Age Owner Branch data TLA 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) 1996-2024, PostgreSQL Global Development Group
10 : : * Portions Copyright (c) 1994, Regents of the University of California
11 : : *
12 : : * IDENTIFICATION
13 : : * src/backend/optimizer/path/equivclass.c
14 : : *
15 : : *-------------------------------------------------------------------------
16 : : */
17 : : #include "postgres.h"
18 : :
19 : : #include <limits.h>
20 : :
21 : : #include "access/stratnum.h"
22 : : #include "catalog/pg_type.h"
23 : : #include "nodes/makefuncs.h"
24 : : #include "nodes/nodeFuncs.h"
25 : : #include "optimizer/appendinfo.h"
26 : : #include "optimizer/clauses.h"
27 : : #include "optimizer/optimizer.h"
28 : : #include "optimizer/pathnode.h"
29 : : #include "optimizer/paths.h"
30 : : #include "optimizer/planmain.h"
31 : : #include "optimizer/restrictinfo.h"
32 : : #include "rewrite/rewriteManip.h"
33 : : #include "utils/lsyscache.h"
34 : :
35 : :
36 : : static EquivalenceMember *add_eq_member(EquivalenceClass *ec,
37 : : Expr *expr, Relids relids,
38 : : JoinDomain *jdomain,
39 : : EquivalenceMember *parent,
40 : : Oid datatype);
41 : : static bool is_exprlist_member(Expr *node, List *exprs);
42 : : static void generate_base_implied_equalities_const(PlannerInfo *root,
43 : : EquivalenceClass *ec);
44 : : static void generate_base_implied_equalities_no_const(PlannerInfo *root,
45 : : EquivalenceClass *ec);
46 : : static void generate_base_implied_equalities_broken(PlannerInfo *root,
47 : : EquivalenceClass *ec);
48 : : static List *generate_join_implied_equalities_normal(PlannerInfo *root,
49 : : EquivalenceClass *ec,
50 : : Relids join_relids,
51 : : Relids outer_relids,
52 : : Relids inner_relids);
53 : : static List *generate_join_implied_equalities_broken(PlannerInfo *root,
54 : : EquivalenceClass *ec,
55 : : Relids nominal_join_relids,
56 : : Relids outer_relids,
57 : : Relids nominal_inner_relids,
58 : : RelOptInfo *inner_rel);
59 : : static Oid select_equality_operator(EquivalenceClass *ec,
60 : : Oid lefttype, Oid righttype);
61 : : static RestrictInfo *create_join_clause(PlannerInfo *root,
62 : : EquivalenceClass *ec, Oid opno,
63 : : EquivalenceMember *leftem,
64 : : EquivalenceMember *rightem,
65 : : EquivalenceClass *parent_ec);
66 : : static bool reconsider_outer_join_clause(PlannerInfo *root,
67 : : OuterJoinClauseInfo *ojcinfo,
68 : : bool outer_on_left);
69 : : static bool reconsider_full_join_clause(PlannerInfo *root,
70 : : OuterJoinClauseInfo *ojcinfo);
71 : : static JoinDomain *find_join_domain(PlannerInfo *root, Relids relids);
72 : : static Bitmapset *get_eclass_indexes_for_relids(PlannerInfo *root,
73 : : Relids relids);
74 : : static Bitmapset *get_common_eclass_indexes(PlannerInfo *root, Relids relids1,
75 : : Relids relids2);
76 : :
77 : :
78 : : /*
79 : : * process_equivalence
80 : : * The given clause has a mergejoinable operator and is not an outer-join
81 : : * qualification, so its two sides can be considered equal
82 : : * anywhere they are both computable; moreover that equality can be
83 : : * extended transitively. Record this knowledge in the EquivalenceClass
84 : : * data structure, if applicable. Returns true if successful, false if not
85 : : * (in which case caller should treat the clause as ordinary, not an
86 : : * equivalence).
87 : : *
88 : : * In some cases, although we cannot convert a clause into EquivalenceClass
89 : : * knowledge, we can still modify it to a more useful form than the original.
90 : : * Then, *p_restrictinfo will be replaced by a new RestrictInfo, which is what
91 : : * the caller should use for further processing.
92 : : *
93 : : * jdomain is the join domain within which the given clause was found.
94 : : * This limits the applicability of deductions from the EquivalenceClass,
95 : : * as described in optimizer/README.
96 : : *
97 : : * We reject proposed equivalence clauses if they contain leaky functions
98 : : * and have security_level above zero. The EC evaluation rules require us to
99 : : * apply certain tests at certain joining levels, and we can't tolerate
100 : : * delaying any test on security_level grounds. By rejecting candidate clauses
101 : : * that might require security delays, we ensure it's safe to apply an EC
102 : : * clause as soon as it's supposed to be applied.
103 : : *
104 : : * On success return, we have also initialized the clause's left_ec/right_ec
105 : : * fields to point to the EquivalenceClass representing it. This saves lookup
106 : : * effort later.
107 : : *
108 : : * Note: constructing merged EquivalenceClasses is a standard UNION-FIND
109 : : * problem, for which there exist better data structures than simple lists.
110 : : * If this code ever proves to be a bottleneck then it could be sped up ---
111 : : * but for now, simple is beautiful.
112 : : *
113 : : * Note: this is only called during planner startup, not during GEQO
114 : : * exploration, so we need not worry about whether we're in the right
115 : : * memory context.
116 : : */
117 : : bool
2380 tgl@sss.pgh.pa.us 118 :CBC 123958 : process_equivalence(PlannerInfo *root,
119 : : RestrictInfo **p_restrictinfo,
120 : : JoinDomain *jdomain)
121 : : {
122 : 123958 : RestrictInfo *restrictinfo = *p_restrictinfo;
6294 123 : 123958 : Expr *clause = restrictinfo->clause;
124 : : Oid opno,
125 : : collation,
126 : : item1_type,
127 : : item2_type;
128 : : Expr *item1;
129 : : Expr *item2;
130 : : Relids item1_relids,
131 : : item2_relids;
132 : : List *opfamilies;
133 : : EquivalenceClass *ec1,
134 : : *ec2;
135 : : EquivalenceMember *em1,
136 : : *em2;
137 : : ListCell *lc1;
138 : : int ec2_idx;
139 : :
140 : : /* Should not already be marked as having generated an eclass */
4916 141 [ - + ]: 123958 : Assert(restrictinfo->left_ec == NULL);
142 [ - + ]: 123958 : Assert(restrictinfo->right_ec == NULL);
143 : :
144 : : /* Reject if it is potentially postponable by security considerations */
2643 145 [ + + + + ]: 123958 : if (restrictinfo->security_level > 0 && !restrictinfo->leakproof)
146 : 101 : return false;
147 : :
148 : : /* Extract info from given clause */
6294 149 [ - + ]: 123857 : Assert(is_opclause(clause));
150 : 123857 : opno = ((OpExpr *) clause)->opno;
4775 151 : 123857 : collation = ((OpExpr *) clause)->inputcollid;
6294 152 : 123857 : item1 = (Expr *) get_leftop(clause);
153 : 123857 : item2 = (Expr *) get_rightop(clause);
154 : 123857 : item1_relids = restrictinfo->left_relids;
155 : 123857 : item2_relids = restrictinfo->right_relids;
156 : :
157 : : /*
158 : : * Ensure both input expressions expose the desired collation (their types
159 : : * should be OK already); see comments for canonicalize_ec_expression.
160 : : */
4775 161 : 123857 : item1 = canonicalize_ec_expression(item1,
162 : : exprType((Node *) item1),
163 : : collation);
164 : 123857 : item2 = canonicalize_ec_expression(item2,
165 : : exprType((Node *) item2),
166 : : collation);
167 : :
168 : : /*
169 : : * Clauses of the form X=X cannot be translated into EquivalenceClasses.
170 : : * We'd either end up with a single-entry EC, losing the knowledge that
171 : : * the clause was present at all, or else make an EC with duplicate
172 : : * entries, causing other issues.
173 : : */
5311 174 [ + + ]: 123857 : if (equal(item1, item2))
175 : : {
176 : : /*
177 : : * If the operator is strict, then the clause can be treated as just
178 : : * "X IS NOT NULL". (Since we know we are considering a top-level
179 : : * qual, we can ignore the difference between FALSE and NULL results.)
180 : : * It's worth making the conversion because we'll typically get a much
181 : : * better selectivity estimate than we would for X=X.
182 : : *
183 : : * If the operator is not strict, we can't be sure what it will do
184 : : * with NULLs, so don't attempt to optimize it.
185 : : */
2380 186 : 27 : set_opfuncid((OpExpr *) clause);
187 [ + - ]: 27 : if (func_strict(((OpExpr *) clause)->opfuncid))
188 : : {
189 : 27 : NullTest *ntest = makeNode(NullTest);
190 : :
191 : 27 : ntest->arg = item1;
192 : 27 : ntest->nulltesttype = IS_NOT_NULL;
193 : 27 : ntest->argisrow = false; /* correct even if composite arg */
194 : 27 : ntest->location = -1;
195 : :
196 : 27 : *p_restrictinfo =
1179 197 : 27 : make_restrictinfo(root,
198 : : (Expr *) ntest,
2380 199 : 27 : restrictinfo->is_pushed_down,
325 200 : 27 : restrictinfo->has_clone,
201 : 27 : restrictinfo->is_clone,
2380 202 : 27 : restrictinfo->pseudoconstant,
203 : : restrictinfo->security_level,
204 : : NULL,
205 : : restrictinfo->incompatible_relids,
206 : : restrictinfo->outer_relids);
207 : : }
208 : 27 : return false;
209 : : }
210 : :
211 : : /*
212 : : * We use the declared input types of the operator, not exprType() of the
213 : : * inputs, as the nominal datatypes for opfamily lookup. This presumes
214 : : * that btree operators are always registered with amoplefttype and
215 : : * amoprighttype equal to their declared input types. We will need this
216 : : * info anyway to build EquivalenceMember nodes, and by extracting it now
217 : : * we can use type comparisons to short-circuit some equal() tests.
218 : : */
6294 219 : 123830 : op_input_types(opno, &item1_type, &item2_type);
220 : :
221 : 123830 : opfamilies = restrictinfo->mergeopfamilies;
222 : :
223 : : /*
224 : : * Sweep through the existing EquivalenceClasses looking for matches to
225 : : * item1 and item2. These are the possible outcomes:
226 : : *
227 : : * 1. We find both in the same EC. The equivalence is already known, so
228 : : * there's nothing to do.
229 : : *
230 : : * 2. We find both in different ECs. Merge the two ECs together.
231 : : *
232 : : * 3. We find just one. Add the other to its EC.
233 : : *
234 : : * 4. We find neither. Make a new, two-entry EC.
235 : : *
236 : : * Note: since all ECs are built through this process or the similar
237 : : * search in get_eclass_for_sort_expr(), it's impossible that we'd match
238 : : * an item in more than one existing nonvolatile EC. So it's okay to stop
239 : : * at the first match.
240 : : */
241 : 123830 : ec1 = ec2 = NULL;
6292 242 : 123830 : em1 = em2 = NULL;
1270 drowley@postgresql.o 243 : 123830 : ec2_idx = -1;
6294 tgl@sss.pgh.pa.us 244 [ + + + + : 196664 : foreach(lc1, root->eq_classes)
+ + ]
245 : : {
246 : 72855 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
247 : : ListCell *lc2;
248 : :
249 : : /* Never match to a volatile EC */
250 [ - + ]: 72855 : if (cur_ec->ec_has_volatile)
6294 tgl@sss.pgh.pa.us 251 :UBC 0 : continue;
252 : :
253 : : /*
254 : : * The collation has to match; check this first since it's cheaper
255 : : * than the opfamily comparison.
256 : : */
4775 tgl@sss.pgh.pa.us 257 [ + + ]:CBC 72855 : if (collation != cur_ec->ec_collation)
258 : 5752 : continue;
259 : :
260 : : /*
261 : : * A "match" requires matching sets of btree opfamilies. Use of
262 : : * equal() for this test has implications discussed in the comments
263 : : * for get_mergejoin_opfamilies().
264 : : */
6294 265 [ + + ]: 67103 : if (!equal(opfamilies, cur_ec->ec_opfamilies))
266 : 16811 : continue;
267 : :
268 [ + - + + : 151873 : foreach(lc2, cur_ec->ec_members)
+ + ]
269 : : {
270 : 101602 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
271 : :
2489 272 [ - + ]: 101602 : Assert(!cur_em->em_is_child); /* no children yet */
273 : :
274 : : /*
275 : : * Match constants only within the same JoinDomain (see
276 : : * optimizer/README).
277 : : */
440 278 [ + + + + ]: 101602 : if (cur_em->em_is_const && cur_em->em_jdomain != jdomain)
6294 279 : 1951 : continue;
280 : :
281 [ + + ]: 99651 : if (!ec1 &&
282 [ + + + + ]: 187788 : item1_type == cur_em->em_datatype &&
283 : 93831 : equal(item1, cur_em->em_expr))
284 : : {
285 : 7641 : ec1 = cur_ec;
6292 286 : 7641 : em1 = cur_em;
6294 287 [ + + ]: 7641 : if (ec2)
288 : 9 : break;
289 : : }
290 : :
291 [ + + ]: 99642 : if (!ec2 &&
292 [ + + + + ]: 196694 : item2_type == cur_em->em_datatype &&
293 : 98251 : equal(item2, cur_em->em_expr))
294 : : {
295 : 2332 : ec2 = cur_ec;
1270 drowley@postgresql.o 296 : 2332 : ec2_idx = foreach_current_index(lc1);
6292 tgl@sss.pgh.pa.us 297 : 2332 : em2 = cur_em;
6294 298 [ + + ]: 2332 : if (ec1)
299 : 12 : break;
300 : : }
301 : : }
302 : :
303 [ + + + + ]: 50292 : if (ec1 && ec2)
304 : 21 : break;
305 : : }
306 : :
307 : : /* Sweep finished, what did we find? */
308 : :
309 [ + + + + ]: 123830 : if (ec1 && ec2)
310 : : {
311 : : /* If case 1, nothing to do, except add to sources */
312 [ + + ]: 21 : if (ec1 == ec2)
313 : : {
314 : 6 : ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
2643 315 : 6 : ec1->ec_min_security = Min(ec1->ec_min_security,
316 : : restrictinfo->security_level);
317 : 6 : ec1->ec_max_security = Max(ec1->ec_max_security,
318 : : restrictinfo->security_level);
319 : : /* mark the RI as associated with this eclass */
4916 320 : 6 : restrictinfo->left_ec = ec1;
321 : 6 : restrictinfo->right_ec = ec1;
322 : : /* mark the RI as usable with this pair of EMs */
6292 323 : 6 : restrictinfo->left_em = em1;
324 : 6 : restrictinfo->right_em = em2;
6294 325 : 6 : return true;
326 : : }
327 : :
328 : : /*
329 : : * Case 2: need to merge ec1 and ec2. This should never happen after
330 : : * the ECs have reached canonical state; otherwise, pathkeys could be
331 : : * rendered non-canonical by the merge, and relation eclass indexes
332 : : * would get broken by removal of an eq_classes list entry.
333 : : */
1729 drowley@postgresql.o 334 [ - + ]: 15 : if (root->ec_merging_done)
4003 tgl@sss.pgh.pa.us 335 [ # # ]:UBC 0 : elog(ERROR, "too late to merge equivalence classes");
336 : :
337 : : /*
338 : : * We add ec2's items to ec1, then set ec2's ec_merged link to point
339 : : * to ec1 and remove ec2 from the eq_classes list. We cannot simply
340 : : * delete ec2 because that could leave dangling pointers in existing
341 : : * PathKeys. We leave it behind with a link so that the merged EC can
342 : : * be found.
343 : : */
6294 tgl@sss.pgh.pa.us 344 :CBC 15 : ec1->ec_members = list_concat(ec1->ec_members, ec2->ec_members);
345 : 15 : ec1->ec_sources = list_concat(ec1->ec_sources, ec2->ec_sources);
6292 346 : 15 : ec1->ec_derives = list_concat(ec1->ec_derives, ec2->ec_derives);
6294 347 : 15 : ec1->ec_relids = bms_join(ec1->ec_relids, ec2->ec_relids);
348 : 15 : ec1->ec_has_const |= ec2->ec_has_const;
349 : : /* can't need to set has_volatile */
2643 350 : 15 : ec1->ec_min_security = Min(ec1->ec_min_security,
351 : : ec2->ec_min_security);
352 : 15 : ec1->ec_max_security = Max(ec1->ec_max_security,
353 : : ec2->ec_max_security);
6294 354 : 15 : ec2->ec_merged = ec1;
1270 drowley@postgresql.o 355 : 15 : root->eq_classes = list_delete_nth_cell(root->eq_classes, ec2_idx);
356 : : /* just to avoid debugging confusion w/ dangling pointers: */
6294 tgl@sss.pgh.pa.us 357 : 15 : ec2->ec_members = NIL;
358 : 15 : ec2->ec_sources = NIL;
6292 359 : 15 : ec2->ec_derives = NIL;
6294 360 : 15 : ec2->ec_relids = NULL;
361 : 15 : ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
2643 362 : 15 : ec1->ec_min_security = Min(ec1->ec_min_security,
363 : : restrictinfo->security_level);
364 : 15 : ec1->ec_max_security = Max(ec1->ec_max_security,
365 : : restrictinfo->security_level);
366 : : /* mark the RI as associated with this eclass */
4916 367 : 15 : restrictinfo->left_ec = ec1;
368 : 15 : restrictinfo->right_ec = ec1;
369 : : /* mark the RI as usable with this pair of EMs */
6292 370 : 15 : restrictinfo->left_em = em1;
371 : 15 : restrictinfo->right_em = em2;
372 : : }
6294 373 [ + + ]: 123809 : else if (ec1)
374 : : {
375 : : /* Case 3: add item2 to ec1 */
440 376 : 7620 : em2 = add_eq_member(ec1, item2, item2_relids,
377 : : jdomain, NULL, item2_type);
6294 378 : 7620 : ec1->ec_sources = lappend(ec1->ec_sources, restrictinfo);
2643 379 : 7620 : ec1->ec_min_security = Min(ec1->ec_min_security,
380 : : restrictinfo->security_level);
381 : 7620 : ec1->ec_max_security = Max(ec1->ec_max_security,
382 : : restrictinfo->security_level);
383 : : /* mark the RI as associated with this eclass */
4916 384 : 7620 : restrictinfo->left_ec = ec1;
385 : 7620 : restrictinfo->right_ec = ec1;
386 : : /* mark the RI as usable with this pair of EMs */
6292 387 : 7620 : restrictinfo->left_em = em1;
388 : 7620 : restrictinfo->right_em = em2;
389 : : }
6294 390 [ + + ]: 116189 : else if (ec2)
391 : : {
392 : : /* Case 3: add item1 to ec2 */
440 393 : 2311 : em1 = add_eq_member(ec2, item1, item1_relids,
394 : : jdomain, NULL, item1_type);
6294 395 : 2311 : ec2->ec_sources = lappend(ec2->ec_sources, restrictinfo);
2643 396 : 2311 : ec2->ec_min_security = Min(ec2->ec_min_security,
397 : : restrictinfo->security_level);
398 : 2311 : ec2->ec_max_security = Max(ec2->ec_max_security,
399 : : restrictinfo->security_level);
400 : : /* mark the RI as associated with this eclass */
4916 401 : 2311 : restrictinfo->left_ec = ec2;
402 : 2311 : restrictinfo->right_ec = ec2;
403 : : /* mark the RI as usable with this pair of EMs */
6292 404 : 2311 : restrictinfo->left_em = em1;
405 : 2311 : restrictinfo->right_em = em2;
406 : : }
407 : : else
408 : : {
409 : : /* Case 4: make a new, two-entry EC */
6294 410 : 113878 : EquivalenceClass *ec = makeNode(EquivalenceClass);
411 : :
412 : 113878 : ec->ec_opfamilies = opfamilies;
4775 413 : 113878 : ec->ec_collation = collation;
6294 414 : 113878 : ec->ec_members = NIL;
415 : 113878 : ec->ec_sources = list_make1(restrictinfo);
6292 416 : 113878 : ec->ec_derives = NIL;
6294 417 : 113878 : ec->ec_relids = NULL;
418 : 113878 : ec->ec_has_const = false;
419 : 113878 : ec->ec_has_volatile = false;
420 : 113878 : ec->ec_broken = false;
6002 421 : 113878 : ec->ec_sortref = 0;
2643 422 : 113878 : ec->ec_min_security = restrictinfo->security_level;
423 : 113878 : ec->ec_max_security = restrictinfo->security_level;
6294 424 : 113878 : ec->ec_merged = NULL;
440 425 : 113878 : em1 = add_eq_member(ec, item1, item1_relids,
426 : : jdomain, NULL, item1_type);
427 : 113878 : em2 = add_eq_member(ec, item2, item2_relids,
428 : : jdomain, NULL, item2_type);
429 : :
6294 430 : 113878 : root->eq_classes = lappend(root->eq_classes, ec);
431 : :
432 : : /* mark the RI as associated with this eclass */
4916 433 : 113878 : restrictinfo->left_ec = ec;
434 : 113878 : restrictinfo->right_ec = ec;
435 : : /* mark the RI as usable with this pair of EMs */
6292 436 : 113878 : restrictinfo->left_em = em1;
437 : 113878 : restrictinfo->right_em = em2;
438 : : }
439 : :
6294 440 : 123824 : return true;
441 : : }
442 : :
443 : : /*
444 : : * canonicalize_ec_expression
445 : : *
446 : : * This function ensures that the expression exposes the expected type and
447 : : * collation, so that it will be equal() to other equivalence-class expressions
448 : : * that it ought to be equal() to.
449 : : *
450 : : * The rule for datatypes is that the exposed type should match what it would
451 : : * be for an input to an operator of the EC's opfamilies; which is usually
452 : : * the declared input type of the operator, but in the case of polymorphic
453 : : * operators no relabeling is wanted (compare the behavior of parse_coerce.c).
454 : : * Expressions coming in from quals will generally have the right type
455 : : * already, but expressions coming from indexkeys may not (because they are
456 : : * represented without any explicit relabel in pg_index), and the same problem
457 : : * occurs for sort expressions (because the parser is likewise cavalier about
458 : : * putting relabels on them). Such cases will be binary-compatible with the
459 : : * real operators, so adding a RelabelType is sufficient.
460 : : *
461 : : * Also, the expression's exposed collation must match the EC's collation.
462 : : * This is important because in comparisons like "foo < bar COLLATE baz",
463 : : * only one of the expressions has the correct exposed collation as we receive
464 : : * it from the parser. Forcing both of them to have it ensures that all
465 : : * variant spellings of such a construct behave the same. Again, we can
466 : : * stick on a RelabelType to force the right exposed collation. (It might
467 : : * work to not label the collation at all in EC members, but this is risky
468 : : * since some parts of the system expect exprCollation() to deliver the
469 : : * right answer for a sort key.)
470 : : */
471 : : Expr *
4775 472 : 1106735 : canonicalize_ec_expression(Expr *expr, Oid req_type, Oid req_collation)
473 : : {
474 : 1106735 : Oid expr_type = exprType((Node *) expr);
475 : :
476 : : /*
477 : : * For a polymorphic-input-type opclass, just keep the same exposed type.
478 : : * RECORD opclasses work like polymorphic-type ones for this purpose.
479 : : */
2160 480 [ + - + + : 1106735 : if (IsPolymorphicType(req_type) || req_type == RECORDOID)
+ - + + +
+ + + + -
+ - + - +
- + - +
+ ]
4775 481 : 1729 : req_type = expr_type;
482 : :
483 : : /*
484 : : * No work if the expression exposes the right type/collation already.
485 : : */
486 [ + + + + ]: 2196446 : if (expr_type != req_type ||
487 : 1089711 : exprCollation((Node *) expr) != req_collation)
488 : : {
489 : : /*
490 : : * If we have to change the type of the expression, set typmod to -1,
491 : : * since the new type may not have the same typmod interpretation.
492 : : * When we only have to change collation, preserve the exposed typmod.
493 : : */
494 : : int32 req_typmod;
495 : :
1334 496 [ + + ]: 17431 : if (expr_type != req_type)
497 : 17024 : req_typmod = -1;
498 : : else
499 : 407 : req_typmod = exprTypmod((Node *) expr);
500 : :
501 : : /*
502 : : * Use applyRelabelType so that we preserve const-flatness. This is
503 : : * important since eval_const_expressions has already been applied.
504 : : */
505 : 17431 : expr = (Expr *) applyRelabelType((Node *) expr,
506 : : req_type, req_typmod, req_collation,
507 : : COERCE_IMPLICIT_CAST, -1, false);
508 : : }
509 : :
4775 510 : 1106735 : return expr;
511 : : }
512 : :
513 : : /*
514 : : * add_eq_member - build a new EquivalenceMember and add it to an EC
515 : : */
516 : : static EquivalenceMember *
5995 bruce@momjian.us 517 : 377760 : add_eq_member(EquivalenceClass *ec, Expr *expr, Relids relids,
518 : : JoinDomain *jdomain, EquivalenceMember *parent, Oid datatype)
519 : : {
6294 tgl@sss.pgh.pa.us 520 : 377760 : EquivalenceMember *em = makeNode(EquivalenceMember);
521 : :
522 : 377760 : em->em_expr = expr;
523 : 377760 : em->em_relids = relids;
524 : 377760 : em->em_is_const = false;
440 525 : 377760 : em->em_is_child = (parent != NULL);
6294 526 : 377760 : em->em_datatype = datatype;
440 527 : 377760 : em->em_jdomain = jdomain;
528 : 377760 : em->em_parent = parent;
529 : :
6294 530 [ + + ]: 377760 : if (bms_is_empty(relids))
531 : : {
532 : : /*
533 : : * No Vars, assume it's a pseudoconstant. This is correct for entries
534 : : * generated from process_equivalence(), because a WHERE clause can't
535 : : * contain aggregates or SRFs, and non-volatility was checked before
536 : : * process_equivalence() ever got called. But
537 : : * get_eclass_for_sort_expr() has to work harder. We put the tests
538 : : * there not here to save cycles in the equivalence case.
539 : : */
440 540 [ - + ]: 94037 : Assert(!parent);
6294 541 : 94037 : em->em_is_const = true;
542 : 94037 : ec->ec_has_const = true;
543 : : /* it can't affect ec_relids */
544 : : }
440 545 [ + + ]: 283723 : else if (!parent) /* child members don't add to ec_relids */
546 : : {
6294 547 : 248449 : ec->ec_relids = bms_add_members(ec->ec_relids, relids);
548 : : }
549 : 377760 : ec->ec_members = lappend(ec->ec_members, em);
550 : :
6292 551 : 377760 : return em;
552 : : }
553 : :
554 : :
555 : : /*
556 : : * get_eclass_for_sort_expr
557 : : * Given an expression and opfamily/collation info, find an existing
558 : : * equivalence class it is a member of; if none, optionally build a new
559 : : * single-member EquivalenceClass for it.
560 : : *
561 : : * sortref is the SortGroupRef of the originating SortGroupClause, if any,
562 : : * or zero if not. (It should never be zero if the expression is volatile!)
563 : : *
564 : : * If rel is not NULL, it identifies a specific relation we're considering
565 : : * a path for, and indicates that child EC members for that relation can be
566 : : * considered. Otherwise child members are ignored. (Note: since child EC
567 : : * members aren't guaranteed unique, a non-NULL value means that there could
568 : : * be more than one EC that matches the expression; if so it's order-dependent
569 : : * which one you get. This is annoying but it only happens in corner cases,
570 : : * so for now we live with just reporting the first match. See also
571 : : * generate_implied_equalities_for_column and match_pathkeys_to_index.)
572 : : *
573 : : * If create_it is true, we'll build a new EquivalenceClass when there is no
574 : : * match. If create_it is false, we just return NULL when no match.
575 : : *
576 : : * This can be used safely both before and after EquivalenceClass merging;
577 : : * since it never causes merging it does not invalidate any existing ECs
578 : : * or PathKeys. However, ECs added after path generation has begun are
579 : : * of limited usefulness, so usually it's best to create them beforehand.
580 : : *
581 : : * Note: opfamilies must be chosen consistently with the way
582 : : * process_equivalence() would do; that is, generated from a mergejoinable
583 : : * equality operator. Else we might fail to detect valid equivalences,
584 : : * generating poor (but not incorrect) plans.
585 : : */
586 : : EquivalenceClass *
6294 587 : 823180 : get_eclass_for_sort_expr(PlannerInfo *root,
588 : : Expr *expr,
589 : : List *opfamilies,
590 : : Oid opcintype,
591 : : Oid collation,
592 : : Index sortref,
593 : : Relids rel,
594 : : bool create_it)
595 : : {
596 : : JoinDomain *jdomain;
597 : : Relids expr_relids;
598 : : EquivalenceClass *newec;
599 : : EquivalenceMember *newem;
600 : : ListCell *lc1;
601 : : MemoryContext oldcontext;
602 : :
603 : : /*
604 : : * Ensure the expression exposes the correct type and collation.
605 : : */
4775 606 : 823180 : expr = canonicalize_ec_expression(expr, opcintype, collation);
607 : :
608 : : /*
609 : : * Since SortGroupClause nodes are top-level expressions (GROUP BY, ORDER
610 : : * BY, etc), they can be presumed to belong to the top JoinDomain.
611 : : */
440 612 : 823180 : jdomain = linitial_node(JoinDomain, root->join_domains);
613 : :
614 : : /*
615 : : * Scan through the existing EquivalenceClasses for a match
616 : : */
6294 617 [ + + + + : 2708619 : foreach(lc1, root->eq_classes)
+ + ]
618 : : {
619 : 2343687 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
620 : : ListCell *lc2;
621 : :
622 : : /*
623 : : * Never match to a volatile EC, except when we are looking at another
624 : : * reference to the same volatile SortGroupClause.
625 : : */
5328 626 [ + + + + ]: 2343687 : if (cur_ec->ec_has_volatile &&
627 [ + + ]: 18 : (sortref == 0 || sortref != cur_ec->ec_sortref))
6002 628 : 262 : continue;
629 : :
4775 630 [ + + ]: 2343425 : if (collation != cur_ec->ec_collation)
631 : 636037 : continue;
6294 632 [ + + ]: 1707388 : if (!equal(opfamilies, cur_ec->ec_opfamilies))
633 : 340821 : continue;
634 : :
635 [ + + + + : 3101295 : foreach(lc2, cur_ec->ec_members)
+ + ]
636 : : {
637 : 2192976 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
638 : :
639 : : /*
640 : : * Ignore child members unless they match the request.
641 : : */
4412 642 [ + + ]: 2192976 : if (cur_em->em_is_child &&
643 [ + + ]: 146168 : !bms_equal(cur_em->em_relids, rel))
644 : 110334 : continue;
645 : :
646 : : /*
647 : : * Match constants only within the same JoinDomain (see
648 : : * optimizer/README).
649 : : */
440 650 [ + + + + ]: 2082642 : if (cur_em->em_is_const && cur_em->em_jdomain != jdomain)
6294 651 : 37637 : continue;
652 : :
4775 653 [ + + + + ]: 4074879 : if (opcintype == cur_em->em_datatype &&
6294 654 : 2029874 : equal(expr, cur_em->em_expr))
655 : : {
656 : : /*
657 : : * Match!
658 : : *
659 : : * Copy the sortref if it wasn't set yet. That may happen if
660 : : * the ec was constructed from a WHERE clause, i.e. it doesn't
661 : : * have a target reference at all.
662 : : */
84 akorotkov@postgresql 663 [ + + + + ]:GNC 458248 : if (cur_ec->ec_sortref == 0 && sortref > 0)
664 : 4119 : cur_ec->ec_sortref = sortref;
665 : 458248 : return cur_ec;
666 : : }
667 : : }
668 : : }
669 : :
670 : : /* No match; does caller want a NULL result? */
4916 tgl@sss.pgh.pa.us 671 [ + + ]:CBC 364932 : if (!create_it)
672 : 260133 : return NULL;
673 : :
674 : : /*
675 : : * OK, build a new single-member EC
676 : : *
677 : : * Here, we must be sure that we construct the EC in the right context.
678 : : */
6294 679 : 104799 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
680 : :
681 : 104799 : newec = makeNode(EquivalenceClass);
682 : 104799 : newec->ec_opfamilies = list_copy(opfamilies);
4775 683 : 104799 : newec->ec_collation = collation;
6294 684 : 104799 : newec->ec_members = NIL;
685 : 104799 : newec->ec_sources = NIL;
6292 686 : 104799 : newec->ec_derives = NIL;
6294 687 : 104799 : newec->ec_relids = NULL;
688 : 104799 : newec->ec_has_const = false;
689 : 104799 : newec->ec_has_volatile = contain_volatile_functions((Node *) expr);
690 : 104799 : newec->ec_broken = false;
6002 691 : 104799 : newec->ec_sortref = sortref;
2643 692 : 104799 : newec->ec_min_security = UINT_MAX;
693 : 104799 : newec->ec_max_security = 0;
6294 694 : 104799 : newec->ec_merged = NULL;
695 : :
5161 bruce@momjian.us 696 [ + + - + ]: 104799 : if (newec->ec_has_volatile && sortref == 0) /* should not happen */
5328 tgl@sss.pgh.pa.us 697 [ # # ]:UBC 0 : elog(ERROR, "volatile EquivalenceClass has no sortref");
698 : :
699 : : /*
700 : : * Get the precise set of relids appearing in the expression.
701 : : */
1179 tgl@sss.pgh.pa.us 702 :CBC 104799 : expr_relids = pull_varnos(root, (Node *) expr);
703 : :
3803 704 : 104799 : newem = add_eq_member(newec, copyObject(expr), expr_relids,
705 : : jdomain, NULL, opcintype);
706 : :
707 : : /*
708 : : * add_eq_member doesn't check for volatile functions, set-returning
709 : : * functions, aggregates, or window functions, but such could appear in
710 : : * sort expressions; so we have to check whether its const-marking was
711 : : * correct.
712 : : */
6294 713 [ + + ]: 104799 : if (newec->ec_has_const)
714 : : {
6126 715 [ + + + + ]: 8613 : if (newec->ec_has_volatile ||
716 [ + + ]: 8477 : expression_returns_set((Node *) expr) ||
5586 717 [ + + ]: 8314 : contain_agg_clause((Node *) expr) ||
718 : 4113 : contain_window_function((Node *) expr))
719 : : {
6294 720 : 227 : newec->ec_has_const = false;
6292 721 : 227 : newem->em_is_const = false;
722 : : }
723 : : }
724 : :
6294 725 : 104799 : root->eq_classes = lappend(root->eq_classes, newec);
726 : :
727 : : /*
728 : : * If EC merging is already complete, we have to mop up by adding the new
729 : : * EC to the eclass_indexes of the relation(s) mentioned in it.
730 : : */
1729 drowley@postgresql.o 731 [ + + ]: 104799 : if (root->ec_merging_done)
732 : : {
733 : 58010 : int ec_index = list_length(root->eq_classes) - 1;
734 : 58010 : int i = -1;
735 : :
736 [ + + ]: 111562 : while ((i = bms_next_member(newec->ec_relids, i)) > 0)
737 : : {
738 : 53552 : RelOptInfo *rel = root->simple_rel_array[i];
739 : :
440 tgl@sss.pgh.pa.us 740 [ + + ]: 53552 : if (rel == NULL) /* must be an outer join */
741 : : {
742 [ - + ]: 2869 : Assert(bms_is_member(i, root->outer_join_rels));
743 : 2869 : continue;
744 : : }
745 : :
426 746 [ - + ]: 50683 : Assert(rel->reloptkind == RELOPT_BASEREL);
747 : :
1729 drowley@postgresql.o 748 : 50683 : rel->eclass_indexes = bms_add_member(rel->eclass_indexes,
749 : : ec_index);
750 : : }
751 : : }
752 : :
6294 tgl@sss.pgh.pa.us 753 : 104799 : MemoryContextSwitchTo(oldcontext);
754 : :
755 : 104799 : return newec;
756 : : }
757 : :
758 : : /*
759 : : * find_ec_member_matching_expr
760 : : * Locate an EquivalenceClass member matching the given expr, if any;
761 : : * return NULL if no match.
762 : : *
763 : : * "Matching" is defined as "equal after stripping RelabelTypes".
764 : : * This is used for identifying sort expressions, and we need to allow
765 : : * binary-compatible relabeling for some cases involving binary-compatible
766 : : * sort operators.
767 : : *
768 : : * Child EC members are ignored unless they belong to given 'relids'.
769 : : */
770 : : EquivalenceMember *
1090 771 : 131281 : find_ec_member_matching_expr(EquivalenceClass *ec,
772 : : Expr *expr,
773 : : Relids relids)
774 : : {
775 : : ListCell *lc;
776 : :
777 : : /* We ignore binary-compatible relabeling on both ends */
778 [ + - + + ]: 140947 : while (expr && IsA(expr, RelabelType))
779 : 9666 : expr = ((RelabelType *) expr)->arg;
780 : :
781 [ + - + + : 284861 : foreach(lc, ec->ec_members)
+ + ]
782 : : {
783 : 211037 : EquivalenceMember *em = (EquivalenceMember *) lfirst(lc);
784 : : Expr *emexpr;
785 : :
786 : : /*
787 : : * We shouldn't be trying to sort by an equivalence class that
788 : : * contains a constant, so no need to consider such cases any further.
789 : : */
790 [ - + ]: 211037 : if (em->em_is_const)
1090 tgl@sss.pgh.pa.us 791 :UBC 0 : continue;
792 : :
793 : : /*
794 : : * Ignore child members unless they belong to the requested rel.
795 : : */
1090 tgl@sss.pgh.pa.us 796 [ + + ]:CBC 211037 : if (em->em_is_child &&
797 [ + + ]: 71258 : !bms_is_subset(em->em_relids, relids))
798 : 68409 : continue;
799 : :
800 : : /*
801 : : * Match if same expression (after stripping relabel).
802 : : */
803 : 142628 : emexpr = em->em_expr;
804 [ + - + + ]: 145011 : while (emexpr && IsA(emexpr, RelabelType))
805 : 2383 : emexpr = ((RelabelType *) emexpr)->arg;
806 : :
807 [ + + ]: 142628 : if (equal(emexpr, expr))
808 : 57457 : return em;
809 : : }
810 : :
811 : 73824 : return NULL;
812 : : }
813 : :
814 : : /*
815 : : * find_computable_ec_member
816 : : * Locate an EquivalenceClass member that can be computed from the
817 : : * expressions appearing in "exprs"; return NULL if no match.
818 : : *
819 : : * "exprs" can be either a list of bare expression trees, or a list of
820 : : * TargetEntry nodes. Either way, it should contain Vars and possibly
821 : : * Aggrefs and WindowFuncs, which are matched to the corresponding elements
822 : : * of the EquivalenceClass's expressions.
823 : : *
824 : : * Unlike find_ec_member_matching_expr, there's no special provision here
825 : : * for binary-compatible relabeling. This is intentional: if we have to
826 : : * compute an expression in this way, setrefs.c is going to insist on exact
827 : : * matches of Vars to the source tlist.
828 : : *
829 : : * Child EC members are ignored unless they belong to given 'relids'.
830 : : * Also, non-parallel-safe expressions are ignored if 'require_parallel_safe'.
831 : : *
832 : : * Note: some callers pass root == NULL for notational reasons. This is OK
833 : : * when require_parallel_safe is false.
834 : : */
835 : : EquivalenceMember *
836 : 1203 : find_computable_ec_member(PlannerInfo *root,
837 : : EquivalenceClass *ec,
838 : : List *exprs,
839 : : Relids relids,
840 : : bool require_parallel_safe)
841 : : {
842 : : ListCell *lc;
843 : :
844 [ + - + + : 3769 : foreach(lc, ec->ec_members)
+ + ]
845 : : {
846 : 2785 : EquivalenceMember *em = (EquivalenceMember *) lfirst(lc);
847 : : List *exprvars;
848 : : ListCell *lc2;
849 : :
850 : : /*
851 : : * We shouldn't be trying to sort by an equivalence class that
852 : : * contains a constant, so no need to consider such cases any further.
853 : : */
854 [ - + ]: 2785 : if (em->em_is_const)
1090 tgl@sss.pgh.pa.us 855 :UBC 0 : continue;
856 : :
857 : : /*
858 : : * Ignore child members unless they belong to the requested rel.
859 : : */
1090 tgl@sss.pgh.pa.us 860 [ + + ]:CBC 2785 : if (em->em_is_child &&
861 [ + + ]: 1467 : !bms_is_subset(em->em_relids, relids))
862 : 1383 : continue;
863 : :
864 : : /*
865 : : * Match if all Vars and quasi-Vars are available in "exprs".
866 : : */
867 : 1402 : exprvars = pull_var_clause((Node *) em->em_expr,
868 : : PVC_INCLUDE_AGGREGATES |
869 : : PVC_INCLUDE_WINDOWFUNCS |
870 : : PVC_INCLUDE_PLACEHOLDERS);
871 [ + + + + : 1697 : foreach(lc2, exprvars)
+ + ]
872 : : {
873 [ + + ]: 1466 : if (!is_exprlist_member(lfirst(lc2), exprs))
874 : 1171 : break;
875 : : }
876 : 1402 : list_free(exprvars);
877 [ + + ]: 1402 : if (lc2)
878 : 1171 : continue; /* we hit a non-available Var */
879 : :
880 : : /*
881 : : * If requested, reject expressions that are not parallel-safe. We
882 : : * check this last because it's a rather expensive test.
883 : : */
884 [ + + ]: 231 : if (require_parallel_safe &&
885 [ + + ]: 61 : !is_parallel_safe(root, (Node *) em->em_expr))
886 : 12 : continue;
887 : :
888 : 219 : return em; /* found usable expression */
889 : : }
890 : :
891 : 984 : return NULL;
892 : : }
893 : :
894 : : /*
895 : : * is_exprlist_member
896 : : * Subroutine for find_computable_ec_member: is "node" in "exprs"?
897 : : *
898 : : * Per the requirements of that function, "exprs" might or might not have
899 : : * TargetEntry superstructure.
900 : : */
901 : : static bool
902 : 1466 : is_exprlist_member(Expr *node, List *exprs)
903 : : {
904 : : ListCell *lc;
905 : :
906 [ + + + + : 4092 : foreach(lc, exprs)
+ + ]
907 : : {
908 : 2921 : Expr *expr = (Expr *) lfirst(lc);
909 : :
910 [ + - + + ]: 2921 : if (expr && IsA(expr, TargetEntry))
911 : 586 : expr = ((TargetEntry *) expr)->expr;
912 : :
913 [ + + ]: 2921 : if (equal(node, expr))
914 : 295 : return true;
915 : : }
916 : 1171 : return false;
917 : : }
918 : :
919 : : /*
920 : : * relation_can_be_sorted_early
921 : : * Can this relation be sorted on this EC before the final output step?
922 : : *
923 : : * To succeed, we must find an EC member that prepare_sort_from_pathkeys knows
924 : : * how to sort on, given the rel's reltarget as input. There are also a few
925 : : * additional constraints based on the fact that the desired sort will be done
926 : : * "early", within the scan/join part of the plan. Also, non-parallel-safe
927 : : * expressions are ignored if 'require_parallel_safe'.
928 : : *
929 : : * At some point we might want to return the identified EquivalenceMember,
930 : : * but for now, callers only want to know if there is one.
931 : : */
932 : : bool
933 : 5011 : relation_can_be_sorted_early(PlannerInfo *root, RelOptInfo *rel,
934 : : EquivalenceClass *ec, bool require_parallel_safe)
935 : : {
936 : 5011 : PathTarget *target = rel->reltarget;
937 : : EquivalenceMember *em;
938 : : ListCell *lc;
939 : :
940 : : /*
941 : : * Reject volatile ECs immediately; such sorts must always be postponed.
942 : : */
943 [ + + ]: 5011 : if (ec->ec_has_volatile)
944 : 36 : return false;
945 : :
946 : : /*
947 : : * Try to find an EM directly matching some reltarget member.
948 : : */
949 [ + + + + : 10176 : foreach(lc, target->exprs)
+ + ]
950 : : {
951 : 9143 : Expr *targetexpr = (Expr *) lfirst(lc);
952 : :
953 : 9143 : em = find_ec_member_matching_expr(ec, targetexpr, rel->relids);
954 [ + + ]: 9143 : if (!em)
1258 tomas.vondra@postgre 955 : 5201 : continue;
956 : :
957 : : /*
958 : : * Reject expressions involving set-returning functions, as those
959 : : * can't be computed early either. (Note: this test and the following
960 : : * one are effectively checking properties of targetexpr, so there's
961 : : * no point in asking whether some other EC member would be better.)
962 : : */
620 tgl@sss.pgh.pa.us 963 [ - + ]: 3942 : if (expression_returns_set((Node *) em->em_expr))
1210 tomas.vondra@postgre 964 :UBC 0 : continue;
965 : :
966 : : /*
967 : : * If requested, reject expressions that are not parallel-safe. We
968 : : * check this last because it's a rather expensive test.
969 : : */
1090 tgl@sss.pgh.pa.us 970 [ + - ]:CBC 3942 : if (require_parallel_safe &&
971 [ - + ]: 3942 : !is_parallel_safe(root, (Node *) em->em_expr))
1210 tomas.vondra@postgre 972 :UBC 0 : continue;
973 : :
1090 tgl@sss.pgh.pa.us 974 :CBC 3942 : return true;
975 : : }
976 : :
977 : : /*
978 : : * Try to find an expression computable from the reltarget.
979 : : */
980 : 1033 : em = find_computable_ec_member(root, ec, target->exprs, rel->relids,
981 : : require_parallel_safe);
982 [ + + ]: 1033 : if (!em)
983 : 984 : return false;
984 : :
985 : : /*
986 : : * Reject expressions involving set-returning functions, as those can't be
987 : : * computed early either. (There's no point in looking for another EC
988 : : * member in this case; since SRFs can't appear in WHERE, they cannot
989 : : * belong to multi-member ECs.)
990 : : */
620 991 [ + + ]: 49 : if (expression_returns_set((Node *) em->em_expr))
1090 992 : 6 : return false;
993 : :
994 : 43 : return true;
995 : : }
996 : :
997 : : /*
998 : : * generate_base_implied_equalities
999 : : * Generate any restriction clauses that we can deduce from equivalence
1000 : : * classes.
1001 : : *
1002 : : * When an EC contains pseudoconstants, our strategy is to generate
1003 : : * "member = const1" clauses where const1 is the first constant member, for
1004 : : * every other member (including other constants). If we are able to do this
1005 : : * then we don't need any "var = var" comparisons because we've successfully
1006 : : * constrained all the vars at their points of creation. If we fail to
1007 : : * generate any of these clauses due to lack of cross-type operators, we fall
1008 : : * back to the "ec_broken" strategy described below. (XXX if there are
1009 : : * multiple constants of different types, it's possible that we might succeed
1010 : : * in forming all the required clauses if we started from a different const
1011 : : * member; but this seems a sufficiently hokey corner case to not be worth
1012 : : * spending lots of cycles on.)
1013 : : *
1014 : : * For ECs that contain no pseudoconstants, we generate derived clauses
1015 : : * "member1 = member2" for each pair of members belonging to the same base
1016 : : * relation (actually, if there are more than two for the same base relation,
1017 : : * we only need enough clauses to link each to each other). This provides
1018 : : * the base case for the recursion: each row emitted by a base relation scan
1019 : : * will constrain all computable members of the EC to be equal. As each
1020 : : * join path is formed, we'll add additional derived clauses on-the-fly
1021 : : * to maintain this invariant (see generate_join_implied_equalities).
1022 : : *
1023 : : * If the opfamilies used by the EC do not provide complete sets of cross-type
1024 : : * equality operators, it is possible that we will fail to generate a clause
1025 : : * that must be generated to maintain the invariant. (An example: given
1026 : : * "WHERE a.x = b.y AND b.y = a.z", the scheme breaks down if we cannot
1027 : : * generate "a.x = a.z" as a restriction clause for A.) In this case we mark
1028 : : * the EC "ec_broken" and fall back to regurgitating its original source
1029 : : * RestrictInfos at appropriate times. We do not try to retract any derived
1030 : : * clauses already generated from the broken EC, so the resulting plan could
1031 : : * be poor due to bad selectivity estimates caused by redundant clauses. But
1032 : : * the correct solution to that is to fix the opfamilies ...
1033 : : *
1034 : : * Equality clauses derived by this function are passed off to
1035 : : * process_implied_equality (in plan/initsplan.c) to be inserted into the
1036 : : * restrictinfo datastructures. Note that this must be called after initial
1037 : : * scanning of the quals and before Path construction begins.
1038 : : *
1039 : : * We make no attempt to avoid generating duplicate RestrictInfos here: we
1040 : : * don't search ec_sources or ec_derives for matches. It doesn't really
1041 : : * seem worth the trouble to do so.
1042 : : */
1043 : : void
6294 1044 : 139537 : generate_base_implied_equalities(PlannerInfo *root)
1045 : : {
1046 : : int ec_index;
1047 : : ListCell *lc;
1048 : :
1049 : : /*
1050 : : * At this point, we're done absorbing knowledge of equivalences in the
1051 : : * query, so no further EC merging should happen, and ECs remaining in the
1052 : : * eq_classes list can be considered canonical. (But note that it's still
1053 : : * possible for new single-member ECs to be added through
1054 : : * get_eclass_for_sort_expr().)
1055 : : */
1729 drowley@postgresql.o 1056 : 139537 : root->ec_merging_done = true;
1057 : :
1058 : 139537 : ec_index = 0;
6294 tgl@sss.pgh.pa.us 1059 [ + + + + : 300189 : foreach(lc, root->eq_classes)
+ + ]
1060 : : {
1061 : 160652 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
1729 drowley@postgresql.o 1062 : 160652 : bool can_generate_joinclause = false;
1063 : : int i;
1064 : :
5995 bruce@momjian.us 1065 [ - + ]: 160652 : Assert(ec->ec_merged == NULL); /* else shouldn't be in list */
1066 [ - + ]: 160652 : Assert(!ec->ec_broken); /* not yet anyway... */
1067 : :
1068 : : /*
1069 : : * Generate implied equalities that are restriction clauses.
1070 : : * Single-member ECs won't generate any deductions, either here or at
1071 : : * the join level.
1072 : : */
1729 drowley@postgresql.o 1073 [ + + ]: 160652 : if (list_length(ec->ec_members) > 1)
1074 : : {
1075 [ + + ]: 114673 : if (ec->ec_has_const)
1076 : 89631 : generate_base_implied_equalities_const(root, ec);
1077 : : else
1078 : 25042 : generate_base_implied_equalities_no_const(root, ec);
1079 : :
1080 : : /* Recover if we failed to generate required derived clauses */
1081 [ + + ]: 114673 : if (ec->ec_broken)
1082 : 15 : generate_base_implied_equalities_broken(root, ec);
1083 : :
1084 : : /* Detect whether this EC might generate join clauses */
1085 : 114673 : can_generate_joinclause =
1086 : 114673 : (bms_membership(ec->ec_relids) == BMS_MULTIPLE);
1087 : : }
1088 : :
1089 : : /*
1090 : : * Mark the base rels cited in each eclass (which should all exist by
1091 : : * now) with the eq_classes indexes of all eclasses mentioning them.
1092 : : * This will let us avoid searching in subsequent lookups. While
1093 : : * we're at it, we can mark base rels that have pending eclass joins;
1094 : : * this is a cheap version of has_relevant_eclass_joinclause().
1095 : : */
1096 : 160652 : i = -1;
1097 [ + + ]: 356850 : while ((i = bms_next_member(ec->ec_relids, i)) > 0)
1098 : : {
1099 : 196198 : RelOptInfo *rel = root->simple_rel_array[i];
1100 : :
440 tgl@sss.pgh.pa.us 1101 [ + + ]: 196198 : if (rel == NULL) /* must be an outer join */
1102 : : {
1103 [ - + ]: 2555 : Assert(bms_is_member(i, root->outer_join_rels));
1104 : 2555 : continue;
1105 : : }
1106 : :
1729 drowley@postgresql.o 1107 [ - + ]: 193643 : Assert(rel->reloptkind == RELOPT_BASEREL);
1108 : :
1109 : 193643 : rel->eclass_indexes = bms_add_member(rel->eclass_indexes,
1110 : : ec_index);
1111 : :
1112 [ + + ]: 193643 : if (can_generate_joinclause)
1113 : 65904 : rel->has_eclass_joins = true;
1114 : : }
1115 : :
1116 : 160652 : ec_index++;
1117 : : }
6294 tgl@sss.pgh.pa.us 1118 : 139537 : }
1119 : :
1120 : : /*
1121 : : * generate_base_implied_equalities when EC contains pseudoconstant(s)
1122 : : */
1123 : : static void
1124 : 89631 : generate_base_implied_equalities_const(PlannerInfo *root,
1125 : : EquivalenceClass *ec)
1126 : : {
1127 : 89631 : EquivalenceMember *const_em = NULL;
1128 : : ListCell *lc;
1129 : :
1130 : : /*
1131 : : * In the trivial case where we just had one "var = const" clause, push
1132 : : * the original clause back into the main planner machinery. There is
1133 : : * nothing to be gained by doing it differently, and we save the effort to
1134 : : * re-build and re-analyze an equality clause that will be exactly
1135 : : * equivalent to the old one.
1136 : : */
5987 1137 [ + + + - ]: 170945 : if (list_length(ec->ec_members) == 2 &&
1138 : 81314 : list_length(ec->ec_sources) == 1)
1139 : : {
1140 : 81314 : RestrictInfo *restrictinfo = (RestrictInfo *) linitial(ec->ec_sources);
1141 : :
440 1142 : 81314 : distribute_restrictinfo_to_rels(root, restrictinfo);
1143 : 81314 : return;
1144 : : }
1145 : :
1146 : : /*
1147 : : * Find the constant member to use. We prefer an actual constant to
1148 : : * pseudo-constants (such as Params), because the constraint exclusion
1149 : : * machinery might be able to exclude relations on the basis of generated
1150 : : * "var = const" equalities, but "var = param" won't work for that.
1151 : : */
6294 1152 [ + - + + : 18645 : foreach(lc, ec->ec_members)
+ + ]
1153 : : {
1154 : 18620 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1155 : :
1156 [ + + ]: 18620 : if (cur_em->em_is_const)
1157 : : {
1158 : 8320 : const_em = cur_em;
4188 1159 [ + + ]: 8320 : if (IsA(cur_em->em_expr, Const))
1160 : 8292 : break;
1161 : : }
1162 : : }
6294 1163 [ - + ]: 8317 : Assert(const_em != NULL);
1164 : :
1165 : : /* Generate a derived equality against each other member */
1166 [ + - + + : 33328 : foreach(lc, ec->ec_members)
+ + ]
1167 : : {
1168 : 25026 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1169 : : Oid eq_op;
1170 : : RestrictInfo *rinfo;
1171 : :
5995 bruce@momjian.us 1172 [ - + ]: 25026 : Assert(!cur_em->em_is_child); /* no children yet */
6294 tgl@sss.pgh.pa.us 1173 [ + + ]: 25026 : if (cur_em == const_em)
1174 : 8305 : continue;
1175 : 16721 : eq_op = select_equality_operator(ec,
1176 : : cur_em->em_datatype,
1177 : : const_em->em_datatype);
1178 [ + + ]: 16721 : if (!OidIsValid(eq_op))
1179 : : {
1180 : : /* failed... */
1181 : 15 : ec->ec_broken = true;
1182 : 15 : break;
1183 : : }
1184 : :
1185 : : /*
1186 : : * We use the constant's em_jdomain as qualscope, so that if the
1187 : : * generated clause is variable-free (i.e, both EMs are consts) it
1188 : : * will be enforced at the join domain level.
1189 : : */
1264 1190 : 16706 : rinfo = process_implied_equality(root, eq_op, ec->ec_collation,
1191 : : cur_em->em_expr, const_em->em_expr,
440 1192 : 16706 : const_em->em_jdomain->jd_relids,
1193 : : ec->ec_min_security,
1264 1194 : 16706 : cur_em->em_is_const);
1195 : :
1196 : : /*
1197 : : * If the clause didn't degenerate to a constant, fill in the correct
1198 : : * markings for a mergejoinable clause, and save it in ec_derives. (We
1199 : : * will not re-use such clauses directly, but selectivity estimation
1200 : : * may consult the list later. Note that this use of ec_derives does
1201 : : * not overlap with its use for join clauses, since we never generate
1202 : : * join clauses from an ec_has_const eclass.)
1203 : : */
1204 [ + - + + ]: 16706 : if (rinfo && rinfo->mergeopfamilies)
1205 : : {
1206 : : /* it's not redundant, so don't set parent_ec */
1207 : 16640 : rinfo->left_ec = rinfo->right_ec = ec;
1208 : 16640 : rinfo->left_em = cur_em;
1209 : 16640 : rinfo->right_em = const_em;
1210 : 16640 : ec->ec_derives = lappend(ec->ec_derives, rinfo);
1211 : : }
1212 : : }
1213 : : }
1214 : :
1215 : : /*
1216 : : * generate_base_implied_equalities when EC contains no pseudoconstants
1217 : : */
1218 : : static void
6294 1219 : 25042 : generate_base_implied_equalities_no_const(PlannerInfo *root,
1220 : : EquivalenceClass *ec)
1221 : : {
1222 : : EquivalenceMember **prev_ems;
1223 : : ListCell *lc;
1224 : :
1225 : : /*
1226 : : * We scan the EC members once and track the last-seen member for each
1227 : : * base relation. When we see another member of the same base relation,
1228 : : * we generate "prev_em = cur_em". This results in the minimum number of
1229 : : * derived clauses, but it's possible that it will fail when a different
1230 : : * ordering would succeed. XXX FIXME: use a UNION-FIND algorithm similar
1231 : : * to the way we build merged ECs. (Use a list-of-lists for each rel.)
1232 : : */
1233 : : prev_ems = (EquivalenceMember **)
1234 : 25042 : palloc0(root->simple_rel_array_size * sizeof(EquivalenceMember *));
1235 : :
1236 [ + - + + : 75864 : foreach(lc, ec->ec_members)
+ + ]
1237 : : {
1238 : 50822 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
1239 : : int relid;
1240 : :
5995 bruce@momjian.us 1241 [ - + ]: 50822 : Assert(!cur_em->em_is_child); /* no children yet */
3425 tgl@sss.pgh.pa.us 1242 [ + + ]: 50822 : if (!bms_get_singleton_member(cur_em->em_relids, &relid))
6294 1243 : 93 : continue;
1244 [ - + ]: 50729 : Assert(relid < root->simple_rel_array_size);
1245 : :
1246 [ + + ]: 50729 : if (prev_ems[relid] != NULL)
1247 : : {
1248 : 155 : EquivalenceMember *prev_em = prev_ems[relid];
1249 : : Oid eq_op;
1250 : : RestrictInfo *rinfo;
1251 : :
1252 : 155 : eq_op = select_equality_operator(ec,
1253 : : prev_em->em_datatype,
1254 : : cur_em->em_datatype);
1255 [ - + ]: 155 : if (!OidIsValid(eq_op))
1256 : : {
1257 : : /* failed... */
6294 tgl@sss.pgh.pa.us 1258 :UBC 0 : ec->ec_broken = true;
1259 : 0 : break;
1260 : : }
1261 : :
1262 : : /*
1263 : : * The expressions aren't constants, so the passed qualscope will
1264 : : * never be used to place the generated clause. We just need to
1265 : : * be sure it covers both expressions, which em_relids should do.
1266 : : */
1264 tgl@sss.pgh.pa.us 1267 :CBC 155 : rinfo = process_implied_equality(root, eq_op, ec->ec_collation,
1268 : : prev_em->em_expr, cur_em->em_expr,
1269 : : cur_em->em_relids,
1270 : : ec->ec_min_security,
1271 : : false);
1272 : :
1273 : : /*
1274 : : * If the clause didn't degenerate to a constant, fill in the
1275 : : * correct markings for a mergejoinable clause. We don't put it
1276 : : * in ec_derives however; we don't currently need to re-find such
1277 : : * clauses, and we don't want to clutter that list with non-join
1278 : : * clauses.
1279 : : */
1280 [ + - + - ]: 155 : if (rinfo && rinfo->mergeopfamilies)
1281 : : {
1282 : : /* it's not redundant, so don't set parent_ec */
1283 : 155 : rinfo->left_ec = rinfo->right_ec = ec;
1284 : 155 : rinfo->left_em = prev_em;
1285 : 155 : rinfo->right_em = cur_em;
1286 : : }
1287 : : }
6294 1288 : 50729 : prev_ems[relid] = cur_em;
1289 : : }
1290 : :
1291 : 25042 : pfree(prev_ems);
1292 : :
1293 : : /*
1294 : : * We also have to make sure that all the Vars used in the member clauses
1295 : : * will be available at any join node we might try to reference them at.
1296 : : * For the moment we force all the Vars to be available at all join nodes
1297 : : * for this eclass. Perhaps this could be improved by doing some
1298 : : * pre-analysis of which members we prefer to join, but it's no worse than
1299 : : * what happened in the pre-8.3 code.
1300 : : */
1301 [ + - + + : 75864 : foreach(lc, ec->ec_members)
+ + ]
1302 : : {
1303 : 50822 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
5474 1304 : 50822 : List *vars = pull_var_clause((Node *) cur_em->em_expr,
1305 : : PVC_RECURSE_AGGREGATES |
1306 : : PVC_RECURSE_WINDOWFUNCS |
1307 : : PVC_INCLUDE_PLACEHOLDERS);
1308 : :
606 1309 : 50822 : add_vars_to_targetlist(root, vars, ec->ec_relids);
6294 1310 : 50822 : list_free(vars);
1311 : : }
1312 : 25042 : }
1313 : :
1314 : : /*
1315 : : * generate_base_implied_equalities cleanup after failure
1316 : : *
1317 : : * What we must do here is push any zero- or one-relation source RestrictInfos
1318 : : * of the EC back into the main restrictinfo datastructures. Multi-relation
1319 : : * clauses will be regurgitated later by generate_join_implied_equalities().
1320 : : * (We do it this way to maintain continuity with the case that ec_broken
1321 : : * becomes set only after we've gone up a join level or two.) However, for
1322 : : * an EC that contains constants, we can adopt a simpler strategy and just
1323 : : * throw back all the source RestrictInfos immediately; that works because
1324 : : * we know that such an EC can't become broken later. (This rule justifies
1325 : : * ignoring ec_has_const ECs in generate_join_implied_equalities, even when
1326 : : * they are broken.)
1327 : : */
1328 : : static void
1329 : 15 : generate_base_implied_equalities_broken(PlannerInfo *root,
1330 : : EquivalenceClass *ec)
1331 : : {
1332 : : ListCell *lc;
1333 : :
1334 [ + - + + : 48 : foreach(lc, ec->ec_sources)
+ + ]
1335 : : {
1336 : 33 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
1337 : :
4378 1338 [ - + - - ]: 33 : if (ec->ec_has_const ||
4378 tgl@sss.pgh.pa.us 1339 :UBC 0 : bms_membership(restrictinfo->required_relids) != BMS_MULTIPLE)
6294 tgl@sss.pgh.pa.us 1340 :CBC 33 : distribute_restrictinfo_to_rels(root, restrictinfo);
1341 : : }
1342 : 15 : }
1343 : :
1344 : :
1345 : : /*
1346 : : * generate_join_implied_equalities
1347 : : * Generate any join clauses that we can deduce from equivalence classes.
1348 : : *
1349 : : * At a join node, we must enforce restriction clauses sufficient to ensure
1350 : : * that all equivalence-class members computable at that node are equal.
1351 : : * Since the set of clauses to enforce can vary depending on which subset
1352 : : * relations are the inputs, we have to compute this afresh for each join
1353 : : * relation pair. Hence a fresh List of RestrictInfo nodes is built and
1354 : : * passed back on each call.
1355 : : *
1356 : : * In addition to its use at join nodes, this can be applied to generate
1357 : : * eclass-based join clauses for use in a parameterized scan of a base rel.
1358 : : * The reason for the asymmetry of specifying the inner rel as a RelOptInfo
1359 : : * and the outer rel by Relids is that this usage occurs before we have
1360 : : * built any join RelOptInfos.
1361 : : *
1362 : : * An annoying special case for parameterized scans is that the inner rel can
1363 : : * be an appendrel child (an "other rel"). In this case we must generate
1364 : : * appropriate clauses using child EC members. add_child_rel_equivalences
1365 : : * must already have been done for the child rel.
1366 : : *
1367 : : * The results are sufficient for use in merge, hash, and plain nestloop join
1368 : : * methods. We do not worry here about selecting clauses that are optimal
1369 : : * for use in a parameterized indexscan. indxpath.c makes its own selections
1370 : : * of clauses to use, and if the ones we pick here are redundant with those,
1371 : : * the extras will be eliminated at createplan time, using the parent_ec
1372 : : * markers that we provide (see is_redundant_derived_clause()).
1373 : : *
1374 : : * Because the same join clauses are likely to be needed multiple times as
1375 : : * we consider different join paths, we avoid generating multiple copies:
1376 : : * whenever we select a particular pair of EquivalenceMembers to join,
1377 : : * we check to see if the pair matches any original clause (in ec_sources)
1378 : : * or previously-built clause (in ec_derives). This saves memory and allows
1379 : : * re-use of information cached in RestrictInfos. We also avoid generating
1380 : : * commutative duplicates, i.e. if the algorithm selects "a.x = b.y" but
1381 : : * we already have "b.y = a.x", we return the existing clause.
1382 : : *
1383 : : * If we are considering an outer join, sjinfo is the associated OJ info,
1384 : : * otherwise it can be NULL.
1385 : : *
1386 : : * join_relids should always equal bms_union(outer_relids, inner_rel->relids)
1387 : : * plus whatever add_outer_joins_to_relids() would add. We could simplify
1388 : : * this function's API by computing it internally, but most callers have the
1389 : : * value at hand anyway.
1390 : : */
1391 : : List *
1392 : 215766 : generate_join_implied_equalities(PlannerInfo *root,
1393 : : Relids join_relids,
1394 : : Relids outer_relids,
1395 : : RelOptInfo *inner_rel,
1396 : : SpecialJoinInfo *sjinfo)
1397 : : {
1729 drowley@postgresql.o 1398 : 215766 : List *result = NIL;
1399 : 215766 : Relids inner_relids = inner_rel->relids;
1400 : : Relids nominal_inner_relids;
1401 : : Relids nominal_join_relids;
1402 : : Bitmapset *matching_ecs;
1403 : : int i;
1404 : :
1405 : : /* If inner rel is a child, extra setup work is needed */
1406 [ + + + + : 215766 : if (IS_OTHER_REL(inner_rel))
- + ]
1407 : : {
1408 [ - + ]: 3416 : Assert(!bms_is_empty(inner_rel->top_parent_relids));
1409 : :
1410 : : /* Fetch relid set for the topmost parent rel */
1411 : 3416 : nominal_inner_relids = inner_rel->top_parent_relids;
1412 : : /* ECs will be marked with the parent's relid, not the child's */
1413 : 3416 : nominal_join_relids = bms_union(outer_relids, nominal_inner_relids);
333 tgl@sss.pgh.pa.us 1414 : 3416 : nominal_join_relids = add_outer_joins_to_relids(root,
1415 : : nominal_join_relids,
1416 : : sjinfo,
1417 : : NULL);
1418 : : }
1419 : : else
1420 : : {
1729 drowley@postgresql.o 1421 : 212350 : nominal_inner_relids = inner_relids;
1422 : 212350 : nominal_join_relids = join_relids;
1423 : : }
1424 : :
1425 : : /*
1426 : : * Examine all potentially-relevant eclasses.
1427 : : *
1428 : : * If we are considering an outer join, we must include "join" clauses
1429 : : * that mention either input rel plus the outer join's relid; these
1430 : : * represent post-join filter clauses that have to be applied at this
1431 : : * join. We don't have infrastructure that would let us identify such
1432 : : * eclasses cheaply, so just fall back to considering all eclasses
1433 : : * mentioning anything in nominal_join_relids.
1434 : : *
1435 : : * At inner joins, we can be smarter: only consider eclasses mentioning
1436 : : * both input rels.
1437 : : */
333 tgl@sss.pgh.pa.us 1438 [ + + + + ]: 215766 : if (sjinfo && sjinfo->ojrelid != 0)
416 1439 : 46582 : matching_ecs = get_eclass_indexes_for_relids(root, nominal_join_relids);
1440 : : else
1441 : 169184 : matching_ecs = get_common_eclass_indexes(root, nominal_inner_relids,
1442 : : outer_relids);
1443 : :
1729 drowley@postgresql.o 1444 : 215766 : i = -1;
1445 [ + + ]: 632327 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
1446 : : {
1447 : 416561 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
1448 : 416561 : List *sublist = NIL;
1449 : :
1450 : : /* ECs containing consts do not need any further enforcement */
1451 [ + + ]: 416561 : if (ec->ec_has_const)
1452 : 58476 : continue;
1453 : :
1454 : : /* Single-member ECs won't generate any deductions */
1455 [ + + ]: 358085 : if (list_length(ec->ec_members) <= 1)
1456 : 214848 : continue;
1457 : :
1458 : : /* Sanity check that this eclass overlaps the join */
1459 [ - + ]: 143237 : Assert(bms_overlap(ec->ec_relids, nominal_join_relids));
1460 : :
1461 [ + + ]: 143237 : if (!ec->ec_broken)
1462 : 143075 : sublist = generate_join_implied_equalities_normal(root,
1463 : : ec,
1464 : : join_relids,
1465 : : outer_relids,
1466 : : inner_relids);
1467 : :
1468 : : /* Recover if we failed to generate required derived clauses */
1469 [ + + ]: 143237 : if (ec->ec_broken)
1470 : 180 : sublist = generate_join_implied_equalities_broken(root,
1471 : : ec,
1472 : : nominal_join_relids,
1473 : : outer_relids,
1474 : : nominal_inner_relids,
1475 : : inner_rel);
1476 : :
1477 : 143237 : result = list_concat(result, sublist);
1478 : : }
1479 : :
1480 : 215766 : return result;
1481 : : }
1482 : :
1483 : : /*
1484 : : * generate_join_implied_equalities_for_ecs
1485 : : * As above, but consider only the listed ECs.
1486 : : *
1487 : : * For the sole current caller, we can assume sjinfo == NULL, that is we are
1488 : : * not interested in outer-join filter clauses. This might need to change
1489 : : * in future.
1490 : : */
1491 : : List *
2907 tgl@sss.pgh.pa.us 1492 : 439 : generate_join_implied_equalities_for_ecs(PlannerInfo *root,
1493 : : List *eclasses,
1494 : : Relids join_relids,
1495 : : Relids outer_relids,
1496 : : RelOptInfo *inner_rel)
1497 : : {
6294 1498 : 439 : List *result = NIL;
4378 1499 : 439 : Relids inner_relids = inner_rel->relids;
1500 : : Relids nominal_inner_relids;
1501 : : Relids nominal_join_relids;
1502 : : ListCell *lc;
1503 : :
1504 : : /* If inner rel is a child, extra setup work is needed */
2568 rhaas@postgresql.org 1505 [ + - + - : 439 : if (IS_OTHER_REL(inner_rel))
- + ]
1506 : : {
2568 rhaas@postgresql.org 1507 [ # # ]:UBC 0 : Assert(!bms_is_empty(inner_rel->top_parent_relids));
1508 : :
1509 : : /* Fetch relid set for the topmost parent rel */
1510 : 0 : nominal_inner_relids = inner_rel->top_parent_relids;
1511 : : /* ECs will be marked with the parent's relid, not the child's */
4378 tgl@sss.pgh.pa.us 1512 : 0 : nominal_join_relids = bms_union(outer_relids, nominal_inner_relids);
1513 : : }
1514 : : else
1515 : : {
4378 tgl@sss.pgh.pa.us 1516 :CBC 439 : nominal_inner_relids = inner_relids;
1517 : 439 : nominal_join_relids = join_relids;
1518 : : }
1519 : :
2907 1520 [ + - + + : 900 : foreach(lc, eclasses)
+ + ]
1521 : : {
6294 1522 : 461 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc);
5995 bruce@momjian.us 1523 : 461 : List *sublist = NIL;
1524 : :
1525 : : /* ECs containing consts do not need any further enforcement */
6294 tgl@sss.pgh.pa.us 1526 [ - + ]: 461 : if (ec->ec_has_const)
6294 tgl@sss.pgh.pa.us 1527 :UBC 0 : continue;
1528 : :
1529 : : /* Single-member ECs won't generate any deductions */
6294 tgl@sss.pgh.pa.us 1530 [ - + ]:CBC 461 : if (list_length(ec->ec_members) <= 1)
6294 tgl@sss.pgh.pa.us 1531 :UBC 0 : continue;
1532 : :
1533 : : /* We can quickly ignore any that don't overlap the join, too */
4378 tgl@sss.pgh.pa.us 1534 [ - + ]:CBC 461 : if (!bms_overlap(ec->ec_relids, nominal_join_relids))
6294 tgl@sss.pgh.pa.us 1535 :UBC 0 : continue;
1536 : :
6294 tgl@sss.pgh.pa.us 1537 [ + - ]:CBC 461 : if (!ec->ec_broken)
1538 : 461 : sublist = generate_join_implied_equalities_normal(root,
1539 : : ec,
1540 : : join_relids,
1541 : : outer_relids,
1542 : : inner_relids);
1543 : :
1544 : : /* Recover if we failed to generate required derived clauses */
1545 [ - + ]: 461 : if (ec->ec_broken)
6294 tgl@sss.pgh.pa.us 1546 :UBC 0 : sublist = generate_join_implied_equalities_broken(root,
1547 : : ec,
1548 : : nominal_join_relids,
1549 : : outer_relids,
1550 : : nominal_inner_relids,
1551 : : inner_rel);
1552 : :
6294 tgl@sss.pgh.pa.us 1553 :CBC 461 : result = list_concat(result, sublist);
1554 : : }
1555 : :
1556 : 439 : return result;
1557 : : }
1558 : :
1559 : : /*
1560 : : * generate_join_implied_equalities for a still-valid EC
1561 : : */
1562 : : static List *
1563 : 143536 : generate_join_implied_equalities_normal(PlannerInfo *root,
1564 : : EquivalenceClass *ec,
1565 : : Relids join_relids,
1566 : : Relids outer_relids,
1567 : : Relids inner_relids)
1568 : : {
1569 : 143536 : List *result = NIL;
1570 : 143536 : List *new_members = NIL;
1571 : 143536 : List *outer_members = NIL;
1572 : 143536 : List *inner_members = NIL;
1573 : : ListCell *lc1;
1574 : :
1575 : : /*
1576 : : * First, scan the EC to identify member values that are computable at the
1577 : : * outer rel, at the inner rel, or at this relation but not in either
1578 : : * input rel. The outer-rel members should already be enforced equal,
1579 : : * likewise for the inner-rel members. We'll need to create clauses to
1580 : : * enforce that any newly computable members are all equal to each other
1581 : : * as well as to at least one input member, plus enforce at least one
1582 : : * outer-rel member equal to at least one inner-rel member.
1583 : : */
1584 [ + - + + : 494168 : foreach(lc1, ec->ec_members)
+ + ]
1585 : : {
1586 : 350632 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc1);
1587 : :
1588 : : /*
1589 : : * We don't need to check explicitly for child EC members. This test
1590 : : * against join_relids will cause them to be ignored except when
1591 : : * considering a child inner rel, which is what we want.
1592 : : */
4378 1593 [ + + ]: 350632 : if (!bms_is_subset(cur_em->em_relids, join_relids))
1594 : 63321 : continue; /* not computable yet, or wrong child */
1595 : :
1596 [ + + ]: 287311 : if (bms_is_subset(cur_em->em_relids, outer_relids))
6294 1597 : 165944 : outer_members = lappend(outer_members, cur_em);
4378 1598 [ + + ]: 121367 : else if (bms_is_subset(cur_em->em_relids, inner_relids))
6294 1599 : 120569 : inner_members = lappend(inner_members, cur_em);
1600 : : else
1601 : 798 : new_members = lappend(new_members, cur_em);
1602 : : }
1603 : :
1604 : : /*
1605 : : * First, select the joinclause if needed. We can equate any one outer
1606 : : * member to any one inner member, but we have to find a datatype
1607 : : * combination for which an opfamily member operator exists. If we have
1608 : : * choices, we prefer simple Var members (possibly with RelabelType) since
1609 : : * these are (a) cheapest to compute at runtime and (b) most likely to
1610 : : * have useful statistics. Also, prefer operators that are also
1611 : : * hashjoinable.
1612 : : */
1613 [ + + + + ]: 143536 : if (outer_members && inner_members)
1614 : : {
1615 : 114338 : EquivalenceMember *best_outer_em = NULL;
1616 : 114338 : EquivalenceMember *best_inner_em = NULL;
5995 bruce@momjian.us 1617 : 114338 : Oid best_eq_op = InvalidOid;
1618 : 114338 : int best_score = -1;
1619 : : RestrictInfo *rinfo;
1620 : :
6294 tgl@sss.pgh.pa.us 1621 [ + - + + : 121080 : foreach(lc1, outer_members)
+ + ]
1622 : : {
1623 : 114374 : EquivalenceMember *outer_em = (EquivalenceMember *) lfirst(lc1);
1624 : : ListCell *lc2;
1625 : :
1626 [ + - + + : 121122 : foreach(lc2, inner_members)
+ + ]
1627 : : {
1628 : 114380 : EquivalenceMember *inner_em = (EquivalenceMember *) lfirst(lc2);
1629 : : Oid eq_op;
1630 : : int score;
1631 : :
1632 : 114380 : eq_op = select_equality_operator(ec,
1633 : : outer_em->em_datatype,
1634 : : inner_em->em_datatype);
1635 [ + + ]: 114380 : if (!OidIsValid(eq_op))
1636 : 18 : continue;
1637 : 114362 : score = 0;
1638 [ + + ]: 114362 : if (IsA(outer_em->em_expr, Var) ||
1639 [ + + ]: 7302 : (IsA(outer_em->em_expr, RelabelType) &&
1640 [ + + ]: 1654 : IsA(((RelabelType *) outer_em->em_expr)->arg, Var)))
1641 : 108690 : score++;
1642 [ + + ]: 114362 : if (IsA(inner_em->em_expr, Var) ||
1643 [ + + ]: 2696 : (IsA(inner_em->em_expr, RelabelType) &&
1644 [ + + ]: 1449 : IsA(((RelabelType *) inner_em->em_expr)->arg, Var)))
1645 : 113106 : score++;
4854 1646 [ + + ]: 114362 : if (op_hashjoinable(eq_op,
4915 1647 : 114362 : exprType((Node *) outer_em->em_expr)))
6294 1648 : 114323 : score++;
1649 [ + + ]: 114362 : if (score > best_score)
1650 : : {
1651 : 114320 : best_outer_em = outer_em;
1652 : 114320 : best_inner_em = inner_em;
1653 : 114320 : best_eq_op = eq_op;
1654 : 114320 : best_score = score;
1655 [ + + ]: 114320 : if (best_score == 3)
5995 bruce@momjian.us 1656 : 107632 : break; /* no need to look further */
1657 : : }
1658 : : }
6294 tgl@sss.pgh.pa.us 1659 [ + + ]: 114374 : if (best_score == 3)
5995 bruce@momjian.us 1660 : 107632 : break; /* no need to look further */
1661 : : }
6294 tgl@sss.pgh.pa.us 1662 [ + + ]: 114338 : if (best_score < 0)
1663 : : {
1664 : : /* failed... */
1665 : 18 : ec->ec_broken = true;
1666 : 18 : return NIL;
1667 : : }
1668 : :
1669 : : /*
1670 : : * Create clause, setting parent_ec to mark it as redundant with other
1671 : : * joinclauses
1672 : : */
6292 1673 : 114320 : rinfo = create_join_clause(root, ec, best_eq_op,
1674 : : best_outer_em, best_inner_em,
1675 : : ec);
1676 : :
6294 1677 : 114320 : result = lappend(result, rinfo);
1678 : : }
1679 : :
1680 : : /*
1681 : : * Now deal with building restrictions for any expressions that involve
1682 : : * Vars from both sides of the join. We have to equate all of these to
1683 : : * each other as well as to at least one old member (if any).
1684 : : *
1685 : : * XXX as in generate_base_implied_equalities_no_const, we could be a lot
1686 : : * smarter here to avoid unnecessary failures in cross-type situations.
1687 : : * For now, use the same left-to-right method used there.
1688 : : */
1689 [ + + ]: 143518 : if (new_members)
1690 : : {
1691 : 780 : List *old_members = list_concat(outer_members, inner_members);
1692 : 780 : EquivalenceMember *prev_em = NULL;
1693 : : RestrictInfo *rinfo;
1694 : :
1695 : : /* For now, arbitrarily take the first old_member as the one to use */
1696 [ + + ]: 780 : if (old_members)
1697 : 549 : new_members = lappend(new_members, linitial(old_members));
1698 : :
1699 [ + - + + : 2127 : foreach(lc1, new_members)
+ + ]
1700 : : {
1701 : 1347 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc1);
1702 : :
1703 [ + + ]: 1347 : if (prev_em != NULL)
1704 : : {
1705 : : Oid eq_op;
1706 : :
1707 : 567 : eq_op = select_equality_operator(ec,
1708 : : prev_em->em_datatype,
1709 : : cur_em->em_datatype);
1710 [ - + ]: 567 : if (!OidIsValid(eq_op))
1711 : : {
1712 : : /* failed... */
6294 tgl@sss.pgh.pa.us 1713 :UBC 0 : ec->ec_broken = true;
1714 : 0 : return NIL;
1715 : : }
1716 : : /* do NOT set parent_ec, this qual is not redundant! */
6292 tgl@sss.pgh.pa.us 1717 :CBC 567 : rinfo = create_join_clause(root, ec, eq_op,
1718 : : prev_em, cur_em,
1719 : : NULL);
1720 : :
6294 1721 : 567 : result = lappend(result, rinfo);
1722 : : }
1723 : 1347 : prev_em = cur_em;
1724 : : }
1725 : : }
1726 : :
1727 : 143518 : return result;
1728 : : }
1729 : :
1730 : : /*
1731 : : * generate_join_implied_equalities cleanup after failure
1732 : : *
1733 : : * Return any original RestrictInfos that are enforceable at this join.
1734 : : *
1735 : : * In the case of a child inner relation, we have to translate the
1736 : : * original RestrictInfos from parent to child Vars.
1737 : : */
1738 : : static List *
1739 : 180 : generate_join_implied_equalities_broken(PlannerInfo *root,
1740 : : EquivalenceClass *ec,
1741 : : Relids nominal_join_relids,
1742 : : Relids outer_relids,
1743 : : Relids nominal_inner_relids,
1744 : : RelOptInfo *inner_rel)
1745 : : {
1746 : 180 : List *result = NIL;
1747 : : ListCell *lc;
1748 : :
1749 [ + - + + : 492 : foreach(lc, ec->ec_sources)
+ + ]
1750 : : {
1751 : 312 : RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(lc);
4378 1752 : 312 : Relids clause_relids = restrictinfo->required_relids;
1753 : :
1754 [ + + ]: 312 : if (bms_is_subset(clause_relids, nominal_join_relids) &&
1755 [ + + ]: 168 : !bms_is_subset(clause_relids, outer_relids) &&
1756 [ + - ]: 156 : !bms_is_subset(clause_relids, nominal_inner_relids))
6294 1757 : 156 : result = lappend(result, restrictinfo);
1758 : : }
1759 : :
1760 : : /*
1761 : : * If we have to translate, just brute-force apply adjust_appendrel_attrs
1762 : : * to all the RestrictInfos at once. This will result in returning
1763 : : * RestrictInfos that are not listed in ec_derives, but there shouldn't be
1764 : : * any duplication, and it's a sufficiently narrow corner case that we
1765 : : * shouldn't sweat too much over it anyway.
1766 : : *
1767 : : * Since inner_rel might be an indirect descendant of the baserel
1768 : : * mentioned in the ec_sources clauses, we have to be prepared to apply
1769 : : * multiple levels of Var translation.
1770 : : */
2568 rhaas@postgresql.org 1771 [ + + + - : 180 : if (IS_OTHER_REL(inner_rel) && result != NIL)
- + + + ]
3483 tgl@sss.pgh.pa.us 1772 : 81 : result = (List *) adjust_appendrel_attrs_multilevel(root,
1773 : : (Node *) result,
1774 : : inner_rel,
605 1775 : 81 : inner_rel->top_parent);
1776 : :
6294 1777 : 180 : return result;
1778 : : }
1779 : :
1780 : :
1781 : : /*
1782 : : * select_equality_operator
1783 : : * Select a suitable equality operator for comparing two EC members
1784 : : *
1785 : : * Returns InvalidOid if no operator can be found for this datatype combination
1786 : : */
1787 : : static Oid
5995 bruce@momjian.us 1788 : 173843 : select_equality_operator(EquivalenceClass *ec, Oid lefttype, Oid righttype)
1789 : : {
1790 : : ListCell *lc;
1791 : :
6294 tgl@sss.pgh.pa.us 1792 [ + - + + : 173876 : foreach(lc, ec->ec_opfamilies)
+ + ]
1793 : : {
5995 bruce@momjian.us 1794 : 173843 : Oid opfamily = lfirst_oid(lc);
1795 : : Oid opno;
1796 : :
6294 tgl@sss.pgh.pa.us 1797 : 173843 : opno = get_opfamily_member(opfamily, lefttype, righttype,
1798 : : BTEqualStrategyNumber);
2643 1799 [ + + ]: 173843 : if (!OidIsValid(opno))
1800 : 33 : continue;
1801 : : /* If no barrier quals in query, don't worry about leaky operators */
1802 [ + + ]: 173810 : if (ec->ec_max_security == 0)
1803 : 173810 : return opno;
1804 : : /* Otherwise, insist that selected operators be leakproof */
1805 [ + - ]: 214 : if (get_func_leakproof(get_opcode(opno)))
6294 1806 : 214 : return opno;
1807 : : }
1808 : 33 : return InvalidOid;
1809 : : }
1810 : :
1811 : :
1812 : : /*
1813 : : * create_join_clause
1814 : : * Find or make a RestrictInfo comparing the two given EC members
1815 : : * with the given operator (or, possibly, its commutator, because
1816 : : * the ordering of the operands in the result is not guaranteed).
1817 : : *
1818 : : * parent_ec is either equal to ec (if the clause is a potentially-redundant
1819 : : * join clause) or NULL (if not). We have to treat this as part of the
1820 : : * match requirements --- it's possible that a clause comparing the same two
1821 : : * EMs is a join clause in one join path and a restriction clause in another.
1822 : : */
1823 : : static RestrictInfo *
6292 1824 : 157926 : create_join_clause(PlannerInfo *root,
1825 : : EquivalenceClass *ec, Oid opno,
1826 : : EquivalenceMember *leftem,
1827 : : EquivalenceMember *rightem,
1828 : : EquivalenceClass *parent_ec)
1829 : : {
1830 : : RestrictInfo *rinfo;
440 1831 : 157926 : RestrictInfo *parent_rinfo = NULL;
1832 : : ListCell *lc;
1833 : : MemoryContext oldcontext;
1834 : :
1835 : : /*
1836 : : * Search to see if we already built a RestrictInfo for this pair of
1837 : : * EquivalenceMembers. We can use either original source clauses or
1838 : : * previously-derived clauses, and a commutator clause is acceptable.
1839 : : *
1840 : : * We used to verify that opno matches, but that seems redundant: even if
1841 : : * it's not identical, it'd better have the same effects, or the operator
1842 : : * families we're using are broken.
1843 : : */
6292 1844 [ + - + + : 343069 : foreach(lc, ec->ec_sources)
+ + ]
1845 : : {
1846 : 185629 : rinfo = (RestrictInfo *) lfirst(lc);
1847 [ + + ]: 185629 : if (rinfo->left_em == leftem &&
1848 [ + + ]: 84198 : rinfo->right_em == rightem &&
535 1849 [ + + ]: 74741 : rinfo->parent_ec == parent_ec)
1850 : 486 : return rinfo;
1851 [ + + ]: 185578 : if (rinfo->left_em == rightem &&
1852 [ + + ]: 81015 : rinfo->right_em == leftem &&
1853 [ + + ]: 73322 : rinfo->parent_ec == parent_ec)
6292 1854 : 435 : return rinfo;
1855 : : }
1856 : :
1857 [ + + + + : 203277 : foreach(lc, ec->ec_derives)
+ + ]
1858 : : {
1859 : 175086 : rinfo = (RestrictInfo *) lfirst(lc);
1860 [ + + ]: 175086 : if (rinfo->left_em == leftem &&
1861 [ + + ]: 68546 : rinfo->right_em == rightem &&
535 1862 [ + + ]: 61135 : rinfo->parent_ec == parent_ec)
1863 : 129249 : return rinfo;
1864 [ + + ]: 113954 : if (rinfo->left_em == rightem &&
1865 [ + + ]: 72267 : rinfo->right_em == leftem &&
1866 [ + - ]: 68117 : rinfo->parent_ec == parent_ec)
6292 1867 : 68117 : return rinfo;
1868 : : }
1869 : :
1870 : : /*
1871 : : * Not there, so build it, in planner context so we can re-use it. (Not
1872 : : * important in normal planning, but definitely so in GEQO.)
1873 : : */
1874 : 28191 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
1875 : :
1876 : : /*
1877 : : * If either EM is a child, recursively create the corresponding
1878 : : * parent-to-parent clause, so that we can duplicate its rinfo_serial.
1879 : : */
440 1880 [ + + + + ]: 28191 : if (leftem->em_is_child || rightem->em_is_child)
1881 : : {
1882 [ + + ]: 1847 : EquivalenceMember *leftp = leftem->em_parent ? leftem->em_parent : leftem;
1883 [ + + ]: 1847 : EquivalenceMember *rightp = rightem->em_parent ? rightem->em_parent : rightem;
1884 : :
1885 : 1847 : parent_rinfo = create_join_clause(root, ec, opno,
1886 : : leftp, rightp,
1887 : : parent_ec);
1888 : : }
1889 : :
1179 1890 : 28191 : rinfo = build_implied_join_equality(root,
1891 : : opno,
1892 : : ec->ec_collation,
1893 : : leftem->em_expr,
1894 : : rightem->em_expr,
5987 1895 : 28191 : bms_union(leftem->em_relids,
4196 1896 : 28191 : rightem->em_relids),
1897 : : ec->ec_min_security);
1898 : :
1899 : : /* If it's a child clause, copy the parent's rinfo_serial */
440 1900 [ + + ]: 28191 : if (parent_rinfo)
1901 : 1847 : rinfo->rinfo_serial = parent_rinfo->rinfo_serial;
1902 : :
1903 : : /* Mark the clause as redundant, or not */
6292 1904 : 28191 : rinfo->parent_ec = parent_ec;
1905 : :
1906 : : /*
1907 : : * We know the correct values for left_ec/right_ec, ie this particular EC,
1908 : : * so we can just set them directly instead of forcing another lookup.
1909 : : */
1910 : 28191 : rinfo->left_ec = ec;
1911 : 28191 : rinfo->right_ec = ec;
1912 : :
1913 : : /* Mark it as usable with these EMs */
1914 : 28191 : rinfo->left_em = leftem;
1915 : 28191 : rinfo->right_em = rightem;
1916 : : /* and save it for possible re-use */
1917 : 28191 : ec->ec_derives = lappend(ec->ec_derives, rinfo);
1918 : :
1919 : 28191 : MemoryContextSwitchTo(oldcontext);
1920 : :
1921 : 28191 : return rinfo;
1922 : : }
1923 : :
1924 : :
1925 : : /*
1926 : : * reconsider_outer_join_clauses
1927 : : * Re-examine any outer-join clauses that were set aside by
1928 : : * distribute_qual_to_rels(), and see if we can derive any
1929 : : * EquivalenceClasses from them. Then, if they were not made
1930 : : * redundant, push them out into the regular join-clause lists.
1931 : : *
1932 : : * When we have mergejoinable clauses A = B that are outer-join clauses,
1933 : : * we can't blindly combine them with other clauses A = C to deduce B = C,
1934 : : * since in fact the "equality" A = B won't necessarily hold above the
1935 : : * outer join (one of the variables might be NULL instead). Nonetheless
1936 : : * there are cases where we can add qual clauses using transitivity.
1937 : : *
1938 : : * One case that we look for here is an outer-join clause OUTERVAR = INNERVAR
1939 : : * for which there is also an equivalence clause OUTERVAR = CONSTANT.
1940 : : * It is safe and useful to push a clause INNERVAR = CONSTANT into the
1941 : : * evaluation of the inner (nullable) relation, because any inner rows not
1942 : : * meeting this condition will not contribute to the outer-join result anyway.
1943 : : * (Any outer rows they could join to will be eliminated by the pushed-down
1944 : : * equivalence clause.)
1945 : : *
1946 : : * Note that the above rule does not work for full outer joins; nor is it
1947 : : * very interesting to consider cases where the generated equivalence clause
1948 : : * would involve relations outside the outer join, since such clauses couldn't
1949 : : * be pushed into the inner side's scan anyway. So the restriction to
1950 : : * outervar = pseudoconstant is not really giving up anything.
1951 : : *
1952 : : * For full-join cases, we can only do something useful if it's a FULL JOIN
1953 : : * USING and a merged column has an equivalence MERGEDVAR = CONSTANT.
1954 : : * By the time it gets here, the merged column will look like
1955 : : * COALESCE(LEFTVAR, RIGHTVAR)
1956 : : * and we will have a full-join clause LEFTVAR = RIGHTVAR that we can match
1957 : : * the COALESCE expression to. In this situation we can push LEFTVAR = CONSTANT
1958 : : * and RIGHTVAR = CONSTANT into the input relations, since any rows not
1959 : : * meeting these conditions cannot contribute to the join result.
1960 : : *
1961 : : * Again, there isn't any traction to be gained by trying to deal with
1962 : : * clauses comparing a mergedvar to a non-pseudoconstant. So we can make
1963 : : * use of the EquivalenceClasses to search for matching variables that were
1964 : : * equivalenced to constants. The interesting outer-join clauses were
1965 : : * accumulated for us by distribute_qual_to_rels.
1966 : : *
1967 : : * When we find one of these cases, we implement the changes we want by
1968 : : * generating a new equivalence clause INNERVAR = CONSTANT (or LEFTVAR, etc)
1969 : : * and pushing it into the EquivalenceClass structures. This is because we
1970 : : * may already know that INNERVAR is equivalenced to some other var(s), and
1971 : : * we'd like the constant to propagate to them too. Note that it would be
1972 : : * unsafe to merge any existing EC for INNERVAR with the OUTERVAR's EC ---
1973 : : * that could result in propagating constant restrictions from
1974 : : * INNERVAR to OUTERVAR, which would be very wrong.
1975 : : *
1976 : : * It's possible that the INNERVAR is also an OUTERVAR for some other
1977 : : * outer-join clause, in which case the process can be repeated. So we repeat
1978 : : * looping over the lists of clauses until no further deductions can be made.
1979 : : * Whenever we do make a deduction, we remove the generating clause from the
1980 : : * lists, since we don't want to make the same deduction twice.
1981 : : *
1982 : : * If we don't find any match for a set-aside outer join clause, we must
1983 : : * throw it back into the regular joinclause processing by passing it to
1984 : : * distribute_restrictinfo_to_rels(). If we do generate a derived clause,
1985 : : * however, the outer-join clause is redundant. We must still put some
1986 : : * clause into the regular processing, because otherwise the join will be
1987 : : * seen as a clauseless join and avoided during join order searching.
1988 : : * We handle this by generating a constant-TRUE clause that is marked with
1989 : : * the same required_relids etc as the removed outer-join clause, thus
1990 : : * making it a join clause between the correct relations.
1991 : : */
1992 : : void
6294 1993 : 139537 : reconsider_outer_join_clauses(PlannerInfo *root)
1994 : : {
1995 : : bool found;
1996 : : ListCell *cell;
1997 : :
1998 : : /* Outer loop repeats until we find no more deductions */
1999 : : do
2000 : : {
5940 2001 : 140335 : found = false;
2002 : :
2003 : : /* Process the LEFT JOIN clauses */
1735 2004 [ + + + + : 156395 : foreach(cell, root->left_join_clauses)
+ + ]
2005 : : {
440 2006 : 16060 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2007 : :
2008 [ + + ]: 16060 : if (reconsider_outer_join_clause(root, ojcinfo, true))
2009 : : {
2010 : 360 : RestrictInfo *rinfo = ojcinfo->rinfo;
2011 : :
5940 2012 : 360 : found = true;
2013 : : /* remove it from the list */
2014 : 360 : root->left_join_clauses =
1735 2015 : 360 : foreach_delete_current(root->left_join_clauses, cell);
2016 : : /* throw back a dummy replacement clause (see notes above) */
440 2017 : 360 : rinfo = make_restrictinfo(root,
2018 : 360 : (Expr *) makeBoolConst(true, false),
325 2019 : 360 : rinfo->is_pushed_down,
2020 : 360 : rinfo->has_clone,
2021 : 360 : rinfo->is_clone,
2022 : : false, /* pseudoconstant */
2023 : : 0, /* security_level */
2024 : : rinfo->required_relids,
2025 : : rinfo->incompatible_relids,
2026 : : rinfo->outer_relids);
5940 2027 : 360 : distribute_restrictinfo_to_rels(root, rinfo);
2028 : : }
2029 : : }
2030 : :
2031 : : /* Process the RIGHT JOIN clauses */
1735 2032 [ + + + + : 150455 : foreach(cell, root->right_join_clauses)
+ + ]
2033 : : {
440 2034 : 10120 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2035 : :
2036 [ + + ]: 10120 : if (reconsider_outer_join_clause(root, ojcinfo, false))
2037 : : {
2038 : 441 : RestrictInfo *rinfo = ojcinfo->rinfo;
2039 : :
5940 2040 : 441 : found = true;
2041 : : /* remove it from the list */
2042 : 441 : root->right_join_clauses =
1735 2043 : 441 : foreach_delete_current(root->right_join_clauses, cell);
2044 : : /* throw back a dummy replacement clause (see notes above) */
440 2045 : 441 : rinfo = make_restrictinfo(root,
2046 : 441 : (Expr *) makeBoolConst(true, false),
325 2047 : 441 : rinfo->is_pushed_down,
2048 : 441 : rinfo->has_clone,
2049 : 441 : rinfo->is_clone,
2050 : : false, /* pseudoconstant */
2051 : : 0, /* security_level */
2052 : : rinfo->required_relids,
2053 : : rinfo->incompatible_relids,
2054 : : rinfo->outer_relids);
5940 2055 : 441 : distribute_restrictinfo_to_rels(root, rinfo);
2056 : : }
2057 : : }
2058 : :
2059 : : /* Process the FULL JOIN clauses */
1735 2060 [ + + + + : 140946 : foreach(cell, root->full_join_clauses)
+ + ]
2061 : : {
440 2062 : 611 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2063 : :
2064 [ + + ]: 611 : if (reconsider_full_join_clause(root, ojcinfo))
2065 : : {
2066 : 3 : RestrictInfo *rinfo = ojcinfo->rinfo;
2067 : :
5940 2068 : 3 : found = true;
2069 : : /* remove it from the list */
2070 : 3 : root->full_join_clauses =
1735 2071 : 3 : foreach_delete_current(root->full_join_clauses, cell);
2072 : : /* throw back a dummy replacement clause (see notes above) */
440 2073 : 3 : rinfo = make_restrictinfo(root,
2074 : 3 : (Expr *) makeBoolConst(true, false),
325 2075 : 3 : rinfo->is_pushed_down,
2076 : 3 : rinfo->has_clone,
2077 : 3 : rinfo->is_clone,
2078 : : false, /* pseudoconstant */
2079 : : 0, /* security_level */
2080 : : rinfo->required_relids,
2081 : : rinfo->incompatible_relids,
2082 : : rinfo->outer_relids);
5940 2083 : 3 : distribute_restrictinfo_to_rels(root, rinfo);
2084 : : }
2085 : : }
2086 [ + + ]: 140335 : } while (found);
2087 : :
2088 : : /* Now, any remaining clauses have to be thrown back */
2089 [ + + + + : 154964 : foreach(cell, root->left_join_clauses)
+ + ]
2090 : : {
440 2091 : 15427 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2092 : :
2093 : 15427 : distribute_restrictinfo_to_rels(root, ojcinfo->rinfo);
2094 : : }
5940 2095 [ + + + + : 148751 : foreach(cell, root->right_join_clauses)
+ + ]
2096 : : {
440 2097 : 9214 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2098 : :
2099 : 9214 : distribute_restrictinfo_to_rels(root, ojcinfo->rinfo);
2100 : : }
5940 2101 [ + + + + : 140145 : foreach(cell, root->full_join_clauses)
+ + ]
2102 : : {
440 2103 : 608 : OuterJoinClauseInfo *ojcinfo = (OuterJoinClauseInfo *) lfirst(cell);
2104 : :
2105 : 608 : distribute_restrictinfo_to_rels(root, ojcinfo->rinfo);
2106 : : }
6294 2107 : 139537 : }
2108 : :
2109 : : /*
2110 : : * reconsider_outer_join_clauses for a single LEFT/RIGHT JOIN clause
2111 : : *
2112 : : * Returns true if we were able to propagate a constant through the clause.
2113 : : */
2114 : : static bool
440 2115 : 26180 : reconsider_outer_join_clause(PlannerInfo *root, OuterJoinClauseInfo *ojcinfo,
2116 : : bool outer_on_left)
2117 : : {
2118 : 26180 : RestrictInfo *rinfo = ojcinfo->rinfo;
2119 : 26180 : SpecialJoinInfo *sjinfo = ojcinfo->sjinfo;
2120 : : Expr *outervar,
2121 : : *innervar;
2122 : : Oid opno,
2123 : : collation,
2124 : : left_type,
2125 : : right_type,
2126 : : inner_datatype;
2127 : : Relids inner_relids;
2128 : : ListCell *lc1;
2129 : :
6294 2130 [ - + ]: 26180 : Assert(is_opclause(rinfo->clause));
5940 2131 : 26180 : opno = ((OpExpr *) rinfo->clause)->opno;
4775 2132 : 26180 : collation = ((OpExpr *) rinfo->clause)->inputcollid;
2133 : :
2134 : : /* Extract needed info from the clause */
5940 2135 : 26180 : op_input_types(opno, &left_type, &right_type);
6294 2136 [ + + ]: 26180 : if (outer_on_left)
2137 : : {
2138 : 16060 : outervar = (Expr *) get_leftop(rinfo->clause);
2139 : 16060 : innervar = (Expr *) get_rightop(rinfo->clause);
2140 : 16060 : inner_datatype = right_type;
5987 2141 : 16060 : inner_relids = rinfo->right_relids;
2142 : : }
2143 : : else
2144 : : {
6294 2145 : 10120 : outervar = (Expr *) get_rightop(rinfo->clause);
2146 : 10120 : innervar = (Expr *) get_leftop(rinfo->clause);
2147 : 10120 : inner_datatype = left_type;
5987 2148 : 10120 : inner_relids = rinfo->left_relids;
2149 : : }
2150 : :
2151 : : /* Scan EquivalenceClasses for a match to outervar */
6294 2152 [ + - + + : 167521 : foreach(lc1, root->eq_classes)
+ + ]
2153 : : {
2154 : 142142 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2155 : : bool match;
2156 : : ListCell *lc2;
2157 : :
2158 : : /* Ignore EC unless it contains pseudoconstants */
2159 [ + + ]: 142142 : if (!cur_ec->ec_has_const)
2160 : 111789 : continue;
2161 : : /* Never match to a volatile EC */
2162 [ - + ]: 30353 : if (cur_ec->ec_has_volatile)
6294 tgl@sss.pgh.pa.us 2163 :UBC 0 : continue;
2164 : : /* It has to match the outer-join clause as to semantics, too */
4775 tgl@sss.pgh.pa.us 2165 [ + + ]:CBC 30353 : if (collation != cur_ec->ec_collation)
2166 : 1131 : continue;
6294 2167 [ + + ]: 29222 : if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
2168 : 6902 : continue;
2169 : : /* Does it contain a match to outervar? */
2170 : 22320 : match = false;
2171 [ + - + + : 68752 : foreach(lc2, cur_ec->ec_members)
+ + ]
2172 : : {
2173 : 47233 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2174 : :
2489 2175 [ - + ]: 47233 : Assert(!cur_em->em_is_child); /* no children yet */
6294 2176 [ + + ]: 47233 : if (equal(outervar, cur_em->em_expr))
2177 : : {
2178 : 801 : match = true;
2179 : 801 : break;
2180 : : }
2181 : : }
2182 [ + + ]: 22320 : if (!match)
2183 : 21519 : continue; /* no match, so ignore this EC */
2184 : :
2185 : : /*
2186 : : * Yes it does! Try to generate a clause INNERVAR = CONSTANT for each
2187 : : * CONSTANT in the EC. Note that we must succeed with at least one
2188 : : * constant before we can decide to throw away the outer-join clause.
2189 : : */
2190 : 801 : match = false;
2191 [ + - + + : 2868 : foreach(lc2, cur_ec->ec_members)
+ + ]
2192 : : {
2193 : 2067 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2194 : : Oid eq_op;
2195 : : RestrictInfo *newrinfo;
2196 : : JoinDomain *jdomain;
2197 : :
2198 [ + + ]: 2067 : if (!cur_em->em_is_const)
2199 : 1245 : continue; /* ignore non-const members */
2200 : 822 : eq_op = select_equality_operator(cur_ec,
2201 : : inner_datatype,
2202 : : cur_em->em_datatype);
2203 [ - + ]: 822 : if (!OidIsValid(eq_op))
6294 tgl@sss.pgh.pa.us 2204 :UBC 0 : continue; /* can't generate equality */
1179 tgl@sss.pgh.pa.us 2205 :CBC 822 : newrinfo = build_implied_join_equality(root,
2206 : : eq_op,
2207 : : cur_ec->ec_collation,
2208 : : innervar,
2209 : : cur_em->em_expr,
2210 : : bms_copy(inner_relids),
2211 : : cur_ec->ec_min_security);
2212 : : /* This equality holds within the OJ's child JoinDomain */
440 2213 : 822 : jdomain = find_join_domain(root, sjinfo->syn_righthand);
2214 [ + - ]: 822 : if (process_equivalence(root, &newrinfo, jdomain))
6294 2215 : 822 : match = true;
2216 : : }
2217 : :
2218 : : /*
2219 : : * If we were able to equate INNERVAR to any constant, report success.
2220 : : * Otherwise, fall out of the search loop, since we know the OUTERVAR
2221 : : * appears in at most one EC.
2222 : : */
2223 [ + - ]: 801 : if (match)
5940 2224 : 801 : return true;
2225 : : else
6294 tgl@sss.pgh.pa.us 2226 :UBC 0 : break;
2227 : : }
2228 : :
5940 tgl@sss.pgh.pa.us 2229 :CBC 25379 : return false; /* failed to make any deduction */
2230 : : }
2231 : :
2232 : : /*
2233 : : * reconsider_outer_join_clauses for a single FULL JOIN clause
2234 : : *
2235 : : * Returns true if we were able to propagate a constant through the clause.
2236 : : */
2237 : : static bool
440 2238 : 611 : reconsider_full_join_clause(PlannerInfo *root, OuterJoinClauseInfo *ojcinfo)
2239 : : {
2240 : 611 : RestrictInfo *rinfo = ojcinfo->rinfo;
2241 : 611 : SpecialJoinInfo *sjinfo = ojcinfo->sjinfo;
2242 : 611 : Relids fjrelids = bms_make_singleton(sjinfo->ojrelid);
2243 : : Expr *leftvar;
2244 : : Expr *rightvar;
2245 : : Oid opno,
2246 : : collation,
2247 : : left_type,
2248 : : right_type;
2249 : : Relids left_relids,
2250 : : right_relids;
2251 : : ListCell *lc1;
2252 : :
2253 : : /* Extract needed info from the clause */
6294 2254 [ - + ]: 611 : Assert(is_opclause(rinfo->clause));
5940 2255 : 611 : opno = ((OpExpr *) rinfo->clause)->opno;
4775 2256 : 611 : collation = ((OpExpr *) rinfo->clause)->inputcollid;
5940 2257 : 611 : op_input_types(opno, &left_type, &right_type);
6294 2258 : 611 : leftvar = (Expr *) get_leftop(rinfo->clause);
2259 : 611 : rightvar = (Expr *) get_rightop(rinfo->clause);
5987 2260 : 611 : left_relids = rinfo->left_relids;
2261 : 611 : right_relids = rinfo->right_relids;
2262 : :
6294 2263 [ + - + + : 3123 : foreach(lc1, root->eq_classes)
+ + ]
2264 : : {
2265 : 2515 : EquivalenceClass *cur_ec = (EquivalenceClass *) lfirst(lc1);
2266 : 2515 : EquivalenceMember *coal_em = NULL;
2267 : : bool match;
2268 : : bool matchleft;
2269 : : bool matchright;
2270 : : ListCell *lc2;
1270 drowley@postgresql.o 2271 : 2515 : int coal_idx = -1;
2272 : :
2273 : : /* Ignore EC unless it contains pseudoconstants */
6294 tgl@sss.pgh.pa.us 2274 [ + + ]: 2515 : if (!cur_ec->ec_has_const)
2275 : 2367 : continue;
2276 : : /* Never match to a volatile EC */
2277 [ - + ]: 148 : if (cur_ec->ec_has_volatile)
6294 tgl@sss.pgh.pa.us 2278 :UBC 0 : continue;
2279 : : /* It has to match the outer-join clause as to semantics, too */
4775 tgl@sss.pgh.pa.us 2280 [ + + ]:CBC 148 : if (collation != cur_ec->ec_collation)
2281 : 18 : continue;
6294 2282 [ - + ]: 130 : if (!equal(rinfo->mergeopfamilies, cur_ec->ec_opfamilies))
6294 tgl@sss.pgh.pa.us 2283 :UBC 0 : continue;
2284 : :
2285 : : /*
2286 : : * Does it contain a COALESCE(leftvar, rightvar) construct?
2287 : : *
2288 : : * We can assume the COALESCE() inputs are in the same order as the
2289 : : * join clause, since both were automatically generated in the cases
2290 : : * we care about.
2291 : : *
2292 : : * XXX currently this may fail to match in cross-type cases because
2293 : : * the COALESCE will contain typecast operations while the join clause
2294 : : * may not (if there is a cross-type mergejoin operator available for
2295 : : * the two column types). Is it OK to strip implicit coercions from
2296 : : * the COALESCE arguments?
2297 : : */
6294 tgl@sss.pgh.pa.us 2298 :CBC 130 : match = false;
2299 [ + - + + : 381 : foreach(lc2, cur_ec->ec_members)
+ + ]
2300 : : {
2301 : 254 : coal_em = (EquivalenceMember *) lfirst(lc2);
2489 2302 [ - + ]: 254 : Assert(!coal_em->em_is_child); /* no children yet */
6294 2303 [ + + ]: 254 : if (IsA(coal_em->em_expr, CoalesceExpr))
2304 : : {
2305 : 9 : CoalesceExpr *cexpr = (CoalesceExpr *) coal_em->em_expr;
2306 : : Node *cfirst;
2307 : : Node *csecond;
2308 : :
2309 [ + + ]: 9 : if (list_length(cexpr->args) != 2)
2310 : 6 : continue;
2311 : 3 : cfirst = (Node *) linitial(cexpr->args);
2312 : 3 : csecond = (Node *) lsecond(cexpr->args);
2313 : :
2314 : : /*
2315 : : * The COALESCE arguments will be marked as possibly nulled by
2316 : : * the full join, while we wish to generate clauses that apply
2317 : : * to the join's inputs. So we must strip the join from the
2318 : : * nullingrels fields of cfirst/csecond before comparing them
2319 : : * to leftvar/rightvar. (Perhaps with a less hokey
2320 : : * representation for FULL JOIN USING output columns, this
2321 : : * wouldn't be needed?)
2322 : : */
440 2323 : 3 : cfirst = remove_nulling_relids(cfirst, fjrelids, NULL);
2324 : 3 : csecond = remove_nulling_relids(csecond, fjrelids, NULL);
2325 : :
6294 2326 [ + - + - ]: 3 : if (equal(leftvar, cfirst) && equal(rightvar, csecond))
2327 : : {
1270 drowley@postgresql.o 2328 : 3 : coal_idx = foreach_current_index(lc2);
6294 tgl@sss.pgh.pa.us 2329 : 3 : match = true;
2330 : 3 : break;
2331 : : }
2332 : : }
2333 : : }
2334 [ + + ]: 130 : if (!match)
2335 : 127 : continue; /* no match, so ignore this EC */
2336 : :
2337 : : /*
2338 : : * Yes it does! Try to generate clauses LEFTVAR = CONSTANT and
2339 : : * RIGHTVAR = CONSTANT for each CONSTANT in the EC. Note that we must
2340 : : * succeed with at least one constant for each var before we can
2341 : : * decide to throw away the outer-join clause.
2342 : : */
2343 : 3 : matchleft = matchright = false;
2344 [ + - + + : 9 : foreach(lc2, cur_ec->ec_members)
+ + ]
2345 : : {
2346 : 6 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc2);
2347 : : Oid eq_op;
2348 : : RestrictInfo *newrinfo;
2349 : : JoinDomain *jdomain;
2350 : :
2351 [ + + ]: 6 : if (!cur_em->em_is_const)
2352 : 3 : continue; /* ignore non-const members */
2353 : 3 : eq_op = select_equality_operator(cur_ec,
2354 : : left_type,
2355 : : cur_em->em_datatype);
2356 [ + - ]: 3 : if (OidIsValid(eq_op))
2357 : : {
1179 2358 : 3 : newrinfo = build_implied_join_equality(root,
2359 : : eq_op,
2360 : : cur_ec->ec_collation,
2361 : : leftvar,
2362 : : cur_em->em_expr,
2363 : : bms_copy(left_relids),
2364 : : cur_ec->ec_min_security);
2365 : : /* This equality holds within the lefthand child JoinDomain */
440 2366 : 3 : jdomain = find_join_domain(root, sjinfo->syn_lefthand);
2367 [ + - ]: 3 : if (process_equivalence(root, &newrinfo, jdomain))
6294 2368 : 3 : matchleft = true;
2369 : : }
2370 : 3 : eq_op = select_equality_operator(cur_ec,
2371 : : right_type,
2372 : : cur_em->em_datatype);
2373 [ + - ]: 3 : if (OidIsValid(eq_op))
2374 : : {
1179 2375 : 3 : newrinfo = build_implied_join_equality(root,
2376 : : eq_op,
2377 : : cur_ec->ec_collation,
2378 : : rightvar,
2379 : : cur_em->em_expr,
2380 : : bms_copy(right_relids),
2381 : : cur_ec->ec_min_security);
2382 : : /* This equality holds within the righthand child JoinDomain */
440 2383 : 3 : jdomain = find_join_domain(root, sjinfo->syn_righthand);
2384 [ + - ]: 3 : if (process_equivalence(root, &newrinfo, jdomain))
6294 2385 : 3 : matchright = true;
2386 : : }
2387 : : }
2388 : :
2389 : : /*
2390 : : * If we were able to equate both vars to constants, we're done, and
2391 : : * we can throw away the full-join clause as redundant. Moreover, we
2392 : : * can remove the COALESCE entry from the EC, since the added
2393 : : * restrictions ensure it will always have the expected value. (We
2394 : : * don't bother trying to update ec_relids or ec_sources.)
2395 : : */
2396 [ + - + - ]: 3 : if (matchleft && matchright)
2397 : : {
1270 drowley@postgresql.o 2398 : 3 : cur_ec->ec_members = list_delete_nth_cell(cur_ec->ec_members, coal_idx);
5940 tgl@sss.pgh.pa.us 2399 : 3 : return true;
2400 : : }
2401 : :
2402 : : /*
2403 : : * Otherwise, fall out of the search loop, since we know the COALESCE
2404 : : * appears in at most one EC (XXX might stop being true if we allow
2405 : : * stripping of coercions above?)
2406 : : */
6294 tgl@sss.pgh.pa.us 2407 :UBC 0 : break;
2408 : : }
2409 : :
5940 tgl@sss.pgh.pa.us 2410 :CBC 608 : return false; /* failed to make any deduction */
2411 : : }
2412 : :
2413 : : /*
2414 : : * find_join_domain
2415 : : * Find the highest JoinDomain enclosed within the given relid set.
2416 : : *
2417 : : * (We could avoid this search at the cost of complicating APIs elsewhere,
2418 : : * which doesn't seem worth it.)
2419 : : */
2420 : : static JoinDomain *
440 2421 : 828 : find_join_domain(PlannerInfo *root, Relids relids)
2422 : : {
2423 : : ListCell *lc;
2424 : :
2425 [ + - + - : 1701 : foreach(lc, root->join_domains)
+ - ]
2426 : : {
2427 : 1701 : JoinDomain *jdomain = (JoinDomain *) lfirst(lc);
2428 : :
2429 [ + + ]: 1701 : if (bms_is_subset(jdomain->jd_relids, relids))
2430 : 828 : return jdomain;
2431 : : }
440 tgl@sss.pgh.pa.us 2432 [ # # ]:UBC 0 : elog(ERROR, "failed to find appropriate JoinDomain");
2433 : : return NULL; /* keep compiler quiet */
2434 : : }
2435 : :
2436 : :
2437 : : /*
2438 : : * exprs_known_equal
2439 : : * Detect whether two expressions are known equal due to equivalence
2440 : : * relationships.
2441 : : *
2442 : : * Actually, this only shows that the expressions are equal according
2443 : : * to some opfamily's notion of equality --- but we only use it for
2444 : : * selectivity estimation, so a fuzzy idea of equality is OK.
2445 : : *
2446 : : * Note: does not bother to check for "equal(item1, item2)"; caller must
2447 : : * check that case if it's possible to pass identical items.
2448 : : */
2449 : : bool
6294 tgl@sss.pgh.pa.us 2450 :CBC 1618 : exprs_known_equal(PlannerInfo *root, Node *item1, Node *item2)
2451 : : {
2452 : : ListCell *lc1;
2453 : :
2454 [ + + + + : 9864 : foreach(lc1, root->eq_classes)
+ + ]
2455 : : {
2456 : 8306 : EquivalenceClass *ec = (EquivalenceClass *) lfirst(lc1);
2457 : 8306 : bool item1member = false;
2458 : 8306 : bool item2member = false;
2459 : : ListCell *lc2;
2460 : :
2461 : : /* Never match to a volatile EC */
2462 [ - + ]: 8306 : if (ec->ec_has_volatile)
6294 tgl@sss.pgh.pa.us 2463 :UBC 0 : continue;
2464 : :
6294 tgl@sss.pgh.pa.us 2465 [ + + + + :CBC 26542 : foreach(lc2, ec->ec_members)
+ + ]
2466 : : {
2467 : 18296 : EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
2468 : :
4412 2469 [ + + ]: 18296 : if (em->em_is_child)
2470 : 6378 : continue; /* ignore children here */
6294 2471 [ + + ]: 11918 : if (equal(item1, em->em_expr))
2472 : 808 : item1member = true;
2473 [ + + ]: 11110 : else if (equal(item2, em->em_expr))
2474 : 880 : item2member = true;
2475 : : /* Exit as soon as equality is proven */
2476 [ + + + + ]: 11918 : if (item1member && item2member)
2477 : 60 : return true;
2478 : : }
2479 : : }
2480 : 1558 : return false;
2481 : : }
2482 : :
2483 : :
2484 : : /*
2485 : : * match_eclasses_to_foreign_key_col
2486 : : * See whether a foreign key column match is proven by any eclass.
2487 : : *
2488 : : * If the referenced and referencing Vars of the fkey's colno'th column are
2489 : : * known equal due to any eclass, return that eclass; otherwise return NULL.
2490 : : * (In principle there might be more than one matching eclass if multiple
2491 : : * collations are involved, but since collation doesn't matter for equality,
2492 : : * we ignore that fine point here.) This is much like exprs_known_equal,
2493 : : * except that we insist on the comparison operator matching the eclass, so
2494 : : * that the result is definite not approximate.
2495 : : *
2496 : : * On success, we also set fkinfo->eclass[colno] to the matching eclass,
2497 : : * and set fkinfo->fk_eclass_member[colno] to the eclass member for the
2498 : : * referencing Var.
2499 : : */
2500 : : EquivalenceClass *
2857 2501 : 1081 : match_eclasses_to_foreign_key_col(PlannerInfo *root,
2502 : : ForeignKeyOptInfo *fkinfo,
2503 : : int colno)
2504 : : {
2505 : 1081 : Index var1varno = fkinfo->con_relid;
2506 : 1081 : AttrNumber var1attno = fkinfo->conkey[colno];
2507 : 1081 : Index var2varno = fkinfo->ref_relid;
2508 : 1081 : AttrNumber var2attno = fkinfo->confkey[colno];
2509 : 1081 : Oid eqop = fkinfo->conpfeqop[colno];
1729 drowley@postgresql.o 2510 : 1081 : RelOptInfo *rel1 = root->simple_rel_array[var1varno];
2511 : 1081 : RelOptInfo *rel2 = root->simple_rel_array[var2varno];
2489 tgl@sss.pgh.pa.us 2512 : 1081 : List *opfamilies = NIL; /* compute only if needed */
2513 : : Bitmapset *matching_ecs;
2514 : : int i;
2515 : :
2516 : : /* Consider only eclasses mentioning both relations */
1729 drowley@postgresql.o 2517 [ - + ]: 1081 : Assert(root->ec_merging_done);
2518 [ - + - - ]: 1081 : Assert(IS_SIMPLE_REL(rel1));
2519 [ - + - - ]: 1081 : Assert(IS_SIMPLE_REL(rel2));
2520 : 1081 : matching_ecs = bms_intersect(rel1->eclass_indexes,
2521 : 1081 : rel2->eclass_indexes);
2522 : :
2523 : 1081 : i = -1;
2524 [ + + ]: 1129 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
2525 : : {
2526 : 219 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes,
2527 : : i);
1264 tgl@sss.pgh.pa.us 2528 : 219 : EquivalenceMember *item1_em = NULL;
2529 : 219 : EquivalenceMember *item2_em = NULL;
2530 : : ListCell *lc2;
2531 : :
2532 : : /* Never match to a volatile EC */
2857 2533 [ - + ]: 219 : if (ec->ec_has_volatile)
2857 tgl@sss.pgh.pa.us 2534 :UBC 0 : continue;
2535 : : /* Note: it seems okay to match to "broken" eclasses here */
2536 : :
2857 tgl@sss.pgh.pa.us 2537 [ + - + + :CBC 537 : foreach(lc2, ec->ec_members)
+ + ]
2538 : : {
2539 : 489 : EquivalenceMember *em = (EquivalenceMember *) lfirst(lc2);
2540 : : Var *var;
2541 : :
2542 [ - + ]: 489 : if (em->em_is_child)
2857 tgl@sss.pgh.pa.us 2543 :UBC 0 : continue; /* ignore children here */
2544 : :
2545 : : /* EM must be a Var, possibly with RelabelType */
2857 tgl@sss.pgh.pa.us 2546 :CBC 489 : var = (Var *) em->em_expr;
2547 [ + - - + ]: 489 : while (var && IsA(var, RelabelType))
2857 tgl@sss.pgh.pa.us 2548 :UBC 0 : var = (Var *) ((RelabelType *) var)->arg;
2857 tgl@sss.pgh.pa.us 2549 [ + - + + ]:CBC 489 : if (!(var && IsA(var, Var)))
2550 : 3 : continue;
2551 : :
2552 : : /* Match? */
2553 [ + + + + ]: 486 : if (var->varno == var1varno && var->varattno == var1attno)
1264 2554 : 171 : item1_em = em;
2857 2555 [ + + + + ]: 315 : else if (var->varno == var2varno && var->varattno == var2attno)
1264 2556 : 171 : item2_em = em;
2557 : :
2558 : : /* Have we found both PK and FK column in this EC? */
2559 [ + + + + ]: 486 : if (item1_em && item2_em)
2560 : : {
2561 : : /*
2562 : : * Succeed if eqop matches EC's opfamilies. We could test
2563 : : * this before scanning the members, but it's probably cheaper
2564 : : * to test for member matches first.
2565 : : */
2857 2566 [ + - ]: 171 : if (opfamilies == NIL) /* compute if we didn't already */
2567 : 171 : opfamilies = get_mergejoin_opfamilies(eqop);
2568 [ + - ]: 171 : if (equal(opfamilies, ec->ec_opfamilies))
2569 : : {
1264 2570 : 171 : fkinfo->eclass[colno] = ec;
2571 : 171 : fkinfo->fk_eclass_member[colno] = item2_em;
2857 2572 : 171 : return ec;
2573 : : }
2574 : : /* Otherwise, done with this EC, move on to the next */
2857 tgl@sss.pgh.pa.us 2575 :UBC 0 : break;
2576 : : }
2577 : : }
2578 : : }
2857 tgl@sss.pgh.pa.us 2579 :CBC 910 : return NULL;
2580 : : }
2581 : :
2582 : : /*
2583 : : * find_derived_clause_for_ec_member
2584 : : * Search for a previously-derived clause mentioning the given EM.
2585 : : *
2586 : : * The eclass should be an ec_has_const EC, of which the EM is a non-const
2587 : : * member. This should ensure there is just one derived clause mentioning
2588 : : * the EM (and equating it to a constant).
2589 : : * Returns NULL if no such clause can be found.
2590 : : */
2591 : : RestrictInfo *
1264 2592 : 3 : find_derived_clause_for_ec_member(EquivalenceClass *ec,
2593 : : EquivalenceMember *em)
2594 : : {
2595 : : ListCell *lc;
2596 : :
2597 [ - + ]: 3 : Assert(ec->ec_has_const);
2598 [ - + ]: 3 : Assert(!em->em_is_const);
2599 [ + - + - : 3 : foreach(lc, ec->ec_derives)
+ - ]
2600 : : {
2601 : 3 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2602 : :
2603 : : /*
2604 : : * generate_base_implied_equalities_const will have put non-const
2605 : : * members on the left side of derived clauses.
2606 : : */
2607 [ + - ]: 3 : if (rinfo->left_em == em)
2608 : 3 : return rinfo;
2609 : : }
1264 tgl@sss.pgh.pa.us 2610 :UBC 0 : return NULL;
2611 : : }
2612 : :
2613 : :
2614 : : /*
2615 : : * add_child_rel_equivalences
2616 : : * Search for EC members that reference the root parent of child_rel, and
2617 : : * add transformed members referencing the child_rel.
2618 : : *
2619 : : * Note that this function won't be called at all unless we have at least some
2620 : : * reason to believe that the EC members it generates will be useful.
2621 : : *
2622 : : * parent_rel and child_rel could be derived from appinfo, but since the
2623 : : * caller has already computed them, we might as well just pass them in.
2624 : : *
2625 : : * The passed-in AppendRelInfo is not used when the parent_rel is not a
2626 : : * top-level baserel, since it shows the mapping from the parent_rel but
2627 : : * we need to translate EC expressions that refer to the top-level parent.
2628 : : * Using it is faster than using adjust_appendrel_attrs_multilevel(), though,
2629 : : * so we prefer it when we can.
2630 : : */
2631 : : void
6294 tgl@sss.pgh.pa.us 2632 :CBC 10809 : add_child_rel_equivalences(PlannerInfo *root,
2633 : : AppendRelInfo *appinfo,
2634 : : RelOptInfo *parent_rel,
2635 : : RelOptInfo *child_rel)
2636 : : {
1622 2637 : 10809 : Relids top_parent_relids = child_rel->top_parent_relids;
2638 : 10809 : Relids child_relids = child_rel->relids;
2639 : : int i;
2640 : :
2641 : : /*
2642 : : * EC merging should be complete already, so we can use the parent rel's
2643 : : * eclass_indexes to avoid searching all of root->eq_classes.
2644 : : */
1729 drowley@postgresql.o 2645 [ - + ]: 10809 : Assert(root->ec_merging_done);
2646 [ + + - + ]: 10809 : Assert(IS_SIMPLE_REL(parent_rel));
2647 : :
2648 : 10809 : i = -1;
2649 [ + + ]: 32044 : while ((i = bms_next_member(parent_rel->eclass_indexes, i)) >= 0)
2650 : : {
2651 : 21235 : EquivalenceClass *cur_ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
2652 : : int num_members;
2653 : :
2654 : : /*
2655 : : * If this EC contains a volatile expression, then generating child
2656 : : * EMs would be downright dangerous, so skip it. We rely on a
2657 : : * volatile EC having only one EM.
2658 : : */
4426 tgl@sss.pgh.pa.us 2659 [ - + ]: 21235 : if (cur_ec->ec_has_volatile)
6294 tgl@sss.pgh.pa.us 2660 :UBC 0 : continue;
2661 : :
2662 : : /* Sanity check eclass_indexes only contain ECs for parent_rel */
1622 tgl@sss.pgh.pa.us 2663 [ - + ]:CBC 21235 : Assert(bms_is_subset(top_parent_relids, cur_ec->ec_relids));
2664 : :
2665 : : /*
2666 : : * We don't use foreach() here because there's no point in scanning
2667 : : * newly-added child members, so we can stop after the last
2668 : : * pre-existing EC member.
2669 : : */
1735 2670 : 21235 : num_members = list_length(cur_ec->ec_members);
2671 [ + + ]: 99468 : for (int pos = 0; pos < num_members; pos++)
2672 : : {
2673 : 78233 : EquivalenceMember *cur_em = (EquivalenceMember *) list_nth(cur_ec->ec_members, pos);
2674 : :
3670 2675 [ + + ]: 78233 : if (cur_em->em_is_const)
2676 : 1584 : continue; /* ignore consts here */
2677 : :
2678 : : /*
2679 : : * We consider only original EC members here, not
2680 : : * already-transformed child members. Otherwise, if some original
2681 : : * member expression references more than one appendrel, we'd get
2682 : : * an O(N^2) explosion of useless derived expressions for
2683 : : * combinations of children. (But add_child_join_rel_equivalences
2684 : : * may add targeted combinations for partitionwise-join purposes.)
2685 : : */
1767 2686 [ + + ]: 76649 : if (cur_em->em_is_child)
2687 : 48640 : continue; /* ignore children here */
2688 : :
2689 : : /*
2690 : : * Consider only members that reference and can be computed at
2691 : : * child's topmost parent rel. In particular we want to exclude
2692 : : * parent-rel Vars that have nonempty varnullingrels. Translating
2693 : : * those might fail, if the transformed expression wouldn't be a
2694 : : * simple Var; and in any case it wouldn't produce a member that
2695 : : * has any use in creating plans for the child rel.
2696 : : */
440 2697 [ + + ]: 28009 : if (bms_is_subset(cur_em->em_relids, top_parent_relids) &&
2698 [ + - ]: 20017 : !bms_is_empty(cur_em->em_relids))
2699 : : {
2700 : : /* OK, generate transformed child version */
2701 : : Expr *child_expr;
2702 : : Relids new_relids;
2703 : :
1767 2704 [ + + ]: 20017 : if (parent_rel->reloptkind == RELOPT_BASEREL)
2705 : : {
2706 : : /* Simple single-level transformation */
2707 : : child_expr = (Expr *)
2708 : 16252 : adjust_appendrel_attrs(root,
2709 : 16252 : (Node *) cur_em->em_expr,
2710 : : 1, &appinfo);
2711 : : }
2712 : : else
2713 : : {
2714 : : /* Must do multi-level transformation */
2715 : : child_expr = (Expr *)
2716 : 3765 : adjust_appendrel_attrs_multilevel(root,
2717 : 3765 : (Node *) cur_em->em_expr,
2718 : : child_rel,
605 2719 : 3765 : child_rel->top_parent);
2720 : : }
2721 : :
2722 : : /*
2723 : : * Transform em_relids to match. Note we do *not* do
2724 : : * pull_varnos(child_expr) here, as for example the
2725 : : * transformation might have substituted a constant, but we
2726 : : * don't want the child member to be marked as constant.
2727 : : */
4378 2728 : 20017 : new_relids = bms_difference(cur_em->em_relids,
2729 : : top_parent_relids);
1622 2730 : 20017 : new_relids = bms_add_members(new_relids, child_relids);
2731 : :
440 2732 : 20017 : (void) add_eq_member(cur_ec, child_expr, new_relids,
2733 : : cur_em->em_jdomain,
2734 : : cur_em, cur_em->em_datatype);
2735 : :
2736 : : /* Record this EC index for the child rel */
1729 drowley@postgresql.o 2737 : 20017 : child_rel->eclass_indexes = bms_add_member(child_rel->eclass_indexes, i);
2738 : : }
2739 : : }
2740 : : }
6294 tgl@sss.pgh.pa.us 2741 : 10809 : }
2742 : :
2743 : : /*
2744 : : * add_child_join_rel_equivalences
2745 : : * Like add_child_rel_equivalences(), but for joinrels
2746 : : *
2747 : : * Here we find the ECs relevant to the top parent joinrel and add transformed
2748 : : * member expressions that refer to this child joinrel.
2749 : : *
2750 : : * Note that this function won't be called at all unless we have at least some
2751 : : * reason to believe that the EC members it generates will be useful.
2752 : : */
2753 : : void
1622 2754 : 2090 : add_child_join_rel_equivalences(PlannerInfo *root,
2755 : : int nappinfos, AppendRelInfo **appinfos,
2756 : : RelOptInfo *parent_joinrel,
2757 : : RelOptInfo *child_joinrel)
2758 : : {
2759 : 2090 : Relids top_parent_relids = child_joinrel->top_parent_relids;
2760 : 2090 : Relids child_relids = child_joinrel->relids;
2761 : : Bitmapset *matching_ecs;
2762 : : MemoryContext oldcontext;
2763 : : int i;
2764 : :
2765 [ + - + - : 2090 : Assert(IS_JOIN_REL(child_joinrel) && IS_JOIN_REL(parent_joinrel));
+ + - + ]
2766 : :
2767 : : /* We need consider only ECs that mention the parent joinrel */
2768 : 2090 : matching_ecs = get_eclass_indexes_for_relids(root, top_parent_relids);
2769 : :
2770 : : /*
2771 : : * If we're being called during GEQO join planning, we still have to
2772 : : * create any new EC members in the main planner context, to avoid having
2773 : : * a corrupt EC data structure after the GEQO context is reset. This is
2774 : : * problematic since we'll leak memory across repeated GEQO cycles. For
2775 : : * now, though, bloat is better than crash. If it becomes a real issue
2776 : : * we'll have to do something to avoid generating duplicate EC members.
2777 : : */
1286 2778 : 2090 : oldcontext = MemoryContextSwitchTo(root->planner_cxt);
2779 : :
1622 2780 : 2090 : i = -1;
2781 [ + + ]: 9843 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
2782 : : {
2783 : 7753 : EquivalenceClass *cur_ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
2784 : : int num_members;
2785 : :
2786 : : /*
2787 : : * If this EC contains a volatile expression, then generating child
2788 : : * EMs would be downright dangerous, so skip it. We rely on a
2789 : : * volatile EC having only one EM.
2790 : : */
2791 [ - + ]: 7753 : if (cur_ec->ec_has_volatile)
1622 tgl@sss.pgh.pa.us 2792 :UBC 0 : continue;
2793 : :
2794 : : /* Sanity check on get_eclass_indexes_for_relids result */
1622 tgl@sss.pgh.pa.us 2795 [ - + ]:CBC 7753 : Assert(bms_overlap(top_parent_relids, cur_ec->ec_relids));
2796 : :
2797 : : /*
2798 : : * We don't use foreach() here because there's no point in scanning
2799 : : * newly-added child members, so we can stop after the last
2800 : : * pre-existing EC member.
2801 : : */
2802 : 7753 : num_members = list_length(cur_ec->ec_members);
2803 [ + + ]: 51721 : for (int pos = 0; pos < num_members; pos++)
2804 : : {
2805 : 43968 : EquivalenceMember *cur_em = (EquivalenceMember *) list_nth(cur_ec->ec_members, pos);
2806 : :
2807 [ + + ]: 43968 : if (cur_em->em_is_const)
2808 : 1116 : continue; /* ignore consts here */
2809 : :
2810 : : /*
2811 : : * We consider only original EC members here, not
2812 : : * already-transformed child members.
2813 : : */
2814 [ + + ]: 42852 : if (cur_em->em_is_child)
2815 : 33037 : continue; /* ignore children here */
2816 : :
2817 : : /*
2818 : : * We may ignore expressions that reference a single baserel,
2819 : : * because add_child_rel_equivalences should have handled them.
2820 : : */
2821 [ + + ]: 9815 : if (bms_membership(cur_em->em_relids) != BMS_MULTIPLE)
2822 : 8504 : continue;
2823 : :
2824 : : /* Does this member reference child's topmost parent rel? */
2825 [ + - ]: 1311 : if (bms_overlap(cur_em->em_relids, top_parent_relids))
2826 : : {
2827 : : /* Yes, generate transformed child version */
2828 : : Expr *child_expr;
2829 : : Relids new_relids;
2830 : :
2831 [ + + ]: 1311 : if (parent_joinrel->reloptkind == RELOPT_JOINREL)
2832 : : {
2833 : : /* Simple single-level transformation */
2834 : : child_expr = (Expr *)
2835 : 1263 : adjust_appendrel_attrs(root,
2836 : 1263 : (Node *) cur_em->em_expr,
2837 : : nappinfos, appinfos);
2838 : : }
2839 : : else
2840 : : {
2841 : : /* Must do multi-level transformation */
2842 [ - + ]: 48 : Assert(parent_joinrel->reloptkind == RELOPT_OTHER_JOINREL);
2843 : : child_expr = (Expr *)
2844 : 48 : adjust_appendrel_attrs_multilevel(root,
2845 : 48 : (Node *) cur_em->em_expr,
2846 : : child_joinrel,
605 2847 : 48 : child_joinrel->top_parent);
2848 : : }
2849 : :
2850 : : /*
2851 : : * Transform em_relids to match. Note we do *not* do
2852 : : * pull_varnos(child_expr) here, as for example the
2853 : : * transformation might have substituted a constant, but we
2854 : : * don't want the child member to be marked as constant.
2855 : : */
1622 2856 : 1311 : new_relids = bms_difference(cur_em->em_relids,
2857 : : top_parent_relids);
2858 : 1311 : new_relids = bms_add_members(new_relids, child_relids);
2859 : :
440 2860 : 1311 : (void) add_eq_member(cur_ec, child_expr, new_relids,
2861 : : cur_em->em_jdomain,
2862 : : cur_em, cur_em->em_datatype);
2863 : : }
2864 : : }
2865 : : }
2866 : :
1286 2867 : 2090 : MemoryContextSwitchTo(oldcontext);
1622 2868 : 2090 : }
2869 : :
2870 : : /*
2871 : : * add_setop_child_rel_equivalences
2872 : : * Add equivalence members for each non-resjunk target in 'child_tlist'
2873 : : * to the EquivalenceClass in the corresponding setop_pathkey's pk_class.
2874 : : *
2875 : : * 'root' is the PlannerInfo belonging to the top-level set operation.
2876 : : * 'child_rel' is the RelOptInfo of the child relation we're adding
2877 : : * EquivalenceMembers for.
2878 : : * 'child_tlist' is the target list for the setop child relation. The target
2879 : : * list expressions are what we add as EquivalenceMembers.
2880 : : * 'setop_pathkeys' is a list of PathKeys which must contain an entry for each
2881 : : * non-resjunk target in 'child_tlist'.
2882 : : */
2883 : : void
20 drowley@postgresql.o 2884 :GNC 4682 : add_setop_child_rel_equivalences(PlannerInfo *root, RelOptInfo *child_rel,
2885 : : List *child_tlist, List *setop_pathkeys)
2886 : : {
2887 : : ListCell *lc;
2888 : 4682 : ListCell *lc2 = list_head(setop_pathkeys);
2889 : :
2890 [ + - + + : 18628 : foreach(lc, child_tlist)
+ + ]
2891 : : {
2892 : 13946 : TargetEntry *tle = lfirst_node(TargetEntry, lc);
2893 : : EquivalenceMember *parent_em;
2894 : : PathKey *pk;
2895 : :
2896 [ - + ]: 13946 : if (tle->resjunk)
20 drowley@postgresql.o 2897 :UNC 0 : continue;
2898 : :
20 drowley@postgresql.o 2899 [ - + ]:GNC 13946 : if (lc2 == NULL)
20 drowley@postgresql.o 2900 [ # # ]:UNC 0 : elog(ERROR, "too few pathkeys for set operation");
2901 : :
20 drowley@postgresql.o 2902 :GNC 13946 : pk = lfirst_node(PathKey, lc2);
2903 : 13946 : parent_em = linitial(pk->pk_eclass->ec_members);
2904 : :
2905 : : /*
2906 : : * We can safely pass the parent member as the first member in the
2907 : : * ec_members list as this is added first in generate_union_paths,
2908 : : * likewise, the JoinDomain can be that of the initial member of the
2909 : : * Pathkey's EquivalenceClass.
2910 : : */
2911 : 13946 : add_eq_member(pk->pk_eclass,
2912 : : tle->expr,
2913 : : child_rel->relids,
2914 : : parent_em->em_jdomain,
2915 : : parent_em,
2916 : 13946 : exprType((Node *) tle->expr));
2917 : :
2918 : 13946 : lc2 = lnext(setop_pathkeys, lc2);
2919 : : }
2920 : :
2921 : : /*
2922 : : * transformSetOperationStmt() ensures that the targetlist never contains
2923 : : * any resjunk columns, so all eclasses that exist in 'root' must have
2924 : : * received a new member in the loop above. Add them to the child_rel's
2925 : : * eclass_indexes.
2926 : : */
2927 : 4682 : child_rel->eclass_indexes = bms_add_range(child_rel->eclass_indexes, 0,
2928 : 4682 : list_length(root->eq_classes) - 1);
2929 : 4682 : }
2930 : :
2931 : :
2932 : : /*
2933 : : * generate_implied_equalities_for_column
2934 : : * Create EC-derived joinclauses usable with a specific column.
2935 : : *
2936 : : * This is used by indxpath.c to extract potentially indexable joinclauses
2937 : : * from ECs, and can be used by foreign data wrappers for similar purposes.
2938 : : * We assume that only expressions in Vars of a single table are of interest,
2939 : : * but the caller provides a callback function to identify exactly which
2940 : : * such expressions it would like to know about.
2941 : : *
2942 : : * We assume that any given table/index column could appear in only one EC.
2943 : : * (This should be true in all but the most pathological cases, and if it
2944 : : * isn't, we stop on the first match anyway.) Therefore, what we return
2945 : : * is a redundant list of clauses equating the table/index column to each of
2946 : : * the other-relation values it is known to be equal to. Any one of
2947 : : * these clauses can be used to create a parameterized path, and there
2948 : : * is no value in using more than one. (But it *is* worthwhile to create
2949 : : * a separate parameterized path for each one, since that leads to different
2950 : : * join orders.)
2951 : : *
2952 : : * The caller can pass a Relids set of rels we aren't interested in joining
2953 : : * to, so as to save the work of creating useless clauses.
2954 : : */
2955 : : List *
4042 tgl@sss.pgh.pa.us 2956 :CBC 232442 : generate_implied_equalities_for_column(PlannerInfo *root,
2957 : : RelOptInfo *rel,
2958 : : ec_matches_callback_type callback,
2959 : : void *callback_arg,
2960 : : Relids prohibited_rels)
2961 : : {
6294 2962 : 232442 : List *result = NIL;
2963 : 232442 : bool is_child_rel = (rel->reloptkind == RELOPT_OTHER_MEMBER_REL);
2964 : : Relids parent_relids;
2965 : : int i;
2966 : :
2967 : : /* Should be OK to rely on eclass_indexes */
1729 drowley@postgresql.o 2968 [ - + ]: 232442 : Assert(root->ec_merging_done);
2969 : :
2970 : : /* Indexes are available only on base or "other" member relations. */
2568 rhaas@postgresql.org 2971 [ + + - + ]: 232442 : Assert(IS_SIMPLE_REL(rel));
2972 : :
2973 : : /* If it's a child rel, we'll need to know what its parent(s) are */
4461 tgl@sss.pgh.pa.us 2974 [ + + ]: 232442 : if (is_child_rel)
3483 2975 : 5263 : parent_relids = find_childrel_parents(root, rel);
2976 : : else
2977 : 227179 : parent_relids = NULL; /* not used, but keep compiler quiet */
2978 : :
1729 drowley@postgresql.o 2979 : 232442 : i = -1;
2980 [ + + ]: 634145 : while ((i = bms_next_member(rel->eclass_indexes, i)) >= 0)
2981 : : {
2982 : 441036 : EquivalenceClass *cur_ec = (EquivalenceClass *) list_nth(root->eq_classes, i);
2983 : : EquivalenceMember *cur_em;
2984 : : ListCell *lc2;
2985 : :
2986 : : /* Sanity check eclass_indexes only contain ECs for rel */
2987 [ + + - + ]: 441036 : Assert(is_child_rel || bms_is_subset(rel->relids, cur_ec->ec_relids));
2988 : :
2989 : : /*
2990 : : * Won't generate joinclauses if const or single-member (the latter
2991 : : * test covers the volatile case too)
2992 : : */
6294 tgl@sss.pgh.pa.us 2993 [ + + + + ]: 441036 : if (cur_ec->ec_has_const || list_length(cur_ec->ec_members) <= 1)
2994 : 243028 : continue;
2995 : :
2996 : : /*
2997 : : * Scan members, looking for a match to the target column. Note that
2998 : : * child EC members are considered, but only when they belong to the
2999 : : * target relation. (Unlike regular members, the same expression
3000 : : * could be a child member of more than one EC. Therefore, it's
3001 : : * potentially order-dependent which EC a child relation's target
3002 : : * column gets matched to. This is annoying but it only happens in
3003 : : * corner cases, so for now we live with just reporting the first
3004 : : * match. See also get_eclass_for_sort_expr.)
3005 : : */
4461 3006 : 198008 : cur_em = NULL;
3007 [ + - + + : 592260 : foreach(lc2, cur_ec->ec_members)
+ + ]
3008 : : {
3009 : 433809 : cur_em = (EquivalenceMember *) lfirst(lc2);
3010 [ + + + + ]: 632256 : if (bms_equal(cur_em->em_relids, rel->relids) &&
4042 3011 : 198447 : callback(root, rel, cur_ec, cur_em, callback_arg))
4461 3012 : 39557 : break;
3013 : 394252 : cur_em = NULL;
3014 : : }
3015 : :
3016 [ + + ]: 198008 : if (!cur_em)
6294 3017 : 158451 : continue;
3018 : :
3019 : : /*
3020 : : * Found our match. Scan the other EC members and attempt to generate
3021 : : * joinclauses.
3022 : : */
3023 [ + - + + : 130799 : foreach(lc2, cur_ec->ec_members)
+ + ]
3024 : : {
4461 3025 : 91242 : EquivalenceMember *other_em = (EquivalenceMember *) lfirst(lc2);
3026 : : Oid eq_op;
3027 : : RestrictInfo *rinfo;
3028 : :
4412 3029 [ + + ]: 91242 : if (other_em->em_is_child)
3030 : 10234 : continue; /* ignore children here */
3031 : :
3032 : : /* Make sure it'll be a join to a different rel */
4461 3033 [ + + + + ]: 123891 : if (other_em == cur_em ||
3034 : 42883 : bms_overlap(other_em->em_relids, rel->relids))
6294 3035 : 38291 : continue;
3036 : :
3037 : : /* Forget it if caller doesn't want joins to this rel */
4245 3038 [ + + ]: 42717 : if (bms_overlap(other_em->em_relids, prohibited_rels))
3039 : 69 : continue;
3040 : :
3041 : : /*
3042 : : * Also, if this is a child rel, avoid generating a useless join
3043 : : * to its parent rel(s).
3044 : : */
4461 3045 [ + + + + ]: 45741 : if (is_child_rel &&
3483 3046 : 3093 : bms_overlap(parent_relids, other_em->em_relids))
4461 3047 : 1456 : continue;
3048 : :
3049 : 41192 : eq_op = select_equality_operator(cur_ec,
3050 : : cur_em->em_datatype,
3051 : : other_em->em_datatype);
3052 [ - + ]: 41192 : if (!OidIsValid(eq_op))
4461 tgl@sss.pgh.pa.us 3053 :UBC 0 : continue;
3054 : :
3055 : : /* set parent_ec to mark as redundant with other joinclauses */
4461 tgl@sss.pgh.pa.us 3056 :CBC 41192 : rinfo = create_join_clause(root, cur_ec, eq_op,
3057 : : cur_em, other_em,
3058 : : cur_ec);
3059 : :
3060 : 41192 : result = lappend(result, rinfo);
3061 : : }
3062 : :
3063 : : /*
3064 : : * If somehow we failed to create any join clauses, we might as well
3065 : : * keep scanning the ECs for another match. But if we did make any,
3066 : : * we're done, because we don't want to return non-redundant clauses.
3067 : : */
3068 [ + + ]: 39557 : if (result)
3069 : 39333 : break;
3070 : : }
3071 : :
6294 3072 : 232442 : return result;
3073 : : }
3074 : :
3075 : : /*
3076 : : * have_relevant_eclass_joinclause
3077 : : * Detect whether there is an EquivalenceClass that could produce
3078 : : * a joinclause involving the two given relations.
3079 : : *
3080 : : * This is essentially a very cut-down version of
3081 : : * generate_join_implied_equalities(). Note it's OK to occasionally say "yes"
3082 : : * incorrectly. Hence we don't bother with details like whether the lack of a
3083 : : * cross-type operator might prevent the clause from actually being generated.
3084 : : * False negatives are not always fatal either: they will discourage, but not
3085 : : * completely prevent, investigation of particular join pathways.
3086 : : */
3087 : : bool
3088 : 70924 : have_relevant_eclass_joinclause(PlannerInfo *root,
3089 : : RelOptInfo *rel1, RelOptInfo *rel2)
3090 : : {
3091 : : Bitmapset *matching_ecs;
3092 : : int i;
3093 : :
3094 : : /*
3095 : : * Examine only eclasses mentioning both rel1 and rel2.
3096 : : *
3097 : : * Note that we do not consider the possibility of an eclass generating
3098 : : * "join" clauses that mention just one of the rels plus an outer join
3099 : : * that could be formed from them. Although such clauses must be
3100 : : * correctly enforced when we form the outer join, they don't seem like
3101 : : * sufficient reason to prioritize this join over other ones. The join
3102 : : * ordering rules will force the join to be made when necessary.
3103 : : */
1729 drowley@postgresql.o 3104 : 70924 : matching_ecs = get_common_eclass_indexes(root, rel1->relids,
3105 : : rel2->relids);
3106 : :
3107 : 70924 : i = -1;
3108 [ + + ]: 70924 : while ((i = bms_next_member(matching_ecs, i)) >= 0)
3109 : : {
3110 : 59041 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes,
3111 : : i);
3112 : :
3113 : : /*
3114 : : * Sanity check that get_common_eclass_indexes gave only ECs
3115 : : * containing both rels.
3116 : : */
3117 [ - + ]: 59041 : Assert(bms_overlap(rel1->relids, ec->ec_relids));
3118 [ - + ]: 59041 : Assert(bms_overlap(rel2->relids, ec->ec_relids));
3119 : :
3120 : : /*
3121 : : * Won't generate joinclauses if single-member (this test covers the
3122 : : * volatile case too)
3123 : : */
5940 tgl@sss.pgh.pa.us 3124 [ - + ]: 59041 : if (list_length(ec->ec_members) <= 1)
6294 tgl@sss.pgh.pa.us 3125 :UBC 0 : continue;
3126 : :
3127 : : /*
3128 : : * We do not need to examine the individual members of the EC, because
3129 : : * all that we care about is whether each rel overlaps the relids of
3130 : : * at least one member, and get_common_eclass_indexes() and the single
3131 : : * member check above are sufficient to prove that. (As with
3132 : : * have_relevant_joinclause(), it is not necessary that the EC be able
3133 : : * to form a joinclause relating exactly the two given rels, only that
3134 : : * it be able to form a joinclause mentioning both, and this will
3135 : : * surely be true if both of them overlap ec_relids.)
3136 : : *
3137 : : * Note we don't test ec_broken; if we did, we'd need a separate code
3138 : : * path to look through ec_sources. Checking the membership anyway is
3139 : : * OK as a possibly-overoptimistic heuristic.
3140 : : *
3141 : : * We don't test ec_has_const either, even though a const eclass won't
3142 : : * generate real join clauses. This is because if we had "WHERE a.x =
3143 : : * b.y and a.x = 42", it is worth considering a join between a and b,
3144 : : * since the join result is likely to be small even though it'll end
3145 : : * up being an unqualified nestloop.
3146 : : */
3147 : :
1729 drowley@postgresql.o 3148 :CBC 59041 : return true;
3149 : : }
3150 : :
6294 tgl@sss.pgh.pa.us 3151 : 11883 : return false;
3152 : : }
3153 : :
3154 : :
3155 : : /*
3156 : : * has_relevant_eclass_joinclause
3157 : : * Detect whether there is an EquivalenceClass that could produce
3158 : : * a joinclause involving the given relation and anything else.
3159 : : *
3160 : : * This is the same as have_relevant_eclass_joinclause with the other rel
3161 : : * implicitly defined as "everything else in the query".
3162 : : */
3163 : : bool
3164 : 88581 : has_relevant_eclass_joinclause(PlannerInfo *root, RelOptInfo *rel1)
3165 : : {
3166 : : Bitmapset *matched_ecs;
3167 : : int i;
3168 : :
3169 : : /* Examine only eclasses mentioning rel1 */
1729 drowley@postgresql.o 3170 : 88581 : matched_ecs = get_eclass_indexes_for_relids(root, rel1->relids);
3171 : :
3172 : 88581 : i = -1;
3173 [ + + ]: 321551 : while ((i = bms_next_member(matched_ecs, i)) >= 0)
3174 : : {
3175 : 262037 : EquivalenceClass *ec = (EquivalenceClass *) list_nth(root->eq_classes,
3176 : : i);
3177 : :
3178 : : /*
3179 : : * Won't generate joinclauses if single-member (this test covers the
3180 : : * volatile case too)
3181 : : */
5940 tgl@sss.pgh.pa.us 3182 [ + + ]: 262037 : if (list_length(ec->ec_members) <= 1)
6294 3183 : 131551 : continue;
3184 : :
3185 : : /*
3186 : : * Per the comment in have_relevant_eclass_joinclause, it's sufficient
3187 : : * to find an EC that mentions both this rel and some other rel.
3188 : : */
1729 drowley@postgresql.o 3189 [ + + ]: 130486 : if (!bms_is_subset(ec->ec_relids, rel1->relids))
6294 tgl@sss.pgh.pa.us 3190 : 29067 : return true;
3191 : : }
3192 : :
3193 : 59514 : return false;
3194 : : }
3195 : :
3196 : :
3197 : : /*
3198 : : * eclass_useful_for_merging
3199 : : * Detect whether the EC could produce any mergejoinable join clauses
3200 : : * against the specified relation.
3201 : : *
3202 : : * This is just a heuristic test and doesn't have to be exact; it's better
3203 : : * to say "yes" incorrectly than "no". Hence we don't bother with details
3204 : : * like whether the lack of a cross-type operator might prevent the clause
3205 : : * from actually being generated.
3206 : : */
3207 : : bool
3174 3208 : 304321 : eclass_useful_for_merging(PlannerInfo *root,
3209 : : EquivalenceClass *eclass,
3210 : : RelOptInfo *rel)
3211 : : {
3212 : : Relids relids;
3213 : : ListCell *lc;
3214 : :
6294 3215 [ - + ]: 304321 : Assert(!eclass->ec_merged);
3216 : :
3217 : : /*
3218 : : * Won't generate joinclauses if const or single-member (the latter test
3219 : : * covers the volatile case too)
3220 : : */
3221 [ + + + + ]: 304321 : if (eclass->ec_has_const || list_length(eclass->ec_members) <= 1)
3222 : 29600 : return false;
3223 : :
3224 : : /*
3225 : : * Note we don't test ec_broken; if we did, we'd need a separate code path
3226 : : * to look through ec_sources. Checking the members anyway is OK as a
3227 : : * possibly-overoptimistic heuristic.
3228 : : */
3229 : :
3230 : : /* If specified rel is a child, we must consider the topmost parent rel */
2568 rhaas@postgresql.org 3231 [ + + + + : 274721 : if (IS_OTHER_REL(rel))
- + ]
3232 : : {
3233 [ - + ]: 5217 : Assert(!bms_is_empty(rel->top_parent_relids));
3234 : 5217 : relids = rel->top_parent_relids;
3235 : : }
3236 : : else
3174 tgl@sss.pgh.pa.us 3237 : 269504 : relids = rel->relids;
3238 : :
3239 : : /* If rel already includes all members of eclass, no point in searching */
3240 [ + + ]: 274721 : if (bms_is_subset(eclass->ec_relids, relids))
6294 3241 : 105674 : return false;
3242 : :
3243 : : /* To join, we need a member not in the given rel */
3244 [ + - + + : 262637 : foreach(lc, eclass->ec_members)
+ + ]
3245 : : {
3246 : 262376 : EquivalenceMember *cur_em = (EquivalenceMember *) lfirst(lc);
3247 : :
4412 3248 [ - + ]: 262376 : if (cur_em->em_is_child)
4412 tgl@sss.pgh.pa.us 3249 :UBC 0 : continue; /* ignore children here */
3250 : :
3174 tgl@sss.pgh.pa.us 3251 [ + + ]:CBC 262376 : if (!bms_overlap(cur_em->em_relids, relids))
6294 3252 : 168786 : return true;
3253 : : }
3254 : :
3255 : 261 : return false;
3256 : : }
3257 : :
3258 : :
3259 : : /*
3260 : : * is_redundant_derived_clause
3261 : : * Test whether rinfo is derived from same EC as any clause in clauselist;
3262 : : * if so, it can be presumed to represent a condition that's redundant
3263 : : * with that member of the list.
3264 : : */
3265 : : bool
4378 3266 : 36 : is_redundant_derived_clause(RestrictInfo *rinfo, List *clauselist)
3267 : : {
3268 : 36 : EquivalenceClass *parent_ec = rinfo->parent_ec;
3269 : : ListCell *lc;
3270 : :
3271 : : /* Fail if it's not a potentially-redundant clause from some EC */
3272 [ + - ]: 36 : if (parent_ec == NULL)
3273 : 36 : return false;
3274 : :
4378 tgl@sss.pgh.pa.us 3275 [ # # # # :UBC 0 : foreach(lc, clauselist)
# # ]
3276 : : {
3277 : 0 : RestrictInfo *otherrinfo = (RestrictInfo *) lfirst(lc);
3278 : :
3279 [ # # ]: 0 : if (otherrinfo->parent_ec == parent_ec)
3280 : 0 : return true;
3281 : : }
3282 : :
3283 : 0 : return false;
3284 : : }
3285 : :
3286 : : /*
3287 : : * is_redundant_with_indexclauses
3288 : : * Test whether rinfo is redundant with any clause in the IndexClause
3289 : : * list. Here, for convenience, we test both simple identity and
3290 : : * whether it is derived from the same EC as any member of the list.
3291 : : */
3292 : : bool
1891 tgl@sss.pgh.pa.us 3293 :CBC 576779 : is_redundant_with_indexclauses(RestrictInfo *rinfo, List *indexclauses)
3294 : : {
3295 : 576779 : EquivalenceClass *parent_ec = rinfo->parent_ec;
3296 : : ListCell *lc;
3297 : :
3298 [ + + + + : 795390 : foreach(lc, indexclauses)
+ + ]
3299 : : {
3300 : 595474 : IndexClause *iclause = lfirst_node(IndexClause, lc);
3301 : 595474 : RestrictInfo *otherrinfo = iclause->rinfo;
3302 : :
3303 : : /* If indexclause is lossy, it won't enforce the condition exactly */
3304 [ + + ]: 595474 : if (iclause->lossy)
3305 : 18094 : continue;
3306 : :
3307 : : /* Match if it's same clause (pointer equality should be enough) */
3308 [ + + ]: 577380 : if (rinfo == otherrinfo)
3309 : 376863 : return true;
3310 : : /* Match if derived from same EC */
3311 [ + + + + ]: 200661 : if (parent_ec && otherrinfo->parent_ec == parent_ec)
3312 : 144 : return true;
3313 : :
3314 : : /*
3315 : : * No need to look at the derived clauses in iclause->indexquals; they
3316 : : * couldn't match if the parent clause didn't.
3317 : : */
3318 : : }
3319 : :
3320 : 199916 : return false;
3321 : : }
3322 : :
3323 : : /*
3324 : : * get_eclass_indexes_for_relids
3325 : : * Build and return a Bitmapset containing the indexes into root's
3326 : : * eq_classes list for all eclasses that mention any of these relids
3327 : : */
3328 : : static Bitmapset *
1729 drowley@postgresql.o 3329 : 428096 : get_eclass_indexes_for_relids(PlannerInfo *root, Relids relids)
3330 : : {
3331 : 428096 : Bitmapset *ec_indexes = NULL;
3332 : 428096 : int i = -1;
3333 : :
3334 : : /* Should be OK to rely on eclass_indexes */
3335 [ - + ]: 428096 : Assert(root->ec_merging_done);
3336 : :
3337 [ + + ]: 1395182 : while ((i = bms_next_member(relids, i)) > 0)
3338 : : {
3339 : 967086 : RelOptInfo *rel = root->simple_rel_array[i];
3340 : :
440 tgl@sss.pgh.pa.us 3341 [ + + ]: 967086 : if (rel == NULL) /* must be an outer join */
3342 : : {
3343 [ - + ]: 152445 : Assert(bms_is_member(i, root->outer_join_rels));
3344 : 152445 : continue;
3345 : : }
3346 : :
1729 drowley@postgresql.o 3347 : 814641 : ec_indexes = bms_add_members(ec_indexes, rel->eclass_indexes);
3348 : : }
3349 : 428096 : return ec_indexes;
3350 : : }
3351 : :
3352 : : /*
3353 : : * get_common_eclass_indexes
3354 : : * Build and return a Bitmapset containing the indexes into root's
3355 : : * eq_classes list for all eclasses that mention rels in both
3356 : : * relids1 and relids2.
3357 : : */
3358 : : static Bitmapset *
3359 : 240108 : get_common_eclass_indexes(PlannerInfo *root, Relids relids1, Relids relids2)
3360 : : {
3361 : : Bitmapset *rel1ecs;
3362 : : Bitmapset *rel2ecs;
3363 : : int relid;
3364 : :
3365 : 240108 : rel1ecs = get_eclass_indexes_for_relids(root, relids1);
3366 : :
3367 : : /*
3368 : : * We can get away with just using the relation's eclass_indexes directly
3369 : : * when relids2 is a singleton set.
3370 : : */
3371 [ + + ]: 240108 : if (bms_get_singleton_member(relids2, &relid))
3372 : 189373 : rel2ecs = root->simple_rel_array[relid]->eclass_indexes;
3373 : : else
3374 : 50735 : rel2ecs = get_eclass_indexes_for_relids(root, relids2);
3375 : :
3376 : : /* Calculate and return the common EC indexes, recycling the left input. */
3377 : 240108 : return bms_int_members(rel1ecs, rel2ecs);
3378 : : }
|
