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) 1996-2026, 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 Node *strip_all_phvs_deep(PlannerInfo *root, Node *node);
246 : static bool contain_placeholder_walker(Node *node, void *context);
247 : static Node *strip_all_phvs_mutator(Node *node, void *context);
248 : static void examine_simple_variable(PlannerInfo *root, Var *var,
249 : VariableStatData *vardata);
250 : static void examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
251 : int indexcol, VariableStatData *vardata);
252 : static bool get_variable_range(PlannerInfo *root, VariableStatData *vardata,
253 : Oid sortop, Oid collation,
254 : Datum *min, Datum *max);
255 : static void get_stats_slot_range(AttStatsSlot *sslot,
256 : Oid opfuncoid, FmgrInfo *opproc,
257 : Oid collation, int16 typLen, bool typByVal,
258 : Datum *min, Datum *max, bool *p_have_data);
259 : static bool get_actual_variable_range(PlannerInfo *root,
260 : VariableStatData *vardata,
261 : Oid sortop, Oid collation,
262 : Datum *min, Datum *max);
263 : static bool get_actual_variable_endpoint(Relation heapRel,
264 : Relation indexRel,
265 : ScanDirection indexscandir,
266 : ScanKey scankeys,
267 : int16 typLen,
268 : bool typByVal,
269 : TupleTableSlot *tableslot,
270 : MemoryContext outercontext,
271 : Datum *endpointDatum);
272 : static RelOptInfo *find_join_input_rel(PlannerInfo *root, Relids relids);
273 : static double btcost_correlation(IndexOptInfo *index,
274 : VariableStatData *vardata);
275 :
276 : /* Define support routines for MCV hash tables */
277 : #define SH_PREFIX MCVHashTable
278 : #define SH_ELEMENT_TYPE MCVHashEntry
279 : #define SH_KEY_TYPE Datum
280 : #define SH_KEY value
281 : #define SH_HASH_KEY(tab,key) hash_mcv(tab, key)
282 : #define SH_EQUAL(tab,key0,key1) mcvs_equal(tab, key0, key1)
283 : #define SH_SCOPE static inline
284 : #define SH_STORE_HASH
285 : #define SH_GET_HASH(tab,ent) (ent)->hash
286 : #define SH_DEFINE
287 : #define SH_DECLARE
288 : #include "lib/simplehash.h"
289 :
290 :
291 : /*
292 : * eqsel - Selectivity of "=" for any data types.
293 : *
294 : * Note: this routine is also used to estimate selectivity for some
295 : * operators that are not "=" but have comparable selectivity behavior,
296 : * such as "~=" (geometric approximate-match). Even for "=", we must
297 : * keep in mind that the left and right datatypes may differ.
298 : */
299 : Datum
300 411510 : eqsel(PG_FUNCTION_ARGS)
301 : {
302 411510 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, false));
303 : }
304 :
305 : /*
306 : * Common code for eqsel() and neqsel()
307 : */
308 : static double
309 436995 : eqsel_internal(PG_FUNCTION_ARGS, bool negate)
310 : {
311 436995 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
312 436995 : Oid operator = PG_GETARG_OID(1);
313 436995 : List *args = (List *) PG_GETARG_POINTER(2);
314 436995 : int varRelid = PG_GETARG_INT32(3);
315 436995 : Oid collation = PG_GET_COLLATION();
316 : VariableStatData vardata;
317 : Node *other;
318 : bool varonleft;
319 : double selec;
320 :
321 : /*
322 : * When asked about <>, we do the estimation using the corresponding =
323 : * operator, then convert to <> via "1.0 - eq_selectivity - nullfrac".
324 : */
325 436995 : if (negate)
326 : {
327 25485 : operator = get_negator(operator);
328 25485 : if (!OidIsValid(operator))
329 : {
330 : /* Use default selectivity (should we raise an error instead?) */
331 0 : return 1.0 - DEFAULT_EQ_SEL;
332 : }
333 : }
334 :
335 : /*
336 : * If expression is not variable = something or something = variable, then
337 : * punt and return a default estimate.
338 : */
339 436995 : if (!get_restriction_variable(root, args, varRelid,
340 : &vardata, &other, &varonleft))
341 2842 : return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;
342 :
343 : /*
344 : * We can do a lot better if the something is a constant. (Note: the
345 : * Const might result from estimation rather than being a simple constant
346 : * in the query.)
347 : */
348 434150 : if (IsA(other, Const))
349 171066 : selec = var_eq_const(&vardata, operator, collation,
350 171066 : ((Const *) other)->constvalue,
351 171066 : ((Const *) other)->constisnull,
352 : varonleft, negate);
353 : else
354 263084 : selec = var_eq_non_const(&vardata, operator, collation, other,
355 : varonleft, negate);
356 :
357 434150 : ReleaseVariableStats(vardata);
358 :
359 434150 : return selec;
360 : }
361 :
362 : /*
363 : * var_eq_const --- eqsel for var = const case
364 : *
365 : * This is exported so that some other estimation functions can use it.
366 : */
367 : double
368 194867 : var_eq_const(VariableStatData *vardata, Oid oproid, Oid collation,
369 : Datum constval, bool constisnull,
370 : bool varonleft, bool negate)
371 : {
372 : double selec;
373 194867 : double nullfrac = 0.0;
374 : bool isdefault;
375 : Oid opfuncoid;
376 :
377 : /*
378 : * If the constant is NULL, assume operator is strict and return zero, ie,
379 : * operator will never return TRUE. (It's zero even for a negator op.)
380 : */
381 194867 : if (constisnull)
382 212 : return 0.0;
383 :
384 : /*
385 : * Grab the nullfrac for use below. Note we allow use of nullfrac
386 : * regardless of security check.
387 : */
388 194655 : if (HeapTupleIsValid(vardata->statsTuple))
389 : {
390 : Form_pg_statistic stats;
391 :
392 141039 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
393 141039 : nullfrac = stats->stanullfrac;
394 : }
395 :
396 : /*
397 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
398 : * assume there is exactly one match regardless of anything else. (This
399 : * is slightly bogus, since the index or clause's equality operator might
400 : * be different from ours, but it's much more likely to be right than
401 : * ignoring the information.)
402 : */
403 194655 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
404 : {
405 44121 : selec = 1.0 / vardata->rel->tuples;
406 : }
407 256635 : else if (HeapTupleIsValid(vardata->statsTuple) &&
408 106101 : statistic_proc_security_check(vardata,
409 106101 : (opfuncoid = get_opcode(oproid))))
410 106101 : {
411 : AttStatsSlot sslot;
412 106101 : bool match = false;
413 : int i;
414 :
415 : /*
416 : * Is the constant "=" to any of the column's most common values?
417 : * (Although the given operator may not really be "=", we will assume
418 : * that seeing whether it returns TRUE is an appropriate test. If you
419 : * don't like this, maybe you shouldn't be using eqsel for your
420 : * operator...)
421 : */
422 106101 : if (get_attstatsslot(&sslot, vardata->statsTuple,
423 : STATISTIC_KIND_MCV, InvalidOid,
424 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
425 : {
426 94198 : LOCAL_FCINFO(fcinfo, 2);
427 : FmgrInfo eqproc;
428 :
429 94198 : fmgr_info(opfuncoid, &eqproc);
430 :
431 : /*
432 : * Save a few cycles by setting up the fcinfo struct just once.
433 : * Using FunctionCallInvoke directly also avoids failure if the
434 : * eqproc returns NULL, though really equality functions should
435 : * never do that.
436 : */
437 94198 : InitFunctionCallInfoData(*fcinfo, &eqproc, 2, collation,
438 : NULL, NULL);
439 94198 : fcinfo->args[0].isnull = false;
440 94198 : fcinfo->args[1].isnull = false;
441 : /* be careful to apply operator right way 'round */
442 94198 : if (varonleft)
443 94182 : fcinfo->args[1].value = constval;
444 : else
445 16 : fcinfo->args[0].value = constval;
446 :
447 1481116 : for (i = 0; i < sslot.nvalues; i++)
448 : {
449 : Datum fresult;
450 :
451 1436165 : if (varonleft)
452 1436137 : fcinfo->args[0].value = sslot.values[i];
453 : else
454 28 : fcinfo->args[1].value = sslot.values[i];
455 1436165 : fcinfo->isnull = false;
456 1436165 : fresult = FunctionCallInvoke(fcinfo);
457 1436165 : if (!fcinfo->isnull && DatumGetBool(fresult))
458 : {
459 49247 : match = true;
460 49247 : break;
461 : }
462 : }
463 : }
464 : else
465 : {
466 : /* no most-common-value info available */
467 11903 : i = 0; /* keep compiler quiet */
468 : }
469 :
470 106101 : if (match)
471 : {
472 : /*
473 : * Constant is "=" to this common value. We know selectivity
474 : * exactly (or as exactly as ANALYZE could calculate it, anyway).
475 : */
476 49247 : selec = sslot.numbers[i];
477 : }
478 : else
479 : {
480 : /*
481 : * Comparison is against a constant that is neither NULL nor any
482 : * of the common values. Its selectivity cannot be more than
483 : * this:
484 : */
485 56854 : double sumcommon = 0.0;
486 : double otherdistinct;
487 :
488 1253988 : for (i = 0; i < sslot.nnumbers; i++)
489 1197134 : sumcommon += sslot.numbers[i];
490 56854 : selec = 1.0 - sumcommon - nullfrac;
491 56854 : CLAMP_PROBABILITY(selec);
492 :
493 : /*
494 : * and in fact it's probably a good deal less. We approximate that
495 : * all the not-common values share this remaining fraction
496 : * equally, so we divide by the number of other distinct values.
497 : */
498 56854 : otherdistinct = get_variable_numdistinct(vardata, &isdefault) -
499 56854 : sslot.nnumbers;
500 56854 : if (otherdistinct > 1)
501 26698 : selec /= otherdistinct;
502 :
503 : /*
504 : * Another cross-check: selectivity shouldn't be estimated as more
505 : * than the least common "most common value".
506 : */
507 56854 : if (sslot.nnumbers > 0 && selec > sslot.numbers[sslot.nnumbers - 1])
508 0 : selec = sslot.numbers[sslot.nnumbers - 1];
509 : }
510 :
511 106101 : free_attstatsslot(&sslot);
512 : }
513 : else
514 : {
515 : /*
516 : * No ANALYZE stats available, so make a guess using estimated number
517 : * of distinct values and assuming they are equally common. (The guess
518 : * is unlikely to be very good, but we do know a few special cases.)
519 : */
520 44433 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
521 : }
522 :
523 : /* now adjust if we wanted <> rather than = */
524 194655 : if (negate)
525 20421 : selec = 1.0 - selec - nullfrac;
526 :
527 : /* result should be in range, but make sure... */
528 194655 : CLAMP_PROBABILITY(selec);
529 :
530 194655 : return selec;
531 : }
532 :
533 : /*
534 : * var_eq_non_const --- eqsel for var = something-other-than-const case
535 : *
536 : * This is exported so that some other estimation functions can use it.
537 : */
538 : double
539 263084 : var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,
540 : Node *other,
541 : bool varonleft, bool negate)
542 : {
543 : double selec;
544 263084 : double nullfrac = 0.0;
545 : bool isdefault;
546 :
547 : /*
548 : * Grab the nullfrac for use below.
549 : */
550 263084 : if (HeapTupleIsValid(vardata->statsTuple))
551 : {
552 : Form_pg_statistic stats;
553 :
554 161650 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
555 161650 : nullfrac = stats->stanullfrac;
556 : }
557 :
558 : /*
559 : * If we matched the var to a unique index, DISTINCT or GROUP-BY clause,
560 : * assume there is exactly one match regardless of anything else. (This
561 : * is slightly bogus, since the index or clause's equality operator might
562 : * be different from ours, but it's much more likely to be right than
563 : * ignoring the information.)
564 : */
565 263084 : if (vardata->isunique && vardata->rel && vardata->rel->tuples >= 1.0)
566 : {
567 92644 : selec = 1.0 / vardata->rel->tuples;
568 : }
569 170440 : else if (HeapTupleIsValid(vardata->statsTuple))
570 : {
571 : double ndistinct;
572 : AttStatsSlot sslot;
573 :
574 : /*
575 : * Search is for a value that we do not know a priori, but we will
576 : * assume it is not NULL. Estimate the selectivity as non-null
577 : * fraction divided by number of distinct values, so that we get a
578 : * result averaged over all possible values whether common or
579 : * uncommon. (Essentially, we are assuming that the not-yet-known
580 : * comparison value is equally likely to be any of the possible
581 : * values, regardless of their frequency in the table. Is that a good
582 : * idea?)
583 : */
584 80498 : selec = 1.0 - nullfrac;
585 80498 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
586 80498 : if (ndistinct > 1)
587 78460 : selec /= ndistinct;
588 :
589 : /*
590 : * Cross-check: selectivity should never be estimated as more than the
591 : * most common value's.
592 : */
593 80498 : if (get_attstatsslot(&sslot, vardata->statsTuple,
594 : STATISTIC_KIND_MCV, InvalidOid,
595 : ATTSTATSSLOT_NUMBERS))
596 : {
597 70290 : if (sslot.nnumbers > 0 && selec > sslot.numbers[0])
598 291 : selec = sslot.numbers[0];
599 70290 : free_attstatsslot(&sslot);
600 : }
601 : }
602 : else
603 : {
604 : /*
605 : * No ANALYZE stats available, so make a guess using estimated number
606 : * of distinct values and assuming they are equally common. (The guess
607 : * is unlikely to be very good, but we do know a few special cases.)
608 : */
609 89942 : selec = 1.0 / get_variable_numdistinct(vardata, &isdefault);
610 : }
611 :
612 : /* now adjust if we wanted <> rather than = */
613 263084 : if (negate)
614 3727 : selec = 1.0 - selec - nullfrac;
615 :
616 : /* result should be in range, but make sure... */
617 263084 : CLAMP_PROBABILITY(selec);
618 :
619 263084 : return selec;
620 : }
621 :
622 : /*
623 : * neqsel - Selectivity of "!=" for any data types.
624 : *
625 : * This routine is also used for some operators that are not "!="
626 : * but have comparable selectivity behavior. See above comments
627 : * for eqsel().
628 : */
629 : Datum
630 25485 : neqsel(PG_FUNCTION_ARGS)
631 : {
632 25485 : PG_RETURN_FLOAT8((float8) eqsel_internal(fcinfo, true));
633 : }
634 :
635 : /*
636 : * scalarineqsel - Selectivity of "<", "<=", ">", ">=" for scalars.
637 : *
638 : * This is the guts of scalarltsel/scalarlesel/scalargtsel/scalargesel.
639 : * The isgt and iseq flags distinguish which of the four cases apply.
640 : *
641 : * The caller has commuted the clause, if necessary, so that we can treat
642 : * the variable as being on the left. The caller must also make sure that
643 : * the other side of the clause is a non-null Const, and dissect that into
644 : * a value and datatype. (This definition simplifies some callers that
645 : * want to estimate against a computed value instead of a Const node.)
646 : *
647 : * This routine works for any datatype (or pair of datatypes) known to
648 : * convert_to_scalar(). If it is applied to some other datatype,
649 : * it will return an approximate estimate based on assuming that the constant
650 : * value falls in the middle of the bin identified by binary search.
651 : */
652 : static double
653 200678 : scalarineqsel(PlannerInfo *root, Oid operator, bool isgt, bool iseq,
654 : Oid collation,
655 : VariableStatData *vardata, Datum constval, Oid consttype)
656 : {
657 : Form_pg_statistic stats;
658 : FmgrInfo opproc;
659 : double mcv_selec,
660 : hist_selec,
661 : sumcommon;
662 : double selec;
663 :
664 200678 : if (!HeapTupleIsValid(vardata->statsTuple))
665 : {
666 : /*
667 : * No stats are available. Typically this means we have to fall back
668 : * on the default estimate; but if the variable is CTID then we can
669 : * make an estimate based on comparing the constant to the table size.
670 : */
671 14322 : if (vardata->var && IsA(vardata->var, Var) &&
672 11838 : ((Var *) vardata->var)->varattno == SelfItemPointerAttributeNumber)
673 : {
674 : ItemPointer itemptr;
675 : double block;
676 : double density;
677 :
678 : /*
679 : * If the relation's empty, we're going to include all of it.
680 : * (This is mostly to avoid divide-by-zero below.)
681 : */
682 1010 : if (vardata->rel->pages == 0)
683 0 : return 1.0;
684 :
685 1010 : itemptr = (ItemPointer) DatumGetPointer(constval);
686 1010 : block = ItemPointerGetBlockNumberNoCheck(itemptr);
687 :
688 : /*
689 : * Determine the average number of tuples per page (density).
690 : *
691 : * Since the last page will, on average, be only half full, we can
692 : * estimate it to have half as many tuples as earlier pages. So
693 : * give it half the weight of a regular page.
694 : */
695 1010 : density = vardata->rel->tuples / (vardata->rel->pages - 0.5);
696 :
697 : /* If target is the last page, use half the density. */
698 1010 : if (block >= vardata->rel->pages - 1)
699 15 : density *= 0.5;
700 :
701 : /*
702 : * Using the average tuples per page, calculate how far into the
703 : * page the itemptr is likely to be and adjust block accordingly,
704 : * by adding that fraction of a whole block (but never more than a
705 : * whole block, no matter how high the itemptr's offset is). Here
706 : * we are ignoring the possibility of dead-tuple line pointers,
707 : * which is fairly bogus, but we lack the info to do better.
708 : */
709 1010 : if (density > 0.0)
710 : {
711 1010 : OffsetNumber offset = ItemPointerGetOffsetNumberNoCheck(itemptr);
712 :
713 1010 : block += Min(offset / density, 1.0);
714 : }
715 :
716 : /*
717 : * Convert relative block number to selectivity. Again, the last
718 : * page has only half weight.
719 : */
720 1010 : selec = block / (vardata->rel->pages - 0.5);
721 :
722 : /*
723 : * The calculation so far gave us a selectivity for the "<=" case.
724 : * We'll have one fewer tuple for "<" and one additional tuple for
725 : * ">=", the latter of which we'll reverse the selectivity for
726 : * below, so we can simply subtract one tuple for both cases. The
727 : * cases that need this adjustment can be identified by iseq being
728 : * equal to isgt.
729 : */
730 1010 : if (iseq == isgt && vardata->rel->tuples >= 1.0)
731 940 : selec -= (1.0 / vardata->rel->tuples);
732 :
733 : /* Finally, reverse the selectivity for the ">", ">=" cases. */
734 1010 : if (isgt)
735 931 : selec = 1.0 - selec;
736 :
737 1010 : CLAMP_PROBABILITY(selec);
738 1010 : return selec;
739 : }
740 :
741 : /* no stats available, so default result */
742 13312 : return DEFAULT_INEQ_SEL;
743 : }
744 186356 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
745 :
746 186356 : fmgr_info(get_opcode(operator), &opproc);
747 :
748 : /*
749 : * If we have most-common-values info, add up the fractions of the MCV
750 : * entries that satisfy MCV OP CONST. These fractions contribute directly
751 : * to the result selectivity. Also add up the total fraction represented
752 : * by MCV entries.
753 : */
754 186356 : mcv_selec = mcv_selectivity(vardata, &opproc, collation, constval, true,
755 : &sumcommon);
756 :
757 : /*
758 : * If there is a histogram, determine which bin the constant falls in, and
759 : * compute the resulting contribution to selectivity.
760 : */
761 186356 : hist_selec = ineq_histogram_selectivity(root, vardata,
762 : operator, &opproc, isgt, iseq,
763 : collation,
764 : constval, consttype);
765 :
766 : /*
767 : * Now merge the results from the MCV and histogram calculations,
768 : * realizing that the histogram covers only the non-null values that are
769 : * not listed in MCV.
770 : */
771 186356 : selec = 1.0 - stats->stanullfrac - sumcommon;
772 :
773 186356 : if (hist_selec >= 0.0)
774 117714 : selec *= hist_selec;
775 : else
776 : {
777 : /*
778 : * If no histogram but there are values not accounted for by MCV,
779 : * arbitrarily assume half of them will match.
780 : */
781 68642 : selec *= 0.5;
782 : }
783 :
784 186356 : selec += mcv_selec;
785 :
786 : /* result should be in range, but make sure... */
787 186356 : CLAMP_PROBABILITY(selec);
788 :
789 186356 : return selec;
790 : }
791 :
792 : /*
793 : * mcv_selectivity - Examine the MCV list for selectivity estimates
794 : *
795 : * Determine the fraction of the variable's MCV population that satisfies
796 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft. Also
797 : * compute the fraction of the total column population represented by the MCV
798 : * list. This code will work for any boolean-returning predicate operator.
799 : *
800 : * The function result is the MCV selectivity, and the fraction of the
801 : * total population is returned into *sumcommonp. Zeroes are returned
802 : * if there is no MCV list.
803 : */
804 : double
805 189548 : mcv_selectivity(VariableStatData *vardata, FmgrInfo *opproc, Oid collation,
806 : Datum constval, bool varonleft,
807 : double *sumcommonp)
808 : {
809 : double mcv_selec,
810 : sumcommon;
811 : AttStatsSlot sslot;
812 : int i;
813 :
814 189548 : mcv_selec = 0.0;
815 189548 : sumcommon = 0.0;
816 :
817 377870 : if (HeapTupleIsValid(vardata->statsTuple) &&
818 376479 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
819 188157 : get_attstatsslot(&sslot, vardata->statsTuple,
820 : STATISTIC_KIND_MCV, InvalidOid,
821 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS))
822 : {
823 103139 : LOCAL_FCINFO(fcinfo, 2);
824 :
825 : /*
826 : * We invoke the opproc "by hand" so that we won't fail on NULL
827 : * results. Such cases won't arise for normal comparison functions,
828 : * but generic_restriction_selectivity could perhaps be used with
829 : * operators that can return NULL. A small side benefit is to not
830 : * need to re-initialize the fcinfo struct from scratch each time.
831 : */
832 103139 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
833 : NULL, NULL);
834 103139 : fcinfo->args[0].isnull = false;
835 103139 : fcinfo->args[1].isnull = false;
836 : /* be careful to apply operator right way 'round */
837 103139 : if (varonleft)
838 103139 : fcinfo->args[1].value = constval;
839 : else
840 0 : fcinfo->args[0].value = constval;
841 :
842 2522823 : for (i = 0; i < sslot.nvalues; i++)
843 : {
844 : Datum fresult;
845 :
846 2419684 : if (varonleft)
847 2419684 : fcinfo->args[0].value = sslot.values[i];
848 : else
849 0 : fcinfo->args[1].value = sslot.values[i];
850 2419684 : fcinfo->isnull = false;
851 2419684 : fresult = FunctionCallInvoke(fcinfo);
852 2419684 : if (!fcinfo->isnull && DatumGetBool(fresult))
853 932846 : mcv_selec += sslot.numbers[i];
854 2419684 : sumcommon += sslot.numbers[i];
855 : }
856 103139 : free_attstatsslot(&sslot);
857 : }
858 :
859 189548 : *sumcommonp = sumcommon;
860 189548 : return mcv_selec;
861 : }
862 :
863 : /*
864 : * histogram_selectivity - Examine the histogram for selectivity estimates
865 : *
866 : * Determine the fraction of the variable's histogram entries that satisfy
867 : * the predicate (VAR OP CONST), or (CONST OP VAR) if !varonleft.
868 : *
869 : * This code will work for any boolean-returning predicate operator, whether
870 : * or not it has anything to do with the histogram sort operator. We are
871 : * essentially using the histogram just as a representative sample. However,
872 : * small histograms are unlikely to be all that representative, so the caller
873 : * should be prepared to fall back on some other estimation approach when the
874 : * histogram is missing or very small. It may also be prudent to combine this
875 : * approach with another one when the histogram is small.
876 : *
877 : * If the actual histogram size is not at least min_hist_size, we won't bother
878 : * to do the calculation at all. Also, if the n_skip parameter is > 0, we
879 : * ignore the first and last n_skip histogram elements, on the grounds that
880 : * they are outliers and hence not very representative. Typical values for
881 : * these parameters are 10 and 1.
882 : *
883 : * The function result is the selectivity, or -1 if there is no histogram
884 : * or it's smaller than min_hist_size.
885 : *
886 : * The output parameter *hist_size receives the actual histogram size,
887 : * or zero if no histogram. Callers may use this number to decide how
888 : * much faith to put in the function result.
889 : *
890 : * Note that the result disregards both the most-common-values (if any) and
891 : * null entries. The caller is expected to combine this result with
892 : * statistics for those portions of the column population. It may also be
893 : * prudent to clamp the result range, ie, disbelieve exact 0 or 1 outputs.
894 : */
895 : double
896 3192 : histogram_selectivity(VariableStatData *vardata,
897 : FmgrInfo *opproc, Oid collation,
898 : Datum constval, bool varonleft,
899 : int min_hist_size, int n_skip,
900 : int *hist_size)
901 : {
902 : double result;
903 : AttStatsSlot sslot;
904 :
905 : /* check sanity of parameters */
906 : Assert(n_skip >= 0);
907 : Assert(min_hist_size > 2 * n_skip);
908 :
909 5158 : if (HeapTupleIsValid(vardata->statsTuple) &&
910 3929 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
911 1963 : get_attstatsslot(&sslot, vardata->statsTuple,
912 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
913 : ATTSTATSSLOT_VALUES))
914 : {
915 1915 : *hist_size = sslot.nvalues;
916 1915 : if (sslot.nvalues >= min_hist_size)
917 : {
918 892 : LOCAL_FCINFO(fcinfo, 2);
919 892 : int nmatch = 0;
920 : int i;
921 :
922 : /*
923 : * We invoke the opproc "by hand" so that we won't fail on NULL
924 : * results. Such cases won't arise for normal comparison
925 : * functions, but generic_restriction_selectivity could perhaps be
926 : * used with operators that can return NULL. A small side benefit
927 : * is to not need to re-initialize the fcinfo struct from scratch
928 : * each time.
929 : */
930 892 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
931 : NULL, NULL);
932 892 : fcinfo->args[0].isnull = false;
933 892 : fcinfo->args[1].isnull = false;
934 : /* be careful to apply operator right way 'round */
935 892 : if (varonleft)
936 892 : fcinfo->args[1].value = constval;
937 : else
938 0 : fcinfo->args[0].value = constval;
939 :
940 70904 : for (i = n_skip; i < sslot.nvalues - n_skip; i++)
941 : {
942 : Datum fresult;
943 :
944 70012 : if (varonleft)
945 70012 : fcinfo->args[0].value = sslot.values[i];
946 : else
947 0 : fcinfo->args[1].value = sslot.values[i];
948 70012 : fcinfo->isnull = false;
949 70012 : fresult = FunctionCallInvoke(fcinfo);
950 70012 : if (!fcinfo->isnull && DatumGetBool(fresult))
951 3974 : nmatch++;
952 : }
953 892 : result = ((double) nmatch) / ((double) (sslot.nvalues - 2 * n_skip));
954 : }
955 : else
956 1023 : result = -1;
957 1915 : free_attstatsslot(&sslot);
958 : }
959 : else
960 : {
961 1277 : *hist_size = 0;
962 1277 : result = -1;
963 : }
964 :
965 3192 : return result;
966 : }
967 :
968 : /*
969 : * generic_restriction_selectivity - Selectivity for almost anything
970 : *
971 : * This function estimates selectivity for operators that we don't have any
972 : * special knowledge about, but are on data types that we collect standard
973 : * MCV and/or histogram statistics for. (Additional assumptions are that
974 : * the operator is strict and immutable, or at least stable.)
975 : *
976 : * If we have "VAR OP CONST" or "CONST OP VAR", selectivity is estimated by
977 : * applying the operator to each element of the column's MCV and/or histogram
978 : * stats, and merging the results using the assumption that the histogram is
979 : * a reasonable random sample of the column's non-MCV population. Note that
980 : * if the operator's semantics are related to the histogram ordering, this
981 : * might not be such a great assumption; other functions such as
982 : * scalarineqsel() are probably a better match in such cases.
983 : *
984 : * Otherwise, fall back to the default selectivity provided by the caller.
985 : */
986 : double
987 565 : generic_restriction_selectivity(PlannerInfo *root, Oid oproid, Oid collation,
988 : List *args, int varRelid,
989 : double default_selectivity)
990 : {
991 : double selec;
992 : VariableStatData vardata;
993 : Node *other;
994 : bool varonleft;
995 :
996 : /*
997 : * If expression is not variable OP something or something OP variable,
998 : * then punt and return the default estimate.
999 : */
1000 565 : if (!get_restriction_variable(root, args, varRelid,
1001 : &vardata, &other, &varonleft))
1002 0 : return default_selectivity;
1003 :
1004 : /*
1005 : * If the something is a NULL constant, assume operator is strict and
1006 : * return zero, ie, operator will never return TRUE.
1007 : */
1008 565 : if (IsA(other, Const) &&
1009 565 : ((Const *) other)->constisnull)
1010 : {
1011 0 : ReleaseVariableStats(vardata);
1012 0 : return 0.0;
1013 : }
1014 :
1015 565 : if (IsA(other, Const))
1016 : {
1017 : /* Variable is being compared to a known non-null constant */
1018 565 : Datum constval = ((Const *) other)->constvalue;
1019 : FmgrInfo opproc;
1020 : double mcvsum;
1021 : double mcvsel;
1022 : double nullfrac;
1023 : int hist_size;
1024 :
1025 565 : fmgr_info(get_opcode(oproid), &opproc);
1026 :
1027 : /*
1028 : * Calculate the selectivity for the column's most common values.
1029 : */
1030 565 : mcvsel = mcv_selectivity(&vardata, &opproc, collation,
1031 : constval, varonleft,
1032 : &mcvsum);
1033 :
1034 : /*
1035 : * If the histogram is large enough, see what fraction of it matches
1036 : * the query, and assume that's representative of the non-MCV
1037 : * population. Otherwise use the default selectivity for the non-MCV
1038 : * population.
1039 : */
1040 565 : selec = histogram_selectivity(&vardata, &opproc, collation,
1041 : constval, varonleft,
1042 : 10, 1, &hist_size);
1043 565 : if (selec < 0)
1044 : {
1045 : /* Nope, fall back on default */
1046 565 : selec = default_selectivity;
1047 : }
1048 0 : else if (hist_size < 100)
1049 : {
1050 : /*
1051 : * For histogram sizes from 10 to 100, we combine the histogram
1052 : * and default selectivities, putting increasingly more trust in
1053 : * the histogram for larger sizes.
1054 : */
1055 0 : double hist_weight = hist_size / 100.0;
1056 :
1057 0 : selec = selec * hist_weight +
1058 0 : default_selectivity * (1.0 - hist_weight);
1059 : }
1060 :
1061 : /* In any case, don't believe extremely small or large estimates. */
1062 565 : if (selec < 0.0001)
1063 0 : selec = 0.0001;
1064 565 : else if (selec > 0.9999)
1065 0 : selec = 0.9999;
1066 :
1067 : /* Don't forget to account for nulls. */
1068 565 : if (HeapTupleIsValid(vardata.statsTuple))
1069 42 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata.statsTuple))->stanullfrac;
1070 : else
1071 523 : nullfrac = 0.0;
1072 :
1073 : /*
1074 : * Now merge the results from the MCV and histogram calculations,
1075 : * realizing that the histogram covers only the non-null values that
1076 : * are not listed in MCV.
1077 : */
1078 565 : selec *= 1.0 - nullfrac - mcvsum;
1079 565 : selec += mcvsel;
1080 : }
1081 : else
1082 : {
1083 : /* Comparison value is not constant, so we can't do anything */
1084 0 : selec = default_selectivity;
1085 : }
1086 :
1087 565 : ReleaseVariableStats(vardata);
1088 :
1089 : /* result should be in range, but make sure... */
1090 565 : CLAMP_PROBABILITY(selec);
1091 :
1092 565 : return selec;
1093 : }
1094 :
1095 : /*
1096 : * ineq_histogram_selectivity - Examine the histogram for scalarineqsel
1097 : *
1098 : * Determine the fraction of the variable's histogram population that
1099 : * satisfies the inequality condition, ie, VAR < (or <=, >, >=) CONST.
1100 : * The isgt and iseq flags distinguish which of the four cases apply.
1101 : *
1102 : * While opproc could be looked up from the operator OID, common callers
1103 : * also need to call it separately, so we make the caller pass both.
1104 : *
1105 : * Returns -1 if there is no histogram (valid results will always be >= 0).
1106 : *
1107 : * Note that the result disregards both the most-common-values (if any) and
1108 : * null entries. The caller is expected to combine this result with
1109 : * statistics for those portions of the column population.
1110 : *
1111 : * This is exported so that some other estimation functions can use it.
1112 : */
1113 : double
1114 189157 : ineq_histogram_selectivity(PlannerInfo *root,
1115 : VariableStatData *vardata,
1116 : Oid opoid, FmgrInfo *opproc, bool isgt, bool iseq,
1117 : Oid collation,
1118 : Datum constval, Oid consttype)
1119 : {
1120 : double hist_selec;
1121 : AttStatsSlot sslot;
1122 :
1123 189157 : hist_selec = -1.0;
1124 :
1125 : /*
1126 : * Someday, ANALYZE might store more than one histogram per rel/att,
1127 : * corresponding to more than one possible sort ordering defined for the
1128 : * column type. Right now, we know there is only one, so just grab it and
1129 : * see if it matches the query.
1130 : *
1131 : * Note that we can't use opoid as search argument; the staop appearing in
1132 : * pg_statistic will be for the relevant '<' operator, but what we have
1133 : * might be some other inequality operator such as '>='. (Even if opoid
1134 : * is a '<' operator, it could be cross-type.) Hence we must use
1135 : * comparison_ops_are_compatible() to see if the operators match.
1136 : */
1137 377973 : if (HeapTupleIsValid(vardata->statsTuple) &&
1138 377470 : statistic_proc_security_check(vardata, opproc->fn_oid) &&
1139 188654 : get_attstatsslot(&sslot, vardata->statsTuple,
1140 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
1141 : ATTSTATSSLOT_VALUES))
1142 : {
1143 120172 : if (sslot.nvalues > 1 &&
1144 240306 : sslot.stacoll == collation &&
1145 120134 : comparison_ops_are_compatible(sslot.staop, opoid))
1146 120080 : {
1147 : /*
1148 : * Use binary search to find the desired location, namely the
1149 : * right end of the histogram bin containing the comparison value,
1150 : * which is the leftmost entry for which the comparison operator
1151 : * succeeds (if isgt) or fails (if !isgt).
1152 : *
1153 : * In this loop, we pay no attention to whether the operator iseq
1154 : * or not; that detail will be mopped up below. (We cannot tell,
1155 : * anyway, whether the operator thinks the values are equal.)
1156 : *
1157 : * If the binary search accesses the first or last histogram
1158 : * entry, we try to replace that endpoint with the true column min
1159 : * or max as found by get_actual_variable_range(). This
1160 : * ameliorates misestimates when the min or max is moving as a
1161 : * result of changes since the last ANALYZE. Note that this could
1162 : * result in effectively including MCVs into the histogram that
1163 : * weren't there before, but we don't try to correct for that.
1164 : */
1165 : double histfrac;
1166 120080 : int lobound = 0; /* first possible slot to search */
1167 120080 : int hibound = sslot.nvalues; /* last+1 slot to search */
1168 120080 : bool have_end = false;
1169 :
1170 : /*
1171 : * If there are only two histogram entries, we'll want up-to-date
1172 : * values for both. (If there are more than two, we need at most
1173 : * one of them to be updated, so we deal with that within the
1174 : * loop.)
1175 : */
1176 120080 : if (sslot.nvalues == 2)
1177 2535 : have_end = get_actual_variable_range(root,
1178 : vardata,
1179 : sslot.staop,
1180 : collation,
1181 : &sslot.values[0],
1182 2535 : &sslot.values[1]);
1183 :
1184 783749 : while (lobound < hibound)
1185 : {
1186 663669 : int probe = (lobound + hibound) / 2;
1187 : bool ltcmp;
1188 :
1189 : /*
1190 : * If we find ourselves about to compare to the first or last
1191 : * histogram entry, first try to replace it with the actual
1192 : * current min or max (unless we already did so above).
1193 : */
1194 663669 : if (probe == 0 && sslot.nvalues > 2)
1195 58487 : have_end = get_actual_variable_range(root,
1196 : vardata,
1197 : sslot.staop,
1198 : collation,
1199 : &sslot.values[0],
1200 : NULL);
1201 605182 : else if (probe == sslot.nvalues - 1 && sslot.nvalues > 2)
1202 40723 : have_end = get_actual_variable_range(root,
1203 : vardata,
1204 : sslot.staop,
1205 : collation,
1206 : NULL,
1207 40723 : &sslot.values[probe]);
1208 :
1209 663669 : ltcmp = DatumGetBool(FunctionCall2Coll(opproc,
1210 : collation,
1211 663669 : sslot.values[probe],
1212 : constval));
1213 663669 : if (isgt)
1214 35638 : ltcmp = !ltcmp;
1215 663669 : if (ltcmp)
1216 253030 : lobound = probe + 1;
1217 : else
1218 410639 : hibound = probe;
1219 : }
1220 :
1221 120080 : if (lobound <= 0)
1222 : {
1223 : /*
1224 : * Constant is below lower histogram boundary. More
1225 : * precisely, we have found that no entry in the histogram
1226 : * satisfies the inequality clause (if !isgt) or they all do
1227 : * (if isgt). We estimate that that's true of the entire
1228 : * table, so set histfrac to 0.0 (which we'll flip to 1.0
1229 : * below, if isgt).
1230 : */
1231 51370 : histfrac = 0.0;
1232 : }
1233 68710 : else if (lobound >= sslot.nvalues)
1234 : {
1235 : /*
1236 : * Inverse case: constant is above upper histogram boundary.
1237 : */
1238 19740 : histfrac = 1.0;
1239 : }
1240 : else
1241 : {
1242 : /* We have values[i-1] <= constant <= values[i]. */
1243 48970 : int i = lobound;
1244 48970 : double eq_selec = 0;
1245 : double val,
1246 : high,
1247 : low;
1248 : double binfrac;
1249 :
1250 : /*
1251 : * In the cases where we'll need it below, obtain an estimate
1252 : * of the selectivity of "x = constval". We use a calculation
1253 : * similar to what var_eq_const() does for a non-MCV constant,
1254 : * ie, estimate that all distinct non-MCV values occur equally
1255 : * often. But multiplication by "1.0 - sumcommon - nullfrac"
1256 : * will be done by our caller, so we shouldn't do that here.
1257 : * Therefore we can't try to clamp the estimate by reference
1258 : * to the least common MCV; the result would be too small.
1259 : *
1260 : * Note: since this is effectively assuming that constval
1261 : * isn't an MCV, it's logically dubious if constval in fact is
1262 : * one. But we have to apply *some* correction for equality,
1263 : * and anyway we cannot tell if constval is an MCV, since we
1264 : * don't have a suitable equality operator at hand.
1265 : */
1266 48970 : if (i == 1 || isgt == iseq)
1267 : {
1268 : double otherdistinct;
1269 : bool isdefault;
1270 : AttStatsSlot mcvslot;
1271 :
1272 : /* Get estimated number of distinct values */
1273 20581 : otherdistinct = get_variable_numdistinct(vardata,
1274 : &isdefault);
1275 :
1276 : /* Subtract off the number of known MCVs */
1277 20581 : if (get_attstatsslot(&mcvslot, vardata->statsTuple,
1278 : STATISTIC_KIND_MCV, InvalidOid,
1279 : ATTSTATSSLOT_NUMBERS))
1280 : {
1281 2592 : otherdistinct -= mcvslot.nnumbers;
1282 2592 : free_attstatsslot(&mcvslot);
1283 : }
1284 :
1285 : /* If result doesn't seem sane, leave eq_selec at 0 */
1286 20581 : if (otherdistinct > 1)
1287 20560 : eq_selec = 1.0 / otherdistinct;
1288 : }
1289 :
1290 : /*
1291 : * Convert the constant and the two nearest bin boundary
1292 : * values to a uniform comparison scale, and do a linear
1293 : * interpolation within this bin.
1294 : */
1295 48970 : if (convert_to_scalar(constval, consttype, collation,
1296 : &val,
1297 48970 : sslot.values[i - 1], sslot.values[i],
1298 : vardata->vartype,
1299 : &low, &high))
1300 : {
1301 48970 : if (high <= low)
1302 : {
1303 : /* cope if bin boundaries appear identical */
1304 0 : binfrac = 0.5;
1305 : }
1306 48970 : else if (val <= low)
1307 10989 : binfrac = 0.0;
1308 37981 : else if (val >= high)
1309 1632 : binfrac = 1.0;
1310 : else
1311 : {
1312 36349 : binfrac = (val - low) / (high - low);
1313 :
1314 : /*
1315 : * Watch out for the possibility that we got a NaN or
1316 : * Infinity from the division. This can happen
1317 : * despite the previous checks, if for example "low"
1318 : * is -Infinity.
1319 : */
1320 36349 : if (isnan(binfrac) ||
1321 36349 : binfrac < 0.0 || binfrac > 1.0)
1322 0 : binfrac = 0.5;
1323 : }
1324 : }
1325 : else
1326 : {
1327 : /*
1328 : * Ideally we'd produce an error here, on the grounds that
1329 : * the given operator shouldn't have scalarXXsel
1330 : * registered as its selectivity func unless we can deal
1331 : * with its operand types. But currently, all manner of
1332 : * stuff is invoking scalarXXsel, so give a default
1333 : * estimate until that can be fixed.
1334 : */
1335 0 : binfrac = 0.5;
1336 : }
1337 :
1338 : /*
1339 : * Now, compute the overall selectivity across the values
1340 : * represented by the histogram. We have i-1 full bins and
1341 : * binfrac partial bin below the constant.
1342 : */
1343 48970 : histfrac = (double) (i - 1) + binfrac;
1344 48970 : histfrac /= (double) (sslot.nvalues - 1);
1345 :
1346 : /*
1347 : * At this point, histfrac is an estimate of the fraction of
1348 : * the population represented by the histogram that satisfies
1349 : * "x <= constval". Somewhat remarkably, this statement is
1350 : * true regardless of which operator we were doing the probes
1351 : * with, so long as convert_to_scalar() delivers reasonable
1352 : * results. If the probe constant is equal to some histogram
1353 : * entry, we would have considered the bin to the left of that
1354 : * entry if probing with "<" or ">=", or the bin to the right
1355 : * if probing with "<=" or ">"; but binfrac would have come
1356 : * out as 1.0 in the first case and 0.0 in the second, leading
1357 : * to the same histfrac in either case. For probe constants
1358 : * between histogram entries, we find the same bin and get the
1359 : * same estimate with any operator.
1360 : *
1361 : * The fact that the estimate corresponds to "x <= constval"
1362 : * and not "x < constval" is because of the way that ANALYZE
1363 : * constructs the histogram: each entry is, effectively, the
1364 : * rightmost value in its sample bucket. So selectivity
1365 : * values that are exact multiples of 1/(histogram_size-1)
1366 : * should be understood as estimates including a histogram
1367 : * entry plus everything to its left.
1368 : *
1369 : * However, that breaks down for the first histogram entry,
1370 : * which necessarily is the leftmost value in its sample
1371 : * bucket. That means the first histogram bin is slightly
1372 : * narrower than the rest, by an amount equal to eq_selec.
1373 : * Another way to say that is that we want "x <= leftmost" to
1374 : * be estimated as eq_selec not zero. So, if we're dealing
1375 : * with the first bin (i==1), rescale to make that true while
1376 : * adjusting the rest of that bin linearly.
1377 : */
1378 48970 : if (i == 1)
1379 9166 : histfrac += eq_selec * (1.0 - binfrac);
1380 :
1381 : /*
1382 : * "x <= constval" is good if we want an estimate for "<=" or
1383 : * ">", but if we are estimating for "<" or ">=", we now need
1384 : * to decrease the estimate by eq_selec.
1385 : */
1386 48970 : if (isgt == iseq)
1387 15360 : histfrac -= eq_selec;
1388 : }
1389 :
1390 : /*
1391 : * Now the estimate is finished for "<" and "<=" cases. If we are
1392 : * estimating for ">" or ">=", flip it.
1393 : */
1394 120080 : hist_selec = isgt ? (1.0 - histfrac) : histfrac;
1395 :
1396 : /*
1397 : * The histogram boundaries are only approximate to begin with,
1398 : * and may well be out of date anyway. Therefore, don't believe
1399 : * extremely small or large selectivity estimates --- unless we
1400 : * got actual current endpoint values from the table, in which
1401 : * case just do the usual sanity clamp. Somewhat arbitrarily, we
1402 : * set the cutoff for other cases at a hundredth of the histogram
1403 : * resolution.
1404 : */
1405 120080 : if (have_end)
1406 67841 : CLAMP_PROBABILITY(hist_selec);
1407 : else
1408 : {
1409 52239 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1410 :
1411 52239 : if (hist_selec < cutoff)
1412 18514 : hist_selec = cutoff;
1413 33725 : else if (hist_selec > 1.0 - cutoff)
1414 11507 : hist_selec = 1.0 - cutoff;
1415 : }
1416 : }
1417 92 : else if (sslot.nvalues > 1)
1418 : {
1419 : /*
1420 : * If we get here, we have a histogram but it's not sorted the way
1421 : * we want. Do a brute-force search to see how many of the
1422 : * entries satisfy the comparison condition, and take that
1423 : * fraction as our estimate. (This is identical to the inner loop
1424 : * of histogram_selectivity; maybe share code?)
1425 : */
1426 92 : LOCAL_FCINFO(fcinfo, 2);
1427 92 : int nmatch = 0;
1428 :
1429 92 : InitFunctionCallInfoData(*fcinfo, opproc, 2, collation,
1430 : NULL, NULL);
1431 92 : fcinfo->args[0].isnull = false;
1432 92 : fcinfo->args[1].isnull = false;
1433 92 : fcinfo->args[1].value = constval;
1434 481050 : for (int i = 0; i < sslot.nvalues; i++)
1435 : {
1436 : Datum fresult;
1437 :
1438 480958 : fcinfo->args[0].value = sslot.values[i];
1439 480958 : fcinfo->isnull = false;
1440 480958 : fresult = FunctionCallInvoke(fcinfo);
1441 480958 : if (!fcinfo->isnull && DatumGetBool(fresult))
1442 966 : nmatch++;
1443 : }
1444 92 : hist_selec = ((double) nmatch) / ((double) sslot.nvalues);
1445 :
1446 : /*
1447 : * As above, clamp to a hundredth of the histogram resolution.
1448 : * This case is surely even less trustworthy than the normal one,
1449 : * so we shouldn't believe exact 0 or 1 selectivity. (Maybe the
1450 : * clamp should be more restrictive in this case?)
1451 : */
1452 : {
1453 92 : double cutoff = 0.01 / (double) (sslot.nvalues - 1);
1454 :
1455 92 : if (hist_selec < cutoff)
1456 6 : hist_selec = cutoff;
1457 86 : else if (hist_selec > 1.0 - cutoff)
1458 6 : hist_selec = 1.0 - cutoff;
1459 : }
1460 : }
1461 :
1462 120172 : free_attstatsslot(&sslot);
1463 : }
1464 :
1465 189157 : return hist_selec;
1466 : }
1467 :
1468 : /*
1469 : * Common wrapper function for the selectivity estimators that simply
1470 : * invoke scalarineqsel().
1471 : */
1472 : static Datum
1473 26002 : scalarineqsel_wrapper(PG_FUNCTION_ARGS, bool isgt, bool iseq)
1474 : {
1475 26002 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
1476 26002 : Oid operator = PG_GETARG_OID(1);
1477 26002 : List *args = (List *) PG_GETARG_POINTER(2);
1478 26002 : int varRelid = PG_GETARG_INT32(3);
1479 26002 : Oid collation = PG_GET_COLLATION();
1480 : VariableStatData vardata;
1481 : Node *other;
1482 : bool varonleft;
1483 : Datum constval;
1484 : Oid consttype;
1485 : double selec;
1486 :
1487 : /*
1488 : * If expression is not variable op something or something op variable,
1489 : * then punt and return a default estimate.
1490 : */
1491 26002 : if (!get_restriction_variable(root, args, varRelid,
1492 : &vardata, &other, &varonleft))
1493 322 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1494 :
1495 : /*
1496 : * Can't do anything useful if the something is not a constant, either.
1497 : */
1498 25680 : if (!IsA(other, Const))
1499 : {
1500 1421 : ReleaseVariableStats(vardata);
1501 1421 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1502 : }
1503 :
1504 : /*
1505 : * If the constant is NULL, assume operator is strict and return zero, ie,
1506 : * operator will never return TRUE.
1507 : */
1508 24259 : if (((Const *) other)->constisnull)
1509 : {
1510 33 : ReleaseVariableStats(vardata);
1511 33 : PG_RETURN_FLOAT8(0.0);
1512 : }
1513 24226 : constval = ((Const *) other)->constvalue;
1514 24226 : consttype = ((Const *) other)->consttype;
1515 :
1516 : /*
1517 : * Force the var to be on the left to simplify logic in scalarineqsel.
1518 : */
1519 24226 : if (!varonleft)
1520 : {
1521 192 : operator = get_commutator(operator);
1522 192 : if (!operator)
1523 : {
1524 : /* Use default selectivity (should we raise an error instead?) */
1525 0 : ReleaseVariableStats(vardata);
1526 0 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
1527 : }
1528 192 : isgt = !isgt;
1529 : }
1530 :
1531 : /* The rest of the work is done by scalarineqsel(). */
1532 24226 : selec = scalarineqsel(root, operator, isgt, iseq, collation,
1533 : &vardata, constval, consttype);
1534 :
1535 24226 : ReleaseVariableStats(vardata);
1536 :
1537 24226 : PG_RETURN_FLOAT8((float8) selec);
1538 : }
1539 :
1540 : /*
1541 : * scalarltsel - Selectivity of "<" for scalars.
1542 : */
1543 : Datum
1544 7615 : scalarltsel(PG_FUNCTION_ARGS)
1545 : {
1546 7615 : return scalarineqsel_wrapper(fcinfo, false, false);
1547 : }
1548 :
1549 : /*
1550 : * scalarlesel - Selectivity of "<=" for scalars.
1551 : */
1552 : Datum
1553 2318 : scalarlesel(PG_FUNCTION_ARGS)
1554 : {
1555 2318 : return scalarineqsel_wrapper(fcinfo, false, true);
1556 : }
1557 :
1558 : /*
1559 : * scalargtsel - Selectivity of ">" for scalars.
1560 : */
1561 : Datum
1562 8130 : scalargtsel(PG_FUNCTION_ARGS)
1563 : {
1564 8130 : return scalarineqsel_wrapper(fcinfo, true, false);
1565 : }
1566 :
1567 : /*
1568 : * scalargesel - Selectivity of ">=" for scalars.
1569 : */
1570 : Datum
1571 7939 : scalargesel(PG_FUNCTION_ARGS)
1572 : {
1573 7939 : return scalarineqsel_wrapper(fcinfo, true, true);
1574 : }
1575 :
1576 : /*
1577 : * boolvarsel - Selectivity of Boolean variable.
1578 : *
1579 : * This can actually be called on any boolean-valued expression. If it
1580 : * involves only Vars of the specified relation, and if there are statistics
1581 : * about the Var or expression (the latter is possible if it's indexed) then
1582 : * we'll produce a real estimate; otherwise it's just a default.
1583 : */
1584 : Selectivity
1585 35115 : boolvarsel(PlannerInfo *root, Node *arg, int varRelid)
1586 : {
1587 : VariableStatData vardata;
1588 : double selec;
1589 :
1590 35115 : examine_variable(root, arg, varRelid, &vardata);
1591 35115 : if (HeapTupleIsValid(vardata.statsTuple))
1592 : {
1593 : /*
1594 : * A boolean variable V is equivalent to the clause V = 't', so we
1595 : * compute the selectivity as if that is what we have.
1596 : */
1597 19210 : selec = var_eq_const(&vardata, BooleanEqualOperator, InvalidOid,
1598 : BoolGetDatum(true), false, true, false);
1599 : }
1600 15905 : else if (is_funcclause(arg))
1601 : {
1602 : /*
1603 : * If we have no stats and it's a function call, estimate 0.3333333.
1604 : * This seems a pretty unprincipled choice, but Postgres has been
1605 : * using that estimate for function calls since 1992. The hoariness
1606 : * of this behavior suggests that we should not be in too much hurry
1607 : * to use another value.
1608 : */
1609 7041 : selec = 0.3333333;
1610 : }
1611 : else
1612 : {
1613 : /* Otherwise, the default estimate is 0.5 */
1614 8864 : selec = 0.5;
1615 : }
1616 35115 : ReleaseVariableStats(vardata);
1617 35115 : return selec;
1618 : }
1619 :
1620 : /*
1621 : * booltestsel - Selectivity of BooleanTest Node.
1622 : */
1623 : Selectivity
1624 507 : booltestsel(PlannerInfo *root, BoolTestType booltesttype, Node *arg,
1625 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1626 : {
1627 : VariableStatData vardata;
1628 : double selec;
1629 :
1630 507 : examine_variable(root, arg, varRelid, &vardata);
1631 :
1632 507 : if (HeapTupleIsValid(vardata.statsTuple))
1633 : {
1634 : Form_pg_statistic stats;
1635 : double freq_null;
1636 : AttStatsSlot sslot;
1637 :
1638 12 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1639 12 : freq_null = stats->stanullfrac;
1640 :
1641 12 : if (get_attstatsslot(&sslot, vardata.statsTuple,
1642 : STATISTIC_KIND_MCV, InvalidOid,
1643 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)
1644 6 : && sslot.nnumbers > 0)
1645 6 : {
1646 : double freq_true;
1647 : double freq_false;
1648 :
1649 : /*
1650 : * Get first MCV frequency and derive frequency for true.
1651 : */
1652 6 : if (DatumGetBool(sslot.values[0]))
1653 0 : freq_true = sslot.numbers[0];
1654 : else
1655 6 : freq_true = 1.0 - sslot.numbers[0] - freq_null;
1656 :
1657 : /*
1658 : * Next derive frequency for false. Then use these as appropriate
1659 : * to derive frequency for each case.
1660 : */
1661 6 : freq_false = 1.0 - freq_true - freq_null;
1662 :
1663 6 : switch (booltesttype)
1664 : {
1665 0 : case IS_UNKNOWN:
1666 : /* select only NULL values */
1667 0 : selec = freq_null;
1668 0 : break;
1669 0 : case IS_NOT_UNKNOWN:
1670 : /* select non-NULL values */
1671 0 : selec = 1.0 - freq_null;
1672 0 : break;
1673 6 : case IS_TRUE:
1674 : /* select only TRUE values */
1675 6 : selec = freq_true;
1676 6 : break;
1677 0 : case IS_NOT_TRUE:
1678 : /* select non-TRUE values */
1679 0 : selec = 1.0 - freq_true;
1680 0 : break;
1681 0 : case IS_FALSE:
1682 : /* select only FALSE values */
1683 0 : selec = freq_false;
1684 0 : break;
1685 0 : case IS_NOT_FALSE:
1686 : /* select non-FALSE values */
1687 0 : selec = 1.0 - freq_false;
1688 0 : break;
1689 0 : default:
1690 0 : elog(ERROR, "unrecognized booltesttype: %d",
1691 : (int) booltesttype);
1692 : selec = 0.0; /* Keep compiler quiet */
1693 : break;
1694 : }
1695 :
1696 6 : free_attstatsslot(&sslot);
1697 : }
1698 : else
1699 : {
1700 : /*
1701 : * No most-common-value info available. Still have null fraction
1702 : * information, so use it for IS [NOT] UNKNOWN. Otherwise adjust
1703 : * for null fraction and assume a 50-50 split of TRUE and FALSE.
1704 : */
1705 6 : switch (booltesttype)
1706 : {
1707 6 : case IS_UNKNOWN:
1708 : /* select only NULL values */
1709 6 : selec = freq_null;
1710 6 : break;
1711 0 : case IS_NOT_UNKNOWN:
1712 : /* select non-NULL values */
1713 0 : selec = 1.0 - freq_null;
1714 0 : break;
1715 0 : case IS_TRUE:
1716 : case IS_FALSE:
1717 : /* Assume we select half of the non-NULL values */
1718 0 : selec = (1.0 - freq_null) / 2.0;
1719 0 : break;
1720 0 : case IS_NOT_TRUE:
1721 : case IS_NOT_FALSE:
1722 : /* Assume we select NULLs plus half of the non-NULLs */
1723 : /* equiv. to freq_null + (1.0 - freq_null) / 2.0 */
1724 0 : selec = (freq_null + 1.0) / 2.0;
1725 0 : break;
1726 0 : default:
1727 0 : elog(ERROR, "unrecognized booltesttype: %d",
1728 : (int) booltesttype);
1729 : selec = 0.0; /* Keep compiler quiet */
1730 : break;
1731 : }
1732 : }
1733 : }
1734 : else
1735 : {
1736 : /*
1737 : * If we can't get variable statistics for the argument, perhaps
1738 : * clause_selectivity can do something with it. We ignore the
1739 : * possibility of a NULL value when using clause_selectivity, and just
1740 : * assume the value is either TRUE or FALSE.
1741 : */
1742 495 : switch (booltesttype)
1743 : {
1744 24 : case IS_UNKNOWN:
1745 24 : selec = DEFAULT_UNK_SEL;
1746 24 : break;
1747 54 : case IS_NOT_UNKNOWN:
1748 54 : selec = DEFAULT_NOT_UNK_SEL;
1749 54 : break;
1750 126 : case IS_TRUE:
1751 : case IS_NOT_FALSE:
1752 126 : selec = (double) clause_selectivity(root, arg,
1753 : varRelid,
1754 : jointype, sjinfo);
1755 126 : break;
1756 291 : case IS_FALSE:
1757 : case IS_NOT_TRUE:
1758 291 : selec = 1.0 - (double) clause_selectivity(root, arg,
1759 : varRelid,
1760 : jointype, sjinfo);
1761 291 : break;
1762 0 : default:
1763 0 : elog(ERROR, "unrecognized booltesttype: %d",
1764 : (int) booltesttype);
1765 : selec = 0.0; /* Keep compiler quiet */
1766 : break;
1767 : }
1768 : }
1769 :
1770 507 : ReleaseVariableStats(vardata);
1771 :
1772 : /* result should be in range, but make sure... */
1773 507 : CLAMP_PROBABILITY(selec);
1774 :
1775 507 : return (Selectivity) selec;
1776 : }
1777 :
1778 : /*
1779 : * nulltestsel - Selectivity of NullTest Node.
1780 : */
1781 : Selectivity
1782 9614 : nulltestsel(PlannerInfo *root, NullTestType nulltesttype, Node *arg,
1783 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
1784 : {
1785 : VariableStatData vardata;
1786 : double selec;
1787 :
1788 9614 : examine_variable(root, arg, varRelid, &vardata);
1789 :
1790 9614 : if (HeapTupleIsValid(vardata.statsTuple))
1791 : {
1792 : Form_pg_statistic stats;
1793 : double freq_null;
1794 :
1795 5113 : stats = (Form_pg_statistic) GETSTRUCT(vardata.statsTuple);
1796 5113 : freq_null = stats->stanullfrac;
1797 :
1798 5113 : switch (nulltesttype)
1799 : {
1800 3638 : case IS_NULL:
1801 :
1802 : /*
1803 : * Use freq_null directly.
1804 : */
1805 3638 : selec = freq_null;
1806 3638 : break;
1807 1475 : case IS_NOT_NULL:
1808 :
1809 : /*
1810 : * Select not unknown (not null) values. Calculate from
1811 : * freq_null.
1812 : */
1813 1475 : selec = 1.0 - freq_null;
1814 1475 : break;
1815 0 : default:
1816 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1817 : (int) nulltesttype);
1818 : return (Selectivity) 0; /* keep compiler quiet */
1819 : }
1820 : }
1821 4501 : else if (vardata.var && IsA(vardata.var, Var) &&
1822 4109 : ((Var *) vardata.var)->varattno < 0)
1823 : {
1824 : /*
1825 : * There are no stats for system columns, but we know they are never
1826 : * NULL.
1827 : */
1828 52 : selec = (nulltesttype == IS_NULL) ? 0.0 : 1.0;
1829 : }
1830 : else
1831 : {
1832 : /*
1833 : * No ANALYZE stats available, so make a guess
1834 : */
1835 4449 : switch (nulltesttype)
1836 : {
1837 1108 : case IS_NULL:
1838 1108 : selec = DEFAULT_UNK_SEL;
1839 1108 : break;
1840 3341 : case IS_NOT_NULL:
1841 3341 : selec = DEFAULT_NOT_UNK_SEL;
1842 3341 : break;
1843 0 : default:
1844 0 : elog(ERROR, "unrecognized nulltesttype: %d",
1845 : (int) nulltesttype);
1846 : return (Selectivity) 0; /* keep compiler quiet */
1847 : }
1848 : }
1849 :
1850 9614 : ReleaseVariableStats(vardata);
1851 :
1852 : /* result should be in range, but make sure... */
1853 9614 : CLAMP_PROBABILITY(selec);
1854 :
1855 9614 : return (Selectivity) selec;
1856 : }
1857 :
1858 : /*
1859 : * strip_array_coercion - strip binary-compatible relabeling from an array expr
1860 : *
1861 : * For array values, the parser normally generates ArrayCoerceExpr conversions,
1862 : * but it seems possible that RelabelType might show up. Also, the planner
1863 : * is not currently tense about collapsing stacked ArrayCoerceExpr nodes,
1864 : * so we need to be ready to deal with more than one level.
1865 : */
1866 : static Node *
1867 72792 : strip_array_coercion(Node *node)
1868 : {
1869 : for (;;)
1870 : {
1871 72848 : if (node && IsA(node, ArrayCoerceExpr))
1872 56 : {
1873 1749 : ArrayCoerceExpr *acoerce = (ArrayCoerceExpr *) node;
1874 :
1875 : /*
1876 : * If the per-element expression is just a RelabelType on top of
1877 : * CaseTestExpr, then we know it's a binary-compatible relabeling.
1878 : */
1879 1749 : if (IsA(acoerce->elemexpr, RelabelType) &&
1880 56 : IsA(((RelabelType *) acoerce->elemexpr)->arg, CaseTestExpr))
1881 56 : node = (Node *) acoerce->arg;
1882 : else
1883 : break;
1884 : }
1885 71099 : else if (node && IsA(node, RelabelType))
1886 : {
1887 : /* We don't really expect this case, but may as well cope */
1888 0 : node = (Node *) ((RelabelType *) node)->arg;
1889 : }
1890 : else
1891 : break;
1892 : }
1893 72792 : return node;
1894 : }
1895 :
1896 : /*
1897 : * scalararraysel - Selectivity of ScalarArrayOpExpr Node.
1898 : */
1899 : Selectivity
1900 13216 : scalararraysel(PlannerInfo *root,
1901 : ScalarArrayOpExpr *clause,
1902 : bool is_join_clause,
1903 : int varRelid,
1904 : JoinType jointype,
1905 : SpecialJoinInfo *sjinfo)
1906 : {
1907 13216 : Oid operator = clause->opno;
1908 13216 : bool useOr = clause->useOr;
1909 13216 : bool isEquality = false;
1910 13216 : bool isInequality = false;
1911 : Node *leftop;
1912 : Node *rightop;
1913 : Oid nominal_element_type;
1914 : Oid nominal_element_collation;
1915 : TypeCacheEntry *typentry;
1916 : RegProcedure oprsel;
1917 : FmgrInfo oprselproc;
1918 : Selectivity s1;
1919 : Selectivity s1disjoint;
1920 :
1921 : /* First, deconstruct the expression */
1922 : Assert(list_length(clause->args) == 2);
1923 13216 : leftop = (Node *) linitial(clause->args);
1924 13216 : rightop = (Node *) lsecond(clause->args);
1925 :
1926 : /* aggressively reduce both sides to constants */
1927 13216 : leftop = estimate_expression_value(root, leftop);
1928 13216 : rightop = estimate_expression_value(root, rightop);
1929 :
1930 : /* get nominal (after relabeling) element type of rightop */
1931 13216 : nominal_element_type = get_base_element_type(exprType(rightop));
1932 13216 : if (!OidIsValid(nominal_element_type))
1933 0 : return (Selectivity) 0.5; /* probably shouldn't happen */
1934 : /* get nominal collation, too, for generating constants */
1935 13216 : nominal_element_collation = exprCollation(rightop);
1936 :
1937 : /* look through any binary-compatible relabeling of rightop */
1938 13216 : rightop = strip_array_coercion(rightop);
1939 :
1940 : /*
1941 : * Detect whether the operator is the default equality or inequality
1942 : * operator of the array element type.
1943 : */
1944 13216 : typentry = lookup_type_cache(nominal_element_type, TYPECACHE_EQ_OPR);
1945 13216 : if (OidIsValid(typentry->eq_opr))
1946 : {
1947 13214 : if (operator == typentry->eq_opr)
1948 11428 : isEquality = true;
1949 1786 : else if (get_negator(operator) == typentry->eq_opr)
1950 1503 : isInequality = true;
1951 : }
1952 :
1953 : /*
1954 : * If it is equality or inequality, we might be able to estimate this as a
1955 : * form of array containment; for instance "const = ANY(column)" can be
1956 : * treated as "ARRAY[const] <@ column". scalararraysel_containment tries
1957 : * that, and returns the selectivity estimate if successful, or -1 if not.
1958 : */
1959 13216 : if ((isEquality || isInequality) && !is_join_clause)
1960 : {
1961 12930 : s1 = scalararraysel_containment(root, leftop, rightop,
1962 : nominal_element_type,
1963 : isEquality, useOr, varRelid);
1964 12930 : if (s1 >= 0.0)
1965 59 : return s1;
1966 : }
1967 :
1968 : /*
1969 : * Look up the underlying operator's selectivity estimator. Punt if it
1970 : * hasn't got one.
1971 : */
1972 13157 : if (is_join_clause)
1973 1 : oprsel = get_oprjoin(operator);
1974 : else
1975 13156 : oprsel = get_oprrest(operator);
1976 13157 : if (!oprsel)
1977 2 : return (Selectivity) 0.5;
1978 13155 : fmgr_info(oprsel, &oprselproc);
1979 :
1980 : /*
1981 : * In the array-containment check above, we must only believe that an
1982 : * operator is equality or inequality if it is the default btree equality
1983 : * operator (or its negator) for the element type, since those are the
1984 : * operators that array containment will use. But in what follows, we can
1985 : * be a little laxer, and also believe that any operators using eqsel() or
1986 : * neqsel() as selectivity estimator act like equality or inequality.
1987 : */
1988 13155 : if (oprsel == F_EQSEL || oprsel == F_EQJOINSEL)
1989 11467 : isEquality = true;
1990 1688 : else if (oprsel == F_NEQSEL || oprsel == F_NEQJOINSEL)
1991 1448 : isInequality = true;
1992 :
1993 : /*
1994 : * We consider three cases:
1995 : *
1996 : * 1. rightop is an Array constant: deconstruct the array, apply the
1997 : * operator's selectivity function for each array element, and merge the
1998 : * results in the same way that clausesel.c does for AND/OR combinations.
1999 : *
2000 : * 2. rightop is an ARRAY[] construct: apply the operator's selectivity
2001 : * function for each element of the ARRAY[] construct, and merge.
2002 : *
2003 : * 3. otherwise, make a guess ...
2004 : */
2005 13155 : if (rightop && IsA(rightop, Const))
2006 10831 : {
2007 10846 : Datum arraydatum = ((Const *) rightop)->constvalue;
2008 10846 : bool arrayisnull = ((Const *) rightop)->constisnull;
2009 : ArrayType *arrayval;
2010 : int16 elmlen;
2011 : bool elmbyval;
2012 : char elmalign;
2013 : int num_elems;
2014 : Datum *elem_values;
2015 : bool *elem_nulls;
2016 : int i;
2017 :
2018 10846 : if (arrayisnull) /* qual can't succeed if null array */
2019 15 : return (Selectivity) 0.0;
2020 10831 : arrayval = DatumGetArrayTypeP(arraydatum);
2021 10831 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
2022 : &elmlen, &elmbyval, &elmalign);
2023 10831 : deconstruct_array(arrayval,
2024 : ARR_ELEMTYPE(arrayval),
2025 : elmlen, elmbyval, elmalign,
2026 : &elem_values, &elem_nulls, &num_elems);
2027 :
2028 : /*
2029 : * For generic operators, we assume the probability of success is
2030 : * independent for each array element. But for "= ANY" or "<> ALL",
2031 : * if the array elements are distinct (which'd typically be the case)
2032 : * then the probabilities are disjoint, and we should just sum them.
2033 : *
2034 : * If we were being really tense we would try to confirm that the
2035 : * elements are all distinct, but that would be expensive and it
2036 : * doesn't seem to be worth the cycles; it would amount to penalizing
2037 : * well-written queries in favor of poorly-written ones. However, we
2038 : * do protect ourselves a little bit by checking whether the
2039 : * disjointness assumption leads to an impossible (out of range)
2040 : * probability; if so, we fall back to the normal calculation.
2041 : */
2042 10831 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2043 :
2044 44733 : for (i = 0; i < num_elems; i++)
2045 : {
2046 : List *args;
2047 : Selectivity s2;
2048 :
2049 33902 : args = list_make2(leftop,
2050 : makeConst(nominal_element_type,
2051 : -1,
2052 : nominal_element_collation,
2053 : elmlen,
2054 : elem_values[i],
2055 : elem_nulls[i],
2056 : elmbyval));
2057 33902 : if (is_join_clause)
2058 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2059 : clause->inputcollid,
2060 : PointerGetDatum(root),
2061 : ObjectIdGetDatum(operator),
2062 : PointerGetDatum(args),
2063 : Int16GetDatum(jointype),
2064 : PointerGetDatum(sjinfo)));
2065 : else
2066 33902 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2067 : clause->inputcollid,
2068 : PointerGetDatum(root),
2069 : ObjectIdGetDatum(operator),
2070 : PointerGetDatum(args),
2071 : Int32GetDatum(varRelid)));
2072 :
2073 33902 : if (useOr)
2074 : {
2075 29215 : s1 = s1 + s2 - s1 * s2;
2076 29215 : if (isEquality)
2077 28693 : s1disjoint += s2;
2078 : }
2079 : else
2080 : {
2081 4687 : s1 = s1 * s2;
2082 4687 : if (isInequality)
2083 4531 : s1disjoint += s2 - 1.0;
2084 : }
2085 : }
2086 :
2087 : /* accept disjoint-probability estimate if in range */
2088 10831 : if ((useOr ? isEquality : isInequality) &&
2089 10519 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2090 10495 : s1 = s1disjoint;
2091 : }
2092 2309 : else if (rightop && IsA(rightop, ArrayExpr) &&
2093 193 : !((ArrayExpr *) rightop)->multidims)
2094 193 : {
2095 193 : ArrayExpr *arrayexpr = (ArrayExpr *) rightop;
2096 : int16 elmlen;
2097 : bool elmbyval;
2098 : ListCell *l;
2099 :
2100 193 : get_typlenbyval(arrayexpr->element_typeid,
2101 : &elmlen, &elmbyval);
2102 :
2103 : /*
2104 : * We use the assumption of disjoint probabilities here too, although
2105 : * the odds of equal array elements are rather higher if the elements
2106 : * are not all constants (which they won't be, else constant folding
2107 : * would have reduced the ArrayExpr to a Const). In this path it's
2108 : * critical to have the sanity check on the s1disjoint estimate.
2109 : */
2110 193 : s1 = s1disjoint = (useOr ? 0.0 : 1.0);
2111 :
2112 714 : foreach(l, arrayexpr->elements)
2113 : {
2114 521 : Node *elem = (Node *) lfirst(l);
2115 : List *args;
2116 : Selectivity s2;
2117 :
2118 : /*
2119 : * Theoretically, if elem isn't of nominal_element_type we should
2120 : * insert a RelabelType, but it seems unlikely that any operator
2121 : * estimation function would really care ...
2122 : */
2123 521 : args = list_make2(leftop, elem);
2124 521 : if (is_join_clause)
2125 3 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2126 : clause->inputcollid,
2127 : PointerGetDatum(root),
2128 : ObjectIdGetDatum(operator),
2129 : PointerGetDatum(args),
2130 : Int16GetDatum(jointype),
2131 : PointerGetDatum(sjinfo)));
2132 : else
2133 518 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2134 : clause->inputcollid,
2135 : PointerGetDatum(root),
2136 : ObjectIdGetDatum(operator),
2137 : PointerGetDatum(args),
2138 : Int32GetDatum(varRelid)));
2139 :
2140 521 : if (useOr)
2141 : {
2142 521 : s1 = s1 + s2 - s1 * s2;
2143 521 : if (isEquality)
2144 521 : s1disjoint += s2;
2145 : }
2146 : else
2147 : {
2148 0 : s1 = s1 * s2;
2149 0 : if (isInequality)
2150 0 : s1disjoint += s2 - 1.0;
2151 : }
2152 : }
2153 :
2154 : /* accept disjoint-probability estimate if in range */
2155 193 : if ((useOr ? isEquality : isInequality) &&
2156 193 : s1disjoint >= 0.0 && s1disjoint <= 1.0)
2157 193 : s1 = s1disjoint;
2158 : }
2159 : else
2160 : {
2161 : CaseTestExpr *dummyexpr;
2162 : List *args;
2163 : Selectivity s2;
2164 : int i;
2165 :
2166 : /*
2167 : * We need a dummy rightop to pass to the operator selectivity
2168 : * routine. It can be pretty much anything that doesn't look like a
2169 : * constant; CaseTestExpr is a convenient choice.
2170 : */
2171 2116 : dummyexpr = makeNode(CaseTestExpr);
2172 2116 : dummyexpr->typeId = nominal_element_type;
2173 2116 : dummyexpr->typeMod = -1;
2174 2116 : dummyexpr->collation = clause->inputcollid;
2175 2116 : args = list_make2(leftop, dummyexpr);
2176 2116 : if (is_join_clause)
2177 0 : s2 = DatumGetFloat8(FunctionCall5Coll(&oprselproc,
2178 : clause->inputcollid,
2179 : PointerGetDatum(root),
2180 : ObjectIdGetDatum(operator),
2181 : PointerGetDatum(args),
2182 : Int16GetDatum(jointype),
2183 : PointerGetDatum(sjinfo)));
2184 : else
2185 2116 : s2 = DatumGetFloat8(FunctionCall4Coll(&oprselproc,
2186 : clause->inputcollid,
2187 : PointerGetDatum(root),
2188 : ObjectIdGetDatum(operator),
2189 : PointerGetDatum(args),
2190 : Int32GetDatum(varRelid)));
2191 2116 : s1 = useOr ? 0.0 : 1.0;
2192 :
2193 : /*
2194 : * Arbitrarily assume 10 elements in the eventual array value (see
2195 : * also estimate_array_length). We don't risk an assumption of
2196 : * disjoint probabilities here.
2197 : */
2198 23276 : for (i = 0; i < 10; i++)
2199 : {
2200 21160 : if (useOr)
2201 21160 : s1 = s1 + s2 - s1 * s2;
2202 : else
2203 0 : s1 = s1 * s2;
2204 : }
2205 : }
2206 :
2207 : /* result should be in range, but make sure... */
2208 13140 : CLAMP_PROBABILITY(s1);
2209 :
2210 13140 : return s1;
2211 : }
2212 :
2213 : /*
2214 : * Estimate number of elements in the array yielded by an expression.
2215 : *
2216 : * Note: the result is integral, but we use "double" to avoid overflow
2217 : * concerns. Most callers will use it in double-type expressions anyway.
2218 : *
2219 : * Note: in some code paths root can be passed as NULL, resulting in
2220 : * slightly worse estimates.
2221 : */
2222 : double
2223 59576 : estimate_array_length(PlannerInfo *root, Node *arrayexpr)
2224 : {
2225 : /* look through any binary-compatible relabeling of arrayexpr */
2226 59576 : arrayexpr = strip_array_coercion(arrayexpr);
2227 :
2228 59576 : if (arrayexpr && IsA(arrayexpr, Const))
2229 : {
2230 28041 : Datum arraydatum = ((Const *) arrayexpr)->constvalue;
2231 28041 : bool arrayisnull = ((Const *) arrayexpr)->constisnull;
2232 : ArrayType *arrayval;
2233 :
2234 28041 : if (arrayisnull)
2235 45 : return 0;
2236 27996 : arrayval = DatumGetArrayTypeP(arraydatum);
2237 27996 : return ArrayGetNItems(ARR_NDIM(arrayval), ARR_DIMS(arrayval));
2238 : }
2239 31535 : else if (arrayexpr && IsA(arrayexpr, ArrayExpr) &&
2240 342 : !((ArrayExpr *) arrayexpr)->multidims)
2241 : {
2242 342 : return list_length(((ArrayExpr *) arrayexpr)->elements);
2243 : }
2244 31193 : else if (arrayexpr && root)
2245 : {
2246 : /* See if we can find any statistics about it */
2247 : VariableStatData vardata;
2248 : AttStatsSlot sslot;
2249 31181 : double nelem = 0;
2250 :
2251 31181 : examine_variable(root, arrayexpr, 0, &vardata);
2252 31181 : if (HeapTupleIsValid(vardata.statsTuple))
2253 : {
2254 : /*
2255 : * Found stats, so use the average element count, which is stored
2256 : * in the last stanumbers element of the DECHIST statistics.
2257 : * Actually that is the average count of *distinct* elements;
2258 : * perhaps we should scale it up somewhat?
2259 : */
2260 4945 : if (get_attstatsslot(&sslot, vardata.statsTuple,
2261 : STATISTIC_KIND_DECHIST, InvalidOid,
2262 : ATTSTATSSLOT_NUMBERS))
2263 : {
2264 4888 : if (sslot.nnumbers > 0)
2265 4888 : nelem = clamp_row_est(sslot.numbers[sslot.nnumbers - 1]);
2266 4888 : free_attstatsslot(&sslot);
2267 : }
2268 : }
2269 31181 : ReleaseVariableStats(vardata);
2270 :
2271 31181 : if (nelem > 0)
2272 4888 : return nelem;
2273 : }
2274 :
2275 : /* Else use a default guess --- this should match scalararraysel */
2276 26305 : return 10;
2277 : }
2278 :
2279 : /*
2280 : * rowcomparesel - Selectivity of RowCompareExpr Node.
2281 : *
2282 : * We estimate RowCompare selectivity by considering just the first (high
2283 : * order) columns, which makes it equivalent to an ordinary OpExpr. While
2284 : * this estimate could be refined by considering additional columns, it
2285 : * seems unlikely that we could do a lot better without multi-column
2286 : * statistics.
2287 : */
2288 : Selectivity
2289 126 : rowcomparesel(PlannerInfo *root,
2290 : RowCompareExpr *clause,
2291 : int varRelid, JoinType jointype, SpecialJoinInfo *sjinfo)
2292 : {
2293 : Selectivity s1;
2294 126 : Oid opno = linitial_oid(clause->opnos);
2295 126 : Oid inputcollid = linitial_oid(clause->inputcollids);
2296 : List *opargs;
2297 : bool is_join_clause;
2298 :
2299 : /* Build equivalent arg list for single operator */
2300 126 : opargs = list_make2(linitial(clause->largs), linitial(clause->rargs));
2301 :
2302 : /*
2303 : * Decide if it's a join clause. This should match clausesel.c's
2304 : * treat_as_join_clause(), except that we intentionally consider only the
2305 : * leading columns and not the rest of the clause.
2306 : */
2307 126 : if (varRelid != 0)
2308 : {
2309 : /*
2310 : * Caller is forcing restriction mode (eg, because we are examining an
2311 : * inner indexscan qual).
2312 : */
2313 27 : is_join_clause = false;
2314 : }
2315 99 : else if (sjinfo == NULL)
2316 : {
2317 : /*
2318 : * It must be a restriction clause, since it's being evaluated at a
2319 : * scan node.
2320 : */
2321 93 : is_join_clause = false;
2322 : }
2323 : else
2324 : {
2325 : /*
2326 : * Otherwise, it's a join if there's more than one base relation used.
2327 : */
2328 6 : is_join_clause = (NumRelids(root, (Node *) opargs) > 1);
2329 : }
2330 :
2331 126 : if (is_join_clause)
2332 : {
2333 : /* Estimate selectivity for a join clause. */
2334 6 : s1 = join_selectivity(root, opno,
2335 : opargs,
2336 : inputcollid,
2337 : jointype,
2338 : sjinfo);
2339 : }
2340 : else
2341 : {
2342 : /* Estimate selectivity for a restriction clause. */
2343 120 : s1 = restriction_selectivity(root, opno,
2344 : opargs,
2345 : inputcollid,
2346 : varRelid);
2347 : }
2348 :
2349 126 : return s1;
2350 : }
2351 :
2352 : /*
2353 : * eqjoinsel - Join selectivity of "="
2354 : */
2355 : Datum
2356 154326 : eqjoinsel(PG_FUNCTION_ARGS)
2357 : {
2358 154326 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
2359 154326 : Oid operator = PG_GETARG_OID(1);
2360 154326 : List *args = (List *) PG_GETARG_POINTER(2);
2361 :
2362 : #ifdef NOT_USED
2363 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
2364 : #endif
2365 154326 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
2366 154326 : Oid collation = PG_GET_COLLATION();
2367 : double selec;
2368 : double selec_inner;
2369 : VariableStatData vardata1;
2370 : VariableStatData vardata2;
2371 : double nd1;
2372 : double nd2;
2373 : bool isdefault1;
2374 : bool isdefault2;
2375 : Oid opfuncoid;
2376 : FmgrInfo eqproc;
2377 154326 : Oid hashLeft = InvalidOid;
2378 154326 : Oid hashRight = InvalidOid;
2379 : AttStatsSlot sslot1;
2380 : AttStatsSlot sslot2;
2381 154326 : Form_pg_statistic stats1 = NULL;
2382 154326 : Form_pg_statistic stats2 = NULL;
2383 154326 : bool have_mcvs1 = false;
2384 154326 : bool have_mcvs2 = false;
2385 154326 : bool *hasmatch1 = NULL;
2386 154326 : bool *hasmatch2 = NULL;
2387 154326 : int nmatches = 0;
2388 : bool get_mcv_stats;
2389 : bool join_is_reversed;
2390 : RelOptInfo *inner_rel;
2391 :
2392 154326 : get_join_variables(root, args, sjinfo,
2393 : &vardata1, &vardata2, &join_is_reversed);
2394 :
2395 154326 : nd1 = get_variable_numdistinct(&vardata1, &isdefault1);
2396 154326 : nd2 = get_variable_numdistinct(&vardata2, &isdefault2);
2397 :
2398 154326 : opfuncoid = get_opcode(operator);
2399 :
2400 154326 : memset(&sslot1, 0, sizeof(sslot1));
2401 154326 : memset(&sslot2, 0, sizeof(sslot2));
2402 :
2403 : /*
2404 : * There is no use in fetching one side's MCVs if we lack MCVs for the
2405 : * other side, so do a quick check to verify that both stats exist.
2406 : */
2407 416766 : get_mcv_stats = (HeapTupleIsValid(vardata1.statsTuple) &&
2408 191494 : HeapTupleIsValid(vardata2.statsTuple) &&
2409 83380 : get_attstatsslot(&sslot1, vardata1.statsTuple,
2410 : STATISTIC_KIND_MCV, InvalidOid,
2411 262440 : 0) &&
2412 39815 : get_attstatsslot(&sslot2, vardata2.statsTuple,
2413 : STATISTIC_KIND_MCV, InvalidOid,
2414 : 0));
2415 :
2416 154326 : if (HeapTupleIsValid(vardata1.statsTuple))
2417 : {
2418 : /* note we allow use of nullfrac regardless of security check */
2419 108114 : stats1 = (Form_pg_statistic) GETSTRUCT(vardata1.statsTuple);
2420 123146 : if (get_mcv_stats &&
2421 15032 : statistic_proc_security_check(&vardata1, opfuncoid))
2422 15032 : have_mcvs1 = get_attstatsslot(&sslot1, vardata1.statsTuple,
2423 : STATISTIC_KIND_MCV, InvalidOid,
2424 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2425 : }
2426 :
2427 154326 : if (HeapTupleIsValid(vardata2.statsTuple))
2428 : {
2429 : /* note we allow use of nullfrac regardless of security check */
2430 97407 : stats2 = (Form_pg_statistic) GETSTRUCT(vardata2.statsTuple);
2431 112439 : if (get_mcv_stats &&
2432 15032 : statistic_proc_security_check(&vardata2, opfuncoid))
2433 15032 : have_mcvs2 = get_attstatsslot(&sslot2, vardata2.statsTuple,
2434 : STATISTIC_KIND_MCV, InvalidOid,
2435 : ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS);
2436 : }
2437 :
2438 : /* Prepare info usable by both eqjoinsel_inner and eqjoinsel_semi */
2439 154326 : if (have_mcvs1 && have_mcvs2)
2440 : {
2441 15032 : fmgr_info(opfuncoid, &eqproc);
2442 15032 : hasmatch1 = (bool *) palloc0(sslot1.nvalues * sizeof(bool));
2443 15032 : hasmatch2 = (bool *) palloc0(sslot2.nvalues * sizeof(bool));
2444 :
2445 : /*
2446 : * If the MCV lists are long enough to justify hashing, try to look up
2447 : * hash functions for the join operator.
2448 : */
2449 15032 : if ((sslot1.nvalues + sslot2.nvalues) >= EQJOINSEL_MCV_HASH_THRESHOLD)
2450 967 : (void) get_op_hash_functions(operator, &hashLeft, &hashRight);
2451 : }
2452 : else
2453 139294 : memset(&eqproc, 0, sizeof(eqproc)); /* silence uninit-var warnings */
2454 :
2455 : /* We need to compute the inner-join selectivity in all cases */
2456 154326 : selec_inner = eqjoinsel_inner(&eqproc, collation,
2457 : hashLeft, hashRight,
2458 : &vardata1, &vardata2,
2459 : nd1, nd2,
2460 : isdefault1, isdefault2,
2461 : &sslot1, &sslot2,
2462 : stats1, stats2,
2463 : have_mcvs1, have_mcvs2,
2464 : hasmatch1, hasmatch2,
2465 : &nmatches);
2466 :
2467 154326 : switch (sjinfo->jointype)
2468 : {
2469 147076 : case JOIN_INNER:
2470 : case JOIN_LEFT:
2471 : case JOIN_FULL:
2472 147076 : selec = selec_inner;
2473 147076 : break;
2474 7250 : case JOIN_SEMI:
2475 : case JOIN_ANTI:
2476 :
2477 : /*
2478 : * Look up the join's inner relation. min_righthand is sufficient
2479 : * information because neither SEMI nor ANTI joins permit any
2480 : * reassociation into or out of their RHS, so the righthand will
2481 : * always be exactly that set of rels.
2482 : */
2483 7250 : inner_rel = find_join_input_rel(root, sjinfo->min_righthand);
2484 :
2485 7250 : if (!join_is_reversed)
2486 3528 : selec = eqjoinsel_semi(&eqproc, collation,
2487 : hashLeft, hashRight,
2488 : false,
2489 : &vardata1, &vardata2,
2490 : nd1, nd2,
2491 : isdefault1, isdefault2,
2492 : &sslot1, &sslot2,
2493 : stats1, stats2,
2494 : have_mcvs1, have_mcvs2,
2495 : hasmatch1, hasmatch2,
2496 : &nmatches,
2497 : inner_rel);
2498 : else
2499 3722 : selec = eqjoinsel_semi(&eqproc, collation,
2500 : hashLeft, hashRight,
2501 : true,
2502 : &vardata2, &vardata1,
2503 : nd2, nd1,
2504 : isdefault2, isdefault1,
2505 : &sslot2, &sslot1,
2506 : stats2, stats1,
2507 : have_mcvs2, have_mcvs1,
2508 : hasmatch2, hasmatch1,
2509 : &nmatches,
2510 : inner_rel);
2511 :
2512 : /*
2513 : * We should never estimate the output of a semijoin to be more
2514 : * rows than we estimate for an inner join with the same input
2515 : * rels and join condition; it's obviously impossible for that to
2516 : * happen. The former estimate is N1 * Ssemi while the latter is
2517 : * N1 * N2 * Sinner, so we may clamp Ssemi <= N2 * Sinner. Doing
2518 : * this is worthwhile because of the shakier estimation rules we
2519 : * use in eqjoinsel_semi, particularly in cases where it has to
2520 : * punt entirely.
2521 : */
2522 7250 : selec = Min(selec, inner_rel->rows * selec_inner);
2523 7250 : break;
2524 0 : default:
2525 : /* other values not expected here */
2526 0 : elog(ERROR, "unrecognized join type: %d",
2527 : (int) sjinfo->jointype);
2528 : selec = 0; /* keep compiler quiet */
2529 : break;
2530 : }
2531 :
2532 154326 : free_attstatsslot(&sslot1);
2533 154326 : free_attstatsslot(&sslot2);
2534 :
2535 154326 : ReleaseVariableStats(vardata1);
2536 154326 : ReleaseVariableStats(vardata2);
2537 :
2538 154326 : if (hasmatch1)
2539 15032 : pfree(hasmatch1);
2540 154326 : if (hasmatch2)
2541 15032 : pfree(hasmatch2);
2542 :
2543 154326 : CLAMP_PROBABILITY(selec);
2544 :
2545 154326 : PG_RETURN_FLOAT8((float8) selec);
2546 : }
2547 :
2548 : /*
2549 : * eqjoinsel_inner --- eqjoinsel for normal inner join
2550 : *
2551 : * In addition to computing the selectivity estimate, this will fill
2552 : * hasmatch1[], hasmatch2[], and *p_nmatches (if have_mcvs1 && have_mcvs2).
2553 : * We may be able to re-use that data in eqjoinsel_semi.
2554 : *
2555 : * We also use this for LEFT/FULL outer joins; it's not presently clear
2556 : * that it's worth trying to distinguish them here.
2557 : */
2558 : static double
2559 154326 : eqjoinsel_inner(FmgrInfo *eqproc, Oid collation,
2560 : Oid hashLeft, Oid hashRight,
2561 : VariableStatData *vardata1, VariableStatData *vardata2,
2562 : double nd1, double nd2,
2563 : bool isdefault1, bool isdefault2,
2564 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2565 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2566 : bool have_mcvs1, bool have_mcvs2,
2567 : bool *hasmatch1, bool *hasmatch2,
2568 : int *p_nmatches)
2569 : {
2570 : double selec;
2571 :
2572 154326 : if (have_mcvs1 && have_mcvs2)
2573 15032 : {
2574 : /*
2575 : * We have most-common-value lists for both relations. Run through
2576 : * the lists to see which MCVs actually join to each other with the
2577 : * given operator. This allows us to determine the exact join
2578 : * selectivity for the portion of the relations represented by the MCV
2579 : * lists. We still have to estimate for the remaining population, but
2580 : * in a skewed distribution this gives us a big leg up in accuracy.
2581 : * For motivation see the analysis in Y. Ioannidis and S.
2582 : * Christodoulakis, "On the propagation of errors in the size of join
2583 : * results", Technical Report 1018, Computer Science Dept., University
2584 : * of Wisconsin, Madison, March 1991 (available from ftp.cs.wisc.edu).
2585 : */
2586 15032 : double nullfrac1 = stats1->stanullfrac;
2587 15032 : double nullfrac2 = stats2->stanullfrac;
2588 : double matchprodfreq,
2589 : matchfreq1,
2590 : matchfreq2,
2591 : unmatchfreq1,
2592 : unmatchfreq2,
2593 : otherfreq1,
2594 : otherfreq2,
2595 : totalsel1,
2596 : totalsel2;
2597 : int i,
2598 : nmatches;
2599 :
2600 : /* Fill the match arrays */
2601 15032 : eqjoinsel_find_matches(eqproc, collation,
2602 : hashLeft, hashRight,
2603 : false,
2604 : sslot1, sslot2,
2605 : sslot1->nvalues, sslot2->nvalues,
2606 : hasmatch1, hasmatch2,
2607 : p_nmatches, &matchprodfreq);
2608 15032 : nmatches = *p_nmatches;
2609 15032 : CLAMP_PROBABILITY(matchprodfreq);
2610 :
2611 : /* Sum up frequencies of matched and unmatched MCVs */
2612 15032 : matchfreq1 = unmatchfreq1 = 0.0;
2613 423931 : for (i = 0; i < sslot1->nvalues; i++)
2614 : {
2615 408899 : if (hasmatch1[i])
2616 163120 : matchfreq1 += sslot1->numbers[i];
2617 : else
2618 245779 : unmatchfreq1 += sslot1->numbers[i];
2619 : }
2620 15032 : CLAMP_PROBABILITY(matchfreq1);
2621 15032 : CLAMP_PROBABILITY(unmatchfreq1);
2622 15032 : matchfreq2 = unmatchfreq2 = 0.0;
2623 302839 : for (i = 0; i < sslot2->nvalues; i++)
2624 : {
2625 287807 : if (hasmatch2[i])
2626 163120 : matchfreq2 += sslot2->numbers[i];
2627 : else
2628 124687 : unmatchfreq2 += sslot2->numbers[i];
2629 : }
2630 15032 : CLAMP_PROBABILITY(matchfreq2);
2631 15032 : CLAMP_PROBABILITY(unmatchfreq2);
2632 :
2633 : /*
2634 : * Compute total frequency of non-null values that are not in the MCV
2635 : * lists.
2636 : */
2637 15032 : otherfreq1 = 1.0 - nullfrac1 - matchfreq1 - unmatchfreq1;
2638 15032 : otherfreq2 = 1.0 - nullfrac2 - matchfreq2 - unmatchfreq2;
2639 15032 : CLAMP_PROBABILITY(otherfreq1);
2640 15032 : CLAMP_PROBABILITY(otherfreq2);
2641 :
2642 : /*
2643 : * We can estimate the total selectivity from the point of view of
2644 : * relation 1 as: the known selectivity for matched MCVs, plus
2645 : * unmatched MCVs that are assumed to match against random members of
2646 : * relation 2's non-MCV population, plus non-MCV values that are
2647 : * assumed to match against random members of relation 2's unmatched
2648 : * MCVs plus non-MCV values.
2649 : */
2650 15032 : totalsel1 = matchprodfreq;
2651 15032 : if (nd2 > sslot2->nvalues)
2652 3387 : totalsel1 += unmatchfreq1 * otherfreq2 / (nd2 - sslot2->nvalues);
2653 15032 : if (nd2 > nmatches)
2654 6352 : totalsel1 += otherfreq1 * (otherfreq2 + unmatchfreq2) /
2655 6352 : (nd2 - nmatches);
2656 : /* Same estimate from the point of view of relation 2. */
2657 15032 : totalsel2 = matchprodfreq;
2658 15032 : if (nd1 > sslot1->nvalues)
2659 4364 : totalsel2 += unmatchfreq2 * otherfreq1 / (nd1 - sslot1->nvalues);
2660 15032 : if (nd1 > nmatches)
2661 5570 : totalsel2 += otherfreq2 * (otherfreq1 + unmatchfreq1) /
2662 5570 : (nd1 - nmatches);
2663 :
2664 : /*
2665 : * Use the smaller of the two estimates. This can be justified in
2666 : * essentially the same terms as given below for the no-stats case: to
2667 : * a first approximation, we are estimating from the point of view of
2668 : * the relation with smaller nd.
2669 : */
2670 15032 : selec = (totalsel1 < totalsel2) ? totalsel1 : totalsel2;
2671 : }
2672 : else
2673 : {
2674 : /*
2675 : * We do not have MCV lists for both sides. Estimate the join
2676 : * selectivity as MIN(1/nd1,1/nd2)*(1-nullfrac1)*(1-nullfrac2). This
2677 : * is plausible if we assume that the join operator is strict and the
2678 : * non-null values are about equally distributed: a given non-null
2679 : * tuple of rel1 will join to either zero or N2*(1-nullfrac2)/nd2 rows
2680 : * of rel2, so total join rows are at most
2681 : * N1*(1-nullfrac1)*N2*(1-nullfrac2)/nd2 giving a join selectivity of
2682 : * not more than (1-nullfrac1)*(1-nullfrac2)/nd2. By the same logic it
2683 : * is not more than (1-nullfrac1)*(1-nullfrac2)/nd1, so the expression
2684 : * with MIN() is an upper bound. Using the MIN() means we estimate
2685 : * from the point of view of the relation with smaller nd (since the
2686 : * larger nd is determining the MIN). It is reasonable to assume that
2687 : * most tuples in this rel will have join partners, so the bound is
2688 : * probably reasonably tight and should be taken as-is.
2689 : *
2690 : * XXX Can we be smarter if we have an MCV list for just one side? It
2691 : * seems that if we assume equal distribution for the other side, we
2692 : * end up with the same answer anyway.
2693 : */
2694 139294 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2695 139294 : double nullfrac2 = stats2 ? stats2->stanullfrac : 0.0;
2696 :
2697 139294 : selec = (1.0 - nullfrac1) * (1.0 - nullfrac2);
2698 139294 : if (nd1 > nd2)
2699 70433 : selec /= nd1;
2700 : else
2701 68861 : selec /= nd2;
2702 : }
2703 :
2704 154326 : return selec;
2705 : }
2706 :
2707 : /*
2708 : * eqjoinsel_semi --- eqjoinsel for semi join
2709 : *
2710 : * (Also used for anti join, which we are supposed to estimate the same way.)
2711 : * Caller has ensured that vardata1 is the LHS variable; however, eqproc
2712 : * is for the original join operator, which might now need to have the inputs
2713 : * swapped in order to apply correctly. Also, if have_mcvs1 && have_mcvs2
2714 : * then hasmatch1[], hasmatch2[], and *p_nmatches were filled by
2715 : * eqjoinsel_inner.
2716 : */
2717 : static double
2718 7250 : eqjoinsel_semi(FmgrInfo *eqproc, Oid collation,
2719 : Oid hashLeft, Oid hashRight,
2720 : bool op_is_reversed,
2721 : VariableStatData *vardata1, VariableStatData *vardata2,
2722 : double nd1, double nd2,
2723 : bool isdefault1, bool isdefault2,
2724 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2725 : Form_pg_statistic stats1, Form_pg_statistic stats2,
2726 : bool have_mcvs1, bool have_mcvs2,
2727 : bool *hasmatch1, bool *hasmatch2,
2728 : int *p_nmatches,
2729 : RelOptInfo *inner_rel)
2730 : {
2731 : double selec;
2732 :
2733 : /*
2734 : * We clamp nd2 to be not more than what we estimate the inner relation's
2735 : * size to be. This is intuitively somewhat reasonable since obviously
2736 : * there can't be more than that many distinct values coming from the
2737 : * inner rel. The reason for the asymmetry (ie, that we don't clamp nd1
2738 : * likewise) is that this is the only pathway by which restriction clauses
2739 : * applied to the inner rel will affect the join result size estimate,
2740 : * since set_joinrel_size_estimates will multiply SEMI/ANTI selectivity by
2741 : * only the outer rel's size. If we clamped nd1 we'd be double-counting
2742 : * the selectivity of outer-rel restrictions.
2743 : *
2744 : * We can apply this clamping both with respect to the base relation from
2745 : * which the join variable comes (if there is just one), and to the
2746 : * immediate inner input relation of the current join.
2747 : *
2748 : * If we clamp, we can treat nd2 as being a non-default estimate; it's not
2749 : * great, maybe, but it didn't come out of nowhere either. This is most
2750 : * helpful when the inner relation is empty and consequently has no stats.
2751 : */
2752 7250 : if (vardata2->rel)
2753 : {
2754 7247 : if (nd2 >= vardata2->rel->rows)
2755 : {
2756 6171 : nd2 = vardata2->rel->rows;
2757 6171 : isdefault2 = false;
2758 : }
2759 : }
2760 7250 : if (nd2 >= inner_rel->rows)
2761 : {
2762 6142 : nd2 = inner_rel->rows;
2763 6142 : isdefault2 = false;
2764 : }
2765 :
2766 7250 : if (have_mcvs1 && have_mcvs2)
2767 309 : {
2768 : /*
2769 : * We have most-common-value lists for both relations. Run through
2770 : * the lists to see which MCVs actually join to each other with the
2771 : * given operator. This allows us to determine the exact join
2772 : * selectivity for the portion of the relations represented by the MCV
2773 : * lists. We still have to estimate for the remaining population, but
2774 : * in a skewed distribution this gives us a big leg up in accuracy.
2775 : */
2776 309 : double nullfrac1 = stats1->stanullfrac;
2777 : double matchprodfreq,
2778 : matchfreq1,
2779 : uncertainfrac,
2780 : uncertain;
2781 : int i,
2782 : nmatches,
2783 : clamped_nvalues2;
2784 :
2785 : /*
2786 : * The clamping above could have resulted in nd2 being less than
2787 : * sslot2->nvalues; in which case, we assume that precisely the nd2
2788 : * most common values in the relation will appear in the join input,
2789 : * and so compare to only the first nd2 members of the MCV list. Of
2790 : * course this is frequently wrong, but it's the best bet we can make.
2791 : */
2792 309 : clamped_nvalues2 = Min(sslot2->nvalues, nd2);
2793 :
2794 : /*
2795 : * If we did not set clamped_nvalues2 to less than sslot2->nvalues,
2796 : * then the hasmatch1[] and hasmatch2[] match flags computed by
2797 : * eqjoinsel_inner are still perfectly applicable, so we need not
2798 : * re-do the matching work. Note that it does not matter if
2799 : * op_is_reversed: we'd get the same answers.
2800 : *
2801 : * If we did clamp, then a different set of sslot2 values is to be
2802 : * compared, so we have to re-do the matching.
2803 : */
2804 309 : if (clamped_nvalues2 != sslot2->nvalues)
2805 : {
2806 : /* Must re-zero the arrays */
2807 0 : memset(hasmatch1, 0, sslot1->nvalues * sizeof(bool));
2808 0 : memset(hasmatch2, 0, clamped_nvalues2 * sizeof(bool));
2809 : /* Re-fill the match arrays */
2810 0 : eqjoinsel_find_matches(eqproc, collation,
2811 : hashLeft, hashRight,
2812 : op_is_reversed,
2813 : sslot1, sslot2,
2814 : sslot1->nvalues, clamped_nvalues2,
2815 : hasmatch1, hasmatch2,
2816 : p_nmatches, &matchprodfreq);
2817 : }
2818 309 : nmatches = *p_nmatches;
2819 :
2820 : /* Sum up frequencies of matched MCVs */
2821 309 : matchfreq1 = 0.0;
2822 6776 : for (i = 0; i < sslot1->nvalues; i++)
2823 : {
2824 6467 : if (hasmatch1[i])
2825 5740 : matchfreq1 += sslot1->numbers[i];
2826 : }
2827 309 : CLAMP_PROBABILITY(matchfreq1);
2828 :
2829 : /*
2830 : * Now we need to estimate the fraction of relation 1 that has at
2831 : * least one join partner. We know for certain that the matched MCVs
2832 : * do, so that gives us a lower bound, but we're really in the dark
2833 : * about everything else. Our crude approach is: if nd1 <= nd2 then
2834 : * assume all non-null rel1 rows have join partners, else assume for
2835 : * the uncertain rows that a fraction nd2/nd1 have join partners. We
2836 : * can discount the known-matched MCVs from the distinct-values counts
2837 : * before doing the division.
2838 : *
2839 : * Crude as the above is, it's completely useless if we don't have
2840 : * reliable ndistinct values for both sides. Hence, if either nd1 or
2841 : * nd2 is default, punt and assume half of the uncertain rows have
2842 : * join partners.
2843 : */
2844 309 : if (!isdefault1 && !isdefault2)
2845 : {
2846 309 : nd1 -= nmatches;
2847 309 : nd2 -= nmatches;
2848 309 : if (nd1 <= nd2 || nd2 < 0)
2849 291 : uncertainfrac = 1.0;
2850 : else
2851 18 : uncertainfrac = nd2 / nd1;
2852 : }
2853 : else
2854 0 : uncertainfrac = 0.5;
2855 309 : uncertain = 1.0 - matchfreq1 - nullfrac1;
2856 309 : CLAMP_PROBABILITY(uncertain);
2857 309 : selec = matchfreq1 + uncertainfrac * uncertain;
2858 : }
2859 : else
2860 : {
2861 : /*
2862 : * Without MCV lists for both sides, we can only use the heuristic
2863 : * about nd1 vs nd2.
2864 : */
2865 6941 : double nullfrac1 = stats1 ? stats1->stanullfrac : 0.0;
2866 :
2867 6941 : if (!isdefault1 && !isdefault2)
2868 : {
2869 5231 : if (nd1 <= nd2 || nd2 < 0)
2870 2568 : selec = 1.0 - nullfrac1;
2871 : else
2872 2663 : selec = (nd2 / nd1) * (1.0 - nullfrac1);
2873 : }
2874 : else
2875 1710 : selec = 0.5 * (1.0 - nullfrac1);
2876 : }
2877 :
2878 7250 : return selec;
2879 : }
2880 :
2881 : /*
2882 : * Identify matching MCVs for eqjoinsel_inner or eqjoinsel_semi.
2883 : *
2884 : * Inputs:
2885 : * eqproc: FmgrInfo for equality function to use (might be reversed)
2886 : * collation: OID of collation to use
2887 : * hashLeft, hashRight: OIDs of hash functions associated with equality op,
2888 : * or InvalidOid if we're not to use hashing
2889 : * op_is_reversed: indicates that eqproc compares right type to left type
2890 : * sslot1, sslot2: MCV values for the lefthand and righthand inputs
2891 : * nvalues1, nvalues2: number of values to be considered (can be less than
2892 : * sslotN->nvalues, but not more)
2893 : * Outputs:
2894 : * hasmatch1[], hasmatch2[]: pre-zeroed arrays of lengths nvalues1, nvalues2;
2895 : * entries are set to true if that MCV has a match on the other side
2896 : * *p_nmatches: receives number of MCV pairs that match
2897 : * *p_matchprodfreq: receives sum(sslot1->numbers[i] * sslot2->numbers[j])
2898 : * for matching MCVs
2899 : *
2900 : * Note that hashLeft is for the eqproc's left-hand input type, hashRight
2901 : * for its right, regardless of op_is_reversed.
2902 : *
2903 : * Note we assume that each MCV will match at most one member of the other
2904 : * MCV list. If the operator isn't really equality, there could be multiple
2905 : * matches --- but we don't look for them, both for speed and because the
2906 : * math wouldn't add up...
2907 : */
2908 : static void
2909 15032 : eqjoinsel_find_matches(FmgrInfo *eqproc, Oid collation,
2910 : Oid hashLeft, Oid hashRight,
2911 : bool op_is_reversed,
2912 : AttStatsSlot *sslot1, AttStatsSlot *sslot2,
2913 : int nvalues1, int nvalues2,
2914 : bool *hasmatch1, bool *hasmatch2,
2915 : int *p_nmatches, double *p_matchprodfreq)
2916 : {
2917 15032 : LOCAL_FCINFO(fcinfo, 2);
2918 15032 : double matchprodfreq = 0.0;
2919 15032 : int nmatches = 0;
2920 :
2921 : /*
2922 : * Save a few cycles by setting up the fcinfo struct just once. Using
2923 : * FunctionCallInvoke directly also avoids failure if the eqproc returns
2924 : * NULL, though really equality functions should never do that.
2925 : */
2926 15032 : InitFunctionCallInfoData(*fcinfo, eqproc, 2, collation,
2927 : NULL, NULL);
2928 15032 : fcinfo->args[0].isnull = false;
2929 15032 : fcinfo->args[1].isnull = false;
2930 :
2931 15032 : if (OidIsValid(hashLeft) && OidIsValid(hashRight))
2932 967 : {
2933 : /* Use a hash table to speed up the matching */
2934 967 : LOCAL_FCINFO(hash_fcinfo, 1);
2935 : FmgrInfo hash_proc;
2936 : MCVHashContext hashContext;
2937 : MCVHashTable_hash *hashTable;
2938 : AttStatsSlot *statsProbe;
2939 : AttStatsSlot *statsHash;
2940 : bool *hasMatchProbe;
2941 : bool *hasMatchHash;
2942 : int nvaluesProbe;
2943 : int nvaluesHash;
2944 :
2945 : /* Make sure we build the hash table on the smaller array. */
2946 967 : if (sslot1->nvalues >= sslot2->nvalues)
2947 : {
2948 967 : statsProbe = sslot1;
2949 967 : statsHash = sslot2;
2950 967 : hasMatchProbe = hasmatch1;
2951 967 : hasMatchHash = hasmatch2;
2952 967 : nvaluesProbe = nvalues1;
2953 967 : nvaluesHash = nvalues2;
2954 : }
2955 : else
2956 : {
2957 : /* We'll have to reverse the direction of use of the operator. */
2958 0 : op_is_reversed = !op_is_reversed;
2959 0 : statsProbe = sslot2;
2960 0 : statsHash = sslot1;
2961 0 : hasMatchProbe = hasmatch2;
2962 0 : hasMatchHash = hasmatch1;
2963 0 : nvaluesProbe = nvalues2;
2964 0 : nvaluesHash = nvalues1;
2965 : }
2966 :
2967 : /*
2968 : * Build the hash table on the smaller array, using the appropriate
2969 : * hash function for its data type.
2970 : */
2971 967 : fmgr_info(op_is_reversed ? hashLeft : hashRight, &hash_proc);
2972 967 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
2973 : NULL, NULL);
2974 967 : hash_fcinfo->args[0].isnull = false;
2975 :
2976 967 : hashContext.equal_fcinfo = fcinfo;
2977 967 : hashContext.hash_fcinfo = hash_fcinfo;
2978 967 : hashContext.op_is_reversed = op_is_reversed;
2979 967 : hashContext.insert_mode = true;
2980 967 : get_typlenbyval(statsHash->valuetype,
2981 : &hashContext.hash_typlen,
2982 : &hashContext.hash_typbyval);
2983 :
2984 967 : hashTable = MCVHashTable_create(CurrentMemoryContext,
2985 : nvaluesHash,
2986 : &hashContext);
2987 :
2988 97667 : for (int i = 0; i < nvaluesHash; i++)
2989 : {
2990 96700 : bool found = false;
2991 96700 : MCVHashEntry *entry = MCVHashTable_insert(hashTable,
2992 96700 : statsHash->values[i],
2993 : &found);
2994 :
2995 : /*
2996 : * MCVHashTable_insert will only report "found" if the new value
2997 : * is equal to some previous one per datum_image_eq(). That
2998 : * probably shouldn't happen, since we're not expecting duplicates
2999 : * in the MCV list. If we do find a dup, just ignore it, leaving
3000 : * the hash entry's index pointing at the first occurrence. That
3001 : * matches the behavior that the non-hashed code path would have.
3002 : */
3003 96700 : if (likely(!found))
3004 96700 : entry->index = i;
3005 : }
3006 :
3007 : /*
3008 : * Prepare to probe the hash table. If the probe values are of a
3009 : * different data type, then we need to change hash functions. (This
3010 : * code relies on the assumption that since we defined SH_STORE_HASH,
3011 : * simplehash.h will never need to compute hash values for existing
3012 : * hash table entries.)
3013 : */
3014 967 : hashContext.insert_mode = false;
3015 967 : if (hashLeft != hashRight)
3016 : {
3017 0 : fmgr_info(op_is_reversed ? hashRight : hashLeft, &hash_proc);
3018 : /* Resetting hash_fcinfo is probably unnecessary, but be safe */
3019 0 : InitFunctionCallInfoData(*hash_fcinfo, &hash_proc, 1, collation,
3020 : NULL, NULL);
3021 0 : hash_fcinfo->args[0].isnull = false;
3022 : }
3023 :
3024 : /* Look up each probe value in turn. */
3025 97667 : for (int i = 0; i < nvaluesProbe; i++)
3026 : {
3027 96700 : MCVHashEntry *entry = MCVHashTable_lookup(hashTable,
3028 96700 : statsProbe->values[i]);
3029 :
3030 : /* As in the other code path, skip already-matched hash entries */
3031 96700 : if (entry != NULL && !hasMatchHash[entry->index])
3032 : {
3033 34444 : hasMatchHash[entry->index] = hasMatchProbe[i] = true;
3034 34444 : nmatches++;
3035 34444 : matchprodfreq += statsHash->numbers[entry->index] * statsProbe->numbers[i];
3036 : }
3037 : }
3038 :
3039 967 : MCVHashTable_destroy(hashTable);
3040 : }
3041 : else
3042 : {
3043 : /* We're not to use hashing, so do it the O(N^2) way */
3044 : int index1,
3045 : index2;
3046 :
3047 : /* Set up to supply the values in the order the operator expects */
3048 14065 : if (op_is_reversed)
3049 : {
3050 0 : index1 = 1;
3051 0 : index2 = 0;
3052 : }
3053 : else
3054 : {
3055 14065 : index1 = 0;
3056 14065 : index2 = 1;
3057 : }
3058 :
3059 326264 : for (int i = 0; i < nvalues1; i++)
3060 : {
3061 312199 : fcinfo->args[index1].value = sslot1->values[i];
3062 :
3063 6907422 : for (int j = 0; j < nvalues2; j++)
3064 : {
3065 : Datum fresult;
3066 :
3067 6723899 : if (hasmatch2[j])
3068 1949424 : continue;
3069 4774475 : fcinfo->args[index2].value = sslot2->values[j];
3070 4774475 : fcinfo->isnull = false;
3071 4774475 : fresult = FunctionCallInvoke(fcinfo);
3072 4774475 : if (!fcinfo->isnull && DatumGetBool(fresult))
3073 : {
3074 128676 : hasmatch1[i] = hasmatch2[j] = true;
3075 128676 : matchprodfreq += sslot1->numbers[i] * sslot2->numbers[j];
3076 128676 : nmatches++;
3077 128676 : break;
3078 : }
3079 : }
3080 : }
3081 : }
3082 :
3083 15032 : *p_nmatches = nmatches;
3084 15032 : *p_matchprodfreq = matchprodfreq;
3085 15032 : }
3086 :
3087 : /*
3088 : * Support functions for the hash tables used by eqjoinsel_find_matches
3089 : */
3090 : static uint32
3091 193400 : hash_mcv(MCVHashTable_hash *tab, Datum key)
3092 : {
3093 193400 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3094 193400 : FunctionCallInfo fcinfo = context->hash_fcinfo;
3095 : Datum fresult;
3096 :
3097 193400 : fcinfo->args[0].value = key;
3098 193400 : fcinfo->isnull = false;
3099 193400 : fresult = FunctionCallInvoke(fcinfo);
3100 : Assert(!fcinfo->isnull);
3101 193400 : return DatumGetUInt32(fresult);
3102 : }
3103 :
3104 : static bool
3105 34444 : mcvs_equal(MCVHashTable_hash *tab, Datum key0, Datum key1)
3106 : {
3107 34444 : MCVHashContext *context = (MCVHashContext *) tab->private_data;
3108 :
3109 34444 : if (context->insert_mode)
3110 : {
3111 : /*
3112 : * During the insertion step, any comparisons will be between two
3113 : * Datums of the hash table's data type, so if the given operator is
3114 : * cross-type it will be the wrong thing to use. Fortunately, we can
3115 : * use datum_image_eq instead. The MCV values should all be distinct
3116 : * anyway, so it's mostly pro-forma to compare them at all.
3117 : */
3118 0 : return datum_image_eq(key0, key1,
3119 0 : context->hash_typbyval, context->hash_typlen);
3120 : }
3121 : else
3122 : {
3123 34444 : FunctionCallInfo fcinfo = context->equal_fcinfo;
3124 : Datum fresult;
3125 :
3126 : /*
3127 : * Apply the operator the correct way around. Although simplehash.h
3128 : * doesn't document this explicitly, during lookups key0 is from the
3129 : * hash table while key1 is the probe value, so we should compare them
3130 : * in that order only if op_is_reversed.
3131 : */
3132 34444 : if (context->op_is_reversed)
3133 : {
3134 0 : fcinfo->args[0].value = key0;
3135 0 : fcinfo->args[1].value = key1;
3136 : }
3137 : else
3138 : {
3139 34444 : fcinfo->args[0].value = key1;
3140 34444 : fcinfo->args[1].value = key0;
3141 : }
3142 34444 : fcinfo->isnull = false;
3143 34444 : fresult = FunctionCallInvoke(fcinfo);
3144 34444 : return (!fcinfo->isnull && DatumGetBool(fresult));
3145 : }
3146 : }
3147 :
3148 : /*
3149 : * neqjoinsel - Join selectivity of "!="
3150 : */
3151 : Datum
3152 2471 : neqjoinsel(PG_FUNCTION_ARGS)
3153 : {
3154 2471 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3155 2471 : Oid operator = PG_GETARG_OID(1);
3156 2471 : List *args = (List *) PG_GETARG_POINTER(2);
3157 2471 : JoinType jointype = (JoinType) PG_GETARG_INT16(3);
3158 2471 : SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) PG_GETARG_POINTER(4);
3159 2471 : Oid collation = PG_GET_COLLATION();
3160 : float8 result;
3161 :
3162 2471 : if (jointype == JOIN_SEMI || jointype == JOIN_ANTI)
3163 827 : {
3164 : /*
3165 : * For semi-joins, if there is more than one distinct value in the RHS
3166 : * relation then every non-null LHS row must find a row to join since
3167 : * it can only be equal to one of them. We'll assume that there is
3168 : * always more than one distinct RHS value for the sake of stability,
3169 : * though in theory we could have special cases for empty RHS
3170 : * (selectivity = 0) and single-distinct-value RHS (selectivity =
3171 : * fraction of LHS that has the same value as the single RHS value).
3172 : *
3173 : * For anti-joins, if we use the same assumption that there is more
3174 : * than one distinct key in the RHS relation, then every non-null LHS
3175 : * row must be suppressed by the anti-join.
3176 : *
3177 : * So either way, the selectivity estimate should be 1 - nullfrac.
3178 : */
3179 : VariableStatData leftvar;
3180 : VariableStatData rightvar;
3181 : bool reversed;
3182 : HeapTuple statsTuple;
3183 : double nullfrac;
3184 :
3185 827 : get_join_variables(root, args, sjinfo, &leftvar, &rightvar, &reversed);
3186 827 : statsTuple = reversed ? rightvar.statsTuple : leftvar.statsTuple;
3187 827 : if (HeapTupleIsValid(statsTuple))
3188 706 : nullfrac = ((Form_pg_statistic) GETSTRUCT(statsTuple))->stanullfrac;
3189 : else
3190 121 : nullfrac = 0.0;
3191 827 : ReleaseVariableStats(leftvar);
3192 827 : ReleaseVariableStats(rightvar);
3193 :
3194 827 : result = 1.0 - nullfrac;
3195 : }
3196 : else
3197 : {
3198 : /*
3199 : * We want 1 - eqjoinsel() where the equality operator is the one
3200 : * associated with this != operator, that is, its negator.
3201 : */
3202 1644 : Oid eqop = get_negator(operator);
3203 :
3204 1644 : if (eqop)
3205 : {
3206 : result =
3207 1644 : DatumGetFloat8(DirectFunctionCall5Coll(eqjoinsel,
3208 : collation,
3209 : PointerGetDatum(root),
3210 : ObjectIdGetDatum(eqop),
3211 : PointerGetDatum(args),
3212 : Int16GetDatum(jointype),
3213 : PointerGetDatum(sjinfo)));
3214 : }
3215 : else
3216 : {
3217 : /* Use default selectivity (should we raise an error instead?) */
3218 0 : result = DEFAULT_EQ_SEL;
3219 : }
3220 1644 : result = 1.0 - result;
3221 : }
3222 :
3223 2471 : PG_RETURN_FLOAT8(result);
3224 : }
3225 :
3226 : /*
3227 : * scalarltjoinsel - Join selectivity of "<" for scalars
3228 : */
3229 : Datum
3230 162 : scalarltjoinsel(PG_FUNCTION_ARGS)
3231 : {
3232 162 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3233 : }
3234 :
3235 : /*
3236 : * scalarlejoinsel - Join selectivity of "<=" for scalars
3237 : */
3238 : Datum
3239 138 : scalarlejoinsel(PG_FUNCTION_ARGS)
3240 : {
3241 138 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3242 : }
3243 :
3244 : /*
3245 : * scalargtjoinsel - Join selectivity of ">" for scalars
3246 : */
3247 : Datum
3248 138 : scalargtjoinsel(PG_FUNCTION_ARGS)
3249 : {
3250 138 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3251 : }
3252 :
3253 : /*
3254 : * scalargejoinsel - Join selectivity of ">=" for scalars
3255 : */
3256 : Datum
3257 92 : scalargejoinsel(PG_FUNCTION_ARGS)
3258 : {
3259 92 : PG_RETURN_FLOAT8(DEFAULT_INEQ_SEL);
3260 : }
3261 :
3262 :
3263 : /*
3264 : * mergejoinscansel - Scan selectivity of merge join.
3265 : *
3266 : * A merge join will stop as soon as it exhausts either input stream.
3267 : * Therefore, if we can estimate the ranges of both input variables,
3268 : * we can estimate how much of the input will actually be read. This
3269 : * can have a considerable impact on the cost when using indexscans.
3270 : *
3271 : * Also, we can estimate how much of each input has to be read before the
3272 : * first join pair is found, which will affect the join's startup time.
3273 : *
3274 : * clause should be a clause already known to be mergejoinable. opfamily,
3275 : * cmptype, and nulls_first specify the sort ordering being used.
3276 : *
3277 : * The outputs are:
3278 : * *leftstart is set to the fraction of the left-hand variable expected
3279 : * to be scanned before the first join pair is found (0 to 1).
3280 : * *leftend is set to the fraction of the left-hand variable expected
3281 : * to be scanned before the join terminates (0 to 1).
3282 : * *rightstart, *rightend similarly for the right-hand variable.
3283 : */
3284 : void
3285 81576 : mergejoinscansel(PlannerInfo *root, Node *clause,
3286 : Oid opfamily, CompareType cmptype, bool nulls_first,
3287 : Selectivity *leftstart, Selectivity *leftend,
3288 : Selectivity *rightstart, Selectivity *rightend)
3289 : {
3290 : Node *left,
3291 : *right;
3292 : VariableStatData leftvar,
3293 : rightvar;
3294 : Oid opmethod;
3295 : int op_strategy;
3296 : Oid op_lefttype;
3297 : Oid op_righttype;
3298 : Oid opno,
3299 : collation,
3300 : lsortop,
3301 : rsortop,
3302 : lstatop,
3303 : rstatop,
3304 : ltop,
3305 : leop,
3306 : revltop,
3307 : revleop;
3308 : StrategyNumber ltstrat,
3309 : lestrat,
3310 : gtstrat,
3311 : gestrat;
3312 : bool isgt;
3313 : Datum leftmin,
3314 : leftmax,
3315 : rightmin,
3316 : rightmax;
3317 : double selec;
3318 :
3319 : /* Set default results if we can't figure anything out. */
3320 : /* XXX should default "start" fraction be a bit more than 0? */
3321 81576 : *leftstart = *rightstart = 0.0;
3322 81576 : *leftend = *rightend = 1.0;
3323 :
3324 : /* Deconstruct the merge clause */
3325 81576 : if (!is_opclause(clause))
3326 0 : return; /* shouldn't happen */
3327 81576 : opno = ((OpExpr *) clause)->opno;
3328 81576 : collation = ((OpExpr *) clause)->inputcollid;
3329 81576 : left = get_leftop((Expr *) clause);
3330 81576 : right = get_rightop((Expr *) clause);
3331 81576 : if (!right)
3332 0 : return; /* shouldn't happen */
3333 :
3334 : /* Look for stats for the inputs */
3335 81576 : examine_variable(root, left, 0, &leftvar);
3336 81576 : examine_variable(root, right, 0, &rightvar);
3337 :
3338 81576 : opmethod = get_opfamily_method(opfamily);
3339 :
3340 : /* Extract the operator's declared left/right datatypes */
3341 81576 : get_op_opfamily_properties(opno, opfamily, false,
3342 : &op_strategy,
3343 : &op_lefttype,
3344 : &op_righttype);
3345 : Assert(IndexAmTranslateStrategy(op_strategy, opmethod, opfamily, true) == COMPARE_EQ);
3346 :
3347 : /*
3348 : * Look up the various operators we need. If we don't find them all, it
3349 : * probably means the opfamily is broken, but we just fail silently.
3350 : *
3351 : * Note: we expect that pg_statistic histograms will be sorted by the '<'
3352 : * operator, regardless of which sort direction we are considering.
3353 : */
3354 81576 : switch (cmptype)
3355 : {
3356 81558 : case COMPARE_LT:
3357 81558 : isgt = false;
3358 81558 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3359 81558 : lestrat = IndexAmTranslateCompareType(COMPARE_LE, opmethod, opfamily, true);
3360 81558 : if (op_lefttype == op_righttype)
3361 : {
3362 : /* easy case */
3363 80532 : ltop = get_opfamily_member(opfamily,
3364 : op_lefttype, op_righttype,
3365 : ltstrat);
3366 80532 : leop = get_opfamily_member(opfamily,
3367 : op_lefttype, op_righttype,
3368 : lestrat);
3369 80532 : lsortop = ltop;
3370 80532 : rsortop = ltop;
3371 80532 : lstatop = lsortop;
3372 80532 : rstatop = rsortop;
3373 80532 : revltop = ltop;
3374 80532 : revleop = leop;
3375 : }
3376 : else
3377 : {
3378 1026 : ltop = get_opfamily_member(opfamily,
3379 : op_lefttype, op_righttype,
3380 : ltstrat);
3381 1026 : leop = get_opfamily_member(opfamily,
3382 : op_lefttype, op_righttype,
3383 : lestrat);
3384 1026 : lsortop = get_opfamily_member(opfamily,
3385 : op_lefttype, op_lefttype,
3386 : ltstrat);
3387 1026 : rsortop = get_opfamily_member(opfamily,
3388 : op_righttype, op_righttype,
3389 : ltstrat);
3390 1026 : lstatop = lsortop;
3391 1026 : rstatop = rsortop;
3392 1026 : revltop = get_opfamily_member(opfamily,
3393 : op_righttype, op_lefttype,
3394 : ltstrat);
3395 1026 : revleop = get_opfamily_member(opfamily,
3396 : op_righttype, op_lefttype,
3397 : lestrat);
3398 : }
3399 81558 : break;
3400 18 : case COMPARE_GT:
3401 : /* descending-order case */
3402 18 : isgt = true;
3403 18 : ltstrat = IndexAmTranslateCompareType(COMPARE_LT, opmethod, opfamily, true);
3404 18 : gtstrat = IndexAmTranslateCompareType(COMPARE_GT, opmethod, opfamily, true);
3405 18 : gestrat = IndexAmTranslateCompareType(COMPARE_GE, opmethod, opfamily, true);
3406 18 : if (op_lefttype == op_righttype)
3407 : {
3408 : /* easy case */
3409 18 : ltop = get_opfamily_member(opfamily,
3410 : op_lefttype, op_righttype,
3411 : gtstrat);
3412 18 : leop = get_opfamily_member(opfamily,
3413 : op_lefttype, op_righttype,
3414 : gestrat);
3415 18 : lsortop = ltop;
3416 18 : rsortop = ltop;
3417 18 : lstatop = get_opfamily_member(opfamily,
3418 : op_lefttype, op_lefttype,
3419 : ltstrat);
3420 18 : rstatop = lstatop;
3421 18 : revltop = ltop;
3422 18 : revleop = leop;
3423 : }
3424 : else
3425 : {
3426 0 : ltop = get_opfamily_member(opfamily,
3427 : op_lefttype, op_righttype,
3428 : gtstrat);
3429 0 : leop = get_opfamily_member(opfamily,
3430 : op_lefttype, op_righttype,
3431 : gestrat);
3432 0 : lsortop = get_opfamily_member(opfamily,
3433 : op_lefttype, op_lefttype,
3434 : gtstrat);
3435 0 : rsortop = get_opfamily_member(opfamily,
3436 : op_righttype, op_righttype,
3437 : gtstrat);
3438 0 : lstatop = get_opfamily_member(opfamily,
3439 : op_lefttype, op_lefttype,
3440 : ltstrat);
3441 0 : rstatop = get_opfamily_member(opfamily,
3442 : op_righttype, op_righttype,
3443 : ltstrat);
3444 0 : revltop = get_opfamily_member(opfamily,
3445 : op_righttype, op_lefttype,
3446 : gtstrat);
3447 0 : revleop = get_opfamily_member(opfamily,
3448 : op_righttype, op_lefttype,
3449 : gestrat);
3450 : }
3451 18 : break;
3452 0 : default:
3453 0 : goto fail; /* shouldn't get here */
3454 : }
3455 :
3456 81576 : if (!OidIsValid(lsortop) ||
3457 81576 : !OidIsValid(rsortop) ||
3458 81576 : !OidIsValid(lstatop) ||
3459 81576 : !OidIsValid(rstatop) ||
3460 81570 : !OidIsValid(ltop) ||
3461 81570 : !OidIsValid(leop) ||
3462 81570 : !OidIsValid(revltop) ||
3463 : !OidIsValid(revleop))
3464 6 : goto fail; /* insufficient info in catalogs */
3465 :
3466 : /* Try to get ranges of both inputs */
3467 81570 : if (!isgt)
3468 : {
3469 81552 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3470 : &leftmin, &leftmax))
3471 24037 : goto fail; /* no range available from stats */
3472 57515 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3473 : &rightmin, &rightmax))
3474 13405 : goto fail; /* no range available from stats */
3475 : }
3476 : else
3477 : {
3478 : /* need to swap the max and min */
3479 18 : if (!get_variable_range(root, &leftvar, lstatop, collation,
3480 : &leftmax, &leftmin))
3481 15 : goto fail; /* no range available from stats */
3482 3 : if (!get_variable_range(root, &rightvar, rstatop, collation,
3483 : &rightmax, &rightmin))
3484 0 : goto fail; /* no range available from stats */
3485 : }
3486 :
3487 : /*
3488 : * Now, the fraction of the left variable that will be scanned is the
3489 : * fraction that's <= the right-side maximum value. But only believe
3490 : * non-default estimates, else stick with our 1.0.
3491 : */
3492 44113 : selec = scalarineqsel(root, leop, isgt, true, collation, &leftvar,
3493 : rightmax, op_righttype);
3494 44113 : if (selec != DEFAULT_INEQ_SEL)
3495 44110 : *leftend = selec;
3496 :
3497 : /* And similarly for the right variable. */
3498 44113 : selec = scalarineqsel(root, revleop, isgt, true, collation, &rightvar,
3499 : leftmax, op_lefttype);
3500 44113 : if (selec != DEFAULT_INEQ_SEL)
3501 44113 : *rightend = selec;
3502 :
3503 : /*
3504 : * Only one of the two "end" fractions can really be less than 1.0;
3505 : * believe the smaller estimate and reset the other one to exactly 1.0. If
3506 : * we get exactly equal estimates (as can easily happen with self-joins),
3507 : * believe neither.
3508 : */
3509 44113 : if (*leftend > *rightend)
3510 13617 : *leftend = 1.0;
3511 30496 : else if (*leftend < *rightend)
3512 17977 : *rightend = 1.0;
3513 : else
3514 12519 : *leftend = *rightend = 1.0;
3515 :
3516 : /*
3517 : * Also, the fraction of the left variable that will be scanned before the
3518 : * first join pair is found is the fraction that's < the right-side
3519 : * minimum value. But only believe non-default estimates, else stick with
3520 : * our own default.
3521 : */
3522 44113 : selec = scalarineqsel(root, ltop, isgt, false, collation, &leftvar,
3523 : rightmin, op_righttype);
3524 44113 : if (selec != DEFAULT_INEQ_SEL)
3525 44113 : *leftstart = selec;
3526 :
3527 : /* And similarly for the right variable. */
3528 44113 : selec = scalarineqsel(root, revltop, isgt, false, collation, &rightvar,
3529 : leftmin, op_lefttype);
3530 44113 : if (selec != DEFAULT_INEQ_SEL)
3531 44113 : *rightstart = selec;
3532 :
3533 : /*
3534 : * Only one of the two "start" fractions can really be more than zero;
3535 : * believe the larger estimate and reset the other one to exactly 0.0. If
3536 : * we get exactly equal estimates (as can easily happen with self-joins),
3537 : * believe neither.
3538 : */
3539 44113 : if (*leftstart < *rightstart)
3540 10020 : *leftstart = 0.0;
3541 34093 : else if (*leftstart > *rightstart)
3542 12624 : *rightstart = 0.0;
3543 : else
3544 21469 : *leftstart = *rightstart = 0.0;
3545 :
3546 : /*
3547 : * If the sort order is nulls-first, we're going to have to skip over any
3548 : * nulls too. These would not have been counted by scalarineqsel, and we
3549 : * can safely add in this fraction regardless of whether we believe
3550 : * scalarineqsel's results or not. But be sure to clamp the sum to 1.0!
3551 : */
3552 44113 : if (nulls_first)
3553 : {
3554 : Form_pg_statistic stats;
3555 :
3556 3 : if (HeapTupleIsValid(leftvar.statsTuple))
3557 : {
3558 3 : stats = (Form_pg_statistic) GETSTRUCT(leftvar.statsTuple);
3559 3 : *leftstart += stats->stanullfrac;
3560 3 : CLAMP_PROBABILITY(*leftstart);
3561 3 : *leftend += stats->stanullfrac;
3562 3 : CLAMP_PROBABILITY(*leftend);
3563 : }
3564 3 : if (HeapTupleIsValid(rightvar.statsTuple))
3565 : {
3566 3 : stats = (Form_pg_statistic) GETSTRUCT(rightvar.statsTuple);
3567 3 : *rightstart += stats->stanullfrac;
3568 3 : CLAMP_PROBABILITY(*rightstart);
3569 3 : *rightend += stats->stanullfrac;
3570 3 : CLAMP_PROBABILITY(*rightend);
3571 : }
3572 : }
3573 :
3574 : /* Disbelieve start >= end, just in case that can happen */
3575 44113 : if (*leftstart >= *leftend)
3576 : {
3577 89 : *leftstart = 0.0;
3578 89 : *leftend = 1.0;
3579 : }
3580 44113 : if (*rightstart >= *rightend)
3581 : {
3582 752 : *rightstart = 0.0;
3583 752 : *rightend = 1.0;
3584 : }
3585 :
3586 43361 : fail:
3587 81576 : ReleaseVariableStats(leftvar);
3588 81576 : ReleaseVariableStats(rightvar);
3589 : }
3590 :
3591 :
3592 : /*
3593 : * matchingsel -- generic matching-operator selectivity support
3594 : *
3595 : * Use these for any operators that (a) are on data types for which we collect
3596 : * standard statistics, and (b) have behavior for which the default estimate
3597 : * (twice DEFAULT_EQ_SEL) is sane. Typically that is good for match-like
3598 : * operators.
3599 : */
3600 :
3601 : Datum
3602 565 : matchingsel(PG_FUNCTION_ARGS)
3603 : {
3604 565 : PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);
3605 565 : Oid operator = PG_GETARG_OID(1);
3606 565 : List *args = (List *) PG_GETARG_POINTER(2);
3607 565 : int varRelid = PG_GETARG_INT32(3);
3608 565 : Oid collation = PG_GET_COLLATION();
3609 : double selec;
3610 :
3611 : /* Use generic restriction selectivity logic. */
3612 565 : selec = generic_restriction_selectivity(root, operator, collation,
3613 : args, varRelid,
3614 : DEFAULT_MATCHING_SEL);
3615 :
3616 565 : PG_RETURN_FLOAT8((float8) selec);
3617 : }
3618 :
3619 : Datum
3620 3 : matchingjoinsel(PG_FUNCTION_ARGS)
3621 : {
3622 : /* Just punt, for the moment. */
3623 3 : PG_RETURN_FLOAT8(DEFAULT_MATCHING_SEL);
3624 : }
3625 :
3626 :
3627 : /*
3628 : * Helper routine for estimate_num_groups: add an item to a list of
3629 : * GroupVarInfos, but only if it's not known equal to any of the existing
3630 : * entries.
3631 : */
3632 : typedef struct
3633 : {
3634 : Node *var; /* might be an expression, not just a Var */
3635 : RelOptInfo *rel; /* relation it belongs to */
3636 : double ndistinct; /* # distinct values */
3637 : bool isdefault; /* true if DEFAULT_NUM_DISTINCT was used */
3638 : } GroupVarInfo;
3639 :
3640 : static List *
3641 232407 : add_unique_group_var(PlannerInfo *root, List *varinfos,
3642 : Node *var, VariableStatData *vardata)
3643 : {
3644 : GroupVarInfo *varinfo;
3645 : double ndistinct;
3646 : bool isdefault;
3647 : ListCell *lc;
3648 :
3649 232407 : ndistinct = get_variable_numdistinct(vardata, &isdefault);
3650 :
3651 : /*
3652 : * The nullingrels bits within the var could cause the same var to be
3653 : * counted multiple times if it's marked with different nullingrels. They
3654 : * could also prevent us from matching the var to the expressions in
3655 : * extended statistics (see estimate_multivariate_ndistinct). So strip
3656 : * them out first.
3657 : */
3658 232407 : var = remove_nulling_relids(var, root->outer_join_rels, NULL);
3659 :
3660 280989 : foreach(lc, varinfos)
3661 : {
3662 49155 : varinfo = (GroupVarInfo *) lfirst(lc);
3663 :
3664 : /* Drop exact duplicates */
3665 49155 : if (equal(var, varinfo->var))
3666 573 : return varinfos;
3667 :
3668 : /*
3669 : * Drop known-equal vars, but only if they belong to different
3670 : * relations (see comments for estimate_num_groups). We aren't too
3671 : * fussy about the semantics of "equal" here.
3672 : */
3673 52255 : if (vardata->rel != varinfo->rel &&
3674 3535 : exprs_known_equal(root, var, varinfo->var, InvalidOid))
3675 : {
3676 150 : if (varinfo->ndistinct <= ndistinct)
3677 : {
3678 : /* Keep older item, forget new one */
3679 138 : return varinfos;
3680 : }
3681 : else
3682 : {
3683 : /* Delete the older item */
3684 12 : varinfos = foreach_delete_current(varinfos, lc);
3685 : }
3686 : }
3687 : }
3688 :
3689 231834 : varinfo = palloc_object(GroupVarInfo);
3690 :
3691 231834 : varinfo->var = var;
3692 231834 : varinfo->rel = vardata->rel;
3693 231834 : varinfo->ndistinct = ndistinct;
3694 231834 : varinfo->isdefault = isdefault;
3695 231834 : varinfos = lappend(varinfos, varinfo);
3696 231834 : return varinfos;
3697 : }
3698 :
3699 : /*
3700 : * estimate_num_groups - Estimate number of groups in a grouped query
3701 : *
3702 : * Given a query having a GROUP BY clause, estimate how many groups there
3703 : * will be --- ie, the number of distinct combinations of the GROUP BY
3704 : * expressions.
3705 : *
3706 : * This routine is also used to estimate the number of rows emitted by
3707 : * a DISTINCT filtering step; that is an isomorphic problem. (Note:
3708 : * actually, we only use it for DISTINCT when there's no grouping or
3709 : * aggregation ahead of the DISTINCT.)
3710 : *
3711 : * Inputs:
3712 : * root - the query
3713 : * groupExprs - list of expressions being grouped by
3714 : * input_rows - number of rows estimated to arrive at the group/unique
3715 : * filter step
3716 : * pgset - NULL, or a List** pointing to a grouping set to filter the
3717 : * groupExprs against
3718 : *
3719 : * Outputs:
3720 : * estinfo - When passed as non-NULL, the function will set bits in the
3721 : * "flags" field in order to provide callers with additional information
3722 : * about the estimation. Currently, we only set the SELFLAG_USED_DEFAULT
3723 : * bit if we used any default values in the estimation.
3724 : *
3725 : * Given the lack of any cross-correlation statistics in the system, it's
3726 : * impossible to do anything really trustworthy with GROUP BY conditions
3727 : * involving multiple Vars. We should however avoid assuming the worst
3728 : * case (all possible cross-product terms actually appear as groups) since
3729 : * very often the grouped-by Vars are highly correlated. Our current approach
3730 : * is as follows:
3731 : * 1. Expressions yielding boolean are assumed to contribute two groups,
3732 : * independently of their content, and are ignored in the subsequent
3733 : * steps. This is mainly because tests like "col IS NULL" break the
3734 : * heuristic used in step 2 especially badly.
3735 : * 2. Reduce the given expressions to a list of unique Vars used. For
3736 : * example, GROUP BY a, a + b is treated the same as GROUP BY a, b.
3737 : * It is clearly correct not to count the same Var more than once.
3738 : * It is also reasonable to treat f(x) the same as x: f() cannot
3739 : * increase the number of distinct values (unless it is volatile,
3740 : * which we consider unlikely for grouping), but it probably won't
3741 : * reduce the number of distinct values much either.
3742 : * As a special case, if a GROUP BY expression can be matched to an
3743 : * expressional index for which we have statistics, then we treat the
3744 : * whole expression as though it were just a Var.
3745 : * 3. If the list contains Vars of different relations that are known equal
3746 : * due to equivalence classes, then drop all but one of the Vars from each
3747 : * known-equal set, keeping the one with smallest estimated # of values
3748 : * (since the extra values of the others can't appear in joined rows).
3749 : * Note the reason we only consider Vars of different relations is that
3750 : * if we considered ones of the same rel, we'd be double-counting the
3751 : * restriction selectivity of the equality in the next step.
3752 : * 4. For Vars within a single source rel, we multiply together the numbers
3753 : * of values, clamp to the number of rows in the rel (divided by 10 if
3754 : * more than one Var), and then multiply by a factor based on the
3755 : * selectivity of the restriction clauses for that rel. When there's
3756 : * more than one Var, the initial product is probably too high (it's the
3757 : * worst case) but clamping to a fraction of the rel's rows seems to be a
3758 : * helpful heuristic for not letting the estimate get out of hand. (The
3759 : * factor of 10 is derived from pre-Postgres-7.4 practice.) The factor
3760 : * we multiply by to adjust for the restriction selectivity assumes that
3761 : * the restriction clauses are independent of the grouping, which may not
3762 : * be a valid assumption, but it's hard to do better.
3763 : * 5. If there are Vars from multiple rels, we repeat step 4 for each such
3764 : * rel, and multiply the results together.
3765 : * Note that rels not containing grouped Vars are ignored completely, as are
3766 : * join clauses. Such rels cannot increase the number of groups, and we
3767 : * assume such clauses do not reduce the number either (somewhat bogus,
3768 : * but we don't have the info to do better).
3769 : */
3770 : double
3771 201306 : estimate_num_groups(PlannerInfo *root, List *groupExprs, double input_rows,
3772 : List **pgset, EstimationInfo *estinfo)
3773 : {
3774 201306 : List *varinfos = NIL;
3775 201306 : double srf_multiplier = 1.0;
3776 : double numdistinct;
3777 : ListCell *l;
3778 : int i;
3779 :
3780 : /* Zero the estinfo output parameter, if non-NULL */
3781 201306 : if (estinfo != NULL)
3782 175291 : memset(estinfo, 0, sizeof(EstimationInfo));
3783 :
3784 : /*
3785 : * We don't ever want to return an estimate of zero groups, as that tends
3786 : * to lead to division-by-zero and other unpleasantness. The input_rows
3787 : * estimate is usually already at least 1, but clamp it just in case it
3788 : * isn't.
3789 : */
3790 201306 : input_rows = clamp_row_est(input_rows);
3791 :
3792 : /*
3793 : * If no grouping columns, there's exactly one group. (This can't happen
3794 : * for normal cases with GROUP BY or DISTINCT, but it is possible for
3795 : * corner cases with set operations.)
3796 : */
3797 201306 : if (groupExprs == NIL || (pgset && *pgset == NIL))
3798 581 : return 1.0;
3799 :
3800 : /*
3801 : * Count groups derived from boolean grouping expressions. For other
3802 : * expressions, find the unique Vars used, treating an expression as a Var
3803 : * if we can find stats for it. For each one, record the statistical
3804 : * estimate of number of distinct values (total in its table, without
3805 : * regard for filtering).
3806 : */
3807 200725 : numdistinct = 1.0;
3808 :
3809 200725 : i = 0;
3810 432452 : foreach(l, groupExprs)
3811 : {
3812 231751 : Node *groupexpr = (Node *) lfirst(l);
3813 : double this_srf_multiplier;
3814 : VariableStatData vardata;
3815 : List *varshere;
3816 : ListCell *l2;
3817 :
3818 : /* is expression in this grouping set? */
3819 231751 : if (pgset && !list_member_int(*pgset, i++))
3820 192379 : continue;
3821 :
3822 : /*
3823 : * Set-returning functions in grouping columns are a bit problematic.
3824 : * The code below will effectively ignore their SRF nature and come up
3825 : * with a numdistinct estimate as though they were scalar functions.
3826 : * We compensate by scaling up the end result by the largest SRF
3827 : * rowcount estimate. (This will be an overestimate if the SRF
3828 : * produces multiple copies of any output value, but it seems best to
3829 : * assume the SRF's outputs are distinct. In any case, it's probably
3830 : * pointless to worry too much about this without much better
3831 : * estimates for SRF output rowcounts than we have today.)
3832 : */
3833 231345 : this_srf_multiplier = expression_returns_set_rows(root, groupexpr);
3834 231345 : if (srf_multiplier < this_srf_multiplier)
3835 78 : srf_multiplier = this_srf_multiplier;
3836 :
3837 : /* Short-circuit for expressions returning boolean */
3838 231345 : if (exprType(groupexpr) == BOOLOID)
3839 : {
3840 183 : numdistinct *= 2.0;
3841 183 : continue;
3842 : }
3843 :
3844 : /*
3845 : * If examine_variable is able to deduce anything about the GROUP BY
3846 : * expression, treat it as a single variable even if it's really more
3847 : * complicated.
3848 : *
3849 : * XXX This has the consequence that if there's a statistics object on
3850 : * the expression, we don't split it into individual Vars. This
3851 : * affects our selection of statistics in
3852 : * estimate_multivariate_ndistinct, because it's probably better to
3853 : * use more accurate estimate for each expression and treat them as
3854 : * independent, than to combine estimates for the extracted variables
3855 : * when we don't know how that relates to the expressions.
3856 : */
3857 231162 : examine_variable(root, groupexpr, 0, &vardata);
3858 231162 : if (HeapTupleIsValid(vardata.statsTuple) || vardata.isunique)
3859 : {
3860 191446 : varinfos = add_unique_group_var(root, varinfos,
3861 : groupexpr, &vardata);
3862 191446 : ReleaseVariableStats(vardata);
3863 191446 : continue;
3864 : }
3865 39716 : ReleaseVariableStats(vardata);
3866 :
3867 : /*
3868 : * Else pull out the component Vars. Handle PlaceHolderVars by
3869 : * recursing into their arguments (effectively assuming that the
3870 : * PlaceHolderVar doesn't change the number of groups, which boils
3871 : * down to ignoring the possible addition of nulls to the result set).
3872 : */
3873 39716 : varshere = pull_var_clause(groupexpr,
3874 : PVC_RECURSE_AGGREGATES |
3875 : PVC_RECURSE_WINDOWFUNCS |
3876 : PVC_RECURSE_PLACEHOLDERS);
3877 :
3878 : /*
3879 : * If we find any variable-free GROUP BY item, then either it is a
3880 : * constant (and we can ignore it) or it contains a volatile function;
3881 : * in the latter case we punt and assume that each input row will
3882 : * yield a distinct group.
3883 : */
3884 39716 : if (varshere == NIL)
3885 : {
3886 368 : if (contain_volatile_functions(groupexpr))
3887 24 : return input_rows;
3888 344 : continue;
3889 : }
3890 :
3891 : /*
3892 : * Else add variables to varinfos list
3893 : */
3894 80309 : foreach(l2, varshere)
3895 : {
3896 40961 : Node *var = (Node *) lfirst(l2);
3897 :
3898 40961 : examine_variable(root, var, 0, &vardata);
3899 40961 : varinfos = add_unique_group_var(root, varinfos, var, &vardata);
3900 40961 : ReleaseVariableStats(vardata);
3901 : }
3902 : }
3903 :
3904 : /*
3905 : * If now no Vars, we must have an all-constant or all-boolean GROUP BY
3906 : * list.
3907 : */
3908 200701 : if (varinfos == NIL)
3909 : {
3910 : /* Apply SRF multiplier as we would do in the long path */
3911 235 : numdistinct *= srf_multiplier;
3912 : /* Round off */
3913 235 : numdistinct = ceil(numdistinct);
3914 : /* Guard against out-of-range answers */
3915 235 : if (numdistinct > input_rows)
3916 43 : numdistinct = input_rows;
3917 235 : if (numdistinct < 1.0)
3918 0 : numdistinct = 1.0;
3919 235 : return numdistinct;
3920 : }
3921 :
3922 : /*
3923 : * Group Vars by relation and estimate total numdistinct.
3924 : *
3925 : * For each iteration of the outer loop, we process the frontmost Var in
3926 : * varinfos, plus all other Vars in the same relation. We remove these
3927 : * Vars from the newvarinfos list for the next iteration. This is the
3928 : * easiest way to group Vars of same rel together.
3929 : */
3930 : do
3931 : {
3932 201946 : GroupVarInfo *varinfo1 = (GroupVarInfo *) linitial(varinfos);
3933 201946 : RelOptInfo *rel = varinfo1->rel;
3934 201946 : double reldistinct = 1;
3935 201946 : double relmaxndistinct = reldistinct;
3936 201946 : int relvarcount = 0;
3937 201946 : List *newvarinfos = NIL;
3938 201946 : List *relvarinfos = NIL;
3939 :
3940 : /*
3941 : * Split the list of varinfos in two - one for the current rel, one
3942 : * for remaining Vars on other rels.
3943 : */
3944 201946 : relvarinfos = lappend(relvarinfos, varinfo1);
3945 234647 : for_each_from(l, varinfos, 1)
3946 : {
3947 32701 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3948 :
3949 32701 : if (varinfo2->rel == varinfo1->rel)
3950 : {
3951 : /* varinfos on current rel */
3952 29876 : relvarinfos = lappend(relvarinfos, varinfo2);
3953 : }
3954 : else
3955 : {
3956 : /* not time to process varinfo2 yet */
3957 2825 : newvarinfos = lappend(newvarinfos, varinfo2);
3958 : }
3959 : }
3960 :
3961 : /*
3962 : * Get the numdistinct estimate for the Vars of this rel. We
3963 : * iteratively search for multivariate n-distinct with maximum number
3964 : * of vars; assuming that each var group is independent of the others,
3965 : * we multiply them together. Any remaining relvarinfos after no more
3966 : * multivariate matches are found are assumed independent too, so
3967 : * their individual ndistinct estimates are multiplied also.
3968 : *
3969 : * While iterating, count how many separate numdistinct values we
3970 : * apply. We apply a fudge factor below, but only if we multiplied
3971 : * more than one such values.
3972 : */
3973 403955 : while (relvarinfos)
3974 : {
3975 : double mvndistinct;
3976 :
3977 202009 : if (estimate_multivariate_ndistinct(root, rel, &relvarinfos,
3978 : &mvndistinct))
3979 : {
3980 207 : reldistinct *= mvndistinct;
3981 207 : if (relmaxndistinct < mvndistinct)
3982 201 : relmaxndistinct = mvndistinct;
3983 207 : relvarcount++;
3984 : }
3985 : else
3986 : {
3987 433186 : foreach(l, relvarinfos)
3988 : {
3989 231384 : GroupVarInfo *varinfo2 = (GroupVarInfo *) lfirst(l);
3990 :
3991 231384 : reldistinct *= varinfo2->ndistinct;
3992 231384 : if (relmaxndistinct < varinfo2->ndistinct)
3993 205562 : relmaxndistinct = varinfo2->ndistinct;
3994 231384 : relvarcount++;
3995 :
3996 : /*
3997 : * When varinfo2's isdefault is set then we'd better set
3998 : * the SELFLAG_USED_DEFAULT bit in the EstimationInfo.
3999 : */
4000 231384 : if (estinfo != NULL && varinfo2->isdefault)
4001 9782 : estinfo->flags |= SELFLAG_USED_DEFAULT;
4002 : }
4003 :
4004 : /* we're done with this relation */
4005 201802 : relvarinfos = NIL;
4006 : }
4007 : }
4008 :
4009 : /*
4010 : * Sanity check --- don't divide by zero if empty relation.
4011 : */
4012 : Assert(IS_SIMPLE_REL(rel));
4013 201946 : if (rel->tuples > 0)
4014 : {
4015 : /*
4016 : * Clamp to size of rel, or size of rel / 10 if multiple Vars. The
4017 : * fudge factor is because the Vars are probably correlated but we
4018 : * don't know by how much. We should never clamp to less than the
4019 : * largest ndistinct value for any of the Vars, though, since
4020 : * there will surely be at least that many groups.
4021 : */
4022 201419 : double clamp = rel->tuples;
4023 :
4024 201419 : if (relvarcount > 1)
4025 : {
4026 27151 : clamp *= 0.1;
4027 27151 : if (clamp < relmaxndistinct)
4028 : {
4029 25812 : clamp = relmaxndistinct;
4030 : /* for sanity in case some ndistinct is too large: */
4031 25812 : if (clamp > rel->tuples)
4032 39 : clamp = rel->tuples;
4033 : }
4034 : }
4035 201419 : if (reldistinct > clamp)
4036 22015 : reldistinct = clamp;
4037 :
4038 : /*
4039 : * Update the estimate based on the restriction selectivity,
4040 : * guarding against division by zero when reldistinct is zero.
4041 : * Also skip this if we know that we are returning all rows.
4042 : */
4043 201419 : if (reldistinct > 0 && rel->rows < rel->tuples)
4044 : {
4045 : /*
4046 : * Given a table containing N rows with n distinct values in a
4047 : * uniform distribution, if we select p rows at random then
4048 : * the expected number of distinct values selected is
4049 : *
4050 : * n * (1 - product((N-N/n-i)/(N-i), i=0..p-1))
4051 : *
4052 : * = n * (1 - (N-N/n)! / (N-N/n-p)! * (N-p)! / N!)
4053 : *
4054 : * See "Approximating block accesses in database
4055 : * organizations", S. B. Yao, Communications of the ACM,
4056 : * Volume 20 Issue 4, April 1977 Pages 260-261.
4057 : *
4058 : * Alternatively, re-arranging the terms from the factorials,
4059 : * this may be written as
4060 : *
4061 : * n * (1 - product((N-p-i)/(N-i), i=0..N/n-1))
4062 : *
4063 : * This form of the formula is more efficient to compute in
4064 : * the common case where p is larger than N/n. Additionally,
4065 : * as pointed out by Dell'Era, if i << N for all terms in the
4066 : * product, it can be approximated by
4067 : *
4068 : * n * (1 - ((N-p)/N)^(N/n))
4069 : *
4070 : * See "Expected distinct values when selecting from a bag
4071 : * without replacement", Alberto Dell'Era,
4072 : * http://www.adellera.it/investigations/distinct_balls/.
4073 : *
4074 : * The condition i << N is equivalent to n >> 1, so this is a
4075 : * good approximation when the number of distinct values in
4076 : * the table is large. It turns out that this formula also
4077 : * works well even when n is small.
4078 : */
4079 66325 : reldistinct *=
4080 66325 : (1 - pow((rel->tuples - rel->rows) / rel->tuples,
4081 66325 : rel->tuples / reldistinct));
4082 : }
4083 201419 : reldistinct = clamp_row_est(reldistinct);
4084 :
4085 : /*
4086 : * Update estimate of total distinct groups.
4087 : */
4088 201419 : numdistinct *= reldistinct;
4089 : }
4090 :
4091 201946 : varinfos = newvarinfos;
4092 201946 : } while (varinfos != NIL);
4093 :
4094 : /* Now we can account for the effects of any SRFs */
4095 200466 : numdistinct *= srf_multiplier;
4096 :
4097 : /* Round off */
4098 200466 : numdistinct = ceil(numdistinct);
4099 :
4100 : /* Guard against out-of-range answers */
4101 200466 : if (numdistinct > input_rows)
4102 42308 : numdistinct = input_rows;
4103 200466 : if (numdistinct < 1.0)
4104 0 : numdistinct = 1.0;
4105 :
4106 200466 : return numdistinct;
4107 : }
4108 :
4109 : /*
4110 : * Try to estimate the bucket size of the hash join inner side when the join
4111 : * condition contains two or more clauses by employing extended statistics.
4112 : *
4113 : * The main idea of this approach is that the distinct value generated by
4114 : * multivariate estimation on two or more columns would provide less bucket size
4115 : * than estimation on one separate column.
4116 : *
4117 : * IMPORTANT: It is crucial to synchronize the approach of combining different
4118 : * estimations with the caller's method.
4119 : *
4120 : * Return a list of clauses that didn't fetch any extended statistics.
4121 : */
4122 : List *
4123 250175 : estimate_multivariate_bucketsize(PlannerInfo *root, RelOptInfo *inner,
4124 : List *hashclauses,
4125 : Selectivity *innerbucketsize)
4126 : {
4127 : List *clauses;
4128 : List *otherclauses;
4129 : double ndistinct;
4130 :
4131 250175 : if (list_length(hashclauses) <= 1)
4132 : {
4133 : /*
4134 : * Nothing to do for a single clause. Could we employ univariate
4135 : * extended stat here?
4136 : */
4137 229955 : return hashclauses;
4138 : }
4139 :
4140 : /* "clauses" is the list of hashclauses we've not dealt with yet */
4141 20220 : clauses = list_copy(hashclauses);
4142 : /* "otherclauses" holds clauses we are going to return to caller */
4143 20220 : otherclauses = NIL;
4144 : /* current estimate of ndistinct */
4145 20220 : ndistinct = 1.0;
4146 40446 : while (clauses != NIL)
4147 : {
4148 : ListCell *lc;
4149 20226 : int relid = -1;
4150 20226 : List *varinfos = NIL;
4151 20226 : List *origin_rinfos = NIL;
4152 : double mvndistinct;
4153 : List *origin_varinfos;
4154 20226 : int group_relid = -1;
4155 20226 : RelOptInfo *group_rel = NULL;
4156 : ListCell *lc1,
4157 : *lc2;
4158 :
4159 : /*
4160 : * Find clauses, referencing the same single base relation and try to
4161 : * estimate such a group with extended statistics. Create varinfo for
4162 : * an approved clause, push it to otherclauses, if it can't be
4163 : * estimated here or ignore to process at the next iteration.
4164 : */
4165 60984 : foreach(lc, clauses)
4166 : {
4167 40758 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
4168 : Node *expr;
4169 : Relids relids;
4170 : GroupVarInfo *varinfo;
4171 :
4172 : /*
4173 : * Find the inner side of the join, which we need to estimate the
4174 : * number of buckets. Use outer_is_left because the
4175 : * clause_sides_match_join routine has called on hash clauses.
4176 : */
4177 81516 : relids = rinfo->outer_is_left ?
4178 40758 : rinfo->right_relids : rinfo->left_relids;
4179 81516 : expr = rinfo->outer_is_left ?
4180 40758 : get_rightop(rinfo->clause) : get_leftop(rinfo->clause);
4181 :
4182 40758 : if (bms_get_singleton_member(relids, &relid) &&
4183 39143 : root->simple_rel_array[relid]->statlist != NIL)
4184 24 : {
4185 30 : bool is_duplicate = false;
4186 :
4187 : /*
4188 : * This inner-side expression references only one relation.
4189 : * Extended statistics on this clause can exist.
4190 : */
4191 30 : if (group_relid < 0)
4192 : {
4193 15 : RangeTblEntry *rte = root->simple_rte_array[relid];
4194 :
4195 15 : if (!rte || (rte->relkind != RELKIND_RELATION &&
4196 0 : rte->relkind != RELKIND_MATVIEW &&
4197 0 : rte->relkind != RELKIND_FOREIGN_TABLE &&
4198 0 : rte->relkind != RELKIND_PARTITIONED_TABLE))
4199 : {
4200 : /* Extended statistics can't exist in principle */
4201 0 : otherclauses = lappend(otherclauses, rinfo);
4202 0 : clauses = foreach_delete_current(clauses, lc);
4203 0 : continue;
4204 : }
4205 :
4206 15 : group_relid = relid;
4207 15 : group_rel = root->simple_rel_array[relid];
4208 : }
4209 15 : else if (group_relid != relid)
4210 : {
4211 : /*
4212 : * Being in the group forming state we don't need other
4213 : * clauses.
4214 : */
4215 0 : continue;
4216 : }
4217 :
4218 : /*
4219 : * We're going to add the new clause to the varinfos list. We
4220 : * might re-use add_unique_group_var(), but we don't do so for
4221 : * two reasons.
4222 : *
4223 : * 1) We must keep the origin_rinfos list ordered exactly the
4224 : * same way as varinfos.
4225 : *
4226 : * 2) add_unique_group_var() is designed for
4227 : * estimate_num_groups(), where a larger number of groups is
4228 : * worse. While estimating the number of hash buckets, we
4229 : * have the opposite: a lesser number of groups is worse.
4230 : * Therefore, we don't have to remove "known equal" vars: the
4231 : * removed var may valuably contribute to the multivariate
4232 : * statistics to grow the number of groups.
4233 : */
4234 :
4235 : /*
4236 : * Clear nullingrels to correctly match hash keys. See
4237 : * add_unique_group_var()'s comment for details.
4238 : */
4239 30 : expr = remove_nulling_relids(expr, root->outer_join_rels, NULL);
4240 :
4241 : /*
4242 : * Detect and exclude exact duplicates from the list of hash
4243 : * keys (like add_unique_group_var does).
4244 : */
4245 42 : foreach(lc1, varinfos)
4246 : {
4247 18 : varinfo = (GroupVarInfo *) lfirst(lc1);
4248 :
4249 18 : if (!equal(expr, varinfo->var))
4250 12 : continue;
4251 :
4252 6 : is_duplicate = true;
4253 6 : break;
4254 : }
4255 :
4256 30 : if (is_duplicate)
4257 : {
4258 : /*
4259 : * Skip exact duplicates. Adding them to the otherclauses
4260 : * list also doesn't make sense.
4261 : */
4262 6 : continue;
4263 : }
4264 :
4265 : /*
4266 : * Initialize GroupVarInfo. We only use it to call
4267 : * estimate_multivariate_ndistinct(), which doesn't care about
4268 : * ndistinct and isdefault fields. Thus, skip these fields.
4269 : */
4270 24 : varinfo = palloc0_object(GroupVarInfo);
4271 24 : varinfo->var = expr;
4272 24 : varinfo->rel = root->simple_rel_array[relid];
4273 24 : varinfos = lappend(varinfos, varinfo);
4274 :
4275 : /*
4276 : * Remember the link to RestrictInfo for the case the clause
4277 : * is failed to be estimated.
4278 : */
4279 24 : origin_rinfos = lappend(origin_rinfos, rinfo);
4280 : }
4281 : else
4282 : {
4283 : /* This clause can't be estimated with extended statistics */
4284 40728 : otherclauses = lappend(otherclauses, rinfo);
4285 : }
4286 :
4287 40752 : clauses = foreach_delete_current(clauses, lc);
4288 : }
4289 :
4290 20226 : if (list_length(varinfos) < 2)
4291 : {
4292 : /*
4293 : * Multivariate statistics doesn't apply to single columns except
4294 : * for expressions, but it has not been implemented yet.
4295 : */
4296 20220 : otherclauses = list_concat(otherclauses, origin_rinfos);
4297 20220 : list_free_deep(varinfos);
4298 20220 : list_free(origin_rinfos);
4299 20220 : continue;
4300 : }
4301 :
4302 : Assert(group_rel != NULL);
4303 :
4304 : /* Employ the extended statistics. */
4305 6 : origin_varinfos = varinfos;
4306 : for (;;)
4307 6 : {
4308 12 : bool estimated = estimate_multivariate_ndistinct(root,
4309 : group_rel,
4310 : &varinfos,
4311 : &mvndistinct);
4312 :
4313 12 : if (!estimated)
4314 6 : break;
4315 :
4316 : /*
4317 : * We've got an estimation. Use ndistinct value in a consistent
4318 : * way - according to the caller's logic (see
4319 : * final_cost_hashjoin).
4320 : */
4321 6 : if (ndistinct < mvndistinct)
4322 6 : ndistinct = mvndistinct;
4323 : Assert(ndistinct >= 1.0);
4324 : }
4325 :
4326 : Assert(list_length(origin_varinfos) == list_length(origin_rinfos));
4327 :
4328 : /* Collect unmatched clauses as otherclauses. */
4329 21 : forboth(lc1, origin_varinfos, lc2, origin_rinfos)
4330 : {
4331 15 : GroupVarInfo *vinfo = lfirst(lc1);
4332 :
4333 15 : if (!list_member_ptr(varinfos, vinfo))
4334 : /* Already estimated */
4335 15 : continue;
4336 :
4337 : /* Can't be estimated here - push to the returning list */
4338 0 : otherclauses = lappend(otherclauses, lfirst(lc2));
4339 : }
4340 : }
4341 :
4342 20220 : *innerbucketsize = 1.0 / ndistinct;
4343 20220 : return otherclauses;
4344 : }
4345 :
4346 : /*
4347 : * Estimate hash bucket statistics when the specified expression is used
4348 : * as a hash key for the given number of buckets.
4349 : *
4350 : * This attempts to determine two values:
4351 : *
4352 : * 1. The frequency of the most common value of the expression (returns
4353 : * zero into *mcv_freq if we can't get that). This will be frequency
4354 : * relative to the entire underlying table.
4355 : *
4356 : * 2. The "bucketsize fraction", ie, average number of entries in a bucket
4357 : * divided by total number of tuples to be hashed.
4358 : *
4359 : * XXX This is really pretty bogus since we're effectively assuming that the
4360 : * distribution of hash keys will be the same after applying restriction
4361 : * clauses as it was in the underlying relation. However, we are not nearly
4362 : * smart enough to figure out how the restrict clauses might change the
4363 : * distribution, so this will have to do for now.
4364 : *
4365 : * We are passed the number of buckets the executor will use for the given
4366 : * input relation. If the data were perfectly distributed, with the same
4367 : * number of tuples going into each available bucket, then the bucketsize
4368 : * fraction would be 1/nbuckets. But this happy state of affairs will occur
4369 : * only if (a) there are at least nbuckets distinct data values, and (b)
4370 : * we have a not-too-skewed data distribution. Otherwise the buckets will
4371 : * be nonuniformly occupied. If the other relation in the join has a key
4372 : * distribution similar to this one's, then the most-loaded buckets are
4373 : * exactly those that will be probed most often. Therefore, the "average"
4374 : * bucket size for costing purposes should really be taken as something close
4375 : * to the "worst case" bucket size. We try to estimate this by adjusting the
4376 : * fraction if there are too few distinct data values, and then clamping to
4377 : * at least the bucket size implied by the most common value's frequency.
4378 : *
4379 : * If no statistics are available, use a default estimate of 0.1. This will
4380 : * discourage use of a hash rather strongly if the inner relation is large,
4381 : * which is what we want. We do not want to hash unless we know that the
4382 : * inner rel is well-dispersed (or the alternatives seem much worse).
4383 : *
4384 : * The caller should also check that the mcv_freq is not so large that the
4385 : * most common value would by itself require an impractically large bucket.
4386 : * In a hash join, the executor can split buckets if they get too big, but
4387 : * obviously that doesn't help for a bucket that contains many duplicates of
4388 : * the same value.
4389 : */
4390 : void
4391 114503 : estimate_hash_bucket_stats(PlannerInfo *root, Node *hashkey, double nbuckets,
4392 : Selectivity *mcv_freq,
4393 : Selectivity *bucketsize_frac)
4394 : {
4395 : VariableStatData vardata;
4396 : double estfract,
4397 : ndistinct;
4398 : bool isdefault;
4399 : AttStatsSlot sslot;
4400 :
4401 114503 : examine_variable(root, hashkey, 0, &vardata);
4402 :
4403 : /* Initialize *mcv_freq to "unknown" */
4404 114503 : *mcv_freq = 0.0;
4405 :
4406 : /* Look up the frequency of the most common value, if available */
4407 114503 : if (HeapTupleIsValid(vardata.statsTuple))
4408 : {
4409 77743 : if (get_attstatsslot(&sslot, vardata.statsTuple,
4410 : STATISTIC_KIND_MCV, InvalidOid,
4411 : ATTSTATSSLOT_NUMBERS))
4412 : {
4413 : /*
4414 : * The first MCV stat is for the most common value.
4415 : */
4416 44355 : if (sslot.nnumbers > 0)
4417 44355 : *mcv_freq = sslot.numbers[0];
4418 44355 : free_attstatsslot(&sslot);
4419 : }
4420 33388 : else if (get_attstatsslot(&sslot, vardata.statsTuple,
4421 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
4422 : 0))
4423 : {
4424 : /*
4425 : * If there are no recorded MCVs, but we do have a histogram, then
4426 : * assume that ANALYZE determined that the column is unique.
4427 : */
4428 32056 : if (vardata.rel && vardata.rel->tuples > 0)
4429 32047 : *mcv_freq = 1.0 / vardata.rel->tuples;
4430 : }
4431 : }
4432 :
4433 : /* Get number of distinct values */
4434 114503 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
4435 :
4436 : /*
4437 : * If ndistinct isn't real, punt. We normally return 0.1, but if the
4438 : * mcv_freq is known to be even higher than that, use it instead.
4439 : */
4440 114503 : if (isdefault)
4441 : {
4442 16913 : *bucketsize_frac = (Selectivity) Max(0.1, *mcv_freq);
4443 16913 : ReleaseVariableStats(vardata);
4444 16913 : return;
4445 : }
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 97590 : if (vardata.rel && vardata.rel->tuples > 0)
4456 : {
4457 97567 : ndistinct *= vardata.rel->rows / vardata.rel->tuples;
4458 97567 : 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 97590 : if (ndistinct > nbuckets)
4467 38 : estfract = 1.0 / nbuckets;
4468 : else
4469 97552 : estfract = 1.0 / ndistinct;
4470 :
4471 : /*
4472 : * Clamp the bucketsize fraction to be not less than the MCV frequency,
4473 : * since whichever bucket the MCV values end up in will have at least that
4474 : * size. This has no effect if *mcv_freq is still zero.
4475 : */
4476 97590 : estfract = Max(estfract, *mcv_freq);
4477 :
4478 97590 : *bucketsize_frac = (Selectivity) estfract;
4479 :
4480 97590 : ReleaseVariableStats(vardata);
4481 : }
4482 :
4483 : /*
4484 : * estimate_hashagg_tablesize
4485 : * estimate the number of bytes that a hash aggregate hashtable will
4486 : * require based on the agg_costs, path width and number of groups.
4487 : *
4488 : * We return the result as "double" to forestall any possible overflow
4489 : * problem in the multiplication by dNumGroups.
4490 : *
4491 : * XXX this may be over-estimating the size now that hashagg knows to omit
4492 : * unneeded columns from the hashtable. Also for mixed-mode grouping sets,
4493 : * grouping columns not in the hashed set are counted here even though hashagg
4494 : * won't store them. Is this a problem?
4495 : */
4496 : double
4497 1311 : estimate_hashagg_tablesize(PlannerInfo *root, Path *path,
4498 : const AggClauseCosts *agg_costs, double dNumGroups)
4499 : {
4500 : Size hashentrysize;
4501 :
4502 1311 : hashentrysize = hash_agg_entry_size(list_length(root->aggtransinfos),
4503 1311 : path->pathtarget->width,
4504 1311 : agg_costs->transitionSpace);
4505 :
4506 : /*
4507 : * Note that this disregards the effect of fill-factor and growth policy
4508 : * of the hash table. That's probably ok, given that the default
4509 : * fill-factor is relatively high. It'd be hard to meaningfully factor in
4510 : * "double-in-size" growth policies here.
4511 : */
4512 1311 : return hashentrysize * dNumGroups;
4513 : }
4514 :
4515 :
4516 : /*-------------------------------------------------------------------------
4517 : *
4518 : * Support routines
4519 : *
4520 : *-------------------------------------------------------------------------
4521 : */
4522 :
4523 : /*
4524 : * Find the best matching ndistinct extended statistics for the given list of
4525 : * GroupVarInfos.
4526 : *
4527 : * Callers must ensure that the given GroupVarInfos all belong to 'rel' and
4528 : * the GroupVarInfos list does not contain any duplicate Vars or expressions.
4529 : *
4530 : * When statistics are found that match > 1 of the given GroupVarInfo, the
4531 : * *ndistinct parameter is set according to the ndistinct estimate and a new
4532 : * list is built with the matching GroupVarInfos removed, which is output via
4533 : * the *varinfos parameter before returning true. When no matching stats are
4534 : * found, false is returned and the *varinfos and *ndistinct parameters are
4535 : * left untouched.
4536 : */
4537 : static bool
4538 202021 : estimate_multivariate_ndistinct(PlannerInfo *root, RelOptInfo *rel,
4539 : List **varinfos, double *ndistinct)
4540 : {
4541 : ListCell *lc;
4542 : int nmatches_vars;
4543 : int nmatches_exprs;
4544 202021 : Oid statOid = InvalidOid;
4545 : MVNDistinct *stats;
4546 202021 : StatisticExtInfo *matched_info = NULL;
4547 202021 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
4548 :
4549 : /* bail out immediately if the table has no extended statistics */
4550 202021 : if (!rel->statlist)
4551 201738 : return false;
4552 :
4553 : /* look for the ndistinct statistics object matching the most vars */
4554 283 : nmatches_vars = 0; /* we require at least two matches */
4555 283 : nmatches_exprs = 0;
4556 1127 : foreach(lc, rel->statlist)
4557 : {
4558 : ListCell *lc2;
4559 844 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(lc);
4560 844 : int nshared_vars = 0;
4561 844 : int nshared_exprs = 0;
4562 :
4563 : /* skip statistics of other kinds */
4564 844 : if (info->kind != STATS_EXT_NDISTINCT)
4565 399 : continue;
4566 :
4567 : /* skip statistics with mismatching stxdinherit value */
4568 445 : if (info->inherit != rte->inh)
4569 15 : continue;
4570 :
4571 : /*
4572 : * Determine how many expressions (and variables in non-matched
4573 : * expressions) match. We'll then use these numbers to pick the
4574 : * statistics object that best matches the clauses.
4575 : */
4576 1361 : foreach(lc2, *varinfos)
4577 : {
4578 : ListCell *lc3;
4579 931 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4580 : AttrNumber attnum;
4581 :
4582 : Assert(varinfo->rel == rel);
4583 :
4584 : /* simple Var, search in statistics keys directly */
4585 931 : if (IsA(varinfo->var, Var))
4586 : {
4587 748 : attnum = ((Var *) varinfo->var)->varattno;
4588 :
4589 : /*
4590 : * Ignore system attributes - we don't support statistics on
4591 : * them, so can't match them (and it'd fail as the values are
4592 : * negative).
4593 : */
4594 748 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4595 6 : continue;
4596 :
4597 742 : if (bms_is_member(attnum, info->keys))
4598 438 : nshared_vars++;
4599 :
4600 742 : continue;
4601 : }
4602 :
4603 : /* expression - see if it's in the statistics object */
4604 330 : foreach(lc3, info->exprs)
4605 : {
4606 264 : Node *expr = (Node *) lfirst(lc3);
4607 :
4608 264 : if (equal(varinfo->var, expr))
4609 : {
4610 117 : nshared_exprs++;
4611 117 : break;
4612 : }
4613 : }
4614 : }
4615 :
4616 : /*
4617 : * The ndistinct extended statistics contain estimates for a minimum
4618 : * of pairs of columns which the statistics are defined on and
4619 : * certainly not single columns. Here we skip unless we managed to
4620 : * match to at least two columns.
4621 : */
4622 430 : if (nshared_vars + nshared_exprs < 2)
4623 199 : continue;
4624 :
4625 : /*
4626 : * Check if these statistics are a better match than the previous best
4627 : * match and if so, take note of the StatisticExtInfo.
4628 : *
4629 : * The statslist is sorted by statOid, so the StatisticExtInfo we
4630 : * select as the best match is deterministic even when multiple sets
4631 : * of statistics match equally as well.
4632 : */
4633 231 : if ((nshared_exprs > nmatches_exprs) ||
4634 177 : (((nshared_exprs == nmatches_exprs)) && (nshared_vars > nmatches_vars)))
4635 : {
4636 219 : statOid = info->statOid;
4637 219 : nmatches_vars = nshared_vars;
4638 219 : nmatches_exprs = nshared_exprs;
4639 219 : matched_info = info;
4640 : }
4641 : }
4642 :
4643 : /* No match? */
4644 283 : if (statOid == InvalidOid)
4645 70 : return false;
4646 :
4647 : Assert(nmatches_vars + nmatches_exprs > 1);
4648 :
4649 213 : stats = statext_ndistinct_load(statOid, rte->inh);
4650 :
4651 : /*
4652 : * If we have a match, search it for the specific item that matches (there
4653 : * must be one), and construct the output values.
4654 : */
4655 213 : if (stats)
4656 : {
4657 : int i;
4658 213 : List *newlist = NIL;
4659 213 : MVNDistinctItem *item = NULL;
4660 : ListCell *lc2;
4661 213 : Bitmapset *matched = NULL;
4662 : AttrNumber attnum_offset;
4663 :
4664 : /*
4665 : * How much we need to offset the attnums? If there are no
4666 : * expressions, no offset is needed. Otherwise offset enough to move
4667 : * the lowest one (which is equal to number of expressions) to 1.
4668 : */
4669 213 : if (matched_info->exprs)
4670 75 : attnum_offset = (list_length(matched_info->exprs) + 1);
4671 : else
4672 138 : attnum_offset = 0;
4673 :
4674 : /* see what actually matched */
4675 744 : foreach(lc2, *varinfos)
4676 : {
4677 : ListCell *lc3;
4678 : int idx;
4679 531 : bool found = false;
4680 :
4681 531 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc2);
4682 :
4683 : /*
4684 : * Process a simple Var expression, by matching it to keys
4685 : * directly. If there's a matching expression, we'll try matching
4686 : * it later.
4687 : */
4688 531 : if (IsA(varinfo->var, Var))
4689 : {
4690 438 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4691 :
4692 : /*
4693 : * Ignore expressions on system attributes. Can't rely on the
4694 : * bms check for negative values.
4695 : */
4696 438 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4697 3 : continue;
4698 :
4699 : /* Is the variable covered by the statistics object? */
4700 435 : if (!bms_is_member(attnum, matched_info->keys))
4701 60 : continue;
4702 :
4703 375 : attnum = attnum + attnum_offset;
4704 :
4705 : /* ensure sufficient offset */
4706 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4707 :
4708 375 : matched = bms_add_member(matched, attnum);
4709 :
4710 375 : found = true;
4711 : }
4712 :
4713 : /*
4714 : * XXX Maybe we should allow searching the expressions even if we
4715 : * found an attribute matching the expression? That would handle
4716 : * trivial expressions like "(a)" but it seems fairly useless.
4717 : */
4718 468 : if (found)
4719 375 : continue;
4720 :
4721 : /* expression - see if it's in the statistics object */
4722 93 : idx = 0;
4723 153 : foreach(lc3, matched_info->exprs)
4724 : {
4725 138 : Node *expr = (Node *) lfirst(lc3);
4726 :
4727 138 : if (equal(varinfo->var, expr))
4728 : {
4729 78 : AttrNumber attnum = -(idx + 1);
4730 :
4731 78 : attnum = attnum + attnum_offset;
4732 :
4733 : /* ensure sufficient offset */
4734 : Assert(AttrNumberIsForUserDefinedAttr(attnum));
4735 :
4736 78 : matched = bms_add_member(matched, attnum);
4737 :
4738 : /* there should be just one matching expression */
4739 78 : break;
4740 : }
4741 :
4742 60 : idx++;
4743 : }
4744 : }
4745 :
4746 : /* Find the specific item that exactly matches the combination */
4747 432 : for (i = 0; i < stats->nitems; i++)
4748 : {
4749 : int j;
4750 432 : MVNDistinctItem *tmpitem = &stats->items[i];
4751 :
4752 432 : if (tmpitem->nattributes != bms_num_members(matched))
4753 81 : continue;
4754 :
4755 : /* assume it's the right item */
4756 351 : item = tmpitem;
4757 :
4758 : /* check that all item attributes/expressions fit the match */
4759 846 : for (j = 0; j < tmpitem->nattributes; j++)
4760 : {
4761 633 : AttrNumber attnum = tmpitem->attributes[j];
4762 :
4763 : /*
4764 : * Thanks to how we constructed the matched bitmap above, we
4765 : * can just offset all attnums the same way.
4766 : */
4767 633 : attnum = attnum + attnum_offset;
4768 :
4769 633 : if (!bms_is_member(attnum, matched))
4770 : {
4771 : /* nah, it's not this item */
4772 138 : item = NULL;
4773 138 : break;
4774 : }
4775 : }
4776 :
4777 : /*
4778 : * If the item has all the matched attributes, we know it's the
4779 : * right one - there can't be a better one. matching more.
4780 : */
4781 351 : if (item)
4782 213 : break;
4783 : }
4784 :
4785 : /*
4786 : * Make sure we found an item. There has to be one, because ndistinct
4787 : * statistics includes all combinations of attributes.
4788 : */
4789 213 : if (!item)
4790 0 : elog(ERROR, "corrupt MVNDistinct entry");
4791 :
4792 : /* Form the output varinfo list, keeping only unmatched ones */
4793 744 : foreach(lc, *varinfos)
4794 : {
4795 531 : GroupVarInfo *varinfo = (GroupVarInfo *) lfirst(lc);
4796 : ListCell *lc3;
4797 531 : bool found = false;
4798 :
4799 : /*
4800 : * Let's look at plain variables first, because it's the most
4801 : * common case and the check is quite cheap. We can simply get the
4802 : * attnum and check (with an offset) matched bitmap.
4803 : */
4804 531 : if (IsA(varinfo->var, Var))
4805 435 : {
4806 438 : AttrNumber attnum = ((Var *) varinfo->var)->varattno;
4807 :
4808 : /*
4809 : * If it's a system attribute, we're done. We don't support
4810 : * extended statistics on system attributes, so it's clearly
4811 : * not matched. Just keep the expression and continue.
4812 : */
4813 438 : if (!AttrNumberIsForUserDefinedAttr(attnum))
4814 : {
4815 3 : newlist = lappend(newlist, varinfo);
4816 3 : continue;
4817 : }
4818 :
4819 : /* apply the same offset as above */
4820 435 : attnum += attnum_offset;
4821 :
4822 : /* if it's not matched, keep the varinfo */
4823 435 : if (!bms_is_member(attnum, matched))
4824 60 : newlist = lappend(newlist, varinfo);
4825 :
4826 : /* The rest of the loop deals with complex expressions. */
4827 435 : continue;
4828 : }
4829 :
4830 : /*
4831 : * Process complex expressions, not just simple Vars.
4832 : *
4833 : * First, we search for an exact match of an expression. If we
4834 : * find one, we can just discard the whole GroupVarInfo, with all
4835 : * the variables we extracted from it.
4836 : *
4837 : * Otherwise we inspect the individual vars, and try matching it
4838 : * to variables in the item.
4839 : */
4840 153 : foreach(lc3, matched_info->exprs)
4841 : {
4842 138 : Node *expr = (Node *) lfirst(lc3);
4843 :
4844 138 : if (equal(varinfo->var, expr))
4845 : {
4846 78 : found = true;
4847 78 : break;
4848 : }
4849 : }
4850 :
4851 : /* found exact match, skip */
4852 93 : if (found)
4853 78 : continue;
4854 :
4855 15 : newlist = lappend(newlist, varinfo);
4856 : }
4857 :
4858 213 : *varinfos = newlist;
4859 213 : *ndistinct = item->ndistinct;
4860 213 : return true;
4861 : }
4862 :
4863 0 : return false;
4864 : }
4865 :
4866 : /*
4867 : * convert_to_scalar
4868 : * Convert non-NULL values of the indicated types to the comparison
4869 : * scale needed by scalarineqsel().
4870 : * Returns "true" if successful.
4871 : *
4872 : * XXX this routine is a hack: ideally we should look up the conversion
4873 : * subroutines in pg_type.
4874 : *
4875 : * All numeric datatypes are simply converted to their equivalent
4876 : * "double" values. (NUMERIC values that are outside the range of "double"
4877 : * are clamped to +/- HUGE_VAL.)
4878 : *
4879 : * String datatypes are converted by convert_string_to_scalar(),
4880 : * which is explained below. The reason why this routine deals with
4881 : * three values at a time, not just one, is that we need it for strings.
4882 : *
4883 : * The bytea datatype is just enough different from strings that it has
4884 : * to be treated separately.
4885 : *
4886 : * The several datatypes representing absolute times are all converted
4887 : * to Timestamp, which is actually an int64, and then we promote that to
4888 : * a double. Note this will give correct results even for the "special"
4889 : * values of Timestamp, since those are chosen to compare correctly;
4890 : * see timestamp_cmp.
4891 : *
4892 : * The several datatypes representing relative times (intervals) are all
4893 : * converted to measurements expressed in seconds.
4894 : */
4895 : static bool
4896 48970 : convert_to_scalar(Datum value, Oid valuetypid, Oid collid, double *scaledvalue,
4897 : Datum lobound, Datum hibound, Oid boundstypid,
4898 : double *scaledlobound, double *scaledhibound)
4899 : {
4900 48970 : bool failure = false;
4901 :
4902 : /*
4903 : * Both the valuetypid and the boundstypid should exactly match the
4904 : * declared input type(s) of the operator we are invoked for. However,
4905 : * extensions might try to use scalarineqsel as estimator for operators
4906 : * with input type(s) we don't handle here; in such cases, we want to
4907 : * return false, not fail. In any case, we mustn't assume that valuetypid
4908 : * and boundstypid are identical.
4909 : *
4910 : * XXX The histogram we are interpolating between points of could belong
4911 : * to a column that's only binary-compatible with the declared type. In
4912 : * essence we are assuming that the semantics of binary-compatible types
4913 : * are enough alike that we can use a histogram generated with one type's
4914 : * operators to estimate selectivity for the other's. This is outright
4915 : * wrong in some cases --- in particular signed versus unsigned
4916 : * interpretation could trip us up. But it's useful enough in the
4917 : * majority of cases that we do it anyway. Should think about more
4918 : * rigorous ways to do it.
4919 : */
4920 48970 : switch (valuetypid)
4921 : {
4922 : /*
4923 : * Built-in numeric types
4924 : */
4925 45327 : case BOOLOID:
4926 : case INT2OID:
4927 : case INT4OID:
4928 : case INT8OID:
4929 : case FLOAT4OID:
4930 : case FLOAT8OID:
4931 : case NUMERICOID:
4932 : case OIDOID:
4933 : case REGPROCOID:
4934 : case REGPROCEDUREOID:
4935 : case REGOPEROID:
4936 : case REGOPERATOROID:
4937 : case REGCLASSOID:
4938 : case REGTYPEOID:
4939 : case REGCOLLATIONOID:
4940 : case REGCONFIGOID:
4941 : case REGDICTIONARYOID:
4942 : case REGROLEOID:
4943 : case REGNAMESPACEOID:
4944 : case REGDATABASEOID:
4945 45327 : *scaledvalue = convert_numeric_to_scalar(value, valuetypid,
4946 : &failure);
4947 45327 : *scaledlobound = convert_numeric_to_scalar(lobound, boundstypid,
4948 : &failure);
4949 45327 : *scaledhibound = convert_numeric_to_scalar(hibound, boundstypid,
4950 : &failure);
4951 45327 : return !failure;
4952 :
4953 : /*
4954 : * Built-in string types
4955 : */
4956 3643 : case CHAROID:
4957 : case BPCHAROID:
4958 : case VARCHAROID:
4959 : case TEXTOID:
4960 : case NAMEOID:
4961 : {
4962 3643 : char *valstr = convert_string_datum(value, valuetypid,
4963 : collid, &failure);
4964 3643 : char *lostr = convert_string_datum(lobound, boundstypid,
4965 : collid, &failure);
4966 3643 : char *histr = convert_string_datum(hibound, boundstypid,
4967 : collid, &failure);
4968 :
4969 : /*
4970 : * Bail out if any of the values is not of string type. We
4971 : * might leak converted strings for the other value(s), but
4972 : * that's not worth troubling over.
4973 : */
4974 3643 : if (failure)
4975 0 : return false;
4976 :
4977 3643 : convert_string_to_scalar(valstr, scaledvalue,
4978 : lostr, scaledlobound,
4979 : histr, scaledhibound);
4980 3643 : pfree(valstr);
4981 3643 : pfree(lostr);
4982 3643 : pfree(histr);
4983 3643 : return true;
4984 : }
4985 :
4986 : /*
4987 : * Built-in bytea type
4988 : */
4989 0 : case BYTEAOID:
4990 : {
4991 : /* We only support bytea vs bytea comparison */
4992 0 : if (boundstypid != BYTEAOID)
4993 0 : return false;
4994 0 : convert_bytea_to_scalar(value, scaledvalue,
4995 : lobound, scaledlobound,
4996 : hibound, scaledhibound);
4997 0 : return true;
4998 : }
4999 :
5000 : /*
5001 : * Built-in time types
5002 : */
5003 0 : case TIMESTAMPOID:
5004 : case TIMESTAMPTZOID:
5005 : case DATEOID:
5006 : case INTERVALOID:
5007 : case TIMEOID:
5008 : case TIMETZOID:
5009 0 : *scaledvalue = convert_timevalue_to_scalar(value, valuetypid,
5010 : &failure);
5011 0 : *scaledlobound = convert_timevalue_to_scalar(lobound, boundstypid,
5012 : &failure);
5013 0 : *scaledhibound = convert_timevalue_to_scalar(hibound, boundstypid,
5014 : &failure);
5015 0 : return !failure;
5016 :
5017 : /*
5018 : * Built-in network types
5019 : */
5020 0 : case INETOID:
5021 : case CIDROID:
5022 : case MACADDROID:
5023 : case MACADDR8OID:
5024 0 : *scaledvalue = convert_network_to_scalar(value, valuetypid,
5025 : &failure);
5026 0 : *scaledlobound = convert_network_to_scalar(lobound, boundstypid,
5027 : &failure);
5028 0 : *scaledhibound = convert_network_to_scalar(hibound, boundstypid,
5029 : &failure);
5030 0 : return !failure;
5031 : }
5032 : /* Don't know how to convert */
5033 0 : *scaledvalue = *scaledlobound = *scaledhibound = 0;
5034 0 : return false;
5035 : }
5036 :
5037 : /*
5038 : * Do convert_to_scalar()'s work for any numeric data type.
5039 : *
5040 : * On failure (e.g., unsupported typid), set *failure to true;
5041 : * otherwise, that variable is not changed.
5042 : */
5043 : static double
5044 135981 : convert_numeric_to_scalar(Datum value, Oid typid, bool *failure)
5045 : {
5046 135981 : switch (typid)
5047 : {
5048 0 : case BOOLOID:
5049 0 : return (double) DatumGetBool(value);
5050 6 : case INT2OID:
5051 6 : return (double) DatumGetInt16(value);
5052 15957 : case INT4OID:
5053 15957 : return (double) DatumGetInt32(value);
5054 0 : case INT8OID:
5055 0 : return (double) DatumGetInt64(value);
5056 0 : case FLOAT4OID:
5057 0 : return (double) DatumGetFloat4(value);
5058 27 : case FLOAT8OID:
5059 27 : return (double) DatumGetFloat8(value);
5060 0 : case NUMERICOID:
5061 : /* Note: out-of-range values will be clamped to +-HUGE_VAL */
5062 0 : return (double)
5063 0 : DatumGetFloat8(DirectFunctionCall1(numeric_float8_no_overflow,
5064 : value));
5065 119991 : case OIDOID:
5066 : case REGPROCOID:
5067 : case REGPROCEDUREOID:
5068 : case REGOPEROID:
5069 : case REGOPERATOROID:
5070 : case REGCLASSOID:
5071 : case REGTYPEOID:
5072 : case REGCOLLATIONOID:
5073 : case REGCONFIGOID:
5074 : case REGDICTIONARYOID:
5075 : case REGROLEOID:
5076 : case REGNAMESPACEOID:
5077 : case REGDATABASEOID:
5078 : /* we can treat OIDs as integers... */
5079 119991 : return (double) DatumGetObjectId(value);
5080 : }
5081 :
5082 0 : *failure = true;
5083 0 : return 0;
5084 : }
5085 :
5086 : /*
5087 : * Do convert_to_scalar()'s work for any character-string data type.
5088 : *
5089 : * String datatypes are converted to a scale that ranges from 0 to 1,
5090 : * where we visualize the bytes of the string as fractional digits.
5091 : *
5092 : * We do not want the base to be 256, however, since that tends to
5093 : * generate inflated selectivity estimates; few databases will have
5094 : * occurrences of all 256 possible byte values at each position.
5095 : * Instead, use the smallest and largest byte values seen in the bounds
5096 : * as the estimated range for each byte, after some fudging to deal with
5097 : * the fact that we probably aren't going to see the full range that way.
5098 : *
5099 : * An additional refinement is that we discard any common prefix of the
5100 : * three strings before computing the scaled values. This allows us to
5101 : * "zoom in" when we encounter a narrow data range. An example is a phone
5102 : * number database where all the values begin with the same area code.
5103 : * (Actually, the bounds will be adjacent histogram-bin-boundary values,
5104 : * so this is more likely to happen than you might think.)
5105 : */
5106 : static void
5107 3643 : convert_string_to_scalar(char *value,
5108 : double *scaledvalue,
5109 : char *lobound,
5110 : double *scaledlobound,
5111 : char *hibound,
5112 : double *scaledhibound)
5113 : {
5114 : int rangelo,
5115 : rangehi;
5116 : char *sptr;
5117 :
5118 3643 : rangelo = rangehi = (unsigned char) hibound[0];
5119 45244 : for (sptr = lobound; *sptr; sptr++)
5120 : {
5121 41601 : if (rangelo > (unsigned char) *sptr)
5122 8626 : rangelo = (unsigned char) *sptr;
5123 41601 : if (rangehi < (unsigned char) *sptr)
5124 4486 : rangehi = (unsigned char) *sptr;
5125 : }
5126 44036 : for (sptr = hibound; *sptr; sptr++)
5127 : {
5128 40393 : if (rangelo > (unsigned char) *sptr)
5129 617 : rangelo = (unsigned char) *sptr;
5130 40393 : if (rangehi < (unsigned char) *sptr)
5131 1680 : rangehi = (unsigned char) *sptr;
5132 : }
5133 : /* If range includes any upper-case ASCII chars, make it include all */
5134 3643 : if (rangelo <= 'Z' && rangehi >= 'A')
5135 : {
5136 649 : if (rangelo > 'A')
5137 111 : rangelo = 'A';
5138 649 : if (rangehi < 'Z')
5139 240 : rangehi = 'Z';
5140 : }
5141 : /* Ditto lower-case */
5142 3643 : if (rangelo <= 'z' && rangehi >= 'a')
5143 : {
5144 3392 : if (rangelo > 'a')
5145 5 : rangelo = 'a';
5146 3392 : if (rangehi < 'z')
5147 3358 : rangehi = 'z';
5148 : }
5149 : /* Ditto digits */
5150 3643 : if (rangelo <= '9' && rangehi >= '0')
5151 : {
5152 309 : if (rangelo > '0')
5153 262 : rangelo = '0';
5154 309 : if (rangehi < '9')
5155 7 : rangehi = '9';
5156 : }
5157 :
5158 : /*
5159 : * If range includes less than 10 chars, assume we have not got enough
5160 : * data, and make it include regular ASCII set.
5161 : */
5162 3643 : if (rangehi - rangelo < 9)
5163 : {
5164 0 : rangelo = ' ';
5165 0 : rangehi = 127;
5166 : }
5167 :
5168 : /*
5169 : * Now strip any common prefix of the three strings.
5170 : */
5171 7556 : while (*lobound)
5172 : {
5173 7556 : if (*lobound != *hibound || *lobound != *value)
5174 : break;
5175 3913 : lobound++, hibound++, value++;
5176 : }
5177 :
5178 : /*
5179 : * Now we can do the conversions.
5180 : */
5181 3643 : *scaledvalue = convert_one_string_to_scalar(value, rangelo, rangehi);
5182 3643 : *scaledlobound = convert_one_string_to_scalar(lobound, rangelo, rangehi);
5183 3643 : *scaledhibound = convert_one_string_to_scalar(hibound, rangelo, rangehi);
5184 3643 : }
5185 :
5186 : static double
5187 10929 : convert_one_string_to_scalar(char *value, int rangelo, int rangehi)
5188 : {
5189 10929 : int slen = strlen(value);
5190 : double num,
5191 : denom,
5192 : base;
5193 :
5194 10929 : if (slen <= 0)
5195 0 : return 0.0; /* empty string has scalar value 0 */
5196 :
5197 : /*
5198 : * There seems little point in considering more than a dozen bytes from
5199 : * the string. Since base is at least 10, that will give us nominal
5200 : * resolution of at least 12 decimal digits, which is surely far more
5201 : * precision than this estimation technique has got anyway (especially in
5202 : * non-C locales). Also, even with the maximum possible base of 256, this
5203 : * ensures denom cannot grow larger than 256^13 = 2.03e31, which will not
5204 : * overflow on any known machine.
5205 : */
5206 10929 : if (slen > 12)
5207 2815 : slen = 12;
5208 :
5209 : /* Convert initial characters to fraction */
5210 10929 : base = rangehi - rangelo + 1;
5211 10929 : num = 0.0;
5212 10929 : denom = base;
5213 91320 : while (slen-- > 0)
5214 : {
5215 80391 : int ch = (unsigned char) *value++;
5216 :
5217 80391 : if (ch < rangelo)
5218 104 : ch = rangelo - 1;
5219 80287 : else if (ch > rangehi)
5220 0 : ch = rangehi + 1;
5221 80391 : num += ((double) (ch - rangelo)) / denom;
5222 80391 : denom *= base;
5223 : }
5224 :
5225 10929 : return num;
5226 : }
5227 :
5228 : /*
5229 : * Convert a string-type Datum into a palloc'd, null-terminated string.
5230 : *
5231 : * On failure (e.g., unsupported typid), set *failure to true;
5232 : * otherwise, that variable is not changed. (We'll return NULL on failure.)
5233 : *
5234 : * When using a non-C locale, we must pass the string through pg_strxfrm()
5235 : * before continuing, so as to generate correct locale-specific results.
5236 : */
5237 : static char *
5238 10929 : convert_string_datum(Datum value, Oid typid, Oid collid, bool *failure)
5239 : {
5240 : char *val;
5241 : pg_locale_t mylocale;
5242 :
5243 10929 : switch (typid)
5244 : {
5245 0 : case CHAROID:
5246 0 : val = (char *) palloc(2);
5247 0 : val[0] = DatumGetChar(value);
5248 0 : val[1] = '\0';
5249 0 : break;
5250 3375 : case BPCHAROID:
5251 : case VARCHAROID:
5252 : case TEXTOID:
5253 3375 : val = TextDatumGetCString(value);
5254 3375 : break;
5255 7554 : case NAMEOID:
5256 : {
5257 7554 : NameData *nm = (NameData *) DatumGetPointer(value);
5258 :
5259 7554 : val = pstrdup(NameStr(*nm));
5260 7554 : break;
5261 : }
5262 0 : default:
5263 0 : *failure = true;
5264 0 : return NULL;
5265 : }
5266 :
5267 10929 : mylocale = pg_newlocale_from_collation(collid);
5268 :
5269 10929 : if (!mylocale->collate_is_c)
5270 : {
5271 : char *xfrmstr;
5272 : size_t xfrmlen;
5273 : size_t xfrmlen2 PG_USED_FOR_ASSERTS_ONLY;
5274 :
5275 : /*
5276 : * XXX: We could guess at a suitable output buffer size and only call
5277 : * pg_strxfrm() twice if our guess is too small.
5278 : *
5279 : * XXX: strxfrm doesn't support UTF-8 encoding on Win32, it can return
5280 : * bogus data or set an error. This is not really a problem unless it
5281 : * crashes since it will only give an estimation error and nothing
5282 : * fatal.
5283 : *
5284 : * XXX: we do not check pg_strxfrm_enabled(). On some platforms and in
5285 : * some cases, libc strxfrm() may return the wrong results, but that
5286 : * will only lead to an estimation error.
5287 : */
5288 36 : xfrmlen = pg_strxfrm(NULL, val, 0, mylocale);
5289 : #ifdef WIN32
5290 :
5291 : /*
5292 : * On Windows, strxfrm returns INT_MAX when an error occurs. Instead
5293 : * of trying to allocate this much memory (and fail), just return the
5294 : * original string unmodified as if we were in the C locale.
5295 : */
5296 : if (xfrmlen == INT_MAX)
5297 : return val;
5298 : #endif
5299 36 : xfrmstr = (char *) palloc(xfrmlen + 1);
5300 36 : xfrmlen2 = pg_strxfrm(xfrmstr, val, xfrmlen + 1, mylocale);
5301 :
5302 : /*
5303 : * Some systems (e.g., glibc) can return a smaller value from the
5304 : * second call than the first; thus the Assert must be <= not ==.
5305 : */
5306 : Assert(xfrmlen2 <= xfrmlen);
5307 36 : pfree(val);
5308 36 : val = xfrmstr;
5309 : }
5310 :
5311 10929 : return val;
5312 : }
5313 :
5314 : /*
5315 : * Do convert_to_scalar()'s work for any bytea data type.
5316 : *
5317 : * Very similar to convert_string_to_scalar except we can't assume
5318 : * null-termination and therefore pass explicit lengths around.
5319 : *
5320 : * Also, assumptions about likely "normal" ranges of characters have been
5321 : * removed - a data range of 0..255 is always used, for now. (Perhaps
5322 : * someday we will add information about actual byte data range to
5323 : * pg_statistic.)
5324 : */
5325 : static void
5326 0 : convert_bytea_to_scalar(Datum value,
5327 : double *scaledvalue,
5328 : Datum lobound,
5329 : double *scaledlobound,
5330 : Datum hibound,
5331 : double *scaledhibound)
5332 : {
5333 0 : bytea *valuep = DatumGetByteaPP(value);
5334 0 : bytea *loboundp = DatumGetByteaPP(lobound);
5335 0 : bytea *hiboundp = DatumGetByteaPP(hibound);
5336 : int rangelo,
5337 : rangehi,
5338 0 : valuelen = VARSIZE_ANY_EXHDR(valuep),
5339 0 : loboundlen = VARSIZE_ANY_EXHDR(loboundp),
5340 0 : hiboundlen = VARSIZE_ANY_EXHDR(hiboundp),
5341 : i,
5342 : minlen;
5343 0 : unsigned char *valstr = (unsigned char *) VARDATA_ANY(valuep);
5344 0 : unsigned char *lostr = (unsigned char *) VARDATA_ANY(loboundp);
5345 0 : unsigned char *histr = (unsigned char *) VARDATA_ANY(hiboundp);
5346 :
5347 : /*
5348 : * Assume bytea data is uniformly distributed across all byte values.
5349 : */
5350 0 : rangelo = 0;
5351 0 : rangehi = 255;
5352 :
5353 : /*
5354 : * Now strip any common prefix of the three strings.
5355 : */
5356 0 : minlen = Min(Min(valuelen, loboundlen), hiboundlen);
5357 0 : for (i = 0; i < minlen; i++)
5358 : {
5359 0 : if (*lostr != *histr || *lostr != *valstr)
5360 : break;
5361 0 : lostr++, histr++, valstr++;
5362 0 : loboundlen--, hiboundlen--, valuelen--;
5363 : }
5364 :
5365 : /*
5366 : * Now we can do the conversions.
5367 : */
5368 0 : *scaledvalue = convert_one_bytea_to_scalar(valstr, valuelen, rangelo, rangehi);
5369 0 : *scaledlobound = convert_one_bytea_to_scalar(lostr, loboundlen, rangelo, rangehi);
5370 0 : *scaledhibound = convert_one_bytea_to_scalar(histr, hiboundlen, rangelo, rangehi);
5371 0 : }
5372 :
5373 : static double
5374 0 : convert_one_bytea_to_scalar(unsigned char *value, int valuelen,
5375 : int rangelo, int rangehi)
5376 : {
5377 : double num,
5378 : denom,
5379 : base;
5380 :
5381 0 : if (valuelen <= 0)
5382 0 : return 0.0; /* empty string has scalar value 0 */
5383 :
5384 : /*
5385 : * Since base is 256, need not consider more than about 10 chars (even
5386 : * this many seems like overkill)
5387 : */
5388 0 : if (valuelen > 10)
5389 0 : valuelen = 10;
5390 :
5391 : /* Convert initial characters to fraction */
5392 0 : base = rangehi - rangelo + 1;
5393 0 : num = 0.0;
5394 0 : denom = base;
5395 0 : while (valuelen-- > 0)
5396 : {
5397 0 : int ch = *value++;
5398 :
5399 0 : if (ch < rangelo)
5400 0 : ch = rangelo - 1;
5401 0 : else if (ch > rangehi)
5402 0 : ch = rangehi + 1;
5403 0 : num += ((double) (ch - rangelo)) / denom;
5404 0 : denom *= base;
5405 : }
5406 :
5407 0 : return num;
5408 : }
5409 :
5410 : /*
5411 : * Do convert_to_scalar()'s work for any timevalue data type.
5412 : *
5413 : * On failure (e.g., unsupported typid), set *failure to true;
5414 : * otherwise, that variable is not changed.
5415 : */
5416 : static double
5417 0 : convert_timevalue_to_scalar(Datum value, Oid typid, bool *failure)
5418 : {
5419 0 : switch (typid)
5420 : {
5421 0 : case TIMESTAMPOID:
5422 0 : return DatumGetTimestamp(value);
5423 0 : case TIMESTAMPTZOID:
5424 0 : return DatumGetTimestampTz(value);
5425 0 : case DATEOID:
5426 0 : return date2timestamp_no_overflow(DatumGetDateADT(value));
5427 0 : case INTERVALOID:
5428 : {
5429 0 : Interval *interval = DatumGetIntervalP(value);
5430 :
5431 : /*
5432 : * Convert the month part of Interval to days using assumed
5433 : * average month length of 365.25/12.0 days. Not too
5434 : * accurate, but plenty good enough for our purposes.
5435 : *
5436 : * This also works for infinite intervals, which just have all
5437 : * fields set to INT_MIN/INT_MAX, and so will produce a result
5438 : * smaller/larger than any finite interval.
5439 : */
5440 0 : return interval->time + interval->day * (double) USECS_PER_DAY +
5441 0 : interval->month * ((DAYS_PER_YEAR / (double) MONTHS_PER_YEAR) * USECS_PER_DAY);
5442 : }
5443 0 : case TIMEOID:
5444 0 : return DatumGetTimeADT(value);
5445 0 : case TIMETZOID:
5446 : {
5447 0 : TimeTzADT *timetz = DatumGetTimeTzADTP(value);
5448 :
5449 : /* use GMT-equivalent time */
5450 0 : return (double) (timetz->time + (timetz->zone * 1000000.0));
5451 : }
5452 : }
5453 :
5454 0 : *failure = true;
5455 0 : return 0;
5456 : }
5457 :
5458 :
5459 : /*
5460 : * get_restriction_variable
5461 : * Examine the args of a restriction clause to see if it's of the
5462 : * form (variable op pseudoconstant) or (pseudoconstant op variable),
5463 : * where "variable" could be either a Var or an expression in vars of a
5464 : * single relation. If so, extract information about the variable,
5465 : * and also indicate which side it was on and the other argument.
5466 : *
5467 : * Inputs:
5468 : * root: the planner info
5469 : * args: clause argument list
5470 : * varRelid: see specs for restriction selectivity functions
5471 : *
5472 : * Outputs: (these are valid only if true is returned)
5473 : * *vardata: gets information about variable (see examine_variable)
5474 : * *other: gets other clause argument, aggressively reduced to a constant
5475 : * *varonleft: set true if variable is on the left, false if on the right
5476 : *
5477 : * Returns true if a variable is identified, otherwise false.
5478 : *
5479 : * Note: if there are Vars on both sides of the clause, we must fail, because
5480 : * callers are expecting that the other side will act like a pseudoconstant.
5481 : */
5482 : bool
5483 472790 : get_restriction_variable(PlannerInfo *root, List *args, int varRelid,
5484 : VariableStatData *vardata, Node **other,
5485 : bool *varonleft)
5486 : {
5487 : Node *left,
5488 : *right;
5489 : VariableStatData rdata;
5490 :
5491 : /* Fail if not a binary opclause (probably shouldn't happen) */
5492 472790 : if (list_length(args) != 2)
5493 0 : return false;
5494 :
5495 472790 : left = (Node *) linitial(args);
5496 472790 : right = (Node *) lsecond(args);
5497 :
5498 : /*
5499 : * Examine both sides. Note that when varRelid is nonzero, Vars of other
5500 : * relations will be treated as pseudoconstants.
5501 : */
5502 472790 : examine_variable(root, left, varRelid, vardata);
5503 472790 : examine_variable(root, right, varRelid, &rdata);
5504 :
5505 : /*
5506 : * If one side is a variable and the other not, we win.
5507 : */
5508 472790 : if (vardata->rel && rdata.rel == NULL)
5509 : {
5510 421001 : *varonleft = true;
5511 421001 : *other = estimate_expression_value(root, rdata.var);
5512 : /* Assume we need no ReleaseVariableStats(rdata) here */
5513 420998 : return true;
5514 : }
5515 :
5516 51789 : if (vardata->rel == NULL && rdata.rel)
5517 : {
5518 48403 : *varonleft = false;
5519 48403 : *other = estimate_expression_value(root, vardata->var);
5520 : /* Assume we need no ReleaseVariableStats(*vardata) here */
5521 48403 : *vardata = rdata;
5522 48403 : return true;
5523 : }
5524 :
5525 : /* Oops, clause has wrong structure (probably var op var) */
5526 3386 : ReleaseVariableStats(*vardata);
5527 3386 : ReleaseVariableStats(rdata);
5528 :
5529 3386 : return false;
5530 : }
5531 :
5532 : /*
5533 : * get_join_variables
5534 : * Apply examine_variable() to each side of a join clause.
5535 : * Also, attempt to identify whether the join clause has the same
5536 : * or reversed sense compared to the SpecialJoinInfo.
5537 : *
5538 : * We consider the join clause "normal" if it is "lhs_var OP rhs_var",
5539 : * or "reversed" if it is "rhs_var OP lhs_var". In complicated cases
5540 : * where we can't tell for sure, we default to assuming it's normal.
5541 : */
5542 : void
5543 155153 : get_join_variables(PlannerInfo *root, List *args, SpecialJoinInfo *sjinfo,
5544 : VariableStatData *vardata1, VariableStatData *vardata2,
5545 : bool *join_is_reversed)
5546 : {
5547 : Node *left,
5548 : *right;
5549 :
5550 155153 : if (list_length(args) != 2)
5551 0 : elog(ERROR, "join operator should take two arguments");
5552 :
5553 155153 : left = (Node *) linitial(args);
5554 155153 : right = (Node *) lsecond(args);
5555 :
5556 155153 : examine_variable(root, left, 0, vardata1);
5557 155153 : examine_variable(root, right, 0, vardata2);
5558 :
5559 310179 : if (vardata1->rel &&
5560 155026 : bms_is_subset(vardata1->rel->relids, sjinfo->syn_righthand))
5561 51751 : *join_is_reversed = true; /* var1 is on RHS */
5562 206679 : else if (vardata2->rel &&
5563 103277 : bms_is_subset(vardata2->rel->relids, sjinfo->syn_lefthand))
5564 169 : *join_is_reversed = true; /* var2 is on LHS */
5565 : else
5566 103233 : *join_is_reversed = false;
5567 155153 : }
5568 :
5569 : /* statext_expressions_load copies the tuple, so just pfree it. */
5570 : static void
5571 831 : ReleaseDummy(HeapTuple tuple)
5572 : {
5573 831 : pfree(tuple);
5574 831 : }
5575 :
5576 : /*
5577 : * examine_variable
5578 : * Try to look up statistical data about an expression.
5579 : * Fill in a VariableStatData struct to describe the expression.
5580 : *
5581 : * Inputs:
5582 : * root: the planner info
5583 : * node: the expression tree to examine
5584 : * varRelid: see specs for restriction selectivity functions
5585 : *
5586 : * Outputs: *vardata is filled as follows:
5587 : * var: the input expression (with any phvs or binary relabeling stripped,
5588 : * if it is or contains a variable; but otherwise unchanged)
5589 : * rel: RelOptInfo for relation containing variable; NULL if expression
5590 : * contains no Vars (NOTE this could point to a RelOptInfo of a
5591 : * subquery, not one in the current query).
5592 : * statsTuple: the pg_statistic entry for the variable, if one exists;
5593 : * otherwise NULL.
5594 : * freefunc: pointer to a function to release statsTuple with.
5595 : * vartype: exposed type of the expression; this should always match
5596 : * the declared input type of the operator we are estimating for.
5597 : * atttype, atttypmod: actual type/typmod of the "var" expression. This is
5598 : * commonly the same as the exposed type of the variable argument,
5599 : * but can be different in binary-compatible-type cases.
5600 : * isunique: true if we were able to match the var to a unique index, a
5601 : * single-column DISTINCT or GROUP-BY clause, implying its values are
5602 : * unique for this query. (Caution: this should be trusted for
5603 : * statistical purposes only, since we do not check indimmediate nor
5604 : * verify that the exact same definition of equality applies.)
5605 : * acl_ok: true if current user has permission to read all table rows from
5606 : * the column(s) underlying the pg_statistic entry. This is consulted by
5607 : * statistic_proc_security_check().
5608 : *
5609 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
5610 : */
5611 : void
5612 1895011 : examine_variable(PlannerInfo *root, Node *node, int varRelid,
5613 : VariableStatData *vardata)
5614 : {
5615 : Node *basenode;
5616 : Relids varnos;
5617 : Relids basevarnos;
5618 : RelOptInfo *onerel;
5619 :
5620 : /* Make sure we don't return dangling pointers in vardata */
5621 13265077 : MemSet(vardata, 0, sizeof(VariableStatData));
5622 :
5623 : /* Save the exposed type of the expression */
5624 1895011 : vardata->vartype = exprType(node);
5625 :
5626 : /*
5627 : * PlaceHolderVars are transparent for the purpose of statistics lookup;
5628 : * they do not alter the value distribution of the underlying expression.
5629 : * However, they can obscure the structure, preventing us from recognizing
5630 : * matches to base columns, index expressions, or extended statistics. So
5631 : * strip them out first.
5632 : */
5633 1895011 : basenode = strip_all_phvs_deep(root, node);
5634 :
5635 : /*
5636 : * Look inside any binary-compatible relabeling. We need to handle nested
5637 : * RelabelType nodes here, because the prior stripping of PlaceHolderVars
5638 : * may have brought separate RelabelTypes into adjacency.
5639 : */
5640 1919779 : while (IsA(basenode, RelabelType))
5641 24768 : basenode = (Node *) ((RelabelType *) basenode)->arg;
5642 :
5643 : /* Fast path for a simple Var */
5644 1895011 : if (IsA(basenode, Var) &&
5645 485905 : (varRelid == 0 || varRelid == ((Var *) basenode)->varno))
5646 : {
5647 1360299 : Var *var = (Var *) basenode;
5648 :
5649 : /* Set up result fields other than the stats tuple */
5650 1360299 : vardata->var = basenode; /* return Var without phvs or relabeling */
5651 1360299 : vardata->rel = find_base_rel(root, var->varno);
5652 1360299 : vardata->atttype = var->vartype;
5653 1360299 : vardata->atttypmod = var->vartypmod;
5654 1360299 : vardata->isunique = has_unique_index(vardata->rel, var->varattno);
5655 :
5656 : /* Try to locate some stats */
5657 1360299 : examine_simple_variable(root, var, vardata);
5658 :
5659 1360299 : 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 534712 : varnos = pull_varnos(root, basenode);
5668 534712 : basevarnos = bms_difference(varnos, root->outer_join_rels);
5669 :
5670 534712 : onerel = NULL;
5671 :
5672 534712 : 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 281792 : if (bms_get_singleton_member(basevarnos, &relid))
5682 : {
5683 279371 : if (varRelid == 0 || varRelid == relid)
5684 : {
5685 36787 : onerel = find_base_rel(root, relid);
5686 36787 : vardata->rel = onerel;
5687 36787 : node = basenode; /* strip any phvs or relabeling */
5688 : }
5689 : /* else treat it as a constant */
5690 : }
5691 : else
5692 : {
5693 : /* varnos has multiple relids */
5694 2421 : if (varRelid == 0)
5695 : {
5696 : /* treat it as a variable of a join relation */
5697 1765 : vardata->rel = find_join_rel(root, varnos);
5698 1765 : node = basenode; /* strip any phvs or relabeling */
5699 : }
5700 656 : else if (bms_is_member(varRelid, varnos))
5701 : {
5702 : /* ignore the vars belonging to other relations */
5703 599 : vardata->rel = find_base_rel(root, varRelid);
5704 599 : node = basenode; /* strip any phvs or relabeling */
5705 : /* note: no point in expressional-index search here */
5706 : }
5707 : /* else treat it as a constant */
5708 : }
5709 : }
5710 :
5711 534712 : bms_free(basevarnos);
5712 :
5713 534712 : vardata->var = node;
5714 534712 : vardata->atttype = exprType(node);
5715 534712 : vardata->atttypmod = exprTypmod(node);
5716 :
5717 534712 : 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 36787 : if (bms_overlap(varnos, root->outer_join_rels))
5741 1289 : node = remove_nulling_relids(node, root->outer_join_rels, NULL);
5742 :
5743 81912 : foreach(ilist, onerel->indexlist)
5744 : {
5745 46616 : IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
5746 : ListCell *indexpr_item;
5747 : int pos;
5748 :
5749 46616 : indexpr_item = list_head(index->indexprs);
5750 46616 : if (indexpr_item == NULL)
5751 44180 : continue; /* no expressions here... */
5752 :
5753 3417 : for (pos = 0; pos < index->ncolumns; pos++)
5754 : {
5755 2472 : if (index->indexkeys[pos] == 0)
5756 : {
5757 : Node *indexkey;
5758 :
5759 2436 : if (indexpr_item == NULL)
5760 0 : elog(ERROR, "too few entries in indexprs list");
5761 2436 : indexkey = (Node *) lfirst(indexpr_item);
5762 2436 : if (indexkey && IsA(indexkey, RelabelType))
5763 0 : indexkey = (Node *) ((RelabelType *) indexkey)->arg;
5764 2436 : 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 1809 : if (index->unique &&
5771 219 : index->nkeycolumns == 1 &&
5772 219 : pos == 0 &&
5773 219 : (index->indpred == NIL || index->predOK))
5774 219 : 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 1809 : 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 1809 : else if (index->indpred == NIL)
5800 : {
5801 1809 : vardata->statsTuple =
5802 3618 : SearchSysCache3(STATRELATTINH,
5803 : ObjectIdGetDatum(index->indexoid),
5804 1809 : Int16GetDatum(pos + 1),
5805 : BoolGetDatum(false));
5806 1809 : vardata->freefunc = ReleaseSysCache;
5807 :
5808 1809 : 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 1491 : vardata->acl_ok =
5833 1491 : all_rows_selectable(root,
5834 1491 : index->rel->relid,
5835 : NULL);
5836 : }
5837 : else
5838 : {
5839 : /* suppress leakproofness checks later */
5840 318 : vardata->acl_ok = true;
5841 : }
5842 : }
5843 1809 : if (vardata->statsTuple)
5844 1491 : break;
5845 : }
5846 945 : indexpr_item = lnext(index->indexprs, indexpr_item);
5847 : }
5848 : }
5849 2436 : if (vardata->statsTuple)
5850 1491 : 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 38873 : foreach(slist, onerel->statlist)
5860 : {
5861 2230 : StatisticExtInfo *info = (StatisticExtInfo *) lfirst(slist);
5862 2230 : 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 2230 : if (vardata->statsTuple)
5871 144 : break;
5872 :
5873 : /* skip stats without per-expression stats */
5874 2086 : if (info->kind != STATS_EXT_EXPRESSIONS)
5875 1074 : continue;
5876 :
5877 : /* skip stats with mismatching stxdinherit value */
5878 1012 : if (info->inherit != rte->inh)
5879 3 : continue;
5880 :
5881 1009 : pos = 0;
5882 1660 : foreach(expr_item, info->exprs)
5883 : {
5884 1483 : Node *expr = (Node *) lfirst(expr_item);
5885 :
5886 : Assert(expr);
5887 :
5888 : /* strip RelabelType before comparing it */
5889 1483 : 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 1483 : 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 832 : vardata->statsTuple =
5900 832 : statext_expressions_load(info->statOid, rte->inh, pos);
5901 :
5902 : /* Nothing to release if no data found */
5903 832 : if (vardata->statsTuple != NULL)
5904 : {
5905 831 : vardata->freefunc = ReleaseDummy;
5906 : }
5907 :
5908 : /*
5909 : * Test if user has permission to access all rows from the
5910 : * table.
5911 : *
5912 : * For simplicity, we insist on the whole table being
5913 : * selectable, rather than trying to identify which
5914 : * column(s) the statistics object depends on.
5915 : *
5916 : * Note that for an inheritance child, permissions are
5917 : * checked on the inheritance root parent, and whole-table
5918 : * select privilege on the parent doesn't quite guarantee
5919 : * that the user could read all columns of the child. But
5920 : * in practice it's unlikely that any interesting security
5921 : * violation could result from allowing access to the
5922 : * expression stats, so we allow it anyway. See similar
5923 : * code in examine_simple_variable() for additional
5924 : * comments.
5925 : */
5926 832 : vardata->acl_ok = all_rows_selectable(root,
5927 : onerel->relid,
5928 : NULL);
5929 :
5930 832 : break;
5931 : }
5932 :
5933 651 : pos++;
5934 : }
5935 : }
5936 : }
5937 :
5938 534712 : bms_free(varnos);
5939 : }
5940 :
5941 : /*
5942 : * strip_all_phvs_deep
5943 : * Deeply strip all PlaceHolderVars in an expression.
5944 :
5945 : * As a performance optimization, we first use a lightweight walker to check
5946 : * for the presence of any PlaceHolderVars. The expensive mutator is invoked
5947 : * only if a PlaceHolderVar is found, avoiding unnecessary memory allocation
5948 : * and tree copying in the common case where no PlaceHolderVars are present.
5949 : */
5950 : static Node *
5951 1895011 : strip_all_phvs_deep(PlannerInfo *root, Node *node)
5952 : {
5953 : /* If there are no PHVs anywhere, we needn't work hard */
5954 1895011 : if (root->glob->lastPHId == 0)
5955 1877713 : return node;
5956 :
5957 17298 : if (!contain_placeholder_walker(node, NULL))
5958 15129 : return node;
5959 2169 : return strip_all_phvs_mutator(node, NULL);
5960 : }
5961 :
5962 : /*
5963 : * contain_placeholder_walker
5964 : * Lightweight walker to check if an expression contains any
5965 : * PlaceHolderVars
5966 : */
5967 : static bool
5968 19597 : contain_placeholder_walker(Node *node, void *context)
5969 : {
5970 19597 : if (node == NULL)
5971 105 : return false;
5972 19492 : if (IsA(node, PlaceHolderVar))
5973 2169 : return true;
5974 :
5975 17323 : return expression_tree_walker(node, contain_placeholder_walker, context);
5976 : }
5977 :
5978 : /*
5979 : * strip_all_phvs_mutator
5980 : * Mutator to deeply strip all PlaceHolderVars
5981 : */
5982 : static Node *
5983 5847 : strip_all_phvs_mutator(Node *node, void *context)
5984 : {
5985 5847 : if (node == NULL)
5986 24 : return NULL;
5987 5823 : if (IsA(node, PlaceHolderVar))
5988 : {
5989 : /* Strip it and recurse into its contained expression */
5990 2244 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
5991 :
5992 2244 : return strip_all_phvs_mutator((Node *) phv->phexpr, context);
5993 : }
5994 :
5995 3579 : return expression_tree_mutator(node, strip_all_phvs_mutator, context);
5996 : }
5997 :
5998 : /*
5999 : * examine_simple_variable
6000 : * Handle a simple Var for examine_variable
6001 : *
6002 : * This is split out as a subroutine so that we can recurse to deal with
6003 : * Vars referencing subqueries (either sub-SELECT-in-FROM or CTE style).
6004 : *
6005 : * We already filled in all the fields of *vardata except for the stats tuple.
6006 : */
6007 : static void
6008 1363492 : examine_simple_variable(PlannerInfo *root, Var *var,
6009 : VariableStatData *vardata)
6010 : {
6011 1363492 : RangeTblEntry *rte = root->simple_rte_array[var->varno];
6012 :
6013 : Assert(IsA(rte, RangeTblEntry));
6014 :
6015 1363492 : if (get_relation_stats_hook &&
6016 0 : (*get_relation_stats_hook) (root, rte, var->varattno, vardata))
6017 : {
6018 : /*
6019 : * The hook took control of acquiring a stats tuple. If it did supply
6020 : * a tuple, it'd better have supplied a freefunc.
6021 : */
6022 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6023 0 : !vardata->freefunc)
6024 0 : elog(ERROR, "no function provided to release variable stats with");
6025 : }
6026 1363492 : else if (rte->rtekind == RTE_RELATION)
6027 : {
6028 : /*
6029 : * Plain table or parent of an inheritance appendrel, so look up the
6030 : * column in pg_statistic
6031 : */
6032 1298329 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6033 : ObjectIdGetDatum(rte->relid),
6034 1298329 : Int16GetDatum(var->varattno),
6035 1298329 : BoolGetDatum(rte->inh));
6036 1298329 : vardata->freefunc = ReleaseSysCache;
6037 :
6038 1298329 : if (HeapTupleIsValid(vardata->statsTuple))
6039 : {
6040 : /*
6041 : * Test if user has permission to read all rows from this column.
6042 : *
6043 : * This requires that the user has the appropriate SELECT
6044 : * privileges and that there are no securityQuals from security
6045 : * barrier views or RLS policies. If that's not the case, then we
6046 : * only permit leakproof functions to be passed pg_statistic data
6047 : * in vardata, otherwise the functions might reveal data that the
6048 : * user doesn't have permission to see --- see
6049 : * statistic_proc_security_check().
6050 : */
6051 899537 : vardata->acl_ok =
6052 899537 : all_rows_selectable(root, var->varno,
6053 899537 : bms_make_singleton(var->varattno - FirstLowInvalidHeapAttributeNumber));
6054 : }
6055 : else
6056 : {
6057 : /* suppress any possible leakproofness checks later */
6058 398792 : vardata->acl_ok = true;
6059 : }
6060 : }
6061 65163 : else if ((rte->rtekind == RTE_SUBQUERY && !rte->inh) ||
6062 58981 : (rte->rtekind == RTE_CTE && !rte->self_reference))
6063 : {
6064 : /*
6065 : * Plain subquery (not one that was converted to an appendrel) or
6066 : * non-recursive CTE. In either case, we can try to find out what the
6067 : * Var refers to within the subquery. We skip this for appendrel and
6068 : * recursive-CTE cases because any column stats we did find would
6069 : * likely not be very relevant.
6070 : */
6071 : PlannerInfo *subroot;
6072 : Query *subquery;
6073 : List *subtlist;
6074 : TargetEntry *ste;
6075 :
6076 : /*
6077 : * Punt if it's a whole-row var rather than a plain column reference.
6078 : */
6079 9889 : if (var->varattno == InvalidAttrNumber)
6080 0 : return;
6081 :
6082 : /*
6083 : * Otherwise, find the subquery's planner subroot.
6084 : */
6085 9889 : if (rte->rtekind == RTE_SUBQUERY)
6086 : {
6087 : RelOptInfo *rel;
6088 :
6089 : /*
6090 : * Fetch RelOptInfo for subquery. Note that we don't change the
6091 : * rel returned in vardata, since caller expects it to be a rel of
6092 : * the caller's query level. Because we might already be
6093 : * recursing, we can't use that rel pointer either, but have to
6094 : * look up the Var's rel afresh.
6095 : */
6096 6182 : rel = find_base_rel(root, var->varno);
6097 :
6098 6182 : subroot = rel->subroot;
6099 : }
6100 : else
6101 : {
6102 : /* CTE case is more difficult */
6103 : PlannerInfo *cteroot;
6104 : Index levelsup;
6105 : int ndx;
6106 : int plan_id;
6107 : ListCell *lc;
6108 :
6109 : /*
6110 : * Find the referenced CTE, and locate the subroot previously made
6111 : * for it.
6112 : */
6113 3707 : levelsup = rte->ctelevelsup;
6114 3707 : cteroot = root;
6115 6922 : while (levelsup-- > 0)
6116 : {
6117 3215 : cteroot = cteroot->parent_root;
6118 3215 : if (!cteroot) /* shouldn't happen */
6119 0 : elog(ERROR, "bad levelsup for CTE \"%s\"", rte->ctename);
6120 : }
6121 :
6122 : /*
6123 : * Note: cte_plan_ids can be shorter than cteList, if we are still
6124 : * working on planning the CTEs (ie, this is a side-reference from
6125 : * another CTE). So we mustn't use forboth here.
6126 : */
6127 3707 : ndx = 0;
6128 4822 : foreach(lc, cteroot->parse->cteList)
6129 : {
6130 4822 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
6131 :
6132 4822 : if (strcmp(cte->ctename, rte->ctename) == 0)
6133 3707 : break;
6134 1115 : ndx++;
6135 : }
6136 3707 : if (lc == NULL) /* shouldn't happen */
6137 0 : elog(ERROR, "could not find CTE \"%s\"", rte->ctename);
6138 3707 : if (ndx >= list_length(cteroot->cte_plan_ids))
6139 0 : elog(ERROR, "could not find plan for CTE \"%s\"", rte->ctename);
6140 3707 : plan_id = list_nth_int(cteroot->cte_plan_ids, ndx);
6141 3707 : if (plan_id <= 0)
6142 0 : elog(ERROR, "no plan was made for CTE \"%s\"", rte->ctename);
6143 3707 : subroot = list_nth(root->glob->subroots, plan_id - 1);
6144 : }
6145 :
6146 : /* If the subquery hasn't been planned yet, we have to punt */
6147 9889 : if (subroot == NULL)
6148 0 : return;
6149 : Assert(IsA(subroot, PlannerInfo));
6150 :
6151 : /*
6152 : * We must use the subquery parsetree as mangled by the planner, not
6153 : * the raw version from the RTE, because we need a Var that will refer
6154 : * to the subroot's live RelOptInfos. For instance, if any subquery
6155 : * pullup happened during planning, Vars in the targetlist might have
6156 : * gotten replaced, and we need to see the replacement expressions.
6157 : */
6158 9889 : subquery = subroot->parse;
6159 : Assert(IsA(subquery, Query));
6160 :
6161 : /*
6162 : * Punt if subquery uses set operations or grouping sets, as these
6163 : * will mash underlying columns' stats beyond recognition. (Set ops
6164 : * are particularly nasty; if we forged ahead, we would return stats
6165 : * relevant to only the leftmost subselect...) DISTINCT is also
6166 : * problematic, but we check that later because there is a possibility
6167 : * of learning something even with it.
6168 : */
6169 9889 : if (subquery->setOperations ||
6170 8676 : subquery->groupingSets)
6171 1261 : return;
6172 :
6173 : /* Get the subquery output expression referenced by the upper Var */
6174 8628 : if (subquery->returningList)
6175 103 : subtlist = subquery->returningList;
6176 : else
6177 8525 : subtlist = subquery->targetList;
6178 8628 : ste = get_tle_by_resno(subtlist, var->varattno);
6179 8628 : if (ste == NULL || ste->resjunk)
6180 0 : elog(ERROR, "subquery %s does not have attribute %d",
6181 : rte->eref->aliasname, var->varattno);
6182 8628 : var = (Var *) ste->expr;
6183 :
6184 : /*
6185 : * If subquery uses DISTINCT, we can't make use of any stats for the
6186 : * variable ... but, if it's the only DISTINCT column, we are entitled
6187 : * to consider it unique. We do the test this way so that it works
6188 : * for cases involving DISTINCT ON.
6189 : */
6190 8628 : if (subquery->distinctClause)
6191 : {
6192 941 : if (list_length(subquery->distinctClause) == 1 &&
6193 319 : targetIsInSortList(ste, InvalidOid, subquery->distinctClause))
6194 158 : vardata->isunique = true;
6195 : /* cannot go further */
6196 622 : return;
6197 : }
6198 :
6199 : /* The same idea as with DISTINCT clause works for a GROUP-BY too */
6200 8006 : if (subquery->groupClause)
6201 : {
6202 564 : if (list_length(subquery->groupClause) == 1 &&
6203 237 : targetIsInSortList(ste, InvalidOid, subquery->groupClause))
6204 181 : vardata->isunique = true;
6205 : /* cannot go further */
6206 327 : return;
6207 : }
6208 :
6209 : /*
6210 : * If the sub-query originated from a view with the security_barrier
6211 : * attribute, we must not look at the variable's statistics, though it
6212 : * seems all right to notice the existence of a DISTINCT clause. So
6213 : * stop here.
6214 : *
6215 : * This is probably a harsher restriction than necessary; it's
6216 : * certainly OK for the selectivity estimator (which is a C function,
6217 : * and therefore omnipotent anyway) to look at the statistics. But
6218 : * many selectivity estimators will happily *invoke the operator
6219 : * function* to try to work out a good estimate - and that's not OK.
6220 : * So for now, don't dig down for stats.
6221 : */
6222 7679 : if (rte->security_barrier)
6223 1431 : return;
6224 :
6225 : /* Can only handle a simple Var of subquery's query level */
6226 6248 : if (var && IsA(var, Var) &&
6227 3193 : var->varlevelsup == 0)
6228 : {
6229 : /*
6230 : * OK, recurse into the subquery. Note that the original setting
6231 : * of vardata->isunique (which will surely be false) is left
6232 : * unchanged in this situation. That's what we want, since even
6233 : * if the underlying column is unique, the subquery may have
6234 : * joined to other tables in a way that creates duplicates.
6235 : */
6236 3193 : examine_simple_variable(subroot, var, vardata);
6237 : }
6238 : }
6239 : else
6240 : {
6241 : /*
6242 : * Otherwise, the Var comes from a FUNCTION or VALUES RTE. (We won't
6243 : * see RTE_JOIN here because join alias Vars have already been
6244 : * flattened.) There's not much we can do with function outputs, but
6245 : * maybe someday try to be smarter about VALUES.
6246 : */
6247 : }
6248 : }
6249 :
6250 : /*
6251 : * all_rows_selectable
6252 : * Test whether the user has permission to select all rows from a given
6253 : * relation.
6254 : *
6255 : * Inputs:
6256 : * root: the planner info
6257 : * varno: the index of the relation (assumed to be an RTE_RELATION)
6258 : * varattnos: the attributes for which permission is required, or NULL if
6259 : * whole-table access is required
6260 : *
6261 : * Returns true if the user has the required select permissions, and there are
6262 : * no securityQuals from security barrier views or RLS policies.
6263 : *
6264 : * Note that if the relation is an inheritance child relation, securityQuals
6265 : * and access permissions are checked against the inheritance root parent (the
6266 : * relation actually mentioned in the query) --- see the comments in
6267 : * expand_single_inheritance_child() for an explanation of why it has to be
6268 : * done this way.
6269 : *
6270 : * If varattnos is non-NULL, its attribute numbers should be offset by
6271 : * FirstLowInvalidHeapAttributeNumber so that system attributes can be
6272 : * checked. If varattnos is NULL, only table-level SELECT privileges are
6273 : * checked, not any column-level privileges.
6274 : *
6275 : * Note: if the relation is accessed via a view, this function actually tests
6276 : * whether the view owner has permission to select from the relation. To
6277 : * ensure that the current user has permission, it is also necessary to check
6278 : * that the current user has permission to select from the view, which we do
6279 : * at planner-startup --- see subquery_planner().
6280 : *
6281 : * This is exported so that other estimation functions can use it.
6282 : */
6283 : bool
6284 901986 : all_rows_selectable(PlannerInfo *root, Index varno, Bitmapset *varattnos)
6285 : {
6286 901986 : RelOptInfo *rel = find_base_rel_noerr(root, varno);
6287 901986 : RangeTblEntry *rte = planner_rt_fetch(varno, root);
6288 : Oid userid;
6289 : int varattno;
6290 :
6291 : Assert(rte->rtekind == RTE_RELATION);
6292 :
6293 : /*
6294 : * Determine the user ID to use for privilege checks (either the current
6295 : * user or the view owner, if we're accessing the table via a view).
6296 : *
6297 : * Normally the relation will have an associated RelOptInfo from which we
6298 : * can find the userid, but it might not if it's a RETURNING Var for an
6299 : * INSERT target relation. In that case use the RTEPermissionInfo
6300 : * associated with the RTE.
6301 : *
6302 : * If we navigate up to a parent relation, we keep using the same userid,
6303 : * since it's the same in all relations of a given inheritance tree.
6304 : */
6305 901986 : if (rel)
6306 901965 : userid = rel->userid;
6307 : else
6308 : {
6309 : RTEPermissionInfo *perminfo;
6310 :
6311 21 : perminfo = getRTEPermissionInfo(root->parse->rteperminfos, rte);
6312 21 : userid = perminfo->checkAsUser;
6313 : }
6314 901986 : if (!OidIsValid(userid))
6315 796242 : userid = GetUserId();
6316 :
6317 : /*
6318 : * Permissions and securityQuals must be checked on the table actually
6319 : * mentioned in the query, so if this is an inheritance child, navigate up
6320 : * to the inheritance root parent. If the user can read the whole table
6321 : * or the required columns there, then they can read from the child table
6322 : * too. For per-column checks, we must find out which of the root
6323 : * parent's attributes the child relation's attributes correspond to.
6324 : */
6325 901986 : if (root->append_rel_array != NULL)
6326 : {
6327 : AppendRelInfo *appinfo;
6328 :
6329 116636 : appinfo = root->append_rel_array[varno];
6330 :
6331 : /*
6332 : * Partitions are mapped to their immediate parent, not the root
6333 : * parent, so must be ready to walk up multiple AppendRelInfos. But
6334 : * stop if we hit a parent that is not RTE_RELATION --- that's a
6335 : * flattened UNION ALL subquery, not an inheritance parent.
6336 : */
6337 217522 : while (appinfo &&
6338 101072 : planner_rt_fetch(appinfo->parent_relid,
6339 101072 : root)->rtekind == RTE_RELATION)
6340 : {
6341 100886 : Bitmapset *parent_varattnos = NULL;
6342 :
6343 : /*
6344 : * For each child attribute, find the corresponding parent
6345 : * attribute. In rare cases, the attribute may be local to the
6346 : * child table, in which case, we've got to live with having no
6347 : * access to this column.
6348 : */
6349 100886 : varattno = -1;
6350 200347 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6351 : {
6352 : AttrNumber attno;
6353 : AttrNumber parent_attno;
6354 :
6355 99461 : attno = varattno + FirstLowInvalidHeapAttributeNumber;
6356 :
6357 99461 : if (attno == InvalidAttrNumber)
6358 : {
6359 : /*
6360 : * Whole-row reference, so must map each column of the
6361 : * child to the parent table.
6362 : */
6363 18 : for (attno = 1; attno <= appinfo->num_child_cols; attno++)
6364 : {
6365 12 : parent_attno = appinfo->parent_colnos[attno - 1];
6366 12 : if (parent_attno == 0)
6367 0 : return false; /* attr is local to child */
6368 : parent_varattnos =
6369 12 : bms_add_member(parent_varattnos,
6370 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6371 : }
6372 : }
6373 : else
6374 : {
6375 99455 : if (attno < 0)
6376 : {
6377 : /* System attnos are the same in all tables */
6378 0 : parent_attno = attno;
6379 : }
6380 : else
6381 : {
6382 99455 : if (attno > appinfo->num_child_cols)
6383 0 : return false; /* safety check */
6384 99455 : parent_attno = appinfo->parent_colnos[attno - 1];
6385 99455 : if (parent_attno == 0)
6386 0 : return false; /* attr is local to child */
6387 : }
6388 : parent_varattnos =
6389 99455 : bms_add_member(parent_varattnos,
6390 : parent_attno - FirstLowInvalidHeapAttributeNumber);
6391 : }
6392 : }
6393 :
6394 : /* If the parent is itself a child, continue up */
6395 100886 : varno = appinfo->parent_relid;
6396 100886 : varattnos = parent_varattnos;
6397 100886 : appinfo = root->append_rel_array[varno];
6398 : }
6399 :
6400 : /* Perform the access check on this parent rel */
6401 116636 : rte = planner_rt_fetch(varno, root);
6402 : Assert(rte->rtekind == RTE_RELATION);
6403 : }
6404 :
6405 : /*
6406 : * For all rows to be accessible, there must be no securityQuals from
6407 : * security barrier views or RLS policies.
6408 : */
6409 901986 : if (rte->securityQuals != NIL)
6410 414 : return false;
6411 :
6412 : /*
6413 : * Test for table-level SELECT privilege.
6414 : *
6415 : * If varattnos is non-NULL, this is sufficient to give access to all
6416 : * requested attributes, even for a child table, since we have verified
6417 : * that all required child columns have matching parent columns.
6418 : *
6419 : * If varattnos is NULL (whole-table access requested), this doesn't
6420 : * necessarily guarantee that the user can read all columns of a child
6421 : * table, but we allow it anyway (see comments in examine_variable()) and
6422 : * don't bother checking any column privileges.
6423 : */
6424 901572 : if (pg_class_aclcheck(rte->relid, userid, ACL_SELECT) == ACLCHECK_OK)
6425 901346 : return true;
6426 :
6427 226 : if (varattnos == NULL)
6428 6 : return false; /* whole-table access requested */
6429 :
6430 : /*
6431 : * Don't have table-level SELECT privilege, so check per-column
6432 : * privileges.
6433 : */
6434 220 : varattno = -1;
6435 323 : while ((varattno = bms_next_member(varattnos, varattno)) >= 0)
6436 : {
6437 220 : AttrNumber attno = varattno + FirstLowInvalidHeapAttributeNumber;
6438 :
6439 220 : if (attno == InvalidAttrNumber)
6440 : {
6441 : /* Whole-row reference, so must have access to all columns */
6442 3 : if (pg_attribute_aclcheck_all(rte->relid, userid, ACL_SELECT,
6443 : ACLMASK_ALL) != ACLCHECK_OK)
6444 3 : return false;
6445 : }
6446 : else
6447 : {
6448 217 : if (pg_attribute_aclcheck(rte->relid, attno, userid,
6449 : ACL_SELECT) != ACLCHECK_OK)
6450 114 : return false;
6451 : }
6452 : }
6453 :
6454 : /* If we reach here, have all required column privileges */
6455 103 : return true;
6456 : }
6457 :
6458 : /*
6459 : * examine_indexcol_variable
6460 : * Try to look up statistical data about an index column/expression.
6461 : * Fill in a VariableStatData struct to describe the column.
6462 : *
6463 : * Inputs:
6464 : * root: the planner info
6465 : * index: the index whose column we're interested in
6466 : * indexcol: 0-based index column number (subscripts index->indexkeys[])
6467 : *
6468 : * Outputs: *vardata is filled as follows:
6469 : * var: the input expression (with any binary relabeling stripped, if
6470 : * it is or contains a variable; but otherwise the type is preserved)
6471 : * rel: RelOptInfo for table relation containing variable.
6472 : * statsTuple: the pg_statistic entry for the variable, if one exists;
6473 : * otherwise NULL.
6474 : * freefunc: pointer to a function to release statsTuple with.
6475 : *
6476 : * Caller is responsible for doing ReleaseVariableStats() before exiting.
6477 : */
6478 : static void
6479 472177 : examine_indexcol_variable(PlannerInfo *root, IndexOptInfo *index,
6480 : int indexcol, VariableStatData *vardata)
6481 : {
6482 : AttrNumber colnum;
6483 : Oid relid;
6484 :
6485 472177 : if (index->indexkeys[indexcol] != 0)
6486 : {
6487 : /* Simple variable --- look to stats for the underlying table */
6488 471070 : RangeTblEntry *rte = planner_rt_fetch(index->rel->relid, root);
6489 :
6490 : Assert(rte->rtekind == RTE_RELATION);
6491 471070 : relid = rte->relid;
6492 : Assert(relid != InvalidOid);
6493 471070 : colnum = index->indexkeys[indexcol];
6494 471070 : vardata->rel = index->rel;
6495 :
6496 471070 : if (get_relation_stats_hook &&
6497 0 : (*get_relation_stats_hook) (root, rte, colnum, vardata))
6498 : {
6499 : /*
6500 : * The hook took control of acquiring a stats tuple. If it did
6501 : * supply a tuple, it'd better have supplied a freefunc.
6502 : */
6503 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6504 0 : !vardata->freefunc)
6505 0 : elog(ERROR, "no function provided to release variable stats with");
6506 : }
6507 : else
6508 : {
6509 471070 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6510 : ObjectIdGetDatum(relid),
6511 : Int16GetDatum(colnum),
6512 471070 : BoolGetDatum(rte->inh));
6513 471070 : vardata->freefunc = ReleaseSysCache;
6514 : }
6515 : }
6516 : else
6517 : {
6518 : /* Expression --- maybe there are stats for the index itself */
6519 1107 : relid = index->indexoid;
6520 1107 : colnum = indexcol + 1;
6521 :
6522 1107 : if (get_index_stats_hook &&
6523 0 : (*get_index_stats_hook) (root, relid, colnum, vardata))
6524 : {
6525 : /*
6526 : * The hook took control of acquiring a stats tuple. If it did
6527 : * supply a tuple, it'd better have supplied a freefunc.
6528 : */
6529 0 : if (HeapTupleIsValid(vardata->statsTuple) &&
6530 0 : !vardata->freefunc)
6531 0 : elog(ERROR, "no function provided to release variable stats with");
6532 : }
6533 : else
6534 : {
6535 1107 : vardata->statsTuple = SearchSysCache3(STATRELATTINH,
6536 : ObjectIdGetDatum(relid),
6537 : Int16GetDatum(colnum),
6538 : BoolGetDatum(false));
6539 1107 : vardata->freefunc = ReleaseSysCache;
6540 : }
6541 : }
6542 472177 : }
6543 :
6544 : /*
6545 : * Check whether it is permitted to call func_oid passing some of the
6546 : * pg_statistic data in vardata. We allow this if either of the following
6547 : * conditions is met: (1) the user has SELECT privileges on the table or
6548 : * column underlying the pg_statistic data and there are no securityQuals from
6549 : * security barrier views or RLS policies, or (2) the function is marked
6550 : * leakproof.
6551 : */
6552 : bool
6553 620298 : statistic_proc_security_check(VariableStatData *vardata, Oid func_oid)
6554 : {
6555 620298 : if (vardata->acl_ok)
6556 619373 : return true; /* have SELECT privs and no securityQuals */
6557 :
6558 925 : if (!OidIsValid(func_oid))
6559 0 : return false;
6560 :
6561 925 : if (get_func_leakproof(func_oid))
6562 458 : return true;
6563 :
6564 467 : ereport(DEBUG2,
6565 : (errmsg_internal("not using statistics because function \"%s\" is not leakproof",
6566 : get_func_name(func_oid))));
6567 467 : return false;
6568 : }
6569 :
6570 : /*
6571 : * get_variable_numdistinct
6572 : * Estimate the number of distinct values of a variable.
6573 : *
6574 : * vardata: results of examine_variable
6575 : * *isdefault: set to true if the result is a default rather than based on
6576 : * anything meaningful.
6577 : *
6578 : * NB: be careful to produce a positive integral result, since callers may
6579 : * compare the result to exact integer counts, or might divide by it.
6580 : */
6581 : double
6582 971664 : get_variable_numdistinct(VariableStatData *vardata, bool *isdefault)
6583 : {
6584 : double stadistinct;
6585 971664 : double stanullfrac = 0.0;
6586 : double ntuples;
6587 :
6588 971664 : *isdefault = false;
6589 :
6590 : /*
6591 : * Determine the stadistinct value to use. There are cases where we can
6592 : * get an estimate even without a pg_statistic entry, or can get a better
6593 : * value than is in pg_statistic. Grab stanullfrac too if we can find it
6594 : * (otherwise, assume no nulls, for lack of any better idea).
6595 : */
6596 971664 : if (HeapTupleIsValid(vardata->statsTuple))
6597 : {
6598 : /* Use the pg_statistic entry */
6599 : Form_pg_statistic stats;
6600 :
6601 637352 : stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);
6602 637352 : stadistinct = stats->stadistinct;
6603 637352 : stanullfrac = stats->stanullfrac;
6604 : }
6605 334312 : else if (vardata->vartype == BOOLOID)
6606 : {
6607 : /*
6608 : * Special-case boolean columns: presumably, two distinct values.
6609 : *
6610 : * Are there any other datatypes we should wire in special estimates
6611 : * for?
6612 : */
6613 480 : stadistinct = 2.0;
6614 : }
6615 333832 : else if (vardata->rel && vardata->rel->rtekind == RTE_VALUES)
6616 : {
6617 : /*
6618 : * If the Var represents a column of a VALUES RTE, assume it's unique.
6619 : * This could of course be very wrong, but it should tend to be true
6620 : * in well-written queries. We could consider examining the VALUES'
6621 : * contents to get some real statistics; but that only works if the
6622 : * entries are all constants, and it would be pretty expensive anyway.
6623 : */
6624 1797 : stadistinct = -1.0; /* unique (and all non null) */
6625 : }
6626 : else
6627 : {
6628 : /*
6629 : * We don't keep statistics for system columns, but in some cases we
6630 : * can infer distinctness anyway.
6631 : */
6632 332035 : if (vardata->var && IsA(vardata->var, Var))
6633 : {
6634 310326 : switch (((Var *) vardata->var)->varattno)
6635 : {
6636 612 : case SelfItemPointerAttributeNumber:
6637 612 : stadistinct = -1.0; /* unique (and all non null) */
6638 612 : break;
6639 17543 : case TableOidAttributeNumber:
6640 17543 : stadistinct = 1.0; /* only 1 value */
6641 17543 : break;
6642 292171 : default:
6643 292171 : stadistinct = 0.0; /* means "unknown" */
6644 292171 : break;
6645 : }
6646 : }
6647 : else
6648 21709 : stadistinct = 0.0; /* means "unknown" */
6649 :
6650 : /*
6651 : * XXX consider using estimate_num_groups on expressions?
6652 : */
6653 : }
6654 :
6655 : /*
6656 : * If there is a unique index, DISTINCT or GROUP-BY clause for the
6657 : * variable, assume it is unique no matter what pg_statistic says; the
6658 : * statistics could be out of date, or we might have found a partial
6659 : * unique index that proves the var is unique for this query. However,
6660 : * we'd better still believe the null-fraction statistic.
6661 : */
6662 971664 : if (vardata->isunique)
6663 245838 : stadistinct = -1.0 * (1.0 - stanullfrac);
6664 :
6665 : /*
6666 : * If we had an absolute estimate, use that.
6667 : */
6668 971664 : if (stadistinct > 0.0)
6669 237545 : return clamp_row_est(stadistinct);
6670 :
6671 : /*
6672 : * Otherwise we need to get the relation size; punt if not available.
6673 : */
6674 734119 : if (vardata->rel == NULL)
6675 : {
6676 354 : *isdefault = true;
6677 354 : return DEFAULT_NUM_DISTINCT;
6678 : }
6679 733765 : ntuples = vardata->rel->tuples;
6680 733765 : if (ntuples <= 0.0)
6681 : {
6682 87935 : *isdefault = true;
6683 87935 : return DEFAULT_NUM_DISTINCT;
6684 : }
6685 :
6686 : /*
6687 : * If we had a relative estimate, use that.
6688 : */
6689 645830 : if (stadistinct < 0.0)
6690 451991 : return clamp_row_est(-stadistinct * ntuples);
6691 :
6692 : /*
6693 : * With no data, estimate ndistinct = ntuples if the table is small, else
6694 : * use default. We use DEFAULT_NUM_DISTINCT as the cutoff for "small" so
6695 : * that the behavior isn't discontinuous.
6696 : */
6697 193839 : if (ntuples < DEFAULT_NUM_DISTINCT)
6698 92600 : return clamp_row_est(ntuples);
6699 :
6700 101239 : *isdefault = true;
6701 101239 : return DEFAULT_NUM_DISTINCT;
6702 : }
6703 :
6704 : /*
6705 : * get_variable_range
6706 : * Estimate the minimum and maximum value of the specified variable.
6707 : * If successful, store values in *min and *max, and return true.
6708 : * If no data available, return false.
6709 : *
6710 : * sortop is the "<" comparison operator to use. This should generally
6711 : * be "<" not ">", as only the former is likely to be found in pg_statistic.
6712 : * The collation must be specified too.
6713 : */
6714 : static bool
6715 139088 : get_variable_range(PlannerInfo *root, VariableStatData *vardata,
6716 : Oid sortop, Oid collation,
6717 : Datum *min, Datum *max)
6718 : {
6719 139088 : Datum tmin = 0;
6720 139088 : Datum tmax = 0;
6721 139088 : bool have_data = false;
6722 : int16 typLen;
6723 : bool typByVal;
6724 : Oid opfuncoid;
6725 : FmgrInfo opproc;
6726 : AttStatsSlot sslot;
6727 :
6728 : /*
6729 : * XXX It's very tempting to try to use the actual column min and max, if
6730 : * we can get them relatively-cheaply with an index probe. However, since
6731 : * this function is called many times during join planning, that could
6732 : * have unpleasant effects on planning speed. Need more investigation
6733 : * before enabling this.
6734 : */
6735 : #ifdef NOT_USED
6736 : if (get_actual_variable_range(root, vardata, sortop, collation, min, max))
6737 : return true;
6738 : #endif
6739 :
6740 139088 : if (!HeapTupleIsValid(vardata->statsTuple))
6741 : {
6742 : /* no stats available, so default result */
6743 36508 : return false;
6744 : }
6745 :
6746 : /*
6747 : * If we can't apply the sortop to the stats data, just fail. In
6748 : * principle, if there's a histogram and no MCVs, we could return the
6749 : * histogram endpoints without ever applying the sortop ... but it's
6750 : * probably not worth trying, because whatever the caller wants to do with
6751 : * the endpoints would likely fail the security check too.
6752 : */
6753 102580 : if (!statistic_proc_security_check(vardata,
6754 102580 : (opfuncoid = get_opcode(sortop))))
6755 0 : return false;
6756 :
6757 102580 : opproc.fn_oid = InvalidOid; /* mark this as not looked up yet */
6758 :
6759 102580 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
6760 :
6761 : /*
6762 : * If there is a histogram with the ordering we want, grab the first and
6763 : * last values.
6764 : */
6765 102580 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6766 : STATISTIC_KIND_HISTOGRAM, sortop,
6767 : ATTSTATSSLOT_VALUES))
6768 : {
6769 65939 : if (sslot.stacoll == collation && sslot.nvalues > 0)
6770 : {
6771 65939 : tmin = datumCopy(sslot.values[0], typByVal, typLen);
6772 65939 : tmax = datumCopy(sslot.values[sslot.nvalues - 1], typByVal, typLen);
6773 65939 : have_data = true;
6774 : }
6775 65939 : free_attstatsslot(&sslot);
6776 : }
6777 :
6778 : /*
6779 : * Otherwise, if there is a histogram with some other ordering, scan it
6780 : * and get the min and max values according to the ordering we want. This
6781 : * of course may not find values that are really extremal according to our
6782 : * ordering, but it beats ignoring available data.
6783 : */
6784 139221 : if (!have_data &&
6785 36641 : get_attstatsslot(&sslot, vardata->statsTuple,
6786 : STATISTIC_KIND_HISTOGRAM, InvalidOid,
6787 : ATTSTATSSLOT_VALUES))
6788 : {
6789 0 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6790 : collation, typLen, typByVal,
6791 : &tmin, &tmax, &have_data);
6792 0 : free_attstatsslot(&sslot);
6793 : }
6794 :
6795 : /*
6796 : * If we have most-common-values info, look for extreme MCVs. This is
6797 : * needed even if we also have a histogram, since the histogram excludes
6798 : * the MCVs. However, if we *only* have MCVs and no histogram, we should
6799 : * be pretty wary of deciding that that is a full representation of the
6800 : * data. Proceed only if the MCVs represent the whole table (to within
6801 : * roundoff error).
6802 : */
6803 102580 : if (get_attstatsslot(&sslot, vardata->statsTuple,
6804 : STATISTIC_KIND_MCV, InvalidOid,
6805 102580 : have_data ? ATTSTATSSLOT_VALUES :
6806 : (ATTSTATSSLOT_VALUES | ATTSTATSSLOT_NUMBERS)))
6807 : {
6808 56925 : bool use_mcvs = have_data;
6809 :
6810 56925 : if (!have_data)
6811 : {
6812 35802 : double sumcommon = 0.0;
6813 : double nullfrac;
6814 : int i;
6815 :
6816 280356 : for (i = 0; i < sslot.nnumbers; i++)
6817 244554 : sumcommon += sslot.numbers[i];
6818 35802 : nullfrac = ((Form_pg_statistic) GETSTRUCT(vardata->statsTuple))->stanullfrac;
6819 35802 : if (sumcommon + nullfrac > 0.99999)
6820 35692 : use_mcvs = true;
6821 : }
6822 :
6823 56925 : if (use_mcvs)
6824 56815 : get_stats_slot_range(&sslot, opfuncoid, &opproc,
6825 : collation, typLen, typByVal,
6826 : &tmin, &tmax, &have_data);
6827 56925 : free_attstatsslot(&sslot);
6828 : }
6829 :
6830 102580 : *min = tmin;
6831 102580 : *max = tmax;
6832 102580 : return have_data;
6833 : }
6834 :
6835 : /*
6836 : * get_stats_slot_range: scan sslot for min/max values
6837 : *
6838 : * Subroutine for get_variable_range: update min/max/have_data according
6839 : * to what we find in the statistics array.
6840 : */
6841 : static void
6842 56815 : get_stats_slot_range(AttStatsSlot *sslot, Oid opfuncoid, FmgrInfo *opproc,
6843 : Oid collation, int16 typLen, bool typByVal,
6844 : Datum *min, Datum *max, bool *p_have_data)
6845 : {
6846 56815 : Datum tmin = *min;
6847 56815 : Datum tmax = *max;
6848 56815 : bool have_data = *p_have_data;
6849 56815 : bool found_tmin = false;
6850 56815 : bool found_tmax = false;
6851 :
6852 : /* Look up the comparison function, if we didn't already do so */
6853 56815 : if (opproc->fn_oid != opfuncoid)
6854 56815 : fmgr_info(opfuncoid, opproc);
6855 :
6856 : /* Scan all the slot's values */
6857 1418780 : for (int i = 0; i < sslot->nvalues; i++)
6858 : {
6859 1361965 : if (!have_data)
6860 : {
6861 35692 : tmin = tmax = sslot->values[i];
6862 35692 : found_tmin = found_tmax = true;
6863 35692 : *p_have_data = have_data = true;
6864 35692 : continue;
6865 : }
6866 1326273 : if (DatumGetBool(FunctionCall2Coll(opproc,
6867 : collation,
6868 1326273 : sslot->values[i], tmin)))
6869 : {
6870 33553 : tmin = sslot->values[i];
6871 33553 : found_tmin = true;
6872 : }
6873 1326273 : if (DatumGetBool(FunctionCall2Coll(opproc,
6874 : collation,
6875 1326273 : tmax, sslot->values[i])))
6876 : {
6877 133427 : tmax = sslot->values[i];
6878 133427 : found_tmax = true;
6879 : }
6880 : }
6881 :
6882 : /*
6883 : * Copy the slot's values, if we found new extreme values.
6884 : */
6885 56815 : if (found_tmin)
6886 48183 : *min = datumCopy(tmin, typByVal, typLen);
6887 56815 : if (found_tmax)
6888 37890 : *max = datumCopy(tmax, typByVal, typLen);
6889 56815 : }
6890 :
6891 :
6892 : /*
6893 : * get_actual_variable_range
6894 : * Attempt to identify the current *actual* minimum and/or maximum
6895 : * of the specified variable, by looking for a suitable btree index
6896 : * and fetching its low and/or high values.
6897 : * If successful, store values in *min and *max, and return true.
6898 : * (Either pointer can be NULL if that endpoint isn't needed.)
6899 : * If unsuccessful, return false.
6900 : *
6901 : * sortop is the "<" comparison operator to use.
6902 : * collation is the required collation.
6903 : */
6904 : static bool
6905 101745 : get_actual_variable_range(PlannerInfo *root, VariableStatData *vardata,
6906 : Oid sortop, Oid collation,
6907 : Datum *min, Datum *max)
6908 : {
6909 101745 : bool have_data = false;
6910 101745 : RelOptInfo *rel = vardata->rel;
6911 : RangeTblEntry *rte;
6912 : ListCell *lc;
6913 :
6914 : /* No hope if no relation or it doesn't have indexes */
6915 101745 : if (rel == NULL || rel->indexlist == NIL)
6916 6910 : return false;
6917 : /* If it has indexes it must be a plain relation */
6918 94835 : rte = root->simple_rte_array[rel->relid];
6919 : Assert(rte->rtekind == RTE_RELATION);
6920 :
6921 : /* ignore partitioned tables. Any indexes here are not real indexes */
6922 94835 : if (rte->relkind == RELKIND_PARTITIONED_TABLE)
6923 378 : return false;
6924 :
6925 : /* Search through the indexes to see if any match our problem */
6926 184508 : foreach(lc, rel->indexlist)
6927 : {
6928 157892 : IndexOptInfo *index = (IndexOptInfo *) lfirst(lc);
6929 : ScanDirection indexscandir;
6930 : StrategyNumber strategy;
6931 :
6932 : /* Ignore non-ordering indexes */
6933 157892 : if (index->sortopfamily == NULL)
6934 0 : continue;
6935 :
6936 : /*
6937 : * Ignore partial indexes --- we only want stats that cover the entire
6938 : * relation.
6939 : */
6940 157892 : if (index->indpred != NIL)
6941 144 : continue;
6942 :
6943 : /*
6944 : * The index list might include hypothetical indexes inserted by a
6945 : * get_relation_info hook --- don't try to access them.
6946 : */
6947 157748 : if (index->hypothetical)
6948 0 : continue;
6949 :
6950 : /*
6951 : * get_actual_variable_endpoint uses the index-only-scan machinery, so
6952 : * ignore indexes that can't use it on their first column.
6953 : */
6954 157748 : if (!index->canreturn[0])
6955 0 : continue;
6956 :
6957 : /*
6958 : * The first index column must match the desired variable, sortop, and
6959 : * collation --- but we can use a descending-order index.
6960 : */
6961 157748 : if (collation != index->indexcollations[0])
6962 20556 : continue; /* test first 'cause it's cheapest */
6963 137192 : if (!match_index_to_operand(vardata->var, 0, index))
6964 69351 : continue;
6965 67841 : strategy = get_op_opfamily_strategy(sortop, index->sortopfamily[0]);
6966 67841 : switch (IndexAmTranslateStrategy(strategy, index->relam, index->sortopfamily[0], true))
6967 : {
6968 67841 : case COMPARE_LT:
6969 67841 : if (index->reverse_sort[0])
6970 0 : indexscandir = BackwardScanDirection;
6971 : else
6972 67841 : indexscandir = ForwardScanDirection;
6973 67841 : break;
6974 0 : case COMPARE_GT:
6975 0 : if (index->reverse_sort[0])
6976 0 : indexscandir = ForwardScanDirection;
6977 : else
6978 0 : indexscandir = BackwardScanDirection;
6979 0 : break;
6980 0 : default:
6981 : /* index doesn't match the sortop */
6982 0 : continue;
6983 : }
6984 :
6985 : /*
6986 : * Found a suitable index to extract data from. Set up some data that
6987 : * can be used by both invocations of get_actual_variable_endpoint.
6988 : */
6989 : {
6990 : MemoryContext tmpcontext;
6991 : MemoryContext oldcontext;
6992 : Relation heapRel;
6993 : Relation indexRel;
6994 : TupleTableSlot *slot;
6995 : int16 typLen;
6996 : bool typByVal;
6997 : ScanKeyData scankeys[1];
6998 :
6999 : /* Make sure any cruft gets recycled when we're done */
7000 67841 : tmpcontext = AllocSetContextCreate(CurrentMemoryContext,
7001 : "get_actual_variable_range workspace",
7002 : ALLOCSET_DEFAULT_SIZES);
7003 67841 : oldcontext = MemoryContextSwitchTo(tmpcontext);
7004 :
7005 : /*
7006 : * Open the table and index so we can read from them. We should
7007 : * already have some type of lock on each.
7008 : */
7009 67841 : heapRel = table_open(rte->relid, NoLock);
7010 67841 : indexRel = index_open(index->indexoid, NoLock);
7011 :
7012 : /* build some stuff needed for indexscan execution */
7013 67841 : slot = table_slot_create(heapRel, NULL);
7014 67841 : get_typlenbyval(vardata->atttype, &typLen, &typByVal);
7015 :
7016 : /* set up an IS NOT NULL scan key so that we ignore nulls */
7017 67841 : ScanKeyEntryInitialize(&scankeys[0],
7018 : SK_ISNULL | SK_SEARCHNOTNULL,
7019 : 1, /* index col to scan */
7020 : InvalidStrategy, /* no strategy */
7021 : InvalidOid, /* no strategy subtype */
7022 : InvalidOid, /* no collation */
7023 : InvalidOid, /* no reg proc for this */
7024 : (Datum) 0); /* constant */
7025 :
7026 : /* If min is requested ... */
7027 67841 : if (min)
7028 : {
7029 38133 : have_data = get_actual_variable_endpoint(heapRel,
7030 : indexRel,
7031 : indexscandir,
7032 : scankeys,
7033 : typLen,
7034 : typByVal,
7035 : slot,
7036 : oldcontext,
7037 : min);
7038 : }
7039 : else
7040 : {
7041 : /* If min not requested, still want to fetch max */
7042 29708 : have_data = true;
7043 : }
7044 :
7045 : /* If max is requested, and we didn't already fail ... */
7046 67841 : if (max && have_data)
7047 : {
7048 : /* scan in the opposite direction; all else is the same */
7049 30774 : have_data = get_actual_variable_endpoint(heapRel,
7050 : indexRel,
7051 30774 : -indexscandir,
7052 : scankeys,
7053 : typLen,
7054 : typByVal,
7055 : slot,
7056 : oldcontext,
7057 : max);
7058 : }
7059 :
7060 : /* Clean everything up */
7061 67841 : ExecDropSingleTupleTableSlot(slot);
7062 :
7063 67841 : index_close(indexRel, NoLock);
7064 67841 : table_close(heapRel, NoLock);
7065 :
7066 67841 : MemoryContextSwitchTo(oldcontext);
7067 67841 : MemoryContextDelete(tmpcontext);
7068 :
7069 : /* And we're done */
7070 67841 : break;
7071 : }
7072 : }
7073 :
7074 94457 : return have_data;
7075 : }
7076 :
7077 : /*
7078 : * Get one endpoint datum (min or max depending on indexscandir) from the
7079 : * specified index. Return true if successful, false if not.
7080 : * On success, endpoint value is stored to *endpointDatum (and copied into
7081 : * outercontext).
7082 : *
7083 : * scankeys is a 1-element scankey array set up to reject nulls.
7084 : * typLen/typByVal describe the datatype of the index's first column.
7085 : * tableslot is a slot suitable to hold table tuples, in case we need
7086 : * to probe the heap.
7087 : * (We could compute these values locally, but that would mean computing them
7088 : * twice when get_actual_variable_range needs both the min and the max.)
7089 : *
7090 : * Failure occurs either when the index is empty, or we decide that it's
7091 : * taking too long to find a suitable tuple.
7092 : */
7093 : static bool
7094 68907 : get_actual_variable_endpoint(Relation heapRel,
7095 : Relation indexRel,
7096 : ScanDirection indexscandir,
7097 : ScanKey scankeys,
7098 : int16 typLen,
7099 : bool typByVal,
7100 : TupleTableSlot *tableslot,
7101 : MemoryContext outercontext,
7102 : Datum *endpointDatum)
7103 : {
7104 68907 : bool have_data = false;
7105 : SnapshotData SnapshotNonVacuumable;
7106 : IndexScanDesc index_scan;
7107 68907 : Buffer vmbuffer = InvalidBuffer;
7108 68907 : BlockNumber last_heap_block = InvalidBlockNumber;
7109 68907 : int n_visited_heap_pages = 0;
7110 : ItemPointer tid;
7111 : Datum values[INDEX_MAX_KEYS];
7112 : bool isnull[INDEX_MAX_KEYS];
7113 : MemoryContext oldcontext;
7114 :
7115 : /*
7116 : * We use the index-only-scan machinery for this. With mostly-static
7117 : * tables that's a win because it avoids a heap visit. It's also a win
7118 : * for dynamic data, but the reason is less obvious; read on for details.
7119 : *
7120 : * In principle, we should scan the index with our current active
7121 : * snapshot, which is the best approximation we've got to what the query
7122 : * will see when executed. But that won't be exact if a new snap is taken
7123 : * before running the query, and it can be very expensive if a lot of
7124 : * recently-dead or uncommitted rows exist at the beginning or end of the
7125 : * index (because we'll laboriously fetch each one and reject it).
7126 : * Instead, we use SnapshotNonVacuumable. That will accept recently-dead
7127 : * and uncommitted rows as well as normal visible rows. On the other
7128 : * hand, it will reject known-dead rows, and thus not give a bogus answer
7129 : * when the extreme value has been deleted (unless the deletion was quite
7130 : * recent); that case motivates not using SnapshotAny here.
7131 : *
7132 : * A crucial point here is that SnapshotNonVacuumable, with
7133 : * GlobalVisTestFor(heapRel) as horizon, yields the inverse of the
7134 : * condition that the indexscan will use to decide that index entries are
7135 : * killable (see heap_hot_search_buffer()). Therefore, if the snapshot
7136 : * rejects a tuple (or more precisely, all tuples of a HOT chain) and we
7137 : * have to continue scanning past it, we know that the indexscan will mark
7138 : * that index entry killed. That means that the next
7139 : * get_actual_variable_endpoint() call will not have to re-consider that
7140 : * index entry. In this way we avoid repetitive work when this function
7141 : * is used a lot during planning.
7142 : *
7143 : * But using SnapshotNonVacuumable creates a hazard of its own. In a
7144 : * recently-created index, some index entries may point at "broken" HOT
7145 : * chains in which not all the tuple versions contain data matching the
7146 : * index entry. The live tuple version(s) certainly do match the index,
7147 : * but SnapshotNonVacuumable can accept recently-dead tuple versions that
7148 : * don't match. Hence, if we took data from the selected heap tuple, we
7149 : * might get a bogus answer that's not close to the index extremal value,
7150 : * or could even be NULL. We avoid this hazard because we take the data
7151 : * from the index entry not the heap.
7152 : *
7153 : * Despite all this care, there are situations where we might find many
7154 : * non-visible tuples near the end of the index. We don't want to expend
7155 : * a huge amount of time here, so we give up once we've read too many heap
7156 : * pages. When we fail for that reason, the caller will end up using
7157 : * whatever extremal value is recorded in pg_statistic.
7158 : */
7159 68907 : InitNonVacuumableSnapshot(SnapshotNonVacuumable,
7160 : GlobalVisTestFor(heapRel));
7161 :
7162 68907 : index_scan = index_beginscan(heapRel, indexRel,
7163 : &SnapshotNonVacuumable, NULL,
7164 : 1, 0);
7165 : /* Set it up for index-only scan */
7166 68907 : index_scan->xs_want_itup = true;
7167 68907 : index_rescan(index_scan, scankeys, 1, NULL, 0);
7168 :
7169 : /* Fetch first/next tuple in specified direction */
7170 86030 : while ((tid = index_getnext_tid(index_scan, indexscandir)) != NULL)
7171 : {
7172 86030 : BlockNumber block = ItemPointerGetBlockNumber(tid);
7173 :
7174 86030 : if (!VM_ALL_VISIBLE(heapRel,
7175 : block,
7176 : &vmbuffer))
7177 : {
7178 : /* Rats, we have to visit the heap to check visibility */
7179 59421 : if (!index_fetch_heap(index_scan, tableslot))
7180 : {
7181 : /*
7182 : * No visible tuple for this index entry, so we need to
7183 : * advance to the next entry. Before doing so, count heap
7184 : * page fetches and give up if we've done too many.
7185 : *
7186 : * We don't charge a page fetch if this is the same heap page
7187 : * as the previous tuple. This is on the conservative side,
7188 : * since other recently-accessed pages are probably still in
7189 : * buffers too; but it's good enough for this heuristic.
7190 : */
7191 : #define VISITED_PAGES_LIMIT 100
7192 :
7193 17123 : if (block != last_heap_block)
7194 : {
7195 1478 : last_heap_block = block;
7196 1478 : n_visited_heap_pages++;
7197 1478 : if (n_visited_heap_pages > VISITED_PAGES_LIMIT)
7198 0 : break;
7199 : }
7200 :
7201 17123 : continue; /* no visible tuple, try next index entry */
7202 : }
7203 :
7204 : /* We don't actually need the heap tuple for anything */
7205 42298 : ExecClearTuple(tableslot);
7206 :
7207 : /*
7208 : * We don't care whether there's more than one visible tuple in
7209 : * the HOT chain; if any are visible, that's good enough.
7210 : */
7211 : }
7212 :
7213 : /*
7214 : * We expect that the index will return data in IndexTuple not
7215 : * HeapTuple format.
7216 : */
7217 68907 : if (!index_scan->xs_itup)
7218 0 : elog(ERROR, "no data returned for index-only scan");
7219 :
7220 : /*
7221 : * We do not yet support recheck here.
7222 : */
7223 68907 : if (index_scan->xs_recheck)
7224 0 : break;
7225 :
7226 : /* OK to deconstruct the index tuple */
7227 68907 : index_deform_tuple(index_scan->xs_itup,
7228 : index_scan->xs_itupdesc,
7229 : values, isnull);
7230 :
7231 : /* Shouldn't have got a null, but be careful */
7232 68907 : if (isnull[0])
7233 0 : elog(ERROR, "found unexpected null value in index \"%s\"",
7234 : RelationGetRelationName(indexRel));
7235 :
7236 : /* Copy the index column value out to caller's context */
7237 68907 : oldcontext = MemoryContextSwitchTo(outercontext);
7238 68907 : *endpointDatum = datumCopy(values[0], typByVal, typLen);
7239 68907 : MemoryContextSwitchTo(oldcontext);
7240 68907 : have_data = true;
7241 68907 : break;
7242 : }
7243 :
7244 68907 : if (vmbuffer != InvalidBuffer)
7245 62528 : ReleaseBuffer(vmbuffer);
7246 68907 : index_endscan(index_scan);
7247 :
7248 68907 : return have_data;
7249 : }
7250 :
7251 : /*
7252 : * find_join_input_rel
7253 : * Look up the input relation for a join.
7254 : *
7255 : * We assume that the input relation's RelOptInfo must have been constructed
7256 : * already.
7257 : */
7258 : static RelOptInfo *
7259 7250 : find_join_input_rel(PlannerInfo *root, Relids relids)
7260 : {
7261 7250 : RelOptInfo *rel = NULL;
7262 :
7263 7250 : if (!bms_is_empty(relids))
7264 : {
7265 : int relid;
7266 :
7267 7250 : if (bms_get_singleton_member(relids, &relid))
7268 7089 : rel = find_base_rel(root, relid);
7269 : else
7270 161 : rel = find_join_rel(root, relids);
7271 : }
7272 :
7273 7250 : if (rel == NULL)
7274 0 : elog(ERROR, "could not find RelOptInfo for given relids");
7275 :
7276 7250 : return rel;
7277 : }
7278 :
7279 :
7280 : /*-------------------------------------------------------------------------
7281 : *
7282 : * Index cost estimation functions
7283 : *
7284 : *-------------------------------------------------------------------------
7285 : */
7286 :
7287 : /*
7288 : * Extract the actual indexquals (as RestrictInfos) from an IndexClause list
7289 : */
7290 : List *
7291 480599 : get_quals_from_indexclauses(List *indexclauses)
7292 : {
7293 480599 : List *result = NIL;
7294 : ListCell *lc;
7295 :
7296 842452 : foreach(lc, indexclauses)
7297 : {
7298 361853 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7299 : ListCell *lc2;
7300 :
7301 725171 : foreach(lc2, iclause->indexquals)
7302 : {
7303 363318 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7304 :
7305 363318 : result = lappend(result, rinfo);
7306 : }
7307 : }
7308 480599 : return result;
7309 : }
7310 :
7311 : /*
7312 : * Compute the total evaluation cost of the comparison operands in a list
7313 : * of index qual expressions. Since we know these will be evaluated just
7314 : * once per scan, there's no need to distinguish startup from per-row cost.
7315 : *
7316 : * This can be used either on the result of get_quals_from_indexclauses(),
7317 : * or directly on an indexorderbys list. In both cases, we expect that the
7318 : * index key expression is on the left side of binary clauses.
7319 : */
7320 : Cost
7321 954695 : index_other_operands_eval_cost(PlannerInfo *root, List *indexquals)
7322 : {
7323 954695 : Cost qual_arg_cost = 0;
7324 : ListCell *lc;
7325 :
7326 1318244 : foreach(lc, indexquals)
7327 : {
7328 363549 : Expr *clause = (Expr *) lfirst(lc);
7329 : Node *other_operand;
7330 : QualCost index_qual_cost;
7331 :
7332 : /*
7333 : * Index quals will have RestrictInfos, indexorderbys won't. Look
7334 : * through RestrictInfo if present.
7335 : */
7336 363549 : if (IsA(clause, RestrictInfo))
7337 363312 : clause = ((RestrictInfo *) clause)->clause;
7338 :
7339 363549 : if (IsA(clause, OpExpr))
7340 : {
7341 353539 : OpExpr *op = (OpExpr *) clause;
7342 :
7343 353539 : other_operand = (Node *) lsecond(op->args);
7344 : }
7345 10010 : else if (IsA(clause, RowCompareExpr))
7346 : {
7347 198 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7348 :
7349 198 : other_operand = (Node *) rc->rargs;
7350 : }
7351 9812 : else if (IsA(clause, ScalarArrayOpExpr))
7352 : {
7353 8349 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7354 :
7355 8349 : other_operand = (Node *) lsecond(saop->args);
7356 : }
7357 1463 : else if (IsA(clause, NullTest))
7358 : {
7359 1463 : other_operand = NULL;
7360 : }
7361 : else
7362 : {
7363 0 : elog(ERROR, "unsupported indexqual type: %d",
7364 : (int) nodeTag(clause));
7365 : other_operand = NULL; /* keep compiler quiet */
7366 : }
7367 :
7368 363549 : cost_qual_eval_node(&index_qual_cost, other_operand, root);
7369 363549 : qual_arg_cost += index_qual_cost.startup + index_qual_cost.per_tuple;
7370 : }
7371 954695 : return qual_arg_cost;
7372 : }
7373 :
7374 : void
7375 474102 : genericcostestimate(PlannerInfo *root,
7376 : IndexPath *path,
7377 : double loop_count,
7378 : GenericCosts *costs)
7379 : {
7380 474102 : IndexOptInfo *index = path->indexinfo;
7381 474102 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
7382 474102 : List *indexOrderBys = path->indexorderbys;
7383 : Cost indexStartupCost;
7384 : Cost indexTotalCost;
7385 : Selectivity indexSelectivity;
7386 : double indexCorrelation;
7387 : double numIndexPages;
7388 : double numIndexTuples;
7389 : double spc_random_page_cost;
7390 : double num_sa_scans;
7391 : double num_outer_scans;
7392 : double num_scans;
7393 : double qual_op_cost;
7394 : double qual_arg_cost;
7395 : List *selectivityQuals;
7396 : ListCell *l;
7397 :
7398 : /*
7399 : * If the index is partial, AND the index predicate with the explicitly
7400 : * given indexquals to produce a more accurate idea of the index
7401 : * selectivity.
7402 : */
7403 474102 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
7404 :
7405 : /*
7406 : * If caller didn't give us an estimate for ScalarArrayOpExpr index scans,
7407 : * just assume that the number of index descents is the number of distinct
7408 : * combinations of array elements from all of the scan's SAOP clauses.
7409 : */
7410 474102 : num_sa_scans = costs->num_sa_scans;
7411 474102 : if (num_sa_scans < 1)
7412 : {
7413 3960 : num_sa_scans = 1;
7414 8312 : foreach(l, indexQuals)
7415 : {
7416 4352 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
7417 :
7418 4352 : if (IsA(rinfo->clause, ScalarArrayOpExpr))
7419 : {
7420 13 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) rinfo->clause;
7421 13 : double alength = estimate_array_length(root, lsecond(saop->args));
7422 :
7423 13 : if (alength > 1)
7424 13 : num_sa_scans *= alength;
7425 : }
7426 : }
7427 : }
7428 :
7429 : /* Estimate the fraction of main-table tuples that will be visited */
7430 474102 : indexSelectivity = clauselist_selectivity(root, selectivityQuals,
7431 474102 : index->rel->relid,
7432 : JOIN_INNER,
7433 : NULL);
7434 :
7435 : /*
7436 : * If caller didn't give us an estimate, estimate the number of index
7437 : * tuples that will be visited. We do it in this rather peculiar-looking
7438 : * way in order to get the right answer for partial indexes.
7439 : */
7440 474102 : numIndexTuples = costs->numIndexTuples;
7441 474102 : if (numIndexTuples <= 0.0)
7442 : {
7443 61460 : numIndexTuples = indexSelectivity * index->rel->tuples;
7444 :
7445 : /*
7446 : * The above calculation counts all the tuples visited across all
7447 : * scans induced by ScalarArrayOpExpr nodes. We want to consider the
7448 : * average per-indexscan number, so adjust. This is a handy place to
7449 : * round to integer, too. (If caller supplied tuple estimate, it's
7450 : * responsible for handling these considerations.)
7451 : */
7452 61460 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
7453 : }
7454 :
7455 : /*
7456 : * We can bound the number of tuples by the index size in any case. Also,
7457 : * always estimate at least one tuple is touched, even when
7458 : * indexSelectivity estimate is tiny.
7459 : */
7460 474102 : if (numIndexTuples > index->tuples)
7461 5653 : numIndexTuples = index->tuples;
7462 474102 : if (numIndexTuples < 1.0)
7463 64186 : numIndexTuples = 1.0;
7464 :
7465 : /*
7466 : * Estimate the number of index pages that will be retrieved.
7467 : *
7468 : * We use the simplistic method of taking a pro-rata fraction of the total
7469 : * number of index pages. In effect, this counts only leaf pages and not
7470 : * any overhead such as index metapage or upper tree levels.
7471 : *
7472 : * In practice access to upper index levels is often nearly free because
7473 : * those tend to stay in cache under load; moreover, the cost involved is
7474 : * highly dependent on index type. We therefore ignore such costs here
7475 : * and leave it to the caller to add a suitable charge if needed.
7476 : */
7477 474102 : if (index->pages > 1 && index->tuples > 1)
7478 425466 : numIndexPages = ceil(numIndexTuples * index->pages / index->tuples);
7479 : else
7480 48636 : numIndexPages = 1.0;
7481 :
7482 : /* fetch estimated page cost for tablespace containing index */
7483 474102 : get_tablespace_page_costs(index->reltablespace,
7484 : &spc_random_page_cost,
7485 : NULL);
7486 :
7487 : /*
7488 : * Now compute the disk access costs.
7489 : *
7490 : * The above calculations are all per-index-scan. However, if we are in a
7491 : * nestloop inner scan, we can expect the scan to be repeated (with
7492 : * different search keys) for each row of the outer relation. Likewise,
7493 : * ScalarArrayOpExpr quals result in multiple index scans. This creates
7494 : * the potential for cache effects to reduce the number of disk page
7495 : * fetches needed. We want to estimate the average per-scan I/O cost in
7496 : * the presence of caching.
7497 : *
7498 : * We use the Mackert-Lohman formula (see costsize.c for details) to
7499 : * estimate the total number of page fetches that occur. While this
7500 : * wasn't what it was designed for, it seems a reasonable model anyway.
7501 : * Note that we are counting pages not tuples anymore, so we take N = T =
7502 : * index size, as if there were one "tuple" per page.
7503 : */
7504 474102 : num_outer_scans = loop_count;
7505 474102 : num_scans = num_sa_scans * num_outer_scans;
7506 :
7507 474102 : if (num_scans > 1)
7508 : {
7509 : double pages_fetched;
7510 :
7511 : /* total page fetches ignoring cache effects */
7512 57612 : pages_fetched = numIndexPages * num_scans;
7513 :
7514 : /* use Mackert and Lohman formula to adjust for cache effects */
7515 57612 : pages_fetched = index_pages_fetched(pages_fetched,
7516 : index->pages,
7517 57612 : (double) index->pages,
7518 : root);
7519 :
7520 : /*
7521 : * Now compute the total disk access cost, and then report a pro-rated
7522 : * share for each outer scan. (Don't pro-rate for ScalarArrayOpExpr,
7523 : * since that's internal to the indexscan.)
7524 : */
7525 57612 : indexTotalCost = (pages_fetched * spc_random_page_cost)
7526 : / num_outer_scans;
7527 : }
7528 : else
7529 : {
7530 : /*
7531 : * For a single index scan, we just charge spc_random_page_cost per
7532 : * page touched.
7533 : */
7534 416490 : indexTotalCost = numIndexPages * spc_random_page_cost;
7535 : }
7536 :
7537 : /*
7538 : * CPU cost: any complex expressions in the indexquals will need to be
7539 : * evaluated once at the start of the scan to reduce them to runtime keys
7540 : * to pass to the index AM (see nodeIndexscan.c). We model the per-tuple
7541 : * CPU costs as cpu_index_tuple_cost plus one cpu_operator_cost per
7542 : * indexqual operator. Because we have numIndexTuples as a per-scan
7543 : * number, we have to multiply by num_sa_scans to get the correct result
7544 : * for ScalarArrayOpExpr cases. Similarly add in costs for any index
7545 : * ORDER BY expressions.
7546 : *
7547 : * Note: this neglects the possible costs of rechecking lossy operators.
7548 : * Detecting that that might be needed seems more expensive than it's
7549 : * worth, though, considering all the other inaccuracies here ...
7550 : */
7551 474102 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals) +
7552 474102 : index_other_operands_eval_cost(root, indexOrderBys);
7553 474102 : qual_op_cost = cpu_operator_cost *
7554 474102 : (list_length(indexQuals) + list_length(indexOrderBys));
7555 :
7556 474102 : indexStartupCost = qual_arg_cost;
7557 474102 : indexTotalCost += qual_arg_cost;
7558 474102 : indexTotalCost += numIndexTuples * num_sa_scans * (cpu_index_tuple_cost + qual_op_cost);
7559 :
7560 : /*
7561 : * Generic assumption about index correlation: there isn't any.
7562 : */
7563 474102 : indexCorrelation = 0.0;
7564 :
7565 : /*
7566 : * Return everything to caller.
7567 : */
7568 474102 : costs->indexStartupCost = indexStartupCost;
7569 474102 : costs->indexTotalCost = indexTotalCost;
7570 474102 : costs->indexSelectivity = indexSelectivity;
7571 474102 : costs->indexCorrelation = indexCorrelation;
7572 474102 : costs->numIndexPages = numIndexPages;
7573 474102 : costs->numIndexTuples = numIndexTuples;
7574 474102 : costs->spc_random_page_cost = spc_random_page_cost;
7575 474102 : costs->num_sa_scans = num_sa_scans;
7576 474102 : }
7577 :
7578 : /*
7579 : * If the index is partial, add its predicate to the given qual list.
7580 : *
7581 : * ANDing the index predicate with the explicitly given indexquals produces
7582 : * a more accurate idea of the index's selectivity. However, we need to be
7583 : * careful not to insert redundant clauses, because clauselist_selectivity()
7584 : * is easily fooled into computing a too-low selectivity estimate. Our
7585 : * approach is to add only the predicate clause(s) that cannot be proven to
7586 : * be implied by the given indexquals. This successfully handles cases such
7587 : * as a qual "x = 42" used with a partial index "WHERE x >= 40 AND x < 50".
7588 : * There are many other cases where we won't detect redundancy, leading to a
7589 : * too-low selectivity estimate, which will bias the system in favor of using
7590 : * partial indexes where possible. That is not necessarily bad though.
7591 : *
7592 : * Note that indexQuals contains RestrictInfo nodes while the indpred
7593 : * does not, so the output list will be mixed. This is OK for both
7594 : * predicate_implied_by() and clauselist_selectivity(), but might be
7595 : * problematic if the result were passed to other things.
7596 : */
7597 : List *
7598 807549 : add_predicate_to_index_quals(IndexOptInfo *index, List *indexQuals)
7599 : {
7600 807549 : List *predExtraQuals = NIL;
7601 : ListCell *lc;
7602 :
7603 807549 : if (index->indpred == NIL)
7604 806540 : return indexQuals;
7605 :
7606 2024 : foreach(lc, index->indpred)
7607 : {
7608 1015 : Node *predQual = (Node *) lfirst(lc);
7609 1015 : List *oneQual = list_make1(predQual);
7610 :
7611 1015 : if (!predicate_implied_by(oneQual, indexQuals, false))
7612 906 : predExtraQuals = list_concat(predExtraQuals, oneQual);
7613 : }
7614 1009 : return list_concat(predExtraQuals, indexQuals);
7615 : }
7616 :
7617 : /*
7618 : * Estimate correlation of btree index's first column.
7619 : *
7620 : * If we can get an estimate of the first column's ordering correlation C
7621 : * from pg_statistic, estimate the index correlation as C for a single-column
7622 : * index, or C * 0.75 for multiple columns. The idea here is that multiple
7623 : * columns dilute the importance of the first column's ordering, but don't
7624 : * negate it entirely.
7625 : *
7626 : * We already filled in the stats tuple for *vardata when called.
7627 : */
7628 : static double
7629 317899 : btcost_correlation(IndexOptInfo *index, VariableStatData *vardata)
7630 : {
7631 : Oid sortop;
7632 : AttStatsSlot sslot;
7633 317899 : double indexCorrelation = 0;
7634 :
7635 : Assert(HeapTupleIsValid(vardata->statsTuple));
7636 :
7637 317899 : sortop = get_opfamily_member(index->opfamily[0],
7638 317899 : index->opcintype[0],
7639 317899 : index->opcintype[0],
7640 : BTLessStrategyNumber);
7641 635798 : if (OidIsValid(sortop) &&
7642 317899 : get_attstatsslot(&sslot, vardata->statsTuple,
7643 : STATISTIC_KIND_CORRELATION, sortop,
7644 : ATTSTATSSLOT_NUMBERS))
7645 : {
7646 : double varCorrelation;
7647 :
7648 : Assert(sslot.nnumbers == 1);
7649 312862 : varCorrelation = sslot.numbers[0];
7650 :
7651 312862 : if (index->reverse_sort[0])
7652 0 : varCorrelation = -varCorrelation;
7653 :
7654 312862 : if (index->nkeycolumns > 1)
7655 108702 : indexCorrelation = varCorrelation * 0.75;
7656 : else
7657 204160 : indexCorrelation = varCorrelation;
7658 :
7659 312862 : free_attstatsslot(&sslot);
7660 : }
7661 :
7662 317899 : return indexCorrelation;
7663 : }
7664 :
7665 : void
7666 470142 : btcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
7667 : Cost *indexStartupCost, Cost *indexTotalCost,
7668 : Selectivity *indexSelectivity, double *indexCorrelation,
7669 : double *indexPages)
7670 : {
7671 470142 : IndexOptInfo *index = path->indexinfo;
7672 470142 : GenericCosts costs = {0};
7673 470142 : VariableStatData vardata = {0};
7674 : double numIndexTuples;
7675 : Cost descentCost;
7676 : List *indexBoundQuals;
7677 : List *indexSkipQuals;
7678 : int indexcol;
7679 : bool eqQualHere;
7680 : bool found_row_compare;
7681 : bool found_array;
7682 : bool found_is_null_op;
7683 470142 : bool have_correlation = false;
7684 : double num_sa_scans;
7685 470142 : double correlation = 0.0;
7686 : ListCell *lc;
7687 :
7688 : /*
7689 : * For a btree scan, only leading '=' quals plus inequality quals for the
7690 : * immediately next attribute contribute to index selectivity (these are
7691 : * the "boundary quals" that determine the starting and stopping points of
7692 : * the index scan). Additional quals can suppress visits to the heap, so
7693 : * it's OK to count them in indexSelectivity, but they should not count
7694 : * for estimating numIndexTuples. So we must examine the given indexquals
7695 : * to find out which ones count as boundary quals. We rely on the
7696 : * knowledge that they are given in index column order. Note that nbtree
7697 : * preprocessing can add skip arrays that act as leading '=' quals in the
7698 : * absence of ordinary input '=' quals, so in practice _most_ input quals
7699 : * are able to act as index bound quals (which we take into account here).
7700 : *
7701 : * For a RowCompareExpr, we consider only the first column, just as
7702 : * rowcomparesel() does.
7703 : *
7704 : * If there's a SAOP or skip array in the quals, we'll actually perform up
7705 : * to N index descents (not just one), but the underlying array key's
7706 : * operator can be considered to act the same as it normally does.
7707 : */
7708 470142 : indexBoundQuals = NIL;
7709 470142 : indexSkipQuals = NIL;
7710 470142 : indexcol = 0;
7711 470142 : eqQualHere = false;
7712 470142 : found_row_compare = false;
7713 470142 : found_array = false;
7714 470142 : found_is_null_op = false;
7715 470142 : num_sa_scans = 1;
7716 796772 : foreach(lc, path->indexclauses)
7717 : {
7718 348798 : IndexClause *iclause = lfirst_node(IndexClause, lc);
7719 : ListCell *lc2;
7720 :
7721 348798 : if (indexcol < iclause->indexcol)
7722 : {
7723 71246 : double num_sa_scans_prev_cols = num_sa_scans;
7724 :
7725 : /*
7726 : * Beginning of a new column's quals.
7727 : *
7728 : * Skip scans use skip arrays, which are ScalarArrayOp style
7729 : * arrays that generate their elements procedurally and on demand.
7730 : * Given a multi-column index on "(a, b)", and an SQL WHERE clause
7731 : * "WHERE b = 42", a skip scan will effectively use an indexqual
7732 : * "WHERE a = ANY('{every col a value}') AND b = 42". (Obviously,
7733 : * the array on "a" must also return "IS NULL" matches, since our
7734 : * WHERE clause used no strict operator on "a").
7735 : *
7736 : * Here we consider how nbtree will backfill skip arrays for any
7737 : * index columns that lacked an '=' qual. This maintains our
7738 : * num_sa_scans estimate, and determines if this new column (the
7739 : * "iclause->indexcol" column, not the prior "indexcol" column)
7740 : * can have its RestrictInfos/quals added to indexBoundQuals.
7741 : *
7742 : * We'll need to handle columns that have inequality quals, where
7743 : * the skip array generates values from a range constrained by the
7744 : * quals (not every possible value). We've been maintaining
7745 : * indexSkipQuals to help with this; it will now contain all of
7746 : * the prior column's quals (that is, indexcol's quals) when they
7747 : * might be used for this.
7748 : */
7749 71246 : if (found_row_compare)
7750 : {
7751 : /*
7752 : * Skip arrays can't be added after a RowCompare input qual
7753 : * due to limitations in nbtree
7754 : */
7755 12 : break;
7756 : }
7757 71234 : if (eqQualHere)
7758 : {
7759 : /*
7760 : * Don't need to add a skip array for an indexcol that already
7761 : * has an '=' qual/equality constraint
7762 : */
7763 49470 : indexcol++;
7764 49470 : indexSkipQuals = NIL;
7765 : }
7766 71234 : eqQualHere = false;
7767 :
7768 72871 : while (indexcol < iclause->indexcol)
7769 : {
7770 : double ndistinct;
7771 23793 : bool isdefault = true;
7772 :
7773 23793 : found_array = true;
7774 :
7775 : /*
7776 : * A skipped attribute's ndistinct forms the basis of our
7777 : * estimate of the total number of "array elements" used by
7778 : * its skip array at runtime. Look that up first.
7779 : */
7780 23793 : examine_indexcol_variable(root, index, indexcol, &vardata);
7781 23793 : ndistinct = get_variable_numdistinct(&vardata, &isdefault);
7782 :
7783 23793 : if (indexcol == 0)
7784 : {
7785 : /*
7786 : * Get an estimate of the leading column's correlation in
7787 : * passing (avoids rereading variable stats below)
7788 : */
7789 21758 : if (HeapTupleIsValid(vardata.statsTuple))
7790 12812 : correlation = btcost_correlation(index, &vardata);
7791 21758 : have_correlation = true;
7792 : }
7793 :
7794 23793 : ReleaseVariableStats(vardata);
7795 :
7796 : /*
7797 : * If ndistinct is a default estimate, conservatively assume
7798 : * that no skipping will happen at runtime
7799 : */
7800 23793 : if (isdefault)
7801 : {
7802 7380 : num_sa_scans = num_sa_scans_prev_cols;
7803 22156 : break; /* done building indexBoundQuals */
7804 : }
7805 :
7806 : /*
7807 : * Apply indexcol's indexSkipQuals selectivity to ndistinct
7808 : */
7809 16413 : if (indexSkipQuals != NIL)
7810 : {
7811 : List *partialSkipQuals;
7812 : Selectivity ndistinctfrac;
7813 :
7814 : /*
7815 : * If the index is partial, AND the index predicate with
7816 : * the index-bound quals to produce a more accurate idea
7817 : * of the number of distinct values for prior indexcol
7818 : */
7819 332 : partialSkipQuals = add_predicate_to_index_quals(index,
7820 : indexSkipQuals);
7821 :
7822 332 : ndistinctfrac = clauselist_selectivity(root, partialSkipQuals,
7823 332 : index->rel->relid,
7824 : JOIN_INNER,
7825 : NULL);
7826 :
7827 : /*
7828 : * If ndistinctfrac is selective (on its own), the scan is
7829 : * unlikely to benefit from repositioning itself using
7830 : * later quals. Do not allow iclause->indexcol's quals to
7831 : * be added to indexBoundQuals (it would increase descent
7832 : * costs, without lowering numIndexTuples costs by much).
7833 : */
7834 332 : if (ndistinctfrac < DEFAULT_RANGE_INEQ_SEL)
7835 : {
7836 187 : num_sa_scans = num_sa_scans_prev_cols;
7837 187 : break; /* done building indexBoundQuals */
7838 : }
7839 :
7840 : /* Adjust ndistinct downward */
7841 145 : ndistinct = rint(ndistinct * ndistinctfrac);
7842 145 : ndistinct = Max(ndistinct, 1);
7843 : }
7844 :
7845 : /*
7846 : * When there's no inequality quals, account for the need to
7847 : * find an initial value by counting -inf/+inf as a value.
7848 : *
7849 : * We don't charge anything extra for possible next/prior key
7850 : * index probes, which are sometimes used to find the next
7851 : * valid skip array element (ahead of using the located
7852 : * element value to relocate the scan to the next position
7853 : * that might contain matching tuples). It seems hard to do
7854 : * better here. Use of the skip support infrastructure often
7855 : * avoids most next/prior key probes. But even when it can't,
7856 : * there's a decent chance that most individual next/prior key
7857 : * probes will locate a leaf page whose key space overlaps all
7858 : * of the scan's keys (even the lower-order keys) -- which
7859 : * also avoids the need for a separate, extra index descent.
7860 : * Note also that these probes are much cheaper than non-probe
7861 : * primitive index scans: they're reliably very selective.
7862 : */
7863 16226 : if (indexSkipQuals == NIL)
7864 16081 : ndistinct += 1;
7865 :
7866 : /*
7867 : * Update num_sa_scans estimate by multiplying by ndistinct.
7868 : *
7869 : * We make the pessimistic assumption that there is no
7870 : * naturally occurring cross-column correlation. This is
7871 : * often wrong, but it seems best to err on the side of not
7872 : * expecting skipping to be helpful...
7873 : */
7874 16226 : num_sa_scans *= ndistinct;
7875 :
7876 : /*
7877 : * ...but back out of adding this latest group of 1 or more
7878 : * skip arrays when num_sa_scans exceeds the total number of
7879 : * index pages (revert to num_sa_scans from before indexcol).
7880 : * This causes a sharp discontinuity in cost (as a function of
7881 : * the indexcol's ndistinct), but that is representative of
7882 : * actual runtime costs.
7883 : *
7884 : * Note that skipping is helpful when each primitive index
7885 : * scan only manages to skip over 1 or 2 irrelevant leaf pages
7886 : * on average. Skip arrays bring savings in CPU costs due to
7887 : * the scan not needing to evaluate indexquals against every
7888 : * tuple, which can greatly exceed any savings in I/O costs.
7889 : * This test is a test of whether num_sa_scans implies that
7890 : * we're past the point where the ability to skip ceases to
7891 : * lower the scan's costs (even qual evaluation CPU costs).
7892 : */
7893 16226 : if (index->pages < num_sa_scans)
7894 : {
7895 14589 : num_sa_scans = num_sa_scans_prev_cols;
7896 14589 : break; /* done building indexBoundQuals */
7897 : }
7898 :
7899 1637 : indexcol++;
7900 1637 : indexSkipQuals = NIL;
7901 : }
7902 :
7903 : /*
7904 : * Finished considering the need to add skip arrays to bridge an
7905 : * initial eqQualHere gap between the old and new index columns
7906 : * (or there was no initial eqQualHere gap in the first place).
7907 : *
7908 : * If an initial gap could not be bridged, then new column's quals
7909 : * (i.e. iclause->indexcol's quals) won't go into indexBoundQuals,
7910 : * and so won't affect our final numIndexTuples estimate.
7911 : */
7912 71234 : if (indexcol != iclause->indexcol)
7913 22156 : break; /* done building indexBoundQuals */
7914 : }
7915 :
7916 : Assert(indexcol == iclause->indexcol);
7917 :
7918 : /* Examine each indexqual associated with this index clause */
7919 654635 : foreach(lc2, iclause->indexquals)
7920 : {
7921 328005 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
7922 328005 : Expr *clause = rinfo->clause;
7923 328005 : Oid clause_op = InvalidOid;
7924 : int op_strategy;
7925 :
7926 328005 : if (IsA(clause, OpExpr))
7927 : {
7928 318530 : OpExpr *op = (OpExpr *) clause;
7929 :
7930 318530 : clause_op = op->opno;
7931 : }
7932 9475 : else if (IsA(clause, RowCompareExpr))
7933 : {
7934 198 : RowCompareExpr *rc = (RowCompareExpr *) clause;
7935 :
7936 198 : clause_op = linitial_oid(rc->opnos);
7937 198 : found_row_compare = true;
7938 : }
7939 9277 : else if (IsA(clause, ScalarArrayOpExpr))
7940 : {
7941 8135 : ScalarArrayOpExpr *saop = (ScalarArrayOpExpr *) clause;
7942 8135 : Node *other_operand = (Node *) lsecond(saop->args);
7943 8135 : double alength = estimate_array_length(root, other_operand);
7944 :
7945 8135 : clause_op = saop->opno;
7946 8135 : found_array = true;
7947 : /* estimate SA descents by indexBoundQuals only */
7948 8135 : if (alength > 1)
7949 7981 : num_sa_scans *= alength;
7950 : }
7951 1142 : else if (IsA(clause, NullTest))
7952 : {
7953 1142 : NullTest *nt = (NullTest *) clause;
7954 :
7955 1142 : if (nt->nulltesttype == IS_NULL)
7956 : {
7957 120 : found_is_null_op = true;
7958 : /* IS NULL is like = for selectivity/skip scan purposes */
7959 120 : eqQualHere = true;
7960 : }
7961 : }
7962 : else
7963 0 : elog(ERROR, "unsupported indexqual type: %d",
7964 : (int) nodeTag(clause));
7965 :
7966 : /* check for equality operator */
7967 328005 : if (OidIsValid(clause_op))
7968 : {
7969 326863 : op_strategy = get_op_opfamily_strategy(clause_op,
7970 326863 : index->opfamily[indexcol]);
7971 : Assert(op_strategy != 0); /* not a member of opfamily?? */
7972 326863 : if (op_strategy == BTEqualStrategyNumber)
7973 310290 : eqQualHere = true;
7974 : }
7975 :
7976 328005 : indexBoundQuals = lappend(indexBoundQuals, rinfo);
7977 :
7978 : /*
7979 : * We apply inequality selectivities to estimate index descent
7980 : * costs with scans that use skip arrays. Save this indexcol's
7981 : * RestrictInfos if it looks like they'll be needed for that.
7982 : */
7983 328005 : if (!eqQualHere && !found_row_compare &&
7984 17046 : indexcol < index->nkeycolumns - 1)
7985 2856 : indexSkipQuals = lappend(indexSkipQuals, rinfo);
7986 : }
7987 : }
7988 :
7989 : /*
7990 : * If index is unique and we found an '=' clause for each column, we can
7991 : * just assume numIndexTuples = 1 and skip the expensive
7992 : * clauselist_selectivity calculations. However, an array or NullTest
7993 : * always invalidates that theory (even when eqQualHere has been set).
7994 : */
7995 470142 : if (index->unique &&
7996 381469 : indexcol == index->nkeycolumns - 1 &&
7997 142022 : eqQualHere &&
7998 142022 : !found_array &&
7999 138183 : !found_is_null_op)
8000 138159 : numIndexTuples = 1.0;
8001 : else
8002 : {
8003 : List *selectivityQuals;
8004 : Selectivity btreeSelectivity;
8005 :
8006 : /*
8007 : * If the index is partial, AND the index predicate with the
8008 : * index-bound quals to produce a more accurate idea of the number of
8009 : * rows covered by the bound conditions.
8010 : */
8011 331983 : selectivityQuals = add_predicate_to_index_quals(index, indexBoundQuals);
8012 :
8013 331983 : btreeSelectivity = clauselist_selectivity(root, selectivityQuals,
8014 331983 : index->rel->relid,
8015 : JOIN_INNER,
8016 : NULL);
8017 331983 : numIndexTuples = btreeSelectivity * index->rel->tuples;
8018 :
8019 : /*
8020 : * btree automatically combines individual array element primitive
8021 : * index scans whenever the tuples covered by the next set of array
8022 : * keys are close to tuples covered by the current set. That puts a
8023 : * natural ceiling on the worst case number of descents -- there
8024 : * cannot possibly be more than one descent per leaf page scanned.
8025 : *
8026 : * Clamp the number of descents to at most 1/3 the number of index
8027 : * pages. This avoids implausibly high estimates with low selectivity
8028 : * paths, where scans usually require only one or two descents. This
8029 : * is most likely to help when there are several SAOP clauses, where
8030 : * naively accepting the total number of distinct combinations of
8031 : * array elements as the number of descents would frequently lead to
8032 : * wild overestimates.
8033 : *
8034 : * We somewhat arbitrarily don't just make the cutoff the total number
8035 : * of leaf pages (we make it 1/3 the total number of pages instead) to
8036 : * give the btree code credit for its ability to continue on the leaf
8037 : * level with low selectivity scans.
8038 : *
8039 : * Note: num_sa_scans includes both ScalarArrayOp array elements and
8040 : * skip array elements whose qual affects our numIndexTuples estimate.
8041 : */
8042 331983 : num_sa_scans = Min(num_sa_scans, ceil(index->pages * 0.3333333));
8043 331983 : num_sa_scans = Max(num_sa_scans, 1);
8044 :
8045 : /*
8046 : * As in genericcostestimate(), we have to adjust for any array quals
8047 : * included in indexBoundQuals, and then round to integer.
8048 : *
8049 : * It is tempting to make genericcostestimate behave as if array
8050 : * clauses work in almost the same way as scalar operators during
8051 : * btree scans, making the top-level scan look like a continuous scan
8052 : * (as opposed to num_sa_scans-many primitive index scans). After
8053 : * all, btree scans mostly work like that at runtime. However, such a
8054 : * scheme would badly bias genericcostestimate's simplistic approach
8055 : * to calculating numIndexPages through prorating.
8056 : *
8057 : * Stick with the approach taken by non-native SAOP scans for now.
8058 : * genericcostestimate will use the Mackert-Lohman formula to
8059 : * compensate for repeat page fetches, even though that definitely
8060 : * won't happen during btree scans (not for leaf pages, at least).
8061 : * We're usually very pessimistic about the number of primitive index
8062 : * scans that will be required, but it's not clear how to do better.
8063 : */
8064 331983 : numIndexTuples = rint(numIndexTuples / num_sa_scans);
8065 : }
8066 :
8067 : /*
8068 : * Now do generic index cost estimation.
8069 : */
8070 470142 : costs.numIndexTuples = numIndexTuples;
8071 470142 : costs.num_sa_scans = num_sa_scans;
8072 :
8073 470142 : genericcostestimate(root, path, loop_count, &costs);
8074 :
8075 : /*
8076 : * Add a CPU-cost component to represent the costs of initial btree
8077 : * descent. We don't charge any I/O cost for touching upper btree levels,
8078 : * since they tend to stay in cache, but we still have to do about log2(N)
8079 : * comparisons to descend a btree of N leaf tuples. We charge one
8080 : * cpu_operator_cost per comparison.
8081 : *
8082 : * If there are SAOP or skip array keys, charge this once per estimated
8083 : * index descent. The ones after the first one are not startup cost so
8084 : * far as the overall plan goes, so just add them to "total" cost.
8085 : */
8086 470142 : if (index->tuples > 1) /* avoid computing log(0) */
8087 : {
8088 425968 : descentCost = ceil(log(index->tuples) / log(2.0)) * cpu_operator_cost;
8089 425968 : costs.indexStartupCost += descentCost;
8090 425968 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8091 : }
8092 :
8093 : /*
8094 : * Even though we're not charging I/O cost for touching upper btree pages,
8095 : * it's still reasonable to charge some CPU cost per page descended
8096 : * through. Moreover, if we had no such charge at all, bloated indexes
8097 : * would appear to have the same search cost as unbloated ones, at least
8098 : * in cases where only a single leaf page is expected to be visited. This
8099 : * cost is somewhat arbitrarily set at 50x cpu_operator_cost per page
8100 : * touched. The number of such pages is btree tree height plus one (ie,
8101 : * we charge for the leaf page too). As above, charge once per estimated
8102 : * SAOP/skip array descent.
8103 : */
8104 470142 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8105 470142 : costs.indexStartupCost += descentCost;
8106 470142 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8107 :
8108 470142 : if (!have_correlation)
8109 : {
8110 448384 : examine_indexcol_variable(root, index, 0, &vardata);
8111 448384 : if (HeapTupleIsValid(vardata.statsTuple))
8112 305087 : costs.indexCorrelation = btcost_correlation(index, &vardata);
8113 448384 : ReleaseVariableStats(vardata);
8114 : }
8115 : else
8116 : {
8117 : /* btcost_correlation already called earlier on */
8118 21758 : costs.indexCorrelation = correlation;
8119 : }
8120 :
8121 470142 : *indexStartupCost = costs.indexStartupCost;
8122 470142 : *indexTotalCost = costs.indexTotalCost;
8123 470142 : *indexSelectivity = costs.indexSelectivity;
8124 470142 : *indexCorrelation = costs.indexCorrelation;
8125 470142 : *indexPages = costs.numIndexPages;
8126 470142 : }
8127 :
8128 : void
8129 215 : hashcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8130 : Cost *indexStartupCost, Cost *indexTotalCost,
8131 : Selectivity *indexSelectivity, double *indexCorrelation,
8132 : double *indexPages)
8133 : {
8134 215 : GenericCosts costs = {0};
8135 :
8136 215 : genericcostestimate(root, path, loop_count, &costs);
8137 :
8138 : /*
8139 : * A hash index has no descent costs as such, since the index AM can go
8140 : * directly to the target bucket after computing the hash value. There
8141 : * are a couple of other hash-specific costs that we could conceivably add
8142 : * here, though:
8143 : *
8144 : * Ideally we'd charge spc_random_page_cost for each page in the target
8145 : * bucket, not just the numIndexPages pages that genericcostestimate
8146 : * thought we'd visit. However in most cases we don't know which bucket
8147 : * that will be. There's no point in considering the average bucket size
8148 : * because the hash AM makes sure that's always one page.
8149 : *
8150 : * Likewise, we could consider charging some CPU for each index tuple in
8151 : * the bucket, if we knew how many there were. But the per-tuple cost is
8152 : * just a hash value comparison, not a general datatype-dependent
8153 : * comparison, so any such charge ought to be quite a bit less than
8154 : * cpu_operator_cost; which makes it probably not worth worrying about.
8155 : *
8156 : * A bigger issue is that chance hash-value collisions will result in
8157 : * wasted probes into the heap. We don't currently attempt to model this
8158 : * cost on the grounds that it's rare, but maybe it's not rare enough.
8159 : * (Any fix for this ought to consider the generic lossy-operator problem,
8160 : * though; it's not entirely hash-specific.)
8161 : */
8162 :
8163 215 : *indexStartupCost = costs.indexStartupCost;
8164 215 : *indexTotalCost = costs.indexTotalCost;
8165 215 : *indexSelectivity = costs.indexSelectivity;
8166 215 : *indexCorrelation = costs.indexCorrelation;
8167 215 : *indexPages = costs.numIndexPages;
8168 215 : }
8169 :
8170 : void
8171 2447 : gistcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8172 : Cost *indexStartupCost, Cost *indexTotalCost,
8173 : Selectivity *indexSelectivity, double *indexCorrelation,
8174 : double *indexPages)
8175 : {
8176 2447 : IndexOptInfo *index = path->indexinfo;
8177 2447 : GenericCosts costs = {0};
8178 : Cost descentCost;
8179 :
8180 2447 : genericcostestimate(root, path, loop_count, &costs);
8181 :
8182 : /*
8183 : * We model index descent costs similarly to those for btree, but to do
8184 : * that we first need an idea of the tree height. We somewhat arbitrarily
8185 : * assume that the fanout is 100, meaning the tree height is at most
8186 : * log100(index->pages).
8187 : *
8188 : * Although this computation isn't really expensive enough to require
8189 : * caching, we might as well use index->tree_height to cache it.
8190 : */
8191 2447 : if (index->tree_height < 0) /* unknown? */
8192 : {
8193 2440 : if (index->pages > 1) /* avoid computing log(0) */
8194 1360 : index->tree_height = (int) (log(index->pages) / log(100.0));
8195 : else
8196 1080 : index->tree_height = 0;
8197 : }
8198 :
8199 : /*
8200 : * Add a CPU-cost component to represent the costs of initial descent. We
8201 : * just use log(N) here not log2(N) since the branching factor isn't
8202 : * necessarily two anyway. As for btree, charge once per SA scan.
8203 : */
8204 2447 : if (index->tuples > 1) /* avoid computing log(0) */
8205 : {
8206 2447 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8207 2447 : costs.indexStartupCost += descentCost;
8208 2447 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8209 : }
8210 :
8211 : /*
8212 : * Likewise add a per-page charge, calculated the same as for btrees.
8213 : */
8214 2447 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8215 2447 : costs.indexStartupCost += descentCost;
8216 2447 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8217 :
8218 2447 : *indexStartupCost = costs.indexStartupCost;
8219 2447 : *indexTotalCost = costs.indexTotalCost;
8220 2447 : *indexSelectivity = costs.indexSelectivity;
8221 2447 : *indexCorrelation = costs.indexCorrelation;
8222 2447 : *indexPages = costs.numIndexPages;
8223 2447 : }
8224 :
8225 : void
8226 892 : spgcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8227 : Cost *indexStartupCost, Cost *indexTotalCost,
8228 : Selectivity *indexSelectivity, double *indexCorrelation,
8229 : double *indexPages)
8230 : {
8231 892 : IndexOptInfo *index = path->indexinfo;
8232 892 : GenericCosts costs = {0};
8233 : Cost descentCost;
8234 :
8235 892 : genericcostestimate(root, path, loop_count, &costs);
8236 :
8237 : /*
8238 : * We model index descent costs similarly to those for btree, but to do
8239 : * that we first need an idea of the tree height. We somewhat arbitrarily
8240 : * assume that the fanout is 100, meaning the tree height is at most
8241 : * log100(index->pages).
8242 : *
8243 : * Although this computation isn't really expensive enough to require
8244 : * caching, we might as well use index->tree_height to cache it.
8245 : */
8246 892 : if (index->tree_height < 0) /* unknown? */
8247 : {
8248 889 : if (index->pages > 1) /* avoid computing log(0) */
8249 889 : index->tree_height = (int) (log(index->pages) / log(100.0));
8250 : else
8251 0 : index->tree_height = 0;
8252 : }
8253 :
8254 : /*
8255 : * Add a CPU-cost component to represent the costs of initial descent. We
8256 : * just use log(N) here not log2(N) since the branching factor isn't
8257 : * necessarily two anyway. As for btree, charge once per SA scan.
8258 : */
8259 892 : if (index->tuples > 1) /* avoid computing log(0) */
8260 : {
8261 892 : descentCost = ceil(log(index->tuples)) * cpu_operator_cost;
8262 892 : costs.indexStartupCost += descentCost;
8263 892 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8264 : }
8265 :
8266 : /*
8267 : * Likewise add a per-page charge, calculated the same as for btrees.
8268 : */
8269 892 : descentCost = (index->tree_height + 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8270 892 : costs.indexStartupCost += descentCost;
8271 892 : costs.indexTotalCost += costs.num_sa_scans * descentCost;
8272 :
8273 892 : *indexStartupCost = costs.indexStartupCost;
8274 892 : *indexTotalCost = costs.indexTotalCost;
8275 892 : *indexSelectivity = costs.indexSelectivity;
8276 892 : *indexCorrelation = costs.indexCorrelation;
8277 892 : *indexPages = costs.numIndexPages;
8278 892 : }
8279 :
8280 :
8281 : /*
8282 : * Support routines for gincostestimate
8283 : */
8284 :
8285 : typedef struct
8286 : {
8287 : bool attHasFullScan[INDEX_MAX_KEYS];
8288 : bool attHasNormalScan[INDEX_MAX_KEYS];
8289 : double partialEntries;
8290 : double exactEntries;
8291 : double searchEntries;
8292 : double arrayScans;
8293 : } GinQualCounts;
8294 :
8295 : /*
8296 : * Estimate the number of index terms that need to be searched for while
8297 : * testing the given GIN query, and increment the counts in *counts
8298 : * appropriately. If the query is unsatisfiable, return false.
8299 : */
8300 : static bool
8301 1240 : gincost_pattern(IndexOptInfo *index, int indexcol,
8302 : Oid clause_op, Datum query,
8303 : GinQualCounts *counts)
8304 : {
8305 : FmgrInfo flinfo;
8306 : Oid extractProcOid;
8307 : Oid collation;
8308 : int strategy_op;
8309 : Oid lefttype,
8310 : righttype;
8311 1240 : int32 nentries = 0;
8312 1240 : bool *partial_matches = NULL;
8313 1240 : Pointer *extra_data = NULL;
8314 1240 : bool *nullFlags = NULL;
8315 1240 : int32 searchMode = GIN_SEARCH_MODE_DEFAULT;
8316 : int32 i;
8317 :
8318 : Assert(indexcol < index->nkeycolumns);
8319 :
8320 : /*
8321 : * Get the operator's strategy number and declared input data types within
8322 : * the index opfamily. (We don't need the latter, but we use
8323 : * get_op_opfamily_properties because it will throw error if it fails to
8324 : * find a matching pg_amop entry.)
8325 : */
8326 1240 : get_op_opfamily_properties(clause_op, index->opfamily[indexcol], false,
8327 : &strategy_op, &lefttype, &righttype);
8328 :
8329 : /*
8330 : * GIN always uses the "default" support functions, which are those with
8331 : * lefttype == righttype == the opclass' opcintype (see
8332 : * IndexSupportInitialize in relcache.c).
8333 : */
8334 1240 : extractProcOid = get_opfamily_proc(index->opfamily[indexcol],
8335 1240 : index->opcintype[indexcol],
8336 1240 : index->opcintype[indexcol],
8337 : GIN_EXTRACTQUERY_PROC);
8338 :
8339 1240 : if (!OidIsValid(extractProcOid))
8340 : {
8341 : /* should not happen; throw same error as index_getprocinfo */
8342 0 : elog(ERROR, "missing support function %d for attribute %d of index \"%s\"",
8343 : GIN_EXTRACTQUERY_PROC, indexcol + 1,
8344 : get_rel_name(index->indexoid));
8345 : }
8346 :
8347 : /*
8348 : * Choose collation to pass to extractProc (should match initGinState).
8349 : */
8350 1240 : if (OidIsValid(index->indexcollations[indexcol]))
8351 207 : collation = index->indexcollations[indexcol];
8352 : else
8353 1033 : collation = DEFAULT_COLLATION_OID;
8354 :
8355 1240 : fmgr_info(extractProcOid, &flinfo);
8356 :
8357 1240 : set_fn_opclass_options(&flinfo, index->opclassoptions[indexcol]);
8358 :
8359 1240 : FunctionCall7Coll(&flinfo,
8360 : collation,
8361 : query,
8362 : PointerGetDatum(&nentries),
8363 : UInt16GetDatum(strategy_op),
8364 : PointerGetDatum(&partial_matches),
8365 : PointerGetDatum(&extra_data),
8366 : PointerGetDatum(&nullFlags),
8367 : PointerGetDatum(&searchMode));
8368 :
8369 1240 : if (nentries <= 0 && searchMode == GIN_SEARCH_MODE_DEFAULT)
8370 : {
8371 : /* No match is possible */
8372 6 : return false;
8373 : }
8374 :
8375 4838 : for (i = 0; i < nentries; i++)
8376 : {
8377 : /*
8378 : * For partial match we haven't any information to estimate number of
8379 : * matched entries in index, so, we just estimate it as 100
8380 : */
8381 3604 : if (partial_matches && partial_matches[i])
8382 347 : counts->partialEntries += 100;
8383 : else
8384 3257 : counts->exactEntries++;
8385 :
8386 3604 : counts->searchEntries++;
8387 : }
8388 :
8389 1234 : if (searchMode == GIN_SEARCH_MODE_DEFAULT)
8390 : {
8391 992 : counts->attHasNormalScan[indexcol] = true;
8392 : }
8393 242 : else if (searchMode == GIN_SEARCH_MODE_INCLUDE_EMPTY)
8394 : {
8395 : /* Treat "include empty" like an exact-match item */
8396 22 : counts->attHasNormalScan[indexcol] = true;
8397 22 : counts->exactEntries++;
8398 22 : counts->searchEntries++;
8399 : }
8400 : else
8401 : {
8402 : /* It's GIN_SEARCH_MODE_ALL */
8403 220 : counts->attHasFullScan[indexcol] = true;
8404 : }
8405 :
8406 1234 : return true;
8407 : }
8408 :
8409 : /*
8410 : * Estimate the number of index terms that need to be searched for while
8411 : * testing the given GIN index clause, and increment the counts in *counts
8412 : * appropriately. If the query is unsatisfiable, return false.
8413 : */
8414 : static bool
8415 1234 : gincost_opexpr(PlannerInfo *root,
8416 : IndexOptInfo *index,
8417 : int indexcol,
8418 : OpExpr *clause,
8419 : GinQualCounts *counts)
8420 : {
8421 1234 : Oid clause_op = clause->opno;
8422 1234 : Node *operand = (Node *) lsecond(clause->args);
8423 :
8424 : /* aggressively reduce to a constant, and look through relabeling */
8425 1234 : operand = estimate_expression_value(root, operand);
8426 :
8427 1234 : if (IsA(operand, RelabelType))
8428 0 : operand = (Node *) ((RelabelType *) operand)->arg;
8429 :
8430 : /*
8431 : * It's impossible to call extractQuery method for unknown operand. So
8432 : * unless operand is a Const we can't do much; just assume there will be
8433 : * one ordinary search entry from the operand at runtime.
8434 : */
8435 1234 : if (!IsA(operand, Const))
8436 : {
8437 0 : counts->exactEntries++;
8438 0 : counts->searchEntries++;
8439 0 : return true;
8440 : }
8441 :
8442 : /* If Const is null, there can be no matches */
8443 1234 : if (((Const *) operand)->constisnull)
8444 0 : return false;
8445 :
8446 : /* Otherwise, apply extractQuery and get the actual term counts */
8447 1234 : return gincost_pattern(index, indexcol, clause_op,
8448 : ((Const *) operand)->constvalue,
8449 : counts);
8450 : }
8451 :
8452 : /*
8453 : * Estimate the number of index terms that need to be searched for while
8454 : * testing the given GIN index clause, and increment the counts in *counts
8455 : * appropriately. If the query is unsatisfiable, return false.
8456 : *
8457 : * A ScalarArrayOpExpr will give rise to N separate indexscans at runtime,
8458 : * each of which involves one value from the RHS array, plus all the
8459 : * non-array quals (if any). To model this, we average the counts across
8460 : * the RHS elements, and add the averages to the counts in *counts (which
8461 : * correspond to per-indexscan costs). We also multiply counts->arrayScans
8462 : * by N, causing gincostestimate to scale up its estimates accordingly.
8463 : */
8464 : static bool
8465 3 : gincost_scalararrayopexpr(PlannerInfo *root,
8466 : IndexOptInfo *index,
8467 : int indexcol,
8468 : ScalarArrayOpExpr *clause,
8469 : double numIndexEntries,
8470 : GinQualCounts *counts)
8471 : {
8472 3 : Oid clause_op = clause->opno;
8473 3 : Node *rightop = (Node *) lsecond(clause->args);
8474 : ArrayType *arrayval;
8475 : int16 elmlen;
8476 : bool elmbyval;
8477 : char elmalign;
8478 : int numElems;
8479 : Datum *elemValues;
8480 : bool *elemNulls;
8481 : GinQualCounts arraycounts;
8482 3 : int numPossible = 0;
8483 : int i;
8484 :
8485 : Assert(clause->useOr);
8486 :
8487 : /* aggressively reduce to a constant, and look through relabeling */
8488 3 : rightop = estimate_expression_value(root, rightop);
8489 :
8490 3 : if (IsA(rightop, RelabelType))
8491 0 : rightop = (Node *) ((RelabelType *) rightop)->arg;
8492 :
8493 : /*
8494 : * It's impossible to call extractQuery method for unknown operand. So
8495 : * unless operand is a Const we can't do much; just assume there will be
8496 : * one ordinary search entry from each array entry at runtime, and fall
8497 : * back on a probably-bad estimate of the number of array entries.
8498 : */
8499 3 : if (!IsA(rightop, Const))
8500 : {
8501 0 : counts->exactEntries++;
8502 0 : counts->searchEntries++;
8503 0 : counts->arrayScans *= estimate_array_length(root, rightop);
8504 0 : return true;
8505 : }
8506 :
8507 : /* If Const is null, there can be no matches */
8508 3 : if (((Const *) rightop)->constisnull)
8509 0 : return false;
8510 :
8511 : /* Otherwise, extract the array elements and iterate over them */
8512 3 : arrayval = DatumGetArrayTypeP(((Const *) rightop)->constvalue);
8513 3 : get_typlenbyvalalign(ARR_ELEMTYPE(arrayval),
8514 : &elmlen, &elmbyval, &elmalign);
8515 3 : deconstruct_array(arrayval,
8516 : ARR_ELEMTYPE(arrayval),
8517 : elmlen, elmbyval, elmalign,
8518 : &elemValues, &elemNulls, &numElems);
8519 :
8520 3 : memset(&arraycounts, 0, sizeof(arraycounts));
8521 :
8522 9 : for (i = 0; i < numElems; i++)
8523 : {
8524 : GinQualCounts elemcounts;
8525 :
8526 : /* NULL can't match anything, so ignore, as the executor will */
8527 6 : if (elemNulls[i])
8528 0 : continue;
8529 :
8530 : /* Otherwise, apply extractQuery and get the actual term counts */
8531 6 : memset(&elemcounts, 0, sizeof(elemcounts));
8532 :
8533 6 : if (gincost_pattern(index, indexcol, clause_op, elemValues[i],
8534 : &elemcounts))
8535 : {
8536 : /* We ignore array elements that are unsatisfiable patterns */
8537 6 : numPossible++;
8538 :
8539 6 : if (elemcounts.attHasFullScan[indexcol] &&
8540 0 : !elemcounts.attHasNormalScan[indexcol])
8541 : {
8542 : /*
8543 : * Full index scan will be required. We treat this as if
8544 : * every key in the index had been listed in the query; is
8545 : * that reasonable?
8546 : */
8547 0 : elemcounts.partialEntries = 0;
8548 0 : elemcounts.exactEntries = numIndexEntries;
8549 0 : elemcounts.searchEntries = numIndexEntries;
8550 : }
8551 6 : arraycounts.partialEntries += elemcounts.partialEntries;
8552 6 : arraycounts.exactEntries += elemcounts.exactEntries;
8553 6 : arraycounts.searchEntries += elemcounts.searchEntries;
8554 : }
8555 : }
8556 :
8557 3 : if (numPossible == 0)
8558 : {
8559 : /* No satisfiable patterns in the array */
8560 0 : return false;
8561 : }
8562 :
8563 : /*
8564 : * Now add the averages to the global counts. This will give us an
8565 : * estimate of the average number of terms searched for in each indexscan,
8566 : * including contributions from both array and non-array quals.
8567 : */
8568 3 : counts->partialEntries += arraycounts.partialEntries / numPossible;
8569 3 : counts->exactEntries += arraycounts.exactEntries / numPossible;
8570 3 : counts->searchEntries += arraycounts.searchEntries / numPossible;
8571 :
8572 3 : counts->arrayScans *= numPossible;
8573 :
8574 3 : return true;
8575 : }
8576 :
8577 : /*
8578 : * GIN has search behavior completely different from other index types
8579 : */
8580 : void
8581 1132 : gincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8582 : Cost *indexStartupCost, Cost *indexTotalCost,
8583 : Selectivity *indexSelectivity, double *indexCorrelation,
8584 : double *indexPages)
8585 : {
8586 1132 : IndexOptInfo *index = path->indexinfo;
8587 1132 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8588 : List *selectivityQuals;
8589 1132 : double numPages = index->pages,
8590 1132 : numTuples = index->tuples;
8591 : double numEntryPages,
8592 : numDataPages,
8593 : numPendingPages,
8594 : numEntries;
8595 : GinQualCounts counts;
8596 : bool matchPossible;
8597 : bool fullIndexScan;
8598 : double partialScale;
8599 : double entryPagesFetched,
8600 : dataPagesFetched,
8601 : dataPagesFetchedBySel;
8602 : double qual_op_cost,
8603 : qual_arg_cost,
8604 : spc_random_page_cost,
8605 : outer_scans;
8606 : Cost descentCost;
8607 : Relation indexRel;
8608 : GinStatsData ginStats;
8609 : ListCell *lc;
8610 : int i;
8611 :
8612 : /*
8613 : * Obtain statistical information from the meta page, if possible. Else
8614 : * set ginStats to zeroes, and we'll cope below.
8615 : */
8616 1132 : if (!index->hypothetical)
8617 : {
8618 : /* Lock should have already been obtained in plancat.c */
8619 1132 : indexRel = index_open(index->indexoid, NoLock);
8620 1132 : ginGetStats(indexRel, &ginStats);
8621 1132 : index_close(indexRel, NoLock);
8622 : }
8623 : else
8624 : {
8625 0 : memset(&ginStats, 0, sizeof(ginStats));
8626 : }
8627 :
8628 : /*
8629 : * Assuming we got valid (nonzero) stats at all, nPendingPages can be
8630 : * trusted, but the other fields are data as of the last VACUUM. We can
8631 : * scale them up to account for growth since then, but that method only
8632 : * goes so far; in the worst case, the stats might be for a completely
8633 : * empty index, and scaling them will produce pretty bogus numbers.
8634 : * Somewhat arbitrarily, set the cutoff for doing scaling at 4X growth; if
8635 : * it's grown more than that, fall back to estimating things only from the
8636 : * assumed-accurate index size. But we'll trust nPendingPages in any case
8637 : * so long as it's not clearly insane, ie, more than the index size.
8638 : */
8639 1132 : if (ginStats.nPendingPages < numPages)
8640 1132 : numPendingPages = ginStats.nPendingPages;
8641 : else
8642 0 : numPendingPages = 0;
8643 :
8644 1132 : if (numPages > 0 && ginStats.nTotalPages <= numPages &&
8645 1132 : ginStats.nTotalPages > numPages / 4 &&
8646 1106 : ginStats.nEntryPages > 0 && ginStats.nEntries > 0)
8647 974 : {
8648 : /*
8649 : * OK, the stats seem close enough to sane to be trusted. But we
8650 : * still need to scale them by the ratio numPages / nTotalPages to
8651 : * account for growth since the last VACUUM.
8652 : */
8653 974 : double scale = numPages / ginStats.nTotalPages;
8654 :
8655 974 : numEntryPages = ceil(ginStats.nEntryPages * scale);
8656 974 : numDataPages = ceil(ginStats.nDataPages * scale);
8657 974 : numEntries = ceil(ginStats.nEntries * scale);
8658 : /* ensure we didn't round up too much */
8659 974 : numEntryPages = Min(numEntryPages, numPages - numPendingPages);
8660 974 : numDataPages = Min(numDataPages,
8661 : numPages - numPendingPages - numEntryPages);
8662 : }
8663 : else
8664 : {
8665 : /*
8666 : * We might get here because it's a hypothetical index, or an index
8667 : * created pre-9.1 and never vacuumed since upgrading (in which case
8668 : * its stats would read as zeroes), or just because it's grown too
8669 : * much since the last VACUUM for us to put our faith in scaling.
8670 : *
8671 : * Invent some plausible internal statistics based on the index page
8672 : * count (and clamp that to at least 10 pages, just in case). We
8673 : * estimate that 90% of the index is entry pages, and the rest is data
8674 : * pages. Estimate 100 entries per entry page; this is rather bogus
8675 : * since it'll depend on the size of the keys, but it's more robust
8676 : * than trying to predict the number of entries per heap tuple.
8677 : */
8678 158 : numPages = Max(numPages, 10);
8679 158 : numEntryPages = floor((numPages - numPendingPages) * 0.90);
8680 158 : numDataPages = numPages - numPendingPages - numEntryPages;
8681 158 : numEntries = floor(numEntryPages * 100);
8682 : }
8683 :
8684 : /* In an empty index, numEntries could be zero. Avoid divide-by-zero */
8685 1132 : if (numEntries < 1)
8686 0 : numEntries = 1;
8687 :
8688 : /*
8689 : * If the index is partial, AND the index predicate with the index-bound
8690 : * quals to produce a more accurate idea of the number of rows covered by
8691 : * the bound conditions.
8692 : */
8693 1132 : selectivityQuals = add_predicate_to_index_quals(index, indexQuals);
8694 :
8695 : /* Estimate the fraction of main-table tuples that will be visited */
8696 2264 : *indexSelectivity = clauselist_selectivity(root, selectivityQuals,
8697 1132 : index->rel->relid,
8698 : JOIN_INNER,
8699 : NULL);
8700 :
8701 : /* fetch estimated page cost for tablespace containing index */
8702 1132 : get_tablespace_page_costs(index->reltablespace,
8703 : &spc_random_page_cost,
8704 : NULL);
8705 :
8706 : /*
8707 : * Generic assumption about index correlation: there isn't any.
8708 : */
8709 1132 : *indexCorrelation = 0.0;
8710 :
8711 : /*
8712 : * Examine quals to estimate number of search entries & partial matches
8713 : */
8714 1132 : memset(&counts, 0, sizeof(counts));
8715 1132 : counts.arrayScans = 1;
8716 1132 : matchPossible = true;
8717 :
8718 2369 : foreach(lc, path->indexclauses)
8719 : {
8720 1237 : IndexClause *iclause = lfirst_node(IndexClause, lc);
8721 : ListCell *lc2;
8722 :
8723 2468 : foreach(lc2, iclause->indexquals)
8724 : {
8725 1237 : RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc2);
8726 1237 : Expr *clause = rinfo->clause;
8727 :
8728 1237 : if (IsA(clause, OpExpr))
8729 : {
8730 1234 : matchPossible = gincost_opexpr(root,
8731 : index,
8732 1234 : iclause->indexcol,
8733 : (OpExpr *) clause,
8734 : &counts);
8735 1234 : if (!matchPossible)
8736 6 : break;
8737 : }
8738 3 : else if (IsA(clause, ScalarArrayOpExpr))
8739 : {
8740 3 : matchPossible = gincost_scalararrayopexpr(root,
8741 : index,
8742 3 : iclause->indexcol,
8743 : (ScalarArrayOpExpr *) clause,
8744 : numEntries,
8745 : &counts);
8746 3 : if (!matchPossible)
8747 0 : break;
8748 : }
8749 : else
8750 : {
8751 : /* shouldn't be anything else for a GIN index */
8752 0 : elog(ERROR, "unsupported GIN indexqual type: %d",
8753 : (int) nodeTag(clause));
8754 : }
8755 : }
8756 : }
8757 :
8758 : /* Fall out if there were any provably-unsatisfiable quals */
8759 1132 : if (!matchPossible)
8760 : {
8761 6 : *indexStartupCost = 0;
8762 6 : *indexTotalCost = 0;
8763 6 : *indexSelectivity = 0;
8764 6 : return;
8765 : }
8766 :
8767 : /*
8768 : * If attribute has a full scan and at the same time doesn't have normal
8769 : * scan, then we'll have to scan all non-null entries of that attribute.
8770 : * Currently, we don't have per-attribute statistics for GIN. Thus, we
8771 : * must assume the whole GIN index has to be scanned in this case.
8772 : */
8773 1126 : fullIndexScan = false;
8774 2197 : for (i = 0; i < index->nkeycolumns; i++)
8775 : {
8776 1240 : if (counts.attHasFullScan[i] && !counts.attHasNormalScan[i])
8777 : {
8778 169 : fullIndexScan = true;
8779 169 : break;
8780 : }
8781 : }
8782 :
8783 1126 : if (fullIndexScan || indexQuals == NIL)
8784 : {
8785 : /*
8786 : * Full index scan will be required. We treat this as if every key in
8787 : * the index had been listed in the query; is that reasonable?
8788 : */
8789 169 : counts.partialEntries = 0;
8790 169 : counts.exactEntries = numEntries;
8791 169 : counts.searchEntries = numEntries;
8792 : }
8793 :
8794 : /* Will we have more than one iteration of a nestloop scan? */
8795 1126 : outer_scans = loop_count;
8796 :
8797 : /*
8798 : * Compute cost to begin scan, first of all, pay attention to pending
8799 : * list.
8800 : */
8801 1126 : entryPagesFetched = numPendingPages;
8802 :
8803 : /*
8804 : * Estimate number of entry pages read. We need to do
8805 : * counts.searchEntries searches. Use a power function as it should be,
8806 : * but tuples on leaf pages usually is much greater. Here we include all
8807 : * searches in entry tree, including search of first entry in partial
8808 : * match algorithm
8809 : */
8810 1126 : entryPagesFetched += ceil(counts.searchEntries * rint(pow(numEntryPages, 0.15)));
8811 :
8812 : /*
8813 : * Add an estimate of entry pages read by partial match algorithm. It's a
8814 : * scan over leaf pages in entry tree. We haven't any useful stats here,
8815 : * so estimate it as proportion. Because counts.partialEntries is really
8816 : * pretty bogus (see code above), it's possible that it is more than
8817 : * numEntries; clamp the proportion to ensure sanity.
8818 : */
8819 1126 : partialScale = counts.partialEntries / numEntries;
8820 1126 : partialScale = Min(partialScale, 1.0);
8821 :
8822 1126 : entryPagesFetched += ceil(numEntryPages * partialScale);
8823 :
8824 : /*
8825 : * Partial match algorithm reads all data pages before doing actual scan,
8826 : * so it's a startup cost. Again, we haven't any useful stats here, so
8827 : * estimate it as proportion.
8828 : */
8829 1126 : dataPagesFetched = ceil(numDataPages * partialScale);
8830 :
8831 1126 : *indexStartupCost = 0;
8832 1126 : *indexTotalCost = 0;
8833 :
8834 : /*
8835 : * Add a CPU-cost component to represent the costs of initial entry btree
8836 : * descent. We don't charge any I/O cost for touching upper btree levels,
8837 : * since they tend to stay in cache, but we still have to do about log2(N)
8838 : * comparisons to descend a btree of N leaf tuples. We charge one
8839 : * cpu_operator_cost per comparison.
8840 : *
8841 : * If there are ScalarArrayOpExprs, charge this once per SA scan. The
8842 : * ones after the first one are not startup cost so far as the overall
8843 : * plan is concerned, so add them only to "total" cost.
8844 : */
8845 1126 : if (numEntries > 1) /* avoid computing log(0) */
8846 : {
8847 1126 : descentCost = ceil(log(numEntries) / log(2.0)) * cpu_operator_cost;
8848 1126 : *indexStartupCost += descentCost * counts.searchEntries;
8849 1126 : *indexTotalCost += counts.arrayScans * descentCost * counts.searchEntries;
8850 : }
8851 :
8852 : /*
8853 : * Add a cpu cost per entry-page fetched. This is not amortized over a
8854 : * loop.
8855 : */
8856 1126 : *indexStartupCost += entryPagesFetched * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8857 1126 : *indexTotalCost += entryPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8858 :
8859 : /*
8860 : * Add a cpu cost per data-page fetched. This is also not amortized over a
8861 : * loop. Since those are the data pages from the partial match algorithm,
8862 : * charge them as startup cost.
8863 : */
8864 1126 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * dataPagesFetched;
8865 :
8866 : /*
8867 : * Since we add the startup cost to the total cost later on, remove the
8868 : * initial arrayscan from the total.
8869 : */
8870 1126 : *indexTotalCost += dataPagesFetched * (counts.arrayScans - 1) * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8871 :
8872 : /*
8873 : * Calculate cache effects if more than one scan due to nestloops or array
8874 : * quals. The result is pro-rated per nestloop scan, but the array qual
8875 : * factor shouldn't be pro-rated (compare genericcostestimate).
8876 : */
8877 1126 : if (outer_scans > 1 || counts.arrayScans > 1)
8878 : {
8879 3 : entryPagesFetched *= outer_scans * counts.arrayScans;
8880 3 : entryPagesFetched = index_pages_fetched(entryPagesFetched,
8881 : (BlockNumber) numEntryPages,
8882 : numEntryPages, root);
8883 3 : entryPagesFetched /= outer_scans;
8884 3 : dataPagesFetched *= outer_scans * counts.arrayScans;
8885 3 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8886 : (BlockNumber) numDataPages,
8887 : numDataPages, root);
8888 3 : dataPagesFetched /= outer_scans;
8889 : }
8890 :
8891 : /*
8892 : * Here we use random page cost because logically-close pages could be far
8893 : * apart on disk.
8894 : */
8895 1126 : *indexStartupCost += (entryPagesFetched + dataPagesFetched) * spc_random_page_cost;
8896 :
8897 : /*
8898 : * Now compute the number of data pages fetched during the scan.
8899 : *
8900 : * We assume every entry to have the same number of items, and that there
8901 : * is no overlap between them. (XXX: tsvector and array opclasses collect
8902 : * statistics on the frequency of individual keys; it would be nice to use
8903 : * those here.)
8904 : */
8905 1126 : dataPagesFetched = ceil(numDataPages * counts.exactEntries / numEntries);
8906 :
8907 : /*
8908 : * If there is a lot of overlap among the entries, in particular if one of
8909 : * the entries is very frequent, the above calculation can grossly
8910 : * under-estimate. As a simple cross-check, calculate a lower bound based
8911 : * on the overall selectivity of the quals. At a minimum, we must read
8912 : * one item pointer for each matching entry.
8913 : *
8914 : * The width of each item pointer varies, based on the level of
8915 : * compression. We don't have statistics on that, but an average of
8916 : * around 3 bytes per item is fairly typical.
8917 : */
8918 1126 : dataPagesFetchedBySel = ceil(*indexSelectivity *
8919 1126 : (numTuples / (BLCKSZ / 3)));
8920 1126 : if (dataPagesFetchedBySel > dataPagesFetched)
8921 933 : dataPagesFetched = dataPagesFetchedBySel;
8922 :
8923 : /* Add one page cpu-cost to the startup cost */
8924 1126 : *indexStartupCost += DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost * counts.searchEntries;
8925 :
8926 : /*
8927 : * Add once again a CPU-cost for those data pages, before amortizing for
8928 : * cache.
8929 : */
8930 1126 : *indexTotalCost += dataPagesFetched * counts.arrayScans * DEFAULT_PAGE_CPU_MULTIPLIER * cpu_operator_cost;
8931 :
8932 : /* Account for cache effects, the same as above */
8933 1126 : if (outer_scans > 1 || counts.arrayScans > 1)
8934 : {
8935 3 : dataPagesFetched *= outer_scans * counts.arrayScans;
8936 3 : dataPagesFetched = index_pages_fetched(dataPagesFetched,
8937 : (BlockNumber) numDataPages,
8938 : numDataPages, root);
8939 3 : dataPagesFetched /= outer_scans;
8940 : }
8941 :
8942 : /* And apply random_page_cost as the cost per page */
8943 1126 : *indexTotalCost += *indexStartupCost +
8944 1126 : dataPagesFetched * spc_random_page_cost;
8945 :
8946 : /*
8947 : * Add on index qual eval costs, much as in genericcostestimate. We charge
8948 : * cpu but we can disregard indexorderbys, since GIN doesn't support
8949 : * those.
8950 : */
8951 1126 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
8952 1126 : qual_op_cost = cpu_operator_cost * list_length(indexQuals);
8953 :
8954 1126 : *indexStartupCost += qual_arg_cost;
8955 1126 : *indexTotalCost += qual_arg_cost;
8956 :
8957 : /*
8958 : * Add a cpu cost per search entry, corresponding to the actual visited
8959 : * entries.
8960 : */
8961 1126 : *indexTotalCost += (counts.searchEntries * counts.arrayScans) * (qual_op_cost);
8962 : /* Now add a cpu cost per tuple in the posting lists / trees */
8963 1126 : *indexTotalCost += (numTuples * *indexSelectivity) * (cpu_index_tuple_cost);
8964 1126 : *indexPages = dataPagesFetched;
8965 : }
8966 :
8967 : /*
8968 : * BRIN has search behavior completely different from other index types
8969 : */
8970 : void
8971 5365 : brincostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
8972 : Cost *indexStartupCost, Cost *indexTotalCost,
8973 : Selectivity *indexSelectivity, double *indexCorrelation,
8974 : double *indexPages)
8975 : {
8976 5365 : IndexOptInfo *index = path->indexinfo;
8977 5365 : List *indexQuals = get_quals_from_indexclauses(path->indexclauses);
8978 5365 : double numPages = index->pages;
8979 5365 : RelOptInfo *baserel = index->rel;
8980 5365 : RangeTblEntry *rte = planner_rt_fetch(baserel->relid, root);
8981 : Cost spc_seq_page_cost;
8982 : Cost spc_random_page_cost;
8983 : double qual_arg_cost;
8984 : double qualSelectivity;
8985 : BrinStatsData statsData;
8986 : double indexRanges;
8987 : double minimalRanges;
8988 : double estimatedRanges;
8989 : double selec;
8990 : Relation indexRel;
8991 : ListCell *l;
8992 : VariableStatData vardata;
8993 :
8994 : Assert(rte->rtekind == RTE_RELATION);
8995 :
8996 : /* fetch estimated page cost for the tablespace containing the index */
8997 5365 : get_tablespace_page_costs(index->reltablespace,
8998 : &spc_random_page_cost,
8999 : &spc_seq_page_cost);
9000 :
9001 : /*
9002 : * Obtain some data from the index itself, if possible. Otherwise invent
9003 : * some plausible internal statistics based on the relation page count.
9004 : */
9005 5365 : if (!index->hypothetical)
9006 : {
9007 : /*
9008 : * A lock should have already been obtained on the index in plancat.c.
9009 : */
9010 5365 : indexRel = index_open(index->indexoid, NoLock);
9011 5365 : brinGetStats(indexRel, &statsData);
9012 5365 : index_close(indexRel, NoLock);
9013 :
9014 : /* work out the actual number of ranges in the index */
9015 5365 : indexRanges = Max(ceil((double) baserel->pages /
9016 : statsData.pagesPerRange), 1.0);
9017 : }
9018 : else
9019 : {
9020 : /*
9021 : * Assume default number of pages per range, and estimate the number
9022 : * of ranges based on that.
9023 : */
9024 0 : indexRanges = Max(ceil((double) baserel->pages /
9025 : BRIN_DEFAULT_PAGES_PER_RANGE), 1.0);
9026 :
9027 0 : statsData.pagesPerRange = BRIN_DEFAULT_PAGES_PER_RANGE;
9028 0 : statsData.revmapNumPages = (indexRanges / REVMAP_PAGE_MAXITEMS) + 1;
9029 : }
9030 :
9031 : /*
9032 : * Compute index correlation
9033 : *
9034 : * Because we can use all index quals equally when scanning, we can use
9035 : * the largest correlation (in absolute value) among columns used by the
9036 : * query. Start at zero, the worst possible case. If we cannot find any
9037 : * correlation statistics, we will keep it as 0.
9038 : */
9039 5365 : *indexCorrelation = 0;
9040 :
9041 10731 : foreach(l, path->indexclauses)
9042 : {
9043 5366 : IndexClause *iclause = lfirst_node(IndexClause, l);
9044 5366 : AttrNumber attnum = index->indexkeys[iclause->indexcol];
9045 :
9046 : /* attempt to lookup stats in relation for this index column */
9047 5366 : if (attnum != 0)
9048 : {
9049 : /* Simple variable -- look to stats for the underlying table */
9050 5366 : if (get_relation_stats_hook &&
9051 0 : (*get_relation_stats_hook) (root, rte, attnum, &vardata))
9052 : {
9053 : /*
9054 : * The hook took control of acquiring a stats tuple. If it
9055 : * did supply a tuple, it'd better have supplied a freefunc.
9056 : */
9057 0 : if (HeapTupleIsValid(vardata.statsTuple) && !vardata.freefunc)
9058 0 : elog(ERROR,
9059 : "no function provided to release variable stats with");
9060 : }
9061 : else
9062 : {
9063 5366 : vardata.statsTuple =
9064 5366 : SearchSysCache3(STATRELATTINH,
9065 : ObjectIdGetDatum(rte->relid),
9066 : Int16GetDatum(attnum),
9067 : BoolGetDatum(false));
9068 5366 : vardata.freefunc = ReleaseSysCache;
9069 : }
9070 : }
9071 : else
9072 : {
9073 : /*
9074 : * Looks like we've found an expression column in the index. Let's
9075 : * see if there's any stats for it.
9076 : */
9077 :
9078 : /* get the attnum from the 0-based index. */
9079 0 : attnum = iclause->indexcol + 1;
9080 :
9081 0 : if (get_index_stats_hook &&
9082 0 : (*get_index_stats_hook) (root, index->indexoid, attnum, &vardata))
9083 : {
9084 : /*
9085 : * The hook took control of acquiring a stats tuple. If it
9086 : * did supply a tuple, it'd better have supplied a freefunc.
9087 : */
9088 0 : if (HeapTupleIsValid(vardata.statsTuple) &&
9089 0 : !vardata.freefunc)
9090 0 : elog(ERROR, "no function provided to release variable stats with");
9091 : }
9092 : else
9093 : {
9094 0 : vardata.statsTuple = SearchSysCache3(STATRELATTINH,
9095 : ObjectIdGetDatum(index->indexoid),
9096 : Int16GetDatum(attnum),
9097 : BoolGetDatum(false));
9098 0 : vardata.freefunc = ReleaseSysCache;
9099 : }
9100 : }
9101 :
9102 5366 : if (HeapTupleIsValid(vardata.statsTuple))
9103 : {
9104 : AttStatsSlot sslot;
9105 :
9106 18 : if (get_attstatsslot(&sslot, vardata.statsTuple,
9107 : STATISTIC_KIND_CORRELATION, InvalidOid,
9108 : ATTSTATSSLOT_NUMBERS))
9109 : {
9110 18 : double varCorrelation = 0.0;
9111 :
9112 18 : if (sslot.nnumbers > 0)
9113 18 : varCorrelation = fabs(sslot.numbers[0]);
9114 :
9115 18 : if (varCorrelation > *indexCorrelation)
9116 18 : *indexCorrelation = varCorrelation;
9117 :
9118 18 : free_attstatsslot(&sslot);
9119 : }
9120 : }
9121 :
9122 5366 : ReleaseVariableStats(vardata);
9123 : }
9124 :
9125 5365 : qualSelectivity = clauselist_selectivity(root, indexQuals,
9126 5365 : baserel->relid,
9127 : JOIN_INNER, NULL);
9128 :
9129 : /*
9130 : * Now calculate the minimum possible ranges we could match with if all of
9131 : * the rows were in the perfect order in the table's heap.
9132 : */
9133 5365 : minimalRanges = ceil(indexRanges * qualSelectivity);
9134 :
9135 : /*
9136 : * Now estimate the number of ranges that we'll touch by using the
9137 : * indexCorrelation from the stats. Careful not to divide by zero (note
9138 : * we're using the absolute value of the correlation).
9139 : */
9140 5365 : if (*indexCorrelation < 1.0e-10)
9141 5347 : estimatedRanges = indexRanges;
9142 : else
9143 18 : estimatedRanges = Min(minimalRanges / *indexCorrelation, indexRanges);
9144 :
9145 : /* we expect to visit this portion of the table */
9146 5365 : selec = estimatedRanges / indexRanges;
9147 :
9148 5365 : CLAMP_PROBABILITY(selec);
9149 :
9150 5365 : *indexSelectivity = selec;
9151 :
9152 : /*
9153 : * Compute the index qual costs, much as in genericcostestimate, to add to
9154 : * the index costs. We can disregard indexorderbys, since BRIN doesn't
9155 : * support those.
9156 : */
9157 5365 : qual_arg_cost = index_other_operands_eval_cost(root, indexQuals);
9158 :
9159 : /*
9160 : * Compute the startup cost as the cost to read the whole revmap
9161 : * sequentially, including the cost to execute the index quals.
9162 : */
9163 5365 : *indexStartupCost =
9164 5365 : spc_seq_page_cost * statsData.revmapNumPages * loop_count;
9165 5365 : *indexStartupCost += qual_arg_cost;
9166 :
9167 : /*
9168 : * To read a BRIN index there might be a bit of back and forth over
9169 : * regular pages, as revmap might point to them out of sequential order;
9170 : * calculate the total cost as reading the whole index in random order.
9171 : */
9172 5365 : *indexTotalCost = *indexStartupCost +
9173 5365 : spc_random_page_cost * (numPages - statsData.revmapNumPages) * loop_count;
9174 :
9175 : /*
9176 : * Charge a small amount per range tuple which we expect to match to. This
9177 : * is meant to reflect the costs of manipulating the bitmap. The BRIN scan
9178 : * will set a bit for each page in the range when we find a matching
9179 : * range, so we must multiply the charge by the number of pages in the
9180 : * range.
9181 : */
9182 5365 : *indexTotalCost += 0.1 * cpu_operator_cost * estimatedRanges *
9183 5365 : statsData.pagesPerRange;
9184 :
9185 5365 : *indexPages = index->pages;
9186 5365 : }
|
