Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * selfuncs.c
4 : * Selectivity functions and index cost estimation functions for
5 : * standard operators and index access methods.
6 : *
7 : * Selectivity routines are registered in the pg_operator catalog
8 : * in the "oprrest" and "oprjoin" attributes.
9 : *
10 : * Index cost functions are located via the index AM's API struct,
11 : * which is obtained from the handler function registered in pg_am.
12 : *
13 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
14 : * Portions Copyright (c) 1994, Regents of the University of California
15 : *
16 : *
17 : * IDENTIFICATION
18 : * src/backend/utils/adt/selfuncs.c
19 : *
20 : *-------------------------------------------------------------------------
21 : */
22 :
23 : /*----------
24 : * Operator selectivity estimation functions are called to estimate the
25 : * selectivity of WHERE clauses whose top-level operator is their operator.
26 : * We divide the problem into two cases:
27 : * Restriction clause estimation: the clause involves vars of just
28 : * one relation.
29 : * Join clause estimation: the clause involves vars of multiple rels.
30 : * Join selectivity estimation is far more difficult and usually less accurate
31 : * than restriction estimation.
32 : *
33 : * When dealing with the inner scan of a nestloop join, we consider the
34 : * join's joinclauses as restriction clauses for the inner relation, and
35 : * treat vars of the outer relation as parameters (a/k/a constants of unknown
36 : * values). So, restriction estimators need to be able to accept an argument
37 : * telling which relation is to be treated as the variable.
38 : *
39 : * The call convention for a restriction estimator (oprrest function) is
40 : *
41 : * Selectivity oprrest (PlannerInfo *root,
42 : * Oid operator,
43 : * List *args,
44 : * int varRelid);
45 : *
46 : * root: general information about the query (rtable and RelOptInfo lists
47 : * are particularly important for the estimator).
48 : * operator: OID of the specific operator in question.
49 : * args: argument list from the operator clause.
50 : * varRelid: if not zero, the relid (rtable index) of the relation to
51 : * be treated as the variable relation. May be zero if the args list
52 : * is known to contain vars of only one relation.
53 : *
54 : * This is represented at the SQL level (in pg_proc) as
55 : *
56 : * float8 oprrest (internal, oid, internal, int4);
57 : *
58 : * The result is a selectivity, that is, a fraction (0 to 1) of the rows
59 : * of the relation that are expected to produce a TRUE result for the
60 : * given operator.
61 : *
62 : * The call convention for a join estimator (oprjoin function) is similar
63 : * except that varRelid is not needed, and instead join information is
64 : * supplied:
65 : *
66 : * Selectivity oprjoin (PlannerInfo *root,
67 : * Oid operator,
68 : * List *args,
69 : * JoinType jointype,
70 : * SpecialJoinInfo *sjinfo);
71 : *
72 : * float8 oprjoin (internal, oid, internal, int2, internal);
73 : *
74 : * (Before Postgres 8.4, join estimators had only the first four of these
75 : * parameters. That signature is still allowed, but deprecated.) The
76 : * relationship between jointype and sjinfo is explained in the comments for
77 : * clause_selectivity() --- the short version is that jointype is usually
78 : * best ignored in favor of examining sjinfo.
79 : *
80 : * Join selectivity for regular inner and outer joins is defined as the
81 : * fraction (0 to 1) of the cross product of the relations that is expected
82 : * to produce a TRUE result for the given operator. For both semi and anti
83 : * joins, however, the selectivity is defined as the fraction of the left-hand
84 : * side relation's rows that are expected to have a match (ie, at least one
85 : * row with a TRUE result) in the right-hand side.
86 : *
87 : * For both oprrest and oprjoin functions, the operator's input collation OID
88 : * (if any) is passed using the standard fmgr mechanism, so that the estimator
89 : * function can fetch it with PG_GET_COLLATION(). Note, however, that all
90 : * statistics in pg_statistic are currently built using the relevant column's
91 : * collation.
92 : *----------
93 : */
94 :
95 : #include "postgres.h"
96 :
97 : #include <ctype.h>
98 : #include <math.h>
99 :
100 : #include "access/brin.h"
101 : #include "access/brin_page.h"
102 : #include "access/gin.h"
103 : #include "access/table.h"
104 : #include "access/tableam.h"
105 : #include "access/visibilitymap.h"
106 : #include "catalog/pg_collation.h"
107 : #include "catalog/pg_operator.h"
108 : #include "catalog/pg_statistic.h"
109 : #include "catalog/pg_statistic_ext.h"
110 : #include "executor/nodeAgg.h"
111 : #include "miscadmin.h"
112 : #include "nodes/makefuncs.h"
113 : #include "nodes/nodeFuncs.h"
114 : #include "optimizer/clauses.h"
115 : #include "optimizer/cost.h"
116 : #include "optimizer/optimizer.h"
117 : #include "optimizer/pathnode.h"
118 : #include "optimizer/paths.h"
119 : #include "optimizer/plancat.h"
120 : #include "parser/parse_clause.h"
121 : #include "parser/parse_relation.h"
122 : #include "parser/parsetree.h"
123 : #include "rewrite/rewriteManip.h"
124 : #include "statistics/statistics.h"
125 : #include "storage/bufmgr.h"
126 : #include "utils/acl.h"
127 : #include "utils/array.h"
128 : #include "utils/builtins.h"
129 : #include "utils/date.h"
130 : #include "utils/datum.h"
131 : #include "utils/fmgroids.h"
132 : #include "utils/index_selfuncs.h"
133 : #include "utils/lsyscache.h"
134 : #include "utils/memutils.h"
135 : #include "utils/pg_locale.h"
136 : #include "utils/rel.h"
137 : #include "utils/selfuncs.h"
138 : #include "utils/snapmgr.h"
139 : #include "utils/spccache.h"
140 : #include "utils/syscache.h"
141 : #include "utils/timestamp.h"
142 : #include "utils/typcache.h"
143 :
144 : #define DEFAULT_PAGE_CPU_MULTIPLIER 50.0
145 :
146 : /*
147 : * In production builds, switch to hash-based MCV matching when the lists are
148 : * large enough to amortize hash setup cost. (This threshold is compared to
149 : * the sum of the lengths of the two MCV lists. This is simplistic but seems
150 : * to work well enough.) In debug builds, we use a smaller threshold so that
151 : * the regression tests cover both paths well.
152 : */
153 : #ifndef USE_ASSERT_CHECKING
154 : #define EQJOINSEL_MCV_HASH_THRESHOLD 200
155 : #else
156 : #define EQJOINSEL_MCV_HASH_THRESHOLD 20
157 : #endif
158 :
159 : /* Entries in the simplehash hash table used by eqjoinsel_find_matches */
160 : typedef struct MCVHashEntry
161 : {
162 : Datum value; /* the value represented by this entry */
163 : int index; /* its index in the relevant AttStatsSlot */
164 : uint32 hash; /* hash code for the Datum */
165 : char status; /* status code used by simplehash.h */
166 : } MCVHashEntry;
167 :
168 : /* private_data for the simplehash hash table */
169 : typedef struct MCVHashContext
170 : {
171 : FunctionCallInfo equal_fcinfo; /* the equality join operator */
172 : FunctionCallInfo hash_fcinfo; /* the hash function to use */
173 : bool op_is_reversed; /* equality compares hash type to probe type */
174 : bool insert_mode; /* doing inserts or lookups? */
175 : bool hash_typbyval; /* typbyval of hashed data type */
176 : int16 hash_typlen; /* typlen of hashed data type */
177 : } MCVHashContext;
178 :
179 : /* forward reference */
180 : typedef struct MCVHashTable_hash MCVHashTable_hash;
181 :
182 : /* Hooks for plugins to get control when we ask for stats */
183 : get_relation_stats_hook_type get_relation_stats_hook = NULL;
184 : get_index_stats_hook_type get_index_stats_hook = NULL;
185 :
186 : static double eqsel_internal(PG_FUNCTION_ARGS, bool negate);
187 : static double eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
188 : Oid hashLeft, Oid hashRight,
189 : VariableStatData *vardata1, VariableStatData *vardata2,
190 : double nd1, double nd2,
191 : bool isdefault1, bool isdefault2,
192 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
193 : Form_pg_statistic stats1, Form_pg_statistic stats2,
194 : bool have_mcvs1, bool have_mcvs2,
195 : bool *hasmatch1, bool *hasmatch2,
196 : int *p_nmatches);
197 : static double eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
198 : Oid hashLeft, Oid hashRight,
199 : bool op_is_reversed,
200 : VariableStatData *vardata1, VariableStatData *vardata2,
201 : double nd1, double nd2,
202 : bool isdefault1, bool isdefault2,
203 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
204 : Form_pg_statistic stats1, Form_pg_statistic stats2,
205 : bool have_mcvs1, bool have_mcvs2,
206 : bool *hasmatch1, bool *hasmatch2,
207 : int *p_nmatches,
208 : RelOptInfo *inner_rel);
209 : static void eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
210 : Oid hashLeft, Oid hashRight,
211 : bool op_is_reversed,
212 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
213 : int nvalues1, int nvalues2,
214 : bool *hasmatch1, bool *hasmatch2,
215 : int *p_nmatches, double *p_matchprodfreq);
216 : static uint32 hash_mcv(MCVHashTable_hash *tab, Datum key);
217 : static bool mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1);
218 : static bool estimate_multivariate_ndistinct(PlannerInfo *root,
219 : RelOptInfo *rel, List **varinfos, double *ndistinct);
220 : static bool convert_to_scalar(Datum value, Oid valuetypid, Oid collid,
221 : double *scaledvalue,
222 : Datum lobound, Datum hibound, Oid boundstypid,
223 : double *scaledlobound, double *scaledhibound);
224 : static double convert_numeric_to_scalar(Datum value, Oid typid, bool *failure);
225 : static void convert_string_to_scalar(char *value,
226 : double *scaledvalue,
227 : char *lobound,
228 : double *scaledlobound,
229 : char *hibound,
230 : double *scaledhibound);
231 : static void convert_bytea_to_scalar(Datum value,
232 : double *scaledvalue,
233 : Datum lobound,
234 : double *scaledlobound,
235 : Datum hibound,
236 : double *scaledhibound);
237 : static double convert_one_string_to_scalar(char *value,
238 : int rangelo, int rangehi);
239 : static double convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
240 : int rangelo, int rangehi);
241 : static char *convert_string_datum(Datum value, Oid typid, Oid collid,
242 : bool *failure);
243 : static double convert_timevalue_to_scalar(Datum value, Oid typid,
244 : bool *failure);
245 : static void examine_simple_variable(PlannerInfo *root, Var *var,
246 : VariableStatData *vardata);
247 : static void examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
248 : int indexcol, VariableStatData *vardata);
249 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
250 : Oid sortop, Oid collation,
251 : Datum *min, Datum *max);
252 : static void get_stats_slot_range(AttStatsSlot *sslot,
253 : Oid opfuncoid, FmgrInfo *opproc,
254 : Oid collation, int16 typLen, bool typByVal,
255 : Datum *min, Datum *max, bool *p_have_data);
256 : static bool get_actual_variable_range(PlannerInfo *root,
257 : VariableStatData *vardata,
258 : Oid sortop, Oid collation,
259 : Datum *min, Datum *max);
260 : static bool get_actual_variable_endpoint(Relation heapRel,
261 : Relation indexRel,
262 : ScanDirection indexscandir,
263 : ScanKey scankeys,
264 : int16 typLen,
265 : bool typByVal,
266 : TupleTableSlot *tableslot,
267 : MemoryContext outercontext,
268 : Datum *endpointDatum);
269 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
270 : static double btcost_correlation(IndexOptInfo *index,
271 : VariableStatData *vardata);
272 :
273 : /* Define support routines for MCV hash tables */
274 : #define SH_PREFIX MCVHashTable
275 : #define SH_ELEMENT_TYPE MCVHashEntry
276 : #define SH_KEY_TYPE Datum
277 : #define SH_KEY value
278 : #define SH_HASH_KEY(tab,key) hash_mcv(tab, key)
279 : #define SH_EQUAL(tab,key0,key1) mcvs_equal(tab, key0, key1)
280 : #define SH_SCOPE static inline
281 : #define SH_STORE_HASH
282 : #define SH_GET_HASH(tab,ent) (ent)->hash
283 : #define SH_DEFINE
284 : #define SH_DECLARE
285 : #include "lib/simplehash.h"
286 :
287 :
288 : /*
289 : * eqsel - Selectivity of "=" for any data types.
290 : *
291 : * Note: this routine is also used to estimate selectivity for some
292 : * operators that are not "=" but have comparable selectivity behavior,
293 : * such as "~=" (geometric approximate-match). Even for "=", we must
294 : * keep in mind that the left and right datatypes may differ.
295 : */
296 : Datum
297 692054 : eqsel(PG_FUNCTION_ARGS)
298 : {
299 692054 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
300 : }
301 :
302 : /*
303 : * Common code for eqsel() and neqsel()
304 : */
305 : static double
306 738118 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
307 : {
308 738118 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
309 738118 : Oid operator = PG_GETARG_OID(1);
310 738118 : List *args = (List *) PG_GETARG_POINTER(2);
311 738118 : int varRelid = PG_GETARG_INT32(3);
312 738118 : Oid collation = PG_GET_COLLATION();
313 : VariableStatData vardata;
314 : Node *other;
315 : bool varonleft;
316 : double selec;
317 :
318 : /*
319 : * When asked about <>, we do the estimation using the corresponding =
320 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
321 : */
322 738118 : if (negate)
323 : {
324 46064 : operator = get_negator(operator);
325 46064 : if (!OidIsValid(operator))
326 : {
327 : /* Use default selectivity (should we raise an error instead?) */
328 0 : return 1.0 - DEFAULT_EQ_SEL;
329 : }
330 : }
331 :
332 : /*
333 : * If expression is not variable = something or something = variable, then
334 : * punt and return a default estimate.
335 : */
336 738118 : if (!get_restriction_variable(root, args, varRelid,
337 : &vardata, &other, &varonleft))
338 5394 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
339 :
340 : /*
341 : * We can do a lot better if the something is a constant. (Note: the
342 : * Const might result from estimation rather than being a simple constant
343 : * in the query.)
344 : */
345 732718 : if (IsA(other, Const))
346 303508 : selec = var_eq_const(&vardata, operator, collation,
347 303508 : ((Const *) other)->constvalue,
348 303508 : ((Const *) other)->constisnull,
349 : varonleft, negate);
350 : else
351 429210 : selec = var_eq_non_const(&vardata, operator, collation, other,
352 : varonleft, negate);
353 :
354 732718 : ReleaseVariableStats(vardata);
355 :
356 732718 : return selec;
357 : }
358 :
359 : /*
360 : * var_eq_const --- eqsel for var = const case
361 : *
362 : * This is exported so that some other estimation functions can use it.
363 : */
364 : double
365 347872 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
366 : Datum constval, bool constisnull,
367 : bool varonleft, bool negate)
368 : {
369 : double selec;
370 347872 : double nullfrac = 0.0;
371 : bool isdefault;
372 : Oid opfuncoid;
373 :
374 : /*
375 : * If the constant is NULL, assume operator is strict and return zero, ie,
376 : * operator will never return TRUE. (It's zero even for a negator op.)
377 : */
378 347872 : if (constisnull)
379 408 : return 0.0;
380 :
381 : /*
382 : * Grab the nullfrac for use below. Note we allow use of nullfrac
383 : * regardless of security check.
384 : */
385 347464 : if (HeapTupleIsValid(vardata->statsTuple))
386 : {
387 : Form_pg_statistic stats;
388 :
389 261502 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
390 261502 : nullfrac = stats->stanullfrac;
391 : }
392 :
393 : /*
394 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
395 : * assume there is exactly one match regardless of anything else. (This
396 : * is slightly bogus, since the index or clause's equality operator might
397 : * be different from ours, but it's much more likely to be right than
398 : * ignoring the information.)
399 : */
400 347464 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
401 : {
402 82908 : selec = 1.0 / vardata->rel->tuples;
403 : }
404 460682 : else if (HeapTupleIsValid(vardata->statsTuple) &&
405 196126 : statistic_proc_security_check(vardata,
406 196126 : (opfuncoid = get_opcode(oproid))))
407 196126 : {
408 : AttStatsSlot sslot;
409 196126 : bool match = false;
410 : int i;
411 :
412 : /*
413 : * Is the constant "=" to any of the column's most common values?
414 : * (Although the given operator may not really be "=", we will assume
415 : * that seeing whether it returns TRUE is an appropriate test. If you
416 : * don't like this, maybe you shouldn't be using eqsel for your
417 : * operator...)
418 : */
419 196126 : if (get_attstatsslot(&sslot, vardata->statsTuple,
420 : STATISTIC_KIND_MCV, InvalidOid,
421 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
422 : {
423 175170 : LOCAL_FCINFO(fcinfo, 2);
424 : FmgrInfo eqproc;
425 :
426 175170 : fmgr_info(opfuncoid, &eqproc);
427 :
428 : /*
429 : * Save a few cycles by setting up the fcinfo struct just once.
430 : * Using FunctionCallInvoke directly also avoids failure if the
431 : * eqproc returns NULL, though really equality functions should
432 : * never do that.
433 : */
434 175170 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
435 : NULL, NULL);
436 175170 : fcinfo->args[0].isnull = false;
437 175170 : fcinfo->args[1].isnull = false;
438 : /* be careful to apply operator right way 'round */
439 175170 : if (varonleft)
440 175138 : fcinfo->args[1].value = constval;
441 : else
442 32 : fcinfo->args[0].value = constval;
443 :
444 2962796 : for (i = 0; i < sslot.nvalues; i++)
445 : {
446 : Datum fresult;
447 :
448 2880162 : if (varonleft)
449 2880106 : fcinfo->args[0].value = sslot.values[i];
450 : else
451 56 : fcinfo->args[1].value = sslot.values[i];
452 2880162 : fcinfo->isnull = false;
453 2880162 : fresult = FunctionCallInvoke(fcinfo);
454 2880162 : if (!fcinfo->isnull && DatumGetBool(fresult))
455 : {
456 92536 : match = true;
457 92536 : break;
458 : }
459 : }
460 : }
461 : else
462 : {
463 : /* no most-common-value info available */
464 20956 : i = 0; /* keep compiler quiet */
465 : }
466 :
467 196126 : if (match)
468 : {
469 : /*
470 : * Constant is "=" to this common value. We know selectivity
471 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
472 : */
473 92536 : selec = sslot.numbers[i];
474 : }
475 : else
476 : {
477 : /*
478 : * Comparison is against a constant that is neither NULL nor any
479 : * of the common values. Its selectivity cannot be more than
480 : * this:
481 : */
482 103590 : double sumcommon = 0.0;
483 : double otherdistinct;
484 :
485 2521716 : for (i = 0; i < sslot.nnumbers; i++)
486 2418126 : sumcommon += sslot.numbers[i];
487 103590 : selec = 1.0 - sumcommon - nullfrac;
488 103590 : CLAMP_PROBABILITY(selec);
489 :
490 : /*
491 : * and in fact it's probably a good deal less. We approximate that
492 : * all the not-common values share this remaining fraction
493 : * equally, so we divide by the number of other distinct values.
494 : */
495 103590 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
496 103590 : sslot.nnumbers;
497 103590 : if (otherdistinct > 1)
498 52066 : selec /= otherdistinct;
499 :
500 : /*
501 : * Another cross-check: selectivity shouldn't be estimated as more
502 : * than the least common "most common value".
503 : */
504 103590 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
505 0 : selec = sslot.numbers[sslot.nnumbers - 1];
506 : }
507 :
508 196126 : free_attstatsslot(&sslot);
509 : }
510 : else
511 : {
512 : /*
513 : * No ANALYZE stats available, so make a guess using estimated number
514 : * of distinct values and assuming they are equally common. (The guess
515 : * is unlikely to be very good, but we do know a few special cases.)
516 : */
517 68430 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
518 : }
519 :
520 : /* now adjust if we wanted <> rather than = */
521 347464 : if (negate)
522 37248 : selec = 1.0 - selec - nullfrac;
523 :
524 : /* result should be in range, but make sure... */
525 347464 : CLAMP_PROBABILITY(selec);
526 :
527 347464 : return selec;
528 : }
529 :
530 : /*
531 : * var_eq_non_const --- eqsel for var = something-other-than-const case
532 : *
533 : * This is exported so that some other estimation functions can use it.
534 : */
535 : double
536 429210 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
537 : Node *other,
538 : bool varonleft, bool negate)
539 : {
540 : double selec;
541 429210 : double nullfrac = 0.0;
542 : bool isdefault;
543 :
544 : /*
545 : * Grab the nullfrac for use below.
546 : */
547 429210 : if (HeapTupleIsValid(vardata->statsTuple))
548 : {
549 : Form_pg_statistic stats;
550 :
551 294114 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
552 294114 : nullfrac = stats->stanullfrac;
553 : }
554 :
555 : /*
556 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
557 : * assume there is exactly one match regardless of anything else. (This
558 : * is slightly bogus, since the index or clause's equality operator might
559 : * be different from ours, but it's much more likely to be right than
560 : * ignoring the information.)
561 : */
562 429210 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
563 : {
564 161634 : selec = 1.0 / vardata->rel->tuples;
565 : }
566 267576 : else if (HeapTupleIsValid(vardata->statsTuple))
567 : {
568 : double ndistinct;
569 : AttStatsSlot sslot;
570 :
571 : /*
572 : * Search is for a value that we do not know a priori, but we will
573 : * assume it is not NULL. Estimate the selectivity as non-null
574 : * fraction divided by number of distinct values, so that we get a
575 : * result averaged over all possible values whether common or
576 : * uncommon. (Essentially, we are assuming that the not-yet-known
577 : * comparison value is equally likely to be any of the possible
578 : * values, regardless of their frequency in the table. Is that a good
579 : * idea?)
580 : */
581 148602 : selec = 1.0 - nullfrac;
582 148602 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
583 148602 : if (ndistinct > 1)
584 144898 : selec /= ndistinct;
585 :
586 : /*
587 : * Cross-check: selectivity should never be estimated as more than the
588 : * most common value's.
589 : */
590 148602 : if (get_attstatsslot(&sslot, vardata->statsTuple,
591 : STATISTIC_KIND_MCV, InvalidOid,
592 : ATTSTATSSLOT_NUMBERS))
593 : {
594 129486 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
595 558 : selec = sslot.numbers[0];
596 129486 : free_attstatsslot(&sslot);
597 : }
598 : }
599 : else
600 : {
601 : /*
602 : * No ANALYZE stats available, so make a guess using estimated number
603 : * of distinct values and assuming they are equally common. (The guess
604 : * is unlikely to be very good, but we do know a few special cases.)
605 : */
606 118974 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
607 : }
608 :
609 : /* now adjust if we wanted <> rather than = */
610 429210 : if (negate)
611 6484 : selec = 1.0 - selec - nullfrac;
612 :
613 : /* result should be in range, but make sure... */
614 429210 : CLAMP_PROBABILITY(selec);
615 :
616 429210 : return selec;
617 : }
618 :
619 : /*
620 : * neqsel - Selectivity of "!=" for any data types.
621 : *
622 : * This routine is also used for some operators that are not "!="
623 : * but have comparable selectivity behavior. See above comments
624 : * for eqsel().
625 : */
626 : Datum
627 46064 : neqsel(PG_FUNCTION_ARGS)
628 : {
629 46064 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
630 : }
631 :
632 : /*
633 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
634 : *
635 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
636 : * The isgt and iseq flags distinguish which of the four cases apply.
637 : *
638 : * The caller has commuted the clause, if necessary, so that we can treat
639 : * the variable as being on the left. The caller must also make sure that
640 : * the other side of the clause is a non-null Const, and dissect that into
641 : * a value and datatype. (This definition simplifies some callers that
642 : * want to estimate against a computed value instead of a Const node.)
643 : *
644 : * This routine works for any datatype (or pair of datatypes) known to
645 : * convert_to_scalar(). If it is applied to some other datatype,
646 : * it will return an approximate estimate based on assuming that the constant
647 : * value falls in the middle of the bin identified by binary search.
648 : */
649 : static double
650 371494 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
651 : Oid collation,
652 : VariableStatData *vardata, Datum constval, Oid consttype)
653 : {
654 : Form_pg_statistic stats;
655 : FmgrInfo opproc;
656 : double mcv_selec,
657 : hist_selec,
658 : sumcommon;
659 : double selec;
660 :
661 371494 : if (!HeapTupleIsValid(vardata->statsTuple))
662 : {
663 : /*
664 : * No stats are available. Typically this means we have to fall back
665 : * on the default estimate; but if the variable is CTID then we can
666 : * make an estimate based on comparing the constant to the table size.
667 : */
668 22662 : if (vardata->var && IsA(vardata->var, Var) &&
669 17710 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
670 : {
671 : ItemPointer itemptr;
672 : double block;
673 : double density;
674 :
675 : /*
676 : * If the relation's empty, we're going to include all of it.
677 : * (This is mostly to avoid divide-by-zero below.)
678 : */
679 1960 : if (vardata->rel->pages == 0)
680 0 : return 1.0;
681 :
682 1960 : itemptr = (ItemPointer) DatumGetPointer(constval);
683 1960 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
684 :
685 : /*
686 : * Determine the average number of tuples per page (density).
687 : *
688 : * Since the last page will, on average, be only half full, we can
689 : * estimate it to have half as many tuples as earlier pages. So
690 : * give it half the weight of a regular page.
691 : */
692 1960 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
693 :
694 : /* If target is the last page, use half the density. */
695 1960 : if (block >= vardata->rel->pages - 1)
696 30 : density *= 0.5;
697 :
698 : /*
699 : * Using the average tuples per page, calculate how far into the
700 : * page the itemptr is likely to be and adjust block accordingly,
701 : * by adding that fraction of a whole block (but never more than a
702 : * whole block, no matter how high the itemptr's offset is). Here
703 : * we are ignoring the possibility of dead-tuple line pointers,
704 : * which is fairly bogus, but we lack the info to do better.
705 : */
706 1960 : if (density > 0.0)
707 : {
708 1960 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
709 :
710 1960 : block += Min(offset / density, 1.0);
711 : }
712 :
713 : /*
714 : * Convert relative block number to selectivity. Again, the last
715 : * page has only half weight.
716 : */
717 1960 : selec = block / (vardata->rel->pages - 0.5);
718 :
719 : /*
720 : * The calculation so far gave us a selectivity for the "<=" case.
721 : * We'll have one fewer tuple for "<" and one additional tuple for
722 : * ">=", the latter of which we'll reverse the selectivity for
723 : * below, so we can simply subtract one tuple for both cases. The
724 : * cases that need this adjustment can be identified by iseq being
725 : * equal to isgt.
726 : */
727 1960 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
728 1844 : selec -= (1.0 / vardata->rel->tuples);
729 :
730 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
731 1960 : if (isgt)
732 1838 : selec = 1.0 - selec;
733 :
734 1960 : CLAMP_PROBABILITY(selec);
735 1960 : return selec;
736 : }
737 :
738 : /* no stats available, so default result */
739 20702 : return DEFAULT_INEQ_SEL;
740 : }
741 348832 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
742 :
743 348832 : fmgr_info(get_opcode(operator), &opproc);
744 :
745 : /*
746 : * If we have most-common-values info, add up the fractions of the MCV
747 : * entries that satisfy MCV OP CONST. These fractions contribute directly
748 : * to the result selectivity. Also add up the total fraction represented
749 : * by MCV entries.
750 : */
751 348832 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
752 : &sumcommon);
753 :
754 : /*
755 : * If there is a histogram, determine which bin the constant falls in, and
756 : * compute the resulting contribution to selectivity.
757 : */
758 348832 : hist_selec = ineq_histogram_selectivity(root, vardata,
759 : operator, &opproc, isgt, iseq,
760 : collation,
761 : constval, consttype);
762 :
763 : /*
764 : * Now merge the results from the MCV and histogram calculations,
765 : * realizing that the histogram covers only the non-null values that are
766 : * not listed in MCV.
767 : */
768 348832 : selec = 1.0 - stats->stanullfrac - sumcommon;
769 :
770 348832 : if (hist_selec >= 0.0)
771 216700 : selec *= hist_selec;
772 : else
773 : {
774 : /*
775 : * If no histogram but there are values not accounted for by MCV,
776 : * arbitrarily assume half of them will match.
777 : */
778 132132 : selec *= 0.5;
779 : }
780 :
781 348832 : selec += mcv_selec;
782 :
783 : /* result should be in range, but make sure... */
784 348832 : CLAMP_PROBABILITY(selec);
785 :
786 348832 : return selec;
787 : }
788 :
789 : /*
790 : * mcv_selectivity - Examine the MCV list for selectivity estimates
791 : *
792 : * Determine the fraction of the variable's MCV population that satisfies
793 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
794 : * compute the fraction of the total column population represented by the MCV
795 : * list. This code will work for any boolean-returning predicate operator.
796 : *
797 : * The function result is the MCV selectivity, and the fraction of the
798 : * total population is returned into *sumcommonp. Zeroes are returned
799 : * if there is no MCV list.
800 : */
801 : double
802 355012 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
803 : Datum constval, bool varonleft,
804 : double *sumcommonp)
805 : {
806 : double mcv_selec,
807 : sumcommon;
808 : AttStatsSlot sslot;
809 : int i;
810 :
811 355012 : mcv_selec = 0.0;
812 355012 : sumcommon = 0.0;
813 :
814 707564 : if (HeapTupleIsValid(vardata->statsTuple) &&
815 704774 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
816 352222 : get_attstatsslot(&sslot, vardata->statsTuple,
817 : STATISTIC_KIND_MCV, InvalidOid,
818 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
819 : {
820 195104 : LOCAL_FCINFO(fcinfo, 2);
821 :
822 : /*
823 : * We invoke the opproc "by hand" so that we won't fail on NULL
824 : * results. Such cases won't arise for normal comparison functions,
825 : * but generic_restriction_selectivity could perhaps be used with
826 : * operators that can return NULL. A small side benefit is to not
827 : * need to re-initialize the fcinfo struct from scratch each time.
828 : */
829 195104 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
830 : NULL, NULL);
831 195104 : fcinfo->args[0].isnull = false;
832 195104 : fcinfo->args[1].isnull = false;
833 : /* be careful to apply operator right way 'round */
834 195104 : if (varonleft)
835 195104 : fcinfo->args[1].value = constval;
836 : else
837 0 : fcinfo->args[0].value = constval;
838 :
839 4608484 : for (i = 0; i < sslot.nvalues; i++)
840 : {
841 : Datum fresult;
842 :
843 4413380 : if (varonleft)
844 4413380 : fcinfo->args[0].value = sslot.values[i];
845 : else
846 0 : fcinfo->args[1].value = sslot.values[i];
847 4413380 : fcinfo->isnull = false;
848 4413380 : fresult = FunctionCallInvoke(fcinfo);
849 4413380 : if (!fcinfo->isnull && DatumGetBool(fresult))
850 1703916 : mcv_selec += sslot.numbers[i];
851 4413380 : sumcommon += sslot.numbers[i];
852 : }
853 195104 : free_attstatsslot(&sslot);
854 : }
855 :
856 355012 : *sumcommonp = sumcommon;
857 355012 : return mcv_selec;
858 : }
859 :
860 : /*
861 : * histogram_selectivity - Examine the histogram for selectivity estimates
862 : *
863 : * Determine the fraction of the variable's histogram entries that satisfy
864 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
865 : *
866 : * This code will work for any boolean-returning predicate operator, whether
867 : * or not it has anything to do with the histogram sort operator. We are
868 : * essentially using the histogram just as a representative sample. However,
869 : * small histograms are unlikely to be all that representative, so the caller
870 : * should be prepared to fall back on some other estimation approach when the
871 : * histogram is missing or very small. It may also be prudent to combine this
872 : * approach with another one when the histogram is small.
873 : *
874 : * If the actual histogram size is not at least min_hist_size, we won't bother
875 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
876 : * ignore the first and last n_skip histogram elements, on the grounds that
877 : * they are outliers and hence not very representative. Typical values for
878 : * these parameters are 10 and 1.
879 : *
880 : * The function result is the selectivity, or -1 if there is no histogram
881 : * or it's smaller than min_hist_size.
882 : *
883 : * The output parameter *hist_size receives the actual histogram size,
884 : * or zero if no histogram. Callers may use this number to decide how
885 : * much faith to put in the function result.
886 : *
887 : * Note that the result disregards both the most-common-values (if any) and
888 : * null entries. The caller is expected to combine this result with
889 : * statistics for those portions of the column population. It may also be
890 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
891 : */
892 : double
893 6180 : histogram_selectivity(VariableStatData *vardata,
894 : FmgrInfo *opproc, Oid collation,
895 : Datum constval, bool varonleft,
896 : int min_hist_size, int n_skip,
897 : int *hist_size)
898 : {
899 : double result;
900 : AttStatsSlot sslot;
901 :
902 : /* check sanity of parameters */
903 : Assert(n_skip >= 0);
904 : Assert(min_hist_size > 2 * n_skip);
905 :
906 9900 : if (HeapTupleIsValid(vardata->statsTuple) &&
907 7434 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
908 3714 : get_attstatsslot(&sslot, vardata->statsTuple,
909 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
910 : ATTSTATSSLOT_VALUES))
911 : {
912 3620 : *hist_size = sslot.nvalues;
913 3620 : if (sslot.nvalues >= min_hist_size)
914 : {
915 1782 : LOCAL_FCINFO(fcinfo, 2);
916 1782 : int nmatch = 0;
917 : int i;
918 :
919 : /*
920 : * We invoke the opproc "by hand" so that we won't fail on NULL
921 : * results. Such cases won't arise for normal comparison
922 : * functions, but generic_restriction_selectivity could perhaps be
923 : * used with operators that can return NULL. A small side benefit
924 : * is to not need to re-initialize the fcinfo struct from scratch
925 : * each time.
926 : */
927 1782 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
928 : NULL, NULL);
929 1782 : fcinfo->args[0].isnull = false;
930 1782 : fcinfo->args[1].isnull = false;
931 : /* be careful to apply operator right way 'round */
932 1782 : if (varonleft)
933 1782 : fcinfo->args[1].value = constval;
934 : else
935 0 : fcinfo->args[0].value = constval;
936 :
937 146900 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
938 : {
939 : Datum fresult;
940 :
941 145118 : if (varonleft)
942 145118 : fcinfo->args[0].value = sslot.values[i];
943 : else
944 0 : fcinfo->args[1].value = sslot.values[i];
945 145118 : fcinfo->isnull = false;
946 145118 : fresult = FunctionCallInvoke(fcinfo);
947 145118 : if (!fcinfo->isnull && DatumGetBool(fresult))
948 9572 : nmatch++;
949 : }
950 1782 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
951 : }
952 : else
953 1838 : result = -1;
954 3620 : free_attstatsslot(&sslot);
955 : }
956 : else
957 : {
958 2560 : *hist_size = 0;
959 2560 : result = -1;
960 : }
961 :
962 6180 : return result;
963 : }
964 :
965 : /*
966 : * generic_restriction_selectivity - Selectivity for almost anything
967 : *
968 : * This function estimates selectivity for operators that we don't have any
969 : * special knowledge about, but are on data types that we collect standard
970 : * MCV and/or histogram statistics for. (Additional assumptions are that
971 : * the operator is strict and immutable, or at least stable.)
972 : *
973 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
974 : * applying the operator to each element of the column's MCV and/or histogram
975 : * stats, and merging the results using the assumption that the histogram is
976 : * a reasonable random sample of the column's non-MCV population. Note that
977 : * if the operator's semantics are related to the histogram ordering, this
978 : * might not be such a great assumption; other functions such as
979 : * scalarineqsel() are probably a better match in such cases.
980 : *
981 : * Otherwise, fall back to the default selectivity provided by the caller.
982 : */
983 : double
984 1130 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
985 : List *args, int varRelid,
986 : double default_selectivity)
987 : {
988 : double selec;
989 : VariableStatData vardata;
990 : Node *other;
991 : bool varonleft;
992 :
993 : /*
994 : * If expression is not variable OP something or something OP variable,
995 : * then punt and return the default estimate.
996 : */
997 1130 : if (!get_restriction_variable(root, args, varRelid,
998 : &vardata, &other, &varonleft))
999 0 : return default_selectivity;
1000 :
1001 : /*
1002 : * If the something is a NULL constant, assume operator is strict and
1003 : * return zero, ie, operator will never return TRUE.
1004 : */
1005 1130 : if (IsA(other, Const) &&
1006 1130 : ((Const *) other)->constisnull)
1007 : {
1008 0 : ReleaseVariableStats(vardata);
1009 0 : return 0.0;
1010 : }
1011 :
1012 1130 : if (IsA(other, Const))
1013 : {
1014 : /* Variable is being compared to a known non-null constant */
1015 1130 : Datum constval = ((Const *) other)->constvalue;
1016 : FmgrInfo opproc;
1017 : double mcvsum;
1018 : double mcvsel;
1019 : double nullfrac;
1020 : int hist_size;
1021 :
1022 1130 : fmgr_info(get_opcode(oproid), &opproc);
1023 :
1024 : /*
1025 : * Calculate the selectivity for the column's most common values.
1026 : */
1027 1130 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
1028 : constval, varonleft,
1029 : &mcvsum);
1030 :
1031 : /*
1032 : * If the histogram is large enough, see what fraction of it matches
1033 : * the query, and assume that's representative of the non-MCV
1034 : * population. Otherwise use the default selectivity for the non-MCV
1035 : * population.
1036 : */
1037 1130 : selec = histogram_selectivity(&vardata, &opproc, collation,
1038 : constval, varonleft,
1039 : 10, 1, &hist_size);
1040 1130 : if (selec < 0)
1041 : {
1042 : /* Nope, fall back on default */
1043 1130 : selec = default_selectivity;
1044 : }
1045 0 : else if (hist_size < 100)
1046 : {
1047 : /*
1048 : * For histogram sizes from 10 to 100, we combine the histogram
1049 : * and default selectivities, putting increasingly more trust in
1050 : * the histogram for larger sizes.
1051 : */
1052 0 : double hist_weight = hist_size / 100.0;
1053 :
1054 0 : selec = selec * hist_weight +
1055 0 : default_selectivity * (1.0 - hist_weight);
1056 : }
1057 :
1058 : /* In any case, don't believe extremely small or large estimates. */
1059 1130 : if (selec < 0.0001)
1060 0 : selec = 0.0001;
1061 1130 : else if (selec > 0.9999)
1062 0 : selec = 0.9999;
1063 :
1064 : /* Don't forget to account for nulls. */
1065 1130 : if (HeapTupleIsValid(vardata.statsTuple))
1066 84 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1067 : else
1068 1046 : nullfrac = 0.0;
1069 :
1070 : /*
1071 : * Now merge the results from the MCV and histogram calculations,
1072 : * realizing that the histogram covers only the non-null values that
1073 : * are not listed in MCV.
1074 : */
1075 1130 : selec *= 1.0 - nullfrac - mcvsum;
1076 1130 : selec += mcvsel;
1077 : }
1078 : else
1079 : {
1080 : /* Comparison value is not constant, so we can't do anything */
1081 0 : selec = default_selectivity;
1082 : }
1083 :
1084 1130 : ReleaseVariableStats(vardata);
1085 :
1086 : /* result should be in range, but make sure... */
1087 1130 : CLAMP_PROBABILITY(selec);
1088 :
1089 1130 : return selec;
1090 : }
1091 :
1092 : /*
1093 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1094 : *
1095 : * Determine the fraction of the variable's histogram population that
1096 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1097 : * The isgt and iseq flags distinguish which of the four cases apply.
1098 : *
1099 : * While opproc could be looked up from the operator OID, common callers
1100 : * also need to call it separately, so we make the caller pass both.
1101 : *
1102 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1103 : *
1104 : * Note that the result disregards both the most-common-values (if any) and
1105 : * null entries. The caller is expected to combine this result with
1106 : * statistics for those portions of the column population.
1107 : *
1108 : * This is exported so that some other estimation functions can use it.
1109 : */
1110 : double
1111 353928 : ineq_histogram_selectivity(PlannerInfo *root,
1112 : VariableStatData *vardata,
1113 : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1114 : Oid collation,
1115 : Datum constval, Oid consttype)
1116 : {
1117 : double hist_selec;
1118 : AttStatsSlot sslot;
1119 :
1120 353928 : hist_selec = -1.0;
1121 :
1122 : /*
1123 : * Someday, ANALYZE might store more than one histogram per rel/att,
1124 : * corresponding to more than one possible sort ordering defined for the
1125 : * column type. Right now, we know there is only one, so just grab it and
1126 : * see if it matches the query.
1127 : *
1128 : * Note that we can't use opoid as search argument; the staop appearing in
1129 : * pg_statistic will be for the relevant '<' operator, but what we have
1130 : * might be some other inequality operator such as '>='. (Even if opoid
1131 : * is a '<' operator, it could be cross-type.) Hence we must use
1132 : * comparison_ops_are_compatible() to see if the operators match.
1133 : */
1134 707162 : if (HeapTupleIsValid(vardata->statsTuple) &&
1135 706144 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1136 352910 : get_attstatsslot(&sslot, vardata->statsTuple,
1137 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1138 : ATTSTATSSLOT_VALUES))
1139 : {
1140 221100 : if (sslot.nvalues > 1 &&
1141 442124 : sslot.stacoll == collation &&
1142 221024 : comparison_ops_are_compatible(sslot.staop, opoid))
1143 220916 : {
1144 : /*
1145 : * Use binary search to find the desired location, namely the
1146 : * right end of the histogram bin containing the comparison value,
1147 : * which is the leftmost entry for which the comparison operator
1148 : * succeeds (if isgt) or fails (if !isgt).
1149 : *
1150 : * In this loop, we pay no attention to whether the operator iseq
1151 : * or not; that detail will be mopped up below. (We cannot tell,
1152 : * anyway, whether the operator thinks the values are equal.)
1153 : *
1154 : * If the binary search accesses the first or last histogram
1155 : * entry, we try to replace that endpoint with the true column min
1156 : * or max as found by get_actual_variable_range(). This
1157 : * ameliorates misestimates when the min or max is moving as a
1158 : * result of changes since the last ANALYZE. Note that this could
1159 : * result in effectively including MCVs into the histogram that
1160 : * weren't there before, but we don't try to correct for that.
1161 : */
1162 : double histfrac;
1163 220916 : int lobound = 0; /* first possible slot to search */
1164 220916 : int hibound = sslot.nvalues; /* last+1 slot to search */
1165 220916 : bool have_end = false;
1166 :
1167 : /*
1168 : * If there are only two histogram entries, we'll want up-to-date
1169 : * values for both. (If there are more than two, we need at most
1170 : * one of them to be updated, so we deal with that within the
1171 : * loop.)
1172 : */
1173 220916 : if (sslot.nvalues == 2)
1174 3946 : have_end = get_actual_variable_range(root,
1175 : vardata,
1176 : sslot.staop,
1177 : collation,
1178 : &sslot.values[0],
1179 3946 : &sslot.values[1]);
1180 :
1181 1460074 : while (lobound < hibound)
1182 : {
1183 1239158 : int probe = (lobound + hibound) / 2;
1184 : bool ltcmp;
1185 :
1186 : /*
1187 : * If we find ourselves about to compare to the first or last
1188 : * histogram entry, first try to replace it with the actual
1189 : * current min or max (unless we already did so above).
1190 : */
1191 1239158 : if (probe == 0 && sslot.nvalues > 2)
1192 108596 : have_end = get_actual_variable_range(root,
1193 : vardata,
1194 : sslot.staop,
1195 : collation,
1196 : &sslot.values[0],
1197 : NULL);
1198 1130562 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1199 74544 : have_end = get_actual_variable_range(root,
1200 : vardata,
1201 : sslot.staop,
1202 : collation,
1203 : NULL,
1204 74544 : &sslot.values[probe]);
1205 :
1206 1239158 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1207 : collation,
1208 1239158 : sslot.values[probe],
1209 : constval));
1210 1239158 : if (isgt)
1211 68478 : ltcmp = !ltcmp;
1212 1239158 : if (ltcmp)
1213 467132 : lobound = probe + 1;
1214 : else
1215 772026 : hibound = probe;
1216 : }
1217 :
1218 220916 : if (lobound <= 0)
1219 : {
1220 : /*
1221 : * Constant is below lower histogram boundary. More
1222 : * precisely, we have found that no entry in the histogram
1223 : * satisfies the inequality clause (if !isgt) or they all do
1224 : * (if isgt). We estimate that that's true of the entire
1225 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1226 : * below, if isgt).
1227 : */
1228 94516 : histfrac = 0.0;
1229 : }
1230 126400 : else if (lobound >= sslot.nvalues)
1231 : {
1232 : /*
1233 : * Inverse case: constant is above upper histogram boundary.
1234 : */
1235 34176 : histfrac = 1.0;
1236 : }
1237 : else
1238 : {
1239 : /* We have values[i-1] <= constant <= values[i]. */
1240 92224 : int i = lobound;
1241 92224 : double eq_selec = 0;
1242 : double val,
1243 : high,
1244 : low;
1245 : double binfrac;
1246 :
1247 : /*
1248 : * In the cases where we'll need it below, obtain an estimate
1249 : * of the selectivity of "x = constval". We use a calculation
1250 : * similar to what var_eq_const() does for a non-MCV constant,
1251 : * ie, estimate that all distinct non-MCV values occur equally
1252 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1253 : * will be done by our caller, so we shouldn't do that here.
1254 : * Therefore we can't try to clamp the estimate by reference
1255 : * to the least common MCV; the result would be too small.
1256 : *
1257 : * Note: since this is effectively assuming that constval
1258 : * isn't an MCV, it's logically dubious if constval in fact is
1259 : * one. But we have to apply *some* correction for equality,
1260 : * and anyway we cannot tell if constval is an MCV, since we
1261 : * don't have a suitable equality operator at hand.
1262 : */
1263 92224 : if (i == 1 || isgt == iseq)
1264 : {
1265 : double otherdistinct;
1266 : bool isdefault;
1267 : AttStatsSlot mcvslot;
1268 :
1269 : /* Get estimated number of distinct values */
1270 38052 : otherdistinct = get_variable_numdistinct(vardata,
1271 : &isdefault);
1272 :
1273 : /* Subtract off the number of known MCVs */
1274 38052 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1275 : STATISTIC_KIND_MCV, InvalidOid,
1276 : ATTSTATSSLOT_NUMBERS))
1277 : {
1278 3916 : otherdistinct -= mcvslot.nnumbers;
1279 3916 : free_attstatsslot(&mcvslot);
1280 : }
1281 :
1282 : /* If result doesn't seem sane, leave eq_selec at 0 */
1283 38052 : if (otherdistinct > 1)
1284 38010 : eq_selec = 1.0 / otherdistinct;
1285 : }
1286 :
1287 : /*
1288 : * Convert the constant and the two nearest bin boundary
1289 : * values to a uniform comparison scale, and do a linear
1290 : * interpolation within this bin.
1291 : */
1292 92224 : if (convert_to_scalar(constval, consttype, collation,
1293 : &val,
1294 92224 : sslot.values[i - 1], sslot.values[i],
1295 : vardata->vartype,
1296 : &low, &high))
1297 : {
1298 92224 : if (high <= low)
1299 : {
1300 : /* cope if bin boundaries appear identical */
1301 0 : binfrac = 0.5;
1302 : }
1303 92224 : else if (val <= low)
1304 20090 : binfrac = 0.0;
1305 72134 : else if (val >= high)
1306 2686 : binfrac = 1.0;
1307 : else
1308 : {
1309 69448 : binfrac = (val - low) / (high - low);
1310 :
1311 : /*
1312 : * Watch out for the possibility that we got a NaN or
1313 : * Infinity from the division. This can happen
1314 : * despite the previous checks, if for example "low"
1315 : * is -Infinity.
1316 : */
1317 69448 : if (isnan(binfrac) ||
1318 69448 : binfrac < 0.0 || binfrac > 1.0)
1319 0 : binfrac = 0.5;
1320 : }
1321 : }
1322 : else
1323 : {
1324 : /*
1325 : * Ideally we'd produce an error here, on the grounds that
1326 : * the given operator shouldn't have scalarXXsel
1327 : * registered as its selectivity func unless we can deal
1328 : * with its operand types. But currently, all manner of
1329 : * stuff is invoking scalarXXsel, so give a default
1330 : * estimate until that can be fixed.
1331 : */
1332 0 : binfrac = 0.5;
1333 : }
1334 :
1335 : /*
1336 : * Now, compute the overall selectivity across the values
1337 : * represented by the histogram. We have i-1 full bins and
1338 : * binfrac partial bin below the constant.
1339 : */
1340 92224 : histfrac = (double) (i - 1) + binfrac;
1341 92224 : histfrac /= (double) (sslot.nvalues - 1);
1342 :
1343 : /*
1344 : * At this point, histfrac is an estimate of the fraction of
1345 : * the population represented by the histogram that satisfies
1346 : * "x <= constval". Somewhat remarkably, this statement is
1347 : * true regardless of which operator we were doing the probes
1348 : * with, so long as convert_to_scalar() delivers reasonable
1349 : * results. If the probe constant is equal to some histogram
1350 : * entry, we would have considered the bin to the left of that
1351 : * entry if probing with "<" or ">=", or the bin to the right
1352 : * if probing with "<=" or ">"; but binfrac would have come
1353 : * out as 1.0 in the first case and 0.0 in the second, leading
1354 : * to the same histfrac in either case. For probe constants
1355 : * between histogram entries, we find the same bin and get the
1356 : * same estimate with any operator.
1357 : *
1358 : * The fact that the estimate corresponds to "x <= constval"
1359 : * and not "x < constval" is because of the way that ANALYZE
1360 : * constructs the histogram: each entry is, effectively, the
1361 : * rightmost value in its sample bucket. So selectivity
1362 : * values that are exact multiples of 1/(histogram_size-1)
1363 : * should be understood as estimates including a histogram
1364 : * entry plus everything to its left.
1365 : *
1366 : * However, that breaks down for the first histogram entry,
1367 : * which necessarily is the leftmost value in its sample
1368 : * bucket. That means the first histogram bin is slightly
1369 : * narrower than the rest, by an amount equal to eq_selec.
1370 : * Another way to say that is that we want "x <= leftmost" to
1371 : * be estimated as eq_selec not zero. So, if we're dealing
1372 : * with the first bin (i==1), rescale to make that true while
1373 : * adjusting the rest of that bin linearly.
1374 : */
1375 92224 : if (i == 1)
1376 16478 : histfrac += eq_selec * (1.0 - binfrac);
1377 :
1378 : /*
1379 : * "x <= constval" is good if we want an estimate for "<=" or
1380 : * ">", but if we are estimating for "<" or ">=", we now need
1381 : * to decrease the estimate by eq_selec.
1382 : */
1383 92224 : if (isgt == iseq)
1384 28808 : histfrac -= eq_selec;
1385 : }
1386 :
1387 : /*
1388 : * Now the estimate is finished for "<" and "<=" cases. If we are
1389 : * estimating for ">" or ">=", flip it.
1390 : */
1391 220916 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1392 :
1393 : /*
1394 : * The histogram boundaries are only approximate to begin with,
1395 : * and may well be out of date anyway. Therefore, don't believe
1396 : * extremely small or large selectivity estimates --- unless we
1397 : * got actual current endpoint values from the table, in which
1398 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1399 : * set the cutoff for other cases at a hundredth of the histogram
1400 : * resolution.
1401 : */
1402 220916 : if (have_end)
1403 125666 : CLAMP_PROBABILITY(hist_selec);
1404 : else
1405 : {
1406 95250 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1407 :
1408 95250 : if (hist_selec < cutoff)
1409 33546 : hist_selec = cutoff;
1410 61704 : else if (hist_selec > 1.0 - cutoff)
1411 22288 : hist_selec = 1.0 - cutoff;
1412 : }
1413 : }
1414 184 : else if (sslot.nvalues > 1)
1415 : {
1416 : /*
1417 : * If we get here, we have a histogram but it's not sorted the way
1418 : * we want. Do a brute-force search to see how many of the
1419 : * entries satisfy the comparison condition, and take that
1420 : * fraction as our estimate. (This is identical to the inner loop
1421 : * of histogram_selectivity; maybe share code?)
1422 : */
1423 184 : LOCAL_FCINFO(fcinfo, 2);
1424 184 : int nmatch = 0;
1425 :
1426 184 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1427 : NULL, NULL);
1428 184 : fcinfo->args[0].isnull = false;
1429 184 : fcinfo->args[1].isnull = false;
1430 184 : fcinfo->args[1].value = constval;
1431 962444 : for (int i = 0; i < sslot.nvalues; i++)
1432 : {
1433 : Datum fresult;
1434 :
1435 962260 : fcinfo->args[0].value = sslot.values[i];
1436 962260 : fcinfo->isnull = false;
1437 962260 : fresult = FunctionCallInvoke(fcinfo);
1438 962260 : if (!fcinfo->isnull && DatumGetBool(fresult))
1439 2188 : nmatch++;
1440 : }
1441 184 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1442 :
1443 : /*
1444 : * As above, clamp to a hundredth of the histogram resolution.
1445 : * This case is surely even less trustworthy than the normal one,
1446 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1447 : * clamp should be more restrictive in this case?)
1448 : */
1449 : {
1450 184 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1451 :
1452 184 : if (hist_selec < cutoff)
1453 12 : hist_selec = cutoff;
1454 172 : else if (hist_selec > 1.0 - cutoff)
1455 12 : hist_selec = 1.0 - cutoff;
1456 : }
1457 : }
1458 :
1459 221100 : free_attstatsslot(&sslot);
1460 : }
1461 :
1462 353928 : return hist_selec;
1463 : }
1464 :
1465 : /*
1466 : * Common wrapper function for the selectivity estimators that simply
1467 : * invoke scalarineqsel().
1468 : */
1469 : static Datum
1470 45052 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1471 : {
1472 45052 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1473 45052 : Oid operator = PG_GETARG_OID(1);
1474 45052 : List *args = (List *) PG_GETARG_POINTER(2);
1475 45052 : int varRelid = PG_GETARG_INT32(3);
1476 45052 : Oid collation = PG_GET_COLLATION();
1477 : VariableStatData vardata;
1478 : Node *other;
1479 : bool varonleft;
1480 : Datum constval;
1481 : Oid consttype;
1482 : double selec;
1483 :
1484 : /*
1485 : * If expression is not variable op something or something op variable,
1486 : * then punt and return a default estimate.
1487 : */
1488 45052 : if (!get_restriction_variable(root, args, varRelid,
1489 : &vardata, &other, &varonleft))
1490 650 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1491 :
1492 : /*
1493 : * Can't do anything useful if the something is not a constant, either.
1494 : */
1495 44402 : if (!IsA(other, Const))
1496 : {
1497 2818 : ReleaseVariableStats(vardata);
1498 2818 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1499 : }
1500 :
1501 : /*
1502 : * If the constant is NULL, assume operator is strict and return zero, ie,
1503 : * operator will never return TRUE.
1504 : */
1505 41584 : if (((Const *) other)->constisnull)
1506 : {
1507 66 : ReleaseVariableStats(vardata);
1508 66 : PG_RETURN_FLOAT8(0.0);
1509 : }
1510 41518 : constval = ((Const *) other)->constvalue;
1511 41518 : consttype = ((Const *) other)->consttype;
1512 :
1513 : /*
1514 : * Force the var to be on the left to simplify logic in scalarineqsel.
1515 : */
1516 41518 : if (!varonleft)
1517 : {
1518 384 : operator = get_commutator(operator);
1519 384 : if (!operator)
1520 : {
1521 : /* Use default selectivity (should we raise an error instead?) */
1522 0 : ReleaseVariableStats(vardata);
1523 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1524 : }
1525 384 : isgt = !isgt;
1526 : }
1527 :
1528 : /* The rest of the work is done by scalarineqsel(). */
1529 41518 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1530 : &vardata, constval, consttype);
1531 :
1532 41518 : ReleaseVariableStats(vardata);
1533 :
1534 41518 : PG_RETURN_FLOAT8((float8) selec);
1535 : }
1536 :
1537 : /*
1538 : * scalarltsel - Selectivity of "<" for scalars.
1539 : */
1540 : Datum
1541 15040 : scalarltsel(PG_FUNCTION_ARGS)
1542 : {
1543 15040 : return scalarineqsel_wrapper(fcinfo, false, false);
1544 : }
1545 :
1546 : /*
1547 : * scalarlesel - Selectivity of "<=" for scalars.
1548 : */
1549 : Datum
1550 4594 : scalarlesel(PG_FUNCTION_ARGS)
1551 : {
1552 4594 : return scalarineqsel_wrapper(fcinfo, false, true);
1553 : }
1554 :
1555 : /*
1556 : * scalargtsel - Selectivity of ">" for scalars.
1557 : */
1558 : Datum
1559 15430 : scalargtsel(PG_FUNCTION_ARGS)
1560 : {
1561 15430 : return scalarineqsel_wrapper(fcinfo, true, false);
1562 : }
1563 :
1564 : /*
1565 : * scalargesel - Selectivity of ">=" for scalars.
1566 : */
1567 : Datum
1568 9988 : scalargesel(PG_FUNCTION_ARGS)
1569 : {
1570 9988 : return scalarineqsel_wrapper(fcinfo, true, true);
1571 : }
1572 :
1573 : /*
1574 : * boolvarsel - Selectivity of Boolean variable.
1575 : *
1576 : * This can actually be called on any boolean-valued expression. If it
1577 : * involves only Vars of the specified relation, and if there are statistics
1578 : * about the Var or expression (the latter is possible if it's indexed) then
1579 : * we'll produce a real estimate; otherwise it's just a default.
1580 : */
1581 : Selectivity
1582 56386 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1583 : {
1584 : VariableStatData vardata;
1585 : double selec;
1586 :
1587 56386 : examine_variable(root, arg, varRelid, &vardata);
1588 56386 : if (HeapTupleIsValid(vardata.statsTuple))
1589 : {
1590 : /*
1591 : * A boolean variable V is equivalent to the clause V = 't', so we
1592 : * compute the selectivity as if that is what we have.
1593 : */
1594 35608 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1595 : BoolGetDatum(true), false, true, false);
1596 : }
1597 20778 : else if (is_funcclause(arg))
1598 : {
1599 : /*
1600 : * If we have no stats and it's a function call, estimate 0.3333333.
1601 : * This seems a pretty unprincipled choice, but Postgres has been
1602 : * using that estimate for function calls since 1992. The hoariness
1603 : * of this behavior suggests that we should not be in too much hurry
1604 : * to use another value.
1605 : */
1606 12334 : selec = 0.3333333;
1607 : }
1608 : else
1609 : {
1610 : /* Otherwise, the default estimate is 0.5 */
1611 8444 : selec = 0.5;
1612 : }
1613 56386 : ReleaseVariableStats(vardata);
1614 56386 : return selec;
1615 : }
1616 :
1617 : /*
1618 : * booltestsel - Selectivity of BooleanTest Node.
1619 : */
1620 : Selectivity
1621 886 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1622 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1623 : {
1624 : VariableStatData vardata;
1625 : double selec;
1626 :
1627 886 : examine_variable(root, arg, varRelid, &vardata);
1628 :
1629 886 : if (HeapTupleIsValid(vardata.statsTuple))
1630 : {
1631 : Form_pg_statistic stats;
1632 : double freq_null;
1633 : AttStatsSlot sslot;
1634 :
1635 0 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1636 0 : freq_null = stats->stanullfrac;
1637 :
1638 0 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1639 : STATISTIC_KIND_MCV, InvalidOid,
1640 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1641 0 : && sslot.nnumbers > 0)
1642 0 : {
1643 : double freq_true;
1644 : double freq_false;
1645 :
1646 : /*
1647 : * Get first MCV frequency and derive frequency for true.
1648 : */
1649 0 : if (DatumGetBool(sslot.values[0]))
1650 0 : freq_true = sslot.numbers[0];
1651 : else
1652 0 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1653 :
1654 : /*
1655 : * Next derive frequency for false. Then use these as appropriate
1656 : * to derive frequency for each case.
1657 : */
1658 0 : freq_false = 1.0 - freq_true - freq_null;
1659 :
1660 0 : switch (booltesttype)
1661 : {
1662 0 : case IS_UNKNOWN:
1663 : /* select only NULL values */
1664 0 : selec = freq_null;
1665 0 : break;
1666 0 : case IS_NOT_UNKNOWN:
1667 : /* select non-NULL values */
1668 0 : selec = 1.0 - freq_null;
1669 0 : break;
1670 0 : case IS_TRUE:
1671 : /* select only TRUE values */
1672 0 : selec = freq_true;
1673 0 : break;
1674 0 : case IS_NOT_TRUE:
1675 : /* select non-TRUE values */
1676 0 : selec = 1.0 - freq_true;
1677 0 : break;
1678 0 : case IS_FALSE:
1679 : /* select only FALSE values */
1680 0 : selec = freq_false;
1681 0 : break;
1682 0 : case IS_NOT_FALSE:
1683 : /* select non-FALSE values */
1684 0 : selec = 1.0 - freq_false;
1685 0 : break;
1686 0 : default:
1687 0 : elog(ERROR, "unrecognized booltesttype: %d",
1688 : (int) booltesttype);
1689 : selec = 0.0; /* Keep compiler quiet */
1690 : break;
1691 : }
1692 :
1693 0 : free_attstatsslot(&sslot);
1694 : }
1695 : else
1696 : {
1697 : /*
1698 : * No most-common-value info available. Still have null fraction
1699 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1700 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1701 : */
1702 0 : switch (booltesttype)
1703 : {
1704 0 : case IS_UNKNOWN:
1705 : /* select only NULL values */
1706 0 : selec = freq_null;
1707 0 : break;
1708 0 : case IS_NOT_UNKNOWN:
1709 : /* select non-NULL values */
1710 0 : selec = 1.0 - freq_null;
1711 0 : break;
1712 0 : case IS_TRUE:
1713 : case IS_FALSE:
1714 : /* Assume we select half of the non-NULL values */
1715 0 : selec = (1.0 - freq_null) / 2.0;
1716 0 : break;
1717 0 : case IS_NOT_TRUE:
1718 : case IS_NOT_FALSE:
1719 : /* Assume we select NULLs plus half of the non-NULLs */
1720 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1721 0 : selec = (freq_null + 1.0) / 2.0;
1722 0 : break;
1723 0 : default:
1724 0 : elog(ERROR, "unrecognized booltesttype: %d",
1725 : (int) booltesttype);
1726 : selec = 0.0; /* Keep compiler quiet */
1727 : break;
1728 : }
1729 : }
1730 : }
1731 : else
1732 : {
1733 : /*
1734 : * If we can't get variable statistics for the argument, perhaps
1735 : * clause_selectivity can do something with it. We ignore the
1736 : * possibility of a NULL value when using clause_selectivity, and just
1737 : * assume the value is either TRUE or FALSE.
1738 : */
1739 886 : switch (booltesttype)
1740 : {
1741 48 : case IS_UNKNOWN:
1742 48 : selec = DEFAULT_UNK_SEL;
1743 48 : break;
1744 108 : case IS_NOT_UNKNOWN:
1745 108 : selec = DEFAULT_NOT_UNK_SEL;
1746 108 : break;
1747 252 : case IS_TRUE:
1748 : case IS_NOT_FALSE:
1749 252 : selec = (double) clause_selectivity(root, arg,
1750 : varRelid,
1751 : jointype, sjinfo);
1752 252 : break;
1753 478 : case IS_FALSE:
1754 : case IS_NOT_TRUE:
1755 478 : selec = 1.0 - (double) clause_selectivity(root, arg,
1756 : varRelid,
1757 : jointype, sjinfo);
1758 478 : break;
1759 0 : default:
1760 0 : elog(ERROR, "unrecognized booltesttype: %d",
1761 : (int) booltesttype);
1762 : selec = 0.0; /* Keep compiler quiet */
1763 : break;
1764 : }
1765 : }
1766 :
1767 886 : ReleaseVariableStats(vardata);
1768 :
1769 : /* result should be in range, but make sure... */
1770 886 : CLAMP_PROBABILITY(selec);
1771 :
1772 886 : return (Selectivity) selec;
1773 : }
1774 :
1775 : /*
1776 : * nulltestsel - Selectivity of NullTest Node.
1777 : */
1778 : Selectivity
1779 17394 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1780 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1781 : {
1782 : VariableStatData vardata;
1783 : double selec;
1784 :
1785 17394 : examine_variable(root, arg, varRelid, &vardata);
1786 :
1787 17394 : if (HeapTupleIsValid(vardata.statsTuple))
1788 : {
1789 : Form_pg_statistic stats;
1790 : double freq_null;
1791 :
1792 9762 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1793 9762 : freq_null = stats->stanullfrac;
1794 :
1795 9762 : switch (nulltesttype)
1796 : {
1797 7232 : case IS_NULL:
1798 :
1799 : /*
1800 : * Use freq_null directly.
1801 : */
1802 7232 : selec = freq_null;
1803 7232 : break;
1804 2530 : case IS_NOT_NULL:
1805 :
1806 : /*
1807 : * Select not unknown (not null) values. Calculate from
1808 : * freq_null.
1809 : */
1810 2530 : selec = 1.0 - freq_null;
1811 2530 : break;
1812 0 : default:
1813 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1814 : (int) nulltesttype);
1815 : return (Selectivity) 0; /* keep compiler quiet */
1816 : }
1817 : }
1818 7632 : else if (vardata.var && IsA(vardata.var, Var) &&
1819 6872 : ((Var *) vardata.var)->varattno < 0)
1820 : {
1821 : /*
1822 : * There are no stats for system columns, but we know they are never
1823 : * NULL.
1824 : */
1825 68 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1826 : }
1827 : else
1828 : {
1829 : /*
1830 : * No ANALYZE stats available, so make a guess
1831 : */
1832 7564 : switch (nulltesttype)
1833 : {
1834 2108 : case IS_NULL:
1835 2108 : selec = DEFAULT_UNK_SEL;
1836 2108 : break;
1837 5456 : case IS_NOT_NULL:
1838 5456 : selec = DEFAULT_NOT_UNK_SEL;
1839 5456 : break;
1840 0 : default:
1841 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1842 : (int) nulltesttype);
1843 : return (Selectivity) 0; /* keep compiler quiet */
1844 : }
1845 : }
1846 :
1847 17394 : ReleaseVariableStats(vardata);
1848 :
1849 : /* result should be in range, but make sure... */
1850 17394 : CLAMP_PROBABILITY(selec);
1851 :
1852 17394 : return (Selectivity) selec;
1853 : }
1854 :
1855 : /*
1856 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1857 : *
1858 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1859 : * but it seems possible that RelabelType might show up. Also, the planner
1860 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1861 : * so we need to be ready to deal with more than one level.
1862 : */
1863 : static Node *
1864 127636 : strip_array_coercion(Node *node)
1865 : {
1866 : for (;;)
1867 : {
1868 127718 : if (node && IsA(node, ArrayCoerceExpr))
1869 82 : {
1870 2976 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1871 :
1872 : /*
1873 : * If the per-element expression is just a RelabelType on top of
1874 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1875 : */
1876 2976 : if (IsA(acoerce->elemexpr, RelabelType) &&
1877 82 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1878 82 : node = (Node *) acoerce->arg;
1879 : else
1880 : break;
1881 : }
1882 124742 : else if (node && IsA(node, RelabelType))
1883 : {
1884 : /* We don't really expect this case, but may as well cope */
1885 0 : node = (Node *) ((RelabelType *) node)->arg;
1886 : }
1887 : else
1888 : break;
1889 : }
1890 127636 : return node;
1891 : }
1892 :
1893 : /*
1894 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1895 : */
1896 : Selectivity
1897 22352 : scalararraysel(PlannerInfo *root,
1898 : ScalarArrayOpExpr *clause,
1899 : bool is_join_clause,
1900 : int varRelid,
1901 : JoinType jointype,
1902 : SpecialJoinInfo *sjinfo)
1903 : {
1904 22352 : Oid operator = clause->opno;
1905 22352 : bool useOr = clause->useOr;
1906 22352 : bool isEquality = false;
1907 22352 : bool isInequality = false;
1908 : Node *leftop;
1909 : Node *rightop;
1910 : Oid nominal_element_type;
1911 : Oid nominal_element_collation;
1912 : TypeCacheEntry *typentry;
1913 : RegProcedure oprsel;
1914 : FmgrInfo oprselproc;
1915 : Selectivity s1;
1916 : Selectivity s1disjoint;
1917 :
1918 : /* First, deconstruct the expression */
1919 : Assert(list_length(clause->args) == 2);
1920 22352 : leftop = (Node *) linitial(clause->args);
1921 22352 : rightop = (Node *) lsecond(clause->args);
1922 :
1923 : /* aggressively reduce both sides to constants */
1924 22352 : leftop = estimate_expression_value(root, leftop);
1925 22352 : rightop = estimate_expression_value(root, rightop);
1926 :
1927 : /* get nominal (after relabeling) element type of rightop */
1928 22352 : nominal_element_type = get_base_element_type(exprType(rightop));
1929 22352 : if (!OidIsValid(nominal_element_type))
1930 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1931 : /* get nominal collation, too, for generating constants */
1932 22352 : nominal_element_collation = exprCollation(rightop);
1933 :
1934 : /* look through any binary-compatible relabeling of rightop */
1935 22352 : rightop = strip_array_coercion(rightop);
1936 :
1937 : /*
1938 : * Detect whether the operator is the default equality or inequality
1939 : * operator of the array element type.
1940 : */
1941 22352 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1942 22352 : if (OidIsValid(typentry->eq_opr))
1943 : {
1944 22348 : if (operator == typentry->eq_opr)
1945 19072 : isEquality = true;
1946 3276 : else if (get_negator(operator) == typentry->eq_opr)
1947 2710 : isInequality = true;
1948 : }
1949 :
1950 : /*
1951 : * If it is equality or inequality, we might be able to estimate this as a
1952 : * form of array containment; for instance "const = ANY(column)" can be
1953 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1954 : * that, and returns the selectivity estimate if successful, or -1 if not.
1955 : */
1956 22352 : if ((isEquality || isInequality) && !is_join_clause)
1957 : {
1958 21782 : s1 = scalararraysel_containment(root, leftop, rightop,
1959 : nominal_element_type,
1960 : isEquality, useOr, varRelid);
1961 21782 : if (s1 >= 0.0)
1962 118 : return s1;
1963 : }
1964 :
1965 : /*
1966 : * Look up the underlying operator's selectivity estimator. Punt if it
1967 : * hasn't got one.
1968 : */
1969 22234 : if (is_join_clause)
1970 0 : oprsel = get_oprjoin(operator);
1971 : else
1972 22234 : oprsel = get_oprrest(operator);
1973 22234 : if (!oprsel)
1974 4 : return (Selectivity) 0.5;
1975 22230 : fmgr_info(oprsel, &oprselproc);
1976 :
1977 : /*
1978 : * In the array-containment check above, we must only believe that an
1979 : * operator is equality or inequality if it is the default btree equality
1980 : * operator (or its negator) for the element type, since those are the
1981 : * operators that array containment will use. But in what follows, we can
1982 : * be a little laxer, and also believe that any operators using eqsel() or
1983 : * neqsel() as selectivity estimator act like equality or inequality.
1984 : */
1985 22230 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1986 19150 : isEquality = true;
1987 3080 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1988 2600 : isInequality = true;
1989 :
1990 : /*
1991 : * We consider three cases:
1992 : *
1993 : * 1. rightop is an Array constant: deconstruct the array, apply the
1994 : * operator's selectivity function for each array element, and merge the
1995 : * results in the same way that clausesel.c does for AND/OR combinations.
1996 : *
1997 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
1998 : * function for each element of the ARRAY[] construct, and merge.
1999 : *
2000 : * 3. otherwise, make a guess ...
2001 : */
2002 22230 : if (rightop && IsA(rightop, Const))
2003 18066 : {
2004 18096 : Datum arraydatum = ((Const *) rightop)->constvalue;
2005 18096 : bool arrayisnull = ((Const *) rightop)->constisnull;
2006 : ArrayType *arrayval;
2007 : int16 elmlen;
2008 : bool elmbyval;
2009 : char elmalign;
2010 : int num_elems;
2011 : Datum *elem_values;
2012 : bool *elem_nulls;
2013 : int i;
2014 :
2015 18096 : if (arrayisnull) /* qual can't succeed if null array */
2016 30 : return (Selectivity) 0.0;
2017 18066 : arrayval = DatumGetArrayTypeP(arraydatum);
2018 18066 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2019 : &elmlen, &elmbyval, &elmalign);
2020 18066 : deconstruct_array(arrayval,
2021 : ARR_ELEMTYPE(arrayval),
2022 : elmlen, elmbyval, elmalign,
2023 : &elem_values, &elem_nulls, &num_elems);
2024 :
2025 : /*
2026 : * For generic operators, we assume the probability of success is
2027 : * independent for each array element. But for "= ANY" or "<> ALL",
2028 : * if the array elements are distinct (which'd typically be the case)
2029 : * then the probabilities are disjoint, and we should just sum them.
2030 : *
2031 : * If we were being really tense we would try to confirm that the
2032 : * elements are all distinct, but that would be expensive and it
2033 : * doesn't seem to be worth the cycles; it would amount to penalizing
2034 : * well-written queries in favor of poorly-written ones. However, we
2035 : * do protect ourselves a little bit by checking whether the
2036 : * disjointness assumption leads to an impossible (out of range)
2037 : * probability; if so, we fall back to the normal calculation.
2038 : */
2039 18066 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2040 :
2041 77022 : for (i = 0; i < num_elems; i++)
2042 : {
2043 : List *args;
2044 : Selectivity s2;
2045 :
2046 58956 : args = list_make2(leftop,
2047 : makeConst(nominal_element_type,
2048 : -1,
2049 : nominal_element_collation,
2050 : elmlen,
2051 : elem_values[i],
2052 : elem_nulls[i],
2053 : elmbyval));
2054 58956 : if (is_join_clause)
2055 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2056 : clause->inputcollid,
2057 : PointerGetDatum(root),
2058 : ObjectIdGetDatum(operator),
2059 : PointerGetDatum(args),
2060 : Int16GetDatum(jointype),
2061 : PointerGetDatum(sjinfo)));
2062 : else
2063 58956 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2064 : clause->inputcollid,
2065 : PointerGetDatum(root),
2066 : ObjectIdGetDatum(operator),
2067 : PointerGetDatum(args),
2068 : Int32GetDatum(varRelid)));
2069 :
2070 58956 : if (useOr)
2071 : {
2072 50558 : s1 = s1 + s2 - s1 * s2;
2073 50558 : if (isEquality)
2074 49514 : s1disjoint += s2;
2075 : }
2076 : else
2077 : {
2078 8398 : s1 = s1 * s2;
2079 8398 : if (isInequality)
2080 8086 : s1disjoint += s2 - 1.0;
2081 : }
2082 : }
2083 :
2084 : /* accept disjoint-probability estimate if in range */
2085 18066 : if ((useOr ? isEquality : isInequality) &&
2086 17404 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2087 17374 : s1 = s1disjoint;
2088 : }
2089 4134 : else if (rightop && IsA(rightop, ArrayExpr) &&
2090 294 : !((ArrayExpr *) rightop)->multidims)
2091 294 : {
2092 294 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2093 : int16 elmlen;
2094 : bool elmbyval;
2095 : ListCell *l;
2096 :
2097 294 : get_typlenbyval(arrayexpr->element_typeid,
2098 : &elmlen, &elmbyval);
2099 :
2100 : /*
2101 : * We use the assumption of disjoint probabilities here too, although
2102 : * the odds of equal array elements are rather higher if the elements
2103 : * are not all constants (which they won't be, else constant folding
2104 : * would have reduced the ArrayExpr to a Const). In this path it's
2105 : * critical to have the sanity check on the s1disjoint estimate.
2106 : */
2107 294 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2108 :
2109 1060 : foreach(l, arrayexpr->elements)
2110 : {
2111 766 : Node *elem = (Node *) lfirst(l);
2112 : List *args;
2113 : Selectivity s2;
2114 :
2115 : /*
2116 : * Theoretically, if elem isn't of nominal_element_type we should
2117 : * insert a RelabelType, but it seems unlikely that any operator
2118 : * estimation function would really care ...
2119 : */
2120 766 : args = list_make2(leftop, elem);
2121 766 : if (is_join_clause)
2122 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2123 : clause->inputcollid,
2124 : PointerGetDatum(root),
2125 : ObjectIdGetDatum(operator),
2126 : PointerGetDatum(args),
2127 : Int16GetDatum(jointype),
2128 : PointerGetDatum(sjinfo)));
2129 : else
2130 766 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2131 : clause->inputcollid,
2132 : PointerGetDatum(root),
2133 : ObjectIdGetDatum(operator),
2134 : PointerGetDatum(args),
2135 : Int32GetDatum(varRelid)));
2136 :
2137 766 : if (useOr)
2138 : {
2139 766 : s1 = s1 + s2 - s1 * s2;
2140 766 : if (isEquality)
2141 766 : s1disjoint += s2;
2142 : }
2143 : else
2144 : {
2145 0 : s1 = s1 * s2;
2146 0 : if (isInequality)
2147 0 : s1disjoint += s2 - 1.0;
2148 : }
2149 : }
2150 :
2151 : /* accept disjoint-probability estimate if in range */
2152 294 : if ((useOr ? isEquality : isInequality) &&
2153 294 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2154 294 : s1 = s1disjoint;
2155 : }
2156 : else
2157 : {
2158 : CaseTestExpr *dummyexpr;
2159 : List *args;
2160 : Selectivity s2;
2161 : int i;
2162 :
2163 : /*
2164 : * We need a dummy rightop to pass to the operator selectivity
2165 : * routine. It can be pretty much anything that doesn't look like a
2166 : * constant; CaseTestExpr is a convenient choice.
2167 : */
2168 3840 : dummyexpr = makeNode(CaseTestExpr);
2169 3840 : dummyexpr->typeId = nominal_element_type;
2170 3840 : dummyexpr->typeMod = -1;
2171 3840 : dummyexpr->collation = clause->inputcollid;
2172 3840 : args = list_make2(leftop, dummyexpr);
2173 3840 : if (is_join_clause)
2174 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2175 : clause->inputcollid,
2176 : PointerGetDatum(root),
2177 : ObjectIdGetDatum(operator),
2178 : PointerGetDatum(args),
2179 : Int16GetDatum(jointype),
2180 : PointerGetDatum(sjinfo)));
2181 : else
2182 3840 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2183 : clause->inputcollid,
2184 : PointerGetDatum(root),
2185 : ObjectIdGetDatum(operator),
2186 : PointerGetDatum(args),
2187 : Int32GetDatum(varRelid)));
2188 3840 : s1 = useOr ? 0.0 : 1.0;
2189 :
2190 : /*
2191 : * Arbitrarily assume 10 elements in the eventual array value (see
2192 : * also estimate_array_length). We don't risk an assumption of
2193 : * disjoint probabilities here.
2194 : */
2195 42240 : for (i = 0; i < 10; i++)
2196 : {
2197 38400 : if (useOr)
2198 38400 : s1 = s1 + s2 - s1 * s2;
2199 : else
2200 0 : s1 = s1 * s2;
2201 : }
2202 : }
2203 :
2204 : /* result should be in range, but make sure... */
2205 22200 : CLAMP_PROBABILITY(s1);
2206 :
2207 22200 : return s1;
2208 : }
2209 :
2210 : /*
2211 : * Estimate number of elements in the array yielded by an expression.
2212 : *
2213 : * Note: the result is integral, but we use "double" to avoid overflow
2214 : * concerns. Most callers will use it in double-type expressions anyway.
2215 : *
2216 : * Note: in some code paths root can be passed as NULL, resulting in
2217 : * slightly worse estimates.
2218 : */
2219 : double
2220 105284 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2221 : {
2222 : /* look through any binary-compatible relabeling of arrayexpr */
2223 105284 : arrayexpr = strip_array_coercion(arrayexpr);
2224 :
2225 105284 : if (arrayexpr && IsA(arrayexpr, Const))
2226 : {
2227 47438 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2228 47438 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2229 : ArrayType *arrayval;
2230 :
2231 47438 : if (arrayisnull)
2232 90 : return 0;
2233 47348 : arrayval = DatumGetArrayTypeP(arraydatum);
2234 47348 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2235 : }
2236 57846 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2237 498 : !((ArrayExpr *) arrayexpr)->multidims)
2238 : {
2239 498 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2240 : }
2241 57348 : else if (arrayexpr && root)
2242 : {
2243 : /* See if we can find any statistics about it */
2244 : VariableStatData vardata;
2245 : AttStatsSlot sslot;
2246 57324 : double nelem = 0;
2247 :
2248 57324 : examine_variable(root, arrayexpr, 0, &vardata);
2249 57324 : if (HeapTupleIsValid(vardata.statsTuple))
2250 : {
2251 : /*
2252 : * Found stats, so use the average element count, which is stored
2253 : * in the last stanumbers element of the DECHIST statistics.
2254 : * Actually that is the average count of *distinct* elements;
2255 : * perhaps we should scale it up somewhat?
2256 : */
2257 10476 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2258 : STATISTIC_KIND_DECHIST, InvalidOid,
2259 : ATTSTATSSLOT_NUMBERS))
2260 : {
2261 10362 : if (sslot.nnumbers > 0)
2262 10362 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2263 10362 : free_attstatsslot(&sslot);
2264 : }
2265 : }
2266 57324 : ReleaseVariableStats(vardata);
2267 :
2268 57324 : if (nelem > 0)
2269 10362 : return nelem;
2270 : }
2271 :
2272 : /* Else use a default guess --- this should match scalararraysel */
2273 46986 : return 10;
2274 : }
2275 :
2276 : /*
2277 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2278 : *
2279 : * We estimate RowCompare selectivity by considering just the first (high
2280 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2281 : * this estimate could be refined by considering additional columns, it
2282 : * seems unlikely that we could do a lot better without multi-column
2283 : * statistics.
2284 : */
2285 : Selectivity
2286 252 : rowcomparesel(PlannerInfo *root,
2287 : RowCompareExpr *clause,
2288 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2289 : {
2290 : Selectivity s1;
2291 252 : Oid opno = linitial_oid(clause->opnos);
2292 252 : Oid inputcollid = linitial_oid(clause->inputcollids);
2293 : List *opargs;
2294 : bool is_join_clause;
2295 :
2296 : /* Build equivalent arg list for single operator */
2297 252 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2298 :
2299 : /*
2300 : * Decide if it's a join clause. This should match clausesel.c's
2301 : * treat_as_join_clause(), except that we intentionally consider only the
2302 : * leading columns and not the rest of the clause.
2303 : */
2304 252 : if (varRelid != 0)
2305 : {
2306 : /*
2307 : * Caller is forcing restriction mode (eg, because we are examining an
2308 : * inner indexscan qual).
2309 : */
2310 54 : is_join_clause = false;
2311 : }
2312 198 : else if (sjinfo == NULL)
2313 : {
2314 : /*
2315 : * It must be a restriction clause, since it's being evaluated at a
2316 : * scan node.
2317 : */
2318 186 : is_join_clause = false;
2319 : }
2320 : else
2321 : {
2322 : /*
2323 : * Otherwise, it's a join if there's more than one base relation used.
2324 : */
2325 12 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2326 : }
2327 :
2328 252 : if (is_join_clause)
2329 : {
2330 : /* Estimate selectivity for a join clause. */
2331 12 : s1 = join_selectivity(root, opno,
2332 : opargs,
2333 : inputcollid,
2334 : jointype,
2335 : sjinfo);
2336 : }
2337 : else
2338 : {
2339 : /* Estimate selectivity for a restriction clause. */
2340 240 : s1 = restriction_selectivity(root, opno,
2341 : opargs,
2342 : inputcollid,
2343 : varRelid);
2344 : }
2345 :
2346 252 : return s1;
2347 : }
2348 :
2349 : /*
2350 : * eqjoinsel - Join selectivity of "="
2351 : */
2352 : Datum
2353 264360 : eqjoinsel(PG_FUNCTION_ARGS)
2354 : {
2355 264360 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2356 264360 : Oid operator = PG_GETARG_OID(1);
2357 264360 : List *args = (List *) PG_GETARG_POINTER(2);
2358 :
2359 : #ifdef NOT_USED
2360 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2361 : #endif
2362 264360 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2363 264360 : Oid collation = PG_GET_COLLATION();
2364 : double selec;
2365 : double selec_inner;
2366 : VariableStatData vardata1;
2367 : VariableStatData vardata2;
2368 : double nd1;
2369 : double nd2;
2370 : bool isdefault1;
2371 : bool isdefault2;
2372 : Oid opfuncoid;
2373 : FmgrInfo eqproc;
2374 264360 : Oid hashLeft = InvalidOid;
2375 264360 : Oid hashRight = InvalidOid;
2376 : AttStatsSlot sslot1;
2377 : AttStatsSlot sslot2;
2378 264360 : Form_pg_statistic stats1 = NULL;
2379 264360 : Form_pg_statistic stats2 = NULL;
2380 264360 : bool have_mcvs1 = false;
2381 264360 : bool have_mcvs2 = false;
2382 264360 : bool *hasmatch1 = NULL;
2383 264360 : bool *hasmatch2 = NULL;
2384 264360 : int nmatches = 0;
2385 : bool get_mcv_stats;
2386 : bool join_is_reversed;
2387 : RelOptInfo *inner_rel;
2388 :
2389 264360 : get_join_variables(root, args, sjinfo,
2390 : &vardata1, &vardata2, &join_is_reversed);
2391 :
2392 264360 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2393 264360 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2394 :
2395 264360 : opfuncoid = get_opcode(operator);
2396 :
2397 264360 : memset(&sslot1, 0, sizeof(sslot1));
2398 264360 : memset(&sslot2, 0, sizeof(sslot2));
2399 :
2400 : /*
2401 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2402 : * other side, so do a quick check to verify that both stats exist.
2403 : */
2404 728968 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2405 357184 : HeapTupleIsValid(vardata2.statsTuple) &&
2406 156936 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2407 : STATISTIC_KIND_MCV, InvalidOid,
2408 464608 : 0) &&
2409 71950 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2410 : STATISTIC_KIND_MCV, InvalidOid,
2411 : 0));
2412 :
2413 264360 : if (HeapTupleIsValid(vardata1.statsTuple))
2414 : {
2415 : /* note we allow use of nullfrac regardless of security check */
2416 200248 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2417 228676 : if (get_mcv_stats &&
2418 28428 : statistic_proc_security_check(&vardata1, opfuncoid))
2419 28428 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2420 : STATISTIC_KIND_MCV, InvalidOid,
2421 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2422 : }
2423 :
2424 264360 : if (HeapTupleIsValid(vardata2.statsTuple))
2425 : {
2426 : /* note we allow use of nullfrac regardless of security check */
2427 177390 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2428 205818 : if (get_mcv_stats &&
2429 28428 : statistic_proc_security_check(&vardata2, opfuncoid))
2430 28428 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2431 : STATISTIC_KIND_MCV, InvalidOid,
2432 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2433 : }
2434 :
2435 : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
2436 264360 : if (have_mcvs1 && have_mcvs2)
2437 : {
2438 28428 : fmgr_info(opfuncoid, &eqproc);
2439 28428 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2440 28428 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2441 :
2442 : /*
2443 : * If the MCV lists are long enough to justify hashing, try to look up
2444 : * hash functions for the join operator.
2445 : */
2446 28428 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2447 1656 : (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
2448 : }
2449 : else
2450 235932 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2451 :
2452 : /* We need to compute the inner-join selectivity in all cases */
2453 264360 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2454 : hashLeft, hashRight,
2455 : &vardata1, &vardata2,
2456 : nd1, nd2,
2457 : isdefault1, isdefault2,
2458 : &sslot1, &sslot2,
2459 : stats1, stats2,
2460 : have_mcvs1, have_mcvs2,
2461 : hasmatch1, hasmatch2,
2462 : &nmatches);
2463 :
2464 264360 : switch (sjinfo->jointype)
2465 : {
2466 253646 : case JOIN_INNER:
2467 : case JOIN_LEFT:
2468 : case JOIN_FULL:
2469 253646 : selec = selec_inner;
2470 253646 : break;
2471 10714 : case JOIN_SEMI:
2472 : case JOIN_ANTI:
2473 :
2474 : /*
2475 : * Look up the join's inner relation. min_righthand is sufficient
2476 : * information because neither SEMI nor ANTI joins permit any
2477 : * reassociation into or out of their RHS, so the righthand will
2478 : * always be exactly that set of rels.
2479 : */
2480 10714 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2481 :
2482 10714 : if (!join_is_reversed)
2483 6598 : selec = eqjoinsel_semi(&eqproc, collation,
2484 : hashLeft, hashRight,
2485 : false,
2486 : &vardata1, &vardata2,
2487 : nd1, nd2,
2488 : isdefault1, isdefault2,
2489 : &sslot1, &sslot2,
2490 : stats1, stats2,
2491 : have_mcvs1, have_mcvs2,
2492 : hasmatch1, hasmatch2,
2493 : &nmatches,
2494 : inner_rel);
2495 : else
2496 4116 : selec = eqjoinsel_semi(&eqproc, collation,
2497 : hashLeft, hashRight,
2498 : true,
2499 : &vardata2, &vardata1,
2500 : nd2, nd1,
2501 : isdefault2, isdefault1,
2502 : &sslot2, &sslot1,
2503 : stats2, stats1,
2504 : have_mcvs2, have_mcvs1,
2505 : hasmatch2, hasmatch1,
2506 : &nmatches,
2507 : inner_rel);
2508 :
2509 : /*
2510 : * We should never estimate the output of a semijoin to be more
2511 : * rows than we estimate for an inner join with the same input
2512 : * rels and join condition; it's obviously impossible for that to
2513 : * happen. The former estimate is N1 * Ssemi while the latter is
2514 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2515 : * this is worthwhile because of the shakier estimation rules we
2516 : * use in eqjoinsel_semi, particularly in cases where it has to
2517 : * punt entirely.
2518 : */
2519 10714 : selec = Min(selec, inner_rel->rows * selec_inner);
2520 10714 : break;
2521 0 : default:
2522 : /* other values not expected here */
2523 0 : elog(ERROR, "unrecognized join type: %d",
2524 : (int) sjinfo->jointype);
2525 : selec = 0; /* keep compiler quiet */
2526 : break;
2527 : }
2528 :
2529 264360 : free_attstatsslot(&sslot1);
2530 264360 : free_attstatsslot(&sslot2);
2531 :
2532 264360 : ReleaseVariableStats(vardata1);
2533 264360 : ReleaseVariableStats(vardata2);
2534 :
2535 264360 : if (hasmatch1)
2536 28428 : pfree(hasmatch1);
2537 264360 : if (hasmatch2)
2538 28428 : pfree(hasmatch2);
2539 :
2540 264360 : CLAMP_PROBABILITY(selec);
2541 :
2542 264360 : PG_RETURN_FLOAT8((float8) selec);
2543 : }
2544 :
2545 : /*
2546 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2547 : *
2548 : * In addition to computing the selectivity estimate, this will fill
2549 : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2550 : * We may be able to re-use that data in eqjoinsel_semi.
2551 : *
2552 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2553 : * that it's worth trying to distinguish them here.
2554 : */
2555 : static double
2556 264360 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2557 : Oid hashLeft, Oid hashRight,
2558 : VariableStatData *vardata1, VariableStatData *vardata2,
2559 : double nd1, double nd2,
2560 : bool isdefault1, bool isdefault2,
2561 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2562 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2563 : bool have_mcvs1, bool have_mcvs2,
2564 : bool *hasmatch1, bool *hasmatch2,
2565 : int *p_nmatches)
2566 : {
2567 : double selec;
2568 :
2569 264360 : if (have_mcvs1 && have_mcvs2)
2570 28428 : {
2571 : /*
2572 : * We have most-common-value lists for both relations. Run through
2573 : * the lists to see which MCVs actually join to each other with the
2574 : * given operator. This allows us to determine the exact join
2575 : * selectivity for the portion of the relations represented by the MCV
2576 : * lists. We still have to estimate for the remaining population, but
2577 : * in a skewed distribution this gives us a big leg up in accuracy.
2578 : * For motivation see the analysis in Y. Ioannidis and S.
2579 : * Christodoulakis, "On the propagation of errors in the size of join
2580 : * results", Technical Report 1018, Computer Science Dept., University
2581 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2582 : */
2583 28428 : double nullfrac1 = stats1->stanullfrac;
2584 28428 : double nullfrac2 = stats2->stanullfrac;
2585 : double matchprodfreq,
2586 : matchfreq1,
2587 : matchfreq2,
2588 : unmatchfreq1,
2589 : unmatchfreq2,
2590 : otherfreq1,
2591 : otherfreq2,
2592 : totalsel1,
2593 : totalsel2;
2594 : int i,
2595 : nmatches;
2596 :
2597 : /* Fill the match arrays */
2598 28428 : eqjoinsel_find_matches(eqproc, collation,
2599 : hashLeft, hashRight,
2600 : false,
2601 : sslot1, sslot2,
2602 : sslot1->nvalues, sslot2->nvalues,
2603 : hasmatch1, hasmatch2,
2604 : p_nmatches, &matchprodfreq);
2605 28428 : nmatches = *p_nmatches;
2606 28428 : CLAMP_PROBABILITY(matchprodfreq);
2607 :
2608 : /* Sum up frequencies of matched and unmatched MCVs */
2609 28428 : matchfreq1 = unmatchfreq1 = 0.0;
2610 707960 : for (i = 0; i < sslot1->nvalues; i++)
2611 : {
2612 679532 : if (hasmatch1[i])
2613 293448 : matchfreq1 += sslot1->numbers[i];
2614 : else
2615 386084 : unmatchfreq1 += sslot1->numbers[i];
2616 : }
2617 28428 : CLAMP_PROBABILITY(matchfreq1);
2618 28428 : CLAMP_PROBABILITY(unmatchfreq1);
2619 28428 : matchfreq2 = unmatchfreq2 = 0.0;
2620 531604 : for (i = 0; i < sslot2->nvalues; i++)
2621 : {
2622 503176 : if (hasmatch2[i])
2623 293448 : matchfreq2 += sslot2->numbers[i];
2624 : else
2625 209728 : unmatchfreq2 += sslot2->numbers[i];
2626 : }
2627 28428 : CLAMP_PROBABILITY(matchfreq2);
2628 28428 : CLAMP_PROBABILITY(unmatchfreq2);
2629 :
2630 : /*
2631 : * Compute total frequency of non-null values that are not in the MCV
2632 : * lists.
2633 : */
2634 28428 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2635 28428 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2636 28428 : CLAMP_PROBABILITY(otherfreq1);
2637 28428 : CLAMP_PROBABILITY(otherfreq2);
2638 :
2639 : /*
2640 : * We can estimate the total selectivity from the point of view of
2641 : * relation 1 as: the known selectivity for matched MCVs, plus
2642 : * unmatched MCVs that are assumed to match against random members of
2643 : * relation 2's non-MCV population, plus non-MCV values that are
2644 : * assumed to match against random members of relation 2's unmatched
2645 : * MCVs plus non-MCV values.
2646 : */
2647 28428 : totalsel1 = matchprodfreq;
2648 28428 : if (nd2 > sslot2->nvalues)
2649 6114 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2650 28428 : if (nd2 > nmatches)
2651 11064 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2652 11064 : (nd2 - nmatches);
2653 : /* Same estimate from the point of view of relation 2. */
2654 28428 : totalsel2 = matchprodfreq;
2655 28428 : if (nd1 > sslot1->nvalues)
2656 7044 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2657 28428 : if (nd1 > nmatches)
2658 9792 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2659 9792 : (nd1 - nmatches);
2660 :
2661 : /*
2662 : * Use the smaller of the two estimates. This can be justified in
2663 : * essentially the same terms as given below for the no-stats case: to
2664 : * a first approximation, we are estimating from the point of view of
2665 : * the relation with smaller nd.
2666 : */
2667 28428 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2668 : }
2669 : else
2670 : {
2671 : /*
2672 : * We do not have MCV lists for both sides. Estimate the join
2673 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2674 : * is plausible if we assume that the join operator is strict and the
2675 : * non-null values are about equally distributed: a given non-null
2676 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2677 : * of rel2, so total join rows are at most
2678 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2679 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2680 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2681 : * with MIN() is an upper bound. Using the MIN() means we estimate
2682 : * from the point of view of the relation with smaller nd (since the
2683 : * larger nd is determining the MIN). It is reasonable to assume that
2684 : * most tuples in this rel will have join partners, so the bound is
2685 : * probably reasonably tight and should be taken as-is.
2686 : *
2687 : * XXX Can we be smarter if we have an MCV list for just one side? It
2688 : * seems that if we assume equal distribution for the other side, we
2689 : * end up with the same answer anyway.
2690 : */
2691 235932 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2692 235932 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2693 :
2694 235932 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2695 235932 : if (nd1 > nd2)
2696 123982 : selec /= nd1;
2697 : else
2698 111950 : selec /= nd2;
2699 : }
2700 :
2701 264360 : return selec;
2702 : }
2703 :
2704 : /*
2705 : * eqjoinsel_semi --- eqjoinsel for semi join
2706 : *
2707 : * (Also used for anti join, which we are supposed to estimate the same way.)
2708 : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2709 : * is for the original join operator, which might now need to have the inputs
2710 : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2711 : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2712 : * eqjoinsel_inner.
2713 : */
2714 : static double
2715 10714 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2716 : Oid hashLeft, Oid hashRight,
2717 : bool op_is_reversed,
2718 : VariableStatData *vardata1, VariableStatData *vardata2,
2719 : double nd1, double nd2,
2720 : bool isdefault1, bool isdefault2,
2721 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2722 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2723 : bool have_mcvs1, bool have_mcvs2,
2724 : bool *hasmatch1, bool *hasmatch2,
2725 : int *p_nmatches,
2726 : RelOptInfo *inner_rel)
2727 : {
2728 : double selec;
2729 :
2730 : /*
2731 : * We clamp nd2 to be not more than what we estimate the inner relation's
2732 : * size to be. This is intuitively somewhat reasonable since obviously
2733 : * there can't be more than that many distinct values coming from the
2734 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2735 : * likewise) is that this is the only pathway by which restriction clauses
2736 : * applied to the inner rel will affect the join result size estimate,
2737 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2738 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2739 : * the selectivity of outer-rel restrictions.
2740 : *
2741 : * We can apply this clamping both with respect to the base relation from
2742 : * which the join variable comes (if there is just one), and to the
2743 : * immediate inner input relation of the current join.
2744 : *
2745 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2746 : * great, maybe, but it didn't come out of nowhere either. This is most
2747 : * helpful when the inner relation is empty and consequently has no stats.
2748 : */
2749 10714 : if (vardata2->rel)
2750 : {
2751 10708 : if (nd2 >= vardata2->rel->rows)
2752 : {
2753 8580 : nd2 = vardata2->rel->rows;
2754 8580 : isdefault2 = false;
2755 : }
2756 : }
2757 10714 : if (nd2 >= inner_rel->rows)
2758 : {
2759 8522 : nd2 = inner_rel->rows;
2760 8522 : isdefault2 = false;
2761 : }
2762 :
2763 10714 : if (have_mcvs1 && have_mcvs2)
2764 624 : {
2765 : /*
2766 : * We have most-common-value lists for both relations. Run through
2767 : * the lists to see which MCVs actually join to each other with the
2768 : * given operator. This allows us to determine the exact join
2769 : * selectivity for the portion of the relations represented by the MCV
2770 : * lists. We still have to estimate for the remaining population, but
2771 : * in a skewed distribution this gives us a big leg up in accuracy.
2772 : */
2773 624 : double nullfrac1 = stats1->stanullfrac;
2774 : double matchprodfreq,
2775 : matchfreq1,
2776 : uncertainfrac,
2777 : uncertain;
2778 : int i,
2779 : nmatches,
2780 : clamped_nvalues2;
2781 :
2782 : /*
2783 : * The clamping above could have resulted in nd2 being less than
2784 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2785 : * most common values in the relation will appear in the join input,
2786 : * and so compare to only the first nd2 members of the MCV list. Of
2787 : * course this is frequently wrong, but it's the best bet we can make.
2788 : */
2789 624 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2790 :
2791 : /*
2792 : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2793 : * then the hasmatch1[] and hasmatch2[] match flags computed by
2794 : * eqjoinsel_inner are still perfectly applicable, so we need not
2795 : * re-do the matching work. Note that it does not matter if
2796 : * op_is_reversed: we'd get the same answers.
2797 : *
2798 : * If we did clamp, then a different set of sslot2 values is to be
2799 : * compared, so we have to re-do the matching.
2800 : */
2801 624 : if (clamped_nvalues2 != sslot2->nvalues)
2802 : {
2803 : /* Must re-zero the arrays */
2804 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2805 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2806 : /* Re-fill the match arrays */
2807 0 : eqjoinsel_find_matches(eqproc, collation,
2808 : hashLeft, hashRight,
2809 : op_is_reversed,
2810 : sslot1, sslot2,
2811 : sslot1->nvalues, clamped_nvalues2,
2812 : hasmatch1, hasmatch2,
2813 : p_nmatches, &matchprodfreq);
2814 : }
2815 624 : nmatches = *p_nmatches;
2816 :
2817 : /* Sum up frequencies of matched MCVs */
2818 624 : matchfreq1 = 0.0;
2819 13700 : for (i = 0; i < sslot1->nvalues; i++)
2820 : {
2821 13076 : if (hasmatch1[i])
2822 11436 : matchfreq1 += sslot1->numbers[i];
2823 : }
2824 624 : CLAMP_PROBABILITY(matchfreq1);
2825 :
2826 : /*
2827 : * Now we need to estimate the fraction of relation 1 that has at
2828 : * least one join partner. We know for certain that the matched MCVs
2829 : * do, so that gives us a lower bound, but we're really in the dark
2830 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2831 : * assume all non-null rel1 rows have join partners, else assume for
2832 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2833 : * can discount the known-matched MCVs from the distinct-values counts
2834 : * before doing the division.
2835 : *
2836 : * Crude as the above is, it's completely useless if we don't have
2837 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2838 : * nd2 is default, punt and assume half of the uncertain rows have
2839 : * join partners.
2840 : */
2841 624 : if (!isdefault1 && !isdefault2)
2842 : {
2843 624 : nd1 -= nmatches;
2844 624 : nd2 -= nmatches;
2845 624 : if (nd1 <= nd2 || nd2 < 0)
2846 594 : uncertainfrac = 1.0;
2847 : else
2848 30 : uncertainfrac = nd2 / nd1;
2849 : }
2850 : else
2851 0 : uncertainfrac = 0.5;
2852 624 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2853 624 : CLAMP_PROBABILITY(uncertain);
2854 624 : selec = matchfreq1 + uncertainfrac * uncertain;
2855 : }
2856 : else
2857 : {
2858 : /*
2859 : * Without MCV lists for both sides, we can only use the heuristic
2860 : * about nd1 vs nd2.
2861 : */
2862 10090 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2863 :
2864 10090 : if (!isdefault1 && !isdefault2)
2865 : {
2866 7742 : if (nd1 <= nd2 || nd2 < 0)
2867 4910 : selec = 1.0 - nullfrac1;
2868 : else
2869 2832 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2870 : }
2871 : else
2872 2348 : selec = 0.5 * (1.0 - nullfrac1);
2873 : }
2874 :
2875 10714 : return selec;
2876 : }
2877 :
2878 : /*
2879 : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2880 : *
2881 : * Inputs:
2882 : * eqproc: FmgrInfo for equality function to use (might be reversed)
2883 : * collation: OID of collation to use
2884 : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2885 : * or InvalidOid if we're not to use hashing
2886 : * op_is_reversed: indicates that eqproc compares right type to left type
2887 : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2888 : * nvalues1, nvalues2: number of values to be considered (can be less than
2889 : * sslotN->nvalues, but not more)
2890 : * Outputs:
2891 : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2892 : * entries are set to true if that MCV has a match on the other side
2893 : * *p_nmatches: receives number of MCV pairs that match
2894 : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2895 : * for matching MCVs
2896 : *
2897 : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2898 : * for its right, regardless of op_is_reversed.
2899 : *
2900 : * Note we assume that each MCV will match at most one member of the other
2901 : * MCV list. If the operator isn't really equality, there could be multiple
2902 : * matches --- but we don't look for them, both for speed and because the
2903 : * math wouldn't add up...
2904 : */
2905 : static void
2906 28428 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2907 : Oid hashLeft, Oid hashRight,
2908 : bool op_is_reversed,
2909 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2910 : int nvalues1, int nvalues2,
2911 : bool *hasmatch1, bool *hasmatch2,
2912 : int *p_nmatches, double *p_matchprodfreq)
2913 : {
2914 28428 : LOCAL_FCINFO(fcinfo, 2);
2915 28428 : double matchprodfreq = 0.0;
2916 28428 : int nmatches = 0;
2917 :
2918 : /*
2919 : * Save a few cycles by setting up the fcinfo struct just once. Using
2920 : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2921 : * NULL, though really equality functions should never do that.
2922 : */
2923 28428 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2924 : NULL, NULL);
2925 28428 : fcinfo->args[0].isnull = false;
2926 28428 : fcinfo->args[1].isnull = false;
2927 :
2928 28428 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2929 1656 : {
2930 : /* Use a hash table to speed up the matching */
2931 1656 : LOCAL_FCINFO(hash_fcinfo, 1);
2932 : FmgrInfo hash_proc;
2933 : MCVHashContext hashContext;
2934 : MCVHashTable_hash *hashTable;
2935 : AttStatsSlot *statsProbe;
2936 : AttStatsSlot *statsHash;
2937 : bool *hasMatchProbe;
2938 : bool *hasMatchHash;
2939 : int nvaluesProbe;
2940 : int nvaluesHash;
2941 :
2942 : /* Make sure we build the hash table on the smaller array. */
2943 1656 : if (sslot1->nvalues >= sslot2->nvalues)
2944 : {
2945 1656 : statsProbe = sslot1;
2946 1656 : statsHash = sslot2;
2947 1656 : hasMatchProbe = hasmatch1;
2948 1656 : hasMatchHash = hasmatch2;
2949 1656 : nvaluesProbe = nvalues1;
2950 1656 : nvaluesHash = nvalues2;
2951 : }
2952 : else
2953 : {
2954 : /* We'll have to reverse the direction of use of the operator. */
2955 0 : op_is_reversed = !op_is_reversed;
2956 0 : statsProbe = sslot2;
2957 0 : statsHash = sslot1;
2958 0 : hasMatchProbe = hasmatch2;
2959 0 : hasMatchHash = hasmatch1;
2960 0 : nvaluesProbe = nvalues2;
2961 0 : nvaluesHash = nvalues1;
2962 : }
2963 :
2964 : /*
2965 : * Build the hash table on the smaller array, using the appropriate
2966 : * hash function for its data type.
2967 : */
2968 1656 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
2969 1656 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
2970 : NULL, NULL);
2971 1656 : hash_fcinfo->args[0].isnull = false;
2972 :
2973 1656 : hashContext.equal_fcinfo = fcinfo;
2974 1656 : hashContext.hash_fcinfo = hash_fcinfo;
2975 1656 : hashContext.op_is_reversed = op_is_reversed;
2976 1656 : hashContext.insert_mode = true;
2977 1656 : get_typlenbyval(statsHash->valuetype,
2978 : &hashContext.hash_typlen,
2979 : &hashContext.hash_typbyval);
2980 :
2981 1656 : hashTable = MCVHashTable_create(CurrentMemoryContext,
2982 : nvaluesHash,
2983 : &hashContext);
2984 :
2985 167256 : for (int i = 0; i < nvaluesHash; i++)
2986 : {
2987 165600 : bool found = false;
2988 165600 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
2989 165600 : statsHash->values[i],
2990 : &found);
2991 :
2992 : /*
2993 : * MCVHashTable_insert will only report "found" if the new value
2994 : * is equal to some previous one per datum_image_eq(). That
2995 : * probably shouldn't happen, since we're not expecting duplicates
2996 : * in the MCV list. If we do find a dup, just ignore it, leaving
2997 : * the hash entry's index pointing at the first occurrence. That
2998 : * matches the behavior that the non-hashed code path would have.
2999 : */
3000 165600 : if (likely(!found))
3001 165600 : entry->index = i;
3002 : }
3003 :
3004 : /*
3005 : * Prepare to probe the hash table. If the probe values are of a
3006 : * different data type, then we need to change hash functions. (This
3007 : * code relies on the assumption that since we defined SH_STORE_HASH,
3008 : * simplehash.h will never need to compute hash values for existing
3009 : * hash table entries.)
3010 : */
3011 1656 : hashContext.insert_mode = false;
3012 1656 : if (hashLeft != hashRight)
3013 : {
3014 0 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3015 : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3016 0 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3017 : NULL, NULL);
3018 0 : hash_fcinfo->args[0].isnull = false;
3019 : }
3020 :
3021 : /* Look up each probe value in turn. */
3022 167256 : for (int i = 0; i < nvaluesProbe; i++)
3023 : {
3024 165600 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3025 165600 : statsProbe->values[i]);
3026 :
3027 : /* As in the other code path, skip already-matched hash entries */
3028 165600 : if (entry != NULL && !hasMatchHash[entry->index])
3029 : {
3030 64856 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3031 64856 : nmatches++;
3032 64856 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3033 : }
3034 : }
3035 :
3036 1656 : MCVHashTable_destroy(hashTable);
3037 : }
3038 : else
3039 : {
3040 : /* We're not to use hashing, so do it the O(N^2) way */
3041 : int index1,
3042 : index2;
3043 :
3044 : /* Set up to supply the values in the order the operator expects */
3045 26772 : if (op_is_reversed)
3046 : {
3047 0 : index1 = 1;
3048 0 : index2 = 0;
3049 : }
3050 : else
3051 : {
3052 26772 : index1 = 0;
3053 26772 : index2 = 1;
3054 : }
3055 :
3056 540704 : for (int i = 0; i < nvalues1; i++)
3057 : {
3058 513932 : fcinfo->args[index1].value = sslot1->values[i];
3059 :
3060 10448060 : for (int j = 0; j < nvalues2; j++)
3061 : {
3062 : Datum fresult;
3063 :
3064 10162720 : if (hasmatch2[j])
3065 3119544 : continue;
3066 7043176 : fcinfo->args[index2].value = sslot2->values[j];
3067 7043176 : fcinfo->isnull = false;
3068 7043176 : fresult = FunctionCallInvoke(fcinfo);
3069 7043176 : if (!fcinfo->isnull && DatumGetBool(fresult))
3070 : {
3071 228592 : hasmatch1[i] = hasmatch2[j] = true;
3072 228592 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3073 228592 : nmatches++;
3074 228592 : break;
3075 : }
3076 : }
3077 : }
3078 : }
3079 :
3080 28428 : *p_nmatches = nmatches;
3081 28428 : *p_matchprodfreq = matchprodfreq;
3082 28428 : }
3083 :
3084 : /*
3085 : * Support functions for the hash tables used by eqjoinsel_find_matches
3086 : */
3087 : static uint32
3088 331200 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3089 : {
3090 331200 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3091 331200 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3092 : Datum fresult;
3093 :
3094 331200 : fcinfo->args[0].value = key;
3095 331200 : fcinfo->isnull = false;
3096 331200 : fresult = FunctionCallInvoke(fcinfo);
3097 : Assert(!fcinfo->isnull);
3098 331200 : return DatumGetUInt32(fresult);
3099 : }
3100 :
3101 : static bool
3102 64856 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3103 : {
3104 64856 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3105 :
3106 64856 : if (context->insert_mode)
3107 : {
3108 : /*
3109 : * During the insertion step, any comparisons will be between two
3110 : * Datums of the hash table's data type, so if the given operator is
3111 : * cross-type it will be the wrong thing to use. Fortunately, we can
3112 : * use datum_image_eq instead. The MCV values should all be distinct
3113 : * anyway, so it's mostly pro-forma to compare them at all.
3114 : */
3115 0 : return datum_image_eq(key0, key1,
3116 0 : context->hash_typbyval, context->hash_typlen);
3117 : }
3118 : else
3119 : {
3120 64856 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3121 : Datum fresult;
3122 :
3123 : /*
3124 : * Apply the operator the correct way around. Although simplehash.h
3125 : * doesn't document this explicitly, during lookups key0 is from the
3126 : * hash table while key1 is the probe value, so we should compare them
3127 : * in that order only if op_is_reversed.
3128 : */
3129 64856 : if (context->op_is_reversed)
3130 : {
3131 0 : fcinfo->args[0].value = key0;
3132 0 : fcinfo->args[1].value = key1;
3133 : }
3134 : else
3135 : {
3136 64856 : fcinfo->args[0].value = key1;
3137 64856 : fcinfo->args[1].value = key0;
3138 : }
3139 64856 : fcinfo->isnull = false;
3140 64856 : fresult = FunctionCallInvoke(fcinfo);
3141 64856 : return (!fcinfo->isnull && DatumGetBool(fresult));
3142 : }
3143 : }
3144 :
3145 : /*
3146 : * neqjoinsel - Join selectivity of "!="
3147 : */
3148 : Datum
3149 3742 : neqjoinsel(PG_FUNCTION_ARGS)
3150 : {
3151 3742 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3152 3742 : Oid operator = PG_GETARG_OID(1);
3153 3742 : List *args = (List *) PG_GETARG_POINTER(2);
3154 3742 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
3155 3742 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
3156 3742 : Oid collation = PG_GET_COLLATION();
3157 : float8 result;
3158 :
3159 3742 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
3160 1266 : {
3161 : /*
3162 : * For semi-joins, if there is more than one distinct value in the RHS
3163 : * relation then every non-null LHS row must find a row to join since
3164 : * it can only be equal to one of them. We'll assume that there is
3165 : * always more than one distinct RHS value for the sake of stability,
3166 : * though in theory we could have special cases for empty RHS
3167 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3168 : * fraction of LHS that has the same value as the single RHS value).
3169 : *
3170 : * For anti-joins, if we use the same assumption that there is more
3171 : * than one distinct key in the RHS relation, then every non-null LHS
3172 : * row must be suppressed by the anti-join.
3173 : *
3174 : * So either way, the selectivity estimate should be 1 - nullfrac.
3175 : */
3176 : VariableStatData leftvar;
3177 : VariableStatData rightvar;
3178 : bool reversed;
3179 : HeapTuple statsTuple;
3180 : double nullfrac;
3181 :
3182 1266 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3183 1266 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3184 1266 : if (HeapTupleIsValid(statsTuple))
3185 1034 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3186 : else
3187 232 : nullfrac = 0.0;
3188 1266 : ReleaseVariableStats(leftvar);
3189 1266 : ReleaseVariableStats(rightvar);
3190 :
3191 1266 : result = 1.0 - nullfrac;
3192 : }
3193 : else
3194 : {
3195 : /*
3196 : * We want 1 - eqjoinsel() where the equality operator is the one
3197 : * associated with this != operator, that is, its negator.
3198 : */
3199 2476 : Oid eqop = get_negator(operator);
3200 :
3201 2476 : if (eqop)
3202 : {
3203 : result =
3204 2476 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3205 : collation,
3206 : PointerGetDatum(root),
3207 : ObjectIdGetDatum(eqop),
3208 : PointerGetDatum(args),
3209 : Int16GetDatum(jointype),
3210 : PointerGetDatum(sjinfo)));
3211 : }
3212 : else
3213 : {
3214 : /* Use default selectivity (should we raise an error instead?) */
3215 0 : result = DEFAULT_EQ_SEL;
3216 : }
3217 2476 : result = 1.0 - result;
3218 : }
3219 :
3220 3742 : PG_RETURN_FLOAT8(result);
3221 : }
3222 :
3223 : /*
3224 : * scalarltjoinsel - Join selectivity of "<" for scalars
3225 : */
3226 : Datum
3227 324 : scalarltjoinsel(PG_FUNCTION_ARGS)
3228 : {
3229 324 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3230 : }
3231 :
3232 : /*
3233 : * scalarlejoinsel - Join selectivity of "<=" for scalars
3234 : */
3235 : Datum
3236 276 : scalarlejoinsel(PG_FUNCTION_ARGS)
3237 : {
3238 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3239 : }
3240 :
3241 : /*
3242 : * scalargtjoinsel - Join selectivity of ">" for scalars
3243 : */
3244 : Datum
3245 276 : scalargtjoinsel(PG_FUNCTION_ARGS)
3246 : {
3247 276 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3248 : }
3249 :
3250 : /*
3251 : * scalargejoinsel - Join selectivity of ">=" for scalars
3252 : */
3253 : Datum
3254 184 : scalargejoinsel(PG_FUNCTION_ARGS)
3255 : {
3256 184 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3257 : }
3258 :
3259 :
3260 : /*
3261 : * mergejoinscansel - Scan selectivity of merge join.
3262 : *
3263 : * A merge join will stop as soon as it exhausts either input stream.
3264 : * Therefore, if we can estimate the ranges of both input variables,
3265 : * we can estimate how much of the input will actually be read. This
3266 : * can have a considerable impact on the cost when using indexscans.
3267 : *
3268 : * Also, we can estimate how much of each input has to be read before the
3269 : * first join pair is found, which will affect the join's startup time.
3270 : *
3271 : * clause should be a clause already known to be mergejoinable. opfamily,
3272 : * cmptype, and nulls_first specify the sort ordering being used.
3273 : *
3274 : * The outputs are:
3275 : * *leftstart is set to the fraction of the left-hand variable expected
3276 : * to be scanned before the first join pair is found (0 to 1).
3277 : * *leftend is set to the fraction of the left-hand variable expected
3278 : * to be scanned before the join terminates (0 to 1).
3279 : * *rightstart, *rightend similarly for the right-hand variable.
3280 : */
3281 : void
3282 141368 : mergejoinscansel(PlannerInfo *root, Node *clause,
3283 : Oid opfamily, CompareType cmptype, bool nulls_first,
3284 : Selectivity *leftstart, Selectivity *leftend,
3285 : Selectivity *rightstart, Selectivity *rightend)
3286 : {
3287 : Node *left,
3288 : *right;
3289 : VariableStatData leftvar,
3290 : rightvar;
3291 : Oid opmethod;
3292 : int op_strategy;
3293 : Oid op_lefttype;
3294 : Oid op_righttype;
3295 : Oid opno,
3296 : collation,
3297 : lsortop,
3298 : rsortop,
3299 : lstatop,
3300 : rstatop,
3301 : ltop,
3302 : leop,
3303 : revltop,
3304 : revleop;
3305 : StrategyNumber ltstrat,
3306 : lestrat,
3307 : gtstrat,
3308 : gestrat;
3309 : bool isgt;
3310 : Datum leftmin,
3311 : leftmax,
3312 : rightmin,
3313 : rightmax;
3314 : double selec;
3315 :
3316 : /* Set default results if we can't figure anything out. */
3317 : /* XXX should default "start" fraction be a bit more than 0? */
3318 141368 : *leftstart = *rightstart = 0.0;
3319 141368 : *leftend = *rightend = 1.0;
3320 :
3321 : /* Deconstruct the merge clause */
3322 141368 : if (!is_opclause(clause))
3323 0 : return; /* shouldn't happen */
3324 141368 : opno = ((OpExpr *) clause)->opno;
3325 141368 : collation = ((OpExpr *) clause)->inputcollid;
3326 141368 : left = get_leftop((Expr *) clause);
3327 141368 : right = get_rightop((Expr *) clause);
3328 141368 : if (!right)
3329 0 : return; /* shouldn't happen */
3330 :
3331 : /* Look for stats for the inputs */
3332 141368 : examine_variable(root, left, 0, &leftvar);
3333 141368 : examine_variable(root, right, 0, &rightvar);
3334 :
3335 141368 : opmethod = get_opfamily_method(opfamily);
3336 :
3337 : /* Extract the operator's declared left/right datatypes */
3338 141368 : get_op_opfamily_properties(opno, opfamily, false,
3339 : &op_strategy,
3340 : &op_lefttype,
3341 : &op_righttype);
3342 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3343 :
3344 : /*
3345 : * Look up the various operators we need. If we don't find them all, it
3346 : * probably means the opfamily is broken, but we just fail silently.
3347 : *
3348 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3349 : * operator, regardless of which sort direction we are considering.
3350 : */
3351 141368 : switch (cmptype)
3352 : {
3353 141296 : case COMPARE_LT:
3354 141296 : isgt = false;
3355 141296 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3356 141296 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3357 141296 : if (op_lefttype == op_righttype)
3358 : {
3359 : /* easy case */
3360 139530 : ltop = get_opfamily_member(opfamily,
3361 : op_lefttype, op_righttype,
3362 : ltstrat);
3363 139530 : leop = get_opfamily_member(opfamily,
3364 : op_lefttype, op_righttype,
3365 : lestrat);
3366 139530 : lsortop = ltop;
3367 139530 : rsortop = ltop;
3368 139530 : lstatop = lsortop;
3369 139530 : rstatop = rsortop;
3370 139530 : revltop = ltop;
3371 139530 : revleop = leop;
3372 : }
3373 : else
3374 : {
3375 1766 : ltop = get_opfamily_member(opfamily,
3376 : op_lefttype, op_righttype,
3377 : ltstrat);
3378 1766 : leop = get_opfamily_member(opfamily,
3379 : op_lefttype, op_righttype,
3380 : lestrat);
3381 1766 : lsortop = get_opfamily_member(opfamily,
3382 : op_lefttype, op_lefttype,
3383 : ltstrat);
3384 1766 : rsortop = get_opfamily_member(opfamily,
3385 : op_righttype, op_righttype,
3386 : ltstrat);
3387 1766 : lstatop = lsortop;
3388 1766 : rstatop = rsortop;
3389 1766 : revltop = get_opfamily_member(opfamily,
3390 : op_righttype, op_lefttype,
3391 : ltstrat);
3392 1766 : revleop = get_opfamily_member(opfamily,
3393 : op_righttype, op_lefttype,
3394 : lestrat);
3395 : }
3396 141296 : break;
3397 72 : case COMPARE_GT:
3398 : /* descending-order case */
3399 72 : isgt = true;
3400 72 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3401 72 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3402 72 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3403 72 : if (op_lefttype == op_righttype)
3404 : {
3405 : /* easy case */
3406 72 : ltop = get_opfamily_member(opfamily,
3407 : op_lefttype, op_righttype,
3408 : gtstrat);
3409 72 : leop = get_opfamily_member(opfamily,
3410 : op_lefttype, op_righttype,
3411 : gestrat);
3412 72 : lsortop = ltop;
3413 72 : rsortop = ltop;
3414 72 : lstatop = get_opfamily_member(opfamily,
3415 : op_lefttype, op_lefttype,
3416 : ltstrat);
3417 72 : rstatop = lstatop;
3418 72 : revltop = ltop;
3419 72 : revleop = leop;
3420 : }
3421 : else
3422 : {
3423 0 : ltop = get_opfamily_member(opfamily,
3424 : op_lefttype, op_righttype,
3425 : gtstrat);
3426 0 : leop = get_opfamily_member(opfamily,
3427 : op_lefttype, op_righttype,
3428 : gestrat);
3429 0 : lsortop = get_opfamily_member(opfamily,
3430 : op_lefttype, op_lefttype,
3431 : gtstrat);
3432 0 : rsortop = get_opfamily_member(opfamily,
3433 : op_righttype, op_righttype,
3434 : gtstrat);
3435 0 : lstatop = get_opfamily_member(opfamily,
3436 : op_lefttype, op_lefttype,
3437 : ltstrat);
3438 0 : rstatop = get_opfamily_member(opfamily,
3439 : op_righttype, op_righttype,
3440 : ltstrat);
3441 0 : revltop = get_opfamily_member(opfamily,
3442 : op_righttype, op_lefttype,
3443 : gtstrat);
3444 0 : revleop = get_opfamily_member(opfamily,
3445 : op_righttype, op_lefttype,
3446 : gestrat);
3447 : }
3448 72 : break;
3449 0 : default:
3450 0 : goto fail; /* shouldn't get here */
3451 : }
3452 :
3453 141368 : if (!OidIsValid(lsortop) ||
3454 141368 : !OidIsValid(rsortop) ||
3455 141368 : !OidIsValid(lstatop) ||
3456 141368 : !OidIsValid(rstatop) ||
3457 141356 : !OidIsValid(ltop) ||
3458 141356 : !OidIsValid(leop) ||
3459 141356 : !OidIsValid(revltop) ||
3460 : !OidIsValid(revleop))
3461 12 : goto fail; /* insufficient info in catalogs */
3462 :
3463 : /* Try to get ranges of both inputs */
3464 141356 : if (!isgt)
3465 : {
3466 141284 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3467 : &leftmin, &leftmax))
3468 34412 : goto fail; /* no range available from stats */
3469 106872 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3470 : &rightmin, &rightmax))
3471 24420 : goto fail; /* no range available from stats */
3472 : }
3473 : else
3474 : {
3475 : /* need to swap the max and min */
3476 72 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3477 : &leftmax, &leftmin))
3478 30 : goto fail; /* no range available from stats */
3479 42 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3480 : &rightmax, &rightmin))
3481 0 : goto fail; /* no range available from stats */
3482 : }
3483 :
3484 : /*
3485 : * Now, the fraction of the left variable that will be scanned is the
3486 : * fraction that's <= the right-side maximum value. But only believe
3487 : * non-default estimates, else stick with our 1.0.
3488 : */
3489 82494 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3490 : rightmax, op_righttype);
3491 82494 : if (selec != DEFAULT_INEQ_SEL)
3492 82488 : *leftend = selec;
3493 :
3494 : /* And similarly for the right variable. */
3495 82494 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3496 : leftmax, op_lefttype);
3497 82494 : if (selec != DEFAULT_INEQ_SEL)
3498 82494 : *rightend = selec;
3499 :
3500 : /*
3501 : * Only one of the two "end" fractions can really be less than 1.0;
3502 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3503 : * we get exactly equal estimates (as can easily happen with self-joins),
3504 : * believe neither.
3505 : */
3506 82494 : if (*leftend > *rightend)
3507 25030 : *leftend = 1.0;
3508 57464 : else if (*leftend < *rightend)
3509 33702 : *rightend = 1.0;
3510 : else
3511 23762 : *leftend = *rightend = 1.0;
3512 :
3513 : /*
3514 : * Also, the fraction of the left variable that will be scanned before the
3515 : * first join pair is found is the fraction that's < the right-side
3516 : * minimum value. But only believe non-default estimates, else stick with
3517 : * our own default.
3518 : */
3519 82494 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3520 : rightmin, op_righttype);
3521 82494 : if (selec != DEFAULT_INEQ_SEL)
3522 82494 : *leftstart = selec;
3523 :
3524 : /* And similarly for the right variable. */
3525 82494 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3526 : leftmin, op_lefttype);
3527 82494 : if (selec != DEFAULT_INEQ_SEL)
3528 82494 : *rightstart = selec;
3529 :
3530 : /*
3531 : * Only one of the two "start" fractions can really be more than zero;
3532 : * believe the larger estimate and reset the other one to exactly 0.0. If
3533 : * we get exactly equal estimates (as can easily happen with self-joins),
3534 : * believe neither.
3535 : */
3536 82494 : if (*leftstart < *rightstart)
3537 17272 : *leftstart = 0.0;
3538 65222 : else if (*leftstart > *rightstart)
3539 23964 : *rightstart = 0.0;
3540 : else
3541 41258 : *leftstart = *rightstart = 0.0;
3542 :
3543 : /*
3544 : * If the sort order is nulls-first, we're going to have to skip over any
3545 : * nulls too. These would not have been counted by scalarineqsel, and we
3546 : * can safely add in this fraction regardless of whether we believe
3547 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3548 : */
3549 82494 : if (nulls_first)
3550 : {
3551 : Form_pg_statistic stats;
3552 :
3553 42 : if (HeapTupleIsValid(leftvar.statsTuple))
3554 : {
3555 42 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3556 42 : *leftstart += stats->stanullfrac;
3557 42 : CLAMP_PROBABILITY(*leftstart);
3558 42 : *leftend += stats->stanullfrac;
3559 42 : CLAMP_PROBABILITY(*leftend);
3560 : }
3561 42 : if (HeapTupleIsValid(rightvar.statsTuple))
3562 : {
3563 42 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3564 42 : *rightstart += stats->stanullfrac;
3565 42 : CLAMP_PROBABILITY(*rightstart);
3566 42 : *rightend += stats->stanullfrac;
3567 42 : CLAMP_PROBABILITY(*rightend);
3568 : }
3569 : }
3570 :
3571 : /* Disbelieve start >= end, just in case that can happen */
3572 82494 : if (*leftstart >= *leftend)
3573 : {
3574 212 : *leftstart = 0.0;
3575 212 : *leftend = 1.0;
3576 : }
3577 82494 : if (*rightstart >= *rightend)
3578 : {
3579 1138 : *rightstart = 0.0;
3580 1138 : *rightend = 1.0;
3581 : }
3582 :
3583 81356 : fail:
3584 141368 : ReleaseVariableStats(leftvar);
3585 141368 : ReleaseVariableStats(rightvar);
3586 : }
3587 :
3588 :
3589 : /*
3590 : * matchingsel -- generic matching-operator selectivity support
3591 : *
3592 : * Use these for any operators that (a) are on data types for which we collect
3593 : * standard statistics, and (b) have behavior for which the default estimate
3594 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3595 : * operators.
3596 : */
3597 :
3598 : Datum
3599 1130 : matchingsel(PG_FUNCTION_ARGS)
3600 : {
3601 1130 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3602 1130 : Oid operator = PG_GETARG_OID(1);
3603 1130 : List *args = (List *) PG_GETARG_POINTER(2);
3604 1130 : int varRelid = PG_GETARG_INT32(3);
3605 1130 : Oid collation = PG_GET_COLLATION();
3606 : double selec;
3607 :
3608 : /* Use generic restriction selectivity logic. */
3609 1130 : selec = generic_restriction_selectivity(root, operator, collation,
3610 : args, varRelid,
3611 : DEFAULT_MATCHING_SEL);
3612 :
3613 1130 : PG_RETURN_FLOAT8((float8) selec);
3614 : }
3615 :
3616 : Datum
3617 6 : matchingjoinsel(PG_FUNCTION_ARGS)
3618 : {
3619 : /* Just punt, for the moment. */
3620 6 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3621 : }
3622 :
3623 :
3624 : /*
3625 : * Helper routine for estimate_num_groups: add an item to a list of
3626 : * GroupVarInfos, but only if it's not known equal to any of the existing
3627 : * entries.
3628 : */
3629 : typedef struct
3630 : {
3631 : Node *var; /* might be an expression, not just a Var */
3632 : RelOptInfo *rel; /* relation it belongs to */
3633 : double ndistinct; /* # distinct values */
3634 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3635 : } GroupVarInfo;
3636 :
3637 : static List *
3638 393444 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3639 : Node *var, VariableStatData *vardata)
3640 : {
3641 : GroupVarInfo *varinfo;
3642 : double ndistinct;
3643 : bool isdefault;
3644 : ListCell *lc;
3645 :
3646 393444 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3647 :
3648 : /*
3649 : * The nullingrels bits within the var could cause the same var to be
3650 : * counted multiple times if it's marked with different nullingrels. They
3651 : * could also prevent us from matching the var to the expressions in
3652 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3653 : * them out first.
3654 : */
3655 393444 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3656 :
3657 476078 : foreach(lc, varinfos)
3658 : {
3659 83750 : varinfo = (GroupVarInfo *) lfirst(lc);
3660 :
3661 : /* Drop exact duplicates */
3662 83750 : if (equal(var, varinfo->var))
3663 1116 : return varinfos;
3664 :
3665 : /*
3666 : * Drop known-equal vars, but only if they belong to different
3667 : * relations (see comments for estimate_num_groups). We aren't too
3668 : * fussy about the semantics of "equal" here.
3669 : */
3670 89746 : if (vardata->rel != varinfo->rel &&
3671 6836 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3672 : {
3673 300 : if (varinfo->ndistinct <= ndistinct)
3674 : {
3675 : /* Keep older item, forget new one */
3676 276 : return varinfos;
3677 : }
3678 : else
3679 : {
3680 : /* Delete the older item */
3681 24 : varinfos = foreach_delete_current(varinfos, lc);
3682 : }
3683 : }
3684 : }
3685 :
3686 392328 : varinfo = (GroupVarInfo *) palloc(sizeof(GroupVarInfo));
3687 :
3688 392328 : varinfo->var = var;
3689 392328 : varinfo->rel = vardata->rel;
3690 392328 : varinfo->ndistinct = ndistinct;
3691 392328 : varinfo->isdefault = isdefault;
3692 392328 : varinfos = lappend(varinfos, varinfo);
3693 392328 : return varinfos;
3694 : }
3695 :
3696 : /*
3697 : * estimate_num_groups - Estimate number of groups in a grouped query
3698 : *
3699 : * Given a query having a GROUP BY clause, estimate how many groups there
3700 : * will be --- ie, the number of distinct combinations of the GROUP BY
3701 : * expressions.
3702 : *
3703 : * This routine is also used to estimate the number of rows emitted by
3704 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3705 : * actually, we only use it for DISTINCT when there's no grouping or
3706 : * aggregation ahead of the DISTINCT.)
3707 : *
3708 : * Inputs:
3709 : * root - the query
3710 : * groupExprs - list of expressions being grouped by
3711 : * input_rows - number of rows estimated to arrive at the group/unique
3712 : * filter step
3713 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3714 : * groupExprs against
3715 : *
3716 : * Outputs:
3717 : * estinfo - When passed as non-NULL, the function will set bits in the
3718 : * "flags" field in order to provide callers with additional information
3719 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3720 : * bit if we used any default values in the estimation.
3721 : *
3722 : * Given the lack of any cross-correlation statistics in the system, it's
3723 : * impossible to do anything really trustworthy with GROUP BY conditions
3724 : * involving multiple Vars. We should however avoid assuming the worst
3725 : * case (all possible cross-product terms actually appear as groups) since
3726 : * very often the grouped-by Vars are highly correlated. Our current approach
3727 : * is as follows:
3728 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3729 : * independently of their content, and are ignored in the subsequent
3730 : * steps. This is mainly because tests like "col IS NULL" break the
3731 : * heuristic used in step 2 especially badly.
3732 : * 2. Reduce the given expressions to a list of unique Vars used. For
3733 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3734 : * It is clearly correct not to count the same Var more than once.
3735 : * It is also reasonable to treat f(x) the same as x: f() cannot
3736 : * increase the number of distinct values (unless it is volatile,
3737 : * which we consider unlikely for grouping), but it probably won't
3738 : * reduce the number of distinct values much either.
3739 : * As a special case, if a GROUP BY expression can be matched to an
3740 : * expressional index for which we have statistics, then we treat the
3741 : * whole expression as though it were just a Var.
3742 : * 3. If the list contains Vars of different relations that are known equal
3743 : * due to equivalence classes, then drop all but one of the Vars from each
3744 : * known-equal set, keeping the one with smallest estimated # of values
3745 : * (since the extra values of the others can't appear in joined rows).
3746 : * Note the reason we only consider Vars of different relations is that
3747 : * if we considered ones of the same rel, we'd be double-counting the
3748 : * restriction selectivity of the equality in the next step.
3749 : * 4. For Vars within a single source rel, we multiply together the numbers
3750 : * of values, clamp to the number of rows in the rel (divided by 10 if
3751 : * more than one Var), and then multiply by a factor based on the
3752 : * selectivity of the restriction clauses for that rel. When there's
3753 : * more than one Var, the initial product is probably too high (it's the
3754 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3755 : * helpful heuristic for not letting the estimate get out of hand. (The
3756 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3757 : * we multiply by to adjust for the restriction selectivity assumes that
3758 : * the restriction clauses are independent of the grouping, which may not
3759 : * be a valid assumption, but it's hard to do better.
3760 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3761 : * rel, and multiply the results together.
3762 : * Note that rels not containing grouped Vars are ignored completely, as are
3763 : * join clauses. Such rels cannot increase the number of groups, and we
3764 : * assume such clauses do not reduce the number either (somewhat bogus,
3765 : * but we don't have the info to do better).
3766 : */
3767 : double
3768 342026 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3769 : List **pgset, EstimationInfo *estinfo)
3770 : {
3771 342026 : List *varinfos = NIL;
3772 342026 : double srf_multiplier = 1.0;
3773 : double numdistinct;
3774 : ListCell *l;
3775 : int i;
3776 :
3777 : /* Zero the estinfo output parameter, if non-NULL */
3778 342026 : if (estinfo != NULL)
3779 293248 : memset(estinfo, 0, sizeof(EstimationInfo));
3780 :
3781 : /*
3782 : * We don't ever want to return an estimate of zero groups, as that tends
3783 : * to lead to division-by-zero and other unpleasantness. The input_rows
3784 : * estimate is usually already at least 1, but clamp it just in case it
3785 : * isn't.
3786 : */
3787 342026 : input_rows = clamp_row_est(input_rows);
3788 :
3789 : /*
3790 : * If no grouping columns, there's exactly one group. (This can't happen
3791 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3792 : * corner cases with set operations.)
3793 : */
3794 342026 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3795 1104 : return 1.0;
3796 :
3797 : /*
3798 : * Count groups derived from boolean grouping expressions. For other
3799 : * expressions, find the unique Vars used, treating an expression as a Var
3800 : * if we can find stats for it. For each one, record the statistical
3801 : * estimate of number of distinct values (total in its table, without
3802 : * regard for filtering).
3803 : */
3804 340922 : numdistinct = 1.0;
3805 :
3806 340922 : i = 0;
3807 732606 : foreach(l, groupExprs)
3808 : {
3809 391732 : Node *groupexpr = (Node *) lfirst(l);
3810 : double this_srf_multiplier;
3811 : VariableStatData vardata;
3812 : List *varshere;
3813 : ListCell *l2;
3814 :
3815 : /* is expression in this grouping set? */
3816 391732 : if (pgset && !list_member_int(*pgset, i++))
3817 322708 : continue;
3818 :
3819 : /*
3820 : * Set-returning functions in grouping columns are a bit problematic.
3821 : * The code below will effectively ignore their SRF nature and come up
3822 : * with a numdistinct estimate as though they were scalar functions.
3823 : * We compensate by scaling up the end result by the largest SRF
3824 : * rowcount estimate. (This will be an overestimate if the SRF
3825 : * produces multiple copies of any output value, but it seems best to
3826 : * assume the SRF's outputs are distinct. In any case, it's probably
3827 : * pointless to worry too much about this without much better
3828 : * estimates for SRF output rowcounts than we have today.)
3829 : */
3830 390932 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3831 390932 : if (srf_multiplier < this_srf_multiplier)
3832 132 : srf_multiplier = this_srf_multiplier;
3833 :
3834 : /* Short-circuit for expressions returning boolean */
3835 390932 : if (exprType(groupexpr) == BOOLOID)
3836 : {
3837 204 : numdistinct *= 2.0;
3838 204 : continue;
3839 : }
3840 :
3841 : /*
3842 : * If examine_variable is able to deduce anything about the GROUP BY
3843 : * expression, treat it as a single variable even if it's really more
3844 : * complicated.
3845 : *
3846 : * XXX This has the consequence that if there's a statistics object on
3847 : * the expression, we don't split it into individual Vars. This
3848 : * affects our selection of statistics in
3849 : * estimate_multivariate_ndistinct, because it's probably better to
3850 : * use more accurate estimate for each expression and treat them as
3851 : * independent, than to combine estimates for the extracted variables
3852 : * when we don't know how that relates to the expressions.
3853 : */
3854 390728 : examine_variable(root, groupexpr, 0, &vardata);
3855 390728 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3856 : {
3857 321026 : varinfos = add_unique_group_var(root, varinfos,
3858 : groupexpr, &vardata);
3859 321026 : ReleaseVariableStats(vardata);
3860 321026 : continue;
3861 : }
3862 69702 : ReleaseVariableStats(vardata);
3863 :
3864 : /*
3865 : * Else pull out the component Vars. Handle PlaceHolderVars by
3866 : * recursing into their arguments (effectively assuming that the
3867 : * PlaceHolderVar doesn't change the number of groups, which boils
3868 : * down to ignoring the possible addition of nulls to the result set).
3869 : */
3870 69702 : varshere = pull_var_clause(groupexpr,
3871 : PVC_RECURSE_AGGREGATES |
3872 : PVC_RECURSE_WINDOWFUNCS |
3873 : PVC_RECURSE_PLACEHOLDERS);
3874 :
3875 : /*
3876 : * If we find any variable-free GROUP BY item, then either it is a
3877 : * constant (and we can ignore it) or it contains a volatile function;
3878 : * in the latter case we punt and assume that each input row will
3879 : * yield a distinct group.
3880 : */
3881 69702 : if (varshere == NIL)
3882 : {
3883 726 : if (contain_volatile_functions(groupexpr))
3884 48 : return input_rows;
3885 678 : continue;
3886 : }
3887 :
3888 : /*
3889 : * Else add variables to varinfos list
3890 : */
3891 141394 : foreach(l2, varshere)
3892 : {
3893 72418 : Node *var = (Node *) lfirst(l2);
3894 :
3895 72418 : examine_variable(root, var, 0, &vardata);
3896 72418 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3897 72418 : ReleaseVariableStats(vardata);
3898 : }
3899 : }
3900 :
3901 : /*
3902 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3903 : * list.
3904 : */
3905 340874 : if (varinfos == NIL)
3906 : {
3907 : /* Apply SRF multiplier as we would do in the long path */
3908 394 : numdistinct *= srf_multiplier;
3909 : /* Round off */
3910 394 : numdistinct = ceil(numdistinct);
3911 : /* Guard against out-of-range answers */
3912 394 : if (numdistinct > input_rows)
3913 44 : numdistinct = input_rows;
3914 394 : if (numdistinct < 1.0)
3915 0 : numdistinct = 1.0;
3916 394 : return numdistinct;
3917 : }
3918 :
3919 : /*
3920 : * Group Vars by relation and estimate total numdistinct.
3921 : *
3922 : * For each iteration of the outer loop, we process the frontmost Var in
3923 : * varinfos, plus all other Vars in the same relation. We remove these
3924 : * Vars from the newvarinfos list for the next iteration. This is the
3925 : * easiest way to group Vars of same rel together.
3926 : */
3927 : do
3928 : {
3929 343386 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3930 343386 : RelOptInfo *rel = varinfo1->rel;
3931 343386 : double reldistinct = 1;
3932 343386 : double relmaxndistinct = reldistinct;
3933 343386 : int relvarcount = 0;
3934 343386 : List *newvarinfos = NIL;
3935 343386 : List *relvarinfos = NIL;
3936 :
3937 : /*
3938 : * Split the list of varinfos in two - one for the current rel, one
3939 : * for remaining Vars on other rels.
3940 : */
3941 343386 : relvarinfos = lappend(relvarinfos, varinfo1);
3942 397838 : for_each_from(l, varinfos, 1)
3943 : {
3944 54452 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3945 :
3946 54452 : if (varinfo2->rel == varinfo1->rel)
3947 : {
3948 : /* varinfos on current rel */
3949 48918 : relvarinfos = lappend(relvarinfos, varinfo2);
3950 : }
3951 : else
3952 : {
3953 : /* not time to process varinfo2 yet */
3954 5534 : newvarinfos = lappend(newvarinfos, varinfo2);
3955 : }
3956 : }
3957 :
3958 : /*
3959 : * Get the numdistinct estimate for the Vars of this rel. We
3960 : * iteratively search for multivariate n-distinct with maximum number
3961 : * of vars; assuming that each var group is independent of the others,
3962 : * we multiply them together. Any remaining relvarinfos after no more
3963 : * multivariate matches are found are assumed independent too, so
3964 : * their individual ndistinct estimates are multiplied also.
3965 : *
3966 : * While iterating, count how many separate numdistinct values we
3967 : * apply. We apply a fudge factor below, but only if we multiplied
3968 : * more than one such values.
3969 : */
3970 686898 : while (relvarinfos)
3971 : {
3972 : double mvndistinct;
3973 :
3974 343512 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3975 : &mvndistinct))
3976 : {
3977 414 : reldistinct *= mvndistinct;
3978 414 : if (relmaxndistinct < mvndistinct)
3979 402 : relmaxndistinct = mvndistinct;
3980 414 : relvarcount++;
3981 : }
3982 : else
3983 : {
3984 734526 : foreach(l, relvarinfos)
3985 : {
3986 391428 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3987 :
3988 391428 : reldistinct *= varinfo2->ndistinct;
3989 391428 : if (relmaxndistinct < varinfo2->ndistinct)
3990 345564 : relmaxndistinct = varinfo2->ndistinct;
3991 391428 : relvarcount++;
3992 :
3993 : /*
3994 : * When varinfo2's isdefault is set then we'd better set
3995 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3996 : */
3997 391428 : if (estinfo != NULL && varinfo2->isdefault)
3998 20200 : estinfo->flags |= SELFLAG_USED_DEFAULT;
3999 : }
4000 :
4001 : /* we're done with this relation */
4002 343098 : relvarinfos = NIL;
4003 : }
4004 : }
4005 :
4006 : /*
4007 : * Sanity check --- don't divide by zero if empty relation.
4008 : */
4009 : Assert(IS_SIMPLE_REL(rel));
4010 343386 : if (rel->tuples > 0)
4011 : {
4012 : /*
4013 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4014 : * fudge factor is because the Vars are probably correlated but we
4015 : * don't know by how much. We should never clamp to less than the
4016 : * largest ndistinct value for any of the Vars, though, since
4017 : * there will surely be at least that many groups.
4018 : */
4019 342342 : double clamp = rel->tuples;
4020 :
4021 342342 : if (relvarcount > 1)
4022 : {
4023 44146 : clamp *= 0.1;
4024 44146 : if (clamp < relmaxndistinct)
4025 : {
4026 41486 : clamp = relmaxndistinct;
4027 : /* for sanity in case some ndistinct is too large: */
4028 41486 : if (clamp > rel->tuples)
4029 78 : clamp = rel->tuples;
4030 : }
4031 : }
4032 342342 : if (reldistinct > clamp)
4033 36134 : reldistinct = clamp;
4034 :
4035 : /*
4036 : * Update the estimate based on the restriction selectivity,
4037 : * guarding against division by zero when reldistinct is zero.
4038 : * Also skip this if we know that we are returning all rows.
4039 : */
4040 342342 : if (reldistinct > 0 && rel->rows < rel->tuples)
4041 : {
4042 : /*
4043 : * Given a table containing N rows with n distinct values in a
4044 : * uniform distribution, if we select p rows at random then
4045 : * the expected number of distinct values selected is
4046 : *
4047 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4048 : *
4049 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4050 : *
4051 : * See "Approximating block accesses in database
4052 : * organizations", S. B. Yao, Communications of the ACM,
4053 : * Volume 20 Issue 4, April 1977 Pages 260-261.
4054 : *
4055 : * Alternatively, re-arranging the terms from the factorials,
4056 : * this may be written as
4057 : *
4058 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4059 : *
4060 : * This form of the formula is more efficient to compute in
4061 : * the common case where p is larger than N/n. Additionally,
4062 : * as pointed out by Dell'Era, if i << N for all terms in the
4063 : * product, it can be approximated by
4064 : *
4065 : * n * (1 - ((N-p)/N)^(N/n))
4066 : *
4067 : * See "Expected distinct values when selecting from a bag
4068 : * without replacement", Alberto Dell'Era,
4069 : * http://www.adellera.it/investigations/distinct_balls/.
4070 : *
4071 : * The condition i << N is equivalent to n >> 1, so this is a
4072 : * good approximation when the number of distinct values in
4073 : * the table is large. It turns out that this formula also
4074 : * works well even when n is small.
4075 : */
4076 108394 : reldistinct *=
4077 108394 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4078 108394 : rel->tuples / reldistinct));
4079 : }
4080 342342 : reldistinct = clamp_row_est(reldistinct);
4081 :
4082 : /*
4083 : * Update estimate of total distinct groups.
4084 : */
4085 342342 : numdistinct *= reldistinct;
4086 : }
4087 :
4088 343386 : varinfos = newvarinfos;
4089 343386 : } while (varinfos != NIL);
4090 :
4091 : /* Now we can account for the effects of any SRFs */
4092 340480 : numdistinct *= srf_multiplier;
4093 :
4094 : /* Round off */
4095 340480 : numdistinct = ceil(numdistinct);
4096 :
4097 : /* Guard against out-of-range answers */
4098 340480 : if (numdistinct > input_rows)
4099 70006 : numdistinct = input_rows;
4100 340480 : if (numdistinct < 1.0)
4101 0 : numdistinct = 1.0;
4102 :
4103 340480 : return numdistinct;
4104 : }
4105 :
4106 : /*
4107 : * Try to estimate the bucket size of the hash join inner side when the join
4108 : * condition contains two or more clauses by employing extended statistics.
4109 : *
4110 : * The main idea of this approach is that the distinct value generated by
4111 : * multivariate estimation on two or more columns would provide less bucket size
4112 : * than estimation on one separate column.
4113 : *
4114 : * IMPORTANT: It is crucial to synchronize the approach of combining different
4115 : * estimations with the caller's method.
4116 : *
4117 : * Return a list of clauses that didn't fetch any extended statistics.
4118 : */
4119 : List *
4120 451216 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4121 : List *hashclauses,
4122 : Selectivity *innerbucketsize)
4123 : {
4124 : List *clauses;
4125 : List *otherclauses;
4126 : double ndistinct;
4127 :
4128 451216 : if (list_length(hashclauses) <= 1)
4129 : {
4130 : /*
4131 : * Nothing to do for a single clause. Could we employ univariate
4132 : * extended stat here?
4133 : */
4134 415154 : return hashclauses;
4135 : }
4136 :
4137 : /* "clauses" is the list of hashclauses we've not dealt with yet */
4138 36062 : clauses = list_copy(hashclauses);
4139 : /* "otherclauses" holds clauses we are going to return to caller */
4140 36062 : otherclauses = NIL;
4141 : /* current estimate of ndistinct */
4142 36062 : ndistinct = 1.0;
4143 72136 : while (clauses != NIL)
4144 : {
4145 : ListCell *lc;
4146 36074 : int relid = -1;
4147 36074 : List *varinfos = NIL;
4148 36074 : List *origin_rinfos = NIL;
4149 : double mvndistinct;
4150 : List *origin_varinfos;
4151 36074 : int group_relid = -1;
4152 36074 : RelOptInfo *group_rel = NULL;
4153 : ListCell *lc1,
4154 : *lc2;
4155 :
4156 : /*
4157 : * Find clauses, referencing the same single base relation and try to
4158 : * estimate such a group with extended statistics. Create varinfo for
4159 : * an approved clause, push it to otherclauses, if it can't be
4160 : * estimated here or ignore to process at the next iteration.
4161 : */
4162 108834 : foreach(lc, clauses)
4163 : {
4164 72760 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4165 : Node *expr;
4166 : Relids relids;
4167 : GroupVarInfo *varinfo;
4168 :
4169 : /*
4170 : * Find the inner side of the join, which we need to estimate the
4171 : * number of buckets. Use outer_is_left because the
4172 : * clause_sides_match_join routine has called on hash clauses.
4173 : */
4174 145520 : relids = rinfo->outer_is_left ?
4175 72760 : rinfo->right_relids : rinfo->left_relids;
4176 145520 : expr = rinfo->outer_is_left ?
4177 72760 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4178 :
4179 72760 : if (bms_get_singleton_member(relids, &relid) &&
4180 70178 : root->simple_rel_array[relid]->statlist != NIL)
4181 48 : {
4182 60 : bool is_duplicate = false;
4183 :
4184 : /*
4185 : * This inner-side expression references only one relation.
4186 : * Extended statistics on this clause can exist.
4187 : */
4188 60 : if (group_relid < 0)
4189 : {
4190 30 : RangeTblEntry *rte = root->simple_rte_array[relid];
4191 :
4192 30 : if (!rte || (rte->relkind != RELKIND_RELATION &&
4193 0 : rte->relkind != RELKIND_MATVIEW &&
4194 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4195 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4196 : {
4197 : /* Extended statistics can't exist in principle */
4198 0 : otherclauses = lappend(otherclauses, rinfo);
4199 0 : clauses = foreach_delete_current(clauses, lc);
4200 0 : continue;
4201 : }
4202 :
4203 30 : group_relid = relid;
4204 30 : group_rel = root->simple_rel_array[relid];
4205 : }
4206 30 : else if (group_relid != relid)
4207 : {
4208 : /*
4209 : * Being in the group forming state we don't need other
4210 : * clauses.
4211 : */
4212 0 : continue;
4213 : }
4214 :
4215 : /*
4216 : * We're going to add the new clause to the varinfos list. We
4217 : * might re-use add_unique_group_var(), but we don't do so for
4218 : * two reasons.
4219 : *
4220 : * 1) We must keep the origin_rinfos list ordered exactly the
4221 : * same way as varinfos.
4222 : *
4223 : * 2) add_unique_group_var() is designed for
4224 : * estimate_num_groups(), where a larger number of groups is
4225 : * worse. While estimating the number of hash buckets, we
4226 : * have the opposite: a lesser number of groups is worse.
4227 : * Therefore, we don't have to remove "known equal" vars: the
4228 : * removed var may valuably contribute to the multivariate
4229 : * statistics to grow the number of groups.
4230 : */
4231 :
4232 : /*
4233 : * Clear nullingrels to correctly match hash keys. See
4234 : * add_unique_group_var()'s comment for details.
4235 : */
4236 60 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4237 :
4238 : /*
4239 : * Detect and exclude exact duplicates from the list of hash
4240 : * keys (like add_unique_group_var does).
4241 : */
4242 84 : foreach(lc1, varinfos)
4243 : {
4244 36 : varinfo = (GroupVarInfo *) lfirst(lc1);
4245 :
4246 36 : if (!equal(expr, varinfo->var))
4247 24 : continue;
4248 :
4249 12 : is_duplicate = true;
4250 12 : break;
4251 : }
4252 :
4253 60 : if (is_duplicate)
4254 : {
4255 : /*
4256 : * Skip exact duplicates. Adding them to the otherclauses
4257 : * list also doesn't make sense.
4258 : */
4259 12 : continue;
4260 : }
4261 :
4262 : /*
4263 : * Initialize GroupVarInfo. We only use it to call
4264 : * estimate_multivariate_ndistinct(), which doesn't care about
4265 : * ndistinct and isdefault fields. Thus, skip these fields.
4266 : */
4267 48 : varinfo = (GroupVarInfo *) palloc0(sizeof(GroupVarInfo));
4268 48 : varinfo->var = expr;
4269 48 : varinfo->rel = root->simple_rel_array[relid];
4270 48 : varinfos = lappend(varinfos, varinfo);
4271 :
4272 : /*
4273 : * Remember the link to RestrictInfo for the case the clause
4274 : * is failed to be estimated.
4275 : */
4276 48 : origin_rinfos = lappend(origin_rinfos, rinfo);
4277 : }
4278 : else
4279 : {
4280 : /* This clause can't be estimated with extended statistics */
4281 72700 : otherclauses = lappend(otherclauses, rinfo);
4282 : }
4283 :
4284 72748 : clauses = foreach_delete_current(clauses, lc);
4285 : }
4286 :
4287 36074 : if (list_length(varinfos) < 2)
4288 : {
4289 : /*
4290 : * Multivariate statistics doesn't apply to single columns except
4291 : * for expressions, but it has not been implemented yet.
4292 : */
4293 36062 : otherclauses = list_concat(otherclauses, origin_rinfos);
4294 36062 : list_free_deep(varinfos);
4295 36062 : list_free(origin_rinfos);
4296 36062 : continue;
4297 : }
4298 :
4299 : Assert(group_rel != NULL);
4300 :
4301 : /* Employ the extended statistics. */
4302 12 : origin_varinfos = varinfos;
4303 : for (;;)
4304 12 : {
4305 24 : bool estimated = estimate_multivariate_ndistinct(root,
4306 : group_rel,
4307 : &varinfos,
4308 : &mvndistinct);
4309 :
4310 24 : if (!estimated)
4311 12 : break;
4312 :
4313 : /*
4314 : * We've got an estimation. Use ndistinct value in a consistent
4315 : * way - according to the caller's logic (see
4316 : * final_cost_hashjoin).
4317 : */
4318 12 : if (ndistinct < mvndistinct)
4319 12 : ndistinct = mvndistinct;
4320 : Assert(ndistinct >= 1.0);
4321 : }
4322 :
4323 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4324 :
4325 : /* Collect unmatched clauses as otherclauses. */
4326 42 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4327 : {
4328 30 : GroupVarInfo *vinfo = lfirst(lc1);
4329 :
4330 30 : if (!list_member_ptr(varinfos, vinfo))
4331 : /* Already estimated */
4332 30 : continue;
4333 :
4334 : /* Can't be estimated here - push to the returning list */
4335 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4336 : }
4337 : }
4338 :
4339 36062 : *innerbucketsize = 1.0 / ndistinct;
4340 36062 : return otherclauses;
4341 : }
4342 :
4343 : /*
4344 : * Estimate hash bucket statistics when the specified expression is used
4345 : * as a hash key for the given number of buckets.
4346 : *
4347 : * This attempts to determine two values:
4348 : *
4349 : * 1. The frequency of the most common value of the expression (returns
4350 : * zero into *mcv_freq if we can't get that).
4351 : *
4352 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4353 : * divided by total tuples in relation.
4354 : *
4355 : * XXX This is really pretty bogus since we're effectively assuming that the
4356 : * distribution of hash keys will be the same after applying restriction
4357 : * clauses as it was in the underlying relation. However, we are not nearly
4358 : * smart enough to figure out how the restrict clauses might change the
4359 : * distribution, so this will have to do for now.
4360 : *
4361 : * We are passed the number of buckets the executor will use for the given
4362 : * input relation. If the data were perfectly distributed, with the same
4363 : * number of tuples going into each available bucket, then the bucketsize
4364 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4365 : * only if (a) there are at least nbuckets distinct data values, and (b)
4366 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4367 : * be nonuniformly occupied. If the other relation in the join has a key
4368 : * distribution similar to this one's, then the most-loaded buckets are
4369 : * exactly those that will be probed most often. Therefore, the "average"
4370 : * bucket size for costing purposes should really be taken as something close
4371 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4372 : * fraction if there are too few distinct data values, and then scaling up
4373 : * by the ratio of the most common value's frequency to the average frequency.
4374 : *
4375 : * If no statistics are available, use a default estimate of 0.1. This will
4376 : * discourage use of a hash rather strongly if the inner relation is large,
4377 : * which is what we want. We do not want to hash unless we know that the
4378 : * inner rel is well-dispersed (or the alternatives seem much worse).
4379 : *
4380 : * The caller should also check that the mcv_freq is not so large that the
4381 : * most common value would by itself require an impractically large bucket.
4382 : * In a hash join, the executor can split buckets if they get too big, but
4383 : * obviously that doesn't help for a bucket that contains many duplicates of
4384 : * the same value.
4385 : */
4386 : void
4387 204624 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4388 : Selectivity *mcv_freq,
4389 : Selectivity *bucketsize_frac)
4390 : {
4391 : VariableStatData vardata;
4392 : double estfract,
4393 : ndistinct,
4394 : stanullfrac,
4395 : avgfreq;
4396 : bool isdefault;
4397 : AttStatsSlot sslot;
4398 :
4399 204624 : examine_variable(root, hashkey, 0, &vardata);
4400 :
4401 : /* Look up the frequency of the most common value, if available */
4402 204624 : *mcv_freq = 0.0;
4403 :
4404 204624 : if (HeapTupleIsValid(vardata.statsTuple))
4405 : {
4406 147800 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4407 : STATISTIC_KIND_MCV, InvalidOid,
4408 : ATTSTATSSLOT_NUMBERS))
4409 : {
4410 : /*
4411 : * The first MCV stat is for the most common value.
4412 : */
4413 85382 : if (sslot.nnumbers > 0)
4414 85382 : *mcv_freq = sslot.numbers[0];
4415 85382 : free_attstatsslot(&sslot);
4416 : }
4417 : }
4418 :
4419 : /* Get number of distinct values */
4420 204624 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4421 :
4422 : /*
4423 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4424 : * mcv_freq is known to be even higher than that, use it instead.
4425 : */
4426 204624 : if (isdefault)
4427 : {
4428 25972 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4429 25972 : ReleaseVariableStats(vardata);
4430 25972 : return;
4431 : }
4432 :
4433 : /* Get fraction that are null */
4434 178652 : if (HeapTupleIsValid(vardata.statsTuple))
4435 : {
4436 : Form_pg_statistic stats;
4437 :
4438 147782 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
4439 147782 : stanullfrac = stats->stanullfrac;
4440 : }
4441 : else
4442 30870 : stanullfrac = 0.0;
4443 :
4444 : /* Compute avg freq of all distinct data values in raw relation */
4445 178652 : avgfreq = (1.0 - stanullfrac) / ndistinct;
4446 :
4447 : /*
4448 : * Adjust ndistinct to account for restriction clauses. Observe we are
4449 : * assuming that the data distribution is affected uniformly by the
4450 : * restriction clauses!
4451 : *
4452 : * XXX Possibly better way, but much more expensive: multiply by
4453 : * selectivity of rel's restriction clauses that mention the target Var.
4454 : */
4455 178652 : if (vardata.rel && vardata.rel->tuples > 0)
4456 : {
4457 178618 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4458 178618 : ndistinct = clamp_row_est(ndistinct);
4459 : }
4460 :
4461 : /*
4462 : * Initial estimate of bucketsize fraction is 1/nbuckets as long as the
4463 : * number of buckets is less than the expected number of distinct values;
4464 : * otherwise it is 1/ndistinct.
4465 : */
4466 178652 : if (ndistinct > nbuckets)
4467 92 : estfract = 1.0 / nbuckets;
4468 : else
4469 178560 : estfract = 1.0 / ndistinct;
4470 :
4471 : /*
4472 : * Adjust estimated bucketsize upward to account for skewed distribution.
4473 : */
4474 178652 : if (avgfreq > 0.0 && *mcv_freq > avgfreq)
4475 78374 : estfract *= *mcv_freq / avgfreq;
4476 :
4477 : /*
4478 : * Clamp bucketsize to sane range (the above adjustment could easily
4479 : * produce an out-of-range result). We set the lower bound a little above
4480 : * zero, since zero isn't a very sane result.
4481 : */
4482 178652 : if (estfract < 1.0e-6)
4483 0 : estfract = 1.0e-6;
4484 178652 : else if (estfract > 1.0)
4485 35166 : estfract = 1.0;
4486 :
4487 178652 : *bucketsize_frac = (Selectivity) estfract;
4488 :
4489 178652 : ReleaseVariableStats(vardata);
4490 : }
4491 :
4492 : /*
4493 : * estimate_hashagg_tablesize
4494 : * estimate the number of bytes that a hash aggregate hashtable will
4495 : * require based on the agg_costs, path width and number of groups.
4496 : *
4497 : * We return the result as "double" to forestall any possible overflow
4498 : * problem in the multiplication by dNumGroups.
4499 : *
4500 : * XXX this may be over-estimating the size now that hashagg knows to omit
4501 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4502 : * grouping columns not in the hashed set are counted here even though hashagg
4503 : * won't store them. Is this a problem?
4504 : */
4505 : double
4506 2434 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4507 : const AggClauseCosts *agg_costs, double dNumGroups)
4508 : {
4509 : Size hashentrysize;
4510 :
4511 2434 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4512 2434 : path->pathtarget->width,
4513 2434 : agg_costs->transitionSpace);
4514 :
4515 : /*
4516 : * Note that this disregards the effect of fill-factor and growth policy
4517 : * of the hash table. That's probably ok, given that the default
4518 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4519 : * "double-in-size" growth policies here.
4520 : */
4521 2434 : return hashentrysize * dNumGroups;
4522 : }
4523 :
4524 :
4525 : /*-------------------------------------------------------------------------
4526 : *
4527 : * Support routines
4528 : *
4529 : *-------------------------------------------------------------------------
4530 : */
4531 :
4532 : /*
4533 : * Find the best matching ndistinct extended statistics for the given list of
4534 : * GroupVarInfos.
4535 : *
4536 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4537 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4538 : *
4539 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4540 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4541 : * list is built with the matching GroupVarInfos removed, which is output via
4542 : * the *varinfos parameter before returning true. When no matching stats are
4543 : * found, false is returned and the *varinfos and *ndistinct parameters are
4544 : * left untouched.
4545 : */
4546 : static bool
4547 343536 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4548 : List **varinfos, double *ndistinct)
4549 : {
4550 : ListCell *lc;
4551 : int nmatches_vars;
4552 : int nmatches_exprs;
4553 343536 : Oid statOid = InvalidOid;
4554 : MVNDistinct *stats;
4555 343536 : StatisticExtInfo *matched_info = NULL;
4556 343536 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4557 :
4558 : /* bail out immediately if the table has no extended statistics */
4559 343536 : if (!rel->statlist)
4560 342972 : return false;
4561 :
4562 : /* look for the ndistinct statistics object matching the most vars */
4563 564 : nmatches_vars = 0; /* we require at least two matches */
4564 564 : nmatches_exprs = 0;
4565 2244 : foreach(lc, rel->statlist)
4566 : {
4567 : ListCell *lc2;
4568 1680 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4569 1680 : int nshared_vars = 0;
4570 1680 : int nshared_exprs = 0;
4571 :
4572 : /* skip statistics of other kinds */
4573 1680 : if (info->kind != STATS_EXT_NDISTINCT)
4574 792 : continue;
4575 :
4576 : /* skip statistics with mismatching stxdinherit value */
4577 888 : if (info->inherit != rte->inh)
4578 30 : continue;
4579 :
4580 : /*
4581 : * Determine how many expressions (and variables in non-matched
4582 : * expressions) match. We'll then use these numbers to pick the
4583 : * statistics object that best matches the clauses.
4584 : */
4585 2718 : foreach(lc2, *varinfos)
4586 : {
4587 : ListCell *lc3;
4588 1860 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4589 : AttrNumber attnum;
4590 :
4591 : Assert(varinfo->rel == rel);
4592 :
4593 : /* simple Var, search in statistics keys directly */
4594 1860 : if (IsA(varinfo->var, Var))
4595 : {
4596 1494 : attnum = ((Var *) varinfo->var)->varattno;
4597 :
4598 : /*
4599 : * Ignore system attributes - we don't support statistics on
4600 : * them, so can't match them (and it'd fail as the values are
4601 : * negative).
4602 : */
4603 1494 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4604 12 : continue;
4605 :
4606 1482 : if (bms_is_member(attnum, info->keys))
4607 876 : nshared_vars++;
4608 :
4609 1482 : continue;
4610 : }
4611 :
4612 : /* expression - see if it's in the statistics object */
4613 660 : foreach(lc3, info->exprs)
4614 : {
4615 528 : Node *expr = (Node *) lfirst(lc3);
4616 :
4617 528 : if (equal(varinfo->var, expr))
4618 : {
4619 234 : nshared_exprs++;
4620 234 : break;
4621 : }
4622 : }
4623 : }
4624 :
4625 : /*
4626 : * The ndistinct extended statistics contain estimates for a minimum
4627 : * of pairs of columns which the statistics are defined on and
4628 : * certainly not single columns. Here we skip unless we managed to
4629 : * match to at least two columns.
4630 : */
4631 858 : if (nshared_vars + nshared_exprs < 2)
4632 396 : continue;
4633 :
4634 : /*
4635 : * Check if these statistics are a better match than the previous best
4636 : * match and if so, take note of the StatisticExtInfo.
4637 : *
4638 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4639 : * select as the best match is deterministic even when multiple sets
4640 : * of statistics match equally as well.
4641 : */
4642 462 : if ((nshared_exprs > nmatches_exprs) ||
4643 354 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4644 : {
4645 438 : statOid = info->statOid;
4646 438 : nmatches_vars = nshared_vars;
4647 438 : nmatches_exprs = nshared_exprs;
4648 438 : matched_info = info;
4649 : }
4650 : }
4651 :
4652 : /* No match? */
4653 564 : if (statOid == InvalidOid)
4654 138 : return false;
4655 :
4656 : Assert(nmatches_vars + nmatches_exprs > 1);
4657 :
4658 426 : stats = statext_ndistinct_load(statOid, rte->inh);
4659 :
4660 : /*
4661 : * If we have a match, search it for the specific item that matches (there
4662 : * must be one), and construct the output values.
4663 : */
4664 426 : if (stats)
4665 : {
4666 : int i;
4667 426 : List *newlist = NIL;
4668 426 : MVNDistinctItem *item = NULL;
4669 : ListCell *lc2;
4670 426 : Bitmapset *matched = NULL;
4671 : AttrNumber attnum_offset;
4672 :
4673 : /*
4674 : * How much we need to offset the attnums? If there are no
4675 : * expressions, no offset is needed. Otherwise offset enough to move
4676 : * the lowest one (which is equal to number of expressions) to 1.
4677 : */
4678 426 : if (matched_info->exprs)
4679 150 : attnum_offset = (list_length(matched_info->exprs) + 1);
4680 : else
4681 276 : attnum_offset = 0;
4682 :
4683 : /* see what actually matched */
4684 1488 : foreach(lc2, *varinfos)
4685 : {
4686 : ListCell *lc3;
4687 : int idx;
4688 1062 : bool found = false;
4689 :
4690 1062 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4691 :
4692 : /*
4693 : * Process a simple Var expression, by matching it to keys
4694 : * directly. If there's a matching expression, we'll try matching
4695 : * it later.
4696 : */
4697 1062 : if (IsA(varinfo->var, Var))
4698 : {
4699 876 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4700 :
4701 : /*
4702 : * Ignore expressions on system attributes. Can't rely on the
4703 : * bms check for negative values.
4704 : */
4705 876 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4706 6 : continue;
4707 :
4708 : /* Is the variable covered by the statistics object? */
4709 870 : if (!bms_is_member(attnum, matched_info->keys))
4710 120 : continue;
4711 :
4712 750 : attnum = attnum + attnum_offset;
4713 :
4714 : /* ensure sufficient offset */
4715 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4716 :
4717 750 : matched = bms_add_member(matched, attnum);
4718 :
4719 750 : found = true;
4720 : }
4721 :
4722 : /*
4723 : * XXX Maybe we should allow searching the expressions even if we
4724 : * found an attribute matching the expression? That would handle
4725 : * trivial expressions like "(a)" but it seems fairly useless.
4726 : */
4727 936 : if (found)
4728 750 : continue;
4729 :
4730 : /* expression - see if it's in the statistics object */
4731 186 : idx = 0;
4732 306 : foreach(lc3, matched_info->exprs)
4733 : {
4734 276 : Node *expr = (Node *) lfirst(lc3);
4735 :
4736 276 : if (equal(varinfo->var, expr))
4737 : {
4738 156 : AttrNumber attnum = -(idx + 1);
4739 :
4740 156 : attnum = attnum + attnum_offset;
4741 :
4742 : /* ensure sufficient offset */
4743 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4744 :
4745 156 : matched = bms_add_member(matched, attnum);
4746 :
4747 : /* there should be just one matching expression */
4748 156 : break;
4749 : }
4750 :
4751 120 : idx++;
4752 : }
4753 : }
4754 :
4755 : /* Find the specific item that exactly matches the combination */
4756 864 : for (i = 0; i < stats->nitems; i++)
4757 : {
4758 : int j;
4759 864 : MVNDistinctItem *tmpitem = &stats->items[i];
4760 :
4761 864 : if (tmpitem->nattributes != bms_num_members(matched))
4762 162 : continue;
4763 :
4764 : /* assume it's the right item */
4765 702 : item = tmpitem;
4766 :
4767 : /* check that all item attributes/expressions fit the match */
4768 1692 : for (j = 0; j < tmpitem->nattributes; j++)
4769 : {
4770 1266 : AttrNumber attnum = tmpitem->attributes[j];
4771 :
4772 : /*
4773 : * Thanks to how we constructed the matched bitmap above, we
4774 : * can just offset all attnums the same way.
4775 : */
4776 1266 : attnum = attnum + attnum_offset;
4777 :
4778 1266 : if (!bms_is_member(attnum, matched))
4779 : {
4780 : /* nah, it's not this item */
4781 276 : item = NULL;
4782 276 : break;
4783 : }
4784 : }
4785 :
4786 : /*
4787 : * If the item has all the matched attributes, we know it's the
4788 : * right one - there can't be a better one. matching more.
4789 : */
4790 702 : if (item)
4791 426 : break;
4792 : }
4793 :
4794 : /*
4795 : * Make sure we found an item. There has to be one, because ndistinct
4796 : * statistics includes all combinations of attributes.
4797 : */
4798 426 : if (!item)
4799 0 : elog(ERROR, "corrupt MVNDistinct entry");
4800 :
4801 : /* Form the output varinfo list, keeping only unmatched ones */
4802 1488 : foreach(lc, *varinfos)
4803 : {
4804 1062 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4805 : ListCell *lc3;
4806 1062 : bool found = false;
4807 :
4808 : /*
4809 : * Let's look at plain variables first, because it's the most
4810 : * common case and the check is quite cheap. We can simply get the
4811 : * attnum and check (with an offset) matched bitmap.
4812 : */
4813 1062 : if (IsA(varinfo->var, Var))
4814 870 : {
4815 876 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4816 :
4817 : /*
4818 : * If it's a system attribute, we're done. We don't support
4819 : * extended statistics on system attributes, so it's clearly
4820 : * not matched. Just keep the expression and continue.
4821 : */
4822 876 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4823 : {
4824 6 : newlist = lappend(newlist, varinfo);
4825 6 : continue;
4826 : }
4827 :
4828 : /* apply the same offset as above */
4829 870 : attnum += attnum_offset;
4830 :
4831 : /* if it's not matched, keep the varinfo */
4832 870 : if (!bms_is_member(attnum, matched))
4833 120 : newlist = lappend(newlist, varinfo);
4834 :
4835 : /* The rest of the loop deals with complex expressions. */
4836 870 : continue;
4837 : }
4838 :
4839 : /*
4840 : * Process complex expressions, not just simple Vars.
4841 : *
4842 : * First, we search for an exact match of an expression. If we
4843 : * find one, we can just discard the whole GroupVarInfo, with all
4844 : * the variables we extracted from it.
4845 : *
4846 : * Otherwise we inspect the individual vars, and try matching it
4847 : * to variables in the item.
4848 : */
4849 306 : foreach(lc3, matched_info->exprs)
4850 : {
4851 276 : Node *expr = (Node *) lfirst(lc3);
4852 :
4853 276 : if (equal(varinfo->var, expr))
4854 : {
4855 156 : found = true;
4856 156 : break;
4857 : }
4858 : }
4859 :
4860 : /* found exact match, skip */
4861 186 : if (found)
4862 156 : continue;
4863 :
4864 30 : newlist = lappend(newlist, varinfo);
4865 : }
4866 :
4867 426 : *varinfos = newlist;
4868 426 : *ndistinct = item->ndistinct;
4869 426 : return true;
4870 : }
4871 :
4872 0 : return false;
4873 : }
4874 :
4875 : /*
4876 : * convert_to_scalar
4877 : * Convert non-NULL values of the indicated types to the comparison
4878 : * scale needed by scalarineqsel().
4879 : * Returns "true" if successful.
4880 : *
4881 : * XXX this routine is a hack: ideally we should look up the conversion
4882 : * subroutines in pg_type.
4883 : *
4884 : * All numeric datatypes are simply converted to their equivalent
4885 : * "double" values. (NUMERIC values that are outside the range of "double"
4886 : * are clamped to +/- HUGE_VAL.)
4887 : *
4888 : * String datatypes are converted by convert_string_to_scalar(),
4889 : * which is explained below. The reason why this routine deals with
4890 : * three values at a time, not just one, is that we need it for strings.
4891 : *
4892 : * The bytea datatype is just enough different from strings that it has
4893 : * to be treated separately.
4894 : *
4895 : * The several datatypes representing absolute times are all converted
4896 : * to Timestamp, which is actually an int64, and then we promote that to
4897 : * a double. Note this will give correct results even for the "special"
4898 : * values of Timestamp, since those are chosen to compare correctly;
4899 : * see timestamp_cmp.
4900 : *
4901 : * The several datatypes representing relative times (intervals) are all
4902 : * converted to measurements expressed in seconds.
4903 : */
4904 : static bool
4905 92224 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4906 : Datum lobound, Datum hibound, Oid boundstypid,
4907 : double *scaledlobound, double *scaledhibound)
4908 : {
4909 92224 : bool failure = false;
4910 :
4911 : /*
4912 : * Both the valuetypid and the boundstypid should exactly match the
4913 : * declared input type(s) of the operator we are invoked for. However,
4914 : * extensions might try to use scalarineqsel as estimator for operators
4915 : * with input type(s) we don't handle here; in such cases, we want to
4916 : * return false, not fail. In any case, we mustn't assume that valuetypid
4917 : * and boundstypid are identical.
4918 : *
4919 : * XXX The histogram we are interpolating between points of could belong
4920 : * to a column that's only binary-compatible with the declared type. In
4921 : * essence we are assuming that the semantics of binary-compatible types
4922 : * are enough alike that we can use a histogram generated with one type's
4923 : * operators to estimate selectivity for the other's. This is outright
4924 : * wrong in some cases --- in particular signed versus unsigned
4925 : * interpretation could trip us up. But it's useful enough in the
4926 : * majority of cases that we do it anyway. Should think about more
4927 : * rigorous ways to do it.
4928 : */
4929 92224 : switch (valuetypid)
4930 : {
4931 : /*
4932 : * Built-in numeric types
4933 : */
4934 85360 : case BOOLOID:
4935 : case INT2OID:
4936 : case INT4OID:
4937 : case INT8OID:
4938 : case FLOAT4OID:
4939 : case FLOAT8OID:
4940 : case NUMERICOID:
4941 : case OIDOID:
4942 : case REGPROCOID:
4943 : case REGPROCEDUREOID:
4944 : case REGOPEROID:
4945 : case REGOPERATOROID:
4946 : case REGCLASSOID:
4947 : case REGTYPEOID:
4948 : case REGCOLLATIONOID:
4949 : case REGCONFIGOID:
4950 : case REGDICTIONARYOID:
4951 : case REGROLEOID:
4952 : case REGNAMESPACEOID:
4953 : case REGDATABASEOID:
4954 85360 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4955 : &failure);
4956 85360 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4957 : &failure);
4958 85360 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4959 : &failure);
4960 85360 : return !failure;
4961 :
4962 : /*
4963 : * Built-in string types
4964 : */
4965 6864 : case CHAROID:
4966 : case BPCHAROID:
4967 : case VARCHAROID:
4968 : case TEXTOID:
4969 : case NAMEOID:
4970 : {
4971 6864 : char *valstr = convert_string_datum(value, valuetypid,
4972 : collid, &failure);
4973 6864 : char *lostr = convert_string_datum(lobound, boundstypid,
4974 : collid, &failure);
4975 6864 : char *histr = convert_string_datum(hibound, boundstypid,
4976 : collid, &failure);
4977 :
4978 : /*
4979 : * Bail out if any of the values is not of string type. We
4980 : * might leak converted strings for the other value(s), but
4981 : * that's not worth troubling over.
4982 : */
4983 6864 : if (failure)
4984 0 : return false;
4985 :
4986 6864 : convert_string_to_scalar(valstr, scaledvalue,
4987 : lostr, scaledlobound,
4988 : histr, scaledhibound);
4989 6864 : pfree(valstr);
4990 6864 : pfree(lostr);
4991 6864 : pfree(histr);
4992 6864 : return true;
4993 : }
4994 :
4995 : /*
4996 : * Built-in bytea type
4997 : */
4998 0 : case BYTEAOID:
4999 : {
5000 : /* We only support bytea vs bytea comparison */
5001 0 : if (boundstypid != BYTEAOID)
5002 0 : return false;
5003 0 : convert_bytea_to_scalar(value, scaledvalue,
5004 : lobound, scaledlobound,
5005 : hibound, scaledhibound);
5006 0 : return true;
5007 : }
5008 :
5009 : /*
5010 : * Built-in time types
5011 : */
5012 0 : case TIMESTAMPOID:
5013 : case TIMESTAMPTZOID:
5014 : case DATEOID:
5015 : case INTERVALOID:
5016 : case TIMEOID:
5017 : case TIMETZOID:
5018 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5019 : &failure);
5020 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5021 : &failure);
5022 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5023 : &failure);
5024 0 : return !failure;
5025 :
5026 : /*
5027 : * Built-in network types
5028 : */
5029 0 : case INETOID:
5030 : case CIDROID:
5031 : case MACADDROID:
5032 : case MACADDR8OID:
5033 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5034 : &failure);
5035 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5036 : &failure);
5037 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5038 : &failure);
5039 0 : return !failure;
5040 : }
5041 : /* Don't know how to convert */
5042 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
5043 0 : return false;
5044 : }
5045 :
5046 : /*
5047 : * Do convert_to_scalar()'s work for any numeric data type.
5048 : *
5049 : * On failure (e.g., unsupported typid), set *failure to true;
5050 : * otherwise, that variable is not changed.
5051 : */
5052 : static double
5053 256080 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5054 : {
5055 256080 : switch (typid)
5056 : {
5057 0 : case BOOLOID:
5058 0 : return (double) DatumGetBool(value);
5059 12 : case INT2OID:
5060 12 : return (double) DatumGetInt16(value);
5061 31770 : case INT4OID:
5062 31770 : return (double) DatumGetInt32(value);
5063 0 : case INT8OID:
5064 0 : return (double) DatumGetInt64(value);
5065 0 : case FLOAT4OID:
5066 0 : return (double) DatumGetFloat4(value);
5067 54 : case FLOAT8OID:
5068 54 : return (double) DatumGetFloat8(value);
5069 0 : case NUMERICOID:
5070 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
5071 0 : return (double)
5072 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5073 : value));
5074 224244 : case OIDOID:
5075 : case REGPROCOID:
5076 : case REGPROCEDUREOID:
5077 : case REGOPEROID:
5078 : case REGOPERATOROID:
5079 : case REGCLASSOID:
5080 : case REGTYPEOID:
5081 : case REGCOLLATIONOID:
5082 : case REGCONFIGOID:
5083 : case REGDICTIONARYOID:
5084 : case REGROLEOID:
5085 : case REGNAMESPACEOID:
5086 : case REGDATABASEOID:
5087 : /* we can treat OIDs as integers... */
5088 224244 : return (double) DatumGetObjectId(value);
5089 : }
5090 :
5091 0 : *failure = true;
5092 0 : return 0;
5093 : }
5094 :
5095 : /*
5096 : * Do convert_to_scalar()'s work for any character-string data type.
5097 : *
5098 : * String datatypes are converted to a scale that ranges from 0 to 1,
5099 : * where we visualize the bytes of the string as fractional digits.
5100 : *
5101 : * We do not want the base to be 256, however, since that tends to
5102 : * generate inflated selectivity estimates; few databases will have
5103 : * occurrences of all 256 possible byte values at each position.
5104 : * Instead, use the smallest and largest byte values seen in the bounds
5105 : * as the estimated range for each byte, after some fudging to deal with
5106 : * the fact that we probably aren't going to see the full range that way.
5107 : *
5108 : * An additional refinement is that we discard any common prefix of the
5109 : * three strings before computing the scaled values. This allows us to
5110 : * "zoom in" when we encounter a narrow data range. An example is a phone
5111 : * number database where all the values begin with the same area code.
5112 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5113 : * so this is more likely to happen than you might think.)
5114 : */
5115 : static void
5116 6864 : convert_string_to_scalar(char *value,
5117 : double *scaledvalue,
5118 : char *lobound,
5119 : double *scaledlobound,
5120 : char *hibound,
5121 : double *scaledhibound)
5122 : {
5123 : int rangelo,
5124 : rangehi;
5125 : char *sptr;
5126 :
5127 6864 : rangelo = rangehi = (unsigned char) hibound[0];
5128 83798 : for (sptr = lobound; *sptr; sptr++)
5129 : {
5130 76934 : if (rangelo > (unsigned char) *sptr)
5131 16152 : rangelo = (unsigned char) *sptr;
5132 76934 : if (rangehi < (unsigned char) *sptr)
5133 8510 : rangehi = (unsigned char) *sptr;
5134 : }
5135 81864 : for (sptr = hibound; *sptr; sptr++)
5136 : {
5137 75000 : if (rangelo > (unsigned char) *sptr)
5138 1304 : rangelo = (unsigned char) *sptr;
5139 75000 : if (rangehi < (unsigned char) *sptr)
5140 3118 : rangehi = (unsigned char) *sptr;
5141 : }
5142 : /* If range includes any upper-case ASCII chars, make it include all */
5143 6864 : if (rangelo <= 'Z' && rangehi >= 'A')
5144 : {
5145 1426 : if (rangelo > 'A')
5146 222 : rangelo = 'A';
5147 1426 : if (rangehi < 'Z')
5148 480 : rangehi = 'Z';
5149 : }
5150 : /* Ditto lower-case */
5151 6864 : if (rangelo <= 'z' && rangehi >= 'a')
5152 : {
5153 6362 : if (rangelo > 'a')
5154 38 : rangelo = 'a';
5155 6362 : if (rangehi < 'z')
5156 6280 : rangehi = 'z';
5157 : }
5158 : /* Ditto digits */
5159 6864 : if (rangelo <= '9' && rangehi >= '0')
5160 : {
5161 746 : if (rangelo > '0')
5162 652 : rangelo = '0';
5163 746 : if (rangehi < '9')
5164 14 : rangehi = '9';
5165 : }
5166 :
5167 : /*
5168 : * If range includes less than 10 chars, assume we have not got enough
5169 : * data, and make it include regular ASCII set.
5170 : */
5171 6864 : if (rangehi - rangelo < 9)
5172 : {
5173 0 : rangelo = ' ';
5174 0 : rangehi = 127;
5175 : }
5176 :
5177 : /*
5178 : * Now strip any common prefix of the three strings.
5179 : */
5180 14286 : while (*lobound)
5181 : {
5182 14286 : if (*lobound != *hibound || *lobound != *value)
5183 : break;
5184 7422 : lobound++, hibound++, value++;
5185 : }
5186 :
5187 : /*
5188 : * Now we can do the conversions.
5189 : */
5190 6864 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5191 6864 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5192 6864 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5193 6864 : }
5194 :
5195 : static double
5196 20592 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5197 : {
5198 20592 : int slen = strlen(value);
5199 : double num,
5200 : denom,
5201 : base;
5202 :
5203 20592 : if (slen <= 0)
5204 0 : return 0.0; /* empty string has scalar value 0 */
5205 :
5206 : /*
5207 : * There seems little point in considering more than a dozen bytes from
5208 : * the string. Since base is at least 10, that will give us nominal
5209 : * resolution of at least 12 decimal digits, which is surely far more
5210 : * precision than this estimation technique has got anyway (especially in
5211 : * non-C locales). Also, even with the maximum possible base of 256, this
5212 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5213 : * overflow on any known machine.
5214 : */
5215 20592 : if (slen > 12)
5216 5210 : slen = 12;
5217 :
5218 : /* Convert initial characters to fraction */
5219 20592 : base = rangehi - rangelo + 1;
5220 20592 : num = 0.0;
5221 20592 : denom = base;
5222 170346 : while (slen-- > 0)
5223 : {
5224 149754 : int ch = (unsigned char) *value++;
5225 :
5226 149754 : if (ch < rangelo)
5227 152 : ch = rangelo - 1;
5228 149602 : else if (ch > rangehi)
5229 0 : ch = rangehi + 1;
5230 149754 : num += ((double) (ch - rangelo)) / denom;
5231 149754 : denom *= base;
5232 : }
5233 :
5234 20592 : return num;
5235 : }
5236 :
5237 : /*
5238 : * Convert a string-type Datum into a palloc'd, null-terminated string.
5239 : *
5240 : * On failure (e.g., unsupported typid), set *failure to true;
5241 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5242 : *
5243 : * When using a non-C locale, we must pass the string through pg_strxfrm()
5244 : * before continuing, so as to generate correct locale-specific results.
5245 : */
5246 : static char *
5247 20592 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5248 : {
5249 : char *val;
5250 : pg_locale_t mylocale;
5251 :
5252 20592 : switch (typid)
5253 : {
5254 0 : case CHAROID:
5255 0 : val = (char *) palloc(2);
5256 0 : val[0] = DatumGetChar(value);
5257 0 : val[1] = '\0';
5258 0 : break;
5259 6328 : case BPCHAROID:
5260 : case VARCHAROID:
5261 : case TEXTOID:
5262 6328 : val = TextDatumGetCString(value);
5263 6328 : break;
5264 14264 : case NAMEOID:
5265 : {
5266 14264 : NameData *nm = (NameData *) DatumGetPointer(value);
5267 :
5268 14264 : val = pstrdup(NameStr(*nm));
5269 14264 : break;
5270 : }
5271 0 : default:
5272 0 : *failure = true;
5273 0 : return NULL;
5274 : }
5275 :
5276 20592 : mylocale = pg_newlocale_from_collation(collid);
5277 :
5278 20592 : if (!mylocale->collate_is_c)
5279 : {
5280 : char *xfrmstr;
5281 : size_t xfrmlen;
5282 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5283 :
5284 : /*
5285 : * XXX: We could guess at a suitable output buffer size and only call
5286 : * pg_strxfrm() twice if our guess is too small.
5287 : *
5288 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5289 : * bogus data or set an error. This is not really a problem unless it
5290 : * crashes since it will only give an estimation error and nothing
5291 : * fatal.
5292 : *
5293 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5294 : * some cases, libc strxfrm() may return the wrong results, but that
5295 : * will only lead to an estimation error.
5296 : */
5297 72 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5298 : #ifdef WIN32
5299 :
5300 : /*
5301 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5302 : * of trying to allocate this much memory (and fail), just return the
5303 : * original string unmodified as if we were in the C locale.
5304 : */
5305 : if (xfrmlen == INT_MAX)
5306 : return val;
5307 : #endif
5308 72 : xfrmstr = (char *) palloc(xfrmlen + 1);
5309 72 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5310 :
5311 : /*
5312 : * Some systems (e.g., glibc) can return a smaller value from the
5313 : * second call than the first; thus the Assert must be <= not ==.
5314 : */
5315 : Assert(xfrmlen2 <= xfrmlen);
5316 72 : pfree(val);
5317 72 : val = xfrmstr;
5318 : }
5319 :
5320 20592 : return val;
5321 : }
5322 :
5323 : /*
5324 : * Do convert_to_scalar()'s work for any bytea data type.
5325 : *
5326 : * Very similar to convert_string_to_scalar except we can't assume
5327 : * null-termination and therefore pass explicit lengths around.
5328 : *
5329 : * Also, assumptions about likely "normal" ranges of characters have been
5330 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5331 : * someday we will add information about actual byte data range to
5332 : * pg_statistic.)
5333 : */
5334 : static void
5335 0 : convert_bytea_to_scalar(Datum value,
5336 : double *scaledvalue,
5337 : Datum lobound,
5338 : double *scaledlobound,
5339 : Datum hibound,
5340 : double *scaledhibound)
5341 : {
5342 0 : bytea *valuep = DatumGetByteaPP(value);
5343 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5344 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5345 : int rangelo,
5346 : rangehi,
5347 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5348 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5349 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5350 : i,
5351 : minlen;
5352 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5353 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5354 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5355 :
5356 : /*
5357 : * Assume bytea data is uniformly distributed across all byte values.
5358 : */
5359 0 : rangelo = 0;
5360 0 : rangehi = 255;
5361 :
5362 : /*
5363 : * Now strip any common prefix of the three strings.
5364 : */
5365 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5366 0 : for (i = 0; i < minlen; i++)
5367 : {
5368 0 : if (*lostr != *histr || *lostr != *valstr)
5369 : break;
5370 0 : lostr++, histr++, valstr++;
5371 0 : loboundlen--, hiboundlen--, valuelen--;
5372 : }
5373 :
5374 : /*
5375 : * Now we can do the conversions.
5376 : */
5377 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5378 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5379 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5380 0 : }
5381 :
5382 : static double
5383 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5384 : int rangelo, int rangehi)
5385 : {
5386 : double num,
5387 : denom,
5388 : base;
5389 :
5390 0 : if (valuelen <= 0)
5391 0 : return 0.0; /* empty string has scalar value 0 */
5392 :
5393 : /*
5394 : * Since base is 256, need not consider more than about 10 chars (even
5395 : * this many seems like overkill)
5396 : */
5397 0 : if (valuelen > 10)
5398 0 : valuelen = 10;
5399 :
5400 : /* Convert initial characters to fraction */
5401 0 : base = rangehi - rangelo + 1;
5402 0 : num = 0.0;
5403 0 : denom = base;
5404 0 : while (valuelen-- > 0)
5405 : {
5406 0 : int ch = *value++;
5407 :
5408 0 : if (ch < rangelo)
5409 0 : ch = rangelo - 1;
5410 0 : else if (ch > rangehi)
5411 0 : ch = rangehi + 1;
5412 0 : num += ((double) (ch - rangelo)) / denom;
5413 0 : denom *= base;
5414 : }
5415 :
5416 0 : return num;
5417 : }
5418 :
5419 : /*
5420 : * Do convert_to_scalar()'s work for any timevalue data type.
5421 : *
5422 : * On failure (e.g., unsupported typid), set *failure to true;
5423 : * otherwise, that variable is not changed.
5424 : */
5425 : static double
5426 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5427 : {
5428 0 : switch (typid)
5429 : {
5430 0 : case TIMESTAMPOID:
5431 0 : return DatumGetTimestamp(value);
5432 0 : case TIMESTAMPTZOID:
5433 0 : return DatumGetTimestampTz(value);
5434 0 : case DATEOID:
5435 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5436 0 : case INTERVALOID:
5437 : {
5438 0 : Interval *interval = DatumGetIntervalP(value);
5439 :
5440 : /*
5441 : * Convert the month part of Interval to days using assumed
5442 : * average month length of 365.25/12.0 days. Not too
5443 : * accurate, but plenty good enough for our purposes.
5444 : *
5445 : * This also works for infinite intervals, which just have all
5446 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5447 : * smaller/larger than any finite interval.
5448 : */
5449 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5450 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5451 : }
5452 0 : case TIMEOID:
5453 0 : return DatumGetTimeADT(value);
5454 0 : case TIMETZOID:
5455 : {
5456 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5457 :
5458 : /* use GMT-equivalent time */
5459 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5460 : }
5461 : }
5462 :
5463 0 : *failure = true;
5464 0 : return 0;
5465 : }
5466 :
5467 :
5468 : /*
5469 : * get_restriction_variable
5470 : * Examine the args of a restriction clause to see if it's of the
5471 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5472 : * where "variable" could be either a Var or an expression in vars of a
5473 : * single relation. If so, extract information about the variable,
5474 : * and also indicate which side it was on and the other argument.
5475 : *
5476 : * Inputs:
5477 : * root: the planner info
5478 : * args: clause argument list
5479 : * varRelid: see specs for restriction selectivity functions
5480 : *
5481 : * Outputs: (these are valid only if true is returned)
5482 : * *vardata: gets information about variable (see examine_variable)
5483 : * *other: gets other clause argument, aggressively reduced to a constant
5484 : * *varonleft: set true if variable is on the left, false if on the right
5485 : *
5486 : * Returns true if a variable is identified, otherwise false.
5487 : *
5488 : * Note: if there are Vars on both sides of the clause, we must fail, because
5489 : * callers are expecting that the other side will act like a pseudoconstant.
5490 : */
5491 : bool
5492 802346 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5493 : VariableStatData *vardata, Node **other,
5494 : bool *varonleft)
5495 : {
5496 : Node *left,
5497 : *right;
5498 : VariableStatData rdata;
5499 :
5500 : /* Fail if not a binary opclause (probably shouldn't happen) */
5501 802346 : if (list_length(args) != 2)
5502 0 : return false;
5503 :
5504 802346 : left = (Node *) linitial(args);
5505 802346 : right = (Node *) lsecond(args);
5506 :
5507 : /*
5508 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5509 : * relations will be treated as pseudoconstants.
5510 : */
5511 802346 : examine_variable(root, left, varRelid, vardata);
5512 802346 : examine_variable(root, right, varRelid, &rdata);
5513 :
5514 : /*
5515 : * If one side is a variable and the other not, we win.
5516 : */
5517 802346 : if (vardata->rel && rdata.rel == NULL)
5518 : {
5519 715128 : *varonleft = true;
5520 715128 : *other = estimate_expression_value(root, rdata.var);
5521 : /* Assume we need no ReleaseVariableStats(rdata) here */
5522 715122 : return true;
5523 : }
5524 :
5525 87218 : if (vardata->rel == NULL && rdata.rel)
5526 : {
5527 80754 : *varonleft = false;
5528 80754 : *other = estimate_expression_value(root, vardata->var);
5529 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5530 80754 : *vardata = rdata;
5531 80754 : return true;
5532 : }
5533 :
5534 : /* Oops, clause has wrong structure (probably var op var) */
5535 6464 : ReleaseVariableStats(*vardata);
5536 6464 : ReleaseVariableStats(rdata);
5537 :
5538 6464 : return false;
5539 : }
5540 :
5541 : /*
5542 : * get_join_variables
5543 : * Apply examine_variable() to each side of a join clause.
5544 : * Also, attempt to identify whether the join clause has the same
5545 : * or reversed sense compared to the SpecialJoinInfo.
5546 : *
5547 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5548 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5549 : * where we can't tell for sure, we default to assuming it's normal.
5550 : */
5551 : void
5552 265626 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5553 : VariableStatData *vardata1, VariableStatData *vardata2,
5554 : bool *join_is_reversed)
5555 : {
5556 : Node *left,
5557 : *right;
5558 :
5559 265626 : if (list_length(args) != 2)
5560 0 : elog(ERROR, "join operator should take two arguments");
5561 :
5562 265626 : left = (Node *) linitial(args);
5563 265626 : right = (Node *) lsecond(args);
5564 :
5565 265626 : examine_variable(root, left, 0, vardata1);
5566 265626 : examine_variable(root, right, 0, vardata2);
5567 :
5568 531072 : if (vardata1->rel &&
5569 265446 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5570 89950 : *join_is_reversed = true; /* var1 is on RHS */
5571 351206 : else if (vardata2->rel &&
5572 175530 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5573 146 : *join_is_reversed = true; /* var2 is on LHS */
5574 : else
5575 175530 : *join_is_reversed = false;
5576 265626 : }
5577 :
5578 : /* statext_expressions_load copies the tuple, so just pfree it. */
5579 : static void
5580 1650 : ReleaseDummy(HeapTuple tuple)
5581 : {
5582 1650 : pfree(tuple);
5583 1650 : }
5584 :
5585 : /*
5586 : * examine_variable
5587 : * Try to look up statistical data about an expression.
5588 : * Fill in a VariableStatData struct to describe the expression.
5589 : *
5590 : * Inputs:
5591 : * root: the planner info
5592 : * node: the expression tree to examine
5593 : * varRelid: see specs for restriction selectivity functions
5594 : *
5595 : * Outputs: *vardata is filled as follows:
5596 : * var: the input expression (with any binary relabeling stripped, if
5597 : * it is or contains a variable; but otherwise the type is preserved)
5598 : * rel: RelOptInfo for relation containing variable; NULL if expression
5599 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5600 : * subquery, not one in the current query).
5601 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5602 : * otherwise NULL.
5603 : * freefunc: pointer to a function to release statsTuple with.
5604 : * vartype: exposed type of the expression; this should always match
5605 : * the declared input type of the operator we are estimating for.
5606 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5607 : * commonly the same as the exposed type of the variable argument,
5608 : * but can be different in binary-compatible-type cases.
5609 : * isunique: true if we were able to match the var to a unique index, a
5610 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5611 : * unique for this query. (Caution: this should be trusted for
5612 : * statistical purposes only, since we do not check indimmediate nor
5613 : * verify that the exact same definition of equality applies.)
5614 : * acl_ok: true if current user has permission to read all table rows from
5615 : * the column(s) underlying the pg_statistic entry. This is consulted by
5616 : * statistic_proc_security_check().
5617 : *
5618 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5619 : */
5620 : void
5621 3240222 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5622 : VariableStatData *vardata)
5623 : {
5624 : Node *basenode;
5625 : Relids varnos;
5626 : Relids basevarnos;
5627 : RelOptInfo *onerel;
5628 :
5629 : /* Make sure we don't return dangling pointers in vardata */
5630 22681554 : MemSet(vardata, 0, sizeof(VariableStatData));
5631 :
5632 : /* Save the exposed type of the expression */
5633 3240222 : vardata->vartype = exprType(node);
5634 :
5635 : /* Look inside any binary-compatible relabeling */
5636 :
5637 3240222 : if (IsA(node, RelabelType))
5638 47036 : basenode = (Node *) ((RelabelType *) node)->arg;
5639 : else
5640 3193186 : basenode = node;
5641 :
5642 : /* Fast path for a simple Var */
5643 :
5644 3240222 : if (IsA(basenode, Var) &&
5645 782510 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5646 : {
5647 2324718 : Var *var = (Var *) basenode;
5648 :
5649 : /* Set up result fields other than the stats tuple */
5650 2324718 : vardata->var = basenode; /* return Var without relabeling */
5651 2324718 : vardata->rel = find_base_rel(root, var->varno);
5652 2324718 : vardata->atttype = var->vartype;
5653 2324718 : vardata->atttypmod = var->vartypmod;
5654 2324718 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5655 :
5656 : /* Try to locate some stats */
5657 2324718 : examine_simple_variable(root, var, vardata);
5658 :
5659 2324718 : return;
5660 : }
5661 :
5662 : /*
5663 : * Okay, it's a more complicated expression. Determine variable
5664 : * membership. Note that when varRelid isn't zero, only vars of that
5665 : * relation are considered "real" vars.
5666 : */
5667 915504 : varnos = pull_varnos(root, basenode);
5668 915504 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5669 :
5670 915504 : onerel = NULL;
5671 :
5672 915504 : if (bms_is_empty(basevarnos))
5673 : {
5674 : /* No Vars at all ... must be pseudo-constant clause */
5675 : }
5676 : else
5677 : {
5678 : int relid;
5679 :
5680 : /* Check if the expression is in vars of a single base relation */
5681 469000 : if (bms_get_singleton_member(basevarnos, &relid))
5682 : {
5683 462132 : if (varRelid == 0 || varRelid == relid)
5684 : {
5685 69374 : onerel = find_base_rel(root, relid);
5686 69374 : vardata->rel = onerel;
5687 69374 : node = basenode; /* strip any relabeling */
5688 : }
5689 : /* else treat it as a constant */
5690 : }
5691 : else
5692 : {
5693 : /* varnos has multiple relids */
5694 6868 : if (varRelid == 0)
5695 : {
5696 : /* treat it as a variable of a join relation */
5697 5416 : vardata->rel = find_join_rel(root, varnos);
5698 5416 : node = basenode; /* strip any relabeling */
5699 : }
5700 1452 : else if (bms_is_member(varRelid, varnos))
5701 : {
5702 : /* ignore the vars belonging to other relations */
5703 1278 : vardata->rel = find_base_rel(root, varRelid);
5704 1278 : node = basenode; /* strip any relabeling */
5705 : /* note: no point in expressional-index search here */
5706 : }
5707 : /* else treat it as a constant */
5708 : }
5709 : }
5710 :
5711 915504 : bms_free(basevarnos);
5712 :
5713 915504 : vardata->var = node;
5714 915504 : vardata->atttype = exprType(node);
5715 915504 : vardata->atttypmod = exprTypmod(node);
5716 :
5717 915504 : if (onerel)
5718 : {
5719 : /*
5720 : * We have an expression in vars of a single relation. Try to match
5721 : * it to expressional index columns, in hopes of finding some
5722 : * statistics.
5723 : *
5724 : * Note that we consider all index columns including INCLUDE columns,
5725 : * since there could be stats for such columns. But the test for
5726 : * uniqueness needs to be warier.
5727 : *
5728 : * XXX it's conceivable that there are multiple matches with different
5729 : * index opfamilies; if so, we need to pick one that matches the
5730 : * operator we are estimating for. FIXME later.
5731 : */
5732 : ListCell *ilist;
5733 : ListCell *slist;
5734 :
5735 : /*
5736 : * The nullingrels bits within the expression could prevent us from
5737 : * matching it to expressional index columns or to the expressions in
5738 : * extended statistics. So strip them out first.
5739 : */
5740 69374 : if (bms_overlap(varnos, root->outer_join_rels))
5741 3088 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5742 :
5743 149404 : foreach(ilist, onerel->indexlist)
5744 : {
5745 82988 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5746 : ListCell *indexpr_item;
5747 : int pos;
5748 :
5749 82988 : indexpr_item = list_head(index->indexprs);
5750 82988 : if (indexpr_item == NULL)
5751 78098 : continue; /* no expressions here... */
5752 :
5753 6894 : for (pos = 0; pos < index->ncolumns; pos++)
5754 : {
5755 4962 : if (index->indexkeys[pos] == 0)
5756 : {
5757 : Node *indexkey;
5758 :
5759 4890 : if (indexpr_item == NULL)
5760 0 : elog(ERROR, "too few entries in indexprs list");
5761 4890 : indexkey = (Node *) lfirst(indexpr_item);
5762 4890 : if (indexkey && IsA(indexkey, RelabelType))
5763 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5764 4890 : if (equal(node, indexkey))
5765 : {
5766 : /*
5767 : * Found a match ... is it a unique index? Tests here
5768 : * should match has_unique_index().
5769 : */
5770 3594 : if (index->unique &&
5771 438 : index->nkeycolumns == 1 &&
5772 438 : pos == 0 &&
5773 438 : (index->indpred == NIL || index->predOK))
5774 438 : vardata->isunique = true;
5775 :
5776 : /*
5777 : * Has it got stats? We only consider stats for
5778 : * non-partial indexes, since partial indexes probably
5779 : * don't reflect whole-relation statistics; the above
5780 : * check for uniqueness is the only info we take from
5781 : * a partial index.
5782 : *
5783 : * An index stats hook, however, must make its own
5784 : * decisions about what to do with partial indexes.
5785 : */
5786 3594 : if (get_index_stats_hook &&
5787 0 : (*get_index_stats_hook) (root, index->indexoid,
5788 0 : pos + 1, vardata))
5789 : {
5790 : /*
5791 : * The hook took control of acquiring a stats
5792 : * tuple. If it did supply a tuple, it'd better
5793 : * have supplied a freefunc.
5794 : */
5795 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5796 0 : !vardata->freefunc)
5797 0 : elog(ERROR, "no function provided to release variable stats with");
5798 : }
5799 3594 : else if (index->indpred == NIL)
5800 : {
5801 3594 : vardata->statsTuple =
5802 7188 : SearchSysCache3(STATRELATTINH,
5803 : ObjectIdGetDatum(index->indexoid),
5804 3594 : Int16GetDatum(pos + 1),
5805 : BoolGetDatum(false));
5806 3594 : vardata->freefunc = ReleaseSysCache;
5807 :
5808 3594 : if (HeapTupleIsValid(vardata->statsTuple))
5809 : {
5810 : /*
5811 : * Test if user has permission to access all
5812 : * rows from the index's table.
5813 : *
5814 : * For simplicity, we insist on the whole
5815 : * table being selectable, rather than trying
5816 : * to identify which column(s) the index
5817 : * depends on.
5818 : *
5819 : * Note that for an inheritance child,
5820 : * permissions are checked on the inheritance
5821 : * root parent, and whole-table select
5822 : * privilege on the parent doesn't quite
5823 : * guarantee that the user could read all
5824 : * columns of the child. But in practice it's
5825 : * unlikely that any interesting security
5826 : * violation could result from allowing access
5827 : * to the expression index's stats, so we
5828 : * allow it anyway. See similar code in
5829 : * examine_simple_variable() for additional
5830 : * comments.
5831 : */
5832 2958 : vardata->acl_ok =
5833 2958 : all_rows_selectable(root,
5834 2958 : index->rel->relid,
5835 : NULL);
5836 : }
5837 : else
5838 : {
5839 : /* suppress leakproofness checks later */
5840 636 : vardata->acl_ok = true;
5841 : }
5842 : }
5843 3594 : if (vardata->statsTuple)
5844 2958 : break;
5845 : }
5846 1932 : indexpr_item = lnext(index->indexprs, indexpr_item);
5847 : }
5848 : }
5849 4890 : if (vardata->statsTuple)
5850 2958 : break;
5851 : }
5852 :
5853 : /*
5854 : * Search extended statistics for one with a matching expression.
5855 : * There might be multiple ones, so just grab the first one. In the
5856 : * future, we might consider the statistics target (and pick the most
5857 : * accurate statistics) and maybe some other parameters.
5858 : */
5859 73490 : foreach(slist, onerel->statlist)
5860 : {
5861 4404 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5862 4404 : RangeTblEntry *rte = planner_rt_fetch(onerel->relid, root);
5863 : ListCell *expr_item;
5864 : int pos;
5865 :
5866 : /*
5867 : * Stop once we've found statistics for the expression (either
5868 : * from extended stats, or for an index in the preceding loop).
5869 : */
5870 4404 : if (vardata->statsTuple)
5871 288 : break;
5872 :
5873 : /* skip stats without per-expression stats */
5874 4116 : if (info->kind != STATS_EXT_EXPRESSIONS)
5875 2106 : continue;
5876 :
5877 : /* skip stats with mismatching stxdinherit value */
5878 2010 : if (info->inherit != rte->inh)
5879 6 : continue;
5880 :
5881 2004 : pos = 0;
5882 3306 : foreach(expr_item, info->exprs)
5883 : {
5884 2952 : Node *expr = (Node *) lfirst(expr_item);
5885 :
5886 : Assert(expr);
5887 :
5888 : /* strip RelabelType before comparing it */
5889 2952 : if (expr && IsA(expr, RelabelType))
5890 0 : expr = (Node *) ((RelabelType *) expr)->arg;
5891 :
5892 : /* found a match, see if we can extract pg_statistic row */
5893 2952 : if (equal(node, expr))
5894 : {
5895 : /*
5896 : * XXX Not sure if we should cache the tuple somewhere.
5897 : * Now we just create a new copy every time.
5898 : */
5899 1650 : vardata->statsTuple =
5900 1650 : statext_expressions_load(info->statOid, rte->inh, pos);
5901 :
5902 1650 : vardata->freefunc = ReleaseDummy;
5903 :
5904 : /*
5905 : * Test if user has permission to access all rows from the
5906 : * table.
5907 : *
5908 : * For simplicity, we insist on the whole table being
5909 : * selectable, rather than trying to identify which
5910 : * column(s) the statistics object depends on.
5911 : *
5912 : * Note that for an inheritance child, permissions are
5913 : * checked on the inheritance root parent, and whole-table
5914 : * select privilege on the parent doesn't quite guarantee
5915 : * that the user could read all columns of the child. But
5916 : * in practice it's unlikely that any interesting security
5917 : * violation could result from allowing access to the
5918 : * expression stats, so we allow it anyway. See similar
5919 : * code in examine_simple_variable() for additional
5920 : * comments.
5921 : */
5922 1650 : vardata->acl_ok = all_rows_selectable(root,
5923 : onerel->relid,
5924 : NULL);
5925 :
5926 1650 : break;
5927 : }
5928 :
5929 1302 : pos++;
5930 : }
5931 : }
5932 : }
5933 :
5934 915504 : bms_free(varnos);
5935 : }
5936 :
5937 : /*
5938 : * examine_simple_variable
5939 : * Handle a simple Var for examine_variable
5940 : *
5941 : * This is split out as a subroutine so that we can recurse to deal with
5942 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
5943 : *
5944 : * We already filled in all the fields of *vardata except for the stats tuple.
5945 : */
5946 : static void
5947 2330958 : examine_simple_variable(PlannerInfo *root, Var *var,
5948 : VariableStatData *vardata)
5949 : {
5950 2330958 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
5951 :
5952 : Assert(IsA(rte, RangeTblEntry));
5953 :
5954 2330958 : if (get_relation_stats_hook &&
5955 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
5956 : {
5957 : /*
5958 : * The hook took control of acquiring a stats tuple. If it did supply
5959 : * a tuple, it'd better have supplied a freefunc.
5960 : */
5961 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
5962 0 : !vardata->freefunc)
5963 0 : elog(ERROR, "no function provided to release variable stats with");
5964 : }
5965 2330958 : else if (rte->rtekind == RTE_RELATION)
5966 : {
5967 : /*
5968 : * Plain table or parent of an inheritance appendrel, so look up the
5969 : * column in pg_statistic
5970 : */
5971 2211768 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
5972 : ObjectIdGetDatum(rte->relid),
5973 2211768 : Int16GetDatum(var->varattno),
5974 2211768 : BoolGetDatum(rte->inh));
5975 2211768 : vardata->freefunc = ReleaseSysCache;
5976 :
5977 2211768 : if (HeapTupleIsValid(vardata->statsTuple))
5978 : {
5979 : /*
5980 : * Test if user has permission to read all rows from this column.
5981 : *
5982 : * This requires that the user has the appropriate SELECT
5983 : * privileges and that there are no securityQuals from security
5984 : * barrier views or RLS policies. If that's not the case, then we
5985 : * only permit leakproof functions to be passed pg_statistic data
5986 : * in vardata, otherwise the functions might reveal data that the
5987 : * user doesn't have permission to see --- see
5988 : * statistic_proc_security_check().
5989 : */
5990 1637188 : vardata->acl_ok =
5991 1637188 : all_rows_selectable(root, var->varno,
5992 1637188 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
5993 : }
5994 : else
5995 : {
5996 : /* suppress any possible leakproofness checks later */
5997 574580 : vardata->acl_ok = true;
5998 : }
5999 : }
6000 119190 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6001 108960 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6002 : {
6003 : /*
6004 : * Plain subquery (not one that was converted to an appendrel) or
6005 : * non-recursive CTE. In either case, we can try to find out what the
6006 : * Var refers to within the subquery. We skip this for appendrel and
6007 : * recursive-CTE cases because any column stats we did find would
6008 : * likely not be very relevant.
6009 : */
6010 : PlannerInfo *subroot;
6011 : Query *subquery;
6012 : List *subtlist;
6013 : TargetEntry *ste;
6014 :
6015 : /*
6016 : * Punt if it's a whole-row var rather than a plain column reference.
6017 : */
6018 17298 : if (var->varattno == InvalidAttrNumber)
6019 0 : return;
6020 :
6021 : /*
6022 : * Otherwise, find the subquery's planner subroot.
6023 : */
6024 17298 : if (rte->rtekind == RTE_SUBQUERY)
6025 : {
6026 : RelOptInfo *rel;
6027 :
6028 : /*
6029 : * Fetch RelOptInfo for subquery. Note that we don't change the
6030 : * rel returned in vardata, since caller expects it to be a rel of
6031 : * the caller's query level. Because we might already be
6032 : * recursing, we can't use that rel pointer either, but have to
6033 : * look up the Var's rel afresh.
6034 : */
6035 10230 : rel = find_base_rel(root, var->varno);
6036 :
6037 10230 : subroot = rel->subroot;
6038 : }
6039 : else
6040 : {
6041 : /* CTE case is more difficult */
6042 : PlannerInfo *cteroot;
6043 : Index levelsup;
6044 : int ndx;
6045 : int plan_id;
6046 : ListCell *lc;
6047 :
6048 : /*
6049 : * Find the referenced CTE, and locate the subroot previously made
6050 : * for it.
6051 : */
6052 7068 : levelsup = rte->ctelevelsup;
6053 7068 : cteroot = root;
6054 13322 : while (levelsup-- > 0)
6055 : {
6056 6254 : cteroot = cteroot->parent_root;
6057 6254 : if (!cteroot) /* shouldn't happen */
6058 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6059 : }
6060 :
6061 : /*
6062 : * Note: cte_plan_ids can be shorter than cteList, if we are still
6063 : * working on planning the CTEs (ie, this is a side-reference from
6064 : * another CTE). So we mustn't use forboth here.
6065 : */
6066 7068 : ndx = 0;
6067 9290 : foreach(lc, cteroot->parse->cteList)
6068 : {
6069 9290 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6070 :
6071 9290 : if (strcmp(cte->ctename, rte->ctename) == 0)
6072 7068 : break;
6073 2222 : ndx++;
6074 : }
6075 7068 : if (lc == NULL) /* shouldn't happen */
6076 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6077 7068 : if (ndx >= list_length(cteroot->cte_plan_ids))
6078 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6079 7068 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6080 7068 : if (plan_id <= 0)
6081 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
6082 7068 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6083 : }
6084 :
6085 : /* If the subquery hasn't been planned yet, we have to punt */
6086 17298 : if (subroot == NULL)
6087 0 : return;
6088 : Assert(IsA(subroot, PlannerInfo));
6089 :
6090 : /*
6091 : * We must use the subquery parsetree as mangled by the planner, not
6092 : * the raw version from the RTE, because we need a Var that will refer
6093 : * to the subroot's live RelOptInfos. For instance, if any subquery
6094 : * pullup happened during planning, Vars in the targetlist might have
6095 : * gotten replaced, and we need to see the replacement expressions.
6096 : */
6097 17298 : subquery = subroot->parse;
6098 : Assert(IsA(subquery, Query));
6099 :
6100 : /*
6101 : * Punt if subquery uses set operations or grouping sets, as these
6102 : * will mash underlying columns' stats beyond recognition. (Set ops
6103 : * are particularly nasty; if we forged ahead, we would return stats
6104 : * relevant to only the leftmost subselect...) DISTINCT is also
6105 : * problematic, but we check that later because there is a possibility
6106 : * of learning something even with it.
6107 : */
6108 17298 : if (subquery->setOperations ||
6109 15076 : subquery->groupingSets)
6110 2246 : return;
6111 :
6112 : /* Get the subquery output expression referenced by the upper Var */
6113 15052 : if (subquery->returningList)
6114 206 : subtlist = subquery->returningList;
6115 : else
6116 14846 : subtlist = subquery->targetList;
6117 15052 : ste = get_tle_by_resno(subtlist, var->varattno);
6118 15052 : if (ste == NULL || ste->resjunk)
6119 0 : elog(ERROR, "subquery %s does not have attribute %d",
6120 : rte->eref->aliasname, var->varattno);
6121 15052 : var = (Var *) ste->expr;
6122 :
6123 : /*
6124 : * If subquery uses DISTINCT, we can't make use of any stats for the
6125 : * variable ... but, if it's the only DISTINCT column, we are entitled
6126 : * to consider it unique. We do the test this way so that it works
6127 : * for cases involving DISTINCT ON.
6128 : */
6129 15052 : if (subquery->distinctClause)
6130 : {
6131 1838 : if (list_length(subquery->distinctClause) == 1 &&
6132 616 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6133 308 : vardata->isunique = true;
6134 : /* cannot go further */
6135 1222 : return;
6136 : }
6137 :
6138 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
6139 13830 : if (subquery->groupClause)
6140 : {
6141 1080 : if (list_length(subquery->groupClause) == 1 &&
6142 450 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6143 338 : vardata->isunique = true;
6144 : /* cannot go further */
6145 630 : return;
6146 : }
6147 :
6148 : /*
6149 : * If the sub-query originated from a view with the security_barrier
6150 : * attribute, we must not look at the variable's statistics, though it
6151 : * seems all right to notice the existence of a DISTINCT clause. So
6152 : * stop here.
6153 : *
6154 : * This is probably a harsher restriction than necessary; it's
6155 : * certainly OK for the selectivity estimator (which is a C function,
6156 : * and therefore omnipotent anyway) to look at the statistics. But
6157 : * many selectivity estimators will happily *invoke the operator
6158 : * function* to try to work out a good estimate - and that's not OK.
6159 : * So for now, don't dig down for stats.
6160 : */
6161 13200 : if (rte->security_barrier)
6162 1362 : return;
6163 :
6164 : /* Can only handle a simple Var of subquery's query level */
6165 11838 : if (var && IsA(var, Var) &&
6166 6240 : var->varlevelsup == 0)
6167 : {
6168 : /*
6169 : * OK, recurse into the subquery. Note that the original setting
6170 : * of vardata->isunique (which will surely be false) is left
6171 : * unchanged in this situation. That's what we want, since even
6172 : * if the underlying column is unique, the subquery may have
6173 : * joined to other tables in a way that creates duplicates.
6174 : */
6175 6240 : examine_simple_variable(subroot, var, vardata);
6176 : }
6177 : }
6178 : else
6179 : {
6180 : /*
6181 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6182 : * see RTE_JOIN here because join alias Vars have already been
6183 : * flattened.) There's not much we can do with function outputs, but
6184 : * maybe someday try to be smarter about VALUES.
6185 : */
6186 : }
6187 : }
6188 :
6189 : /*
6190 : * all_rows_selectable
6191 : * Test whether the user has permission to select all rows from a given
6192 : * relation.
6193 : *
6194 : * Inputs:
6195 : * root: the planner info
6196 : * varno: the index of the relation (assumed to be an RTE_RELATION)
6197 : * varattnos: the attributes for which permission is required, or NULL if
6198 : * whole-table access is required
6199 : *
6200 : * Returns true if the user has the required select permissions, and there are
6201 : * no securityQuals from security barrier views or RLS policies.
6202 : *
6203 : * Note that if the relation is an inheritance child relation, securityQuals
6204 : * and access permissions are checked against the inheritance root parent (the
6205 : * relation actually mentioned in the query) --- see the comments in
6206 : * expand_single_inheritance_child() for an explanation of why it has to be
6207 : * done this way.
6208 : *
6209 : * If varattnos is non-NULL, its attribute numbers should be offset by
6210 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6211 : * checked. If varattnos is NULL, only table-level SELECT privileges are
6212 : * checked, not any column-level privileges.
6213 : *
6214 : * Note: if the relation is accessed via a view, this function actually tests
6215 : * whether the view owner has permission to select from the relation. To
6216 : * ensure that the current user has permission, it is also necessary to check
6217 : * that the current user has permission to select from the view, which we do
6218 : * at planner-startup --- see subquery_planner().
6219 : *
6220 : * This is exported so that other estimation functions can use it.
6221 : */
6222 : bool
6223 1642048 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6224 : {
6225 1642048 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6226 1642048 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6227 : Oid userid;
6228 : int varattno;
6229 :
6230 : Assert(rte->rtekind == RTE_RELATION);
6231 :
6232 : /*
6233 : * Determine the user ID to use for privilege checks (either the current
6234 : * user or the view owner, if we're accessing the table via a view).
6235 : *
6236 : * Normally the relation will have an associated RelOptInfo from which we
6237 : * can find the userid, but it might not if it's a RETURNING Var for an
6238 : * INSERT target relation. In that case use the RTEPermissionInfo
6239 : * associated with the RTE.
6240 : *
6241 : * If we navigate up to a parent relation, we keep using the same userid,
6242 : * since it's the same in all relations of a given inheritance tree.
6243 : */
6244 1642048 : if (rel)
6245 1642006 : userid = rel->userid;
6246 : else
6247 : {
6248 : RTEPermissionInfo *perminfo;
6249 :
6250 42 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6251 42 : userid = perminfo->checkAsUser;
6252 : }
6253 1642048 : if (!OidIsValid(userid))
6254 1467124 : userid = GetUserId();
6255 :
6256 : /*
6257 : * Permissions and securityQuals must be checked on the table actually
6258 : * mentioned in the query, so if this is an inheritance child, navigate up
6259 : * to the inheritance root parent. If the user can read the whole table
6260 : * or the required columns there, then they can read from the child table
6261 : * too. For per-column checks, we must find out which of the root
6262 : * parent's attributes the child relation's attributes correspond to.
6263 : */
6264 1642048 : if (root->append_rel_array != NULL)
6265 : {
6266 : AppendRelInfo *appinfo;
6267 :
6268 232650 : appinfo = root->append_rel_array[varno];
6269 :
6270 : /*
6271 : * Partitions are mapped to their immediate parent, not the root
6272 : * parent, so must be ready to walk up multiple AppendRelInfos. But
6273 : * stop if we hit a parent that is not RTE_RELATION --- that's a
6274 : * flattened UNION ALL subquery, not an inheritance parent.
6275 : */
6276 434430 : while (appinfo &&
6277 202152 : planner_rt_fetch(appinfo->parent_relid,
6278 202152 : root)->rtekind == RTE_RELATION)
6279 : {
6280 201780 : Bitmapset *parent_varattnos = NULL;
6281 :
6282 : /*
6283 : * For each child attribute, find the corresponding parent
6284 : * attribute. In rare cases, the attribute may be local to the
6285 : * child table, in which case, we've got to live with having no
6286 : * access to this column.
6287 : */
6288 201780 : varattno = -1;
6289 400710 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6290 : {
6291 : AttrNumber attno;
6292 : AttrNumber parent_attno;
6293 :
6294 198930 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6295 :
6296 198930 : if (attno == InvalidAttrNumber)
6297 : {
6298 : /*
6299 : * Whole-row reference, so must map each column of the
6300 : * child to the parent table.
6301 : */
6302 36 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6303 : {
6304 24 : parent_attno = appinfo->parent_colnos[attno - 1];
6305 24 : if (parent_attno == 0)
6306 0 : return false; /* attr is local to child */
6307 : parent_varattnos =
6308 24 : bms_add_member(parent_varattnos,
6309 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6310 : }
6311 : }
6312 : else
6313 : {
6314 198918 : if (attno < 0)
6315 : {
6316 : /* System attnos are the same in all tables */
6317 0 : parent_attno = attno;
6318 : }
6319 : else
6320 : {
6321 198918 : if (attno > appinfo->num_child_cols)
6322 0 : return false; /* safety check */
6323 198918 : parent_attno = appinfo->parent_colnos[attno - 1];
6324 198918 : if (parent_attno == 0)
6325 0 : return false; /* attr is local to child */
6326 : }
6327 : parent_varattnos =
6328 198918 : bms_add_member(parent_varattnos,
6329 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6330 : }
6331 : }
6332 :
6333 : /* If the parent is itself a child, continue up */
6334 201780 : varno = appinfo->parent_relid;
6335 201780 : varattnos = parent_varattnos;
6336 201780 : appinfo = root->append_rel_array[varno];
6337 : }
6338 :
6339 : /* Perform the access check on this parent rel */
6340 232650 : rte = planner_rt_fetch(varno, root);
6341 : Assert(rte->rtekind == RTE_RELATION);
6342 : }
6343 :
6344 : /*
6345 : * For all rows to be accessible, there must be no securityQuals from
6346 : * security barrier views or RLS policies.
6347 : */
6348 1642048 : if (rte->securityQuals != NIL)
6349 828 : return false;
6350 :
6351 : /*
6352 : * Test for table-level SELECT privilege.
6353 : *
6354 : * If varattnos is non-NULL, this is sufficient to give access to all
6355 : * requested attributes, even for a child table, since we have verified
6356 : * that all required child columns have matching parent columns.
6357 : *
6358 : * If varattnos is NULL (whole-table access requested), this doesn't
6359 : * necessarily guarantee that the user can read all columns of a child
6360 : * table, but we allow it anyway (see comments in examine_variable()) and
6361 : * don't bother checking any column privileges.
6362 : */
6363 1641220 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6364 1640768 : return true;
6365 :
6366 452 : if (varattnos == NULL)
6367 12 : return false; /* whole-table access requested */
6368 :
6369 : /*
6370 : * Don't have table-level SELECT privilege, so check per-column
6371 : * privileges.
6372 : */
6373 440 : varattno = -1;
6374 646 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6375 : {
6376 440 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6377 :
6378 440 : if (attno == InvalidAttrNumber)
6379 : {
6380 : /* Whole-row reference, so must have access to all columns */
6381 6 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6382 : ACLMASK_ALL) != ACLCHECK_OK)
6383 6 : return false;
6384 : }
6385 : else
6386 : {
6387 434 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6388 : ACL_SELECT) != ACLCHECK_OK)
6389 228 : return false;
6390 : }
6391 : }
6392 :
6393 : /* If we reach here, have all required column privileges */
6394 206 : return true;
6395 : }
6396 :
6397 : /*
6398 : * examine_indexcol_variable
6399 : * Try to look up statistical data about an index column/expression.
6400 : * Fill in a VariableStatData struct to describe the column.
6401 : *
6402 : * Inputs:
6403 : * root: the planner info
6404 : * index: the index whose column we're interested in
6405 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6406 : *
6407 : * Outputs: *vardata is filled as follows:
6408 : * var: the input expression (with any binary relabeling stripped, if
6409 : * it is or contains a variable; but otherwise the type is preserved)
6410 : * rel: RelOptInfo for table relation containing variable.
6411 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6412 : * otherwise NULL.
6413 : * freefunc: pointer to a function to release statsTuple with.
6414 : *
6415 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6416 : */
6417 : static void
6418 786624 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6419 : int indexcol, VariableStatData *vardata)
6420 : {
6421 : AttrNumber colnum;
6422 : Oid relid;
6423 :
6424 786624 : if (index->indexkeys[indexcol] != 0)
6425 : {
6426 : /* Simple variable --- look to stats for the underlying table */
6427 784434 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6428 :
6429 : Assert(rte->rtekind == RTE_RELATION);
6430 784434 : relid = rte->relid;
6431 : Assert(relid != InvalidOid);
6432 784434 : colnum = index->indexkeys[indexcol];
6433 784434 : vardata->rel = index->rel;
6434 :
6435 784434 : if (get_relation_stats_hook &&
6436 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6437 : {
6438 : /*
6439 : * The hook took control of acquiring a stats tuple. If it did
6440 : * supply a tuple, it'd better have supplied a freefunc.
6441 : */
6442 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6443 0 : !vardata->freefunc)
6444 0 : elog(ERROR, "no function provided to release variable stats with");
6445 : }
6446 : else
6447 : {
6448 784434 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6449 : ObjectIdGetDatum(relid),
6450 : Int16GetDatum(colnum),
6451 784434 : BoolGetDatum(rte->inh));
6452 784434 : vardata->freefunc = ReleaseSysCache;
6453 : }
6454 : }
6455 : else
6456 : {
6457 : /* Expression --- maybe there are stats for the index itself */
6458 2190 : relid = index->indexoid;
6459 2190 : colnum = indexcol + 1;
6460 :
6461 2190 : if (get_index_stats_hook &&
6462 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6463 : {
6464 : /*
6465 : * The hook took control of acquiring a stats tuple. If it did
6466 : * supply a tuple, it'd better have supplied a freefunc.
6467 : */
6468 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6469 0 : !vardata->freefunc)
6470 0 : elog(ERROR, "no function provided to release variable stats with");
6471 : }
6472 : else
6473 : {
6474 2190 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6475 : ObjectIdGetDatum(relid),
6476 : Int16GetDatum(colnum),
6477 : BoolGetDatum(false));
6478 2190 : vardata->freefunc = ReleaseSysCache;
6479 : }
6480 : }
6481 786624 : }
6482 :
6483 : /*
6484 : * Check whether it is permitted to call func_oid passing some of the
6485 : * pg_statistic data in vardata. We allow this if either of the following
6486 : * conditions is met: (1) the user has SELECT privileges on the table or
6487 : * column underlying the pg_statistic data and there are no securityQuals from
6488 : * security barrier views or RLS policies, or (2) the function is marked
6489 : * leakproof.
6490 : */
6491 : bool
6492 1160164 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6493 : {
6494 1160164 : if (vardata->acl_ok)
6495 1158298 : return true; /* have SELECT privs and no securityQuals */
6496 :
6497 1866 : if (!OidIsValid(func_oid))
6498 0 : return false;
6499 :
6500 1866 : if (get_func_leakproof(func_oid))
6501 924 : return true;
6502 :
6503 942 : ereport(DEBUG2,
6504 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6505 : get_func_name(func_oid))));
6506 942 : return false;
6507 : }
6508 :
6509 : /*
6510 : * get_variable_numdistinct
6511 : * Estimate the number of distinct values of a variable.
6512 : *
6513 : * vardata: results of examine_variable
6514 : * *isdefault: set to true if the result is a default rather than based on
6515 : * anything meaningful.
6516 : *
6517 : * NB: be careful to produce a positive integral result, since callers may
6518 : * compare the result to exact integer counts, or might divide by it.
6519 : */
6520 : double
6521 1644318 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6522 : {
6523 : double stadistinct;
6524 1644318 : double stanullfrac = 0.0;
6525 : double ntuples;
6526 :
6527 1644318 : *isdefault = false;
6528 :
6529 : /*
6530 : * Determine the stadistinct value to use. There are cases where we can
6531 : * get an estimate even without a pg_statistic entry, or can get a better
6532 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6533 : * (otherwise, assume no nulls, for lack of any better idea).
6534 : */
6535 1644318 : if (HeapTupleIsValid(vardata->statsTuple))
6536 : {
6537 : /* Use the pg_statistic entry */
6538 : Form_pg_statistic stats;
6539 :
6540 1151530 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6541 1151530 : stadistinct = stats->stadistinct;
6542 1151530 : stanullfrac = stats->stanullfrac;
6543 : }
6544 492788 : else if (vardata->vartype == BOOLOID)
6545 : {
6546 : /*
6547 : * Special-case boolean columns: presumably, two distinct values.
6548 : *
6549 : * Are there any other datatypes we should wire in special estimates
6550 : * for?
6551 : */
6552 592 : stadistinct = 2.0;
6553 : }
6554 492196 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6555 : {
6556 : /*
6557 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6558 : * This could of course be very wrong, but it should tend to be true
6559 : * in well-written queries. We could consider examining the VALUES'
6560 : * contents to get some real statistics; but that only works if the
6561 : * entries are all constants, and it would be pretty expensive anyway.
6562 : */
6563 3528 : stadistinct = -1.0; /* unique (and all non null) */
6564 : }
6565 : else
6566 : {
6567 : /*
6568 : * We don't keep statistics for system columns, but in some cases we
6569 : * can infer distinctness anyway.
6570 : */
6571 488668 : if (vardata->var && IsA(vardata->var, Var))
6572 : {
6573 451378 : switch (((Var *) vardata->var)->varattno)
6574 : {
6575 1200 : case SelfItemPointerAttributeNumber:
6576 1200 : stadistinct = -1.0; /* unique (and all non null) */
6577 1200 : break;
6578 26096 : case TableOidAttributeNumber:
6579 26096 : stadistinct = 1.0; /* only 1 value */
6580 26096 : break;
6581 424082 : default:
6582 424082 : stadistinct = 0.0; /* means "unknown" */
6583 424082 : break;
6584 : }
6585 : }
6586 : else
6587 37290 : stadistinct = 0.0; /* means "unknown" */
6588 :
6589 : /*
6590 : * XXX consider using estimate_num_groups on expressions?
6591 : */
6592 : }
6593 :
6594 : /*
6595 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6596 : * variable, assume it is unique no matter what pg_statistic says; the
6597 : * statistics could be out of date, or we might have found a partial
6598 : * unique index that proves the var is unique for this query. However,
6599 : * we'd better still believe the null-fraction statistic.
6600 : */
6601 1644318 : if (vardata->isunique)
6602 408162 : stadistinct = -1.0 * (1.0 - stanullfrac);
6603 :
6604 : /*
6605 : * If we had an absolute estimate, use that.
6606 : */
6607 1644318 : if (stadistinct > 0.0)
6608 427204 : return clamp_row_est(stadistinct);
6609 :
6610 : /*
6611 : * Otherwise we need to get the relation size; punt if not available.
6612 : */
6613 1217114 : if (vardata->rel == NULL)
6614 : {
6615 416 : *isdefault = true;
6616 416 : return DEFAULT_NUM_DISTINCT;
6617 : }
6618 1216698 : ntuples = vardata->rel->tuples;
6619 1216698 : if (ntuples <= 0.0)
6620 : {
6621 117400 : *isdefault = true;
6622 117400 : return DEFAULT_NUM_DISTINCT;
6623 : }
6624 :
6625 : /*
6626 : * If we had a relative estimate, use that.
6627 : */
6628 1099298 : if (stadistinct < 0.0)
6629 796382 : return clamp_row_est(-stadistinct * ntuples);
6630 :
6631 : /*
6632 : * With no data, estimate ndistinct = ntuples if the table is small, else
6633 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6634 : * that the behavior isn't discontinuous.
6635 : */
6636 302916 : if (ntuples < DEFAULT_NUM_DISTINCT)
6637 137444 : return clamp_row_est(ntuples);
6638 :
6639 165472 : *isdefault = true;
6640 165472 : return DEFAULT_NUM_DISTINCT;
6641 : }
6642 :
6643 : /*
6644 : * get_variable_range
6645 : * Estimate the minimum and maximum value of the specified variable.
6646 : * If successful, store values in *min and *max, and return true.
6647 : * If no data available, return false.
6648 : *
6649 : * sortop is the "<" comparison operator to use. This should generally
6650 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6651 : * The collation must be specified too.
6652 : */
6653 : static bool
6654 248270 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6655 : Oid sortop, Oid collation,
6656 : Datum *min, Datum *max)
6657 : {
6658 248270 : Datum tmin = 0;
6659 248270 : Datum tmax = 0;
6660 248270 : bool have_data = false;
6661 : int16 typLen;
6662 : bool typByVal;
6663 : Oid opfuncoid;
6664 : FmgrInfo opproc;
6665 : AttStatsSlot sslot;
6666 :
6667 : /*
6668 : * XXX It's very tempting to try to use the actual column min and max, if
6669 : * we can get them relatively-cheaply with an index probe. However, since
6670 : * this function is called many times during join planning, that could
6671 : * have unpleasant effects on planning speed. Need more investigation
6672 : * before enabling this.
6673 : */
6674 : #ifdef NOT_USED
6675 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6676 : return true;
6677 : #endif
6678 :
6679 248270 : if (!HeapTupleIsValid(vardata->statsTuple))
6680 : {
6681 : /* no stats available, so default result */
6682 55498 : return false;
6683 : }
6684 :
6685 : /*
6686 : * If we can't apply the sortop to the stats data, just fail. In
6687 : * principle, if there's a histogram and no MCVs, we could return the
6688 : * histogram endpoints without ever applying the sortop ... but it's
6689 : * probably not worth trying, because whatever the caller wants to do with
6690 : * the endpoints would likely fail the security check too.
6691 : */
6692 192772 : if (!statistic_proc_security_check(vardata,
6693 192772 : (opfuncoid = get_opcode(sortop))))
6694 0 : return false;
6695 :
6696 192772 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6697 :
6698 192772 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6699 :
6700 : /*
6701 : * If there is a histogram with the ordering we want, grab the first and
6702 : * last values.
6703 : */
6704 192772 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6705 : STATISTIC_KIND_HISTOGRAM, sortop,
6706 : ATTSTATSSLOT_VALUES))
6707 : {
6708 121228 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6709 : {
6710 121228 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6711 121228 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6712 121228 : have_data = true;
6713 : }
6714 121228 : free_attstatsslot(&sslot);
6715 : }
6716 :
6717 : /*
6718 : * Otherwise, if there is a histogram with some other ordering, scan it
6719 : * and get the min and max values according to the ordering we want. This
6720 : * of course may not find values that are really extremal according to our
6721 : * ordering, but it beats ignoring available data.
6722 : */
6723 264316 : if (!have_data &&
6724 71544 : get_attstatsslot(&sslot, vardata->statsTuple,
6725 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6726 : ATTSTATSSLOT_VALUES))
6727 : {
6728 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6729 : collation, typLen, typByVal,
6730 : &tmin, &tmax, &have_data);
6731 0 : free_attstatsslot(&sslot);
6732 : }
6733 :
6734 : /*
6735 : * If we have most-common-values info, look for extreme MCVs. This is
6736 : * needed even if we also have a histogram, since the histogram excludes
6737 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6738 : * be pretty wary of deciding that that is a full representation of the
6739 : * data. Proceed only if the MCVs represent the whole table (to within
6740 : * roundoff error).
6741 : */
6742 192772 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6743 : STATISTIC_KIND_MCV, InvalidOid,
6744 192772 : have_data ? ATTSTATSSLOT_VALUES :
6745 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6746 : {
6747 108272 : bool use_mcvs = have_data;
6748 :
6749 108272 : if (!have_data)
6750 : {
6751 70140 : double sumcommon = 0.0;
6752 : double nullfrac;
6753 : int i;
6754 :
6755 526996 : for (i = 0; i < sslot.nnumbers; i++)
6756 456856 : sumcommon += sslot.numbers[i];
6757 70140 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6758 70140 : if (sumcommon + nullfrac > 0.99999)
6759 68180 : use_mcvs = true;
6760 : }
6761 :
6762 108272 : if (use_mcvs)
6763 106312 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6764 : collation, typLen, typByVal,
6765 : &tmin, &tmax, &have_data);
6766 108272 : free_attstatsslot(&sslot);
6767 : }
6768 :
6769 192772 : *min = tmin;
6770 192772 : *max = tmax;
6771 192772 : return have_data;
6772 : }
6773 :
6774 : /*
6775 : * get_stats_slot_range: scan sslot for min/max values
6776 : *
6777 : * Subroutine for get_variable_range: update min/max/have_data according
6778 : * to what we find in the statistics array.
6779 : */
6780 : static void
6781 106312 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6782 : Oid collation, int16 typLen, bool typByVal,
6783 : Datum *min, Datum *max, bool *p_have_data)
6784 : {
6785 106312 : Datum tmin = *min;
6786 106312 : Datum tmax = *max;
6787 106312 : bool have_data = *p_have_data;
6788 106312 : bool found_tmin = false;
6789 106312 : bool found_tmax = false;
6790 :
6791 : /* Look up the comparison function, if we didn't already do so */
6792 106312 : if (opproc->fn_oid != opfuncoid)
6793 106312 : fmgr_info(opfuncoid, opproc);
6794 :
6795 : /* Scan all the slot's values */
6796 2568104 : for (int i = 0; i < sslot->nvalues; i++)
6797 : {
6798 2461792 : if (!have_data)
6799 : {
6800 68180 : tmin = tmax = sslot->values[i];
6801 68180 : found_tmin = found_tmax = true;
6802 68180 : *p_have_data = have_data = true;
6803 68180 : continue;
6804 : }
6805 2393612 : if (DatumGetBool(FunctionCall2Coll(opproc,
6806 : collation,
6807 2393612 : sslot->values[i], tmin)))
6808 : {
6809 59380 : tmin = sslot->values[i];
6810 59380 : found_tmin = true;
6811 : }
6812 2393612 : if (DatumGetBool(FunctionCall2Coll(opproc,
6813 : collation,
6814 2393612 : tmax, sslot->values[i])))
6815 : {
6816 263030 : tmax = sslot->values[i];
6817 263030 : found_tmax = true;
6818 : }
6819 : }
6820 :
6821 : /*
6822 : * Copy the slot's values, if we found new extreme values.
6823 : */
6824 106312 : if (found_tmin)
6825 91028 : *min = datumCopy(tmin, typByVal, typLen);
6826 106312 : if (found_tmax)
6827 72842 : *max = datumCopy(tmax, typByVal, typLen);
6828 106312 : }
6829 :
6830 :
6831 : /*
6832 : * get_actual_variable_range
6833 : * Attempt to identify the current *actual* minimum and/or maximum
6834 : * of the specified variable, by looking for a suitable btree index
6835 : * and fetching its low and/or high values.
6836 : * If successful, store values in *min and *max, and return true.
6837 : * (Either pointer can be NULL if that endpoint isn't needed.)
6838 : * If unsuccessful, return false.
6839 : *
6840 : * sortop is the "<" comparison operator to use.
6841 : * collation is the required collation.
6842 : */
6843 : static bool
6844 187086 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6845 : Oid sortop, Oid collation,
6846 : Datum *min, Datum *max)
6847 : {
6848 187086 : bool have_data = false;
6849 187086 : RelOptInfo *rel = vardata->rel;
6850 : RangeTblEntry *rte;
6851 : ListCell *lc;
6852 :
6853 : /* No hope if no relation or it doesn't have indexes */
6854 187086 : if (rel == NULL || rel->indexlist == NIL)
6855 13710 : return false;
6856 : /* If it has indexes it must be a plain relation */
6857 173376 : rte = root->simple_rte_array[rel->relid];
6858 : Assert(rte->rtekind == RTE_RELATION);
6859 :
6860 : /* ignore partitioned tables. Any indexes here are not real indexes */
6861 173376 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6862 876 : return false;
6863 :
6864 : /* Search through the indexes to see if any match our problem */
6865 333370 : foreach(lc, rel->indexlist)
6866 : {
6867 286536 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6868 : ScanDirection indexscandir;
6869 : StrategyNumber strategy;
6870 :
6871 : /* Ignore non-ordering indexes */
6872 286536 : if (index->sortopfamily == NULL)
6873 0 : continue;
6874 :
6875 : /*
6876 : * Ignore partial indexes --- we only want stats that cover the entire
6877 : * relation.
6878 : */
6879 286536 : if (index->indpred != NIL)
6880 288 : continue;
6881 :
6882 : /*
6883 : * The index list might include hypothetical indexes inserted by a
6884 : * get_relation_info hook --- don't try to access them.
6885 : */
6886 286248 : if (index->hypothetical)
6887 0 : continue;
6888 :
6889 : /*
6890 : * get_actual_variable_endpoint uses the index-only-scan machinery, so
6891 : * ignore indexes that can't use it on their first column.
6892 : */
6893 286248 : if (!index->canreturn[0])
6894 0 : continue;
6895 :
6896 : /*
6897 : * The first index column must match the desired variable, sortop, and
6898 : * collation --- but we can use a descending-order index.
6899 : */
6900 286248 : if (collation != index->indexcollations[0])
6901 36810 : continue; /* test first 'cause it's cheapest */
6902 249438 : if (!match_index_to_operand(vardata->var, 0, index))
6903 123772 : continue;
6904 125666 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6905 125666 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6906 : {
6907 125666 : case COMPARE_LT:
6908 125666 : if (index->reverse_sort[0])
6909 0 : indexscandir = BackwardScanDirection;
6910 : else
6911 125666 : indexscandir = ForwardScanDirection;
6912 125666 : break;
6913 0 : case COMPARE_GT:
6914 0 : if (index->reverse_sort[0])
6915 0 : indexscandir = ForwardScanDirection;
6916 : else
6917 0 : indexscandir = BackwardScanDirection;
6918 0 : break;
6919 0 : default:
6920 : /* index doesn't match the sortop */
6921 0 : continue;
6922 : }
6923 :
6924 : /*
6925 : * Found a suitable index to extract data from. Set up some data that
6926 : * can be used by both invocations of get_actual_variable_endpoint.
6927 : */
6928 : {
6929 : MemoryContext tmpcontext;
6930 : MemoryContext oldcontext;
6931 : Relation heapRel;
6932 : Relation indexRel;
6933 : TupleTableSlot *slot;
6934 : int16 typLen;
6935 : bool typByVal;
6936 : ScanKeyData scankeys[1];
6937 :
6938 : /* Make sure any cruft gets recycled when we're done */
6939 125666 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
6940 : "get_actual_variable_range workspace",
6941 : ALLOCSET_DEFAULT_SIZES);
6942 125666 : oldcontext = MemoryContextSwitchTo(tmpcontext);
6943 :
6944 : /*
6945 : * Open the table and index so we can read from them. We should
6946 : * already have some type of lock on each.
6947 : */
6948 125666 : heapRel = table_open(rte->relid, NoLock);
6949 125666 : indexRel = index_open(index->indexoid, NoLock);
6950 :
6951 : /* build some stuff needed for indexscan execution */
6952 125666 : slot = table_slot_create(heapRel, NULL);
6953 125666 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6954 :
6955 : /* set up an IS NOT NULL scan key so that we ignore nulls */
6956 125666 : ScanKeyEntryInitialize(&scankeys[0],
6957 : SK_ISNULL | SK_SEARCHNOTNULL,
6958 : 1, /* index col to scan */
6959 : InvalidStrategy, /* no strategy */
6960 : InvalidOid, /* no strategy subtype */
6961 : InvalidOid, /* no collation */
6962 : InvalidOid, /* no reg proc for this */
6963 : (Datum) 0); /* constant */
6964 :
6965 : /* If min is requested ... */
6966 125666 : if (min)
6967 : {
6968 71058 : have_data = get_actual_variable_endpoint(heapRel,
6969 : indexRel,
6970 : indexscandir,
6971 : scankeys,
6972 : typLen,
6973 : typByVal,
6974 : slot,
6975 : oldcontext,
6976 : min);
6977 : }
6978 : else
6979 : {
6980 : /* If min not requested, still want to fetch max */
6981 54608 : have_data = true;
6982 : }
6983 :
6984 : /* If max is requested, and we didn't already fail ... */
6985 125666 : if (max && have_data)
6986 : {
6987 : /* scan in the opposite direction; all else is the same */
6988 56256 : have_data = get_actual_variable_endpoint(heapRel,
6989 : indexRel,
6990 56256 : -indexscandir,
6991 : scankeys,
6992 : typLen,
6993 : typByVal,
6994 : slot,
6995 : oldcontext,
6996 : max);
6997 : }
6998 :
6999 : /* Clean everything up */
7000 125666 : ExecDropSingleTupleTableSlot(slot);
7001 :
7002 125666 : index_close(indexRel, NoLock);
7003 125666 : table_close(heapRel, NoLock);
7004 :
7005 125666 : MemoryContextSwitchTo(oldcontext);
7006 125666 : MemoryContextDelete(tmpcontext);
7007 :
7008 : /* And we're done */
7009 125666 : break;
7010 : }
7011 : }
7012 :
7013 172500 : return have_data;
7014 : }
7015 :
7016 : /*
7017 : * Get one endpoint datum (min or max depending on indexscandir) from the
7018 : * specified index. Return true if successful, false if not.
7019 : * On success, endpoint value is stored to *endpointDatum (and copied into
7020 : * outercontext).
7021 : *
7022 : * scankeys is a 1-element scankey array set up to reject nulls.
7023 : * typLen/typByVal describe the datatype of the index's first column.
7024 : * tableslot is a slot suitable to hold table tuples, in case we need
7025 : * to probe the heap.
7026 : * (We could compute these values locally, but that would mean computing them
7027 : * twice when get_actual_variable_range needs both the min and the max.)
7028 : *
7029 : * Failure occurs either when the index is empty, or we decide that it's
7030 : * taking too long to find a suitable tuple.
7031 : */
7032 : static bool
7033 127314 : get_actual_variable_endpoint(Relation heapRel,
7034 : Relation indexRel,
7035 : ScanDirection indexscandir,
7036 : ScanKey scankeys,
7037 : int16 typLen,
7038 : bool typByVal,
7039 : TupleTableSlot *tableslot,
7040 : MemoryContext outercontext,
7041 : Datum *endpointDatum)
7042 : {
7043 127314 : bool have_data = false;
7044 : SnapshotData SnapshotNonVacuumable;
7045 : IndexScanDesc index_scan;
7046 127314 : Buffer vmbuffer = InvalidBuffer;
7047 127314 : BlockNumber last_heap_block = InvalidBlockNumber;
7048 127314 : int n_visited_heap_pages = 0;
7049 : ItemPointer tid;
7050 : Datum values[INDEX_MAX_KEYS];
7051 : bool isnull[INDEX_MAX_KEYS];
7052 : MemoryContext oldcontext;
7053 :
7054 : /*
7055 : * We use the index-only-scan machinery for this. With mostly-static
7056 : * tables that's a win because it avoids a heap visit. It's also a win
7057 : * for dynamic data, but the reason is less obvious; read on for details.
7058 : *
7059 : * In principle, we should scan the index with our current active
7060 : * snapshot, which is the best approximation we've got to what the query
7061 : * will see when executed. But that won't be exact if a new snap is taken
7062 : * before running the query, and it can be very expensive if a lot of
7063 : * recently-dead or uncommitted rows exist at the beginning or end of the
7064 : * index (because we'll laboriously fetch each one and reject it).
7065 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7066 : * and uncommitted rows as well as normal visible rows. On the other
7067 : * hand, it will reject known-dead rows, and thus not give a bogus answer
7068 : * when the extreme value has been deleted (unless the deletion was quite
7069 : * recent); that case motivates not using SnapshotAny here.
7070 : *
7071 : * A crucial point here is that SnapshotNonVacuumable, with
7072 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7073 : * condition that the indexscan will use to decide that index entries are
7074 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7075 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7076 : * have to continue scanning past it, we know that the indexscan will mark
7077 : * that index entry killed. That means that the next
7078 : * get_actual_variable_endpoint() call will not have to re-consider that
7079 : * index entry. In this way we avoid repetitive work when this function
7080 : * is used a lot during planning.
7081 : *
7082 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7083 : * recently-created index, some index entries may point at "broken" HOT
7084 : * chains in which not all the tuple versions contain data matching the
7085 : * index entry. The live tuple version(s) certainly do match the index,
7086 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7087 : * don't match. Hence, if we took data from the selected heap tuple, we
7088 : * might get a bogus answer that's not close to the index extremal value,
7089 : * or could even be NULL. We avoid this hazard because we take the data
7090 : * from the index entry not the heap.
7091 : *
7092 : * Despite all this care, there are situations where we might find many
7093 : * non-visible tuples near the end of the index. We don't want to expend
7094 : * a huge amount of time here, so we give up once we've read too many heap
7095 : * pages. When we fail for that reason, the caller will end up using
7096 : * whatever extremal value is recorded in pg_statistic.
7097 : */
7098 127314 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7099 : GlobalVisTestFor(heapRel));
7100 :
7101 127314 : index_scan = index_beginscan(heapRel, indexRel,
7102 : &SnapshotNonVacuumable, NULL,
7103 : 1, 0);
7104 : /* Set it up for index-only scan */
7105 127314 : index_scan->xs_want_itup = true;
7106 127314 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7107 :
7108 : /* Fetch first/next tuple in specified direction */
7109 164714 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7110 : {
7111 164714 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7112 :
7113 164714 : if (!VM_ALL_VISIBLE(heapRel,
7114 : block,
7115 : &vmbuffer))
7116 : {
7117 : /* Rats, we have to visit the heap to check visibility */
7118 119280 : if (!index_fetch_heap(index_scan, tableslot))
7119 : {
7120 : /*
7121 : * No visible tuple for this index entry, so we need to
7122 : * advance to the next entry. Before doing so, count heap
7123 : * page fetches and give up if we've done too many.
7124 : *
7125 : * We don't charge a page fetch if this is the same heap page
7126 : * as the previous tuple. This is on the conservative side,
7127 : * since other recently-accessed pages are probably still in
7128 : * buffers too; but it's good enough for this heuristic.
7129 : */
7130 : #define VISITED_PAGES_LIMIT 100
7131 :
7132 37400 : if (block != last_heap_block)
7133 : {
7134 3674 : last_heap_block = block;
7135 3674 : n_visited_heap_pages++;
7136 3674 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
7137 0 : break;
7138 : }
7139 :
7140 37400 : continue; /* no visible tuple, try next index entry */
7141 : }
7142 :
7143 : /* We don't actually need the heap tuple for anything */
7144 81880 : ExecClearTuple(tableslot);
7145 :
7146 : /*
7147 : * We don't care whether there's more than one visible tuple in
7148 : * the HOT chain; if any are visible, that's good enough.
7149 : */
7150 : }
7151 :
7152 : /*
7153 : * We expect that the index will return data in IndexTuple not
7154 : * HeapTuple format.
7155 : */
7156 127314 : if (!index_scan->xs_itup)
7157 0 : elog(ERROR, "no data returned for index-only scan");
7158 :
7159 : /*
7160 : * We do not yet support recheck here.
7161 : */
7162 127314 : if (index_scan->xs_recheck)
7163 0 : break;
7164 :
7165 : /* OK to deconstruct the index tuple */
7166 127314 : index_deform_tuple(index_scan->xs_itup,
7167 : index_scan->xs_itupdesc,
7168 : values, isnull);
7169 :
7170 : /* Shouldn't have got a null, but be careful */
7171 127314 : if (isnull[0])
7172 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7173 : RelationGetRelationName(indexRel));
7174 :
7175 : /* Copy the index column value out to caller's context */
7176 127314 : oldcontext = MemoryContextSwitchTo(outercontext);
7177 127314 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7178 127314 : MemoryContextSwitchTo(oldcontext);
7179 127314 : have_data = true;
7180 127314 : break;
7181 : }
7182 :
7183 127314 : if (vmbuffer != InvalidBuffer)
7184 114270 : ReleaseBuffer(vmbuffer);
7185 127314 : index_endscan(index_scan);
7186 :
7187 127314 : return have_data;
7188 : }
7189 :
7190 : /*
7191 : * find_join_input_rel
7192 : * Look up the input relation for a join.
7193 : *
7194 : * We assume that the input relation's RelOptInfo must have been constructed
7195 : * already.
7196 : */
7197 : static RelOptInfo *
7198 10714 : find_join_input_rel(PlannerInfo *root, Relids relids)
7199 : {
7200 10714 : RelOptInfo *rel = NULL;
7201 :
7202 10714 : if (!bms_is_empty(relids))
7203 : {
7204 : int relid;
7205 :
7206 10714 : if (bms_get_singleton_member(relids, &relid))
7207 10398 : rel = find_base_rel(root, relid);
7208 : else
7209 316 : rel = find_join_rel(root, relids);
7210 : }
7211 :
7212 10714 : if (rel == NULL)
7213 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7214 :
7215 10714 : return rel;
7216 : }
7217 :
7218 :
7219 : /*-------------------------------------------------------------------------
7220 : *
7221 : * Index cost estimation functions
7222 : *
7223 : *-------------------------------------------------------------------------
7224 : */
7225 :
7226 : /*
7227 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7228 : */
7229 : List *
7230 803878 : get_quals_from_indexclauses(List *indexclauses)
7231 : {
7232 803878 : List *result = NIL;
7233 : ListCell *lc;
7234 :
7235 1414280 : foreach(lc, indexclauses)
7236 : {
7237 610402 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7238 : ListCell *lc2;
7239 :
7240 1223720 : foreach(lc2, iclause->indexquals)
7241 : {
7242 613318 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7243 :
7244 613318 : result = lappend(result, rinfo);
7245 : }
7246 : }
7247 803878 : return result;
7248 : }
7249 :
7250 : /*
7251 : * Compute the total evaluation cost of the comparison operands in a list
7252 : * of index qual expressions. Since we know these will be evaluated just
7253 : * once per scan, there's no need to distinguish startup from per-row cost.
7254 : *
7255 : * This can be used either on the result of get_quals_from_indexclauses(),
7256 : * or directly on an indexorderbys list. In both cases, we expect that the
7257 : * index key expression is on the left side of binary clauses.
7258 : */
7259 : Cost
7260 1594750 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7261 : {
7262 1594750 : Cost qual_arg_cost = 0;
7263 : ListCell *lc;
7264 :
7265 2208530 : foreach(lc, indexquals)
7266 : {
7267 613780 : Expr *clause = (Expr *) lfirst(lc);
7268 : Node *other_operand;
7269 : QualCost index_qual_cost;
7270 :
7271 : /*
7272 : * Index quals will have RestrictInfos, indexorderbys won't. Look
7273 : * through RestrictInfo if present.
7274 : */
7275 613780 : if (IsA(clause, RestrictInfo))
7276 613306 : clause = ((RestrictInfo *) clause)->clause;
7277 :
7278 613780 : if (IsA(clause, OpExpr))
7279 : {
7280 599112 : OpExpr *op = (OpExpr *) clause;
7281 :
7282 599112 : other_operand = (Node *) lsecond(op->args);
7283 : }
7284 14668 : else if (IsA(clause, RowCompareExpr))
7285 : {
7286 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7287 :
7288 396 : other_operand = (Node *) rc->rargs;
7289 : }
7290 14272 : else if (IsA(clause, ScalarArrayOpExpr))
7291 : {
7292 11346 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7293 :
7294 11346 : other_operand = (Node *) lsecond(saop->args);
7295 : }
7296 2926 : else if (IsA(clause, NullTest))
7297 : {
7298 2926 : other_operand = NULL;
7299 : }
7300 : else
7301 : {
7302 0 : elog(ERROR, "unsupported indexqual type: %d",
7303 : (int) nodeTag(clause));
7304 : other_operand = NULL; /* keep compiler quiet */
7305 : }
7306 :
7307 613780 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7308 613780 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7309 : }
7310 1594750 : return qual_arg_cost;
7311 : }
7312 :
7313 : void
7314 790884 : genericcostestimate(PlannerInfo *root,
7315 : IndexPath *path,
7316 : double loop_count,
7317 : GenericCosts *costs)
7318 : {
7319 790884 : IndexOptInfo *index = path->indexinfo;
7320 790884 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7321 790884 : List *indexOrderBys = path->indexorderbys;
7322 : Cost indexStartupCost;
7323 : Cost indexTotalCost;
7324 : Selectivity indexSelectivity;
7325 : double indexCorrelation;
7326 : double numIndexPages;
7327 : double numIndexTuples;
7328 : double spc_random_page_cost;
7329 : double num_sa_scans;
7330 : double num_outer_scans;
7331 : double num_scans;
7332 : double qual_op_cost;
7333 : double qual_arg_cost;
7334 : List *selectivityQuals;
7335 : ListCell *l;
7336 :
7337 : /*
7338 : * If the index is partial, AND the index predicate with the explicitly
7339 : * given indexquals to produce a more accurate idea of the index
7340 : * selectivity.
7341 : */
7342 790884 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7343 :
7344 : /*
7345 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7346 : * just assume that the number of index descents is the number of distinct
7347 : * combinations of array elements from all of the scan's SAOP clauses.
7348 : */
7349 790884 : num_sa_scans = costs->num_sa_scans;
7350 790884 : if (num_sa_scans < 1)
7351 : {
7352 7890 : num_sa_scans = 1;
7353 16562 : foreach(l, indexQuals)
7354 : {
7355 8672 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7356 :
7357 8672 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7358 : {
7359 26 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7360 26 : double alength = estimate_array_length(root, lsecond(saop->args));
7361 :
7362 26 : if (alength > 1)
7363 26 : num_sa_scans *= alength;
7364 : }
7365 : }
7366 : }
7367 :
7368 : /* Estimate the fraction of main-table tuples that will be visited */
7369 790884 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7370 790884 : index->rel->relid,
7371 : JOIN_INNER,
7372 : NULL);
7373 :
7374 : /*
7375 : * If caller didn't give us an estimate, estimate the number of index
7376 : * tuples that will be visited. We do it in this rather peculiar-looking
7377 : * way in order to get the right answer for partial indexes.
7378 : */
7379 790884 : numIndexTuples = costs->numIndexTuples;
7380 790884 : if (numIndexTuples <= 0.0)
7381 : {
7382 91564 : numIndexTuples = indexSelectivity * index->rel->tuples;
7383 :
7384 : /*
7385 : * The above calculation counts all the tuples visited across all
7386 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7387 : * average per-indexscan number, so adjust. This is a handy place to
7388 : * round to integer, too. (If caller supplied tuple estimate, it's
7389 : * responsible for handling these considerations.)
7390 : */
7391 91564 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7392 : }
7393 :
7394 : /*
7395 : * We can bound the number of tuples by the index size in any case. Also,
7396 : * always estimate at least one tuple is touched, even when
7397 : * indexSelectivity estimate is tiny.
7398 : */
7399 790884 : if (numIndexTuples > index->tuples)
7400 6536 : numIndexTuples = index->tuples;
7401 790884 : if (numIndexTuples < 1.0)
7402 92258 : numIndexTuples = 1.0;
7403 :
7404 : /*
7405 : * Estimate the number of index pages that will be retrieved.
7406 : *
7407 : * We use the simplistic method of taking a pro-rata fraction of the total
7408 : * number of index pages. In effect, this counts only leaf pages and not
7409 : * any overhead such as index metapage or upper tree levels.
7410 : *
7411 : * In practice access to upper index levels is often nearly free because
7412 : * those tend to stay in cache under load; moreover, the cost involved is
7413 : * highly dependent on index type. We therefore ignore such costs here
7414 : * and leave it to the caller to add a suitable charge if needed.
7415 : */
7416 790884 : if (index->pages > 1 && index->tuples > 1)
7417 726726 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
7418 : else
7419 64158 : numIndexPages = 1.0;
7420 :
7421 : /* fetch estimated page cost for tablespace containing index */
7422 790884 : get_tablespace_page_costs(index->reltablespace,
7423 : &spc_random_page_cost,
7424 : NULL);
7425 :
7426 : /*
7427 : * Now compute the disk access costs.
7428 : *
7429 : * The above calculations are all per-index-scan. However, if we are in a
7430 : * nestloop inner scan, we can expect the scan to be repeated (with
7431 : * different search keys) for each row of the outer relation. Likewise,
7432 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7433 : * the potential for cache effects to reduce the number of disk page
7434 : * fetches needed. We want to estimate the average per-scan I/O cost in
7435 : * the presence of caching.
7436 : *
7437 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7438 : * estimate the total number of page fetches that occur. While this
7439 : * wasn't what it was designed for, it seems a reasonable model anyway.
7440 : * Note that we are counting pages not tuples anymore, so we take N = T =
7441 : * index size, as if there were one "tuple" per page.
7442 : */
7443 790884 : num_outer_scans = loop_count;
7444 790884 : num_scans = num_sa_scans * num_outer_scans;
7445 :
7446 790884 : if (num_scans > 1)
7447 : {
7448 : double pages_fetched;
7449 :
7450 : /* total page fetches ignoring cache effects */
7451 94566 : pages_fetched = numIndexPages * num_scans;
7452 :
7453 : /* use Mackert and Lohman formula to adjust for cache effects */
7454 94566 : pages_fetched = index_pages_fetched(pages_fetched,
7455 : index->pages,
7456 94566 : (double) index->pages,
7457 : root);
7458 :
7459 : /*
7460 : * Now compute the total disk access cost, and then report a pro-rated
7461 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7462 : * since that's internal to the indexscan.)
7463 : */
7464 94566 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7465 : / num_outer_scans;
7466 : }
7467 : else
7468 : {
7469 : /*
7470 : * For a single index scan, we just charge spc_random_page_cost per
7471 : * page touched.
7472 : */
7473 696318 : indexTotalCost = numIndexPages * spc_random_page_cost;
7474 : }
7475 :
7476 : /*
7477 : * CPU cost: any complex expressions in the indexquals will need to be
7478 : * evaluated once at the start of the scan to reduce them to runtime keys
7479 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7480 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7481 : * indexqual operator. Because we have numIndexTuples as a per-scan
7482 : * number, we have to multiply by num_sa_scans to get the correct result
7483 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7484 : * ORDER BY expressions.
7485 : *
7486 : * Note: this neglects the possible costs of rechecking lossy operators.
7487 : * Detecting that that might be needed seems more expensive than it's
7488 : * worth, though, considering all the other inaccuracies here ...
7489 : */
7490 790884 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7491 790884 : index_other_operands_eval_cost(root, indexOrderBys);
7492 790884 : qual_op_cost = cpu_operator_cost *
7493 790884 : (list_length(indexQuals) + list_length(indexOrderBys));
7494 :
7495 790884 : indexStartupCost = qual_arg_cost;
7496 790884 : indexTotalCost += qual_arg_cost;
7497 790884 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7498 :
7499 : /*
7500 : * Generic assumption about index correlation: there isn't any.
7501 : */
7502 790884 : indexCorrelation = 0.0;
7503 :
7504 : /*
7505 : * Return everything to caller.
7506 : */
7507 790884 : costs->indexStartupCost = indexStartupCost;
7508 790884 : costs->indexTotalCost = indexTotalCost;
7509 790884 : costs->indexSelectivity = indexSelectivity;
7510 790884 : costs->indexCorrelation = indexCorrelation;
7511 790884 : costs->numIndexPages = numIndexPages;
7512 790884 : costs->numIndexTuples = numIndexTuples;
7513 790884 : costs->spc_random_page_cost = spc_random_page_cost;
7514 790884 : costs->num_sa_scans = num_sa_scans;
7515 790884 : }
7516 :
7517 : /*
7518 : * If the index is partial, add its predicate to the given qual list.
7519 : *
7520 : * ANDing the index predicate with the explicitly given indexquals produces
7521 : * a more accurate idea of the index's selectivity. However, we need to be
7522 : * careful not to insert redundant clauses, because clauselist_selectivity()
7523 : * is easily fooled into computing a too-low selectivity estimate. Our
7524 : * approach is to add only the predicate clause(s) that cannot be proven to
7525 : * be implied by the given indexquals. This successfully handles cases such
7526 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7527 : * There are many other cases where we won't detect redundancy, leading to a
7528 : * too-low selectivity estimate, which will bias the system in favor of using
7529 : * partial indexes where possible. That is not necessarily bad though.
7530 : *
7531 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7532 : * does not, so the output list will be mixed. This is OK for both
7533 : * predicate_implied_by() and clauselist_selectivity(), but might be
7534 : * problematic if the result were passed to other things.
7535 : */
7536 : List *
7537 1333776 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7538 : {
7539 1333776 : List *predExtraQuals = NIL;
7540 : ListCell *lc;
7541 :
7542 1333776 : if (index->indpred == NIL)
7543 1331750 : return indexQuals;
7544 :
7545 4064 : foreach(lc, index->indpred)
7546 : {
7547 2038 : Node *predQual = (Node *) lfirst(lc);
7548 2038 : List *oneQual = list_make1(predQual);
7549 :
7550 2038 : if (!predicate_implied_by(oneQual, indexQuals, false))
7551 1836 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7552 : }
7553 2026 : return list_concat(predExtraQuals, indexQuals);
7554 : }
7555 :
7556 : /*
7557 : * Estimate correlation of btree index's first column.
7558 : *
7559 : * If we can get an estimate of the first column's ordering correlation C
7560 : * from pg_statistic, estimate the index correlation as C for a single-column
7561 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7562 : * columns dilute the importance of the first column's ordering, but don't
7563 : * negate it entirely.
7564 : *
7565 : * We already filled in the stats tuple for *vardata when called.
7566 : */
7567 : static double
7568 588624 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7569 : {
7570 : Oid sortop;
7571 : AttStatsSlot sslot;
7572 588624 : double indexCorrelation = 0;
7573 :
7574 : Assert(HeapTupleIsValid(vardata->statsTuple));
7575 :
7576 588624 : sortop = get_opfamily_member(index->opfamily[0],
7577 588624 : index->opcintype[0],
7578 588624 : index->opcintype[0],
7579 : BTLessStrategyNumber);
7580 1177248 : if (OidIsValid(sortop) &&
7581 588624 : get_attstatsslot(&sslot, vardata->statsTuple,
7582 : STATISTIC_KIND_CORRELATION, sortop,
7583 : ATTSTATSSLOT_NUMBERS))
7584 : {
7585 : double varCorrelation;
7586 :
7587 : Assert(sslot.nnumbers == 1);
7588 580630 : varCorrelation = sslot.numbers[0];
7589 :
7590 580630 : if (index->reverse_sort[0])
7591 0 : varCorrelation = -varCorrelation;
7592 :
7593 580630 : if (index->nkeycolumns > 1)
7594 202838 : indexCorrelation = varCorrelation * 0.75;
7595 : else
7596 377792 : indexCorrelation = varCorrelation;
7597 :
7598 580630 : free_attstatsslot(&sslot);
7599 : }
7600 :
7601 588624 : return indexCorrelation;
7602 : }
7603 :
7604 : void
7605 782994 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7606 : Cost *indexStartupCost, Cost *indexTotalCost,
7607 : Selectivity *indexSelectivity, double *indexCorrelation,
7608 : double *indexPages)
7609 : {
7610 782994 : IndexOptInfo *index = path->indexinfo;
7611 782994 : GenericCosts costs = {0};
7612 782994 : VariableStatData vardata = {0};
7613 : double numIndexTuples;
7614 : Cost descentCost;
7615 : List *indexBoundQuals;
7616 : List *indexSkipQuals;
7617 : int indexcol;
7618 : bool eqQualHere;
7619 : bool found_row_compare;
7620 : bool found_array;
7621 : bool found_is_null_op;
7622 782994 : bool have_correlation = false;
7623 : double num_sa_scans;
7624 782994 : double correlation = 0.0;
7625 : ListCell *lc;
7626 :
7627 : /*
7628 : * For a btree scan, only leading '=' quals plus inequality quals for the
7629 : * immediately next attribute contribute to index selectivity (these are
7630 : * the "boundary quals" that determine the starting and stopping points of
7631 : * the index scan). Additional quals can suppress visits to the heap, so
7632 : * it's OK to count them in indexSelectivity, but they should not count
7633 : * for estimating numIndexTuples. So we must examine the given indexquals
7634 : * to find out which ones count as boundary quals. We rely on the
7635 : * knowledge that they are given in index column order. Note that nbtree
7636 : * preprocessing can add skip arrays that act as leading '=' quals in the
7637 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7638 : * are able to act as index bound quals (which we take into account here).
7639 : *
7640 : * For a RowCompareExpr, we consider only the first column, just as
7641 : * rowcomparesel() does.
7642 : *
7643 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7644 : * to N index descents (not just one), but the underlying array key's
7645 : * operator can be considered to act the same as it normally does.
7646 : */
7647 782994 : indexBoundQuals = NIL;
7648 782994 : indexSkipQuals = NIL;
7649 782994 : indexcol = 0;
7650 782994 : eqQualHere = false;
7651 782994 : found_row_compare = false;
7652 782994 : found_array = false;
7653 782994 : found_is_null_op = false;
7654 782994 : num_sa_scans = 1;
7655 1331570 : foreach(lc, path->indexclauses)
7656 : {
7657 585468 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7658 : ListCell *lc2;
7659 :
7660 585468 : if (indexcol < iclause->indexcol)
7661 : {
7662 115054 : double num_sa_scans_prev_cols = num_sa_scans;
7663 :
7664 : /*
7665 : * Beginning of a new column's quals.
7666 : *
7667 : * Skip scans use skip arrays, which are ScalarArrayOp style
7668 : * arrays that generate their elements procedurally and on demand.
7669 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7670 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7671 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7672 : * the array on "a" must also return "IS NULL" matches, since our
7673 : * WHERE clause used no strict operator on "a").
7674 : *
7675 : * Here we consider how nbtree will backfill skip arrays for any
7676 : * index columns that lacked an '=' qual. This maintains our
7677 : * num_sa_scans estimate, and determines if this new column (the
7678 : * "iclause->indexcol" column, not the prior "indexcol" column)
7679 : * can have its RestrictInfos/quals added to indexBoundQuals.
7680 : *
7681 : * We'll need to handle columns that have inequality quals, where
7682 : * the skip array generates values from a range constrained by the
7683 : * quals (not every possible value). We've been maintaining
7684 : * indexSkipQuals to help with this; it will now contain all of
7685 : * the prior column's quals (that is, indexcol's quals) when they
7686 : * might be used for this.
7687 : */
7688 115054 : if (found_row_compare)
7689 : {
7690 : /*
7691 : * Skip arrays can't be added after a RowCompare input qual
7692 : * due to limitations in nbtree
7693 : */
7694 24 : break;
7695 : }
7696 115030 : if (eqQualHere)
7697 : {
7698 : /*
7699 : * Don't need to add a skip array for an indexcol that already
7700 : * has an '=' qual/equality constraint
7701 : */
7702 78766 : indexcol++;
7703 78766 : indexSkipQuals = NIL;
7704 : }
7705 115030 : eqQualHere = false;
7706 :
7707 118044 : while (indexcol < iclause->indexcol)
7708 : {
7709 : double ndistinct;
7710 39882 : bool isdefault = true;
7711 :
7712 39882 : found_array = true;
7713 :
7714 : /*
7715 : * A skipped attribute's ndistinct forms the basis of our
7716 : * estimate of the total number of "array elements" used by
7717 : * its skip array at runtime. Look that up first.
7718 : */
7719 39882 : examine_indexcol_variable(root, index, indexcol, &vardata);
7720 39882 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7721 :
7722 39882 : if (indexcol == 0)
7723 : {
7724 : /*
7725 : * Get an estimate of the leading column's correlation in
7726 : * passing (avoids rereading variable stats below)
7727 : */
7728 36252 : if (HeapTupleIsValid(vardata.statsTuple))
7729 23500 : correlation = btcost_correlation(index, &vardata);
7730 36252 : have_correlation = true;
7731 : }
7732 :
7733 39882 : ReleaseVariableStats(vardata);
7734 :
7735 : /*
7736 : * If ndistinct is a default estimate, conservatively assume
7737 : * that no skipping will happen at runtime
7738 : */
7739 39882 : if (isdefault)
7740 : {
7741 11556 : num_sa_scans = num_sa_scans_prev_cols;
7742 36868 : break; /* done building indexBoundQuals */
7743 : }
7744 :
7745 : /*
7746 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7747 : */
7748 28326 : if (indexSkipQuals != NIL)
7749 : {
7750 : List *partialSkipQuals;
7751 : Selectivity ndistinctfrac;
7752 :
7753 : /*
7754 : * If the index is partial, AND the index predicate with
7755 : * the index-bound quals to produce a more accurate idea
7756 : * of the number of distinct values for prior indexcol
7757 : */
7758 664 : partialSkipQuals = add_predicate_to_index_quals(index,
7759 : indexSkipQuals);
7760 :
7761 664 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7762 664 : index->rel->relid,
7763 : JOIN_INNER,
7764 : NULL);
7765 :
7766 : /*
7767 : * If ndistinctfrac is selective (on its own), the scan is
7768 : * unlikely to benefit from repositioning itself using
7769 : * later quals. Do not allow iclause->indexcol's quals to
7770 : * be added to indexBoundQuals (it would increase descent
7771 : * costs, without lowering numIndexTuples costs by much).
7772 : */
7773 664 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7774 : {
7775 374 : num_sa_scans = num_sa_scans_prev_cols;
7776 374 : break; /* done building indexBoundQuals */
7777 : }
7778 :
7779 : /* Adjust ndistinct downward */
7780 290 : ndistinct = rint(ndistinct * ndistinctfrac);
7781 290 : ndistinct = Max(ndistinct, 1);
7782 : }
7783 :
7784 : /*
7785 : * When there's no inequality quals, account for the need to
7786 : * find an initial value by counting -inf/+inf as a value.
7787 : *
7788 : * We don't charge anything extra for possible next/prior key
7789 : * index probes, which are sometimes used to find the next
7790 : * valid skip array element (ahead of using the located
7791 : * element value to relocate the scan to the next position
7792 : * that might contain matching tuples). It seems hard to do
7793 : * better here. Use of the skip support infrastructure often
7794 : * avoids most next/prior key probes. But even when it can't,
7795 : * there's a decent chance that most individual next/prior key
7796 : * probes will locate a leaf page whose key space overlaps all
7797 : * of the scan's keys (even the lower-order keys) -- which
7798 : * also avoids the need for a separate, extra index descent.
7799 : * Note also that these probes are much cheaper than non-probe
7800 : * primitive index scans: they're reliably very selective.
7801 : */
7802 27952 : if (indexSkipQuals == NIL)
7803 27662 : ndistinct += 1;
7804 :
7805 : /*
7806 : * Update num_sa_scans estimate by multiplying by ndistinct.
7807 : *
7808 : * We make the pessimistic assumption that there is no
7809 : * naturally occurring cross-column correlation. This is
7810 : * often wrong, but it seems best to err on the side of not
7811 : * expecting skipping to be helpful...
7812 : */
7813 27952 : num_sa_scans *= ndistinct;
7814 :
7815 : /*
7816 : * ...but back out of adding this latest group of 1 or more
7817 : * skip arrays when num_sa_scans exceeds the total number of
7818 : * index pages (revert to num_sa_scans from before indexcol).
7819 : * This causes a sharp discontinuity in cost (as a function of
7820 : * the indexcol's ndistinct), but that is representative of
7821 : * actual runtime costs.
7822 : *
7823 : * Note that skipping is helpful when each primitive index
7824 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7825 : * on average. Skip arrays bring savings in CPU costs due to
7826 : * the scan not needing to evaluate indexquals against every
7827 : * tuple, which can greatly exceed any savings in I/O costs.
7828 : * This test is a test of whether num_sa_scans implies that
7829 : * we're past the point where the ability to skip ceases to
7830 : * lower the scan's costs (even qual evaluation CPU costs).
7831 : */
7832 27952 : if (index->pages < num_sa_scans)
7833 : {
7834 24938 : num_sa_scans = num_sa_scans_prev_cols;
7835 24938 : break; /* done building indexBoundQuals */
7836 : }
7837 :
7838 3014 : indexcol++;
7839 3014 : indexSkipQuals = NIL;
7840 : }
7841 :
7842 : /*
7843 : * Finished considering the need to add skip arrays to bridge an
7844 : * initial eqQualHere gap between the old and new index columns
7845 : * (or there was no initial eqQualHere gap in the first place).
7846 : *
7847 : * If an initial gap could not be bridged, then new column's quals
7848 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7849 : * and so won't affect our final numIndexTuples estimate.
7850 : */
7851 115030 : if (indexcol != iclause->indexcol)
7852 36868 : break; /* done building indexBoundQuals */
7853 : }
7854 :
7855 : Assert(indexcol == iclause->indexcol);
7856 :
7857 : /* Examine each indexqual associated with this index clause */
7858 1099896 : foreach(lc2, iclause->indexquals)
7859 : {
7860 551320 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7861 551320 : Expr *clause = rinfo->clause;
7862 551320 : Oid clause_op = InvalidOid;
7863 : int op_strategy;
7864 :
7865 551320 : if (IsA(clause, OpExpr))
7866 : {
7867 537722 : OpExpr *op = (OpExpr *) clause;
7868 :
7869 537722 : clause_op = op->opno;
7870 : }
7871 13598 : else if (IsA(clause, RowCompareExpr))
7872 : {
7873 396 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7874 :
7875 396 : clause_op = linitial_oid(rc->opnos);
7876 396 : found_row_compare = true;
7877 : }
7878 13202 : else if (IsA(clause, ScalarArrayOpExpr))
7879 : {
7880 10918 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7881 10918 : Node *other_operand = (Node *) lsecond(saop->args);
7882 10918 : double alength = estimate_array_length(root, other_operand);
7883 :
7884 10918 : clause_op = saop->opno;
7885 10918 : found_array = true;
7886 : /* estimate SA descents by indexBoundQuals only */
7887 10918 : if (alength > 1)
7888 10718 : num_sa_scans *= alength;
7889 : }
7890 2284 : else if (IsA(clause, NullTest))
7891 : {
7892 2284 : NullTest *nt = (NullTest *) clause;
7893 :
7894 2284 : if (nt->nulltesttype == IS_NULL)
7895 : {
7896 240 : found_is_null_op = true;
7897 : /* IS NULL is like = for selectivity/skip scan purposes */
7898 240 : eqQualHere = true;
7899 : }
7900 : }
7901 : else
7902 0 : elog(ERROR, "unsupported indexqual type: %d",
7903 : (int) nodeTag(clause));
7904 :
7905 : /* check for equality operator */
7906 551320 : if (OidIsValid(clause_op))
7907 : {
7908 549036 : op_strategy = get_op_opfamily_strategy(clause_op,
7909 549036 : index->opfamily[indexcol]);
7910 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7911 549036 : if (op_strategy == BTEqualStrategyNumber)
7912 522678 : eqQualHere = true;
7913 : }
7914 :
7915 551320 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7916 :
7917 : /*
7918 : * We apply inequality selectivities to estimate index descent
7919 : * costs with scans that use skip arrays. Save this indexcol's
7920 : * RestrictInfos if it looks like they'll be needed for that.
7921 : */
7922 551320 : if (!eqQualHere && !found_row_compare &&
7923 27304 : indexcol < index->nkeycolumns - 1)
7924 5712 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
7925 : }
7926 : }
7927 :
7928 : /*
7929 : * If index is unique and we found an '=' clause for each column, we can
7930 : * just assume numIndexTuples = 1 and skip the expensive
7931 : * clauselist_selectivity calculations. However, an array or NullTest
7932 : * always invalidates that theory (even when eqQualHere has been set).
7933 : */
7934 782994 : if (index->unique &&
7935 638402 : indexcol == index->nkeycolumns - 1 &&
7936 249188 : eqQualHere &&
7937 249188 : !found_array &&
7938 243078 : !found_is_null_op)
7939 243030 : numIndexTuples = 1.0;
7940 : else
7941 : {
7942 : List *selectivityQuals;
7943 : Selectivity btreeSelectivity;
7944 :
7945 : /*
7946 : * If the index is partial, AND the index predicate with the
7947 : * index-bound quals to produce a more accurate idea of the number of
7948 : * rows covered by the bound conditions.
7949 : */
7950 539964 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
7951 :
7952 539964 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
7953 539964 : index->rel->relid,
7954 : JOIN_INNER,
7955 : NULL);
7956 539964 : numIndexTuples = btreeSelectivity * index->rel->tuples;
7957 :
7958 : /*
7959 : * btree automatically combines individual array element primitive
7960 : * index scans whenever the tuples covered by the next set of array
7961 : * keys are close to tuples covered by the current set. That puts a
7962 : * natural ceiling on the worst case number of descents -- there
7963 : * cannot possibly be more than one descent per leaf page scanned.
7964 : *
7965 : * Clamp the number of descents to at most 1/3 the number of index
7966 : * pages. This avoids implausibly high estimates with low selectivity
7967 : * paths, where scans usually require only one or two descents. This
7968 : * is most likely to help when there are several SAOP clauses, where
7969 : * naively accepting the total number of distinct combinations of
7970 : * array elements as the number of descents would frequently lead to
7971 : * wild overestimates.
7972 : *
7973 : * We somewhat arbitrarily don't just make the cutoff the total number
7974 : * of leaf pages (we make it 1/3 the total number of pages instead) to
7975 : * give the btree code credit for its ability to continue on the leaf
7976 : * level with low selectivity scans.
7977 : *
7978 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
7979 : * skip array elements whose qual affects our numIndexTuples estimate.
7980 : */
7981 539964 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
7982 539964 : num_sa_scans = Max(num_sa_scans, 1);
7983 :
7984 : /*
7985 : * As in genericcostestimate(), we have to adjust for any array quals
7986 : * included in indexBoundQuals, and then round to integer.
7987 : *
7988 : * It is tempting to make genericcostestimate behave as if array
7989 : * clauses work in almost the same way as scalar operators during
7990 : * btree scans, making the top-level scan look like a continuous scan
7991 : * (as opposed to num_sa_scans-many primitive index scans). After
7992 : * all, btree scans mostly work like that at runtime. However, such a
7993 : * scheme would badly bias genericcostestimate's simplistic approach
7994 : * to calculating numIndexPages through prorating.
7995 : *
7996 : * Stick with the approach taken by non-native SAOP scans for now.
7997 : * genericcostestimate will use the Mackert-Lohman formula to
7998 : * compensate for repeat page fetches, even though that definitely
7999 : * won't happen during btree scans (not for leaf pages, at least).
8000 : * We're usually very pessimistic about the number of primitive index
8001 : * scans that will be required, but it's not clear how to do better.
8002 : */
8003 539964 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8004 : }
8005 :
8006 : /*
8007 : * Now do generic index cost estimation.
8008 : */
8009 782994 : costs.numIndexTuples = numIndexTuples;
8010 782994 : costs.num_sa_scans = num_sa_scans;
8011 :
8012 782994 : genericcostestimate(root, path, loop_count, &costs);
8013 :
8014 : /*
8015 : * Add a CPU-cost component to represent the costs of initial btree
8016 : * descent. We don't charge any I/O cost for touching upper btree levels,
8017 : * since they tend to stay in cache, but we still have to do about log2(N)
8018 : * comparisons to descend a btree of N leaf tuples. We charge one
8019 : * cpu_operator_cost per comparison.
8020 : *
8021 : * If there are SAOP or skip array keys, charge this once per estimated
8022 : * index descent. The ones after the first one are not startup cost so
8023 : * far as the overall plan goes, so just add them to "total" cost.
8024 : */
8025 782994 : if (index->tuples > 1) /* avoid computing log(0) */
8026 : {
8027 727638 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8028 727638 : costs.indexStartupCost += descentCost;
8029 727638 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8030 : }
8031 :
8032 : /*
8033 : * Even though we're not charging I/O cost for touching upper btree pages,
8034 : * it's still reasonable to charge some CPU cost per page descended
8035 : * through. Moreover, if we had no such charge at all, bloated indexes
8036 : * would appear to have the same search cost as unbloated ones, at least
8037 : * in cases where only a single leaf page is expected to be visited. This
8038 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8039 : * touched. The number of such pages is btree tree height plus one (ie,
8040 : * we charge for the leaf page too). As above, charge once per estimated
8041 : * SAOP/skip array descent.
8042 : */
8043 782994 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8044 782994 : costs.indexStartupCost += descentCost;
8045 782994 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8046 :
8047 782994 : if (!have_correlation)
8048 : {
8049 746742 : examine_indexcol_variable(root, index, 0, &vardata);
8050 746742 : if (HeapTupleIsValid(vardata.statsTuple))
8051 565124 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8052 746742 : ReleaseVariableStats(vardata);
8053 : }
8054 : else
8055 : {
8056 : /* btcost_correlation already called earlier on */
8057 36252 : costs.indexCorrelation = correlation;
8058 : }
8059 :
8060 782994 : *indexStartupCost = costs.indexStartupCost;
8061 782994 : *indexTotalCost = costs.indexTotalCost;
8062 782994 : *indexSelectivity = costs.indexSelectivity;
8063 782994 : *indexCorrelation = costs.indexCorrelation;
8064 782994 : *indexPages = costs.numIndexPages;
8065 782994 : }
8066 :
8067 : void
8068 418 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8069 : Cost *indexStartupCost, Cost *indexTotalCost,
8070 : Selectivity *indexSelectivity, double *indexCorrelation,
8071 : double *indexPages)
8072 : {
8073 418 : GenericCosts costs = {0};
8074 :
8075 418 : genericcostestimate(root, path, loop_count, &costs);
8076 :
8077 : /*
8078 : * A hash index has no descent costs as such, since the index AM can go
8079 : * directly to the target bucket after computing the hash value. There
8080 : * are a couple of other hash-specific costs that we could conceivably add
8081 : * here, though:
8082 : *
8083 : * Ideally we'd charge spc_random_page_cost for each page in the target
8084 : * bucket, not just the numIndexPages pages that genericcostestimate
8085 : * thought we'd visit. However in most cases we don't know which bucket
8086 : * that will be. There's no point in considering the average bucket size
8087 : * because the hash AM makes sure that's always one page.
8088 : *
8089 : * Likewise, we could consider charging some CPU for each index tuple in
8090 : * the bucket, if we knew how many there were. But the per-tuple cost is
8091 : * just a hash value comparison, not a general datatype-dependent
8092 : * comparison, so any such charge ought to be quite a bit less than
8093 : * cpu_operator_cost; which makes it probably not worth worrying about.
8094 : *
8095 : * A bigger issue is that chance hash-value collisions will result in
8096 : * wasted probes into the heap. We don't currently attempt to model this
8097 : * cost on the grounds that it's rare, but maybe it's not rare enough.
8098 : * (Any fix for this ought to consider the generic lossy-operator problem,
8099 : * though; it's not entirely hash-specific.)
8100 : */
8101 :
8102 418 : *indexStartupCost = costs.indexStartupCost;
8103 418 : *indexTotalCost = costs.indexTotalCost;
8104 418 : *indexSelectivity = costs.indexSelectivity;
8105 418 : *indexCorrelation = costs.indexCorrelation;
8106 418 : *indexPages = costs.numIndexPages;
8107 418 : }
8108 :
8109 : void
8110 4876 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8111 : Cost *indexStartupCost, Cost *indexTotalCost,
8112 : Selectivity *indexSelectivity, double *indexCorrelation,
8113 : double *indexPages)
8114 : {
8115 4876 : IndexOptInfo *index = path->indexinfo;
8116 4876 : GenericCosts costs = {0};
8117 : Cost descentCost;
8118 :
8119 4876 : genericcostestimate(root, path, loop_count, &costs);
8120 :
8121 : /*
8122 : * We model index descent costs similarly to those for btree, but to do
8123 : * that we first need an idea of the tree height. We somewhat arbitrarily
8124 : * assume that the fanout is 100, meaning the tree height is at most
8125 : * log100(index->pages).
8126 : *
8127 : * Although this computation isn't really expensive enough to require
8128 : * caching, we might as well use index->tree_height to cache it.
8129 : */
8130 4876 : if (index->tree_height < 0) /* unknown? */
8131 : {
8132 4862 : if (index->pages > 1) /* avoid computing log(0) */
8133 2720 : index->tree_height = (int) (log(index->pages) / log(100.0));
8134 : else
8135 2142 : index->tree_height = 0;
8136 : }
8137 :
8138 : /*
8139 : * Add a CPU-cost component to represent the costs of initial descent. We
8140 : * just use log(N) here not log2(N) since the branching factor isn't
8141 : * necessarily two anyway. As for btree, charge once per SA scan.
8142 : */
8143 4876 : if (index->tuples > 1) /* avoid computing log(0) */
8144 : {
8145 4876 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8146 4876 : costs.indexStartupCost += descentCost;
8147 4876 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8148 : }
8149 :
8150 : /*
8151 : * Likewise add a per-page charge, calculated the same as for btrees.
8152 : */
8153 4876 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8154 4876 : costs.indexStartupCost += descentCost;
8155 4876 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8156 :
8157 4876 : *indexStartupCost = costs.indexStartupCost;
8158 4876 : *indexTotalCost = costs.indexTotalCost;
8159 4876 : *indexSelectivity = costs.indexSelectivity;
8160 4876 : *indexCorrelation = costs.indexCorrelation;
8161 4876 : *indexPages = costs.numIndexPages;
8162 4876 : }
8163 :
8164 : void
8165 1784 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8166 : Cost *indexStartupCost, Cost *indexTotalCost,
8167 : Selectivity *indexSelectivity, double *indexCorrelation,
8168 : double *indexPages)
8169 : {
8170 1784 : IndexOptInfo *index = path->indexinfo;
8171 1784 : GenericCosts costs = {0};
8172 : Cost descentCost;
8173 :
8174 1784 : genericcostestimate(root, path, loop_count, &costs);
8175 :
8176 : /*
8177 : * We model index descent costs similarly to those for btree, but to do
8178 : * that we first need an idea of the tree height. We somewhat arbitrarily
8179 : * assume that the fanout is 100, meaning the tree height is at most
8180 : * log100(index->pages).
8181 : *
8182 : * Although this computation isn't really expensive enough to require
8183 : * caching, we might as well use index->tree_height to cache it.
8184 : */
8185 1784 : if (index->tree_height < 0) /* unknown? */
8186 : {
8187 1778 : if (index->pages > 1) /* avoid computing log(0) */
8188 1778 : index->tree_height = (int) (log(index->pages) / log(100.0));
8189 : else
8190 0 : index->tree_height = 0;
8191 : }
8192 :
8193 : /*
8194 : * Add a CPU-cost component to represent the costs of initial descent. We
8195 : * just use log(N) here not log2(N) since the branching factor isn't
8196 : * necessarily two anyway. As for btree, charge once per SA scan.
8197 : */
8198 1784 : if (index->tuples > 1) /* avoid computing log(0) */
8199 : {
8200 1784 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8201 1784 : costs.indexStartupCost += descentCost;
8202 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8203 : }
8204 :
8205 : /*
8206 : * Likewise add a per-page charge, calculated the same as for btrees.
8207 : */
8208 1784 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8209 1784 : costs.indexStartupCost += descentCost;
8210 1784 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8211 :
8212 1784 : *indexStartupCost = costs.indexStartupCost;
8213 1784 : *indexTotalCost = costs.indexTotalCost;
8214 1784 : *indexSelectivity = costs.indexSelectivity;
8215 1784 : *indexCorrelation = costs.indexCorrelation;
8216 1784 : *indexPages = costs.numIndexPages;
8217 1784 : }
8218 :
8219 :
8220 : /*
8221 : * Support routines for gincostestimate
8222 : */
8223 :
8224 : typedef struct
8225 : {
8226 : bool attHasFullScan[INDEX_MAX_KEYS];
8227 : bool attHasNormalScan[INDEX_MAX_KEYS];
8228 : double partialEntries;
8229 : double exactEntries;
8230 : double searchEntries;
8231 : double arrayScans;
8232 : } GinQualCounts;
8233 :
8234 : /*
8235 : * Estimate the number of index terms that need to be searched for while
8236 : * testing the given GIN query, and increment the counts in *counts
8237 : * appropriately. If the query is unsatisfiable, return false.
8238 : */
8239 : static bool
8240 2480 : gincost_pattern(IndexOptInfo *index, int indexcol,
8241 : Oid clause_op, Datum query,
8242 : GinQualCounts *counts)
8243 : {
8244 : FmgrInfo flinfo;
8245 : Oid extractProcOid;
8246 : Oid collation;
8247 : int strategy_op;
8248 : Oid lefttype,
8249 : righttype;
8250 2480 : int32 nentries = 0;
8251 2480 : bool *partial_matches = NULL;
8252 2480 : Pointer *extra_data = NULL;
8253 2480 : bool *nullFlags = NULL;
8254 2480 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8255 : int32 i;
8256 :
8257 : Assert(indexcol < index->nkeycolumns);
8258 :
8259 : /*
8260 : * Get the operator's strategy number and declared input data types within
8261 : * the index opfamily. (We don't need the latter, but we use
8262 : * get_op_opfamily_properties because it will throw error if it fails to
8263 : * find a matching pg_amop entry.)
8264 : */
8265 2480 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8266 : &strategy_op, &lefttype, &righttype);
8267 :
8268 : /*
8269 : * GIN always uses the "default" support functions, which are those with
8270 : * lefttype == righttype == the opclass' opcintype (see
8271 : * IndexSupportInitialize in relcache.c).
8272 : */
8273 2480 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8274 2480 : index->opcintype[indexcol],
8275 2480 : index->opcintype[indexcol],
8276 : GIN_EXTRACTQUERY_PROC);
8277 :
8278 2480 : if (!OidIsValid(extractProcOid))
8279 : {
8280 : /* should not happen; throw same error as index_getprocinfo */
8281 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8282 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8283 : get_rel_name(index->indexoid));
8284 : }
8285 :
8286 : /*
8287 : * Choose collation to pass to extractProc (should match initGinState).
8288 : */
8289 2480 : if (OidIsValid(index->indexcollations[indexcol]))
8290 414 : collation = index->indexcollations[indexcol];
8291 : else
8292 2066 : collation = DEFAULT_COLLATION_OID;
8293 :
8294 2480 : fmgr_info(extractProcOid, &flinfo);
8295 :
8296 2480 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8297 :
8298 2480 : FunctionCall7Coll(&flinfo,
8299 : collation,
8300 : query,
8301 : PointerGetDatum(&nentries),
8302 : UInt16GetDatum(strategy_op),
8303 : PointerGetDatum(&partial_matches),
8304 : PointerGetDatum(&extra_data),
8305 : PointerGetDatum(&nullFlags),
8306 : PointerGetDatum(&searchMode));
8307 :
8308 2480 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8309 : {
8310 : /* No match is possible */
8311 12 : return false;
8312 : }
8313 :
8314 9676 : for (i = 0; i < nentries; i++)
8315 : {
8316 : /*
8317 : * For partial match we haven't any information to estimate number of
8318 : * matched entries in index, so, we just estimate it as 100
8319 : */
8320 7208 : if (partial_matches && partial_matches[i])
8321 694 : counts->partialEntries += 100;
8322 : else
8323 6514 : counts->exactEntries++;
8324 :
8325 7208 : counts->searchEntries++;
8326 : }
8327 :
8328 2468 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8329 : {
8330 1984 : counts->attHasNormalScan[indexcol] = true;
8331 : }
8332 484 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8333 : {
8334 : /* Treat "include empty" like an exact-match item */
8335 44 : counts->attHasNormalScan[indexcol] = true;
8336 44 : counts->exactEntries++;
8337 44 : counts->searchEntries++;
8338 : }
8339 : else
8340 : {
8341 : /* It's GIN_SEARCH_MODE_ALL */
8342 440 : counts->attHasFullScan[indexcol] = true;
8343 : }
8344 :
8345 2468 : return true;
8346 : }
8347 :
8348 : /*
8349 : * Estimate the number of index terms that need to be searched for while
8350 : * testing the given GIN index clause, and increment the counts in *counts
8351 : * appropriately. If the query is unsatisfiable, return false.
8352 : */
8353 : static bool
8354 2468 : gincost_opexpr(PlannerInfo *root,
8355 : IndexOptInfo *index,
8356 : int indexcol,
8357 : OpExpr *clause,
8358 : GinQualCounts *counts)
8359 : {
8360 2468 : Oid clause_op = clause->opno;
8361 2468 : Node *operand = (Node *) lsecond(clause->args);
8362 :
8363 : /* aggressively reduce to a constant, and look through relabeling */
8364 2468 : operand = estimate_expression_value(root, operand);
8365 :
8366 2468 : if (IsA(operand, RelabelType))
8367 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8368 :
8369 : /*
8370 : * It's impossible to call extractQuery method for unknown operand. So
8371 : * unless operand is a Const we can't do much; just assume there will be
8372 : * one ordinary search entry from the operand at runtime.
8373 : */
8374 2468 : if (!IsA(operand, Const))
8375 : {
8376 0 : counts->exactEntries++;
8377 0 : counts->searchEntries++;
8378 0 : return true;
8379 : }
8380 :
8381 : /* If Const is null, there can be no matches */
8382 2468 : if (((Const *) operand)->constisnull)
8383 0 : return false;
8384 :
8385 : /* Otherwise, apply extractQuery and get the actual term counts */
8386 2468 : return gincost_pattern(index, indexcol, clause_op,
8387 : ((Const *) operand)->constvalue,
8388 : counts);
8389 : }
8390 :
8391 : /*
8392 : * Estimate the number of index terms that need to be searched for while
8393 : * testing the given GIN index clause, and increment the counts in *counts
8394 : * appropriately. If the query is unsatisfiable, return false.
8395 : *
8396 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8397 : * each of which involves one value from the RHS array, plus all the
8398 : * non-array quals (if any). To model this, we average the counts across
8399 : * the RHS elements, and add the averages to the counts in *counts (which
8400 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8401 : * by N, causing gincostestimate to scale up its estimates accordingly.
8402 : */
8403 : static bool
8404 6 : gincost_scalararrayopexpr(PlannerInfo *root,
8405 : IndexOptInfo *index,
8406 : int indexcol,
8407 : ScalarArrayOpExpr *clause,
8408 : double numIndexEntries,
8409 : GinQualCounts *counts)
8410 : {
8411 6 : Oid clause_op = clause->opno;
8412 6 : Node *rightop = (Node *) lsecond(clause->args);
8413 : ArrayType *arrayval;
8414 : int16 elmlen;
8415 : bool elmbyval;
8416 : char elmalign;
8417 : int numElems;
8418 : Datum *elemValues;
8419 : bool *elemNulls;
8420 : GinQualCounts arraycounts;
8421 6 : int numPossible = 0;
8422 : int i;
8423 :
8424 : Assert(clause->useOr);
8425 :
8426 : /* aggressively reduce to a constant, and look through relabeling */
8427 6 : rightop = estimate_expression_value(root, rightop);
8428 :
8429 6 : if (IsA(rightop, RelabelType))
8430 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8431 :
8432 : /*
8433 : * It's impossible to call extractQuery method for unknown operand. So
8434 : * unless operand is a Const we can't do much; just assume there will be
8435 : * one ordinary search entry from each array entry at runtime, and fall
8436 : * back on a probably-bad estimate of the number of array entries.
8437 : */
8438 6 : if (!IsA(rightop, Const))
8439 : {
8440 0 : counts->exactEntries++;
8441 0 : counts->searchEntries++;
8442 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8443 0 : return true;
8444 : }
8445 :
8446 : /* If Const is null, there can be no matches */
8447 6 : if (((Const *) rightop)->constisnull)
8448 0 : return false;
8449 :
8450 : /* Otherwise, extract the array elements and iterate over them */
8451 6 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8452 6 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8453 : &elmlen, &elmbyval, &elmalign);
8454 6 : deconstruct_array(arrayval,
8455 : ARR_ELEMTYPE(arrayval),
8456 : elmlen, elmbyval, elmalign,
8457 : &elemValues, &elemNulls, &numElems);
8458 :
8459 6 : memset(&arraycounts, 0, sizeof(arraycounts));
8460 :
8461 18 : for (i = 0; i < numElems; i++)
8462 : {
8463 : GinQualCounts elemcounts;
8464 :
8465 : /* NULL can't match anything, so ignore, as the executor will */
8466 12 : if (elemNulls[i])
8467 0 : continue;
8468 :
8469 : /* Otherwise, apply extractQuery and get the actual term counts */
8470 12 : memset(&elemcounts, 0, sizeof(elemcounts));
8471 :
8472 12 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8473 : &elemcounts))
8474 : {
8475 : /* We ignore array elements that are unsatisfiable patterns */
8476 12 : numPossible++;
8477 :
8478 12 : if (elemcounts.attHasFullScan[indexcol] &&
8479 0 : !elemcounts.attHasNormalScan[indexcol])
8480 : {
8481 : /*
8482 : * Full index scan will be required. We treat this as if
8483 : * every key in the index had been listed in the query; is
8484 : * that reasonable?
8485 : */
8486 0 : elemcounts.partialEntries = 0;
8487 0 : elemcounts.exactEntries = numIndexEntries;
8488 0 : elemcounts.searchEntries = numIndexEntries;
8489 : }
8490 12 : arraycounts.partialEntries += elemcounts.partialEntries;
8491 12 : arraycounts.exactEntries += elemcounts.exactEntries;
8492 12 : arraycounts.searchEntries += elemcounts.searchEntries;
8493 : }
8494 : }
8495 :
8496 6 : if (numPossible == 0)
8497 : {
8498 : /* No satisfiable patterns in the array */
8499 0 : return false;
8500 : }
8501 :
8502 : /*
8503 : * Now add the averages to the global counts. This will give us an
8504 : * estimate of the average number of terms searched for in each indexscan,
8505 : * including contributions from both array and non-array quals.
8506 : */
8507 6 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8508 6 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8509 6 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8510 :
8511 6 : counts->arrayScans *= numPossible;
8512 :
8513 6 : return true;
8514 : }
8515 :
8516 : /*
8517 : * GIN has search behavior completely different from other index types
8518 : */
8519 : void
8520 2264 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8521 : Cost *indexStartupCost, Cost *indexTotalCost,
8522 : Selectivity *indexSelectivity, double *indexCorrelation,
8523 : double *indexPages)
8524 : {
8525 2264 : IndexOptInfo *index = path->indexinfo;
8526 2264 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8527 : List *selectivityQuals;
8528 2264 : double numPages = index->pages,
8529 2264 : numTuples = index->tuples;
8530 : double numEntryPages,
8531 : numDataPages,
8532 : numPendingPages,
8533 : numEntries;
8534 : GinQualCounts counts;
8535 : bool matchPossible;
8536 : bool fullIndexScan;
8537 : double partialScale;
8538 : double entryPagesFetched,
8539 : dataPagesFetched,
8540 : dataPagesFetchedBySel;
8541 : double qual_op_cost,
8542 : qual_arg_cost,
8543 : spc_random_page_cost,
8544 : outer_scans;
8545 : Cost descentCost;
8546 : Relation indexRel;
8547 : GinStatsData ginStats;
8548 : ListCell *lc;
8549 : int i;
8550 :
8551 : /*
8552 : * Obtain statistical information from the meta page, if possible. Else
8553 : * set ginStats to zeroes, and we'll cope below.
8554 : */
8555 2264 : if (!index->hypothetical)
8556 : {
8557 : /* Lock should have already been obtained in plancat.c */
8558 2264 : indexRel = index_open(index->indexoid, NoLock);
8559 2264 : ginGetStats(indexRel, &ginStats);
8560 2264 : index_close(indexRel, NoLock);
8561 : }
8562 : else
8563 : {
8564 0 : memset(&ginStats, 0, sizeof(ginStats));
8565 : }
8566 :
8567 : /*
8568 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8569 : * trusted, but the other fields are data as of the last VACUUM. We can
8570 : * scale them up to account for growth since then, but that method only
8571 : * goes so far; in the worst case, the stats might be for a completely
8572 : * empty index, and scaling them will produce pretty bogus numbers.
8573 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8574 : * it's grown more than that, fall back to estimating things only from the
8575 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8576 : * so long as it's not clearly insane, ie, more than the index size.
8577 : */
8578 2264 : if (ginStats.nPendingPages < numPages)
8579 2264 : numPendingPages = ginStats.nPendingPages;
8580 : else
8581 0 : numPendingPages = 0;
8582 :
8583 2264 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8584 2264 : ginStats.nTotalPages > numPages / 4 &&
8585 2212 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8586 1948 : {
8587 : /*
8588 : * OK, the stats seem close enough to sane to be trusted. But we
8589 : * still need to scale them by the ratio numPages / nTotalPages to
8590 : * account for growth since the last VACUUM.
8591 : */
8592 1948 : double scale = numPages / ginStats.nTotalPages;
8593 :
8594 1948 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8595 1948 : numDataPages = ceil(ginStats.nDataPages * scale);
8596 1948 : numEntries = ceil(ginStats.nEntries * scale);
8597 : /* ensure we didn't round up too much */
8598 1948 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8599 1948 : numDataPages = Min(numDataPages,
8600 : numPages - numPendingPages - numEntryPages);
8601 : }
8602 : else
8603 : {
8604 : /*
8605 : * We might get here because it's a hypothetical index, or an index
8606 : * created pre-9.1 and never vacuumed since upgrading (in which case
8607 : * its stats would read as zeroes), or just because it's grown too
8608 : * much since the last VACUUM for us to put our faith in scaling.
8609 : *
8610 : * Invent some plausible internal statistics based on the index page
8611 : * count (and clamp that to at least 10 pages, just in case). We
8612 : * estimate that 90% of the index is entry pages, and the rest is data
8613 : * pages. Estimate 100 entries per entry page; this is rather bogus
8614 : * since it'll depend on the size of the keys, but it's more robust
8615 : * than trying to predict the number of entries per heap tuple.
8616 : */
8617 316 : numPages = Max(numPages, 10);
8618 316 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8619 316 : numDataPages = numPages - numPendingPages - numEntryPages;
8620 316 : numEntries = floor(numEntryPages * 100);
8621 : }
8622 :
8623 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8624 2264 : if (numEntries < 1)
8625 0 : numEntries = 1;
8626 :
8627 : /*
8628 : * If the index is partial, AND the index predicate with the index-bound
8629 : * quals to produce a more accurate idea of the number of rows covered by
8630 : * the bound conditions.
8631 : */
8632 2264 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8633 :
8634 : /* Estimate the fraction of main-table tuples that will be visited */
8635 4528 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8636 2264 : index->rel->relid,
8637 : JOIN_INNER,
8638 : NULL);
8639 :
8640 : /* fetch estimated page cost for tablespace containing index */
8641 2264 : get_tablespace_page_costs(index->reltablespace,
8642 : &spc_random_page_cost,
8643 : NULL);
8644 :
8645 : /*
8646 : * Generic assumption about index correlation: there isn't any.
8647 : */
8648 2264 : *indexCorrelation = 0.0;
8649 :
8650 : /*
8651 : * Examine quals to estimate number of search entries & partial matches
8652 : */
8653 2264 : memset(&counts, 0, sizeof(counts));
8654 2264 : counts.arrayScans = 1;
8655 2264 : matchPossible = true;
8656 :
8657 4738 : foreach(lc, path->indexclauses)
8658 : {
8659 2474 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8660 : ListCell *lc2;
8661 :
8662 4936 : foreach(lc2, iclause->indexquals)
8663 : {
8664 2474 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8665 2474 : Expr *clause = rinfo->clause;
8666 :
8667 2474 : if (IsA(clause, OpExpr))
8668 : {
8669 2468 : matchPossible = gincost_opexpr(root,
8670 : index,
8671 2468 : iclause->indexcol,
8672 : (OpExpr *) clause,
8673 : &counts);
8674 2468 : if (!matchPossible)
8675 12 : break;
8676 : }
8677 6 : else if (IsA(clause, ScalarArrayOpExpr))
8678 : {
8679 6 : matchPossible = gincost_scalararrayopexpr(root,
8680 : index,
8681 6 : iclause->indexcol,
8682 : (ScalarArrayOpExpr *) clause,
8683 : numEntries,
8684 : &counts);
8685 6 : if (!matchPossible)
8686 0 : break;
8687 : }
8688 : else
8689 : {
8690 : /* shouldn't be anything else for a GIN index */
8691 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8692 : (int) nodeTag(clause));
8693 : }
8694 : }
8695 : }
8696 :
8697 : /* Fall out if there were any provably-unsatisfiable quals */
8698 2264 : if (!matchPossible)
8699 : {
8700 12 : *indexStartupCost = 0;
8701 12 : *indexTotalCost = 0;
8702 12 : *indexSelectivity = 0;
8703 12 : return;
8704 : }
8705 :
8706 : /*
8707 : * If attribute has a full scan and at the same time doesn't have normal
8708 : * scan, then we'll have to scan all non-null entries of that attribute.
8709 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8710 : * must assume the whole GIN index has to be scanned in this case.
8711 : */
8712 2252 : fullIndexScan = false;
8713 4394 : for (i = 0; i < index->nkeycolumns; i++)
8714 : {
8715 2480 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8716 : {
8717 338 : fullIndexScan = true;
8718 338 : break;
8719 : }
8720 : }
8721 :
8722 2252 : if (fullIndexScan || indexQuals == NIL)
8723 : {
8724 : /*
8725 : * Full index scan will be required. We treat this as if every key in
8726 : * the index had been listed in the query; is that reasonable?
8727 : */
8728 338 : counts.partialEntries = 0;
8729 338 : counts.exactEntries = numEntries;
8730 338 : counts.searchEntries = numEntries;
8731 : }
8732 :
8733 : /* Will we have more than one iteration of a nestloop scan? */
8734 2252 : outer_scans = loop_count;
8735 :
8736 : /*
8737 : * Compute cost to begin scan, first of all, pay attention to pending
8738 : * list.
8739 : */
8740 2252 : entryPagesFetched = numPendingPages;
8741 :
8742 : /*
8743 : * Estimate number of entry pages read. We need to do
8744 : * counts.searchEntries searches. Use a power function as it should be,
8745 : * but tuples on leaf pages usually is much greater. Here we include all
8746 : * searches in entry tree, including search of first entry in partial
8747 : * match algorithm
8748 : */
8749 2252 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8750 :
8751 : /*
8752 : * Add an estimate of entry pages read by partial match algorithm. It's a
8753 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8754 : * so estimate it as proportion. Because counts.partialEntries is really
8755 : * pretty bogus (see code above), it's possible that it is more than
8756 : * numEntries; clamp the proportion to ensure sanity.
8757 : */
8758 2252 : partialScale = counts.partialEntries / numEntries;
8759 2252 : partialScale = Min(partialScale, 1.0);
8760 :
8761 2252 : entryPagesFetched += ceil(numEntryPages * partialScale);
8762 :
8763 : /*
8764 : * Partial match algorithm reads all data pages before doing actual scan,
8765 : * so it's a startup cost. Again, we haven't any useful stats here, so
8766 : * estimate it as proportion.
8767 : */
8768 2252 : dataPagesFetched = ceil(numDataPages * partialScale);
8769 :
8770 2252 : *indexStartupCost = 0;
8771 2252 : *indexTotalCost = 0;
8772 :
8773 : /*
8774 : * Add a CPU-cost component to represent the costs of initial entry btree
8775 : * descent. We don't charge any I/O cost for touching upper btree levels,
8776 : * since they tend to stay in cache, but we still have to do about log2(N)
8777 : * comparisons to descend a btree of N leaf tuples. We charge one
8778 : * cpu_operator_cost per comparison.
8779 : *
8780 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8781 : * ones after the first one are not startup cost so far as the overall
8782 : * plan is concerned, so add them only to "total" cost.
8783 : */
8784 2252 : if (numEntries > 1) /* avoid computing log(0) */
8785 : {
8786 2252 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8787 2252 : *indexStartupCost += descentCost * counts.searchEntries;
8788 2252 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8789 : }
8790 :
8791 : /*
8792 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8793 : * loop.
8794 : */
8795 2252 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8796 2252 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8797 :
8798 : /*
8799 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8800 : * loop. Since those are the data pages from the partial match algorithm,
8801 : * charge them as startup cost.
8802 : */
8803 2252 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8804 :
8805 : /*
8806 : * Since we add the startup cost to the total cost later on, remove the
8807 : * initial arrayscan from the total.
8808 : */
8809 2252 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8810 :
8811 : /*
8812 : * Calculate cache effects if more than one scan due to nestloops or array
8813 : * quals. The result is pro-rated per nestloop scan, but the array qual
8814 : * factor shouldn't be pro-rated (compare genericcostestimate).
8815 : */
8816 2252 : if (outer_scans > 1 || counts.arrayScans > 1)
8817 : {
8818 6 : entryPagesFetched *= outer_scans * counts.arrayScans;
8819 6 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8820 : (BlockNumber) numEntryPages,
8821 : numEntryPages, root);
8822 6 : entryPagesFetched /= outer_scans;
8823 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8824 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8825 : (BlockNumber) numDataPages,
8826 : numDataPages, root);
8827 6 : dataPagesFetched /= outer_scans;
8828 : }
8829 :
8830 : /*
8831 : * Here we use random page cost because logically-close pages could be far
8832 : * apart on disk.
8833 : */
8834 2252 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8835 :
8836 : /*
8837 : * Now compute the number of data pages fetched during the scan.
8838 : *
8839 : * We assume every entry to have the same number of items, and that there
8840 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8841 : * statistics on the frequency of individual keys; it would be nice to use
8842 : * those here.)
8843 : */
8844 2252 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8845 :
8846 : /*
8847 : * If there is a lot of overlap among the entries, in particular if one of
8848 : * the entries is very frequent, the above calculation can grossly
8849 : * under-estimate. As a simple cross-check, calculate a lower bound based
8850 : * on the overall selectivity of the quals. At a minimum, we must read
8851 : * one item pointer for each matching entry.
8852 : *
8853 : * The width of each item pointer varies, based on the level of
8854 : * compression. We don't have statistics on that, but an average of
8855 : * around 3 bytes per item is fairly typical.
8856 : */
8857 2252 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8858 2252 : (numTuples / (BLCKSZ / 3)));
8859 2252 : if (dataPagesFetchedBySel > dataPagesFetched)
8860 1866 : dataPagesFetched = dataPagesFetchedBySel;
8861 :
8862 : /* Add one page cpu-cost to the startup cost */
8863 2252 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8864 :
8865 : /*
8866 : * Add once again a CPU-cost for those data pages, before amortizing for
8867 : * cache.
8868 : */
8869 2252 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8870 :
8871 : /* Account for cache effects, the same as above */
8872 2252 : if (outer_scans > 1 || counts.arrayScans > 1)
8873 : {
8874 6 : dataPagesFetched *= outer_scans * counts.arrayScans;
8875 6 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8876 : (BlockNumber) numDataPages,
8877 : numDataPages, root);
8878 6 : dataPagesFetched /= outer_scans;
8879 : }
8880 :
8881 : /* And apply random_page_cost as the cost per page */
8882 2252 : *indexTotalCost += *indexStartupCost +
8883 2252 : dataPagesFetched * spc_random_page_cost;
8884 :
8885 : /*
8886 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8887 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8888 : * those.
8889 : */
8890 2252 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8891 2252 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8892 :
8893 2252 : *indexStartupCost += qual_arg_cost;
8894 2252 : *indexTotalCost += qual_arg_cost;
8895 :
8896 : /*
8897 : * Add a cpu cost per search entry, corresponding to the actual visited
8898 : * entries.
8899 : */
8900 2252 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8901 : /* Now add a cpu cost per tuple in the posting lists / trees */
8902 2252 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8903 2252 : *indexPages = dataPagesFetched;
8904 : }
8905 :
8906 : /*
8907 : * BRIN has search behavior completely different from other index types
8908 : */
8909 : void
8910 10730 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8911 : Cost *indexStartupCost, Cost *indexTotalCost,
8912 : Selectivity *indexSelectivity, double *indexCorrelation,
8913 : double *indexPages)
8914 : {
8915 10730 : IndexOptInfo *index = path->indexinfo;
8916 10730 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8917 10730 : double numPages = index->pages;
8918 10730 : RelOptInfo *baserel = index->rel;
8919 10730 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
8920 : Cost spc_seq_page_cost;
8921 : Cost spc_random_page_cost;
8922 : double qual_arg_cost;
8923 : double qualSelectivity;
8924 : BrinStatsData statsData;
8925 : double indexRanges;
8926 : double minimalRanges;
8927 : double estimatedRanges;
8928 : double selec;
8929 : Relation indexRel;
8930 : ListCell *l;
8931 : VariableStatData vardata;
8932 :
8933 : Assert(rte->rtekind == RTE_RELATION);
8934 :
8935 : /* fetch estimated page cost for the tablespace containing the index */
8936 10730 : get_tablespace_page_costs(index->reltablespace,
8937 : &spc_random_page_cost,
8938 : &spc_seq_page_cost);
8939 :
8940 : /*
8941 : * Obtain some data from the index itself, if possible. Otherwise invent
8942 : * some plausible internal statistics based on the relation page count.
8943 : */
8944 10730 : if (!index->hypothetical)
8945 : {
8946 : /*
8947 : * A lock should have already been obtained on the index in plancat.c.
8948 : */
8949 10730 : indexRel = index_open(index->indexoid, NoLock);
8950 10730 : brinGetStats(indexRel, &statsData);
8951 10730 : index_close(indexRel, NoLock);
8952 :
8953 : /* work out the actual number of ranges in the index */
8954 10730 : indexRanges = Max(ceil((double) baserel->pages /
8955 : statsData.pagesPerRange), 1.0);
8956 : }
8957 : else
8958 : {
8959 : /*
8960 : * Assume default number of pages per range, and estimate the number
8961 : * of ranges based on that.
8962 : */
8963 0 : indexRanges = Max(ceil((double) baserel->pages /
8964 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
8965 :
8966 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
8967 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
8968 : }
8969 :
8970 : /*
8971 : * Compute index correlation
8972 : *
8973 : * Because we can use all index quals equally when scanning, we can use
8974 : * the largest correlation (in absolute value) among columns used by the
8975 : * query. Start at zero, the worst possible case. If we cannot find any
8976 : * correlation statistics, we will keep it as 0.
8977 : */
8978 10730 : *indexCorrelation = 0;
8979 :
8980 21462 : foreach(l, path->indexclauses)
8981 : {
8982 10732 : IndexClause *iclause = lfirst_node(IndexClause, l);
8983 10732 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
8984 :
8985 : /* attempt to lookup stats in relation for this index column */
8986 10732 : if (attnum != 0)
8987 : {
8988 : /* Simple variable -- look to stats for the underlying table */
8989 10732 : if (get_relation_stats_hook &&
8990 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
8991 : {
8992 : /*
8993 : * The hook took control of acquiring a stats tuple. If it
8994 : * did supply a tuple, it'd better have supplied a freefunc.
8995 : */
8996 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
8997 0 : elog(ERROR,
8998 : "no function provided to release variable stats with");
8999 : }
9000 : else
9001 : {
9002 10732 : vardata.statsTuple =
9003 10732 : SearchSysCache3(STATRELATTINH,
9004 : ObjectIdGetDatum(rte->relid),
9005 : Int16GetDatum(attnum),
9006 : BoolGetDatum(false));
9007 10732 : vardata.freefunc = ReleaseSysCache;
9008 : }
9009 : }
9010 : else
9011 : {
9012 : /*
9013 : * Looks like we've found an expression column in the index. Let's
9014 : * see if there's any stats for it.
9015 : */
9016 :
9017 : /* get the attnum from the 0-based index. */
9018 0 : attnum = iclause->indexcol + 1;
9019 :
9020 0 : if (get_index_stats_hook &&
9021 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9022 : {
9023 : /*
9024 : * The hook took control of acquiring a stats tuple. If it
9025 : * did supply a tuple, it'd better have supplied a freefunc.
9026 : */
9027 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9028 0 : !vardata.freefunc)
9029 0 : elog(ERROR, "no function provided to release variable stats with");
9030 : }
9031 : else
9032 : {
9033 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9034 : ObjectIdGetDatum(index->indexoid),
9035 : Int16GetDatum(attnum),
9036 : BoolGetDatum(false));
9037 0 : vardata.freefunc = ReleaseSysCache;
9038 : }
9039 : }
9040 :
9041 10732 : if (HeapTupleIsValid(vardata.statsTuple))
9042 : {
9043 : AttStatsSlot sslot;
9044 :
9045 36 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9046 : STATISTIC_KIND_CORRELATION, InvalidOid,
9047 : ATTSTATSSLOT_NUMBERS))
9048 : {
9049 36 : double varCorrelation = 0.0;
9050 :
9051 36 : if (sslot.nnumbers > 0)
9052 36 : varCorrelation = fabs(sslot.numbers[0]);
9053 :
9054 36 : if (varCorrelation > *indexCorrelation)
9055 36 : *indexCorrelation = varCorrelation;
9056 :
9057 36 : free_attstatsslot(&sslot);
9058 : }
9059 : }
9060 :
9061 10732 : ReleaseVariableStats(vardata);
9062 : }
9063 :
9064 10730 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9065 10730 : baserel->relid,
9066 : JOIN_INNER, NULL);
9067 :
9068 : /*
9069 : * Now calculate the minimum possible ranges we could match with if all of
9070 : * the rows were in the perfect order in the table's heap.
9071 : */
9072 10730 : minimalRanges = ceil(indexRanges * qualSelectivity);
9073 :
9074 : /*
9075 : * Now estimate the number of ranges that we'll touch by using the
9076 : * indexCorrelation from the stats. Careful not to divide by zero (note
9077 : * we're using the absolute value of the correlation).
9078 : */
9079 10730 : if (*indexCorrelation < 1.0e-10)
9080 10694 : estimatedRanges = indexRanges;
9081 : else
9082 36 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9083 :
9084 : /* we expect to visit this portion of the table */
9085 10730 : selec = estimatedRanges / indexRanges;
9086 :
9087 10730 : CLAMP_PROBABILITY(selec);
9088 :
9089 10730 : *indexSelectivity = selec;
9090 :
9091 : /*
9092 : * Compute the index qual costs, much as in genericcostestimate, to add to
9093 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9094 : * support those.
9095 : */
9096 10730 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9097 :
9098 : /*
9099 : * Compute the startup cost as the cost to read the whole revmap
9100 : * sequentially, including the cost to execute the index quals.
9101 : */
9102 10730 : *indexStartupCost =
9103 10730 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
9104 10730 : *indexStartupCost += qual_arg_cost;
9105 :
9106 : /*
9107 : * To read a BRIN index there might be a bit of back and forth over
9108 : * regular pages, as revmap might point to them out of sequential order;
9109 : * calculate the total cost as reading the whole index in random order.
9110 : */
9111 10730 : *indexTotalCost = *indexStartupCost +
9112 10730 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9113 :
9114 : /*
9115 : * Charge a small amount per range tuple which we expect to match to. This
9116 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9117 : * will set a bit for each page in the range when we find a matching
9118 : * range, so we must multiply the charge by the number of pages in the
9119 : * range.
9120 : */
9121 10730 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9122 10730 : statsData.pagesPerRange;
9123 :
9124 10730 : *indexPages = index->pages;
9125 10730 : }
|
