 :
3395 : /*
3396 : * Sanity check that get_common_eclass_indexes gave only ECs
3397 : * containing both rels.
3398 : */
3399 : Assert(bms_overlap(rel1->relids, ec->ec_relids));
3400 : Assert(bms_overlap(rel2->relids, ec->ec_relids));
3401 :
3402 : /*
3403 : * Won't generate joinclauses if single-member (this test covers the
3404 : * volatile case too)
3405 : */
3406 154778 : if (list_length(ec->ec_members) <= 1)
3407 18 : continue;
3408 :
3409 : /*
3410 : * We do not need to examine the individual members of the EC, because
3411 : * all that we care about is whether each rel overlaps the relids of
3412 : * at least one member, and get_common_eclass_indexes() and the single
3413 : * member check above are sufficient to prove that. (As with
3414 : * have_relevant_joinclause(), it is not necessary that the EC be able
3415 : * to form a joinclause relating exactly the two given rels, only that
3416 : * it be able to form a joinclause mentioning both, and this will
3417 : * surely be true if both of them overlap ec_relids.)
3418 : *
3419 : * Note we don't test ec_broken; if we did, we'd need a separate code
3420 : * path to look through ec_sources. Checking the membership anyway is
3421 : * OK as a possibly-overoptimistic heuristic.
3422 : *
3423 : * We don't test ec_has_const either, even though a const eclass won't
3424 : * generate real join clauses. This is because if we had "WHERE a.x =
3425 : * b.y and a.x = 42", it is worth considering a join between a and b,
3426 : * since the join result is likely to be small even though it'll end
3427 : * up being an unqualified nestloop.
3428 : */
3429 :
3430 154760 : return true;
3431 : }
3432 :
3433 27934 : return false;
3434 : }
3435 :
3436 :
3437 : /*
3438 : * has_relevant_eclass_joinclause
3439 : * Detect whether there is an EquivalenceClass that could produce
3440 : * a joinclause involving the given relation and anything else.
3441 : *
3442 : * This is the same as have_relevant_eclass_joinclause with the other rel
3443 : * implicitly defined as "everything else in the query".
3444 : */
3445 : bool
3446 212424 : has_relevant_eclass_joinclause(PlannerInfo *root, RelOptInfo *rel1)
3447 : {
3448 : Bitmapset *matched_ecs;
3449 : int i;
3450 :
3451 : /* Examine only eclasses mentioning rel1 */
3452 212424 : matched_ecs = get_eclass_indexes_for_relids(root, rel1->relids);
3453 :
3454 212424 : i = -1;
3455 753696 : while ((i = bms_next_member(matched_ecs, i)) >= 0)
3456 : {
3457 618876 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes,
3458 : i);
3459 :
3460 : /*
3461 : * Won't generate joinclauses if single-member (this test covers the
3462 : * volatile case too)
3463 : */
3464 618876 : if (list_length(ec->ec_members) <= 1)
3465 286354 : continue;
3466 :
3467 : /*
3468 : * Per the comment in have_relevant_eclass_joinclause, it's sufficient
3469 : * to find an EC that mentions both this rel and some other rel.
3470 : */
3471 332522 : if (!bms_is_subset(ec->ec_relids, rel1->relids))
3472 77604 : return true;
3473 : }
3474 :
3475 134820 : return false;
3476 : }
3477 :
3478 :
3479 : /*
3480 : * eclass_useful_for_merging
3481 : * Detect whether the EC could produce any mergejoinable join clauses
3482 : * against the specified relation.
3483 : *
3484 : * This is just a heuristic test and doesn't have to be exact; it's better
3485 : * to say "yes" incorrectly than "no". Hence we don't bother with details
3486 : * like whether the lack of a cross-type operator might prevent the clause
3487 : * from actually being generated.
3488 : */
3489 : bool
3490 845322 : eclass_useful_for_merging(PlannerInfo *root,
3491 : EquivalenceClass *eclass,
3492 : RelOptInfo *rel)
3493 : {
3494 : Relids relids;
3495 : ListCell *lc;
3496 :
3497 : Assert(!eclass->ec_merged);
3498 :
3499 : /*
3500 : * Won't generate joinclauses if const or single-member (the latter test
3501 : * covers the volatile case too)
3502 : */
3503 845322 : if (eclass->ec_has_const || list_length(eclass->ec_members) <= 1)
3504 79992 : return false;
3505 :
3506 : /*
3507 : * Note we don't test ec_broken; if we did, we'd need a separate code path
3508 : * to look through ec_sources. Checking the members anyway is OK as a
3509 : * possibly-overoptimistic heuristic.
3510 : */
3511 :
3512 : /* If specified rel is a child, we must consider the topmost parent rel */
3513 765330 : if (IS_OTHER_REL(rel))
3514 : {
3515 : Assert(!bms_is_empty(rel->top_parent_relids));
3516 10934 : relids = rel->top_parent_relids;
3517 : }
3518 : else
3519 754396 : relids = rel->relids;
3520 :
3521 : /* If rel already includes all members of eclass, no point in searching */
3522 765330 : if (bms_is_subset(eclass->ec_relids, relids))
3523 278094 : return false;
3524 :
3525 : /*
3526 : * To join, we need a member not in the given rel. Ignore children here.
3527 : */
3528 765914 : foreach(lc, eclass->ec_members)
3529 : {
3530 765392 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
3531 :
3532 : /* Child members should not exist in ec_members */
3533 : Assert(!cur_em->em_is_child);
3534 :
3535 765392 : if (!bms_overlap(cur_em->em_relids, relids))
3536 486714 : return true;
3537 : }
3538 :
3539 522 : return false;
3540 : }
3541 :
3542 :
3543 : /*
3544 : * is_redundant_derived_clause
3545 : * Test whether rinfo is derived from same EC as any clause in clauselist;
3546 : * if so, it can be presumed to represent a condition that's redundant
3547 : * with that member of the list.
3548 : */
3549 : bool
3550 84 : is_redundant_derived_clause(RestrictInfo *rinfo, List *clauselist)
3551 : {
3552 84 : EquivalenceClass *parent_ec = rinfo->parent_ec;
3553 : ListCell *lc;
3554 :
3555 : /* Fail if it's not a potentially-redundant clause from some EC */
3556 84 : if (parent_ec == NULL)
3557 84 : return false;
3558 :
3559 0 : foreach(lc, clauselist)
3560 : {
3561 0 : RestrictInfo *otherrinfo = (RestrictInfo *) lfirst(lc);
3562 :
3563 0 : if (otherrinfo->parent_ec == parent_ec)
3564 0 : return true;
3565 : }
3566 :
3567 0 : return false;
3568 : }
3569 :
3570 : /*
3571 : * is_redundant_with_indexclauses
3572 : * Test whether rinfo is redundant with any clause in the IndexClause
3573 : * list. Here, for convenience, we test both simple identity and
3574 : * whether it is derived from the same EC as any member of the list.
3575 : */
3576 : bool
3577 1431560 : is_redundant_with_indexclauses(RestrictInfo *rinfo, List *indexclauses)
3578 : {
3579 1431560 : EquivalenceClass *parent_ec = rinfo->parent_ec;
3580 : ListCell *lc;
3581 :
3582 1978272 : foreach(lc, indexclauses)
3583 : {
3584 1465166 : IndexClause *iclause = lfirst_node(IndexClause, lc);
3585 1465166 : RestrictInfo *otherrinfo = iclause->rinfo;
3586 :
3587 : /* If indexclause is lossy, it won't enforce the condition exactly */
3588 1465166 : if (iclause->lossy)
3589 7220 : continue;
3590 :
3591 : /* Match if it's same clause (pointer equality should be enough) */
3592 1457946 : if (rinfo == otherrinfo)
3593 918454 : return true;
3594 : /* Match if derived from same EC */
3595 539852 : if (parent_ec && otherrinfo->parent_ec == parent_ec)
3596 360 : return true;
3597 :
3598 : /*
3599 : * No need to look at the derived clauses in iclause->indexquals; they
3600 : * couldn't match if the parent clause didn't.
3601 : */
3602 : }
3603 :
3604 513106 : return false;
3605 : }
3606 :
3607 : /*
3608 : * get_eclass_indexes_for_relids
3609 : * Build and return a Bitmapset containing the indexes into root's
3610 : * eq_classes list for all eclasses that mention any of these relids
3611 : */
3612 : static Bitmapset *
3613 1053636 : get_eclass_indexes_for_relids(PlannerInfo *root, Relids relids)
3614 : {
3615 1053636 : Bitmapset *ec_indexes = NULL;
3616 1053636 : int i = -1;
3617 :
3618 : /* Should be OK to rely on eclass_indexes */
3619 : Assert(root->ec_merging_done);
3620 :
3621 3386002 : while ((i = bms_next_member(relids, i)) > 0)
3622 : {
3623 2332366 : RelOptInfo *rel = root->simple_rel_array[i];
3624 :
3625 : /* ignore the RTE_GROUP RTE */
3626 2332366 : if (i == root->group_rtindex)
3627 0 : continue;
3628 :
3629 2332366 : if (rel == NULL) /* must be an outer join */
3630 : {
3631 : Assert(bms_is_member(i, root->outer_join_rels));
3632 348018 : continue;
3633 : }
3634 :
3635 1984348 : ec_indexes = bms_add_members(ec_indexes, rel->eclass_indexes);
3636 : }
3637 1053636 : return ec_indexes;
3638 : }
3639 :
3640 : /*
3641 : * get_common_eclass_indexes
3642 : * Build and return a Bitmapset containing the indexes into root's
3643 : * eq_classes list for all eclasses that mention rels in both
3644 : * relids1 and relids2.
3645 : */
3646 : static Bitmapset *
3647 603700 : get_common_eclass_indexes(PlannerInfo *root, Relids relids1, Relids relids2)
3648 : {
3649 : Bitmapset *rel1ecs;
3650 : Bitmapset *rel2ecs;
3651 : int relid;
3652 :
3653 603700 : rel1ecs = get_eclass_indexes_for_relids(root, relids1);
3654 :
3655 : /*
3656 : * We can get away with just using the relation's eclass_indexes directly
3657 : * when relids2 is a singleton set.
3658 : */
3659 603700 : if (bms_get_singleton_member(relids2, &relid))
3660 471220 : rel2ecs = root->simple_rel_array[relid]->eclass_indexes;
3661 : else
3662 132480 : rel2ecs = get_eclass_indexes_for_relids(root, relids2);
3663 :
3664 : /* Calculate and return the common EC indexes, recycling the left input. */
3665 603700 : return bms_int_members(rel1ecs, rel2ecs);
3666 : }
3667 :
3668 : /*
3669 : * ec_build_derives_hash
3670 : * Construct the auxiliary hash table for derived clause lookups.
3671 : */
3672 : static void
3673 0 : ec_build_derives_hash(PlannerInfo *root, EquivalenceClass *ec)
3674 : {
3675 : Assert(!ec->ec_derives_hash);
3676 :
3677 : /*
3678 : * Create the hash table.
3679 : *
3680 : * We pass list_length(ec->ec_derives_list) as the initial size.
3681 : * Simplehash will divide this by the fillfactor (typically 0.9) and round
3682 : * up to the next power of two, so this will usually give us at least 64
3683 : * buckets around the threshold. That avoids immediate resizing without
3684 : * hardcoding a specific size.
3685 : */
3686 0 : ec->ec_derives_hash = derives_create(root->planner_cxt,
3687 0 : list_length(ec->ec_derives_list),
3688 : NULL);
3689 :
3690 0 : foreach_node(RestrictInfo, rinfo, ec->ec_derives_list)
3691 0 : ec_add_clause_to_derives_hash(ec, rinfo);
3692 0 : }
3693 :
3694 : /*
3695 : * ec_add_derived_clause
3696 : * Add a clause to the set of derived clauses for the given
3697 : * EquivalenceClass. Always appends to ec_derives_list; also adds
3698 : * to ec_derives_hash if it exists.
3699 : *
3700 : * Also asserts expected invariants of derived clauses.
3701 : */
3702 : static void
3703 113224 : ec_add_derived_clause(EquivalenceClass *ec, RestrictInfo *clause)
3704 : {
3705 : /*
3706 : * Constant, if present, is always placed on the RHS; see
3707 : * generate_base_implied_equalities_const(). LHS is never a constant.
3708 : */
3709 : Assert(!clause->left_em->em_is_const);
3710 :
3711 : /*
3712 : * Clauses containing a constant are never considered redundant, so
3713 : * parent_ec is not set.
3714 : */
3715 : Assert(!clause->parent_ec || !clause->right_em->em_is_const);
3716 :
3717 113224 : ec->ec_derives_list = lappend(ec->ec_derives_list, clause);
3718 113224 : if (ec->ec_derives_hash)
3719 0 : ec_add_clause_to_derives_hash(ec, clause);
3720 113224 : }
3721 :
3722 : /*
3723 : * ec_add_derived_clauses
3724 : * Add a list of clauses to the set of clauses derived from the given
3725 : * EquivalenceClass; adding to the list and hash table if needed.
3726 : *
3727 : * This function is similar to ec_add_derived_clause() but optimized for adding
3728 : * multiple clauses at a time to the ec_derives_list. The assertions from
3729 : * ec_add_derived_clause() are not repeated here, as the input clauses are
3730 : * assumed to have already been validated.
3731 : */
3732 : static void
3733 42 : ec_add_derived_clauses(EquivalenceClass *ec, List *clauses)
3734 : {
3735 42 : ec->ec_derives_list = list_concat(ec->ec_derives_list, clauses);
3736 42 : if (ec->ec_derives_hash)
3737 0 : foreach_node(RestrictInfo, rinfo, clauses)
3738 0 : ec_add_clause_to_derives_hash(ec, rinfo);
3739 42 : }
3740 :
3741 : /*
3742 : * fill_ec_derives_key
3743 : * Compute a canonical key for ec_derives_hash lookup or insertion.
3744 : *
3745 : * Derived clauses are looked up using a pair of EquivalenceMembers and a
3746 : * parent EquivalenceClass. To avoid storing or searching for both EM orderings,
3747 : * we canonicalize the key:
3748 : *
3749 : * - For clauses involving two non-constant EMs, em1 is set to the EM with lower
3750 : * memory address and em2 is set to the other one.
3751 : * - For clauses involving a constant EM, the caller must pass the non-constant
3752 : * EM as leftem and NULL as rightem; we then set em1 = NULL and em2 = leftem.
3753 : */
3754 : static inline void
3755 0 : fill_ec_derives_key(ECDerivesKey *key,
3756 : EquivalenceMember *leftem,
3757 : EquivalenceMember *rightem,
3758 : EquivalenceClass *parent_ec)
3759 : {
3760 : Assert(leftem); /* Always required for lookup or insertion */
3761 :
3762 0 : if (rightem == NULL)
3763 : {
3764 0 : key->em1 = NULL;
3765 0 : key->em2 = leftem;
3766 : }
3767 0 : else if (leftem < rightem)
3768 : {
3769 0 : key->em1 = leftem;
3770 0 : key->em2 = rightem;
3771 : }
3772 : else
3773 : {
3774 0 : key->em1 = rightem;
3775 0 : key->em2 = leftem;
3776 : }
3777 0 : key->parent_ec = parent_ec;
3778 0 : }
3779 :
3780 : /*
3781 : * ec_add_clause_to_derives_hash
3782 : * Add a derived clause to ec_derives_hash in the given EquivalenceClass.
3783 : *
3784 : * Each clause is associated with a canonicalized key. For constant-containing
3785 : * clauses, only the non-constant EM is used for lookup; see comments in
3786 : * fill_ec_derives_key().
3787 : */
3788 : static void
3789 0 : ec_add_clause_to_derives_hash(EquivalenceClass *ec, RestrictInfo *rinfo)
3790 : {
3791 : ECDerivesKey key;
3792 : ECDerivesEntry *entry;
3793 : bool found;
3794 :
3795 : /*
3796 : * Constants are always placed on the RHS; see
3797 : * generate_base_implied_equalities_const().
3798 : */
3799 : Assert(!rinfo->left_em->em_is_const);
3800 :
3801 : /*
3802 : * Clauses containing a constant are never considered redundant, so
3803 : * parent_ec is not set.
3804 : */
3805 : Assert(!rinfo->parent_ec || !rinfo->right_em->em_is_const);
3806 :
3807 : /*
3808 : * See fill_ec_derives_key() for details: we use a canonicalized key to
3809 : * avoid storing both EM orderings. For constant EMs, only the
3810 : * non-constant EM is included in the key.
3811 : */
3812 0 : fill_ec_derives_key(&key,
3813 : rinfo->left_em,
3814 0 : rinfo->right_em->em_is_const ? NULL : rinfo->right_em,
3815 : rinfo->parent_ec);
3816 0 : entry = derives_insert(ec->ec_derives_hash, key, &found);
3817 : Assert(!found);
3818 0 : entry->rinfo = rinfo;
3819 0 : }
3820 :
3821 : /*
3822 : * ec_clear_derived_clauses
3823 : * Reset ec_derives_list and ec_derives_hash.
3824 : *
3825 : * We destroy the hash table explicitly, since it may consume significant
3826 : * space. The list holds the same set of entries and can become equally large
3827 : * when thousands of partitions are involved, so we free it as well -- even
3828 : * though we do not typically free lists.
3829 : */
3830 : void
3831 17482 : ec_clear_derived_clauses(EquivalenceClass *ec)
3832 : {
3833 17482 : list_free(ec->ec_derives_list);
3834 17482 : ec->ec_derives_list = NIL;
3835 :
3836 17482 : if (ec->ec_derives_hash)
3837 : {
3838 0 : derives_destroy(ec->ec_derives_hash);
3839 0 : ec->ec_derives_hash = NULL;
3840 : }
3841 17482 : }
3842 :
3843 : /*
3844 : * ec_search_clause_for_ems
3845 : * Search for an existing RestrictInfo that equates the given pair
3846 : * of EquivalenceMembers, either from ec_sources or ec_derives.
3847 : *
3848 : * Returns a clause with matching operands in either given order or commuted
3849 : * order. We used to require matching operator OIDs, but dropped that since any
3850 : * semantically different operator here would indicate a broken operator family.
3851 : *
3852 : * Returns NULL if no matching clause is found.
3853 : */
3854 : static RestrictInfo *
3855 423874 : ec_search_clause_for_ems(PlannerInfo *root, EquivalenceClass *ec,
3856 : EquivalenceMember *leftem, EquivalenceMember *rightem,
3857 : EquivalenceClass *parent_ec)
3858 : {
3859 : /* Check original source clauses */
3860 1357244 : foreach_node(RestrictInfo, rinfo, ec->ec_sources)
3861 : {
3862 511452 : if (rinfo->left_em == leftem &&
3863 229014 : rinfo->right_em == rightem &&
3864 198512 : rinfo->parent_ec == parent_ec)
3865 978 : return rinfo;
3866 511344 : if (rinfo->left_em == rightem &&
3867 221938 : rinfo->right_em == leftem &&
3868 197386 : rinfo->parent_ec == parent_ec)
3869 870 : return rinfo;
3870 : }
3871 :
3872 : /* Not found in ec_sources; search derived clauses */
3873 422896 : return ec_search_derived_clause_for_ems(root, ec, leftem, rightem,
3874 : parent_ec);
3875 : }
3876 :
3877 : /*
3878 : * ec_search_derived_clause_for_ems
3879 : * Search for an existing derived clause between two EquivalenceMembers.
3880 : *
3881 : * If the number of derived clauses exceeds a threshold, switch to hash table
3882 : * lookup; otherwise, scan ec_derives_list linearly.
3883 : *
3884 : * Clauses involving constants are looked up by passing the non-constant EM
3885 : * as leftem and setting rightem to NULL. In that case, we expect to find a
3886 : * clause with a constant on the RHS.
3887 : *
3888 : * While searching the list, we compare each given EM with both sides of each
3889 : * clause. But for hash table lookups, we construct a canonicalized key and
3890 : * perform a single lookup.
3891 : */
3892 : static RestrictInfo *
3893 422902 : ec_search_derived_clause_for_ems(PlannerInfo *root, EquivalenceClass *ec,
3894 : EquivalenceMember *leftem,
3895 : EquivalenceMember *rightem,
3896 : EquivalenceClass *parent_ec)
3897 : {
3898 : /* Switch to using hash lookup when list grows "too long". */
3899 845804 : if (!ec->ec_derives_hash &&
3900 422902 : list_length(ec->ec_derives_list) >= EC_DERIVES_HASH_THRESHOLD)
3901 0 : ec_build_derives_hash(root, ec);
3902 :
3903 : /* Perform hash table lookup if available */
3904 422902 : if (ec->ec_derives_hash)
3905 : {
3906 : ECDerivesKey key;
3907 : RestrictInfo *rinfo;
3908 : ECDerivesEntry *entry;
3909 :
3910 0 : fill_ec_derives_key(&key, leftem, rightem, parent_ec);
3911 0 : entry = derives_lookup(ec->ec_derives_hash, key);
3912 0 : if (entry)
3913 : {
3914 0 : rinfo = entry->rinfo;
3915 : Assert(rinfo);
3916 : Assert(rightem || rinfo->right_em->em_is_const);
3917 0 : return rinfo;
3918 : }
3919 : }
3920 : else
3921 : {
3922 : /* Fallback to linear search over ec_derives_list */
3923 623066 : foreach_node(RestrictInfo, rinfo, ec->ec_derives_list)
3924 : {
3925 : /* Handle special case: lookup by non-const EM alone */
3926 472450 : if (!rightem &&
3927 6 : rinfo->left_em == leftem)
3928 : {
3929 : Assert(rinfo->right_em->em_is_const);
3930 347594 : return rinfo;
3931 : }
3932 472444 : if (rinfo->left_em == leftem &&
3933 189946 : rinfo->right_em == rightem &&
3934 169410 : rinfo->parent_ec == parent_ec)
3935 169404 : return rinfo;
3936 303040 : if (rinfo->left_em == rightem &&
3937 190086 : rinfo->right_em == leftem &&
3938 178184 : rinfo->parent_ec == parent_ec)
3939 178184 : return rinfo;
3940 : }
3941 : }
3942 :
3943 75308 : return NULL;
3944 : }
|
